File: | lib/Target/X86/X86ISelLowering.cpp |
Warning: | line 27192, column 5 Value stored to 'AllowIntDomain' is never read |
1 | //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===// |
2 | // |
3 | // The LLVM Compiler Infrastructure |
4 | // |
5 | // This file is distributed under the University of Illinois Open Source |
6 | // License. See LICENSE.TXT for details. |
7 | // |
8 | //===----------------------------------------------------------------------===// |
9 | // |
10 | // This file defines the interfaces that X86 uses to lower LLVM code into a |
11 | // selection DAG. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "X86ISelLowering.h" |
16 | #include "Utils/X86ShuffleDecode.h" |
17 | #include "X86CallingConv.h" |
18 | #include "X86FrameLowering.h" |
19 | #include "X86InstrBuilder.h" |
20 | #include "X86IntrinsicsInfo.h" |
21 | #include "X86MachineFunctionInfo.h" |
22 | #include "X86ShuffleDecodeConstantPool.h" |
23 | #include "X86TargetMachine.h" |
24 | #include "X86TargetObjectFile.h" |
25 | #include "llvm/ADT/SmallBitVector.h" |
26 | #include "llvm/ADT/SmallSet.h" |
27 | #include "llvm/ADT/Statistic.h" |
28 | #include "llvm/ADT/StringExtras.h" |
29 | #include "llvm/ADT/StringSwitch.h" |
30 | #include "llvm/Analysis/EHPersonalities.h" |
31 | #include "llvm/CodeGen/IntrinsicLowering.h" |
32 | #include "llvm/CodeGen/MachineFrameInfo.h" |
33 | #include "llvm/CodeGen/MachineFunction.h" |
34 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
35 | #include "llvm/CodeGen/MachineJumpTableInfo.h" |
36 | #include "llvm/CodeGen/MachineModuleInfo.h" |
37 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
38 | #include "llvm/CodeGen/WinEHFuncInfo.h" |
39 | #include "llvm/IR/CallSite.h" |
40 | #include "llvm/IR/CallingConv.h" |
41 | #include "llvm/IR/Constants.h" |
42 | #include "llvm/IR/DerivedTypes.h" |
43 | #include "llvm/IR/DiagnosticInfo.h" |
44 | #include "llvm/IR/Function.h" |
45 | #include "llvm/IR/GlobalAlias.h" |
46 | #include "llvm/IR/GlobalVariable.h" |
47 | #include "llvm/IR/Instructions.h" |
48 | #include "llvm/IR/Intrinsics.h" |
49 | #include "llvm/MC/MCAsmInfo.h" |
50 | #include "llvm/MC/MCContext.h" |
51 | #include "llvm/MC/MCExpr.h" |
52 | #include "llvm/MC/MCSymbol.h" |
53 | #include "llvm/Support/CommandLine.h" |
54 | #include "llvm/Support/Debug.h" |
55 | #include "llvm/Support/ErrorHandling.h" |
56 | #include "llvm/Support/KnownBits.h" |
57 | #include "llvm/Support/MathExtras.h" |
58 | #include "llvm/Target/TargetLowering.h" |
59 | #include "llvm/Target/TargetOptions.h" |
60 | #include <algorithm> |
61 | #include <bitset> |
62 | #include <cctype> |
63 | #include <numeric> |
64 | using namespace llvm; |
65 | |
66 | #define DEBUG_TYPE"x86-isel" "x86-isel" |
67 | |
68 | STATISTIC(NumTailCalls, "Number of tail calls")static llvm::Statistic NumTailCalls = {"x86-isel", "NumTailCalls" , "Number of tail calls", {0}, false}; |
69 | |
70 | static cl::opt<bool> ExperimentalVectorWideningLegalization( |
71 | "x86-experimental-vector-widening-legalization", cl::init(false), |
72 | cl::desc("Enable an experimental vector type legalization through widening " |
73 | "rather than promotion."), |
74 | cl::Hidden); |
75 | |
76 | static cl::opt<int> ExperimentalPrefLoopAlignment( |
77 | "x86-experimental-pref-loop-alignment", cl::init(4), |
78 | cl::desc("Sets the preferable loop alignment for experiments " |
79 | "(the last x86-experimental-pref-loop-alignment bits" |
80 | " of the loop header PC will be 0)."), |
81 | cl::Hidden); |
82 | |
83 | static cl::opt<bool> MulConstantOptimization( |
84 | "mul-constant-optimization", cl::init(true), |
85 | cl::desc("Replace 'mul x, Const' with more effective instructions like " |
86 | "SHIFT, LEA, etc."), |
87 | cl::Hidden); |
88 | |
89 | /// Call this when the user attempts to do something unsupported, like |
90 | /// returning a double without SSE2 enabled on x86_64. This is not fatal, unlike |
91 | /// report_fatal_error, so calling code should attempt to recover without |
92 | /// crashing. |
93 | static void errorUnsupported(SelectionDAG &DAG, const SDLoc &dl, |
94 | const char *Msg) { |
95 | MachineFunction &MF = DAG.getMachineFunction(); |
96 | DAG.getContext()->diagnose( |
97 | DiagnosticInfoUnsupported(*MF.getFunction(), Msg, dl.getDebugLoc())); |
98 | } |
99 | |
100 | X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM, |
101 | const X86Subtarget &STI) |
102 | : TargetLowering(TM), Subtarget(STI) { |
103 | bool UseX87 = !Subtarget.useSoftFloat() && Subtarget.hasX87(); |
104 | X86ScalarSSEf64 = Subtarget.hasSSE2(); |
105 | X86ScalarSSEf32 = Subtarget.hasSSE1(); |
106 | MVT PtrVT = MVT::getIntegerVT(8 * TM.getPointerSize()); |
107 | |
108 | // Set up the TargetLowering object. |
109 | |
110 | // X86 is weird. It always uses i8 for shift amounts and setcc results. |
111 | setBooleanContents(ZeroOrOneBooleanContent); |
112 | // X86-SSE is even stranger. It uses -1 or 0 for vector masks. |
113 | setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); |
114 | |
115 | // For 64-bit, since we have so many registers, use the ILP scheduler. |
116 | // For 32-bit, use the register pressure specific scheduling. |
117 | // For Atom, always use ILP scheduling. |
118 | if (Subtarget.isAtom()) |
119 | setSchedulingPreference(Sched::ILP); |
120 | else if (Subtarget.is64Bit()) |
121 | setSchedulingPreference(Sched::ILP); |
122 | else |
123 | setSchedulingPreference(Sched::RegPressure); |
124 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
125 | setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister()); |
126 | |
127 | // Bypass expensive divides and use cheaper ones. |
128 | if (TM.getOptLevel() >= CodeGenOpt::Default) { |
129 | if (Subtarget.hasSlowDivide32()) |
130 | addBypassSlowDiv(32, 8); |
131 | if (Subtarget.hasSlowDivide64() && Subtarget.is64Bit()) |
132 | addBypassSlowDiv(64, 32); |
133 | } |
134 | |
135 | if (Subtarget.isTargetKnownWindowsMSVC() || |
136 | Subtarget.isTargetWindowsItanium()) { |
137 | // Setup Windows compiler runtime calls. |
138 | setLibcallName(RTLIB::SDIV_I64, "_alldiv"); |
139 | setLibcallName(RTLIB::UDIV_I64, "_aulldiv"); |
140 | setLibcallName(RTLIB::SREM_I64, "_allrem"); |
141 | setLibcallName(RTLIB::UREM_I64, "_aullrem"); |
142 | setLibcallName(RTLIB::MUL_I64, "_allmul"); |
143 | setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall); |
144 | setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall); |
145 | setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall); |
146 | setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall); |
147 | setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall); |
148 | } |
149 | |
150 | if (Subtarget.isTargetDarwin()) { |
151 | // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp. |
152 | setUseUnderscoreSetJmp(false); |
153 | setUseUnderscoreLongJmp(false); |
154 | } else if (Subtarget.isTargetWindowsGNU()) { |
155 | // MS runtime is weird: it exports _setjmp, but longjmp! |
156 | setUseUnderscoreSetJmp(true); |
157 | setUseUnderscoreLongJmp(false); |
158 | } else { |
159 | setUseUnderscoreSetJmp(true); |
160 | setUseUnderscoreLongJmp(true); |
161 | } |
162 | |
163 | // Set up the register classes. |
164 | addRegisterClass(MVT::i8, &X86::GR8RegClass); |
165 | addRegisterClass(MVT::i16, &X86::GR16RegClass); |
166 | addRegisterClass(MVT::i32, &X86::GR32RegClass); |
167 | if (Subtarget.is64Bit()) |
168 | addRegisterClass(MVT::i64, &X86::GR64RegClass); |
169 | |
170 | for (MVT VT : MVT::integer_valuetypes()) |
171 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote); |
172 | |
173 | // We don't accept any truncstore of integer registers. |
174 | setTruncStoreAction(MVT::i64, MVT::i32, Expand); |
175 | setTruncStoreAction(MVT::i64, MVT::i16, Expand); |
176 | setTruncStoreAction(MVT::i64, MVT::i8 , Expand); |
177 | setTruncStoreAction(MVT::i32, MVT::i16, Expand); |
178 | setTruncStoreAction(MVT::i32, MVT::i8 , Expand); |
179 | setTruncStoreAction(MVT::i16, MVT::i8, Expand); |
180 | |
181 | setTruncStoreAction(MVT::f64, MVT::f32, Expand); |
182 | |
183 | // SETOEQ and SETUNE require checking two conditions. |
184 | setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand); |
185 | setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand); |
186 | setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand); |
187 | setCondCodeAction(ISD::SETUNE, MVT::f32, Expand); |
188 | setCondCodeAction(ISD::SETUNE, MVT::f64, Expand); |
189 | setCondCodeAction(ISD::SETUNE, MVT::f80, Expand); |
190 | |
191 | // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this |
192 | // operation. |
193 | setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote); |
194 | setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote); |
195 | setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote); |
196 | |
197 | if (Subtarget.is64Bit()) { |
198 | if (!Subtarget.useSoftFloat() && Subtarget.hasAVX512()) |
199 | // f32/f64 are legal, f80 is custom. |
200 | setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom); |
201 | else |
202 | setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote); |
203 | setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom); |
204 | } else if (!Subtarget.useSoftFloat()) { |
205 | // We have an algorithm for SSE2->double, and we turn this into a |
206 | // 64-bit FILD followed by conditional FADD for other targets. |
207 | setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom); |
208 | // We have an algorithm for SSE2, and we turn this into a 64-bit |
209 | // FILD or VCVTUSI2SS/SD for other targets. |
210 | setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom); |
211 | } |
212 | |
213 | // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have |
214 | // this operation. |
215 | setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote); |
216 | setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote); |
217 | |
218 | if (!Subtarget.useSoftFloat()) { |
219 | // SSE has no i16 to fp conversion, only i32. |
220 | if (X86ScalarSSEf32) { |
221 | setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote); |
222 | // f32 and f64 cases are Legal, f80 case is not |
223 | setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom); |
224 | } else { |
225 | setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom); |
226 | setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom); |
227 | } |
228 | } else { |
229 | setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote); |
230 | setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote); |
231 | } |
232 | |
233 | // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have |
234 | // this operation. |
235 | setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote); |
236 | setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote); |
237 | |
238 | if (!Subtarget.useSoftFloat()) { |
239 | // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64 |
240 | // are Legal, f80 is custom lowered. |
241 | setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom); |
242 | setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom); |
243 | |
244 | if (X86ScalarSSEf32) { |
245 | setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote); |
246 | // f32 and f64 cases are Legal, f80 case is not |
247 | setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom); |
248 | } else { |
249 | setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom); |
250 | setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom); |
251 | } |
252 | } else { |
253 | setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote); |
254 | setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Expand); |
255 | setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Expand); |
256 | } |
257 | |
258 | // Handle FP_TO_UINT by promoting the destination to a larger signed |
259 | // conversion. |
260 | setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote); |
261 | setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote); |
262 | setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote); |
263 | |
264 | if (Subtarget.is64Bit()) { |
265 | if (!Subtarget.useSoftFloat() && Subtarget.hasAVX512()) { |
266 | // FP_TO_UINT-i32/i64 is legal for f32/f64, but custom for f80. |
267 | setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom); |
268 | setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom); |
269 | } else { |
270 | setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote); |
271 | setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand); |
272 | } |
273 | } else if (!Subtarget.useSoftFloat()) { |
274 | // Since AVX is a superset of SSE3, only check for SSE here. |
275 | if (Subtarget.hasSSE1() && !Subtarget.hasSSE3()) |
276 | // Expand FP_TO_UINT into a select. |
277 | // FIXME: We would like to use a Custom expander here eventually to do |
278 | // the optimal thing for SSE vs. the default expansion in the legalizer. |
279 | setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand); |
280 | else |
281 | // With AVX512 we can use vcvts[ds]2usi for f32/f64->i32, f80 is custom. |
282 | // With SSE3 we can use fisttpll to convert to a signed i64; without |
283 | // SSE, we're stuck with a fistpll. |
284 | setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom); |
285 | |
286 | setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom); |
287 | } |
288 | |
289 | // TODO: when we have SSE, these could be more efficient, by using movd/movq. |
290 | if (!X86ScalarSSEf64) { |
291 | setOperationAction(ISD::BITCAST , MVT::f32 , Expand); |
292 | setOperationAction(ISD::BITCAST , MVT::i32 , Expand); |
293 | if (Subtarget.is64Bit()) { |
294 | setOperationAction(ISD::BITCAST , MVT::f64 , Expand); |
295 | // Without SSE, i64->f64 goes through memory. |
296 | setOperationAction(ISD::BITCAST , MVT::i64 , Expand); |
297 | } |
298 | } else if (!Subtarget.is64Bit()) |
299 | setOperationAction(ISD::BITCAST , MVT::i64 , Custom); |
300 | |
301 | // Scalar integer divide and remainder are lowered to use operations that |
302 | // produce two results, to match the available instructions. This exposes |
303 | // the two-result form to trivial CSE, which is able to combine x/y and x%y |
304 | // into a single instruction. |
305 | // |
306 | // Scalar integer multiply-high is also lowered to use two-result |
307 | // operations, to match the available instructions. However, plain multiply |
308 | // (low) operations are left as Legal, as there are single-result |
309 | // instructions for this in x86. Using the two-result multiply instructions |
310 | // when both high and low results are needed must be arranged by dagcombine. |
311 | for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) { |
312 | setOperationAction(ISD::MULHS, VT, Expand); |
313 | setOperationAction(ISD::MULHU, VT, Expand); |
314 | setOperationAction(ISD::SDIV, VT, Expand); |
315 | setOperationAction(ISD::UDIV, VT, Expand); |
316 | setOperationAction(ISD::SREM, VT, Expand); |
317 | setOperationAction(ISD::UREM, VT, Expand); |
318 | } |
319 | |
320 | setOperationAction(ISD::BR_JT , MVT::Other, Expand); |
321 | setOperationAction(ISD::BRCOND , MVT::Other, Custom); |
322 | for (auto VT : { MVT::f32, MVT::f64, MVT::f80, MVT::f128, |
323 | MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) { |
324 | setOperationAction(ISD::BR_CC, VT, Expand); |
325 | setOperationAction(ISD::SELECT_CC, VT, Expand); |
326 | } |
327 | if (Subtarget.is64Bit()) |
328 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal); |
329 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal); |
330 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal); |
331 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand); |
332 | setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand); |
333 | |
334 | setOperationAction(ISD::FREM , MVT::f32 , Expand); |
335 | setOperationAction(ISD::FREM , MVT::f64 , Expand); |
336 | setOperationAction(ISD::FREM , MVT::f80 , Expand); |
337 | setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom); |
338 | |
339 | // Promote the i8 variants and force them on up to i32 which has a shorter |
340 | // encoding. |
341 | setOperationPromotedToType(ISD::CTTZ , MVT::i8 , MVT::i32); |
342 | setOperationPromotedToType(ISD::CTTZ_ZERO_UNDEF, MVT::i8 , MVT::i32); |
343 | if (!Subtarget.hasBMI()) { |
344 | setOperationAction(ISD::CTTZ , MVT::i16 , Custom); |
345 | setOperationAction(ISD::CTTZ , MVT::i32 , Custom); |
346 | setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Legal); |
347 | setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Legal); |
348 | if (Subtarget.is64Bit()) { |
349 | setOperationAction(ISD::CTTZ , MVT::i64 , Custom); |
350 | setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Legal); |
351 | } |
352 | } |
353 | |
354 | if (Subtarget.hasLZCNT()) { |
355 | // When promoting the i8 variants, force them to i32 for a shorter |
356 | // encoding. |
357 | setOperationPromotedToType(ISD::CTLZ , MVT::i8 , MVT::i32); |
358 | setOperationPromotedToType(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32); |
359 | } else { |
360 | setOperationAction(ISD::CTLZ , MVT::i8 , Custom); |
361 | setOperationAction(ISD::CTLZ , MVT::i16 , Custom); |
362 | setOperationAction(ISD::CTLZ , MVT::i32 , Custom); |
363 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom); |
364 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom); |
365 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom); |
366 | if (Subtarget.is64Bit()) { |
367 | setOperationAction(ISD::CTLZ , MVT::i64 , Custom); |
368 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom); |
369 | } |
370 | } |
371 | |
372 | // Special handling for half-precision floating point conversions. |
373 | // If we don't have F16C support, then lower half float conversions |
374 | // into library calls. |
375 | if (Subtarget.useSoftFloat() || |
376 | (!Subtarget.hasF16C() && !Subtarget.hasAVX512())) { |
377 | setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand); |
378 | setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand); |
379 | } |
380 | |
381 | // There's never any support for operations beyond MVT::f32. |
382 | setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand); |
383 | setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand); |
384 | setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand); |
385 | setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand); |
386 | |
387 | setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand); |
388 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand); |
389 | setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand); |
390 | setTruncStoreAction(MVT::f32, MVT::f16, Expand); |
391 | setTruncStoreAction(MVT::f64, MVT::f16, Expand); |
392 | setTruncStoreAction(MVT::f80, MVT::f16, Expand); |
393 | |
394 | if (Subtarget.hasPOPCNT()) { |
395 | setOperationAction(ISD::CTPOP , MVT::i8 , Promote); |
396 | } else { |
397 | setOperationAction(ISD::CTPOP , MVT::i8 , Expand); |
398 | setOperationAction(ISD::CTPOP , MVT::i16 , Expand); |
399 | setOperationAction(ISD::CTPOP , MVT::i32 , Expand); |
400 | if (Subtarget.is64Bit()) |
401 | setOperationAction(ISD::CTPOP , MVT::i64 , Expand); |
402 | } |
403 | |
404 | setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom); |
405 | |
406 | if (!Subtarget.hasMOVBE()) |
407 | setOperationAction(ISD::BSWAP , MVT::i16 , Expand); |
408 | |
409 | // These should be promoted to a larger select which is supported. |
410 | setOperationAction(ISD::SELECT , MVT::i1 , Promote); |
411 | // X86 wants to expand cmov itself. |
412 | for (auto VT : { MVT::f32, MVT::f64, MVT::f80, MVT::f128 }) { |
413 | setOperationAction(ISD::SELECT, VT, Custom); |
414 | setOperationAction(ISD::SETCC, VT, Custom); |
415 | } |
416 | for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) { |
417 | if (VT == MVT::i64 && !Subtarget.is64Bit()) |
418 | continue; |
419 | setOperationAction(ISD::SELECT, VT, Custom); |
420 | setOperationAction(ISD::SETCC, VT, Custom); |
421 | } |
422 | setOperationAction(ISD::EH_RETURN , MVT::Other, Custom); |
423 | // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support |
424 | // SjLj exception handling but a light-weight setjmp/longjmp replacement to |
425 | // support continuation, user-level threading, and etc.. As a result, no |
426 | // other SjLj exception interfaces are implemented and please don't build |
427 | // your own exception handling based on them. |
428 | // LLVM/Clang supports zero-cost DWARF exception handling. |
429 | setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom); |
430 | setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom); |
431 | setOperationAction(ISD::EH_SJLJ_SETUP_DISPATCH, MVT::Other, Custom); |
432 | if (TM.Options.ExceptionModel == ExceptionHandling::SjLj) |
433 | setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume"); |
434 | |
435 | // Darwin ABI issue. |
436 | for (auto VT : { MVT::i32, MVT::i64 }) { |
437 | if (VT == MVT::i64 && !Subtarget.is64Bit()) |
438 | continue; |
439 | setOperationAction(ISD::ConstantPool , VT, Custom); |
440 | setOperationAction(ISD::JumpTable , VT, Custom); |
441 | setOperationAction(ISD::GlobalAddress , VT, Custom); |
442 | setOperationAction(ISD::GlobalTLSAddress, VT, Custom); |
443 | setOperationAction(ISD::ExternalSymbol , VT, Custom); |
444 | setOperationAction(ISD::BlockAddress , VT, Custom); |
445 | } |
446 | |
447 | // 64-bit shl, sra, srl (iff 32-bit x86) |
448 | for (auto VT : { MVT::i32, MVT::i64 }) { |
449 | if (VT == MVT::i64 && !Subtarget.is64Bit()) |
450 | continue; |
451 | setOperationAction(ISD::SHL_PARTS, VT, Custom); |
452 | setOperationAction(ISD::SRA_PARTS, VT, Custom); |
453 | setOperationAction(ISD::SRL_PARTS, VT, Custom); |
454 | } |
455 | |
456 | if (Subtarget.hasSSE1()) |
457 | setOperationAction(ISD::PREFETCH , MVT::Other, Legal); |
458 | |
459 | setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom); |
460 | |
461 | // Expand certain atomics |
462 | for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) { |
463 | setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom); |
464 | setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom); |
465 | setOperationAction(ISD::ATOMIC_LOAD_ADD, VT, Custom); |
466 | setOperationAction(ISD::ATOMIC_LOAD_OR, VT, Custom); |
467 | setOperationAction(ISD::ATOMIC_LOAD_XOR, VT, Custom); |
468 | setOperationAction(ISD::ATOMIC_LOAD_AND, VT, Custom); |
469 | setOperationAction(ISD::ATOMIC_STORE, VT, Custom); |
470 | } |
471 | |
472 | if (Subtarget.hasCmpxchg16b()) { |
473 | setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom); |
474 | } |
475 | |
476 | // FIXME - use subtarget debug flags |
477 | if (!Subtarget.isTargetDarwin() && !Subtarget.isTargetELF() && |
478 | !Subtarget.isTargetCygMing() && !Subtarget.isTargetWin64() && |
479 | TM.Options.ExceptionModel != ExceptionHandling::SjLj) { |
480 | setOperationAction(ISD::EH_LABEL, MVT::Other, Expand); |
481 | } |
482 | |
483 | setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom); |
484 | setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom); |
485 | |
486 | setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom); |
487 | setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom); |
488 | |
489 | setOperationAction(ISD::TRAP, MVT::Other, Legal); |
490 | setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal); |
491 | |
492 | // VASTART needs to be custom lowered to use the VarArgsFrameIndex |
493 | setOperationAction(ISD::VASTART , MVT::Other, Custom); |
494 | setOperationAction(ISD::VAEND , MVT::Other, Expand); |
495 | bool Is64Bit = Subtarget.is64Bit(); |
496 | setOperationAction(ISD::VAARG, MVT::Other, Is64Bit ? Custom : Expand); |
497 | setOperationAction(ISD::VACOPY, MVT::Other, Is64Bit ? Custom : Expand); |
498 | |
499 | setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); |
500 | setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); |
501 | |
502 | setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom); |
503 | |
504 | // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering. |
505 | setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom); |
506 | setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom); |
507 | |
508 | if (!Subtarget.useSoftFloat() && X86ScalarSSEf64) { |
509 | // f32 and f64 use SSE. |
510 | // Set up the FP register classes. |
511 | addRegisterClass(MVT::f32, Subtarget.hasAVX512() ? &X86::FR32XRegClass |
512 | : &X86::FR32RegClass); |
513 | addRegisterClass(MVT::f64, Subtarget.hasAVX512() ? &X86::FR64XRegClass |
514 | : &X86::FR64RegClass); |
515 | |
516 | for (auto VT : { MVT::f32, MVT::f64 }) { |
517 | // Use ANDPD to simulate FABS. |
518 | setOperationAction(ISD::FABS, VT, Custom); |
519 | |
520 | // Use XORP to simulate FNEG. |
521 | setOperationAction(ISD::FNEG, VT, Custom); |
522 | |
523 | // Use ANDPD and ORPD to simulate FCOPYSIGN. |
524 | setOperationAction(ISD::FCOPYSIGN, VT, Custom); |
525 | |
526 | // We don't support sin/cos/fmod |
527 | setOperationAction(ISD::FSIN , VT, Expand); |
528 | setOperationAction(ISD::FCOS , VT, Expand); |
529 | setOperationAction(ISD::FSINCOS, VT, Expand); |
530 | } |
531 | |
532 | // Lower this to MOVMSK plus an AND. |
533 | setOperationAction(ISD::FGETSIGN, MVT::i64, Custom); |
534 | setOperationAction(ISD::FGETSIGN, MVT::i32, Custom); |
535 | |
536 | // Expand FP immediates into loads from the stack, except for the special |
537 | // cases we handle. |
538 | addLegalFPImmediate(APFloat(+0.0)); // xorpd |
539 | addLegalFPImmediate(APFloat(+0.0f)); // xorps |
540 | } else if (UseX87 && X86ScalarSSEf32) { |
541 | // Use SSE for f32, x87 for f64. |
542 | // Set up the FP register classes. |
543 | addRegisterClass(MVT::f32, Subtarget.hasAVX512() ? &X86::FR32XRegClass |
544 | : &X86::FR32RegClass); |
545 | addRegisterClass(MVT::f64, &X86::RFP64RegClass); |
546 | |
547 | // Use ANDPS to simulate FABS. |
548 | setOperationAction(ISD::FABS , MVT::f32, Custom); |
549 | |
550 | // Use XORP to simulate FNEG. |
551 | setOperationAction(ISD::FNEG , MVT::f32, Custom); |
552 | |
553 | setOperationAction(ISD::UNDEF, MVT::f64, Expand); |
554 | |
555 | // Use ANDPS and ORPS to simulate FCOPYSIGN. |
556 | setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand); |
557 | setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); |
558 | |
559 | // We don't support sin/cos/fmod |
560 | setOperationAction(ISD::FSIN , MVT::f32, Expand); |
561 | setOperationAction(ISD::FCOS , MVT::f32, Expand); |
562 | setOperationAction(ISD::FSINCOS, MVT::f32, Expand); |
563 | |
564 | // Special cases we handle for FP constants. |
565 | addLegalFPImmediate(APFloat(+0.0f)); // xorps |
566 | addLegalFPImmediate(APFloat(+0.0)); // FLD0 |
567 | addLegalFPImmediate(APFloat(+1.0)); // FLD1 |
568 | addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS |
569 | addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS |
570 | |
571 | if (!TM.Options.UnsafeFPMath) { |
572 | setOperationAction(ISD::FSIN , MVT::f64, Expand); |
573 | setOperationAction(ISD::FCOS , MVT::f64, Expand); |
574 | setOperationAction(ISD::FSINCOS, MVT::f64, Expand); |
575 | } |
576 | } else if (UseX87) { |
577 | // f32 and f64 in x87. |
578 | // Set up the FP register classes. |
579 | addRegisterClass(MVT::f64, &X86::RFP64RegClass); |
580 | addRegisterClass(MVT::f32, &X86::RFP32RegClass); |
581 | |
582 | for (auto VT : { MVT::f32, MVT::f64 }) { |
583 | setOperationAction(ISD::UNDEF, VT, Expand); |
584 | setOperationAction(ISD::FCOPYSIGN, VT, Expand); |
585 | |
586 | if (!TM.Options.UnsafeFPMath) { |
587 | setOperationAction(ISD::FSIN , VT, Expand); |
588 | setOperationAction(ISD::FCOS , VT, Expand); |
589 | setOperationAction(ISD::FSINCOS, VT, Expand); |
590 | } |
591 | } |
592 | addLegalFPImmediate(APFloat(+0.0)); // FLD0 |
593 | addLegalFPImmediate(APFloat(+1.0)); // FLD1 |
594 | addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS |
595 | addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS |
596 | addLegalFPImmediate(APFloat(+0.0f)); // FLD0 |
597 | addLegalFPImmediate(APFloat(+1.0f)); // FLD1 |
598 | addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS |
599 | addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS |
600 | } |
601 | |
602 | // We don't support FMA. |
603 | setOperationAction(ISD::FMA, MVT::f64, Expand); |
604 | setOperationAction(ISD::FMA, MVT::f32, Expand); |
605 | |
606 | // Long double always uses X87, except f128 in MMX. |
607 | if (UseX87) { |
608 | if (Subtarget.is64Bit() && Subtarget.hasMMX()) { |
609 | addRegisterClass(MVT::f128, &X86::FR128RegClass); |
610 | ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); |
611 | setOperationAction(ISD::FABS , MVT::f128, Custom); |
612 | setOperationAction(ISD::FNEG , MVT::f128, Custom); |
613 | setOperationAction(ISD::FCOPYSIGN, MVT::f128, Custom); |
614 | } |
615 | |
616 | addRegisterClass(MVT::f80, &X86::RFP80RegClass); |
617 | setOperationAction(ISD::UNDEF, MVT::f80, Expand); |
618 | setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand); |
619 | { |
620 | APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended()); |
621 | addLegalFPImmediate(TmpFlt); // FLD0 |
622 | TmpFlt.changeSign(); |
623 | addLegalFPImmediate(TmpFlt); // FLD0/FCHS |
624 | |
625 | bool ignored; |
626 | APFloat TmpFlt2(+1.0); |
627 | TmpFlt2.convert(APFloat::x87DoubleExtended(), APFloat::rmNearestTiesToEven, |
628 | &ignored); |
629 | addLegalFPImmediate(TmpFlt2); // FLD1 |
630 | TmpFlt2.changeSign(); |
631 | addLegalFPImmediate(TmpFlt2); // FLD1/FCHS |
632 | } |
633 | |
634 | if (!TM.Options.UnsafeFPMath) { |
635 | setOperationAction(ISD::FSIN , MVT::f80, Expand); |
636 | setOperationAction(ISD::FCOS , MVT::f80, Expand); |
637 | setOperationAction(ISD::FSINCOS, MVT::f80, Expand); |
638 | } |
639 | |
640 | setOperationAction(ISD::FFLOOR, MVT::f80, Expand); |
641 | setOperationAction(ISD::FCEIL, MVT::f80, Expand); |
642 | setOperationAction(ISD::FTRUNC, MVT::f80, Expand); |
643 | setOperationAction(ISD::FRINT, MVT::f80, Expand); |
644 | setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand); |
645 | setOperationAction(ISD::FMA, MVT::f80, Expand); |
646 | } |
647 | |
648 | // Always use a library call for pow. |
649 | setOperationAction(ISD::FPOW , MVT::f32 , Expand); |
650 | setOperationAction(ISD::FPOW , MVT::f64 , Expand); |
651 | setOperationAction(ISD::FPOW , MVT::f80 , Expand); |
652 | |
653 | setOperationAction(ISD::FLOG, MVT::f80, Expand); |
654 | setOperationAction(ISD::FLOG2, MVT::f80, Expand); |
655 | setOperationAction(ISD::FLOG10, MVT::f80, Expand); |
656 | setOperationAction(ISD::FEXP, MVT::f80, Expand); |
657 | setOperationAction(ISD::FEXP2, MVT::f80, Expand); |
658 | setOperationAction(ISD::FMINNUM, MVT::f80, Expand); |
659 | setOperationAction(ISD::FMAXNUM, MVT::f80, Expand); |
660 | |
661 | // Some FP actions are always expanded for vector types. |
662 | for (auto VT : { MVT::v4f32, MVT::v8f32, MVT::v16f32, |
663 | MVT::v2f64, MVT::v4f64, MVT::v8f64 }) { |
664 | setOperationAction(ISD::FSIN, VT, Expand); |
665 | setOperationAction(ISD::FSINCOS, VT, Expand); |
666 | setOperationAction(ISD::FCOS, VT, Expand); |
667 | setOperationAction(ISD::FREM, VT, Expand); |
668 | setOperationAction(ISD::FCOPYSIGN, VT, Expand); |
669 | setOperationAction(ISD::FPOW, VT, Expand); |
670 | setOperationAction(ISD::FLOG, VT, Expand); |
671 | setOperationAction(ISD::FLOG2, VT, Expand); |
672 | setOperationAction(ISD::FLOG10, VT, Expand); |
673 | setOperationAction(ISD::FEXP, VT, Expand); |
674 | setOperationAction(ISD::FEXP2, VT, Expand); |
675 | } |
676 | |
677 | // First set operation action for all vector types to either promote |
678 | // (for widening) or expand (for scalarization). Then we will selectively |
679 | // turn on ones that can be effectively codegen'd. |
680 | for (MVT VT : MVT::vector_valuetypes()) { |
681 | setOperationAction(ISD::SDIV, VT, Expand); |
682 | setOperationAction(ISD::UDIV, VT, Expand); |
683 | setOperationAction(ISD::SREM, VT, Expand); |
684 | setOperationAction(ISD::UREM, VT, Expand); |
685 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand); |
686 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand); |
687 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand); |
688 | setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand); |
689 | setOperationAction(ISD::FMA, VT, Expand); |
690 | setOperationAction(ISD::FFLOOR, VT, Expand); |
691 | setOperationAction(ISD::FCEIL, VT, Expand); |
692 | setOperationAction(ISD::FTRUNC, VT, Expand); |
693 | setOperationAction(ISD::FRINT, VT, Expand); |
694 | setOperationAction(ISD::FNEARBYINT, VT, Expand); |
695 | setOperationAction(ISD::SMUL_LOHI, VT, Expand); |
696 | setOperationAction(ISD::MULHS, VT, Expand); |
697 | setOperationAction(ISD::UMUL_LOHI, VT, Expand); |
698 | setOperationAction(ISD::MULHU, VT, Expand); |
699 | setOperationAction(ISD::SDIVREM, VT, Expand); |
700 | setOperationAction(ISD::UDIVREM, VT, Expand); |
701 | setOperationAction(ISD::CTPOP, VT, Expand); |
702 | setOperationAction(ISD::CTTZ, VT, Expand); |
703 | setOperationAction(ISD::CTLZ, VT, Expand); |
704 | setOperationAction(ISD::ROTL, VT, Expand); |
705 | setOperationAction(ISD::ROTR, VT, Expand); |
706 | setOperationAction(ISD::BSWAP, VT, Expand); |
707 | setOperationAction(ISD::SETCC, VT, Expand); |
708 | setOperationAction(ISD::FP_TO_UINT, VT, Expand); |
709 | setOperationAction(ISD::FP_TO_SINT, VT, Expand); |
710 | setOperationAction(ISD::UINT_TO_FP, VT, Expand); |
711 | setOperationAction(ISD::SINT_TO_FP, VT, Expand); |
712 | setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand); |
713 | setOperationAction(ISD::TRUNCATE, VT, Expand); |
714 | setOperationAction(ISD::SIGN_EXTEND, VT, Expand); |
715 | setOperationAction(ISD::ZERO_EXTEND, VT, Expand); |
716 | setOperationAction(ISD::ANY_EXTEND, VT, Expand); |
717 | setOperationAction(ISD::SELECT_CC, VT, Expand); |
718 | for (MVT InnerVT : MVT::vector_valuetypes()) { |
719 | setTruncStoreAction(InnerVT, VT, Expand); |
720 | |
721 | setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand); |
722 | setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand); |
723 | |
724 | // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like |
725 | // types, we have to deal with them whether we ask for Expansion or not. |
726 | // Setting Expand causes its own optimisation problems though, so leave |
727 | // them legal. |
728 | if (VT.getVectorElementType() == MVT::i1) |
729 | setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand); |
730 | |
731 | // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are |
732 | // split/scalarized right now. |
733 | if (VT.getVectorElementType() == MVT::f16) |
734 | setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand); |
735 | } |
736 | } |
737 | |
738 | // FIXME: In order to prevent SSE instructions being expanded to MMX ones |
739 | // with -msoft-float, disable use of MMX as well. |
740 | if (!Subtarget.useSoftFloat() && Subtarget.hasMMX()) { |
741 | addRegisterClass(MVT::x86mmx, &X86::VR64RegClass); |
742 | // No operations on x86mmx supported, everything uses intrinsics. |
743 | } |
744 | |
745 | if (!Subtarget.useSoftFloat() && Subtarget.hasSSE1()) { |
746 | addRegisterClass(MVT::v4f32, Subtarget.hasVLX() ? &X86::VR128XRegClass |
747 | : &X86::VR128RegClass); |
748 | |
749 | setOperationAction(ISD::FNEG, MVT::v4f32, Custom); |
750 | setOperationAction(ISD::FABS, MVT::v4f32, Custom); |
751 | setOperationAction(ISD::FCOPYSIGN, MVT::v4f32, Custom); |
752 | setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom); |
753 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom); |
754 | setOperationAction(ISD::VSELECT, MVT::v4f32, Custom); |
755 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom); |
756 | setOperationAction(ISD::SELECT, MVT::v4f32, Custom); |
757 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom); |
758 | } |
759 | |
760 | if (!Subtarget.useSoftFloat() && Subtarget.hasSSE2()) { |
761 | addRegisterClass(MVT::v2f64, Subtarget.hasVLX() ? &X86::VR128XRegClass |
762 | : &X86::VR128RegClass); |
763 | |
764 | // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM |
765 | // registers cannot be used even for integer operations. |
766 | addRegisterClass(MVT::v16i8, Subtarget.hasVLX() ? &X86::VR128XRegClass |
767 | : &X86::VR128RegClass); |
768 | addRegisterClass(MVT::v8i16, Subtarget.hasVLX() ? &X86::VR128XRegClass |
769 | : &X86::VR128RegClass); |
770 | addRegisterClass(MVT::v4i32, Subtarget.hasVLX() ? &X86::VR128XRegClass |
771 | : &X86::VR128RegClass); |
772 | addRegisterClass(MVT::v2i64, Subtarget.hasVLX() ? &X86::VR128XRegClass |
773 | : &X86::VR128RegClass); |
774 | |
775 | setOperationAction(ISD::MUL, MVT::v16i8, Custom); |
776 | setOperationAction(ISD::MUL, MVT::v4i32, Custom); |
777 | setOperationAction(ISD::MUL, MVT::v2i64, Custom); |
778 | setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom); |
779 | setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom); |
780 | setOperationAction(ISD::MULHU, MVT::v16i8, Custom); |
781 | setOperationAction(ISD::MULHS, MVT::v16i8, Custom); |
782 | setOperationAction(ISD::MULHU, MVT::v8i16, Legal); |
783 | setOperationAction(ISD::MULHS, MVT::v8i16, Legal); |
784 | setOperationAction(ISD::MUL, MVT::v8i16, Legal); |
785 | setOperationAction(ISD::FNEG, MVT::v2f64, Custom); |
786 | setOperationAction(ISD::FABS, MVT::v2f64, Custom); |
787 | setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Custom); |
788 | |
789 | setOperationAction(ISD::SMAX, MVT::v8i16, Legal); |
790 | setOperationAction(ISD::UMAX, MVT::v16i8, Legal); |
791 | setOperationAction(ISD::SMIN, MVT::v8i16, Legal); |
792 | setOperationAction(ISD::UMIN, MVT::v16i8, Legal); |
793 | |
794 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom); |
795 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom); |
796 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom); |
797 | |
798 | for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64 }) { |
799 | setOperationAction(ISD::SETCC, VT, Custom); |
800 | setOperationAction(ISD::CTPOP, VT, Custom); |
801 | setOperationAction(ISD::CTTZ, VT, Custom); |
802 | } |
803 | |
804 | for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) { |
805 | setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom); |
806 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); |
807 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); |
808 | setOperationAction(ISD::VSELECT, VT, Custom); |
809 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); |
810 | } |
811 | |
812 | // We support custom legalizing of sext and anyext loads for specific |
813 | // memory vector types which we can load as a scalar (or sequence of |
814 | // scalars) and extend in-register to a legal 128-bit vector type. For sext |
815 | // loads these must work with a single scalar load. |
816 | for (MVT VT : MVT::integer_vector_valuetypes()) { |
817 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom); |
818 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom); |
819 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom); |
820 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom); |
821 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom); |
822 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom); |
823 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom); |
824 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom); |
825 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom); |
826 | } |
827 | |
828 | for (auto VT : { MVT::v2f64, MVT::v2i64 }) { |
829 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); |
830 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); |
831 | setOperationAction(ISD::VSELECT, VT, Custom); |
832 | |
833 | if (VT == MVT::v2i64 && !Subtarget.is64Bit()) |
834 | continue; |
835 | |
836 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); |
837 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); |
838 | } |
839 | |
840 | // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64. |
841 | for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) { |
842 | setOperationPromotedToType(ISD::AND, VT, MVT::v2i64); |
843 | setOperationPromotedToType(ISD::OR, VT, MVT::v2i64); |
844 | setOperationPromotedToType(ISD::XOR, VT, MVT::v2i64); |
845 | setOperationPromotedToType(ISD::LOAD, VT, MVT::v2i64); |
846 | setOperationPromotedToType(ISD::SELECT, VT, MVT::v2i64); |
847 | } |
848 | |
849 | // Custom lower v2i64 and v2f64 selects. |
850 | setOperationAction(ISD::SELECT, MVT::v2f64, Custom); |
851 | setOperationAction(ISD::SELECT, MVT::v2i64, Custom); |
852 | |
853 | setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal); |
854 | setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom); |
855 | |
856 | setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal); |
857 | setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom); |
858 | |
859 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom); |
860 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom); |
861 | setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom); |
862 | |
863 | // Fast v2f32 UINT_TO_FP( v2i32 ) custom conversion. |
864 | setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom); |
865 | |
866 | setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom); |
867 | setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom); |
868 | |
869 | for (MVT VT : MVT::fp_vector_valuetypes()) |
870 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal); |
871 | |
872 | setOperationAction(ISD::BITCAST, MVT::v2i32, Custom); |
873 | setOperationAction(ISD::BITCAST, MVT::v4i16, Custom); |
874 | setOperationAction(ISD::BITCAST, MVT::v8i8, Custom); |
875 | |
876 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom); |
877 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom); |
878 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom); |
879 | |
880 | // In the customized shift lowering, the legal v4i32/v2i64 cases |
881 | // in AVX2 will be recognized. |
882 | for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64 }) { |
883 | setOperationAction(ISD::SRL, VT, Custom); |
884 | setOperationAction(ISD::SHL, VT, Custom); |
885 | setOperationAction(ISD::SRA, VT, Custom); |
886 | } |
887 | } |
888 | |
889 | if (!Subtarget.useSoftFloat() && Subtarget.hasSSSE3()) { |
890 | setOperationAction(ISD::ABS, MVT::v16i8, Legal); |
891 | setOperationAction(ISD::ABS, MVT::v8i16, Legal); |
892 | setOperationAction(ISD::ABS, MVT::v4i32, Legal); |
893 | setOperationAction(ISD::BITREVERSE, MVT::v16i8, Custom); |
894 | setOperationAction(ISD::CTLZ, MVT::v16i8, Custom); |
895 | setOperationAction(ISD::CTLZ, MVT::v8i16, Custom); |
896 | setOperationAction(ISD::CTLZ, MVT::v4i32, Custom); |
897 | setOperationAction(ISD::CTLZ, MVT::v2i64, Custom); |
898 | } |
899 | |
900 | if (!Subtarget.useSoftFloat() && Subtarget.hasSSE41()) { |
901 | for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) { |
902 | setOperationAction(ISD::FFLOOR, RoundedTy, Legal); |
903 | setOperationAction(ISD::FCEIL, RoundedTy, Legal); |
904 | setOperationAction(ISD::FTRUNC, RoundedTy, Legal); |
905 | setOperationAction(ISD::FRINT, RoundedTy, Legal); |
906 | setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal); |
907 | } |
908 | |
909 | setOperationAction(ISD::SMAX, MVT::v16i8, Legal); |
910 | setOperationAction(ISD::SMAX, MVT::v4i32, Legal); |
911 | setOperationAction(ISD::UMAX, MVT::v8i16, Legal); |
912 | setOperationAction(ISD::UMAX, MVT::v4i32, Legal); |
913 | setOperationAction(ISD::SMIN, MVT::v16i8, Legal); |
914 | setOperationAction(ISD::SMIN, MVT::v4i32, Legal); |
915 | setOperationAction(ISD::UMIN, MVT::v8i16, Legal); |
916 | setOperationAction(ISD::UMIN, MVT::v4i32, Legal); |
917 | |
918 | // FIXME: Do we need to handle scalar-to-vector here? |
919 | setOperationAction(ISD::MUL, MVT::v4i32, Legal); |
920 | |
921 | // We directly match byte blends in the backend as they match the VSELECT |
922 | // condition form. |
923 | setOperationAction(ISD::VSELECT, MVT::v16i8, Legal); |
924 | |
925 | // SSE41 brings specific instructions for doing vector sign extend even in |
926 | // cases where we don't have SRA. |
927 | for (auto VT : { MVT::v8i16, MVT::v4i32, MVT::v2i64 }) { |
928 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Legal); |
929 | setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Legal); |
930 | } |
931 | |
932 | for (MVT VT : MVT::integer_vector_valuetypes()) { |
933 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom); |
934 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom); |
935 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom); |
936 | } |
937 | |
938 | // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X |
939 | for (auto LoadExtOp : { ISD::SEXTLOAD, ISD::ZEXTLOAD }) { |
940 | setLoadExtAction(LoadExtOp, MVT::v8i16, MVT::v8i8, Legal); |
941 | setLoadExtAction(LoadExtOp, MVT::v4i32, MVT::v4i8, Legal); |
942 | setLoadExtAction(LoadExtOp, MVT::v2i64, MVT::v2i8, Legal); |
943 | setLoadExtAction(LoadExtOp, MVT::v4i32, MVT::v4i16, Legal); |
944 | setLoadExtAction(LoadExtOp, MVT::v2i64, MVT::v2i16, Legal); |
945 | setLoadExtAction(LoadExtOp, MVT::v2i64, MVT::v2i32, Legal); |
946 | } |
947 | |
948 | // i8 vectors are custom because the source register and source |
949 | // source memory operand types are not the same width. |
950 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom); |
951 | } |
952 | |
953 | if (!Subtarget.useSoftFloat() && Subtarget.hasXOP()) { |
954 | for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, |
955 | MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 }) |
956 | setOperationAction(ISD::ROTL, VT, Custom); |
957 | |
958 | // XOP can efficiently perform BITREVERSE with VPPERM. |
959 | for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) |
960 | setOperationAction(ISD::BITREVERSE, VT, Custom); |
961 | |
962 | for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, |
963 | MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 }) |
964 | setOperationAction(ISD::BITREVERSE, VT, Custom); |
965 | } |
966 | |
967 | if (!Subtarget.useSoftFloat() && Subtarget.hasFp256()) { |
968 | bool HasInt256 = Subtarget.hasInt256(); |
969 | |
970 | addRegisterClass(MVT::v32i8, Subtarget.hasVLX() ? &X86::VR256XRegClass |
971 | : &X86::VR256RegClass); |
972 | addRegisterClass(MVT::v16i16, Subtarget.hasVLX() ? &X86::VR256XRegClass |
973 | : &X86::VR256RegClass); |
974 | addRegisterClass(MVT::v8i32, Subtarget.hasVLX() ? &X86::VR256XRegClass |
975 | : &X86::VR256RegClass); |
976 | addRegisterClass(MVT::v8f32, Subtarget.hasVLX() ? &X86::VR256XRegClass |
977 | : &X86::VR256RegClass); |
978 | addRegisterClass(MVT::v4i64, Subtarget.hasVLX() ? &X86::VR256XRegClass |
979 | : &X86::VR256RegClass); |
980 | addRegisterClass(MVT::v4f64, Subtarget.hasVLX() ? &X86::VR256XRegClass |
981 | : &X86::VR256RegClass); |
982 | |
983 | for (auto VT : { MVT::v8f32, MVT::v4f64 }) { |
984 | setOperationAction(ISD::FFLOOR, VT, Legal); |
985 | setOperationAction(ISD::FCEIL, VT, Legal); |
986 | setOperationAction(ISD::FTRUNC, VT, Legal); |
987 | setOperationAction(ISD::FRINT, VT, Legal); |
988 | setOperationAction(ISD::FNEARBYINT, VT, Legal); |
989 | setOperationAction(ISD::FNEG, VT, Custom); |
990 | setOperationAction(ISD::FABS, VT, Custom); |
991 | setOperationAction(ISD::FCOPYSIGN, VT, Custom); |
992 | } |
993 | |
994 | // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted |
995 | // even though v8i16 is a legal type. |
996 | setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote); |
997 | setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote); |
998 | setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal); |
999 | |
1000 | setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote); |
1001 | setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal); |
1002 | setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal); |
1003 | |
1004 | setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom); |
1005 | setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom); |
1006 | |
1007 | for (MVT VT : MVT::fp_vector_valuetypes()) |
1008 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal); |
1009 | |
1010 | // In the customized shift lowering, the legal v8i32/v4i64 cases |
1011 | // in AVX2 will be recognized. |
1012 | for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 }) { |
1013 | setOperationAction(ISD::SRL, VT, Custom); |
1014 | setOperationAction(ISD::SHL, VT, Custom); |
1015 | setOperationAction(ISD::SRA, VT, Custom); |
1016 | } |
1017 | |
1018 | setOperationAction(ISD::SELECT, MVT::v4f64, Custom); |
1019 | setOperationAction(ISD::SELECT, MVT::v4i64, Custom); |
1020 | setOperationAction(ISD::SELECT, MVT::v8f32, Custom); |
1021 | |
1022 | for (auto VT : { MVT::v16i16, MVT::v8i32, MVT::v4i64 }) { |
1023 | setOperationAction(ISD::SIGN_EXTEND, VT, Custom); |
1024 | setOperationAction(ISD::ZERO_EXTEND, VT, Custom); |
1025 | setOperationAction(ISD::ANY_EXTEND, VT, Custom); |
1026 | } |
1027 | |
1028 | setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom); |
1029 | setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom); |
1030 | setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom); |
1031 | setOperationAction(ISD::BITREVERSE, MVT::v32i8, Custom); |
1032 | |
1033 | for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 }) { |
1034 | setOperationAction(ISD::SETCC, VT, Custom); |
1035 | setOperationAction(ISD::CTPOP, VT, Custom); |
1036 | setOperationAction(ISD::CTTZ, VT, Custom); |
1037 | setOperationAction(ISD::CTLZ, VT, Custom); |
1038 | } |
1039 | |
1040 | if (Subtarget.hasAnyFMA()) { |
1041 | for (auto VT : { MVT::f32, MVT::f64, MVT::v4f32, MVT::v8f32, |
1042 | MVT::v2f64, MVT::v4f64 }) |
1043 | setOperationAction(ISD::FMA, VT, Legal); |
1044 | } |
1045 | |
1046 | for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 }) { |
1047 | setOperationAction(ISD::ADD, VT, HasInt256 ? Legal : Custom); |
1048 | setOperationAction(ISD::SUB, VT, HasInt256 ? Legal : Custom); |
1049 | } |
1050 | |
1051 | setOperationAction(ISD::MUL, MVT::v4i64, Custom); |
1052 | setOperationAction(ISD::MUL, MVT::v8i32, HasInt256 ? Legal : Custom); |
1053 | setOperationAction(ISD::MUL, MVT::v16i16, HasInt256 ? Legal : Custom); |
1054 | setOperationAction(ISD::MUL, MVT::v32i8, Custom); |
1055 | |
1056 | setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom); |
1057 | setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom); |
1058 | |
1059 | setOperationAction(ISD::MULHU, MVT::v16i16, HasInt256 ? Legal : Custom); |
1060 | setOperationAction(ISD::MULHS, MVT::v16i16, HasInt256 ? Legal : Custom); |
1061 | setOperationAction(ISD::MULHU, MVT::v32i8, Custom); |
1062 | setOperationAction(ISD::MULHS, MVT::v32i8, Custom); |
1063 | |
1064 | for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32 }) { |
1065 | setOperationAction(ISD::ABS, VT, HasInt256 ? Legal : Custom); |
1066 | setOperationAction(ISD::SMAX, VT, HasInt256 ? Legal : Custom); |
1067 | setOperationAction(ISD::UMAX, VT, HasInt256 ? Legal : Custom); |
1068 | setOperationAction(ISD::SMIN, VT, HasInt256 ? Legal : Custom); |
1069 | setOperationAction(ISD::UMIN, VT, HasInt256 ? Legal : Custom); |
1070 | } |
1071 | |
1072 | if (HasInt256) { |
1073 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i64, Custom); |
1074 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i32, Custom); |
1075 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v16i16, Custom); |
1076 | |
1077 | // The custom lowering for UINT_TO_FP for v8i32 becomes interesting |
1078 | // when we have a 256bit-wide blend with immediate. |
1079 | setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom); |
1080 | |
1081 | // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X |
1082 | for (auto LoadExtOp : { ISD::SEXTLOAD, ISD::ZEXTLOAD }) { |
1083 | setLoadExtAction(LoadExtOp, MVT::v16i16, MVT::v16i8, Legal); |
1084 | setLoadExtAction(LoadExtOp, MVT::v8i32, MVT::v8i8, Legal); |
1085 | setLoadExtAction(LoadExtOp, MVT::v4i64, MVT::v4i8, Legal); |
1086 | setLoadExtAction(LoadExtOp, MVT::v8i32, MVT::v8i16, Legal); |
1087 | setLoadExtAction(LoadExtOp, MVT::v4i64, MVT::v4i16, Legal); |
1088 | setLoadExtAction(LoadExtOp, MVT::v4i64, MVT::v4i32, Legal); |
1089 | } |
1090 | } |
1091 | |
1092 | for (auto VT : { MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64, |
1093 | MVT::v4f32, MVT::v8f32, MVT::v2f64, MVT::v4f64 }) { |
1094 | setOperationAction(ISD::MLOAD, VT, Legal); |
1095 | setOperationAction(ISD::MSTORE, VT, Legal); |
1096 | } |
1097 | |
1098 | // Extract subvector is special because the value type |
1099 | // (result) is 128-bit but the source is 256-bit wide. |
1100 | for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, |
1101 | MVT::v4f32, MVT::v2f64 }) { |
1102 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom); |
1103 | } |
1104 | |
1105 | // Custom lower several nodes for 256-bit types. |
1106 | for (MVT VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64, |
1107 | MVT::v8f32, MVT::v4f64 }) { |
1108 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); |
1109 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); |
1110 | setOperationAction(ISD::VSELECT, VT, Custom); |
1111 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); |
1112 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); |
1113 | setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom); |
1114 | setOperationAction(ISD::INSERT_SUBVECTOR, VT, Legal); |
1115 | setOperationAction(ISD::CONCAT_VECTORS, VT, Custom); |
1116 | } |
1117 | |
1118 | if (HasInt256) |
1119 | setOperationAction(ISD::VSELECT, MVT::v32i8, Legal); |
1120 | |
1121 | // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64. |
1122 | for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32 }) { |
1123 | setOperationPromotedToType(ISD::AND, VT, MVT::v4i64); |
1124 | setOperationPromotedToType(ISD::OR, VT, MVT::v4i64); |
1125 | setOperationPromotedToType(ISD::XOR, VT, MVT::v4i64); |
1126 | setOperationPromotedToType(ISD::LOAD, VT, MVT::v4i64); |
1127 | setOperationPromotedToType(ISD::SELECT, VT, MVT::v4i64); |
1128 | } |
1129 | } |
1130 | |
1131 | if (!Subtarget.useSoftFloat() && Subtarget.hasAVX512()) { |
1132 | addRegisterClass(MVT::v16i32, &X86::VR512RegClass); |
1133 | addRegisterClass(MVT::v16f32, &X86::VR512RegClass); |
1134 | addRegisterClass(MVT::v8i64, &X86::VR512RegClass); |
1135 | addRegisterClass(MVT::v8f64, &X86::VR512RegClass); |
1136 | |
1137 | addRegisterClass(MVT::v1i1, &X86::VK1RegClass); |
1138 | addRegisterClass(MVT::v8i1, &X86::VK8RegClass); |
1139 | addRegisterClass(MVT::v16i1, &X86::VK16RegClass); |
1140 | |
1141 | for (MVT VT : MVT::fp_vector_valuetypes()) |
1142 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal); |
1143 | |
1144 | for (auto ExtType : {ISD::ZEXTLOAD, ISD::SEXTLOAD, ISD::EXTLOAD}) { |
1145 | setLoadExtAction(ExtType, MVT::v16i32, MVT::v16i8, Legal); |
1146 | setLoadExtAction(ExtType, MVT::v16i32, MVT::v16i16, Legal); |
1147 | setLoadExtAction(ExtType, MVT::v32i16, MVT::v32i8, Legal); |
1148 | setLoadExtAction(ExtType, MVT::v8i64, MVT::v8i8, Legal); |
1149 | setLoadExtAction(ExtType, MVT::v8i64, MVT::v8i16, Legal); |
1150 | setLoadExtAction(ExtType, MVT::v8i64, MVT::v8i32, Legal); |
1151 | } |
1152 | |
1153 | for (MVT VT : {MVT::v2i64, MVT::v4i32, MVT::v8i32, MVT::v4i64, MVT::v8i16, |
1154 | MVT::v16i8, MVT::v16i16, MVT::v32i8, MVT::v16i32, |
1155 | MVT::v8i64, MVT::v32i16, MVT::v64i8}) { |
1156 | MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements()); |
1157 | setLoadExtAction(ISD::SEXTLOAD, VT, MaskVT, Custom); |
1158 | setLoadExtAction(ISD::ZEXTLOAD, VT, MaskVT, Custom); |
1159 | setLoadExtAction(ISD::EXTLOAD, VT, MaskVT, Custom); |
1160 | setTruncStoreAction(VT, MaskVT, Custom); |
1161 | } |
1162 | |
1163 | for (MVT VT : { MVT::v16f32, MVT::v8f64 }) { |
1164 | setOperationAction(ISD::FNEG, VT, Custom); |
1165 | setOperationAction(ISD::FABS, VT, Custom); |
1166 | setOperationAction(ISD::FMA, VT, Legal); |
1167 | setOperationAction(ISD::FCOPYSIGN, VT, Custom); |
1168 | } |
1169 | |
1170 | setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal); |
1171 | setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal); |
1172 | setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal); |
1173 | setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal); |
1174 | setOperationAction(ISD::FP_TO_UINT, MVT::v2i32, Custom); |
1175 | setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal); |
1176 | setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom); |
1177 | setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom); |
1178 | setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote); |
1179 | setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote); |
1180 | setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal); |
1181 | setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal); |
1182 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal); |
1183 | setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom); |
1184 | setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom); |
1185 | setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom); |
1186 | setOperationAction(ISD::UINT_TO_FP, MVT::v16i1, Custom); |
1187 | setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom); |
1188 | setOperationAction(ISD::UINT_TO_FP, MVT::v8i1, Custom); |
1189 | setOperationAction(ISD::SINT_TO_FP, MVT::v4i1, Custom); |
1190 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i1, Custom); |
1191 | setOperationAction(ISD::SINT_TO_FP, MVT::v2i1, Custom); |
1192 | setOperationAction(ISD::UINT_TO_FP, MVT::v2i1, Custom); |
1193 | setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal); |
1194 | setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal); |
1195 | |
1196 | setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal); |
1197 | setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal); |
1198 | setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal); |
1199 | setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal); |
1200 | setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal); |
1201 | if (Subtarget.hasVLX()){ |
1202 | setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal); |
1203 | setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal); |
1204 | setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal); |
1205 | setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal); |
1206 | setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal); |
1207 | |
1208 | setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal); |
1209 | setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal); |
1210 | setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal); |
1211 | setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal); |
1212 | setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal); |
1213 | } else { |
1214 | for (auto VT : {MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64, |
1215 | MVT::v4f32, MVT::v8f32, MVT::v2f64, MVT::v4f64}) { |
1216 | setOperationAction(ISD::MLOAD, VT, Custom); |
1217 | setOperationAction(ISD::MSTORE, VT, Custom); |
1218 | } |
1219 | } |
1220 | setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom); |
1221 | setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom); |
1222 | |
1223 | if (Subtarget.hasDQI()) { |
1224 | for (auto VT : { MVT::v2i64, MVT::v4i64, MVT::v8i64 }) { |
1225 | setOperationAction(ISD::SINT_TO_FP, VT, Legal); |
1226 | setOperationAction(ISD::UINT_TO_FP, VT, Legal); |
1227 | setOperationAction(ISD::FP_TO_SINT, VT, Legal); |
1228 | setOperationAction(ISD::FP_TO_UINT, VT, Legal); |
1229 | } |
1230 | if (Subtarget.hasVLX()) { |
1231 | // Fast v2f32 SINT_TO_FP( v2i32 ) custom conversion. |
1232 | setOperationAction(ISD::SINT_TO_FP, MVT::v2f32, Custom); |
1233 | setOperationAction(ISD::FP_TO_SINT, MVT::v2f32, Custom); |
1234 | setOperationAction(ISD::FP_TO_UINT, MVT::v2f32, Custom); |
1235 | } |
1236 | } |
1237 | if (Subtarget.hasVLX()) { |
1238 | setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal); |
1239 | setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal); |
1240 | setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal); |
1241 | setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal); |
1242 | setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal); |
1243 | setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal); |
1244 | setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal); |
1245 | setOperationAction(ISD::ZERO_EXTEND, MVT::v4i32, Custom); |
1246 | setOperationAction(ISD::ZERO_EXTEND, MVT::v2i64, Custom); |
1247 | setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom); |
1248 | setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom); |
1249 | |
1250 | // FIXME. This commands are available on SSE/AVX2, add relevant patterns. |
1251 | setLoadExtAction(ISD::EXTLOAD, MVT::v8i32, MVT::v8i8, Legal); |
1252 | setLoadExtAction(ISD::EXTLOAD, MVT::v8i32, MVT::v8i16, Legal); |
1253 | setLoadExtAction(ISD::EXTLOAD, MVT::v4i32, MVT::v4i8, Legal); |
1254 | setLoadExtAction(ISD::EXTLOAD, MVT::v4i32, MVT::v4i16, Legal); |
1255 | setLoadExtAction(ISD::EXTLOAD, MVT::v4i64, MVT::v4i8, Legal); |
1256 | setLoadExtAction(ISD::EXTLOAD, MVT::v4i64, MVT::v4i16, Legal); |
1257 | setLoadExtAction(ISD::EXTLOAD, MVT::v4i64, MVT::v4i32, Legal); |
1258 | setLoadExtAction(ISD::EXTLOAD, MVT::v2i64, MVT::v2i8, Legal); |
1259 | setLoadExtAction(ISD::EXTLOAD, MVT::v2i64, MVT::v2i16, Legal); |
1260 | setLoadExtAction(ISD::EXTLOAD, MVT::v2i64, MVT::v2i32, Legal); |
1261 | } |
1262 | |
1263 | setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom); |
1264 | setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom); |
1265 | setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom); |
1266 | setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom); |
1267 | setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom); |
1268 | setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom); |
1269 | setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom); |
1270 | setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom); |
1271 | setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom); |
1272 | setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom); |
1273 | |
1274 | for (auto VT : { MVT::v16f32, MVT::v8f64 }) { |
1275 | setOperationAction(ISD::FFLOOR, VT, Legal); |
1276 | setOperationAction(ISD::FCEIL, VT, Legal); |
1277 | setOperationAction(ISD::FTRUNC, VT, Legal); |
1278 | setOperationAction(ISD::FRINT, VT, Legal); |
1279 | setOperationAction(ISD::FNEARBYINT, VT, Legal); |
1280 | } |
1281 | |
1282 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i64, Custom); |
1283 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v16i32, Custom); |
1284 | |
1285 | // Without BWI we need to use custom lowering to handle MVT::v64i8 input. |
1286 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v64i8, Custom); |
1287 | setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, MVT::v64i8, Custom); |
1288 | |
1289 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom); |
1290 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom); |
1291 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom); |
1292 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom); |
1293 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Custom); |
1294 | |
1295 | setOperationAction(ISD::MUL, MVT::v8i64, Custom); |
1296 | |
1297 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v1i1, Custom); |
1298 | setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v16i1, Custom); |
1299 | setOperationAction(ISD::BUILD_VECTOR, MVT::v1i1, Custom); |
1300 | setOperationAction(ISD::SELECT, MVT::v8f64, Custom); |
1301 | setOperationAction(ISD::SELECT, MVT::v8i64, Custom); |
1302 | setOperationAction(ISD::SELECT, MVT::v16f32, Custom); |
1303 | |
1304 | setOperationAction(ISD::MUL, MVT::v16i32, Legal); |
1305 | |
1306 | // NonVLX sub-targets extend 128/256 vectors to use the 512 version. |
1307 | setOperationAction(ISD::ABS, MVT::v4i64, Legal); |
1308 | setOperationAction(ISD::ABS, MVT::v2i64, Legal); |
1309 | |
1310 | for (auto VT : { MVT::v8i1, MVT::v16i1 }) { |
1311 | setOperationAction(ISD::ADD, VT, Custom); |
1312 | setOperationAction(ISD::SUB, VT, Custom); |
1313 | setOperationAction(ISD::MUL, VT, Custom); |
1314 | setOperationAction(ISD::SETCC, VT, Custom); |
1315 | setOperationAction(ISD::SELECT, VT, Custom); |
1316 | setOperationAction(ISD::TRUNCATE, VT, Custom); |
1317 | |
1318 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); |
1319 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); |
1320 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); |
1321 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); |
1322 | setOperationAction(ISD::VSELECT, VT, Expand); |
1323 | } |
1324 | |
1325 | for (auto VT : { MVT::v16i32, MVT::v8i64 }) { |
1326 | setOperationAction(ISD::SMAX, VT, Legal); |
1327 | setOperationAction(ISD::UMAX, VT, Legal); |
1328 | setOperationAction(ISD::SMIN, VT, Legal); |
1329 | setOperationAction(ISD::UMIN, VT, Legal); |
1330 | setOperationAction(ISD::ABS, VT, Legal); |
1331 | setOperationAction(ISD::SRL, VT, Custom); |
1332 | setOperationAction(ISD::SHL, VT, Custom); |
1333 | setOperationAction(ISD::SRA, VT, Custom); |
1334 | setOperationAction(ISD::CTPOP, VT, Custom); |
1335 | setOperationAction(ISD::CTTZ, VT, Custom); |
1336 | } |
1337 | |
1338 | // Need to promote to 64-bit even though we have 32-bit masked instructions |
1339 | // because the IR optimizers rearrange bitcasts around logic ops leaving |
1340 | // too many variations to handle if we don't promote them. |
1341 | setOperationPromotedToType(ISD::AND, MVT::v16i32, MVT::v8i64); |
1342 | setOperationPromotedToType(ISD::OR, MVT::v16i32, MVT::v8i64); |
1343 | setOperationPromotedToType(ISD::XOR, MVT::v16i32, MVT::v8i64); |
1344 | |
1345 | if (Subtarget.hasCDI()) { |
1346 | // NonVLX sub-targets extend 128/256 vectors to use the 512 version. |
1347 | for (auto VT : {MVT::v4i32, MVT::v8i32, MVT::v16i32, MVT::v2i64, |
1348 | MVT::v4i64, MVT::v8i64}) { |
1349 | setOperationAction(ISD::CTLZ, VT, Legal); |
1350 | setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Custom); |
1351 | } |
1352 | } // Subtarget.hasCDI() |
1353 | |
1354 | if (Subtarget.hasDQI()) { |
1355 | // NonVLX sub-targets extend 128/256 vectors to use the 512 version. |
1356 | setOperationAction(ISD::MUL, MVT::v2i64, Legal); |
1357 | setOperationAction(ISD::MUL, MVT::v4i64, Legal); |
1358 | setOperationAction(ISD::MUL, MVT::v8i64, Legal); |
1359 | } |
1360 | |
1361 | if (Subtarget.hasVPOPCNTDQ()) { |
1362 | // VPOPCNTDQ sub-targets extend 128/256 vectors to use the avx512 |
1363 | // version of popcntd/q. |
1364 | for (auto VT : {MVT::v16i32, MVT::v8i64, MVT::v8i32, MVT::v4i64, |
1365 | MVT::v4i32, MVT::v2i64}) |
1366 | setOperationAction(ISD::CTPOP, VT, Legal); |
1367 | } |
1368 | |
1369 | // Custom lower several nodes. |
1370 | for (auto VT : { MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64, |
1371 | MVT::v4f32, MVT::v8f32, MVT::v2f64, MVT::v4f64 }) { |
1372 | setOperationAction(ISD::MGATHER, VT, Custom); |
1373 | setOperationAction(ISD::MSCATTER, VT, Custom); |
1374 | } |
1375 | // Extract subvector is special because the value type |
1376 | // (result) is 256-bit but the source is 512-bit wide. |
1377 | // 128-bit was made Custom under AVX1. |
1378 | for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64, |
1379 | MVT::v8f32, MVT::v4f64 }) |
1380 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom); |
1381 | for (auto VT : { MVT::v2i1, MVT::v4i1, MVT::v8i1, |
1382 | MVT::v16i1, MVT::v32i1, MVT::v64i1 }) |
1383 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal); |
1384 | |
1385 | for (auto VT : { MVT::v16i32, MVT::v8i64, MVT::v16f32, MVT::v8f64 }) { |
1386 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); |
1387 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); |
1388 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); |
1389 | setOperationAction(ISD::VSELECT, VT, Custom); |
1390 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); |
1391 | setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom); |
1392 | setOperationAction(ISD::INSERT_SUBVECTOR, VT, Legal); |
1393 | setOperationAction(ISD::MLOAD, VT, Legal); |
1394 | setOperationAction(ISD::MSTORE, VT, Legal); |
1395 | setOperationAction(ISD::MGATHER, VT, Legal); |
1396 | setOperationAction(ISD::MSCATTER, VT, Custom); |
1397 | } |
1398 | for (auto VT : { MVT::v64i8, MVT::v32i16, MVT::v16i32 }) { |
1399 | setOperationPromotedToType(ISD::LOAD, VT, MVT::v8i64); |
1400 | setOperationPromotedToType(ISD::SELECT, VT, MVT::v8i64); |
1401 | } |
1402 | }// has AVX-512 |
1403 | |
1404 | if (!Subtarget.useSoftFloat() && Subtarget.hasBWI()) { |
1405 | addRegisterClass(MVT::v32i16, &X86::VR512RegClass); |
1406 | addRegisterClass(MVT::v64i8, &X86::VR512RegClass); |
1407 | |
1408 | addRegisterClass(MVT::v32i1, &X86::VK32RegClass); |
1409 | addRegisterClass(MVT::v64i1, &X86::VK64RegClass); |
1410 | |
1411 | setOperationAction(ISD::ADD, MVT::v32i1, Custom); |
1412 | setOperationAction(ISD::ADD, MVT::v64i1, Custom); |
1413 | setOperationAction(ISD::SUB, MVT::v32i1, Custom); |
1414 | setOperationAction(ISD::SUB, MVT::v64i1, Custom); |
1415 | setOperationAction(ISD::MUL, MVT::v32i1, Custom); |
1416 | setOperationAction(ISD::MUL, MVT::v64i1, Custom); |
1417 | |
1418 | setOperationAction(ISD::SETCC, MVT::v32i1, Custom); |
1419 | setOperationAction(ISD::SETCC, MVT::v64i1, Custom); |
1420 | setOperationAction(ISD::MUL, MVT::v32i16, Legal); |
1421 | setOperationAction(ISD::MUL, MVT::v64i8, Custom); |
1422 | setOperationAction(ISD::MULHS, MVT::v32i16, Legal); |
1423 | setOperationAction(ISD::MULHU, MVT::v32i16, Legal); |
1424 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom); |
1425 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom); |
1426 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i16, Custom); |
1427 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i8, Custom); |
1428 | setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom); |
1429 | setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom); |
1430 | setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i16, Legal); |
1431 | setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i8, Legal); |
1432 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v32i16, Custom); |
1433 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v64i8, Custom); |
1434 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v32i1, Custom); |
1435 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v64i1, Custom); |
1436 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i16, Custom); |
1437 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v64i8, Custom); |
1438 | setOperationAction(ISD::SELECT, MVT::v32i1, Custom); |
1439 | setOperationAction(ISD::SELECT, MVT::v64i1, Custom); |
1440 | setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom); |
1441 | setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom); |
1442 | setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom); |
1443 | setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom); |
1444 | setOperationAction(ISD::ANY_EXTEND, MVT::v32i16, Custom); |
1445 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i16, Custom); |
1446 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i8, Custom); |
1447 | setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom); |
1448 | setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom); |
1449 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom); |
1450 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom); |
1451 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i16, Custom); |
1452 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i8, Custom); |
1453 | setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom); |
1454 | setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom); |
1455 | setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom); |
1456 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i1, Custom); |
1457 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i1, Custom); |
1458 | setOperationAction(ISD::BUILD_VECTOR, MVT::v32i1, Custom); |
1459 | setOperationAction(ISD::BUILD_VECTOR, MVT::v64i1, Custom); |
1460 | setOperationAction(ISD::VSELECT, MVT::v32i1, Expand); |
1461 | setOperationAction(ISD::VSELECT, MVT::v64i1, Expand); |
1462 | setOperationAction(ISD::BITREVERSE, MVT::v64i8, Custom); |
1463 | |
1464 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v32i16, Custom); |
1465 | |
1466 | setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal); |
1467 | if (Subtarget.hasVLX()) { |
1468 | setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal); |
1469 | setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal); |
1470 | } |
1471 | |
1472 | LegalizeAction Action = Subtarget.hasVLX() ? Legal : Custom; |
1473 | for (auto VT : { MVT::v32i8, MVT::v16i8, MVT::v16i16, MVT::v8i16 }) { |
1474 | setOperationAction(ISD::MLOAD, VT, Action); |
1475 | setOperationAction(ISD::MSTORE, VT, Action); |
1476 | } |
1477 | |
1478 | if (Subtarget.hasCDI()) { |
1479 | setOperationAction(ISD::CTLZ, MVT::v32i16, Custom); |
1480 | setOperationAction(ISD::CTLZ, MVT::v64i8, Custom); |
1481 | } |
1482 | |
1483 | for (auto VT : { MVT::v64i8, MVT::v32i16 }) { |
1484 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); |
1485 | setOperationAction(ISD::VSELECT, VT, Custom); |
1486 | setOperationAction(ISD::ABS, VT, Legal); |
1487 | setOperationAction(ISD::SRL, VT, Custom); |
1488 | setOperationAction(ISD::SHL, VT, Custom); |
1489 | setOperationAction(ISD::SRA, VT, Custom); |
1490 | setOperationAction(ISD::MLOAD, VT, Legal); |
1491 | setOperationAction(ISD::MSTORE, VT, Legal); |
1492 | setOperationAction(ISD::CTPOP, VT, Custom); |
1493 | setOperationAction(ISD::CTTZ, VT, Custom); |
1494 | setOperationAction(ISD::SMAX, VT, Legal); |
1495 | setOperationAction(ISD::UMAX, VT, Legal); |
1496 | setOperationAction(ISD::SMIN, VT, Legal); |
1497 | setOperationAction(ISD::UMIN, VT, Legal); |
1498 | |
1499 | setOperationPromotedToType(ISD::AND, VT, MVT::v8i64); |
1500 | setOperationPromotedToType(ISD::OR, VT, MVT::v8i64); |
1501 | setOperationPromotedToType(ISD::XOR, VT, MVT::v8i64); |
1502 | } |
1503 | |
1504 | for (auto ExtType : {ISD::ZEXTLOAD, ISD::SEXTLOAD, ISD::EXTLOAD}) { |
1505 | setLoadExtAction(ExtType, MVT::v32i16, MVT::v32i8, Legal); |
1506 | if (Subtarget.hasVLX()) { |
1507 | // FIXME. This commands are available on SSE/AVX2, add relevant patterns. |
1508 | setLoadExtAction(ExtType, MVT::v16i16, MVT::v16i8, Legal); |
1509 | setLoadExtAction(ExtType, MVT::v8i16, MVT::v8i8, Legal); |
1510 | } |
1511 | } |
1512 | } |
1513 | |
1514 | if (!Subtarget.useSoftFloat() && Subtarget.hasVLX()) { |
1515 | addRegisterClass(MVT::v4i1, &X86::VK4RegClass); |
1516 | addRegisterClass(MVT::v2i1, &X86::VK2RegClass); |
1517 | |
1518 | for (auto VT : { MVT::v2i1, MVT::v4i1 }) { |
1519 | setOperationAction(ISD::ADD, VT, Custom); |
1520 | setOperationAction(ISD::SUB, VT, Custom); |
1521 | setOperationAction(ISD::MUL, VT, Custom); |
1522 | setOperationAction(ISD::VSELECT, VT, Expand); |
1523 | |
1524 | setOperationAction(ISD::TRUNCATE, VT, Custom); |
1525 | setOperationAction(ISD::SETCC, VT, Custom); |
1526 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); |
1527 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); |
1528 | setOperationAction(ISD::SELECT, VT, Custom); |
1529 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); |
1530 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); |
1531 | } |
1532 | |
1533 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom); |
1534 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom); |
1535 | setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom); |
1536 | setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom); |
1537 | |
1538 | for (auto VT : { MVT::v2i64, MVT::v4i64 }) { |
1539 | setOperationAction(ISD::SMAX, VT, Legal); |
1540 | setOperationAction(ISD::UMAX, VT, Legal); |
1541 | setOperationAction(ISD::SMIN, VT, Legal); |
1542 | setOperationAction(ISD::UMIN, VT, Legal); |
1543 | } |
1544 | } |
1545 | |
1546 | // We want to custom lower some of our intrinsics. |
1547 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); |
1548 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom); |
1549 | setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom); |
1550 | if (!Subtarget.is64Bit()) { |
1551 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom); |
1552 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom); |
1553 | } |
1554 | |
1555 | // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't |
1556 | // handle type legalization for these operations here. |
1557 | // |
1558 | // FIXME: We really should do custom legalization for addition and |
1559 | // subtraction on x86-32 once PR3203 is fixed. We really can't do much better |
1560 | // than generic legalization for 64-bit multiplication-with-overflow, though. |
1561 | for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) { |
1562 | if (VT == MVT::i64 && !Subtarget.is64Bit()) |
1563 | continue; |
1564 | // Add/Sub/Mul with overflow operations are custom lowered. |
1565 | setOperationAction(ISD::SADDO, VT, Custom); |
1566 | setOperationAction(ISD::UADDO, VT, Custom); |
1567 | setOperationAction(ISD::SSUBO, VT, Custom); |
1568 | setOperationAction(ISD::USUBO, VT, Custom); |
1569 | setOperationAction(ISD::SMULO, VT, Custom); |
1570 | setOperationAction(ISD::UMULO, VT, Custom); |
1571 | |
1572 | // Support carry in as value rather than glue. |
1573 | setOperationAction(ISD::ADDCARRY, VT, Custom); |
1574 | setOperationAction(ISD::SUBCARRY, VT, Custom); |
1575 | setOperationAction(ISD::SETCCCARRY, VT, Custom); |
1576 | } |
1577 | |
1578 | if (!Subtarget.is64Bit()) { |
1579 | // These libcalls are not available in 32-bit. |
1580 | setLibcallName(RTLIB::SHL_I128, nullptr); |
1581 | setLibcallName(RTLIB::SRL_I128, nullptr); |
1582 | setLibcallName(RTLIB::SRA_I128, nullptr); |
1583 | } |
1584 | |
1585 | // Combine sin / cos into one node or libcall if possible. |
1586 | if (Subtarget.hasSinCos()) { |
1587 | setLibcallName(RTLIB::SINCOS_F32, "sincosf"); |
1588 | setLibcallName(RTLIB::SINCOS_F64, "sincos"); |
1589 | if (Subtarget.isTargetDarwin()) { |
1590 | // For MacOSX, we don't want the normal expansion of a libcall to sincos. |
1591 | // We want to issue a libcall to __sincos_stret to avoid memory traffic. |
1592 | setOperationAction(ISD::FSINCOS, MVT::f64, Custom); |
1593 | setOperationAction(ISD::FSINCOS, MVT::f32, Custom); |
1594 | } |
1595 | } |
1596 | |
1597 | if (Subtarget.isTargetWin64()) { |
1598 | setOperationAction(ISD::SDIV, MVT::i128, Custom); |
1599 | setOperationAction(ISD::UDIV, MVT::i128, Custom); |
1600 | setOperationAction(ISD::SREM, MVT::i128, Custom); |
1601 | setOperationAction(ISD::UREM, MVT::i128, Custom); |
1602 | setOperationAction(ISD::SDIVREM, MVT::i128, Custom); |
1603 | setOperationAction(ISD::UDIVREM, MVT::i128, Custom); |
1604 | } |
1605 | |
1606 | // On 32 bit MSVC, `fmodf(f32)` is not defined - only `fmod(f64)` |
1607 | // is. We should promote the value to 64-bits to solve this. |
1608 | // This is what the CRT headers do - `fmodf` is an inline header |
1609 | // function casting to f64 and calling `fmod`. |
1610 | if (Subtarget.is32Bit() && (Subtarget.isTargetKnownWindowsMSVC() || |
1611 | Subtarget.isTargetWindowsItanium())) |
1612 | for (ISD::NodeType Op : |
1613 | {ISD::FCEIL, ISD::FCOS, ISD::FEXP, ISD::FFLOOR, ISD::FREM, ISD::FLOG, |
1614 | ISD::FLOG10, ISD::FPOW, ISD::FSIN}) |
1615 | if (isOperationExpand(Op, MVT::f32)) |
1616 | setOperationAction(Op, MVT::f32, Promote); |
1617 | |
1618 | // We have target-specific dag combine patterns for the following nodes: |
1619 | setTargetDAGCombine(ISD::VECTOR_SHUFFLE); |
1620 | setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT); |
1621 | setTargetDAGCombine(ISD::INSERT_SUBVECTOR); |
1622 | setTargetDAGCombine(ISD::BITCAST); |
1623 | setTargetDAGCombine(ISD::VSELECT); |
1624 | setTargetDAGCombine(ISD::SELECT); |
1625 | setTargetDAGCombine(ISD::SHL); |
1626 | setTargetDAGCombine(ISD::SRA); |
1627 | setTargetDAGCombine(ISD::SRL); |
1628 | setTargetDAGCombine(ISD::OR); |
1629 | setTargetDAGCombine(ISD::AND); |
1630 | setTargetDAGCombine(ISD::ADD); |
1631 | setTargetDAGCombine(ISD::FADD); |
1632 | setTargetDAGCombine(ISD::FSUB); |
1633 | setTargetDAGCombine(ISD::FNEG); |
1634 | setTargetDAGCombine(ISD::FMA); |
1635 | setTargetDAGCombine(ISD::FMINNUM); |
1636 | setTargetDAGCombine(ISD::FMAXNUM); |
1637 | setTargetDAGCombine(ISD::SUB); |
1638 | setTargetDAGCombine(ISD::LOAD); |
1639 | setTargetDAGCombine(ISD::MLOAD); |
1640 | setTargetDAGCombine(ISD::STORE); |
1641 | setTargetDAGCombine(ISD::MSTORE); |
1642 | setTargetDAGCombine(ISD::TRUNCATE); |
1643 | setTargetDAGCombine(ISD::ZERO_EXTEND); |
1644 | setTargetDAGCombine(ISD::ANY_EXTEND); |
1645 | setTargetDAGCombine(ISD::SIGN_EXTEND); |
1646 | setTargetDAGCombine(ISD::SIGN_EXTEND_INREG); |
1647 | setTargetDAGCombine(ISD::SIGN_EXTEND_VECTOR_INREG); |
1648 | setTargetDAGCombine(ISD::ZERO_EXTEND_VECTOR_INREG); |
1649 | setTargetDAGCombine(ISD::SINT_TO_FP); |
1650 | setTargetDAGCombine(ISD::UINT_TO_FP); |
1651 | setTargetDAGCombine(ISD::SETCC); |
1652 | setTargetDAGCombine(ISD::MUL); |
1653 | setTargetDAGCombine(ISD::XOR); |
1654 | setTargetDAGCombine(ISD::MSCATTER); |
1655 | setTargetDAGCombine(ISD::MGATHER); |
1656 | |
1657 | computeRegisterProperties(Subtarget.getRegisterInfo()); |
1658 | |
1659 | MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores |
1660 | MaxStoresPerMemsetOptSize = 8; |
1661 | MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores |
1662 | MaxStoresPerMemcpyOptSize = 4; |
1663 | MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores |
1664 | MaxStoresPerMemmoveOptSize = 4; |
1665 | |
1666 | // TODO: These control memcmp expansion in CGP and are set low to prevent |
1667 | // altering the vector expansion for 16/32 byte memcmp in SelectionDAGBuilder. |
1668 | MaxLoadsPerMemcmp = 1; |
1669 | MaxLoadsPerMemcmpOptSize = 1; |
1670 | |
1671 | // Set loop alignment to 2^ExperimentalPrefLoopAlignment bytes (default: 2^4). |
1672 | setPrefLoopAlignment(ExperimentalPrefLoopAlignment); |
1673 | |
1674 | // An out-of-order CPU can speculatively execute past a predictable branch, |
1675 | // but a conditional move could be stalled by an expensive earlier operation. |
1676 | PredictableSelectIsExpensive = Subtarget.getSchedModel().isOutOfOrder(); |
1677 | EnableExtLdPromotion = true; |
1678 | setPrefFunctionAlignment(4); // 2^4 bytes. |
1679 | |
1680 | verifyIntrinsicTables(); |
1681 | } |
1682 | |
1683 | // This has so far only been implemented for 64-bit MachO. |
1684 | bool X86TargetLowering::useLoadStackGuardNode() const { |
1685 | return Subtarget.isTargetMachO() && Subtarget.is64Bit(); |
1686 | } |
1687 | |
1688 | TargetLoweringBase::LegalizeTypeAction |
1689 | X86TargetLowering::getPreferredVectorAction(EVT VT) const { |
1690 | if (ExperimentalVectorWideningLegalization && |
1691 | VT.getVectorNumElements() != 1 && |
1692 | VT.getVectorElementType().getSimpleVT() != MVT::i1) |
1693 | return TypeWidenVector; |
1694 | |
1695 | return TargetLoweringBase::getPreferredVectorAction(VT); |
1696 | } |
1697 | |
1698 | EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, |
1699 | LLVMContext& Context, |
1700 | EVT VT) const { |
1701 | if (!VT.isVector()) |
1702 | return MVT::i8; |
1703 | |
1704 | if (VT.isSimple()) { |
1705 | MVT VVT = VT.getSimpleVT(); |
1706 | const unsigned NumElts = VVT.getVectorNumElements(); |
1707 | MVT EltVT = VVT.getVectorElementType(); |
1708 | if (VVT.is512BitVector()) { |
1709 | if (Subtarget.hasAVX512()) |
1710 | if (EltVT == MVT::i32 || EltVT == MVT::i64 || |
1711 | EltVT == MVT::f32 || EltVT == MVT::f64) |
1712 | switch(NumElts) { |
1713 | case 8: return MVT::v8i1; |
1714 | case 16: return MVT::v16i1; |
1715 | } |
1716 | if (Subtarget.hasBWI()) |
1717 | if (EltVT == MVT::i8 || EltVT == MVT::i16) |
1718 | switch(NumElts) { |
1719 | case 32: return MVT::v32i1; |
1720 | case 64: return MVT::v64i1; |
1721 | } |
1722 | } |
1723 | |
1724 | if (Subtarget.hasBWI() && Subtarget.hasVLX()) |
1725 | return MVT::getVectorVT(MVT::i1, NumElts); |
1726 | |
1727 | if (!isTypeLegal(VT) && getTypeAction(Context, VT) == TypePromoteInteger) { |
1728 | EVT LegalVT = getTypeToTransformTo(Context, VT); |
1729 | EltVT = LegalVT.getVectorElementType().getSimpleVT(); |
1730 | } |
1731 | |
1732 | if (Subtarget.hasVLX() && EltVT.getSizeInBits() >= 32) |
1733 | switch(NumElts) { |
1734 | case 2: return MVT::v2i1; |
1735 | case 4: return MVT::v4i1; |
1736 | case 8: return MVT::v8i1; |
1737 | } |
1738 | } |
1739 | |
1740 | return VT.changeVectorElementTypeToInteger(); |
1741 | } |
1742 | |
1743 | /// Helper for getByValTypeAlignment to determine |
1744 | /// the desired ByVal argument alignment. |
1745 | static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) { |
1746 | if (MaxAlign == 16) |
1747 | return; |
1748 | if (VectorType *VTy = dyn_cast<VectorType>(Ty)) { |
1749 | if (VTy->getBitWidth() == 128) |
1750 | MaxAlign = 16; |
1751 | } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { |
1752 | unsigned EltAlign = 0; |
1753 | getMaxByValAlign(ATy->getElementType(), EltAlign); |
1754 | if (EltAlign > MaxAlign) |
1755 | MaxAlign = EltAlign; |
1756 | } else if (StructType *STy = dyn_cast<StructType>(Ty)) { |
1757 | for (auto *EltTy : STy->elements()) { |
1758 | unsigned EltAlign = 0; |
1759 | getMaxByValAlign(EltTy, EltAlign); |
1760 | if (EltAlign > MaxAlign) |
1761 | MaxAlign = EltAlign; |
1762 | if (MaxAlign == 16) |
1763 | break; |
1764 | } |
1765 | } |
1766 | } |
1767 | |
1768 | /// Return the desired alignment for ByVal aggregate |
1769 | /// function arguments in the caller parameter area. For X86, aggregates |
1770 | /// that contain SSE vectors are placed at 16-byte boundaries while the rest |
1771 | /// are at 4-byte boundaries. |
1772 | unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty, |
1773 | const DataLayout &DL) const { |
1774 | if (Subtarget.is64Bit()) { |
1775 | // Max of 8 and alignment of type. |
1776 | unsigned TyAlign = DL.getABITypeAlignment(Ty); |
1777 | if (TyAlign > 8) |
1778 | return TyAlign; |
1779 | return 8; |
1780 | } |
1781 | |
1782 | unsigned Align = 4; |
1783 | if (Subtarget.hasSSE1()) |
1784 | getMaxByValAlign(Ty, Align); |
1785 | return Align; |
1786 | } |
1787 | |
1788 | /// Returns the target specific optimal type for load |
1789 | /// and store operations as a result of memset, memcpy, and memmove |
1790 | /// lowering. If DstAlign is zero that means it's safe to destination |
1791 | /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it |
1792 | /// means there isn't a need to check it against alignment requirement, |
1793 | /// probably because the source does not need to be loaded. If 'IsMemset' is |
1794 | /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that |
1795 | /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy |
1796 | /// source is constant so it does not need to be loaded. |
1797 | /// It returns EVT::Other if the type should be determined using generic |
1798 | /// target-independent logic. |
1799 | EVT |
1800 | X86TargetLowering::getOptimalMemOpType(uint64_t Size, |
1801 | unsigned DstAlign, unsigned SrcAlign, |
1802 | bool IsMemset, bool ZeroMemset, |
1803 | bool MemcpyStrSrc, |
1804 | MachineFunction &MF) const { |
1805 | const Function *F = MF.getFunction(); |
1806 | if (!F->hasFnAttribute(Attribute::NoImplicitFloat)) { |
1807 | if (Size >= 16 && |
1808 | (!Subtarget.isUnalignedMem16Slow() || |
1809 | ((DstAlign == 0 || DstAlign >= 16) && |
1810 | (SrcAlign == 0 || SrcAlign >= 16)))) { |
1811 | // FIXME: Check if unaligned 32-byte accesses are slow. |
1812 | if (Size >= 32 && Subtarget.hasAVX()) { |
1813 | // Although this isn't a well-supported type for AVX1, we'll let |
1814 | // legalization and shuffle lowering produce the optimal codegen. If we |
1815 | // choose an optimal type with a vector element larger than a byte, |
1816 | // getMemsetStores() may create an intermediate splat (using an integer |
1817 | // multiply) before we splat as a vector. |
1818 | return MVT::v32i8; |
1819 | } |
1820 | if (Subtarget.hasSSE2()) |
1821 | return MVT::v16i8; |
1822 | // TODO: Can SSE1 handle a byte vector? |
1823 | if (Subtarget.hasSSE1()) |
1824 | return MVT::v4f32; |
1825 | } else if ((!IsMemset || ZeroMemset) && !MemcpyStrSrc && Size >= 8 && |
1826 | !Subtarget.is64Bit() && Subtarget.hasSSE2()) { |
1827 | // Do not use f64 to lower memcpy if source is string constant. It's |
1828 | // better to use i32 to avoid the loads. |
1829 | // Also, do not use f64 to lower memset unless this is a memset of zeros. |
1830 | // The gymnastics of splatting a byte value into an XMM register and then |
1831 | // only using 8-byte stores (because this is a CPU with slow unaligned |
1832 | // 16-byte accesses) makes that a loser. |
1833 | return MVT::f64; |
1834 | } |
1835 | } |
1836 | // This is a compromise. If we reach here, unaligned accesses may be slow on |
1837 | // this target. However, creating smaller, aligned accesses could be even |
1838 | // slower and would certainly be a lot more code. |
1839 | if (Subtarget.is64Bit() && Size >= 8) |
1840 | return MVT::i64; |
1841 | return MVT::i32; |
1842 | } |
1843 | |
1844 | bool X86TargetLowering::isSafeMemOpType(MVT VT) const { |
1845 | if (VT == MVT::f32) |
1846 | return X86ScalarSSEf32; |
1847 | else if (VT == MVT::f64) |
1848 | return X86ScalarSSEf64; |
1849 | return true; |
1850 | } |
1851 | |
1852 | bool |
1853 | X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT, |
1854 | unsigned, |
1855 | unsigned, |
1856 | bool *Fast) const { |
1857 | if (Fast) { |
1858 | switch (VT.getSizeInBits()) { |
1859 | default: |
1860 | // 8-byte and under are always assumed to be fast. |
1861 | *Fast = true; |
1862 | break; |
1863 | case 128: |
1864 | *Fast = !Subtarget.isUnalignedMem16Slow(); |
1865 | break; |
1866 | case 256: |
1867 | *Fast = !Subtarget.isUnalignedMem32Slow(); |
1868 | break; |
1869 | // TODO: What about AVX-512 (512-bit) accesses? |
1870 | } |
1871 | } |
1872 | // Misaligned accesses of any size are always allowed. |
1873 | return true; |
1874 | } |
1875 | |
1876 | /// Return the entry encoding for a jump table in the |
1877 | /// current function. The returned value is a member of the |
1878 | /// MachineJumpTableInfo::JTEntryKind enum. |
1879 | unsigned X86TargetLowering::getJumpTableEncoding() const { |
1880 | // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF |
1881 | // symbol. |
1882 | if (isPositionIndependent() && Subtarget.isPICStyleGOT()) |
1883 | return MachineJumpTableInfo::EK_Custom32; |
1884 | |
1885 | // Otherwise, use the normal jump table encoding heuristics. |
1886 | return TargetLowering::getJumpTableEncoding(); |
1887 | } |
1888 | |
1889 | bool X86TargetLowering::useSoftFloat() const { |
1890 | return Subtarget.useSoftFloat(); |
1891 | } |
1892 | |
1893 | void X86TargetLowering::markLibCallAttributes(MachineFunction *MF, unsigned CC, |
1894 | ArgListTy &Args) const { |
1895 | |
1896 | // Only relabel X86-32 for C / Stdcall CCs. |
1897 | if (Subtarget.is64Bit()) |
1898 | return; |
1899 | if (CC != CallingConv::C && CC != CallingConv::X86_StdCall) |
1900 | return; |
1901 | unsigned ParamRegs = 0; |
1902 | if (auto *M = MF->getFunction()->getParent()) |
1903 | ParamRegs = M->getNumberRegisterParameters(); |
1904 | |
1905 | // Mark the first N int arguments as having reg |
1906 | for (unsigned Idx = 0; Idx < Args.size(); Idx++) { |
1907 | Type *T = Args[Idx].Ty; |
1908 | if (T->isPointerTy() || T->isIntegerTy()) |
1909 | if (MF->getDataLayout().getTypeAllocSize(T) <= 8) { |
1910 | unsigned numRegs = 1; |
1911 | if (MF->getDataLayout().getTypeAllocSize(T) > 4) |
1912 | numRegs = 2; |
1913 | if (ParamRegs < numRegs) |
1914 | return; |
1915 | ParamRegs -= numRegs; |
1916 | Args[Idx].IsInReg = true; |
1917 | } |
1918 | } |
1919 | } |
1920 | |
1921 | const MCExpr * |
1922 | X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI, |
1923 | const MachineBasicBlock *MBB, |
1924 | unsigned uid,MCContext &Ctx) const{ |
1925 | assert(isPositionIndependent() && Subtarget.isPICStyleGOT())((isPositionIndependent() && Subtarget.isPICStyleGOT( )) ? static_cast<void> (0) : __assert_fail ("isPositionIndependent() && Subtarget.isPICStyleGOT()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 1925, __PRETTY_FUNCTION__)); |
1926 | // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF |
1927 | // entries. |
1928 | return MCSymbolRefExpr::create(MBB->getSymbol(), |
1929 | MCSymbolRefExpr::VK_GOTOFF, Ctx); |
1930 | } |
1931 | |
1932 | /// Returns relocation base for the given PIC jumptable. |
1933 | SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table, |
1934 | SelectionDAG &DAG) const { |
1935 | if (!Subtarget.is64Bit()) |
1936 | // This doesn't have SDLoc associated with it, but is not really the |
1937 | // same as a Register. |
1938 | return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), |
1939 | getPointerTy(DAG.getDataLayout())); |
1940 | return Table; |
1941 | } |
1942 | |
1943 | /// This returns the relocation base for the given PIC jumptable, |
1944 | /// the same as getPICJumpTableRelocBase, but as an MCExpr. |
1945 | const MCExpr *X86TargetLowering:: |
1946 | getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI, |
1947 | MCContext &Ctx) const { |
1948 | // X86-64 uses RIP relative addressing based on the jump table label. |
1949 | if (Subtarget.isPICStyleRIPRel()) |
1950 | return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx); |
1951 | |
1952 | // Otherwise, the reference is relative to the PIC base. |
1953 | return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx); |
1954 | } |
1955 | |
1956 | std::pair<const TargetRegisterClass *, uint8_t> |
1957 | X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI, |
1958 | MVT VT) const { |
1959 | const TargetRegisterClass *RRC = nullptr; |
1960 | uint8_t Cost = 1; |
1961 | switch (VT.SimpleTy) { |
1962 | default: |
1963 | return TargetLowering::findRepresentativeClass(TRI, VT); |
1964 | case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64: |
1965 | RRC = Subtarget.is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass; |
1966 | break; |
1967 | case MVT::x86mmx: |
1968 | RRC = &X86::VR64RegClass; |
1969 | break; |
1970 | case MVT::f32: case MVT::f64: |
1971 | case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64: |
1972 | case MVT::v4f32: case MVT::v2f64: |
1973 | case MVT::v32i8: case MVT::v16i16: case MVT::v8i32: case MVT::v4i64: |
1974 | case MVT::v8f32: case MVT::v4f64: |
1975 | case MVT::v64i8: case MVT::v32i16: case MVT::v16i32: case MVT::v8i64: |
1976 | case MVT::v16f32: case MVT::v8f64: |
1977 | RRC = &X86::VR128XRegClass; |
1978 | break; |
1979 | } |
1980 | return std::make_pair(RRC, Cost); |
1981 | } |
1982 | |
1983 | unsigned X86TargetLowering::getAddressSpace() const { |
1984 | if (Subtarget.is64Bit()) |
1985 | return (getTargetMachine().getCodeModel() == CodeModel::Kernel) ? 256 : 257; |
1986 | return 256; |
1987 | } |
1988 | |
1989 | static bool hasStackGuardSlotTLS(const Triple &TargetTriple) { |
1990 | return TargetTriple.isOSGlibc() || TargetTriple.isOSFuchsia() || |
1991 | (TargetTriple.isAndroid() && !TargetTriple.isAndroidVersionLT(17)); |
1992 | } |
1993 | |
1994 | static Constant* SegmentOffset(IRBuilder<> &IRB, |
1995 | unsigned Offset, unsigned AddressSpace) { |
1996 | return ConstantExpr::getIntToPtr( |
1997 | ConstantInt::get(Type::getInt32Ty(IRB.getContext()), Offset), |
1998 | Type::getInt8PtrTy(IRB.getContext())->getPointerTo(AddressSpace)); |
1999 | } |
2000 | |
2001 | Value *X86TargetLowering::getIRStackGuard(IRBuilder<> &IRB) const { |
2002 | // glibc, bionic, and Fuchsia have a special slot for the stack guard in |
2003 | // tcbhead_t; use it instead of the usual global variable (see |
2004 | // sysdeps/{i386,x86_64}/nptl/tls.h) |
2005 | if (hasStackGuardSlotTLS(Subtarget.getTargetTriple())) { |
2006 | if (Subtarget.isTargetFuchsia()) { |
2007 | // <magenta/tls.h> defines MX_TLS_STACK_GUARD_OFFSET with this value. |
2008 | return SegmentOffset(IRB, 0x10, getAddressSpace()); |
2009 | } else { |
2010 | // %fs:0x28, unless we're using a Kernel code model, in which case |
2011 | // it's %gs:0x28. gs:0x14 on i386. |
2012 | unsigned Offset = (Subtarget.is64Bit()) ? 0x28 : 0x14; |
2013 | return SegmentOffset(IRB, Offset, getAddressSpace()); |
2014 | } |
2015 | } |
2016 | |
2017 | return TargetLowering::getIRStackGuard(IRB); |
2018 | } |
2019 | |
2020 | void X86TargetLowering::insertSSPDeclarations(Module &M) const { |
2021 | // MSVC CRT provides functionalities for stack protection. |
2022 | if (Subtarget.getTargetTriple().isOSMSVCRT()) { |
2023 | // MSVC CRT has a global variable holding security cookie. |
2024 | M.getOrInsertGlobal("__security_cookie", |
2025 | Type::getInt8PtrTy(M.getContext())); |
2026 | |
2027 | // MSVC CRT has a function to validate security cookie. |
2028 | auto *SecurityCheckCookie = cast<Function>( |
2029 | M.getOrInsertFunction("__security_check_cookie", |
2030 | Type::getVoidTy(M.getContext()), |
2031 | Type::getInt8PtrTy(M.getContext()))); |
2032 | SecurityCheckCookie->setCallingConv(CallingConv::X86_FastCall); |
2033 | SecurityCheckCookie->addAttribute(1, Attribute::AttrKind::InReg); |
2034 | return; |
2035 | } |
2036 | // glibc, bionic, and Fuchsia have a special slot for the stack guard. |
2037 | if (hasStackGuardSlotTLS(Subtarget.getTargetTriple())) |
2038 | return; |
2039 | TargetLowering::insertSSPDeclarations(M); |
2040 | } |
2041 | |
2042 | Value *X86TargetLowering::getSDagStackGuard(const Module &M) const { |
2043 | // MSVC CRT has a global variable holding security cookie. |
2044 | if (Subtarget.getTargetTriple().isOSMSVCRT()) |
2045 | return M.getGlobalVariable("__security_cookie"); |
2046 | return TargetLowering::getSDagStackGuard(M); |
2047 | } |
2048 | |
2049 | Value *X86TargetLowering::getSSPStackGuardCheck(const Module &M) const { |
2050 | // MSVC CRT has a function to validate security cookie. |
2051 | if (Subtarget.getTargetTriple().isOSMSVCRT()) |
2052 | return M.getFunction("__security_check_cookie"); |
2053 | return TargetLowering::getSSPStackGuardCheck(M); |
2054 | } |
2055 | |
2056 | Value *X86TargetLowering::getSafeStackPointerLocation(IRBuilder<> &IRB) const { |
2057 | if (Subtarget.getTargetTriple().isOSContiki()) |
2058 | return getDefaultSafeStackPointerLocation(IRB, false); |
2059 | |
2060 | // Android provides a fixed TLS slot for the SafeStack pointer. See the |
2061 | // definition of TLS_SLOT_SAFESTACK in |
2062 | // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h |
2063 | if (Subtarget.isTargetAndroid()) { |
2064 | // %fs:0x48, unless we're using a Kernel code model, in which case it's %gs: |
2065 | // %gs:0x24 on i386 |
2066 | unsigned Offset = (Subtarget.is64Bit()) ? 0x48 : 0x24; |
2067 | return SegmentOffset(IRB, Offset, getAddressSpace()); |
2068 | } |
2069 | |
2070 | // Fuchsia is similar. |
2071 | if (Subtarget.isTargetFuchsia()) { |
2072 | // <magenta/tls.h> defines MX_TLS_UNSAFE_SP_OFFSET with this value. |
2073 | return SegmentOffset(IRB, 0x18, getAddressSpace()); |
2074 | } |
2075 | |
2076 | return TargetLowering::getSafeStackPointerLocation(IRB); |
2077 | } |
2078 | |
2079 | bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS, |
2080 | unsigned DestAS) const { |
2081 | assert(SrcAS != DestAS && "Expected different address spaces!")((SrcAS != DestAS && "Expected different address spaces!" ) ? static_cast<void> (0) : __assert_fail ("SrcAS != DestAS && \"Expected different address spaces!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2081, __PRETTY_FUNCTION__)); |
2082 | |
2083 | return SrcAS < 256 && DestAS < 256; |
2084 | } |
2085 | |
2086 | //===----------------------------------------------------------------------===// |
2087 | // Return Value Calling Convention Implementation |
2088 | //===----------------------------------------------------------------------===// |
2089 | |
2090 | #include "X86GenCallingConv.inc" |
2091 | |
2092 | bool X86TargetLowering::CanLowerReturn( |
2093 | CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg, |
2094 | const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const { |
2095 | SmallVector<CCValAssign, 16> RVLocs; |
2096 | CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context); |
2097 | return CCInfo.CheckReturn(Outs, RetCC_X86); |
2098 | } |
2099 | |
2100 | const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const { |
2101 | static const MCPhysReg ScratchRegs[] = { X86::R11, 0 }; |
2102 | return ScratchRegs; |
2103 | } |
2104 | |
2105 | /// Lowers masks values (v*i1) to the local register values |
2106 | /// \returns DAG node after lowering to register type |
2107 | static SDValue lowerMasksToReg(const SDValue &ValArg, const EVT &ValLoc, |
2108 | const SDLoc &Dl, SelectionDAG &DAG) { |
2109 | EVT ValVT = ValArg.getValueType(); |
2110 | |
2111 | if ((ValVT == MVT::v8i1 && (ValLoc == MVT::i8 || ValLoc == MVT::i32)) || |
2112 | (ValVT == MVT::v16i1 && (ValLoc == MVT::i16 || ValLoc == MVT::i32))) { |
2113 | // Two stage lowering might be required |
2114 | // bitcast: v8i1 -> i8 / v16i1 -> i16 |
2115 | // anyextend: i8 -> i32 / i16 -> i32 |
2116 | EVT TempValLoc = ValVT == MVT::v8i1 ? MVT::i8 : MVT::i16; |
2117 | SDValue ValToCopy = DAG.getBitcast(TempValLoc, ValArg); |
2118 | if (ValLoc == MVT::i32) |
2119 | ValToCopy = DAG.getNode(ISD::ANY_EXTEND, Dl, ValLoc, ValToCopy); |
2120 | return ValToCopy; |
2121 | } else if ((ValVT == MVT::v32i1 && ValLoc == MVT::i32) || |
2122 | (ValVT == MVT::v64i1 && ValLoc == MVT::i64)) { |
2123 | // One stage lowering is required |
2124 | // bitcast: v32i1 -> i32 / v64i1 -> i64 |
2125 | return DAG.getBitcast(ValLoc, ValArg); |
2126 | } else |
2127 | return DAG.getNode(ISD::SIGN_EXTEND, Dl, ValLoc, ValArg); |
2128 | } |
2129 | |
2130 | /// Breaks v64i1 value into two registers and adds the new node to the DAG |
2131 | static void Passv64i1ArgInRegs( |
2132 | const SDLoc &Dl, SelectionDAG &DAG, SDValue Chain, SDValue &Arg, |
2133 | SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass, CCValAssign &VA, |
2134 | CCValAssign &NextVA, const X86Subtarget &Subtarget) { |
2135 | assert((Subtarget.hasBWI() || Subtarget.hasBMI()) &&(((Subtarget.hasBWI() || Subtarget.hasBMI()) && "Expected AVX512BW or AVX512BMI target!" ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasBWI() || Subtarget.hasBMI()) && \"Expected AVX512BW or AVX512BMI target!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2136, __PRETTY_FUNCTION__)) |
2136 | "Expected AVX512BW or AVX512BMI target!")(((Subtarget.hasBWI() || Subtarget.hasBMI()) && "Expected AVX512BW or AVX512BMI target!" ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasBWI() || Subtarget.hasBMI()) && \"Expected AVX512BW or AVX512BMI target!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2136, __PRETTY_FUNCTION__)); |
2137 | assert(Subtarget.is32Bit() && "Expecting 32 bit target")((Subtarget.is32Bit() && "Expecting 32 bit target") ? static_cast<void> (0) : __assert_fail ("Subtarget.is32Bit() && \"Expecting 32 bit target\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2137, __PRETTY_FUNCTION__)); |
2138 | assert(Arg.getValueType() == MVT::i64 && "Expecting 64 bit value")((Arg.getValueType() == MVT::i64 && "Expecting 64 bit value" ) ? static_cast<void> (0) : __assert_fail ("Arg.getValueType() == MVT::i64 && \"Expecting 64 bit value\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2138, __PRETTY_FUNCTION__)); |
2139 | assert(VA.isRegLoc() && NextVA.isRegLoc() &&((VA.isRegLoc() && NextVA.isRegLoc() && "The value should reside in two registers" ) ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && NextVA.isRegLoc() && \"The value should reside in two registers\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2140, __PRETTY_FUNCTION__)) |
2140 | "The value should reside in two registers")((VA.isRegLoc() && NextVA.isRegLoc() && "The value should reside in two registers" ) ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && NextVA.isRegLoc() && \"The value should reside in two registers\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2140, __PRETTY_FUNCTION__)); |
2141 | |
2142 | // Before splitting the value we cast it to i64 |
2143 | Arg = DAG.getBitcast(MVT::i64, Arg); |
2144 | |
2145 | // Splitting the value into two i32 types |
2146 | SDValue Lo, Hi; |
2147 | Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, Dl, MVT::i32, Arg, |
2148 | DAG.getConstant(0, Dl, MVT::i32)); |
2149 | Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, Dl, MVT::i32, Arg, |
2150 | DAG.getConstant(1, Dl, MVT::i32)); |
2151 | |
2152 | // Attach the two i32 types into corresponding registers |
2153 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), Lo)); |
2154 | RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), Hi)); |
2155 | } |
2156 | |
2157 | SDValue |
2158 | X86TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, |
2159 | bool isVarArg, |
2160 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
2161 | const SmallVectorImpl<SDValue> &OutVals, |
2162 | const SDLoc &dl, SelectionDAG &DAG) const { |
2163 | MachineFunction &MF = DAG.getMachineFunction(); |
2164 | X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); |
2165 | |
2166 | // In some cases we need to disable registers from the default CSR list. |
2167 | // For example, when they are used for argument passing. |
2168 | bool ShouldDisableCalleeSavedRegister = |
2169 | CallConv == CallingConv::X86_RegCall || |
2170 | MF.getFunction()->hasFnAttribute("no_caller_saved_registers"); |
2171 | |
2172 | if (CallConv == CallingConv::X86_INTR && !Outs.empty()) |
2173 | report_fatal_error("X86 interrupts may not return any value"); |
2174 | |
2175 | SmallVector<CCValAssign, 16> RVLocs; |
2176 | CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext()); |
2177 | CCInfo.AnalyzeReturn(Outs, RetCC_X86); |
2178 | |
2179 | SDValue Flag; |
2180 | SmallVector<SDValue, 6> RetOps; |
2181 | RetOps.push_back(Chain); // Operand #0 = Chain (updated below) |
2182 | // Operand #1 = Bytes To Pop |
2183 | RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl, |
2184 | MVT::i32)); |
2185 | |
2186 | // Copy the result values into the output registers. |
2187 | for (unsigned I = 0, OutsIndex = 0, E = RVLocs.size(); I != E; |
2188 | ++I, ++OutsIndex) { |
2189 | CCValAssign &VA = RVLocs[I]; |
2190 | assert(VA.isRegLoc() && "Can only return in registers!")((VA.isRegLoc() && "Can only return in registers!") ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && \"Can only return in registers!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2190, __PRETTY_FUNCTION__)); |
2191 | |
2192 | // Add the register to the CalleeSaveDisableRegs list. |
2193 | if (ShouldDisableCalleeSavedRegister) |
2194 | MF.getRegInfo().disableCalleeSavedRegister(VA.getLocReg()); |
2195 | |
2196 | SDValue ValToCopy = OutVals[OutsIndex]; |
2197 | EVT ValVT = ValToCopy.getValueType(); |
2198 | |
2199 | // Promote values to the appropriate types. |
2200 | if (VA.getLocInfo() == CCValAssign::SExt) |
2201 | ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy); |
2202 | else if (VA.getLocInfo() == CCValAssign::ZExt) |
2203 | ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy); |
2204 | else if (VA.getLocInfo() == CCValAssign::AExt) { |
2205 | if (ValVT.isVector() && ValVT.getVectorElementType() == MVT::i1) |
2206 | ValToCopy = lowerMasksToReg(ValToCopy, VA.getLocVT(), dl, DAG); |
2207 | else |
2208 | ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy); |
2209 | } |
2210 | else if (VA.getLocInfo() == CCValAssign::BCvt) |
2211 | ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy); |
2212 | |
2213 | assert(VA.getLocInfo() != CCValAssign::FPExt &&((VA.getLocInfo() != CCValAssign::FPExt && "Unexpected FP-extend for return value." ) ? static_cast<void> (0) : __assert_fail ("VA.getLocInfo() != CCValAssign::FPExt && \"Unexpected FP-extend for return value.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2214, __PRETTY_FUNCTION__)) |
2214 | "Unexpected FP-extend for return value.")((VA.getLocInfo() != CCValAssign::FPExt && "Unexpected FP-extend for return value." ) ? static_cast<void> (0) : __assert_fail ("VA.getLocInfo() != CCValAssign::FPExt && \"Unexpected FP-extend for return value.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2214, __PRETTY_FUNCTION__)); |
2215 | |
2216 | // If this is x86-64, and we disabled SSE, we can't return FP values, |
2217 | // or SSE or MMX vectors. |
2218 | if ((ValVT == MVT::f32 || ValVT == MVT::f64 || |
2219 | VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) && |
2220 | (Subtarget.is64Bit() && !Subtarget.hasSSE1())) { |
2221 | errorUnsupported(DAG, dl, "SSE register return with SSE disabled"); |
2222 | VA.convertToReg(X86::FP0); // Set reg to FP0, avoid hitting asserts. |
2223 | } else if (ValVT == MVT::f64 && |
2224 | (Subtarget.is64Bit() && !Subtarget.hasSSE2())) { |
2225 | // Likewise we can't return F64 values with SSE1 only. gcc does so, but |
2226 | // llvm-gcc has never done it right and no one has noticed, so this |
2227 | // should be OK for now. |
2228 | errorUnsupported(DAG, dl, "SSE2 register return with SSE2 disabled"); |
2229 | VA.convertToReg(X86::FP0); // Set reg to FP0, avoid hitting asserts. |
2230 | } |
2231 | |
2232 | // Returns in ST0/ST1 are handled specially: these are pushed as operands to |
2233 | // the RET instruction and handled by the FP Stackifier. |
2234 | if (VA.getLocReg() == X86::FP0 || |
2235 | VA.getLocReg() == X86::FP1) { |
2236 | // If this is a copy from an xmm register to ST(0), use an FPExtend to |
2237 | // change the value to the FP stack register class. |
2238 | if (isScalarFPTypeInSSEReg(VA.getValVT())) |
2239 | ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy); |
2240 | RetOps.push_back(ValToCopy); |
2241 | // Don't emit a copytoreg. |
2242 | continue; |
2243 | } |
2244 | |
2245 | // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64 |
2246 | // which is returned in RAX / RDX. |
2247 | if (Subtarget.is64Bit()) { |
2248 | if (ValVT == MVT::x86mmx) { |
2249 | if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) { |
2250 | ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy); |
2251 | ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, |
2252 | ValToCopy); |
2253 | // If we don't have SSE2 available, convert to v4f32 so the generated |
2254 | // register is legal. |
2255 | if (!Subtarget.hasSSE2()) |
2256 | ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy); |
2257 | } |
2258 | } |
2259 | } |
2260 | |
2261 | SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; |
2262 | |
2263 | if (VA.needsCustom()) { |
2264 | assert(VA.getValVT() == MVT::v64i1 &&((VA.getValVT() == MVT::v64i1 && "Currently the only custom case is when we split v64i1 to 2 regs" ) ? static_cast<void> (0) : __assert_fail ("VA.getValVT() == MVT::v64i1 && \"Currently the only custom case is when we split v64i1 to 2 regs\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2265, __PRETTY_FUNCTION__)) |
2265 | "Currently the only custom case is when we split v64i1 to 2 regs")((VA.getValVT() == MVT::v64i1 && "Currently the only custom case is when we split v64i1 to 2 regs" ) ? static_cast<void> (0) : __assert_fail ("VA.getValVT() == MVT::v64i1 && \"Currently the only custom case is when we split v64i1 to 2 regs\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2265, __PRETTY_FUNCTION__)); |
2266 | |
2267 | Passv64i1ArgInRegs(dl, DAG, Chain, ValToCopy, RegsToPass, VA, RVLocs[++I], |
2268 | Subtarget); |
2269 | |
2270 | assert(2 == RegsToPass.size() &&((2 == RegsToPass.size() && "Expecting two registers after Pass64BitArgInRegs" ) ? static_cast<void> (0) : __assert_fail ("2 == RegsToPass.size() && \"Expecting two registers after Pass64BitArgInRegs\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2271, __PRETTY_FUNCTION__)) |
2271 | "Expecting two registers after Pass64BitArgInRegs")((2 == RegsToPass.size() && "Expecting two registers after Pass64BitArgInRegs" ) ? static_cast<void> (0) : __assert_fail ("2 == RegsToPass.size() && \"Expecting two registers after Pass64BitArgInRegs\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2271, __PRETTY_FUNCTION__)); |
2272 | |
2273 | // Add the second register to the CalleeSaveDisableRegs list. |
2274 | if (ShouldDisableCalleeSavedRegister) |
2275 | MF.getRegInfo().disableCalleeSavedRegister(RVLocs[I].getLocReg()); |
2276 | } else { |
2277 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), ValToCopy)); |
2278 | } |
2279 | |
2280 | // Add nodes to the DAG and add the values into the RetOps list |
2281 | for (auto &Reg : RegsToPass) { |
2282 | Chain = DAG.getCopyToReg(Chain, dl, Reg.first, Reg.second, Flag); |
2283 | Flag = Chain.getValue(1); |
2284 | RetOps.push_back(DAG.getRegister(Reg.first, Reg.second.getValueType())); |
2285 | } |
2286 | } |
2287 | |
2288 | // Swift calling convention does not require we copy the sret argument |
2289 | // into %rax/%eax for the return, and SRetReturnReg is not set for Swift. |
2290 | |
2291 | // All x86 ABIs require that for returning structs by value we copy |
2292 | // the sret argument into %rax/%eax (depending on ABI) for the return. |
2293 | // We saved the argument into a virtual register in the entry block, |
2294 | // so now we copy the value out and into %rax/%eax. |
2295 | // |
2296 | // Checking Function.hasStructRetAttr() here is insufficient because the IR |
2297 | // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is |
2298 | // false, then an sret argument may be implicitly inserted in the SelDAG. In |
2299 | // either case FuncInfo->setSRetReturnReg() will have been called. |
2300 | if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) { |
2301 | // When we have both sret and another return value, we should use the |
2302 | // original Chain stored in RetOps[0], instead of the current Chain updated |
2303 | // in the above loop. If we only have sret, RetOps[0] equals to Chain. |
2304 | |
2305 | // For the case of sret and another return value, we have |
2306 | // Chain_0 at the function entry |
2307 | // Chain_1 = getCopyToReg(Chain_0) in the above loop |
2308 | // If we use Chain_1 in getCopyFromReg, we will have |
2309 | // Val = getCopyFromReg(Chain_1) |
2310 | // Chain_2 = getCopyToReg(Chain_1, Val) from below |
2311 | |
2312 | // getCopyToReg(Chain_0) will be glued together with |
2313 | // getCopyToReg(Chain_1, Val) into Unit A, getCopyFromReg(Chain_1) will be |
2314 | // in Unit B, and we will have cyclic dependency between Unit A and Unit B: |
2315 | // Data dependency from Unit B to Unit A due to usage of Val in |
2316 | // getCopyToReg(Chain_1, Val) |
2317 | // Chain dependency from Unit A to Unit B |
2318 | |
2319 | // So here, we use RetOps[0] (i.e Chain_0) for getCopyFromReg. |
2320 | SDValue Val = DAG.getCopyFromReg(RetOps[0], dl, SRetReg, |
2321 | getPointerTy(MF.getDataLayout())); |
2322 | |
2323 | unsigned RetValReg |
2324 | = (Subtarget.is64Bit() && !Subtarget.isTarget64BitILP32()) ? |
2325 | X86::RAX : X86::EAX; |
2326 | Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag); |
2327 | Flag = Chain.getValue(1); |
2328 | |
2329 | // RAX/EAX now acts like a return value. |
2330 | RetOps.push_back( |
2331 | DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout()))); |
2332 | |
2333 | // Add the returned register to the CalleeSaveDisableRegs list. |
2334 | if (ShouldDisableCalleeSavedRegister) |
2335 | MF.getRegInfo().disableCalleeSavedRegister(RetValReg); |
2336 | } |
2337 | |
2338 | const X86RegisterInfo *TRI = Subtarget.getRegisterInfo(); |
2339 | const MCPhysReg *I = |
2340 | TRI->getCalleeSavedRegsViaCopy(&DAG.getMachineFunction()); |
2341 | if (I) { |
2342 | for (; *I; ++I) { |
2343 | if (X86::GR64RegClass.contains(*I)) |
2344 | RetOps.push_back(DAG.getRegister(*I, MVT::i64)); |
2345 | else |
2346 | llvm_unreachable("Unexpected register class in CSRsViaCopy!")::llvm::llvm_unreachable_internal("Unexpected register class in CSRsViaCopy!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2346); |
2347 | } |
2348 | } |
2349 | |
2350 | RetOps[0] = Chain; // Update chain. |
2351 | |
2352 | // Add the flag if we have it. |
2353 | if (Flag.getNode()) |
2354 | RetOps.push_back(Flag); |
2355 | |
2356 | X86ISD::NodeType opcode = X86ISD::RET_FLAG; |
2357 | if (CallConv == CallingConv::X86_INTR) |
2358 | opcode = X86ISD::IRET; |
2359 | return DAG.getNode(opcode, dl, MVT::Other, RetOps); |
2360 | } |
2361 | |
2362 | bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const { |
2363 | if (N->getNumValues() != 1 || !N->hasNUsesOfValue(1, 0)) |
2364 | return false; |
2365 | |
2366 | SDValue TCChain = Chain; |
2367 | SDNode *Copy = *N->use_begin(); |
2368 | if (Copy->getOpcode() == ISD::CopyToReg) { |
2369 | // If the copy has a glue operand, we conservatively assume it isn't safe to |
2370 | // perform a tail call. |
2371 | if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue) |
2372 | return false; |
2373 | TCChain = Copy->getOperand(0); |
2374 | } else if (Copy->getOpcode() != ISD::FP_EXTEND) |
2375 | return false; |
2376 | |
2377 | bool HasRet = false; |
2378 | for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end(); |
2379 | UI != UE; ++UI) { |
2380 | if (UI->getOpcode() != X86ISD::RET_FLAG) |
2381 | return false; |
2382 | // If we are returning more than one value, we can definitely |
2383 | // not make a tail call see PR19530 |
2384 | if (UI->getNumOperands() > 4) |
2385 | return false; |
2386 | if (UI->getNumOperands() == 4 && |
2387 | UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue) |
2388 | return false; |
2389 | HasRet = true; |
2390 | } |
2391 | |
2392 | if (!HasRet) |
2393 | return false; |
2394 | |
2395 | Chain = TCChain; |
2396 | return true; |
2397 | } |
2398 | |
2399 | EVT X86TargetLowering::getTypeForExtReturn(LLVMContext &Context, EVT VT, |
2400 | ISD::NodeType ExtendKind) const { |
2401 | MVT ReturnMVT = MVT::i32; |
2402 | |
2403 | bool Darwin = Subtarget.getTargetTriple().isOSDarwin(); |
2404 | if (VT == MVT::i1 || (!Darwin && (VT == MVT::i8 || VT == MVT::i16))) { |
2405 | // The ABI does not require i1, i8 or i16 to be extended. |
2406 | // |
2407 | // On Darwin, there is code in the wild relying on Clang's old behaviour of |
2408 | // always extending i8/i16 return values, so keep doing that for now. |
2409 | // (PR26665). |
2410 | ReturnMVT = MVT::i8; |
2411 | } |
2412 | |
2413 | EVT MinVT = getRegisterType(Context, ReturnMVT); |
2414 | return VT.bitsLT(MinVT) ? MinVT : VT; |
2415 | } |
2416 | |
2417 | /// Reads two 32 bit registers and creates a 64 bit mask value. |
2418 | /// \param VA The current 32 bit value that need to be assigned. |
2419 | /// \param NextVA The next 32 bit value that need to be assigned. |
2420 | /// \param Root The parent DAG node. |
2421 | /// \param [in,out] InFlag Represents SDvalue in the parent DAG node for |
2422 | /// glue purposes. In the case the DAG is already using |
2423 | /// physical register instead of virtual, we should glue |
2424 | /// our new SDValue to InFlag SDvalue. |
2425 | /// \return a new SDvalue of size 64bit. |
2426 | static SDValue getv64i1Argument(CCValAssign &VA, CCValAssign &NextVA, |
2427 | SDValue &Root, SelectionDAG &DAG, |
2428 | const SDLoc &Dl, const X86Subtarget &Subtarget, |
2429 | SDValue *InFlag = nullptr) { |
2430 | assert((Subtarget.hasBWI()) && "Expected AVX512BW target!")(((Subtarget.hasBWI()) && "Expected AVX512BW target!" ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasBWI()) && \"Expected AVX512BW target!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2430, __PRETTY_FUNCTION__)); |
2431 | assert(Subtarget.is32Bit() && "Expecting 32 bit target")((Subtarget.is32Bit() && "Expecting 32 bit target") ? static_cast<void> (0) : __assert_fail ("Subtarget.is32Bit() && \"Expecting 32 bit target\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2431, __PRETTY_FUNCTION__)); |
2432 | assert(VA.getValVT() == MVT::v64i1 &&((VA.getValVT() == MVT::v64i1 && "Expecting first location of 64 bit width type" ) ? static_cast<void> (0) : __assert_fail ("VA.getValVT() == MVT::v64i1 && \"Expecting first location of 64 bit width type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2433, __PRETTY_FUNCTION__)) |
2433 | "Expecting first location of 64 bit width type")((VA.getValVT() == MVT::v64i1 && "Expecting first location of 64 bit width type" ) ? static_cast<void> (0) : __assert_fail ("VA.getValVT() == MVT::v64i1 && \"Expecting first location of 64 bit width type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2433, __PRETTY_FUNCTION__)); |
2434 | assert(NextVA.getValVT() == VA.getValVT() &&((NextVA.getValVT() == VA.getValVT() && "The locations should have the same type" ) ? static_cast<void> (0) : __assert_fail ("NextVA.getValVT() == VA.getValVT() && \"The locations should have the same type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2435, __PRETTY_FUNCTION__)) |
2435 | "The locations should have the same type")((NextVA.getValVT() == VA.getValVT() && "The locations should have the same type" ) ? static_cast<void> (0) : __assert_fail ("NextVA.getValVT() == VA.getValVT() && \"The locations should have the same type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2435, __PRETTY_FUNCTION__)); |
2436 | assert(VA.isRegLoc() && NextVA.isRegLoc() &&((VA.isRegLoc() && NextVA.isRegLoc() && "The values should reside in two registers" ) ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && NextVA.isRegLoc() && \"The values should reside in two registers\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2437, __PRETTY_FUNCTION__)) |
2437 | "The values should reside in two registers")((VA.isRegLoc() && NextVA.isRegLoc() && "The values should reside in two registers" ) ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && NextVA.isRegLoc() && \"The values should reside in two registers\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2437, __PRETTY_FUNCTION__)); |
2438 | |
2439 | SDValue Lo, Hi; |
2440 | unsigned Reg; |
2441 | SDValue ArgValueLo, ArgValueHi; |
2442 | |
2443 | MachineFunction &MF = DAG.getMachineFunction(); |
2444 | const TargetRegisterClass *RC = &X86::GR32RegClass; |
2445 | |
2446 | // Read a 32 bit value from the registers |
2447 | if (nullptr == InFlag) { |
2448 | // When no physical register is present, |
2449 | // create an intermediate virtual register |
2450 | Reg = MF.addLiveIn(VA.getLocReg(), RC); |
2451 | ArgValueLo = DAG.getCopyFromReg(Root, Dl, Reg, MVT::i32); |
2452 | Reg = MF.addLiveIn(NextVA.getLocReg(), RC); |
2453 | ArgValueHi = DAG.getCopyFromReg(Root, Dl, Reg, MVT::i32); |
2454 | } else { |
2455 | // When a physical register is available read the value from it and glue |
2456 | // the reads together. |
2457 | ArgValueLo = |
2458 | DAG.getCopyFromReg(Root, Dl, VA.getLocReg(), MVT::i32, *InFlag); |
2459 | *InFlag = ArgValueLo.getValue(2); |
2460 | ArgValueHi = |
2461 | DAG.getCopyFromReg(Root, Dl, NextVA.getLocReg(), MVT::i32, *InFlag); |
2462 | *InFlag = ArgValueHi.getValue(2); |
2463 | } |
2464 | |
2465 | // Convert the i32 type into v32i1 type |
2466 | Lo = DAG.getBitcast(MVT::v32i1, ArgValueLo); |
2467 | |
2468 | // Convert the i32 type into v32i1 type |
2469 | Hi = DAG.getBitcast(MVT::v32i1, ArgValueHi); |
2470 | |
2471 | // Concatenate the two values together |
2472 | return DAG.getNode(ISD::CONCAT_VECTORS, Dl, MVT::v64i1, Lo, Hi); |
2473 | } |
2474 | |
2475 | /// The function will lower a register of various sizes (8/16/32/64) |
2476 | /// to a mask value of the expected size (v8i1/v16i1/v32i1/v64i1) |
2477 | /// \returns a DAG node contains the operand after lowering to mask type. |
2478 | static SDValue lowerRegToMasks(const SDValue &ValArg, const EVT &ValVT, |
2479 | const EVT &ValLoc, const SDLoc &Dl, |
2480 | SelectionDAG &DAG) { |
2481 | SDValue ValReturned = ValArg; |
2482 | |
2483 | if (ValVT == MVT::v1i1) |
2484 | return DAG.getNode(ISD::SCALAR_TO_VECTOR, Dl, MVT::v1i1, ValReturned); |
2485 | |
2486 | if (ValVT == MVT::v64i1) { |
2487 | // In 32 bit machine, this case is handled by getv64i1Argument |
2488 | assert(ValLoc == MVT::i64 && "Expecting only i64 locations")((ValLoc == MVT::i64 && "Expecting only i64 locations" ) ? static_cast<void> (0) : __assert_fail ("ValLoc == MVT::i64 && \"Expecting only i64 locations\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2488, __PRETTY_FUNCTION__)); |
2489 | // In 64 bit machine, There is no need to truncate the value only bitcast |
2490 | } else { |
2491 | MVT maskLen; |
2492 | switch (ValVT.getSimpleVT().SimpleTy) { |
2493 | case MVT::v8i1: |
2494 | maskLen = MVT::i8; |
2495 | break; |
2496 | case MVT::v16i1: |
2497 | maskLen = MVT::i16; |
2498 | break; |
2499 | case MVT::v32i1: |
2500 | maskLen = MVT::i32; |
2501 | break; |
2502 | default: |
2503 | llvm_unreachable("Expecting a vector of i1 types")::llvm::llvm_unreachable_internal("Expecting a vector of i1 types" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2503); |
2504 | } |
2505 | |
2506 | ValReturned = DAG.getNode(ISD::TRUNCATE, Dl, maskLen, ValReturned); |
2507 | } |
2508 | return DAG.getBitcast(ValVT, ValReturned); |
2509 | } |
2510 | |
2511 | /// Lower the result values of a call into the |
2512 | /// appropriate copies out of appropriate physical registers. |
2513 | /// |
2514 | SDValue X86TargetLowering::LowerCallResult( |
2515 | SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg, |
2516 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
2517 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, |
2518 | uint32_t *RegMask) const { |
2519 | |
2520 | const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); |
2521 | // Assign locations to each value returned by this call. |
2522 | SmallVector<CCValAssign, 16> RVLocs; |
2523 | bool Is64Bit = Subtarget.is64Bit(); |
2524 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, |
2525 | *DAG.getContext()); |
2526 | CCInfo.AnalyzeCallResult(Ins, RetCC_X86); |
2527 | |
2528 | // Copy all of the result registers out of their specified physreg. |
2529 | for (unsigned I = 0, InsIndex = 0, E = RVLocs.size(); I != E; |
2530 | ++I, ++InsIndex) { |
2531 | CCValAssign &VA = RVLocs[I]; |
2532 | EVT CopyVT = VA.getLocVT(); |
2533 | |
2534 | // In some calling conventions we need to remove the used registers |
2535 | // from the register mask. |
2536 | if (RegMask) { |
2537 | for (MCSubRegIterator SubRegs(VA.getLocReg(), TRI, /*IncludeSelf=*/true); |
2538 | SubRegs.isValid(); ++SubRegs) |
2539 | RegMask[*SubRegs / 32] &= ~(1u << (*SubRegs % 32)); |
2540 | } |
2541 | |
2542 | // If this is x86-64, and we disabled SSE, we can't return FP values |
2543 | if ((CopyVT == MVT::f32 || CopyVT == MVT::f64 || CopyVT == MVT::f128) && |
2544 | ((Is64Bit || Ins[InsIndex].Flags.isInReg()) && !Subtarget.hasSSE1())) { |
2545 | errorUnsupported(DAG, dl, "SSE register return with SSE disabled"); |
2546 | VA.convertToReg(X86::FP0); // Set reg to FP0, avoid hitting asserts. |
2547 | } |
2548 | |
2549 | // If we prefer to use the value in xmm registers, copy it out as f80 and |
2550 | // use a truncate to move it from fp stack reg to xmm reg. |
2551 | bool RoundAfterCopy = false; |
2552 | if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) && |
2553 | isScalarFPTypeInSSEReg(VA.getValVT())) { |
2554 | if (!Subtarget.hasX87()) |
2555 | report_fatal_error("X87 register return with X87 disabled"); |
2556 | CopyVT = MVT::f80; |
2557 | RoundAfterCopy = (CopyVT != VA.getLocVT()); |
2558 | } |
2559 | |
2560 | SDValue Val; |
2561 | if (VA.needsCustom()) { |
2562 | assert(VA.getValVT() == MVT::v64i1 &&((VA.getValVT() == MVT::v64i1 && "Currently the only custom case is when we split v64i1 to 2 regs" ) ? static_cast<void> (0) : __assert_fail ("VA.getValVT() == MVT::v64i1 && \"Currently the only custom case is when we split v64i1 to 2 regs\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2563, __PRETTY_FUNCTION__)) |
2563 | "Currently the only custom case is when we split v64i1 to 2 regs")((VA.getValVT() == MVT::v64i1 && "Currently the only custom case is when we split v64i1 to 2 regs" ) ? static_cast<void> (0) : __assert_fail ("VA.getValVT() == MVT::v64i1 && \"Currently the only custom case is when we split v64i1 to 2 regs\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2563, __PRETTY_FUNCTION__)); |
2564 | Val = |
2565 | getv64i1Argument(VA, RVLocs[++I], Chain, DAG, dl, Subtarget, &InFlag); |
2566 | } else { |
2567 | Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), CopyVT, InFlag) |
2568 | .getValue(1); |
2569 | Val = Chain.getValue(0); |
2570 | InFlag = Chain.getValue(2); |
2571 | } |
2572 | |
2573 | if (RoundAfterCopy) |
2574 | Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val, |
2575 | // This truncation won't change the value. |
2576 | DAG.getIntPtrConstant(1, dl)); |
2577 | |
2578 | if (VA.isExtInLoc() && (VA.getValVT().getScalarType() == MVT::i1)) { |
2579 | if (VA.getValVT().isVector() && |
2580 | ((VA.getLocVT() == MVT::i64) || (VA.getLocVT() == MVT::i32) || |
2581 | (VA.getLocVT() == MVT::i16) || (VA.getLocVT() == MVT::i8))) { |
2582 | // promoting a mask type (v*i1) into a register of type i64/i32/i16/i8 |
2583 | Val = lowerRegToMasks(Val, VA.getValVT(), VA.getLocVT(), dl, DAG); |
2584 | } else |
2585 | Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val); |
2586 | } |
2587 | |
2588 | InVals.push_back(Val); |
2589 | } |
2590 | |
2591 | return Chain; |
2592 | } |
2593 | |
2594 | //===----------------------------------------------------------------------===// |
2595 | // C & StdCall & Fast Calling Convention implementation |
2596 | //===----------------------------------------------------------------------===// |
2597 | // StdCall calling convention seems to be standard for many Windows' API |
2598 | // routines and around. It differs from C calling convention just a little: |
2599 | // callee should clean up the stack, not caller. Symbols should be also |
2600 | // decorated in some fancy way :) It doesn't support any vector arguments. |
2601 | // For info on fast calling convention see Fast Calling Convention (tail call) |
2602 | // implementation LowerX86_32FastCCCallTo. |
2603 | |
2604 | /// CallIsStructReturn - Determines whether a call uses struct return |
2605 | /// semantics. |
2606 | enum StructReturnType { |
2607 | NotStructReturn, |
2608 | RegStructReturn, |
2609 | StackStructReturn |
2610 | }; |
2611 | static StructReturnType |
2612 | callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsMCU) { |
2613 | if (Outs.empty()) |
2614 | return NotStructReturn; |
2615 | |
2616 | const ISD::ArgFlagsTy &Flags = Outs[0].Flags; |
2617 | if (!Flags.isSRet()) |
2618 | return NotStructReturn; |
2619 | if (Flags.isInReg() || IsMCU) |
2620 | return RegStructReturn; |
2621 | return StackStructReturn; |
2622 | } |
2623 | |
2624 | /// Determines whether a function uses struct return semantics. |
2625 | static StructReturnType |
2626 | argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins, bool IsMCU) { |
2627 | if (Ins.empty()) |
2628 | return NotStructReturn; |
2629 | |
2630 | const ISD::ArgFlagsTy &Flags = Ins[0].Flags; |
2631 | if (!Flags.isSRet()) |
2632 | return NotStructReturn; |
2633 | if (Flags.isInReg() || IsMCU) |
2634 | return RegStructReturn; |
2635 | return StackStructReturn; |
2636 | } |
2637 | |
2638 | /// Make a copy of an aggregate at address specified by "Src" to address |
2639 | /// "Dst" with size and alignment information specified by the specific |
2640 | /// parameter attribute. The copy will be passed as a byval function parameter. |
2641 | static SDValue CreateCopyOfByValArgument(SDValue Src, SDValue Dst, |
2642 | SDValue Chain, ISD::ArgFlagsTy Flags, |
2643 | SelectionDAG &DAG, const SDLoc &dl) { |
2644 | SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32); |
2645 | |
2646 | return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(), |
2647 | /*isVolatile*/false, /*AlwaysInline=*/true, |
2648 | /*isTailCall*/false, |
2649 | MachinePointerInfo(), MachinePointerInfo()); |
2650 | } |
2651 | |
2652 | /// Return true if the calling convention is one that we can guarantee TCO for. |
2653 | static bool canGuaranteeTCO(CallingConv::ID CC) { |
2654 | return (CC == CallingConv::Fast || CC == CallingConv::GHC || |
2655 | CC == CallingConv::X86_RegCall || CC == CallingConv::HiPE || |
2656 | CC == CallingConv::HHVM); |
2657 | } |
2658 | |
2659 | /// Return true if we might ever do TCO for calls with this calling convention. |
2660 | static bool mayTailCallThisCC(CallingConv::ID CC) { |
2661 | switch (CC) { |
2662 | // C calling conventions: |
2663 | case CallingConv::C: |
2664 | case CallingConv::X86_64_Win64: |
2665 | case CallingConv::X86_64_SysV: |
2666 | // Callee pop conventions: |
2667 | case CallingConv::X86_ThisCall: |
2668 | case CallingConv::X86_StdCall: |
2669 | case CallingConv::X86_VectorCall: |
2670 | case CallingConv::X86_FastCall: |
2671 | return true; |
2672 | default: |
2673 | return canGuaranteeTCO(CC); |
2674 | } |
2675 | } |
2676 | |
2677 | /// Return true if the function is being made into a tailcall target by |
2678 | /// changing its ABI. |
2679 | static bool shouldGuaranteeTCO(CallingConv::ID CC, bool GuaranteedTailCallOpt) { |
2680 | return GuaranteedTailCallOpt && canGuaranteeTCO(CC); |
2681 | } |
2682 | |
2683 | bool X86TargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const { |
2684 | auto Attr = |
2685 | CI->getParent()->getParent()->getFnAttribute("disable-tail-calls"); |
2686 | if (!CI->isTailCall() || Attr.getValueAsString() == "true") |
2687 | return false; |
2688 | |
2689 | ImmutableCallSite CS(CI); |
2690 | CallingConv::ID CalleeCC = CS.getCallingConv(); |
2691 | if (!mayTailCallThisCC(CalleeCC)) |
2692 | return false; |
2693 | |
2694 | return true; |
2695 | } |
2696 | |
2697 | SDValue |
2698 | X86TargetLowering::LowerMemArgument(SDValue Chain, CallingConv::ID CallConv, |
2699 | const SmallVectorImpl<ISD::InputArg> &Ins, |
2700 | const SDLoc &dl, SelectionDAG &DAG, |
2701 | const CCValAssign &VA, |
2702 | MachineFrameInfo &MFI, unsigned i) const { |
2703 | // Create the nodes corresponding to a load from this parameter slot. |
2704 | ISD::ArgFlagsTy Flags = Ins[i].Flags; |
2705 | bool AlwaysUseMutable = shouldGuaranteeTCO( |
2706 | CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt); |
2707 | bool isImmutable = !AlwaysUseMutable && !Flags.isByVal(); |
2708 | EVT ValVT; |
2709 | MVT PtrVT = getPointerTy(DAG.getDataLayout()); |
2710 | |
2711 | // If value is passed by pointer we have address passed instead of the value |
2712 | // itself. No need to extend if the mask value and location share the same |
2713 | // absolute size. |
2714 | bool ExtendedInMem = |
2715 | VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1 && |
2716 | VA.getValVT().getSizeInBits() != VA.getLocVT().getSizeInBits(); |
2717 | |
2718 | if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem) |
2719 | ValVT = VA.getLocVT(); |
2720 | else |
2721 | ValVT = VA.getValVT(); |
2722 | |
2723 | // Calculate SP offset of interrupt parameter, re-arrange the slot normally |
2724 | // taken by a return address. |
2725 | int Offset = 0; |
2726 | if (CallConv == CallingConv::X86_INTR) { |
2727 | // X86 interrupts may take one or two arguments. |
2728 | // On the stack there will be no return address as in regular call. |
2729 | // Offset of last argument need to be set to -4/-8 bytes. |
2730 | // Where offset of the first argument out of two, should be set to 0 bytes. |
2731 | Offset = (Subtarget.is64Bit() ? 8 : 4) * ((i + 1) % Ins.size() - 1); |
2732 | if (Subtarget.is64Bit() && Ins.size() == 2) { |
2733 | // The stack pointer needs to be realigned for 64 bit handlers with error |
2734 | // code, so the argument offset changes by 8 bytes. |
2735 | Offset += 8; |
2736 | } |
2737 | } |
2738 | |
2739 | // FIXME: For now, all byval parameter objects are marked mutable. This can be |
2740 | // changed with more analysis. |
2741 | // In case of tail call optimization mark all arguments mutable. Since they |
2742 | // could be overwritten by lowering of arguments in case of a tail call. |
2743 | if (Flags.isByVal()) { |
2744 | unsigned Bytes = Flags.getByValSize(); |
2745 | if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects. |
2746 | int FI = MFI.CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable); |
2747 | // Adjust SP offset of interrupt parameter. |
2748 | if (CallConv == CallingConv::X86_INTR) { |
2749 | MFI.setObjectOffset(FI, Offset); |
2750 | } |
2751 | return DAG.getFrameIndex(FI, PtrVT); |
2752 | } |
2753 | |
2754 | // This is an argument in memory. We might be able to perform copy elision. |
2755 | if (Flags.isCopyElisionCandidate()) { |
2756 | EVT ArgVT = Ins[i].ArgVT; |
2757 | SDValue PartAddr; |
2758 | if (Ins[i].PartOffset == 0) { |
2759 | // If this is a one-part value or the first part of a multi-part value, |
2760 | // create a stack object for the entire argument value type and return a |
2761 | // load from our portion of it. This assumes that if the first part of an |
2762 | // argument is in memory, the rest will also be in memory. |
2763 | int FI = MFI.CreateFixedObject(ArgVT.getStoreSize(), VA.getLocMemOffset(), |
2764 | /*Immutable=*/false); |
2765 | PartAddr = DAG.getFrameIndex(FI, PtrVT); |
2766 | return DAG.getLoad( |
2767 | ValVT, dl, Chain, PartAddr, |
2768 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI)); |
2769 | } else { |
2770 | // This is not the first piece of an argument in memory. See if there is |
2771 | // already a fixed stack object including this offset. If so, assume it |
2772 | // was created by the PartOffset == 0 branch above and create a load from |
2773 | // the appropriate offset into it. |
2774 | int64_t PartBegin = VA.getLocMemOffset(); |
2775 | int64_t PartEnd = PartBegin + ValVT.getSizeInBits() / 8; |
2776 | int FI = MFI.getObjectIndexBegin(); |
2777 | for (; MFI.isFixedObjectIndex(FI); ++FI) { |
2778 | int64_t ObjBegin = MFI.getObjectOffset(FI); |
2779 | int64_t ObjEnd = ObjBegin + MFI.getObjectSize(FI); |
2780 | if (ObjBegin <= PartBegin && PartEnd <= ObjEnd) |
2781 | break; |
2782 | } |
2783 | if (MFI.isFixedObjectIndex(FI)) { |
2784 | SDValue Addr = |
2785 | DAG.getNode(ISD::ADD, dl, PtrVT, DAG.getFrameIndex(FI, PtrVT), |
2786 | DAG.getIntPtrConstant(Ins[i].PartOffset, dl)); |
2787 | return DAG.getLoad( |
2788 | ValVT, dl, Chain, Addr, |
2789 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI, |
2790 | Ins[i].PartOffset)); |
2791 | } |
2792 | } |
2793 | } |
2794 | |
2795 | int FI = MFI.CreateFixedObject(ValVT.getSizeInBits() / 8, |
2796 | VA.getLocMemOffset(), isImmutable); |
2797 | |
2798 | // Set SExt or ZExt flag. |
2799 | if (VA.getLocInfo() == CCValAssign::ZExt) { |
2800 | MFI.setObjectZExt(FI, true); |
2801 | } else if (VA.getLocInfo() == CCValAssign::SExt) { |
2802 | MFI.setObjectSExt(FI, true); |
2803 | } |
2804 | |
2805 | // Adjust SP offset of interrupt parameter. |
2806 | if (CallConv == CallingConv::X86_INTR) { |
2807 | MFI.setObjectOffset(FI, Offset); |
2808 | } |
2809 | |
2810 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
2811 | SDValue Val = DAG.getLoad( |
2812 | ValVT, dl, Chain, FIN, |
2813 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI)); |
2814 | return ExtendedInMem |
2815 | ? (VA.getValVT().isVector() |
2816 | ? DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VA.getValVT(), Val) |
2817 | : DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val)) |
2818 | : Val; |
2819 | } |
2820 | |
2821 | // FIXME: Get this from tablegen. |
2822 | static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv, |
2823 | const X86Subtarget &Subtarget) { |
2824 | assert(Subtarget.is64Bit())((Subtarget.is64Bit()) ? static_cast<void> (0) : __assert_fail ("Subtarget.is64Bit()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2824, __PRETTY_FUNCTION__)); |
2825 | |
2826 | if (Subtarget.isCallingConvWin64(CallConv)) { |
2827 | static const MCPhysReg GPR64ArgRegsWin64[] = { |
2828 | X86::RCX, X86::RDX, X86::R8, X86::R9 |
2829 | }; |
2830 | return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64)); |
2831 | } |
2832 | |
2833 | static const MCPhysReg GPR64ArgRegs64Bit[] = { |
2834 | X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9 |
2835 | }; |
2836 | return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit)); |
2837 | } |
2838 | |
2839 | // FIXME: Get this from tablegen. |
2840 | static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF, |
2841 | CallingConv::ID CallConv, |
2842 | const X86Subtarget &Subtarget) { |
2843 | assert(Subtarget.is64Bit())((Subtarget.is64Bit()) ? static_cast<void> (0) : __assert_fail ("Subtarget.is64Bit()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2843, __PRETTY_FUNCTION__)); |
2844 | if (Subtarget.isCallingConvWin64(CallConv)) { |
2845 | // The XMM registers which might contain var arg parameters are shadowed |
2846 | // in their paired GPR. So we only need to save the GPR to their home |
2847 | // slots. |
2848 | // TODO: __vectorcall will change this. |
2849 | return None; |
2850 | } |
2851 | |
2852 | const Function *Fn = MF.getFunction(); |
2853 | bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat); |
2854 | bool isSoftFloat = Subtarget.useSoftFloat(); |
2855 | assert(!(isSoftFloat && NoImplicitFloatOps) &&((!(isSoftFloat && NoImplicitFloatOps) && "SSE register cannot be used when SSE is disabled!" ) ? static_cast<void> (0) : __assert_fail ("!(isSoftFloat && NoImplicitFloatOps) && \"SSE register cannot be used when SSE is disabled!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2856, __PRETTY_FUNCTION__)) |
2856 | "SSE register cannot be used when SSE is disabled!")((!(isSoftFloat && NoImplicitFloatOps) && "SSE register cannot be used when SSE is disabled!" ) ? static_cast<void> (0) : __assert_fail ("!(isSoftFloat && NoImplicitFloatOps) && \"SSE register cannot be used when SSE is disabled!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2856, __PRETTY_FUNCTION__)); |
2857 | if (isSoftFloat || NoImplicitFloatOps || !Subtarget.hasSSE1()) |
2858 | // Kernel mode asks for SSE to be disabled, so there are no XMM argument |
2859 | // registers. |
2860 | return None; |
2861 | |
2862 | static const MCPhysReg XMMArgRegs64Bit[] = { |
2863 | X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, |
2864 | X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7 |
2865 | }; |
2866 | return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit)); |
2867 | } |
2868 | |
2869 | #ifndef NDEBUG |
2870 | static bool isSortedByValueNo(const SmallVectorImpl<CCValAssign> &ArgLocs) { |
2871 | return std::is_sorted(ArgLocs.begin(), ArgLocs.end(), |
2872 | [](const CCValAssign &A, const CCValAssign &B) -> bool { |
2873 | return A.getValNo() < B.getValNo(); |
2874 | }); |
2875 | } |
2876 | #endif |
2877 | |
2878 | SDValue X86TargetLowering::LowerFormalArguments( |
2879 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
2880 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
2881 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
2882 | MachineFunction &MF = DAG.getMachineFunction(); |
2883 | X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); |
2884 | const TargetFrameLowering &TFI = *Subtarget.getFrameLowering(); |
2885 | |
2886 | const Function *Fn = MF.getFunction(); |
2887 | if (Fn->hasExternalLinkage() && |
2888 | Subtarget.isTargetCygMing() && |
2889 | Fn->getName() == "main") |
2890 | FuncInfo->setForceFramePointer(true); |
2891 | |
2892 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
2893 | bool Is64Bit = Subtarget.is64Bit(); |
2894 | bool IsWin64 = Subtarget.isCallingConvWin64(CallConv); |
2895 | |
2896 | assert(((!(isVarArg && canGuaranteeTCO(CallConv)) && "Var args not supported with calling conv' regcall, fastcc, ghc or hipe" ) ? static_cast<void> (0) : __assert_fail ("!(isVarArg && canGuaranteeTCO(CallConv)) && \"Var args not supported with calling conv' regcall, fastcc, ghc or hipe\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2898, __PRETTY_FUNCTION__)) |
2897 | !(isVarArg && canGuaranteeTCO(CallConv)) &&((!(isVarArg && canGuaranteeTCO(CallConv)) && "Var args not supported with calling conv' regcall, fastcc, ghc or hipe" ) ? static_cast<void> (0) : __assert_fail ("!(isVarArg && canGuaranteeTCO(CallConv)) && \"Var args not supported with calling conv' regcall, fastcc, ghc or hipe\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2898, __PRETTY_FUNCTION__)) |
2898 | "Var args not supported with calling conv' regcall, fastcc, ghc or hipe")((!(isVarArg && canGuaranteeTCO(CallConv)) && "Var args not supported with calling conv' regcall, fastcc, ghc or hipe" ) ? static_cast<void> (0) : __assert_fail ("!(isVarArg && canGuaranteeTCO(CallConv)) && \"Var args not supported with calling conv' regcall, fastcc, ghc or hipe\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2898, __PRETTY_FUNCTION__)); |
2899 | |
2900 | if (CallConv == CallingConv::X86_INTR) { |
2901 | bool isLegal = Ins.size() == 1 || |
2902 | (Ins.size() == 2 && ((Is64Bit && Ins[1].VT == MVT::i64) || |
2903 | (!Is64Bit && Ins[1].VT == MVT::i32))); |
2904 | if (!isLegal) |
2905 | report_fatal_error("X86 interrupts may take one or two arguments"); |
2906 | } |
2907 | |
2908 | // Assign locations to all of the incoming arguments. |
2909 | SmallVector<CCValAssign, 16> ArgLocs; |
2910 | CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext()); |
2911 | |
2912 | // Allocate shadow area for Win64. |
2913 | if (IsWin64) |
2914 | CCInfo.AllocateStack(32, 8); |
2915 | |
2916 | CCInfo.AnalyzeArguments(Ins, CC_X86); |
2917 | |
2918 | // In vectorcall calling convention a second pass is required for the HVA |
2919 | // types. |
2920 | if (CallingConv::X86_VectorCall == CallConv) { |
2921 | CCInfo.AnalyzeArgumentsSecondPass(Ins, CC_X86); |
2922 | } |
2923 | |
2924 | // The next loop assumes that the locations are in the same order of the |
2925 | // input arguments. |
2926 | assert(isSortedByValueNo(ArgLocs) &&((isSortedByValueNo(ArgLocs) && "Argument Location list must be sorted before lowering" ) ? static_cast<void> (0) : __assert_fail ("isSortedByValueNo(ArgLocs) && \"Argument Location list must be sorted before lowering\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2927, __PRETTY_FUNCTION__)) |
2927 | "Argument Location list must be sorted before lowering")((isSortedByValueNo(ArgLocs) && "Argument Location list must be sorted before lowering" ) ? static_cast<void> (0) : __assert_fail ("isSortedByValueNo(ArgLocs) && \"Argument Location list must be sorted before lowering\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2927, __PRETTY_FUNCTION__)); |
2928 | |
2929 | SDValue ArgValue; |
2930 | for (unsigned I = 0, InsIndex = 0, E = ArgLocs.size(); I != E; |
2931 | ++I, ++InsIndex) { |
2932 | assert(InsIndex < Ins.size() && "Invalid Ins index")((InsIndex < Ins.size() && "Invalid Ins index") ? static_cast <void> (0) : __assert_fail ("InsIndex < Ins.size() && \"Invalid Ins index\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2932, __PRETTY_FUNCTION__)); |
2933 | CCValAssign &VA = ArgLocs[I]; |
2934 | |
2935 | if (VA.isRegLoc()) { |
2936 | EVT RegVT = VA.getLocVT(); |
2937 | if (VA.needsCustom()) { |
2938 | assert(((VA.getValVT() == MVT::v64i1 && "Currently the only custom case is when we split v64i1 to 2 regs" ) ? static_cast<void> (0) : __assert_fail ("VA.getValVT() == MVT::v64i1 && \"Currently the only custom case is when we split v64i1 to 2 regs\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2940, __PRETTY_FUNCTION__)) |
2939 | VA.getValVT() == MVT::v64i1 &&((VA.getValVT() == MVT::v64i1 && "Currently the only custom case is when we split v64i1 to 2 regs" ) ? static_cast<void> (0) : __assert_fail ("VA.getValVT() == MVT::v64i1 && \"Currently the only custom case is when we split v64i1 to 2 regs\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2940, __PRETTY_FUNCTION__)) |
2940 | "Currently the only custom case is when we split v64i1 to 2 regs")((VA.getValVT() == MVT::v64i1 && "Currently the only custom case is when we split v64i1 to 2 regs" ) ? static_cast<void> (0) : __assert_fail ("VA.getValVT() == MVT::v64i1 && \"Currently the only custom case is when we split v64i1 to 2 regs\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2940, __PRETTY_FUNCTION__)); |
2941 | |
2942 | // v64i1 values, in regcall calling convention, that are |
2943 | // compiled to 32 bit arch, are split up into two registers. |
2944 | ArgValue = |
2945 | getv64i1Argument(VA, ArgLocs[++I], Chain, DAG, dl, Subtarget); |
2946 | } else { |
2947 | const TargetRegisterClass *RC; |
2948 | if (RegVT == MVT::i32) |
2949 | RC = &X86::GR32RegClass; |
2950 | else if (Is64Bit && RegVT == MVT::i64) |
2951 | RC = &X86::GR64RegClass; |
2952 | else if (RegVT == MVT::f32) |
2953 | RC = Subtarget.hasAVX512() ? &X86::FR32XRegClass : &X86::FR32RegClass; |
2954 | else if (RegVT == MVT::f64) |
2955 | RC = Subtarget.hasAVX512() ? &X86::FR64XRegClass : &X86::FR64RegClass; |
2956 | else if (RegVT == MVT::f80) |
2957 | RC = &X86::RFP80RegClass; |
2958 | else if (RegVT == MVT::f128) |
2959 | RC = &X86::FR128RegClass; |
2960 | else if (RegVT.is512BitVector()) |
2961 | RC = &X86::VR512RegClass; |
2962 | else if (RegVT.is256BitVector()) |
2963 | RC = Subtarget.hasVLX() ? &X86::VR256XRegClass : &X86::VR256RegClass; |
2964 | else if (RegVT.is128BitVector()) |
2965 | RC = Subtarget.hasVLX() ? &X86::VR128XRegClass : &X86::VR128RegClass; |
2966 | else if (RegVT == MVT::x86mmx) |
2967 | RC = &X86::VR64RegClass; |
2968 | else if (RegVT == MVT::v1i1) |
2969 | RC = &X86::VK1RegClass; |
2970 | else if (RegVT == MVT::v8i1) |
2971 | RC = &X86::VK8RegClass; |
2972 | else if (RegVT == MVT::v16i1) |
2973 | RC = &X86::VK16RegClass; |
2974 | else if (RegVT == MVT::v32i1) |
2975 | RC = &X86::VK32RegClass; |
2976 | else if (RegVT == MVT::v64i1) |
2977 | RC = &X86::VK64RegClass; |
2978 | else |
2979 | llvm_unreachable("Unknown argument type!")::llvm::llvm_unreachable_internal("Unknown argument type!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 2979); |
2980 | |
2981 | unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); |
2982 | ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT); |
2983 | } |
2984 | |
2985 | // If this is an 8 or 16-bit value, it is really passed promoted to 32 |
2986 | // bits. Insert an assert[sz]ext to capture this, then truncate to the |
2987 | // right size. |
2988 | if (VA.getLocInfo() == CCValAssign::SExt) |
2989 | ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue, |
2990 | DAG.getValueType(VA.getValVT())); |
2991 | else if (VA.getLocInfo() == CCValAssign::ZExt) |
2992 | ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue, |
2993 | DAG.getValueType(VA.getValVT())); |
2994 | else if (VA.getLocInfo() == CCValAssign::BCvt) |
2995 | ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue); |
2996 | |
2997 | if (VA.isExtInLoc()) { |
2998 | // Handle MMX values passed in XMM regs. |
2999 | if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1) |
3000 | ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue); |
3001 | else if (VA.getValVT().isVector() && |
3002 | VA.getValVT().getScalarType() == MVT::i1 && |
3003 | ((VA.getLocVT() == MVT::i64) || (VA.getLocVT() == MVT::i32) || |
3004 | (VA.getLocVT() == MVT::i16) || (VA.getLocVT() == MVT::i8))) { |
3005 | // Promoting a mask type (v*i1) into a register of type i64/i32/i16/i8 |
3006 | ArgValue = lowerRegToMasks(ArgValue, VA.getValVT(), RegVT, dl, DAG); |
3007 | } else |
3008 | ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); |
3009 | } |
3010 | } else { |
3011 | assert(VA.isMemLoc())((VA.isMemLoc()) ? static_cast<void> (0) : __assert_fail ("VA.isMemLoc()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3011, __PRETTY_FUNCTION__)); |
3012 | ArgValue = |
3013 | LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, InsIndex); |
3014 | } |
3015 | |
3016 | // If value is passed via pointer - do a load. |
3017 | if (VA.getLocInfo() == CCValAssign::Indirect) |
3018 | ArgValue = |
3019 | DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, MachinePointerInfo()); |
3020 | |
3021 | InVals.push_back(ArgValue); |
3022 | } |
3023 | |
3024 | for (unsigned I = 0, E = Ins.size(); I != E; ++I) { |
3025 | // Swift calling convention does not require we copy the sret argument |
3026 | // into %rax/%eax for the return. We don't set SRetReturnReg for Swift. |
3027 | if (CallConv == CallingConv::Swift) |
3028 | continue; |
3029 | |
3030 | // All x86 ABIs require that for returning structs by value we copy the |
3031 | // sret argument into %rax/%eax (depending on ABI) for the return. Save |
3032 | // the argument into a virtual register so that we can access it from the |
3033 | // return points. |
3034 | if (Ins[I].Flags.isSRet()) { |
3035 | unsigned Reg = FuncInfo->getSRetReturnReg(); |
3036 | if (!Reg) { |
3037 | MVT PtrTy = getPointerTy(DAG.getDataLayout()); |
3038 | Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy)); |
3039 | FuncInfo->setSRetReturnReg(Reg); |
3040 | } |
3041 | SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[I]); |
3042 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain); |
3043 | break; |
3044 | } |
3045 | } |
3046 | |
3047 | unsigned StackSize = CCInfo.getNextStackOffset(); |
3048 | // Align stack specially for tail calls. |
3049 | if (shouldGuaranteeTCO(CallConv, |
3050 | MF.getTarget().Options.GuaranteedTailCallOpt)) |
3051 | StackSize = GetAlignedArgumentStackSize(StackSize, DAG); |
3052 | |
3053 | // If the function takes variable number of arguments, make a frame index for |
3054 | // the start of the first vararg value... for expansion of llvm.va_start. We |
3055 | // can skip this if there are no va_start calls. |
3056 | if (MFI.hasVAStart() && |
3057 | (Is64Bit || (CallConv != CallingConv::X86_FastCall && |
3058 | CallConv != CallingConv::X86_ThisCall))) { |
3059 | FuncInfo->setVarArgsFrameIndex(MFI.CreateFixedObject(1, StackSize, true)); |
3060 | } |
3061 | |
3062 | // Figure out if XMM registers are in use. |
3063 | assert(!(Subtarget.useSoftFloat() &&((!(Subtarget.useSoftFloat() && Fn->hasFnAttribute (Attribute::NoImplicitFloat)) && "SSE register cannot be used when SSE is disabled!" ) ? static_cast<void> (0) : __assert_fail ("!(Subtarget.useSoftFloat() && Fn->hasFnAttribute(Attribute::NoImplicitFloat)) && \"SSE register cannot be used when SSE is disabled!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3065, __PRETTY_FUNCTION__)) |
3064 | Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&((!(Subtarget.useSoftFloat() && Fn->hasFnAttribute (Attribute::NoImplicitFloat)) && "SSE register cannot be used when SSE is disabled!" ) ? static_cast<void> (0) : __assert_fail ("!(Subtarget.useSoftFloat() && Fn->hasFnAttribute(Attribute::NoImplicitFloat)) && \"SSE register cannot be used when SSE is disabled!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3065, __PRETTY_FUNCTION__)) |
3065 | "SSE register cannot be used when SSE is disabled!")((!(Subtarget.useSoftFloat() && Fn->hasFnAttribute (Attribute::NoImplicitFloat)) && "SSE register cannot be used when SSE is disabled!" ) ? static_cast<void> (0) : __assert_fail ("!(Subtarget.useSoftFloat() && Fn->hasFnAttribute(Attribute::NoImplicitFloat)) && \"SSE register cannot be used when SSE is disabled!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3065, __PRETTY_FUNCTION__)); |
3066 | |
3067 | // 64-bit calling conventions support varargs and register parameters, so we |
3068 | // have to do extra work to spill them in the prologue. |
3069 | if (Is64Bit && isVarArg && MFI.hasVAStart()) { |
3070 | // Find the first unallocated argument registers. |
3071 | ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget); |
3072 | ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget); |
3073 | unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs); |
3074 | unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs); |
3075 | assert(!(NumXMMRegs && !Subtarget.hasSSE1()) &&((!(NumXMMRegs && !Subtarget.hasSSE1()) && "SSE register cannot be used when SSE is disabled!" ) ? static_cast<void> (0) : __assert_fail ("!(NumXMMRegs && !Subtarget.hasSSE1()) && \"SSE register cannot be used when SSE is disabled!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3076, __PRETTY_FUNCTION__)) |
3076 | "SSE register cannot be used when SSE is disabled!")((!(NumXMMRegs && !Subtarget.hasSSE1()) && "SSE register cannot be used when SSE is disabled!" ) ? static_cast<void> (0) : __assert_fail ("!(NumXMMRegs && !Subtarget.hasSSE1()) && \"SSE register cannot be used when SSE is disabled!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3076, __PRETTY_FUNCTION__)); |
3077 | |
3078 | // Gather all the live in physical registers. |
3079 | SmallVector<SDValue, 6> LiveGPRs; |
3080 | SmallVector<SDValue, 8> LiveXMMRegs; |
3081 | SDValue ALVal; |
3082 | for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) { |
3083 | unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass); |
3084 | LiveGPRs.push_back( |
3085 | DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64)); |
3086 | } |
3087 | if (!ArgXMMs.empty()) { |
3088 | unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass); |
3089 | ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8); |
3090 | for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) { |
3091 | unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass); |
3092 | LiveXMMRegs.push_back( |
3093 | DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32)); |
3094 | } |
3095 | } |
3096 | |
3097 | if (IsWin64) { |
3098 | // Get to the caller-allocated home save location. Add 8 to account |
3099 | // for the return address. |
3100 | int HomeOffset = TFI.getOffsetOfLocalArea() + 8; |
3101 | FuncInfo->setRegSaveFrameIndex( |
3102 | MFI.CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false)); |
3103 | // Fixup to set vararg frame on shadow area (4 x i64). |
3104 | if (NumIntRegs < 4) |
3105 | FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex()); |
3106 | } else { |
3107 | // For X86-64, if there are vararg parameters that are passed via |
3108 | // registers, then we must store them to their spots on the stack so |
3109 | // they may be loaded by dereferencing the result of va_next. |
3110 | FuncInfo->setVarArgsGPOffset(NumIntRegs * 8); |
3111 | FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16); |
3112 | FuncInfo->setRegSaveFrameIndex(MFI.CreateStackObject( |
3113 | ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false)); |
3114 | } |
3115 | |
3116 | // Store the integer parameter registers. |
3117 | SmallVector<SDValue, 8> MemOps; |
3118 | SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), |
3119 | getPointerTy(DAG.getDataLayout())); |
3120 | unsigned Offset = FuncInfo->getVarArgsGPOffset(); |
3121 | for (SDValue Val : LiveGPRs) { |
3122 | SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()), |
3123 | RSFIN, DAG.getIntPtrConstant(Offset, dl)); |
3124 | SDValue Store = |
3125 | DAG.getStore(Val.getValue(1), dl, Val, FIN, |
3126 | MachinePointerInfo::getFixedStack( |
3127 | DAG.getMachineFunction(), |
3128 | FuncInfo->getRegSaveFrameIndex(), Offset)); |
3129 | MemOps.push_back(Store); |
3130 | Offset += 8; |
3131 | } |
3132 | |
3133 | if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) { |
3134 | // Now store the XMM (fp + vector) parameter registers. |
3135 | SmallVector<SDValue, 12> SaveXMMOps; |
3136 | SaveXMMOps.push_back(Chain); |
3137 | SaveXMMOps.push_back(ALVal); |
3138 | SaveXMMOps.push_back(DAG.getIntPtrConstant( |
3139 | FuncInfo->getRegSaveFrameIndex(), dl)); |
3140 | SaveXMMOps.push_back(DAG.getIntPtrConstant( |
3141 | FuncInfo->getVarArgsFPOffset(), dl)); |
3142 | SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(), |
3143 | LiveXMMRegs.end()); |
3144 | MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl, |
3145 | MVT::Other, SaveXMMOps)); |
3146 | } |
3147 | |
3148 | if (!MemOps.empty()) |
3149 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps); |
3150 | } |
3151 | |
3152 | if (isVarArg && MFI.hasMustTailInVarArgFunc()) { |
3153 | // Find the largest legal vector type. |
3154 | MVT VecVT = MVT::Other; |
3155 | // FIXME: Only some x86_32 calling conventions support AVX512. |
3156 | if (Subtarget.hasAVX512() && |
3157 | (Is64Bit || (CallConv == CallingConv::X86_VectorCall || |
3158 | CallConv == CallingConv::Intel_OCL_BI))) |
3159 | VecVT = MVT::v16f32; |
3160 | else if (Subtarget.hasAVX()) |
3161 | VecVT = MVT::v8f32; |
3162 | else if (Subtarget.hasSSE2()) |
3163 | VecVT = MVT::v4f32; |
3164 | |
3165 | // We forward some GPRs and some vector types. |
3166 | SmallVector<MVT, 2> RegParmTypes; |
3167 | MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32; |
3168 | RegParmTypes.push_back(IntVT); |
3169 | if (VecVT != MVT::Other) |
3170 | RegParmTypes.push_back(VecVT); |
3171 | |
3172 | // Compute the set of forwarded registers. The rest are scratch. |
3173 | SmallVectorImpl<ForwardedRegister> &Forwards = |
3174 | FuncInfo->getForwardedMustTailRegParms(); |
3175 | CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86); |
3176 | |
3177 | // Conservatively forward AL on x86_64, since it might be used for varargs. |
3178 | if (Is64Bit && !CCInfo.isAllocated(X86::AL)) { |
3179 | unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass); |
3180 | Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8)); |
3181 | } |
3182 | |
3183 | // Copy all forwards from physical to virtual registers. |
3184 | for (ForwardedRegister &F : Forwards) { |
3185 | // FIXME: Can we use a less constrained schedule? |
3186 | SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT); |
3187 | F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT)); |
3188 | Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal); |
3189 | } |
3190 | } |
3191 | |
3192 | // Some CCs need callee pop. |
3193 | if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, |
3194 | MF.getTarget().Options.GuaranteedTailCallOpt)) { |
3195 | FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything. |
3196 | } else if (CallConv == CallingConv::X86_INTR && Ins.size() == 2) { |
3197 | // X86 interrupts must pop the error code (and the alignment padding) if |
3198 | // present. |
3199 | FuncInfo->setBytesToPopOnReturn(Is64Bit ? 16 : 4); |
3200 | } else { |
3201 | FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing. |
3202 | // If this is an sret function, the return should pop the hidden pointer. |
3203 | if (!Is64Bit && !canGuaranteeTCO(CallConv) && |
3204 | !Subtarget.getTargetTriple().isOSMSVCRT() && |
3205 | argsAreStructReturn(Ins, Subtarget.isTargetMCU()) == StackStructReturn) |
3206 | FuncInfo->setBytesToPopOnReturn(4); |
3207 | } |
3208 | |
3209 | if (!Is64Bit) { |
3210 | // RegSaveFrameIndex is X86-64 only. |
3211 | FuncInfo->setRegSaveFrameIndex(0xAAAAAAA); |
3212 | if (CallConv == CallingConv::X86_FastCall || |
3213 | CallConv == CallingConv::X86_ThisCall) |
3214 | // fastcc functions can't have varargs. |
3215 | FuncInfo->setVarArgsFrameIndex(0xAAAAAAA); |
3216 | } |
3217 | |
3218 | FuncInfo->setArgumentStackSize(StackSize); |
3219 | |
3220 | if (WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo()) { |
3221 | EHPersonality Personality = classifyEHPersonality(Fn->getPersonalityFn()); |
3222 | if (Personality == EHPersonality::CoreCLR) { |
3223 | assert(Is64Bit)((Is64Bit) ? static_cast<void> (0) : __assert_fail ("Is64Bit" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3223, __PRETTY_FUNCTION__)); |
3224 | // TODO: Add a mechanism to frame lowering that will allow us to indicate |
3225 | // that we'd prefer this slot be allocated towards the bottom of the frame |
3226 | // (i.e. near the stack pointer after allocating the frame). Every |
3227 | // funclet needs a copy of this slot in its (mostly empty) frame, and the |
3228 | // offset from the bottom of this and each funclet's frame must be the |
3229 | // same, so the size of funclets' (mostly empty) frames is dictated by |
3230 | // how far this slot is from the bottom (since they allocate just enough |
3231 | // space to accommodate holding this slot at the correct offset). |
3232 | int PSPSymFI = MFI.CreateStackObject(8, 8, /*isSS=*/false); |
3233 | EHInfo->PSPSymFrameIdx = PSPSymFI; |
3234 | } |
3235 | } |
3236 | |
3237 | if (CallConv == CallingConv::X86_RegCall || |
3238 | Fn->hasFnAttribute("no_caller_saved_registers")) { |
3239 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
3240 | for (const auto &Pair : make_range(MRI.livein_begin(), MRI.livein_end())) |
3241 | MF.getRegInfo().disableCalleeSavedRegister(Pair.first); |
3242 | } |
3243 | |
3244 | return Chain; |
3245 | } |
3246 | |
3247 | SDValue X86TargetLowering::LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, |
3248 | SDValue Arg, const SDLoc &dl, |
3249 | SelectionDAG &DAG, |
3250 | const CCValAssign &VA, |
3251 | ISD::ArgFlagsTy Flags) const { |
3252 | unsigned LocMemOffset = VA.getLocMemOffset(); |
3253 | SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl); |
3254 | PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()), |
3255 | StackPtr, PtrOff); |
3256 | if (Flags.isByVal()) |
3257 | return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl); |
3258 | |
3259 | return DAG.getStore( |
3260 | Chain, dl, Arg, PtrOff, |
3261 | MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset)); |
3262 | } |
3263 | |
3264 | /// Emit a load of return address if tail call |
3265 | /// optimization is performed and it is required. |
3266 | SDValue X86TargetLowering::EmitTailCallLoadRetAddr( |
3267 | SelectionDAG &DAG, SDValue &OutRetAddr, SDValue Chain, bool IsTailCall, |
3268 | bool Is64Bit, int FPDiff, const SDLoc &dl) const { |
3269 | // Adjust the Return address stack slot. |
3270 | EVT VT = getPointerTy(DAG.getDataLayout()); |
3271 | OutRetAddr = getReturnAddressFrameIndex(DAG); |
3272 | |
3273 | // Load the "old" Return address. |
3274 | OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo()); |
3275 | return SDValue(OutRetAddr.getNode(), 1); |
3276 | } |
3277 | |
3278 | /// Emit a store of the return address if tail call |
3279 | /// optimization is performed and it is required (FPDiff!=0). |
3280 | static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF, |
3281 | SDValue Chain, SDValue RetAddrFrIdx, |
3282 | EVT PtrVT, unsigned SlotSize, |
3283 | int FPDiff, const SDLoc &dl) { |
3284 | // Store the return address to the appropriate stack slot. |
3285 | if (!FPDiff) return Chain; |
3286 | // Calculate the new stack slot for the return address. |
3287 | int NewReturnAddrFI = |
3288 | MF.getFrameInfo().CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize, |
3289 | false); |
3290 | SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT); |
3291 | Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx, |
3292 | MachinePointerInfo::getFixedStack( |
3293 | DAG.getMachineFunction(), NewReturnAddrFI)); |
3294 | return Chain; |
3295 | } |
3296 | |
3297 | /// Returns a vector_shuffle mask for an movs{s|d}, movd |
3298 | /// operation of specified width. |
3299 | static SDValue getMOVL(SelectionDAG &DAG, const SDLoc &dl, MVT VT, SDValue V1, |
3300 | SDValue V2) { |
3301 | unsigned NumElems = VT.getVectorNumElements(); |
3302 | SmallVector<int, 8> Mask; |
3303 | Mask.push_back(NumElems); |
3304 | for (unsigned i = 1; i != NumElems; ++i) |
3305 | Mask.push_back(i); |
3306 | return DAG.getVectorShuffle(VT, dl, V1, V2, Mask); |
3307 | } |
3308 | |
3309 | SDValue |
3310 | X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, |
3311 | SmallVectorImpl<SDValue> &InVals) const { |
3312 | SelectionDAG &DAG = CLI.DAG; |
3313 | SDLoc &dl = CLI.DL; |
3314 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; |
3315 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; |
3316 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; |
3317 | SDValue Chain = CLI.Chain; |
3318 | SDValue Callee = CLI.Callee; |
3319 | CallingConv::ID CallConv = CLI.CallConv; |
3320 | bool &isTailCall = CLI.IsTailCall; |
3321 | bool isVarArg = CLI.IsVarArg; |
3322 | |
3323 | MachineFunction &MF = DAG.getMachineFunction(); |
3324 | bool Is64Bit = Subtarget.is64Bit(); |
3325 | bool IsWin64 = Subtarget.isCallingConvWin64(CallConv); |
3326 | StructReturnType SR = callIsStructReturn(Outs, Subtarget.isTargetMCU()); |
3327 | bool IsSibcall = false; |
3328 | X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>(); |
3329 | auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls"); |
3330 | const CallInst *CI = |
3331 | CLI.CS ? dyn_cast<CallInst>(CLI.CS->getInstruction()) : nullptr; |
3332 | const Function *Fn = CI ? CI->getCalledFunction() : nullptr; |
3333 | bool HasNCSR = (CI && CI->hasFnAttr("no_caller_saved_registers")) || |
3334 | (Fn && Fn->hasFnAttribute("no_caller_saved_registers")); |
3335 | |
3336 | if (CallConv == CallingConv::X86_INTR) |
3337 | report_fatal_error("X86 interrupts may not be called directly"); |
3338 | |
3339 | if (Attr.getValueAsString() == "true") |
3340 | isTailCall = false; |
3341 | |
3342 | if (Subtarget.isPICStyleGOT() && |
3343 | !MF.getTarget().Options.GuaranteedTailCallOpt) { |
3344 | // If we are using a GOT, disable tail calls to external symbols with |
3345 | // default visibility. Tail calling such a symbol requires using a GOT |
3346 | // relocation, which forces early binding of the symbol. This breaks code |
3347 | // that require lazy function symbol resolution. Using musttail or |
3348 | // GuaranteedTailCallOpt will override this. |
3349 | GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee); |
3350 | if (!G || (!G->getGlobal()->hasLocalLinkage() && |
3351 | G->getGlobal()->hasDefaultVisibility())) |
3352 | isTailCall = false; |
3353 | } |
3354 | |
3355 | bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall(); |
3356 | if (IsMustTail) { |
3357 | // Force this to be a tail call. The verifier rules are enough to ensure |
3358 | // that we can lower this successfully without moving the return address |
3359 | // around. |
3360 | isTailCall = true; |
3361 | } else if (isTailCall) { |
3362 | // Check if it's really possible to do a tail call. |
3363 | isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, |
3364 | isVarArg, SR != NotStructReturn, |
3365 | MF.getFunction()->hasStructRetAttr(), CLI.RetTy, |
3366 | Outs, OutVals, Ins, DAG); |
3367 | |
3368 | // Sibcalls are automatically detected tailcalls which do not require |
3369 | // ABI changes. |
3370 | if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall) |
3371 | IsSibcall = true; |
3372 | |
3373 | if (isTailCall) |
3374 | ++NumTailCalls; |
3375 | } |
3376 | |
3377 | assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&((!(isVarArg && canGuaranteeTCO(CallConv)) && "Var args not supported with calling convention fastcc, ghc or hipe" ) ? static_cast<void> (0) : __assert_fail ("!(isVarArg && canGuaranteeTCO(CallConv)) && \"Var args not supported with calling convention fastcc, ghc or hipe\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3378, __PRETTY_FUNCTION__)) |
3378 | "Var args not supported with calling convention fastcc, ghc or hipe")((!(isVarArg && canGuaranteeTCO(CallConv)) && "Var args not supported with calling convention fastcc, ghc or hipe" ) ? static_cast<void> (0) : __assert_fail ("!(isVarArg && canGuaranteeTCO(CallConv)) && \"Var args not supported with calling convention fastcc, ghc or hipe\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3378, __PRETTY_FUNCTION__)); |
3379 | |
3380 | // Analyze operands of the call, assigning locations to each operand. |
3381 | SmallVector<CCValAssign, 16> ArgLocs; |
3382 | CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext()); |
3383 | |
3384 | // Allocate shadow area for Win64. |
3385 | if (IsWin64) |
3386 | CCInfo.AllocateStack(32, 8); |
3387 | |
3388 | CCInfo.AnalyzeArguments(Outs, CC_X86); |
3389 | |
3390 | // In vectorcall calling convention a second pass is required for the HVA |
3391 | // types. |
3392 | if (CallingConv::X86_VectorCall == CallConv) { |
3393 | CCInfo.AnalyzeArgumentsSecondPass(Outs, CC_X86); |
3394 | } |
3395 | |
3396 | // Get a count of how many bytes are to be pushed on the stack. |
3397 | unsigned NumBytes = CCInfo.getAlignedCallFrameSize(); |
3398 | if (IsSibcall) |
3399 | // This is a sibcall. The memory operands are available in caller's |
3400 | // own caller's stack. |
3401 | NumBytes = 0; |
3402 | else if (MF.getTarget().Options.GuaranteedTailCallOpt && |
3403 | canGuaranteeTCO(CallConv)) |
3404 | NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG); |
3405 | |
3406 | int FPDiff = 0; |
3407 | if (isTailCall && !IsSibcall && !IsMustTail) { |
3408 | // Lower arguments at fp - stackoffset + fpdiff. |
3409 | unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn(); |
3410 | |
3411 | FPDiff = NumBytesCallerPushed - NumBytes; |
3412 | |
3413 | // Set the delta of movement of the returnaddr stackslot. |
3414 | // But only set if delta is greater than previous delta. |
3415 | if (FPDiff < X86Info->getTCReturnAddrDelta()) |
3416 | X86Info->setTCReturnAddrDelta(FPDiff); |
3417 | } |
3418 | |
3419 | unsigned NumBytesToPush = NumBytes; |
3420 | unsigned NumBytesToPop = NumBytes; |
3421 | |
3422 | // If we have an inalloca argument, all stack space has already been allocated |
3423 | // for us and be right at the top of the stack. We don't support multiple |
3424 | // arguments passed in memory when using inalloca. |
3425 | if (!Outs.empty() && Outs.back().Flags.isInAlloca()) { |
3426 | NumBytesToPush = 0; |
3427 | if (!ArgLocs.back().isMemLoc()) |
3428 | report_fatal_error("cannot use inalloca attribute on a register " |
3429 | "parameter"); |
3430 | if (ArgLocs.back().getLocMemOffset() != 0) |
3431 | report_fatal_error("any parameter with the inalloca attribute must be " |
3432 | "the only memory argument"); |
3433 | } |
3434 | |
3435 | if (!IsSibcall) |
3436 | Chain = DAG.getCALLSEQ_START(Chain, NumBytesToPush, |
3437 | NumBytes - NumBytesToPush, dl); |
3438 | |
3439 | SDValue RetAddrFrIdx; |
3440 | // Load return address for tail calls. |
3441 | if (isTailCall && FPDiff) |
3442 | Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, |
3443 | Is64Bit, FPDiff, dl); |
3444 | |
3445 | SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; |
3446 | SmallVector<SDValue, 8> MemOpChains; |
3447 | SDValue StackPtr; |
3448 | |
3449 | // The next loop assumes that the locations are in the same order of the |
3450 | // input arguments. |
3451 | assert(isSortedByValueNo(ArgLocs) &&((isSortedByValueNo(ArgLocs) && "Argument Location list must be sorted before lowering" ) ? static_cast<void> (0) : __assert_fail ("isSortedByValueNo(ArgLocs) && \"Argument Location list must be sorted before lowering\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3452, __PRETTY_FUNCTION__)) |
3452 | "Argument Location list must be sorted before lowering")((isSortedByValueNo(ArgLocs) && "Argument Location list must be sorted before lowering" ) ? static_cast<void> (0) : __assert_fail ("isSortedByValueNo(ArgLocs) && \"Argument Location list must be sorted before lowering\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3452, __PRETTY_FUNCTION__)); |
3453 | |
3454 | // Walk the register/memloc assignments, inserting copies/loads. In the case |
3455 | // of tail call optimization arguments are handle later. |
3456 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
3457 | for (unsigned I = 0, OutIndex = 0, E = ArgLocs.size(); I != E; |
3458 | ++I, ++OutIndex) { |
3459 | assert(OutIndex < Outs.size() && "Invalid Out index")((OutIndex < Outs.size() && "Invalid Out index") ? static_cast<void> (0) : __assert_fail ("OutIndex < Outs.size() && \"Invalid Out index\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3459, __PRETTY_FUNCTION__)); |
3460 | // Skip inalloca arguments, they have already been written. |
3461 | ISD::ArgFlagsTy Flags = Outs[OutIndex].Flags; |
3462 | if (Flags.isInAlloca()) |
3463 | continue; |
3464 | |
3465 | CCValAssign &VA = ArgLocs[I]; |
3466 | EVT RegVT = VA.getLocVT(); |
3467 | SDValue Arg = OutVals[OutIndex]; |
3468 | bool isByVal = Flags.isByVal(); |
3469 | |
3470 | // Promote the value if needed. |
3471 | switch (VA.getLocInfo()) { |
3472 | default: llvm_unreachable("Unknown loc info!")::llvm::llvm_unreachable_internal("Unknown loc info!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3472); |
3473 | case CCValAssign::Full: break; |
3474 | case CCValAssign::SExt: |
3475 | Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg); |
3476 | break; |
3477 | case CCValAssign::ZExt: |
3478 | Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg); |
3479 | break; |
3480 | case CCValAssign::AExt: |
3481 | if (Arg.getValueType().isVector() && |
3482 | Arg.getValueType().getVectorElementType() == MVT::i1) |
3483 | Arg = lowerMasksToReg(Arg, RegVT, dl, DAG); |
3484 | else if (RegVT.is128BitVector()) { |
3485 | // Special case: passing MMX values in XMM registers. |
3486 | Arg = DAG.getBitcast(MVT::i64, Arg); |
3487 | Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg); |
3488 | Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg); |
3489 | } else |
3490 | Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg); |
3491 | break; |
3492 | case CCValAssign::BCvt: |
3493 | Arg = DAG.getBitcast(RegVT, Arg); |
3494 | break; |
3495 | case CCValAssign::Indirect: { |
3496 | // Store the argument. |
3497 | SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT()); |
3498 | int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex(); |
3499 | Chain = DAG.getStore( |
3500 | Chain, dl, Arg, SpillSlot, |
3501 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI)); |
3502 | Arg = SpillSlot; |
3503 | break; |
3504 | } |
3505 | } |
3506 | |
3507 | if (VA.needsCustom()) { |
3508 | assert(VA.getValVT() == MVT::v64i1 &&((VA.getValVT() == MVT::v64i1 && "Currently the only custom case is when we split v64i1 to 2 regs" ) ? static_cast<void> (0) : __assert_fail ("VA.getValVT() == MVT::v64i1 && \"Currently the only custom case is when we split v64i1 to 2 regs\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3509, __PRETTY_FUNCTION__)) |
3509 | "Currently the only custom case is when we split v64i1 to 2 regs")((VA.getValVT() == MVT::v64i1 && "Currently the only custom case is when we split v64i1 to 2 regs" ) ? static_cast<void> (0) : __assert_fail ("VA.getValVT() == MVT::v64i1 && \"Currently the only custom case is when we split v64i1 to 2 regs\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3509, __PRETTY_FUNCTION__)); |
3510 | // Split v64i1 value into two registers |
3511 | Passv64i1ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++I], |
3512 | Subtarget); |
3513 | } else if (VA.isRegLoc()) { |
3514 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); |
3515 | if (isVarArg && IsWin64) { |
3516 | // Win64 ABI requires argument XMM reg to be copied to the corresponding |
3517 | // shadow reg if callee is a varargs function. |
3518 | unsigned ShadowReg = 0; |
3519 | switch (VA.getLocReg()) { |
3520 | case X86::XMM0: ShadowReg = X86::RCX; break; |
3521 | case X86::XMM1: ShadowReg = X86::RDX; break; |
3522 | case X86::XMM2: ShadowReg = X86::R8; break; |
3523 | case X86::XMM3: ShadowReg = X86::R9; break; |
3524 | } |
3525 | if (ShadowReg) |
3526 | RegsToPass.push_back(std::make_pair(ShadowReg, Arg)); |
3527 | } |
3528 | } else if (!IsSibcall && (!isTailCall || isByVal)) { |
3529 | assert(VA.isMemLoc())((VA.isMemLoc()) ? static_cast<void> (0) : __assert_fail ("VA.isMemLoc()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3529, __PRETTY_FUNCTION__)); |
3530 | if (!StackPtr.getNode()) |
3531 | StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(), |
3532 | getPointerTy(DAG.getDataLayout())); |
3533 | MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg, |
3534 | dl, DAG, VA, Flags)); |
3535 | } |
3536 | } |
3537 | |
3538 | if (!MemOpChains.empty()) |
3539 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains); |
3540 | |
3541 | if (Subtarget.isPICStyleGOT()) { |
3542 | // ELF / PIC requires GOT in the EBX register before function calls via PLT |
3543 | // GOT pointer. |
3544 | if (!isTailCall) { |
3545 | RegsToPass.push_back(std::make_pair( |
3546 | unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), |
3547 | getPointerTy(DAG.getDataLayout())))); |
3548 | } else { |
3549 | // If we are tail calling and generating PIC/GOT style code load the |
3550 | // address of the callee into ECX. The value in ecx is used as target of |
3551 | // the tail jump. This is done to circumvent the ebx/callee-saved problem |
3552 | // for tail calls on PIC/GOT architectures. Normally we would just put the |
3553 | // address of GOT into ebx and then call target@PLT. But for tail calls |
3554 | // ebx would be restored (since ebx is callee saved) before jumping to the |
3555 | // target@PLT. |
3556 | |
3557 | // Note: The actual moving to ECX is done further down. |
3558 | GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee); |
3559 | if (G && !G->getGlobal()->hasLocalLinkage() && |
3560 | G->getGlobal()->hasDefaultVisibility()) |
3561 | Callee = LowerGlobalAddress(Callee, DAG); |
3562 | else if (isa<ExternalSymbolSDNode>(Callee)) |
3563 | Callee = LowerExternalSymbol(Callee, DAG); |
3564 | } |
3565 | } |
3566 | |
3567 | if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) { |
3568 | // From AMD64 ABI document: |
3569 | // For calls that may call functions that use varargs or stdargs |
3570 | // (prototype-less calls or calls to functions containing ellipsis (...) in |
3571 | // the declaration) %al is used as hidden argument to specify the number |
3572 | // of SSE registers used. The contents of %al do not need to match exactly |
3573 | // the number of registers, but must be an ubound on the number of SSE |
3574 | // registers used and is in the range 0 - 8 inclusive. |
3575 | |
3576 | // Count the number of XMM registers allocated. |
3577 | static const MCPhysReg XMMArgRegs[] = { |
3578 | X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, |
3579 | X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7 |
3580 | }; |
3581 | unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs); |
3582 | assert((Subtarget.hasSSE1() || !NumXMMRegs)(((Subtarget.hasSSE1() || !NumXMMRegs) && "SSE registers cannot be used when SSE is disabled" ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasSSE1() || !NumXMMRegs) && \"SSE registers cannot be used when SSE is disabled\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3583, __PRETTY_FUNCTION__)) |
3583 | && "SSE registers cannot be used when SSE is disabled")(((Subtarget.hasSSE1() || !NumXMMRegs) && "SSE registers cannot be used when SSE is disabled" ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasSSE1() || !NumXMMRegs) && \"SSE registers cannot be used when SSE is disabled\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3583, __PRETTY_FUNCTION__)); |
3584 | |
3585 | RegsToPass.push_back(std::make_pair(unsigned(X86::AL), |
3586 | DAG.getConstant(NumXMMRegs, dl, |
3587 | MVT::i8))); |
3588 | } |
3589 | |
3590 | if (isVarArg && IsMustTail) { |
3591 | const auto &Forwards = X86Info->getForwardedMustTailRegParms(); |
3592 | for (const auto &F : Forwards) { |
3593 | SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT); |
3594 | RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val)); |
3595 | } |
3596 | } |
3597 | |
3598 | // For tail calls lower the arguments to the 'real' stack slots. Sibcalls |
3599 | // don't need this because the eligibility check rejects calls that require |
3600 | // shuffling arguments passed in memory. |
3601 | if (!IsSibcall && isTailCall) { |
3602 | // Force all the incoming stack arguments to be loaded from the stack |
3603 | // before any new outgoing arguments are stored to the stack, because the |
3604 | // outgoing stack slots may alias the incoming argument stack slots, and |
3605 | // the alias isn't otherwise explicit. This is slightly more conservative |
3606 | // than necessary, because it means that each store effectively depends |
3607 | // on every argument instead of just those arguments it would clobber. |
3608 | SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain); |
3609 | |
3610 | SmallVector<SDValue, 8> MemOpChains2; |
3611 | SDValue FIN; |
3612 | int FI = 0; |
3613 | for (unsigned I = 0, OutsIndex = 0, E = ArgLocs.size(); I != E; |
3614 | ++I, ++OutsIndex) { |
3615 | CCValAssign &VA = ArgLocs[I]; |
3616 | |
3617 | if (VA.isRegLoc()) { |
3618 | if (VA.needsCustom()) { |
3619 | assert((CallConv == CallingConv::X86_RegCall) &&(((CallConv == CallingConv::X86_RegCall) && "Expecting custom case only in regcall calling convention" ) ? static_cast<void> (0) : __assert_fail ("(CallConv == CallingConv::X86_RegCall) && \"Expecting custom case only in regcall calling convention\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3620, __PRETTY_FUNCTION__)) |
3620 | "Expecting custom case only in regcall calling convention")(((CallConv == CallingConv::X86_RegCall) && "Expecting custom case only in regcall calling convention" ) ? static_cast<void> (0) : __assert_fail ("(CallConv == CallingConv::X86_RegCall) && \"Expecting custom case only in regcall calling convention\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3620, __PRETTY_FUNCTION__)); |
3621 | // This means that we are in special case where one argument was |
3622 | // passed through two register locations - Skip the next location |
3623 | ++I; |
3624 | } |
3625 | |
3626 | continue; |
3627 | } |
3628 | |
3629 | assert(VA.isMemLoc())((VA.isMemLoc()) ? static_cast<void> (0) : __assert_fail ("VA.isMemLoc()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3629, __PRETTY_FUNCTION__)); |
3630 | SDValue Arg = OutVals[OutsIndex]; |
3631 | ISD::ArgFlagsTy Flags = Outs[OutsIndex].Flags; |
3632 | // Skip inalloca arguments. They don't require any work. |
3633 | if (Flags.isInAlloca()) |
3634 | continue; |
3635 | // Create frame index. |
3636 | int32_t Offset = VA.getLocMemOffset()+FPDiff; |
3637 | uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8; |
3638 | FI = MF.getFrameInfo().CreateFixedObject(OpSize, Offset, true); |
3639 | FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); |
3640 | |
3641 | if (Flags.isByVal()) { |
3642 | // Copy relative to framepointer. |
3643 | SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl); |
3644 | if (!StackPtr.getNode()) |
3645 | StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(), |
3646 | getPointerTy(DAG.getDataLayout())); |
3647 | Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()), |
3648 | StackPtr, Source); |
3649 | |
3650 | MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, |
3651 | ArgChain, |
3652 | Flags, DAG, dl)); |
3653 | } else { |
3654 | // Store relative to framepointer. |
3655 | MemOpChains2.push_back(DAG.getStore( |
3656 | ArgChain, dl, Arg, FIN, |
3657 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI))); |
3658 | } |
3659 | } |
3660 | |
3661 | if (!MemOpChains2.empty()) |
3662 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2); |
3663 | |
3664 | // Store the return address to the appropriate stack slot. |
3665 | Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, |
3666 | getPointerTy(DAG.getDataLayout()), |
3667 | RegInfo->getSlotSize(), FPDiff, dl); |
3668 | } |
3669 | |
3670 | // Build a sequence of copy-to-reg nodes chained together with token chain |
3671 | // and flag operands which copy the outgoing args into registers. |
3672 | SDValue InFlag; |
3673 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { |
3674 | Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, |
3675 | RegsToPass[i].second, InFlag); |
3676 | InFlag = Chain.getValue(1); |
3677 | } |
3678 | |
3679 | if (DAG.getTarget().getCodeModel() == CodeModel::Large) { |
3680 | assert(Is64Bit && "Large code model is only legal in 64-bit mode.")((Is64Bit && "Large code model is only legal in 64-bit mode." ) ? static_cast<void> (0) : __assert_fail ("Is64Bit && \"Large code model is only legal in 64-bit mode.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3680, __PRETTY_FUNCTION__)); |
3681 | // In the 64-bit large code model, we have to make all calls |
3682 | // through a register, since the call instruction's 32-bit |
3683 | // pc-relative offset may not be large enough to hold the whole |
3684 | // address. |
3685 | } else if (Callee->getOpcode() == ISD::GlobalAddress) { |
3686 | // If the callee is a GlobalAddress node (quite common, every direct call |
3687 | // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack |
3688 | // it. |
3689 | GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee); |
3690 | |
3691 | // We should use extra load for direct calls to dllimported functions in |
3692 | // non-JIT mode. |
3693 | const GlobalValue *GV = G->getGlobal(); |
3694 | if (!GV->hasDLLImportStorageClass()) { |
3695 | unsigned char OpFlags = Subtarget.classifyGlobalFunctionReference(GV); |
3696 | |
3697 | Callee = DAG.getTargetGlobalAddress( |
3698 | GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags); |
3699 | |
3700 | if (OpFlags == X86II::MO_GOTPCREL) { |
3701 | // Add a wrapper. |
3702 | Callee = DAG.getNode(X86ISD::WrapperRIP, dl, |
3703 | getPointerTy(DAG.getDataLayout()), Callee); |
3704 | // Add extra indirection |
3705 | Callee = DAG.getLoad( |
3706 | getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee, |
3707 | MachinePointerInfo::getGOT(DAG.getMachineFunction())); |
3708 | } |
3709 | } |
3710 | } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { |
3711 | const Module *Mod = DAG.getMachineFunction().getFunction()->getParent(); |
3712 | unsigned char OpFlags = |
3713 | Subtarget.classifyGlobalFunctionReference(nullptr, *Mod); |
3714 | |
3715 | Callee = DAG.getTargetExternalSymbol( |
3716 | S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags); |
3717 | } else if (Subtarget.isTarget64BitILP32() && |
3718 | Callee->getValueType(0) == MVT::i32) { |
3719 | // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI |
3720 | Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee); |
3721 | } |
3722 | |
3723 | // Returns a chain & a flag for retval copy to use. |
3724 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); |
3725 | SmallVector<SDValue, 8> Ops; |
3726 | |
3727 | if (!IsSibcall && isTailCall) { |
3728 | Chain = DAG.getCALLSEQ_END(Chain, |
3729 | DAG.getIntPtrConstant(NumBytesToPop, dl, true), |
3730 | DAG.getIntPtrConstant(0, dl, true), InFlag, dl); |
3731 | InFlag = Chain.getValue(1); |
3732 | } |
3733 | |
3734 | Ops.push_back(Chain); |
3735 | Ops.push_back(Callee); |
3736 | |
3737 | if (isTailCall) |
3738 | Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32)); |
3739 | |
3740 | // Add argument registers to the end of the list so that they are known live |
3741 | // into the call. |
3742 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) |
3743 | Ops.push_back(DAG.getRegister(RegsToPass[i].first, |
3744 | RegsToPass[i].second.getValueType())); |
3745 | |
3746 | // Add a register mask operand representing the call-preserved registers. |
3747 | // If HasNCSR is asserted (attribute NoCallerSavedRegisters exists) then we |
3748 | // set X86_INTR calling convention because it has the same CSR mask |
3749 | // (same preserved registers). |
3750 | const uint32_t *Mask = RegInfo->getCallPreservedMask( |
3751 | MF, HasNCSR ? (CallingConv::ID)CallingConv::X86_INTR : CallConv); |
3752 | assert(Mask && "Missing call preserved mask for calling convention")((Mask && "Missing call preserved mask for calling convention" ) ? static_cast<void> (0) : __assert_fail ("Mask && \"Missing call preserved mask for calling convention\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3752, __PRETTY_FUNCTION__)); |
3753 | |
3754 | // If this is an invoke in a 32-bit function using a funclet-based |
3755 | // personality, assume the function clobbers all registers. If an exception |
3756 | // is thrown, the runtime will not restore CSRs. |
3757 | // FIXME: Model this more precisely so that we can register allocate across |
3758 | // the normal edge and spill and fill across the exceptional edge. |
3759 | if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) { |
3760 | const Function *CallerFn = MF.getFunction(); |
3761 | EHPersonality Pers = |
3762 | CallerFn->hasPersonalityFn() |
3763 | ? classifyEHPersonality(CallerFn->getPersonalityFn()) |
3764 | : EHPersonality::Unknown; |
3765 | if (isFuncletEHPersonality(Pers)) |
3766 | Mask = RegInfo->getNoPreservedMask(); |
3767 | } |
3768 | |
3769 | // Define a new register mask from the existing mask. |
3770 | uint32_t *RegMask = nullptr; |
3771 | |
3772 | // In some calling conventions we need to remove the used physical registers |
3773 | // from the reg mask. |
3774 | if (CallConv == CallingConv::X86_RegCall || HasNCSR) { |
3775 | const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); |
3776 | |
3777 | // Allocate a new Reg Mask and copy Mask. |
3778 | RegMask = MF.allocateRegisterMask(TRI->getNumRegs()); |
3779 | unsigned RegMaskSize = (TRI->getNumRegs() + 31) / 32; |
3780 | memcpy(RegMask, Mask, sizeof(uint32_t) * RegMaskSize); |
3781 | |
3782 | // Make sure all sub registers of the argument registers are reset |
3783 | // in the RegMask. |
3784 | for (auto const &RegPair : RegsToPass) |
3785 | for (MCSubRegIterator SubRegs(RegPair.first, TRI, /*IncludeSelf=*/true); |
3786 | SubRegs.isValid(); ++SubRegs) |
3787 | RegMask[*SubRegs / 32] &= ~(1u << (*SubRegs % 32)); |
3788 | |
3789 | // Create the RegMask Operand according to our updated mask. |
3790 | Ops.push_back(DAG.getRegisterMask(RegMask)); |
3791 | } else { |
3792 | // Create the RegMask Operand according to the static mask. |
3793 | Ops.push_back(DAG.getRegisterMask(Mask)); |
3794 | } |
3795 | |
3796 | if (InFlag.getNode()) |
3797 | Ops.push_back(InFlag); |
3798 | |
3799 | if (isTailCall) { |
3800 | // We used to do: |
3801 | //// If this is the first return lowered for this function, add the regs |
3802 | //// to the liveout set for the function. |
3803 | // This isn't right, although it's probably harmless on x86; liveouts |
3804 | // should be computed from returns not tail calls. Consider a void |
3805 | // function making a tail call to a function returning int. |
3806 | MF.getFrameInfo().setHasTailCall(); |
3807 | return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops); |
3808 | } |
3809 | |
3810 | Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops); |
3811 | InFlag = Chain.getValue(1); |
3812 | |
3813 | // Create the CALLSEQ_END node. |
3814 | unsigned NumBytesForCalleeToPop; |
3815 | if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, |
3816 | DAG.getTarget().Options.GuaranteedTailCallOpt)) |
3817 | NumBytesForCalleeToPop = NumBytes; // Callee pops everything |
3818 | else if (!Is64Bit && !canGuaranteeTCO(CallConv) && |
3819 | !Subtarget.getTargetTriple().isOSMSVCRT() && |
3820 | SR == StackStructReturn) |
3821 | // If this is a call to a struct-return function, the callee |
3822 | // pops the hidden struct pointer, so we have to push it back. |
3823 | // This is common for Darwin/X86, Linux & Mingw32 targets. |
3824 | // For MSVC Win32 targets, the caller pops the hidden struct pointer. |
3825 | NumBytesForCalleeToPop = 4; |
3826 | else |
3827 | NumBytesForCalleeToPop = 0; // Callee pops nothing. |
3828 | |
3829 | if (CLI.DoesNotReturn && !getTargetMachine().Options.TrapUnreachable) { |
3830 | // No need to reset the stack after the call if the call doesn't return. To |
3831 | // make the MI verify, we'll pretend the callee does it for us. |
3832 | NumBytesForCalleeToPop = NumBytes; |
3833 | } |
3834 | |
3835 | // Returns a flag for retval copy to use. |
3836 | if (!IsSibcall) { |
3837 | Chain = DAG.getCALLSEQ_END(Chain, |
3838 | DAG.getIntPtrConstant(NumBytesToPop, dl, true), |
3839 | DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl, |
3840 | true), |
3841 | InFlag, dl); |
3842 | InFlag = Chain.getValue(1); |
3843 | } |
3844 | |
3845 | // Handle result values, copying them out of physregs into vregs that we |
3846 | // return. |
3847 | return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG, |
3848 | InVals, RegMask); |
3849 | } |
3850 | |
3851 | //===----------------------------------------------------------------------===// |
3852 | // Fast Calling Convention (tail call) implementation |
3853 | //===----------------------------------------------------------------------===// |
3854 | |
3855 | // Like std call, callee cleans arguments, convention except that ECX is |
3856 | // reserved for storing the tail called function address. Only 2 registers are |
3857 | // free for argument passing (inreg). Tail call optimization is performed |
3858 | // provided: |
3859 | // * tailcallopt is enabled |
3860 | // * caller/callee are fastcc |
3861 | // On X86_64 architecture with GOT-style position independent code only local |
3862 | // (within module) calls are supported at the moment. |
3863 | // To keep the stack aligned according to platform abi the function |
3864 | // GetAlignedArgumentStackSize ensures that argument delta is always multiples |
3865 | // of stack alignment. (Dynamic linkers need this - darwin's dyld for example) |
3866 | // If a tail called function callee has more arguments than the caller the |
3867 | // caller needs to make sure that there is room to move the RETADDR to. This is |
3868 | // achieved by reserving an area the size of the argument delta right after the |
3869 | // original RETADDR, but before the saved framepointer or the spilled registers |
3870 | // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4) |
3871 | // stack layout: |
3872 | // arg1 |
3873 | // arg2 |
3874 | // RETADDR |
3875 | // [ new RETADDR |
3876 | // move area ] |
3877 | // (possible EBP) |
3878 | // ESI |
3879 | // EDI |
3880 | // local1 .. |
3881 | |
3882 | /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align |
3883 | /// requirement. |
3884 | unsigned |
3885 | X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize, |
3886 | SelectionDAG& DAG) const { |
3887 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
3888 | const TargetFrameLowering &TFI = *Subtarget.getFrameLowering(); |
3889 | unsigned StackAlignment = TFI.getStackAlignment(); |
3890 | uint64_t AlignMask = StackAlignment - 1; |
3891 | int64_t Offset = StackSize; |
3892 | unsigned SlotSize = RegInfo->getSlotSize(); |
3893 | if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) { |
3894 | // Number smaller than 12 so just add the difference. |
3895 | Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask)); |
3896 | } else { |
3897 | // Mask out lower bits, add stackalignment once plus the 12 bytes. |
3898 | Offset = ((~AlignMask) & Offset) + StackAlignment + |
3899 | (StackAlignment-SlotSize); |
3900 | } |
3901 | return Offset; |
3902 | } |
3903 | |
3904 | /// Return true if the given stack call argument is already available in the |
3905 | /// same position (relatively) of the caller's incoming argument stack. |
3906 | static |
3907 | bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags, |
3908 | MachineFrameInfo &MFI, const MachineRegisterInfo *MRI, |
3909 | const X86InstrInfo *TII, const CCValAssign &VA) { |
3910 | unsigned Bytes = Arg.getValueSizeInBits() / 8; |
3911 | |
3912 | for (;;) { |
3913 | // Look through nodes that don't alter the bits of the incoming value. |
3914 | unsigned Op = Arg.getOpcode(); |
3915 | if (Op == ISD::ZERO_EXTEND || Op == ISD::ANY_EXTEND || Op == ISD::BITCAST) { |
3916 | Arg = Arg.getOperand(0); |
3917 | continue; |
3918 | } |
3919 | if (Op == ISD::TRUNCATE) { |
3920 | const SDValue &TruncInput = Arg.getOperand(0); |
3921 | if (TruncInput.getOpcode() == ISD::AssertZext && |
3922 | cast<VTSDNode>(TruncInput.getOperand(1))->getVT() == |
3923 | Arg.getValueType()) { |
3924 | Arg = TruncInput.getOperand(0); |
3925 | continue; |
3926 | } |
3927 | } |
3928 | break; |
3929 | } |
3930 | |
3931 | int FI = INT_MAX2147483647; |
3932 | if (Arg.getOpcode() == ISD::CopyFromReg) { |
3933 | unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg(); |
3934 | if (!TargetRegisterInfo::isVirtualRegister(VR)) |
3935 | return false; |
3936 | MachineInstr *Def = MRI->getVRegDef(VR); |
3937 | if (!Def) |
3938 | return false; |
3939 | if (!Flags.isByVal()) { |
3940 | if (!TII->isLoadFromStackSlot(*Def, FI)) |
3941 | return false; |
3942 | } else { |
3943 | unsigned Opcode = Def->getOpcode(); |
3944 | if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r || |
3945 | Opcode == X86::LEA64_32r) && |
3946 | Def->getOperand(1).isFI()) { |
3947 | FI = Def->getOperand(1).getIndex(); |
3948 | Bytes = Flags.getByValSize(); |
3949 | } else |
3950 | return false; |
3951 | } |
3952 | } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) { |
3953 | if (Flags.isByVal()) |
3954 | // ByVal argument is passed in as a pointer but it's now being |
3955 | // dereferenced. e.g. |
3956 | // define @foo(%struct.X* %A) { |
3957 | // tail call @bar(%struct.X* byval %A) |
3958 | // } |
3959 | return false; |
3960 | SDValue Ptr = Ld->getBasePtr(); |
3961 | FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr); |
3962 | if (!FINode) |
3963 | return false; |
3964 | FI = FINode->getIndex(); |
3965 | } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) { |
3966 | FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg); |
3967 | FI = FINode->getIndex(); |
3968 | Bytes = Flags.getByValSize(); |
3969 | } else |
3970 | return false; |
3971 | |
3972 | assert(FI != INT_MAX)((FI != 2147483647) ? static_cast<void> (0) : __assert_fail ("FI != INT_MAX", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 3972, __PRETTY_FUNCTION__)); |
3973 | if (!MFI.isFixedObjectIndex(FI)) |
3974 | return false; |
3975 | |
3976 | if (Offset != MFI.getObjectOffset(FI)) |
3977 | return false; |
3978 | |
3979 | if (VA.getLocVT().getSizeInBits() > Arg.getValueSizeInBits()) { |
3980 | // If the argument location is wider than the argument type, check that any |
3981 | // extension flags match. |
3982 | if (Flags.isZExt() != MFI.isObjectZExt(FI) || |
3983 | Flags.isSExt() != MFI.isObjectSExt(FI)) { |
3984 | return false; |
3985 | } |
3986 | } |
3987 | |
3988 | return Bytes == MFI.getObjectSize(FI); |
3989 | } |
3990 | |
3991 | /// Check whether the call is eligible for tail call optimization. Targets |
3992 | /// that want to do tail call optimization should implement this function. |
3993 | bool X86TargetLowering::IsEligibleForTailCallOptimization( |
3994 | SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg, |
3995 | bool isCalleeStructRet, bool isCallerStructRet, Type *RetTy, |
3996 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
3997 | const SmallVectorImpl<SDValue> &OutVals, |
3998 | const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const { |
3999 | if (!mayTailCallThisCC(CalleeCC)) |
4000 | return false; |
4001 | |
4002 | // If -tailcallopt is specified, make fastcc functions tail-callable. |
4003 | MachineFunction &MF = DAG.getMachineFunction(); |
4004 | const Function *CallerF = MF.getFunction(); |
4005 | |
4006 | // If the function return type is x86_fp80 and the callee return type is not, |
4007 | // then the FP_EXTEND of the call result is not a nop. It's not safe to |
4008 | // perform a tailcall optimization here. |
4009 | if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty()) |
4010 | return false; |
4011 | |
4012 | CallingConv::ID CallerCC = CallerF->getCallingConv(); |
4013 | bool CCMatch = CallerCC == CalleeCC; |
4014 | bool IsCalleeWin64 = Subtarget.isCallingConvWin64(CalleeCC); |
4015 | bool IsCallerWin64 = Subtarget.isCallingConvWin64(CallerCC); |
4016 | |
4017 | // Win64 functions have extra shadow space for argument homing. Don't do the |
4018 | // sibcall if the caller and callee have mismatched expectations for this |
4019 | // space. |
4020 | if (IsCalleeWin64 != IsCallerWin64) |
4021 | return false; |
4022 | |
4023 | if (DAG.getTarget().Options.GuaranteedTailCallOpt) { |
4024 | if (canGuaranteeTCO(CalleeCC) && CCMatch) |
4025 | return true; |
4026 | return false; |
4027 | } |
4028 | |
4029 | // Look for obvious safe cases to perform tail call optimization that do not |
4030 | // require ABI changes. This is what gcc calls sibcall. |
4031 | |
4032 | // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to |
4033 | // emit a special epilogue. |
4034 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
4035 | if (RegInfo->needsStackRealignment(MF)) |
4036 | return false; |
4037 | |
4038 | // Also avoid sibcall optimization if either caller or callee uses struct |
4039 | // return semantics. |
4040 | if (isCalleeStructRet || isCallerStructRet) |
4041 | return false; |
4042 | |
4043 | // Do not sibcall optimize vararg calls unless all arguments are passed via |
4044 | // registers. |
4045 | LLVMContext &C = *DAG.getContext(); |
4046 | if (isVarArg && !Outs.empty()) { |
4047 | // Optimizing for varargs on Win64 is unlikely to be safe without |
4048 | // additional testing. |
4049 | if (IsCalleeWin64 || IsCallerWin64) |
4050 | return false; |
4051 | |
4052 | SmallVector<CCValAssign, 16> ArgLocs; |
4053 | CCState CCInfo(CalleeCC, isVarArg, MF, ArgLocs, C); |
4054 | |
4055 | CCInfo.AnalyzeCallOperands(Outs, CC_X86); |
4056 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) |
4057 | if (!ArgLocs[i].isRegLoc()) |
4058 | return false; |
4059 | } |
4060 | |
4061 | // If the call result is in ST0 / ST1, it needs to be popped off the x87 |
4062 | // stack. Therefore, if it's not used by the call it is not safe to optimize |
4063 | // this into a sibcall. |
4064 | bool Unused = false; |
4065 | for (unsigned i = 0, e = Ins.size(); i != e; ++i) { |
4066 | if (!Ins[i].Used) { |
4067 | Unused = true; |
4068 | break; |
4069 | } |
4070 | } |
4071 | if (Unused) { |
4072 | SmallVector<CCValAssign, 16> RVLocs; |
4073 | CCState CCInfo(CalleeCC, false, MF, RVLocs, C); |
4074 | CCInfo.AnalyzeCallResult(Ins, RetCC_X86); |
4075 | for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) { |
4076 | CCValAssign &VA = RVLocs[i]; |
4077 | if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) |
4078 | return false; |
4079 | } |
4080 | } |
4081 | |
4082 | // Check that the call results are passed in the same way. |
4083 | if (!CCState::resultsCompatible(CalleeCC, CallerCC, MF, C, Ins, |
4084 | RetCC_X86, RetCC_X86)) |
4085 | return false; |
4086 | // The callee has to preserve all registers the caller needs to preserve. |
4087 | const X86RegisterInfo *TRI = Subtarget.getRegisterInfo(); |
4088 | const uint32_t *CallerPreserved = TRI->getCallPreservedMask(MF, CallerCC); |
4089 | if (!CCMatch) { |
4090 | const uint32_t *CalleePreserved = TRI->getCallPreservedMask(MF, CalleeCC); |
4091 | if (!TRI->regmaskSubsetEqual(CallerPreserved, CalleePreserved)) |
4092 | return false; |
4093 | } |
4094 | |
4095 | unsigned StackArgsSize = 0; |
4096 | |
4097 | // If the callee takes no arguments then go on to check the results of the |
4098 | // call. |
4099 | if (!Outs.empty()) { |
4100 | // Check if stack adjustment is needed. For now, do not do this if any |
4101 | // argument is passed on the stack. |
4102 | SmallVector<CCValAssign, 16> ArgLocs; |
4103 | CCState CCInfo(CalleeCC, isVarArg, MF, ArgLocs, C); |
4104 | |
4105 | // Allocate shadow area for Win64 |
4106 | if (IsCalleeWin64) |
4107 | CCInfo.AllocateStack(32, 8); |
4108 | |
4109 | CCInfo.AnalyzeCallOperands(Outs, CC_X86); |
4110 | StackArgsSize = CCInfo.getNextStackOffset(); |
4111 | |
4112 | if (CCInfo.getNextStackOffset()) { |
4113 | // Check if the arguments are already laid out in the right way as |
4114 | // the caller's fixed stack objects. |
4115 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
4116 | const MachineRegisterInfo *MRI = &MF.getRegInfo(); |
4117 | const X86InstrInfo *TII = Subtarget.getInstrInfo(); |
4118 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { |
4119 | CCValAssign &VA = ArgLocs[i]; |
4120 | SDValue Arg = OutVals[i]; |
4121 | ISD::ArgFlagsTy Flags = Outs[i].Flags; |
4122 | if (VA.getLocInfo() == CCValAssign::Indirect) |
4123 | return false; |
4124 | if (!VA.isRegLoc()) { |
4125 | if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags, |
4126 | MFI, MRI, TII, VA)) |
4127 | return false; |
4128 | } |
4129 | } |
4130 | } |
4131 | |
4132 | bool PositionIndependent = isPositionIndependent(); |
4133 | // If the tailcall address may be in a register, then make sure it's |
4134 | // possible to register allocate for it. In 32-bit, the call address can |
4135 | // only target EAX, EDX, or ECX since the tail call must be scheduled after |
4136 | // callee-saved registers are restored. These happen to be the same |
4137 | // registers used to pass 'inreg' arguments so watch out for those. |
4138 | if (!Subtarget.is64Bit() && ((!isa<GlobalAddressSDNode>(Callee) && |
4139 | !isa<ExternalSymbolSDNode>(Callee)) || |
4140 | PositionIndependent)) { |
4141 | unsigned NumInRegs = 0; |
4142 | // In PIC we need an extra register to formulate the address computation |
4143 | // for the callee. |
4144 | unsigned MaxInRegs = PositionIndependent ? 2 : 3; |
4145 | |
4146 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { |
4147 | CCValAssign &VA = ArgLocs[i]; |
4148 | if (!VA.isRegLoc()) |
4149 | continue; |
4150 | unsigned Reg = VA.getLocReg(); |
4151 | switch (Reg) { |
4152 | default: break; |
4153 | case X86::EAX: case X86::EDX: case X86::ECX: |
4154 | if (++NumInRegs == MaxInRegs) |
4155 | return false; |
4156 | break; |
4157 | } |
4158 | } |
4159 | } |
4160 | |
4161 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
4162 | if (!parametersInCSRMatch(MRI, CallerPreserved, ArgLocs, OutVals)) |
4163 | return false; |
4164 | } |
4165 | |
4166 | bool CalleeWillPop = |
4167 | X86::isCalleePop(CalleeCC, Subtarget.is64Bit(), isVarArg, |
4168 | MF.getTarget().Options.GuaranteedTailCallOpt); |
4169 | |
4170 | if (unsigned BytesToPop = |
4171 | MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn()) { |
4172 | // If we have bytes to pop, the callee must pop them. |
4173 | bool CalleePopMatches = CalleeWillPop && BytesToPop == StackArgsSize; |
4174 | if (!CalleePopMatches) |
4175 | return false; |
4176 | } else if (CalleeWillPop && StackArgsSize > 0) { |
4177 | // If we don't have bytes to pop, make sure the callee doesn't pop any. |
4178 | return false; |
4179 | } |
4180 | |
4181 | return true; |
4182 | } |
4183 | |
4184 | FastISel * |
4185 | X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo, |
4186 | const TargetLibraryInfo *libInfo) const { |
4187 | return X86::createFastISel(funcInfo, libInfo); |
4188 | } |
4189 | |
4190 | //===----------------------------------------------------------------------===// |
4191 | // Other Lowering Hooks |
4192 | //===----------------------------------------------------------------------===// |
4193 | |
4194 | static bool MayFoldLoad(SDValue Op) { |
4195 | return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode()); |
4196 | } |
4197 | |
4198 | static bool MayFoldIntoStore(SDValue Op) { |
4199 | return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin()); |
4200 | } |
4201 | |
4202 | static bool MayFoldIntoZeroExtend(SDValue Op) { |
4203 | if (Op.hasOneUse()) { |
4204 | unsigned Opcode = Op.getNode()->use_begin()->getOpcode(); |
4205 | return (ISD::ZERO_EXTEND == Opcode); |
4206 | } |
4207 | return false; |
4208 | } |
4209 | |
4210 | static bool isTargetShuffle(unsigned Opcode) { |
4211 | switch(Opcode) { |
4212 | default: return false; |
4213 | case X86ISD::BLENDI: |
4214 | case X86ISD::PSHUFB: |
4215 | case X86ISD::PSHUFD: |
4216 | case X86ISD::PSHUFHW: |
4217 | case X86ISD::PSHUFLW: |
4218 | case X86ISD::SHUFP: |
4219 | case X86ISD::INSERTPS: |
4220 | case X86ISD::PALIGNR: |
4221 | case X86ISD::VSHLDQ: |
4222 | case X86ISD::VSRLDQ: |
4223 | case X86ISD::MOVLHPS: |
4224 | case X86ISD::MOVLHPD: |
4225 | case X86ISD::MOVHLPS: |
4226 | case X86ISD::MOVLPS: |
4227 | case X86ISD::MOVLPD: |
4228 | case X86ISD::MOVSHDUP: |
4229 | case X86ISD::MOVSLDUP: |
4230 | case X86ISD::MOVDDUP: |
4231 | case X86ISD::MOVSS: |
4232 | case X86ISD::MOVSD: |
4233 | case X86ISD::UNPCKL: |
4234 | case X86ISD::UNPCKH: |
4235 | case X86ISD::VBROADCAST: |
4236 | case X86ISD::VPERMILPI: |
4237 | case X86ISD::VPERMILPV: |
4238 | case X86ISD::VPERM2X128: |
4239 | case X86ISD::VPERMIL2: |
4240 | case X86ISD::VPERMI: |
4241 | case X86ISD::VPPERM: |
4242 | case X86ISD::VPERMV: |
4243 | case X86ISD::VPERMV3: |
4244 | case X86ISD::VPERMIV3: |
4245 | case X86ISD::VZEXT_MOVL: |
4246 | return true; |
4247 | } |
4248 | } |
4249 | |
4250 | static bool isTargetShuffleVariableMask(unsigned Opcode) { |
4251 | switch (Opcode) { |
4252 | default: return false; |
4253 | // Target Shuffles. |
4254 | case X86ISD::PSHUFB: |
4255 | case X86ISD::VPERMILPV: |
4256 | case X86ISD::VPERMIL2: |
4257 | case X86ISD::VPPERM: |
4258 | case X86ISD::VPERMV: |
4259 | case X86ISD::VPERMV3: |
4260 | case X86ISD::VPERMIV3: |
4261 | return true; |
4262 | // 'Faux' Target Shuffles. |
4263 | case ISD::AND: |
4264 | case X86ISD::ANDNP: |
4265 | return true; |
4266 | } |
4267 | } |
4268 | |
4269 | SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const { |
4270 | MachineFunction &MF = DAG.getMachineFunction(); |
4271 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
4272 | X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); |
4273 | int ReturnAddrIndex = FuncInfo->getRAIndex(); |
4274 | |
4275 | if (ReturnAddrIndex == 0) { |
4276 | // Set up a frame object for the return address. |
4277 | unsigned SlotSize = RegInfo->getSlotSize(); |
4278 | ReturnAddrIndex = MF.getFrameInfo().CreateFixedObject(SlotSize, |
4279 | -(int64_t)SlotSize, |
4280 | false); |
4281 | FuncInfo->setRAIndex(ReturnAddrIndex); |
4282 | } |
4283 | |
4284 | return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout())); |
4285 | } |
4286 | |
4287 | bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M, |
4288 | bool hasSymbolicDisplacement) { |
4289 | // Offset should fit into 32 bit immediate field. |
4290 | if (!isInt<32>(Offset)) |
4291 | return false; |
4292 | |
4293 | // If we don't have a symbolic displacement - we don't have any extra |
4294 | // restrictions. |
4295 | if (!hasSymbolicDisplacement) |
4296 | return true; |
4297 | |
4298 | // FIXME: Some tweaks might be needed for medium code model. |
4299 | if (M != CodeModel::Small && M != CodeModel::Kernel) |
4300 | return false; |
4301 | |
4302 | // For small code model we assume that latest object is 16MB before end of 31 |
4303 | // bits boundary. We may also accept pretty large negative constants knowing |
4304 | // that all objects are in the positive half of address space. |
4305 | if (M == CodeModel::Small && Offset < 16*1024*1024) |
4306 | return true; |
4307 | |
4308 | // For kernel code model we know that all object resist in the negative half |
4309 | // of 32bits address space. We may not accept negative offsets, since they may |
4310 | // be just off and we may accept pretty large positive ones. |
4311 | if (M == CodeModel::Kernel && Offset >= 0) |
4312 | return true; |
4313 | |
4314 | return false; |
4315 | } |
4316 | |
4317 | /// Determines whether the callee is required to pop its own arguments. |
4318 | /// Callee pop is necessary to support tail calls. |
4319 | bool X86::isCalleePop(CallingConv::ID CallingConv, |
4320 | bool is64Bit, bool IsVarArg, bool GuaranteeTCO) { |
4321 | // If GuaranteeTCO is true, we force some calls to be callee pop so that we |
4322 | // can guarantee TCO. |
4323 | if (!IsVarArg && shouldGuaranteeTCO(CallingConv, GuaranteeTCO)) |
4324 | return true; |
4325 | |
4326 | switch (CallingConv) { |
4327 | default: |
4328 | return false; |
4329 | case CallingConv::X86_StdCall: |
4330 | case CallingConv::X86_FastCall: |
4331 | case CallingConv::X86_ThisCall: |
4332 | case CallingConv::X86_VectorCall: |
4333 | return !is64Bit; |
4334 | } |
4335 | } |
4336 | |
4337 | /// \brief Return true if the condition is an unsigned comparison operation. |
4338 | static bool isX86CCUnsigned(unsigned X86CC) { |
4339 | switch (X86CC) { |
4340 | default: |
4341 | llvm_unreachable("Invalid integer condition!")::llvm::llvm_unreachable_internal("Invalid integer condition!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4341); |
4342 | case X86::COND_E: |
4343 | case X86::COND_NE: |
4344 | case X86::COND_B: |
4345 | case X86::COND_A: |
4346 | case X86::COND_BE: |
4347 | case X86::COND_AE: |
4348 | return true; |
4349 | case X86::COND_G: |
4350 | case X86::COND_GE: |
4351 | case X86::COND_L: |
4352 | case X86::COND_LE: |
4353 | return false; |
4354 | } |
4355 | } |
4356 | |
4357 | static X86::CondCode TranslateIntegerX86CC(ISD::CondCode SetCCOpcode) { |
4358 | switch (SetCCOpcode) { |
4359 | default: llvm_unreachable("Invalid integer condition!")::llvm::llvm_unreachable_internal("Invalid integer condition!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4359); |
4360 | case ISD::SETEQ: return X86::COND_E; |
4361 | case ISD::SETGT: return X86::COND_G; |
4362 | case ISD::SETGE: return X86::COND_GE; |
4363 | case ISD::SETLT: return X86::COND_L; |
4364 | case ISD::SETLE: return X86::COND_LE; |
4365 | case ISD::SETNE: return X86::COND_NE; |
4366 | case ISD::SETULT: return X86::COND_B; |
4367 | case ISD::SETUGT: return X86::COND_A; |
4368 | case ISD::SETULE: return X86::COND_BE; |
4369 | case ISD::SETUGE: return X86::COND_AE; |
4370 | } |
4371 | } |
4372 | |
4373 | /// Do a one-to-one translation of a ISD::CondCode to the X86-specific |
4374 | /// condition code, returning the condition code and the LHS/RHS of the |
4375 | /// comparison to make. |
4376 | static X86::CondCode TranslateX86CC(ISD::CondCode SetCCOpcode, const SDLoc &DL, |
4377 | bool isFP, SDValue &LHS, SDValue &RHS, |
4378 | SelectionDAG &DAG) { |
4379 | if (!isFP) { |
4380 | if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) { |
4381 | if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) { |
4382 | // X > -1 -> X == 0, jump !sign. |
4383 | RHS = DAG.getConstant(0, DL, RHS.getValueType()); |
4384 | return X86::COND_NS; |
4385 | } |
4386 | if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) { |
4387 | // X < 0 -> X == 0, jump on sign. |
4388 | return X86::COND_S; |
4389 | } |
4390 | if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) { |
4391 | // X < 1 -> X <= 0 |
4392 | RHS = DAG.getConstant(0, DL, RHS.getValueType()); |
4393 | return X86::COND_LE; |
4394 | } |
4395 | } |
4396 | |
4397 | return TranslateIntegerX86CC(SetCCOpcode); |
4398 | } |
4399 | |
4400 | // First determine if it is required or is profitable to flip the operands. |
4401 | |
4402 | // If LHS is a foldable load, but RHS is not, flip the condition. |
4403 | if (ISD::isNON_EXTLoad(LHS.getNode()) && |
4404 | !ISD::isNON_EXTLoad(RHS.getNode())) { |
4405 | SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode); |
4406 | std::swap(LHS, RHS); |
4407 | } |
4408 | |
4409 | switch (SetCCOpcode) { |
4410 | default: break; |
4411 | case ISD::SETOLT: |
4412 | case ISD::SETOLE: |
4413 | case ISD::SETUGT: |
4414 | case ISD::SETUGE: |
4415 | std::swap(LHS, RHS); |
4416 | break; |
4417 | } |
4418 | |
4419 | // On a floating point condition, the flags are set as follows: |
4420 | // ZF PF CF op |
4421 | // 0 | 0 | 0 | X > Y |
4422 | // 0 | 0 | 1 | X < Y |
4423 | // 1 | 0 | 0 | X == Y |
4424 | // 1 | 1 | 1 | unordered |
4425 | switch (SetCCOpcode) { |
4426 | default: llvm_unreachable("Condcode should be pre-legalized away")::llvm::llvm_unreachable_internal("Condcode should be pre-legalized away" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4426); |
4427 | case ISD::SETUEQ: |
4428 | case ISD::SETEQ: return X86::COND_E; |
4429 | case ISD::SETOLT: // flipped |
4430 | case ISD::SETOGT: |
4431 | case ISD::SETGT: return X86::COND_A; |
4432 | case ISD::SETOLE: // flipped |
4433 | case ISD::SETOGE: |
4434 | case ISD::SETGE: return X86::COND_AE; |
4435 | case ISD::SETUGT: // flipped |
4436 | case ISD::SETULT: |
4437 | case ISD::SETLT: return X86::COND_B; |
4438 | case ISD::SETUGE: // flipped |
4439 | case ISD::SETULE: |
4440 | case ISD::SETLE: return X86::COND_BE; |
4441 | case ISD::SETONE: |
4442 | case ISD::SETNE: return X86::COND_NE; |
4443 | case ISD::SETUO: return X86::COND_P; |
4444 | case ISD::SETO: return X86::COND_NP; |
4445 | case ISD::SETOEQ: |
4446 | case ISD::SETUNE: return X86::COND_INVALID; |
4447 | } |
4448 | } |
4449 | |
4450 | /// Is there a floating point cmov for the specific X86 condition code? |
4451 | /// Current x86 isa includes the following FP cmov instructions: |
4452 | /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu. |
4453 | static bool hasFPCMov(unsigned X86CC) { |
4454 | switch (X86CC) { |
4455 | default: |
4456 | return false; |
4457 | case X86::COND_B: |
4458 | case X86::COND_BE: |
4459 | case X86::COND_E: |
4460 | case X86::COND_P: |
4461 | case X86::COND_A: |
4462 | case X86::COND_AE: |
4463 | case X86::COND_NE: |
4464 | case X86::COND_NP: |
4465 | return true; |
4466 | } |
4467 | } |
4468 | |
4469 | |
4470 | bool X86TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info, |
4471 | const CallInst &I, |
4472 | unsigned Intrinsic) const { |
4473 | |
4474 | const IntrinsicData* IntrData = getIntrinsicWithChain(Intrinsic); |
4475 | if (!IntrData) |
4476 | return false; |
4477 | |
4478 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
4479 | Info.readMem = false; |
4480 | Info.writeMem = false; |
4481 | Info.vol = false; |
4482 | Info.offset = 0; |
4483 | |
4484 | switch (IntrData->Type) { |
4485 | case EXPAND_FROM_MEM: { |
4486 | Info.ptrVal = I.getArgOperand(0); |
4487 | Info.memVT = MVT::getVT(I.getType()); |
4488 | Info.align = 1; |
4489 | Info.readMem = true; |
4490 | break; |
4491 | } |
4492 | case COMPRESS_TO_MEM: { |
4493 | Info.ptrVal = I.getArgOperand(0); |
4494 | Info.memVT = MVT::getVT(I.getArgOperand(1)->getType()); |
4495 | Info.align = 1; |
4496 | Info.writeMem = true; |
4497 | break; |
4498 | } |
4499 | case TRUNCATE_TO_MEM_VI8: |
4500 | case TRUNCATE_TO_MEM_VI16: |
4501 | case TRUNCATE_TO_MEM_VI32: { |
4502 | Info.ptrVal = I.getArgOperand(0); |
4503 | MVT VT = MVT::getVT(I.getArgOperand(1)->getType()); |
4504 | MVT ScalarVT = MVT::INVALID_SIMPLE_VALUE_TYPE; |
4505 | if (IntrData->Type == TRUNCATE_TO_MEM_VI8) |
4506 | ScalarVT = MVT::i8; |
4507 | else if (IntrData->Type == TRUNCATE_TO_MEM_VI16) |
4508 | ScalarVT = MVT::i16; |
4509 | else if (IntrData->Type == TRUNCATE_TO_MEM_VI32) |
4510 | ScalarVT = MVT::i32; |
4511 | |
4512 | Info.memVT = MVT::getVectorVT(ScalarVT, VT.getVectorNumElements()); |
4513 | Info.align = 1; |
4514 | Info.writeMem = true; |
4515 | break; |
4516 | } |
4517 | default: |
4518 | return false; |
4519 | } |
4520 | |
4521 | return true; |
4522 | } |
4523 | |
4524 | /// Returns true if the target can instruction select the |
4525 | /// specified FP immediate natively. If false, the legalizer will |
4526 | /// materialize the FP immediate as a load from a constant pool. |
4527 | bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const { |
4528 | for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) { |
4529 | if (Imm.bitwiseIsEqual(LegalFPImmediates[i])) |
4530 | return true; |
4531 | } |
4532 | return false; |
4533 | } |
4534 | |
4535 | bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load, |
4536 | ISD::LoadExtType ExtTy, |
4537 | EVT NewVT) const { |
4538 | // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF |
4539 | // relocation target a movq or addq instruction: don't let the load shrink. |
4540 | SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr(); |
4541 | if (BasePtr.getOpcode() == X86ISD::WrapperRIP) |
4542 | if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0))) |
4543 | return GA->getTargetFlags() != X86II::MO_GOTTPOFF; |
4544 | return true; |
4545 | } |
4546 | |
4547 | /// \brief Returns true if it is beneficial to convert a load of a constant |
4548 | /// to just the constant itself. |
4549 | bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm, |
4550 | Type *Ty) const { |
4551 | assert(Ty->isIntegerTy())((Ty->isIntegerTy()) ? static_cast<void> (0) : __assert_fail ("Ty->isIntegerTy()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4551, __PRETTY_FUNCTION__)); |
4552 | |
4553 | unsigned BitSize = Ty->getPrimitiveSizeInBits(); |
4554 | if (BitSize == 0 || BitSize > 64) |
4555 | return false; |
4556 | return true; |
4557 | } |
4558 | |
4559 | bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT, |
4560 | unsigned Index) const { |
4561 | if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT)) |
4562 | return false; |
4563 | |
4564 | return (Index == 0 || Index == ResVT.getVectorNumElements()); |
4565 | } |
4566 | |
4567 | bool X86TargetLowering::isCheapToSpeculateCttz() const { |
4568 | // Speculate cttz only if we can directly use TZCNT. |
4569 | return Subtarget.hasBMI(); |
4570 | } |
4571 | |
4572 | bool X86TargetLowering::isCheapToSpeculateCtlz() const { |
4573 | // Speculate ctlz only if we can directly use LZCNT. |
4574 | return Subtarget.hasLZCNT(); |
4575 | } |
4576 | |
4577 | bool X86TargetLowering::isCtlzFast() const { |
4578 | return Subtarget.hasFastLZCNT(); |
4579 | } |
4580 | |
4581 | bool X86TargetLowering::isMaskAndCmp0FoldingBeneficial( |
4582 | const Instruction &AndI) const { |
4583 | return true; |
4584 | } |
4585 | |
4586 | bool X86TargetLowering::hasAndNotCompare(SDValue Y) const { |
4587 | if (!Subtarget.hasBMI()) |
4588 | return false; |
4589 | |
4590 | // There are only 32-bit and 64-bit forms for 'andn'. |
4591 | EVT VT = Y.getValueType(); |
4592 | if (VT != MVT::i32 && VT != MVT::i64) |
4593 | return false; |
4594 | |
4595 | return true; |
4596 | } |
4597 | |
4598 | MVT X86TargetLowering::hasFastEqualityCompare(unsigned NumBits) const { |
4599 | MVT VT = MVT::getIntegerVT(NumBits); |
4600 | if (isTypeLegal(VT)) |
4601 | return VT; |
4602 | |
4603 | // PMOVMSKB can handle this. |
4604 | if (NumBits == 128 && isTypeLegal(MVT::v16i8)) |
4605 | return MVT::v16i8; |
4606 | |
4607 | // VPMOVMSKB can handle this. |
4608 | if (NumBits == 256 && isTypeLegal(MVT::v32i8)) |
4609 | return MVT::v32i8; |
4610 | |
4611 | // TODO: Allow 64-bit type for 32-bit target. |
4612 | // TODO: 512-bit types should be allowed, but make sure that those |
4613 | // cases are handled in combineVectorSizedSetCCEquality(). |
4614 | |
4615 | return MVT::INVALID_SIMPLE_VALUE_TYPE; |
4616 | } |
4617 | |
4618 | /// Val is the undef sentinel value or equal to the specified value. |
4619 | static bool isUndefOrEqual(int Val, int CmpVal) { |
4620 | return ((Val == SM_SentinelUndef) || (Val == CmpVal)); |
4621 | } |
4622 | |
4623 | /// Val is either the undef or zero sentinel value. |
4624 | static bool isUndefOrZero(int Val) { |
4625 | return ((Val == SM_SentinelUndef) || (Val == SM_SentinelZero)); |
4626 | } |
4627 | |
4628 | /// Return true if every element in Mask, beginning |
4629 | /// from position Pos and ending in Pos+Size is the undef sentinel value. |
4630 | static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) { |
4631 | for (unsigned i = Pos, e = Pos + Size; i != e; ++i) |
4632 | if (Mask[i] != SM_SentinelUndef) |
4633 | return false; |
4634 | return true; |
4635 | } |
4636 | |
4637 | /// Return true if Val is undef or if its value falls within the |
4638 | /// specified range (L, H]. |
4639 | static bool isUndefOrInRange(int Val, int Low, int Hi) { |
4640 | return (Val == SM_SentinelUndef) || (Val >= Low && Val < Hi); |
4641 | } |
4642 | |
4643 | /// Return true if every element in Mask is undef or if its value |
4644 | /// falls within the specified range (L, H]. |
4645 | static bool isUndefOrInRange(ArrayRef<int> Mask, |
4646 | int Low, int Hi) { |
4647 | for (int M : Mask) |
4648 | if (!isUndefOrInRange(M, Low, Hi)) |
4649 | return false; |
4650 | return true; |
4651 | } |
4652 | |
4653 | /// Return true if Val is undef, zero or if its value falls within the |
4654 | /// specified range (L, H]. |
4655 | static bool isUndefOrZeroOrInRange(int Val, int Low, int Hi) { |
4656 | return isUndefOrZero(Val) || (Val >= Low && Val < Hi); |
4657 | } |
4658 | |
4659 | /// Return true if every element in Mask is undef, zero or if its value |
4660 | /// falls within the specified range (L, H]. |
4661 | static bool isUndefOrZeroOrInRange(ArrayRef<int> Mask, int Low, int Hi) { |
4662 | for (int M : Mask) |
4663 | if (!isUndefOrZeroOrInRange(M, Low, Hi)) |
4664 | return false; |
4665 | return true; |
4666 | } |
4667 | |
4668 | /// Return true if every element in Mask, beginning |
4669 | /// from position Pos and ending in Pos+Size, falls within the specified |
4670 | /// sequential range (Low, Low+Size]. or is undef. |
4671 | static bool isSequentialOrUndefInRange(ArrayRef<int> Mask, |
4672 | unsigned Pos, unsigned Size, int Low) { |
4673 | for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low) |
4674 | if (!isUndefOrEqual(Mask[i], Low)) |
4675 | return false; |
4676 | return true; |
4677 | } |
4678 | |
4679 | /// Return true if every element in Mask, beginning |
4680 | /// from position Pos and ending in Pos+Size, falls within the specified |
4681 | /// sequential range (Low, Low+Size], or is undef or is zero. |
4682 | static bool isSequentialOrUndefOrZeroInRange(ArrayRef<int> Mask, unsigned Pos, |
4683 | unsigned Size, int Low) { |
4684 | for (unsigned i = Pos, e = Pos + Size; i != e; ++i, ++Low) |
4685 | if (!isUndefOrZero(Mask[i]) && Mask[i] != Low) |
4686 | return false; |
4687 | return true; |
4688 | } |
4689 | |
4690 | /// Return true if every element in Mask, beginning |
4691 | /// from position Pos and ending in Pos+Size is undef or is zero. |
4692 | static bool isUndefOrZeroInRange(ArrayRef<int> Mask, unsigned Pos, |
4693 | unsigned Size) { |
4694 | for (unsigned i = Pos, e = Pos + Size; i != e; ++i) |
4695 | if (!isUndefOrZero(Mask[i])) |
4696 | return false; |
4697 | return true; |
4698 | } |
4699 | |
4700 | /// \brief Helper function to test whether a shuffle mask could be |
4701 | /// simplified by widening the elements being shuffled. |
4702 | /// |
4703 | /// Appends the mask for wider elements in WidenedMask if valid. Otherwise |
4704 | /// leaves it in an unspecified state. |
4705 | /// |
4706 | /// NOTE: This must handle normal vector shuffle masks and *target* vector |
4707 | /// shuffle masks. The latter have the special property of a '-2' representing |
4708 | /// a zero-ed lane of a vector. |
4709 | static bool canWidenShuffleElements(ArrayRef<int> Mask, |
4710 | SmallVectorImpl<int> &WidenedMask) { |
4711 | WidenedMask.assign(Mask.size() / 2, 0); |
4712 | for (int i = 0, Size = Mask.size(); i < Size; i += 2) { |
4713 | int M0 = Mask[i]; |
4714 | int M1 = Mask[i + 1]; |
4715 | |
4716 | // If both elements are undef, its trivial. |
4717 | if (M0 == SM_SentinelUndef && M1 == SM_SentinelUndef) { |
4718 | WidenedMask[i / 2] = SM_SentinelUndef; |
4719 | continue; |
4720 | } |
4721 | |
4722 | // Check for an undef mask and a mask value properly aligned to fit with |
4723 | // a pair of values. If we find such a case, use the non-undef mask's value. |
4724 | if (M0 == SM_SentinelUndef && M1 >= 0 && (M1 % 2) == 1) { |
4725 | WidenedMask[i / 2] = M1 / 2; |
4726 | continue; |
4727 | } |
4728 | if (M1 == SM_SentinelUndef && M0 >= 0 && (M0 % 2) == 0) { |
4729 | WidenedMask[i / 2] = M0 / 2; |
4730 | continue; |
4731 | } |
4732 | |
4733 | // When zeroing, we need to spread the zeroing across both lanes to widen. |
4734 | if (M0 == SM_SentinelZero || M1 == SM_SentinelZero) { |
4735 | if ((M0 == SM_SentinelZero || M0 == SM_SentinelUndef) && |
4736 | (M1 == SM_SentinelZero || M1 == SM_SentinelUndef)) { |
4737 | WidenedMask[i / 2] = SM_SentinelZero; |
4738 | continue; |
4739 | } |
4740 | return false; |
4741 | } |
4742 | |
4743 | // Finally check if the two mask values are adjacent and aligned with |
4744 | // a pair. |
4745 | if (M0 != SM_SentinelUndef && (M0 % 2) == 0 && (M0 + 1) == M1) { |
4746 | WidenedMask[i / 2] = M0 / 2; |
4747 | continue; |
4748 | } |
4749 | |
4750 | // Otherwise we can't safely widen the elements used in this shuffle. |
4751 | return false; |
4752 | } |
4753 | assert(WidenedMask.size() == Mask.size() / 2 &&((WidenedMask.size() == Mask.size() / 2 && "Incorrect size of mask after widening the elements!" ) ? static_cast<void> (0) : __assert_fail ("WidenedMask.size() == Mask.size() / 2 && \"Incorrect size of mask after widening the elements!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4754, __PRETTY_FUNCTION__)) |
4754 | "Incorrect size of mask after widening the elements!")((WidenedMask.size() == Mask.size() / 2 && "Incorrect size of mask after widening the elements!" ) ? static_cast<void> (0) : __assert_fail ("WidenedMask.size() == Mask.size() / 2 && \"Incorrect size of mask after widening the elements!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4754, __PRETTY_FUNCTION__)); |
4755 | |
4756 | return true; |
4757 | } |
4758 | |
4759 | /// Helper function to scale a shuffle or target shuffle mask, replacing each |
4760 | /// mask index with the scaled sequential indices for an equivalent narrowed |
4761 | /// mask. This is the reverse process to canWidenShuffleElements, but can always |
4762 | /// succeed. |
4763 | static void scaleShuffleMask(int Scale, ArrayRef<int> Mask, |
4764 | SmallVectorImpl<int> &ScaledMask) { |
4765 | assert(0 < Scale && "Unexpected scaling factor")((0 < Scale && "Unexpected scaling factor") ? static_cast <void> (0) : __assert_fail ("0 < Scale && \"Unexpected scaling factor\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4765, __PRETTY_FUNCTION__)); |
4766 | int NumElts = Mask.size(); |
4767 | ScaledMask.assign(static_cast<size_t>(NumElts * Scale), -1); |
4768 | |
4769 | for (int i = 0; i != NumElts; ++i) { |
4770 | int M = Mask[i]; |
4771 | |
4772 | // Repeat sentinel values in every mask element. |
4773 | if (M < 0) { |
4774 | for (int s = 0; s != Scale; ++s) |
4775 | ScaledMask[(Scale * i) + s] = M; |
4776 | continue; |
4777 | } |
4778 | |
4779 | // Scale mask element and increment across each mask element. |
4780 | for (int s = 0; s != Scale; ++s) |
4781 | ScaledMask[(Scale * i) + s] = (Scale * M) + s; |
4782 | } |
4783 | } |
4784 | |
4785 | /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector |
4786 | /// extract that is suitable for instruction that extract 128 or 256 bit vectors |
4787 | static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) { |
4788 | assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width")(((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width" ) ? static_cast<void> (0) : __assert_fail ("(vecWidth == 128 || vecWidth == 256) && \"Unexpected vector width\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4788, __PRETTY_FUNCTION__)); |
4789 | if (!isa<ConstantSDNode>(N->getOperand(1).getNode())) |
4790 | return false; |
4791 | |
4792 | // The index should be aligned on a vecWidth-bit boundary. |
4793 | uint64_t Index = N->getConstantOperandVal(1); |
4794 | MVT VT = N->getSimpleValueType(0); |
4795 | unsigned ElSize = VT.getScalarSizeInBits(); |
4796 | return (Index * ElSize) % vecWidth == 0; |
4797 | } |
4798 | |
4799 | /// Return true if the specified INSERT_SUBVECTOR |
4800 | /// operand specifies a subvector insert that is suitable for input to |
4801 | /// insertion of 128 or 256-bit subvectors |
4802 | static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) { |
4803 | assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width")(((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width" ) ? static_cast<void> (0) : __assert_fail ("(vecWidth == 128 || vecWidth == 256) && \"Unexpected vector width\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4803, __PRETTY_FUNCTION__)); |
4804 | if (!isa<ConstantSDNode>(N->getOperand(2).getNode())) |
4805 | return false; |
4806 | |
4807 | // The index should be aligned on a vecWidth-bit boundary. |
4808 | uint64_t Index = N->getConstantOperandVal(2); |
4809 | MVT VT = N->getSimpleValueType(0); |
4810 | unsigned ElSize = VT.getScalarSizeInBits(); |
4811 | return (Index * ElSize) % vecWidth == 0; |
4812 | } |
4813 | |
4814 | bool X86::isVINSERT128Index(SDNode *N) { |
4815 | return isVINSERTIndex(N, 128); |
4816 | } |
4817 | |
4818 | bool X86::isVINSERT256Index(SDNode *N) { |
4819 | return isVINSERTIndex(N, 256); |
4820 | } |
4821 | |
4822 | bool X86::isVEXTRACT128Index(SDNode *N) { |
4823 | return isVEXTRACTIndex(N, 128); |
4824 | } |
4825 | |
4826 | bool X86::isVEXTRACT256Index(SDNode *N) { |
4827 | return isVEXTRACTIndex(N, 256); |
4828 | } |
4829 | |
4830 | static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) { |
4831 | assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width")(((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width" ) ? static_cast<void> (0) : __assert_fail ("(vecWidth == 128 || vecWidth == 256) && \"Unsupported vector width\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4831, __PRETTY_FUNCTION__)); |
4832 | assert(isa<ConstantSDNode>(N->getOperand(1).getNode()) &&((isa<ConstantSDNode>(N->getOperand(1).getNode()) && "Illegal extract subvector for VEXTRACT") ? static_cast<void > (0) : __assert_fail ("isa<ConstantSDNode>(N->getOperand(1).getNode()) && \"Illegal extract subvector for VEXTRACT\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4833, __PRETTY_FUNCTION__)) |
4833 | "Illegal extract subvector for VEXTRACT")((isa<ConstantSDNode>(N->getOperand(1).getNode()) && "Illegal extract subvector for VEXTRACT") ? static_cast<void > (0) : __assert_fail ("isa<ConstantSDNode>(N->getOperand(1).getNode()) && \"Illegal extract subvector for VEXTRACT\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4833, __PRETTY_FUNCTION__)); |
4834 | |
4835 | uint64_t Index = N->getConstantOperandVal(1); |
4836 | MVT VecVT = N->getOperand(0).getSimpleValueType(); |
4837 | unsigned NumElemsPerChunk = vecWidth / VecVT.getScalarSizeInBits(); |
4838 | return Index / NumElemsPerChunk; |
4839 | } |
4840 | |
4841 | static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) { |
4842 | assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width")(((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width" ) ? static_cast<void> (0) : __assert_fail ("(vecWidth == 128 || vecWidth == 256) && \"Unsupported vector width\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4842, __PRETTY_FUNCTION__)); |
4843 | assert(isa<ConstantSDNode>(N->getOperand(2).getNode()) &&((isa<ConstantSDNode>(N->getOperand(2).getNode()) && "Illegal insert subvector for VINSERT") ? static_cast<void > (0) : __assert_fail ("isa<ConstantSDNode>(N->getOperand(2).getNode()) && \"Illegal insert subvector for VINSERT\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4844, __PRETTY_FUNCTION__)) |
4844 | "Illegal insert subvector for VINSERT")((isa<ConstantSDNode>(N->getOperand(2).getNode()) && "Illegal insert subvector for VINSERT") ? static_cast<void > (0) : __assert_fail ("isa<ConstantSDNode>(N->getOperand(2).getNode()) && \"Illegal insert subvector for VINSERT\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4844, __PRETTY_FUNCTION__)); |
4845 | |
4846 | uint64_t Index = N->getConstantOperandVal(2); |
4847 | MVT VecVT = N->getSimpleValueType(0); |
4848 | unsigned NumElemsPerChunk = vecWidth / VecVT.getScalarSizeInBits(); |
4849 | return Index / NumElemsPerChunk; |
4850 | } |
4851 | |
4852 | /// Return the appropriate immediate to extract the specified |
4853 | /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions. |
4854 | unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) { |
4855 | return getExtractVEXTRACTImmediate(N, 128); |
4856 | } |
4857 | |
4858 | /// Return the appropriate immediate to extract the specified |
4859 | /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions. |
4860 | unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) { |
4861 | return getExtractVEXTRACTImmediate(N, 256); |
4862 | } |
4863 | |
4864 | /// Return the appropriate immediate to insert at the specified |
4865 | /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions. |
4866 | unsigned X86::getInsertVINSERT128Immediate(SDNode *N) { |
4867 | return getInsertVINSERTImmediate(N, 128); |
4868 | } |
4869 | |
4870 | /// Return the appropriate immediate to insert at the specified |
4871 | /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions. |
4872 | unsigned X86::getInsertVINSERT256Immediate(SDNode *N) { |
4873 | return getInsertVINSERTImmediate(N, 256); |
4874 | } |
4875 | |
4876 | /// Returns true if Elt is a constant zero or a floating point constant +0.0. |
4877 | bool X86::isZeroNode(SDValue Elt) { |
4878 | return isNullConstant(Elt) || isNullFPConstant(Elt); |
4879 | } |
4880 | |
4881 | // Build a vector of constants. |
4882 | // Use an UNDEF node if MaskElt == -1. |
4883 | // Split 64-bit constants in the 32-bit mode. |
4884 | static SDValue getConstVector(ArrayRef<int> Values, MVT VT, SelectionDAG &DAG, |
4885 | const SDLoc &dl, bool IsMask = false) { |
4886 | |
4887 | SmallVector<SDValue, 32> Ops; |
4888 | bool Split = false; |
4889 | |
4890 | MVT ConstVecVT = VT; |
4891 | unsigned NumElts = VT.getVectorNumElements(); |
4892 | bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64); |
4893 | if (!In64BitMode && VT.getVectorElementType() == MVT::i64) { |
4894 | ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2); |
4895 | Split = true; |
4896 | } |
4897 | |
4898 | MVT EltVT = ConstVecVT.getVectorElementType(); |
4899 | for (unsigned i = 0; i < NumElts; ++i) { |
4900 | bool IsUndef = Values[i] < 0 && IsMask; |
4901 | SDValue OpNode = IsUndef ? DAG.getUNDEF(EltVT) : |
4902 | DAG.getConstant(Values[i], dl, EltVT); |
4903 | Ops.push_back(OpNode); |
4904 | if (Split) |
4905 | Ops.push_back(IsUndef ? DAG.getUNDEF(EltVT) : |
4906 | DAG.getConstant(0, dl, EltVT)); |
4907 | } |
4908 | SDValue ConstsNode = DAG.getBuildVector(ConstVecVT, dl, Ops); |
4909 | if (Split) |
4910 | ConstsNode = DAG.getBitcast(VT, ConstsNode); |
4911 | return ConstsNode; |
4912 | } |
4913 | |
4914 | static SDValue getConstVector(ArrayRef<APInt> Bits, APInt &Undefs, |
4915 | MVT VT, SelectionDAG &DAG, const SDLoc &dl) { |
4916 | assert(Bits.size() == Undefs.getBitWidth() &&((Bits.size() == Undefs.getBitWidth() && "Unequal constant and undef arrays" ) ? static_cast<void> (0) : __assert_fail ("Bits.size() == Undefs.getBitWidth() && \"Unequal constant and undef arrays\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4917, __PRETTY_FUNCTION__)) |
4917 | "Unequal constant and undef arrays")((Bits.size() == Undefs.getBitWidth() && "Unequal constant and undef arrays" ) ? static_cast<void> (0) : __assert_fail ("Bits.size() == Undefs.getBitWidth() && \"Unequal constant and undef arrays\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4917, __PRETTY_FUNCTION__)); |
4918 | SmallVector<SDValue, 32> Ops; |
4919 | bool Split = false; |
4920 | |
4921 | MVT ConstVecVT = VT; |
4922 | unsigned NumElts = VT.getVectorNumElements(); |
4923 | bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64); |
4924 | if (!In64BitMode && VT.getVectorElementType() == MVT::i64) { |
4925 | ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2); |
4926 | Split = true; |
4927 | } |
4928 | |
4929 | MVT EltVT = ConstVecVT.getVectorElementType(); |
4930 | for (unsigned i = 0, e = Bits.size(); i != e; ++i) { |
4931 | if (Undefs[i]) { |
4932 | Ops.append(Split ? 2 : 1, DAG.getUNDEF(EltVT)); |
4933 | continue; |
4934 | } |
4935 | const APInt &V = Bits[i]; |
4936 | assert(V.getBitWidth() == VT.getScalarSizeInBits() && "Unexpected sizes")((V.getBitWidth() == VT.getScalarSizeInBits() && "Unexpected sizes" ) ? static_cast<void> (0) : __assert_fail ("V.getBitWidth() == VT.getScalarSizeInBits() && \"Unexpected sizes\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4936, __PRETTY_FUNCTION__)); |
4937 | if (Split) { |
4938 | Ops.push_back(DAG.getConstant(V.trunc(32), dl, EltVT)); |
4939 | Ops.push_back(DAG.getConstant(V.lshr(32).trunc(32), dl, EltVT)); |
4940 | } else if (EltVT == MVT::f32) { |
4941 | APFloat FV(APFloat::IEEEsingle(), V); |
4942 | Ops.push_back(DAG.getConstantFP(FV, dl, EltVT)); |
4943 | } else if (EltVT == MVT::f64) { |
4944 | APFloat FV(APFloat::IEEEdouble(), V); |
4945 | Ops.push_back(DAG.getConstantFP(FV, dl, EltVT)); |
4946 | } else { |
4947 | Ops.push_back(DAG.getConstant(V, dl, EltVT)); |
4948 | } |
4949 | } |
4950 | |
4951 | SDValue ConstsNode = DAG.getBuildVector(ConstVecVT, dl, Ops); |
4952 | return DAG.getBitcast(VT, ConstsNode); |
4953 | } |
4954 | |
4955 | /// Returns a vector of specified type with all zero elements. |
4956 | static SDValue getZeroVector(MVT VT, const X86Subtarget &Subtarget, |
4957 | SelectionDAG &DAG, const SDLoc &dl) { |
4958 | assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector() ||(((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector () || VT.getVectorElementType() == MVT::i1) && "Unexpected vector type" ) ? static_cast<void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector() || VT.getVectorElementType() == MVT::i1) && \"Unexpected vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4960, __PRETTY_FUNCTION__)) |
4959 | VT.getVectorElementType() == MVT::i1) &&(((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector () || VT.getVectorElementType() == MVT::i1) && "Unexpected vector type" ) ? static_cast<void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector() || VT.getVectorElementType() == MVT::i1) && \"Unexpected vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4960, __PRETTY_FUNCTION__)) |
4960 | "Unexpected vector type")(((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector () || VT.getVectorElementType() == MVT::i1) && "Unexpected vector type" ) ? static_cast<void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector() || VT.getVectorElementType() == MVT::i1) && \"Unexpected vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4960, __PRETTY_FUNCTION__)); |
4961 | |
4962 | // Try to build SSE/AVX zero vectors as <N x i32> bitcasted to their dest |
4963 | // type. This ensures they get CSE'd. But if the integer type is not |
4964 | // available, use a floating-point +0.0 instead. |
4965 | SDValue Vec; |
4966 | if (!Subtarget.hasSSE2() && VT.is128BitVector()) { |
4967 | Vec = DAG.getConstantFP(+0.0, dl, MVT::v4f32); |
4968 | } else if (VT.getVectorElementType() == MVT::i1) { |
4969 | assert((Subtarget.hasBWI() || VT.getVectorNumElements() <= 16) &&(((Subtarget.hasBWI() || VT.getVectorNumElements() <= 16) && "Unexpected vector type") ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasBWI() || VT.getVectorNumElements() <= 16) && \"Unexpected vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4970, __PRETTY_FUNCTION__)) |
4970 | "Unexpected vector type")(((Subtarget.hasBWI() || VT.getVectorNumElements() <= 16) && "Unexpected vector type") ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasBWI() || VT.getVectorNumElements() <= 16) && \"Unexpected vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4970, __PRETTY_FUNCTION__)); |
4971 | assert((Subtarget.hasVLX() || VT.getVectorNumElements() >= 8) &&(((Subtarget.hasVLX() || VT.getVectorNumElements() >= 8) && "Unexpected vector type") ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasVLX() || VT.getVectorNumElements() >= 8) && \"Unexpected vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4972, __PRETTY_FUNCTION__)) |
4972 | "Unexpected vector type")(((Subtarget.hasVLX() || VT.getVectorNumElements() >= 8) && "Unexpected vector type") ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasVLX() || VT.getVectorNumElements() >= 8) && \"Unexpected vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4972, __PRETTY_FUNCTION__)); |
4973 | Vec = DAG.getConstant(0, dl, VT); |
4974 | } else { |
4975 | unsigned Num32BitElts = VT.getSizeInBits() / 32; |
4976 | Vec = DAG.getConstant(0, dl, MVT::getVectorVT(MVT::i32, Num32BitElts)); |
4977 | } |
4978 | return DAG.getBitcast(VT, Vec); |
4979 | } |
4980 | |
4981 | static SDValue extractSubVector(SDValue Vec, unsigned IdxVal, SelectionDAG &DAG, |
4982 | const SDLoc &dl, unsigned vectorWidth) { |
4983 | EVT VT = Vec.getValueType(); |
4984 | EVT ElVT = VT.getVectorElementType(); |
4985 | unsigned Factor = VT.getSizeInBits()/vectorWidth; |
4986 | EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT, |
4987 | VT.getVectorNumElements()/Factor); |
4988 | |
4989 | // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR |
4990 | unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits(); |
4991 | assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2")((isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(ElemsPerChunk) && \"Elements per chunk not power of 2\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 4991, __PRETTY_FUNCTION__)); |
4992 | |
4993 | // This is the index of the first element of the vectorWidth-bit chunk |
4994 | // we want. Since ElemsPerChunk is a power of 2 just need to clear bits. |
4995 | IdxVal &= ~(ElemsPerChunk - 1); |
4996 | |
4997 | // If the input is a buildvector just emit a smaller one. |
4998 | if (Vec.getOpcode() == ISD::BUILD_VECTOR) |
4999 | return DAG.getBuildVector( |
5000 | ResultVT, dl, makeArrayRef(Vec->op_begin() + IdxVal, ElemsPerChunk)); |
5001 | |
5002 | SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl); |
5003 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx); |
5004 | } |
5005 | |
5006 | /// Generate a DAG to grab 128-bits from a vector > 128 bits. This |
5007 | /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128 |
5008 | /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4 |
5009 | /// instructions or a simple subregister reference. Idx is an index in the |
5010 | /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes |
5011 | /// lowering EXTRACT_VECTOR_ELT operations easier. |
5012 | static SDValue extract128BitVector(SDValue Vec, unsigned IdxVal, |
5013 | SelectionDAG &DAG, const SDLoc &dl) { |
5014 | assert((Vec.getValueType().is256BitVector() ||(((Vec.getValueType().is256BitVector() || Vec.getValueType(). is512BitVector()) && "Unexpected vector size!") ? static_cast <void> (0) : __assert_fail ("(Vec.getValueType().is256BitVector() || Vec.getValueType().is512BitVector()) && \"Unexpected vector size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5015, __PRETTY_FUNCTION__)) |
5015 | Vec.getValueType().is512BitVector()) && "Unexpected vector size!")(((Vec.getValueType().is256BitVector() || Vec.getValueType(). is512BitVector()) && "Unexpected vector size!") ? static_cast <void> (0) : __assert_fail ("(Vec.getValueType().is256BitVector() || Vec.getValueType().is512BitVector()) && \"Unexpected vector size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5015, __PRETTY_FUNCTION__)); |
5016 | return extractSubVector(Vec, IdxVal, DAG, dl, 128); |
5017 | } |
5018 | |
5019 | /// Generate a DAG to grab 256-bits from a 512-bit vector. |
5020 | static SDValue extract256BitVector(SDValue Vec, unsigned IdxVal, |
5021 | SelectionDAG &DAG, const SDLoc &dl) { |
5022 | assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!")((Vec.getValueType().is512BitVector() && "Unexpected vector size!" ) ? static_cast<void> (0) : __assert_fail ("Vec.getValueType().is512BitVector() && \"Unexpected vector size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5022, __PRETTY_FUNCTION__)); |
5023 | return extractSubVector(Vec, IdxVal, DAG, dl, 256); |
5024 | } |
5025 | |
5026 | static SDValue insertSubVector(SDValue Result, SDValue Vec, unsigned IdxVal, |
5027 | SelectionDAG &DAG, const SDLoc &dl, |
5028 | unsigned vectorWidth) { |
5029 | assert((vectorWidth == 128 || vectorWidth == 256) &&(((vectorWidth == 128 || vectorWidth == 256) && "Unsupported vector width" ) ? static_cast<void> (0) : __assert_fail ("(vectorWidth == 128 || vectorWidth == 256) && \"Unsupported vector width\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5030, __PRETTY_FUNCTION__)) |
5030 | "Unsupported vector width")(((vectorWidth == 128 || vectorWidth == 256) && "Unsupported vector width" ) ? static_cast<void> (0) : __assert_fail ("(vectorWidth == 128 || vectorWidth == 256) && \"Unsupported vector width\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5030, __PRETTY_FUNCTION__)); |
5031 | // Inserting UNDEF is Result |
5032 | if (Vec.isUndef()) |
5033 | return Result; |
5034 | EVT VT = Vec.getValueType(); |
5035 | EVT ElVT = VT.getVectorElementType(); |
5036 | EVT ResultVT = Result.getValueType(); |
5037 | |
5038 | // Insert the relevant vectorWidth bits. |
5039 | unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits(); |
5040 | assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2")((isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(ElemsPerChunk) && \"Elements per chunk not power of 2\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5040, __PRETTY_FUNCTION__)); |
5041 | |
5042 | // This is the index of the first element of the vectorWidth-bit chunk |
5043 | // we want. Since ElemsPerChunk is a power of 2 just need to clear bits. |
5044 | IdxVal &= ~(ElemsPerChunk - 1); |
5045 | |
5046 | SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl); |
5047 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx); |
5048 | } |
5049 | |
5050 | /// Generate a DAG to put 128-bits into a vector > 128 bits. This |
5051 | /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or |
5052 | /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a |
5053 | /// simple superregister reference. Idx is an index in the 128 bits |
5054 | /// we want. It need not be aligned to a 128-bit boundary. That makes |
5055 | /// lowering INSERT_VECTOR_ELT operations easier. |
5056 | static SDValue insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal, |
5057 | SelectionDAG &DAG, const SDLoc &dl) { |
5058 | assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!")((Vec.getValueType().is128BitVector() && "Unexpected vector size!" ) ? static_cast<void> (0) : __assert_fail ("Vec.getValueType().is128BitVector() && \"Unexpected vector size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5058, __PRETTY_FUNCTION__)); |
5059 | return insertSubVector(Result, Vec, IdxVal, DAG, dl, 128); |
5060 | } |
5061 | |
5062 | static SDValue insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal, |
5063 | SelectionDAG &DAG, const SDLoc &dl) { |
5064 | assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!")((Vec.getValueType().is256BitVector() && "Unexpected vector size!" ) ? static_cast<void> (0) : __assert_fail ("Vec.getValueType().is256BitVector() && \"Unexpected vector size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5064, __PRETTY_FUNCTION__)); |
5065 | return insertSubVector(Result, Vec, IdxVal, DAG, dl, 256); |
5066 | } |
5067 | |
5068 | // Return true if the instruction zeroes the unused upper part of the |
5069 | // destination and accepts mask. |
5070 | static bool isMaskedZeroUpperBitsvXi1(unsigned int Opcode) { |
5071 | switch (Opcode) { |
5072 | default: |
5073 | return false; |
5074 | case X86ISD::PCMPEQM: |
5075 | case X86ISD::PCMPGTM: |
5076 | case X86ISD::CMPM: |
5077 | case X86ISD::CMPMU: |
5078 | return true; |
5079 | } |
5080 | } |
5081 | |
5082 | /// Insert i1-subvector to i1-vector. |
5083 | static SDValue insert1BitVector(SDValue Op, SelectionDAG &DAG, |
5084 | const X86Subtarget &Subtarget) { |
5085 | |
5086 | SDLoc dl(Op); |
5087 | SDValue Vec = Op.getOperand(0); |
5088 | SDValue SubVec = Op.getOperand(1); |
5089 | SDValue Idx = Op.getOperand(2); |
5090 | |
5091 | if (!isa<ConstantSDNode>(Idx)) |
5092 | return SDValue(); |
5093 | |
5094 | unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); |
5095 | if (IdxVal == 0 && Vec.isUndef()) // the operation is legal |
5096 | return Op; |
5097 | |
5098 | MVT OpVT = Op.getSimpleValueType(); |
5099 | MVT SubVecVT = SubVec.getSimpleValueType(); |
5100 | unsigned NumElems = OpVT.getVectorNumElements(); |
5101 | unsigned SubVecNumElems = SubVecVT.getVectorNumElements(); |
5102 | |
5103 | assert(IdxVal + SubVecNumElems <= NumElems &&((IdxVal + SubVecNumElems <= NumElems && IdxVal % SubVecVT .getSizeInBits() == 0 && "Unexpected index value in INSERT_SUBVECTOR" ) ? static_cast<void> (0) : __assert_fail ("IdxVal + SubVecNumElems <= NumElems && IdxVal % SubVecVT.getSizeInBits() == 0 && \"Unexpected index value in INSERT_SUBVECTOR\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5105, __PRETTY_FUNCTION__)) |
5104 | IdxVal % SubVecVT.getSizeInBits() == 0 &&((IdxVal + SubVecNumElems <= NumElems && IdxVal % SubVecVT .getSizeInBits() == 0 && "Unexpected index value in INSERT_SUBVECTOR" ) ? static_cast<void> (0) : __assert_fail ("IdxVal + SubVecNumElems <= NumElems && IdxVal % SubVecVT.getSizeInBits() == 0 && \"Unexpected index value in INSERT_SUBVECTOR\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5105, __PRETTY_FUNCTION__)) |
5105 | "Unexpected index value in INSERT_SUBVECTOR")((IdxVal + SubVecNumElems <= NumElems && IdxVal % SubVecVT .getSizeInBits() == 0 && "Unexpected index value in INSERT_SUBVECTOR" ) ? static_cast<void> (0) : __assert_fail ("IdxVal + SubVecNumElems <= NumElems && IdxVal % SubVecVT.getSizeInBits() == 0 && \"Unexpected index value in INSERT_SUBVECTOR\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5105, __PRETTY_FUNCTION__)); |
5106 | |
5107 | // There are 3 possible cases: |
5108 | // 1. Subvector should be inserted in the lower part (IdxVal == 0) |
5109 | // 2. Subvector should be inserted in the upper part |
5110 | // (IdxVal + SubVecNumElems == NumElems) |
5111 | // 3. Subvector should be inserted in the middle (for example v2i1 |
5112 | // to v16i1, index 2) |
5113 | |
5114 | // If this node widens - by concatenating zeroes - the type of the result |
5115 | // of a node with instruction that zeroes all upper (irrelevant) bits of the |
5116 | // output register, mark this node as legal to enable replacing them with |
5117 | // the v8i1 version of the previous instruction during instruction selection. |
5118 | // For example, VPCMPEQDZ128rr instruction stores its v4i1 result in a k-reg, |
5119 | // while zeroing all the upper remaining 60 bits of the register. if the |
5120 | // result of such instruction is inserted into an allZeroVector, then we can |
5121 | // safely remove insert_vector (in instruction selection) as the cmp instr |
5122 | // already zeroed the rest of the register. |
5123 | if (ISD::isBuildVectorAllZeros(Vec.getNode()) && IdxVal == 0 && |
5124 | (isMaskedZeroUpperBitsvXi1(SubVec.getOpcode()) || |
5125 | (SubVec.getOpcode() == ISD::AND && |
5126 | (isMaskedZeroUpperBitsvXi1(SubVec.getOperand(0).getOpcode()) || |
5127 | isMaskedZeroUpperBitsvXi1(SubVec.getOperand(1).getOpcode()))))) |
5128 | return Op; |
5129 | |
5130 | // extend to natively supported kshift |
5131 | MVT MinVT = Subtarget.hasDQI() ? MVT::v8i1 : MVT::v16i1; |
5132 | MVT WideOpVT = OpVT; |
5133 | if (OpVT.getSizeInBits() < MinVT.getStoreSizeInBits()) |
5134 | WideOpVT = MinVT; |
5135 | |
5136 | SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl); |
5137 | SDValue Undef = DAG.getUNDEF(WideOpVT); |
5138 | SDValue WideSubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT, |
5139 | Undef, SubVec, ZeroIdx); |
5140 | |
5141 | // Extract sub-vector if require. |
5142 | auto ExtractSubVec = [&](SDValue V) { |
5143 | return (WideOpVT == OpVT) ? V : DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, |
5144 | OpVT, V, ZeroIdx); |
5145 | }; |
5146 | |
5147 | if (Vec.isUndef()) { |
5148 | if (IdxVal != 0) { |
5149 | SDValue ShiftBits = DAG.getConstant(IdxVal, dl, MVT::i8); |
5150 | WideSubVec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, WideSubVec, |
5151 | ShiftBits); |
5152 | } |
5153 | return ExtractSubVec(WideSubVec); |
5154 | } |
5155 | |
5156 | if (ISD::isBuildVectorAllZeros(Vec.getNode())) { |
5157 | NumElems = WideOpVT.getVectorNumElements(); |
5158 | unsigned ShiftLeft = NumElems - SubVecNumElems; |
5159 | unsigned ShiftRight = NumElems - SubVecNumElems - IdxVal; |
5160 | Vec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, WideSubVec, |
5161 | DAG.getConstant(ShiftLeft, dl, MVT::i8)); |
5162 | Vec = ShiftRight ? DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, Vec, |
5163 | DAG.getConstant(ShiftRight, dl, MVT::i8)) : Vec; |
5164 | return ExtractSubVec(Vec); |
5165 | } |
5166 | |
5167 | if (IdxVal == 0) { |
5168 | // Zero lower bits of the Vec |
5169 | SDValue ShiftBits = DAG.getConstant(SubVecNumElems, dl, MVT::i8); |
5170 | Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT, Undef, Vec, ZeroIdx); |
5171 | Vec = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, Vec, ShiftBits); |
5172 | Vec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, Vec, ShiftBits); |
5173 | // Merge them together, SubVec should be zero extended. |
5174 | WideSubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT, |
5175 | getZeroVector(WideOpVT, Subtarget, DAG, dl), |
5176 | SubVec, ZeroIdx); |
5177 | Vec = DAG.getNode(ISD::OR, dl, WideOpVT, Vec, WideSubVec); |
5178 | return ExtractSubVec(Vec); |
5179 | } |
5180 | |
5181 | // Simple case when we put subvector in the upper part |
5182 | if (IdxVal + SubVecNumElems == NumElems) { |
5183 | // Zero upper bits of the Vec |
5184 | WideSubVec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, WideSubVec, |
5185 | DAG.getConstant(IdxVal, dl, MVT::i8)); |
5186 | SDValue ShiftBits = DAG.getConstant(SubVecNumElems, dl, MVT::i8); |
5187 | Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT, Undef, Vec, ZeroIdx); |
5188 | Vec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, Vec, ShiftBits); |
5189 | Vec = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, Vec, ShiftBits); |
5190 | Vec = DAG.getNode(ISD::OR, dl, WideOpVT, Vec, WideSubVec); |
5191 | return ExtractSubVec(Vec); |
5192 | } |
5193 | // Subvector should be inserted in the middle - use shuffle |
5194 | WideSubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef, |
5195 | SubVec, ZeroIdx); |
5196 | SmallVector<int, 64> Mask; |
5197 | for (unsigned i = 0; i < NumElems; ++i) |
5198 | Mask.push_back(i >= IdxVal && i < IdxVal + SubVecNumElems ? |
5199 | i : i + NumElems); |
5200 | return DAG.getVectorShuffle(OpVT, dl, WideSubVec, Vec, Mask); |
5201 | } |
5202 | |
5203 | /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128 |
5204 | /// instructions. This is used because creating CONCAT_VECTOR nodes of |
5205 | /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower |
5206 | /// large BUILD_VECTORS. |
5207 | static SDValue concat128BitVectors(SDValue V1, SDValue V2, EVT VT, |
5208 | unsigned NumElems, SelectionDAG &DAG, |
5209 | const SDLoc &dl) { |
5210 | SDValue V = insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl); |
5211 | return insert128BitVector(V, V2, NumElems / 2, DAG, dl); |
5212 | } |
5213 | |
5214 | static SDValue concat256BitVectors(SDValue V1, SDValue V2, EVT VT, |
5215 | unsigned NumElems, SelectionDAG &DAG, |
5216 | const SDLoc &dl) { |
5217 | SDValue V = insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl); |
5218 | return insert256BitVector(V, V2, NumElems / 2, DAG, dl); |
5219 | } |
5220 | |
5221 | /// Returns a vector of specified type with all bits set. |
5222 | /// Always build ones vectors as <4 x i32>, <8 x i32> or <16 x i32>. |
5223 | /// Then bitcast to their original type, ensuring they get CSE'd. |
5224 | static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, const SDLoc &dl) { |
5225 | assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&(((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector ()) && "Expected a 128/256/512-bit vector type") ? static_cast <void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) && \"Expected a 128/256/512-bit vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5226, __PRETTY_FUNCTION__)) |
5226 | "Expected a 128/256/512-bit vector type")(((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector ()) && "Expected a 128/256/512-bit vector type") ? static_cast <void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) && \"Expected a 128/256/512-bit vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5226, __PRETTY_FUNCTION__)); |
5227 | |
5228 | APInt Ones = APInt::getAllOnesValue(32); |
5229 | unsigned NumElts = VT.getSizeInBits() / 32; |
5230 | SDValue Vec = DAG.getConstant(Ones, dl, MVT::getVectorVT(MVT::i32, NumElts)); |
5231 | return DAG.getBitcast(VT, Vec); |
5232 | } |
5233 | |
5234 | static SDValue getExtendInVec(unsigned Opc, const SDLoc &DL, EVT VT, SDValue In, |
5235 | SelectionDAG &DAG) { |
5236 | EVT InVT = In.getValueType(); |
5237 | assert((X86ISD::VSEXT == Opc || X86ISD::VZEXT == Opc) && "Unexpected opcode")(((X86ISD::VSEXT == Opc || X86ISD::VZEXT == Opc) && "Unexpected opcode" ) ? static_cast<void> (0) : __assert_fail ("(X86ISD::VSEXT == Opc || X86ISD::VZEXT == Opc) && \"Unexpected opcode\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5237, __PRETTY_FUNCTION__)); |
5238 | |
5239 | if (VT.is128BitVector() && InVT.is128BitVector()) |
5240 | return X86ISD::VSEXT == Opc ? DAG.getSignExtendVectorInReg(In, DL, VT) |
5241 | : DAG.getZeroExtendVectorInReg(In, DL, VT); |
5242 | |
5243 | // For 256-bit vectors, we only need the lower (128-bit) input half. |
5244 | // For 512-bit vectors, we only need the lower input half or quarter. |
5245 | if (VT.getSizeInBits() > 128 && InVT.getSizeInBits() > 128) { |
5246 | int Scale = VT.getScalarSizeInBits() / InVT.getScalarSizeInBits(); |
5247 | In = extractSubVector(In, 0, DAG, DL, |
5248 | std::max(128, (int)VT.getSizeInBits() / Scale)); |
5249 | } |
5250 | |
5251 | return DAG.getNode(Opc, DL, VT, In); |
5252 | } |
5253 | |
5254 | /// Generate unpacklo/unpackhi shuffle mask. |
5255 | static void createUnpackShuffleMask(MVT VT, SmallVectorImpl<int> &Mask, bool Lo, |
5256 | bool Unary) { |
5257 | assert(Mask.empty() && "Expected an empty shuffle mask vector")((Mask.empty() && "Expected an empty shuffle mask vector" ) ? static_cast<void> (0) : __assert_fail ("Mask.empty() && \"Expected an empty shuffle mask vector\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5257, __PRETTY_FUNCTION__)); |
5258 | int NumElts = VT.getVectorNumElements(); |
5259 | int NumEltsInLane = 128 / VT.getScalarSizeInBits(); |
5260 | |
5261 | for (int i = 0; i < NumElts; ++i) { |
5262 | unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane; |
5263 | int Pos = (i % NumEltsInLane) / 2 + LaneStart; |
5264 | Pos += (Unary ? 0 : NumElts * (i % 2)); |
5265 | Pos += (Lo ? 0 : NumEltsInLane / 2); |
5266 | Mask.push_back(Pos); |
5267 | } |
5268 | } |
5269 | |
5270 | /// Returns a vector_shuffle node for an unpackl operation. |
5271 | static SDValue getUnpackl(SelectionDAG &DAG, const SDLoc &dl, MVT VT, |
5272 | SDValue V1, SDValue V2) { |
5273 | SmallVector<int, 8> Mask; |
5274 | createUnpackShuffleMask(VT, Mask, /* Lo = */ true, /* Unary = */ false); |
5275 | return DAG.getVectorShuffle(VT, dl, V1, V2, Mask); |
5276 | } |
5277 | |
5278 | /// Returns a vector_shuffle node for an unpackh operation. |
5279 | static SDValue getUnpackh(SelectionDAG &DAG, const SDLoc &dl, MVT VT, |
5280 | SDValue V1, SDValue V2) { |
5281 | SmallVector<int, 8> Mask; |
5282 | createUnpackShuffleMask(VT, Mask, /* Lo = */ false, /* Unary = */ false); |
5283 | return DAG.getVectorShuffle(VT, dl, V1, V2, Mask); |
5284 | } |
5285 | |
5286 | /// Return a vector_shuffle of the specified vector of zero or undef vector. |
5287 | /// This produces a shuffle where the low element of V2 is swizzled into the |
5288 | /// zero/undef vector, landing at element Idx. |
5289 | /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3). |
5290 | static SDValue getShuffleVectorZeroOrUndef(SDValue V2, int Idx, |
5291 | bool IsZero, |
5292 | const X86Subtarget &Subtarget, |
5293 | SelectionDAG &DAG) { |
5294 | MVT VT = V2.getSimpleValueType(); |
5295 | SDValue V1 = IsZero |
5296 | ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT); |
5297 | int NumElems = VT.getVectorNumElements(); |
5298 | SmallVector<int, 16> MaskVec(NumElems); |
5299 | for (int i = 0; i != NumElems; ++i) |
5300 | // If this is the insertion idx, put the low elt of V2 here. |
5301 | MaskVec[i] = (i == Idx) ? NumElems : i; |
5302 | return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, MaskVec); |
5303 | } |
5304 | |
5305 | static SDValue peekThroughBitcasts(SDValue V) { |
5306 | while (V.getNode() && V.getOpcode() == ISD::BITCAST) |
5307 | V = V.getOperand(0); |
5308 | return V; |
5309 | } |
5310 | |
5311 | static SDValue peekThroughOneUseBitcasts(SDValue V) { |
5312 | while (V.getNode() && V.getOpcode() == ISD::BITCAST && |
5313 | V.getOperand(0).hasOneUse()) |
5314 | V = V.getOperand(0); |
5315 | return V; |
5316 | } |
5317 | |
5318 | static const Constant *getTargetConstantFromNode(SDValue Op) { |
5319 | Op = peekThroughBitcasts(Op); |
5320 | |
5321 | auto *Load = dyn_cast<LoadSDNode>(Op); |
5322 | if (!Load) |
5323 | return nullptr; |
5324 | |
5325 | SDValue Ptr = Load->getBasePtr(); |
5326 | if (Ptr->getOpcode() == X86ISD::Wrapper || |
5327 | Ptr->getOpcode() == X86ISD::WrapperRIP) |
5328 | Ptr = Ptr->getOperand(0); |
5329 | |
5330 | auto *CNode = dyn_cast<ConstantPoolSDNode>(Ptr); |
5331 | if (!CNode || CNode->isMachineConstantPoolEntry()) |
5332 | return nullptr; |
5333 | |
5334 | return dyn_cast<Constant>(CNode->getConstVal()); |
5335 | } |
5336 | |
5337 | // Extract raw constant bits from constant pools. |
5338 | static bool getTargetConstantBitsFromNode(SDValue Op, unsigned EltSizeInBits, |
5339 | APInt &UndefElts, |
5340 | SmallVectorImpl<APInt> &EltBits, |
5341 | bool AllowWholeUndefs = true, |
5342 | bool AllowPartialUndefs = true) { |
5343 | assert(EltBits.empty() && "Expected an empty EltBits vector")((EltBits.empty() && "Expected an empty EltBits vector" ) ? static_cast<void> (0) : __assert_fail ("EltBits.empty() && \"Expected an empty EltBits vector\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5343, __PRETTY_FUNCTION__)); |
5344 | |
5345 | Op = peekThroughBitcasts(Op); |
5346 | |
5347 | EVT VT = Op.getValueType(); |
5348 | unsigned SizeInBits = VT.getSizeInBits(); |
5349 | assert((SizeInBits % EltSizeInBits) == 0 && "Can't split constant!")(((SizeInBits % EltSizeInBits) == 0 && "Can't split constant!" ) ? static_cast<void> (0) : __assert_fail ("(SizeInBits % EltSizeInBits) == 0 && \"Can't split constant!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5349, __PRETTY_FUNCTION__)); |
5350 | unsigned NumElts = SizeInBits / EltSizeInBits; |
5351 | |
5352 | // Bitcast a source array of element bits to the target size. |
5353 | auto CastBitData = [&](APInt &UndefSrcElts, ArrayRef<APInt> SrcEltBits) { |
5354 | unsigned NumSrcElts = UndefSrcElts.getBitWidth(); |
5355 | unsigned SrcEltSizeInBits = SrcEltBits[0].getBitWidth(); |
5356 | assert((NumSrcElts * SrcEltSizeInBits) == SizeInBits &&(((NumSrcElts * SrcEltSizeInBits) == SizeInBits && "Constant bit sizes don't match" ) ? static_cast<void> (0) : __assert_fail ("(NumSrcElts * SrcEltSizeInBits) == SizeInBits && \"Constant bit sizes don't match\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5357, __PRETTY_FUNCTION__)) |
5357 | "Constant bit sizes don't match")(((NumSrcElts * SrcEltSizeInBits) == SizeInBits && "Constant bit sizes don't match" ) ? static_cast<void> (0) : __assert_fail ("(NumSrcElts * SrcEltSizeInBits) == SizeInBits && \"Constant bit sizes don't match\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5357, __PRETTY_FUNCTION__)); |
5358 | |
5359 | // Don't split if we don't allow undef bits. |
5360 | bool AllowUndefs = AllowWholeUndefs || AllowPartialUndefs; |
5361 | if (UndefSrcElts.getBoolValue() && !AllowUndefs) |
5362 | return false; |
5363 | |
5364 | // If we're already the right size, don't bother bitcasting. |
5365 | if (NumSrcElts == NumElts) { |
5366 | UndefElts = UndefSrcElts; |
5367 | EltBits.assign(SrcEltBits.begin(), SrcEltBits.end()); |
5368 | return true; |
5369 | } |
5370 | |
5371 | // Extract all the undef/constant element data and pack into single bitsets. |
5372 | APInt UndefBits(SizeInBits, 0); |
5373 | APInt MaskBits(SizeInBits, 0); |
5374 | |
5375 | for (unsigned i = 0; i != NumSrcElts; ++i) { |
5376 | unsigned BitOffset = i * SrcEltSizeInBits; |
5377 | if (UndefSrcElts[i]) |
5378 | UndefBits.setBits(BitOffset, BitOffset + SrcEltSizeInBits); |
5379 | MaskBits.insertBits(SrcEltBits[i], BitOffset); |
5380 | } |
5381 | |
5382 | // Split the undef/constant single bitset data into the target elements. |
5383 | UndefElts = APInt(NumElts, 0); |
5384 | EltBits.resize(NumElts, APInt(EltSizeInBits, 0)); |
5385 | |
5386 | for (unsigned i = 0; i != NumElts; ++i) { |
5387 | unsigned BitOffset = i * EltSizeInBits; |
5388 | APInt UndefEltBits = UndefBits.extractBits(EltSizeInBits, BitOffset); |
5389 | |
5390 | // Only treat an element as UNDEF if all bits are UNDEF. |
5391 | if (UndefEltBits.isAllOnesValue()) { |
5392 | if (!AllowWholeUndefs) |
5393 | return false; |
5394 | UndefElts.setBit(i); |
5395 | continue; |
5396 | } |
5397 | |
5398 | // If only some bits are UNDEF then treat them as zero (or bail if not |
5399 | // supported). |
5400 | if (UndefEltBits.getBoolValue() && !AllowPartialUndefs) |
5401 | return false; |
5402 | |
5403 | APInt Bits = MaskBits.extractBits(EltSizeInBits, BitOffset); |
5404 | EltBits[i] = Bits.getZExtValue(); |
5405 | } |
5406 | return true; |
5407 | }; |
5408 | |
5409 | // Collect constant bits and insert into mask/undef bit masks. |
5410 | auto CollectConstantBits = [](const Constant *Cst, APInt &Mask, APInt &Undefs, |
5411 | unsigned UndefBitIndex) { |
5412 | if (!Cst) |
5413 | return false; |
5414 | if (isa<UndefValue>(Cst)) { |
5415 | Undefs.setBit(UndefBitIndex); |
5416 | return true; |
5417 | } |
5418 | if (auto *CInt = dyn_cast<ConstantInt>(Cst)) { |
5419 | Mask = CInt->getValue(); |
5420 | return true; |
5421 | } |
5422 | if (auto *CFP = dyn_cast<ConstantFP>(Cst)) { |
5423 | Mask = CFP->getValueAPF().bitcastToAPInt(); |
5424 | return true; |
5425 | } |
5426 | return false; |
5427 | }; |
5428 | |
5429 | // Extract constant bits from build vector. |
5430 | if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) { |
5431 | unsigned SrcEltSizeInBits = VT.getScalarSizeInBits(); |
5432 | unsigned NumSrcElts = SizeInBits / SrcEltSizeInBits; |
5433 | |
5434 | APInt UndefSrcElts(NumSrcElts, 0); |
5435 | SmallVector<APInt, 64> SrcEltBits(NumSrcElts, APInt(SrcEltSizeInBits, 0)); |
5436 | for (unsigned i = 0, e = Op.getNumOperands(); i != e; ++i) { |
5437 | const SDValue &Src = Op.getOperand(i); |
5438 | if (Src.isUndef()) { |
5439 | UndefSrcElts.setBit(i); |
5440 | continue; |
5441 | } |
5442 | auto *Cst = cast<ConstantSDNode>(Src); |
5443 | SrcEltBits[i] = Cst->getAPIntValue().zextOrTrunc(SrcEltSizeInBits); |
5444 | } |
5445 | return CastBitData(UndefSrcElts, SrcEltBits); |
5446 | } |
5447 | |
5448 | // Extract constant bits from constant pool vector. |
5449 | if (auto *Cst = getTargetConstantFromNode(Op)) { |
5450 | Type *CstTy = Cst->getType(); |
5451 | if (!CstTy->isVectorTy() || (SizeInBits != CstTy->getPrimitiveSizeInBits())) |
5452 | return false; |
5453 | |
5454 | unsigned SrcEltSizeInBits = CstTy->getScalarSizeInBits(); |
5455 | unsigned NumSrcElts = CstTy->getVectorNumElements(); |
5456 | |
5457 | APInt UndefSrcElts(NumSrcElts, 0); |
5458 | SmallVector<APInt, 64> SrcEltBits(NumSrcElts, APInt(SrcEltSizeInBits, 0)); |
5459 | for (unsigned i = 0; i != NumSrcElts; ++i) |
5460 | if (!CollectConstantBits(Cst->getAggregateElement(i), SrcEltBits[i], |
5461 | UndefSrcElts, i)) |
5462 | return false; |
5463 | |
5464 | return CastBitData(UndefSrcElts, SrcEltBits); |
5465 | } |
5466 | |
5467 | // Extract constant bits from a broadcasted constant pool scalar. |
5468 | if (Op.getOpcode() == X86ISD::VBROADCAST && |
5469 | EltSizeInBits <= VT.getScalarSizeInBits()) { |
5470 | if (auto *Broadcast = getTargetConstantFromNode(Op.getOperand(0))) { |
5471 | unsigned SrcEltSizeInBits = Broadcast->getType()->getScalarSizeInBits(); |
5472 | unsigned NumSrcElts = SizeInBits / SrcEltSizeInBits; |
5473 | |
5474 | APInt UndefSrcElts(NumSrcElts, 0); |
5475 | SmallVector<APInt, 64> SrcEltBits(1, APInt(SrcEltSizeInBits, 0)); |
5476 | if (CollectConstantBits(Broadcast, SrcEltBits[0], UndefSrcElts, 0)) { |
5477 | if (UndefSrcElts[0]) |
5478 | UndefSrcElts.setBits(0, NumSrcElts); |
5479 | SrcEltBits.append(NumSrcElts - 1, SrcEltBits[0]); |
5480 | return CastBitData(UndefSrcElts, SrcEltBits); |
5481 | } |
5482 | } |
5483 | } |
5484 | |
5485 | // Extract a rematerialized scalar constant insertion. |
5486 | if (Op.getOpcode() == X86ISD::VZEXT_MOVL && |
5487 | Op.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR && |
5488 | isa<ConstantSDNode>(Op.getOperand(0).getOperand(0))) { |
5489 | unsigned SrcEltSizeInBits = VT.getScalarSizeInBits(); |
5490 | unsigned NumSrcElts = SizeInBits / SrcEltSizeInBits; |
5491 | |
5492 | APInt UndefSrcElts(NumSrcElts, 0); |
5493 | SmallVector<APInt, 64> SrcEltBits; |
5494 | auto *CN = cast<ConstantSDNode>(Op.getOperand(0).getOperand(0)); |
5495 | SrcEltBits.push_back(CN->getAPIntValue().zextOrTrunc(SrcEltSizeInBits)); |
5496 | SrcEltBits.append(NumSrcElts - 1, APInt(SrcEltSizeInBits, 0)); |
5497 | return CastBitData(UndefSrcElts, SrcEltBits); |
5498 | } |
5499 | |
5500 | return false; |
5501 | } |
5502 | |
5503 | static bool getTargetShuffleMaskIndices(SDValue MaskNode, |
5504 | unsigned MaskEltSizeInBits, |
5505 | SmallVectorImpl<uint64_t> &RawMask) { |
5506 | APInt UndefElts; |
5507 | SmallVector<APInt, 64> EltBits; |
5508 | |
5509 | // Extract the raw target constant bits. |
5510 | // FIXME: We currently don't support UNDEF bits or mask entries. |
5511 | if (!getTargetConstantBitsFromNode(MaskNode, MaskEltSizeInBits, UndefElts, |
5512 | EltBits, /* AllowWholeUndefs */ false, |
5513 | /* AllowPartialUndefs */ false)) |
5514 | return false; |
5515 | |
5516 | // Insert the extracted elements into the mask. |
5517 | for (APInt Elt : EltBits) |
5518 | RawMask.push_back(Elt.getZExtValue()); |
5519 | |
5520 | return true; |
5521 | } |
5522 | |
5523 | /// Calculates the shuffle mask corresponding to the target-specific opcode. |
5524 | /// If the mask could be calculated, returns it in \p Mask, returns the shuffle |
5525 | /// operands in \p Ops, and returns true. |
5526 | /// Sets \p IsUnary to true if only one source is used. Note that this will set |
5527 | /// IsUnary for shuffles which use a single input multiple times, and in those |
5528 | /// cases it will adjust the mask to only have indices within that single input. |
5529 | /// It is an error to call this with non-empty Mask/Ops vectors. |
5530 | static bool getTargetShuffleMask(SDNode *N, MVT VT, bool AllowSentinelZero, |
5531 | SmallVectorImpl<SDValue> &Ops, |
5532 | SmallVectorImpl<int> &Mask, bool &IsUnary) { |
5533 | unsigned NumElems = VT.getVectorNumElements(); |
5534 | SDValue ImmN; |
5535 | |
5536 | assert(Mask.empty() && "getTargetShuffleMask expects an empty Mask vector")((Mask.empty() && "getTargetShuffleMask expects an empty Mask vector" ) ? static_cast<void> (0) : __assert_fail ("Mask.empty() && \"getTargetShuffleMask expects an empty Mask vector\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5536, __PRETTY_FUNCTION__)); |
5537 | assert(Ops.empty() && "getTargetShuffleMask expects an empty Ops vector")((Ops.empty() && "getTargetShuffleMask expects an empty Ops vector" ) ? static_cast<void> (0) : __assert_fail ("Ops.empty() && \"getTargetShuffleMask expects an empty Ops vector\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5537, __PRETTY_FUNCTION__)); |
5538 | |
5539 | IsUnary = false; |
5540 | bool IsFakeUnary = false; |
5541 | switch(N->getOpcode()) { |
5542 | case X86ISD::BLENDI: |
5543 | ImmN = N->getOperand(N->getNumOperands()-1); |
5544 | DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); |
5545 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); |
5546 | break; |
5547 | case X86ISD::SHUFP: |
5548 | ImmN = N->getOperand(N->getNumOperands()-1); |
5549 | DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); |
5550 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); |
5551 | break; |
5552 | case X86ISD::INSERTPS: |
5553 | ImmN = N->getOperand(N->getNumOperands()-1); |
5554 | DecodeINSERTPSMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); |
5555 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); |
5556 | break; |
5557 | case X86ISD::UNPCKH: |
5558 | DecodeUNPCKHMask(VT, Mask); |
5559 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); |
5560 | break; |
5561 | case X86ISD::UNPCKL: |
5562 | DecodeUNPCKLMask(VT, Mask); |
5563 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); |
5564 | break; |
5565 | case X86ISD::MOVHLPS: |
5566 | DecodeMOVHLPSMask(NumElems, Mask); |
5567 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); |
5568 | break; |
5569 | case X86ISD::MOVLHPS: |
5570 | DecodeMOVLHPSMask(NumElems, Mask); |
5571 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); |
5572 | break; |
5573 | case X86ISD::PALIGNR: |
5574 | assert(VT.getScalarType() == MVT::i8 && "Byte vector expected")((VT.getScalarType() == MVT::i8 && "Byte vector expected" ) ? static_cast<void> (0) : __assert_fail ("VT.getScalarType() == MVT::i8 && \"Byte vector expected\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5574, __PRETTY_FUNCTION__)); |
5575 | ImmN = N->getOperand(N->getNumOperands()-1); |
5576 | DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); |
5577 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); |
5578 | Ops.push_back(N->getOperand(1)); |
5579 | Ops.push_back(N->getOperand(0)); |
5580 | break; |
5581 | case X86ISD::VSHLDQ: |
5582 | assert(VT.getScalarType() == MVT::i8 && "Byte vector expected")((VT.getScalarType() == MVT::i8 && "Byte vector expected" ) ? static_cast<void> (0) : __assert_fail ("VT.getScalarType() == MVT::i8 && \"Byte vector expected\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5582, __PRETTY_FUNCTION__)); |
5583 | ImmN = N->getOperand(N->getNumOperands() - 1); |
5584 | DecodePSLLDQMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); |
5585 | IsUnary = true; |
5586 | break; |
5587 | case X86ISD::VSRLDQ: |
5588 | assert(VT.getScalarType() == MVT::i8 && "Byte vector expected")((VT.getScalarType() == MVT::i8 && "Byte vector expected" ) ? static_cast<void> (0) : __assert_fail ("VT.getScalarType() == MVT::i8 && \"Byte vector expected\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5588, __PRETTY_FUNCTION__)); |
5589 | ImmN = N->getOperand(N->getNumOperands() - 1); |
5590 | DecodePSRLDQMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); |
5591 | IsUnary = true; |
5592 | break; |
5593 | case X86ISD::PSHUFD: |
5594 | case X86ISD::VPERMILPI: |
5595 | ImmN = N->getOperand(N->getNumOperands()-1); |
5596 | DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); |
5597 | IsUnary = true; |
5598 | break; |
5599 | case X86ISD::PSHUFHW: |
5600 | ImmN = N->getOperand(N->getNumOperands()-1); |
5601 | DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); |
5602 | IsUnary = true; |
5603 | break; |
5604 | case X86ISD::PSHUFLW: |
5605 | ImmN = N->getOperand(N->getNumOperands()-1); |
5606 | DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); |
5607 | IsUnary = true; |
5608 | break; |
5609 | case X86ISD::VZEXT_MOVL: |
5610 | DecodeZeroMoveLowMask(VT, Mask); |
5611 | IsUnary = true; |
5612 | break; |
5613 | case X86ISD::VBROADCAST: { |
5614 | SDValue N0 = N->getOperand(0); |
5615 | // See if we're broadcasting from index 0 of an EXTRACT_SUBVECTOR. If so, |
5616 | // add the pre-extracted value to the Ops vector. |
5617 | if (N0.getOpcode() == ISD::EXTRACT_SUBVECTOR && |
5618 | N0.getOperand(0).getValueType() == VT && |
5619 | N0.getConstantOperandVal(1) == 0) |
5620 | Ops.push_back(N0.getOperand(0)); |
5621 | |
5622 | // We only decode broadcasts of same-sized vectors, unless the broadcast |
5623 | // came from an extract from the original width. If we found one, we |
5624 | // pushed it the Ops vector above. |
5625 | if (N0.getValueType() == VT || !Ops.empty()) { |
5626 | DecodeVectorBroadcast(VT, Mask); |
5627 | IsUnary = true; |
5628 | break; |
5629 | } |
5630 | return false; |
5631 | } |
5632 | case X86ISD::VPERMILPV: { |
5633 | IsUnary = true; |
5634 | SDValue MaskNode = N->getOperand(1); |
5635 | unsigned MaskEltSize = VT.getScalarSizeInBits(); |
5636 | SmallVector<uint64_t, 32> RawMask; |
5637 | if (getTargetShuffleMaskIndices(MaskNode, MaskEltSize, RawMask)) { |
5638 | DecodeVPERMILPMask(VT, RawMask, Mask); |
5639 | break; |
5640 | } |
5641 | if (auto *C = getTargetConstantFromNode(MaskNode)) { |
5642 | DecodeVPERMILPMask(C, MaskEltSize, Mask); |
5643 | break; |
5644 | } |
5645 | return false; |
5646 | } |
5647 | case X86ISD::PSHUFB: { |
5648 | IsUnary = true; |
5649 | SDValue MaskNode = N->getOperand(1); |
5650 | SmallVector<uint64_t, 32> RawMask; |
5651 | if (getTargetShuffleMaskIndices(MaskNode, 8, RawMask)) { |
5652 | DecodePSHUFBMask(RawMask, Mask); |
5653 | break; |
5654 | } |
5655 | if (auto *C = getTargetConstantFromNode(MaskNode)) { |
5656 | DecodePSHUFBMask(C, Mask); |
5657 | break; |
5658 | } |
5659 | return false; |
5660 | } |
5661 | case X86ISD::VPERMI: |
5662 | ImmN = N->getOperand(N->getNumOperands()-1); |
5663 | DecodeVPERMMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); |
5664 | IsUnary = true; |
5665 | break; |
5666 | case X86ISD::MOVSS: |
5667 | case X86ISD::MOVSD: |
5668 | DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask); |
5669 | break; |
5670 | case X86ISD::VPERM2X128: |
5671 | ImmN = N->getOperand(N->getNumOperands()-1); |
5672 | DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); |
5673 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); |
5674 | break; |
5675 | case X86ISD::MOVSLDUP: |
5676 | DecodeMOVSLDUPMask(VT, Mask); |
5677 | IsUnary = true; |
5678 | break; |
5679 | case X86ISD::MOVSHDUP: |
5680 | DecodeMOVSHDUPMask(VT, Mask); |
5681 | IsUnary = true; |
5682 | break; |
5683 | case X86ISD::MOVDDUP: |
5684 | DecodeMOVDDUPMask(VT, Mask); |
5685 | IsUnary = true; |
5686 | break; |
5687 | case X86ISD::MOVLHPD: |
5688 | case X86ISD::MOVLPD: |
5689 | case X86ISD::MOVLPS: |
5690 | // Not yet implemented |
5691 | return false; |
5692 | case X86ISD::VPERMIL2: { |
5693 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); |
5694 | unsigned MaskEltSize = VT.getScalarSizeInBits(); |
5695 | SDValue MaskNode = N->getOperand(2); |
5696 | SDValue CtrlNode = N->getOperand(3); |
5697 | if (ConstantSDNode *CtrlOp = dyn_cast<ConstantSDNode>(CtrlNode)) { |
5698 | unsigned CtrlImm = CtrlOp->getZExtValue(); |
5699 | SmallVector<uint64_t, 32> RawMask; |
5700 | if (getTargetShuffleMaskIndices(MaskNode, MaskEltSize, RawMask)) { |
5701 | DecodeVPERMIL2PMask(VT, CtrlImm, RawMask, Mask); |
5702 | break; |
5703 | } |
5704 | if (auto *C = getTargetConstantFromNode(MaskNode)) { |
5705 | DecodeVPERMIL2PMask(C, CtrlImm, MaskEltSize, Mask); |
5706 | break; |
5707 | } |
5708 | } |
5709 | return false; |
5710 | } |
5711 | case X86ISD::VPPERM: { |
5712 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); |
5713 | SDValue MaskNode = N->getOperand(2); |
5714 | SmallVector<uint64_t, 32> RawMask; |
5715 | if (getTargetShuffleMaskIndices(MaskNode, 8, RawMask)) { |
5716 | DecodeVPPERMMask(RawMask, Mask); |
5717 | break; |
5718 | } |
5719 | if (auto *C = getTargetConstantFromNode(MaskNode)) { |
5720 | DecodeVPPERMMask(C, Mask); |
5721 | break; |
5722 | } |
5723 | return false; |
5724 | } |
5725 | case X86ISD::VPERMV: { |
5726 | IsUnary = true; |
5727 | // Unlike most shuffle nodes, VPERMV's mask operand is operand 0. |
5728 | Ops.push_back(N->getOperand(1)); |
5729 | SDValue MaskNode = N->getOperand(0); |
5730 | SmallVector<uint64_t, 32> RawMask; |
5731 | unsigned MaskEltSize = VT.getScalarSizeInBits(); |
5732 | if (getTargetShuffleMaskIndices(MaskNode, MaskEltSize, RawMask)) { |
5733 | DecodeVPERMVMask(RawMask, Mask); |
5734 | break; |
5735 | } |
5736 | if (auto *C = getTargetConstantFromNode(MaskNode)) { |
5737 | DecodeVPERMVMask(C, MaskEltSize, Mask); |
5738 | break; |
5739 | } |
5740 | return false; |
5741 | } |
5742 | case X86ISD::VPERMV3: { |
5743 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(2); |
5744 | // Unlike most shuffle nodes, VPERMV3's mask operand is the middle one. |
5745 | Ops.push_back(N->getOperand(0)); |
5746 | Ops.push_back(N->getOperand(2)); |
5747 | SDValue MaskNode = N->getOperand(1); |
5748 | unsigned MaskEltSize = VT.getScalarSizeInBits(); |
5749 | if (auto *C = getTargetConstantFromNode(MaskNode)) { |
5750 | DecodeVPERMV3Mask(C, MaskEltSize, Mask); |
5751 | break; |
5752 | } |
5753 | return false; |
5754 | } |
5755 | case X86ISD::VPERMIV3: { |
5756 | IsUnary = IsFakeUnary = N->getOperand(1) == N->getOperand(2); |
5757 | // Unlike most shuffle nodes, VPERMIV3's mask operand is the first one. |
5758 | Ops.push_back(N->getOperand(1)); |
5759 | Ops.push_back(N->getOperand(2)); |
5760 | SDValue MaskNode = N->getOperand(0); |
5761 | unsigned MaskEltSize = VT.getScalarSizeInBits(); |
5762 | if (auto *C = getTargetConstantFromNode(MaskNode)) { |
5763 | DecodeVPERMV3Mask(C, MaskEltSize, Mask); |
5764 | break; |
5765 | } |
5766 | return false; |
5767 | } |
5768 | default: llvm_unreachable("unknown target shuffle node")::llvm::llvm_unreachable_internal("unknown target shuffle node" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5768); |
5769 | } |
5770 | |
5771 | // Empty mask indicates the decode failed. |
5772 | if (Mask.empty()) |
5773 | return false; |
5774 | |
5775 | // Check if we're getting a shuffle mask with zero'd elements. |
5776 | if (!AllowSentinelZero) |
5777 | if (any_of(Mask, [](int M) { return M == SM_SentinelZero; })) |
5778 | return false; |
5779 | |
5780 | // If we have a fake unary shuffle, the shuffle mask is spread across two |
5781 | // inputs that are actually the same node. Re-map the mask to always point |
5782 | // into the first input. |
5783 | if (IsFakeUnary) |
5784 | for (int &M : Mask) |
5785 | if (M >= (int)Mask.size()) |
5786 | M -= Mask.size(); |
5787 | |
5788 | // If we didn't already add operands in the opcode-specific code, default to |
5789 | // adding 1 or 2 operands starting at 0. |
5790 | if (Ops.empty()) { |
5791 | Ops.push_back(N->getOperand(0)); |
5792 | if (!IsUnary || IsFakeUnary) |
5793 | Ops.push_back(N->getOperand(1)); |
5794 | } |
5795 | |
5796 | return true; |
5797 | } |
5798 | |
5799 | /// Check a target shuffle mask's inputs to see if we can set any values to |
5800 | /// SM_SentinelZero - this is for elements that are known to be zero |
5801 | /// (not just zeroable) from their inputs. |
5802 | /// Returns true if the target shuffle mask was decoded. |
5803 | static bool setTargetShuffleZeroElements(SDValue N, |
5804 | SmallVectorImpl<int> &Mask, |
5805 | SmallVectorImpl<SDValue> &Ops) { |
5806 | bool IsUnary; |
5807 | if (!isTargetShuffle(N.getOpcode())) |
5808 | return false; |
5809 | |
5810 | MVT VT = N.getSimpleValueType(); |
5811 | if (!getTargetShuffleMask(N.getNode(), VT, true, Ops, Mask, IsUnary)) |
5812 | return false; |
5813 | |
5814 | SDValue V1 = Ops[0]; |
5815 | SDValue V2 = IsUnary ? V1 : Ops[1]; |
5816 | |
5817 | V1 = peekThroughBitcasts(V1); |
5818 | V2 = peekThroughBitcasts(V2); |
5819 | |
5820 | assert((VT.getSizeInBits() % Mask.size()) == 0 &&(((VT.getSizeInBits() % Mask.size()) == 0 && "Illegal split of shuffle value type" ) ? static_cast<void> (0) : __assert_fail ("(VT.getSizeInBits() % Mask.size()) == 0 && \"Illegal split of shuffle value type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5821, __PRETTY_FUNCTION__)) |
5821 | "Illegal split of shuffle value type")(((VT.getSizeInBits() % Mask.size()) == 0 && "Illegal split of shuffle value type" ) ? static_cast<void> (0) : __assert_fail ("(VT.getSizeInBits() % Mask.size()) == 0 && \"Illegal split of shuffle value type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5821, __PRETTY_FUNCTION__)); |
5822 | unsigned EltSizeInBits = VT.getSizeInBits() / Mask.size(); |
5823 | |
5824 | // Extract known constant input data. |
5825 | APInt UndefSrcElts[2]; |
5826 | SmallVector<APInt, 32> SrcEltBits[2]; |
5827 | bool IsSrcConstant[2] = { |
5828 | getTargetConstantBitsFromNode(V1, EltSizeInBits, UndefSrcElts[0], |
5829 | SrcEltBits[0], true, false), |
5830 | getTargetConstantBitsFromNode(V2, EltSizeInBits, UndefSrcElts[1], |
5831 | SrcEltBits[1], true, false)}; |
5832 | |
5833 | for (int i = 0, Size = Mask.size(); i < Size; ++i) { |
5834 | int M = Mask[i]; |
5835 | |
5836 | // Already decoded as SM_SentinelZero / SM_SentinelUndef. |
5837 | if (M < 0) |
5838 | continue; |
5839 | |
5840 | // Determine shuffle input and normalize the mask. |
5841 | unsigned SrcIdx = M / Size; |
5842 | SDValue V = M < Size ? V1 : V2; |
5843 | M %= Size; |
5844 | |
5845 | // We are referencing an UNDEF input. |
5846 | if (V.isUndef()) { |
5847 | Mask[i] = SM_SentinelUndef; |
5848 | continue; |
5849 | } |
5850 | |
5851 | // SCALAR_TO_VECTOR - only the first element is defined, and the rest UNDEF. |
5852 | // TODO: We currently only set UNDEF for integer types - floats use the same |
5853 | // registers as vectors and many of the scalar folded loads rely on the |
5854 | // SCALAR_TO_VECTOR pattern. |
5855 | if (V.getOpcode() == ISD::SCALAR_TO_VECTOR && |
5856 | (Size % V.getValueType().getVectorNumElements()) == 0) { |
5857 | int Scale = Size / V.getValueType().getVectorNumElements(); |
5858 | int Idx = M / Scale; |
5859 | if (Idx != 0 && !VT.isFloatingPoint()) |
5860 | Mask[i] = SM_SentinelUndef; |
5861 | else if (Idx == 0 && X86::isZeroNode(V.getOperand(0))) |
5862 | Mask[i] = SM_SentinelZero; |
5863 | continue; |
5864 | } |
5865 | |
5866 | // Attempt to extract from the source's constant bits. |
5867 | if (IsSrcConstant[SrcIdx]) { |
5868 | if (UndefSrcElts[SrcIdx][M]) |
5869 | Mask[i] = SM_SentinelUndef; |
5870 | else if (SrcEltBits[SrcIdx][M] == 0) |
5871 | Mask[i] = SM_SentinelZero; |
5872 | } |
5873 | } |
5874 | |
5875 | assert(VT.getVectorNumElements() == Mask.size() &&((VT.getVectorNumElements() == Mask.size() && "Different mask size from vector size!" ) ? static_cast<void> (0) : __assert_fail ("VT.getVectorNumElements() == Mask.size() && \"Different mask size from vector size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5876, __PRETTY_FUNCTION__)) |
5876 | "Different mask size from vector size!")((VT.getVectorNumElements() == Mask.size() && "Different mask size from vector size!" ) ? static_cast<void> (0) : __assert_fail ("VT.getVectorNumElements() == Mask.size() && \"Different mask size from vector size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5876, __PRETTY_FUNCTION__)); |
5877 | return true; |
5878 | } |
5879 | |
5880 | // Attempt to decode ops that could be represented as a shuffle mask. |
5881 | // The decoded shuffle mask may contain a different number of elements to the |
5882 | // destination value type. |
5883 | static bool getFauxShuffleMask(SDValue N, SmallVectorImpl<int> &Mask, |
5884 | SmallVectorImpl<SDValue> &Ops, |
5885 | SelectionDAG &DAG) { |
5886 | Mask.clear(); |
5887 | Ops.clear(); |
5888 | |
5889 | MVT VT = N.getSimpleValueType(); |
5890 | unsigned NumElts = VT.getVectorNumElements(); |
5891 | unsigned NumSizeInBits = VT.getSizeInBits(); |
5892 | unsigned NumBitsPerElt = VT.getScalarSizeInBits(); |
5893 | assert((NumBitsPerElt % 8) == 0 && (NumSizeInBits % 8) == 0 &&(((NumBitsPerElt % 8) == 0 && (NumSizeInBits % 8) == 0 && "Expected byte aligned value types") ? static_cast <void> (0) : __assert_fail ("(NumBitsPerElt % 8) == 0 && (NumSizeInBits % 8) == 0 && \"Expected byte aligned value types\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5894, __PRETTY_FUNCTION__)) |
5894 | "Expected byte aligned value types")(((NumBitsPerElt % 8) == 0 && (NumSizeInBits % 8) == 0 && "Expected byte aligned value types") ? static_cast <void> (0) : __assert_fail ("(NumBitsPerElt % 8) == 0 && (NumSizeInBits % 8) == 0 && \"Expected byte aligned value types\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5894, __PRETTY_FUNCTION__)); |
5895 | |
5896 | unsigned Opcode = N.getOpcode(); |
5897 | switch (Opcode) { |
5898 | case ISD::AND: |
5899 | case X86ISD::ANDNP: { |
5900 | // Attempt to decode as a per-byte mask. |
5901 | APInt UndefElts; |
5902 | SmallVector<APInt, 32> EltBits; |
5903 | SDValue N0 = N.getOperand(0); |
5904 | SDValue N1 = N.getOperand(1); |
5905 | bool IsAndN = (X86ISD::ANDNP == Opcode); |
5906 | uint64_t ZeroMask = IsAndN ? 255 : 0; |
5907 | if (!getTargetConstantBitsFromNode(IsAndN ? N0 : N1, 8, UndefElts, EltBits)) |
5908 | return false; |
5909 | for (int i = 0, e = (int)EltBits.size(); i != e; ++i) { |
5910 | if (UndefElts[i]) { |
5911 | Mask.push_back(SM_SentinelUndef); |
5912 | continue; |
5913 | } |
5914 | uint64_t ByteBits = EltBits[i].getZExtValue(); |
5915 | if (ByteBits != 0 && ByteBits != 255) |
5916 | return false; |
5917 | Mask.push_back(ByteBits == ZeroMask ? SM_SentinelZero : i); |
5918 | } |
5919 | Ops.push_back(IsAndN ? N1 : N0); |
5920 | return true; |
5921 | } |
5922 | case ISD::SCALAR_TO_VECTOR: { |
5923 | // Match against a scalar_to_vector of an extract from a vector, |
5924 | // for PEXTRW/PEXTRB we must handle the implicit zext of the scalar. |
5925 | SDValue N0 = N.getOperand(0); |
5926 | SDValue SrcExtract; |
5927 | |
5928 | if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT && |
5929 | N0.getOperand(0).getValueType() == VT) { |
5930 | SrcExtract = N0; |
5931 | } else if (N0.getOpcode() == ISD::AssertZext && |
5932 | N0.getOperand(0).getOpcode() == X86ISD::PEXTRW && |
5933 | cast<VTSDNode>(N0.getOperand(1))->getVT() == MVT::i16) { |
5934 | SrcExtract = N0.getOperand(0); |
5935 | assert(SrcExtract.getOperand(0).getValueType() == MVT::v8i16)((SrcExtract.getOperand(0).getValueType() == MVT::v8i16) ? static_cast <void> (0) : __assert_fail ("SrcExtract.getOperand(0).getValueType() == MVT::v8i16" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5935, __PRETTY_FUNCTION__)); |
5936 | } else if (N0.getOpcode() == ISD::AssertZext && |
5937 | N0.getOperand(0).getOpcode() == X86ISD::PEXTRB && |
5938 | cast<VTSDNode>(N0.getOperand(1))->getVT() == MVT::i8) { |
5939 | SrcExtract = N0.getOperand(0); |
5940 | assert(SrcExtract.getOperand(0).getValueType() == MVT::v16i8)((SrcExtract.getOperand(0).getValueType() == MVT::v16i8) ? static_cast <void> (0) : __assert_fail ("SrcExtract.getOperand(0).getValueType() == MVT::v16i8" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5940, __PRETTY_FUNCTION__)); |
5941 | } |
5942 | |
5943 | if (!SrcExtract || !isa<ConstantSDNode>(SrcExtract.getOperand(1))) |
5944 | return false; |
5945 | |
5946 | SDValue SrcVec = SrcExtract.getOperand(0); |
5947 | EVT SrcVT = SrcVec.getValueType(); |
5948 | unsigned NumSrcElts = SrcVT.getVectorNumElements(); |
5949 | unsigned NumZeros = (NumBitsPerElt / SrcVT.getScalarSizeInBits()) - 1; |
5950 | |
5951 | unsigned SrcIdx = SrcExtract.getConstantOperandVal(1); |
5952 | if (NumSrcElts <= SrcIdx) |
5953 | return false; |
5954 | |
5955 | Ops.push_back(SrcVec); |
5956 | Mask.push_back(SrcIdx); |
5957 | Mask.append(NumZeros, SM_SentinelZero); |
5958 | Mask.append(NumSrcElts - Mask.size(), SM_SentinelUndef); |
5959 | return true; |
5960 | } |
5961 | case X86ISD::PINSRB: |
5962 | case X86ISD::PINSRW: { |
5963 | SDValue InVec = N.getOperand(0); |
5964 | SDValue InScl = N.getOperand(1); |
5965 | uint64_t InIdx = N.getConstantOperandVal(2); |
5966 | assert(InIdx < NumElts && "Illegal insertion index")((InIdx < NumElts && "Illegal insertion index") ? static_cast <void> (0) : __assert_fail ("InIdx < NumElts && \"Illegal insertion index\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5966, __PRETTY_FUNCTION__)); |
5967 | |
5968 | // Attempt to recognise a PINSR*(VEC, 0, Idx) shuffle pattern. |
5969 | if (X86::isZeroNode(InScl)) { |
5970 | Ops.push_back(InVec); |
5971 | for (unsigned i = 0; i != NumElts; ++i) |
5972 | Mask.push_back(i == InIdx ? SM_SentinelZero : (int)i); |
5973 | return true; |
5974 | } |
5975 | |
5976 | // Attempt to recognise a PINSR*(ASSERTZEXT(PEXTR*)) shuffle pattern. |
5977 | // TODO: Expand this to support INSERT_VECTOR_ELT/etc. |
5978 | unsigned ExOp = |
5979 | (X86ISD::PINSRB == Opcode ? X86ISD::PEXTRB : X86ISD::PEXTRW); |
5980 | if (InScl.getOpcode() != ISD::AssertZext || |
5981 | InScl.getOperand(0).getOpcode() != ExOp) |
5982 | return false; |
5983 | |
5984 | SDValue ExVec = InScl.getOperand(0).getOperand(0); |
5985 | uint64_t ExIdx = InScl.getOperand(0).getConstantOperandVal(1); |
5986 | assert(ExIdx < NumElts && "Illegal extraction index")((ExIdx < NumElts && "Illegal extraction index") ? static_cast<void> (0) : __assert_fail ("ExIdx < NumElts && \"Illegal extraction index\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 5986, __PRETTY_FUNCTION__)); |
5987 | Ops.push_back(InVec); |
5988 | Ops.push_back(ExVec); |
5989 | for (unsigned i = 0; i != NumElts; ++i) |
5990 | Mask.push_back(i == InIdx ? NumElts + ExIdx : i); |
5991 | return true; |
5992 | } |
5993 | case X86ISD::PACKSS: { |
5994 | // If we know input saturation won't happen we can treat this |
5995 | // as a truncation shuffle. |
5996 | if (DAG.ComputeNumSignBits(N.getOperand(0)) <= NumBitsPerElt || |
5997 | DAG.ComputeNumSignBits(N.getOperand(1)) <= NumBitsPerElt) |
5998 | return false; |
5999 | |
6000 | Ops.push_back(N.getOperand(0)); |
6001 | Ops.push_back(N.getOperand(1)); |
6002 | for (unsigned i = 0; i != NumElts; ++i) |
6003 | Mask.push_back(i * 2); |
6004 | return true; |
6005 | } |
6006 | case X86ISD::VSHLI: |
6007 | case X86ISD::VSRLI: { |
6008 | uint64_t ShiftVal = N.getConstantOperandVal(1); |
6009 | // Out of range bit shifts are guaranteed to be zero. |
6010 | if (NumBitsPerElt <= ShiftVal) { |
6011 | Mask.append(NumElts, SM_SentinelZero); |
6012 | return true; |
6013 | } |
6014 | |
6015 | // We can only decode 'whole byte' bit shifts as shuffles. |
6016 | if ((ShiftVal % 8) != 0) |
6017 | break; |
6018 | |
6019 | uint64_t ByteShift = ShiftVal / 8; |
6020 | unsigned NumBytes = NumSizeInBits / 8; |
6021 | unsigned NumBytesPerElt = NumBitsPerElt / 8; |
6022 | Ops.push_back(N.getOperand(0)); |
6023 | |
6024 | // Clear mask to all zeros and insert the shifted byte indices. |
6025 | Mask.append(NumBytes, SM_SentinelZero); |
6026 | |
6027 | if (X86ISD::VSHLI == Opcode) { |
6028 | for (unsigned i = 0; i != NumBytes; i += NumBytesPerElt) |
6029 | for (unsigned j = ByteShift; j != NumBytesPerElt; ++j) |
6030 | Mask[i + j] = i + j - ByteShift; |
6031 | } else { |
6032 | for (unsigned i = 0; i != NumBytes; i += NumBytesPerElt) |
6033 | for (unsigned j = ByteShift; j != NumBytesPerElt; ++j) |
6034 | Mask[i + j - ByteShift] = i + j; |
6035 | } |
6036 | return true; |
6037 | } |
6038 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
6039 | case X86ISD::VZEXT: { |
6040 | // TODO - add support for VPMOVZX with smaller input vector types. |
6041 | SDValue Src = N.getOperand(0); |
6042 | MVT SrcVT = Src.getSimpleValueType(); |
6043 | if (NumSizeInBits != SrcVT.getSizeInBits()) |
6044 | break; |
6045 | DecodeZeroExtendMask(SrcVT.getScalarType(), VT, Mask); |
6046 | Ops.push_back(Src); |
6047 | return true; |
6048 | } |
6049 | } |
6050 | |
6051 | return false; |
6052 | } |
6053 | |
6054 | /// Removes unused shuffle source inputs and adjusts the shuffle mask accordingly. |
6055 | static void resolveTargetShuffleInputsAndMask(SmallVectorImpl<SDValue> &Inputs, |
6056 | SmallVectorImpl<int> &Mask) { |
6057 | int MaskWidth = Mask.size(); |
6058 | SmallVector<SDValue, 16> UsedInputs; |
6059 | for (int i = 0, e = Inputs.size(); i < e; ++i) { |
6060 | int lo = UsedInputs.size() * MaskWidth; |
6061 | int hi = lo + MaskWidth; |
6062 | if (any_of(Mask, [lo, hi](int i) { return (lo <= i) && (i < hi); })) { |
6063 | UsedInputs.push_back(Inputs[i]); |
6064 | continue; |
6065 | } |
6066 | for (int &M : Mask) |
6067 | if (lo <= M) |
6068 | M -= MaskWidth; |
6069 | } |
6070 | Inputs = UsedInputs; |
6071 | } |
6072 | |
6073 | /// Calls setTargetShuffleZeroElements to resolve a target shuffle mask's inputs |
6074 | /// and set the SM_SentinelUndef and SM_SentinelZero values. Then check the |
6075 | /// remaining input indices in case we now have a unary shuffle and adjust the |
6076 | /// inputs accordingly. |
6077 | /// Returns true if the target shuffle mask was decoded. |
6078 | static bool resolveTargetShuffleInputs(SDValue Op, |
6079 | SmallVectorImpl<SDValue> &Inputs, |
6080 | SmallVectorImpl<int> &Mask, |
6081 | SelectionDAG &DAG) { |
6082 | if (!setTargetShuffleZeroElements(Op, Mask, Inputs)) |
6083 | if (!getFauxShuffleMask(Op, Mask, Inputs, DAG)) |
6084 | return false; |
6085 | |
6086 | resolveTargetShuffleInputsAndMask(Inputs, Mask); |
6087 | return true; |
6088 | } |
6089 | |
6090 | /// Returns the scalar element that will make up the ith |
6091 | /// element of the result of the vector shuffle. |
6092 | static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG, |
6093 | unsigned Depth) { |
6094 | if (Depth == 6) |
6095 | return SDValue(); // Limit search depth. |
6096 | |
6097 | SDValue V = SDValue(N, 0); |
6098 | EVT VT = V.getValueType(); |
6099 | unsigned Opcode = V.getOpcode(); |
6100 | |
6101 | // Recurse into ISD::VECTOR_SHUFFLE node to find scalars. |
6102 | if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) { |
6103 | int Elt = SV->getMaskElt(Index); |
6104 | |
6105 | if (Elt < 0) |
6106 | return DAG.getUNDEF(VT.getVectorElementType()); |
6107 | |
6108 | unsigned NumElems = VT.getVectorNumElements(); |
6109 | SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0) |
6110 | : SV->getOperand(1); |
6111 | return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1); |
6112 | } |
6113 | |
6114 | // Recurse into target specific vector shuffles to find scalars. |
6115 | if (isTargetShuffle(Opcode)) { |
6116 | MVT ShufVT = V.getSimpleValueType(); |
6117 | MVT ShufSVT = ShufVT.getVectorElementType(); |
6118 | int NumElems = (int)ShufVT.getVectorNumElements(); |
6119 | SmallVector<int, 16> ShuffleMask; |
6120 | SmallVector<SDValue, 16> ShuffleOps; |
6121 | bool IsUnary; |
6122 | |
6123 | if (!getTargetShuffleMask(N, ShufVT, true, ShuffleOps, ShuffleMask, IsUnary)) |
6124 | return SDValue(); |
6125 | |
6126 | int Elt = ShuffleMask[Index]; |
6127 | if (Elt == SM_SentinelZero) |
6128 | return ShufSVT.isInteger() ? DAG.getConstant(0, SDLoc(N), ShufSVT) |
6129 | : DAG.getConstantFP(+0.0, SDLoc(N), ShufSVT); |
6130 | if (Elt == SM_SentinelUndef) |
6131 | return DAG.getUNDEF(ShufSVT); |
6132 | |
6133 | assert(0 <= Elt && Elt < (2*NumElems) && "Shuffle index out of range")((0 <= Elt && Elt < (2*NumElems) && "Shuffle index out of range" ) ? static_cast<void> (0) : __assert_fail ("0 <= Elt && Elt < (2*NumElems) && \"Shuffle index out of range\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6133, __PRETTY_FUNCTION__)); |
6134 | SDValue NewV = (Elt < NumElems) ? ShuffleOps[0] : ShuffleOps[1]; |
6135 | return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, |
6136 | Depth+1); |
6137 | } |
6138 | |
6139 | // Actual nodes that may contain scalar elements |
6140 | if (Opcode == ISD::BITCAST) { |
6141 | V = V.getOperand(0); |
6142 | EVT SrcVT = V.getValueType(); |
6143 | unsigned NumElems = VT.getVectorNumElements(); |
6144 | |
6145 | if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems) |
6146 | return SDValue(); |
6147 | } |
6148 | |
6149 | if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) |
6150 | return (Index == 0) ? V.getOperand(0) |
6151 | : DAG.getUNDEF(VT.getVectorElementType()); |
6152 | |
6153 | if (V.getOpcode() == ISD::BUILD_VECTOR) |
6154 | return V.getOperand(Index); |
6155 | |
6156 | return SDValue(); |
6157 | } |
6158 | |
6159 | /// Custom lower build_vector of v16i8. |
6160 | static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros, |
6161 | unsigned NumNonZero, unsigned NumZero, |
6162 | SelectionDAG &DAG, |
6163 | const X86Subtarget &Subtarget) { |
6164 | if (NumNonZero > 8 && !Subtarget.hasSSE41()) |
6165 | return SDValue(); |
6166 | |
6167 | SDLoc dl(Op); |
6168 | SDValue V; |
6169 | bool First = true; |
6170 | |
6171 | // SSE4.1 - use PINSRB to insert each byte directly. |
6172 | if (Subtarget.hasSSE41()) { |
6173 | for (unsigned i = 0; i < 16; ++i) { |
6174 | bool IsNonZero = (NonZeros & (1 << i)) != 0; |
6175 | if (IsNonZero) { |
6176 | // If the build vector contains zeros or our first insertion is not the |
6177 | // first index then insert into zero vector to break any register |
6178 | // dependency else use SCALAR_TO_VECTOR/VZEXT_MOVL. |
6179 | if (First) { |
6180 | First = false; |
6181 | if (NumZero || 0 != i) |
6182 | V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl); |
6183 | else { |
6184 | assert(0 == i && "Expected insertion into zero-index")((0 == i && "Expected insertion into zero-index") ? static_cast <void> (0) : __assert_fail ("0 == i && \"Expected insertion into zero-index\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6184, __PRETTY_FUNCTION__)); |
6185 | V = DAG.getAnyExtOrTrunc(Op.getOperand(i), dl, MVT::i32); |
6186 | V = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, V); |
6187 | V = DAG.getNode(X86ISD::VZEXT_MOVL, dl, MVT::v4i32, V); |
6188 | V = DAG.getBitcast(MVT::v16i8, V); |
6189 | continue; |
6190 | } |
6191 | } |
6192 | V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v16i8, V, |
6193 | Op.getOperand(i), DAG.getIntPtrConstant(i, dl)); |
6194 | } |
6195 | } |
6196 | |
6197 | return V; |
6198 | } |
6199 | |
6200 | // Pre-SSE4.1 - merge byte pairs and insert with PINSRW. |
6201 | for (unsigned i = 0; i < 16; ++i) { |
6202 | bool ThisIsNonZero = (NonZeros & (1 << i)) != 0; |
6203 | if (ThisIsNonZero && First) { |
6204 | if (NumZero) |
6205 | V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl); |
6206 | else |
6207 | V = DAG.getUNDEF(MVT::v8i16); |
6208 | First = false; |
6209 | } |
6210 | |
6211 | if ((i & 1) != 0) { |
6212 | // FIXME: Investigate extending to i32 instead of just i16. |
6213 | // FIXME: Investigate combining the first 4 bytes as a i32 instead. |
6214 | SDValue ThisElt, LastElt; |
6215 | bool LastIsNonZero = (NonZeros & (1 << (i - 1))) != 0; |
6216 | if (LastIsNonZero) { |
6217 | LastElt = |
6218 | DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i - 1)); |
6219 | } |
6220 | if (ThisIsNonZero) { |
6221 | ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i)); |
6222 | ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16, ThisElt, |
6223 | DAG.getConstant(8, dl, MVT::i8)); |
6224 | if (LastIsNonZero) |
6225 | ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt); |
6226 | } else |
6227 | ThisElt = LastElt; |
6228 | |
6229 | if (ThisElt) { |
6230 | if (1 == i) { |
6231 | V = NumZero ? DAG.getZExtOrTrunc(ThisElt, dl, MVT::i32) |
6232 | : DAG.getAnyExtOrTrunc(ThisElt, dl, MVT::i32); |
6233 | V = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, V); |
6234 | V = DAG.getNode(X86ISD::VZEXT_MOVL, dl, MVT::v4i32, V); |
6235 | V = DAG.getBitcast(MVT::v8i16, V); |
6236 | } else { |
6237 | V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt, |
6238 | DAG.getIntPtrConstant(i / 2, dl)); |
6239 | } |
6240 | } |
6241 | } |
6242 | } |
6243 | |
6244 | return DAG.getBitcast(MVT::v16i8, V); |
6245 | } |
6246 | |
6247 | /// Custom lower build_vector of v8i16. |
6248 | static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros, |
6249 | unsigned NumNonZero, unsigned NumZero, |
6250 | SelectionDAG &DAG, |
6251 | const X86Subtarget &Subtarget) { |
6252 | if (NumNonZero > 4 && !Subtarget.hasSSE41()) |
6253 | return SDValue(); |
6254 | |
6255 | SDLoc dl(Op); |
6256 | SDValue V; |
6257 | bool First = true; |
6258 | for (unsigned i = 0; i < 8; ++i) { |
6259 | bool IsNonZero = (NonZeros & (1 << i)) != 0; |
6260 | if (IsNonZero) { |
6261 | // If the build vector contains zeros or our first insertion is not the |
6262 | // first index then insert into zero vector to break any register |
6263 | // dependency else use SCALAR_TO_VECTOR/VZEXT_MOVL. |
6264 | if (First) { |
6265 | First = false; |
6266 | if (NumZero || 0 != i) |
6267 | V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl); |
6268 | else { |
6269 | assert(0 == i && "Expected insertion into zero-index")((0 == i && "Expected insertion into zero-index") ? static_cast <void> (0) : __assert_fail ("0 == i && \"Expected insertion into zero-index\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6269, __PRETTY_FUNCTION__)); |
6270 | V = DAG.getAnyExtOrTrunc(Op.getOperand(i), dl, MVT::i32); |
6271 | V = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, V); |
6272 | V = DAG.getNode(X86ISD::VZEXT_MOVL, dl, MVT::v4i32, V); |
6273 | V = DAG.getBitcast(MVT::v8i16, V); |
6274 | continue; |
6275 | } |
6276 | } |
6277 | V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, |
6278 | Op.getOperand(i), DAG.getIntPtrConstant(i, dl)); |
6279 | } |
6280 | } |
6281 | |
6282 | return V; |
6283 | } |
6284 | |
6285 | /// Custom lower build_vector of v4i32 or v4f32. |
6286 | static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG, |
6287 | const X86Subtarget &Subtarget) { |
6288 | // Find all zeroable elements. |
6289 | std::bitset<4> Zeroable; |
6290 | for (int i=0; i < 4; ++i) { |
6291 | SDValue Elt = Op->getOperand(i); |
6292 | Zeroable[i] = (Elt.isUndef() || X86::isZeroNode(Elt)); |
6293 | } |
6294 | assert(Zeroable.size() - Zeroable.count() > 1 &&((Zeroable.size() - Zeroable.count() > 1 && "We expect at least two non-zero elements!" ) ? static_cast<void> (0) : __assert_fail ("Zeroable.size() - Zeroable.count() > 1 && \"We expect at least two non-zero elements!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6295, __PRETTY_FUNCTION__)) |
6295 | "We expect at least two non-zero elements!")((Zeroable.size() - Zeroable.count() > 1 && "We expect at least two non-zero elements!" ) ? static_cast<void> (0) : __assert_fail ("Zeroable.size() - Zeroable.count() > 1 && \"We expect at least two non-zero elements!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6295, __PRETTY_FUNCTION__)); |
6296 | |
6297 | // We only know how to deal with build_vector nodes where elements are either |
6298 | // zeroable or extract_vector_elt with constant index. |
6299 | SDValue FirstNonZero; |
6300 | unsigned FirstNonZeroIdx; |
6301 | for (unsigned i=0; i < 4; ++i) { |
6302 | if (Zeroable[i]) |
6303 | continue; |
6304 | SDValue Elt = Op->getOperand(i); |
6305 | if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT || |
6306 | !isa<ConstantSDNode>(Elt.getOperand(1))) |
6307 | return SDValue(); |
6308 | // Make sure that this node is extracting from a 128-bit vector. |
6309 | MVT VT = Elt.getOperand(0).getSimpleValueType(); |
6310 | if (!VT.is128BitVector()) |
6311 | return SDValue(); |
6312 | if (!FirstNonZero.getNode()) { |
6313 | FirstNonZero = Elt; |
6314 | FirstNonZeroIdx = i; |
6315 | } |
6316 | } |
6317 | |
6318 | assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!")((FirstNonZero.getNode() && "Unexpected build vector of all zeros!" ) ? static_cast<void> (0) : __assert_fail ("FirstNonZero.getNode() && \"Unexpected build vector of all zeros!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6318, __PRETTY_FUNCTION__)); |
6319 | SDValue V1 = FirstNonZero.getOperand(0); |
6320 | MVT VT = V1.getSimpleValueType(); |
6321 | |
6322 | // See if this build_vector can be lowered as a blend with zero. |
6323 | SDValue Elt; |
6324 | unsigned EltMaskIdx, EltIdx; |
6325 | int Mask[4]; |
6326 | for (EltIdx = 0; EltIdx < 4; ++EltIdx) { |
6327 | if (Zeroable[EltIdx]) { |
6328 | // The zero vector will be on the right hand side. |
6329 | Mask[EltIdx] = EltIdx+4; |
6330 | continue; |
6331 | } |
6332 | |
6333 | Elt = Op->getOperand(EltIdx); |
6334 | // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index. |
6335 | EltMaskIdx = Elt.getConstantOperandVal(1); |
6336 | if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx) |
6337 | break; |
6338 | Mask[EltIdx] = EltIdx; |
6339 | } |
6340 | |
6341 | if (EltIdx == 4) { |
6342 | // Let the shuffle legalizer deal with blend operations. |
6343 | SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op)); |
6344 | if (V1.getSimpleValueType() != VT) |
6345 | V1 = DAG.getBitcast(VT, V1); |
6346 | return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, Mask); |
6347 | } |
6348 | |
6349 | // See if we can lower this build_vector to a INSERTPS. |
6350 | if (!Subtarget.hasSSE41()) |
6351 | return SDValue(); |
6352 | |
6353 | SDValue V2 = Elt.getOperand(0); |
6354 | if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx) |
6355 | V1 = SDValue(); |
6356 | |
6357 | bool CanFold = true; |
6358 | for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) { |
6359 | if (Zeroable[i]) |
6360 | continue; |
6361 | |
6362 | SDValue Current = Op->getOperand(i); |
6363 | SDValue SrcVector = Current->getOperand(0); |
6364 | if (!V1.getNode()) |
6365 | V1 = SrcVector; |
6366 | CanFold = (SrcVector == V1) && (Current.getConstantOperandVal(1) == i); |
6367 | } |
6368 | |
6369 | if (!CanFold) |
6370 | return SDValue(); |
6371 | |
6372 | assert(V1.getNode() && "Expected at least two non-zero elements!")((V1.getNode() && "Expected at least two non-zero elements!" ) ? static_cast<void> (0) : __assert_fail ("V1.getNode() && \"Expected at least two non-zero elements!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6372, __PRETTY_FUNCTION__)); |
6373 | if (V1.getSimpleValueType() != MVT::v4f32) |
6374 | V1 = DAG.getBitcast(MVT::v4f32, V1); |
6375 | if (V2.getSimpleValueType() != MVT::v4f32) |
6376 | V2 = DAG.getBitcast(MVT::v4f32, V2); |
6377 | |
6378 | // Ok, we can emit an INSERTPS instruction. |
6379 | unsigned ZMask = Zeroable.to_ulong(); |
6380 | |
6381 | unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask; |
6382 | assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!")(((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!" ) ? static_cast<void> (0) : __assert_fail ("(InsertPSMask & ~0xFFu) == 0 && \"Invalid mask!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6382, __PRETTY_FUNCTION__)); |
6383 | SDLoc DL(Op); |
6384 | SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2, |
6385 | DAG.getIntPtrConstant(InsertPSMask, DL)); |
6386 | return DAG.getBitcast(VT, Result); |
6387 | } |
6388 | |
6389 | /// Return a vector logical shift node. |
6390 | static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp, unsigned NumBits, |
6391 | SelectionDAG &DAG, const TargetLowering &TLI, |
6392 | const SDLoc &dl) { |
6393 | assert(VT.is128BitVector() && "Unknown type for VShift")((VT.is128BitVector() && "Unknown type for VShift") ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"Unknown type for VShift\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6393, __PRETTY_FUNCTION__)); |
6394 | MVT ShVT = MVT::v16i8; |
6395 | unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ; |
6396 | SrcOp = DAG.getBitcast(ShVT, SrcOp); |
6397 | MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT); |
6398 | assert(NumBits % 8 == 0 && "Only support byte sized shifts")((NumBits % 8 == 0 && "Only support byte sized shifts" ) ? static_cast<void> (0) : __assert_fail ("NumBits % 8 == 0 && \"Only support byte sized shifts\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6398, __PRETTY_FUNCTION__)); |
6399 | SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy); |
6400 | return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal)); |
6401 | } |
6402 | |
6403 | static SDValue LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, const SDLoc &dl, |
6404 | SelectionDAG &DAG) { |
6405 | |
6406 | // Check if the scalar load can be widened into a vector load. And if |
6407 | // the address is "base + cst" see if the cst can be "absorbed" into |
6408 | // the shuffle mask. |
6409 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) { |
6410 | SDValue Ptr = LD->getBasePtr(); |
6411 | if (!ISD::isNormalLoad(LD) || LD->isVolatile()) |
6412 | return SDValue(); |
6413 | EVT PVT = LD->getValueType(0); |
6414 | if (PVT != MVT::i32 && PVT != MVT::f32) |
6415 | return SDValue(); |
6416 | |
6417 | int FI = -1; |
6418 | int64_t Offset = 0; |
6419 | if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) { |
6420 | FI = FINode->getIndex(); |
6421 | Offset = 0; |
6422 | } else if (DAG.isBaseWithConstantOffset(Ptr) && |
6423 | isa<FrameIndexSDNode>(Ptr.getOperand(0))) { |
6424 | FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex(); |
6425 | Offset = Ptr.getConstantOperandVal(1); |
6426 | Ptr = Ptr.getOperand(0); |
6427 | } else { |
6428 | return SDValue(); |
6429 | } |
6430 | |
6431 | // FIXME: 256-bit vector instructions don't require a strict alignment, |
6432 | // improve this code to support it better. |
6433 | unsigned RequiredAlign = VT.getSizeInBits()/8; |
6434 | SDValue Chain = LD->getChain(); |
6435 | // Make sure the stack object alignment is at least 16 or 32. |
6436 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
6437 | if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) { |
6438 | if (MFI.isFixedObjectIndex(FI)) { |
6439 | // Can't change the alignment. FIXME: It's possible to compute |
6440 | // the exact stack offset and reference FI + adjust offset instead. |
6441 | // If someone *really* cares about this. That's the way to implement it. |
6442 | return SDValue(); |
6443 | } else { |
6444 | MFI.setObjectAlignment(FI, RequiredAlign); |
6445 | } |
6446 | } |
6447 | |
6448 | // (Offset % 16 or 32) must be multiple of 4. Then address is then |
6449 | // Ptr + (Offset & ~15). |
6450 | if (Offset < 0) |
6451 | return SDValue(); |
6452 | if ((Offset % RequiredAlign) & 3) |
6453 | return SDValue(); |
6454 | int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1); |
6455 | if (StartOffset) { |
6456 | SDLoc DL(Ptr); |
6457 | Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr, |
6458 | DAG.getConstant(StartOffset, DL, Ptr.getValueType())); |
6459 | } |
6460 | |
6461 | int EltNo = (Offset - StartOffset) >> 2; |
6462 | unsigned NumElems = VT.getVectorNumElements(); |
6463 | |
6464 | EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems); |
6465 | SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr, |
6466 | LD->getPointerInfo().getWithOffset(StartOffset)); |
6467 | |
6468 | SmallVector<int, 8> Mask(NumElems, EltNo); |
6469 | |
6470 | return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), Mask); |
6471 | } |
6472 | |
6473 | return SDValue(); |
6474 | } |
6475 | |
6476 | /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the |
6477 | /// elements can be replaced by a single large load which has the same value as |
6478 | /// a build_vector or insert_subvector whose loaded operands are 'Elts'. |
6479 | /// |
6480 | /// Example: <load i32 *a, load i32 *a+4, zero, undef> -> zextload a |
6481 | static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts, |
6482 | const SDLoc &DL, SelectionDAG &DAG, |
6483 | const X86Subtarget &Subtarget, |
6484 | bool isAfterLegalize) { |
6485 | unsigned NumElems = Elts.size(); |
6486 | |
6487 | int LastLoadedElt = -1; |
6488 | SmallBitVector LoadMask(NumElems, false); |
6489 | SmallBitVector ZeroMask(NumElems, false); |
6490 | SmallBitVector UndefMask(NumElems, false); |
6491 | |
6492 | // For each element in the initializer, see if we've found a load, zero or an |
6493 | // undef. |
6494 | for (unsigned i = 0; i < NumElems; ++i) { |
6495 | SDValue Elt = peekThroughBitcasts(Elts[i]); |
6496 | if (!Elt.getNode()) |
6497 | return SDValue(); |
6498 | |
6499 | if (Elt.isUndef()) |
6500 | UndefMask[i] = true; |
6501 | else if (X86::isZeroNode(Elt) || ISD::isBuildVectorAllZeros(Elt.getNode())) |
6502 | ZeroMask[i] = true; |
6503 | else if (ISD::isNON_EXTLoad(Elt.getNode())) { |
6504 | LoadMask[i] = true; |
6505 | LastLoadedElt = i; |
6506 | // Each loaded element must be the correct fractional portion of the |
6507 | // requested vector load. |
6508 | if ((NumElems * Elt.getValueSizeInBits()) != VT.getSizeInBits()) |
6509 | return SDValue(); |
6510 | } else |
6511 | return SDValue(); |
6512 | } |
6513 | assert((ZeroMask | UndefMask | LoadMask).count() == NumElems &&(((ZeroMask | UndefMask | LoadMask).count() == NumElems && "Incomplete element masks") ? static_cast<void> (0) : __assert_fail ("(ZeroMask | UndefMask | LoadMask).count() == NumElems && \"Incomplete element masks\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6514, __PRETTY_FUNCTION__)) |
6514 | "Incomplete element masks")(((ZeroMask | UndefMask | LoadMask).count() == NumElems && "Incomplete element masks") ? static_cast<void> (0) : __assert_fail ("(ZeroMask | UndefMask | LoadMask).count() == NumElems && \"Incomplete element masks\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6514, __PRETTY_FUNCTION__)); |
6515 | |
6516 | // Handle Special Cases - all undef or undef/zero. |
6517 | if (UndefMask.count() == NumElems) |
6518 | return DAG.getUNDEF(VT); |
6519 | |
6520 | // FIXME: Should we return this as a BUILD_VECTOR instead? |
6521 | if ((ZeroMask | UndefMask).count() == NumElems) |
6522 | return VT.isInteger() ? DAG.getConstant(0, DL, VT) |
6523 | : DAG.getConstantFP(0.0, DL, VT); |
6524 | |
6525 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
6526 | int FirstLoadedElt = LoadMask.find_first(); |
6527 | SDValue EltBase = peekThroughBitcasts(Elts[FirstLoadedElt]); |
6528 | LoadSDNode *LDBase = cast<LoadSDNode>(EltBase); |
6529 | EVT LDBaseVT = EltBase.getValueType(); |
6530 | |
6531 | // Consecutive loads can contain UNDEFS but not ZERO elements. |
6532 | // Consecutive loads with UNDEFs and ZEROs elements require a |
6533 | // an additional shuffle stage to clear the ZERO elements. |
6534 | bool IsConsecutiveLoad = true; |
6535 | bool IsConsecutiveLoadWithZeros = true; |
6536 | for (int i = FirstLoadedElt + 1; i <= LastLoadedElt; ++i) { |
6537 | if (LoadMask[i]) { |
6538 | SDValue Elt = peekThroughBitcasts(Elts[i]); |
6539 | LoadSDNode *LD = cast<LoadSDNode>(Elt); |
6540 | if (!DAG.areNonVolatileConsecutiveLoads( |
6541 | LD, LDBase, Elt.getValueType().getStoreSizeInBits() / 8, |
6542 | i - FirstLoadedElt)) { |
6543 | IsConsecutiveLoad = false; |
6544 | IsConsecutiveLoadWithZeros = false; |
6545 | break; |
6546 | } |
6547 | } else if (ZeroMask[i]) { |
6548 | IsConsecutiveLoad = false; |
6549 | } |
6550 | } |
6551 | |
6552 | auto CreateLoad = [&DAG, &DL](EVT VT, LoadSDNode *LDBase) { |
6553 | auto MMOFlags = LDBase->getMemOperand()->getFlags(); |
6554 | assert(!(MMOFlags & MachineMemOperand::MOVolatile) &&((!(MMOFlags & MachineMemOperand::MOVolatile) && "Cannot merge volatile loads." ) ? static_cast<void> (0) : __assert_fail ("!(MMOFlags & MachineMemOperand::MOVolatile) && \"Cannot merge volatile loads.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6555, __PRETTY_FUNCTION__)) |
6555 | "Cannot merge volatile loads.")((!(MMOFlags & MachineMemOperand::MOVolatile) && "Cannot merge volatile loads." ) ? static_cast<void> (0) : __assert_fail ("!(MMOFlags & MachineMemOperand::MOVolatile) && \"Cannot merge volatile loads.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6555, __PRETTY_FUNCTION__)); |
6556 | SDValue NewLd = |
6557 | DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(), |
6558 | LDBase->getPointerInfo(), LDBase->getAlignment(), MMOFlags); |
6559 | DAG.makeEquivalentMemoryOrdering(LDBase, NewLd); |
6560 | return NewLd; |
6561 | }; |
6562 | |
6563 | // LOAD - all consecutive load/undefs (must start/end with a load). |
6564 | // If we have found an entire vector of loads and undefs, then return a large |
6565 | // load of the entire vector width starting at the base pointer. |
6566 | // If the vector contains zeros, then attempt to shuffle those elements. |
6567 | if (FirstLoadedElt == 0 && LastLoadedElt == (int)(NumElems - 1) && |
6568 | (IsConsecutiveLoad || IsConsecutiveLoadWithZeros)) { |
6569 | assert(LDBase && "Did not find base load for merging consecutive loads")((LDBase && "Did not find base load for merging consecutive loads" ) ? static_cast<void> (0) : __assert_fail ("LDBase && \"Did not find base load for merging consecutive loads\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6569, __PRETTY_FUNCTION__)); |
6570 | EVT EltVT = LDBase->getValueType(0); |
6571 | // Ensure that the input vector size for the merged loads matches the |
6572 | // cumulative size of the input elements. |
6573 | if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems) |
6574 | return SDValue(); |
6575 | |
6576 | if (isAfterLegalize && !TLI.isOperationLegal(ISD::LOAD, VT)) |
6577 | return SDValue(); |
6578 | |
6579 | // Don't create 256-bit non-temporal aligned loads without AVX2 as these |
6580 | // will lower to regular temporal loads and use the cache. |
6581 | if (LDBase->isNonTemporal() && LDBase->getAlignment() >= 32 && |
6582 | VT.is256BitVector() && !Subtarget.hasInt256()) |
6583 | return SDValue(); |
6584 | |
6585 | if (IsConsecutiveLoad) |
6586 | return CreateLoad(VT, LDBase); |
6587 | |
6588 | // IsConsecutiveLoadWithZeros - we need to create a shuffle of the loaded |
6589 | // vector and a zero vector to clear out the zero elements. |
6590 | if (!isAfterLegalize && NumElems == VT.getVectorNumElements()) { |
6591 | SmallVector<int, 4> ClearMask(NumElems, -1); |
6592 | for (unsigned i = 0; i < NumElems; ++i) { |
6593 | if (ZeroMask[i]) |
6594 | ClearMask[i] = i + NumElems; |
6595 | else if (LoadMask[i]) |
6596 | ClearMask[i] = i; |
6597 | } |
6598 | SDValue V = CreateLoad(VT, LDBase); |
6599 | SDValue Z = VT.isInteger() ? DAG.getConstant(0, DL, VT) |
6600 | : DAG.getConstantFP(0.0, DL, VT); |
6601 | return DAG.getVectorShuffle(VT, DL, V, Z, ClearMask); |
6602 | } |
6603 | } |
6604 | |
6605 | int LoadSize = |
6606 | (1 + LastLoadedElt - FirstLoadedElt) * LDBaseVT.getStoreSizeInBits(); |
6607 | |
6608 | // VZEXT_LOAD - consecutive 32/64-bit load/undefs followed by zeros/undefs. |
6609 | if (IsConsecutiveLoad && FirstLoadedElt == 0 && |
6610 | (LoadSize == 32 || LoadSize == 64) && |
6611 | ((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()))) { |
6612 | MVT VecSVT = VT.isFloatingPoint() ? MVT::getFloatingPointVT(LoadSize) |
6613 | : MVT::getIntegerVT(LoadSize); |
6614 | MVT VecVT = MVT::getVectorVT(VecSVT, VT.getSizeInBits() / LoadSize); |
6615 | if (TLI.isTypeLegal(VecVT)) { |
6616 | SDVTList Tys = DAG.getVTList(VecVT, MVT::Other); |
6617 | SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() }; |
6618 | SDValue ResNode = |
6619 | DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, VecSVT, |
6620 | LDBase->getPointerInfo(), |
6621 | LDBase->getAlignment(), |
6622 | false/*isVolatile*/, true/*ReadMem*/, |
6623 | false/*WriteMem*/); |
6624 | DAG.makeEquivalentMemoryOrdering(LDBase, ResNode); |
6625 | return DAG.getBitcast(VT, ResNode); |
6626 | } |
6627 | } |
6628 | |
6629 | return SDValue(); |
6630 | } |
6631 | |
6632 | static Constant *getConstantVector(MVT VT, const APInt &SplatValue, |
6633 | unsigned SplatBitSize, LLVMContext &C) { |
6634 | unsigned ScalarSize = VT.getScalarSizeInBits(); |
6635 | unsigned NumElm = SplatBitSize / ScalarSize; |
6636 | |
6637 | SmallVector<Constant *, 32> ConstantVec; |
6638 | for (unsigned i = 0; i < NumElm; i++) { |
6639 | APInt Val = SplatValue.extractBits(ScalarSize, ScalarSize * i); |
6640 | Constant *Const; |
6641 | if (VT.isFloatingPoint()) { |
6642 | if (ScalarSize == 32) { |
6643 | Const = ConstantFP::get(C, APFloat(APFloat::IEEEsingle(), Val)); |
6644 | } else { |
6645 | assert(ScalarSize == 64 && "Unsupported floating point scalar size")((ScalarSize == 64 && "Unsupported floating point scalar size" ) ? static_cast<void> (0) : __assert_fail ("ScalarSize == 64 && \"Unsupported floating point scalar size\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6645, __PRETTY_FUNCTION__)); |
6646 | Const = ConstantFP::get(C, APFloat(APFloat::IEEEdouble(), Val)); |
6647 | } |
6648 | } else |
6649 | Const = Constant::getIntegerValue(Type::getIntNTy(C, ScalarSize), Val); |
6650 | ConstantVec.push_back(Const); |
6651 | } |
6652 | return ConstantVector::get(ArrayRef<Constant *>(ConstantVec)); |
6653 | } |
6654 | |
6655 | static bool isUseOfShuffle(SDNode *N) { |
6656 | for (auto *U : N->uses()) { |
6657 | if (isTargetShuffle(U->getOpcode())) |
6658 | return true; |
6659 | if (U->getOpcode() == ISD::BITCAST) // Ignore bitcasts |
6660 | return isUseOfShuffle(U); |
6661 | } |
6662 | return false; |
6663 | } |
6664 | |
6665 | /// Attempt to use the vbroadcast instruction to generate a splat value |
6666 | /// from a splat BUILD_VECTOR which uses: |
6667 | /// a. A single scalar load, or a constant. |
6668 | /// b. Repeated pattern of constants (e.g. <0,1,0,1> or <0,1,2,3,0,1,2,3>). |
6669 | /// |
6670 | /// The VBROADCAST node is returned when a pattern is found, |
6671 | /// or SDValue() otherwise. |
6672 | static SDValue lowerBuildVectorAsBroadcast(BuildVectorSDNode *BVOp, |
6673 | const X86Subtarget &Subtarget, |
6674 | SelectionDAG &DAG) { |
6675 | // VBROADCAST requires AVX. |
6676 | // TODO: Splats could be generated for non-AVX CPUs using SSE |
6677 | // instructions, but there's less potential gain for only 128-bit vectors. |
6678 | if (!Subtarget.hasAVX()) |
6679 | return SDValue(); |
6680 | |
6681 | MVT VT = BVOp->getSimpleValueType(0); |
6682 | SDLoc dl(BVOp); |
6683 | |
6684 | assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&(((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector ()) && "Unsupported vector type for broadcast.") ? static_cast <void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) && \"Unsupported vector type for broadcast.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6685, __PRETTY_FUNCTION__)) |
6685 | "Unsupported vector type for broadcast.")(((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector ()) && "Unsupported vector type for broadcast.") ? static_cast <void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) && \"Unsupported vector type for broadcast.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6685, __PRETTY_FUNCTION__)); |
6686 | |
6687 | BitVector UndefElements; |
6688 | SDValue Ld = BVOp->getSplatValue(&UndefElements); |
6689 | |
6690 | // We need a splat of a single value to use broadcast, and it doesn't |
6691 | // make any sense if the value is only in one element of the vector. |
6692 | if (!Ld || (VT.getVectorNumElements() - UndefElements.count()) <= 1) { |
6693 | APInt SplatValue, Undef; |
6694 | unsigned SplatBitSize; |
6695 | bool HasUndef; |
6696 | // Check if this is a repeated constant pattern suitable for broadcasting. |
6697 | if (BVOp->isConstantSplat(SplatValue, Undef, SplatBitSize, HasUndef) && |
6698 | SplatBitSize > VT.getScalarSizeInBits() && |
6699 | SplatBitSize < VT.getSizeInBits()) { |
6700 | // Avoid replacing with broadcast when it's a use of a shuffle |
6701 | // instruction to preserve the present custom lowering of shuffles. |
6702 | if (isUseOfShuffle(BVOp) || BVOp->hasOneUse()) |
6703 | return SDValue(); |
6704 | // replace BUILD_VECTOR with broadcast of the repeated constants. |
6705 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
6706 | LLVMContext *Ctx = DAG.getContext(); |
6707 | MVT PVT = TLI.getPointerTy(DAG.getDataLayout()); |
6708 | if (Subtarget.hasAVX()) { |
6709 | if (SplatBitSize <= 64 && Subtarget.hasAVX2() && |
6710 | !(SplatBitSize == 64 && Subtarget.is32Bit())) { |
6711 | // Splatted value can fit in one INTEGER constant in constant pool. |
6712 | // Load the constant and broadcast it. |
6713 | MVT CVT = MVT::getIntegerVT(SplatBitSize); |
6714 | Type *ScalarTy = Type::getIntNTy(*Ctx, SplatBitSize); |
6715 | Constant *C = Constant::getIntegerValue(ScalarTy, SplatValue); |
6716 | SDValue CP = DAG.getConstantPool(C, PVT); |
6717 | unsigned Repeat = VT.getSizeInBits() / SplatBitSize; |
6718 | |
6719 | unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment(); |
6720 | Ld = DAG.getLoad( |
6721 | CVT, dl, DAG.getEntryNode(), CP, |
6722 | MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), |
6723 | Alignment); |
6724 | SDValue Brdcst = DAG.getNode(X86ISD::VBROADCAST, dl, |
6725 | MVT::getVectorVT(CVT, Repeat), Ld); |
6726 | return DAG.getBitcast(VT, Brdcst); |
6727 | } else if (SplatBitSize == 32 || SplatBitSize == 64) { |
6728 | // Splatted value can fit in one FLOAT constant in constant pool. |
6729 | // Load the constant and broadcast it. |
6730 | // AVX have support for 32 and 64 bit broadcast for floats only. |
6731 | // No 64bit integer in 32bit subtarget. |
6732 | MVT CVT = MVT::getFloatingPointVT(SplatBitSize); |
6733 | // Lower the splat via APFloat directly, to avoid any conversion. |
6734 | Constant *C = |
6735 | SplatBitSize == 32 |
6736 | ? ConstantFP::get(*Ctx, |
6737 | APFloat(APFloat::IEEEsingle(), SplatValue)) |
6738 | : ConstantFP::get(*Ctx, |
6739 | APFloat(APFloat::IEEEdouble(), SplatValue)); |
6740 | SDValue CP = DAG.getConstantPool(C, PVT); |
6741 | unsigned Repeat = VT.getSizeInBits() / SplatBitSize; |
6742 | |
6743 | unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment(); |
6744 | Ld = DAG.getLoad( |
6745 | CVT, dl, DAG.getEntryNode(), CP, |
6746 | MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), |
6747 | Alignment); |
6748 | SDValue Brdcst = DAG.getNode(X86ISD::VBROADCAST, dl, |
6749 | MVT::getVectorVT(CVT, Repeat), Ld); |
6750 | return DAG.getBitcast(VT, Brdcst); |
6751 | } else if (SplatBitSize > 64) { |
6752 | // Load the vector of constants and broadcast it. |
6753 | MVT CVT = VT.getScalarType(); |
6754 | Constant *VecC = getConstantVector(VT, SplatValue, SplatBitSize, |
6755 | *Ctx); |
6756 | SDValue VCP = DAG.getConstantPool(VecC, PVT); |
6757 | unsigned NumElm = SplatBitSize / VT.getScalarSizeInBits(); |
6758 | unsigned Alignment = cast<ConstantPoolSDNode>(VCP)->getAlignment(); |
6759 | Ld = DAG.getLoad( |
6760 | MVT::getVectorVT(CVT, NumElm), dl, DAG.getEntryNode(), VCP, |
6761 | MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), |
6762 | Alignment); |
6763 | SDValue Brdcst = DAG.getNode(X86ISD::SUBV_BROADCAST, dl, VT, Ld); |
6764 | return DAG.getBitcast(VT, Brdcst); |
6765 | } |
6766 | } |
6767 | } |
6768 | return SDValue(); |
6769 | } |
6770 | |
6771 | bool ConstSplatVal = |
6772 | (Ld.getOpcode() == ISD::Constant || Ld.getOpcode() == ISD::ConstantFP); |
6773 | |
6774 | // Make sure that all of the users of a non-constant load are from the |
6775 | // BUILD_VECTOR node. |
6776 | if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode())) |
6777 | return SDValue(); |
6778 | |
6779 | unsigned ScalarSize = Ld.getValueSizeInBits(); |
6780 | bool IsGE256 = (VT.getSizeInBits() >= 256); |
6781 | |
6782 | // When optimizing for size, generate up to 5 extra bytes for a broadcast |
6783 | // instruction to save 8 or more bytes of constant pool data. |
6784 | // TODO: If multiple splats are generated to load the same constant, |
6785 | // it may be detrimental to overall size. There needs to be a way to detect |
6786 | // that condition to know if this is truly a size win. |
6787 | bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize(); |
6788 | |
6789 | // Handle broadcasting a single constant scalar from the constant pool |
6790 | // into a vector. |
6791 | // On Sandybridge (no AVX2), it is still better to load a constant vector |
6792 | // from the constant pool and not to broadcast it from a scalar. |
6793 | // But override that restriction when optimizing for size. |
6794 | // TODO: Check if splatting is recommended for other AVX-capable CPUs. |
6795 | if (ConstSplatVal && (Subtarget.hasAVX2() || OptForSize)) { |
6796 | EVT CVT = Ld.getValueType(); |
6797 | assert(!CVT.isVector() && "Must not broadcast a vector type")((!CVT.isVector() && "Must not broadcast a vector type" ) ? static_cast<void> (0) : __assert_fail ("!CVT.isVector() && \"Must not broadcast a vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6797, __PRETTY_FUNCTION__)); |
6798 | |
6799 | // Splat f32, i32, v4f64, v4i64 in all cases with AVX2. |
6800 | // For size optimization, also splat v2f64 and v2i64, and for size opt |
6801 | // with AVX2, also splat i8 and i16. |
6802 | // With pattern matching, the VBROADCAST node may become a VMOVDDUP. |
6803 | if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) || |
6804 | (OptForSize && (ScalarSize == 64 || Subtarget.hasAVX2()))) { |
6805 | const Constant *C = nullptr; |
6806 | if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld)) |
6807 | C = CI->getConstantIntValue(); |
6808 | else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld)) |
6809 | C = CF->getConstantFPValue(); |
6810 | |
6811 | assert(C && "Invalid constant type")((C && "Invalid constant type") ? static_cast<void > (0) : __assert_fail ("C && \"Invalid constant type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6811, __PRETTY_FUNCTION__)); |
6812 | |
6813 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
6814 | SDValue CP = |
6815 | DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout())); |
6816 | unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment(); |
6817 | Ld = DAG.getLoad( |
6818 | CVT, dl, DAG.getEntryNode(), CP, |
6819 | MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), |
6820 | Alignment); |
6821 | |
6822 | return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld); |
6823 | } |
6824 | } |
6825 | |
6826 | bool IsLoad = ISD::isNormalLoad(Ld.getNode()); |
6827 | |
6828 | // Handle AVX2 in-register broadcasts. |
6829 | if (!IsLoad && Subtarget.hasInt256() && |
6830 | (ScalarSize == 32 || (IsGE256 && ScalarSize == 64))) |
6831 | return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld); |
6832 | |
6833 | // The scalar source must be a normal load. |
6834 | if (!IsLoad) |
6835 | return SDValue(); |
6836 | |
6837 | if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) || |
6838 | (Subtarget.hasVLX() && ScalarSize == 64)) |
6839 | return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld); |
6840 | |
6841 | // The integer check is needed for the 64-bit into 128-bit so it doesn't match |
6842 | // double since there is no vbroadcastsd xmm |
6843 | if (Subtarget.hasInt256() && Ld.getValueType().isInteger()) { |
6844 | if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64) |
6845 | return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld); |
6846 | } |
6847 | |
6848 | // Unsupported broadcast. |
6849 | return SDValue(); |
6850 | } |
6851 | |
6852 | /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real |
6853 | /// underlying vector and index. |
6854 | /// |
6855 | /// Modifies \p ExtractedFromVec to the real vector and returns the real |
6856 | /// index. |
6857 | static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec, |
6858 | SDValue ExtIdx) { |
6859 | int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue(); |
6860 | if (!isa<ShuffleVectorSDNode>(ExtractedFromVec)) |
6861 | return Idx; |
6862 | |
6863 | // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already |
6864 | // lowered this: |
6865 | // (extract_vector_elt (v8f32 %vreg1), Constant<6>) |
6866 | // to: |
6867 | // (extract_vector_elt (vector_shuffle<2,u,u,u> |
6868 | // (extract_subvector (v8f32 %vreg0), Constant<4>), |
6869 | // undef) |
6870 | // Constant<0>) |
6871 | // In this case the vector is the extract_subvector expression and the index |
6872 | // is 2, as specified by the shuffle. |
6873 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec); |
6874 | SDValue ShuffleVec = SVOp->getOperand(0); |
6875 | MVT ShuffleVecVT = ShuffleVec.getSimpleValueType(); |
6876 | assert(ShuffleVecVT.getVectorElementType() ==((ShuffleVecVT.getVectorElementType() == ExtractedFromVec.getSimpleValueType ().getVectorElementType()) ? static_cast<void> (0) : __assert_fail ("ShuffleVecVT.getVectorElementType() == ExtractedFromVec.getSimpleValueType().getVectorElementType()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6877, __PRETTY_FUNCTION__)) |
6877 | ExtractedFromVec.getSimpleValueType().getVectorElementType())((ShuffleVecVT.getVectorElementType() == ExtractedFromVec.getSimpleValueType ().getVectorElementType()) ? static_cast<void> (0) : __assert_fail ("ShuffleVecVT.getVectorElementType() == ExtractedFromVec.getSimpleValueType().getVectorElementType()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6877, __PRETTY_FUNCTION__)); |
6878 | |
6879 | int ShuffleIdx = SVOp->getMaskElt(Idx); |
6880 | if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) { |
6881 | ExtractedFromVec = ShuffleVec; |
6882 | return ShuffleIdx; |
6883 | } |
6884 | return Idx; |
6885 | } |
6886 | |
6887 | static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) { |
6888 | MVT VT = Op.getSimpleValueType(); |
6889 | |
6890 | // Skip if insert_vec_elt is not supported. |
6891 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
6892 | if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT)) |
6893 | return SDValue(); |
6894 | |
6895 | SDLoc DL(Op); |
6896 | unsigned NumElems = Op.getNumOperands(); |
6897 | |
6898 | SDValue VecIn1; |
6899 | SDValue VecIn2; |
6900 | SmallVector<unsigned, 4> InsertIndices; |
6901 | SmallVector<int, 8> Mask(NumElems, -1); |
6902 | |
6903 | for (unsigned i = 0; i != NumElems; ++i) { |
6904 | unsigned Opc = Op.getOperand(i).getOpcode(); |
6905 | |
6906 | if (Opc == ISD::UNDEF) |
6907 | continue; |
6908 | |
6909 | if (Opc != ISD::EXTRACT_VECTOR_ELT) { |
6910 | // Quit if more than 1 elements need inserting. |
6911 | if (InsertIndices.size() > 1) |
6912 | return SDValue(); |
6913 | |
6914 | InsertIndices.push_back(i); |
6915 | continue; |
6916 | } |
6917 | |
6918 | SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0); |
6919 | SDValue ExtIdx = Op.getOperand(i).getOperand(1); |
6920 | |
6921 | // Quit if non-constant index. |
6922 | if (!isa<ConstantSDNode>(ExtIdx)) |
6923 | return SDValue(); |
6924 | int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx); |
6925 | |
6926 | // Quit if extracted from vector of different type. |
6927 | if (ExtractedFromVec.getValueType() != VT) |
6928 | return SDValue(); |
6929 | |
6930 | if (!VecIn1.getNode()) |
6931 | VecIn1 = ExtractedFromVec; |
6932 | else if (VecIn1 != ExtractedFromVec) { |
6933 | if (!VecIn2.getNode()) |
6934 | VecIn2 = ExtractedFromVec; |
6935 | else if (VecIn2 != ExtractedFromVec) |
6936 | // Quit if more than 2 vectors to shuffle |
6937 | return SDValue(); |
6938 | } |
6939 | |
6940 | if (ExtractedFromVec == VecIn1) |
6941 | Mask[i] = Idx; |
6942 | else if (ExtractedFromVec == VecIn2) |
6943 | Mask[i] = Idx + NumElems; |
6944 | } |
6945 | |
6946 | if (!VecIn1.getNode()) |
6947 | return SDValue(); |
6948 | |
6949 | VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT); |
6950 | SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, Mask); |
6951 | |
6952 | for (unsigned Idx : InsertIndices) |
6953 | NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx), |
6954 | DAG.getIntPtrConstant(Idx, DL)); |
6955 | |
6956 | return NV; |
6957 | } |
6958 | |
6959 | static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) { |
6960 | assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&((ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) && Op.getScalarValueSizeInBits() == 1 && "Can not convert non-constant vector" ) ? static_cast<void> (0) : __assert_fail ("ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) && Op.getScalarValueSizeInBits() == 1 && \"Can not convert non-constant vector\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6962, __PRETTY_FUNCTION__)) |
6961 | Op.getScalarValueSizeInBits() == 1 &&((ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) && Op.getScalarValueSizeInBits() == 1 && "Can not convert non-constant vector" ) ? static_cast<void> (0) : __assert_fail ("ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) && Op.getScalarValueSizeInBits() == 1 && \"Can not convert non-constant vector\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6962, __PRETTY_FUNCTION__)) |
6962 | "Can not convert non-constant vector")((ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) && Op.getScalarValueSizeInBits() == 1 && "Can not convert non-constant vector" ) ? static_cast<void> (0) : __assert_fail ("ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) && Op.getScalarValueSizeInBits() == 1 && \"Can not convert non-constant vector\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6962, __PRETTY_FUNCTION__)); |
6963 | uint64_t Immediate = 0; |
6964 | for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) { |
6965 | SDValue In = Op.getOperand(idx); |
6966 | if (!In.isUndef()) |
6967 | Immediate |= (cast<ConstantSDNode>(In)->getZExtValue() & 0x1) << idx; |
6968 | } |
6969 | SDLoc dl(Op); |
6970 | MVT VT = MVT::getIntegerVT(std::max((int)Op.getValueSizeInBits(), 8)); |
6971 | return DAG.getConstant(Immediate, dl, VT); |
6972 | } |
6973 | // Lower BUILD_VECTOR operation for v8i1 and v16i1 types. |
6974 | SDValue |
6975 | X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const { |
6976 | |
6977 | MVT VT = Op.getSimpleValueType(); |
6978 | assert((VT.getVectorElementType() == MVT::i1) &&(((VT.getVectorElementType() == MVT::i1) && "Unexpected type in LowerBUILD_VECTORvXi1!" ) ? static_cast<void> (0) : __assert_fail ("(VT.getVectorElementType() == MVT::i1) && \"Unexpected type in LowerBUILD_VECTORvXi1!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6979, __PRETTY_FUNCTION__)) |
6979 | "Unexpected type in LowerBUILD_VECTORvXi1!")(((VT.getVectorElementType() == MVT::i1) && "Unexpected type in LowerBUILD_VECTORvXi1!" ) ? static_cast<void> (0) : __assert_fail ("(VT.getVectorElementType() == MVT::i1) && \"Unexpected type in LowerBUILD_VECTORvXi1!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 6979, __PRETTY_FUNCTION__)); |
6980 | |
6981 | SDLoc dl(Op); |
6982 | if (ISD::isBuildVectorAllZeros(Op.getNode())) |
6983 | return DAG.getTargetConstant(0, dl, VT); |
6984 | |
6985 | if (ISD::isBuildVectorAllOnes(Op.getNode())) |
6986 | return DAG.getTargetConstant(1, dl, VT); |
6987 | |
6988 | if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) { |
6989 | SDValue Imm = ConvertI1VectorToInteger(Op, DAG); |
6990 | if (Imm.getValueSizeInBits() == VT.getSizeInBits()) |
6991 | return DAG.getBitcast(VT, Imm); |
6992 | SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm); |
6993 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec, |
6994 | DAG.getIntPtrConstant(0, dl)); |
6995 | } |
6996 | |
6997 | // Vector has one or more non-const elements |
6998 | uint64_t Immediate = 0; |
6999 | SmallVector<unsigned, 16> NonConstIdx; |
7000 | bool IsSplat = true; |
7001 | bool HasConstElts = false; |
7002 | int SplatIdx = -1; |
7003 | for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) { |
7004 | SDValue In = Op.getOperand(idx); |
7005 | if (In.isUndef()) |
7006 | continue; |
7007 | if (!isa<ConstantSDNode>(In)) |
7008 | NonConstIdx.push_back(idx); |
7009 | else { |
7010 | Immediate |= (cast<ConstantSDNode>(In)->getZExtValue() & 0x1) << idx; |
7011 | HasConstElts = true; |
7012 | } |
7013 | if (SplatIdx < 0) |
7014 | SplatIdx = idx; |
7015 | else if (In != Op.getOperand(SplatIdx)) |
7016 | IsSplat = false; |
7017 | } |
7018 | |
7019 | // for splat use " (select i1 splat_elt, all-ones, all-zeroes)" |
7020 | if (IsSplat) |
7021 | return DAG.getSelect(dl, VT, Op.getOperand(SplatIdx), |
7022 | DAG.getConstant(1, dl, VT), |
7023 | DAG.getConstant(0, dl, VT)); |
7024 | |
7025 | // insert elements one by one |
7026 | SDValue DstVec; |
7027 | SDValue Imm; |
7028 | if (Immediate) { |
7029 | MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8)); |
7030 | Imm = DAG.getConstant(Immediate, dl, ImmVT); |
7031 | } |
7032 | else if (HasConstElts) |
7033 | Imm = DAG.getConstant(0, dl, VT); |
7034 | else |
7035 | Imm = DAG.getUNDEF(VT); |
7036 | if (Imm.getValueSizeInBits() == VT.getSizeInBits()) |
7037 | DstVec = DAG.getBitcast(VT, Imm); |
7038 | else { |
7039 | SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm); |
7040 | DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec, |
7041 | DAG.getIntPtrConstant(0, dl)); |
7042 | } |
7043 | |
7044 | for (unsigned i = 0, e = NonConstIdx.size(); i != e; ++i) { |
7045 | unsigned InsertIdx = NonConstIdx[i]; |
7046 | DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec, |
7047 | Op.getOperand(InsertIdx), |
7048 | DAG.getIntPtrConstant(InsertIdx, dl)); |
7049 | } |
7050 | return DstVec; |
7051 | } |
7052 | |
7053 | /// \brief Return true if \p N implements a horizontal binop and return the |
7054 | /// operands for the horizontal binop into V0 and V1. |
7055 | /// |
7056 | /// This is a helper function of LowerToHorizontalOp(). |
7057 | /// This function checks that the build_vector \p N in input implements a |
7058 | /// horizontal operation. Parameter \p Opcode defines the kind of horizontal |
7059 | /// operation to match. |
7060 | /// For example, if \p Opcode is equal to ISD::ADD, then this function |
7061 | /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode |
7062 | /// is equal to ISD::SUB, then this function checks if this is a horizontal |
7063 | /// arithmetic sub. |
7064 | /// |
7065 | /// This function only analyzes elements of \p N whose indices are |
7066 | /// in range [BaseIdx, LastIdx). |
7067 | static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode, |
7068 | SelectionDAG &DAG, |
7069 | unsigned BaseIdx, unsigned LastIdx, |
7070 | SDValue &V0, SDValue &V1) { |
7071 | EVT VT = N->getValueType(0); |
7072 | |
7073 | assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!")((BaseIdx * 2 <= LastIdx && "Invalid Indices in input!" ) ? static_cast<void> (0) : __assert_fail ("BaseIdx * 2 <= LastIdx && \"Invalid Indices in input!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7073, __PRETTY_FUNCTION__)); |
7074 | assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&((VT.isVector() && VT.getVectorNumElements() >= LastIdx && "Invalid Vector in input!") ? static_cast<void > (0) : __assert_fail ("VT.isVector() && VT.getVectorNumElements() >= LastIdx && \"Invalid Vector in input!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7075, __PRETTY_FUNCTION__)) |
7075 | "Invalid Vector in input!")((VT.isVector() && VT.getVectorNumElements() >= LastIdx && "Invalid Vector in input!") ? static_cast<void > (0) : __assert_fail ("VT.isVector() && VT.getVectorNumElements() >= LastIdx && \"Invalid Vector in input!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7075, __PRETTY_FUNCTION__)); |
7076 | |
7077 | bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD); |
7078 | bool CanFold = true; |
7079 | unsigned ExpectedVExtractIdx = BaseIdx; |
7080 | unsigned NumElts = LastIdx - BaseIdx; |
7081 | V0 = DAG.getUNDEF(VT); |
7082 | V1 = DAG.getUNDEF(VT); |
7083 | |
7084 | // Check if N implements a horizontal binop. |
7085 | for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) { |
7086 | SDValue Op = N->getOperand(i + BaseIdx); |
7087 | |
7088 | // Skip UNDEFs. |
7089 | if (Op->isUndef()) { |
7090 | // Update the expected vector extract index. |
7091 | if (i * 2 == NumElts) |
7092 | ExpectedVExtractIdx = BaseIdx; |
7093 | ExpectedVExtractIdx += 2; |
7094 | continue; |
7095 | } |
7096 | |
7097 | CanFold = Op->getOpcode() == Opcode && Op->hasOneUse(); |
7098 | |
7099 | if (!CanFold) |
7100 | break; |
7101 | |
7102 | SDValue Op0 = Op.getOperand(0); |
7103 | SDValue Op1 = Op.getOperand(1); |
7104 | |
7105 | // Try to match the following pattern: |
7106 | // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1)) |
7107 | CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT && |
7108 | Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT && |
7109 | Op0.getOperand(0) == Op1.getOperand(0) && |
7110 | isa<ConstantSDNode>(Op0.getOperand(1)) && |
7111 | isa<ConstantSDNode>(Op1.getOperand(1))); |
7112 | if (!CanFold) |
7113 | break; |
7114 | |
7115 | unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue(); |
7116 | unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue(); |
7117 | |
7118 | if (i * 2 < NumElts) { |
7119 | if (V0.isUndef()) { |
7120 | V0 = Op0.getOperand(0); |
7121 | if (V0.getValueType() != VT) |
7122 | return false; |
7123 | } |
7124 | } else { |
7125 | if (V1.isUndef()) { |
7126 | V1 = Op0.getOperand(0); |
7127 | if (V1.getValueType() != VT) |
7128 | return false; |
7129 | } |
7130 | if (i * 2 == NumElts) |
7131 | ExpectedVExtractIdx = BaseIdx; |
7132 | } |
7133 | |
7134 | SDValue Expected = (i * 2 < NumElts) ? V0 : V1; |
7135 | if (I0 == ExpectedVExtractIdx) |
7136 | CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected; |
7137 | else if (IsCommutable && I1 == ExpectedVExtractIdx) { |
7138 | // Try to match the following dag sequence: |
7139 | // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I)) |
7140 | CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected; |
7141 | } else |
7142 | CanFold = false; |
7143 | |
7144 | ExpectedVExtractIdx += 2; |
7145 | } |
7146 | |
7147 | return CanFold; |
7148 | } |
7149 | |
7150 | /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by |
7151 | /// a concat_vector. |
7152 | /// |
7153 | /// This is a helper function of LowerToHorizontalOp(). |
7154 | /// This function expects two 256-bit vectors called V0 and V1. |
7155 | /// At first, each vector is split into two separate 128-bit vectors. |
7156 | /// Then, the resulting 128-bit vectors are used to implement two |
7157 | /// horizontal binary operations. |
7158 | /// |
7159 | /// The kind of horizontal binary operation is defined by \p X86Opcode. |
7160 | /// |
7161 | /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to |
7162 | /// the two new horizontal binop. |
7163 | /// When Mode is set, the first horizontal binop dag node would take as input |
7164 | /// the lower 128-bit of V0 and the upper 128-bit of V0. The second |
7165 | /// horizontal binop dag node would take as input the lower 128-bit of V1 |
7166 | /// and the upper 128-bit of V1. |
7167 | /// Example: |
7168 | /// HADD V0_LO, V0_HI |
7169 | /// HADD V1_LO, V1_HI |
7170 | /// |
7171 | /// Otherwise, the first horizontal binop dag node takes as input the lower |
7172 | /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop |
7173 | /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1. |
7174 | /// Example: |
7175 | /// HADD V0_LO, V1_LO |
7176 | /// HADD V0_HI, V1_HI |
7177 | /// |
7178 | /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower |
7179 | /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to |
7180 | /// the upper 128-bits of the result. |
7181 | static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1, |
7182 | const SDLoc &DL, SelectionDAG &DAG, |
7183 | unsigned X86Opcode, bool Mode, |
7184 | bool isUndefLO, bool isUndefHI) { |
7185 | MVT VT = V0.getSimpleValueType(); |
7186 | assert(VT.is256BitVector() && VT == V1.getSimpleValueType() &&((VT.is256BitVector() && VT == V1.getSimpleValueType( ) && "Invalid nodes in input!") ? static_cast<void > (0) : __assert_fail ("VT.is256BitVector() && VT == V1.getSimpleValueType() && \"Invalid nodes in input!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7187, __PRETTY_FUNCTION__)) |
7187 | "Invalid nodes in input!")((VT.is256BitVector() && VT == V1.getSimpleValueType( ) && "Invalid nodes in input!") ? static_cast<void > (0) : __assert_fail ("VT.is256BitVector() && VT == V1.getSimpleValueType() && \"Invalid nodes in input!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7187, __PRETTY_FUNCTION__)); |
7188 | |
7189 | unsigned NumElts = VT.getVectorNumElements(); |
7190 | SDValue V0_LO = extract128BitVector(V0, 0, DAG, DL); |
7191 | SDValue V0_HI = extract128BitVector(V0, NumElts/2, DAG, DL); |
7192 | SDValue V1_LO = extract128BitVector(V1, 0, DAG, DL); |
7193 | SDValue V1_HI = extract128BitVector(V1, NumElts/2, DAG, DL); |
7194 | MVT NewVT = V0_LO.getSimpleValueType(); |
7195 | |
7196 | SDValue LO = DAG.getUNDEF(NewVT); |
7197 | SDValue HI = DAG.getUNDEF(NewVT); |
7198 | |
7199 | if (Mode) { |
7200 | // Don't emit a horizontal binop if the result is expected to be UNDEF. |
7201 | if (!isUndefLO && !V0->isUndef()) |
7202 | LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI); |
7203 | if (!isUndefHI && !V1->isUndef()) |
7204 | HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI); |
7205 | } else { |
7206 | // Don't emit a horizontal binop if the result is expected to be UNDEF. |
7207 | if (!isUndefLO && (!V0_LO->isUndef() || !V1_LO->isUndef())) |
7208 | LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO); |
7209 | |
7210 | if (!isUndefHI && (!V0_HI->isUndef() || !V1_HI->isUndef())) |
7211 | HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI); |
7212 | } |
7213 | |
7214 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI); |
7215 | } |
7216 | |
7217 | /// Returns true iff \p BV builds a vector with the result equivalent to |
7218 | /// the result of ADDSUB operation. |
7219 | /// If true is returned then the operands of ADDSUB = Opnd0 +- Opnd1 operation |
7220 | /// are written to the parameters \p Opnd0 and \p Opnd1. |
7221 | static bool isAddSub(const BuildVectorSDNode *BV, |
7222 | const X86Subtarget &Subtarget, SelectionDAG &DAG, |
7223 | SDValue &Opnd0, SDValue &Opnd1) { |
7224 | |
7225 | MVT VT = BV->getSimpleValueType(0); |
7226 | if ((!Subtarget.hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) && |
7227 | (!Subtarget.hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)) && |
7228 | (!Subtarget.hasAVX512() || (VT != MVT::v16f32 && VT != MVT::v8f64))) |
7229 | return false; |
7230 | |
7231 | unsigned NumElts = VT.getVectorNumElements(); |
7232 | SDValue InVec0 = DAG.getUNDEF(VT); |
7233 | SDValue InVec1 = DAG.getUNDEF(VT); |
7234 | |
7235 | // Odd-numbered elements in the input build vector are obtained from |
7236 | // adding two integer/float elements. |
7237 | // Even-numbered elements in the input build vector are obtained from |
7238 | // subtracting two integer/float elements. |
7239 | unsigned ExpectedOpcode = ISD::FSUB; |
7240 | unsigned NextExpectedOpcode = ISD::FADD; |
7241 | bool AddFound = false; |
7242 | bool SubFound = false; |
7243 | |
7244 | for (unsigned i = 0, e = NumElts; i != e; ++i) { |
7245 | SDValue Op = BV->getOperand(i); |
7246 | |
7247 | // Skip 'undef' values. |
7248 | unsigned Opcode = Op.getOpcode(); |
7249 | if (Opcode == ISD::UNDEF) { |
7250 | std::swap(ExpectedOpcode, NextExpectedOpcode); |
7251 | continue; |
7252 | } |
7253 | |
7254 | // Early exit if we found an unexpected opcode. |
7255 | if (Opcode != ExpectedOpcode) |
7256 | return false; |
7257 | |
7258 | SDValue Op0 = Op.getOperand(0); |
7259 | SDValue Op1 = Op.getOperand(1); |
7260 | |
7261 | // Try to match the following pattern: |
7262 | // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i)) |
7263 | // Early exit if we cannot match that sequence. |
7264 | if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT || |
7265 | Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT || |
7266 | !isa<ConstantSDNode>(Op0.getOperand(1)) || |
7267 | !isa<ConstantSDNode>(Op1.getOperand(1)) || |
7268 | Op0.getOperand(1) != Op1.getOperand(1)) |
7269 | return false; |
7270 | |
7271 | unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue(); |
7272 | if (I0 != i) |
7273 | return false; |
7274 | |
7275 | // We found a valid add/sub node. Update the information accordingly. |
7276 | if (i & 1) |
7277 | AddFound = true; |
7278 | else |
7279 | SubFound = true; |
7280 | |
7281 | // Update InVec0 and InVec1. |
7282 | if (InVec0.isUndef()) { |
7283 | InVec0 = Op0.getOperand(0); |
7284 | if (InVec0.getSimpleValueType() != VT) |
7285 | return false; |
7286 | } |
7287 | if (InVec1.isUndef()) { |
7288 | InVec1 = Op1.getOperand(0); |
7289 | if (InVec1.getSimpleValueType() != VT) |
7290 | return false; |
7291 | } |
7292 | |
7293 | // Make sure that operands in input to each add/sub node always |
7294 | // come from a same pair of vectors. |
7295 | if (InVec0 != Op0.getOperand(0)) { |
7296 | if (ExpectedOpcode == ISD::FSUB) |
7297 | return false; |
7298 | |
7299 | // FADD is commutable. Try to commute the operands |
7300 | // and then test again. |
7301 | std::swap(Op0, Op1); |
7302 | if (InVec0 != Op0.getOperand(0)) |
7303 | return false; |
7304 | } |
7305 | |
7306 | if (InVec1 != Op1.getOperand(0)) |
7307 | return false; |
7308 | |
7309 | // Update the pair of expected opcodes. |
7310 | std::swap(ExpectedOpcode, NextExpectedOpcode); |
7311 | } |
7312 | |
7313 | // Don't try to fold this build_vector into an ADDSUB if the inputs are undef. |
7314 | if (!AddFound || !SubFound || InVec0.isUndef() || InVec1.isUndef()) |
7315 | return false; |
7316 | |
7317 | Opnd0 = InVec0; |
7318 | Opnd1 = InVec1; |
7319 | return true; |
7320 | } |
7321 | |
7322 | /// Returns true if is possible to fold MUL and an idiom that has already been |
7323 | /// recognized as ADDSUB(\p Opnd0, \p Opnd1) into FMADDSUB(x, y, \p Opnd1). |
7324 | /// If (and only if) true is returned, the operands of FMADDSUB are written to |
7325 | /// parameters \p Opnd0, \p Opnd1, \p Opnd2. |
7326 | /// |
7327 | /// Prior to calling this function it should be known that there is some |
7328 | /// SDNode that potentially can be replaced with an X86ISD::ADDSUB operation |
7329 | /// using \p Opnd0 and \p Opnd1 as operands. Also, this method is called |
7330 | /// before replacement of such SDNode with ADDSUB operation. Thus the number |
7331 | /// of \p Opnd0 uses is expected to be equal to 2. |
7332 | /// For example, this function may be called for the following IR: |
7333 | /// %AB = fmul fast <2 x double> %A, %B |
7334 | /// %Sub = fsub fast <2 x double> %AB, %C |
7335 | /// %Add = fadd fast <2 x double> %AB, %C |
7336 | /// %Addsub = shufflevector <2 x double> %Sub, <2 x double> %Add, |
7337 | /// <2 x i32> <i32 0, i32 3> |
7338 | /// There is a def for %Addsub here, which potentially can be replaced by |
7339 | /// X86ISD::ADDSUB operation: |
7340 | /// %Addsub = X86ISD::ADDSUB %AB, %C |
7341 | /// and such ADDSUB can further be replaced with FMADDSUB: |
7342 | /// %Addsub = FMADDSUB %A, %B, %C. |
7343 | /// |
7344 | /// The main reason why this method is called before the replacement of the |
7345 | /// recognized ADDSUB idiom with ADDSUB operation is that such replacement |
7346 | /// is illegal sometimes. E.g. 512-bit ADDSUB is not available, while 512-bit |
7347 | /// FMADDSUB is. |
7348 | static bool isFMAddSub(const X86Subtarget &Subtarget, SelectionDAG &DAG, |
7349 | SDValue &Opnd0, SDValue &Opnd1, SDValue &Opnd2) { |
7350 | if (Opnd0.getOpcode() != ISD::FMUL || Opnd0->use_size() != 2 || |
7351 | !Subtarget.hasAnyFMA()) |
7352 | return false; |
7353 | |
7354 | // FIXME: These checks must match the similar ones in |
7355 | // DAGCombiner::visitFADDForFMACombine. It would be good to have one |
7356 | // function that would answer if it is Ok to fuse MUL + ADD to FMADD |
7357 | // or MUL + ADDSUB to FMADDSUB. |
7358 | const TargetOptions &Options = DAG.getTarget().Options; |
7359 | bool AllowFusion = |
7360 | (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath); |
7361 | if (!AllowFusion) |
7362 | return false; |
7363 | |
7364 | Opnd2 = Opnd1; |
7365 | Opnd1 = Opnd0.getOperand(1); |
7366 | Opnd0 = Opnd0.getOperand(0); |
7367 | |
7368 | return true; |
7369 | } |
7370 | |
7371 | /// Try to fold a build_vector that performs an 'addsub' or 'fmaddsub' operation |
7372 | /// accordingly to X86ISD::ADDSUB or X86ISD::FMADDSUB node. |
7373 | static SDValue lowerToAddSubOrFMAddSub(const BuildVectorSDNode *BV, |
7374 | const X86Subtarget &Subtarget, |
7375 | SelectionDAG &DAG) { |
7376 | SDValue Opnd0, Opnd1; |
7377 | if (!isAddSub(BV, Subtarget, DAG, Opnd0, Opnd1)) |
7378 | return SDValue(); |
7379 | |
7380 | MVT VT = BV->getSimpleValueType(0); |
7381 | SDLoc DL(BV); |
7382 | |
7383 | // Try to generate X86ISD::FMADDSUB node here. |
7384 | SDValue Opnd2; |
7385 | if (isFMAddSub(Subtarget, DAG, Opnd0, Opnd1, Opnd2)) |
7386 | return DAG.getNode(X86ISD::FMADDSUB, DL, VT, Opnd0, Opnd1, Opnd2); |
7387 | |
7388 | // Do not generate X86ISD::ADDSUB node for 512-bit types even though |
7389 | // the ADDSUB idiom has been successfully recognized. There are no known |
7390 | // X86 targets with 512-bit ADDSUB instructions! |
7391 | // 512-bit ADDSUB idiom recognition was needed only as part of FMADDSUB idiom |
7392 | // recognition. |
7393 | if (VT.is512BitVector()) |
7394 | return SDValue(); |
7395 | |
7396 | return DAG.getNode(X86ISD::ADDSUB, DL, VT, Opnd0, Opnd1); |
7397 | } |
7398 | |
7399 | /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible. |
7400 | static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV, |
7401 | const X86Subtarget &Subtarget, |
7402 | SelectionDAG &DAG) { |
7403 | MVT VT = BV->getSimpleValueType(0); |
7404 | unsigned NumElts = VT.getVectorNumElements(); |
7405 | unsigned NumUndefsLO = 0; |
7406 | unsigned NumUndefsHI = 0; |
7407 | unsigned Half = NumElts/2; |
7408 | |
7409 | // Count the number of UNDEF operands in the build_vector in input. |
7410 | for (unsigned i = 0, e = Half; i != e; ++i) |
7411 | if (BV->getOperand(i)->isUndef()) |
7412 | NumUndefsLO++; |
7413 | |
7414 | for (unsigned i = Half, e = NumElts; i != e; ++i) |
7415 | if (BV->getOperand(i)->isUndef()) |
7416 | NumUndefsHI++; |
7417 | |
7418 | // Early exit if this is either a build_vector of all UNDEFs or all the |
7419 | // operands but one are UNDEF. |
7420 | if (NumUndefsLO + NumUndefsHI + 1 >= NumElts) |
7421 | return SDValue(); |
7422 | |
7423 | SDLoc DL(BV); |
7424 | SDValue InVec0, InVec1; |
7425 | if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget.hasSSE3()) { |
7426 | // Try to match an SSE3 float HADD/HSUB. |
7427 | if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1)) |
7428 | return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1); |
7429 | |
7430 | if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1)) |
7431 | return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1); |
7432 | } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget.hasSSSE3()) { |
7433 | // Try to match an SSSE3 integer HADD/HSUB. |
7434 | if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1)) |
7435 | return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1); |
7436 | |
7437 | if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1)) |
7438 | return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1); |
7439 | } |
7440 | |
7441 | if (!Subtarget.hasAVX()) |
7442 | return SDValue(); |
7443 | |
7444 | if ((VT == MVT::v8f32 || VT == MVT::v4f64)) { |
7445 | // Try to match an AVX horizontal add/sub of packed single/double |
7446 | // precision floating point values from 256-bit vectors. |
7447 | SDValue InVec2, InVec3; |
7448 | if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) && |
7449 | isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) && |
7450 | ((InVec0.isUndef() || InVec2.isUndef()) || InVec0 == InVec2) && |
7451 | ((InVec1.isUndef() || InVec3.isUndef()) || InVec1 == InVec3)) |
7452 | return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1); |
7453 | |
7454 | if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) && |
7455 | isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) && |
7456 | ((InVec0.isUndef() || InVec2.isUndef()) || InVec0 == InVec2) && |
7457 | ((InVec1.isUndef() || InVec3.isUndef()) || InVec1 == InVec3)) |
7458 | return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1); |
7459 | } else if (VT == MVT::v8i32 || VT == MVT::v16i16) { |
7460 | // Try to match an AVX2 horizontal add/sub of signed integers. |
7461 | SDValue InVec2, InVec3; |
7462 | unsigned X86Opcode; |
7463 | bool CanFold = true; |
7464 | |
7465 | if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) && |
7466 | isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) && |
7467 | ((InVec0.isUndef() || InVec2.isUndef()) || InVec0 == InVec2) && |
7468 | ((InVec1.isUndef() || InVec3.isUndef()) || InVec1 == InVec3)) |
7469 | X86Opcode = X86ISD::HADD; |
7470 | else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) && |
7471 | isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) && |
7472 | ((InVec0.isUndef() || InVec2.isUndef()) || InVec0 == InVec2) && |
7473 | ((InVec1.isUndef() || InVec3.isUndef()) || InVec1 == InVec3)) |
7474 | X86Opcode = X86ISD::HSUB; |
7475 | else |
7476 | CanFold = false; |
7477 | |
7478 | if (CanFold) { |
7479 | // Fold this build_vector into a single horizontal add/sub. |
7480 | // Do this only if the target has AVX2. |
7481 | if (Subtarget.hasAVX2()) |
7482 | return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1); |
7483 | |
7484 | // Do not try to expand this build_vector into a pair of horizontal |
7485 | // add/sub if we can emit a pair of scalar add/sub. |
7486 | if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half) |
7487 | return SDValue(); |
7488 | |
7489 | // Convert this build_vector into a pair of horizontal binop followed by |
7490 | // a concat vector. |
7491 | bool isUndefLO = NumUndefsLO == Half; |
7492 | bool isUndefHI = NumUndefsHI == Half; |
7493 | return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false, |
7494 | isUndefLO, isUndefHI); |
7495 | } |
7496 | } |
7497 | |
7498 | if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 || |
7499 | VT == MVT::v16i16) && Subtarget.hasAVX()) { |
7500 | unsigned X86Opcode; |
7501 | if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1)) |
7502 | X86Opcode = X86ISD::HADD; |
7503 | else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1)) |
7504 | X86Opcode = X86ISD::HSUB; |
7505 | else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1)) |
7506 | X86Opcode = X86ISD::FHADD; |
7507 | else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1)) |
7508 | X86Opcode = X86ISD::FHSUB; |
7509 | else |
7510 | return SDValue(); |
7511 | |
7512 | // Don't try to expand this build_vector into a pair of horizontal add/sub |
7513 | // if we can simply emit a pair of scalar add/sub. |
7514 | if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half) |
7515 | return SDValue(); |
7516 | |
7517 | // Convert this build_vector into two horizontal add/sub followed by |
7518 | // a concat vector. |
7519 | bool isUndefLO = NumUndefsLO == Half; |
7520 | bool isUndefHI = NumUndefsHI == Half; |
7521 | return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true, |
7522 | isUndefLO, isUndefHI); |
7523 | } |
7524 | |
7525 | return SDValue(); |
7526 | } |
7527 | |
7528 | /// If a BUILD_VECTOR's source elements all apply the same bit operation and |
7529 | /// one of their operands is constant, lower to a pair of BUILD_VECTOR and |
7530 | /// just apply the bit to the vectors. |
7531 | /// NOTE: Its not in our interest to start make a general purpose vectorizer |
7532 | /// from this, but enough scalar bit operations are created from the later |
7533 | /// legalization + scalarization stages to need basic support. |
7534 | static SDValue lowerBuildVectorToBitOp(BuildVectorSDNode *Op, |
7535 | SelectionDAG &DAG) { |
7536 | SDLoc DL(Op); |
7537 | MVT VT = Op->getSimpleValueType(0); |
7538 | unsigned NumElems = VT.getVectorNumElements(); |
7539 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
7540 | |
7541 | // Check that all elements have the same opcode. |
7542 | // TODO: Should we allow UNDEFS and if so how many? |
7543 | unsigned Opcode = Op->getOperand(0).getOpcode(); |
7544 | for (unsigned i = 1; i < NumElems; ++i) |
7545 | if (Opcode != Op->getOperand(i).getOpcode()) |
7546 | return SDValue(); |
7547 | |
7548 | // TODO: We may be able to add support for other Ops (ADD/SUB + shifts). |
7549 | switch (Opcode) { |
7550 | default: |
7551 | return SDValue(); |
7552 | case ISD::AND: |
7553 | case ISD::XOR: |
7554 | case ISD::OR: |
7555 | if (!TLI.isOperationLegalOrPromote(Opcode, VT)) |
7556 | return SDValue(); |
7557 | break; |
7558 | } |
7559 | |
7560 | SmallVector<SDValue, 4> LHSElts, RHSElts; |
7561 | for (SDValue Elt : Op->ops()) { |
7562 | SDValue LHS = Elt.getOperand(0); |
7563 | SDValue RHS = Elt.getOperand(1); |
7564 | |
7565 | // We expect the canonicalized RHS operand to be the constant. |
7566 | if (!isa<ConstantSDNode>(RHS)) |
7567 | return SDValue(); |
7568 | LHSElts.push_back(LHS); |
7569 | RHSElts.push_back(RHS); |
7570 | } |
7571 | |
7572 | SDValue LHS = DAG.getBuildVector(VT, DL, LHSElts); |
7573 | SDValue RHS = DAG.getBuildVector(VT, DL, RHSElts); |
7574 | return DAG.getNode(Opcode, DL, VT, LHS, RHS); |
7575 | } |
7576 | |
7577 | /// Create a vector constant without a load. SSE/AVX provide the bare minimum |
7578 | /// functionality to do this, so it's all zeros, all ones, or some derivation |
7579 | /// that is cheap to calculate. |
7580 | static SDValue materializeVectorConstant(SDValue Op, SelectionDAG &DAG, |
7581 | const X86Subtarget &Subtarget) { |
7582 | SDLoc DL(Op); |
7583 | MVT VT = Op.getSimpleValueType(); |
7584 | |
7585 | // Vectors containing all zeros can be matched by pxor and xorps. |
7586 | if (ISD::isBuildVectorAllZeros(Op.getNode())) { |
7587 | // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd |
7588 | // and 2) ensure that i64 scalars are eliminated on x86-32 hosts. |
7589 | if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32) |
7590 | return Op; |
7591 | |
7592 | return getZeroVector(VT, Subtarget, DAG, DL); |
7593 | } |
7594 | |
7595 | // Vectors containing all ones can be matched by pcmpeqd on 128-bit width |
7596 | // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use |
7597 | // vpcmpeqd on 256-bit vectors. |
7598 | if (Subtarget.hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) { |
7599 | if (VT == MVT::v4i32 || VT == MVT::v16i32 || |
7600 | (VT == MVT::v8i32 && Subtarget.hasInt256())) |
7601 | return Op; |
7602 | |
7603 | return getOnesVector(VT, DAG, DL); |
7604 | } |
7605 | |
7606 | return SDValue(); |
7607 | } |
7608 | |
7609 | SDValue |
7610 | X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const { |
7611 | SDLoc dl(Op); |
7612 | |
7613 | MVT VT = Op.getSimpleValueType(); |
7614 | MVT ExtVT = VT.getVectorElementType(); |
7615 | unsigned NumElems = Op.getNumOperands(); |
7616 | |
7617 | // Generate vectors for predicate vectors. |
7618 | if (VT.getVectorElementType() == MVT::i1 && Subtarget.hasAVX512()) |
7619 | return LowerBUILD_VECTORvXi1(Op, DAG); |
7620 | |
7621 | if (SDValue VectorConstant = materializeVectorConstant(Op, DAG, Subtarget)) |
7622 | return VectorConstant; |
7623 | |
7624 | BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode()); |
7625 | if (SDValue AddSub = lowerToAddSubOrFMAddSub(BV, Subtarget, DAG)) |
7626 | return AddSub; |
7627 | if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG)) |
7628 | return HorizontalOp; |
7629 | if (SDValue Broadcast = lowerBuildVectorAsBroadcast(BV, Subtarget, DAG)) |
7630 | return Broadcast; |
7631 | if (SDValue BitOp = lowerBuildVectorToBitOp(BV, DAG)) |
7632 | return BitOp; |
7633 | |
7634 | unsigned EVTBits = ExtVT.getSizeInBits(); |
7635 | |
7636 | unsigned NumZero = 0; |
7637 | unsigned NumNonZero = 0; |
7638 | uint64_t NonZeros = 0; |
7639 | bool IsAllConstants = true; |
7640 | SmallSet<SDValue, 8> Values; |
7641 | for (unsigned i = 0; i < NumElems; ++i) { |
7642 | SDValue Elt = Op.getOperand(i); |
7643 | if (Elt.isUndef()) |
7644 | continue; |
7645 | Values.insert(Elt); |
7646 | if (Elt.getOpcode() != ISD::Constant && |
7647 | Elt.getOpcode() != ISD::ConstantFP) |
7648 | IsAllConstants = false; |
7649 | if (X86::isZeroNode(Elt)) |
7650 | NumZero++; |
7651 | else { |
7652 | assert(i < sizeof(NonZeros) * 8)((i < sizeof(NonZeros) * 8) ? static_cast<void> (0) : __assert_fail ("i < sizeof(NonZeros) * 8", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7652, __PRETTY_FUNCTION__)); // Make sure the shift is within range. |
7653 | NonZeros |= ((uint64_t)1 << i); |
7654 | NumNonZero++; |
7655 | } |
7656 | } |
7657 | |
7658 | // All undef vector. Return an UNDEF. All zero vectors were handled above. |
7659 | if (NumNonZero == 0) |
7660 | return DAG.getUNDEF(VT); |
7661 | |
7662 | // Special case for single non-zero, non-undef, element. |
7663 | if (NumNonZero == 1) { |
7664 | unsigned Idx = countTrailingZeros(NonZeros); |
7665 | SDValue Item = Op.getOperand(Idx); |
7666 | |
7667 | // If this is an insertion of an i64 value on x86-32, and if the top bits of |
7668 | // the value are obviously zero, truncate the value to i32 and do the |
7669 | // insertion that way. Only do this if the value is non-constant or if the |
7670 | // value is a constant being inserted into element 0. It is cheaper to do |
7671 | // a constant pool load than it is to do a movd + shuffle. |
7672 | if (ExtVT == MVT::i64 && !Subtarget.is64Bit() && |
7673 | (!IsAllConstants || Idx == 0)) { |
7674 | if (DAG.MaskedValueIsZero(Item, APInt::getHighBitsSet(64, 32))) { |
7675 | // Handle SSE only. |
7676 | assert(VT == MVT::v2i64 && "Expected an SSE value type!")((VT == MVT::v2i64 && "Expected an SSE value type!") ? static_cast<void> (0) : __assert_fail ("VT == MVT::v2i64 && \"Expected an SSE value type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7676, __PRETTY_FUNCTION__)); |
7677 | MVT VecVT = MVT::v4i32; |
7678 | |
7679 | // Truncate the value (which may itself be a constant) to i32, and |
7680 | // convert it to a vector with movd (S2V+shuffle to zero extend). |
7681 | Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item); |
7682 | Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item); |
7683 | return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef( |
7684 | Item, Idx * 2, true, Subtarget, DAG)); |
7685 | } |
7686 | } |
7687 | |
7688 | // If we have a constant or non-constant insertion into the low element of |
7689 | // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into |
7690 | // the rest of the elements. This will be matched as movd/movq/movss/movsd |
7691 | // depending on what the source datatype is. |
7692 | if (Idx == 0) { |
7693 | if (NumZero == 0) |
7694 | return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item); |
7695 | |
7696 | if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 || |
7697 | (ExtVT == MVT::i64 && Subtarget.is64Bit())) { |
7698 | assert((VT.is128BitVector() || VT.is256BitVector() ||(((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector ()) && "Expected an SSE value type!") ? static_cast< void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) && \"Expected an SSE value type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7700, __PRETTY_FUNCTION__)) |
7699 | VT.is512BitVector()) &&(((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector ()) && "Expected an SSE value type!") ? static_cast< void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) && \"Expected an SSE value type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7700, __PRETTY_FUNCTION__)) |
7700 | "Expected an SSE value type!")(((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector ()) && "Expected an SSE value type!") ? static_cast< void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) && \"Expected an SSE value type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7700, __PRETTY_FUNCTION__)); |
7701 | Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item); |
7702 | // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector. |
7703 | return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG); |
7704 | } |
7705 | |
7706 | // We can't directly insert an i8 or i16 into a vector, so zero extend |
7707 | // it to i32 first. |
7708 | if (ExtVT == MVT::i16 || ExtVT == MVT::i8) { |
7709 | Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item); |
7710 | if (VT.getSizeInBits() >= 256) { |
7711 | MVT ShufVT = MVT::getVectorVT(MVT::i32, VT.getSizeInBits()/32); |
7712 | if (Subtarget.hasAVX()) { |
7713 | Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShufVT, Item); |
7714 | Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG); |
7715 | } else { |
7716 | // Without AVX, we need to extend to a 128-bit vector and then |
7717 | // insert into the 256-bit vector. |
7718 | Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item); |
7719 | SDValue ZeroVec = getZeroVector(ShufVT, Subtarget, DAG, dl); |
7720 | Item = insert128BitVector(ZeroVec, Item, 0, DAG, dl); |
7721 | } |
7722 | } else { |
7723 | assert(VT.is128BitVector() && "Expected an SSE value type!")((VT.is128BitVector() && "Expected an SSE value type!" ) ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"Expected an SSE value type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7723, __PRETTY_FUNCTION__)); |
7724 | Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item); |
7725 | Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG); |
7726 | } |
7727 | return DAG.getBitcast(VT, Item); |
7728 | } |
7729 | } |
7730 | |
7731 | // Is it a vector logical left shift? |
7732 | if (NumElems == 2 && Idx == 1 && |
7733 | X86::isZeroNode(Op.getOperand(0)) && |
7734 | !X86::isZeroNode(Op.getOperand(1))) { |
7735 | unsigned NumBits = VT.getSizeInBits(); |
7736 | return getVShift(true, VT, |
7737 | DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, |
7738 | VT, Op.getOperand(1)), |
7739 | NumBits/2, DAG, *this, dl); |
7740 | } |
7741 | |
7742 | if (IsAllConstants) // Otherwise, it's better to do a constpool load. |
7743 | return SDValue(); |
7744 | |
7745 | // Otherwise, if this is a vector with i32 or f32 elements, and the element |
7746 | // is a non-constant being inserted into an element other than the low one, |
7747 | // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka |
7748 | // movd/movss) to move this into the low element, then shuffle it into |
7749 | // place. |
7750 | if (EVTBits == 32) { |
7751 | Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item); |
7752 | return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG); |
7753 | } |
7754 | } |
7755 | |
7756 | // Splat is obviously ok. Let legalizer expand it to a shuffle. |
7757 | if (Values.size() == 1) { |
7758 | if (EVTBits == 32) { |
7759 | // Instead of a shuffle like this: |
7760 | // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0> |
7761 | // Check if it's possible to issue this instead. |
7762 | // shuffle (vload ptr)), undef, <1, 1, 1, 1> |
7763 | unsigned Idx = countTrailingZeros(NonZeros); |
7764 | SDValue Item = Op.getOperand(Idx); |
7765 | if (Op.getNode()->isOnlyUserOf(Item.getNode())) |
7766 | return LowerAsSplatVectorLoad(Item, VT, dl, DAG); |
7767 | } |
7768 | return SDValue(); |
7769 | } |
7770 | |
7771 | // A vector full of immediates; various special cases are already |
7772 | // handled, so this is best done with a single constant-pool load. |
7773 | if (IsAllConstants) |
7774 | return SDValue(); |
7775 | |
7776 | // See if we can use a vector load to get all of the elements. |
7777 | if (VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) { |
7778 | SmallVector<SDValue, 64> Ops(Op->op_begin(), Op->op_begin() + NumElems); |
7779 | if (SDValue LD = |
7780 | EltsFromConsecutiveLoads(VT, Ops, dl, DAG, Subtarget, false)) |
7781 | return LD; |
7782 | } |
7783 | |
7784 | // For AVX-length vectors, build the individual 128-bit pieces and use |
7785 | // shuffles to put them in place. |
7786 | if (VT.is256BitVector() || VT.is512BitVector()) { |
7787 | SmallVector<SDValue, 64> Ops(Op->op_begin(), Op->op_begin() + NumElems); |
7788 | |
7789 | EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2); |
7790 | |
7791 | // Build both the lower and upper subvector. |
7792 | SDValue Lower = |
7793 | DAG.getBuildVector(HVT, dl, makeArrayRef(&Ops[0], NumElems / 2)); |
7794 | SDValue Upper = DAG.getBuildVector( |
7795 | HVT, dl, makeArrayRef(&Ops[NumElems / 2], NumElems / 2)); |
7796 | |
7797 | // Recreate the wider vector with the lower and upper part. |
7798 | if (VT.is256BitVector()) |
7799 | return concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl); |
7800 | return concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl); |
7801 | } |
7802 | |
7803 | // Let legalizer expand 2-wide build_vectors. |
7804 | if (EVTBits == 64) { |
7805 | if (NumNonZero == 1) { |
7806 | // One half is zero or undef. |
7807 | unsigned Idx = countTrailingZeros(NonZeros); |
7808 | SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, |
7809 | Op.getOperand(Idx)); |
7810 | return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG); |
7811 | } |
7812 | return SDValue(); |
7813 | } |
7814 | |
7815 | // If element VT is < 32 bits, convert it to inserts into a zero vector. |
7816 | if (EVTBits == 8 && NumElems == 16) |
7817 | if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros, NumNonZero, NumZero, |
7818 | DAG, Subtarget)) |
7819 | return V; |
7820 | |
7821 | if (EVTBits == 16 && NumElems == 8) |
7822 | if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros, NumNonZero, NumZero, |
7823 | DAG, Subtarget)) |
7824 | return V; |
7825 | |
7826 | // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS |
7827 | if (EVTBits == 32 && NumElems == 4) |
7828 | if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget)) |
7829 | return V; |
7830 | |
7831 | // If element VT is == 32 bits, turn it into a number of shuffles. |
7832 | if (NumElems == 4 && NumZero > 0) { |
7833 | SmallVector<SDValue, 8> Ops(NumElems); |
7834 | for (unsigned i = 0; i < 4; ++i) { |
7835 | bool isZero = !(NonZeros & (1ULL << i)); |
7836 | if (isZero) |
7837 | Ops[i] = getZeroVector(VT, Subtarget, DAG, dl); |
7838 | else |
7839 | Ops[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i)); |
7840 | } |
7841 | |
7842 | for (unsigned i = 0; i < 2; ++i) { |
7843 | switch ((NonZeros & (0x3 << i*2)) >> (i*2)) { |
7844 | default: break; |
7845 | case 0: |
7846 | Ops[i] = Ops[i*2]; // Must be a zero vector. |
7847 | break; |
7848 | case 1: |
7849 | Ops[i] = getMOVL(DAG, dl, VT, Ops[i*2+1], Ops[i*2]); |
7850 | break; |
7851 | case 2: |
7852 | Ops[i] = getMOVL(DAG, dl, VT, Ops[i*2], Ops[i*2+1]); |
7853 | break; |
7854 | case 3: |
7855 | Ops[i] = getUnpackl(DAG, dl, VT, Ops[i*2], Ops[i*2+1]); |
7856 | break; |
7857 | } |
7858 | } |
7859 | |
7860 | bool Reverse1 = (NonZeros & 0x3) == 2; |
7861 | bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2; |
7862 | int MaskVec[] = { |
7863 | Reverse1 ? 1 : 0, |
7864 | Reverse1 ? 0 : 1, |
7865 | static_cast<int>(Reverse2 ? NumElems+1 : NumElems), |
7866 | static_cast<int>(Reverse2 ? NumElems : NumElems+1) |
7867 | }; |
7868 | return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], MaskVec); |
7869 | } |
7870 | |
7871 | if (Values.size() > 1 && VT.is128BitVector()) { |
7872 | // Check for a build vector from mostly shuffle plus few inserting. |
7873 | if (SDValue Sh = buildFromShuffleMostly(Op, DAG)) |
7874 | return Sh; |
7875 | |
7876 | // For SSE 4.1, use insertps to put the high elements into the low element. |
7877 | if (Subtarget.hasSSE41()) { |
7878 | SDValue Result; |
7879 | if (!Op.getOperand(0).isUndef()) |
7880 | Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0)); |
7881 | else |
7882 | Result = DAG.getUNDEF(VT); |
7883 | |
7884 | for (unsigned i = 1; i < NumElems; ++i) { |
7885 | if (Op.getOperand(i).isUndef()) continue; |
7886 | Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result, |
7887 | Op.getOperand(i), DAG.getIntPtrConstant(i, dl)); |
7888 | } |
7889 | return Result; |
7890 | } |
7891 | |
7892 | // Otherwise, expand into a number of unpckl*, start by extending each of |
7893 | // our (non-undef) elements to the full vector width with the element in the |
7894 | // bottom slot of the vector (which generates no code for SSE). |
7895 | SmallVector<SDValue, 8> Ops(NumElems); |
7896 | for (unsigned i = 0; i < NumElems; ++i) { |
7897 | if (!Op.getOperand(i).isUndef()) |
7898 | Ops[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i)); |
7899 | else |
7900 | Ops[i] = DAG.getUNDEF(VT); |
7901 | } |
7902 | |
7903 | // Next, we iteratively mix elements, e.g. for v4f32: |
7904 | // Step 1: unpcklps 0, 1 ==> X: <?, ?, 1, 0> |
7905 | // : unpcklps 2, 3 ==> Y: <?, ?, 3, 2> |
7906 | // Step 2: unpcklpd X, Y ==> <3, 2, 1, 0> |
7907 | for (unsigned Scale = 1; Scale < NumElems; Scale *= 2) { |
7908 | // Generate scaled UNPCKL shuffle mask. |
7909 | SmallVector<int, 16> Mask; |
7910 | for(unsigned i = 0; i != Scale; ++i) |
7911 | Mask.push_back(i); |
7912 | for (unsigned i = 0; i != Scale; ++i) |
7913 | Mask.push_back(NumElems+i); |
7914 | Mask.append(NumElems - Mask.size(), SM_SentinelUndef); |
7915 | |
7916 | for (unsigned i = 0, e = NumElems / (2 * Scale); i != e; ++i) |
7917 | Ops[i] = DAG.getVectorShuffle(VT, dl, Ops[2*i], Ops[(2*i)+1], Mask); |
7918 | } |
7919 | return Ops[0]; |
7920 | } |
7921 | return SDValue(); |
7922 | } |
7923 | |
7924 | // 256-bit AVX can use the vinsertf128 instruction |
7925 | // to create 256-bit vectors from two other 128-bit ones. |
7926 | static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) { |
7927 | SDLoc dl(Op); |
7928 | MVT ResVT = Op.getSimpleValueType(); |
7929 | |
7930 | assert((ResVT.is256BitVector() ||(((ResVT.is256BitVector() || ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide") ? static_cast<void > (0) : __assert_fail ("(ResVT.is256BitVector() || ResVT.is512BitVector()) && \"Value type must be 256-/512-bit wide\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7931, __PRETTY_FUNCTION__)) |
7931 | ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide")(((ResVT.is256BitVector() || ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide") ? static_cast<void > (0) : __assert_fail ("(ResVT.is256BitVector() || ResVT.is512BitVector()) && \"Value type must be 256-/512-bit wide\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7931, __PRETTY_FUNCTION__)); |
7932 | |
7933 | SDValue V1 = Op.getOperand(0); |
7934 | SDValue V2 = Op.getOperand(1); |
7935 | unsigned NumElems = ResVT.getVectorNumElements(); |
7936 | if (ResVT.is256BitVector()) |
7937 | return concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl); |
7938 | |
7939 | if (Op.getNumOperands() == 4) { |
7940 | MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(), |
7941 | ResVT.getVectorNumElements()/2); |
7942 | SDValue V3 = Op.getOperand(2); |
7943 | SDValue V4 = Op.getOperand(3); |
7944 | return concat256BitVectors( |
7945 | concat128BitVectors(V1, V2, HalfVT, NumElems / 2, DAG, dl), |
7946 | concat128BitVectors(V3, V4, HalfVT, NumElems / 2, DAG, dl), ResVT, |
7947 | NumElems, DAG, dl); |
7948 | } |
7949 | return concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl); |
7950 | } |
7951 | |
7952 | // Return true if all the operands of the given CONCAT_VECTORS node are zeros |
7953 | // except for the first one. (CONCAT_VECTORS Op, 0, 0,...,0) |
7954 | static bool isExpandWithZeros(const SDValue &Op) { |
7955 | assert(Op.getOpcode() == ISD::CONCAT_VECTORS &&((Op.getOpcode() == ISD::CONCAT_VECTORS && "Expand with zeros only possible in CONCAT_VECTORS nodes!" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::CONCAT_VECTORS && \"Expand with zeros only possible in CONCAT_VECTORS nodes!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7956, __PRETTY_FUNCTION__)) |
7956 | "Expand with zeros only possible in CONCAT_VECTORS nodes!")((Op.getOpcode() == ISD::CONCAT_VECTORS && "Expand with zeros only possible in CONCAT_VECTORS nodes!" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::CONCAT_VECTORS && \"Expand with zeros only possible in CONCAT_VECTORS nodes!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7956, __PRETTY_FUNCTION__)); |
7957 | |
7958 | for (unsigned i = 1; i < Op.getNumOperands(); i++) |
7959 | if (!ISD::isBuildVectorAllZeros(Op.getOperand(i).getNode())) |
7960 | return false; |
7961 | |
7962 | return true; |
7963 | } |
7964 | |
7965 | // Returns true if the given node is a type promotion (by concatenating i1 |
7966 | // zeros) of the result of a node that already zeros all upper bits of |
7967 | // k-register. |
7968 | static SDValue isTypePromotionOfi1ZeroUpBits(SDValue Op) { |
7969 | unsigned Opc = Op.getOpcode(); |
7970 | |
7971 | assert(Opc == ISD::CONCAT_VECTORS &&((Opc == ISD::CONCAT_VECTORS && Op.getSimpleValueType ().getVectorElementType() == MVT::i1 && "Unexpected node to check for type promotion!" ) ? static_cast<void> (0) : __assert_fail ("Opc == ISD::CONCAT_VECTORS && Op.getSimpleValueType().getVectorElementType() == MVT::i1 && \"Unexpected node to check for type promotion!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7973, __PRETTY_FUNCTION__)) |
7972 | Op.getSimpleValueType().getVectorElementType() == MVT::i1 &&((Opc == ISD::CONCAT_VECTORS && Op.getSimpleValueType ().getVectorElementType() == MVT::i1 && "Unexpected node to check for type promotion!" ) ? static_cast<void> (0) : __assert_fail ("Opc == ISD::CONCAT_VECTORS && Op.getSimpleValueType().getVectorElementType() == MVT::i1 && \"Unexpected node to check for type promotion!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7973, __PRETTY_FUNCTION__)) |
7973 | "Unexpected node to check for type promotion!")((Opc == ISD::CONCAT_VECTORS && Op.getSimpleValueType ().getVectorElementType() == MVT::i1 && "Unexpected node to check for type promotion!" ) ? static_cast<void> (0) : __assert_fail ("Opc == ISD::CONCAT_VECTORS && Op.getSimpleValueType().getVectorElementType() == MVT::i1 && \"Unexpected node to check for type promotion!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 7973, __PRETTY_FUNCTION__)); |
7974 | |
7975 | // As long as we are concatenating zeros to the upper part of a previous node |
7976 | // result, climb up the tree until a node with different opcode is |
7977 | // encountered |
7978 | while (Opc == ISD::INSERT_SUBVECTOR || Opc == ISD::CONCAT_VECTORS) { |
7979 | if (Opc == ISD::INSERT_SUBVECTOR) { |
7980 | if (ISD::isBuildVectorAllZeros(Op.getOperand(0).getNode()) && |
7981 | Op.getConstantOperandVal(2) == 0) |
7982 | Op = Op.getOperand(1); |
7983 | else |
7984 | return SDValue(); |
7985 | } else { // Opc == ISD::CONCAT_VECTORS |
7986 | if (isExpandWithZeros(Op)) |
7987 | Op = Op.getOperand(0); |
7988 | else |
7989 | return SDValue(); |
7990 | } |
7991 | Opc = Op.getOpcode(); |
7992 | } |
7993 | |
7994 | // Check if the first inserted node zeroes the upper bits, or an 'and' result |
7995 | // of a node that zeros the upper bits (its masked version). |
7996 | if (isMaskedZeroUpperBitsvXi1(Op.getOpcode()) || |
7997 | (Op.getOpcode() == ISD::AND && |
7998 | (isMaskedZeroUpperBitsvXi1(Op.getOperand(0).getOpcode()) || |
7999 | isMaskedZeroUpperBitsvXi1(Op.getOperand(1).getOpcode())))) { |
8000 | return Op; |
8001 | } |
8002 | |
8003 | return SDValue(); |
8004 | } |
8005 | |
8006 | static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op, |
8007 | const X86Subtarget &Subtarget, |
8008 | SelectionDAG & DAG) { |
8009 | SDLoc dl(Op); |
8010 | MVT ResVT = Op.getSimpleValueType(); |
8011 | unsigned NumOfOperands = Op.getNumOperands(); |
8012 | |
8013 | assert(isPowerOf2_32(NumOfOperands) &&((isPowerOf2_32(NumOfOperands) && "Unexpected number of operands in CONCAT_VECTORS" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(NumOfOperands) && \"Unexpected number of operands in CONCAT_VECTORS\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8014, __PRETTY_FUNCTION__)) |
8014 | "Unexpected number of operands in CONCAT_VECTORS")((isPowerOf2_32(NumOfOperands) && "Unexpected number of operands in CONCAT_VECTORS" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(NumOfOperands) && \"Unexpected number of operands in CONCAT_VECTORS\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8014, __PRETTY_FUNCTION__)); |
8015 | |
8016 | // If this node promotes - by concatenating zeroes - the type of the result |
8017 | // of a node with instruction that zeroes all upper (irrelevant) bits of the |
8018 | // output register, mark it as legal and catch the pattern in instruction |
8019 | // selection to avoid emitting extra insturctions (for zeroing upper bits). |
8020 | if (SDValue Promoted = isTypePromotionOfi1ZeroUpBits(Op)) { |
8021 | SDValue ZeroC = DAG.getConstant(0, dl, MVT::i64); |
8022 | SDValue AllZeros = DAG.getSplatBuildVector(ResVT, dl, ZeroC); |
8023 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, AllZeros, Promoted, |
8024 | ZeroC); |
8025 | } |
8026 | |
8027 | SDValue Undef = DAG.getUNDEF(ResVT); |
8028 | if (NumOfOperands > 2) { |
8029 | // Specialize the cases when all, or all but one, of the operands are undef. |
8030 | unsigned NumOfDefinedOps = 0; |
8031 | unsigned OpIdx = 0; |
8032 | for (unsigned i = 0; i < NumOfOperands; i++) |
8033 | if (!Op.getOperand(i).isUndef()) { |
8034 | NumOfDefinedOps++; |
8035 | OpIdx = i; |
8036 | } |
8037 | if (NumOfDefinedOps == 0) |
8038 | return Undef; |
8039 | if (NumOfDefinedOps == 1) { |
8040 | unsigned SubVecNumElts = |
8041 | Op.getOperand(OpIdx).getValueType().getVectorNumElements(); |
8042 | SDValue IdxVal = DAG.getIntPtrConstant(SubVecNumElts * OpIdx, dl); |
8043 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, |
8044 | Op.getOperand(OpIdx), IdxVal); |
8045 | } |
8046 | |
8047 | MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(), |
8048 | ResVT.getVectorNumElements()/2); |
8049 | SmallVector<SDValue, 2> Ops; |
8050 | for (unsigned i = 0; i < NumOfOperands/2; i++) |
8051 | Ops.push_back(Op.getOperand(i)); |
8052 | SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops); |
8053 | Ops.clear(); |
8054 | for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++) |
8055 | Ops.push_back(Op.getOperand(i)); |
8056 | SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops); |
8057 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi); |
8058 | } |
8059 | |
8060 | // 2 operands |
8061 | SDValue V1 = Op.getOperand(0); |
8062 | SDValue V2 = Op.getOperand(1); |
8063 | unsigned NumElems = ResVT.getVectorNumElements(); |
8064 | assert(V1.getValueType() == V2.getValueType() &&((V1.getValueType() == V2.getValueType() && V1.getValueType ().getVectorNumElements() == NumElems/2 && "Unexpected operands in CONCAT_VECTORS" ) ? static_cast<void> (0) : __assert_fail ("V1.getValueType() == V2.getValueType() && V1.getValueType().getVectorNumElements() == NumElems/2 && \"Unexpected operands in CONCAT_VECTORS\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8066, __PRETTY_FUNCTION__)) |
8065 | V1.getValueType().getVectorNumElements() == NumElems/2 &&((V1.getValueType() == V2.getValueType() && V1.getValueType ().getVectorNumElements() == NumElems/2 && "Unexpected operands in CONCAT_VECTORS" ) ? static_cast<void> (0) : __assert_fail ("V1.getValueType() == V2.getValueType() && V1.getValueType().getVectorNumElements() == NumElems/2 && \"Unexpected operands in CONCAT_VECTORS\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8066, __PRETTY_FUNCTION__)) |
8066 | "Unexpected operands in CONCAT_VECTORS")((V1.getValueType() == V2.getValueType() && V1.getValueType ().getVectorNumElements() == NumElems/2 && "Unexpected operands in CONCAT_VECTORS" ) ? static_cast<void> (0) : __assert_fail ("V1.getValueType() == V2.getValueType() && V1.getValueType().getVectorNumElements() == NumElems/2 && \"Unexpected operands in CONCAT_VECTORS\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8066, __PRETTY_FUNCTION__)); |
8067 | |
8068 | if (ResVT.getSizeInBits() >= 16) |
8069 | return Op; // The operation is legal with KUNPCK |
8070 | |
8071 | bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode()); |
8072 | bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode()); |
8073 | SDValue ZeroVec = getZeroVector(ResVT, Subtarget, DAG, dl); |
8074 | if (IsZeroV1 && IsZeroV2) |
8075 | return ZeroVec; |
8076 | |
8077 | SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl); |
8078 | if (V2.isUndef()) |
8079 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx); |
8080 | if (IsZeroV2) |
8081 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, ZeroVec, V1, ZeroIdx); |
8082 | |
8083 | SDValue IdxVal = DAG.getIntPtrConstant(NumElems/2, dl); |
8084 | if (V1.isUndef()) |
8085 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, IdxVal); |
8086 | |
8087 | if (IsZeroV1) |
8088 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, ZeroVec, V2, IdxVal); |
8089 | |
8090 | V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx); |
8091 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, V1, V2, IdxVal); |
8092 | } |
8093 | |
8094 | static SDValue LowerCONCAT_VECTORS(SDValue Op, |
8095 | const X86Subtarget &Subtarget, |
8096 | SelectionDAG &DAG) { |
8097 | MVT VT = Op.getSimpleValueType(); |
8098 | if (VT.getVectorElementType() == MVT::i1) |
8099 | return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG); |
8100 | |
8101 | assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||(((VT.is256BitVector() && Op.getNumOperands() == 2) || (VT.is512BitVector() && (Op.getNumOperands() == 2 || Op.getNumOperands() == 4))) ? static_cast<void> (0) : __assert_fail ("(VT.is256BitVector() && Op.getNumOperands() == 2) || (VT.is512BitVector() && (Op.getNumOperands() == 2 || Op.getNumOperands() == 4))" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8103, __PRETTY_FUNCTION__)) |
8102 | (VT.is512BitVector() && (Op.getNumOperands() == 2 ||(((VT.is256BitVector() && Op.getNumOperands() == 2) || (VT.is512BitVector() && (Op.getNumOperands() == 2 || Op.getNumOperands() == 4))) ? static_cast<void> (0) : __assert_fail ("(VT.is256BitVector() && Op.getNumOperands() == 2) || (VT.is512BitVector() && (Op.getNumOperands() == 2 || Op.getNumOperands() == 4))" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8103, __PRETTY_FUNCTION__)) |
8103 | Op.getNumOperands() == 4)))(((VT.is256BitVector() && Op.getNumOperands() == 2) || (VT.is512BitVector() && (Op.getNumOperands() == 2 || Op.getNumOperands() == 4))) ? static_cast<void> (0) : __assert_fail ("(VT.is256BitVector() && Op.getNumOperands() == 2) || (VT.is512BitVector() && (Op.getNumOperands() == 2 || Op.getNumOperands() == 4))" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8103, __PRETTY_FUNCTION__)); |
8104 | |
8105 | // AVX can use the vinsertf128 instruction to create 256-bit vectors |
8106 | // from two other 128-bit ones. |
8107 | |
8108 | // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors |
8109 | return LowerAVXCONCAT_VECTORS(Op, DAG); |
8110 | } |
8111 | |
8112 | //===----------------------------------------------------------------------===// |
8113 | // Vector shuffle lowering |
8114 | // |
8115 | // This is an experimental code path for lowering vector shuffles on x86. It is |
8116 | // designed to handle arbitrary vector shuffles and blends, gracefully |
8117 | // degrading performance as necessary. It works hard to recognize idiomatic |
8118 | // shuffles and lower them to optimal instruction patterns without leaving |
8119 | // a framework that allows reasonably efficient handling of all vector shuffle |
8120 | // patterns. |
8121 | //===----------------------------------------------------------------------===// |
8122 | |
8123 | /// \brief Tiny helper function to identify a no-op mask. |
8124 | /// |
8125 | /// This is a somewhat boring predicate function. It checks whether the mask |
8126 | /// array input, which is assumed to be a single-input shuffle mask of the kind |
8127 | /// used by the X86 shuffle instructions (not a fully general |
8128 | /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an |
8129 | /// in-place shuffle are 'no-op's. |
8130 | static bool isNoopShuffleMask(ArrayRef<int> Mask) { |
8131 | for (int i = 0, Size = Mask.size(); i < Size; ++i) { |
8132 | assert(Mask[i] >= -1 && "Out of bound mask element!")((Mask[i] >= -1 && "Out of bound mask element!") ? static_cast<void> (0) : __assert_fail ("Mask[i] >= -1 && \"Out of bound mask element!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8132, __PRETTY_FUNCTION__)); |
8133 | if (Mask[i] >= 0 && Mask[i] != i) |
8134 | return false; |
8135 | } |
8136 | return true; |
8137 | } |
8138 | |
8139 | /// \brief Test whether there are elements crossing 128-bit lanes in this |
8140 | /// shuffle mask. |
8141 | /// |
8142 | /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations |
8143 | /// and we routinely test for these. |
8144 | static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) { |
8145 | int LaneSize = 128 / VT.getScalarSizeInBits(); |
8146 | int Size = Mask.size(); |
8147 | for (int i = 0; i < Size; ++i) |
8148 | if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize) |
8149 | return true; |
8150 | return false; |
8151 | } |
8152 | |
8153 | /// \brief Test whether a shuffle mask is equivalent within each sub-lane. |
8154 | /// |
8155 | /// This checks a shuffle mask to see if it is performing the same |
8156 | /// lane-relative shuffle in each sub-lane. This trivially implies |
8157 | /// that it is also not lane-crossing. It may however involve a blend from the |
8158 | /// same lane of a second vector. |
8159 | /// |
8160 | /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is |
8161 | /// non-trivial to compute in the face of undef lanes. The representation is |
8162 | /// suitable for use with existing 128-bit shuffles as entries from the second |
8163 | /// vector have been remapped to [LaneSize, 2*LaneSize). |
8164 | static bool isRepeatedShuffleMask(unsigned LaneSizeInBits, MVT VT, |
8165 | ArrayRef<int> Mask, |
8166 | SmallVectorImpl<int> &RepeatedMask) { |
8167 | auto LaneSize = LaneSizeInBits / VT.getScalarSizeInBits(); |
8168 | RepeatedMask.assign(LaneSize, -1); |
8169 | int Size = Mask.size(); |
8170 | for (int i = 0; i < Size; ++i) { |
8171 | assert(Mask[i] == SM_SentinelUndef || Mask[i] >= 0)((Mask[i] == SM_SentinelUndef || Mask[i] >= 0) ? static_cast <void> (0) : __assert_fail ("Mask[i] == SM_SentinelUndef || Mask[i] >= 0" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8171, __PRETTY_FUNCTION__)); |
8172 | if (Mask[i] < 0) |
8173 | continue; |
8174 | if ((Mask[i] % Size) / LaneSize != i / LaneSize) |
8175 | // This entry crosses lanes, so there is no way to model this shuffle. |
8176 | return false; |
8177 | |
8178 | // Ok, handle the in-lane shuffles by detecting if and when they repeat. |
8179 | // Adjust second vector indices to start at LaneSize instead of Size. |
8180 | int LocalM = Mask[i] < Size ? Mask[i] % LaneSize |
8181 | : Mask[i] % LaneSize + LaneSize; |
8182 | if (RepeatedMask[i % LaneSize] < 0) |
8183 | // This is the first non-undef entry in this slot of a 128-bit lane. |
8184 | RepeatedMask[i % LaneSize] = LocalM; |
8185 | else if (RepeatedMask[i % LaneSize] != LocalM) |
8186 | // Found a mismatch with the repeated mask. |
8187 | return false; |
8188 | } |
8189 | return true; |
8190 | } |
8191 | |
8192 | /// Test whether a shuffle mask is equivalent within each 128-bit lane. |
8193 | static bool |
8194 | is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask, |
8195 | SmallVectorImpl<int> &RepeatedMask) { |
8196 | return isRepeatedShuffleMask(128, VT, Mask, RepeatedMask); |
8197 | } |
8198 | |
8199 | /// Test whether a shuffle mask is equivalent within each 256-bit lane. |
8200 | static bool |
8201 | is256BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask, |
8202 | SmallVectorImpl<int> &RepeatedMask) { |
8203 | return isRepeatedShuffleMask(256, VT, Mask, RepeatedMask); |
8204 | } |
8205 | |
8206 | /// Test whether a target shuffle mask is equivalent within each sub-lane. |
8207 | /// Unlike isRepeatedShuffleMask we must respect SM_SentinelZero. |
8208 | static bool isRepeatedTargetShuffleMask(unsigned LaneSizeInBits, MVT VT, |
8209 | ArrayRef<int> Mask, |
8210 | SmallVectorImpl<int> &RepeatedMask) { |
8211 | int LaneSize = LaneSizeInBits / VT.getScalarSizeInBits(); |
8212 | RepeatedMask.assign(LaneSize, SM_SentinelUndef); |
8213 | int Size = Mask.size(); |
8214 | for (int i = 0; i < Size; ++i) { |
8215 | assert(isUndefOrZero(Mask[i]) || (Mask[i] >= 0))((isUndefOrZero(Mask[i]) || (Mask[i] >= 0)) ? static_cast< void> (0) : __assert_fail ("isUndefOrZero(Mask[i]) || (Mask[i] >= 0)" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8215, __PRETTY_FUNCTION__)); |
8216 | if (Mask[i] == SM_SentinelUndef) |
8217 | continue; |
8218 | if (Mask[i] == SM_SentinelZero) { |
8219 | if (!isUndefOrZero(RepeatedMask[i % LaneSize])) |
8220 | return false; |
8221 | RepeatedMask[i % LaneSize] = SM_SentinelZero; |
8222 | continue; |
8223 | } |
8224 | if ((Mask[i] % Size) / LaneSize != i / LaneSize) |
8225 | // This entry crosses lanes, so there is no way to model this shuffle. |
8226 | return false; |
8227 | |
8228 | // Ok, handle the in-lane shuffles by detecting if and when they repeat. |
8229 | // Adjust second vector indices to start at LaneSize instead of Size. |
8230 | int LocalM = |
8231 | Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + LaneSize; |
8232 | if (RepeatedMask[i % LaneSize] == SM_SentinelUndef) |
8233 | // This is the first non-undef entry in this slot of a 128-bit lane. |
8234 | RepeatedMask[i % LaneSize] = LocalM; |
8235 | else if (RepeatedMask[i % LaneSize] != LocalM) |
8236 | // Found a mismatch with the repeated mask. |
8237 | return false; |
8238 | } |
8239 | return true; |
8240 | } |
8241 | |
8242 | /// \brief Checks whether a shuffle mask is equivalent to an explicit list of |
8243 | /// arguments. |
8244 | /// |
8245 | /// This is a fast way to test a shuffle mask against a fixed pattern: |
8246 | /// |
8247 | /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... } |
8248 | /// |
8249 | /// It returns true if the mask is exactly as wide as the argument list, and |
8250 | /// each element of the mask is either -1 (signifying undef) or the value given |
8251 | /// in the argument. |
8252 | static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask, |
8253 | ArrayRef<int> ExpectedMask) { |
8254 | if (Mask.size() != ExpectedMask.size()) |
8255 | return false; |
8256 | |
8257 | int Size = Mask.size(); |
8258 | |
8259 | // If the values are build vectors, we can look through them to find |
8260 | // equivalent inputs that make the shuffles equivalent. |
8261 | auto *BV1 = dyn_cast<BuildVectorSDNode>(V1); |
8262 | auto *BV2 = dyn_cast<BuildVectorSDNode>(V2); |
8263 | |
8264 | for (int i = 0; i < Size; ++i) { |
8265 | assert(Mask[i] >= -1 && "Out of bound mask element!")((Mask[i] >= -1 && "Out of bound mask element!") ? static_cast<void> (0) : __assert_fail ("Mask[i] >= -1 && \"Out of bound mask element!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8265, __PRETTY_FUNCTION__)); |
8266 | if (Mask[i] >= 0 && Mask[i] != ExpectedMask[i]) { |
8267 | auto *MaskBV = Mask[i] < Size ? BV1 : BV2; |
8268 | auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2; |
8269 | if (!MaskBV || !ExpectedBV || |
8270 | MaskBV->getOperand(Mask[i] % Size) != |
8271 | ExpectedBV->getOperand(ExpectedMask[i] % Size)) |
8272 | return false; |
8273 | } |
8274 | } |
8275 | |
8276 | return true; |
8277 | } |
8278 | |
8279 | /// Checks whether a target shuffle mask is equivalent to an explicit pattern. |
8280 | /// |
8281 | /// The masks must be exactly the same width. |
8282 | /// |
8283 | /// If an element in Mask matches SM_SentinelUndef (-1) then the corresponding |
8284 | /// value in ExpectedMask is always accepted. Otherwise the indices must match. |
8285 | /// |
8286 | /// SM_SentinelZero is accepted as a valid negative index but must match in both. |
8287 | static bool isTargetShuffleEquivalent(ArrayRef<int> Mask, |
8288 | ArrayRef<int> ExpectedMask) { |
8289 | int Size = Mask.size(); |
8290 | if (Size != (int)ExpectedMask.size()) |
8291 | return false; |
8292 | |
8293 | for (int i = 0; i < Size; ++i) |
8294 | if (Mask[i] == SM_SentinelUndef) |
8295 | continue; |
8296 | else if (Mask[i] < 0 && Mask[i] != SM_SentinelZero) |
8297 | return false; |
8298 | else if (Mask[i] != ExpectedMask[i]) |
8299 | return false; |
8300 | |
8301 | return true; |
8302 | } |
8303 | |
8304 | // Merges a general DAG shuffle mask and zeroable bit mask into a target shuffle |
8305 | // mask. |
8306 | static SmallVector<int, 64> createTargetShuffleMask(ArrayRef<int> Mask, |
8307 | const APInt &Zeroable) { |
8308 | int NumElts = Mask.size(); |
8309 | assert(NumElts == (int)Zeroable.getBitWidth() && "Mismatch mask sizes")((NumElts == (int)Zeroable.getBitWidth() && "Mismatch mask sizes" ) ? static_cast<void> (0) : __assert_fail ("NumElts == (int)Zeroable.getBitWidth() && \"Mismatch mask sizes\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8309, __PRETTY_FUNCTION__)); |
8310 | |
8311 | SmallVector<int, 64> TargetMask(NumElts, SM_SentinelUndef); |
8312 | for (int i = 0; i != NumElts; ++i) { |
8313 | int M = Mask[i]; |
8314 | if (M == SM_SentinelUndef) |
8315 | continue; |
8316 | assert(0 <= M && M < (2 * NumElts) && "Out of range shuffle index")((0 <= M && M < (2 * NumElts) && "Out of range shuffle index" ) ? static_cast<void> (0) : __assert_fail ("0 <= M && M < (2 * NumElts) && \"Out of range shuffle index\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8316, __PRETTY_FUNCTION__)); |
8317 | TargetMask[i] = (Zeroable[i] ? SM_SentinelZero : M); |
8318 | } |
8319 | return TargetMask; |
8320 | } |
8321 | |
8322 | // Check if the shuffle mask is suitable for the AVX vpunpcklwd or vpunpckhwd |
8323 | // instructions. |
8324 | static bool isUnpackWdShuffleMask(ArrayRef<int> Mask, MVT VT) { |
8325 | if (VT != MVT::v8i32 && VT != MVT::v8f32) |
8326 | return false; |
8327 | |
8328 | SmallVector<int, 8> Unpcklwd; |
8329 | createUnpackShuffleMask(MVT::v8i16, Unpcklwd, /* Lo = */ true, |
8330 | /* Unary = */ false); |
8331 | SmallVector<int, 8> Unpckhwd; |
8332 | createUnpackShuffleMask(MVT::v8i16, Unpckhwd, /* Lo = */ false, |
8333 | /* Unary = */ false); |
8334 | bool IsUnpackwdMask = (isTargetShuffleEquivalent(Mask, Unpcklwd) || |
8335 | isTargetShuffleEquivalent(Mask, Unpckhwd)); |
8336 | return IsUnpackwdMask; |
8337 | } |
8338 | |
8339 | /// \brief Get a 4-lane 8-bit shuffle immediate for a mask. |
8340 | /// |
8341 | /// This helper function produces an 8-bit shuffle immediate corresponding to |
8342 | /// the ubiquitous shuffle encoding scheme used in x86 instructions for |
8343 | /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for |
8344 | /// example. |
8345 | /// |
8346 | /// NB: We rely heavily on "undef" masks preserving the input lane. |
8347 | static unsigned getV4X86ShuffleImm(ArrayRef<int> Mask) { |
8348 | assert(Mask.size() == 4 && "Only 4-lane shuffle masks")((Mask.size() == 4 && "Only 4-lane shuffle masks") ? static_cast <void> (0) : __assert_fail ("Mask.size() == 4 && \"Only 4-lane shuffle masks\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8348, __PRETTY_FUNCTION__)); |
8349 | assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!")((Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!" ) ? static_cast<void> (0) : __assert_fail ("Mask[0] >= -1 && Mask[0] < 4 && \"Out of bound mask element!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8349, __PRETTY_FUNCTION__)); |
8350 | assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!")((Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!" ) ? static_cast<void> (0) : __assert_fail ("Mask[1] >= -1 && Mask[1] < 4 && \"Out of bound mask element!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8350, __PRETTY_FUNCTION__)); |
8351 | assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!")((Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!" ) ? static_cast<void> (0) : __assert_fail ("Mask[2] >= -1 && Mask[2] < 4 && \"Out of bound mask element!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8351, __PRETTY_FUNCTION__)); |
8352 | assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!")((Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!" ) ? static_cast<void> (0) : __assert_fail ("Mask[3] >= -1 && Mask[3] < 4 && \"Out of bound mask element!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8352, __PRETTY_FUNCTION__)); |
8353 | |
8354 | unsigned Imm = 0; |
8355 | Imm |= (Mask[0] < 0 ? 0 : Mask[0]) << 0; |
8356 | Imm |= (Mask[1] < 0 ? 1 : Mask[1]) << 2; |
8357 | Imm |= (Mask[2] < 0 ? 2 : Mask[2]) << 4; |
8358 | Imm |= (Mask[3] < 0 ? 3 : Mask[3]) << 6; |
8359 | return Imm; |
8360 | } |
8361 | |
8362 | static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, const SDLoc &DL, |
8363 | SelectionDAG &DAG) { |
8364 | return DAG.getConstant(getV4X86ShuffleImm(Mask), DL, MVT::i8); |
8365 | } |
8366 | |
8367 | /// \brief Compute whether each element of a shuffle is zeroable. |
8368 | /// |
8369 | /// A "zeroable" vector shuffle element is one which can be lowered to zero. |
8370 | /// Either it is an undef element in the shuffle mask, the element of the input |
8371 | /// referenced is undef, or the element of the input referenced is known to be |
8372 | /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle |
8373 | /// as many lanes with this technique as possible to simplify the remaining |
8374 | /// shuffle. |
8375 | static APInt computeZeroableShuffleElements(ArrayRef<int> Mask, |
8376 | SDValue V1, SDValue V2) { |
8377 | APInt Zeroable(Mask.size(), 0); |
8378 | V1 = peekThroughBitcasts(V1); |
8379 | V2 = peekThroughBitcasts(V2); |
8380 | |
8381 | bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode()); |
8382 | bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode()); |
8383 | |
8384 | int VectorSizeInBits = V1.getValueSizeInBits(); |
8385 | int ScalarSizeInBits = VectorSizeInBits / Mask.size(); |
8386 | assert(!(VectorSizeInBits % ScalarSizeInBits) && "Illegal shuffle mask size")((!(VectorSizeInBits % ScalarSizeInBits) && "Illegal shuffle mask size" ) ? static_cast<void> (0) : __assert_fail ("!(VectorSizeInBits % ScalarSizeInBits) && \"Illegal shuffle mask size\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8386, __PRETTY_FUNCTION__)); |
8387 | |
8388 | for (int i = 0, Size = Mask.size(); i < Size; ++i) { |
8389 | int M = Mask[i]; |
8390 | // Handle the easy cases. |
8391 | if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) { |
8392 | Zeroable.setBit(i); |
8393 | continue; |
8394 | } |
8395 | |
8396 | // Determine shuffle input and normalize the mask. |
8397 | SDValue V = M < Size ? V1 : V2; |
8398 | M %= Size; |
8399 | |
8400 | // Currently we can only search BUILD_VECTOR for UNDEF/ZERO elements. |
8401 | if (V.getOpcode() != ISD::BUILD_VECTOR) |
8402 | continue; |
8403 | |
8404 | // If the BUILD_VECTOR has fewer elements then the bitcasted portion of |
8405 | // the (larger) source element must be UNDEF/ZERO. |
8406 | if ((Size % V.getNumOperands()) == 0) { |
8407 | int Scale = Size / V->getNumOperands(); |
8408 | SDValue Op = V.getOperand(M / Scale); |
8409 | if (Op.isUndef() || X86::isZeroNode(Op)) |
8410 | Zeroable.setBit(i); |
8411 | else if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op)) { |
8412 | APInt Val = Cst->getAPIntValue(); |
8413 | Val.lshrInPlace((M % Scale) * ScalarSizeInBits); |
8414 | Val = Val.getLoBits(ScalarSizeInBits); |
8415 | if (Val == 0) |
8416 | Zeroable.setBit(i); |
8417 | } else if (ConstantFPSDNode *Cst = dyn_cast<ConstantFPSDNode>(Op)) { |
8418 | APInt Val = Cst->getValueAPF().bitcastToAPInt(); |
8419 | Val.lshrInPlace((M % Scale) * ScalarSizeInBits); |
8420 | Val = Val.getLoBits(ScalarSizeInBits); |
8421 | if (Val == 0) |
8422 | Zeroable.setBit(i); |
8423 | } |
8424 | continue; |
8425 | } |
8426 | |
8427 | // If the BUILD_VECTOR has more elements then all the (smaller) source |
8428 | // elements must be UNDEF or ZERO. |
8429 | if ((V.getNumOperands() % Size) == 0) { |
8430 | int Scale = V->getNumOperands() / Size; |
8431 | bool AllZeroable = true; |
8432 | for (int j = 0; j < Scale; ++j) { |
8433 | SDValue Op = V.getOperand((M * Scale) + j); |
8434 | AllZeroable &= (Op.isUndef() || X86::isZeroNode(Op)); |
8435 | } |
8436 | if (AllZeroable) |
8437 | Zeroable.setBit(i); |
8438 | continue; |
8439 | } |
8440 | } |
8441 | |
8442 | return Zeroable; |
8443 | } |
8444 | |
8445 | // The Shuffle result is as follow: |
8446 | // 0*a[0]0*a[1]...0*a[n] , n >=0 where a[] elements in a ascending order. |
8447 | // Each Zeroable's element correspond to a particular Mask's element. |
8448 | // As described in computeZeroableShuffleElements function. |
8449 | // |
8450 | // The function looks for a sub-mask that the nonzero elements are in |
8451 | // increasing order. If such sub-mask exist. The function returns true. |
8452 | static bool isNonZeroElementsInOrder(const APInt &Zeroable, |
8453 | ArrayRef<int> Mask, const EVT &VectorType, |
8454 | bool &IsZeroSideLeft) { |
8455 | int NextElement = -1; |
8456 | // Check if the Mask's nonzero elements are in increasing order. |
8457 | for (int i = 0, e = Mask.size(); i < e; i++) { |
8458 | // Checks if the mask's zeros elements are built from only zeros. |
8459 | assert(Mask[i] >= -1 && "Out of bound mask element!")((Mask[i] >= -1 && "Out of bound mask element!") ? static_cast<void> (0) : __assert_fail ("Mask[i] >= -1 && \"Out of bound mask element!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8459, __PRETTY_FUNCTION__)); |
8460 | if (Mask[i] < 0) |
8461 | return false; |
8462 | if (Zeroable[i]) |
8463 | continue; |
8464 | // Find the lowest non zero element |
8465 | if (NextElement < 0) { |
8466 | NextElement = Mask[i] != 0 ? VectorType.getVectorNumElements() : 0; |
8467 | IsZeroSideLeft = NextElement != 0; |
8468 | } |
8469 | // Exit if the mask's non zero elements are not in increasing order. |
8470 | if (NextElement != Mask[i]) |
8471 | return false; |
8472 | NextElement++; |
8473 | } |
8474 | return true; |
8475 | } |
8476 | |
8477 | /// Try to lower a shuffle with a single PSHUFB of V1 or V2. |
8478 | static SDValue lowerVectorShuffleWithPSHUFB(const SDLoc &DL, MVT VT, |
8479 | ArrayRef<int> Mask, SDValue V1, |
8480 | SDValue V2, |
8481 | const APInt &Zeroable, |
8482 | const X86Subtarget &Subtarget, |
8483 | SelectionDAG &DAG) { |
8484 | int Size = Mask.size(); |
8485 | int LaneSize = 128 / VT.getScalarSizeInBits(); |
8486 | const int NumBytes = VT.getSizeInBits() / 8; |
8487 | const int NumEltBytes = VT.getScalarSizeInBits() / 8; |
8488 | |
8489 | assert((Subtarget.hasSSSE3() && VT.is128BitVector()) ||(((Subtarget.hasSSSE3() && VT.is128BitVector()) || (Subtarget .hasAVX2() && VT.is256BitVector()) || (Subtarget.hasBWI () && VT.is512BitVector())) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasSSSE3() && VT.is128BitVector()) || (Subtarget.hasAVX2() && VT.is256BitVector()) || (Subtarget.hasBWI() && VT.is512BitVector())" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8491, __PRETTY_FUNCTION__)) |
8490 | (Subtarget.hasAVX2() && VT.is256BitVector()) ||(((Subtarget.hasSSSE3() && VT.is128BitVector()) || (Subtarget .hasAVX2() && VT.is256BitVector()) || (Subtarget.hasBWI () && VT.is512BitVector())) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasSSSE3() && VT.is128BitVector()) || (Subtarget.hasAVX2() && VT.is256BitVector()) || (Subtarget.hasBWI() && VT.is512BitVector())" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8491, __PRETTY_FUNCTION__)) |
8491 | (Subtarget.hasBWI() && VT.is512BitVector()))(((Subtarget.hasSSSE3() && VT.is128BitVector()) || (Subtarget .hasAVX2() && VT.is256BitVector()) || (Subtarget.hasBWI () && VT.is512BitVector())) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasSSSE3() && VT.is128BitVector()) || (Subtarget.hasAVX2() && VT.is256BitVector()) || (Subtarget.hasBWI() && VT.is512BitVector())" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8491, __PRETTY_FUNCTION__)); |
8492 | |
8493 | SmallVector<SDValue, 64> PSHUFBMask(NumBytes); |
8494 | // Sign bit set in i8 mask means zero element. |
8495 | SDValue ZeroMask = DAG.getConstant(0x80, DL, MVT::i8); |
8496 | |
8497 | SDValue V; |
8498 | for (int i = 0; i < NumBytes; ++i) { |
8499 | int M = Mask[i / NumEltBytes]; |
8500 | if (M < 0) { |
8501 | PSHUFBMask[i] = DAG.getUNDEF(MVT::i8); |
8502 | continue; |
8503 | } |
8504 | if (Zeroable[i / NumEltBytes]) { |
8505 | PSHUFBMask[i] = ZeroMask; |
8506 | continue; |
8507 | } |
8508 | |
8509 | // We can only use a single input of V1 or V2. |
8510 | SDValue SrcV = (M >= Size ? V2 : V1); |
8511 | if (V && V != SrcV) |
8512 | return SDValue(); |
8513 | V = SrcV; |
8514 | M %= Size; |
8515 | |
8516 | // PSHUFB can't cross lanes, ensure this doesn't happen. |
8517 | if ((M / LaneSize) != ((i / NumEltBytes) / LaneSize)) |
8518 | return SDValue(); |
8519 | |
8520 | M = M % LaneSize; |
8521 | M = M * NumEltBytes + (i % NumEltBytes); |
8522 | PSHUFBMask[i] = DAG.getConstant(M, DL, MVT::i8); |
8523 | } |
8524 | assert(V && "Failed to find a source input")((V && "Failed to find a source input") ? static_cast <void> (0) : __assert_fail ("V && \"Failed to find a source input\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8524, __PRETTY_FUNCTION__)); |
8525 | |
8526 | MVT I8VT = MVT::getVectorVT(MVT::i8, NumBytes); |
8527 | return DAG.getBitcast( |
8528 | VT, DAG.getNode(X86ISD::PSHUFB, DL, I8VT, DAG.getBitcast(I8VT, V), |
8529 | DAG.getBuildVector(I8VT, DL, PSHUFBMask))); |
8530 | } |
8531 | |
8532 | static SDValue getMaskNode(SDValue Mask, MVT MaskVT, |
8533 | const X86Subtarget &Subtarget, SelectionDAG &DAG, |
8534 | const SDLoc &dl); |
8535 | |
8536 | // X86 has dedicated shuffle that can be lowered to VEXPAND |
8537 | static SDValue lowerVectorShuffleToEXPAND(const SDLoc &DL, MVT VT, |
8538 | const APInt &Zeroable, |
8539 | ArrayRef<int> Mask, SDValue &V1, |
8540 | SDValue &V2, SelectionDAG &DAG, |
8541 | const X86Subtarget &Subtarget) { |
8542 | bool IsLeftZeroSide = true; |
8543 | if (!isNonZeroElementsInOrder(Zeroable, Mask, V1.getValueType(), |
8544 | IsLeftZeroSide)) |
8545 | return SDValue(); |
8546 | unsigned VEXPANDMask = (~Zeroable).getZExtValue(); |
8547 | MVT IntegerType = |
8548 | MVT::getIntegerVT(std::max((int)VT.getVectorNumElements(), 8)); |
8549 | SDValue MaskNode = DAG.getConstant(VEXPANDMask, DL, IntegerType); |
8550 | unsigned NumElts = VT.getVectorNumElements(); |
8551 | assert((NumElts == 4 || NumElts == 8 || NumElts == 16) &&(((NumElts == 4 || NumElts == 8 || NumElts == 16) && "Unexpected number of vector elements" ) ? static_cast<void> (0) : __assert_fail ("(NumElts == 4 || NumElts == 8 || NumElts == 16) && \"Unexpected number of vector elements\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8552, __PRETTY_FUNCTION__)) |
8552 | "Unexpected number of vector elements")(((NumElts == 4 || NumElts == 8 || NumElts == 16) && "Unexpected number of vector elements" ) ? static_cast<void> (0) : __assert_fail ("(NumElts == 4 || NumElts == 8 || NumElts == 16) && \"Unexpected number of vector elements\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8552, __PRETTY_FUNCTION__)); |
8553 | SDValue VMask = getMaskNode(MaskNode, MVT::getVectorVT(MVT::i1, NumElts), |
8554 | Subtarget, DAG, DL); |
8555 | SDValue ZeroVector = getZeroVector(VT, Subtarget, DAG, DL); |
8556 | SDValue ExpandedVector = IsLeftZeroSide ? V2 : V1; |
8557 | return DAG.getSelect(DL, VT, VMask, |
8558 | DAG.getNode(X86ISD::EXPAND, DL, VT, ExpandedVector), |
8559 | ZeroVector); |
8560 | } |
8561 | |
8562 | static bool matchVectorShuffleWithUNPCK(MVT VT, SDValue &V1, SDValue &V2, |
8563 | unsigned &UnpackOpcode, bool IsUnary, |
8564 | ArrayRef<int> TargetMask, SDLoc &DL, |
8565 | SelectionDAG &DAG, |
8566 | const X86Subtarget &Subtarget) { |
8567 | int NumElts = VT.getVectorNumElements(); |
8568 | |
8569 | bool Undef1 = true, Undef2 = true, Zero1 = true, Zero2 = true; |
8570 | for (int i = 0; i != NumElts; i += 2) { |
8571 | int M1 = TargetMask[i + 0]; |
8572 | int M2 = TargetMask[i + 1]; |
8573 | Undef1 &= (SM_SentinelUndef == M1); |
8574 | Undef2 &= (SM_SentinelUndef == M2); |
8575 | Zero1 &= isUndefOrZero(M1); |
8576 | Zero2 &= isUndefOrZero(M2); |
8577 | } |
8578 | assert(!((Undef1 || Zero1) && (Undef2 || Zero2)) &&((!((Undef1 || Zero1) && (Undef2 || Zero2)) && "Zeroable shuffle detected") ? static_cast<void> (0) : __assert_fail ("!((Undef1 || Zero1) && (Undef2 || Zero2)) && \"Zeroable shuffle detected\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8579, __PRETTY_FUNCTION__)) |
8579 | "Zeroable shuffle detected")((!((Undef1 || Zero1) && (Undef2 || Zero2)) && "Zeroable shuffle detected") ? static_cast<void> (0) : __assert_fail ("!((Undef1 || Zero1) && (Undef2 || Zero2)) && \"Zeroable shuffle detected\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8579, __PRETTY_FUNCTION__)); |
8580 | |
8581 | // Attempt to match the target mask against the unpack lo/hi mask patterns. |
8582 | SmallVector<int, 64> Unpckl, Unpckh; |
8583 | createUnpackShuffleMask(VT, Unpckl, /* Lo = */ true, IsUnary); |
8584 | if (isTargetShuffleEquivalent(TargetMask, Unpckl)) { |
8585 | UnpackOpcode = X86ISD::UNPCKL; |
8586 | V2 = (Undef2 ? DAG.getUNDEF(VT) : (IsUnary ? V1 : V2)); |
8587 | V1 = (Undef1 ? DAG.getUNDEF(VT) : V1); |
8588 | return true; |
8589 | } |
8590 | |
8591 | createUnpackShuffleMask(VT, Unpckh, /* Lo = */ false, IsUnary); |
8592 | if (isTargetShuffleEquivalent(TargetMask, Unpckh)) { |
8593 | UnpackOpcode = X86ISD::UNPCKH; |
8594 | V2 = (Undef2 ? DAG.getUNDEF(VT) : (IsUnary ? V1 : V2)); |
8595 | V1 = (Undef1 ? DAG.getUNDEF(VT) : V1); |
8596 | return true; |
8597 | } |
8598 | |
8599 | // If an unary shuffle, attempt to match as an unpack lo/hi with zero. |
8600 | if (IsUnary && (Zero1 || Zero2)) { |
8601 | // Don't bother if we can blend instead. |
8602 | if ((Subtarget.hasSSE41() || VT == MVT::v2i64 || VT == MVT::v2f64) && |
8603 | isSequentialOrUndefOrZeroInRange(TargetMask, 0, NumElts, 0)) |
8604 | return false; |
8605 | |
8606 | bool MatchLo = true, MatchHi = true; |
8607 | for (int i = 0; (i != NumElts) && (MatchLo || MatchHi); ++i) { |
8608 | int M = TargetMask[i]; |
8609 | |
8610 | // Ignore if the input is known to be zero or the index is undef. |
8611 | if ((((i & 1) == 0) && Zero1) || (((i & 1) == 1) && Zero2) || |
8612 | (M == SM_SentinelUndef)) |
8613 | continue; |
8614 | |
8615 | MatchLo &= (M == Unpckl[i]); |
8616 | MatchHi &= (M == Unpckh[i]); |
8617 | } |
8618 | |
8619 | if (MatchLo || MatchHi) { |
8620 | UnpackOpcode = MatchLo ? X86ISD::UNPCKL : X86ISD::UNPCKH; |
8621 | V2 = Zero2 ? getZeroVector(VT, Subtarget, DAG, DL) : V1; |
8622 | V1 = Zero1 ? getZeroVector(VT, Subtarget, DAG, DL) : V1; |
8623 | return true; |
8624 | } |
8625 | } |
8626 | |
8627 | // If a binary shuffle, commute and try again. |
8628 | if (!IsUnary) { |
8629 | ShuffleVectorSDNode::commuteMask(Unpckl); |
8630 | if (isTargetShuffleEquivalent(TargetMask, Unpckl)) { |
8631 | UnpackOpcode = X86ISD::UNPCKL; |
8632 | std::swap(V1, V2); |
8633 | return true; |
8634 | } |
8635 | |
8636 | ShuffleVectorSDNode::commuteMask(Unpckh); |
8637 | if (isTargetShuffleEquivalent(TargetMask, Unpckh)) { |
8638 | UnpackOpcode = X86ISD::UNPCKH; |
8639 | std::swap(V1, V2); |
8640 | return true; |
8641 | } |
8642 | } |
8643 | |
8644 | return false; |
8645 | } |
8646 | |
8647 | // X86 has dedicated unpack instructions that can handle specific blend |
8648 | // operations: UNPCKH and UNPCKL. |
8649 | static SDValue lowerVectorShuffleWithUNPCK(const SDLoc &DL, MVT VT, |
8650 | ArrayRef<int> Mask, SDValue V1, |
8651 | SDValue V2, SelectionDAG &DAG) { |
8652 | SmallVector<int, 8> Unpckl; |
8653 | createUnpackShuffleMask(VT, Unpckl, /* Lo = */ true, /* Unary = */ false); |
8654 | if (isShuffleEquivalent(V1, V2, Mask, Unpckl)) |
8655 | return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2); |
8656 | |
8657 | SmallVector<int, 8> Unpckh; |
8658 | createUnpackShuffleMask(VT, Unpckh, /* Lo = */ false, /* Unary = */ false); |
8659 | if (isShuffleEquivalent(V1, V2, Mask, Unpckh)) |
8660 | return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2); |
8661 | |
8662 | // Commute and try again. |
8663 | ShuffleVectorSDNode::commuteMask(Unpckl); |
8664 | if (isShuffleEquivalent(V1, V2, Mask, Unpckl)) |
8665 | return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1); |
8666 | |
8667 | ShuffleVectorSDNode::commuteMask(Unpckh); |
8668 | if (isShuffleEquivalent(V1, V2, Mask, Unpckh)) |
8669 | return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1); |
8670 | |
8671 | return SDValue(); |
8672 | } |
8673 | |
8674 | /// \brief Try to emit a bitmask instruction for a shuffle. |
8675 | /// |
8676 | /// This handles cases where we can model a blend exactly as a bitmask due to |
8677 | /// one of the inputs being zeroable. |
8678 | static SDValue lowerVectorShuffleAsBitMask(const SDLoc &DL, MVT VT, SDValue V1, |
8679 | SDValue V2, ArrayRef<int> Mask, |
8680 | const APInt &Zeroable, |
8681 | SelectionDAG &DAG) { |
8682 | assert(!VT.isFloatingPoint() && "Floating point types are not supported")((!VT.isFloatingPoint() && "Floating point types are not supported" ) ? static_cast<void> (0) : __assert_fail ("!VT.isFloatingPoint() && \"Floating point types are not supported\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8682, __PRETTY_FUNCTION__)); |
8683 | MVT EltVT = VT.getVectorElementType(); |
8684 | SDValue Zero = DAG.getConstant(0, DL, EltVT); |
8685 | SDValue AllOnes = DAG.getAllOnesConstant(DL, EltVT); |
8686 | SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero); |
8687 | SDValue V; |
8688 | for (int i = 0, Size = Mask.size(); i < Size; ++i) { |
8689 | if (Zeroable[i]) |
8690 | continue; |
8691 | if (Mask[i] % Size != i) |
8692 | return SDValue(); // Not a blend. |
8693 | if (!V) |
8694 | V = Mask[i] < Size ? V1 : V2; |
8695 | else if (V != (Mask[i] < Size ? V1 : V2)) |
8696 | return SDValue(); // Can only let one input through the mask. |
8697 | |
8698 | VMaskOps[i] = AllOnes; |
8699 | } |
8700 | if (!V) |
8701 | return SDValue(); // No non-zeroable elements! |
8702 | |
8703 | SDValue VMask = DAG.getBuildVector(VT, DL, VMaskOps); |
8704 | return DAG.getNode(ISD::AND, DL, VT, V, VMask); |
8705 | } |
8706 | |
8707 | /// \brief Try to emit a blend instruction for a shuffle using bit math. |
8708 | /// |
8709 | /// This is used as a fallback approach when first class blend instructions are |
8710 | /// unavailable. Currently it is only suitable for integer vectors, but could |
8711 | /// be generalized for floating point vectors if desirable. |
8712 | static SDValue lowerVectorShuffleAsBitBlend(const SDLoc &DL, MVT VT, SDValue V1, |
8713 | SDValue V2, ArrayRef<int> Mask, |
8714 | SelectionDAG &DAG) { |
8715 | assert(VT.isInteger() && "Only supports integer vector types!")((VT.isInteger() && "Only supports integer vector types!" ) ? static_cast<void> (0) : __assert_fail ("VT.isInteger() && \"Only supports integer vector types!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8715, __PRETTY_FUNCTION__)); |
8716 | MVT EltVT = VT.getVectorElementType(); |
8717 | SDValue Zero = DAG.getConstant(0, DL, EltVT); |
8718 | SDValue AllOnes = DAG.getAllOnesConstant(DL, EltVT); |
8719 | SmallVector<SDValue, 16> MaskOps; |
8720 | for (int i = 0, Size = Mask.size(); i < Size; ++i) { |
8721 | if (Mask[i] >= 0 && Mask[i] != i && Mask[i] != i + Size) |
8722 | return SDValue(); // Shuffled input! |
8723 | MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero); |
8724 | } |
8725 | |
8726 | SDValue V1Mask = DAG.getBuildVector(VT, DL, MaskOps); |
8727 | V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask); |
8728 | // We have to cast V2 around. |
8729 | MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64); |
8730 | V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT, |
8731 | DAG.getBitcast(MaskVT, V1Mask), |
8732 | DAG.getBitcast(MaskVT, V2))); |
8733 | return DAG.getNode(ISD::OR, DL, VT, V1, V2); |
8734 | } |
8735 | |
8736 | static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask, |
8737 | SDValue PreservedSrc, |
8738 | const X86Subtarget &Subtarget, |
8739 | SelectionDAG &DAG); |
8740 | |
8741 | static bool matchVectorShuffleAsBlend(SDValue V1, SDValue V2, |
8742 | MutableArrayRef<int> TargetMask, |
8743 | bool &ForceV1Zero, bool &ForceV2Zero, |
8744 | uint64_t &BlendMask) { |
8745 | bool V1IsZeroOrUndef = |
8746 | V1.isUndef() || ISD::isBuildVectorAllZeros(V1.getNode()); |
8747 | bool V2IsZeroOrUndef = |
8748 | V2.isUndef() || ISD::isBuildVectorAllZeros(V2.getNode()); |
8749 | |
8750 | BlendMask = 0; |
8751 | ForceV1Zero = false, ForceV2Zero = false; |
8752 | assert(TargetMask.size() <= 64 && "Shuffle mask too big for blend mask")((TargetMask.size() <= 64 && "Shuffle mask too big for blend mask" ) ? static_cast<void> (0) : __assert_fail ("TargetMask.size() <= 64 && \"Shuffle mask too big for blend mask\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8752, __PRETTY_FUNCTION__)); |
8753 | |
8754 | // Attempt to generate the binary blend mask. If an input is zero then |
8755 | // we can use any lane. |
8756 | // TODO: generalize the zero matching to any scalar like isShuffleEquivalent. |
8757 | for (int i = 0, Size = TargetMask.size(); i < Size; ++i) { |
8758 | int M = TargetMask[i]; |
8759 | if (M == SM_SentinelUndef) |
8760 | continue; |
8761 | if (M == i) |
8762 | continue; |
8763 | if (M == i + Size) { |
8764 | BlendMask |= 1ull << i; |
8765 | continue; |
8766 | } |
8767 | if (M == SM_SentinelZero) { |
8768 | if (V1IsZeroOrUndef) { |
8769 | ForceV1Zero = true; |
8770 | TargetMask[i] = i; |
8771 | continue; |
8772 | } |
8773 | if (V2IsZeroOrUndef) { |
8774 | ForceV2Zero = true; |
8775 | BlendMask |= 1ull << i; |
8776 | TargetMask[i] = i + Size; |
8777 | continue; |
8778 | } |
8779 | } |
8780 | return false; |
8781 | } |
8782 | return true; |
8783 | } |
8784 | |
8785 | uint64_t scaleVectorShuffleBlendMask(uint64_t BlendMask, int Size, int Scale) { |
8786 | uint64_t ScaledMask = 0; |
8787 | for (int i = 0; i != Size; ++i) |
8788 | if (BlendMask & (1ull << i)) |
8789 | ScaledMask |= ((1ull << Scale) - 1) << (i * Scale); |
8790 | return ScaledMask; |
8791 | } |
8792 | |
8793 | /// \brief Try to emit a blend instruction for a shuffle. |
8794 | /// |
8795 | /// This doesn't do any checks for the availability of instructions for blending |
8796 | /// these values. It relies on the availability of the X86ISD::BLENDI pattern to |
8797 | /// be matched in the backend with the type given. What it does check for is |
8798 | /// that the shuffle mask is a blend, or convertible into a blend with zero. |
8799 | static SDValue lowerVectorShuffleAsBlend(const SDLoc &DL, MVT VT, SDValue V1, |
8800 | SDValue V2, ArrayRef<int> Original, |
8801 | const APInt &Zeroable, |
8802 | const X86Subtarget &Subtarget, |
8803 | SelectionDAG &DAG) { |
8804 | SmallVector<int, 64> Mask = createTargetShuffleMask(Original, Zeroable); |
8805 | |
8806 | uint64_t BlendMask = 0; |
8807 | bool ForceV1Zero = false, ForceV2Zero = false; |
8808 | if (!matchVectorShuffleAsBlend(V1, V2, Mask, ForceV1Zero, ForceV2Zero, |
8809 | BlendMask)) |
8810 | return SDValue(); |
8811 | |
8812 | // Create a REAL zero vector - ISD::isBuildVectorAllZeros allows UNDEFs. |
8813 | if (ForceV1Zero) |
8814 | V1 = getZeroVector(VT, Subtarget, DAG, DL); |
8815 | if (ForceV2Zero) |
8816 | V2 = getZeroVector(VT, Subtarget, DAG, DL); |
8817 | |
8818 | switch (VT.SimpleTy) { |
8819 | case MVT::v2f64: |
8820 | case MVT::v4f32: |
8821 | case MVT::v4f64: |
8822 | case MVT::v8f32: |
8823 | return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2, |
8824 | DAG.getConstant(BlendMask, DL, MVT::i8)); |
8825 | |
8826 | case MVT::v4i64: |
8827 | case MVT::v8i32: |
8828 | assert(Subtarget.hasAVX2() && "256-bit integer blends require AVX2!")((Subtarget.hasAVX2() && "256-bit integer blends require AVX2!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX2() && \"256-bit integer blends require AVX2!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8828, __PRETTY_FUNCTION__)); |
8829 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
8830 | case MVT::v2i64: |
8831 | case MVT::v4i32: |
8832 | // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into |
8833 | // that instruction. |
8834 | if (Subtarget.hasAVX2()) { |
8835 | // Scale the blend by the number of 32-bit dwords per element. |
8836 | int Scale = VT.getScalarSizeInBits() / 32; |
8837 | BlendMask = scaleVectorShuffleBlendMask(BlendMask, Mask.size(), Scale); |
8838 | MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32; |
8839 | V1 = DAG.getBitcast(BlendVT, V1); |
8840 | V2 = DAG.getBitcast(BlendVT, V2); |
8841 | return DAG.getBitcast( |
8842 | VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2, |
8843 | DAG.getConstant(BlendMask, DL, MVT::i8))); |
8844 | } |
8845 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
8846 | case MVT::v8i16: { |
8847 | // For integer shuffles we need to expand the mask and cast the inputs to |
8848 | // v8i16s prior to blending. |
8849 | int Scale = 8 / VT.getVectorNumElements(); |
8850 | BlendMask = scaleVectorShuffleBlendMask(BlendMask, Mask.size(), Scale); |
8851 | V1 = DAG.getBitcast(MVT::v8i16, V1); |
8852 | V2 = DAG.getBitcast(MVT::v8i16, V2); |
8853 | return DAG.getBitcast(VT, |
8854 | DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2, |
8855 | DAG.getConstant(BlendMask, DL, MVT::i8))); |
8856 | } |
8857 | |
8858 | case MVT::v16i16: { |
8859 | assert(Subtarget.hasAVX2() && "256-bit integer blends require AVX2!")((Subtarget.hasAVX2() && "256-bit integer blends require AVX2!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX2() && \"256-bit integer blends require AVX2!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8859, __PRETTY_FUNCTION__)); |
8860 | SmallVector<int, 8> RepeatedMask; |
8861 | if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) { |
8862 | // We can lower these with PBLENDW which is mirrored across 128-bit lanes. |
8863 | assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!")((RepeatedMask.size() == 8 && "Repeated mask size doesn't match!" ) ? static_cast<void> (0) : __assert_fail ("RepeatedMask.size() == 8 && \"Repeated mask size doesn't match!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8863, __PRETTY_FUNCTION__)); |
8864 | BlendMask = 0; |
8865 | for (int i = 0; i < 8; ++i) |
8866 | if (RepeatedMask[i] >= 8) |
8867 | BlendMask |= 1ull << i; |
8868 | return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2, |
8869 | DAG.getConstant(BlendMask, DL, MVT::i8)); |
8870 | } |
8871 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
8872 | } |
8873 | case MVT::v16i8: |
8874 | case MVT::v32i8: { |
8875 | assert((VT.is128BitVector() || Subtarget.hasAVX2()) &&(((VT.is128BitVector() || Subtarget.hasAVX2()) && "256-bit byte-blends require AVX2 support!" ) ? static_cast<void> (0) : __assert_fail ("(VT.is128BitVector() || Subtarget.hasAVX2()) && \"256-bit byte-blends require AVX2 support!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8876, __PRETTY_FUNCTION__)) |
8876 | "256-bit byte-blends require AVX2 support!")(((VT.is128BitVector() || Subtarget.hasAVX2()) && "256-bit byte-blends require AVX2 support!" ) ? static_cast<void> (0) : __assert_fail ("(VT.is128BitVector() || Subtarget.hasAVX2()) && \"256-bit byte-blends require AVX2 support!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8876, __PRETTY_FUNCTION__)); |
8877 | |
8878 | if (Subtarget.hasBWI() && Subtarget.hasVLX()) { |
8879 | MVT IntegerType = |
8880 | MVT::getIntegerVT(std::max((int)VT.getVectorNumElements(), 8)); |
8881 | SDValue MaskNode = DAG.getConstant(BlendMask, DL, IntegerType); |
8882 | return getVectorMaskingNode(V2, MaskNode, V1, Subtarget, DAG); |
8883 | } |
8884 | |
8885 | // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB. |
8886 | if (SDValue Masked = |
8887 | lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, Zeroable, DAG)) |
8888 | return Masked; |
8889 | |
8890 | // Scale the blend by the number of bytes per element. |
8891 | int Scale = VT.getScalarSizeInBits() / 8; |
8892 | |
8893 | // This form of blend is always done on bytes. Compute the byte vector |
8894 | // type. |
8895 | MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8); |
8896 | |
8897 | // Compute the VSELECT mask. Note that VSELECT is really confusing in the |
8898 | // mix of LLVM's code generator and the x86 backend. We tell the code |
8899 | // generator that boolean values in the elements of an x86 vector register |
8900 | // are -1 for true and 0 for false. We then use the LLVM semantics of 'true' |
8901 | // mapping a select to operand #1, and 'false' mapping to operand #2. The |
8902 | // reality in x86 is that vector masks (pre-AVX-512) use only the high bit |
8903 | // of the element (the remaining are ignored) and 0 in that high bit would |
8904 | // mean operand #1 while 1 in the high bit would mean operand #2. So while |
8905 | // the LLVM model for boolean values in vector elements gets the relevant |
8906 | // bit set, it is set backwards and over constrained relative to x86's |
8907 | // actual model. |
8908 | SmallVector<SDValue, 32> VSELECTMask; |
8909 | for (int i = 0, Size = Mask.size(); i < Size; ++i) |
8910 | for (int j = 0; j < Scale; ++j) |
8911 | VSELECTMask.push_back( |
8912 | Mask[i] < 0 ? DAG.getUNDEF(MVT::i8) |
8913 | : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL, |
8914 | MVT::i8)); |
8915 | |
8916 | V1 = DAG.getBitcast(BlendVT, V1); |
8917 | V2 = DAG.getBitcast(BlendVT, V2); |
8918 | return DAG.getBitcast( |
8919 | VT, |
8920 | DAG.getSelect(DL, BlendVT, DAG.getBuildVector(BlendVT, DL, VSELECTMask), |
8921 | V1, V2)); |
8922 | } |
8923 | case MVT::v16f32: |
8924 | case MVT::v8f64: |
8925 | case MVT::v8i64: |
8926 | case MVT::v16i32: |
8927 | case MVT::v32i16: |
8928 | case MVT::v64i8: { |
8929 | MVT IntegerType = |
8930 | MVT::getIntegerVT(std::max((int)VT.getVectorNumElements(), 8)); |
8931 | SDValue MaskNode = DAG.getConstant(BlendMask, DL, IntegerType); |
8932 | return getVectorMaskingNode(V2, MaskNode, V1, Subtarget, DAG); |
8933 | } |
8934 | default: |
8935 | llvm_unreachable("Not a supported integer vector type!")::llvm::llvm_unreachable_internal("Not a supported integer vector type!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8935); |
8936 | } |
8937 | } |
8938 | |
8939 | /// \brief Try to lower as a blend of elements from two inputs followed by |
8940 | /// a single-input permutation. |
8941 | /// |
8942 | /// This matches the pattern where we can blend elements from two inputs and |
8943 | /// then reduce the shuffle to a single-input permutation. |
8944 | static SDValue lowerVectorShuffleAsBlendAndPermute(const SDLoc &DL, MVT VT, |
8945 | SDValue V1, SDValue V2, |
8946 | ArrayRef<int> Mask, |
8947 | SelectionDAG &DAG) { |
8948 | // We build up the blend mask while checking whether a blend is a viable way |
8949 | // to reduce the shuffle. |
8950 | SmallVector<int, 32> BlendMask(Mask.size(), -1); |
8951 | SmallVector<int, 32> PermuteMask(Mask.size(), -1); |
8952 | |
8953 | for (int i = 0, Size = Mask.size(); i < Size; ++i) { |
8954 | if (Mask[i] < 0) |
8955 | continue; |
8956 | |
8957 | assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.")((Mask[i] < Size * 2 && "Shuffle input is out of bounds." ) ? static_cast<void> (0) : __assert_fail ("Mask[i] < Size * 2 && \"Shuffle input is out of bounds.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 8957, __PRETTY_FUNCTION__)); |
8958 | |
8959 | if (BlendMask[Mask[i] % Size] < 0) |
8960 | BlendMask[Mask[i] % Size] = Mask[i]; |
8961 | else if (BlendMask[Mask[i] % Size] != Mask[i]) |
8962 | return SDValue(); // Can't blend in the needed input! |
8963 | |
8964 | PermuteMask[i] = Mask[i] % Size; |
8965 | } |
8966 | |
8967 | SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask); |
8968 | return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask); |
8969 | } |
8970 | |
8971 | /// \brief Generic routine to decompose a shuffle and blend into independent |
8972 | /// blends and permutes. |
8973 | /// |
8974 | /// This matches the extremely common pattern for handling combined |
8975 | /// shuffle+blend operations on newer X86 ISAs where we have very fast blend |
8976 | /// operations. It will try to pick the best arrangement of shuffles and |
8977 | /// blends. |
8978 | static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(const SDLoc &DL, |
8979 | MVT VT, SDValue V1, |
8980 | SDValue V2, |
8981 | ArrayRef<int> Mask, |
8982 | SelectionDAG &DAG) { |
8983 | // Shuffle the input elements into the desired positions in V1 and V2 and |
8984 | // blend them together. |
8985 | SmallVector<int, 32> V1Mask(Mask.size(), -1); |
8986 | SmallVector<int, 32> V2Mask(Mask.size(), -1); |
8987 | SmallVector<int, 32> BlendMask(Mask.size(), -1); |
8988 | for (int i = 0, Size = Mask.size(); i < Size; ++i) |
8989 | if (Mask[i] >= 0 && Mask[i] < Size) { |
8990 | V1Mask[i] = Mask[i]; |
8991 | BlendMask[i] = i; |
8992 | } else if (Mask[i] >= Size) { |
8993 | V2Mask[i] = Mask[i] - Size; |
8994 | BlendMask[i] = i + Size; |
8995 | } |
8996 | |
8997 | // Try to lower with the simpler initial blend strategy unless one of the |
8998 | // input shuffles would be a no-op. We prefer to shuffle inputs as the |
8999 | // shuffle may be able to fold with a load or other benefit. However, when |
9000 | // we'll have to do 2x as many shuffles in order to achieve this, blending |
9001 | // first is a better strategy. |
9002 | if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask)) |
9003 | if (SDValue BlendPerm = |
9004 | lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG)) |
9005 | return BlendPerm; |
9006 | |
9007 | V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask); |
9008 | V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask); |
9009 | return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask); |
9010 | } |
9011 | |
9012 | /// \brief Try to lower a vector shuffle as a rotation. |
9013 | /// |
9014 | /// This is used for support PALIGNR for SSSE3 or VALIGND/Q for AVX512. |
9015 | static int matchVectorShuffleAsRotate(SDValue &V1, SDValue &V2, |
9016 | ArrayRef<int> Mask) { |
9017 | int NumElts = Mask.size(); |
9018 | |
9019 | // We need to detect various ways of spelling a rotation: |
9020 | // [11, 12, 13, 14, 15, 0, 1, 2] |
9021 | // [-1, 12, 13, 14, -1, -1, 1, -1] |
9022 | // [-1, -1, -1, -1, -1, -1, 1, 2] |
9023 | // [ 3, 4, 5, 6, 7, 8, 9, 10] |
9024 | // [-1, 4, 5, 6, -1, -1, 9, -1] |
9025 | // [-1, 4, 5, 6, -1, -1, -1, -1] |
9026 | int Rotation = 0; |
9027 | SDValue Lo, Hi; |
9028 | for (int i = 0; i < NumElts; ++i) { |
9029 | int M = Mask[i]; |
9030 | assert((M == SM_SentinelUndef || (0 <= M && M < (2*NumElts))) &&(((M == SM_SentinelUndef || (0 <= M && M < (2*NumElts ))) && "Unexpected mask index.") ? static_cast<void > (0) : __assert_fail ("(M == SM_SentinelUndef || (0 <= M && M < (2*NumElts))) && \"Unexpected mask index.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9031, __PRETTY_FUNCTION__)) |
9031 | "Unexpected mask index.")(((M == SM_SentinelUndef || (0 <= M && M < (2*NumElts ))) && "Unexpected mask index.") ? static_cast<void > (0) : __assert_fail ("(M == SM_SentinelUndef || (0 <= M && M < (2*NumElts))) && \"Unexpected mask index.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9031, __PRETTY_FUNCTION__)); |
9032 | if (M < 0) |
9033 | continue; |
9034 | |
9035 | // Determine where a rotated vector would have started. |
9036 | int StartIdx = i - (M % NumElts); |
9037 | if (StartIdx == 0) |
9038 | // The identity rotation isn't interesting, stop. |
9039 | return -1; |
9040 | |
9041 | // If we found the tail of a vector the rotation must be the missing |
9042 | // front. If we found the head of a vector, it must be how much of the |
9043 | // head. |
9044 | int CandidateRotation = StartIdx < 0 ? -StartIdx : NumElts - StartIdx; |
9045 | |
9046 | if (Rotation == 0) |
9047 | Rotation = CandidateRotation; |
9048 | else if (Rotation != CandidateRotation) |
9049 | // The rotations don't match, so we can't match this mask. |
9050 | return -1; |
9051 | |
9052 | // Compute which value this mask is pointing at. |
9053 | SDValue MaskV = M < NumElts ? V1 : V2; |
9054 | |
9055 | // Compute which of the two target values this index should be assigned |
9056 | // to. This reflects whether the high elements are remaining or the low |
9057 | // elements are remaining. |
9058 | SDValue &TargetV = StartIdx < 0 ? Hi : Lo; |
9059 | |
9060 | // Either set up this value if we've not encountered it before, or check |
9061 | // that it remains consistent. |
9062 | if (!TargetV) |
9063 | TargetV = MaskV; |
9064 | else if (TargetV != MaskV) |
9065 | // This may be a rotation, but it pulls from the inputs in some |
9066 | // unsupported interleaving. |
9067 | return -1; |
9068 | } |
9069 | |
9070 | // Check that we successfully analyzed the mask, and normalize the results. |
9071 | assert(Rotation != 0 && "Failed to locate a viable rotation!")((Rotation != 0 && "Failed to locate a viable rotation!" ) ? static_cast<void> (0) : __assert_fail ("Rotation != 0 && \"Failed to locate a viable rotation!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9071, __PRETTY_FUNCTION__)); |
9072 | assert((Lo || Hi) && "Failed to find a rotated input vector!")(((Lo || Hi) && "Failed to find a rotated input vector!" ) ? static_cast<void> (0) : __assert_fail ("(Lo || Hi) && \"Failed to find a rotated input vector!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9072, __PRETTY_FUNCTION__)); |
9073 | if (!Lo) |
9074 | Lo = Hi; |
9075 | else if (!Hi) |
9076 | Hi = Lo; |
9077 | |
9078 | V1 = Lo; |
9079 | V2 = Hi; |
9080 | |
9081 | return Rotation; |
9082 | } |
9083 | |
9084 | /// \brief Try to lower a vector shuffle as a byte rotation. |
9085 | /// |
9086 | /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary |
9087 | /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use |
9088 | /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will |
9089 | /// try to generically lower a vector shuffle through such an pattern. It |
9090 | /// does not check for the profitability of lowering either as PALIGNR or |
9091 | /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form. |
9092 | /// This matches shuffle vectors that look like: |
9093 | /// |
9094 | /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2] |
9095 | /// |
9096 | /// Essentially it concatenates V1 and V2, shifts right by some number of |
9097 | /// elements, and takes the low elements as the result. Note that while this is |
9098 | /// specified as a *right shift* because x86 is little-endian, it is a *left |
9099 | /// rotate* of the vector lanes. |
9100 | static int matchVectorShuffleAsByteRotate(MVT VT, SDValue &V1, SDValue &V2, |
9101 | ArrayRef<int> Mask) { |
9102 | // Don't accept any shuffles with zero elements. |
9103 | if (any_of(Mask, [](int M) { return M == SM_SentinelZero; })) |
9104 | return -1; |
9105 | |
9106 | // PALIGNR works on 128-bit lanes. |
9107 | SmallVector<int, 16> RepeatedMask; |
9108 | if (!is128BitLaneRepeatedShuffleMask(VT, Mask, RepeatedMask)) |
9109 | return -1; |
9110 | |
9111 | int Rotation = matchVectorShuffleAsRotate(V1, V2, RepeatedMask); |
9112 | if (Rotation <= 0) |
9113 | return -1; |
9114 | |
9115 | // PALIGNR rotates bytes, so we need to scale the |
9116 | // rotation based on how many bytes are in the vector lane. |
9117 | int NumElts = RepeatedMask.size(); |
9118 | int Scale = 16 / NumElts; |
9119 | return Rotation * Scale; |
9120 | } |
9121 | |
9122 | static SDValue lowerVectorShuffleAsByteRotate(const SDLoc &DL, MVT VT, |
9123 | SDValue V1, SDValue V2, |
9124 | ArrayRef<int> Mask, |
9125 | const X86Subtarget &Subtarget, |
9126 | SelectionDAG &DAG) { |
9127 | assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!")((!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!" ) ? static_cast<void> (0) : __assert_fail ("!isNoopShuffleMask(Mask) && \"We shouldn't lower no-op shuffles!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9127, __PRETTY_FUNCTION__)); |
9128 | |
9129 | SDValue Lo = V1, Hi = V2; |
9130 | int ByteRotation = matchVectorShuffleAsByteRotate(VT, Lo, Hi, Mask); |
9131 | if (ByteRotation <= 0) |
9132 | return SDValue(); |
9133 | |
9134 | // Cast the inputs to i8 vector of correct length to match PALIGNR or |
9135 | // PSLLDQ/PSRLDQ. |
9136 | MVT ByteVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8); |
9137 | Lo = DAG.getBitcast(ByteVT, Lo); |
9138 | Hi = DAG.getBitcast(ByteVT, Hi); |
9139 | |
9140 | // SSSE3 targets can use the palignr instruction. |
9141 | if (Subtarget.hasSSSE3()) { |
9142 | assert((!VT.is512BitVector() || Subtarget.hasBWI()) &&(((!VT.is512BitVector() || Subtarget.hasBWI()) && "512-bit PALIGNR requires BWI instructions" ) ? static_cast<void> (0) : __assert_fail ("(!VT.is512BitVector() || Subtarget.hasBWI()) && \"512-bit PALIGNR requires BWI instructions\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9143, __PRETTY_FUNCTION__)) |
9143 | "512-bit PALIGNR requires BWI instructions")(((!VT.is512BitVector() || Subtarget.hasBWI()) && "512-bit PALIGNR requires BWI instructions" ) ? static_cast<void> (0) : __assert_fail ("(!VT.is512BitVector() || Subtarget.hasBWI()) && \"512-bit PALIGNR requires BWI instructions\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9143, __PRETTY_FUNCTION__)); |
9144 | return DAG.getBitcast( |
9145 | VT, DAG.getNode(X86ISD::PALIGNR, DL, ByteVT, Lo, Hi, |
9146 | DAG.getConstant(ByteRotation, DL, MVT::i8))); |
9147 | } |
9148 | |
9149 | assert(VT.is128BitVector() &&((VT.is128BitVector() && "Rotate-based lowering only supports 128-bit lowering!" ) ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"Rotate-based lowering only supports 128-bit lowering!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9150, __PRETTY_FUNCTION__)) |
9150 | "Rotate-based lowering only supports 128-bit lowering!")((VT.is128BitVector() && "Rotate-based lowering only supports 128-bit lowering!" ) ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"Rotate-based lowering only supports 128-bit lowering!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9150, __PRETTY_FUNCTION__)); |
9151 | assert(Mask.size() <= 16 &&((Mask.size() <= 16 && "Can shuffle at most 16 bytes in a 128-bit vector!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() <= 16 && \"Can shuffle at most 16 bytes in a 128-bit vector!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9152, __PRETTY_FUNCTION__)) |
9152 | "Can shuffle at most 16 bytes in a 128-bit vector!")((Mask.size() <= 16 && "Can shuffle at most 16 bytes in a 128-bit vector!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() <= 16 && \"Can shuffle at most 16 bytes in a 128-bit vector!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9152, __PRETTY_FUNCTION__)); |
9153 | assert(ByteVT == MVT::v16i8 &&((ByteVT == MVT::v16i8 && "SSE2 rotate lowering only needed for v16i8!" ) ? static_cast<void> (0) : __assert_fail ("ByteVT == MVT::v16i8 && \"SSE2 rotate lowering only needed for v16i8!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9154, __PRETTY_FUNCTION__)) |
9154 | "SSE2 rotate lowering only needed for v16i8!")((ByteVT == MVT::v16i8 && "SSE2 rotate lowering only needed for v16i8!" ) ? static_cast<void> (0) : __assert_fail ("ByteVT == MVT::v16i8 && \"SSE2 rotate lowering only needed for v16i8!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9154, __PRETTY_FUNCTION__)); |
9155 | |
9156 | // Default SSE2 implementation |
9157 | int LoByteShift = 16 - ByteRotation; |
9158 | int HiByteShift = ByteRotation; |
9159 | |
9160 | SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v16i8, Lo, |
9161 | DAG.getConstant(LoByteShift, DL, MVT::i8)); |
9162 | SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v16i8, Hi, |
9163 | DAG.getConstant(HiByteShift, DL, MVT::i8)); |
9164 | return DAG.getBitcast(VT, |
9165 | DAG.getNode(ISD::OR, DL, MVT::v16i8, LoShift, HiShift)); |
9166 | } |
9167 | |
9168 | /// \brief Try to lower a vector shuffle as a dword/qword rotation. |
9169 | /// |
9170 | /// AVX512 has a VALIGND/VALIGNQ instructions that will do an arbitrary |
9171 | /// rotation of the concatenation of two vectors; This routine will |
9172 | /// try to generically lower a vector shuffle through such an pattern. |
9173 | /// |
9174 | /// Essentially it concatenates V1 and V2, shifts right by some number of |
9175 | /// elements, and takes the low elements as the result. Note that while this is |
9176 | /// specified as a *right shift* because x86 is little-endian, it is a *left |
9177 | /// rotate* of the vector lanes. |
9178 | static SDValue lowerVectorShuffleAsRotate(const SDLoc &DL, MVT VT, |
9179 | SDValue V1, SDValue V2, |
9180 | ArrayRef<int> Mask, |
9181 | const X86Subtarget &Subtarget, |
9182 | SelectionDAG &DAG) { |
9183 | assert((VT.getScalarType() == MVT::i32 || VT.getScalarType() == MVT::i64) &&(((VT.getScalarType() == MVT::i32 || VT.getScalarType() == MVT ::i64) && "Only 32-bit and 64-bit elements are supported!" ) ? static_cast<void> (0) : __assert_fail ("(VT.getScalarType() == MVT::i32 || VT.getScalarType() == MVT::i64) && \"Only 32-bit and 64-bit elements are supported!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9184, __PRETTY_FUNCTION__)) |
9184 | "Only 32-bit and 64-bit elements are supported!")(((VT.getScalarType() == MVT::i32 || VT.getScalarType() == MVT ::i64) && "Only 32-bit and 64-bit elements are supported!" ) ? static_cast<void> (0) : __assert_fail ("(VT.getScalarType() == MVT::i32 || VT.getScalarType() == MVT::i64) && \"Only 32-bit and 64-bit elements are supported!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9184, __PRETTY_FUNCTION__)); |
9185 | |
9186 | // 128/256-bit vectors are only supported with VLX. |
9187 | assert((Subtarget.hasVLX() || (!VT.is128BitVector() && !VT.is256BitVector()))(((Subtarget.hasVLX() || (!VT.is128BitVector() && !VT .is256BitVector())) && "VLX required for 128/256-bit vectors" ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasVLX() || (!VT.is128BitVector() && !VT.is256BitVector())) && \"VLX required for 128/256-bit vectors\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9188, __PRETTY_FUNCTION__)) |
9188 | && "VLX required for 128/256-bit vectors")(((Subtarget.hasVLX() || (!VT.is128BitVector() && !VT .is256BitVector())) && "VLX required for 128/256-bit vectors" ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasVLX() || (!VT.is128BitVector() && !VT.is256BitVector())) && \"VLX required for 128/256-bit vectors\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9188, __PRETTY_FUNCTION__)); |
9189 | |
9190 | SDValue Lo = V1, Hi = V2; |
9191 | int Rotation = matchVectorShuffleAsRotate(Lo, Hi, Mask); |
9192 | if (Rotation <= 0) |
9193 | return SDValue(); |
9194 | |
9195 | return DAG.getNode(X86ISD::VALIGN, DL, VT, Lo, Hi, |
9196 | DAG.getConstant(Rotation, DL, MVT::i8)); |
9197 | } |
9198 | |
9199 | /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros). |
9200 | /// |
9201 | /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and |
9202 | /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function |
9203 | /// matches elements from one of the input vectors shuffled to the left or |
9204 | /// right with zeroable elements 'shifted in'. It handles both the strictly |
9205 | /// bit-wise element shifts and the byte shift across an entire 128-bit double |
9206 | /// quad word lane. |
9207 | /// |
9208 | /// PSHL : (little-endian) left bit shift. |
9209 | /// [ zz, 0, zz, 2 ] |
9210 | /// [ -1, 4, zz, -1 ] |
9211 | /// PSRL : (little-endian) right bit shift. |
9212 | /// [ 1, zz, 3, zz] |
9213 | /// [ -1, -1, 7, zz] |
9214 | /// PSLLDQ : (little-endian) left byte shift |
9215 | /// [ zz, 0, 1, 2, 3, 4, 5, 6] |
9216 | /// [ zz, zz, -1, -1, 2, 3, 4, -1] |
9217 | /// [ zz, zz, zz, zz, zz, zz, -1, 1] |
9218 | /// PSRLDQ : (little-endian) right byte shift |
9219 | /// [ 5, 6, 7, zz, zz, zz, zz, zz] |
9220 | /// [ -1, 5, 6, 7, zz, zz, zz, zz] |
9221 | /// [ 1, 2, -1, -1, -1, -1, zz, zz] |
9222 | static int matchVectorShuffleAsShift(MVT &ShiftVT, unsigned &Opcode, |
9223 | unsigned ScalarSizeInBits, |
9224 | ArrayRef<int> Mask, int MaskOffset, |
9225 | const APInt &Zeroable, |
9226 | const X86Subtarget &Subtarget) { |
9227 | int Size = Mask.size(); |
9228 | unsigned SizeInBits = Size * ScalarSizeInBits; |
9229 | |
9230 | auto CheckZeros = [&](int Shift, int Scale, bool Left) { |
9231 | for (int i = 0; i < Size; i += Scale) |
9232 | for (int j = 0; j < Shift; ++j) |
9233 | if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))]) |
9234 | return false; |
9235 | |
9236 | return true; |
9237 | }; |
9238 | |
9239 | auto MatchShift = [&](int Shift, int Scale, bool Left) { |
9240 | for (int i = 0; i != Size; i += Scale) { |
9241 | unsigned Pos = Left ? i + Shift : i; |
9242 | unsigned Low = Left ? i : i + Shift; |
9243 | unsigned Len = Scale - Shift; |
9244 | if (!isSequentialOrUndefInRange(Mask, Pos, Len, Low + MaskOffset)) |
9245 | return -1; |
9246 | } |
9247 | |
9248 | int ShiftEltBits = ScalarSizeInBits * Scale; |
9249 | bool ByteShift = ShiftEltBits > 64; |
9250 | Opcode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI) |
9251 | : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI); |
9252 | int ShiftAmt = Shift * ScalarSizeInBits / (ByteShift ? 8 : 1); |
9253 | |
9254 | // Normalize the scale for byte shifts to still produce an i64 element |
9255 | // type. |
9256 | Scale = ByteShift ? Scale / 2 : Scale; |
9257 | |
9258 | // We need to round trip through the appropriate type for the shift. |
9259 | MVT ShiftSVT = MVT::getIntegerVT(ScalarSizeInBits * Scale); |
9260 | ShiftVT = ByteShift ? MVT::getVectorVT(MVT::i8, SizeInBits / 8) |
9261 | : MVT::getVectorVT(ShiftSVT, Size / Scale); |
9262 | return (int)ShiftAmt; |
9263 | }; |
9264 | |
9265 | // SSE/AVX supports logical shifts up to 64-bit integers - so we can just |
9266 | // keep doubling the size of the integer elements up to that. We can |
9267 | // then shift the elements of the integer vector by whole multiples of |
9268 | // their width within the elements of the larger integer vector. Test each |
9269 | // multiple to see if we can find a match with the moved element indices |
9270 | // and that the shifted in elements are all zeroable. |
9271 | unsigned MaxWidth = ((SizeInBits == 512) && !Subtarget.hasBWI() ? 64 : 128); |
9272 | for (int Scale = 2; Scale * ScalarSizeInBits <= MaxWidth; Scale *= 2) |
9273 | for (int Shift = 1; Shift != Scale; ++Shift) |
9274 | for (bool Left : {true, false}) |
9275 | if (CheckZeros(Shift, Scale, Left)) { |
9276 | int ShiftAmt = MatchShift(Shift, Scale, Left); |
9277 | if (0 < ShiftAmt) |
9278 | return ShiftAmt; |
9279 | } |
9280 | |
9281 | // no match |
9282 | return -1; |
9283 | } |
9284 | |
9285 | static SDValue lowerVectorShuffleAsShift(const SDLoc &DL, MVT VT, SDValue V1, |
9286 | SDValue V2, ArrayRef<int> Mask, |
9287 | const APInt &Zeroable, |
9288 | const X86Subtarget &Subtarget, |
9289 | SelectionDAG &DAG) { |
9290 | int Size = Mask.size(); |
9291 | assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size")((Size == (int)VT.getVectorNumElements() && "Unexpected mask size" ) ? static_cast<void> (0) : __assert_fail ("Size == (int)VT.getVectorNumElements() && \"Unexpected mask size\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9291, __PRETTY_FUNCTION__)); |
9292 | |
9293 | MVT ShiftVT; |
9294 | SDValue V = V1; |
9295 | unsigned Opcode; |
9296 | |
9297 | // Try to match shuffle against V1 shift. |
9298 | int ShiftAmt = matchVectorShuffleAsShift( |
9299 | ShiftVT, Opcode, VT.getScalarSizeInBits(), Mask, 0, Zeroable, Subtarget); |
9300 | |
9301 | // If V1 failed, try to match shuffle against V2 shift. |
9302 | if (ShiftAmt < 0) { |
9303 | ShiftAmt = |
9304 | matchVectorShuffleAsShift(ShiftVT, Opcode, VT.getScalarSizeInBits(), |
9305 | Mask, Size, Zeroable, Subtarget); |
9306 | V = V2; |
9307 | } |
9308 | |
9309 | if (ShiftAmt < 0) |
9310 | return SDValue(); |
9311 | |
9312 | assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&((DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) && "Illegal integer vector type") ? static_cast<void> (0) : __assert_fail ("DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) && \"Illegal integer vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9313, __PRETTY_FUNCTION__)) |
9313 | "Illegal integer vector type")((DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) && "Illegal integer vector type") ? static_cast<void> (0) : __assert_fail ("DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) && \"Illegal integer vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9313, __PRETTY_FUNCTION__)); |
9314 | V = DAG.getBitcast(ShiftVT, V); |
9315 | V = DAG.getNode(Opcode, DL, ShiftVT, V, |
9316 | DAG.getConstant(ShiftAmt, DL, MVT::i8)); |
9317 | return DAG.getBitcast(VT, V); |
9318 | } |
9319 | |
9320 | /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ. |
9321 | static SDValue lowerVectorShuffleWithSSE4A(const SDLoc &DL, MVT VT, SDValue V1, |
9322 | SDValue V2, ArrayRef<int> Mask, |
9323 | const APInt &Zeroable, |
9324 | SelectionDAG &DAG) { |
9325 | int Size = Mask.size(); |
9326 | int HalfSize = Size / 2; |
9327 | assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size")((Size == (int)VT.getVectorNumElements() && "Unexpected mask size" ) ? static_cast<void> (0) : __assert_fail ("Size == (int)VT.getVectorNumElements() && \"Unexpected mask size\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9327, __PRETTY_FUNCTION__)); |
9328 | assert(!Zeroable.isAllOnesValue() && "Fully zeroable shuffle mask")((!Zeroable.isAllOnesValue() && "Fully zeroable shuffle mask" ) ? static_cast<void> (0) : __assert_fail ("!Zeroable.isAllOnesValue() && \"Fully zeroable shuffle mask\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9328, __PRETTY_FUNCTION__)); |
9329 | |
9330 | // Upper half must be undefined. |
9331 | if (!isUndefInRange(Mask, HalfSize, HalfSize)) |
9332 | return SDValue(); |
9333 | |
9334 | // EXTRQ: Extract Len elements from lower half of source, starting at Idx. |
9335 | // Remainder of lower half result is zero and upper half is all undef. |
9336 | auto LowerAsEXTRQ = [&]() { |
9337 | // Determine the extraction length from the part of the |
9338 | // lower half that isn't zeroable. |
9339 | int Len = HalfSize; |
9340 | for (; Len > 0; --Len) |
9341 | if (!Zeroable[Len - 1]) |
9342 | break; |
9343 | assert(Len > 0 && "Zeroable shuffle mask")((Len > 0 && "Zeroable shuffle mask") ? static_cast <void> (0) : __assert_fail ("Len > 0 && \"Zeroable shuffle mask\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9343, __PRETTY_FUNCTION__)); |
9344 | |
9345 | // Attempt to match first Len sequential elements from the lower half. |
9346 | SDValue Src; |
9347 | int Idx = -1; |
9348 | for (int i = 0; i != Len; ++i) { |
9349 | int M = Mask[i]; |
9350 | if (M < 0) |
9351 | continue; |
9352 | SDValue &V = (M < Size ? V1 : V2); |
9353 | M = M % Size; |
9354 | |
9355 | // The extracted elements must start at a valid index and all mask |
9356 | // elements must be in the lower half. |
9357 | if (i > M || M >= HalfSize) |
9358 | return SDValue(); |
9359 | |
9360 | if (Idx < 0 || (Src == V && Idx == (M - i))) { |
9361 | Src = V; |
9362 | Idx = M - i; |
9363 | continue; |
9364 | } |
9365 | return SDValue(); |
9366 | } |
9367 | |
9368 | if (Idx < 0) |
9369 | return SDValue(); |
9370 | |
9371 | assert((Idx + Len) <= HalfSize && "Illegal extraction mask")(((Idx + Len) <= HalfSize && "Illegal extraction mask" ) ? static_cast<void> (0) : __assert_fail ("(Idx + Len) <= HalfSize && \"Illegal extraction mask\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9371, __PRETTY_FUNCTION__)); |
9372 | int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f; |
9373 | int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f; |
9374 | return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src, |
9375 | DAG.getConstant(BitLen, DL, MVT::i8), |
9376 | DAG.getConstant(BitIdx, DL, MVT::i8)); |
9377 | }; |
9378 | |
9379 | if (SDValue ExtrQ = LowerAsEXTRQ()) |
9380 | return ExtrQ; |
9381 | |
9382 | // INSERTQ: Extract lowest Len elements from lower half of second source and |
9383 | // insert over first source, starting at Idx. |
9384 | // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... } |
9385 | auto LowerAsInsertQ = [&]() { |
9386 | for (int Idx = 0; Idx != HalfSize; ++Idx) { |
9387 | SDValue Base; |
9388 | |
9389 | // Attempt to match first source from mask before insertion point. |
9390 | if (isUndefInRange(Mask, 0, Idx)) { |
9391 | /* EMPTY */ |
9392 | } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) { |
9393 | Base = V1; |
9394 | } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) { |
9395 | Base = V2; |
9396 | } else { |
9397 | continue; |
9398 | } |
9399 | |
9400 | // Extend the extraction length looking to match both the insertion of |
9401 | // the second source and the remaining elements of the first. |
9402 | for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) { |
9403 | SDValue Insert; |
9404 | int Len = Hi - Idx; |
9405 | |
9406 | // Match insertion. |
9407 | if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) { |
9408 | Insert = V1; |
9409 | } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) { |
9410 | Insert = V2; |
9411 | } else { |
9412 | continue; |
9413 | } |
9414 | |
9415 | // Match the remaining elements of the lower half. |
9416 | if (isUndefInRange(Mask, Hi, HalfSize - Hi)) { |
9417 | /* EMPTY */ |
9418 | } else if ((!Base || (Base == V1)) && |
9419 | isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) { |
9420 | Base = V1; |
9421 | } else if ((!Base || (Base == V2)) && |
9422 | isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, |
9423 | Size + Hi)) { |
9424 | Base = V2; |
9425 | } else { |
9426 | continue; |
9427 | } |
9428 | |
9429 | // We may not have a base (first source) - this can safely be undefined. |
9430 | if (!Base) |
9431 | Base = DAG.getUNDEF(VT); |
9432 | |
9433 | int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f; |
9434 | int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f; |
9435 | return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert, |
9436 | DAG.getConstant(BitLen, DL, MVT::i8), |
9437 | DAG.getConstant(BitIdx, DL, MVT::i8)); |
9438 | } |
9439 | } |
9440 | |
9441 | return SDValue(); |
9442 | }; |
9443 | |
9444 | if (SDValue InsertQ = LowerAsInsertQ()) |
9445 | return InsertQ; |
9446 | |
9447 | return SDValue(); |
9448 | } |
9449 | |
9450 | /// \brief Lower a vector shuffle as a zero or any extension. |
9451 | /// |
9452 | /// Given a specific number of elements, element bit width, and extension |
9453 | /// stride, produce either a zero or any extension based on the available |
9454 | /// features of the subtarget. The extended elements are consecutive and |
9455 | /// begin and can start from an offsetted element index in the input; to |
9456 | /// avoid excess shuffling the offset must either being in the bottom lane |
9457 | /// or at the start of a higher lane. All extended elements must be from |
9458 | /// the same lane. |
9459 | static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend( |
9460 | const SDLoc &DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV, |
9461 | ArrayRef<int> Mask, const X86Subtarget &Subtarget, SelectionDAG &DAG) { |
9462 | assert(Scale > 1 && "Need a scale to extend.")((Scale > 1 && "Need a scale to extend.") ? static_cast <void> (0) : __assert_fail ("Scale > 1 && \"Need a scale to extend.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9462, __PRETTY_FUNCTION__)); |
9463 | int EltBits = VT.getScalarSizeInBits(); |
9464 | int NumElements = VT.getVectorNumElements(); |
9465 | int NumEltsPerLane = 128 / EltBits; |
9466 | int OffsetLane = Offset / NumEltsPerLane; |
9467 | assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&(((EltBits == 8 || EltBits == 16 || EltBits == 32) && "Only 8, 16, and 32 bit elements can be extended.") ? static_cast <void> (0) : __assert_fail ("(EltBits == 8 || EltBits == 16 || EltBits == 32) && \"Only 8, 16, and 32 bit elements can be extended.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9468, __PRETTY_FUNCTION__)) |
9468 | "Only 8, 16, and 32 bit elements can be extended.")(((EltBits == 8 || EltBits == 16 || EltBits == 32) && "Only 8, 16, and 32 bit elements can be extended.") ? static_cast <void> (0) : __assert_fail ("(EltBits == 8 || EltBits == 16 || EltBits == 32) && \"Only 8, 16, and 32 bit elements can be extended.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9468, __PRETTY_FUNCTION__)); |
9469 | assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.")((Scale * EltBits <= 64 && "Cannot zero extend past 64 bits." ) ? static_cast<void> (0) : __assert_fail ("Scale * EltBits <= 64 && \"Cannot zero extend past 64 bits.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9469, __PRETTY_FUNCTION__)); |
9470 | assert(0 <= Offset && "Extension offset must be positive.")((0 <= Offset && "Extension offset must be positive." ) ? static_cast<void> (0) : __assert_fail ("0 <= Offset && \"Extension offset must be positive.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9470, __PRETTY_FUNCTION__)); |
9471 | assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&(((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0 ) && "Extension offset must be in the first lane or start an upper lane." ) ? static_cast<void> (0) : __assert_fail ("(Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) && \"Extension offset must be in the first lane or start an upper lane.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9472, __PRETTY_FUNCTION__)) |
9472 | "Extension offset must be in the first lane or start an upper lane.")(((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0 ) && "Extension offset must be in the first lane or start an upper lane." ) ? static_cast<void> (0) : __assert_fail ("(Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) && \"Extension offset must be in the first lane or start an upper lane.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9472, __PRETTY_FUNCTION__)); |
9473 | |
9474 | // Check that an index is in same lane as the base offset. |
9475 | auto SafeOffset = [&](int Idx) { |
9476 | return OffsetLane == (Idx / NumEltsPerLane); |
9477 | }; |
9478 | |
9479 | // Shift along an input so that the offset base moves to the first element. |
9480 | auto ShuffleOffset = [&](SDValue V) { |
9481 | if (!Offset) |
9482 | return V; |
9483 | |
9484 | SmallVector<int, 8> ShMask((unsigned)NumElements, -1); |
9485 | for (int i = 0; i * Scale < NumElements; ++i) { |
9486 | int SrcIdx = i + Offset; |
9487 | ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1; |
9488 | } |
9489 | return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask); |
9490 | }; |
9491 | |
9492 | // Found a valid zext mask! Try various lowering strategies based on the |
9493 | // input type and available ISA extensions. |
9494 | if (Subtarget.hasSSE41()) { |
9495 | // Not worth offsetting 128-bit vectors if scale == 2, a pattern using |
9496 | // PUNPCK will catch this in a later shuffle match. |
9497 | if (Offset && Scale == 2 && VT.is128BitVector()) |
9498 | return SDValue(); |
9499 | MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale), |
9500 | NumElements / Scale); |
9501 | InputV = ShuffleOffset(InputV); |
9502 | InputV = getExtendInVec(X86ISD::VZEXT, DL, ExtVT, InputV, DAG); |
9503 | return DAG.getBitcast(VT, InputV); |
9504 | } |
9505 | |
9506 | assert(VT.is128BitVector() && "Only 128-bit vectors can be extended.")((VT.is128BitVector() && "Only 128-bit vectors can be extended." ) ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"Only 128-bit vectors can be extended.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9506, __PRETTY_FUNCTION__)); |
9507 | |
9508 | // For any extends we can cheat for larger element sizes and use shuffle |
9509 | // instructions that can fold with a load and/or copy. |
9510 | if (AnyExt && EltBits == 32) { |
9511 | int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1, |
9512 | -1}; |
9513 | return DAG.getBitcast( |
9514 | VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, |
9515 | DAG.getBitcast(MVT::v4i32, InputV), |
9516 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG))); |
9517 | } |
9518 | if (AnyExt && EltBits == 16 && Scale > 2) { |
9519 | int PSHUFDMask[4] = {Offset / 2, -1, |
9520 | SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1}; |
9521 | InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, |
9522 | DAG.getBitcast(MVT::v4i32, InputV), |
9523 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)); |
9524 | int PSHUFWMask[4] = {1, -1, -1, -1}; |
9525 | unsigned OddEvenOp = (Offset & 1 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW); |
9526 | return DAG.getBitcast( |
9527 | VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16, |
9528 | DAG.getBitcast(MVT::v8i16, InputV), |
9529 | getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG))); |
9530 | } |
9531 | |
9532 | // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes |
9533 | // to 64-bits. |
9534 | if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget.hasSSE4A()) { |
9535 | assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!")((NumElements == (int)Mask.size() && "Unexpected shuffle mask size!" ) ? static_cast<void> (0) : __assert_fail ("NumElements == (int)Mask.size() && \"Unexpected shuffle mask size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9535, __PRETTY_FUNCTION__)); |
9536 | assert(VT.is128BitVector() && "Unexpected vector width!")((VT.is128BitVector() && "Unexpected vector width!") ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"Unexpected vector width!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9536, __PRETTY_FUNCTION__)); |
9537 | |
9538 | int LoIdx = Offset * EltBits; |
9539 | SDValue Lo = DAG.getBitcast( |
9540 | MVT::v2i64, DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV, |
9541 | DAG.getConstant(EltBits, DL, MVT::i8), |
9542 | DAG.getConstant(LoIdx, DL, MVT::i8))); |
9543 | |
9544 | if (isUndefInRange(Mask, NumElements / 2, NumElements / 2) || |
9545 | !SafeOffset(Offset + 1)) |
9546 | return DAG.getBitcast(VT, Lo); |
9547 | |
9548 | int HiIdx = (Offset + 1) * EltBits; |
9549 | SDValue Hi = DAG.getBitcast( |
9550 | MVT::v2i64, DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV, |
9551 | DAG.getConstant(EltBits, DL, MVT::i8), |
9552 | DAG.getConstant(HiIdx, DL, MVT::i8))); |
9553 | return DAG.getBitcast(VT, |
9554 | DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi)); |
9555 | } |
9556 | |
9557 | // If this would require more than 2 unpack instructions to expand, use |
9558 | // pshufb when available. We can only use more than 2 unpack instructions |
9559 | // when zero extending i8 elements which also makes it easier to use pshufb. |
9560 | if (Scale > 4 && EltBits == 8 && Subtarget.hasSSSE3()) { |
9561 | assert(NumElements == 16 && "Unexpected byte vector width!")((NumElements == 16 && "Unexpected byte vector width!" ) ? static_cast<void> (0) : __assert_fail ("NumElements == 16 && \"Unexpected byte vector width!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9561, __PRETTY_FUNCTION__)); |
9562 | SDValue PSHUFBMask[16]; |
9563 | for (int i = 0; i < 16; ++i) { |
9564 | int Idx = Offset + (i / Scale); |
9565 | PSHUFBMask[i] = DAG.getConstant( |
9566 | (i % Scale == 0 && SafeOffset(Idx)) ? Idx : 0x80, DL, MVT::i8); |
9567 | } |
9568 | InputV = DAG.getBitcast(MVT::v16i8, InputV); |
9569 | return DAG.getBitcast( |
9570 | VT, DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV, |
9571 | DAG.getBuildVector(MVT::v16i8, DL, PSHUFBMask))); |
9572 | } |
9573 | |
9574 | // If we are extending from an offset, ensure we start on a boundary that |
9575 | // we can unpack from. |
9576 | int AlignToUnpack = Offset % (NumElements / Scale); |
9577 | if (AlignToUnpack) { |
9578 | SmallVector<int, 8> ShMask((unsigned)NumElements, -1); |
9579 | for (int i = AlignToUnpack; i < NumElements; ++i) |
9580 | ShMask[i - AlignToUnpack] = i; |
9581 | InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask); |
9582 | Offset -= AlignToUnpack; |
9583 | } |
9584 | |
9585 | // Otherwise emit a sequence of unpacks. |
9586 | do { |
9587 | unsigned UnpackLoHi = X86ISD::UNPCKL; |
9588 | if (Offset >= (NumElements / 2)) { |
9589 | UnpackLoHi = X86ISD::UNPCKH; |
9590 | Offset -= (NumElements / 2); |
9591 | } |
9592 | |
9593 | MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements); |
9594 | SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT) |
9595 | : getZeroVector(InputVT, Subtarget, DAG, DL); |
9596 | InputV = DAG.getBitcast(InputVT, InputV); |
9597 | InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext); |
9598 | Scale /= 2; |
9599 | EltBits *= 2; |
9600 | NumElements /= 2; |
9601 | } while (Scale > 1); |
9602 | return DAG.getBitcast(VT, InputV); |
9603 | } |
9604 | |
9605 | /// \brief Try to lower a vector shuffle as a zero extension on any microarch. |
9606 | /// |
9607 | /// This routine will try to do everything in its power to cleverly lower |
9608 | /// a shuffle which happens to match the pattern of a zero extend. It doesn't |
9609 | /// check for the profitability of this lowering, it tries to aggressively |
9610 | /// match this pattern. It will use all of the micro-architectural details it |
9611 | /// can to emit an efficient lowering. It handles both blends with all-zero |
9612 | /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to |
9613 | /// masking out later). |
9614 | /// |
9615 | /// The reason we have dedicated lowering for zext-style shuffles is that they |
9616 | /// are both incredibly common and often quite performance sensitive. |
9617 | static SDValue lowerVectorShuffleAsZeroOrAnyExtend( |
9618 | const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, |
9619 | const APInt &Zeroable, const X86Subtarget &Subtarget, |
9620 | SelectionDAG &DAG) { |
9621 | int Bits = VT.getSizeInBits(); |
9622 | int NumLanes = Bits / 128; |
9623 | int NumElements = VT.getVectorNumElements(); |
9624 | int NumEltsPerLane = NumElements / NumLanes; |
9625 | assert(VT.getScalarSizeInBits() <= 32 &&((VT.getScalarSizeInBits() <= 32 && "Exceeds 32-bit integer zero extension limit" ) ? static_cast<void> (0) : __assert_fail ("VT.getScalarSizeInBits() <= 32 && \"Exceeds 32-bit integer zero extension limit\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9626, __PRETTY_FUNCTION__)) |
9626 | "Exceeds 32-bit integer zero extension limit")((VT.getScalarSizeInBits() <= 32 && "Exceeds 32-bit integer zero extension limit" ) ? static_cast<void> (0) : __assert_fail ("VT.getScalarSizeInBits() <= 32 && \"Exceeds 32-bit integer zero extension limit\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9626, __PRETTY_FUNCTION__)); |
9627 | assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size")(((int)Mask.size() == NumElements && "Unexpected shuffle mask size" ) ? static_cast<void> (0) : __assert_fail ("(int)Mask.size() == NumElements && \"Unexpected shuffle mask size\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9627, __PRETTY_FUNCTION__)); |
9628 | |
9629 | // Define a helper function to check a particular ext-scale and lower to it if |
9630 | // valid. |
9631 | auto Lower = [&](int Scale) -> SDValue { |
9632 | SDValue InputV; |
9633 | bool AnyExt = true; |
9634 | int Offset = 0; |
9635 | int Matches = 0; |
9636 | for (int i = 0; i < NumElements; ++i) { |
9637 | int M = Mask[i]; |
9638 | if (M < 0) |
9639 | continue; // Valid anywhere but doesn't tell us anything. |
9640 | if (i % Scale != 0) { |
9641 | // Each of the extended elements need to be zeroable. |
9642 | if (!Zeroable[i]) |
9643 | return SDValue(); |
9644 | |
9645 | // We no longer are in the anyext case. |
9646 | AnyExt = false; |
9647 | continue; |
9648 | } |
9649 | |
9650 | // Each of the base elements needs to be consecutive indices into the |
9651 | // same input vector. |
9652 | SDValue V = M < NumElements ? V1 : V2; |
9653 | M = M % NumElements; |
9654 | if (!InputV) { |
9655 | InputV = V; |
9656 | Offset = M - (i / Scale); |
9657 | } else if (InputV != V) |
9658 | return SDValue(); // Flip-flopping inputs. |
9659 | |
9660 | // Offset must start in the lowest 128-bit lane or at the start of an |
9661 | // upper lane. |
9662 | // FIXME: Is it ever worth allowing a negative base offset? |
9663 | if (!((0 <= Offset && Offset < NumEltsPerLane) || |
9664 | (Offset % NumEltsPerLane) == 0)) |
9665 | return SDValue(); |
9666 | |
9667 | // If we are offsetting, all referenced entries must come from the same |
9668 | // lane. |
9669 | if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane)) |
9670 | return SDValue(); |
9671 | |
9672 | if ((M % NumElements) != (Offset + (i / Scale))) |
9673 | return SDValue(); // Non-consecutive strided elements. |
9674 | Matches++; |
9675 | } |
9676 | |
9677 | // If we fail to find an input, we have a zero-shuffle which should always |
9678 | // have already been handled. |
9679 | // FIXME: Maybe handle this here in case during blending we end up with one? |
9680 | if (!InputV) |
9681 | return SDValue(); |
9682 | |
9683 | // If we are offsetting, don't extend if we only match a single input, we |
9684 | // can always do better by using a basic PSHUF or PUNPCK. |
9685 | if (Offset != 0 && Matches < 2) |
9686 | return SDValue(); |
9687 | |
9688 | return lowerVectorShuffleAsSpecificZeroOrAnyExtend( |
9689 | DL, VT, Scale, Offset, AnyExt, InputV, Mask, Subtarget, DAG); |
9690 | }; |
9691 | |
9692 | // The widest scale possible for extending is to a 64-bit integer. |
9693 | assert(Bits % 64 == 0 &&((Bits % 64 == 0 && "The number of bits in a vector must be divisible by 64 on x86!" ) ? static_cast<void> (0) : __assert_fail ("Bits % 64 == 0 && \"The number of bits in a vector must be divisible by 64 on x86!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9694, __PRETTY_FUNCTION__)) |
9694 | "The number of bits in a vector must be divisible by 64 on x86!")((Bits % 64 == 0 && "The number of bits in a vector must be divisible by 64 on x86!" ) ? static_cast<void> (0) : __assert_fail ("Bits % 64 == 0 && \"The number of bits in a vector must be divisible by 64 on x86!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9694, __PRETTY_FUNCTION__)); |
9695 | int NumExtElements = Bits / 64; |
9696 | |
9697 | // Each iteration, try extending the elements half as much, but into twice as |
9698 | // many elements. |
9699 | for (; NumExtElements < NumElements; NumExtElements *= 2) { |
9700 | assert(NumElements % NumExtElements == 0 &&((NumElements % NumExtElements == 0 && "The input vector size must be divisible by the extended size." ) ? static_cast<void> (0) : __assert_fail ("NumElements % NumExtElements == 0 && \"The input vector size must be divisible by the extended size.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9701, __PRETTY_FUNCTION__)) |
9701 | "The input vector size must be divisible by the extended size.")((NumElements % NumExtElements == 0 && "The input vector size must be divisible by the extended size." ) ? static_cast<void> (0) : __assert_fail ("NumElements % NumExtElements == 0 && \"The input vector size must be divisible by the extended size.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9701, __PRETTY_FUNCTION__)); |
9702 | if (SDValue V = Lower(NumElements / NumExtElements)) |
9703 | return V; |
9704 | } |
9705 | |
9706 | // General extends failed, but 128-bit vectors may be able to use MOVQ. |
9707 | if (Bits != 128) |
9708 | return SDValue(); |
9709 | |
9710 | // Returns one of the source operands if the shuffle can be reduced to a |
9711 | // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits. |
9712 | auto CanZExtLowHalf = [&]() { |
9713 | for (int i = NumElements / 2; i != NumElements; ++i) |
9714 | if (!Zeroable[i]) |
9715 | return SDValue(); |
9716 | if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0)) |
9717 | return V1; |
9718 | if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements)) |
9719 | return V2; |
9720 | return SDValue(); |
9721 | }; |
9722 | |
9723 | if (SDValue V = CanZExtLowHalf()) { |
9724 | V = DAG.getBitcast(MVT::v2i64, V); |
9725 | V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V); |
9726 | return DAG.getBitcast(VT, V); |
9727 | } |
9728 | |
9729 | // No viable ext lowering found. |
9730 | return SDValue(); |
9731 | } |
9732 | |
9733 | /// \brief Try to get a scalar value for a specific element of a vector. |
9734 | /// |
9735 | /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar. |
9736 | static SDValue getScalarValueForVectorElement(SDValue V, int Idx, |
9737 | SelectionDAG &DAG) { |
9738 | MVT VT = V.getSimpleValueType(); |
9739 | MVT EltVT = VT.getVectorElementType(); |
9740 | V = peekThroughBitcasts(V); |
9741 | |
9742 | // If the bitcasts shift the element size, we can't extract an equivalent |
9743 | // element from it. |
9744 | MVT NewVT = V.getSimpleValueType(); |
9745 | if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits()) |
9746 | return SDValue(); |
9747 | |
9748 | if (V.getOpcode() == ISD::BUILD_VECTOR || |
9749 | (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) { |
9750 | // Ensure the scalar operand is the same size as the destination. |
9751 | // FIXME: Add support for scalar truncation where possible. |
9752 | SDValue S = V.getOperand(Idx); |
9753 | if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits()) |
9754 | return DAG.getBitcast(EltVT, S); |
9755 | } |
9756 | |
9757 | return SDValue(); |
9758 | } |
9759 | |
9760 | /// \brief Helper to test for a load that can be folded with x86 shuffles. |
9761 | /// |
9762 | /// This is particularly important because the set of instructions varies |
9763 | /// significantly based on whether the operand is a load or not. |
9764 | static bool isShuffleFoldableLoad(SDValue V) { |
9765 | V = peekThroughBitcasts(V); |
9766 | return ISD::isNON_EXTLoad(V.getNode()); |
9767 | } |
9768 | |
9769 | /// \brief Try to lower insertion of a single element into a zero vector. |
9770 | /// |
9771 | /// This is a common pattern that we have especially efficient patterns to lower |
9772 | /// across all subtarget feature sets. |
9773 | static SDValue lowerVectorShuffleAsElementInsertion( |
9774 | const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, |
9775 | const APInt &Zeroable, const X86Subtarget &Subtarget, |
9776 | SelectionDAG &DAG) { |
9777 | MVT ExtVT = VT; |
9778 | MVT EltVT = VT.getVectorElementType(); |
9779 | |
9780 | int V2Index = |
9781 | find_if(Mask, [&Mask](int M) { return M >= (int)Mask.size(); }) - |
9782 | Mask.begin(); |
9783 | bool IsV1Zeroable = true; |
9784 | for (int i = 0, Size = Mask.size(); i < Size; ++i) |
9785 | if (i != V2Index && !Zeroable[i]) { |
9786 | IsV1Zeroable = false; |
9787 | break; |
9788 | } |
9789 | |
9790 | // Check for a single input from a SCALAR_TO_VECTOR node. |
9791 | // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and |
9792 | // all the smarts here sunk into that routine. However, the current |
9793 | // lowering of BUILD_VECTOR makes that nearly impossible until the old |
9794 | // vector shuffle lowering is dead. |
9795 | SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(), |
9796 | DAG); |
9797 | if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) { |
9798 | // We need to zext the scalar if it is smaller than an i32. |
9799 | V2S = DAG.getBitcast(EltVT, V2S); |
9800 | if (EltVT == MVT::i8 || EltVT == MVT::i16) { |
9801 | // Using zext to expand a narrow element won't work for non-zero |
9802 | // insertions. |
9803 | if (!IsV1Zeroable) |
9804 | return SDValue(); |
9805 | |
9806 | // Zero-extend directly to i32. |
9807 | ExtVT = MVT::v4i32; |
9808 | V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S); |
9809 | } |
9810 | V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S); |
9811 | } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 || |
9812 | EltVT == MVT::i16) { |
9813 | // Either not inserting from the low element of the input or the input |
9814 | // element size is too small to use VZEXT_MOVL to clear the high bits. |
9815 | return SDValue(); |
9816 | } |
9817 | |
9818 | if (!IsV1Zeroable) { |
9819 | // If V1 can't be treated as a zero vector we have fewer options to lower |
9820 | // this. We can't support integer vectors or non-zero targets cheaply, and |
9821 | // the V1 elements can't be permuted in any way. |
9822 | assert(VT == ExtVT && "Cannot change extended type when non-zeroable!")((VT == ExtVT && "Cannot change extended type when non-zeroable!" ) ? static_cast<void> (0) : __assert_fail ("VT == ExtVT && \"Cannot change extended type when non-zeroable!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9822, __PRETTY_FUNCTION__)); |
9823 | if (!VT.isFloatingPoint() || V2Index != 0) |
9824 | return SDValue(); |
9825 | SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end()); |
9826 | V1Mask[V2Index] = -1; |
9827 | if (!isNoopShuffleMask(V1Mask)) |
9828 | return SDValue(); |
9829 | // This is essentially a special case blend operation, but if we have |
9830 | // general purpose blend operations, they are always faster. Bail and let |
9831 | // the rest of the lowering handle these as blends. |
9832 | if (Subtarget.hasSSE41()) |
9833 | return SDValue(); |
9834 | |
9835 | // Otherwise, use MOVSD or MOVSS. |
9836 | assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&(((EltVT == MVT::f32 || EltVT == MVT::f64) && "Only two types of floating point element types to handle!" ) ? static_cast<void> (0) : __assert_fail ("(EltVT == MVT::f32 || EltVT == MVT::f64) && \"Only two types of floating point element types to handle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9837, __PRETTY_FUNCTION__)) |
9837 | "Only two types of floating point element types to handle!")(((EltVT == MVT::f32 || EltVT == MVT::f64) && "Only two types of floating point element types to handle!" ) ? static_cast<void> (0) : __assert_fail ("(EltVT == MVT::f32 || EltVT == MVT::f64) && \"Only two types of floating point element types to handle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9837, __PRETTY_FUNCTION__)); |
9838 | return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL, |
9839 | ExtVT, V1, V2); |
9840 | } |
9841 | |
9842 | // This lowering only works for the low element with floating point vectors. |
9843 | if (VT.isFloatingPoint() && V2Index != 0) |
9844 | return SDValue(); |
9845 | |
9846 | V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2); |
9847 | if (ExtVT != VT) |
9848 | V2 = DAG.getBitcast(VT, V2); |
9849 | |
9850 | if (V2Index != 0) { |
9851 | // If we have 4 or fewer lanes we can cheaply shuffle the element into |
9852 | // the desired position. Otherwise it is more efficient to do a vector |
9853 | // shift left. We know that we can do a vector shift left because all |
9854 | // the inputs are zero. |
9855 | if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) { |
9856 | SmallVector<int, 4> V2Shuffle(Mask.size(), 1); |
9857 | V2Shuffle[V2Index] = 0; |
9858 | V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle); |
9859 | } else { |
9860 | V2 = DAG.getBitcast(MVT::v16i8, V2); |
9861 | V2 = DAG.getNode( |
9862 | X86ISD::VSHLDQ, DL, MVT::v16i8, V2, |
9863 | DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL, |
9864 | DAG.getTargetLoweringInfo().getScalarShiftAmountTy( |
9865 | DAG.getDataLayout(), VT))); |
9866 | V2 = DAG.getBitcast(VT, V2); |
9867 | } |
9868 | } |
9869 | return V2; |
9870 | } |
9871 | |
9872 | /// Try to lower broadcast of a single - truncated - integer element, |
9873 | /// coming from a scalar_to_vector/build_vector node \p V0 with larger elements. |
9874 | /// |
9875 | /// This assumes we have AVX2. |
9876 | static SDValue lowerVectorShuffleAsTruncBroadcast(const SDLoc &DL, MVT VT, |
9877 | SDValue V0, int BroadcastIdx, |
9878 | const X86Subtarget &Subtarget, |
9879 | SelectionDAG &DAG) { |
9880 | assert(Subtarget.hasAVX2() &&((Subtarget.hasAVX2() && "We can only lower integer broadcasts with AVX2!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX2() && \"We can only lower integer broadcasts with AVX2!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9881, __PRETTY_FUNCTION__)) |
9881 | "We can only lower integer broadcasts with AVX2!")((Subtarget.hasAVX2() && "We can only lower integer broadcasts with AVX2!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX2() && \"We can only lower integer broadcasts with AVX2!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9881, __PRETTY_FUNCTION__)); |
9882 | |
9883 | EVT EltVT = VT.getVectorElementType(); |
9884 | EVT V0VT = V0.getValueType(); |
9885 | |
9886 | assert(VT.isInteger() && "Unexpected non-integer trunc broadcast!")((VT.isInteger() && "Unexpected non-integer trunc broadcast!" ) ? static_cast<void> (0) : __assert_fail ("VT.isInteger() && \"Unexpected non-integer trunc broadcast!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9886, __PRETTY_FUNCTION__)); |
9887 | assert(V0VT.isVector() && "Unexpected non-vector vector-sized value!")((V0VT.isVector() && "Unexpected non-vector vector-sized value!" ) ? static_cast<void> (0) : __assert_fail ("V0VT.isVector() && \"Unexpected non-vector vector-sized value!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9887, __PRETTY_FUNCTION__)); |
9888 | |
9889 | EVT V0EltVT = V0VT.getVectorElementType(); |
9890 | if (!V0EltVT.isInteger()) |
9891 | return SDValue(); |
9892 | |
9893 | const unsigned EltSize = EltVT.getSizeInBits(); |
9894 | const unsigned V0EltSize = V0EltVT.getSizeInBits(); |
9895 | |
9896 | // This is only a truncation if the original element type is larger. |
9897 | if (V0EltSize <= EltSize) |
9898 | return SDValue(); |
9899 | |
9900 | assert(((V0EltSize % EltSize) == 0) &&((((V0EltSize % EltSize) == 0) && "Scalar type sizes must all be powers of 2 on x86!" ) ? static_cast<void> (0) : __assert_fail ("((V0EltSize % EltSize) == 0) && \"Scalar type sizes must all be powers of 2 on x86!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9901, __PRETTY_FUNCTION__)) |
9901 | "Scalar type sizes must all be powers of 2 on x86!")((((V0EltSize % EltSize) == 0) && "Scalar type sizes must all be powers of 2 on x86!" ) ? static_cast<void> (0) : __assert_fail ("((V0EltSize % EltSize) == 0) && \"Scalar type sizes must all be powers of 2 on x86!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9901, __PRETTY_FUNCTION__)); |
9902 | |
9903 | const unsigned V0Opc = V0.getOpcode(); |
9904 | const unsigned Scale = V0EltSize / EltSize; |
9905 | const unsigned V0BroadcastIdx = BroadcastIdx / Scale; |
9906 | |
9907 | if ((V0Opc != ISD::SCALAR_TO_VECTOR || V0BroadcastIdx != 0) && |
9908 | V0Opc != ISD::BUILD_VECTOR) |
9909 | return SDValue(); |
9910 | |
9911 | SDValue Scalar = V0.getOperand(V0BroadcastIdx); |
9912 | |
9913 | // If we're extracting non-least-significant bits, shift so we can truncate. |
9914 | // Hopefully, we can fold away the trunc/srl/load into the broadcast. |
9915 | // Even if we can't (and !isShuffleFoldableLoad(Scalar)), prefer |
9916 | // vpbroadcast+vmovd+shr to vpshufb(m)+vmovd. |
9917 | if (const int OffsetIdx = BroadcastIdx % Scale) |
9918 | Scalar = DAG.getNode(ISD::SRL, DL, Scalar.getValueType(), Scalar, |
9919 | DAG.getConstant(OffsetIdx * EltSize, DL, Scalar.getValueType())); |
9920 | |
9921 | return DAG.getNode(X86ISD::VBROADCAST, DL, VT, |
9922 | DAG.getNode(ISD::TRUNCATE, DL, EltVT, Scalar)); |
9923 | } |
9924 | |
9925 | /// \brief Try to lower broadcast of a single element. |
9926 | /// |
9927 | /// For convenience, this code also bundles all of the subtarget feature set |
9928 | /// filtering. While a little annoying to re-dispatch on type here, there isn't |
9929 | /// a convenient way to factor it out. |
9930 | static SDValue lowerVectorShuffleAsBroadcast(const SDLoc &DL, MVT VT, |
9931 | SDValue V1, SDValue V2, |
9932 | ArrayRef<int> Mask, |
9933 | const X86Subtarget &Subtarget, |
9934 | SelectionDAG &DAG) { |
9935 | if (!((Subtarget.hasSSE3() && VT == MVT::v2f64) || |
9936 | (Subtarget.hasAVX() && VT.isFloatingPoint()) || |
9937 | (Subtarget.hasAVX2() && VT.isInteger()))) |
9938 | return SDValue(); |
9939 | |
9940 | // With MOVDDUP (v2f64) we can broadcast from a register or a load, otherwise |
9941 | // we can only broadcast from a register with AVX2. |
9942 | unsigned NumElts = Mask.size(); |
9943 | unsigned Opcode = VT == MVT::v2f64 ? X86ISD::MOVDDUP : X86ISD::VBROADCAST; |
9944 | bool BroadcastFromReg = (Opcode == X86ISD::MOVDDUP) || Subtarget.hasAVX2(); |
9945 | |
9946 | // Check that the mask is a broadcast. |
9947 | int BroadcastIdx = -1; |
9948 | for (int i = 0; i != (int)NumElts; ++i) { |
9949 | SmallVector<int, 8> BroadcastMask(NumElts, i); |
9950 | if (isShuffleEquivalent(V1, V2, Mask, BroadcastMask)) { |
9951 | BroadcastIdx = i; |
9952 | break; |
9953 | } |
9954 | } |
9955 | |
9956 | if (BroadcastIdx < 0) |
9957 | return SDValue(); |
9958 | assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "((BroadcastIdx < (int)Mask.size() && "We only expect to be called with " "a sorted mask where the broadcast " "comes from V1.") ? static_cast <void> (0) : __assert_fail ("BroadcastIdx < (int)Mask.size() && \"We only expect to be called with \" \"a sorted mask where the broadcast \" \"comes from V1.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9960, __PRETTY_FUNCTION__)) |
9959 | "a sorted mask where the broadcast "((BroadcastIdx < (int)Mask.size() && "We only expect to be called with " "a sorted mask where the broadcast " "comes from V1.") ? static_cast <void> (0) : __assert_fail ("BroadcastIdx < (int)Mask.size() && \"We only expect to be called with \" \"a sorted mask where the broadcast \" \"comes from V1.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9960, __PRETTY_FUNCTION__)) |
9960 | "comes from V1.")((BroadcastIdx < (int)Mask.size() && "We only expect to be called with " "a sorted mask where the broadcast " "comes from V1.") ? static_cast <void> (0) : __assert_fail ("BroadcastIdx < (int)Mask.size() && \"We only expect to be called with \" \"a sorted mask where the broadcast \" \"comes from V1.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 9960, __PRETTY_FUNCTION__)); |
9961 | |
9962 | // Go up the chain of (vector) values to find a scalar load that we can |
9963 | // combine with the broadcast. |
9964 | SDValue V = V1; |
9965 | for (;;) { |
9966 | switch (V.getOpcode()) { |
9967 | case ISD::BITCAST: { |
9968 | SDValue VSrc = V.getOperand(0); |
9969 | MVT SrcVT = VSrc.getSimpleValueType(); |
9970 | if (VT.getScalarSizeInBits() != SrcVT.getScalarSizeInBits()) |
9971 | break; |
9972 | V = VSrc; |
9973 | continue; |
9974 | } |
9975 | case ISD::CONCAT_VECTORS: { |
9976 | int OperandSize = Mask.size() / V.getNumOperands(); |
9977 | V = V.getOperand(BroadcastIdx / OperandSize); |
9978 | BroadcastIdx %= OperandSize; |
9979 | continue; |
9980 | } |
9981 | case ISD::INSERT_SUBVECTOR: { |
9982 | SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1); |
9983 | auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2)); |
9984 | if (!ConstantIdx) |
9985 | break; |
9986 | |
9987 | int BeginIdx = (int)ConstantIdx->getZExtValue(); |
9988 | int EndIdx = |
9989 | BeginIdx + (int)VInner.getSimpleValueType().getVectorNumElements(); |
9990 | if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) { |
9991 | BroadcastIdx -= BeginIdx; |
9992 | V = VInner; |
9993 | } else { |
9994 | V = VOuter; |
9995 | } |
9996 | continue; |
9997 | } |
9998 | } |
9999 | break; |
10000 | } |
10001 | |
10002 | // Check if this is a broadcast of a scalar. We special case lowering |
10003 | // for scalars so that we can more effectively fold with loads. |
10004 | // First, look through bitcast: if the original value has a larger element |
10005 | // type than the shuffle, the broadcast element is in essence truncated. |
10006 | // Make that explicit to ease folding. |
10007 | if (V.getOpcode() == ISD::BITCAST && VT.isInteger()) |
10008 | if (SDValue TruncBroadcast = lowerVectorShuffleAsTruncBroadcast( |
10009 | DL, VT, V.getOperand(0), BroadcastIdx, Subtarget, DAG)) |
10010 | return TruncBroadcast; |
10011 | |
10012 | MVT BroadcastVT = VT; |
10013 | |
10014 | // Peek through any bitcast (only useful for loads). |
10015 | SDValue BC = peekThroughBitcasts(V); |
10016 | |
10017 | // Also check the simpler case, where we can directly reuse the scalar. |
10018 | if (V.getOpcode() == ISD::BUILD_VECTOR || |
10019 | (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) { |
10020 | V = V.getOperand(BroadcastIdx); |
10021 | |
10022 | // If we can't broadcast from a register, check that the input is a load. |
10023 | if (!BroadcastFromReg && !isShuffleFoldableLoad(V)) |
10024 | return SDValue(); |
10025 | } else if (MayFoldLoad(BC) && !cast<LoadSDNode>(BC)->isVolatile()) { |
10026 | // 32-bit targets need to load i64 as a f64 and then bitcast the result. |
10027 | if (!Subtarget.is64Bit() && VT.getScalarType() == MVT::i64) { |
10028 | BroadcastVT = MVT::getVectorVT(MVT::f64, VT.getVectorNumElements()); |
10029 | Opcode = (BroadcastVT.is128BitVector() ? X86ISD::MOVDDUP : Opcode); |
10030 | } |
10031 | |
10032 | // If we are broadcasting a load that is only used by the shuffle |
10033 | // then we can reduce the vector load to the broadcasted scalar load. |
10034 | LoadSDNode *Ld = cast<LoadSDNode>(BC); |
10035 | SDValue BaseAddr = Ld->getOperand(1); |
10036 | EVT SVT = BroadcastVT.getScalarType(); |
10037 | unsigned Offset = BroadcastIdx * SVT.getStoreSize(); |
10038 | SDValue NewAddr = DAG.getMemBasePlusOffset(BaseAddr, Offset, DL); |
10039 | V = DAG.getLoad(SVT, DL, Ld->getChain(), NewAddr, |
10040 | DAG.getMachineFunction().getMachineMemOperand( |
10041 | Ld->getMemOperand(), Offset, SVT.getStoreSize())); |
10042 | DAG.makeEquivalentMemoryOrdering(Ld, V); |
10043 | } else if (!BroadcastFromReg) { |
10044 | // We can't broadcast from a vector register. |
10045 | return SDValue(); |
10046 | } else if (BroadcastIdx != 0) { |
10047 | // We can only broadcast from the zero-element of a vector register, |
10048 | // but it can be advantageous to broadcast from the zero-element of a |
10049 | // subvector. |
10050 | if (!VT.is256BitVector() && !VT.is512BitVector()) |
10051 | return SDValue(); |
10052 | |
10053 | // VPERMQ/VPERMPD can perform the cross-lane shuffle directly. |
10054 | if (VT == MVT::v4f64 || VT == MVT::v4i64) |
10055 | return SDValue(); |
10056 | |
10057 | // Only broadcast the zero-element of a 128-bit subvector. |
10058 | unsigned EltSize = VT.getScalarSizeInBits(); |
10059 | if (((BroadcastIdx * EltSize) % 128) != 0) |
10060 | return SDValue(); |
10061 | |
10062 | // The shuffle input might have been a bitcast we looked through; look at |
10063 | // the original input vector. Emit an EXTRACT_SUBVECTOR of that type; we'll |
10064 | // later bitcast it to BroadcastVT. |
10065 | MVT SrcVT = V.getSimpleValueType(); |
10066 | assert(SrcVT.getScalarSizeInBits() == BroadcastVT.getScalarSizeInBits() &&((SrcVT.getScalarSizeInBits() == BroadcastVT.getScalarSizeInBits () && "Unexpected vector element size") ? static_cast <void> (0) : __assert_fail ("SrcVT.getScalarSizeInBits() == BroadcastVT.getScalarSizeInBits() && \"Unexpected vector element size\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10067, __PRETTY_FUNCTION__)) |
10067 | "Unexpected vector element size")((SrcVT.getScalarSizeInBits() == BroadcastVT.getScalarSizeInBits () && "Unexpected vector element size") ? static_cast <void> (0) : __assert_fail ("SrcVT.getScalarSizeInBits() == BroadcastVT.getScalarSizeInBits() && \"Unexpected vector element size\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10067, __PRETTY_FUNCTION__)); |
10068 | assert((SrcVT.is256BitVector() || SrcVT.is512BitVector()) &&(((SrcVT.is256BitVector() || SrcVT.is512BitVector()) && "Unexpected vector size") ? static_cast<void> (0) : __assert_fail ("(SrcVT.is256BitVector() || SrcVT.is512BitVector()) && \"Unexpected vector size\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10069, __PRETTY_FUNCTION__)) |
10069 | "Unexpected vector size")(((SrcVT.is256BitVector() || SrcVT.is512BitVector()) && "Unexpected vector size") ? static_cast<void> (0) : __assert_fail ("(SrcVT.is256BitVector() || SrcVT.is512BitVector()) && \"Unexpected vector size\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10069, __PRETTY_FUNCTION__)); |
10070 | |
10071 | MVT ExtVT = MVT::getVectorVT(SrcVT.getScalarType(), 128 / EltSize); |
10072 | V = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ExtVT, V, |
10073 | DAG.getIntPtrConstant(BroadcastIdx, DL)); |
10074 | } |
10075 | |
10076 | if (Opcode == X86ISD::MOVDDUP && !V.getValueType().isVector()) |
10077 | V = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, |
10078 | DAG.getBitcast(MVT::f64, V)); |
10079 | |
10080 | // Bitcast back to the same scalar type as BroadcastVT. |
10081 | MVT SrcVT = V.getSimpleValueType(); |
10082 | if (SrcVT.getScalarType() != BroadcastVT.getScalarType()) { |
10083 | assert(SrcVT.getScalarSizeInBits() == BroadcastVT.getScalarSizeInBits() &&((SrcVT.getScalarSizeInBits() == BroadcastVT.getScalarSizeInBits () && "Unexpected vector element size") ? static_cast <void> (0) : __assert_fail ("SrcVT.getScalarSizeInBits() == BroadcastVT.getScalarSizeInBits() && \"Unexpected vector element size\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10084, __PRETTY_FUNCTION__)) |
10084 | "Unexpected vector element size")((SrcVT.getScalarSizeInBits() == BroadcastVT.getScalarSizeInBits () && "Unexpected vector element size") ? static_cast <void> (0) : __assert_fail ("SrcVT.getScalarSizeInBits() == BroadcastVT.getScalarSizeInBits() && \"Unexpected vector element size\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10084, __PRETTY_FUNCTION__)); |
10085 | if (SrcVT.isVector()) { |
10086 | unsigned NumSrcElts = SrcVT.getVectorNumElements(); |
10087 | SrcVT = MVT::getVectorVT(BroadcastVT.getScalarType(), NumSrcElts); |
10088 | } else { |
10089 | SrcVT = BroadcastVT.getScalarType(); |
10090 | } |
10091 | V = DAG.getBitcast(SrcVT, V); |
10092 | } |
10093 | |
10094 | // 32-bit targets need to load i64 as a f64 and then bitcast the result. |
10095 | if (!Subtarget.is64Bit() && SrcVT == MVT::i64) { |
10096 | V = DAG.getBitcast(MVT::f64, V); |
10097 | unsigned NumBroadcastElts = BroadcastVT.getVectorNumElements(); |
10098 | BroadcastVT = MVT::getVectorVT(MVT::f64, NumBroadcastElts); |
10099 | } |
10100 | |
10101 | // We only support broadcasting from 128-bit vectors to minimize the |
10102 | // number of patterns we need to deal with in isel. So extract down to |
10103 | // 128-bits. |
10104 | if (SrcVT.getSizeInBits() > 128) |
10105 | V = extract128BitVector(V, 0, DAG, DL); |
10106 | |
10107 | return DAG.getBitcast(VT, DAG.getNode(Opcode, DL, BroadcastVT, V)); |
10108 | } |
10109 | |
10110 | // Check for whether we can use INSERTPS to perform the shuffle. We only use |
10111 | // INSERTPS when the V1 elements are already in the correct locations |
10112 | // because otherwise we can just always use two SHUFPS instructions which |
10113 | // are much smaller to encode than a SHUFPS and an INSERTPS. We can also |
10114 | // perform INSERTPS if a single V1 element is out of place and all V2 |
10115 | // elements are zeroable. |
10116 | static bool matchVectorShuffleAsInsertPS(SDValue &V1, SDValue &V2, |
10117 | unsigned &InsertPSMask, |
10118 | const APInt &Zeroable, |
10119 | ArrayRef<int> Mask, |
10120 | SelectionDAG &DAG) { |
10121 | assert(V1.getSimpleValueType().is128BitVector() && "Bad operand type!")((V1.getSimpleValueType().is128BitVector() && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType().is128BitVector() && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10121, __PRETTY_FUNCTION__)); |
10122 | assert(V2.getSimpleValueType().is128BitVector() && "Bad operand type!")((V2.getSimpleValueType().is128BitVector() && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType().is128BitVector() && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10122, __PRETTY_FUNCTION__)); |
10123 | assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!")((Mask.size() == 4 && "Unexpected mask size for v4 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 4 && \"Unexpected mask size for v4 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10123, __PRETTY_FUNCTION__)); |
10124 | |
10125 | // Attempt to match INSERTPS with one element from VA or VB being |
10126 | // inserted into VA (or undef). If successful, V1, V2 and InsertPSMask |
10127 | // are updated. |
10128 | auto matchAsInsertPS = [&](SDValue VA, SDValue VB, |
10129 | ArrayRef<int> CandidateMask) { |
10130 | unsigned ZMask = 0; |
10131 | int VADstIndex = -1; |
10132 | int VBDstIndex = -1; |
10133 | bool VAUsedInPlace = false; |
10134 | |
10135 | for (int i = 0; i < 4; ++i) { |
10136 | // Synthesize a zero mask from the zeroable elements (includes undefs). |
10137 | if (Zeroable[i]) { |
10138 | ZMask |= 1 << i; |
10139 | continue; |
10140 | } |
10141 | |
10142 | // Flag if we use any VA inputs in place. |
10143 | if (i == CandidateMask[i]) { |
10144 | VAUsedInPlace = true; |
10145 | continue; |
10146 | } |
10147 | |
10148 | // We can only insert a single non-zeroable element. |
10149 | if (VADstIndex >= 0 || VBDstIndex >= 0) |
10150 | return false; |
10151 | |
10152 | if (CandidateMask[i] < 4) { |
10153 | // VA input out of place for insertion. |
10154 | VADstIndex = i; |
10155 | } else { |
10156 | // VB input for insertion. |
10157 | VBDstIndex = i; |
10158 | } |
10159 | } |
10160 | |
10161 | // Don't bother if we have no (non-zeroable) element for insertion. |
10162 | if (VADstIndex < 0 && VBDstIndex < 0) |
10163 | return false; |
10164 | |
10165 | // Determine element insertion src/dst indices. The src index is from the |
10166 | // start of the inserted vector, not the start of the concatenated vector. |
10167 | unsigned VBSrcIndex = 0; |
10168 | if (VADstIndex >= 0) { |
10169 | // If we have a VA input out of place, we use VA as the V2 element |
10170 | // insertion and don't use the original V2 at all. |
10171 | VBSrcIndex = CandidateMask[VADstIndex]; |
10172 | VBDstIndex = VADstIndex; |
10173 | VB = VA; |
10174 | } else { |
10175 | VBSrcIndex = CandidateMask[VBDstIndex] - 4; |
10176 | } |
10177 | |
10178 | // If no V1 inputs are used in place, then the result is created only from |
10179 | // the zero mask and the V2 insertion - so remove V1 dependency. |
10180 | if (!VAUsedInPlace) |
10181 | VA = DAG.getUNDEF(MVT::v4f32); |
10182 | |
10183 | // Update V1, V2 and InsertPSMask accordingly. |
10184 | V1 = VA; |
10185 | V2 = VB; |
10186 | |
10187 | // Insert the V2 element into the desired position. |
10188 | InsertPSMask = VBSrcIndex << 6 | VBDstIndex << 4 | ZMask; |
10189 | assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!")(((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!" ) ? static_cast<void> (0) : __assert_fail ("(InsertPSMask & ~0xFFu) == 0 && \"Invalid mask!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10189, __PRETTY_FUNCTION__)); |
10190 | return true; |
10191 | }; |
10192 | |
10193 | if (matchAsInsertPS(V1, V2, Mask)) |
10194 | return true; |
10195 | |
10196 | // Commute and try again. |
10197 | SmallVector<int, 4> CommutedMask(Mask.begin(), Mask.end()); |
10198 | ShuffleVectorSDNode::commuteMask(CommutedMask); |
10199 | if (matchAsInsertPS(V2, V1, CommutedMask)) |
10200 | return true; |
10201 | |
10202 | return false; |
10203 | } |
10204 | |
10205 | static SDValue lowerVectorShuffleAsInsertPS(const SDLoc &DL, SDValue V1, |
10206 | SDValue V2, ArrayRef<int> Mask, |
10207 | const APInt &Zeroable, |
10208 | SelectionDAG &DAG) { |
10209 | assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v4f32 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10209, __PRETTY_FUNCTION__)); |
10210 | assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v4f32 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10210, __PRETTY_FUNCTION__)); |
10211 | |
10212 | // Attempt to match the insertps pattern. |
10213 | unsigned InsertPSMask; |
10214 | if (!matchVectorShuffleAsInsertPS(V1, V2, InsertPSMask, Zeroable, Mask, DAG)) |
10215 | return SDValue(); |
10216 | |
10217 | // Insert the V2 element into the desired position. |
10218 | return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2, |
10219 | DAG.getConstant(InsertPSMask, DL, MVT::i8)); |
10220 | } |
10221 | |
10222 | /// \brief Try to lower a shuffle as a permute of the inputs followed by an |
10223 | /// UNPCK instruction. |
10224 | /// |
10225 | /// This specifically targets cases where we end up with alternating between |
10226 | /// the two inputs, and so can permute them into something that feeds a single |
10227 | /// UNPCK instruction. Note that this routine only targets integer vectors |
10228 | /// because for floating point vectors we have a generalized SHUFPS lowering |
10229 | /// strategy that handles everything that doesn't *exactly* match an unpack, |
10230 | /// making this clever lowering unnecessary. |
10231 | static SDValue lowerVectorShuffleAsPermuteAndUnpack(const SDLoc &DL, MVT VT, |
10232 | SDValue V1, SDValue V2, |
10233 | ArrayRef<int> Mask, |
10234 | SelectionDAG &DAG) { |
10235 | assert(!VT.isFloatingPoint() &&((!VT.isFloatingPoint() && "This routine only supports integer vectors." ) ? static_cast<void> (0) : __assert_fail ("!VT.isFloatingPoint() && \"This routine only supports integer vectors.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10236, __PRETTY_FUNCTION__)) |
10236 | "This routine only supports integer vectors.")((!VT.isFloatingPoint() && "This routine only supports integer vectors." ) ? static_cast<void> (0) : __assert_fail ("!VT.isFloatingPoint() && \"This routine only supports integer vectors.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10236, __PRETTY_FUNCTION__)); |
10237 | assert(VT.is128BitVector() &&((VT.is128BitVector() && "This routine only works on 128-bit vectors." ) ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"This routine only works on 128-bit vectors.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10238, __PRETTY_FUNCTION__)) |
10238 | "This routine only works on 128-bit vectors.")((VT.is128BitVector() && "This routine only works on 128-bit vectors." ) ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"This routine only works on 128-bit vectors.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10238, __PRETTY_FUNCTION__)); |
10239 | assert(!V2.isUndef() &&((!V2.isUndef() && "This routine should only be used when blending two inputs." ) ? static_cast<void> (0) : __assert_fail ("!V2.isUndef() && \"This routine should only be used when blending two inputs.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10240, __PRETTY_FUNCTION__)) |
10240 | "This routine should only be used when blending two inputs.")((!V2.isUndef() && "This routine should only be used when blending two inputs." ) ? static_cast<void> (0) : __assert_fail ("!V2.isUndef() && \"This routine should only be used when blending two inputs.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10240, __PRETTY_FUNCTION__)); |
10241 | assert(Mask.size() >= 2 && "Single element masks are invalid.")((Mask.size() >= 2 && "Single element masks are invalid." ) ? static_cast<void> (0) : __assert_fail ("Mask.size() >= 2 && \"Single element masks are invalid.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10241, __PRETTY_FUNCTION__)); |
10242 | |
10243 | int Size = Mask.size(); |
10244 | |
10245 | int NumLoInputs = |
10246 | count_if(Mask, [Size](int M) { return M >= 0 && M % Size < Size / 2; }); |
10247 | int NumHiInputs = |
10248 | count_if(Mask, [Size](int M) { return M % Size >= Size / 2; }); |
10249 | |
10250 | bool UnpackLo = NumLoInputs >= NumHiInputs; |
10251 | |
10252 | auto TryUnpack = [&](int ScalarSize, int Scale) { |
10253 | SmallVector<int, 16> V1Mask((unsigned)Size, -1); |
10254 | SmallVector<int, 16> V2Mask((unsigned)Size, -1); |
10255 | |
10256 | for (int i = 0; i < Size; ++i) { |
10257 | if (Mask[i] < 0) |
10258 | continue; |
10259 | |
10260 | // Each element of the unpack contains Scale elements from this mask. |
10261 | int UnpackIdx = i / Scale; |
10262 | |
10263 | // We only handle the case where V1 feeds the first slots of the unpack. |
10264 | // We rely on canonicalization to ensure this is the case. |
10265 | if ((UnpackIdx % 2 == 0) != (Mask[i] < Size)) |
10266 | return SDValue(); |
10267 | |
10268 | // Setup the mask for this input. The indexing is tricky as we have to |
10269 | // handle the unpack stride. |
10270 | SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask; |
10271 | VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] = |
10272 | Mask[i] % Size; |
10273 | } |
10274 | |
10275 | // If we will have to shuffle both inputs to use the unpack, check whether |
10276 | // we can just unpack first and shuffle the result. If so, skip this unpack. |
10277 | if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) && |
10278 | !isNoopShuffleMask(V2Mask)) |
10279 | return SDValue(); |
10280 | |
10281 | // Shuffle the inputs into place. |
10282 | V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask); |
10283 | V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask); |
10284 | |
10285 | // Cast the inputs to the type we will use to unpack them. |
10286 | MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), Size / Scale); |
10287 | V1 = DAG.getBitcast(UnpackVT, V1); |
10288 | V2 = DAG.getBitcast(UnpackVT, V2); |
10289 | |
10290 | // Unpack the inputs and cast the result back to the desired type. |
10291 | return DAG.getBitcast( |
10292 | VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL, |
10293 | UnpackVT, V1, V2)); |
10294 | }; |
10295 | |
10296 | // We try each unpack from the largest to the smallest to try and find one |
10297 | // that fits this mask. |
10298 | int OrigScalarSize = VT.getScalarSizeInBits(); |
10299 | for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) |
10300 | if (SDValue Unpack = TryUnpack(ScalarSize, ScalarSize / OrigScalarSize)) |
10301 | return Unpack; |
10302 | |
10303 | // If none of the unpack-rooted lowerings worked (or were profitable) try an |
10304 | // initial unpack. |
10305 | if (NumLoInputs == 0 || NumHiInputs == 0) { |
10306 | assert((NumLoInputs > 0 || NumHiInputs > 0) &&(((NumLoInputs > 0 || NumHiInputs > 0) && "We have to have *some* inputs!" ) ? static_cast<void> (0) : __assert_fail ("(NumLoInputs > 0 || NumHiInputs > 0) && \"We have to have *some* inputs!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10307, __PRETTY_FUNCTION__)) |
10307 | "We have to have *some* inputs!")(((NumLoInputs > 0 || NumHiInputs > 0) && "We have to have *some* inputs!" ) ? static_cast<void> (0) : __assert_fail ("(NumLoInputs > 0 || NumHiInputs > 0) && \"We have to have *some* inputs!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10307, __PRETTY_FUNCTION__)); |
10308 | int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0; |
10309 | |
10310 | // FIXME: We could consider the total complexity of the permute of each |
10311 | // possible unpacking. Or at the least we should consider how many |
10312 | // half-crossings are created. |
10313 | // FIXME: We could consider commuting the unpacks. |
10314 | |
10315 | SmallVector<int, 32> PermMask((unsigned)Size, -1); |
10316 | for (int i = 0; i < Size; ++i) { |
10317 | if (Mask[i] < 0) |
10318 | continue; |
10319 | |
10320 | assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!")((Mask[i] % Size >= HalfOffset && "Found input from wrong half!" ) ? static_cast<void> (0) : __assert_fail ("Mask[i] % Size >= HalfOffset && \"Found input from wrong half!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10320, __PRETTY_FUNCTION__)); |
10321 | |
10322 | PermMask[i] = |
10323 | 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1); |
10324 | } |
10325 | return DAG.getVectorShuffle( |
10326 | VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL, |
10327 | DL, VT, V1, V2), |
10328 | DAG.getUNDEF(VT), PermMask); |
10329 | } |
10330 | |
10331 | return SDValue(); |
10332 | } |
10333 | |
10334 | /// \brief Handle lowering of 2-lane 64-bit floating point shuffles. |
10335 | /// |
10336 | /// This is the basis function for the 2-lane 64-bit shuffles as we have full |
10337 | /// support for floating point shuffles but not integer shuffles. These |
10338 | /// instructions will incur a domain crossing penalty on some chips though so |
10339 | /// it is better to avoid lowering through this for integer vectors where |
10340 | /// possible. |
10341 | static SDValue lowerV2F64VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
10342 | const APInt &Zeroable, |
10343 | SDValue V1, SDValue V2, |
10344 | const X86Subtarget &Subtarget, |
10345 | SelectionDAG &DAG) { |
10346 | assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v2f64 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10346, __PRETTY_FUNCTION__)); |
10347 | assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v2f64 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10347, __PRETTY_FUNCTION__)); |
10348 | assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!")((Mask.size() == 2 && "Unexpected mask size for v2 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 2 && \"Unexpected mask size for v2 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10348, __PRETTY_FUNCTION__)); |
10349 | |
10350 | if (V2.isUndef()) { |
10351 | // Check for being able to broadcast a single element. |
10352 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast( |
10353 | DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG)) |
10354 | return Broadcast; |
10355 | |
10356 | // Straight shuffle of a single input vector. Simulate this by using the |
10357 | // single input as both of the "inputs" to this instruction.. |
10358 | unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1); |
10359 | |
10360 | if (Subtarget.hasAVX()) { |
10361 | // If we have AVX, we can use VPERMILPS which will allow folding a load |
10362 | // into the shuffle. |
10363 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1, |
10364 | DAG.getConstant(SHUFPDMask, DL, MVT::i8)); |
10365 | } |
10366 | |
10367 | return DAG.getNode( |
10368 | X86ISD::SHUFP, DL, MVT::v2f64, |
10369 | Mask[0] == SM_SentinelUndef ? DAG.getUNDEF(MVT::v2f64) : V1, |
10370 | Mask[1] == SM_SentinelUndef ? DAG.getUNDEF(MVT::v2f64) : V1, |
10371 | DAG.getConstant(SHUFPDMask, DL, MVT::i8)); |
10372 | } |
10373 | assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!")((Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!" ) ? static_cast<void> (0) : __assert_fail ("Mask[0] >= 0 && Mask[0] < 2 && \"Non-canonicalized blend!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10373, __PRETTY_FUNCTION__)); |
10374 | assert(Mask[1] >= 2 && "Non-canonicalized blend!")((Mask[1] >= 2 && "Non-canonicalized blend!") ? static_cast <void> (0) : __assert_fail ("Mask[1] >= 2 && \"Non-canonicalized blend!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10374, __PRETTY_FUNCTION__)); |
10375 | |
10376 | // If we have a single input, insert that into V1 if we can do so cheaply. |
10377 | if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) { |
10378 | if (SDValue Insertion = lowerVectorShuffleAsElementInsertion( |
10379 | DL, MVT::v2f64, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
10380 | return Insertion; |
10381 | // Try inverting the insertion since for v2 masks it is easy to do and we |
10382 | // can't reliably sort the mask one way or the other. |
10383 | int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2), |
10384 | Mask[1] < 0 ? -1 : (Mask[1] ^ 2)}; |
10385 | if (SDValue Insertion = lowerVectorShuffleAsElementInsertion( |
10386 | DL, MVT::v2f64, V2, V1, InverseMask, Zeroable, Subtarget, DAG)) |
10387 | return Insertion; |
10388 | } |
10389 | |
10390 | // Try to use one of the special instruction patterns to handle two common |
10391 | // blend patterns if a zero-blend above didn't work. |
10392 | if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) || |
10393 | isShuffleEquivalent(V1, V2, Mask, {1, 3})) |
10394 | if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG)) |
10395 | // We can either use a special instruction to load over the low double or |
10396 | // to move just the low double. |
10397 | return DAG.getNode( |
10398 | isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD, |
10399 | DL, MVT::v2f64, V2, |
10400 | DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S)); |
10401 | |
10402 | if (Subtarget.hasSSE41()) |
10403 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask, |
10404 | Zeroable, Subtarget, DAG)) |
10405 | return Blend; |
10406 | |
10407 | // Use dedicated unpack instructions for masks that match their pattern. |
10408 | if (SDValue V = |
10409 | lowerVectorShuffleWithUNPCK(DL, MVT::v2f64, Mask, V1, V2, DAG)) |
10410 | return V; |
10411 | |
10412 | unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1); |
10413 | return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2, |
10414 | DAG.getConstant(SHUFPDMask, DL, MVT::i8)); |
10415 | } |
10416 | |
10417 | /// \brief Handle lowering of 2-lane 64-bit integer shuffles. |
10418 | /// |
10419 | /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by |
10420 | /// the integer unit to minimize domain crossing penalties. However, for blends |
10421 | /// it falls back to the floating point shuffle operation with appropriate bit |
10422 | /// casting. |
10423 | static SDValue lowerV2I64VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
10424 | const APInt &Zeroable, |
10425 | SDValue V1, SDValue V2, |
10426 | const X86Subtarget &Subtarget, |
10427 | SelectionDAG &DAG) { |
10428 | assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v2i64 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10428, __PRETTY_FUNCTION__)); |
10429 | assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v2i64 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10429, __PRETTY_FUNCTION__)); |
10430 | assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!")((Mask.size() == 2 && "Unexpected mask size for v2 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 2 && \"Unexpected mask size for v2 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10430, __PRETTY_FUNCTION__)); |
10431 | |
10432 | if (V2.isUndef()) { |
10433 | // Check for being able to broadcast a single element. |
10434 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast( |
10435 | DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG)) |
10436 | return Broadcast; |
10437 | |
10438 | // Straight shuffle of a single input vector. For everything from SSE2 |
10439 | // onward this has a single fast instruction with no scary immediates. |
10440 | // We have to map the mask as it is actually a v4i32 shuffle instruction. |
10441 | V1 = DAG.getBitcast(MVT::v4i32, V1); |
10442 | int WidenedMask[4] = { |
10443 | std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1, |
10444 | std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1}; |
10445 | return DAG.getBitcast( |
10446 | MVT::v2i64, |
10447 | DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1, |
10448 | getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG))); |
10449 | } |
10450 | assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!")((Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!" ) ? static_cast<void> (0) : __assert_fail ("Mask[0] != -1 && \"No undef lanes in multi-input v2 shuffles!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10450, __PRETTY_FUNCTION__)); |
10451 | assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!")((Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!" ) ? static_cast<void> (0) : __assert_fail ("Mask[1] != -1 && \"No undef lanes in multi-input v2 shuffles!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10451, __PRETTY_FUNCTION__)); |
10452 | assert(Mask[0] < 2 && "We sort V1 to be the first input.")((Mask[0] < 2 && "We sort V1 to be the first input." ) ? static_cast<void> (0) : __assert_fail ("Mask[0] < 2 && \"We sort V1 to be the first input.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10452, __PRETTY_FUNCTION__)); |
10453 | assert(Mask[1] >= 2 && "We sort V2 to be the second input.")((Mask[1] >= 2 && "We sort V2 to be the second input." ) ? static_cast<void> (0) : __assert_fail ("Mask[1] >= 2 && \"We sort V2 to be the second input.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10453, __PRETTY_FUNCTION__)); |
10454 | |
10455 | // If we have a blend of two same-type PACKUS operations and the blend aligns |
10456 | // with the low and high halves, we can just merge the PACKUS operations. |
10457 | // This is particularly important as it lets us merge shuffles that this |
10458 | // routine itself creates. |
10459 | auto GetPackNode = [](SDValue V) { |
10460 | V = peekThroughBitcasts(V); |
10461 | return V.getOpcode() == X86ISD::PACKUS ? V : SDValue(); |
10462 | }; |
10463 | if (SDValue V1Pack = GetPackNode(V1)) |
10464 | if (SDValue V2Pack = GetPackNode(V2)) { |
10465 | EVT PackVT = V1Pack.getValueType(); |
10466 | if (PackVT == V2Pack.getValueType()) |
10467 | return DAG.getBitcast(MVT::v2i64, |
10468 | DAG.getNode(X86ISD::PACKUS, DL, PackVT, |
10469 | Mask[0] == 0 ? V1Pack.getOperand(0) |
10470 | : V1Pack.getOperand(1), |
10471 | Mask[1] == 2 ? V2Pack.getOperand(0) |
10472 | : V2Pack.getOperand(1))); |
10473 | } |
10474 | |
10475 | // Try to use shift instructions. |
10476 | if (SDValue Shift = lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, |
10477 | Zeroable, Subtarget, DAG)) |
10478 | return Shift; |
10479 | |
10480 | // When loading a scalar and then shuffling it into a vector we can often do |
10481 | // the insertion cheaply. |
10482 | if (SDValue Insertion = lowerVectorShuffleAsElementInsertion( |
10483 | DL, MVT::v2i64, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
10484 | return Insertion; |
10485 | // Try inverting the insertion since for v2 masks it is easy to do and we |
10486 | // can't reliably sort the mask one way or the other. |
10487 | int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2}; |
10488 | if (SDValue Insertion = lowerVectorShuffleAsElementInsertion( |
10489 | DL, MVT::v2i64, V2, V1, InverseMask, Zeroable, Subtarget, DAG)) |
10490 | return Insertion; |
10491 | |
10492 | // We have different paths for blend lowering, but they all must use the |
10493 | // *exact* same predicate. |
10494 | bool IsBlendSupported = Subtarget.hasSSE41(); |
10495 | if (IsBlendSupported) |
10496 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask, |
10497 | Zeroable, Subtarget, DAG)) |
10498 | return Blend; |
10499 | |
10500 | // Use dedicated unpack instructions for masks that match their pattern. |
10501 | if (SDValue V = |
10502 | lowerVectorShuffleWithUNPCK(DL, MVT::v2i64, Mask, V1, V2, DAG)) |
10503 | return V; |
10504 | |
10505 | // Try to use byte rotation instructions. |
10506 | // Its more profitable for pre-SSSE3 to use shuffles/unpacks. |
10507 | if (Subtarget.hasSSSE3()) |
10508 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate( |
10509 | DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG)) |
10510 | return Rotate; |
10511 | |
10512 | // If we have direct support for blends, we should lower by decomposing into |
10513 | // a permute. That will be faster than the domain cross. |
10514 | if (IsBlendSupported) |
10515 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2, |
10516 | Mask, DAG); |
10517 | |
10518 | // We implement this with SHUFPD which is pretty lame because it will likely |
10519 | // incur 2 cycles of stall for integer vectors on Nehalem and older chips. |
10520 | // However, all the alternatives are still more cycles and newer chips don't |
10521 | // have this problem. It would be really nice if x86 had better shuffles here. |
10522 | V1 = DAG.getBitcast(MVT::v2f64, V1); |
10523 | V2 = DAG.getBitcast(MVT::v2f64, V2); |
10524 | return DAG.getBitcast(MVT::v2i64, |
10525 | DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask)); |
10526 | } |
10527 | |
10528 | /// \brief Test whether this can be lowered with a single SHUFPS instruction. |
10529 | /// |
10530 | /// This is used to disable more specialized lowerings when the shufps lowering |
10531 | /// will happen to be efficient. |
10532 | static bool isSingleSHUFPSMask(ArrayRef<int> Mask) { |
10533 | // This routine only handles 128-bit shufps. |
10534 | assert(Mask.size() == 4 && "Unsupported mask size!")((Mask.size() == 4 && "Unsupported mask size!") ? static_cast <void> (0) : __assert_fail ("Mask.size() == 4 && \"Unsupported mask size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10534, __PRETTY_FUNCTION__)); |
10535 | assert(Mask[0] >= -1 && Mask[0] < 8 && "Out of bound mask element!")((Mask[0] >= -1 && Mask[0] < 8 && "Out of bound mask element!" ) ? static_cast<void> (0) : __assert_fail ("Mask[0] >= -1 && Mask[0] < 8 && \"Out of bound mask element!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10535, __PRETTY_FUNCTION__)); |
10536 | assert(Mask[1] >= -1 && Mask[1] < 8 && "Out of bound mask element!")((Mask[1] >= -1 && Mask[1] < 8 && "Out of bound mask element!" ) ? static_cast<void> (0) : __assert_fail ("Mask[1] >= -1 && Mask[1] < 8 && \"Out of bound mask element!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10536, __PRETTY_FUNCTION__)); |
10537 | assert(Mask[2] >= -1 && Mask[2] < 8 && "Out of bound mask element!")((Mask[2] >= -1 && Mask[2] < 8 && "Out of bound mask element!" ) ? static_cast<void> (0) : __assert_fail ("Mask[2] >= -1 && Mask[2] < 8 && \"Out of bound mask element!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10537, __PRETTY_FUNCTION__)); |
10538 | assert(Mask[3] >= -1 && Mask[3] < 8 && "Out of bound mask element!")((Mask[3] >= -1 && Mask[3] < 8 && "Out of bound mask element!" ) ? static_cast<void> (0) : __assert_fail ("Mask[3] >= -1 && Mask[3] < 8 && \"Out of bound mask element!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10538, __PRETTY_FUNCTION__)); |
10539 | |
10540 | // To lower with a single SHUFPS we need to have the low half and high half |
10541 | // each requiring a single input. |
10542 | if (Mask[0] >= 0 && Mask[1] >= 0 && (Mask[0] < 4) != (Mask[1] < 4)) |
10543 | return false; |
10544 | if (Mask[2] >= 0 && Mask[3] >= 0 && (Mask[2] < 4) != (Mask[3] < 4)) |
10545 | return false; |
10546 | |
10547 | return true; |
10548 | } |
10549 | |
10550 | /// \brief Lower a vector shuffle using the SHUFPS instruction. |
10551 | /// |
10552 | /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS. |
10553 | /// It makes no assumptions about whether this is the *best* lowering, it simply |
10554 | /// uses it. |
10555 | static SDValue lowerVectorShuffleWithSHUFPS(const SDLoc &DL, MVT VT, |
10556 | ArrayRef<int> Mask, SDValue V1, |
10557 | SDValue V2, SelectionDAG &DAG) { |
10558 | SDValue LowV = V1, HighV = V2; |
10559 | int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]}; |
10560 | |
10561 | int NumV2Elements = count_if(Mask, [](int M) { return M >= 4; }); |
10562 | |
10563 | if (NumV2Elements == 1) { |
10564 | int V2Index = find_if(Mask, [](int M) { return M >= 4; }) - Mask.begin(); |
10565 | |
10566 | // Compute the index adjacent to V2Index and in the same half by toggling |
10567 | // the low bit. |
10568 | int V2AdjIndex = V2Index ^ 1; |
10569 | |
10570 | if (Mask[V2AdjIndex] < 0) { |
10571 | // Handles all the cases where we have a single V2 element and an undef. |
10572 | // This will only ever happen in the high lanes because we commute the |
10573 | // vector otherwise. |
10574 | if (V2Index < 2) |
10575 | std::swap(LowV, HighV); |
10576 | NewMask[V2Index] -= 4; |
10577 | } else { |
10578 | // Handle the case where the V2 element ends up adjacent to a V1 element. |
10579 | // To make this work, blend them together as the first step. |
10580 | int V1Index = V2AdjIndex; |
10581 | int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0}; |
10582 | V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1, |
10583 | getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG)); |
10584 | |
10585 | // Now proceed to reconstruct the final blend as we have the necessary |
10586 | // high or low half formed. |
10587 | if (V2Index < 2) { |
10588 | LowV = V2; |
10589 | HighV = V1; |
10590 | } else { |
10591 | HighV = V2; |
10592 | } |
10593 | NewMask[V1Index] = 2; // We put the V1 element in V2[2]. |
10594 | NewMask[V2Index] = 0; // We shifted the V2 element into V2[0]. |
10595 | } |
10596 | } else if (NumV2Elements == 2) { |
10597 | if (Mask[0] < 4 && Mask[1] < 4) { |
10598 | // Handle the easy case where we have V1 in the low lanes and V2 in the |
10599 | // high lanes. |
10600 | NewMask[2] -= 4; |
10601 | NewMask[3] -= 4; |
10602 | } else if (Mask[2] < 4 && Mask[3] < 4) { |
10603 | // We also handle the reversed case because this utility may get called |
10604 | // when we detect a SHUFPS pattern but can't easily commute the shuffle to |
10605 | // arrange things in the right direction. |
10606 | NewMask[0] -= 4; |
10607 | NewMask[1] -= 4; |
10608 | HighV = V1; |
10609 | LowV = V2; |
10610 | } else { |
10611 | // We have a mixture of V1 and V2 in both low and high lanes. Rather than |
10612 | // trying to place elements directly, just blend them and set up the final |
10613 | // shuffle to place them. |
10614 | |
10615 | // The first two blend mask elements are for V1, the second two are for |
10616 | // V2. |
10617 | int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1], |
10618 | Mask[2] < 4 ? Mask[2] : Mask[3], |
10619 | (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4, |
10620 | (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4}; |
10621 | V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2, |
10622 | getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG)); |
10623 | |
10624 | // Now we do a normal shuffle of V1 by giving V1 as both operands to |
10625 | // a blend. |
10626 | LowV = HighV = V1; |
10627 | NewMask[0] = Mask[0] < 4 ? 0 : 2; |
10628 | NewMask[1] = Mask[0] < 4 ? 2 : 0; |
10629 | NewMask[2] = Mask[2] < 4 ? 1 : 3; |
10630 | NewMask[3] = Mask[2] < 4 ? 3 : 1; |
10631 | } |
10632 | } |
10633 | return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV, |
10634 | getV4X86ShuffleImm8ForMask(NewMask, DL, DAG)); |
10635 | } |
10636 | |
10637 | /// \brief Lower 4-lane 32-bit floating point shuffles. |
10638 | /// |
10639 | /// Uses instructions exclusively from the floating point unit to minimize |
10640 | /// domain crossing penalties, as these are sufficient to implement all v4f32 |
10641 | /// shuffles. |
10642 | static SDValue lowerV4F32VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
10643 | const APInt &Zeroable, |
10644 | SDValue V1, SDValue V2, |
10645 | const X86Subtarget &Subtarget, |
10646 | SelectionDAG &DAG) { |
10647 | assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v4f32 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10647, __PRETTY_FUNCTION__)); |
10648 | assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v4f32 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10648, __PRETTY_FUNCTION__)); |
10649 | assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!")((Mask.size() == 4 && "Unexpected mask size for v4 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 4 && \"Unexpected mask size for v4 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10649, __PRETTY_FUNCTION__)); |
10650 | |
10651 | int NumV2Elements = count_if(Mask, [](int M) { return M >= 4; }); |
10652 | |
10653 | if (NumV2Elements == 0) { |
10654 | // Check for being able to broadcast a single element. |
10655 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast( |
10656 | DL, MVT::v4f32, V1, V2, Mask, Subtarget, DAG)) |
10657 | return Broadcast; |
10658 | |
10659 | // Use even/odd duplicate instructions for masks that match their pattern. |
10660 | if (Subtarget.hasSSE3()) { |
10661 | if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2})) |
10662 | return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1); |
10663 | if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3})) |
10664 | return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1); |
10665 | } |
10666 | |
10667 | if (Subtarget.hasAVX()) { |
10668 | // If we have AVX, we can use VPERMILPS which will allow folding a load |
10669 | // into the shuffle. |
10670 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1, |
10671 | getV4X86ShuffleImm8ForMask(Mask, DL, DAG)); |
10672 | } |
10673 | |
10674 | // Otherwise, use a straight shuffle of a single input vector. We pass the |
10675 | // input vector to both operands to simulate this with a SHUFPS. |
10676 | return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1, |
10677 | getV4X86ShuffleImm8ForMask(Mask, DL, DAG)); |
10678 | } |
10679 | |
10680 | // There are special ways we can lower some single-element blends. However, we |
10681 | // have custom ways we can lower more complex single-element blends below that |
10682 | // we defer to if both this and BLENDPS fail to match, so restrict this to |
10683 | // when the V2 input is targeting element 0 of the mask -- that is the fast |
10684 | // case here. |
10685 | if (NumV2Elements == 1 && Mask[0] >= 4) |
10686 | if (SDValue V = lowerVectorShuffleAsElementInsertion( |
10687 | DL, MVT::v4f32, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
10688 | return V; |
10689 | |
10690 | if (Subtarget.hasSSE41()) { |
10691 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask, |
10692 | Zeroable, Subtarget, DAG)) |
10693 | return Blend; |
10694 | |
10695 | // Use INSERTPS if we can complete the shuffle efficiently. |
10696 | if (SDValue V = |
10697 | lowerVectorShuffleAsInsertPS(DL, V1, V2, Mask, Zeroable, DAG)) |
10698 | return V; |
10699 | |
10700 | if (!isSingleSHUFPSMask(Mask)) |
10701 | if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute( |
10702 | DL, MVT::v4f32, V1, V2, Mask, DAG)) |
10703 | return BlendPerm; |
10704 | } |
10705 | |
10706 | // Use low/high mov instructions. |
10707 | if (isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) |
10708 | return DAG.getNode(X86ISD::MOVLHPS, DL, MVT::v4f32, V1, V2); |
10709 | if (isShuffleEquivalent(V1, V2, Mask, {2, 3, 6, 7})) |
10710 | return DAG.getNode(X86ISD::MOVHLPS, DL, MVT::v4f32, V2, V1); |
10711 | |
10712 | // Use dedicated unpack instructions for masks that match their pattern. |
10713 | if (SDValue V = |
10714 | lowerVectorShuffleWithUNPCK(DL, MVT::v4f32, Mask, V1, V2, DAG)) |
10715 | return V; |
10716 | |
10717 | // Otherwise fall back to a SHUFPS lowering strategy. |
10718 | return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG); |
10719 | } |
10720 | |
10721 | /// \brief Lower 4-lane i32 vector shuffles. |
10722 | /// |
10723 | /// We try to handle these with integer-domain shuffles where we can, but for |
10724 | /// blends we use the floating point domain blend instructions. |
10725 | static SDValue lowerV4I32VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
10726 | const APInt &Zeroable, |
10727 | SDValue V1, SDValue V2, |
10728 | const X86Subtarget &Subtarget, |
10729 | SelectionDAG &DAG) { |
10730 | assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v4i32 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10730, __PRETTY_FUNCTION__)); |
10731 | assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v4i32 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10731, __PRETTY_FUNCTION__)); |
10732 | assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!")((Mask.size() == 4 && "Unexpected mask size for v4 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 4 && \"Unexpected mask size for v4 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10732, __PRETTY_FUNCTION__)); |
10733 | |
10734 | // Whenever we can lower this as a zext, that instruction is strictly faster |
10735 | // than any alternative. It also allows us to fold memory operands into the |
10736 | // shuffle in many cases. |
10737 | if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend( |
10738 | DL, MVT::v4i32, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
10739 | return ZExt; |
10740 | |
10741 | int NumV2Elements = count_if(Mask, [](int M) { return M >= 4; }); |
10742 | |
10743 | if (NumV2Elements == 0) { |
10744 | // Check for being able to broadcast a single element. |
10745 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast( |
10746 | DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG)) |
10747 | return Broadcast; |
10748 | |
10749 | // Straight shuffle of a single input vector. For everything from SSE2 |
10750 | // onward this has a single fast instruction with no scary immediates. |
10751 | // We coerce the shuffle pattern to be compatible with UNPCK instructions |
10752 | // but we aren't actually going to use the UNPCK instruction because doing |
10753 | // so prevents folding a load into this instruction or making a copy. |
10754 | const int UnpackLoMask[] = {0, 0, 1, 1}; |
10755 | const int UnpackHiMask[] = {2, 2, 3, 3}; |
10756 | if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1})) |
10757 | Mask = UnpackLoMask; |
10758 | else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3})) |
10759 | Mask = UnpackHiMask; |
10760 | |
10761 | return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1, |
10762 | getV4X86ShuffleImm8ForMask(Mask, DL, DAG)); |
10763 | } |
10764 | |
10765 | // Try to use shift instructions. |
10766 | if (SDValue Shift = lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, |
10767 | Zeroable, Subtarget, DAG)) |
10768 | return Shift; |
10769 | |
10770 | // There are special ways we can lower some single-element blends. |
10771 | if (NumV2Elements == 1) |
10772 | if (SDValue V = lowerVectorShuffleAsElementInsertion( |
10773 | DL, MVT::v4i32, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
10774 | return V; |
10775 | |
10776 | // We have different paths for blend lowering, but they all must use the |
10777 | // *exact* same predicate. |
10778 | bool IsBlendSupported = Subtarget.hasSSE41(); |
10779 | if (IsBlendSupported) |
10780 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask, |
10781 | Zeroable, Subtarget, DAG)) |
10782 | return Blend; |
10783 | |
10784 | if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, |
10785 | Zeroable, DAG)) |
10786 | return Masked; |
10787 | |
10788 | // Use dedicated unpack instructions for masks that match their pattern. |
10789 | if (SDValue V = |
10790 | lowerVectorShuffleWithUNPCK(DL, MVT::v4i32, Mask, V1, V2, DAG)) |
10791 | return V; |
10792 | |
10793 | // Try to use byte rotation instructions. |
10794 | // Its more profitable for pre-SSSE3 to use shuffles/unpacks. |
10795 | if (Subtarget.hasSSSE3()) |
10796 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate( |
10797 | DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG)) |
10798 | return Rotate; |
10799 | |
10800 | // Assume that a single SHUFPS is faster than an alternative sequence of |
10801 | // multiple instructions (even if the CPU has a domain penalty). |
10802 | // If some CPU is harmed by the domain switch, we can fix it in a later pass. |
10803 | if (!isSingleSHUFPSMask(Mask)) { |
10804 | // If we have direct support for blends, we should lower by decomposing into |
10805 | // a permute. That will be faster than the domain cross. |
10806 | if (IsBlendSupported) |
10807 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2, |
10808 | Mask, DAG); |
10809 | |
10810 | // Try to lower by permuting the inputs into an unpack instruction. |
10811 | if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack( |
10812 | DL, MVT::v4i32, V1, V2, Mask, DAG)) |
10813 | return Unpack; |
10814 | } |
10815 | |
10816 | // We implement this with SHUFPS because it can blend from two vectors. |
10817 | // Because we're going to eventually use SHUFPS, we use SHUFPS even to build |
10818 | // up the inputs, bypassing domain shift penalties that we would incur if we |
10819 | // directly used PSHUFD on Nehalem and older. For newer chips, this isn't |
10820 | // relevant. |
10821 | SDValue CastV1 = DAG.getBitcast(MVT::v4f32, V1); |
10822 | SDValue CastV2 = DAG.getBitcast(MVT::v4f32, V2); |
10823 | SDValue ShufPS = DAG.getVectorShuffle(MVT::v4f32, DL, CastV1, CastV2, Mask); |
10824 | return DAG.getBitcast(MVT::v4i32, ShufPS); |
10825 | } |
10826 | |
10827 | /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2 |
10828 | /// shuffle lowering, and the most complex part. |
10829 | /// |
10830 | /// The lowering strategy is to try to form pairs of input lanes which are |
10831 | /// targeted at the same half of the final vector, and then use a dword shuffle |
10832 | /// to place them onto the right half, and finally unpack the paired lanes into |
10833 | /// their final position. |
10834 | /// |
10835 | /// The exact breakdown of how to form these dword pairs and align them on the |
10836 | /// correct sides is really tricky. See the comments within the function for |
10837 | /// more of the details. |
10838 | /// |
10839 | /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each |
10840 | /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to |
10841 | /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16 |
10842 | /// vector, form the analogous 128-bit 8-element Mask. |
10843 | static SDValue lowerV8I16GeneralSingleInputVectorShuffle( |
10844 | const SDLoc &DL, MVT VT, SDValue V, MutableArrayRef<int> Mask, |
10845 | const X86Subtarget &Subtarget, SelectionDAG &DAG) { |
10846 | assert(VT.getVectorElementType() == MVT::i16 && "Bad input type!")((VT.getVectorElementType() == MVT::i16 && "Bad input type!" ) ? static_cast<void> (0) : __assert_fail ("VT.getVectorElementType() == MVT::i16 && \"Bad input type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10846, __PRETTY_FUNCTION__)); |
10847 | MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2); |
10848 | |
10849 | assert(Mask.size() == 8 && "Shuffle mask length doesn't match!")((Mask.size() == 8 && "Shuffle mask length doesn't match!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 8 && \"Shuffle mask length doesn't match!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10849, __PRETTY_FUNCTION__)); |
10850 | MutableArrayRef<int> LoMask = Mask.slice(0, 4); |
10851 | MutableArrayRef<int> HiMask = Mask.slice(4, 4); |
10852 | |
10853 | SmallVector<int, 4> LoInputs; |
10854 | copy_if(LoMask, std::back_inserter(LoInputs), [](int M) { return M >= 0; }); |
10855 | std::sort(LoInputs.begin(), LoInputs.end()); |
10856 | LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end()); |
10857 | SmallVector<int, 4> HiInputs; |
10858 | copy_if(HiMask, std::back_inserter(HiInputs), [](int M) { return M >= 0; }); |
10859 | std::sort(HiInputs.begin(), HiInputs.end()); |
10860 | HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end()); |
10861 | int NumLToL = |
10862 | std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin(); |
10863 | int NumHToL = LoInputs.size() - NumLToL; |
10864 | int NumLToH = |
10865 | std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin(); |
10866 | int NumHToH = HiInputs.size() - NumLToH; |
10867 | MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL); |
10868 | MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH); |
10869 | MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL); |
10870 | MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH); |
10871 | |
10872 | // If we are splatting two values from one half - one to each half, then |
10873 | // we can shuffle that half so each is splatted to a dword, then splat those |
10874 | // to their respective halves. |
10875 | auto SplatHalfs = [&](int LoInput, int HiInput, unsigned ShufWOp, |
10876 | int DOffset) { |
10877 | int PSHUFHalfMask[] = {LoInput % 4, LoInput % 4, HiInput % 4, HiInput % 4}; |
10878 | int PSHUFDMask[] = {DOffset + 0, DOffset + 0, DOffset + 1, DOffset + 1}; |
10879 | V = DAG.getNode(ShufWOp, DL, VT, V, |
10880 | getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG)); |
10881 | V = DAG.getBitcast(PSHUFDVT, V); |
10882 | V = DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, V, |
10883 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)); |
10884 | return DAG.getBitcast(VT, V); |
10885 | }; |
10886 | |
10887 | if (NumLToL == 1 && NumLToH == 1 && (NumHToL + NumHToH) == 0) |
10888 | return SplatHalfs(LToLInputs[0], LToHInputs[0], X86ISD::PSHUFLW, 0); |
10889 | if (NumHToL == 1 && NumHToH == 1 && (NumLToL + NumLToH) == 0) |
10890 | return SplatHalfs(HToLInputs[0], HToHInputs[0], X86ISD::PSHUFHW, 2); |
10891 | |
10892 | // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all |
10893 | // such inputs we can swap two of the dwords across the half mark and end up |
10894 | // with <=2 inputs to each half in each half. Once there, we can fall through |
10895 | // to the generic code below. For example: |
10896 | // |
10897 | // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h] |
10898 | // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5] |
10899 | // |
10900 | // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half |
10901 | // and an existing 2-into-2 on the other half. In this case we may have to |
10902 | // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or |
10903 | // 1-into-3 which could cause us to cycle endlessly fixing each side in turn. |
10904 | // Fortunately, we don't have to handle anything but a 2-into-2 pattern |
10905 | // because any other situation (including a 3-into-1 or 1-into-3 in the other |
10906 | // half than the one we target for fixing) will be fixed when we re-enter this |
10907 | // path. We will also combine away any sequence of PSHUFD instructions that |
10908 | // result into a single instruction. Here is an example of the tricky case: |
10909 | // |
10910 | // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h] |
10911 | // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3] |
10912 | // |
10913 | // This now has a 1-into-3 in the high half! Instead, we do two shuffles: |
10914 | // |
10915 | // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h] |
10916 | // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6] |
10917 | // |
10918 | // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h] |
10919 | // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6] |
10920 | // |
10921 | // The result is fine to be handled by the generic logic. |
10922 | auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs, |
10923 | ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs, |
10924 | int AOffset, int BOffset) { |
10925 | assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&(((AToAInputs.size() == 3 || AToAInputs.size() == 1) && "Must call this with A having 3 or 1 inputs from the A half." ) ? static_cast<void> (0) : __assert_fail ("(AToAInputs.size() == 3 || AToAInputs.size() == 1) && \"Must call this with A having 3 or 1 inputs from the A half.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10926, __PRETTY_FUNCTION__)) |
10926 | "Must call this with A having 3 or 1 inputs from the A half.")(((AToAInputs.size() == 3 || AToAInputs.size() == 1) && "Must call this with A having 3 or 1 inputs from the A half." ) ? static_cast<void> (0) : __assert_fail ("(AToAInputs.size() == 3 || AToAInputs.size() == 1) && \"Must call this with A having 3 or 1 inputs from the A half.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10926, __PRETTY_FUNCTION__)); |
10927 | assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&(((BToAInputs.size() == 1 || BToAInputs.size() == 3) && "Must call this with B having 1 or 3 inputs from the B half." ) ? static_cast<void> (0) : __assert_fail ("(BToAInputs.size() == 1 || BToAInputs.size() == 3) && \"Must call this with B having 1 or 3 inputs from the B half.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10928, __PRETTY_FUNCTION__)) |
10928 | "Must call this with B having 1 or 3 inputs from the B half.")(((BToAInputs.size() == 1 || BToAInputs.size() == 3) && "Must call this with B having 1 or 3 inputs from the B half." ) ? static_cast<void> (0) : __assert_fail ("(BToAInputs.size() == 1 || BToAInputs.size() == 3) && \"Must call this with B having 1 or 3 inputs from the B half.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10928, __PRETTY_FUNCTION__)); |
10929 | assert(AToAInputs.size() + BToAInputs.size() == 4 &&((AToAInputs.size() + BToAInputs.size() == 4 && "Must call this with either 3:1 or 1:3 inputs (summing to 4)." ) ? static_cast<void> (0) : __assert_fail ("AToAInputs.size() + BToAInputs.size() == 4 && \"Must call this with either 3:1 or 1:3 inputs (summing to 4).\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10930, __PRETTY_FUNCTION__)) |
10930 | "Must call this with either 3:1 or 1:3 inputs (summing to 4).")((AToAInputs.size() + BToAInputs.size() == 4 && "Must call this with either 3:1 or 1:3 inputs (summing to 4)." ) ? static_cast<void> (0) : __assert_fail ("AToAInputs.size() + BToAInputs.size() == 4 && \"Must call this with either 3:1 or 1:3 inputs (summing to 4).\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10930, __PRETTY_FUNCTION__)); |
10931 | |
10932 | bool ThreeAInputs = AToAInputs.size() == 3; |
10933 | |
10934 | // Compute the index of dword with only one word among the three inputs in |
10935 | // a half by taking the sum of the half with three inputs and subtracting |
10936 | // the sum of the actual three inputs. The difference is the remaining |
10937 | // slot. |
10938 | int ADWord, BDWord; |
10939 | int &TripleDWord = ThreeAInputs ? ADWord : BDWord; |
10940 | int &OneInputDWord = ThreeAInputs ? BDWord : ADWord; |
10941 | int TripleInputOffset = ThreeAInputs ? AOffset : BOffset; |
10942 | ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs; |
10943 | int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0]; |
10944 | int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset); |
10945 | int TripleNonInputIdx = |
10946 | TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0); |
10947 | TripleDWord = TripleNonInputIdx / 2; |
10948 | |
10949 | // We use xor with one to compute the adjacent DWord to whichever one the |
10950 | // OneInput is in. |
10951 | OneInputDWord = (OneInput / 2) ^ 1; |
10952 | |
10953 | // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA |
10954 | // and BToA inputs. If there is also such a problem with the BToB and AToB |
10955 | // inputs, we don't try to fix it necessarily -- we'll recurse and see it in |
10956 | // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it |
10957 | // is essential that we don't *create* a 3<-1 as then we might oscillate. |
10958 | if (BToBInputs.size() == 2 && AToBInputs.size() == 2) { |
10959 | // Compute how many inputs will be flipped by swapping these DWords. We |
10960 | // need |
10961 | // to balance this to ensure we don't form a 3-1 shuffle in the other |
10962 | // half. |
10963 | int NumFlippedAToBInputs = |
10964 | std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) + |
10965 | std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1); |
10966 | int NumFlippedBToBInputs = |
10967 | std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) + |
10968 | std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1); |
10969 | if ((NumFlippedAToBInputs == 1 && |
10970 | (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) || |
10971 | (NumFlippedBToBInputs == 1 && |
10972 | (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) { |
10973 | // We choose whether to fix the A half or B half based on whether that |
10974 | // half has zero flipped inputs. At zero, we may not be able to fix it |
10975 | // with that half. We also bias towards fixing the B half because that |
10976 | // will more commonly be the high half, and we have to bias one way. |
10977 | auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord, |
10978 | ArrayRef<int> Inputs) { |
10979 | int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot. |
10980 | bool IsFixIdxInput = is_contained(Inputs, PinnedIdx ^ 1); |
10981 | // Determine whether the free index is in the flipped dword or the |
10982 | // unflipped dword based on where the pinned index is. We use this bit |
10983 | // in an xor to conditionally select the adjacent dword. |
10984 | int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord)); |
10985 | bool IsFixFreeIdxInput = is_contained(Inputs, FixFreeIdx); |
10986 | if (IsFixIdxInput == IsFixFreeIdxInput) |
10987 | FixFreeIdx += 1; |
10988 | IsFixFreeIdxInput = is_contained(Inputs, FixFreeIdx); |
10989 | assert(IsFixIdxInput != IsFixFreeIdxInput &&((IsFixIdxInput != IsFixFreeIdxInput && "We need to be changing the number of flipped inputs!" ) ? static_cast<void> (0) : __assert_fail ("IsFixIdxInput != IsFixFreeIdxInput && \"We need to be changing the number of flipped inputs!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10990, __PRETTY_FUNCTION__)) |
10990 | "We need to be changing the number of flipped inputs!")((IsFixIdxInput != IsFixFreeIdxInput && "We need to be changing the number of flipped inputs!" ) ? static_cast<void> (0) : __assert_fail ("IsFixIdxInput != IsFixFreeIdxInput && \"We need to be changing the number of flipped inputs!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 10990, __PRETTY_FUNCTION__)); |
10991 | int PSHUFHalfMask[] = {0, 1, 2, 3}; |
10992 | std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]); |
10993 | V = DAG.getNode( |
10994 | FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL, |
10995 | MVT::getVectorVT(MVT::i16, V.getValueSizeInBits() / 16), V, |
10996 | getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG)); |
10997 | |
10998 | for (int &M : Mask) |
10999 | if (M >= 0 && M == FixIdx) |
11000 | M = FixFreeIdx; |
11001 | else if (M >= 0 && M == FixFreeIdx) |
11002 | M = FixIdx; |
11003 | }; |
11004 | if (NumFlippedBToBInputs != 0) { |
11005 | int BPinnedIdx = |
11006 | BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput; |
11007 | FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs); |
11008 | } else { |
11009 | assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!")((NumFlippedAToBInputs != 0 && "Impossible given predicates!" ) ? static_cast<void> (0) : __assert_fail ("NumFlippedAToBInputs != 0 && \"Impossible given predicates!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11009, __PRETTY_FUNCTION__)); |
11010 | int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput; |
11011 | FixFlippedInputs(APinnedIdx, ADWord, AToBInputs); |
11012 | } |
11013 | } |
11014 | } |
11015 | |
11016 | int PSHUFDMask[] = {0, 1, 2, 3}; |
11017 | PSHUFDMask[ADWord] = BDWord; |
11018 | PSHUFDMask[BDWord] = ADWord; |
11019 | V = DAG.getBitcast( |
11020 | VT, |
11021 | DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V), |
11022 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG))); |
11023 | |
11024 | // Adjust the mask to match the new locations of A and B. |
11025 | for (int &M : Mask) |
11026 | if (M >= 0 && M/2 == ADWord) |
11027 | M = 2 * BDWord + M % 2; |
11028 | else if (M >= 0 && M/2 == BDWord) |
11029 | M = 2 * ADWord + M % 2; |
11030 | |
11031 | // Recurse back into this routine to re-compute state now that this isn't |
11032 | // a 3 and 1 problem. |
11033 | return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget, |
11034 | DAG); |
11035 | }; |
11036 | if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3)) |
11037 | return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4); |
11038 | if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3)) |
11039 | return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0); |
11040 | |
11041 | // At this point there are at most two inputs to the low and high halves from |
11042 | // each half. That means the inputs can always be grouped into dwords and |
11043 | // those dwords can then be moved to the correct half with a dword shuffle. |
11044 | // We use at most one low and one high word shuffle to collect these paired |
11045 | // inputs into dwords, and finally a dword shuffle to place them. |
11046 | int PSHUFLMask[4] = {-1, -1, -1, -1}; |
11047 | int PSHUFHMask[4] = {-1, -1, -1, -1}; |
11048 | int PSHUFDMask[4] = {-1, -1, -1, -1}; |
11049 | |
11050 | // First fix the masks for all the inputs that are staying in their |
11051 | // original halves. This will then dictate the targets of the cross-half |
11052 | // shuffles. |
11053 | auto fixInPlaceInputs = |
11054 | [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs, |
11055 | MutableArrayRef<int> SourceHalfMask, |
11056 | MutableArrayRef<int> HalfMask, int HalfOffset) { |
11057 | if (InPlaceInputs.empty()) |
11058 | return; |
11059 | if (InPlaceInputs.size() == 1) { |
11060 | SourceHalfMask[InPlaceInputs[0] - HalfOffset] = |
11061 | InPlaceInputs[0] - HalfOffset; |
11062 | PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2; |
11063 | return; |
11064 | } |
11065 | if (IncomingInputs.empty()) { |
11066 | // Just fix all of the in place inputs. |
11067 | for (int Input : InPlaceInputs) { |
11068 | SourceHalfMask[Input - HalfOffset] = Input - HalfOffset; |
11069 | PSHUFDMask[Input / 2] = Input / 2; |
11070 | } |
11071 | return; |
11072 | } |
11073 | |
11074 | assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!")((InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!" ) ? static_cast<void> (0) : __assert_fail ("InPlaceInputs.size() == 2 && \"Cannot handle 3 or 4 inputs!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11074, __PRETTY_FUNCTION__)); |
11075 | SourceHalfMask[InPlaceInputs[0] - HalfOffset] = |
11076 | InPlaceInputs[0] - HalfOffset; |
11077 | // Put the second input next to the first so that they are packed into |
11078 | // a dword. We find the adjacent index by toggling the low bit. |
11079 | int AdjIndex = InPlaceInputs[0] ^ 1; |
11080 | SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset; |
11081 | std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex); |
11082 | PSHUFDMask[AdjIndex / 2] = AdjIndex / 2; |
11083 | }; |
11084 | fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0); |
11085 | fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4); |
11086 | |
11087 | // Now gather the cross-half inputs and place them into a free dword of |
11088 | // their target half. |
11089 | // FIXME: This operation could almost certainly be simplified dramatically to |
11090 | // look more like the 3-1 fixing operation. |
11091 | auto moveInputsToRightHalf = [&PSHUFDMask]( |
11092 | MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs, |
11093 | MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask, |
11094 | MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset, |
11095 | int DestOffset) { |
11096 | auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) { |
11097 | return SourceHalfMask[Word] >= 0 && SourceHalfMask[Word] != Word; |
11098 | }; |
11099 | auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask, |
11100 | int Word) { |
11101 | int LowWord = Word & ~1; |
11102 | int HighWord = Word | 1; |
11103 | return isWordClobbered(SourceHalfMask, LowWord) || |
11104 | isWordClobbered(SourceHalfMask, HighWord); |
11105 | }; |
11106 | |
11107 | if (IncomingInputs.empty()) |
11108 | return; |
11109 | |
11110 | if (ExistingInputs.empty()) { |
11111 | // Map any dwords with inputs from them into the right half. |
11112 | for (int Input : IncomingInputs) { |
11113 | // If the source half mask maps over the inputs, turn those into |
11114 | // swaps and use the swapped lane. |
11115 | if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) { |
11116 | if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] < 0) { |
11117 | SourceHalfMask[SourceHalfMask[Input - SourceOffset]] = |
11118 | Input - SourceOffset; |
11119 | // We have to swap the uses in our half mask in one sweep. |
11120 | for (int &M : HalfMask) |
11121 | if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset) |
11122 | M = Input; |
11123 | else if (M == Input) |
11124 | M = SourceHalfMask[Input - SourceOffset] + SourceOffset; |
11125 | } else { |
11126 | assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==((SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == Input - SourceOffset && "Previous placement doesn't match!" ) ? static_cast<void> (0) : __assert_fail ("SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == Input - SourceOffset && \"Previous placement doesn't match!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11128, __PRETTY_FUNCTION__)) |
11127 | Input - SourceOffset &&((SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == Input - SourceOffset && "Previous placement doesn't match!" ) ? static_cast<void> (0) : __assert_fail ("SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == Input - SourceOffset && \"Previous placement doesn't match!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11128, __PRETTY_FUNCTION__)) |
11128 | "Previous placement doesn't match!")((SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == Input - SourceOffset && "Previous placement doesn't match!" ) ? static_cast<void> (0) : __assert_fail ("SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == Input - SourceOffset && \"Previous placement doesn't match!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11128, __PRETTY_FUNCTION__)); |
11129 | } |
11130 | // Note that this correctly re-maps both when we do a swap and when |
11131 | // we observe the other side of the swap above. We rely on that to |
11132 | // avoid swapping the members of the input list directly. |
11133 | Input = SourceHalfMask[Input - SourceOffset] + SourceOffset; |
11134 | } |
11135 | |
11136 | // Map the input's dword into the correct half. |
11137 | if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] < 0) |
11138 | PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2; |
11139 | else |
11140 | assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==((PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == Input / 2 && "Previous placement doesn't match!") ? static_cast <void> (0) : __assert_fail ("PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == Input / 2 && \"Previous placement doesn't match!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11142, __PRETTY_FUNCTION__)) |
11141 | Input / 2 &&((PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == Input / 2 && "Previous placement doesn't match!") ? static_cast <void> (0) : __assert_fail ("PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == Input / 2 && \"Previous placement doesn't match!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11142, __PRETTY_FUNCTION__)) |
11142 | "Previous placement doesn't match!")((PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == Input / 2 && "Previous placement doesn't match!") ? static_cast <void> (0) : __assert_fail ("PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == Input / 2 && \"Previous placement doesn't match!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11142, __PRETTY_FUNCTION__)); |
11143 | } |
11144 | |
11145 | // And just directly shift any other-half mask elements to be same-half |
11146 | // as we will have mirrored the dword containing the element into the |
11147 | // same position within that half. |
11148 | for (int &M : HalfMask) |
11149 | if (M >= SourceOffset && M < SourceOffset + 4) { |
11150 | M = M - SourceOffset + DestOffset; |
11151 | assert(M >= 0 && "This should never wrap below zero!")((M >= 0 && "This should never wrap below zero!") ? static_cast<void> (0) : __assert_fail ("M >= 0 && \"This should never wrap below zero!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11151, __PRETTY_FUNCTION__)); |
11152 | } |
11153 | return; |
11154 | } |
11155 | |
11156 | // Ensure we have the input in a viable dword of its current half. This |
11157 | // is particularly tricky because the original position may be clobbered |
11158 | // by inputs being moved and *staying* in that half. |
11159 | if (IncomingInputs.size() == 1) { |
11160 | if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) { |
11161 | int InputFixed = find(SourceHalfMask, -1) - std::begin(SourceHalfMask) + |
11162 | SourceOffset; |
11163 | SourceHalfMask[InputFixed - SourceOffset] = |
11164 | IncomingInputs[0] - SourceOffset; |
11165 | std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0], |
11166 | InputFixed); |
11167 | IncomingInputs[0] = InputFixed; |
11168 | } |
11169 | } else if (IncomingInputs.size() == 2) { |
11170 | if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 || |
11171 | isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) { |
11172 | // We have two non-adjacent or clobbered inputs we need to extract from |
11173 | // the source half. To do this, we need to map them into some adjacent |
11174 | // dword slot in the source mask. |
11175 | int InputsFixed[2] = {IncomingInputs[0] - SourceOffset, |
11176 | IncomingInputs[1] - SourceOffset}; |
11177 | |
11178 | // If there is a free slot in the source half mask adjacent to one of |
11179 | // the inputs, place the other input in it. We use (Index XOR 1) to |
11180 | // compute an adjacent index. |
11181 | if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) && |
11182 | SourceHalfMask[InputsFixed[0] ^ 1] < 0) { |
11183 | SourceHalfMask[InputsFixed[0]] = InputsFixed[0]; |
11184 | SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1]; |
11185 | InputsFixed[1] = InputsFixed[0] ^ 1; |
11186 | } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) && |
11187 | SourceHalfMask[InputsFixed[1] ^ 1] < 0) { |
11188 | SourceHalfMask[InputsFixed[1]] = InputsFixed[1]; |
11189 | SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0]; |
11190 | InputsFixed[0] = InputsFixed[1] ^ 1; |
11191 | } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] < 0 && |
11192 | SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] < 0) { |
11193 | // The two inputs are in the same DWord but it is clobbered and the |
11194 | // adjacent DWord isn't used at all. Move both inputs to the free |
11195 | // slot. |
11196 | SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0]; |
11197 | SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1]; |
11198 | InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1); |
11199 | InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1; |
11200 | } else { |
11201 | // The only way we hit this point is if there is no clobbering |
11202 | // (because there are no off-half inputs to this half) and there is no |
11203 | // free slot adjacent to one of the inputs. In this case, we have to |
11204 | // swap an input with a non-input. |
11205 | for (int i = 0; i < 4; ++i) |
11206 | assert((SourceHalfMask[i] < 0 || SourceHalfMask[i] == i) &&(((SourceHalfMask[i] < 0 || SourceHalfMask[i] == i) && "We can't handle any clobbers here!") ? static_cast<void> (0) : __assert_fail ("(SourceHalfMask[i] < 0 || SourceHalfMask[i] == i) && \"We can't handle any clobbers here!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11207, __PRETTY_FUNCTION__)) |
11207 | "We can't handle any clobbers here!")(((SourceHalfMask[i] < 0 || SourceHalfMask[i] == i) && "We can't handle any clobbers here!") ? static_cast<void> (0) : __assert_fail ("(SourceHalfMask[i] < 0 || SourceHalfMask[i] == i) && \"We can't handle any clobbers here!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11207, __PRETTY_FUNCTION__)); |
11208 | assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&((InputsFixed[1] != (InputsFixed[0] ^ 1) && "Cannot have adjacent inputs here!" ) ? static_cast<void> (0) : __assert_fail ("InputsFixed[1] != (InputsFixed[0] ^ 1) && \"Cannot have adjacent inputs here!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11209, __PRETTY_FUNCTION__)) |
11209 | "Cannot have adjacent inputs here!")((InputsFixed[1] != (InputsFixed[0] ^ 1) && "Cannot have adjacent inputs here!" ) ? static_cast<void> (0) : __assert_fail ("InputsFixed[1] != (InputsFixed[0] ^ 1) && \"Cannot have adjacent inputs here!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11209, __PRETTY_FUNCTION__)); |
11210 | |
11211 | SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1]; |
11212 | SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1; |
11213 | |
11214 | // We also have to update the final source mask in this case because |
11215 | // it may need to undo the above swap. |
11216 | for (int &M : FinalSourceHalfMask) |
11217 | if (M == (InputsFixed[0] ^ 1) + SourceOffset) |
11218 | M = InputsFixed[1] + SourceOffset; |
11219 | else if (M == InputsFixed[1] + SourceOffset) |
11220 | M = (InputsFixed[0] ^ 1) + SourceOffset; |
11221 | |
11222 | InputsFixed[1] = InputsFixed[0] ^ 1; |
11223 | } |
11224 | |
11225 | // Point everything at the fixed inputs. |
11226 | for (int &M : HalfMask) |
11227 | if (M == IncomingInputs[0]) |
11228 | M = InputsFixed[0] + SourceOffset; |
11229 | else if (M == IncomingInputs[1]) |
11230 | M = InputsFixed[1] + SourceOffset; |
11231 | |
11232 | IncomingInputs[0] = InputsFixed[0] + SourceOffset; |
11233 | IncomingInputs[1] = InputsFixed[1] + SourceOffset; |
11234 | } |
11235 | } else { |
11236 | llvm_unreachable("Unhandled input size!")::llvm::llvm_unreachable_internal("Unhandled input size!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11236); |
11237 | } |
11238 | |
11239 | // Now hoist the DWord down to the right half. |
11240 | int FreeDWord = (PSHUFDMask[DestOffset / 2] < 0 ? 0 : 1) + DestOffset / 2; |
11241 | assert(PSHUFDMask[FreeDWord] < 0 && "DWord not free")((PSHUFDMask[FreeDWord] < 0 && "DWord not free") ? static_cast<void> (0) : __assert_fail ("PSHUFDMask[FreeDWord] < 0 && \"DWord not free\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11241, __PRETTY_FUNCTION__)); |
11242 | PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2; |
11243 | for (int &M : HalfMask) |
11244 | for (int Input : IncomingInputs) |
11245 | if (M == Input) |
11246 | M = FreeDWord * 2 + Input % 2; |
11247 | }; |
11248 | moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask, |
11249 | /*SourceOffset*/ 4, /*DestOffset*/ 0); |
11250 | moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask, |
11251 | /*SourceOffset*/ 0, /*DestOffset*/ 4); |
11252 | |
11253 | // Now enact all the shuffles we've computed to move the inputs into their |
11254 | // target half. |
11255 | if (!isNoopShuffleMask(PSHUFLMask)) |
11256 | V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V, |
11257 | getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG)); |
11258 | if (!isNoopShuffleMask(PSHUFHMask)) |
11259 | V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V, |
11260 | getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG)); |
11261 | if (!isNoopShuffleMask(PSHUFDMask)) |
11262 | V = DAG.getBitcast( |
11263 | VT, |
11264 | DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V), |
11265 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG))); |
11266 | |
11267 | // At this point, each half should contain all its inputs, and we can then |
11268 | // just shuffle them into their final position. |
11269 | assert(count_if(LoMask, [](int M) { return M >= 4; }) == 0 &&((count_if(LoMask, [](int M) { return M >= 4; }) == 0 && "Failed to lift all the high half inputs to the low mask!") ? static_cast<void> (0) : __assert_fail ("count_if(LoMask, [](int M) { return M >= 4; }) == 0 && \"Failed to lift all the high half inputs to the low mask!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11270, __PRETTY_FUNCTION__)) |
11270 | "Failed to lift all the high half inputs to the low mask!")((count_if(LoMask, [](int M) { return M >= 4; }) == 0 && "Failed to lift all the high half inputs to the low mask!") ? static_cast<void> (0) : __assert_fail ("count_if(LoMask, [](int M) { return M >= 4; }) == 0 && \"Failed to lift all the high half inputs to the low mask!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11270, __PRETTY_FUNCTION__)); |
11271 | assert(count_if(HiMask, [](int M) { return M >= 0 && M < 4; }) == 0 &&((count_if(HiMask, [](int M) { return M >= 0 && M < 4; }) == 0 && "Failed to lift all the low half inputs to the high mask!" ) ? static_cast<void> (0) : __assert_fail ("count_if(HiMask, [](int M) { return M >= 0 && M < 4; }) == 0 && \"Failed to lift all the low half inputs to the high mask!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11272, __PRETTY_FUNCTION__)) |
11272 | "Failed to lift all the low half inputs to the high mask!")((count_if(HiMask, [](int M) { return M >= 0 && M < 4; }) == 0 && "Failed to lift all the low half inputs to the high mask!" ) ? static_cast<void> (0) : __assert_fail ("count_if(HiMask, [](int M) { return M >= 0 && M < 4; }) == 0 && \"Failed to lift all the low half inputs to the high mask!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11272, __PRETTY_FUNCTION__)); |
11273 | |
11274 | // Do a half shuffle for the low mask. |
11275 | if (!isNoopShuffleMask(LoMask)) |
11276 | V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V, |
11277 | getV4X86ShuffleImm8ForMask(LoMask, DL, DAG)); |
11278 | |
11279 | // Do a half shuffle with the high mask after shifting its values down. |
11280 | for (int &M : HiMask) |
11281 | if (M >= 0) |
11282 | M -= 4; |
11283 | if (!isNoopShuffleMask(HiMask)) |
11284 | V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V, |
11285 | getV4X86ShuffleImm8ForMask(HiMask, DL, DAG)); |
11286 | |
11287 | return V; |
11288 | } |
11289 | |
11290 | /// Helper to form a PSHUFB-based shuffle+blend, opportunistically avoiding the |
11291 | /// blend if only one input is used. |
11292 | static SDValue lowerVectorShuffleAsBlendOfPSHUFBs( |
11293 | const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, |
11294 | const APInt &Zeroable, SelectionDAG &DAG, bool &V1InUse, |
11295 | bool &V2InUse) { |
11296 | SDValue V1Mask[16]; |
11297 | SDValue V2Mask[16]; |
11298 | V1InUse = false; |
11299 | V2InUse = false; |
11300 | |
11301 | int Size = Mask.size(); |
11302 | int Scale = 16 / Size; |
11303 | for (int i = 0; i < 16; ++i) { |
11304 | if (Mask[i / Scale] < 0) { |
11305 | V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8); |
11306 | } else { |
11307 | const int ZeroMask = 0x80; |
11308 | int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale |
11309 | : ZeroMask; |
11310 | int V2Idx = Mask[i / Scale] < Size |
11311 | ? ZeroMask |
11312 | : (Mask[i / Scale] - Size) * Scale + i % Scale; |
11313 | if (Zeroable[i / Scale]) |
11314 | V1Idx = V2Idx = ZeroMask; |
11315 | V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8); |
11316 | V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8); |
11317 | V1InUse |= (ZeroMask != V1Idx); |
11318 | V2InUse |= (ZeroMask != V2Idx); |
11319 | } |
11320 | } |
11321 | |
11322 | if (V1InUse) |
11323 | V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, |
11324 | DAG.getBitcast(MVT::v16i8, V1), |
11325 | DAG.getBuildVector(MVT::v16i8, DL, V1Mask)); |
11326 | if (V2InUse) |
11327 | V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, |
11328 | DAG.getBitcast(MVT::v16i8, V2), |
11329 | DAG.getBuildVector(MVT::v16i8, DL, V2Mask)); |
11330 | |
11331 | // If we need shuffled inputs from both, blend the two. |
11332 | SDValue V; |
11333 | if (V1InUse && V2InUse) |
11334 | V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2); |
11335 | else |
11336 | V = V1InUse ? V1 : V2; |
11337 | |
11338 | // Cast the result back to the correct type. |
11339 | return DAG.getBitcast(VT, V); |
11340 | } |
11341 | |
11342 | /// \brief Generic lowering of 8-lane i16 shuffles. |
11343 | /// |
11344 | /// This handles both single-input shuffles and combined shuffle/blends with |
11345 | /// two inputs. The single input shuffles are immediately delegated to |
11346 | /// a dedicated lowering routine. |
11347 | /// |
11348 | /// The blends are lowered in one of three fundamental ways. If there are few |
11349 | /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle |
11350 | /// of the input is significantly cheaper when lowered as an interleaving of |
11351 | /// the two inputs, try to interleave them. Otherwise, blend the low and high |
11352 | /// halves of the inputs separately (making them have relatively few inputs) |
11353 | /// and then concatenate them. |
11354 | static SDValue lowerV8I16VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
11355 | const APInt &Zeroable, |
11356 | SDValue V1, SDValue V2, |
11357 | const X86Subtarget &Subtarget, |
11358 | SelectionDAG &DAG) { |
11359 | assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v8i16 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11359, __PRETTY_FUNCTION__)); |
11360 | assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v8i16 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11360, __PRETTY_FUNCTION__)); |
11361 | assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!")((Mask.size() == 8 && "Unexpected mask size for v8 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 8 && \"Unexpected mask size for v8 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11361, __PRETTY_FUNCTION__)); |
11362 | |
11363 | // Whenever we can lower this as a zext, that instruction is strictly faster |
11364 | // than any alternative. |
11365 | if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend( |
11366 | DL, MVT::v8i16, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
11367 | return ZExt; |
11368 | |
11369 | int NumV2Inputs = count_if(Mask, [](int M) { return M >= 8; }); |
11370 | |
11371 | if (NumV2Inputs == 0) { |
11372 | // Check for being able to broadcast a single element. |
11373 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast( |
11374 | DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG)) |
11375 | return Broadcast; |
11376 | |
11377 | // Try to use shift instructions. |
11378 | if (SDValue Shift = lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, |
11379 | Zeroable, Subtarget, DAG)) |
11380 | return Shift; |
11381 | |
11382 | // Use dedicated unpack instructions for masks that match their pattern. |
11383 | if (SDValue V = |
11384 | lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG)) |
11385 | return V; |
11386 | |
11387 | // Try to use byte rotation instructions. |
11388 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1, |
11389 | Mask, Subtarget, DAG)) |
11390 | return Rotate; |
11391 | |
11392 | // Make a copy of the mask so it can be modified. |
11393 | SmallVector<int, 8> MutableMask(Mask.begin(), Mask.end()); |
11394 | return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, |
11395 | MutableMask, Subtarget, |
11396 | DAG); |
11397 | } |
11398 | |
11399 | assert(llvm::any_of(Mask, [](int M) { return M >= 0 && M < 8; }) &&((llvm::any_of(Mask, [](int M) { return M >= 0 && M < 8; }) && "All single-input shuffles should be canonicalized to be V1-input " "shuffles.") ? static_cast<void> (0) : __assert_fail ( "llvm::any_of(Mask, [](int M) { return M >= 0 && M < 8; }) && \"All single-input shuffles should be canonicalized to be V1-input \" \"shuffles.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11401, __PRETTY_FUNCTION__)) |
11400 | "All single-input shuffles should be canonicalized to be V1-input "((llvm::any_of(Mask, [](int M) { return M >= 0 && M < 8; }) && "All single-input shuffles should be canonicalized to be V1-input " "shuffles.") ? static_cast<void> (0) : __assert_fail ( "llvm::any_of(Mask, [](int M) { return M >= 0 && M < 8; }) && \"All single-input shuffles should be canonicalized to be V1-input \" \"shuffles.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11401, __PRETTY_FUNCTION__)) |
11401 | "shuffles.")((llvm::any_of(Mask, [](int M) { return M >= 0 && M < 8; }) && "All single-input shuffles should be canonicalized to be V1-input " "shuffles.") ? static_cast<void> (0) : __assert_fail ( "llvm::any_of(Mask, [](int M) { return M >= 0 && M < 8; }) && \"All single-input shuffles should be canonicalized to be V1-input \" \"shuffles.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11401, __PRETTY_FUNCTION__)); |
11402 | |
11403 | // Try to use shift instructions. |
11404 | if (SDValue Shift = lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, |
11405 | Zeroable, Subtarget, DAG)) |
11406 | return Shift; |
11407 | |
11408 | // See if we can use SSE4A Extraction / Insertion. |
11409 | if (Subtarget.hasSSE4A()) |
11410 | if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, |
11411 | Zeroable, DAG)) |
11412 | return V; |
11413 | |
11414 | // There are special ways we can lower some single-element blends. |
11415 | if (NumV2Inputs == 1) |
11416 | if (SDValue V = lowerVectorShuffleAsElementInsertion( |
11417 | DL, MVT::v8i16, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
11418 | return V; |
11419 | |
11420 | // We have different paths for blend lowering, but they all must use the |
11421 | // *exact* same predicate. |
11422 | bool IsBlendSupported = Subtarget.hasSSE41(); |
11423 | if (IsBlendSupported) |
11424 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask, |
11425 | Zeroable, Subtarget, DAG)) |
11426 | return Blend; |
11427 | |
11428 | if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, |
11429 | Zeroable, DAG)) |
11430 | return Masked; |
11431 | |
11432 | // Use dedicated unpack instructions for masks that match their pattern. |
11433 | if (SDValue V = |
11434 | lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG)) |
11435 | return V; |
11436 | |
11437 | // Try to use byte rotation instructions. |
11438 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate( |
11439 | DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG)) |
11440 | return Rotate; |
11441 | |
11442 | if (SDValue BitBlend = |
11443 | lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG)) |
11444 | return BitBlend; |
11445 | |
11446 | // Try to lower by permuting the inputs into an unpack instruction. |
11447 | if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1, |
11448 | V2, Mask, DAG)) |
11449 | return Unpack; |
11450 | |
11451 | // If we can't directly blend but can use PSHUFB, that will be better as it |
11452 | // can both shuffle and set up the inefficient blend. |
11453 | if (!IsBlendSupported && Subtarget.hasSSSE3()) { |
11454 | bool V1InUse, V2InUse; |
11455 | return lowerVectorShuffleAsBlendOfPSHUFBs(DL, MVT::v8i16, V1, V2, Mask, |
11456 | Zeroable, DAG, V1InUse, V2InUse); |
11457 | } |
11458 | |
11459 | // We can always bit-blend if we have to so the fallback strategy is to |
11460 | // decompose into single-input permutes and blends. |
11461 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2, |
11462 | Mask, DAG); |
11463 | } |
11464 | |
11465 | /// \brief Check whether a compaction lowering can be done by dropping even |
11466 | /// elements and compute how many times even elements must be dropped. |
11467 | /// |
11468 | /// This handles shuffles which take every Nth element where N is a power of |
11469 | /// two. Example shuffle masks: |
11470 | /// |
11471 | /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14 |
11472 | /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30 |
11473 | /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12 |
11474 | /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28 |
11475 | /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8 |
11476 | /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24 |
11477 | /// |
11478 | /// Any of these lanes can of course be undef. |
11479 | /// |
11480 | /// This routine only supports N <= 3. |
11481 | /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here |
11482 | /// for larger N. |
11483 | /// |
11484 | /// \returns N above, or the number of times even elements must be dropped if |
11485 | /// there is such a number. Otherwise returns zero. |
11486 | static int canLowerByDroppingEvenElements(ArrayRef<int> Mask, |
11487 | bool IsSingleInput) { |
11488 | // The modulus for the shuffle vector entries is based on whether this is |
11489 | // a single input or not. |
11490 | int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2); |
11491 | assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&((isPowerOf2_32((uint32_t)ShuffleModulus) && "We should only be called with masks with a power-of-2 size!" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32((uint32_t)ShuffleModulus) && \"We should only be called with masks with a power-of-2 size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11492, __PRETTY_FUNCTION__)) |
11492 | "We should only be called with masks with a power-of-2 size!")((isPowerOf2_32((uint32_t)ShuffleModulus) && "We should only be called with masks with a power-of-2 size!" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32((uint32_t)ShuffleModulus) && \"We should only be called with masks with a power-of-2 size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11492, __PRETTY_FUNCTION__)); |
11493 | |
11494 | uint64_t ModMask = (uint64_t)ShuffleModulus - 1; |
11495 | |
11496 | // We track whether the input is viable for all power-of-2 strides 2^1, 2^2, |
11497 | // and 2^3 simultaneously. This is because we may have ambiguity with |
11498 | // partially undef inputs. |
11499 | bool ViableForN[3] = {true, true, true}; |
11500 | |
11501 | for (int i = 0, e = Mask.size(); i < e; ++i) { |
11502 | // Ignore undef lanes, we'll optimistically collapse them to the pattern we |
11503 | // want. |
11504 | if (Mask[i] < 0) |
11505 | continue; |
11506 | |
11507 | bool IsAnyViable = false; |
11508 | for (unsigned j = 0; j != array_lengthof(ViableForN); ++j) |
11509 | if (ViableForN[j]) { |
11510 | uint64_t N = j + 1; |
11511 | |
11512 | // The shuffle mask must be equal to (i * 2^N) % M. |
11513 | if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask)) |
11514 | IsAnyViable = true; |
11515 | else |
11516 | ViableForN[j] = false; |
11517 | } |
11518 | // Early exit if we exhaust the possible powers of two. |
11519 | if (!IsAnyViable) |
11520 | break; |
11521 | } |
11522 | |
11523 | for (unsigned j = 0; j != array_lengthof(ViableForN); ++j) |
11524 | if (ViableForN[j]) |
11525 | return j + 1; |
11526 | |
11527 | // Return 0 as there is no viable power of two. |
11528 | return 0; |
11529 | } |
11530 | |
11531 | /// \brief Generic lowering of v16i8 shuffles. |
11532 | /// |
11533 | /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to |
11534 | /// detect any complexity reducing interleaving. If that doesn't help, it uses |
11535 | /// UNPCK to spread the i8 elements across two i16-element vectors, and uses |
11536 | /// the existing lowering for v8i16 blends on each half, finally PACK-ing them |
11537 | /// back together. |
11538 | static SDValue lowerV16I8VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
11539 | const APInt &Zeroable, |
11540 | SDValue V1, SDValue V2, |
11541 | const X86Subtarget &Subtarget, |
11542 | SelectionDAG &DAG) { |
11543 | assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v16i8 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11543, __PRETTY_FUNCTION__)); |
11544 | assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v16i8 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11544, __PRETTY_FUNCTION__)); |
11545 | assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!")((Mask.size() == 16 && "Unexpected mask size for v16 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 16 && \"Unexpected mask size for v16 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11545, __PRETTY_FUNCTION__)); |
11546 | |
11547 | // Try to use shift instructions. |
11548 | if (SDValue Shift = lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, |
11549 | Zeroable, Subtarget, DAG)) |
11550 | return Shift; |
11551 | |
11552 | // Try to use byte rotation instructions. |
11553 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate( |
11554 | DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG)) |
11555 | return Rotate; |
11556 | |
11557 | // Try to use a zext lowering. |
11558 | if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend( |
11559 | DL, MVT::v16i8, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
11560 | return ZExt; |
11561 | |
11562 | // See if we can use SSE4A Extraction / Insertion. |
11563 | if (Subtarget.hasSSE4A()) |
11564 | if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, |
11565 | Zeroable, DAG)) |
11566 | return V; |
11567 | |
11568 | int NumV2Elements = count_if(Mask, [](int M) { return M >= 16; }); |
11569 | |
11570 | // For single-input shuffles, there are some nicer lowering tricks we can use. |
11571 | if (NumV2Elements == 0) { |
11572 | // Check for being able to broadcast a single element. |
11573 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast( |
11574 | DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG)) |
11575 | return Broadcast; |
11576 | |
11577 | // Check whether we can widen this to an i16 shuffle by duplicating bytes. |
11578 | // Notably, this handles splat and partial-splat shuffles more efficiently. |
11579 | // However, it only makes sense if the pre-duplication shuffle simplifies |
11580 | // things significantly. Currently, this means we need to be able to |
11581 | // express the pre-duplication shuffle as an i16 shuffle. |
11582 | // |
11583 | // FIXME: We should check for other patterns which can be widened into an |
11584 | // i16 shuffle as well. |
11585 | auto canWidenViaDuplication = [](ArrayRef<int> Mask) { |
11586 | for (int i = 0; i < 16; i += 2) |
11587 | if (Mask[i] >= 0 && Mask[i + 1] >= 0 && Mask[i] != Mask[i + 1]) |
11588 | return false; |
11589 | |
11590 | return true; |
11591 | }; |
11592 | auto tryToWidenViaDuplication = [&]() -> SDValue { |
11593 | if (!canWidenViaDuplication(Mask)) |
11594 | return SDValue(); |
11595 | SmallVector<int, 4> LoInputs; |
11596 | copy_if(Mask, std::back_inserter(LoInputs), |
11597 | [](int M) { return M >= 0 && M < 8; }); |
11598 | std::sort(LoInputs.begin(), LoInputs.end()); |
11599 | LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), |
11600 | LoInputs.end()); |
11601 | SmallVector<int, 4> HiInputs; |
11602 | copy_if(Mask, std::back_inserter(HiInputs), [](int M) { return M >= 8; }); |
11603 | std::sort(HiInputs.begin(), HiInputs.end()); |
11604 | HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), |
11605 | HiInputs.end()); |
11606 | |
11607 | bool TargetLo = LoInputs.size() >= HiInputs.size(); |
11608 | ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs; |
11609 | ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs; |
11610 | |
11611 | int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1}; |
11612 | SmallDenseMap<int, int, 8> LaneMap; |
11613 | for (int I : InPlaceInputs) { |
11614 | PreDupI16Shuffle[I/2] = I/2; |
11615 | LaneMap[I] = I; |
11616 | } |
11617 | int j = TargetLo ? 0 : 4, je = j + 4; |
11618 | for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) { |
11619 | // Check if j is already a shuffle of this input. This happens when |
11620 | // there are two adjacent bytes after we move the low one. |
11621 | if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) { |
11622 | // If we haven't yet mapped the input, search for a slot into which |
11623 | // we can map it. |
11624 | while (j < je && PreDupI16Shuffle[j] >= 0) |
11625 | ++j; |
11626 | |
11627 | if (j == je) |
11628 | // We can't place the inputs into a single half with a simple i16 shuffle, so bail. |
11629 | return SDValue(); |
11630 | |
11631 | // Map this input with the i16 shuffle. |
11632 | PreDupI16Shuffle[j] = MovingInputs[i] / 2; |
11633 | } |
11634 | |
11635 | // Update the lane map based on the mapping we ended up with. |
11636 | LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2; |
11637 | } |
11638 | V1 = DAG.getBitcast( |
11639 | MVT::v16i8, |
11640 | DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1), |
11641 | DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle)); |
11642 | |
11643 | // Unpack the bytes to form the i16s that will be shuffled into place. |
11644 | V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL, |
11645 | MVT::v16i8, V1, V1); |
11646 | |
11647 | int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1}; |
11648 | for (int i = 0; i < 16; ++i) |
11649 | if (Mask[i] >= 0) { |
11650 | int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8); |
11651 | assert(MappedMask < 8 && "Invalid v8 shuffle mask!")((MappedMask < 8 && "Invalid v8 shuffle mask!") ? static_cast <void> (0) : __assert_fail ("MappedMask < 8 && \"Invalid v8 shuffle mask!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11651, __PRETTY_FUNCTION__)); |
11652 | if (PostDupI16Shuffle[i / 2] < 0) |
11653 | PostDupI16Shuffle[i / 2] = MappedMask; |
11654 | else |
11655 | assert(PostDupI16Shuffle[i / 2] == MappedMask &&((PostDupI16Shuffle[i / 2] == MappedMask && "Conflicting entries in the original shuffle!" ) ? static_cast<void> (0) : __assert_fail ("PostDupI16Shuffle[i / 2] == MappedMask && \"Conflicting entries in the original shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11656, __PRETTY_FUNCTION__)) |
11656 | "Conflicting entries in the original shuffle!")((PostDupI16Shuffle[i / 2] == MappedMask && "Conflicting entries in the original shuffle!" ) ? static_cast<void> (0) : __assert_fail ("PostDupI16Shuffle[i / 2] == MappedMask && \"Conflicting entries in the original shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11656, __PRETTY_FUNCTION__)); |
11657 | } |
11658 | return DAG.getBitcast( |
11659 | MVT::v16i8, |
11660 | DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1), |
11661 | DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle)); |
11662 | }; |
11663 | if (SDValue V = tryToWidenViaDuplication()) |
11664 | return V; |
11665 | } |
11666 | |
11667 | if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, |
11668 | Zeroable, DAG)) |
11669 | return Masked; |
11670 | |
11671 | // Use dedicated unpack instructions for masks that match their pattern. |
11672 | if (SDValue V = |
11673 | lowerVectorShuffleWithUNPCK(DL, MVT::v16i8, Mask, V1, V2, DAG)) |
11674 | return V; |
11675 | |
11676 | // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly |
11677 | // with PSHUFB. It is important to do this before we attempt to generate any |
11678 | // blends but after all of the single-input lowerings. If the single input |
11679 | // lowerings can find an instruction sequence that is faster than a PSHUFB, we |
11680 | // want to preserve that and we can DAG combine any longer sequences into |
11681 | // a PSHUFB in the end. But once we start blending from multiple inputs, |
11682 | // the complexity of DAG combining bad patterns back into PSHUFB is too high, |
11683 | // and there are *very* few patterns that would actually be faster than the |
11684 | // PSHUFB approach because of its ability to zero lanes. |
11685 | // |
11686 | // FIXME: The only exceptions to the above are blends which are exact |
11687 | // interleavings with direct instructions supporting them. We currently don't |
11688 | // handle those well here. |
11689 | if (Subtarget.hasSSSE3()) { |
11690 | bool V1InUse = false; |
11691 | bool V2InUse = false; |
11692 | |
11693 | SDValue PSHUFB = lowerVectorShuffleAsBlendOfPSHUFBs( |
11694 | DL, MVT::v16i8, V1, V2, Mask, Zeroable, DAG, V1InUse, V2InUse); |
11695 | |
11696 | // If both V1 and V2 are in use and we can use a direct blend or an unpack, |
11697 | // do so. This avoids using them to handle blends-with-zero which is |
11698 | // important as a single pshufb is significantly faster for that. |
11699 | if (V1InUse && V2InUse) { |
11700 | if (Subtarget.hasSSE41()) |
11701 | if (SDValue Blend = lowerVectorShuffleAsBlend( |
11702 | DL, MVT::v16i8, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
11703 | return Blend; |
11704 | |
11705 | // We can use an unpack to do the blending rather than an or in some |
11706 | // cases. Even though the or may be (very minorly) more efficient, we |
11707 | // preference this lowering because there are common cases where part of |
11708 | // the complexity of the shuffles goes away when we do the final blend as |
11709 | // an unpack. |
11710 | // FIXME: It might be worth trying to detect if the unpack-feeding |
11711 | // shuffles will both be pshufb, in which case we shouldn't bother with |
11712 | // this. |
11713 | if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack( |
11714 | DL, MVT::v16i8, V1, V2, Mask, DAG)) |
11715 | return Unpack; |
11716 | } |
11717 | |
11718 | return PSHUFB; |
11719 | } |
11720 | |
11721 | // There are special ways we can lower some single-element blends. |
11722 | if (NumV2Elements == 1) |
11723 | if (SDValue V = lowerVectorShuffleAsElementInsertion( |
11724 | DL, MVT::v16i8, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
11725 | return V; |
11726 | |
11727 | if (SDValue BitBlend = |
11728 | lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG)) |
11729 | return BitBlend; |
11730 | |
11731 | // Check whether a compaction lowering can be done. This handles shuffles |
11732 | // which take every Nth element for some even N. See the helper function for |
11733 | // details. |
11734 | // |
11735 | // We special case these as they can be particularly efficiently handled with |
11736 | // the PACKUSB instruction on x86 and they show up in common patterns of |
11737 | // rearranging bytes to truncate wide elements. |
11738 | bool IsSingleInput = V2.isUndef(); |
11739 | if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask, IsSingleInput)) { |
11740 | // NumEvenDrops is the power of two stride of the elements. Another way of |
11741 | // thinking about it is that we need to drop the even elements this many |
11742 | // times to get the original input. |
11743 | |
11744 | // First we need to zero all the dropped bytes. |
11745 | assert(NumEvenDrops <= 3 &&((NumEvenDrops <= 3 && "No support for dropping even elements more than 3 times." ) ? static_cast<void> (0) : __assert_fail ("NumEvenDrops <= 3 && \"No support for dropping even elements more than 3 times.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11746, __PRETTY_FUNCTION__)) |
11746 | "No support for dropping even elements more than 3 times.")((NumEvenDrops <= 3 && "No support for dropping even elements more than 3 times." ) ? static_cast<void> (0) : __assert_fail ("NumEvenDrops <= 3 && \"No support for dropping even elements more than 3 times.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11746, __PRETTY_FUNCTION__)); |
11747 | // We use the mask type to pick which bytes are preserved based on how many |
11748 | // elements are dropped. |
11749 | MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 }; |
11750 | SDValue ByteClearMask = DAG.getBitcast( |
11751 | MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1])); |
11752 | V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask); |
11753 | if (!IsSingleInput) |
11754 | V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask); |
11755 | |
11756 | // Now pack things back together. |
11757 | V1 = DAG.getBitcast(MVT::v8i16, V1); |
11758 | V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2); |
11759 | SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2); |
11760 | for (int i = 1; i < NumEvenDrops; ++i) { |
11761 | Result = DAG.getBitcast(MVT::v8i16, Result); |
11762 | Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result); |
11763 | } |
11764 | |
11765 | return Result; |
11766 | } |
11767 | |
11768 | // Handle multi-input cases by blending single-input shuffles. |
11769 | if (NumV2Elements > 0) |
11770 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2, |
11771 | Mask, DAG); |
11772 | |
11773 | // The fallback path for single-input shuffles widens this into two v8i16 |
11774 | // vectors with unpacks, shuffles those, and then pulls them back together |
11775 | // with a pack. |
11776 | SDValue V = V1; |
11777 | |
11778 | std::array<int, 8> LoBlendMask = {{-1, -1, -1, -1, -1, -1, -1, -1}}; |
11779 | std::array<int, 8> HiBlendMask = {{-1, -1, -1, -1, -1, -1, -1, -1}}; |
11780 | for (int i = 0; i < 16; ++i) |
11781 | if (Mask[i] >= 0) |
11782 | (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i]; |
11783 | |
11784 | SDValue VLoHalf, VHiHalf; |
11785 | // Check if any of the odd lanes in the v16i8 are used. If not, we can mask |
11786 | // them out and avoid using UNPCK{L,H} to extract the elements of V as |
11787 | // i16s. |
11788 | if (none_of(LoBlendMask, [](int M) { return M >= 0 && M % 2 == 1; }) && |
11789 | none_of(HiBlendMask, [](int M) { return M >= 0 && M % 2 == 1; })) { |
11790 | // Use a mask to drop the high bytes. |
11791 | VLoHalf = DAG.getBitcast(MVT::v8i16, V); |
11792 | VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf, |
11793 | DAG.getConstant(0x00FF, DL, MVT::v8i16)); |
11794 | |
11795 | // This will be a single vector shuffle instead of a blend so nuke VHiHalf. |
11796 | VHiHalf = DAG.getUNDEF(MVT::v8i16); |
11797 | |
11798 | // Squash the masks to point directly into VLoHalf. |
11799 | for (int &M : LoBlendMask) |
11800 | if (M >= 0) |
11801 | M /= 2; |
11802 | for (int &M : HiBlendMask) |
11803 | if (M >= 0) |
11804 | M /= 2; |
11805 | } else { |
11806 | // Otherwise just unpack the low half of V into VLoHalf and the high half into |
11807 | // VHiHalf so that we can blend them as i16s. |
11808 | SDValue Zero = getZeroVector(MVT::v16i8, Subtarget, DAG, DL); |
11809 | |
11810 | VLoHalf = DAG.getBitcast( |
11811 | MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero)); |
11812 | VHiHalf = DAG.getBitcast( |
11813 | MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero)); |
11814 | } |
11815 | |
11816 | SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask); |
11817 | SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask); |
11818 | |
11819 | return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV); |
11820 | } |
11821 | |
11822 | /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles. |
11823 | /// |
11824 | /// This routine breaks down the specific type of 128-bit shuffle and |
11825 | /// dispatches to the lowering routines accordingly. |
11826 | static SDValue lower128BitVectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
11827 | MVT VT, SDValue V1, SDValue V2, |
11828 | const APInt &Zeroable, |
11829 | const X86Subtarget &Subtarget, |
11830 | SelectionDAG &DAG) { |
11831 | switch (VT.SimpleTy) { |
11832 | case MVT::v2i64: |
11833 | return lowerV2I64VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
11834 | case MVT::v2f64: |
11835 | return lowerV2F64VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
11836 | case MVT::v4i32: |
11837 | return lowerV4I32VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
11838 | case MVT::v4f32: |
11839 | return lowerV4F32VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
11840 | case MVT::v8i16: |
11841 | return lowerV8I16VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
11842 | case MVT::v16i8: |
11843 | return lowerV16I8VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
11844 | |
11845 | default: |
11846 | llvm_unreachable("Unimplemented!")::llvm::llvm_unreachable_internal("Unimplemented!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11846); |
11847 | } |
11848 | } |
11849 | |
11850 | /// \brief Generic routine to split vector shuffle into half-sized shuffles. |
11851 | /// |
11852 | /// This routine just extracts two subvectors, shuffles them independently, and |
11853 | /// then concatenates them back together. This should work effectively with all |
11854 | /// AVX vector shuffle types. |
11855 | static SDValue splitAndLowerVectorShuffle(const SDLoc &DL, MVT VT, SDValue V1, |
11856 | SDValue V2, ArrayRef<int> Mask, |
11857 | SelectionDAG &DAG) { |
11858 | assert(VT.getSizeInBits() >= 256 &&((VT.getSizeInBits() >= 256 && "Only for 256-bit or wider vector shuffles!" ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() >= 256 && \"Only for 256-bit or wider vector shuffles!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11859, __PRETTY_FUNCTION__)) |
11859 | "Only for 256-bit or wider vector shuffles!")((VT.getSizeInBits() >= 256 && "Only for 256-bit or wider vector shuffles!" ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() >= 256 && \"Only for 256-bit or wider vector shuffles!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11859, __PRETTY_FUNCTION__)); |
11860 | assert(V1.getSimpleValueType() == VT && "Bad operand type!")((V1.getSimpleValueType() == VT && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == VT && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11860, __PRETTY_FUNCTION__)); |
11861 | assert(V2.getSimpleValueType() == VT && "Bad operand type!")((V2.getSimpleValueType() == VT && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == VT && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11861, __PRETTY_FUNCTION__)); |
11862 | |
11863 | ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2); |
11864 | ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2); |
11865 | |
11866 | int NumElements = VT.getVectorNumElements(); |
11867 | int SplitNumElements = NumElements / 2; |
11868 | MVT ScalarVT = VT.getVectorElementType(); |
11869 | MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2); |
11870 | |
11871 | // Rather than splitting build-vectors, just build two narrower build |
11872 | // vectors. This helps shuffling with splats and zeros. |
11873 | auto SplitVector = [&](SDValue V) { |
11874 | V = peekThroughBitcasts(V); |
11875 | |
11876 | MVT OrigVT = V.getSimpleValueType(); |
11877 | int OrigNumElements = OrigVT.getVectorNumElements(); |
11878 | int OrigSplitNumElements = OrigNumElements / 2; |
11879 | MVT OrigScalarVT = OrigVT.getVectorElementType(); |
11880 | MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2); |
11881 | |
11882 | SDValue LoV, HiV; |
11883 | |
11884 | auto *BV = dyn_cast<BuildVectorSDNode>(V); |
11885 | if (!BV) { |
11886 | LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V, |
11887 | DAG.getIntPtrConstant(0, DL)); |
11888 | HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V, |
11889 | DAG.getIntPtrConstant(OrigSplitNumElements, DL)); |
11890 | } else { |
11891 | |
11892 | SmallVector<SDValue, 16> LoOps, HiOps; |
11893 | for (int i = 0; i < OrigSplitNumElements; ++i) { |
11894 | LoOps.push_back(BV->getOperand(i)); |
11895 | HiOps.push_back(BV->getOperand(i + OrigSplitNumElements)); |
11896 | } |
11897 | LoV = DAG.getBuildVector(OrigSplitVT, DL, LoOps); |
11898 | HiV = DAG.getBuildVector(OrigSplitVT, DL, HiOps); |
11899 | } |
11900 | return std::make_pair(DAG.getBitcast(SplitVT, LoV), |
11901 | DAG.getBitcast(SplitVT, HiV)); |
11902 | }; |
11903 | |
11904 | SDValue LoV1, HiV1, LoV2, HiV2; |
11905 | std::tie(LoV1, HiV1) = SplitVector(V1); |
11906 | std::tie(LoV2, HiV2) = SplitVector(V2); |
11907 | |
11908 | // Now create two 4-way blends of these half-width vectors. |
11909 | auto HalfBlend = [&](ArrayRef<int> HalfMask) { |
11910 | bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false; |
11911 | SmallVector<int, 32> V1BlendMask((unsigned)SplitNumElements, -1); |
11912 | SmallVector<int, 32> V2BlendMask((unsigned)SplitNumElements, -1); |
11913 | SmallVector<int, 32> BlendMask((unsigned)SplitNumElements, -1); |
11914 | for (int i = 0; i < SplitNumElements; ++i) { |
11915 | int M = HalfMask[i]; |
11916 | if (M >= NumElements) { |
11917 | if (M >= NumElements + SplitNumElements) |
11918 | UseHiV2 = true; |
11919 | else |
11920 | UseLoV2 = true; |
11921 | V2BlendMask[i] = M - NumElements; |
11922 | BlendMask[i] = SplitNumElements + i; |
11923 | } else if (M >= 0) { |
11924 | if (M >= SplitNumElements) |
11925 | UseHiV1 = true; |
11926 | else |
11927 | UseLoV1 = true; |
11928 | V1BlendMask[i] = M; |
11929 | BlendMask[i] = i; |
11930 | } |
11931 | } |
11932 | |
11933 | // Because the lowering happens after all combining takes place, we need to |
11934 | // manually combine these blend masks as much as possible so that we create |
11935 | // a minimal number of high-level vector shuffle nodes. |
11936 | |
11937 | // First try just blending the halves of V1 or V2. |
11938 | if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2) |
11939 | return DAG.getUNDEF(SplitVT); |
11940 | if (!UseLoV2 && !UseHiV2) |
11941 | return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask); |
11942 | if (!UseLoV1 && !UseHiV1) |
11943 | return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask); |
11944 | |
11945 | SDValue V1Blend, V2Blend; |
11946 | if (UseLoV1 && UseHiV1) { |
11947 | V1Blend = |
11948 | DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask); |
11949 | } else { |
11950 | // We only use half of V1 so map the usage down into the final blend mask. |
11951 | V1Blend = UseLoV1 ? LoV1 : HiV1; |
11952 | for (int i = 0; i < SplitNumElements; ++i) |
11953 | if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements) |
11954 | BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements); |
11955 | } |
11956 | if (UseLoV2 && UseHiV2) { |
11957 | V2Blend = |
11958 | DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask); |
11959 | } else { |
11960 | // We only use half of V2 so map the usage down into the final blend mask. |
11961 | V2Blend = UseLoV2 ? LoV2 : HiV2; |
11962 | for (int i = 0; i < SplitNumElements; ++i) |
11963 | if (BlendMask[i] >= SplitNumElements) |
11964 | BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0); |
11965 | } |
11966 | return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask); |
11967 | }; |
11968 | SDValue Lo = HalfBlend(LoMask); |
11969 | SDValue Hi = HalfBlend(HiMask); |
11970 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi); |
11971 | } |
11972 | |
11973 | /// \brief Either split a vector in halves or decompose the shuffles and the |
11974 | /// blend. |
11975 | /// |
11976 | /// This is provided as a good fallback for many lowerings of non-single-input |
11977 | /// shuffles with more than one 128-bit lane. In those cases, we want to select |
11978 | /// between splitting the shuffle into 128-bit components and stitching those |
11979 | /// back together vs. extracting the single-input shuffles and blending those |
11980 | /// results. |
11981 | static SDValue lowerVectorShuffleAsSplitOrBlend(const SDLoc &DL, MVT VT, |
11982 | SDValue V1, SDValue V2, |
11983 | ArrayRef<int> Mask, |
11984 | SelectionDAG &DAG) { |
11985 | assert(!V2.isUndef() && "This routine must not be used to lower single-input "((!V2.isUndef() && "This routine must not be used to lower single-input " "shuffles as it could then recurse on itself.") ? static_cast <void> (0) : __assert_fail ("!V2.isUndef() && \"This routine must not be used to lower single-input \" \"shuffles as it could then recurse on itself.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11986, __PRETTY_FUNCTION__)) |
11986 | "shuffles as it could then recurse on itself.")((!V2.isUndef() && "This routine must not be used to lower single-input " "shuffles as it could then recurse on itself.") ? static_cast <void> (0) : __assert_fail ("!V2.isUndef() && \"This routine must not be used to lower single-input \" \"shuffles as it could then recurse on itself.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 11986, __PRETTY_FUNCTION__)); |
11987 | int Size = Mask.size(); |
11988 | |
11989 | // If this can be modeled as a broadcast of two elements followed by a blend, |
11990 | // prefer that lowering. This is especially important because broadcasts can |
11991 | // often fold with memory operands. |
11992 | auto DoBothBroadcast = [&] { |
11993 | int V1BroadcastIdx = -1, V2BroadcastIdx = -1; |
11994 | for (int M : Mask) |
11995 | if (M >= Size) { |
11996 | if (V2BroadcastIdx < 0) |
11997 | V2BroadcastIdx = M - Size; |
11998 | else if (M - Size != V2BroadcastIdx) |
11999 | return false; |
12000 | } else if (M >= 0) { |
12001 | if (V1BroadcastIdx < 0) |
12002 | V1BroadcastIdx = M; |
12003 | else if (M != V1BroadcastIdx) |
12004 | return false; |
12005 | } |
12006 | return true; |
12007 | }; |
12008 | if (DoBothBroadcast()) |
12009 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, |
12010 | DAG); |
12011 | |
12012 | // If the inputs all stem from a single 128-bit lane of each input, then we |
12013 | // split them rather than blending because the split will decompose to |
12014 | // unusually few instructions. |
12015 | int LaneCount = VT.getSizeInBits() / 128; |
12016 | int LaneSize = Size / LaneCount; |
12017 | SmallBitVector LaneInputs[2]; |
12018 | LaneInputs[0].resize(LaneCount, false); |
12019 | LaneInputs[1].resize(LaneCount, false); |
12020 | for (int i = 0; i < Size; ++i) |
12021 | if (Mask[i] >= 0) |
12022 | LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true; |
12023 | if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1) |
12024 | return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG); |
12025 | |
12026 | // Otherwise, just fall back to decomposed shuffles and a blend. This requires |
12027 | // that the decomposed single-input shuffles don't end up here. |
12028 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG); |
12029 | } |
12030 | |
12031 | /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as |
12032 | /// a permutation and blend of those lanes. |
12033 | /// |
12034 | /// This essentially blends the out-of-lane inputs to each lane into the lane |
12035 | /// from a permuted copy of the vector. This lowering strategy results in four |
12036 | /// instructions in the worst case for a single-input cross lane shuffle which |
12037 | /// is lower than any other fully general cross-lane shuffle strategy I'm aware |
12038 | /// of. Special cases for each particular shuffle pattern should be handled |
12039 | /// prior to trying this lowering. |
12040 | static SDValue lowerVectorShuffleAsLanePermuteAndBlend(const SDLoc &DL, MVT VT, |
12041 | SDValue V1, SDValue V2, |
12042 | ArrayRef<int> Mask, |
12043 | SelectionDAG &DAG) { |
12044 | // FIXME: This should probably be generalized for 512-bit vectors as well. |
12045 | assert(VT.is256BitVector() && "Only for 256-bit vector shuffles!")((VT.is256BitVector() && "Only for 256-bit vector shuffles!" ) ? static_cast<void> (0) : __assert_fail ("VT.is256BitVector() && \"Only for 256-bit vector shuffles!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12045, __PRETTY_FUNCTION__)); |
12046 | int Size = Mask.size(); |
12047 | int LaneSize = Size / 2; |
12048 | |
12049 | // If there are only inputs from one 128-bit lane, splitting will in fact be |
12050 | // less expensive. The flags track whether the given lane contains an element |
12051 | // that crosses to another lane. |
12052 | bool LaneCrossing[2] = {false, false}; |
12053 | for (int i = 0; i < Size; ++i) |
12054 | if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize) |
12055 | LaneCrossing[(Mask[i] % Size) / LaneSize] = true; |
12056 | if (!LaneCrossing[0] || !LaneCrossing[1]) |
12057 | return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG); |
12058 | |
12059 | assert(V2.isUndef() &&((V2.isUndef() && "This last part of this routine only works on single input shuffles" ) ? static_cast<void> (0) : __assert_fail ("V2.isUndef() && \"This last part of this routine only works on single input shuffles\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12060, __PRETTY_FUNCTION__)) |
12060 | "This last part of this routine only works on single input shuffles")((V2.isUndef() && "This last part of this routine only works on single input shuffles" ) ? static_cast<void> (0) : __assert_fail ("V2.isUndef() && \"This last part of this routine only works on single input shuffles\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12060, __PRETTY_FUNCTION__)); |
12061 | |
12062 | SmallVector<int, 32> FlippedBlendMask(Size); |
12063 | for (int i = 0; i < Size; ++i) |
12064 | FlippedBlendMask[i] = |
12065 | Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize) |
12066 | ? Mask[i] |
12067 | : Mask[i] % LaneSize + |
12068 | (i / LaneSize) * LaneSize + Size); |
12069 | |
12070 | // Flip the vector, and blend the results which should now be in-lane. The |
12071 | // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and |
12072 | // 5 for the high source. The value 3 selects the high half of source 2 and |
12073 | // the value 2 selects the low half of source 2. We only use source 2 to |
12074 | // allow folding it into a memory operand. |
12075 | unsigned PERMMask = 3 | 2 << 4; |
12076 | SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT), |
12077 | V1, DAG.getConstant(PERMMask, DL, MVT::i8)); |
12078 | return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask); |
12079 | } |
12080 | |
12081 | /// \brief Handle lowering 2-lane 128-bit shuffles. |
12082 | static SDValue lowerV2X128VectorShuffle(const SDLoc &DL, MVT VT, SDValue V1, |
12083 | SDValue V2, ArrayRef<int> Mask, |
12084 | const APInt &Zeroable, |
12085 | const X86Subtarget &Subtarget, |
12086 | SelectionDAG &DAG) { |
12087 | SmallVector<int, 4> WidenedMask; |
12088 | if (!canWidenShuffleElements(Mask, WidenedMask)) |
12089 | return SDValue(); |
12090 | |
12091 | // TODO: If minimizing size and one of the inputs is a zero vector and the |
12092 | // the zero vector has only one use, we could use a VPERM2X128 to save the |
12093 | // instruction bytes needed to explicitly generate the zero vector. |
12094 | |
12095 | // Blends are faster and handle all the non-lane-crossing cases. |
12096 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask, |
12097 | Zeroable, Subtarget, DAG)) |
12098 | return Blend; |
12099 | |
12100 | bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode()); |
12101 | bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode()); |
12102 | |
12103 | // If either input operand is a zero vector, use VPERM2X128 because its mask |
12104 | // allows us to replace the zero input with an implicit zero. |
12105 | if (!IsV1Zero && !IsV2Zero) { |
12106 | // Check for patterns which can be matched with a single insert of a 128-bit |
12107 | // subvector. |
12108 | bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1}); |
12109 | if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) { |
12110 | // With AVX2, use VPERMQ/VPERMPD to allow memory folding. |
12111 | if (Subtarget.hasAVX2() && V2.isUndef()) |
12112 | return SDValue(); |
12113 | |
12114 | // With AVX1, use vperm2f128 (below) to allow load folding. Otherwise, |
12115 | // this will likely become vinsertf128 which can't fold a 256-bit memop. |
12116 | if (!isa<LoadSDNode>(peekThroughBitcasts(V1))) { |
12117 | MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(), |
12118 | VT.getVectorNumElements() / 2); |
12119 | SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1, |
12120 | DAG.getIntPtrConstant(0, DL)); |
12121 | SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, |
12122 | OnlyUsesV1 ? V1 : V2, |
12123 | DAG.getIntPtrConstant(0, DL)); |
12124 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV); |
12125 | } |
12126 | } |
12127 | } |
12128 | |
12129 | // Otherwise form a 128-bit permutation. After accounting for undefs, |
12130 | // convert the 64-bit shuffle mask selection values into 128-bit |
12131 | // selection bits by dividing the indexes by 2 and shifting into positions |
12132 | // defined by a vperm2*128 instruction's immediate control byte. |
12133 | |
12134 | // The immediate permute control byte looks like this: |
12135 | // [1:0] - select 128 bits from sources for low half of destination |
12136 | // [2] - ignore |
12137 | // [3] - zero low half of destination |
12138 | // [5:4] - select 128 bits from sources for high half of destination |
12139 | // [6] - ignore |
12140 | // [7] - zero high half of destination |
12141 | |
12142 | int MaskLO = WidenedMask[0] < 0 ? 0 : WidenedMask[0]; |
12143 | int MaskHI = WidenedMask[1] < 0 ? 0 : WidenedMask[1]; |
12144 | |
12145 | unsigned PermMask = MaskLO | (MaskHI << 4); |
12146 | |
12147 | // If either input is a zero vector, replace it with an undef input. |
12148 | // Shuffle mask values < 4 are selecting elements of V1. |
12149 | // Shuffle mask values >= 4 are selecting elements of V2. |
12150 | // Adjust each half of the permute mask by clearing the half that was |
12151 | // selecting the zero vector and setting the zero mask bit. |
12152 | if (IsV1Zero) { |
12153 | V1 = DAG.getUNDEF(VT); |
12154 | if (MaskLO < 2) |
12155 | PermMask = (PermMask & 0xf0) | 0x08; |
12156 | if (MaskHI < 2) |
12157 | PermMask = (PermMask & 0x0f) | 0x80; |
12158 | } |
12159 | if (IsV2Zero) { |
12160 | V2 = DAG.getUNDEF(VT); |
12161 | if (MaskLO >= 2) |
12162 | PermMask = (PermMask & 0xf0) | 0x08; |
12163 | if (MaskHI >= 2) |
12164 | PermMask = (PermMask & 0x0f) | 0x80; |
12165 | } |
12166 | |
12167 | return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2, |
12168 | DAG.getConstant(PermMask, DL, MVT::i8)); |
12169 | } |
12170 | |
12171 | /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then |
12172 | /// shuffling each lane. |
12173 | /// |
12174 | /// This will only succeed when the result of fixing the 128-bit lanes results |
12175 | /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in |
12176 | /// each 128-bit lanes. This handles many cases where we can quickly blend away |
12177 | /// the lane crosses early and then use simpler shuffles within each lane. |
12178 | /// |
12179 | /// FIXME: It might be worthwhile at some point to support this without |
12180 | /// requiring the 128-bit lane-relative shuffles to be repeating, but currently |
12181 | /// in x86 only floating point has interesting non-repeating shuffles, and even |
12182 | /// those are still *marginally* more expensive. |
12183 | static SDValue lowerVectorShuffleByMerging128BitLanes( |
12184 | const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, |
12185 | const X86Subtarget &Subtarget, SelectionDAG &DAG) { |
12186 | assert(!V2.isUndef() && "This is only useful with multiple inputs.")((!V2.isUndef() && "This is only useful with multiple inputs." ) ? static_cast<void> (0) : __assert_fail ("!V2.isUndef() && \"This is only useful with multiple inputs.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12186, __PRETTY_FUNCTION__)); |
12187 | |
12188 | int Size = Mask.size(); |
12189 | int LaneSize = 128 / VT.getScalarSizeInBits(); |
12190 | int NumLanes = Size / LaneSize; |
12191 | assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.")((NumLanes > 1 && "Only handles 256-bit and wider shuffles." ) ? static_cast<void> (0) : __assert_fail ("NumLanes > 1 && \"Only handles 256-bit and wider shuffles.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12191, __PRETTY_FUNCTION__)); |
12192 | |
12193 | // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also |
12194 | // check whether the in-128-bit lane shuffles share a repeating pattern. |
12195 | SmallVector<int, 4> Lanes((unsigned)NumLanes, -1); |
12196 | SmallVector<int, 4> InLaneMask((unsigned)LaneSize, -1); |
12197 | for (int i = 0; i < Size; ++i) { |
12198 | if (Mask[i] < 0) |
12199 | continue; |
12200 | |
12201 | int j = i / LaneSize; |
12202 | |
12203 | if (Lanes[j] < 0) { |
12204 | // First entry we've seen for this lane. |
12205 | Lanes[j] = Mask[i] / LaneSize; |
12206 | } else if (Lanes[j] != Mask[i] / LaneSize) { |
12207 | // This doesn't match the lane selected previously! |
12208 | return SDValue(); |
12209 | } |
12210 | |
12211 | // Check that within each lane we have a consistent shuffle mask. |
12212 | int k = i % LaneSize; |
12213 | if (InLaneMask[k] < 0) { |
12214 | InLaneMask[k] = Mask[i] % LaneSize; |
12215 | } else if (InLaneMask[k] != Mask[i] % LaneSize) { |
12216 | // This doesn't fit a repeating in-lane mask. |
12217 | return SDValue(); |
12218 | } |
12219 | } |
12220 | |
12221 | // First shuffle the lanes into place. |
12222 | MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64, |
12223 | VT.getSizeInBits() / 64); |
12224 | SmallVector<int, 8> LaneMask((unsigned)NumLanes * 2, -1); |
12225 | for (int i = 0; i < NumLanes; ++i) |
12226 | if (Lanes[i] >= 0) { |
12227 | LaneMask[2 * i + 0] = 2*Lanes[i] + 0; |
12228 | LaneMask[2 * i + 1] = 2*Lanes[i] + 1; |
12229 | } |
12230 | |
12231 | V1 = DAG.getBitcast(LaneVT, V1); |
12232 | V2 = DAG.getBitcast(LaneVT, V2); |
12233 | SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask); |
12234 | |
12235 | // Cast it back to the type we actually want. |
12236 | LaneShuffle = DAG.getBitcast(VT, LaneShuffle); |
12237 | |
12238 | // Now do a simple shuffle that isn't lane crossing. |
12239 | SmallVector<int, 8> NewMask((unsigned)Size, -1); |
12240 | for (int i = 0; i < Size; ++i) |
12241 | if (Mask[i] >= 0) |
12242 | NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize; |
12243 | assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&((!is128BitLaneCrossingShuffleMask(VT, NewMask) && "Must not introduce lane crosses at this point!" ) ? static_cast<void> (0) : __assert_fail ("!is128BitLaneCrossingShuffleMask(VT, NewMask) && \"Must not introduce lane crosses at this point!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12244, __PRETTY_FUNCTION__)) |
12244 | "Must not introduce lane crosses at this point!")((!is128BitLaneCrossingShuffleMask(VT, NewMask) && "Must not introduce lane crosses at this point!" ) ? static_cast<void> (0) : __assert_fail ("!is128BitLaneCrossingShuffleMask(VT, NewMask) && \"Must not introduce lane crosses at this point!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12244, __PRETTY_FUNCTION__)); |
12245 | |
12246 | return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask); |
12247 | } |
12248 | |
12249 | /// Lower shuffles where an entire half of a 256-bit vector is UNDEF. |
12250 | /// This allows for fast cases such as subvector extraction/insertion |
12251 | /// or shuffling smaller vector types which can lower more efficiently. |
12252 | static SDValue lowerVectorShuffleWithUndefHalf(const SDLoc &DL, MVT VT, |
12253 | SDValue V1, SDValue V2, |
12254 | ArrayRef<int> Mask, |
12255 | const X86Subtarget &Subtarget, |
12256 | SelectionDAG &DAG) { |
12257 | assert(VT.is256BitVector() && "Expected 256-bit vector")((VT.is256BitVector() && "Expected 256-bit vector") ? static_cast<void> (0) : __assert_fail ("VT.is256BitVector() && \"Expected 256-bit vector\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12257, __PRETTY_FUNCTION__)); |
12258 | |
12259 | unsigned NumElts = VT.getVectorNumElements(); |
12260 | unsigned HalfNumElts = NumElts / 2; |
12261 | MVT HalfVT = MVT::getVectorVT(VT.getVectorElementType(), HalfNumElts); |
12262 | |
12263 | bool UndefLower = isUndefInRange(Mask, 0, HalfNumElts); |
12264 | bool UndefUpper = isUndefInRange(Mask, HalfNumElts, HalfNumElts); |
12265 | if (!UndefLower && !UndefUpper) |
12266 | return SDValue(); |
12267 | |
12268 | // Upper half is undef and lower half is whole upper subvector. |
12269 | // e.g. vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u> |
12270 | if (UndefUpper && |
12271 | isSequentialOrUndefInRange(Mask, 0, HalfNumElts, HalfNumElts)) { |
12272 | SDValue Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V1, |
12273 | DAG.getIntPtrConstant(HalfNumElts, DL)); |
12274 | return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), Hi, |
12275 | DAG.getIntPtrConstant(0, DL)); |
12276 | } |
12277 | |
12278 | // Lower half is undef and upper half is whole lower subvector. |
12279 | // e.g. vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1> |
12280 | if (UndefLower && |
12281 | isSequentialOrUndefInRange(Mask, HalfNumElts, HalfNumElts, 0)) { |
12282 | SDValue Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V1, |
12283 | DAG.getIntPtrConstant(0, DL)); |
12284 | return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), Hi, |
12285 | DAG.getIntPtrConstant(HalfNumElts, DL)); |
12286 | } |
12287 | |
12288 | // If the shuffle only uses two of the four halves of the input operands, |
12289 | // then extract them and perform the 'half' shuffle at half width. |
12290 | // e.g. vector_shuffle <X, X, X, X, u, u, u, u> or <X, X, u, u> |
12291 | int HalfIdx1 = -1, HalfIdx2 = -1; |
12292 | SmallVector<int, 8> HalfMask(HalfNumElts); |
12293 | unsigned Offset = UndefLower ? HalfNumElts : 0; |
12294 | for (unsigned i = 0; i != HalfNumElts; ++i) { |
12295 | int M = Mask[i + Offset]; |
12296 | if (M < 0) { |
12297 | HalfMask[i] = M; |
12298 | continue; |
12299 | } |
12300 | |
12301 | // Determine which of the 4 half vectors this element is from. |
12302 | // i.e. 0 = Lower V1, 1 = Upper V1, 2 = Lower V2, 3 = Upper V2. |
12303 | int HalfIdx = M / HalfNumElts; |
12304 | |
12305 | // Determine the element index into its half vector source. |
12306 | int HalfElt = M % HalfNumElts; |
12307 | |
12308 | // We can shuffle with up to 2 half vectors, set the new 'half' |
12309 | // shuffle mask accordingly. |
12310 | if (HalfIdx1 < 0 || HalfIdx1 == HalfIdx) { |
12311 | HalfMask[i] = HalfElt; |
12312 | HalfIdx1 = HalfIdx; |
12313 | continue; |
12314 | } |
12315 | if (HalfIdx2 < 0 || HalfIdx2 == HalfIdx) { |
12316 | HalfMask[i] = HalfElt + HalfNumElts; |
12317 | HalfIdx2 = HalfIdx; |
12318 | continue; |
12319 | } |
12320 | |
12321 | // Too many half vectors referenced. |
12322 | return SDValue(); |
12323 | } |
12324 | assert(HalfMask.size() == HalfNumElts && "Unexpected shuffle mask length")((HalfMask.size() == HalfNumElts && "Unexpected shuffle mask length" ) ? static_cast<void> (0) : __assert_fail ("HalfMask.size() == HalfNumElts && \"Unexpected shuffle mask length\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12324, __PRETTY_FUNCTION__)); |
12325 | |
12326 | // Only shuffle the halves of the inputs when useful. |
12327 | int NumLowerHalves = |
12328 | (HalfIdx1 == 0 || HalfIdx1 == 2) + (HalfIdx2 == 0 || HalfIdx2 == 2); |
12329 | int NumUpperHalves = |
12330 | (HalfIdx1 == 1 || HalfIdx1 == 3) + (HalfIdx2 == 1 || HalfIdx2 == 3); |
12331 | |
12332 | // uuuuXXXX - don't extract uppers just to insert again. |
12333 | if (UndefLower && NumUpperHalves != 0) |
12334 | return SDValue(); |
12335 | |
12336 | // XXXXuuuu - don't extract both uppers, instead shuffle and then extract. |
12337 | if (UndefUpper && NumUpperHalves == 2) |
12338 | return SDValue(); |
12339 | |
12340 | // AVX2 - XXXXuuuu - always extract lowers. |
12341 | if (Subtarget.hasAVX2() && !(UndefUpper && NumUpperHalves == 0)) { |
12342 | // AVX2 supports efficient immediate 64-bit element cross-lane shuffles. |
12343 | if (VT == MVT::v4f64 || VT == MVT::v4i64) |
12344 | return SDValue(); |
12345 | // AVX2 supports variable 32-bit element cross-lane shuffles. |
12346 | if (VT == MVT::v8f32 || VT == MVT::v8i32) { |
12347 | // XXXXuuuu - don't extract lowers and uppers. |
12348 | if (UndefUpper && NumLowerHalves != 0 && NumUpperHalves != 0) |
12349 | return SDValue(); |
12350 | } |
12351 | } |
12352 | |
12353 | auto GetHalfVector = [&](int HalfIdx) { |
12354 | if (HalfIdx < 0) |
12355 | return DAG.getUNDEF(HalfVT); |
12356 | SDValue V = (HalfIdx < 2 ? V1 : V2); |
12357 | HalfIdx = (HalfIdx % 2) * HalfNumElts; |
12358 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V, |
12359 | DAG.getIntPtrConstant(HalfIdx, DL)); |
12360 | }; |
12361 | |
12362 | SDValue Half1 = GetHalfVector(HalfIdx1); |
12363 | SDValue Half2 = GetHalfVector(HalfIdx2); |
12364 | SDValue V = DAG.getVectorShuffle(HalfVT, DL, Half1, Half2, HalfMask); |
12365 | return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), V, |
12366 | DAG.getIntPtrConstant(Offset, DL)); |
12367 | } |
12368 | |
12369 | /// \brief Test whether the specified input (0 or 1) is in-place blended by the |
12370 | /// given mask. |
12371 | /// |
12372 | /// This returns true if the elements from a particular input are already in the |
12373 | /// slot required by the given mask and require no permutation. |
12374 | static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) { |
12375 | assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.")(((Input == 0 || Input == 1) && "Only two inputs to shuffles." ) ? static_cast<void> (0) : __assert_fail ("(Input == 0 || Input == 1) && \"Only two inputs to shuffles.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12375, __PRETTY_FUNCTION__)); |
12376 | int Size = Mask.size(); |
12377 | for (int i = 0; i < Size; ++i) |
12378 | if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i) |
12379 | return false; |
12380 | |
12381 | return true; |
12382 | } |
12383 | |
12384 | /// Handle case where shuffle sources are coming from the same 128-bit lane and |
12385 | /// every lane can be represented as the same repeating mask - allowing us to |
12386 | /// shuffle the sources with the repeating shuffle and then permute the result |
12387 | /// to the destination lanes. |
12388 | static SDValue lowerShuffleAsRepeatedMaskAndLanePermute( |
12389 | const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, |
12390 | const X86Subtarget &Subtarget, SelectionDAG &DAG) { |
12391 | int NumElts = VT.getVectorNumElements(); |
12392 | int NumLanes = VT.getSizeInBits() / 128; |
12393 | int NumLaneElts = NumElts / NumLanes; |
12394 | |
12395 | // On AVX2 we may be able to just shuffle the lowest elements and then |
12396 | // broadcast the result. |
12397 | if (Subtarget.hasAVX2()) { |
12398 | for (unsigned BroadcastSize : {16, 32, 64}) { |
12399 | if (BroadcastSize <= VT.getScalarSizeInBits()) |
12400 | continue; |
12401 | int NumBroadcastElts = BroadcastSize / VT.getScalarSizeInBits(); |
12402 | |
12403 | // Attempt to match a repeating pattern every NumBroadcastElts, |
12404 | // accounting for UNDEFs but only references the lowest 128-bit |
12405 | // lane of the inputs. |
12406 | auto FindRepeatingBroadcastMask = [&](SmallVectorImpl<int> &RepeatMask) { |
12407 | for (int i = 0; i != NumElts; i += NumBroadcastElts) |
12408 | for (int j = 0; j != NumBroadcastElts; ++j) { |
12409 | int M = Mask[i + j]; |
12410 | if (M < 0) |
12411 | continue; |
12412 | int &R = RepeatMask[j]; |
12413 | if (0 != ((M % NumElts) / NumLaneElts)) |
12414 | return false; |
12415 | if (0 <= R && R != M) |
12416 | return false; |
12417 | R = M; |
12418 | } |
12419 | return true; |
12420 | }; |
12421 | |
12422 | SmallVector<int, 8> RepeatMask((unsigned)NumElts, -1); |
12423 | if (!FindRepeatingBroadcastMask(RepeatMask)) |
12424 | continue; |
12425 | |
12426 | // Shuffle the (lowest) repeated elements in place for broadcast. |
12427 | SDValue RepeatShuf = DAG.getVectorShuffle(VT, DL, V1, V2, RepeatMask); |
12428 | |
12429 | // Shuffle the actual broadcast. |
12430 | SmallVector<int, 8> BroadcastMask((unsigned)NumElts, -1); |
12431 | for (int i = 0; i != NumElts; i += NumBroadcastElts) |
12432 | for (int j = 0; j != NumBroadcastElts; ++j) |
12433 | BroadcastMask[i + j] = j; |
12434 | return DAG.getVectorShuffle(VT, DL, RepeatShuf, DAG.getUNDEF(VT), |
12435 | BroadcastMask); |
12436 | } |
12437 | } |
12438 | |
12439 | // Bail if the shuffle mask doesn't cross 128-bit lanes. |
12440 | if (!is128BitLaneCrossingShuffleMask(VT, Mask)) |
12441 | return SDValue(); |
12442 | |
12443 | // Bail if we already have a repeated lane shuffle mask. |
12444 | SmallVector<int, 8> RepeatedShuffleMask; |
12445 | if (is128BitLaneRepeatedShuffleMask(VT, Mask, RepeatedShuffleMask)) |
12446 | return SDValue(); |
12447 | |
12448 | // On AVX2 targets we can permute 256-bit vectors as 64-bit sub-lanes |
12449 | // (with PERMQ/PERMPD), otherwise we can only permute whole 128-bit lanes. |
12450 | int SubLaneScale = Subtarget.hasAVX2() && VT.is256BitVector() ? 2 : 1; |
12451 | int NumSubLanes = NumLanes * SubLaneScale; |
12452 | int NumSubLaneElts = NumLaneElts / SubLaneScale; |
12453 | |
12454 | // Check that all the sources are coming from the same lane and see if we can |
12455 | // form a repeating shuffle mask (local to each sub-lane). At the same time, |
12456 | // determine the source sub-lane for each destination sub-lane. |
12457 | int TopSrcSubLane = -1; |
12458 | SmallVector<int, 8> Dst2SrcSubLanes((unsigned)NumSubLanes, -1); |
12459 | SmallVector<int, 8> RepeatedSubLaneMasks[2] = { |
12460 | SmallVector<int, 8>((unsigned)NumSubLaneElts, SM_SentinelUndef), |
12461 | SmallVector<int, 8>((unsigned)NumSubLaneElts, SM_SentinelUndef)}; |
12462 | |
12463 | for (int DstSubLane = 0; DstSubLane != NumSubLanes; ++DstSubLane) { |
12464 | // Extract the sub-lane mask, check that it all comes from the same lane |
12465 | // and normalize the mask entries to come from the first lane. |
12466 | int SrcLane = -1; |
12467 | SmallVector<int, 8> SubLaneMask((unsigned)NumSubLaneElts, -1); |
12468 | for (int Elt = 0; Elt != NumSubLaneElts; ++Elt) { |
12469 | int M = Mask[(DstSubLane * NumSubLaneElts) + Elt]; |
12470 | if (M < 0) |
12471 | continue; |
12472 | int Lane = (M % NumElts) / NumLaneElts; |
12473 | if ((0 <= SrcLane) && (SrcLane != Lane)) |
12474 | return SDValue(); |
12475 | SrcLane = Lane; |
12476 | int LocalM = (M % NumLaneElts) + (M < NumElts ? 0 : NumElts); |
12477 | SubLaneMask[Elt] = LocalM; |
12478 | } |
12479 | |
12480 | // Whole sub-lane is UNDEF. |
12481 | if (SrcLane < 0) |
12482 | continue; |
12483 | |
12484 | // Attempt to match against the candidate repeated sub-lane masks. |
12485 | for (int SubLane = 0; SubLane != SubLaneScale; ++SubLane) { |
12486 | auto MatchMasks = [NumSubLaneElts](ArrayRef<int> M1, ArrayRef<int> M2) { |
12487 | for (int i = 0; i != NumSubLaneElts; ++i) { |
12488 | if (M1[i] < 0 || M2[i] < 0) |
12489 | continue; |
12490 | if (M1[i] != M2[i]) |
12491 | return false; |
12492 | } |
12493 | return true; |
12494 | }; |
12495 | |
12496 | auto &RepeatedSubLaneMask = RepeatedSubLaneMasks[SubLane]; |
12497 | if (!MatchMasks(SubLaneMask, RepeatedSubLaneMask)) |
12498 | continue; |
12499 | |
12500 | // Merge the sub-lane mask into the matching repeated sub-lane mask. |
12501 | for (int i = 0; i != NumSubLaneElts; ++i) { |
12502 | int M = SubLaneMask[i]; |
12503 | if (M < 0) |
12504 | continue; |
12505 | assert((RepeatedSubLaneMask[i] < 0 || RepeatedSubLaneMask[i] == M) &&(((RepeatedSubLaneMask[i] < 0 || RepeatedSubLaneMask[i] == M) && "Unexpected mask element") ? static_cast<void > (0) : __assert_fail ("(RepeatedSubLaneMask[i] < 0 || RepeatedSubLaneMask[i] == M) && \"Unexpected mask element\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12506, __PRETTY_FUNCTION__)) |
12506 | "Unexpected mask element")(((RepeatedSubLaneMask[i] < 0 || RepeatedSubLaneMask[i] == M) && "Unexpected mask element") ? static_cast<void > (0) : __assert_fail ("(RepeatedSubLaneMask[i] < 0 || RepeatedSubLaneMask[i] == M) && \"Unexpected mask element\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12506, __PRETTY_FUNCTION__)); |
12507 | RepeatedSubLaneMask[i] = M; |
12508 | } |
12509 | |
12510 | // Track the top most source sub-lane - by setting the remaining to UNDEF |
12511 | // we can greatly simplify shuffle matching. |
12512 | int SrcSubLane = (SrcLane * SubLaneScale) + SubLane; |
12513 | TopSrcSubLane = std::max(TopSrcSubLane, SrcSubLane); |
12514 | Dst2SrcSubLanes[DstSubLane] = SrcSubLane; |
12515 | break; |
12516 | } |
12517 | |
12518 | // Bail if we failed to find a matching repeated sub-lane mask. |
12519 | if (Dst2SrcSubLanes[DstSubLane] < 0) |
12520 | return SDValue(); |
12521 | } |
12522 | assert(0 <= TopSrcSubLane && TopSrcSubLane < NumSubLanes &&((0 <= TopSrcSubLane && TopSrcSubLane < NumSubLanes && "Unexpected source lane") ? static_cast<void> (0) : __assert_fail ("0 <= TopSrcSubLane && TopSrcSubLane < NumSubLanes && \"Unexpected source lane\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12523, __PRETTY_FUNCTION__)) |
12523 | "Unexpected source lane")((0 <= TopSrcSubLane && TopSrcSubLane < NumSubLanes && "Unexpected source lane") ? static_cast<void> (0) : __assert_fail ("0 <= TopSrcSubLane && TopSrcSubLane < NumSubLanes && \"Unexpected source lane\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12523, __PRETTY_FUNCTION__)); |
12524 | |
12525 | // Create a repeating shuffle mask for the entire vector. |
12526 | SmallVector<int, 8> RepeatedMask((unsigned)NumElts, -1); |
12527 | for (int SubLane = 0; SubLane <= TopSrcSubLane; ++SubLane) { |
12528 | int Lane = SubLane / SubLaneScale; |
12529 | auto &RepeatedSubLaneMask = RepeatedSubLaneMasks[SubLane % SubLaneScale]; |
12530 | for (int Elt = 0; Elt != NumSubLaneElts; ++Elt) { |
12531 | int M = RepeatedSubLaneMask[Elt]; |
12532 | if (M < 0) |
12533 | continue; |
12534 | int Idx = (SubLane * NumSubLaneElts) + Elt; |
12535 | RepeatedMask[Idx] = M + (Lane * NumLaneElts); |
12536 | } |
12537 | } |
12538 | SDValue RepeatedShuffle = DAG.getVectorShuffle(VT, DL, V1, V2, RepeatedMask); |
12539 | |
12540 | // Shuffle each source sub-lane to its destination. |
12541 | SmallVector<int, 8> SubLaneMask((unsigned)NumElts, -1); |
12542 | for (int i = 0; i != NumElts; i += NumSubLaneElts) { |
12543 | int SrcSubLane = Dst2SrcSubLanes[i / NumSubLaneElts]; |
12544 | if (SrcSubLane < 0) |
12545 | continue; |
12546 | for (int j = 0; j != NumSubLaneElts; ++j) |
12547 | SubLaneMask[i + j] = j + (SrcSubLane * NumSubLaneElts); |
12548 | } |
12549 | |
12550 | return DAG.getVectorShuffle(VT, DL, RepeatedShuffle, DAG.getUNDEF(VT), |
12551 | SubLaneMask); |
12552 | } |
12553 | |
12554 | static bool matchVectorShuffleWithSHUFPD(MVT VT, SDValue &V1, SDValue &V2, |
12555 | unsigned &ShuffleImm, |
12556 | ArrayRef<int> Mask) { |
12557 | int NumElts = VT.getVectorNumElements(); |
12558 | assert(VT.getScalarSizeInBits() == 64 &&((VT.getScalarSizeInBits() == 64 && (NumElts == 2 || NumElts == 4 || NumElts == 8) && "Unexpected data type for VSHUFPD" ) ? static_cast<void> (0) : __assert_fail ("VT.getScalarSizeInBits() == 64 && (NumElts == 2 || NumElts == 4 || NumElts == 8) && \"Unexpected data type for VSHUFPD\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12560, __PRETTY_FUNCTION__)) |
12559 | (NumElts == 2 || NumElts == 4 || NumElts == 8) &&((VT.getScalarSizeInBits() == 64 && (NumElts == 2 || NumElts == 4 || NumElts == 8) && "Unexpected data type for VSHUFPD" ) ? static_cast<void> (0) : __assert_fail ("VT.getScalarSizeInBits() == 64 && (NumElts == 2 || NumElts == 4 || NumElts == 8) && \"Unexpected data type for VSHUFPD\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12560, __PRETTY_FUNCTION__)) |
12560 | "Unexpected data type for VSHUFPD")((VT.getScalarSizeInBits() == 64 && (NumElts == 2 || NumElts == 4 || NumElts == 8) && "Unexpected data type for VSHUFPD" ) ? static_cast<void> (0) : __assert_fail ("VT.getScalarSizeInBits() == 64 && (NumElts == 2 || NumElts == 4 || NumElts == 8) && \"Unexpected data type for VSHUFPD\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12560, __PRETTY_FUNCTION__)); |
12561 | |
12562 | // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, .. |
12563 | // Mask for V4F64; 0/1, 4/5, 2/3, 6/7.. |
12564 | ShuffleImm = 0; |
12565 | bool ShufpdMask = true; |
12566 | bool CommutableMask = true; |
12567 | for (int i = 0; i < NumElts; ++i) { |
12568 | if (Mask[i] == SM_SentinelUndef) |
12569 | continue; |
12570 | if (Mask[i] < 0) |
12571 | return false; |
12572 | int Val = (i & 6) + NumElts * (i & 1); |
12573 | int CommutVal = (i & 0xe) + NumElts * ((i & 1) ^ 1); |
12574 | if (Mask[i] < Val || Mask[i] > Val + 1) |
12575 | ShufpdMask = false; |
12576 | if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1) |
12577 | CommutableMask = false; |
12578 | ShuffleImm |= (Mask[i] % 2) << i; |
12579 | } |
12580 | |
12581 | if (ShufpdMask) |
12582 | return true; |
12583 | if (CommutableMask) { |
12584 | std::swap(V1, V2); |
12585 | return true; |
12586 | } |
12587 | |
12588 | return false; |
12589 | } |
12590 | |
12591 | static SDValue lowerVectorShuffleWithSHUFPD(const SDLoc &DL, MVT VT, |
12592 | ArrayRef<int> Mask, SDValue V1, |
12593 | SDValue V2, SelectionDAG &DAG) { |
12594 | assert((VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v8f64)&&(((VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v8f64)&& "Unexpected data type for VSHUFPD") ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v8f64)&& \"Unexpected data type for VSHUFPD\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12595, __PRETTY_FUNCTION__)) |
12595 | "Unexpected data type for VSHUFPD")(((VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v8f64)&& "Unexpected data type for VSHUFPD") ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v8f64)&& \"Unexpected data type for VSHUFPD\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12595, __PRETTY_FUNCTION__)); |
12596 | |
12597 | unsigned Immediate = 0; |
12598 | if (!matchVectorShuffleWithSHUFPD(VT, V1, V2, Immediate, Mask)) |
12599 | return SDValue(); |
12600 | |
12601 | return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2, |
12602 | DAG.getConstant(Immediate, DL, MVT::i8)); |
12603 | } |
12604 | |
12605 | static SDValue lowerVectorShuffleWithPERMV(const SDLoc &DL, MVT VT, |
12606 | ArrayRef<int> Mask, SDValue V1, |
12607 | SDValue V2, SelectionDAG &DAG) { |
12608 | MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits()); |
12609 | MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements()); |
12610 | |
12611 | SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true); |
12612 | if (V2.isUndef()) |
12613 | return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1); |
12614 | |
12615 | return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2); |
12616 | } |
12617 | |
12618 | /// \brief Handle lowering of 4-lane 64-bit floating point shuffles. |
12619 | /// |
12620 | /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2 |
12621 | /// isn't available. |
12622 | static SDValue lowerV4F64VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
12623 | const APInt &Zeroable, |
12624 | SDValue V1, SDValue V2, |
12625 | const X86Subtarget &Subtarget, |
12626 | SelectionDAG &DAG) { |
12627 | assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v4f64 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12627, __PRETTY_FUNCTION__)); |
12628 | assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v4f64 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12628, __PRETTY_FUNCTION__)); |
12629 | assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!")((Mask.size() == 4 && "Unexpected mask size for v4 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 4 && \"Unexpected mask size for v4 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12629, __PRETTY_FUNCTION__)); |
12630 | |
12631 | if (SDValue V = lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, |
12632 | Zeroable, Subtarget, DAG)) |
12633 | return V; |
12634 | |
12635 | if (V2.isUndef()) { |
12636 | // Check for being able to broadcast a single element. |
12637 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast( |
12638 | DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG)) |
12639 | return Broadcast; |
12640 | |
12641 | // Use low duplicate instructions for masks that match their pattern. |
12642 | if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2})) |
12643 | return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1); |
12644 | |
12645 | if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) { |
12646 | // Non-half-crossing single input shuffles can be lowered with an |
12647 | // interleaved permutation. |
12648 | unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) | |
12649 | ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3); |
12650 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1, |
12651 | DAG.getConstant(VPERMILPMask, DL, MVT::i8)); |
12652 | } |
12653 | |
12654 | // With AVX2 we have direct support for this permutation. |
12655 | if (Subtarget.hasAVX2()) |
12656 | return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1, |
12657 | getV4X86ShuffleImm8ForMask(Mask, DL, DAG)); |
12658 | |
12659 | // Try to create an in-lane repeating shuffle mask and then shuffle the |
12660 | // the results into the target lanes. |
12661 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( |
12662 | DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG)) |
12663 | return V; |
12664 | |
12665 | // Otherwise, fall back. |
12666 | return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask, |
12667 | DAG); |
12668 | } |
12669 | |
12670 | // Use dedicated unpack instructions for masks that match their pattern. |
12671 | if (SDValue V = |
12672 | lowerVectorShuffleWithUNPCK(DL, MVT::v4f64, Mask, V1, V2, DAG)) |
12673 | return V; |
12674 | |
12675 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask, |
12676 | Zeroable, Subtarget, DAG)) |
12677 | return Blend; |
12678 | |
12679 | // Check if the blend happens to exactly fit that of SHUFPD. |
12680 | if (SDValue Op = |
12681 | lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG)) |
12682 | return Op; |
12683 | |
12684 | // Try to create an in-lane repeating shuffle mask and then shuffle the |
12685 | // the results into the target lanes. |
12686 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( |
12687 | DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG)) |
12688 | return V; |
12689 | |
12690 | // Try to simplify this by merging 128-bit lanes to enable a lane-based |
12691 | // shuffle. However, if we have AVX2 and either inputs are already in place, |
12692 | // we will be able to shuffle even across lanes the other input in a single |
12693 | // instruction so skip this pattern. |
12694 | if (!(Subtarget.hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) || |
12695 | isShuffleMaskInputInPlace(1, Mask)))) |
12696 | if (SDValue Result = lowerVectorShuffleByMerging128BitLanes( |
12697 | DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG)) |
12698 | return Result; |
12699 | // If we have VLX support, we can use VEXPAND. |
12700 | if (Subtarget.hasVLX()) |
12701 | if (SDValue V = lowerVectorShuffleToEXPAND(DL, MVT::v4f64, Zeroable, Mask, |
12702 | V1, V2, DAG, Subtarget)) |
12703 | return V; |
12704 | |
12705 | // If we have AVX2 then we always want to lower with a blend because an v4 we |
12706 | // can fully permute the elements. |
12707 | if (Subtarget.hasAVX2()) |
12708 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2, |
12709 | Mask, DAG); |
12710 | |
12711 | // Otherwise fall back on generic lowering. |
12712 | return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG); |
12713 | } |
12714 | |
12715 | /// \brief Handle lowering of 4-lane 64-bit integer shuffles. |
12716 | /// |
12717 | /// This routine is only called when we have AVX2 and thus a reasonable |
12718 | /// instruction set for v4i64 shuffling.. |
12719 | static SDValue lowerV4I64VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
12720 | const APInt &Zeroable, |
12721 | SDValue V1, SDValue V2, |
12722 | const X86Subtarget &Subtarget, |
12723 | SelectionDAG &DAG) { |
12724 | assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v4i64 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12724, __PRETTY_FUNCTION__)); |
12725 | assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v4i64 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12725, __PRETTY_FUNCTION__)); |
12726 | assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!")((Mask.size() == 4 && "Unexpected mask size for v4 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 4 && \"Unexpected mask size for v4 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12726, __PRETTY_FUNCTION__)); |
12727 | assert(Subtarget.hasAVX2() && "We can only lower v4i64 with AVX2!")((Subtarget.hasAVX2() && "We can only lower v4i64 with AVX2!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX2() && \"We can only lower v4i64 with AVX2!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12727, __PRETTY_FUNCTION__)); |
12728 | |
12729 | if (SDValue V = lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, |
12730 | Zeroable, Subtarget, DAG)) |
12731 | return V; |
12732 | |
12733 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask, |
12734 | Zeroable, Subtarget, DAG)) |
12735 | return Blend; |
12736 | |
12737 | // Check for being able to broadcast a single element. |
12738 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1, V2, |
12739 | Mask, Subtarget, DAG)) |
12740 | return Broadcast; |
12741 | |
12742 | if (V2.isUndef()) { |
12743 | // When the shuffle is mirrored between the 128-bit lanes of the unit, we |
12744 | // can use lower latency instructions that will operate on both lanes. |
12745 | SmallVector<int, 2> RepeatedMask; |
12746 | if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) { |
12747 | SmallVector<int, 4> PSHUFDMask; |
12748 | scaleShuffleMask(2, RepeatedMask, PSHUFDMask); |
12749 | return DAG.getBitcast( |
12750 | MVT::v4i64, |
12751 | DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, |
12752 | DAG.getBitcast(MVT::v8i32, V1), |
12753 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG))); |
12754 | } |
12755 | |
12756 | // AVX2 provides a direct instruction for permuting a single input across |
12757 | // lanes. |
12758 | return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1, |
12759 | getV4X86ShuffleImm8ForMask(Mask, DL, DAG)); |
12760 | } |
12761 | |
12762 | // Try to use shift instructions. |
12763 | if (SDValue Shift = lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, |
12764 | Zeroable, Subtarget, DAG)) |
12765 | return Shift; |
12766 | |
12767 | // If we have VLX support, we can use VALIGN or VEXPAND. |
12768 | if (Subtarget.hasVLX()) { |
12769 | if (SDValue Rotate = lowerVectorShuffleAsRotate(DL, MVT::v4i64, V1, V2, |
12770 | Mask, Subtarget, DAG)) |
12771 | return Rotate; |
12772 | |
12773 | if (SDValue V = lowerVectorShuffleToEXPAND(DL, MVT::v4i64, Zeroable, Mask, |
12774 | V1, V2, DAG, Subtarget)) |
12775 | return V; |
12776 | } |
12777 | |
12778 | // Try to use PALIGNR. |
12779 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v4i64, V1, V2, |
12780 | Mask, Subtarget, DAG)) |
12781 | return Rotate; |
12782 | |
12783 | // Use dedicated unpack instructions for masks that match their pattern. |
12784 | if (SDValue V = |
12785 | lowerVectorShuffleWithUNPCK(DL, MVT::v4i64, Mask, V1, V2, DAG)) |
12786 | return V; |
12787 | |
12788 | // Try to simplify this by merging 128-bit lanes to enable a lane-based |
12789 | // shuffle. However, if we have AVX2 and either inputs are already in place, |
12790 | // we will be able to shuffle even across lanes the other input in a single |
12791 | // instruction so skip this pattern. |
12792 | if (!isShuffleMaskInputInPlace(0, Mask) && |
12793 | !isShuffleMaskInputInPlace(1, Mask)) |
12794 | if (SDValue Result = lowerVectorShuffleByMerging128BitLanes( |
12795 | DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG)) |
12796 | return Result; |
12797 | |
12798 | // Otherwise fall back on generic blend lowering. |
12799 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2, |
12800 | Mask, DAG); |
12801 | } |
12802 | |
12803 | /// \brief Handle lowering of 8-lane 32-bit floating point shuffles. |
12804 | /// |
12805 | /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2 |
12806 | /// isn't available. |
12807 | static SDValue lowerV8F32VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
12808 | const APInt &Zeroable, |
12809 | SDValue V1, SDValue V2, |
12810 | const X86Subtarget &Subtarget, |
12811 | SelectionDAG &DAG) { |
12812 | assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v8f32 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12812, __PRETTY_FUNCTION__)); |
12813 | assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v8f32 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12813, __PRETTY_FUNCTION__)); |
12814 | assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!")((Mask.size() == 8 && "Unexpected mask size for v8 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 8 && \"Unexpected mask size for v8 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12814, __PRETTY_FUNCTION__)); |
12815 | |
12816 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask, |
12817 | Zeroable, Subtarget, DAG)) |
12818 | return Blend; |
12819 | |
12820 | // Check for being able to broadcast a single element. |
12821 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1, V2, |
12822 | Mask, Subtarget, DAG)) |
12823 | return Broadcast; |
12824 | |
12825 | // If the shuffle mask is repeated in each 128-bit lane, we have many more |
12826 | // options to efficiently lower the shuffle. |
12827 | SmallVector<int, 4> RepeatedMask; |
12828 | if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) { |
12829 | assert(RepeatedMask.size() == 4 &&((RepeatedMask.size() == 4 && "Repeated masks must be half the mask width!" ) ? static_cast<void> (0) : __assert_fail ("RepeatedMask.size() == 4 && \"Repeated masks must be half the mask width!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12830, __PRETTY_FUNCTION__)) |
12830 | "Repeated masks must be half the mask width!")((RepeatedMask.size() == 4 && "Repeated masks must be half the mask width!" ) ? static_cast<void> (0) : __assert_fail ("RepeatedMask.size() == 4 && \"Repeated masks must be half the mask width!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12830, __PRETTY_FUNCTION__)); |
12831 | |
12832 | // Use even/odd duplicate instructions for masks that match their pattern. |
12833 | if (isShuffleEquivalent(V1, V2, RepeatedMask, {0, 0, 2, 2})) |
12834 | return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1); |
12835 | if (isShuffleEquivalent(V1, V2, RepeatedMask, {1, 1, 3, 3})) |
12836 | return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1); |
12837 | |
12838 | if (V2.isUndef()) |
12839 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1, |
12840 | getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG)); |
12841 | |
12842 | // Use dedicated unpack instructions for masks that match their pattern. |
12843 | if (SDValue V = |
12844 | lowerVectorShuffleWithUNPCK(DL, MVT::v8f32, Mask, V1, V2, DAG)) |
12845 | return V; |
12846 | |
12847 | // Otherwise, fall back to a SHUFPS sequence. Here it is important that we |
12848 | // have already handled any direct blends. |
12849 | return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG); |
12850 | } |
12851 | |
12852 | // Try to create an in-lane repeating shuffle mask and then shuffle the |
12853 | // the results into the target lanes. |
12854 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( |
12855 | DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG)) |
12856 | return V; |
12857 | |
12858 | // If we have a single input shuffle with different shuffle patterns in the |
12859 | // two 128-bit lanes use the variable mask to VPERMILPS. |
12860 | if (V2.isUndef()) { |
12861 | SDValue VPermMask = getConstVector(Mask, MVT::v8i32, DAG, DL, true); |
12862 | if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask)) |
12863 | return DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v8f32, V1, VPermMask); |
12864 | |
12865 | if (Subtarget.hasAVX2()) |
12866 | return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32, VPermMask, V1); |
12867 | |
12868 | // Otherwise, fall back. |
12869 | return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask, |
12870 | DAG); |
12871 | } |
12872 | |
12873 | // Try to simplify this by merging 128-bit lanes to enable a lane-based |
12874 | // shuffle. |
12875 | if (SDValue Result = lowerVectorShuffleByMerging128BitLanes( |
12876 | DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG)) |
12877 | return Result; |
12878 | // If we have VLX support, we can use VEXPAND. |
12879 | if (Subtarget.hasVLX()) |
12880 | if (SDValue V = lowerVectorShuffleToEXPAND(DL, MVT::v8f32, Zeroable, Mask, |
12881 | V1, V2, DAG, Subtarget)) |
12882 | return V; |
12883 | |
12884 | // For non-AVX512 if the Mask is of 16bit elements in lane then try to split |
12885 | // since after split we get a more efficient code using vpunpcklwd and |
12886 | // vpunpckhwd instrs than vblend. |
12887 | if (!Subtarget.hasAVX512() && isUnpackWdShuffleMask(Mask, MVT::v8f32)) |
12888 | if (SDValue V = lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, |
12889 | Mask, DAG)) |
12890 | return V; |
12891 | |
12892 | // If we have AVX2 then we always want to lower with a blend because at v8 we |
12893 | // can fully permute the elements. |
12894 | if (Subtarget.hasAVX2()) |
12895 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2, |
12896 | Mask, DAG); |
12897 | |
12898 | // Otherwise fall back on generic lowering. |
12899 | return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG); |
12900 | } |
12901 | |
12902 | /// \brief Handle lowering of 8-lane 32-bit integer shuffles. |
12903 | /// |
12904 | /// This routine is only called when we have AVX2 and thus a reasonable |
12905 | /// instruction set for v8i32 shuffling.. |
12906 | static SDValue lowerV8I32VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
12907 | const APInt &Zeroable, |
12908 | SDValue V1, SDValue V2, |
12909 | const X86Subtarget &Subtarget, |
12910 | SelectionDAG &DAG) { |
12911 | assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v8i32 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12911, __PRETTY_FUNCTION__)); |
12912 | assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v8i32 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12912, __PRETTY_FUNCTION__)); |
12913 | assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!")((Mask.size() == 8 && "Unexpected mask size for v8 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 8 && \"Unexpected mask size for v8 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12913, __PRETTY_FUNCTION__)); |
12914 | assert(Subtarget.hasAVX2() && "We can only lower v8i32 with AVX2!")((Subtarget.hasAVX2() && "We can only lower v8i32 with AVX2!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX2() && \"We can only lower v8i32 with AVX2!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12914, __PRETTY_FUNCTION__)); |
12915 | |
12916 | // Whenever we can lower this as a zext, that instruction is strictly faster |
12917 | // than any alternative. It also allows us to fold memory operands into the |
12918 | // shuffle in many cases. |
12919 | if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend( |
12920 | DL, MVT::v8i32, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
12921 | return ZExt; |
12922 | |
12923 | // For non-AVX512 if the Mask is of 16bit elements in lane then try to split |
12924 | // since after split we get a more efficient code than vblend by using |
12925 | // vpunpcklwd and vpunpckhwd instrs. |
12926 | if (isUnpackWdShuffleMask(Mask, MVT::v8i32) && !V2.isUndef() && |
12927 | !Subtarget.hasAVX512()) |
12928 | if (SDValue V = |
12929 | lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8i32, V1, V2, Mask, DAG)) |
12930 | return V; |
12931 | |
12932 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask, |
12933 | Zeroable, Subtarget, DAG)) |
12934 | return Blend; |
12935 | |
12936 | // Check for being able to broadcast a single element. |
12937 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1, V2, |
12938 | Mask, Subtarget, DAG)) |
12939 | return Broadcast; |
12940 | |
12941 | // If the shuffle mask is repeated in each 128-bit lane we can use more |
12942 | // efficient instructions that mirror the shuffles across the two 128-bit |
12943 | // lanes. |
12944 | SmallVector<int, 4> RepeatedMask; |
12945 | bool Is128BitLaneRepeatedShuffle = |
12946 | is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask); |
12947 | if (Is128BitLaneRepeatedShuffle) { |
12948 | assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!")((RepeatedMask.size() == 4 && "Unexpected repeated mask size!" ) ? static_cast<void> (0) : __assert_fail ("RepeatedMask.size() == 4 && \"Unexpected repeated mask size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 12948, __PRETTY_FUNCTION__)); |
12949 | if (V2.isUndef()) |
12950 | return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1, |
12951 | getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG)); |
12952 | |
12953 | // Use dedicated unpack instructions for masks that match their pattern. |
12954 | if (SDValue V = |
12955 | lowerVectorShuffleWithUNPCK(DL, MVT::v8i32, Mask, V1, V2, DAG)) |
12956 | return V; |
12957 | } |
12958 | |
12959 | // Try to use shift instructions. |
12960 | if (SDValue Shift = lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, |
12961 | Zeroable, Subtarget, DAG)) |
12962 | return Shift; |
12963 | |
12964 | // If we have VLX support, we can use VALIGN or EXPAND. |
12965 | if (Subtarget.hasVLX()) { |
12966 | if (SDValue Rotate = lowerVectorShuffleAsRotate(DL, MVT::v8i32, V1, V2, |
12967 | Mask, Subtarget, DAG)) |
12968 | return Rotate; |
12969 | |
12970 | if (SDValue V = lowerVectorShuffleToEXPAND(DL, MVT::v8i32, Zeroable, Mask, |
12971 | V1, V2, DAG, Subtarget)) |
12972 | return V; |
12973 | } |
12974 | |
12975 | // Try to use byte rotation instructions. |
12976 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate( |
12977 | DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG)) |
12978 | return Rotate; |
12979 | |
12980 | // Try to create an in-lane repeating shuffle mask and then shuffle the |
12981 | // results into the target lanes. |
12982 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( |
12983 | DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG)) |
12984 | return V; |
12985 | |
12986 | // If the shuffle patterns aren't repeated but it is a single input, directly |
12987 | // generate a cross-lane VPERMD instruction. |
12988 | if (V2.isUndef()) { |
12989 | SDValue VPermMask = getConstVector(Mask, MVT::v8i32, DAG, DL, true); |
12990 | return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8i32, VPermMask, V1); |
12991 | } |
12992 | |
12993 | // Assume that a single SHUFPS is faster than an alternative sequence of |
12994 | // multiple instructions (even if the CPU has a domain penalty). |
12995 | // If some CPU is harmed by the domain switch, we can fix it in a later pass. |
12996 | if (Is128BitLaneRepeatedShuffle && isSingleSHUFPSMask(RepeatedMask)) { |
12997 | SDValue CastV1 = DAG.getBitcast(MVT::v8f32, V1); |
12998 | SDValue CastV2 = DAG.getBitcast(MVT::v8f32, V2); |
12999 | SDValue ShufPS = lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, |
13000 | CastV1, CastV2, DAG); |
13001 | return DAG.getBitcast(MVT::v8i32, ShufPS); |
13002 | } |
13003 | |
13004 | // Try to simplify this by merging 128-bit lanes to enable a lane-based |
13005 | // shuffle. |
13006 | if (SDValue Result = lowerVectorShuffleByMerging128BitLanes( |
13007 | DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG)) |
13008 | return Result; |
13009 | |
13010 | // Otherwise fall back on generic blend lowering. |
13011 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2, |
13012 | Mask, DAG); |
13013 | } |
13014 | |
13015 | /// \brief Handle lowering of 16-lane 16-bit integer shuffles. |
13016 | /// |
13017 | /// This routine is only called when we have AVX2 and thus a reasonable |
13018 | /// instruction set for v16i16 shuffling.. |
13019 | static SDValue lowerV16I16VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
13020 | const APInt &Zeroable, |
13021 | SDValue V1, SDValue V2, |
13022 | const X86Subtarget &Subtarget, |
13023 | SelectionDAG &DAG) { |
13024 | assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v16i16 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13024, __PRETTY_FUNCTION__)); |
13025 | assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v16i16 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13025, __PRETTY_FUNCTION__)); |
13026 | assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!")((Mask.size() == 16 && "Unexpected mask size for v16 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 16 && \"Unexpected mask size for v16 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13026, __PRETTY_FUNCTION__)); |
13027 | assert(Subtarget.hasAVX2() && "We can only lower v16i16 with AVX2!")((Subtarget.hasAVX2() && "We can only lower v16i16 with AVX2!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX2() && \"We can only lower v16i16 with AVX2!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13027, __PRETTY_FUNCTION__)); |
13028 | |
13029 | // Whenever we can lower this as a zext, that instruction is strictly faster |
13030 | // than any alternative. It also allows us to fold memory operands into the |
13031 | // shuffle in many cases. |
13032 | if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend( |
13033 | DL, MVT::v16i16, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
13034 | return ZExt; |
13035 | |
13036 | // Check for being able to broadcast a single element. |
13037 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1, V2, |
13038 | Mask, Subtarget, DAG)) |
13039 | return Broadcast; |
13040 | |
13041 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask, |
13042 | Zeroable, Subtarget, DAG)) |
13043 | return Blend; |
13044 | |
13045 | // Use dedicated unpack instructions for masks that match their pattern. |
13046 | if (SDValue V = |
13047 | lowerVectorShuffleWithUNPCK(DL, MVT::v16i16, Mask, V1, V2, DAG)) |
13048 | return V; |
13049 | |
13050 | // Try to use shift instructions. |
13051 | if (SDValue Shift = lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, |
13052 | Zeroable, Subtarget, DAG)) |
13053 | return Shift; |
13054 | |
13055 | // Try to use byte rotation instructions. |
13056 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate( |
13057 | DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG)) |
13058 | return Rotate; |
13059 | |
13060 | // Try to create an in-lane repeating shuffle mask and then shuffle the |
13061 | // the results into the target lanes. |
13062 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( |
13063 | DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG)) |
13064 | return V; |
13065 | |
13066 | if (V2.isUndef()) { |
13067 | // There are no generalized cross-lane shuffle operations available on i16 |
13068 | // element types. |
13069 | if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask)) |
13070 | return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2, |
13071 | Mask, DAG); |
13072 | |
13073 | SmallVector<int, 8> RepeatedMask; |
13074 | if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) { |
13075 | // As this is a single-input shuffle, the repeated mask should be |
13076 | // a strictly valid v8i16 mask that we can pass through to the v8i16 |
13077 | // lowering to handle even the v16 case. |
13078 | return lowerV8I16GeneralSingleInputVectorShuffle( |
13079 | DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG); |
13080 | } |
13081 | } |
13082 | |
13083 | if (SDValue PSHUFB = lowerVectorShuffleWithPSHUFB( |
13084 | DL, MVT::v16i16, Mask, V1, V2, Zeroable, Subtarget, DAG)) |
13085 | return PSHUFB; |
13086 | |
13087 | // AVX512BWVL can lower to VPERMW. |
13088 | if (Subtarget.hasBWI() && Subtarget.hasVLX()) |
13089 | return lowerVectorShuffleWithPERMV(DL, MVT::v16i16, Mask, V1, V2, DAG); |
13090 | |
13091 | // Try to simplify this by merging 128-bit lanes to enable a lane-based |
13092 | // shuffle. |
13093 | if (SDValue Result = lowerVectorShuffleByMerging128BitLanes( |
13094 | DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG)) |
13095 | return Result; |
13096 | |
13097 | // Otherwise fall back on generic lowering. |
13098 | return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG); |
13099 | } |
13100 | |
13101 | /// \brief Handle lowering of 32-lane 8-bit integer shuffles. |
13102 | /// |
13103 | /// This routine is only called when we have AVX2 and thus a reasonable |
13104 | /// instruction set for v32i8 shuffling.. |
13105 | static SDValue lowerV32I8VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
13106 | const APInt &Zeroable, |
13107 | SDValue V1, SDValue V2, |
13108 | const X86Subtarget &Subtarget, |
13109 | SelectionDAG &DAG) { |
13110 | assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v32i8 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13110, __PRETTY_FUNCTION__)); |
13111 | assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v32i8 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13111, __PRETTY_FUNCTION__)); |
13112 | assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!")((Mask.size() == 32 && "Unexpected mask size for v32 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 32 && \"Unexpected mask size for v32 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13112, __PRETTY_FUNCTION__)); |
13113 | assert(Subtarget.hasAVX2() && "We can only lower v32i8 with AVX2!")((Subtarget.hasAVX2() && "We can only lower v32i8 with AVX2!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX2() && \"We can only lower v32i8 with AVX2!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13113, __PRETTY_FUNCTION__)); |
13114 | |
13115 | // Whenever we can lower this as a zext, that instruction is strictly faster |
13116 | // than any alternative. It also allows us to fold memory operands into the |
13117 | // shuffle in many cases. |
13118 | if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend( |
13119 | DL, MVT::v32i8, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
13120 | return ZExt; |
13121 | |
13122 | // Check for being able to broadcast a single element. |
13123 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1, V2, |
13124 | Mask, Subtarget, DAG)) |
13125 | return Broadcast; |
13126 | |
13127 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask, |
13128 | Zeroable, Subtarget, DAG)) |
13129 | return Blend; |
13130 | |
13131 | // Use dedicated unpack instructions for masks that match their pattern. |
13132 | if (SDValue V = |
13133 | lowerVectorShuffleWithUNPCK(DL, MVT::v32i8, Mask, V1, V2, DAG)) |
13134 | return V; |
13135 | |
13136 | // Try to use shift instructions. |
13137 | if (SDValue Shift = lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, |
13138 | Zeroable, Subtarget, DAG)) |
13139 | return Shift; |
13140 | |
13141 | // Try to use byte rotation instructions. |
13142 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate( |
13143 | DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG)) |
13144 | return Rotate; |
13145 | |
13146 | // Try to create an in-lane repeating shuffle mask and then shuffle the |
13147 | // the results into the target lanes. |
13148 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( |
13149 | DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG)) |
13150 | return V; |
13151 | |
13152 | // There are no generalized cross-lane shuffle operations available on i8 |
13153 | // element types. |
13154 | if (V2.isUndef() && is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask)) |
13155 | return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2, Mask, |
13156 | DAG); |
13157 | |
13158 | if (SDValue PSHUFB = lowerVectorShuffleWithPSHUFB( |
13159 | DL, MVT::v32i8, Mask, V1, V2, Zeroable, Subtarget, DAG)) |
13160 | return PSHUFB; |
13161 | |
13162 | // Try to simplify this by merging 128-bit lanes to enable a lane-based |
13163 | // shuffle. |
13164 | if (SDValue Result = lowerVectorShuffleByMerging128BitLanes( |
13165 | DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG)) |
13166 | return Result; |
13167 | |
13168 | // Otherwise fall back on generic lowering. |
13169 | return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG); |
13170 | } |
13171 | |
13172 | /// \brief High-level routine to lower various 256-bit x86 vector shuffles. |
13173 | /// |
13174 | /// This routine either breaks down the specific type of a 256-bit x86 vector |
13175 | /// shuffle or splits it into two 128-bit shuffles and fuses the results back |
13176 | /// together based on the available instructions. |
13177 | static SDValue lower256BitVectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
13178 | MVT VT, SDValue V1, SDValue V2, |
13179 | const APInt &Zeroable, |
13180 | const X86Subtarget &Subtarget, |
13181 | SelectionDAG &DAG) { |
13182 | // If we have a single input to the zero element, insert that into V1 if we |
13183 | // can do so cheaply. |
13184 | int NumElts = VT.getVectorNumElements(); |
13185 | int NumV2Elements = count_if(Mask, [NumElts](int M) { return M >= NumElts; }); |
13186 | |
13187 | if (NumV2Elements == 1 && Mask[0] >= NumElts) |
13188 | if (SDValue Insertion = lowerVectorShuffleAsElementInsertion( |
13189 | DL, VT, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
13190 | return Insertion; |
13191 | |
13192 | // Handle special cases where the lower or upper half is UNDEF. |
13193 | if (SDValue V = |
13194 | lowerVectorShuffleWithUndefHalf(DL, VT, V1, V2, Mask, Subtarget, DAG)) |
13195 | return V; |
13196 | |
13197 | // There is a really nice hard cut-over between AVX1 and AVX2 that means we |
13198 | // can check for those subtargets here and avoid much of the subtarget |
13199 | // querying in the per-vector-type lowering routines. With AVX1 we have |
13200 | // essentially *zero* ability to manipulate a 256-bit vector with integer |
13201 | // types. Since we'll use floating point types there eventually, just |
13202 | // immediately cast everything to a float and operate entirely in that domain. |
13203 | if (VT.isInteger() && !Subtarget.hasAVX2()) { |
13204 | int ElementBits = VT.getScalarSizeInBits(); |
13205 | if (ElementBits < 32) { |
13206 | // No floating point type available, if we can't use the bit operations |
13207 | // for masking/blending then decompose into 128-bit vectors. |
13208 | if (SDValue V = |
13209 | lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, Zeroable, DAG)) |
13210 | return V; |
13211 | if (SDValue V = lowerVectorShuffleAsBitBlend(DL, VT, V1, V2, Mask, DAG)) |
13212 | return V; |
13213 | return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG); |
13214 | } |
13215 | |
13216 | MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits), |
13217 | VT.getVectorNumElements()); |
13218 | V1 = DAG.getBitcast(FpVT, V1); |
13219 | V2 = DAG.getBitcast(FpVT, V2); |
13220 | return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask)); |
13221 | } |
13222 | |
13223 | switch (VT.SimpleTy) { |
13224 | case MVT::v4f64: |
13225 | return lowerV4F64VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
13226 | case MVT::v4i64: |
13227 | return lowerV4I64VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
13228 | case MVT::v8f32: |
13229 | return lowerV8F32VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
13230 | case MVT::v8i32: |
13231 | return lowerV8I32VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
13232 | case MVT::v16i16: |
13233 | return lowerV16I16VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
13234 | case MVT::v32i8: |
13235 | return lowerV32I8VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
13236 | |
13237 | default: |
13238 | llvm_unreachable("Not a valid 256-bit x86 vector type!")::llvm::llvm_unreachable_internal("Not a valid 256-bit x86 vector type!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13238); |
13239 | } |
13240 | } |
13241 | |
13242 | /// \brief Try to lower a vector shuffle as a 128-bit shuffles. |
13243 | static SDValue lowerV4X128VectorShuffle(const SDLoc &DL, MVT VT, |
13244 | ArrayRef<int> Mask, SDValue V1, |
13245 | SDValue V2, SelectionDAG &DAG) { |
13246 | assert(VT.getScalarSizeInBits() == 64 &&((VT.getScalarSizeInBits() == 64 && "Unexpected element type size for 128bit shuffle." ) ? static_cast<void> (0) : __assert_fail ("VT.getScalarSizeInBits() == 64 && \"Unexpected element type size for 128bit shuffle.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13247, __PRETTY_FUNCTION__)) |
13247 | "Unexpected element type size for 128bit shuffle.")((VT.getScalarSizeInBits() == 64 && "Unexpected element type size for 128bit shuffle." ) ? static_cast<void> (0) : __assert_fail ("VT.getScalarSizeInBits() == 64 && \"Unexpected element type size for 128bit shuffle.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13247, __PRETTY_FUNCTION__)); |
13248 | |
13249 | // To handle 256 bit vector requires VLX and most probably |
13250 | // function lowerV2X128VectorShuffle() is better solution. |
13251 | assert(VT.is512BitVector() && "Unexpected vector size for 512bit shuffle.")((VT.is512BitVector() && "Unexpected vector size for 512bit shuffle." ) ? static_cast<void> (0) : __assert_fail ("VT.is512BitVector() && \"Unexpected vector size for 512bit shuffle.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13251, __PRETTY_FUNCTION__)); |
13252 | |
13253 | SmallVector<int, 4> WidenedMask; |
13254 | if (!canWidenShuffleElements(Mask, WidenedMask)) |
13255 | return SDValue(); |
13256 | |
13257 | // Check for patterns which can be matched with a single insert of a 256-bit |
13258 | // subvector. |
13259 | bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, |
13260 | {0, 1, 2, 3, 0, 1, 2, 3}); |
13261 | if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, |
13262 | {0, 1, 2, 3, 8, 9, 10, 11})) { |
13263 | MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(), 4); |
13264 | SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1, |
13265 | DAG.getIntPtrConstant(0, DL)); |
13266 | SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, |
13267 | OnlyUsesV1 ? V1 : V2, |
13268 | DAG.getIntPtrConstant(0, DL)); |
13269 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV); |
13270 | } |
13271 | |
13272 | assert(WidenedMask.size() == 4)((WidenedMask.size() == 4) ? static_cast<void> (0) : __assert_fail ("WidenedMask.size() == 4", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13272, __PRETTY_FUNCTION__)); |
13273 | |
13274 | // See if this is an insertion of the lower 128-bits of V2 into V1. |
13275 | bool IsInsert = true; |
13276 | int V2Index = -1; |
13277 | for (int i = 0; i < 4; ++i) { |
13278 | assert(WidenedMask[i] >= -1)((WidenedMask[i] >= -1) ? static_cast<void> (0) : __assert_fail ("WidenedMask[i] >= -1", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13278, __PRETTY_FUNCTION__)); |
13279 | if (WidenedMask[i] < 0) |
13280 | continue; |
13281 | |
13282 | // Make sure all V1 subvectors are in place. |
13283 | if (WidenedMask[i] < 4) { |
13284 | if (WidenedMask[i] != i) { |
13285 | IsInsert = false; |
13286 | break; |
13287 | } |
13288 | } else { |
13289 | // Make sure we only have a single V2 index and its the lowest 128-bits. |
13290 | if (V2Index >= 0 || WidenedMask[i] != 4) { |
13291 | IsInsert = false; |
13292 | break; |
13293 | } |
13294 | V2Index = i; |
13295 | } |
13296 | } |
13297 | if (IsInsert && V2Index >= 0) { |
13298 | MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(), 2); |
13299 | SDValue Subvec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V2, |
13300 | DAG.getIntPtrConstant(0, DL)); |
13301 | return insert128BitVector(V1, Subvec, V2Index * 2, DAG, DL); |
13302 | } |
13303 | |
13304 | // Try to lower to to vshuf64x2/vshuf32x4. |
13305 | SDValue Ops[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT)}; |
13306 | unsigned PermMask = 0; |
13307 | // Insure elements came from the same Op. |
13308 | for (int i = 0; i < 4; ++i) { |
13309 | assert(WidenedMask[i] >= -1)((WidenedMask[i] >= -1) ? static_cast<void> (0) : __assert_fail ("WidenedMask[i] >= -1", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13309, __PRETTY_FUNCTION__)); |
13310 | if (WidenedMask[i] < 0) |
13311 | continue; |
13312 | |
13313 | SDValue Op = WidenedMask[i] >= 4 ? V2 : V1; |
13314 | unsigned OpIndex = i / 2; |
13315 | if (Ops[OpIndex].isUndef()) |
13316 | Ops[OpIndex] = Op; |
13317 | else if (Ops[OpIndex] != Op) |
13318 | return SDValue(); |
13319 | |
13320 | // Convert the 128-bit shuffle mask selection values into 128-bit selection |
13321 | // bits defined by a vshuf64x2 instruction's immediate control byte. |
13322 | PermMask |= (WidenedMask[i] % 4) << (i * 2); |
13323 | } |
13324 | |
13325 | return DAG.getNode(X86ISD::SHUF128, DL, VT, Ops[0], Ops[1], |
13326 | DAG.getConstant(PermMask, DL, MVT::i8)); |
13327 | } |
13328 | |
13329 | /// \brief Handle lowering of 8-lane 64-bit floating point shuffles. |
13330 | static SDValue lowerV8F64VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
13331 | const APInt &Zeroable, |
13332 | SDValue V1, SDValue V2, |
13333 | const X86Subtarget &Subtarget, |
13334 | SelectionDAG &DAG) { |
13335 | assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v8f64 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13335, __PRETTY_FUNCTION__)); |
13336 | assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v8f64 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13336, __PRETTY_FUNCTION__)); |
13337 | assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!")((Mask.size() == 8 && "Unexpected mask size for v8 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 8 && \"Unexpected mask size for v8 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13337, __PRETTY_FUNCTION__)); |
13338 | |
13339 | if (V2.isUndef()) { |
13340 | // Use low duplicate instructions for masks that match their pattern. |
13341 | if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6})) |
13342 | return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v8f64, V1); |
13343 | |
13344 | if (!is128BitLaneCrossingShuffleMask(MVT::v8f64, Mask)) { |
13345 | // Non-half-crossing single input shuffles can be lowered with an |
13346 | // interleaved permutation. |
13347 | unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) | |
13348 | ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3) | |
13349 | ((Mask[4] == 5) << 4) | ((Mask[5] == 5) << 5) | |
13350 | ((Mask[6] == 7) << 6) | ((Mask[7] == 7) << 7); |
13351 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f64, V1, |
13352 | DAG.getConstant(VPERMILPMask, DL, MVT::i8)); |
13353 | } |
13354 | |
13355 | SmallVector<int, 4> RepeatedMask; |
13356 | if (is256BitLaneRepeatedShuffleMask(MVT::v8f64, Mask, RepeatedMask)) |
13357 | return DAG.getNode(X86ISD::VPERMI, DL, MVT::v8f64, V1, |
13358 | getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG)); |
13359 | } |
13360 | |
13361 | if (SDValue Shuf128 = |
13362 | lowerV4X128VectorShuffle(DL, MVT::v8f64, Mask, V1, V2, DAG)) |
13363 | return Shuf128; |
13364 | |
13365 | if (SDValue Unpck = |
13366 | lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG)) |
13367 | return Unpck; |
13368 | |
13369 | // Check if the blend happens to exactly fit that of SHUFPD. |
13370 | if (SDValue Op = |
13371 | lowerVectorShuffleWithSHUFPD(DL, MVT::v8f64, Mask, V1, V2, DAG)) |
13372 | return Op; |
13373 | |
13374 | if (SDValue V = lowerVectorShuffleToEXPAND(DL, MVT::v8f64, Zeroable, Mask, V1, |
13375 | V2, DAG, Subtarget)) |
13376 | return V; |
13377 | |
13378 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f64, V1, V2, Mask, |
13379 | Zeroable, Subtarget, DAG)) |
13380 | return Blend; |
13381 | |
13382 | return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG); |
13383 | } |
13384 | |
13385 | /// \brief Handle lowering of 16-lane 32-bit floating point shuffles. |
13386 | static SDValue lowerV16F32VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
13387 | const APInt &Zeroable, |
13388 | SDValue V1, SDValue V2, |
13389 | const X86Subtarget &Subtarget, |
13390 | SelectionDAG &DAG) { |
13391 | assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v16f32 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13391, __PRETTY_FUNCTION__)); |
13392 | assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v16f32 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13392, __PRETTY_FUNCTION__)); |
13393 | assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!")((Mask.size() == 16 && "Unexpected mask size for v16 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 16 && \"Unexpected mask size for v16 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13393, __PRETTY_FUNCTION__)); |
13394 | |
13395 | // If the shuffle mask is repeated in each 128-bit lane, we have many more |
13396 | // options to efficiently lower the shuffle. |
13397 | SmallVector<int, 4> RepeatedMask; |
13398 | if (is128BitLaneRepeatedShuffleMask(MVT::v16f32, Mask, RepeatedMask)) { |
13399 | assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!")((RepeatedMask.size() == 4 && "Unexpected repeated mask size!" ) ? static_cast<void> (0) : __assert_fail ("RepeatedMask.size() == 4 && \"Unexpected repeated mask size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13399, __PRETTY_FUNCTION__)); |
13400 | |
13401 | // Use even/odd duplicate instructions for masks that match their pattern. |
13402 | if (isShuffleEquivalent(V1, V2, RepeatedMask, {0, 0, 2, 2})) |
13403 | return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v16f32, V1); |
13404 | if (isShuffleEquivalent(V1, V2, RepeatedMask, {1, 1, 3, 3})) |
13405 | return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v16f32, V1); |
13406 | |
13407 | if (V2.isUndef()) |
13408 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v16f32, V1, |
13409 | getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG)); |
13410 | |
13411 | // Use dedicated unpack instructions for masks that match their pattern. |
13412 | if (SDValue Unpck = |
13413 | lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG)) |
13414 | return Unpck; |
13415 | |
13416 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16f32, V1, V2, Mask, |
13417 | Zeroable, Subtarget, DAG)) |
13418 | return Blend; |
13419 | |
13420 | // Otherwise, fall back to a SHUFPS sequence. |
13421 | return lowerVectorShuffleWithSHUFPS(DL, MVT::v16f32, RepeatedMask, V1, V2, DAG); |
13422 | } |
13423 | // If we have AVX512F support, we can use VEXPAND. |
13424 | if (SDValue V = lowerVectorShuffleToEXPAND(DL, MVT::v16f32, Zeroable, Mask, |
13425 | V1, V2, DAG, Subtarget)) |
13426 | return V; |
13427 | |
13428 | return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG); |
13429 | } |
13430 | |
13431 | /// \brief Handle lowering of 8-lane 64-bit integer shuffles. |
13432 | static SDValue lowerV8I64VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
13433 | const APInt &Zeroable, |
13434 | SDValue V1, SDValue V2, |
13435 | const X86Subtarget &Subtarget, |
13436 | SelectionDAG &DAG) { |
13437 | assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v8i64 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13437, __PRETTY_FUNCTION__)); |
13438 | assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v8i64 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13438, __PRETTY_FUNCTION__)); |
13439 | assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!")((Mask.size() == 8 && "Unexpected mask size for v8 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 8 && \"Unexpected mask size for v8 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13439, __PRETTY_FUNCTION__)); |
13440 | |
13441 | if (SDValue Shuf128 = |
13442 | lowerV4X128VectorShuffle(DL, MVT::v8i64, Mask, V1, V2, DAG)) |
13443 | return Shuf128; |
13444 | |
13445 | if (V2.isUndef()) { |
13446 | // When the shuffle is mirrored between the 128-bit lanes of the unit, we |
13447 | // can use lower latency instructions that will operate on all four |
13448 | // 128-bit lanes. |
13449 | SmallVector<int, 2> Repeated128Mask; |
13450 | if (is128BitLaneRepeatedShuffleMask(MVT::v8i64, Mask, Repeated128Mask)) { |
13451 | SmallVector<int, 4> PSHUFDMask; |
13452 | scaleShuffleMask(2, Repeated128Mask, PSHUFDMask); |
13453 | return DAG.getBitcast( |
13454 | MVT::v8i64, |
13455 | DAG.getNode(X86ISD::PSHUFD, DL, MVT::v16i32, |
13456 | DAG.getBitcast(MVT::v16i32, V1), |
13457 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG))); |
13458 | } |
13459 | |
13460 | SmallVector<int, 4> Repeated256Mask; |
13461 | if (is256BitLaneRepeatedShuffleMask(MVT::v8i64, Mask, Repeated256Mask)) |
13462 | return DAG.getNode(X86ISD::VPERMI, DL, MVT::v8i64, V1, |
13463 | getV4X86ShuffleImm8ForMask(Repeated256Mask, DL, DAG)); |
13464 | } |
13465 | |
13466 | // Try to use shift instructions. |
13467 | if (SDValue Shift = lowerVectorShuffleAsShift(DL, MVT::v8i64, V1, V2, Mask, |
13468 | Zeroable, Subtarget, DAG)) |
13469 | return Shift; |
13470 | |
13471 | // Try to use VALIGN. |
13472 | if (SDValue Rotate = lowerVectorShuffleAsRotate(DL, MVT::v8i64, V1, V2, |
13473 | Mask, Subtarget, DAG)) |
13474 | return Rotate; |
13475 | |
13476 | // Try to use PALIGNR. |
13477 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i64, V1, V2, |
13478 | Mask, Subtarget, DAG)) |
13479 | return Rotate; |
13480 | |
13481 | if (SDValue Unpck = |
13482 | lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG)) |
13483 | return Unpck; |
13484 | // If we have AVX512F support, we can use VEXPAND. |
13485 | if (SDValue V = lowerVectorShuffleToEXPAND(DL, MVT::v8i64, Zeroable, Mask, V1, |
13486 | V2, DAG, Subtarget)) |
13487 | return V; |
13488 | |
13489 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i64, V1, V2, Mask, |
13490 | Zeroable, Subtarget, DAG)) |
13491 | return Blend; |
13492 | |
13493 | return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG); |
13494 | } |
13495 | |
13496 | /// \brief Handle lowering of 16-lane 32-bit integer shuffles. |
13497 | static SDValue lowerV16I32VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
13498 | const APInt &Zeroable, |
13499 | SDValue V1, SDValue V2, |
13500 | const X86Subtarget &Subtarget, |
13501 | SelectionDAG &DAG) { |
13502 | assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v16i32 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13502, __PRETTY_FUNCTION__)); |
13503 | assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v16i32 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13503, __PRETTY_FUNCTION__)); |
13504 | assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!")((Mask.size() == 16 && "Unexpected mask size for v16 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 16 && \"Unexpected mask size for v16 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13504, __PRETTY_FUNCTION__)); |
13505 | |
13506 | // Whenever we can lower this as a zext, that instruction is strictly faster |
13507 | // than any alternative. It also allows us to fold memory operands into the |
13508 | // shuffle in many cases. |
13509 | if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend( |
13510 | DL, MVT::v16i32, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
13511 | return ZExt; |
13512 | |
13513 | // If the shuffle mask is repeated in each 128-bit lane we can use more |
13514 | // efficient instructions that mirror the shuffles across the four 128-bit |
13515 | // lanes. |
13516 | SmallVector<int, 4> RepeatedMask; |
13517 | bool Is128BitLaneRepeatedShuffle = |
13518 | is128BitLaneRepeatedShuffleMask(MVT::v16i32, Mask, RepeatedMask); |
13519 | if (Is128BitLaneRepeatedShuffle) { |
13520 | assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!")((RepeatedMask.size() == 4 && "Unexpected repeated mask size!" ) ? static_cast<void> (0) : __assert_fail ("RepeatedMask.size() == 4 && \"Unexpected repeated mask size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13520, __PRETTY_FUNCTION__)); |
13521 | if (V2.isUndef()) |
13522 | return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v16i32, V1, |
13523 | getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG)); |
13524 | |
13525 | // Use dedicated unpack instructions for masks that match their pattern. |
13526 | if (SDValue V = |
13527 | lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG)) |
13528 | return V; |
13529 | } |
13530 | |
13531 | // Try to use shift instructions. |
13532 | if (SDValue Shift = lowerVectorShuffleAsShift(DL, MVT::v16i32, V1, V2, Mask, |
13533 | Zeroable, Subtarget, DAG)) |
13534 | return Shift; |
13535 | |
13536 | // Try to use VALIGN. |
13537 | if (SDValue Rotate = lowerVectorShuffleAsRotate(DL, MVT::v16i32, V1, V2, |
13538 | Mask, Subtarget, DAG)) |
13539 | return Rotate; |
13540 | |
13541 | // Try to use byte rotation instructions. |
13542 | if (Subtarget.hasBWI()) |
13543 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate( |
13544 | DL, MVT::v16i32, V1, V2, Mask, Subtarget, DAG)) |
13545 | return Rotate; |
13546 | |
13547 | // Assume that a single SHUFPS is faster than using a permv shuffle. |
13548 | // If some CPU is harmed by the domain switch, we can fix it in a later pass. |
13549 | if (Is128BitLaneRepeatedShuffle && isSingleSHUFPSMask(RepeatedMask)) { |
13550 | SDValue CastV1 = DAG.getBitcast(MVT::v16f32, V1); |
13551 | SDValue CastV2 = DAG.getBitcast(MVT::v16f32, V2); |
13552 | SDValue ShufPS = lowerVectorShuffleWithSHUFPS(DL, MVT::v16f32, RepeatedMask, |
13553 | CastV1, CastV2, DAG); |
13554 | return DAG.getBitcast(MVT::v16i32, ShufPS); |
13555 | } |
13556 | // If we have AVX512F support, we can use VEXPAND. |
13557 | if (SDValue V = lowerVectorShuffleToEXPAND(DL, MVT::v16i32, Zeroable, Mask, |
13558 | V1, V2, DAG, Subtarget)) |
13559 | return V; |
13560 | |
13561 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i32, V1, V2, Mask, |
13562 | Zeroable, Subtarget, DAG)) |
13563 | return Blend; |
13564 | return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG); |
13565 | } |
13566 | |
13567 | /// \brief Handle lowering of 32-lane 16-bit integer shuffles. |
13568 | static SDValue lowerV32I16VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
13569 | const APInt &Zeroable, |
13570 | SDValue V1, SDValue V2, |
13571 | const X86Subtarget &Subtarget, |
13572 | SelectionDAG &DAG) { |
13573 | assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v32i16 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13573, __PRETTY_FUNCTION__)); |
13574 | assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v32i16 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13574, __PRETTY_FUNCTION__)); |
13575 | assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!")((Mask.size() == 32 && "Unexpected mask size for v32 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 32 && \"Unexpected mask size for v32 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13575, __PRETTY_FUNCTION__)); |
13576 | assert(Subtarget.hasBWI() && "We can only lower v32i16 with AVX-512-BWI!")((Subtarget.hasBWI() && "We can only lower v32i16 with AVX-512-BWI!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasBWI() && \"We can only lower v32i16 with AVX-512-BWI!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13576, __PRETTY_FUNCTION__)); |
13577 | |
13578 | // Whenever we can lower this as a zext, that instruction is strictly faster |
13579 | // than any alternative. It also allows us to fold memory operands into the |
13580 | // shuffle in many cases. |
13581 | if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend( |
13582 | DL, MVT::v32i16, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
13583 | return ZExt; |
13584 | |
13585 | // Use dedicated unpack instructions for masks that match their pattern. |
13586 | if (SDValue V = |
13587 | lowerVectorShuffleWithUNPCK(DL, MVT::v32i16, Mask, V1, V2, DAG)) |
13588 | return V; |
13589 | |
13590 | // Try to use shift instructions. |
13591 | if (SDValue Shift = lowerVectorShuffleAsShift(DL, MVT::v32i16, V1, V2, Mask, |
13592 | Zeroable, Subtarget, DAG)) |
13593 | return Shift; |
13594 | |
13595 | // Try to use byte rotation instructions. |
13596 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate( |
13597 | DL, MVT::v32i16, V1, V2, Mask, Subtarget, DAG)) |
13598 | return Rotate; |
13599 | |
13600 | if (V2.isUndef()) { |
13601 | SmallVector<int, 8> RepeatedMask; |
13602 | if (is128BitLaneRepeatedShuffleMask(MVT::v32i16, Mask, RepeatedMask)) { |
13603 | // As this is a single-input shuffle, the repeated mask should be |
13604 | // a strictly valid v8i16 mask that we can pass through to the v8i16 |
13605 | // lowering to handle even the v32 case. |
13606 | return lowerV8I16GeneralSingleInputVectorShuffle( |
13607 | DL, MVT::v32i16, V1, RepeatedMask, Subtarget, DAG); |
13608 | } |
13609 | } |
13610 | |
13611 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i16, V1, V2, Mask, |
13612 | Zeroable, Subtarget, DAG)) |
13613 | return Blend; |
13614 | |
13615 | return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG); |
13616 | } |
13617 | |
13618 | /// \brief Handle lowering of 64-lane 8-bit integer shuffles. |
13619 | static SDValue lowerV64I8VectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
13620 | const APInt &Zeroable, |
13621 | SDValue V1, SDValue V2, |
13622 | const X86Subtarget &Subtarget, |
13623 | SelectionDAG &DAG) { |
13624 | assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!")((V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v64i8 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13624, __PRETTY_FUNCTION__)); |
13625 | assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!")((V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v64i8 && \"Bad operand type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13625, __PRETTY_FUNCTION__)); |
13626 | assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!")((Mask.size() == 64 && "Unexpected mask size for v64 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 64 && \"Unexpected mask size for v64 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13626, __PRETTY_FUNCTION__)); |
13627 | assert(Subtarget.hasBWI() && "We can only lower v64i8 with AVX-512-BWI!")((Subtarget.hasBWI() && "We can only lower v64i8 with AVX-512-BWI!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasBWI() && \"We can only lower v64i8 with AVX-512-BWI!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13627, __PRETTY_FUNCTION__)); |
13628 | |
13629 | // Whenever we can lower this as a zext, that instruction is strictly faster |
13630 | // than any alternative. It also allows us to fold memory operands into the |
13631 | // shuffle in many cases. |
13632 | if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend( |
13633 | DL, MVT::v64i8, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
13634 | return ZExt; |
13635 | |
13636 | // Use dedicated unpack instructions for masks that match their pattern. |
13637 | if (SDValue V = |
13638 | lowerVectorShuffleWithUNPCK(DL, MVT::v64i8, Mask, V1, V2, DAG)) |
13639 | return V; |
13640 | |
13641 | // Try to use shift instructions. |
13642 | if (SDValue Shift = lowerVectorShuffleAsShift(DL, MVT::v64i8, V1, V2, Mask, |
13643 | Zeroable, Subtarget, DAG)) |
13644 | return Shift; |
13645 | |
13646 | // Try to use byte rotation instructions. |
13647 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate( |
13648 | DL, MVT::v64i8, V1, V2, Mask, Subtarget, DAG)) |
13649 | return Rotate; |
13650 | |
13651 | if (SDValue PSHUFB = lowerVectorShuffleWithPSHUFB( |
13652 | DL, MVT::v64i8, Mask, V1, V2, Zeroable, Subtarget, DAG)) |
13653 | return PSHUFB; |
13654 | |
13655 | // VBMI can use VPERMV/VPERMV3 byte shuffles. |
13656 | if (Subtarget.hasVBMI()) |
13657 | return lowerVectorShuffleWithPERMV(DL, MVT::v64i8, Mask, V1, V2, DAG); |
13658 | |
13659 | // Try to create an in-lane repeating shuffle mask and then shuffle the |
13660 | // the results into the target lanes. |
13661 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( |
13662 | DL, MVT::v64i8, V1, V2, Mask, Subtarget, DAG)) |
13663 | return V; |
13664 | |
13665 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v64i8, V1, V2, Mask, |
13666 | Zeroable, Subtarget, DAG)) |
13667 | return Blend; |
13668 | |
13669 | // FIXME: Implement direct support for this type! |
13670 | return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG); |
13671 | } |
13672 | |
13673 | /// \brief High-level routine to lower various 512-bit x86 vector shuffles. |
13674 | /// |
13675 | /// This routine either breaks down the specific type of a 512-bit x86 vector |
13676 | /// shuffle or splits it into two 256-bit shuffles and fuses the results back |
13677 | /// together based on the available instructions. |
13678 | static SDValue lower512BitVectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
13679 | MVT VT, SDValue V1, SDValue V2, |
13680 | const APInt &Zeroable, |
13681 | const X86Subtarget &Subtarget, |
13682 | SelectionDAG &DAG) { |
13683 | assert(Subtarget.hasAVX512() &&((Subtarget.hasAVX512() && "Cannot lower 512-bit vectors w/ basic ISA!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX512() && \"Cannot lower 512-bit vectors w/ basic ISA!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13684, __PRETTY_FUNCTION__)) |
13684 | "Cannot lower 512-bit vectors w/ basic ISA!")((Subtarget.hasAVX512() && "Cannot lower 512-bit vectors w/ basic ISA!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX512() && \"Cannot lower 512-bit vectors w/ basic ISA!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13684, __PRETTY_FUNCTION__)); |
13685 | |
13686 | // If we have a single input to the zero element, insert that into V1 if we |
13687 | // can do so cheaply. |
13688 | int NumElts = Mask.size(); |
13689 | int NumV2Elements = count_if(Mask, [NumElts](int M) { return M >= NumElts; }); |
13690 | |
13691 | if (NumV2Elements == 1 && Mask[0] >= NumElts) |
13692 | if (SDValue Insertion = lowerVectorShuffleAsElementInsertion( |
13693 | DL, VT, V1, V2, Mask, Zeroable, Subtarget, DAG)) |
13694 | return Insertion; |
13695 | |
13696 | // Check for being able to broadcast a single element. |
13697 | if (SDValue Broadcast = |
13698 | lowerVectorShuffleAsBroadcast(DL, VT, V1, V2, Mask, Subtarget, DAG)) |
13699 | return Broadcast; |
13700 | |
13701 | // Dispatch to each element type for lowering. If we don't have support for |
13702 | // specific element type shuffles at 512 bits, immediately split them and |
13703 | // lower them. Each lowering routine of a given type is allowed to assume that |
13704 | // the requisite ISA extensions for that element type are available. |
13705 | switch (VT.SimpleTy) { |
13706 | case MVT::v8f64: |
13707 | return lowerV8F64VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
13708 | case MVT::v16f32: |
13709 | return lowerV16F32VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
13710 | case MVT::v8i64: |
13711 | return lowerV8I64VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
13712 | case MVT::v16i32: |
13713 | return lowerV16I32VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
13714 | case MVT::v32i16: |
13715 | return lowerV32I16VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
13716 | case MVT::v64i8: |
13717 | return lowerV64I8VectorShuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); |
13718 | |
13719 | default: |
13720 | llvm_unreachable("Not a valid 512-bit x86 vector type!")::llvm::llvm_unreachable_internal("Not a valid 512-bit x86 vector type!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13720); |
13721 | } |
13722 | } |
13723 | |
13724 | // Lower vXi1 vector shuffles. |
13725 | // There is no a dedicated instruction on AVX-512 that shuffles the masks. |
13726 | // The only way to shuffle bits is to sign-extend the mask vector to SIMD |
13727 | // vector, shuffle and then truncate it back. |
13728 | static SDValue lower1BitVectorShuffle(const SDLoc &DL, ArrayRef<int> Mask, |
13729 | MVT VT, SDValue V1, SDValue V2, |
13730 | const X86Subtarget &Subtarget, |
13731 | SelectionDAG &DAG) { |
13732 | assert(Subtarget.hasAVX512() &&((Subtarget.hasAVX512() && "Cannot lower 512-bit vectors w/o basic ISA!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX512() && \"Cannot lower 512-bit vectors w/o basic ISA!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13733, __PRETTY_FUNCTION__)) |
13733 | "Cannot lower 512-bit vectors w/o basic ISA!")((Subtarget.hasAVX512() && "Cannot lower 512-bit vectors w/o basic ISA!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX512() && \"Cannot lower 512-bit vectors w/o basic ISA!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13733, __PRETTY_FUNCTION__)); |
13734 | MVT ExtVT; |
13735 | switch (VT.SimpleTy) { |
13736 | default: |
13737 | llvm_unreachable("Expected a vector of i1 elements")::llvm::llvm_unreachable_internal("Expected a vector of i1 elements" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13737); |
13738 | case MVT::v2i1: |
13739 | ExtVT = MVT::v2i64; |
13740 | break; |
13741 | case MVT::v4i1: |
13742 | ExtVT = MVT::v4i32; |
13743 | break; |
13744 | case MVT::v8i1: |
13745 | ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL |
13746 | break; |
13747 | case MVT::v16i1: |
13748 | ExtVT = MVT::v16i32; |
13749 | break; |
13750 | case MVT::v32i1: |
13751 | ExtVT = MVT::v32i16; |
13752 | break; |
13753 | case MVT::v64i1: |
13754 | ExtVT = MVT::v64i8; |
13755 | break; |
13756 | } |
13757 | |
13758 | if (ISD::isBuildVectorAllZeros(V1.getNode())) |
13759 | V1 = getZeroVector(ExtVT, Subtarget, DAG, DL); |
13760 | else if (ISD::isBuildVectorAllOnes(V1.getNode())) |
13761 | V1 = getOnesVector(ExtVT, DAG, DL); |
13762 | else |
13763 | V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1); |
13764 | |
13765 | if (V2.isUndef()) |
13766 | V2 = DAG.getUNDEF(ExtVT); |
13767 | else if (ISD::isBuildVectorAllZeros(V2.getNode())) |
13768 | V2 = getZeroVector(ExtVT, Subtarget, DAG, DL); |
13769 | else if (ISD::isBuildVectorAllOnes(V2.getNode())) |
13770 | V2 = getOnesVector(ExtVT, DAG, DL); |
13771 | else |
13772 | V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2); |
13773 | |
13774 | SDValue Shuffle = DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask); |
13775 | // i1 was sign extended we can use X86ISD::CVT2MASK. |
13776 | int NumElems = VT.getVectorNumElements(); |
13777 | if ((Subtarget.hasBWI() && (NumElems >= 32)) || |
13778 | (Subtarget.hasDQI() && (NumElems < 32))) |
13779 | return DAG.getNode(X86ISD::CVT2MASK, DL, VT, Shuffle); |
13780 | |
13781 | return DAG.getNode(ISD::TRUNCATE, DL, VT, Shuffle); |
13782 | } |
13783 | |
13784 | /// Helper function that returns true if the shuffle mask should be |
13785 | /// commuted to improve canonicalization. |
13786 | static bool canonicalizeShuffleMaskWithCommute(ArrayRef<int> Mask) { |
13787 | int NumElements = Mask.size(); |
13788 | |
13789 | int NumV1Elements = 0, NumV2Elements = 0; |
13790 | for (int M : Mask) |
13791 | if (M < 0) |
13792 | continue; |
13793 | else if (M < NumElements) |
13794 | ++NumV1Elements; |
13795 | else |
13796 | ++NumV2Elements; |
13797 | |
13798 | // Commute the shuffle as needed such that more elements come from V1 than |
13799 | // V2. This allows us to match the shuffle pattern strictly on how many |
13800 | // elements come from V1 without handling the symmetric cases. |
13801 | if (NumV2Elements > NumV1Elements) |
13802 | return true; |
13803 | |
13804 | assert(NumV1Elements > 0 && "No V1 indices")((NumV1Elements > 0 && "No V1 indices") ? static_cast <void> (0) : __assert_fail ("NumV1Elements > 0 && \"No V1 indices\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13804, __PRETTY_FUNCTION__)); |
13805 | |
13806 | if (NumV2Elements == 0) |
13807 | return false; |
13808 | |
13809 | // When the number of V1 and V2 elements are the same, try to minimize the |
13810 | // number of uses of V2 in the low half of the vector. When that is tied, |
13811 | // ensure that the sum of indices for V1 is equal to or lower than the sum |
13812 | // indices for V2. When those are equal, try to ensure that the number of odd |
13813 | // indices for V1 is lower than the number of odd indices for V2. |
13814 | if (NumV1Elements == NumV2Elements) { |
13815 | int LowV1Elements = 0, LowV2Elements = 0; |
13816 | for (int M : Mask.slice(0, NumElements / 2)) |
13817 | if (M >= NumElements) |
13818 | ++LowV2Elements; |
13819 | else if (M >= 0) |
13820 | ++LowV1Elements; |
13821 | if (LowV2Elements > LowV1Elements) |
13822 | return true; |
13823 | if (LowV2Elements == LowV1Elements) { |
13824 | int SumV1Indices = 0, SumV2Indices = 0; |
13825 | for (int i = 0, Size = Mask.size(); i < Size; ++i) |
13826 | if (Mask[i] >= NumElements) |
13827 | SumV2Indices += i; |
13828 | else if (Mask[i] >= 0) |
13829 | SumV1Indices += i; |
13830 | if (SumV2Indices < SumV1Indices) |
13831 | return true; |
13832 | if (SumV2Indices == SumV1Indices) { |
13833 | int NumV1OddIndices = 0, NumV2OddIndices = 0; |
13834 | for (int i = 0, Size = Mask.size(); i < Size; ++i) |
13835 | if (Mask[i] >= NumElements) |
13836 | NumV2OddIndices += i % 2; |
13837 | else if (Mask[i] >= 0) |
13838 | NumV1OddIndices += i % 2; |
13839 | if (NumV2OddIndices < NumV1OddIndices) |
13840 | return true; |
13841 | } |
13842 | } |
13843 | } |
13844 | |
13845 | return false; |
13846 | } |
13847 | |
13848 | /// \brief Top-level lowering for x86 vector shuffles. |
13849 | /// |
13850 | /// This handles decomposition, canonicalization, and lowering of all x86 |
13851 | /// vector shuffles. Most of the specific lowering strategies are encapsulated |
13852 | /// above in helper routines. The canonicalization attempts to widen shuffles |
13853 | /// to involve fewer lanes of wider elements, consolidate symmetric patterns |
13854 | /// s.t. only one of the two inputs needs to be tested, etc. |
13855 | static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget &Subtarget, |
13856 | SelectionDAG &DAG) { |
13857 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); |
13858 | ArrayRef<int> Mask = SVOp->getMask(); |
13859 | SDValue V1 = Op.getOperand(0); |
13860 | SDValue V2 = Op.getOperand(1); |
13861 | MVT VT = Op.getSimpleValueType(); |
13862 | int NumElements = VT.getVectorNumElements(); |
13863 | SDLoc DL(Op); |
13864 | bool Is1BitVector = (VT.getVectorElementType() == MVT::i1); |
13865 | |
13866 | assert((VT.getSizeInBits() != 64 || Is1BitVector) &&(((VT.getSizeInBits() != 64 || Is1BitVector) && "Can't lower MMX shuffles" ) ? static_cast<void> (0) : __assert_fail ("(VT.getSizeInBits() != 64 || Is1BitVector) && \"Can't lower MMX shuffles\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13867, __PRETTY_FUNCTION__)) |
13867 | "Can't lower MMX shuffles")(((VT.getSizeInBits() != 64 || Is1BitVector) && "Can't lower MMX shuffles" ) ? static_cast<void> (0) : __assert_fail ("(VT.getSizeInBits() != 64 || Is1BitVector) && \"Can't lower MMX shuffles\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13867, __PRETTY_FUNCTION__)); |
13868 | |
13869 | bool V1IsUndef = V1.isUndef(); |
13870 | bool V2IsUndef = V2.isUndef(); |
13871 | if (V1IsUndef && V2IsUndef) |
13872 | return DAG.getUNDEF(VT); |
13873 | |
13874 | // When we create a shuffle node we put the UNDEF node to second operand, |
13875 | // but in some cases the first operand may be transformed to UNDEF. |
13876 | // In this case we should just commute the node. |
13877 | if (V1IsUndef) |
13878 | return DAG.getCommutedVectorShuffle(*SVOp); |
13879 | |
13880 | // Check for non-undef masks pointing at an undef vector and make the masks |
13881 | // undef as well. This makes it easier to match the shuffle based solely on |
13882 | // the mask. |
13883 | if (V2IsUndef) |
13884 | for (int M : Mask) |
13885 | if (M >= NumElements) { |
13886 | SmallVector<int, 8> NewMask(Mask.begin(), Mask.end()); |
13887 | for (int &M : NewMask) |
13888 | if (M >= NumElements) |
13889 | M = -1; |
13890 | return DAG.getVectorShuffle(VT, DL, V1, V2, NewMask); |
13891 | } |
13892 | |
13893 | // Check for illegal shuffle mask element index values. |
13894 | int MaskUpperLimit = Mask.size() * (V2IsUndef ? 1 : 2); (void)MaskUpperLimit; |
13895 | assert(llvm::all_of(Mask,((llvm::all_of(Mask, [&](int M) { return -1 <= M && M < MaskUpperLimit; }) && "Out of bounds shuffle index" ) ? static_cast<void> (0) : __assert_fail ("llvm::all_of(Mask, [&](int M) { return -1 <= M && M < MaskUpperLimit; }) && \"Out of bounds shuffle index\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13897, __PRETTY_FUNCTION__)) |
13896 | [&](int M) { return -1 <= M && M < MaskUpperLimit; }) &&((llvm::all_of(Mask, [&](int M) { return -1 <= M && M < MaskUpperLimit; }) && "Out of bounds shuffle index" ) ? static_cast<void> (0) : __assert_fail ("llvm::all_of(Mask, [&](int M) { return -1 <= M && M < MaskUpperLimit; }) && \"Out of bounds shuffle index\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13897, __PRETTY_FUNCTION__)) |
13897 | "Out of bounds shuffle index")((llvm::all_of(Mask, [&](int M) { return -1 <= M && M < MaskUpperLimit; }) && "Out of bounds shuffle index" ) ? static_cast<void> (0) : __assert_fail ("llvm::all_of(Mask, [&](int M) { return -1 <= M && M < MaskUpperLimit; }) && \"Out of bounds shuffle index\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13897, __PRETTY_FUNCTION__)); |
13898 | |
13899 | // We actually see shuffles that are entirely re-arrangements of a set of |
13900 | // zero inputs. This mostly happens while decomposing complex shuffles into |
13901 | // simple ones. Directly lower these as a buildvector of zeros. |
13902 | APInt Zeroable = computeZeroableShuffleElements(Mask, V1, V2); |
13903 | if (Zeroable.isAllOnesValue()) |
13904 | return getZeroVector(VT, Subtarget, DAG, DL); |
13905 | |
13906 | // Try to collapse shuffles into using a vector type with fewer elements but |
13907 | // wider element types. We cap this to not form integers or floating point |
13908 | // elements wider than 64 bits, but it might be interesting to form i128 |
13909 | // integers to handle flipping the low and high halves of AVX 256-bit vectors. |
13910 | SmallVector<int, 16> WidenedMask; |
13911 | if (VT.getScalarSizeInBits() < 64 && !Is1BitVector && |
13912 | canWidenShuffleElements(Mask, WidenedMask)) { |
13913 | MVT NewEltVT = VT.isFloatingPoint() |
13914 | ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2) |
13915 | : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2); |
13916 | MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2); |
13917 | // Make sure that the new vector type is legal. For example, v2f64 isn't |
13918 | // legal on SSE1. |
13919 | if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) { |
13920 | V1 = DAG.getBitcast(NewVT, V1); |
13921 | V2 = DAG.getBitcast(NewVT, V2); |
13922 | return DAG.getBitcast( |
13923 | VT, DAG.getVectorShuffle(NewVT, DL, V1, V2, WidenedMask)); |
13924 | } |
13925 | } |
13926 | |
13927 | // Commute the shuffle if it will improve canonicalization. |
13928 | if (canonicalizeShuffleMaskWithCommute(Mask)) |
13929 | return DAG.getCommutedVectorShuffle(*SVOp); |
13930 | |
13931 | // For each vector width, delegate to a specialized lowering routine. |
13932 | if (VT.is128BitVector()) |
13933 | return lower128BitVectorShuffle(DL, Mask, VT, V1, V2, Zeroable, Subtarget, |
13934 | DAG); |
13935 | |
13936 | if (VT.is256BitVector()) |
13937 | return lower256BitVectorShuffle(DL, Mask, VT, V1, V2, Zeroable, Subtarget, |
13938 | DAG); |
13939 | |
13940 | if (VT.is512BitVector()) |
13941 | return lower512BitVectorShuffle(DL, Mask, VT, V1, V2, Zeroable, Subtarget, |
13942 | DAG); |
13943 | |
13944 | if (Is1BitVector) |
13945 | return lower1BitVectorShuffle(DL, Mask, VT, V1, V2, Subtarget, DAG); |
13946 | |
13947 | llvm_unreachable("Unimplemented!")::llvm::llvm_unreachable_internal("Unimplemented!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 13947); |
13948 | } |
13949 | |
13950 | /// \brief Try to lower a VSELECT instruction to a vector shuffle. |
13951 | static SDValue lowerVSELECTtoVectorShuffle(SDValue Op, |
13952 | const X86Subtarget &Subtarget, |
13953 | SelectionDAG &DAG) { |
13954 | SDValue Cond = Op.getOperand(0); |
13955 | SDValue LHS = Op.getOperand(1); |
13956 | SDValue RHS = Op.getOperand(2); |
13957 | SDLoc dl(Op); |
13958 | MVT VT = Op.getSimpleValueType(); |
13959 | |
13960 | if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) |
13961 | return SDValue(); |
13962 | auto *CondBV = cast<BuildVectorSDNode>(Cond); |
13963 | |
13964 | // Only non-legal VSELECTs reach this lowering, convert those into generic |
13965 | // shuffles and re-use the shuffle lowering path for blends. |
13966 | SmallVector<int, 32> Mask; |
13967 | for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) { |
13968 | SDValue CondElt = CondBV->getOperand(i); |
13969 | Mask.push_back( |
13970 | isa<ConstantSDNode>(CondElt) ? i + (isNullConstant(CondElt) ? Size : 0) |
13971 | : -1); |
13972 | } |
13973 | return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask); |
13974 | } |
13975 | |
13976 | SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const { |
13977 | // A vselect where all conditions and data are constants can be optimized into |
13978 | // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR(). |
13979 | if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) && |
13980 | ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) && |
13981 | ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode())) |
13982 | return SDValue(); |
13983 | |
13984 | // If this VSELECT has a vector if i1 as a mask, it will be directly matched |
13985 | // with patterns on the mask registers on AVX-512. |
13986 | if (Op->getOperand(0).getValueType().getScalarSizeInBits() == 1) |
13987 | return Op; |
13988 | |
13989 | // Try to lower this to a blend-style vector shuffle. This can handle all |
13990 | // constant condition cases. |
13991 | if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG)) |
13992 | return BlendOp; |
13993 | |
13994 | // Variable blends are only legal from SSE4.1 onward. |
13995 | if (!Subtarget.hasSSE41()) |
13996 | return SDValue(); |
13997 | |
13998 | SDLoc dl(Op); |
13999 | MVT VT = Op.getSimpleValueType(); |
14000 | |
14001 | // If the VSELECT is on a 512-bit type, we have to convert a non-i1 condition |
14002 | // into an i1 condition so that we can use the mask-based 512-bit blend |
14003 | // instructions. |
14004 | if (VT.getSizeInBits() == 512) { |
14005 | SDValue Cond = Op.getOperand(0); |
14006 | // The vNi1 condition case should be handled above as it can be trivially |
14007 | // lowered. |
14008 | assert(Cond.getValueType().getScalarSizeInBits() ==((Cond.getValueType().getScalarSizeInBits() == VT.getScalarSizeInBits () && "Should have a size-matched integer condition!" ) ? static_cast<void> (0) : __assert_fail ("Cond.getValueType().getScalarSizeInBits() == VT.getScalarSizeInBits() && \"Should have a size-matched integer condition!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14010, __PRETTY_FUNCTION__)) |
14009 | VT.getScalarSizeInBits() &&((Cond.getValueType().getScalarSizeInBits() == VT.getScalarSizeInBits () && "Should have a size-matched integer condition!" ) ? static_cast<void> (0) : __assert_fail ("Cond.getValueType().getScalarSizeInBits() == VT.getScalarSizeInBits() && \"Should have a size-matched integer condition!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14010, __PRETTY_FUNCTION__)) |
14010 | "Should have a size-matched integer condition!")((Cond.getValueType().getScalarSizeInBits() == VT.getScalarSizeInBits () && "Should have a size-matched integer condition!" ) ? static_cast<void> (0) : __assert_fail ("Cond.getValueType().getScalarSizeInBits() == VT.getScalarSizeInBits() && \"Should have a size-matched integer condition!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14010, __PRETTY_FUNCTION__)); |
14011 | // Build a mask by testing the condition against itself (tests for zero). |
14012 | MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements()); |
14013 | SDValue Mask = DAG.getNode(X86ISD::TESTM, dl, MaskVT, Cond, Cond); |
14014 | // Now return a new VSELECT using the mask. |
14015 | return DAG.getSelect(dl, VT, Mask, Op.getOperand(1), Op.getOperand(2)); |
14016 | } |
14017 | |
14018 | // Only some types will be legal on some subtargets. If we can emit a legal |
14019 | // VSELECT-matching blend, return Op, and but if we need to expand, return |
14020 | // a null value. |
14021 | switch (VT.SimpleTy) { |
14022 | default: |
14023 | // Most of the vector types have blends past SSE4.1. |
14024 | return Op; |
14025 | |
14026 | case MVT::v32i8: |
14027 | // The byte blends for AVX vectors were introduced only in AVX2. |
14028 | if (Subtarget.hasAVX2()) |
14029 | return Op; |
14030 | |
14031 | return SDValue(); |
14032 | |
14033 | case MVT::v8i16: |
14034 | case MVT::v16i16: |
14035 | // AVX-512 BWI and VLX features support VSELECT with i16 elements. |
14036 | if (Subtarget.hasBWI() && Subtarget.hasVLX()) |
14037 | return Op; |
14038 | |
14039 | // FIXME: We should custom lower this by fixing the condition and using i8 |
14040 | // blends. |
14041 | return SDValue(); |
14042 | } |
14043 | } |
14044 | |
14045 | static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) { |
14046 | MVT VT = Op.getSimpleValueType(); |
14047 | SDLoc dl(Op); |
14048 | |
14049 | if (!Op.getOperand(0).getSimpleValueType().is128BitVector()) |
14050 | return SDValue(); |
14051 | |
14052 | if (VT.getSizeInBits() == 8) { |
14053 | SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32, |
14054 | Op.getOperand(0), Op.getOperand(1)); |
14055 | SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract, |
14056 | DAG.getValueType(VT)); |
14057 | return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert); |
14058 | } |
14059 | |
14060 | if (VT == MVT::f32) { |
14061 | // EXTRACTPS outputs to a GPR32 register which will require a movd to copy |
14062 | // the result back to FR32 register. It's only worth matching if the |
14063 | // result has a single use which is a store or a bitcast to i32. And in |
14064 | // the case of a store, it's not worth it if the index is a constant 0, |
14065 | // because a MOVSSmr can be used instead, which is smaller and faster. |
14066 | if (!Op.hasOneUse()) |
14067 | return SDValue(); |
14068 | SDNode *User = *Op.getNode()->use_begin(); |
14069 | if ((User->getOpcode() != ISD::STORE || |
14070 | isNullConstant(Op.getOperand(1))) && |
14071 | (User->getOpcode() != ISD::BITCAST || |
14072 | User->getValueType(0) != MVT::i32)) |
14073 | return SDValue(); |
14074 | SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, |
14075 | DAG.getBitcast(MVT::v4i32, Op.getOperand(0)), |
14076 | Op.getOperand(1)); |
14077 | return DAG.getBitcast(MVT::f32, Extract); |
14078 | } |
14079 | |
14080 | if (VT == MVT::i32 || VT == MVT::i64) { |
14081 | // ExtractPS/pextrq works with constant index. |
14082 | if (isa<ConstantSDNode>(Op.getOperand(1))) |
14083 | return Op; |
14084 | } |
14085 | |
14086 | return SDValue(); |
14087 | } |
14088 | |
14089 | /// Extract one bit from mask vector, like v16i1 or v8i1. |
14090 | /// AVX-512 feature. |
14091 | SDValue |
14092 | X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const { |
14093 | SDValue Vec = Op.getOperand(0); |
14094 | SDLoc dl(Vec); |
14095 | MVT VecVT = Vec.getSimpleValueType(); |
14096 | SDValue Idx = Op.getOperand(1); |
14097 | MVT EltVT = Op.getSimpleValueType(); |
14098 | |
14099 | assert((VecVT.getVectorNumElements() <= 16 || Subtarget.hasBWI()) &&(((VecVT.getVectorNumElements() <= 16 || Subtarget.hasBWI( )) && "Unexpected vector type in ExtractBitFromMaskVector" ) ? static_cast<void> (0) : __assert_fail ("(VecVT.getVectorNumElements() <= 16 || Subtarget.hasBWI()) && \"Unexpected vector type in ExtractBitFromMaskVector\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14100, __PRETTY_FUNCTION__)) |
14100 | "Unexpected vector type in ExtractBitFromMaskVector")(((VecVT.getVectorNumElements() <= 16 || Subtarget.hasBWI( )) && "Unexpected vector type in ExtractBitFromMaskVector" ) ? static_cast<void> (0) : __assert_fail ("(VecVT.getVectorNumElements() <= 16 || Subtarget.hasBWI()) && \"Unexpected vector type in ExtractBitFromMaskVector\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14100, __PRETTY_FUNCTION__)); |
14101 | |
14102 | // variable index can't be handled in mask registers, |
14103 | // extend vector to VR512/128 |
14104 | if (!isa<ConstantSDNode>(Idx)) { |
14105 | unsigned NumElts = VecVT.getVectorNumElements(); |
14106 | // Extending v8i1/v16i1 to 512-bit get better performance on KNL |
14107 | // than extending to 128/256bit. |
14108 | unsigned VecSize = (NumElts <= 4 ? 128 : 512); |
14109 | MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(VecSize/NumElts), NumElts); |
14110 | SDValue Ext = DAG.getNode(ISD::SIGN_EXTEND, dl, ExtVT, Vec); |
14111 | SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, |
14112 | ExtVT.getVectorElementType(), Ext, Idx); |
14113 | return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt); |
14114 | } |
14115 | |
14116 | unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); |
14117 | if ((!Subtarget.hasDQI() && (VecVT.getVectorNumElements() == 8)) || |
14118 | (VecVT.getVectorNumElements() < 8)) { |
14119 | // Use kshiftlw/rw instruction. |
14120 | VecVT = MVT::v16i1; |
14121 | Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, VecVT, |
14122 | DAG.getUNDEF(VecVT), |
14123 | Vec, |
14124 | DAG.getIntPtrConstant(0, dl)); |
14125 | } |
14126 | unsigned MaxSift = VecVT.getVectorNumElements() - 1; |
14127 | if (MaxSift - IdxVal) |
14128 | Vec = DAG.getNode(X86ISD::KSHIFTL, dl, VecVT, Vec, |
14129 | DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8)); |
14130 | Vec = DAG.getNode(X86ISD::KSHIFTR, dl, VecVT, Vec, |
14131 | DAG.getConstant(MaxSift, dl, MVT::i8)); |
14132 | return DAG.getNode(X86ISD::VEXTRACT, dl, Op.getSimpleValueType(), Vec, |
14133 | DAG.getIntPtrConstant(0, dl)); |
14134 | } |
14135 | |
14136 | SDValue |
14137 | X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, |
14138 | SelectionDAG &DAG) const { |
14139 | SDLoc dl(Op); |
14140 | SDValue Vec = Op.getOperand(0); |
14141 | MVT VecVT = Vec.getSimpleValueType(); |
14142 | SDValue Idx = Op.getOperand(1); |
14143 | |
14144 | if (VecVT.getVectorElementType() == MVT::i1) |
14145 | return ExtractBitFromMaskVector(Op, DAG); |
14146 | |
14147 | if (!isa<ConstantSDNode>(Idx)) { |
14148 | // Its more profitable to go through memory (1 cycles throughput) |
14149 | // than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) |
14150 | // IACA tool was used to get performance estimation |
14151 | // (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) |
14152 | // |
14153 | // example : extractelement <16 x i8> %a, i32 %i |
14154 | // |
14155 | // Block Throughput: 3.00 Cycles |
14156 | // Throughput Bottleneck: Port5 |
14157 | // |
14158 | // | Num Of | Ports pressure in cycles | | |
14159 | // | Uops | 0 - DV | 5 | 6 | 7 | | |
14160 | // --------------------------------------------- |
14161 | // | 1 | | 1.0 | | | CP | vmovd xmm1, edi |
14162 | // | 1 | | 1.0 | | | CP | vpshufb xmm0, xmm0, xmm1 |
14163 | // | 2 | 1.0 | 1.0 | | | CP | vpextrb eax, xmm0, 0x0 |
14164 | // Total Num Of Uops: 4 |
14165 | // |
14166 | // |
14167 | // Block Throughput: 1.00 Cycles |
14168 | // Throughput Bottleneck: PORT2_AGU, PORT3_AGU, Port4 |
14169 | // |
14170 | // | | Ports pressure in cycles | | |
14171 | // |Uops| 1 | 2 - D |3 - D | 4 | 5 | | |
14172 | // --------------------------------------------------------- |
14173 | // |2^ | | 0.5 | 0.5 |1.0| |CP| vmovaps xmmword ptr [rsp-0x18], xmm0 |
14174 | // |1 |0.5| | | |0.5| | lea rax, ptr [rsp-0x18] |
14175 | // |1 | |0.5, 0.5|0.5, 0.5| | |CP| mov al, byte ptr [rdi+rax*1] |
14176 | // Total Num Of Uops: 4 |
14177 | |
14178 | return SDValue(); |
14179 | } |
14180 | |
14181 | unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); |
14182 | |
14183 | // If this is a 256-bit vector result, first extract the 128-bit vector and |
14184 | // then extract the element from the 128-bit vector. |
14185 | if (VecVT.is256BitVector() || VecVT.is512BitVector()) { |
14186 | // Get the 128-bit vector. |
14187 | Vec = extract128BitVector(Vec, IdxVal, DAG, dl); |
14188 | MVT EltVT = VecVT.getVectorElementType(); |
14189 | |
14190 | unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits(); |
14191 | assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2")((isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(ElemsPerChunk) && \"Elements per chunk not power of 2\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14191, __PRETTY_FUNCTION__)); |
14192 | |
14193 | // Find IdxVal modulo ElemsPerChunk. Since ElemsPerChunk is a power of 2 |
14194 | // this can be done with a mask. |
14195 | IdxVal &= ElemsPerChunk - 1; |
14196 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec, |
14197 | DAG.getConstant(IdxVal, dl, MVT::i32)); |
14198 | } |
14199 | |
14200 | assert(VecVT.is128BitVector() && "Unexpected vector length")((VecVT.is128BitVector() && "Unexpected vector length" ) ? static_cast<void> (0) : __assert_fail ("VecVT.is128BitVector() && \"Unexpected vector length\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14200, __PRETTY_FUNCTION__)); |
14201 | |
14202 | MVT VT = Op.getSimpleValueType(); |
14203 | |
14204 | if (VT.getSizeInBits() == 16) { |
14205 | // If IdxVal is 0, it's cheaper to do a move instead of a pextrw, unless |
14206 | // we're going to zero extend the register or fold the store (SSE41 only). |
14207 | if (IdxVal == 0 && !MayFoldIntoZeroExtend(Op) && |
14208 | !(Subtarget.hasSSE41() && MayFoldIntoStore(Op))) |
14209 | return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, |
14210 | DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, |
14211 | DAG.getBitcast(MVT::v4i32, Vec), Idx)); |
14212 | |
14213 | // Transform it so it match pextrw which produces a 32-bit result. |
14214 | SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32, |
14215 | Op.getOperand(0), Op.getOperand(1)); |
14216 | SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract, |
14217 | DAG.getValueType(VT)); |
14218 | return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert); |
14219 | } |
14220 | |
14221 | if (Subtarget.hasSSE41()) |
14222 | if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG)) |
14223 | return Res; |
14224 | |
14225 | // TODO: We only extract a single element from v16i8, we can probably afford |
14226 | // to be more aggressive here before using the default approach of spilling to |
14227 | // stack. |
14228 | if (VT.getSizeInBits() == 8 && Op->isOnlyUserOf(Vec.getNode())) { |
14229 | // Extract either the lowest i32 or any i16, and extract the sub-byte. |
14230 | int DWordIdx = IdxVal / 4; |
14231 | if (DWordIdx == 0) { |
14232 | SDValue Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, |
14233 | DAG.getBitcast(MVT::v4i32, Vec), |
14234 | DAG.getIntPtrConstant(DWordIdx, dl)); |
14235 | int ShiftVal = (IdxVal % 4) * 8; |
14236 | if (ShiftVal != 0) |
14237 | Res = DAG.getNode(ISD::SRL, dl, MVT::i32, Res, |
14238 | DAG.getConstant(ShiftVal, dl, MVT::i32)); |
14239 | return DAG.getNode(ISD::TRUNCATE, dl, VT, Res); |
14240 | } |
14241 | |
14242 | int WordIdx = IdxVal / 2; |
14243 | SDValue Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, |
14244 | DAG.getBitcast(MVT::v8i16, Vec), |
14245 | DAG.getIntPtrConstant(WordIdx, dl)); |
14246 | int ShiftVal = (IdxVal % 2) * 8; |
14247 | if (ShiftVal != 0) |
14248 | Res = DAG.getNode(ISD::SRL, dl, MVT::i16, Res, |
14249 | DAG.getConstant(ShiftVal, dl, MVT::i16)); |
14250 | return DAG.getNode(ISD::TRUNCATE, dl, VT, Res); |
14251 | } |
14252 | |
14253 | if (VT.getSizeInBits() == 32) { |
14254 | if (IdxVal == 0) |
14255 | return Op; |
14256 | |
14257 | // SHUFPS the element to the lowest double word, then movss. |
14258 | int Mask[4] = { static_cast<int>(IdxVal), -1, -1, -1 }; |
14259 | Vec = DAG.getVectorShuffle(VecVT, dl, Vec, DAG.getUNDEF(VecVT), Mask); |
14260 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec, |
14261 | DAG.getIntPtrConstant(0, dl)); |
14262 | } |
14263 | |
14264 | if (VT.getSizeInBits() == 64) { |
14265 | // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b |
14266 | // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught |
14267 | // to match extract_elt for f64. |
14268 | if (IdxVal == 0) |
14269 | return Op; |
14270 | |
14271 | // UNPCKHPD the element to the lowest double word, then movsd. |
14272 | // Note if the lower 64 bits of the result of the UNPCKHPD is then stored |
14273 | // to a f64mem, the whole operation is folded into a single MOVHPDmr. |
14274 | int Mask[2] = { 1, -1 }; |
14275 | Vec = DAG.getVectorShuffle(VecVT, dl, Vec, DAG.getUNDEF(VecVT), Mask); |
14276 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec, |
14277 | DAG.getIntPtrConstant(0, dl)); |
14278 | } |
14279 | |
14280 | return SDValue(); |
14281 | } |
14282 | |
14283 | /// Insert one bit to mask vector, like v16i1 or v8i1. |
14284 | /// AVX-512 feature. |
14285 | SDValue |
14286 | X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const { |
14287 | SDLoc dl(Op); |
14288 | SDValue Vec = Op.getOperand(0); |
14289 | SDValue Elt = Op.getOperand(1); |
14290 | SDValue Idx = Op.getOperand(2); |
14291 | MVT VecVT = Vec.getSimpleValueType(); |
14292 | |
14293 | if (!isa<ConstantSDNode>(Idx)) { |
14294 | // Non constant index. Extend source and destination, |
14295 | // insert element and then truncate the result. |
14296 | MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32); |
14297 | MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32); |
14298 | SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT, |
14299 | DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec), |
14300 | DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx); |
14301 | return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp); |
14302 | } |
14303 | |
14304 | unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); |
14305 | SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt); |
14306 | unsigned NumElems = VecVT.getVectorNumElements(); |
14307 | |
14308 | if(Vec.isUndef()) { |
14309 | if (IdxVal) |
14310 | EltInVec = DAG.getNode(X86ISD::KSHIFTL, dl, VecVT, EltInVec, |
14311 | DAG.getConstant(IdxVal, dl, MVT::i8)); |
14312 | return EltInVec; |
14313 | } |
14314 | |
14315 | // Insertion of one bit into first position |
14316 | if (IdxVal == 0 ) { |
14317 | // Clean top bits of vector. |
14318 | EltInVec = DAG.getNode(X86ISD::KSHIFTL, dl, VecVT, EltInVec, |
14319 | DAG.getConstant(NumElems - 1, dl, MVT::i8)); |
14320 | EltInVec = DAG.getNode(X86ISD::KSHIFTR, dl, VecVT, EltInVec, |
14321 | DAG.getConstant(NumElems - 1, dl, MVT::i8)); |
14322 | // Clean the first bit in source vector. |
14323 | Vec = DAG.getNode(X86ISD::KSHIFTR, dl, VecVT, Vec, |
14324 | DAG.getConstant(1 , dl, MVT::i8)); |
14325 | Vec = DAG.getNode(X86ISD::KSHIFTL, dl, VecVT, Vec, |
14326 | DAG.getConstant(1, dl, MVT::i8)); |
14327 | |
14328 | return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec); |
14329 | } |
14330 | // Insertion of one bit into last position |
14331 | if (IdxVal == NumElems -1) { |
14332 | // Move the bit to the last position inside the vector. |
14333 | EltInVec = DAG.getNode(X86ISD::KSHIFTL, dl, VecVT, EltInVec, |
14334 | DAG.getConstant(IdxVal, dl, MVT::i8)); |
14335 | // Clean the last bit in the source vector. |
14336 | Vec = DAG.getNode(X86ISD::KSHIFTL, dl, VecVT, Vec, |
14337 | DAG.getConstant(1, dl, MVT::i8)); |
14338 | Vec = DAG.getNode(X86ISD::KSHIFTR, dl, VecVT, Vec, |
14339 | DAG.getConstant(1 , dl, MVT::i8)); |
14340 | |
14341 | return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec); |
14342 | } |
14343 | |
14344 | // Use shuffle to insert element. |
14345 | SmallVector<int, 64> MaskVec(NumElems); |
14346 | for (unsigned i = 0; i != NumElems; ++i) |
14347 | MaskVec[i] = (i == IdxVal) ? NumElems : i; |
14348 | |
14349 | return DAG.getVectorShuffle(VecVT, dl, Vec, EltInVec, MaskVec); |
14350 | } |
14351 | |
14352 | SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, |
14353 | SelectionDAG &DAG) const { |
14354 | MVT VT = Op.getSimpleValueType(); |
14355 | MVT EltVT = VT.getVectorElementType(); |
14356 | unsigned NumElts = VT.getVectorNumElements(); |
14357 | |
14358 | if (EltVT == MVT::i1) |
14359 | return InsertBitToMaskVector(Op, DAG); |
14360 | |
14361 | SDLoc dl(Op); |
14362 | SDValue N0 = Op.getOperand(0); |
14363 | SDValue N1 = Op.getOperand(1); |
14364 | SDValue N2 = Op.getOperand(2); |
14365 | if (!isa<ConstantSDNode>(N2)) |
14366 | return SDValue(); |
14367 | auto *N2C = cast<ConstantSDNode>(N2); |
14368 | unsigned IdxVal = N2C->getZExtValue(); |
14369 | |
14370 | bool IsZeroElt = X86::isZeroNode(N1); |
14371 | bool IsAllOnesElt = VT.isInteger() && llvm::isAllOnesConstant(N1); |
14372 | |
14373 | // If we are inserting a element, see if we can do this more efficiently with |
14374 | // a blend shuffle with a rematerializable vector than a costly integer |
14375 | // insertion. |
14376 | if ((IsZeroElt || IsAllOnesElt) && Subtarget.hasSSE41() && |
14377 | 16 <= EltVT.getSizeInBits()) { |
14378 | SmallVector<int, 8> BlendMask; |
14379 | for (unsigned i = 0; i != NumElts; ++i) |
14380 | BlendMask.push_back(i == IdxVal ? i + NumElts : i); |
14381 | SDValue CstVector = IsZeroElt ? getZeroVector(VT, Subtarget, DAG, dl) |
14382 | : DAG.getConstant(-1, dl, VT); |
14383 | return DAG.getVectorShuffle(VT, dl, N0, CstVector, BlendMask); |
14384 | } |
14385 | |
14386 | // If the vector is wider than 128 bits, extract the 128-bit subvector, insert |
14387 | // into that, and then insert the subvector back into the result. |
14388 | if (VT.is256BitVector() || VT.is512BitVector()) { |
14389 | // With a 256-bit vector, we can insert into the zero element efficiently |
14390 | // using a blend if we have AVX or AVX2 and the right data type. |
14391 | if (VT.is256BitVector() && IdxVal == 0) { |
14392 | // TODO: It is worthwhile to cast integer to floating point and back |
14393 | // and incur a domain crossing penalty if that's what we'll end up |
14394 | // doing anyway after extracting to a 128-bit vector. |
14395 | if ((Subtarget.hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) || |
14396 | (Subtarget.hasAVX2() && EltVT == MVT::i32)) { |
14397 | SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1); |
14398 | N2 = DAG.getIntPtrConstant(1, dl); |
14399 | return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2); |
14400 | } |
14401 | } |
14402 | |
14403 | // Get the desired 128-bit vector chunk. |
14404 | SDValue V = extract128BitVector(N0, IdxVal, DAG, dl); |
14405 | |
14406 | // Insert the element into the desired chunk. |
14407 | unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits(); |
14408 | assert(isPowerOf2_32(NumEltsIn128))((isPowerOf2_32(NumEltsIn128)) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(NumEltsIn128)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14408, __PRETTY_FUNCTION__)); |
14409 | // Since NumEltsIn128 is a power of 2 we can use mask instead of modulo. |
14410 | unsigned IdxIn128 = IdxVal & (NumEltsIn128 - 1); |
14411 | |
14412 | V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1, |
14413 | DAG.getConstant(IdxIn128, dl, MVT::i32)); |
14414 | |
14415 | // Insert the changed part back into the bigger vector |
14416 | return insert128BitVector(N0, V, IdxVal, DAG, dl); |
14417 | } |
14418 | assert(VT.is128BitVector() && "Only 128-bit vector types should be left!")((VT.is128BitVector() && "Only 128-bit vector types should be left!" ) ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"Only 128-bit vector types should be left!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14418, __PRETTY_FUNCTION__)); |
14419 | |
14420 | // Transform it so it match pinsr{b,w} which expects a GR32 as its second |
14421 | // argument. SSE41 required for pinsrb. |
14422 | if (VT == MVT::v8i16 || (VT == MVT::v16i8 && Subtarget.hasSSE41())) { |
14423 | unsigned Opc; |
14424 | if (VT == MVT::v8i16) { |
14425 | assert(Subtarget.hasSSE2() && "SSE2 required for PINSRW")((Subtarget.hasSSE2() && "SSE2 required for PINSRW") ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSE2() && \"SSE2 required for PINSRW\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14425, __PRETTY_FUNCTION__)); |
14426 | Opc = X86ISD::PINSRW; |
14427 | } else { |
14428 | assert(VT == MVT::v16i8 && "PINSRB requires v16i8 vector")((VT == MVT::v16i8 && "PINSRB requires v16i8 vector") ? static_cast<void> (0) : __assert_fail ("VT == MVT::v16i8 && \"PINSRB requires v16i8 vector\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14428, __PRETTY_FUNCTION__)); |
14429 | assert(Subtarget.hasSSE41() && "SSE41 required for PINSRB")((Subtarget.hasSSE41() && "SSE41 required for PINSRB" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSE41() && \"SSE41 required for PINSRB\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14429, __PRETTY_FUNCTION__)); |
14430 | Opc = X86ISD::PINSRB; |
14431 | } |
14432 | |
14433 | if (N1.getValueType() != MVT::i32) |
14434 | N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1); |
14435 | if (N2.getValueType() != MVT::i32) |
14436 | N2 = DAG.getIntPtrConstant(IdxVal, dl); |
14437 | return DAG.getNode(Opc, dl, VT, N0, N1, N2); |
14438 | } |
14439 | |
14440 | if (Subtarget.hasSSE41()) { |
14441 | if (EltVT == MVT::f32) { |
14442 | // Bits [7:6] of the constant are the source select. This will always be |
14443 | // zero here. The DAG Combiner may combine an extract_elt index into |
14444 | // these bits. For example (insert (extract, 3), 2) could be matched by |
14445 | // putting the '3' into bits [7:6] of X86ISD::INSERTPS. |
14446 | // Bits [5:4] of the constant are the destination select. This is the |
14447 | // value of the incoming immediate. |
14448 | // Bits [3:0] of the constant are the zero mask. The DAG Combiner may |
14449 | // combine either bitwise AND or insert of float 0.0 to set these bits. |
14450 | |
14451 | bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize(); |
14452 | if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) { |
14453 | // If this is an insertion of 32-bits into the low 32-bits of |
14454 | // a vector, we prefer to generate a blend with immediate rather |
14455 | // than an insertps. Blends are simpler operations in hardware and so |
14456 | // will always have equal or better performance than insertps. |
14457 | // But if optimizing for size and there's a load folding opportunity, |
14458 | // generate insertps because blendps does not have a 32-bit memory |
14459 | // operand form. |
14460 | N2 = DAG.getIntPtrConstant(1, dl); |
14461 | N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1); |
14462 | return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2); |
14463 | } |
14464 | N2 = DAG.getIntPtrConstant(IdxVal << 4, dl); |
14465 | // Create this as a scalar to vector.. |
14466 | N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1); |
14467 | return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2); |
14468 | } |
14469 | |
14470 | // PINSR* works with constant index. |
14471 | if (EltVT == MVT::i32 || EltVT == MVT::i64) |
14472 | return Op; |
14473 | } |
14474 | |
14475 | return SDValue(); |
14476 | } |
14477 | |
14478 | static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, const X86Subtarget &Subtarget, |
14479 | SelectionDAG &DAG) { |
14480 | SDLoc dl(Op); |
14481 | MVT OpVT = Op.getSimpleValueType(); |
14482 | |
14483 | // It's always cheaper to replace a xor+movd with xorps and simplifies further |
14484 | // combines. |
14485 | if (X86::isZeroNode(Op.getOperand(0))) |
14486 | return getZeroVector(OpVT, Subtarget, DAG, dl); |
14487 | |
14488 | // If this is a 256-bit vector result, first insert into a 128-bit |
14489 | // vector and then insert into the 256-bit vector. |
14490 | if (!OpVT.is128BitVector()) { |
14491 | // Insert into a 128-bit vector. |
14492 | unsigned SizeFactor = OpVT.getSizeInBits() / 128; |
14493 | MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(), |
14494 | OpVT.getVectorNumElements() / SizeFactor); |
14495 | |
14496 | Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0)); |
14497 | |
14498 | // Insert the 128-bit vector. |
14499 | return insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl); |
14500 | } |
14501 | assert(OpVT.is128BitVector() && "Expected an SSE type!")((OpVT.is128BitVector() && "Expected an SSE type!") ? static_cast<void> (0) : __assert_fail ("OpVT.is128BitVector() && \"Expected an SSE type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14501, __PRETTY_FUNCTION__)); |
14502 | |
14503 | // Pass through a v4i32 SCALAR_TO_VECTOR as that's what we use in tblgen. |
14504 | if (OpVT == MVT::v4i32) |
14505 | return Op; |
14506 | |
14507 | SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0)); |
14508 | return DAG.getBitcast( |
14509 | OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt)); |
14510 | } |
14511 | |
14512 | // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in |
14513 | // a simple subregister reference or explicit instructions to grab |
14514 | // upper bits of a vector. |
14515 | static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget &Subtarget, |
14516 | SelectionDAG &DAG) { |
14517 | assert(Subtarget.hasAVX() && "EXTRACT_SUBVECTOR requires AVX")((Subtarget.hasAVX() && "EXTRACT_SUBVECTOR requires AVX" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX() && \"EXTRACT_SUBVECTOR requires AVX\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14517, __PRETTY_FUNCTION__)); |
14518 | |
14519 | SDLoc dl(Op); |
14520 | SDValue In = Op.getOperand(0); |
14521 | SDValue Idx = Op.getOperand(1); |
14522 | unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); |
14523 | MVT ResVT = Op.getSimpleValueType(); |
14524 | |
14525 | assert((In.getSimpleValueType().is256BitVector() ||(((In.getSimpleValueType().is256BitVector() || In.getSimpleValueType ().is512BitVector()) && "Can only extract from 256-bit or 512-bit vectors" ) ? static_cast<void> (0) : __assert_fail ("(In.getSimpleValueType().is256BitVector() || In.getSimpleValueType().is512BitVector()) && \"Can only extract from 256-bit or 512-bit vectors\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14527, __PRETTY_FUNCTION__)) |
14526 | In.getSimpleValueType().is512BitVector()) &&(((In.getSimpleValueType().is256BitVector() || In.getSimpleValueType ().is512BitVector()) && "Can only extract from 256-bit or 512-bit vectors" ) ? static_cast<void> (0) : __assert_fail ("(In.getSimpleValueType().is256BitVector() || In.getSimpleValueType().is512BitVector()) && \"Can only extract from 256-bit or 512-bit vectors\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14527, __PRETTY_FUNCTION__)) |
14527 | "Can only extract from 256-bit or 512-bit vectors")(((In.getSimpleValueType().is256BitVector() || In.getSimpleValueType ().is512BitVector()) && "Can only extract from 256-bit or 512-bit vectors" ) ? static_cast<void> (0) : __assert_fail ("(In.getSimpleValueType().is256BitVector() || In.getSimpleValueType().is512BitVector()) && \"Can only extract from 256-bit or 512-bit vectors\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14527, __PRETTY_FUNCTION__)); |
14528 | |
14529 | // If the input is a buildvector just emit a smaller one. |
14530 | unsigned ElemsPerChunk = ResVT.getVectorNumElements(); |
14531 | if (In.getOpcode() == ISD::BUILD_VECTOR) |
14532 | return DAG.getBuildVector( |
14533 | ResVT, dl, makeArrayRef(In->op_begin() + IdxVal, ElemsPerChunk)); |
14534 | |
14535 | // Everything else is legal. |
14536 | return Op; |
14537 | } |
14538 | |
14539 | // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a |
14540 | // simple superregister reference or explicit instructions to insert |
14541 | // the upper bits of a vector. |
14542 | static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget &Subtarget, |
14543 | SelectionDAG &DAG) { |
14544 | assert(Op.getSimpleValueType().getVectorElementType() == MVT::i1)((Op.getSimpleValueType().getVectorElementType() == MVT::i1) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().getVectorElementType() == MVT::i1" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14544, __PRETTY_FUNCTION__)); |
14545 | |
14546 | return insert1BitVector(Op, DAG, Subtarget); |
14547 | } |
14548 | |
14549 | // Returns the appropriate wrapper opcode for a global reference. |
14550 | unsigned X86TargetLowering::getGlobalWrapperKind(const GlobalValue *GV) const { |
14551 | // References to absolute symbols are never PC-relative. |
14552 | if (GV && GV->isAbsoluteSymbolRef()) |
14553 | return X86ISD::Wrapper; |
14554 | |
14555 | CodeModel::Model M = getTargetMachine().getCodeModel(); |
14556 | if (Subtarget.isPICStyleRIPRel() && |
14557 | (M == CodeModel::Small || M == CodeModel::Kernel)) |
14558 | return X86ISD::WrapperRIP; |
14559 | |
14560 | return X86ISD::Wrapper; |
14561 | } |
14562 | |
14563 | // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as |
14564 | // their target counterpart wrapped in the X86ISD::Wrapper node. Suppose N is |
14565 | // one of the above mentioned nodes. It has to be wrapped because otherwise |
14566 | // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only |
14567 | // be used to form addressing mode. These wrapped nodes will be selected |
14568 | // into MOV32ri. |
14569 | SDValue |
14570 | X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const { |
14571 | ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); |
14572 | |
14573 | // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the |
14574 | // global base reg. |
14575 | unsigned char OpFlag = Subtarget.classifyLocalReference(nullptr); |
14576 | |
14577 | auto PtrVT = getPointerTy(DAG.getDataLayout()); |
14578 | SDValue Result = DAG.getTargetConstantPool( |
14579 | CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag); |
14580 | SDLoc DL(CP); |
14581 | Result = DAG.getNode(getGlobalWrapperKind(), DL, PtrVT, Result); |
14582 | // With PIC, the address is actually $g + Offset. |
14583 | if (OpFlag) { |
14584 | Result = |
14585 | DAG.getNode(ISD::ADD, DL, PtrVT, |
14586 | DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result); |
14587 | } |
14588 | |
14589 | return Result; |
14590 | } |
14591 | |
14592 | SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const { |
14593 | JumpTableSDNode *JT = cast<JumpTableSDNode>(Op); |
14594 | |
14595 | // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the |
14596 | // global base reg. |
14597 | unsigned char OpFlag = Subtarget.classifyLocalReference(nullptr); |
14598 | |
14599 | auto PtrVT = getPointerTy(DAG.getDataLayout()); |
14600 | SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag); |
14601 | SDLoc DL(JT); |
14602 | Result = DAG.getNode(getGlobalWrapperKind(), DL, PtrVT, Result); |
14603 | |
14604 | // With PIC, the address is actually $g + Offset. |
14605 | if (OpFlag) |
14606 | Result = |
14607 | DAG.getNode(ISD::ADD, DL, PtrVT, |
14608 | DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result); |
14609 | |
14610 | return Result; |
14611 | } |
14612 | |
14613 | SDValue |
14614 | X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const { |
14615 | const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol(); |
14616 | |
14617 | // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the |
14618 | // global base reg. |
14619 | const Module *Mod = DAG.getMachineFunction().getFunction()->getParent(); |
14620 | unsigned char OpFlag = Subtarget.classifyGlobalReference(nullptr, *Mod); |
14621 | |
14622 | auto PtrVT = getPointerTy(DAG.getDataLayout()); |
14623 | SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag); |
14624 | |
14625 | SDLoc DL(Op); |
14626 | Result = DAG.getNode(getGlobalWrapperKind(), DL, PtrVT, Result); |
14627 | |
14628 | // With PIC, the address is actually $g + Offset. |
14629 | if (isPositionIndependent() && !Subtarget.is64Bit()) { |
14630 | Result = |
14631 | DAG.getNode(ISD::ADD, DL, PtrVT, |
14632 | DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result); |
14633 | } |
14634 | |
14635 | // For symbols that require a load from a stub to get the address, emit the |
14636 | // load. |
14637 | if (isGlobalStubReference(OpFlag)) |
14638 | Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result, |
14639 | MachinePointerInfo::getGOT(DAG.getMachineFunction())); |
14640 | |
14641 | return Result; |
14642 | } |
14643 | |
14644 | SDValue |
14645 | X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const { |
14646 | // Create the TargetBlockAddressAddress node. |
14647 | unsigned char OpFlags = |
14648 | Subtarget.classifyBlockAddressReference(); |
14649 | const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress(); |
14650 | int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset(); |
14651 | SDLoc dl(Op); |
14652 | auto PtrVT = getPointerTy(DAG.getDataLayout()); |
14653 | SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags); |
14654 | Result = DAG.getNode(getGlobalWrapperKind(), dl, PtrVT, Result); |
14655 | |
14656 | // With PIC, the address is actually $g + Offset. |
14657 | if (isGlobalRelativeToPICBase(OpFlags)) { |
14658 | Result = DAG.getNode(ISD::ADD, dl, PtrVT, |
14659 | DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result); |
14660 | } |
14661 | |
14662 | return Result; |
14663 | } |
14664 | |
14665 | SDValue X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, |
14666 | const SDLoc &dl, int64_t Offset, |
14667 | SelectionDAG &DAG) const { |
14668 | // Create the TargetGlobalAddress node, folding in the constant |
14669 | // offset if it is legal. |
14670 | unsigned char OpFlags = Subtarget.classifyGlobalReference(GV); |
14671 | CodeModel::Model M = DAG.getTarget().getCodeModel(); |
14672 | auto PtrVT = getPointerTy(DAG.getDataLayout()); |
14673 | SDValue Result; |
14674 | if (OpFlags == X86II::MO_NO_FLAG && |
14675 | X86::isOffsetSuitableForCodeModel(Offset, M)) { |
14676 | // A direct static reference to a global. |
14677 | Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset); |
14678 | Offset = 0; |
14679 | } else { |
14680 | Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags); |
14681 | } |
14682 | |
14683 | Result = DAG.getNode(getGlobalWrapperKind(GV), dl, PtrVT, Result); |
14684 | |
14685 | // With PIC, the address is actually $g + Offset. |
14686 | if (isGlobalRelativeToPICBase(OpFlags)) { |
14687 | Result = DAG.getNode(ISD::ADD, dl, PtrVT, |
14688 | DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result); |
14689 | } |
14690 | |
14691 | // For globals that require a load from a stub to get the address, emit the |
14692 | // load. |
14693 | if (isGlobalStubReference(OpFlags)) |
14694 | Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result, |
14695 | MachinePointerInfo::getGOT(DAG.getMachineFunction())); |
14696 | |
14697 | // If there was a non-zero offset that we didn't fold, create an explicit |
14698 | // addition for it. |
14699 | if (Offset != 0) |
14700 | Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, |
14701 | DAG.getConstant(Offset, dl, PtrVT)); |
14702 | |
14703 | return Result; |
14704 | } |
14705 | |
14706 | SDValue |
14707 | X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const { |
14708 | const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); |
14709 | int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset(); |
14710 | return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG); |
14711 | } |
14712 | |
14713 | static SDValue |
14714 | GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA, |
14715 | SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg, |
14716 | unsigned char OperandFlags, bool LocalDynamic = false) { |
14717 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
14718 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); |
14719 | SDLoc dl(GA); |
14720 | SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, |
14721 | GA->getValueType(0), |
14722 | GA->getOffset(), |
14723 | OperandFlags); |
14724 | |
14725 | X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR |
14726 | : X86ISD::TLSADDR; |
14727 | |
14728 | if (InFlag) { |
14729 | SDValue Ops[] = { Chain, TGA, *InFlag }; |
14730 | Chain = DAG.getNode(CallType, dl, NodeTys, Ops); |
14731 | } else { |
14732 | SDValue Ops[] = { Chain, TGA }; |
14733 | Chain = DAG.getNode(CallType, dl, NodeTys, Ops); |
14734 | } |
14735 | |
14736 | // TLSADDR will be codegen'ed as call. Inform MFI that function has calls. |
14737 | MFI.setAdjustsStack(true); |
14738 | MFI.setHasCalls(true); |
14739 | |
14740 | SDValue Flag = Chain.getValue(1); |
14741 | return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag); |
14742 | } |
14743 | |
14744 | // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit |
14745 | static SDValue |
14746 | LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG, |
14747 | const EVT PtrVT) { |
14748 | SDValue InFlag; |
14749 | SDLoc dl(GA); // ? function entry point might be better |
14750 | SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX, |
14751 | DAG.getNode(X86ISD::GlobalBaseReg, |
14752 | SDLoc(), PtrVT), InFlag); |
14753 | InFlag = Chain.getValue(1); |
14754 | |
14755 | return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD); |
14756 | } |
14757 | |
14758 | // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit |
14759 | static SDValue |
14760 | LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG, |
14761 | const EVT PtrVT) { |
14762 | return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, |
14763 | X86::RAX, X86II::MO_TLSGD); |
14764 | } |
14765 | |
14766 | static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA, |
14767 | SelectionDAG &DAG, |
14768 | const EVT PtrVT, |
14769 | bool is64Bit) { |
14770 | SDLoc dl(GA); |
14771 | |
14772 | // Get the start address of the TLS block for this module. |
14773 | X86MachineFunctionInfo *MFI = DAG.getMachineFunction() |
14774 | .getInfo<X86MachineFunctionInfo>(); |
14775 | MFI->incNumLocalDynamicTLSAccesses(); |
14776 | |
14777 | SDValue Base; |
14778 | if (is64Bit) { |
14779 | Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX, |
14780 | X86II::MO_TLSLD, /*LocalDynamic=*/true); |
14781 | } else { |
14782 | SDValue InFlag; |
14783 | SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX, |
14784 | DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag); |
14785 | InFlag = Chain.getValue(1); |
14786 | Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, |
14787 | X86II::MO_TLSLDM, /*LocalDynamic=*/true); |
14788 | } |
14789 | |
14790 | // Note: the CleanupLocalDynamicTLSPass will remove redundant computations |
14791 | // of Base. |
14792 | |
14793 | // Build x@dtpoff. |
14794 | unsigned char OperandFlags = X86II::MO_DTPOFF; |
14795 | unsigned WrapperKind = X86ISD::Wrapper; |
14796 | SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, |
14797 | GA->getValueType(0), |
14798 | GA->getOffset(), OperandFlags); |
14799 | SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA); |
14800 | |
14801 | // Add x@dtpoff with the base. |
14802 | return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base); |
14803 | } |
14804 | |
14805 | // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model. |
14806 | static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG, |
14807 | const EVT PtrVT, TLSModel::Model model, |
14808 | bool is64Bit, bool isPIC) { |
14809 | SDLoc dl(GA); |
14810 | |
14811 | // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit). |
14812 | Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(), |
14813 | is64Bit ? 257 : 256)); |
14814 | |
14815 | SDValue ThreadPointer = |
14816 | DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl), |
14817 | MachinePointerInfo(Ptr)); |
14818 | |
14819 | unsigned char OperandFlags = 0; |
14820 | // Most TLS accesses are not RIP relative, even on x86-64. One exception is |
14821 | // initialexec. |
14822 | unsigned WrapperKind = X86ISD::Wrapper; |
14823 | if (model == TLSModel::LocalExec) { |
14824 | OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF; |
14825 | } else if (model == TLSModel::InitialExec) { |
14826 | if (is64Bit) { |
14827 | OperandFlags = X86II::MO_GOTTPOFF; |
14828 | WrapperKind = X86ISD::WrapperRIP; |
14829 | } else { |
14830 | OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF; |
14831 | } |
14832 | } else { |
14833 | llvm_unreachable("Unexpected model")::llvm::llvm_unreachable_internal("Unexpected model", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14833); |
14834 | } |
14835 | |
14836 | // emit "addl x@ntpoff,%eax" (local exec) |
14837 | // or "addl x@indntpoff,%eax" (initial exec) |
14838 | // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic) |
14839 | SDValue TGA = |
14840 | DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0), |
14841 | GA->getOffset(), OperandFlags); |
14842 | SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA); |
14843 | |
14844 | if (model == TLSModel::InitialExec) { |
14845 | if (isPIC && !is64Bit) { |
14846 | Offset = DAG.getNode(ISD::ADD, dl, PtrVT, |
14847 | DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), |
14848 | Offset); |
14849 | } |
14850 | |
14851 | Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset, |
14852 | MachinePointerInfo::getGOT(DAG.getMachineFunction())); |
14853 | } |
14854 | |
14855 | // The address of the thread local variable is the add of the thread |
14856 | // pointer with the offset of the variable. |
14857 | return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset); |
14858 | } |
14859 | |
14860 | SDValue |
14861 | X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const { |
14862 | |
14863 | GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); |
14864 | |
14865 | if (DAG.getTarget().Options.EmulatedTLS) |
14866 | return LowerToTLSEmulatedModel(GA, DAG); |
14867 | |
14868 | const GlobalValue *GV = GA->getGlobal(); |
14869 | auto PtrVT = getPointerTy(DAG.getDataLayout()); |
14870 | bool PositionIndependent = isPositionIndependent(); |
14871 | |
14872 | if (Subtarget.isTargetELF()) { |
14873 | TLSModel::Model model = DAG.getTarget().getTLSModel(GV); |
14874 | switch (model) { |
14875 | case TLSModel::GeneralDynamic: |
14876 | if (Subtarget.is64Bit()) |
14877 | return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT); |
14878 | return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT); |
14879 | case TLSModel::LocalDynamic: |
14880 | return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT, |
14881 | Subtarget.is64Bit()); |
14882 | case TLSModel::InitialExec: |
14883 | case TLSModel::LocalExec: |
14884 | return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget.is64Bit(), |
14885 | PositionIndependent); |
14886 | } |
14887 | llvm_unreachable("Unknown TLS model.")::llvm::llvm_unreachable_internal("Unknown TLS model.", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 14887); |
14888 | } |
14889 | |
14890 | if (Subtarget.isTargetDarwin()) { |
14891 | // Darwin only has one model of TLS. Lower to that. |
14892 | unsigned char OpFlag = 0; |
14893 | unsigned WrapperKind = Subtarget.isPICStyleRIPRel() ? |
14894 | X86ISD::WrapperRIP : X86ISD::Wrapper; |
14895 | |
14896 | // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the |
14897 | // global base reg. |
14898 | bool PIC32 = PositionIndependent && !Subtarget.is64Bit(); |
14899 | if (PIC32) |
14900 | OpFlag = X86II::MO_TLVP_PIC_BASE; |
14901 | else |
14902 | OpFlag = X86II::MO_TLVP; |
14903 | SDLoc DL(Op); |
14904 | SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL, |
14905 | GA->getValueType(0), |
14906 | GA->getOffset(), OpFlag); |
14907 | SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result); |
14908 | |
14909 | // With PIC32, the address is actually $g + Offset. |
14910 | if (PIC32) |
14911 | Offset = DAG.getNode(ISD::ADD, DL, PtrVT, |
14912 | DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), |
14913 | Offset); |
14914 | |
14915 | // Lowering the machine isd will make sure everything is in the right |
14916 | // location. |
14917 | SDValue Chain = DAG.getEntryNode(); |
14918 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); |
14919 | Chain = DAG.getCALLSEQ_START(Chain, 0, 0, DL); |
14920 | SDValue Args[] = { Chain, Offset }; |
14921 | Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args); |
14922 | Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, DL, true), |
14923 | DAG.getIntPtrConstant(0, DL, true), |
14924 | Chain.getValue(1), DL); |
14925 | |
14926 | // TLSCALL will be codegen'ed as call. Inform MFI that function has calls. |
14927 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
14928 | MFI.setAdjustsStack(true); |
14929 | |
14930 | // And our return value (tls address) is in the standard call return value |
14931 | // location. |
14932 | unsigned Reg = Subtarget.is64Bit() ? X86::RAX : X86::EAX; |
14933 | return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1)); |
14934 | } |
14935 | |
14936 | if (Subtarget.isTargetKnownWindowsMSVC() || |
14937 | Subtarget.isTargetWindowsItanium() || |
14938 | Subtarget.isTargetWindowsGNU()) { |
14939 | // Just use the implicit TLS architecture |
14940 | // Need to generate something similar to: |
14941 | // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage |
14942 | // ; from TEB |
14943 | // mov ecx, dword [rel _tls_index]: Load index (from C runtime) |
14944 | // mov rcx, qword [rdx+rcx*8] |
14945 | // mov eax, .tls$:tlsvar |
14946 | // [rax+rcx] contains the address |
14947 | // Windows 64bit: gs:0x58 |
14948 | // Windows 32bit: fs:__tls_array |
14949 | |
14950 | SDLoc dl(GA); |
14951 | SDValue Chain = DAG.getEntryNode(); |
14952 | |
14953 | // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or |
14954 | // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly |
14955 | // use its literal value of 0x2C. |
14956 | Value *Ptr = Constant::getNullValue(Subtarget.is64Bit() |
14957 | ? Type::getInt8PtrTy(*DAG.getContext(), |
14958 | 256) |
14959 | : Type::getInt32PtrTy(*DAG.getContext(), |
14960 | 257)); |
14961 | |
14962 | SDValue TlsArray = Subtarget.is64Bit() |
14963 | ? DAG.getIntPtrConstant(0x58, dl) |
14964 | : (Subtarget.isTargetWindowsGNU() |
14965 | ? DAG.getIntPtrConstant(0x2C, dl) |
14966 | : DAG.getExternalSymbol("_tls_array", PtrVT)); |
14967 | |
14968 | SDValue ThreadPointer = |
14969 | DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr)); |
14970 | |
14971 | SDValue res; |
14972 | if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) { |
14973 | res = ThreadPointer; |
14974 | } else { |
14975 | // Load the _tls_index variable |
14976 | SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT); |
14977 | if (Subtarget.is64Bit()) |
14978 | IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX, |
14979 | MachinePointerInfo(), MVT::i32); |
14980 | else |
14981 | IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo()); |
14982 | |
14983 | auto &DL = DAG.getDataLayout(); |
14984 | SDValue Scale = |
14985 | DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT); |
14986 | IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale); |
14987 | |
14988 | res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX); |
14989 | } |
14990 | |
14991 | res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo()); |
14992 | |
14993 | // Get the offset of start of .tls section |
14994 | SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, |
14995 | GA->getValueType(0), |
14996 | GA->getOffset(), X86II::MO_SECREL); |
14997 | SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA); |
14998 | |
14999 | // The address of the thread local variable is the add of the thread |
15000 | // pointer with the offset of the variable. |
15001 | return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset); |
15002 | } |
15003 | |
15004 | llvm_unreachable("TLS not implemented for this target.")::llvm::llvm_unreachable_internal("TLS not implemented for this target." , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15004); |
15005 | } |
15006 | |
15007 | /// Lower SRA_PARTS and friends, which return two i32 values |
15008 | /// and take a 2 x i32 value to shift plus a shift amount. |
15009 | static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) { |
15010 | assert(Op.getNumOperands() == 3 && "Not a double-shift!")((Op.getNumOperands() == 3 && "Not a double-shift!") ? static_cast<void> (0) : __assert_fail ("Op.getNumOperands() == 3 && \"Not a double-shift!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15010, __PRETTY_FUNCTION__)); |
15011 | MVT VT = Op.getSimpleValueType(); |
15012 | unsigned VTBits = VT.getSizeInBits(); |
15013 | SDLoc dl(Op); |
15014 | bool isSRA = Op.getOpcode() == ISD::SRA_PARTS; |
15015 | SDValue ShOpLo = Op.getOperand(0); |
15016 | SDValue ShOpHi = Op.getOperand(1); |
15017 | SDValue ShAmt = Op.getOperand(2); |
15018 | // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the |
15019 | // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away |
15020 | // during isel. |
15021 | SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt, |
15022 | DAG.getConstant(VTBits - 1, dl, MVT::i8)); |
15023 | SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi, |
15024 | DAG.getConstant(VTBits - 1, dl, MVT::i8)) |
15025 | : DAG.getConstant(0, dl, VT); |
15026 | |
15027 | SDValue Tmp2, Tmp3; |
15028 | if (Op.getOpcode() == ISD::SHL_PARTS) { |
15029 | Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt); |
15030 | Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt); |
15031 | } else { |
15032 | Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt); |
15033 | Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt); |
15034 | } |
15035 | |
15036 | // If the shift amount is larger or equal than the width of a part we can't |
15037 | // rely on the results of shld/shrd. Insert a test and select the appropriate |
15038 | // values for large shift amounts. |
15039 | SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt, |
15040 | DAG.getConstant(VTBits, dl, MVT::i8)); |
15041 | SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32, |
15042 | AndNode, DAG.getConstant(0, dl, MVT::i8)); |
15043 | |
15044 | SDValue Hi, Lo; |
15045 | SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8); |
15046 | SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond }; |
15047 | SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond }; |
15048 | |
15049 | if (Op.getOpcode() == ISD::SHL_PARTS) { |
15050 | Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0); |
15051 | Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1); |
15052 | } else { |
15053 | Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0); |
15054 | Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1); |
15055 | } |
15056 | |
15057 | SDValue Ops[2] = { Lo, Hi }; |
15058 | return DAG.getMergeValues(Ops, dl); |
15059 | } |
15060 | |
15061 | SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, |
15062 | SelectionDAG &DAG) const { |
15063 | SDValue Src = Op.getOperand(0); |
15064 | MVT SrcVT = Src.getSimpleValueType(); |
15065 | MVT VT = Op.getSimpleValueType(); |
15066 | SDLoc dl(Op); |
15067 | |
15068 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
15069 | if (SrcVT.isVector()) { |
15070 | if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) { |
15071 | return DAG.getNode(X86ISD::CVTSI2P, dl, VT, |
15072 | DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src, |
15073 | DAG.getUNDEF(SrcVT))); |
15074 | } |
15075 | if (SrcVT.getVectorElementType() == MVT::i1) { |
15076 | if (SrcVT == MVT::v2i1 && TLI.isTypeLegal(SrcVT)) |
15077 | return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), |
15078 | DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v2i64, Src)); |
15079 | MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements()); |
15080 | return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), |
15081 | DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src)); |
15082 | } |
15083 | return SDValue(); |
15084 | } |
15085 | |
15086 | assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&((SrcVT <= MVT::i64 && SrcVT >= MVT::i16 && "Unknown SINT_TO_FP to lower!") ? static_cast<void> (0 ) : __assert_fail ("SrcVT <= MVT::i64 && SrcVT >= MVT::i16 && \"Unknown SINT_TO_FP to lower!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15087, __PRETTY_FUNCTION__)) |
15087 | "Unknown SINT_TO_FP to lower!")((SrcVT <= MVT::i64 && SrcVT >= MVT::i16 && "Unknown SINT_TO_FP to lower!") ? static_cast<void> (0 ) : __assert_fail ("SrcVT <= MVT::i64 && SrcVT >= MVT::i16 && \"Unknown SINT_TO_FP to lower!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15087, __PRETTY_FUNCTION__)); |
15088 | |
15089 | // These are really Legal; return the operand so the caller accepts it as |
15090 | // Legal. |
15091 | if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType())) |
15092 | return Op; |
15093 | if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) && |
15094 | Subtarget.is64Bit()) { |
15095 | return Op; |
15096 | } |
15097 | |
15098 | SDValue ValueToStore = Op.getOperand(0); |
15099 | if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) && |
15100 | !Subtarget.is64Bit()) |
15101 | // Bitcasting to f64 here allows us to do a single 64-bit store from |
15102 | // an SSE register, avoiding the store forwarding penalty that would come |
15103 | // with two 32-bit stores. |
15104 | ValueToStore = DAG.getBitcast(MVT::f64, ValueToStore); |
15105 | |
15106 | unsigned Size = SrcVT.getSizeInBits()/8; |
15107 | MachineFunction &MF = DAG.getMachineFunction(); |
15108 | auto PtrVT = getPointerTy(MF.getDataLayout()); |
15109 | int SSFI = MF.getFrameInfo().CreateStackObject(Size, Size, false); |
15110 | SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT); |
15111 | SDValue Chain = DAG.getStore( |
15112 | DAG.getEntryNode(), dl, ValueToStore, StackSlot, |
15113 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI)); |
15114 | return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG); |
15115 | } |
15116 | |
15117 | SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, |
15118 | SDValue StackSlot, |
15119 | SelectionDAG &DAG) const { |
15120 | // Build the FILD |
15121 | SDLoc DL(Op); |
15122 | SDVTList Tys; |
15123 | bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType()); |
15124 | if (useSSE) |
15125 | Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue); |
15126 | else |
15127 | Tys = DAG.getVTList(Op.getValueType(), MVT::Other); |
15128 | |
15129 | unsigned ByteSize = SrcVT.getSizeInBits()/8; |
15130 | |
15131 | FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot); |
15132 | MachineMemOperand *MMO; |
15133 | if (FI) { |
15134 | int SSFI = FI->getIndex(); |
15135 | MMO = DAG.getMachineFunction().getMachineMemOperand( |
15136 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), |
15137 | MachineMemOperand::MOLoad, ByteSize, ByteSize); |
15138 | } else { |
15139 | MMO = cast<LoadSDNode>(StackSlot)->getMemOperand(); |
15140 | StackSlot = StackSlot.getOperand(1); |
15141 | } |
15142 | SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) }; |
15143 | SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG : |
15144 | X86ISD::FILD, DL, |
15145 | Tys, Ops, SrcVT, MMO); |
15146 | |
15147 | if (useSSE) { |
15148 | Chain = Result.getValue(1); |
15149 | SDValue InFlag = Result.getValue(2); |
15150 | |
15151 | // FIXME: Currently the FST is flagged to the FILD_FLAG. This |
15152 | // shouldn't be necessary except that RFP cannot be live across |
15153 | // multiple blocks. When stackifier is fixed, they can be uncoupled. |
15154 | MachineFunction &MF = DAG.getMachineFunction(); |
15155 | unsigned SSFISize = Op.getValueSizeInBits()/8; |
15156 | int SSFI = MF.getFrameInfo().CreateStackObject(SSFISize, SSFISize, false); |
15157 | auto PtrVT = getPointerTy(MF.getDataLayout()); |
15158 | SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT); |
15159 | Tys = DAG.getVTList(MVT::Other); |
15160 | SDValue Ops[] = { |
15161 | Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag |
15162 | }; |
15163 | MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand( |
15164 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), |
15165 | MachineMemOperand::MOStore, SSFISize, SSFISize); |
15166 | |
15167 | Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys, |
15168 | Ops, Op.getValueType(), MMO); |
15169 | Result = DAG.getLoad( |
15170 | Op.getValueType(), DL, Chain, StackSlot, |
15171 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI)); |
15172 | } |
15173 | |
15174 | return Result; |
15175 | } |
15176 | |
15177 | /// 64-bit unsigned integer to double expansion. |
15178 | SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, |
15179 | SelectionDAG &DAG) const { |
15180 | // This algorithm is not obvious. Here it is what we're trying to output: |
15181 | /* |
15182 | movq %rax, %xmm0 |
15183 | punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U } |
15184 | subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 } |
15185 | #ifdef __SSE3__ |
15186 | haddpd %xmm0, %xmm0 |
15187 | #else |
15188 | pshufd $0x4e, %xmm0, %xmm1 |
15189 | addpd %xmm1, %xmm0 |
15190 | #endif |
15191 | */ |
15192 | |
15193 | SDLoc dl(Op); |
15194 | LLVMContext *Context = DAG.getContext(); |
15195 | |
15196 | // Build some magic constants. |
15197 | static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 }; |
15198 | Constant *C0 = ConstantDataVector::get(*Context, CV0); |
15199 | auto PtrVT = getPointerTy(DAG.getDataLayout()); |
15200 | SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16); |
15201 | |
15202 | SmallVector<Constant*,2> CV1; |
15203 | CV1.push_back( |
15204 | ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble(), |
15205 | APInt(64, 0x4330000000000000ULL)))); |
15206 | CV1.push_back( |
15207 | ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble(), |
15208 | APInt(64, 0x4530000000000000ULL)))); |
15209 | Constant *C1 = ConstantVector::get(CV1); |
15210 | SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16); |
15211 | |
15212 | // Load the 64-bit value into an XMM register. |
15213 | SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, |
15214 | Op.getOperand(0)); |
15215 | SDValue CLod0 = |
15216 | DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0, |
15217 | MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), |
15218 | /* Alignment = */ 16); |
15219 | SDValue Unpck1 = |
15220 | getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0); |
15221 | |
15222 | SDValue CLod1 = |
15223 | DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1, |
15224 | MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), |
15225 | /* Alignment = */ 16); |
15226 | SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1); |
15227 | // TODO: Are there any fast-math-flags to propagate here? |
15228 | SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1); |
15229 | SDValue Result; |
15230 | |
15231 | if (Subtarget.hasSSE3()) { |
15232 | // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'. |
15233 | Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub); |
15234 | } else { |
15235 | SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub); |
15236 | SDValue Shuffle = DAG.getVectorShuffle(MVT::v4i32, dl, S2F, S2F, {2,3,0,1}); |
15237 | Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64, |
15238 | DAG.getBitcast(MVT::v2f64, Shuffle), Sub); |
15239 | } |
15240 | |
15241 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result, |
15242 | DAG.getIntPtrConstant(0, dl)); |
15243 | } |
15244 | |
15245 | /// 32-bit unsigned integer to float expansion. |
15246 | SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, |
15247 | SelectionDAG &DAG) const { |
15248 | SDLoc dl(Op); |
15249 | // FP constant to bias correct the final result. |
15250 | SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl, |
15251 | MVT::f64); |
15252 | |
15253 | // Load the 32-bit value into an XMM register. |
15254 | SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, |
15255 | Op.getOperand(0)); |
15256 | |
15257 | // Zero out the upper parts of the register. |
15258 | Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG); |
15259 | |
15260 | Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, |
15261 | DAG.getBitcast(MVT::v2f64, Load), |
15262 | DAG.getIntPtrConstant(0, dl)); |
15263 | |
15264 | // Or the load with the bias. |
15265 | SDValue Or = DAG.getNode( |
15266 | ISD::OR, dl, MVT::v2i64, |
15267 | DAG.getBitcast(MVT::v2i64, |
15268 | DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)), |
15269 | DAG.getBitcast(MVT::v2i64, |
15270 | DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias))); |
15271 | Or = |
15272 | DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, |
15273 | DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl)); |
15274 | |
15275 | // Subtract the bias. |
15276 | // TODO: Are there any fast-math-flags to propagate here? |
15277 | SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias); |
15278 | |
15279 | // Handle final rounding. |
15280 | MVT DestVT = Op.getSimpleValueType(); |
15281 | |
15282 | if (DestVT.bitsLT(MVT::f64)) |
15283 | return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub, |
15284 | DAG.getIntPtrConstant(0, dl)); |
15285 | if (DestVT.bitsGT(MVT::f64)) |
15286 | return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub); |
15287 | |
15288 | // Handle final rounding. |
15289 | return Sub; |
15290 | } |
15291 | |
15292 | static SDValue lowerUINT_TO_FP_v2i32(SDValue Op, SelectionDAG &DAG, |
15293 | const X86Subtarget &Subtarget, SDLoc &DL) { |
15294 | if (Op.getSimpleValueType() != MVT::v2f64) |
15295 | return SDValue(); |
15296 | |
15297 | SDValue N0 = Op.getOperand(0); |
15298 | assert(N0.getSimpleValueType() == MVT::v2i32 && "Unexpected input type")((N0.getSimpleValueType() == MVT::v2i32 && "Unexpected input type" ) ? static_cast<void> (0) : __assert_fail ("N0.getSimpleValueType() == MVT::v2i32 && \"Unexpected input type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15298, __PRETTY_FUNCTION__)); |
15299 | |
15300 | // Legalize to v4i32 type. |
15301 | N0 = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4i32, N0, |
15302 | DAG.getUNDEF(MVT::v2i32)); |
15303 | |
15304 | if (Subtarget.hasAVX512()) |
15305 | return DAG.getNode(X86ISD::CVTUI2P, DL, MVT::v2f64, N0); |
15306 | |
15307 | // Same implementation as VectorLegalizer::ExpandUINT_TO_FLOAT, |
15308 | // but using v2i32 to v2f64 with X86ISD::CVTSI2P. |
15309 | SDValue HalfWord = DAG.getConstant(16, DL, MVT::v4i32); |
15310 | SDValue HalfWordMask = DAG.getConstant(0x0000FFFF, DL, MVT::v4i32); |
15311 | |
15312 | // Two to the power of half-word-size. |
15313 | SDValue TWOHW = DAG.getConstantFP(1 << 16, DL, MVT::v2f64); |
15314 | |
15315 | // Clear upper part of LO, lower HI. |
15316 | SDValue HI = DAG.getNode(ISD::SRL, DL, MVT::v4i32, N0, HalfWord); |
15317 | SDValue LO = DAG.getNode(ISD::AND, DL, MVT::v4i32, N0, HalfWordMask); |
15318 | |
15319 | SDValue fHI = DAG.getNode(X86ISD::CVTSI2P, DL, MVT::v2f64, HI); |
15320 | fHI = DAG.getNode(ISD::FMUL, DL, MVT::v2f64, fHI, TWOHW); |
15321 | SDValue fLO = DAG.getNode(X86ISD::CVTSI2P, DL, MVT::v2f64, LO); |
15322 | |
15323 | // Add the two halves. |
15324 | return DAG.getNode(ISD::FADD, DL, MVT::v2f64, fHI, fLO); |
15325 | } |
15326 | |
15327 | static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG, |
15328 | const X86Subtarget &Subtarget) { |
15329 | // The algorithm is the following: |
15330 | // #ifdef __SSE4_1__ |
15331 | // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa); |
15332 | // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16), |
15333 | // (uint4) 0x53000000, 0xaa); |
15334 | // #else |
15335 | // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000; |
15336 | // uint4 hi = (v >> 16) | (uint4) 0x53000000; |
15337 | // #endif |
15338 | // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f); |
15339 | // return (float4) lo + fhi; |
15340 | |
15341 | // We shouldn't use it when unsafe-fp-math is enabled though: we might later |
15342 | // reassociate the two FADDs, and if we do that, the algorithm fails |
15343 | // spectacularly (PR24512). |
15344 | // FIXME: If we ever have some kind of Machine FMF, this should be marked |
15345 | // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because |
15346 | // there's also the MachineCombiner reassociations happening on Machine IR. |
15347 | if (DAG.getTarget().Options.UnsafeFPMath) |
15348 | return SDValue(); |
15349 | |
15350 | SDLoc DL(Op); |
15351 | SDValue V = Op->getOperand(0); |
15352 | MVT VecIntVT = V.getSimpleValueType(); |
15353 | bool Is128 = VecIntVT == MVT::v4i32; |
15354 | MVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32; |
15355 | // If we convert to something else than the supported type, e.g., to v4f64, |
15356 | // abort early. |
15357 | if (VecFloatVT != Op->getSimpleValueType(0)) |
15358 | return SDValue(); |
15359 | |
15360 | assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&(((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) && "Unsupported custom type") ? static_cast<void> (0) : __assert_fail ("(VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) && \"Unsupported custom type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15361, __PRETTY_FUNCTION__)) |
15361 | "Unsupported custom type")(((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) && "Unsupported custom type") ? static_cast<void> (0) : __assert_fail ("(VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) && \"Unsupported custom type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15361, __PRETTY_FUNCTION__)); |
15362 | |
15363 | // In the #idef/#else code, we have in common: |
15364 | // - The vector of constants: |
15365 | // -- 0x4b000000 |
15366 | // -- 0x53000000 |
15367 | // - A shift: |
15368 | // -- v >> 16 |
15369 | |
15370 | // Create the splat vector for 0x4b000000. |
15371 | SDValue VecCstLow = DAG.getConstant(0x4b000000, DL, VecIntVT); |
15372 | // Create the splat vector for 0x53000000. |
15373 | SDValue VecCstHigh = DAG.getConstant(0x53000000, DL, VecIntVT); |
15374 | |
15375 | // Create the right shift. |
15376 | SDValue VecCstShift = DAG.getConstant(16, DL, VecIntVT); |
15377 | SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift); |
15378 | |
15379 | SDValue Low, High; |
15380 | if (Subtarget.hasSSE41()) { |
15381 | MVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16; |
15382 | // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa); |
15383 | SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow); |
15384 | SDValue VecBitcast = DAG.getBitcast(VecI16VT, V); |
15385 | // Low will be bitcasted right away, so do not bother bitcasting back to its |
15386 | // original type. |
15387 | Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast, |
15388 | VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32)); |
15389 | // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16), |
15390 | // (uint4) 0x53000000, 0xaa); |
15391 | SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh); |
15392 | SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift); |
15393 | // High will be bitcasted right away, so do not bother bitcasting back to |
15394 | // its original type. |
15395 | High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast, |
15396 | VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32)); |
15397 | } else { |
15398 | SDValue VecCstMask = DAG.getConstant(0xffff, DL, VecIntVT); |
15399 | // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000; |
15400 | SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask); |
15401 | Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow); |
15402 | |
15403 | // uint4 hi = (v >> 16) | (uint4) 0x53000000; |
15404 | High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh); |
15405 | } |
15406 | |
15407 | // Create the vector constant for -(0x1.0p39f + 0x1.0p23f). |
15408 | SDValue VecCstFAdd = DAG.getConstantFP( |
15409 | APFloat(APFloat::IEEEsingle(), APInt(32, 0xD3000080)), DL, VecFloatVT); |
15410 | |
15411 | // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f); |
15412 | SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High); |
15413 | // TODO: Are there any fast-math-flags to propagate here? |
15414 | SDValue FHigh = |
15415 | DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd); |
15416 | // return (float4) lo + fhi; |
15417 | SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low); |
15418 | return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh); |
15419 | } |
15420 | |
15421 | SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op, |
15422 | SelectionDAG &DAG) const { |
15423 | SDValue N0 = Op.getOperand(0); |
15424 | MVT SrcVT = N0.getSimpleValueType(); |
15425 | SDLoc dl(Op); |
15426 | |
15427 | if (SrcVT.getVectorElementType() == MVT::i1) { |
15428 | if (SrcVT == MVT::v2i1) |
15429 | return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(), |
15430 | DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64, N0)); |
15431 | MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements()); |
15432 | return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(), |
15433 | DAG.getNode(ISD::ZERO_EXTEND, dl, IntegerVT, N0)); |
15434 | } |
15435 | |
15436 | switch (SrcVT.SimpleTy) { |
15437 | default: |
15438 | llvm_unreachable("Custom UINT_TO_FP is not supported!")::llvm::llvm_unreachable_internal("Custom UINT_TO_FP is not supported!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15438); |
15439 | case MVT::v4i8: |
15440 | case MVT::v4i16: |
15441 | case MVT::v8i8: |
15442 | case MVT::v8i16: { |
15443 | MVT NVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements()); |
15444 | return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), |
15445 | DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0)); |
15446 | } |
15447 | case MVT::v2i32: |
15448 | return lowerUINT_TO_FP_v2i32(Op, DAG, Subtarget, dl); |
15449 | case MVT::v4i32: |
15450 | case MVT::v8i32: |
15451 | return lowerUINT_TO_FP_vXi32(Op, DAG, Subtarget); |
15452 | case MVT::v16i8: |
15453 | case MVT::v16i16: |
15454 | assert(Subtarget.hasAVX512())((Subtarget.hasAVX512()) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX512()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15454, __PRETTY_FUNCTION__)); |
15455 | return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(), |
15456 | DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0)); |
15457 | } |
15458 | } |
15459 | |
15460 | SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, |
15461 | SelectionDAG &DAG) const { |
15462 | SDValue N0 = Op.getOperand(0); |
15463 | SDLoc dl(Op); |
15464 | auto PtrVT = getPointerTy(DAG.getDataLayout()); |
15465 | |
15466 | // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't |
15467 | // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform |
15468 | // the optimization here. |
15469 | if (DAG.SignBitIsZero(N0)) |
15470 | return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0); |
15471 | |
15472 | if (Op.getSimpleValueType().isVector()) |
15473 | return lowerUINT_TO_FP_vec(Op, DAG); |
15474 | |
15475 | MVT SrcVT = N0.getSimpleValueType(); |
15476 | MVT DstVT = Op.getSimpleValueType(); |
15477 | |
15478 | if (Subtarget.hasAVX512() && isScalarFPTypeInSSEReg(DstVT) && |
15479 | (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget.is64Bit()))) { |
15480 | // Conversions from unsigned i32 to f32/f64 are legal, |
15481 | // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode. |
15482 | return Op; |
15483 | } |
15484 | |
15485 | if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64) |
15486 | return LowerUINT_TO_FP_i64(Op, DAG); |
15487 | if (SrcVT == MVT::i32 && X86ScalarSSEf64) |
15488 | return LowerUINT_TO_FP_i32(Op, DAG); |
15489 | if (Subtarget.is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32) |
15490 | return SDValue(); |
15491 | |
15492 | // Make a 64-bit buffer, and use it to build an FILD. |
15493 | SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64); |
15494 | if (SrcVT == MVT::i32) { |
15495 | SDValue OffsetSlot = DAG.getMemBasePlusOffset(StackSlot, 4, dl); |
15496 | SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0), |
15497 | StackSlot, MachinePointerInfo()); |
15498 | SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32), |
15499 | OffsetSlot, MachinePointerInfo()); |
15500 | SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG); |
15501 | return Fild; |
15502 | } |
15503 | |
15504 | assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP")((SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP" ) ? static_cast<void> (0) : __assert_fail ("SrcVT == MVT::i64 && \"Unexpected type in UINT_TO_FP\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15504, __PRETTY_FUNCTION__)); |
15505 | SDValue ValueToStore = Op.getOperand(0); |
15506 | if (isScalarFPTypeInSSEReg(Op.getValueType()) && !Subtarget.is64Bit()) |
15507 | // Bitcasting to f64 here allows us to do a single 64-bit store from |
15508 | // an SSE register, avoiding the store forwarding penalty that would come |
15509 | // with two 32-bit stores. |
15510 | ValueToStore = DAG.getBitcast(MVT::f64, ValueToStore); |
15511 | SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, ValueToStore, StackSlot, |
15512 | MachinePointerInfo()); |
15513 | // For i64 source, we need to add the appropriate power of 2 if the input |
15514 | // was negative. This is the same as the optimization in |
15515 | // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here, |
15516 | // we must be careful to do the computation in x87 extended precision, not |
15517 | // in SSE. (The generic code can't know it's OK to do this, or how to.) |
15518 | int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex(); |
15519 | MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand( |
15520 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), |
15521 | MachineMemOperand::MOLoad, 8, 8); |
15522 | |
15523 | SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other); |
15524 | SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) }; |
15525 | SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, |
15526 | MVT::i64, MMO); |
15527 | |
15528 | APInt FF(32, 0x5F800000ULL); |
15529 | |
15530 | // Check whether the sign bit is set. |
15531 | SDValue SignSet = DAG.getSetCC( |
15532 | dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64), |
15533 | Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT); |
15534 | |
15535 | // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits. |
15536 | SDValue FudgePtr = DAG.getConstantPool( |
15537 | ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT); |
15538 | |
15539 | // Get a pointer to FF if the sign bit was set, or to 0 otherwise. |
15540 | SDValue Zero = DAG.getIntPtrConstant(0, dl); |
15541 | SDValue Four = DAG.getIntPtrConstant(4, dl); |
15542 | SDValue Offset = DAG.getSelect(dl, Zero.getValueType(), SignSet, Zero, Four); |
15543 | FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset); |
15544 | |
15545 | // Load the value out, extending it from f32 to f80. |
15546 | // FIXME: Avoid the extend by constructing the right constant pool? |
15547 | SDValue Fudge = DAG.getExtLoad( |
15548 | ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr, |
15549 | MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32, |
15550 | /* Alignment = */ 4); |
15551 | // Extend everything to 80 bits to force it to be done on x87. |
15552 | // TODO: Are there any fast-math-flags to propagate here? |
15553 | SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge); |
15554 | return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, |
15555 | DAG.getIntPtrConstant(0, dl)); |
15556 | } |
15557 | |
15558 | // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation |
15559 | // is legal, or has an fp128 or f16 source (which needs to be promoted to f32), |
15560 | // just return an <SDValue(), SDValue()> pair. |
15561 | // Otherwise it is assumed to be a conversion from one of f32, f64 or f80 |
15562 | // to i16, i32 or i64, and we lower it to a legal sequence. |
15563 | // If lowered to the final integer result we return a <result, SDValue()> pair. |
15564 | // Otherwise we lower it to a sequence ending with a FIST, return a |
15565 | // <FIST, StackSlot> pair, and the caller is responsible for loading |
15566 | // the final integer result from StackSlot. |
15567 | std::pair<SDValue,SDValue> |
15568 | X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, |
15569 | bool IsSigned, bool IsReplace) const { |
15570 | SDLoc DL(Op); |
15571 | |
15572 | EVT DstTy = Op.getValueType(); |
15573 | EVT TheVT = Op.getOperand(0).getValueType(); |
15574 | auto PtrVT = getPointerTy(DAG.getDataLayout()); |
15575 | |
15576 | if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) { |
15577 | // f16 must be promoted before using the lowering in this routine. |
15578 | // fp128 does not use this lowering. |
15579 | return std::make_pair(SDValue(), SDValue()); |
15580 | } |
15581 | |
15582 | // If using FIST to compute an unsigned i64, we'll need some fixup |
15583 | // to handle values above the maximum signed i64. A FIST is always |
15584 | // used for the 32-bit subtarget, but also for f80 on a 64-bit target. |
15585 | bool UnsignedFixup = !IsSigned && |
15586 | DstTy == MVT::i64 && |
15587 | (!Subtarget.is64Bit() || |
15588 | !isScalarFPTypeInSSEReg(TheVT)); |
15589 | |
15590 | if (!IsSigned && DstTy != MVT::i64 && !Subtarget.hasAVX512()) { |
15591 | // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST. |
15592 | // The low 32 bits of the fist result will have the correct uint32 result. |
15593 | assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT")((DstTy == MVT::i32 && "Unexpected FP_TO_UINT") ? static_cast <void> (0) : __assert_fail ("DstTy == MVT::i32 && \"Unexpected FP_TO_UINT\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15593, __PRETTY_FUNCTION__)); |
15594 | DstTy = MVT::i64; |
15595 | } |
15596 | |
15597 | assert(DstTy.getSimpleVT() <= MVT::i64 &&((DstTy.getSimpleVT() <= MVT::i64 && DstTy.getSimpleVT () >= MVT::i16 && "Unknown FP_TO_INT to lower!") ? static_cast<void> (0) : __assert_fail ("DstTy.getSimpleVT() <= MVT::i64 && DstTy.getSimpleVT() >= MVT::i16 && \"Unknown FP_TO_INT to lower!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15599, __PRETTY_FUNCTION__)) |
15598 | DstTy.getSimpleVT() >= MVT::i16 &&((DstTy.getSimpleVT() <= MVT::i64 && DstTy.getSimpleVT () >= MVT::i16 && "Unknown FP_TO_INT to lower!") ? static_cast<void> (0) : __assert_fail ("DstTy.getSimpleVT() <= MVT::i64 && DstTy.getSimpleVT() >= MVT::i16 && \"Unknown FP_TO_INT to lower!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15599, __PRETTY_FUNCTION__)) |
15599 | "Unknown FP_TO_INT to lower!")((DstTy.getSimpleVT() <= MVT::i64 && DstTy.getSimpleVT () >= MVT::i16 && "Unknown FP_TO_INT to lower!") ? static_cast<void> (0) : __assert_fail ("DstTy.getSimpleVT() <= MVT::i64 && DstTy.getSimpleVT() >= MVT::i16 && \"Unknown FP_TO_INT to lower!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15599, __PRETTY_FUNCTION__)); |
15600 | |
15601 | // These are really Legal. |
15602 | if (DstTy == MVT::i32 && |
15603 | isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) |
15604 | return std::make_pair(SDValue(), SDValue()); |
15605 | if (Subtarget.is64Bit() && |
15606 | DstTy == MVT::i64 && |
15607 | isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) |
15608 | return std::make_pair(SDValue(), SDValue()); |
15609 | |
15610 | // We lower FP->int64 into FISTP64 followed by a load from a temporary |
15611 | // stack slot. |
15612 | MachineFunction &MF = DAG.getMachineFunction(); |
15613 | unsigned MemSize = DstTy.getSizeInBits()/8; |
15614 | int SSFI = MF.getFrameInfo().CreateStackObject(MemSize, MemSize, false); |
15615 | SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT); |
15616 | |
15617 | unsigned Opc; |
15618 | switch (DstTy.getSimpleVT().SimpleTy) { |
15619 | default: llvm_unreachable("Invalid FP_TO_SINT to lower!")::llvm::llvm_unreachable_internal("Invalid FP_TO_SINT to lower!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15619); |
15620 | case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break; |
15621 | case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break; |
15622 | case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break; |
15623 | } |
15624 | |
15625 | SDValue Chain = DAG.getEntryNode(); |
15626 | SDValue Value = Op.getOperand(0); |
15627 | SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment. |
15628 | |
15629 | if (UnsignedFixup) { |
15630 | // |
15631 | // Conversion to unsigned i64 is implemented with a select, |
15632 | // depending on whether the source value fits in the range |
15633 | // of a signed i64. Let Thresh be the FP equivalent of |
15634 | // 0x8000000000000000ULL. |
15635 | // |
15636 | // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000; |
15637 | // FistSrc = (Value < Thresh) ? Value : (Value - Thresh); |
15638 | // Fist-to-mem64 FistSrc |
15639 | // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent |
15640 | // to XOR'ing the high 32 bits with Adjust. |
15641 | // |
15642 | // Being a power of 2, Thresh is exactly representable in all FP formats. |
15643 | // For X87 we'd like to use the smallest FP type for this constant, but |
15644 | // for DAG type consistency we have to match the FP operand type. |
15645 | |
15646 | APFloat Thresh(APFloat::IEEEsingle(), APInt(32, 0x5f000000)); |
15647 | LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) APFloat::opStatus Status = APFloat::opOK; |
15648 | bool LosesInfo = false; |
15649 | if (TheVT == MVT::f64) |
15650 | // The rounding mode is irrelevant as the conversion should be exact. |
15651 | Status = Thresh.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven, |
15652 | &LosesInfo); |
15653 | else if (TheVT == MVT::f80) |
15654 | Status = Thresh.convert(APFloat::x87DoubleExtended(), |
15655 | APFloat::rmNearestTiesToEven, &LosesInfo); |
15656 | |
15657 | assert(Status == APFloat::opOK && !LosesInfo &&((Status == APFloat::opOK && !LosesInfo && "FP conversion should have been exact" ) ? static_cast<void> (0) : __assert_fail ("Status == APFloat::opOK && !LosesInfo && \"FP conversion should have been exact\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15658, __PRETTY_FUNCTION__)) |
15658 | "FP conversion should have been exact")((Status == APFloat::opOK && !LosesInfo && "FP conversion should have been exact" ) ? static_cast<void> (0) : __assert_fail ("Status == APFloat::opOK && !LosesInfo && \"FP conversion should have been exact\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15658, __PRETTY_FUNCTION__)); |
15659 | |
15660 | SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT); |
15661 | |
15662 | SDValue Cmp = DAG.getSetCC(DL, |
15663 | getSetCCResultType(DAG.getDataLayout(), |
15664 | *DAG.getContext(), TheVT), |
15665 | Value, ThreshVal, ISD::SETLT); |
15666 | Adjust = DAG.getSelect(DL, MVT::i32, Cmp, |
15667 | DAG.getConstant(0, DL, MVT::i32), |
15668 | DAG.getConstant(0x80000000, DL, MVT::i32)); |
15669 | SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal); |
15670 | Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(), |
15671 | *DAG.getContext(), TheVT), |
15672 | Value, ThreshVal, ISD::SETLT); |
15673 | Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub); |
15674 | } |
15675 | |
15676 | // FIXME This causes a redundant load/store if the SSE-class value is already |
15677 | // in memory, such as if it is on the callstack. |
15678 | if (isScalarFPTypeInSSEReg(TheVT)) { |
15679 | assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!")((DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!" ) ? static_cast<void> (0) : __assert_fail ("DstTy == MVT::i64 && \"Invalid FP_TO_SINT to lower!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15679, __PRETTY_FUNCTION__)); |
15680 | Chain = DAG.getStore(Chain, DL, Value, StackSlot, |
15681 | MachinePointerInfo::getFixedStack(MF, SSFI)); |
15682 | SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other); |
15683 | SDValue Ops[] = { |
15684 | Chain, StackSlot, DAG.getValueType(TheVT) |
15685 | }; |
15686 | |
15687 | MachineMemOperand *MMO = |
15688 | MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI), |
15689 | MachineMemOperand::MOLoad, MemSize, MemSize); |
15690 | Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO); |
15691 | Chain = Value.getValue(1); |
15692 | SSFI = MF.getFrameInfo().CreateStackObject(MemSize, MemSize, false); |
15693 | StackSlot = DAG.getFrameIndex(SSFI, PtrVT); |
15694 | } |
15695 | |
15696 | MachineMemOperand *MMO = |
15697 | MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI), |
15698 | MachineMemOperand::MOStore, MemSize, MemSize); |
15699 | |
15700 | if (UnsignedFixup) { |
15701 | |
15702 | // Insert the FIST, load its result as two i32's, |
15703 | // and XOR the high i32 with Adjust. |
15704 | |
15705 | SDValue FistOps[] = { Chain, Value, StackSlot }; |
15706 | SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other), |
15707 | FistOps, DstTy, MMO); |
15708 | |
15709 | SDValue Low32 = |
15710 | DAG.getLoad(MVT::i32, DL, FIST, StackSlot, MachinePointerInfo()); |
15711 | SDValue HighAddr = DAG.getMemBasePlusOffset(StackSlot, 4, DL); |
15712 | |
15713 | SDValue High32 = |
15714 | DAG.getLoad(MVT::i32, DL, FIST, HighAddr, MachinePointerInfo()); |
15715 | High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust); |
15716 | |
15717 | if (Subtarget.is64Bit()) { |
15718 | // Join High32 and Low32 into a 64-bit result. |
15719 | // (High32 << 32) | Low32 |
15720 | Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32); |
15721 | High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32); |
15722 | High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32, |
15723 | DAG.getConstant(32, DL, MVT::i8)); |
15724 | SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32); |
15725 | return std::make_pair(Result, SDValue()); |
15726 | } |
15727 | |
15728 | SDValue ResultOps[] = { Low32, High32 }; |
15729 | |
15730 | SDValue pair = IsReplace |
15731 | ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps) |
15732 | : DAG.getMergeValues(ResultOps, DL); |
15733 | return std::make_pair(pair, SDValue()); |
15734 | } else { |
15735 | // Build the FP_TO_INT*_IN_MEM |
15736 | SDValue Ops[] = { Chain, Value, StackSlot }; |
15737 | SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other), |
15738 | Ops, DstTy, MMO); |
15739 | return std::make_pair(FIST, StackSlot); |
15740 | } |
15741 | } |
15742 | |
15743 | static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG, |
15744 | const X86Subtarget &Subtarget) { |
15745 | MVT VT = Op->getSimpleValueType(0); |
15746 | SDValue In = Op->getOperand(0); |
15747 | MVT InVT = In.getSimpleValueType(); |
15748 | SDLoc dl(Op); |
15749 | |
15750 | if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1) |
15751 | return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In); |
15752 | |
15753 | // Optimize vectors in AVX mode: |
15754 | // |
15755 | // v8i16 -> v8i32 |
15756 | // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32. |
15757 | // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32. |
15758 | // Concat upper and lower parts. |
15759 | // |
15760 | // v4i32 -> v4i64 |
15761 | // Use vpunpckldq for 4 lower elements v4i32 -> v2i64. |
15762 | // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64. |
15763 | // Concat upper and lower parts. |
15764 | // |
15765 | |
15766 | if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) && |
15767 | ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) && |
15768 | ((VT != MVT::v4i64) || (InVT != MVT::v4i32))) |
15769 | return SDValue(); |
15770 | |
15771 | if (Subtarget.hasInt256()) |
15772 | return DAG.getNode(X86ISD::VZEXT, dl, VT, In); |
15773 | |
15774 | SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl); |
15775 | SDValue Undef = DAG.getUNDEF(InVT); |
15776 | bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND; |
15777 | SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef); |
15778 | SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef); |
15779 | |
15780 | MVT HVT = MVT::getVectorVT(VT.getVectorElementType(), |
15781 | VT.getVectorNumElements()/2); |
15782 | |
15783 | OpLo = DAG.getBitcast(HVT, OpLo); |
15784 | OpHi = DAG.getBitcast(HVT, OpHi); |
15785 | |
15786 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi); |
15787 | } |
15788 | |
15789 | static SDValue LowerZERO_EXTEND_AVX512(SDValue Op, |
15790 | const X86Subtarget &Subtarget, SelectionDAG &DAG) { |
15791 | MVT VT = Op->getSimpleValueType(0); |
15792 | SDValue In = Op->getOperand(0); |
15793 | MVT InVT = In.getSimpleValueType(); |
15794 | SDLoc DL(Op); |
15795 | unsigned NumElts = VT.getVectorNumElements(); |
15796 | |
15797 | if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1 && |
15798 | (NumElts == 8 || NumElts == 16 || Subtarget.hasBWI())) |
15799 | return DAG.getNode(X86ISD::VZEXT, DL, VT, In); |
15800 | |
15801 | if (InVT.getVectorElementType() != MVT::i1) |
15802 | return SDValue(); |
15803 | |
15804 | // Extend VT if the target is 256 or 128bit vector and VLX is not supported. |
15805 | MVT ExtVT = VT; |
15806 | if (!VT.is512BitVector() && !Subtarget.hasVLX()) |
15807 | ExtVT = MVT::getVectorVT(MVT::getIntegerVT(512/NumElts), NumElts); |
15808 | |
15809 | SDValue One = |
15810 | DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT); |
15811 | SDValue Zero = |
15812 | DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT); |
15813 | |
15814 | SDValue SelectedVal = DAG.getSelect(DL, ExtVT, In, One, Zero); |
15815 | if (VT == ExtVT) |
15816 | return SelectedVal; |
15817 | return DAG.getNode(X86ISD::VTRUNC, DL, VT, SelectedVal); |
15818 | } |
15819 | |
15820 | static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget &Subtarget, |
15821 | SelectionDAG &DAG) { |
15822 | if (Subtarget.hasFp256()) |
15823 | if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget)) |
15824 | return Res; |
15825 | |
15826 | return SDValue(); |
15827 | } |
15828 | |
15829 | static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget &Subtarget, |
15830 | SelectionDAG &DAG) { |
15831 | SDLoc DL(Op); |
15832 | MVT VT = Op.getSimpleValueType(); |
15833 | SDValue In = Op.getOperand(0); |
15834 | MVT SVT = In.getSimpleValueType(); |
15835 | |
15836 | if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1) |
15837 | return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG); |
15838 | |
15839 | if (Subtarget.hasFp256()) |
15840 | if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget)) |
15841 | return Res; |
15842 | |
15843 | assert(!VT.is256BitVector() || !SVT.is128BitVector() ||((!VT.is256BitVector() || !SVT.is128BitVector() || VT.getVectorNumElements () != SVT.getVectorNumElements()) ? static_cast<void> ( 0) : __assert_fail ("!VT.is256BitVector() || !SVT.is128BitVector() || VT.getVectorNumElements() != SVT.getVectorNumElements()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15844, __PRETTY_FUNCTION__)) |
15844 | VT.getVectorNumElements() != SVT.getVectorNumElements())((!VT.is256BitVector() || !SVT.is128BitVector() || VT.getVectorNumElements () != SVT.getVectorNumElements()) ? static_cast<void> ( 0) : __assert_fail ("!VT.is256BitVector() || !SVT.is128BitVector() || VT.getVectorNumElements() != SVT.getVectorNumElements()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15844, __PRETTY_FUNCTION__)); |
15845 | return SDValue(); |
15846 | } |
15847 | |
15848 | /// Helper to recursively truncate vector elements in half with PACKSS. |
15849 | /// It makes use of the fact that vector comparison results will be all-zeros |
15850 | /// or all-ones to use (vXi8 PACKSS(vYi16, vYi16)) instead of matching types. |
15851 | /// AVX2 (Int256) sub-targets require extra shuffling as the PACKSS operates |
15852 | /// within each 128-bit lane. |
15853 | static SDValue truncateVectorCompareWithPACKSS(EVT DstVT, SDValue In, |
15854 | const SDLoc &DL, |
15855 | SelectionDAG &DAG, |
15856 | const X86Subtarget &Subtarget) { |
15857 | // Requires SSE2 but AVX512 has fast truncate. |
15858 | if (!Subtarget.hasSSE2() || Subtarget.hasAVX512()) |
15859 | return SDValue(); |
15860 | |
15861 | EVT SrcVT = In.getValueType(); |
15862 | |
15863 | // No truncation required, we might get here due to recursive calls. |
15864 | if (SrcVT == DstVT) |
15865 | return In; |
15866 | |
15867 | // We only support vector truncation to 128bits or greater from a |
15868 | // 256bits or greater source. |
15869 | if ((DstVT.getSizeInBits() % 128) != 0) |
15870 | return SDValue(); |
15871 | if ((SrcVT.getSizeInBits() % 256) != 0) |
15872 | return SDValue(); |
15873 | |
15874 | unsigned NumElems = SrcVT.getVectorNumElements(); |
15875 | assert(DstVT.getVectorNumElements() == NumElems && "Illegal truncation")((DstVT.getVectorNumElements() == NumElems && "Illegal truncation" ) ? static_cast<void> (0) : __assert_fail ("DstVT.getVectorNumElements() == NumElems && \"Illegal truncation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15875, __PRETTY_FUNCTION__)); |
15876 | assert(SrcVT.getSizeInBits() > DstVT.getSizeInBits() && "Illegal truncation")((SrcVT.getSizeInBits() > DstVT.getSizeInBits() && "Illegal truncation") ? static_cast<void> (0) : __assert_fail ("SrcVT.getSizeInBits() > DstVT.getSizeInBits() && \"Illegal truncation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15876, __PRETTY_FUNCTION__)); |
15877 | |
15878 | EVT PackedSVT = |
15879 | EVT::getIntegerVT(*DAG.getContext(), SrcVT.getScalarSizeInBits() / 2); |
15880 | |
15881 | // Extract lower/upper subvectors. |
15882 | unsigned NumSubElts = NumElems / 2; |
15883 | unsigned SrcSizeInBits = SrcVT.getSizeInBits(); |
15884 | SDValue Lo = extractSubVector(In, 0 * NumSubElts, DAG, DL, SrcSizeInBits / 2); |
15885 | SDValue Hi = extractSubVector(In, 1 * NumSubElts, DAG, DL, SrcSizeInBits / 2); |
15886 | |
15887 | // 256bit -> 128bit truncate - PACKSS lower/upper 128-bit subvectors. |
15888 | if (SrcVT.is256BitVector()) { |
15889 | Lo = DAG.getBitcast(MVT::v8i16, Lo); |
15890 | Hi = DAG.getBitcast(MVT::v8i16, Hi); |
15891 | SDValue Res = DAG.getNode(X86ISD::PACKSS, DL, MVT::v16i8, Lo, Hi); |
15892 | return DAG.getBitcast(DstVT, Res); |
15893 | } |
15894 | |
15895 | // AVX2: 512bit -> 256bit truncate - PACKSS lower/upper 256-bit subvectors. |
15896 | // AVX2: 512bit -> 128bit truncate - PACKSS(PACKSS, PACKSS). |
15897 | if (SrcVT.is512BitVector() && Subtarget.hasInt256()) { |
15898 | Lo = DAG.getBitcast(MVT::v16i16, Lo); |
15899 | Hi = DAG.getBitcast(MVT::v16i16, Hi); |
15900 | SDValue Res = DAG.getNode(X86ISD::PACKSS, DL, MVT::v32i8, Lo, Hi); |
15901 | |
15902 | // 256-bit PACKSS(ARG0, ARG1) leaves us with ((LO0,LO1),(HI0,HI1)), |
15903 | // so we need to shuffle to get ((LO0,HI0),(LO1,HI1)). |
15904 | Res = DAG.getBitcast(MVT::v4i64, Res); |
15905 | Res = DAG.getVectorShuffle(MVT::v4i64, DL, Res, Res, {0, 2, 1, 3}); |
15906 | |
15907 | if (DstVT.is256BitVector()) |
15908 | return DAG.getBitcast(DstVT, Res); |
15909 | |
15910 | // If 512bit -> 128bit truncate another stage. |
15911 | EVT PackedVT = EVT::getVectorVT(*DAG.getContext(), PackedSVT, NumElems); |
15912 | Res = DAG.getBitcast(PackedVT, Res); |
15913 | return truncateVectorCompareWithPACKSS(DstVT, Res, DL, DAG, Subtarget); |
15914 | } |
15915 | |
15916 | // Recursively pack lower/upper subvectors, concat result and pack again. |
15917 | assert(SrcVT.getSizeInBits() >= 512 && "Expected 512-bit vector or greater")((SrcVT.getSizeInBits() >= 512 && "Expected 512-bit vector or greater" ) ? static_cast<void> (0) : __assert_fail ("SrcVT.getSizeInBits() >= 512 && \"Expected 512-bit vector or greater\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15917, __PRETTY_FUNCTION__)); |
15918 | EVT PackedVT = EVT::getVectorVT(*DAG.getContext(), PackedSVT, NumElems / 2); |
15919 | Lo = truncateVectorCompareWithPACKSS(PackedVT, Lo, DL, DAG, Subtarget); |
15920 | Hi = truncateVectorCompareWithPACKSS(PackedVT, Hi, DL, DAG, Subtarget); |
15921 | |
15922 | PackedVT = EVT::getVectorVT(*DAG.getContext(), PackedSVT, NumElems); |
15923 | SDValue Res = DAG.getNode(ISD::CONCAT_VECTORS, DL, PackedVT, Lo, Hi); |
15924 | return truncateVectorCompareWithPACKSS(DstVT, Res, DL, DAG, Subtarget); |
15925 | } |
15926 | |
15927 | static SDValue LowerTruncateVecI1(SDValue Op, SelectionDAG &DAG, |
15928 | const X86Subtarget &Subtarget) { |
15929 | |
15930 | SDLoc DL(Op); |
15931 | MVT VT = Op.getSimpleValueType(); |
15932 | SDValue In = Op.getOperand(0); |
15933 | MVT InVT = In.getSimpleValueType(); |
15934 | |
15935 | assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type.")((VT.getVectorElementType() == MVT::i1 && "Unexpected vector type." ) ? static_cast<void> (0) : __assert_fail ("VT.getVectorElementType() == MVT::i1 && \"Unexpected vector type.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15935, __PRETTY_FUNCTION__)); |
15936 | |
15937 | // Shift LSB to MSB and use VPMOVB/W2M or TESTD/Q. |
15938 | unsigned ShiftInx = InVT.getScalarSizeInBits() - 1; |
15939 | if (InVT.getScalarSizeInBits() <= 16) { |
15940 | if (Subtarget.hasBWI()) { |
15941 | // legal, will go to VPMOVB2M, VPMOVW2M |
15942 | // Shift packed bytes not supported natively, bitcast to word |
15943 | MVT ExtVT = MVT::getVectorVT(MVT::i16, InVT.getSizeInBits()/16); |
15944 | SDValue ShiftNode = DAG.getNode(ISD::SHL, DL, ExtVT, |
15945 | DAG.getBitcast(ExtVT, In), |
15946 | DAG.getConstant(ShiftInx, DL, ExtVT)); |
15947 | ShiftNode = DAG.getBitcast(InVT, ShiftNode); |
15948 | return DAG.getNode(X86ISD::CVT2MASK, DL, VT, ShiftNode); |
15949 | } |
15950 | // Use TESTD/Q, extended vector to packed dword/qword. |
15951 | assert((InVT.is256BitVector() || InVT.is128BitVector()) &&(((InVT.is256BitVector() || InVT.is128BitVector()) && "Unexpected vector type.") ? static_cast<void> (0) : __assert_fail ("(InVT.is256BitVector() || InVT.is128BitVector()) && \"Unexpected vector type.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15952, __PRETTY_FUNCTION__)) |
15952 | "Unexpected vector type.")(((InVT.is256BitVector() || InVT.is128BitVector()) && "Unexpected vector type.") ? static_cast<void> (0) : __assert_fail ("(InVT.is256BitVector() || InVT.is128BitVector()) && \"Unexpected vector type.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15952, __PRETTY_FUNCTION__)); |
15953 | unsigned NumElts = InVT.getVectorNumElements(); |
15954 | MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(512/NumElts), NumElts); |
15955 | In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In); |
15956 | InVT = ExtVT; |
15957 | ShiftInx = InVT.getScalarSizeInBits() - 1; |
15958 | } |
15959 | |
15960 | SDValue ShiftNode = DAG.getNode(ISD::SHL, DL, InVT, In, |
15961 | DAG.getConstant(ShiftInx, DL, InVT)); |
15962 | return DAG.getNode(X86ISD::TESTM, DL, VT, ShiftNode, ShiftNode); |
15963 | } |
15964 | |
15965 | SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const { |
15966 | SDLoc DL(Op); |
15967 | MVT VT = Op.getSimpleValueType(); |
15968 | SDValue In = Op.getOperand(0); |
15969 | MVT InVT = In.getSimpleValueType(); |
15970 | |
15971 | if (VT == MVT::i1) { |
15972 | assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&(((InVT.isInteger() && (InVT.getSizeInBits() <= 64 )) && "Invalid scalar TRUNCATE operation") ? static_cast <void> (0) : __assert_fail ("(InVT.isInteger() && (InVT.getSizeInBits() <= 64)) && \"Invalid scalar TRUNCATE operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15973, __PRETTY_FUNCTION__)) |
15973 | "Invalid scalar TRUNCATE operation")(((InVT.isInteger() && (InVT.getSizeInBits() <= 64 )) && "Invalid scalar TRUNCATE operation") ? static_cast <void> (0) : __assert_fail ("(InVT.isInteger() && (InVT.getSizeInBits() <= 64)) && \"Invalid scalar TRUNCATE operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15973, __PRETTY_FUNCTION__)); |
15974 | if (InVT.getSizeInBits() >= 32) |
15975 | return SDValue(); |
15976 | In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In); |
15977 | return DAG.getNode(ISD::TRUNCATE, DL, VT, In); |
15978 | } |
15979 | assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&((VT.getVectorNumElements() == InVT.getVectorNumElements() && "Invalid TRUNCATE operation") ? static_cast<void> (0) : __assert_fail ("VT.getVectorNumElements() == InVT.getVectorNumElements() && \"Invalid TRUNCATE operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15980, __PRETTY_FUNCTION__)) |
15980 | "Invalid TRUNCATE operation")((VT.getVectorNumElements() == InVT.getVectorNumElements() && "Invalid TRUNCATE operation") ? static_cast<void> (0) : __assert_fail ("VT.getVectorNumElements() == InVT.getVectorNumElements() && \"Invalid TRUNCATE operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 15980, __PRETTY_FUNCTION__)); |
15981 | |
15982 | if (VT.getVectorElementType() == MVT::i1) |
15983 | return LowerTruncateVecI1(Op, DAG, Subtarget); |
15984 | |
15985 | // vpmovqb/w/d, vpmovdb/w, vpmovwb |
15986 | if (Subtarget.hasAVX512()) { |
15987 | // word to byte only under BWI |
15988 | if (InVT == MVT::v16i16 && !Subtarget.hasBWI()) // v16i16 -> v16i8 |
15989 | return DAG.getNode(X86ISD::VTRUNC, DL, VT, |
15990 | getExtendInVec(X86ISD::VSEXT, DL, MVT::v16i32, In, DAG)); |
15991 | return DAG.getNode(X86ISD::VTRUNC, DL, VT, In); |
15992 | } |
15993 | |
15994 | // Truncate with PACKSS if we are truncating a vector zero/all-bits result. |
15995 | if (InVT.getScalarSizeInBits() == DAG.ComputeNumSignBits(In)) |
15996 | if (SDValue V = truncateVectorCompareWithPACKSS(VT, In, DL, DAG, Subtarget)) |
15997 | return V; |
15998 | |
15999 | if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) { |
16000 | // On AVX2, v4i64 -> v4i32 becomes VPERMD. |
16001 | if (Subtarget.hasInt256()) { |
16002 | static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1}; |
16003 | In = DAG.getBitcast(MVT::v8i32, In); |
16004 | In = DAG.getVectorShuffle(MVT::v8i32, DL, In, In, ShufMask); |
16005 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In, |
16006 | DAG.getIntPtrConstant(0, DL)); |
16007 | } |
16008 | |
16009 | SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In, |
16010 | DAG.getIntPtrConstant(0, DL)); |
16011 | SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In, |
16012 | DAG.getIntPtrConstant(2, DL)); |
16013 | OpLo = DAG.getBitcast(MVT::v4i32, OpLo); |
16014 | OpHi = DAG.getBitcast(MVT::v4i32, OpHi); |
16015 | static const int ShufMask[] = {0, 2, 4, 6}; |
16016 | return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask); |
16017 | } |
16018 | |
16019 | if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) { |
16020 | // On AVX2, v8i32 -> v8i16 becomes PSHUFB. |
16021 | if (Subtarget.hasInt256()) { |
16022 | In = DAG.getBitcast(MVT::v32i8, In); |
16023 | |
16024 | // The PSHUFB mask: |
16025 | static const int ShufMask1[] = { 0, 1, 4, 5, 8, 9, 12, 13, |
16026 | -1, -1, -1, -1, -1, -1, -1, -1, |
16027 | 16, 17, 20, 21, 24, 25, 28, 29, |
16028 | -1, -1, -1, -1, -1, -1, -1, -1 }; |
16029 | In = DAG.getVectorShuffle(MVT::v32i8, DL, In, In, ShufMask1); |
16030 | In = DAG.getBitcast(MVT::v4i64, In); |
16031 | |
16032 | static const int ShufMask2[] = {0, 2, -1, -1}; |
16033 | In = DAG.getVectorShuffle(MVT::v4i64, DL, In, In, ShufMask2); |
16034 | In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In, |
16035 | DAG.getIntPtrConstant(0, DL)); |
16036 | return DAG.getBitcast(VT, In); |
16037 | } |
16038 | |
16039 | SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In, |
16040 | DAG.getIntPtrConstant(0, DL)); |
16041 | |
16042 | SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In, |
16043 | DAG.getIntPtrConstant(4, DL)); |
16044 | |
16045 | OpLo = DAG.getBitcast(MVT::v16i8, OpLo); |
16046 | OpHi = DAG.getBitcast(MVT::v16i8, OpHi); |
16047 | |
16048 | // The PSHUFB mask: |
16049 | static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13, |
16050 | -1, -1, -1, -1, -1, -1, -1, -1}; |
16051 | |
16052 | OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, OpLo, ShufMask1); |
16053 | OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, OpHi, ShufMask1); |
16054 | |
16055 | OpLo = DAG.getBitcast(MVT::v4i32, OpLo); |
16056 | OpHi = DAG.getBitcast(MVT::v4i32, OpHi); |
16057 | |
16058 | // The MOVLHPS Mask: |
16059 | static const int ShufMask2[] = {0, 1, 4, 5}; |
16060 | SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2); |
16061 | return DAG.getBitcast(MVT::v8i16, res); |
16062 | } |
16063 | |
16064 | // Handle truncation of V256 to V128 using shuffles. |
16065 | if (!VT.is128BitVector() || !InVT.is256BitVector()) |
16066 | return SDValue(); |
16067 | |
16068 | assert(Subtarget.hasFp256() && "256-bit vector without AVX!")((Subtarget.hasFp256() && "256-bit vector without AVX!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasFp256() && \"256-bit vector without AVX!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16068, __PRETTY_FUNCTION__)); |
16069 | |
16070 | unsigned NumElems = VT.getVectorNumElements(); |
16071 | MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2); |
16072 | |
16073 | SmallVector<int, 16> MaskVec(NumElems * 2, -1); |
16074 | // Prepare truncation shuffle mask |
16075 | for (unsigned i = 0; i != NumElems; ++i) |
16076 | MaskVec[i] = i * 2; |
16077 | In = DAG.getBitcast(NVT, In); |
16078 | SDValue V = DAG.getVectorShuffle(NVT, DL, In, In, MaskVec); |
16079 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V, |
16080 | DAG.getIntPtrConstant(0, DL)); |
16081 | } |
16082 | |
16083 | SDValue X86TargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const { |
16084 | bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT; |
16085 | MVT VT = Op.getSimpleValueType(); |
16086 | |
16087 | if (VT.isVector()) { |
16088 | assert(Subtarget.hasDQI() && Subtarget.hasVLX() && "Requires AVX512DQVL!")((Subtarget.hasDQI() && Subtarget.hasVLX() && "Requires AVX512DQVL!") ? static_cast<void> (0) : __assert_fail ("Subtarget.hasDQI() && Subtarget.hasVLX() && \"Requires AVX512DQVL!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16088, __PRETTY_FUNCTION__)); |
16089 | SDValue Src = Op.getOperand(0); |
16090 | SDLoc dl(Op); |
16091 | if (VT == MVT::v2i64 && Src.getSimpleValueType() == MVT::v2f32) { |
16092 | return DAG.getNode(IsSigned ? X86ISD::CVTTP2SI : X86ISD::CVTTP2UI, dl, VT, |
16093 | DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32, Src, |
16094 | DAG.getUNDEF(MVT::v2f32))); |
16095 | } |
16096 | |
16097 | return SDValue(); |
16098 | } |
16099 | |
16100 | assert(!VT.isVector())((!VT.isVector()) ? static_cast<void> (0) : __assert_fail ("!VT.isVector()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16100, __PRETTY_FUNCTION__)); |
16101 | |
16102 | std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, |
16103 | IsSigned, /*IsReplace=*/ false); |
16104 | SDValue FIST = Vals.first, StackSlot = Vals.second; |
16105 | // If FP_TO_INTHelper failed, the node is actually supposed to be Legal. |
16106 | if (!FIST.getNode()) |
16107 | return Op; |
16108 | |
16109 | if (StackSlot.getNode()) |
16110 | // Load the result. |
16111 | return DAG.getLoad(VT, SDLoc(Op), FIST, StackSlot, MachinePointerInfo()); |
16112 | |
16113 | // The node is the result. |
16114 | return FIST; |
16115 | } |
16116 | |
16117 | static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) { |
16118 | SDLoc DL(Op); |
16119 | MVT VT = Op.getSimpleValueType(); |
16120 | SDValue In = Op.getOperand(0); |
16121 | MVT SVT = In.getSimpleValueType(); |
16122 | |
16123 | assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!")((SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!" ) ? static_cast<void> (0) : __assert_fail ("SVT == MVT::v2f32 && \"Only customize MVT::v2f32 type legalization!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16123, __PRETTY_FUNCTION__)); |
16124 | |
16125 | return DAG.getNode(X86ISD::VFPEXT, DL, VT, |
16126 | DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32, |
16127 | In, DAG.getUNDEF(SVT))); |
16128 | } |
16129 | |
16130 | /// The only differences between FABS and FNEG are the mask and the logic op. |
16131 | /// FNEG also has a folding opportunity for FNEG(FABS(x)). |
16132 | static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) { |
16133 | assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&(((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG ) && "Wrong opcode for lowering FABS or FNEG.") ? static_cast <void> (0) : __assert_fail ("(Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) && \"Wrong opcode for lowering FABS or FNEG.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16134, __PRETTY_FUNCTION__)) |
16134 | "Wrong opcode for lowering FABS or FNEG.")(((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG ) && "Wrong opcode for lowering FABS or FNEG.") ? static_cast <void> (0) : __assert_fail ("(Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) && \"Wrong opcode for lowering FABS or FNEG.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16134, __PRETTY_FUNCTION__)); |
16135 | |
16136 | bool IsFABS = (Op.getOpcode() == ISD::FABS); |
16137 | |
16138 | // If this is a FABS and it has an FNEG user, bail out to fold the combination |
16139 | // into an FNABS. We'll lower the FABS after that if it is still in use. |
16140 | if (IsFABS) |
16141 | for (SDNode *User : Op->uses()) |
16142 | if (User->getOpcode() == ISD::FNEG) |
16143 | return Op; |
16144 | |
16145 | SDLoc dl(Op); |
16146 | MVT VT = Op.getSimpleValueType(); |
16147 | |
16148 | bool IsF128 = (VT == MVT::f128); |
16149 | |
16150 | // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to |
16151 | // decide if we should generate a 16-byte constant mask when we only need 4 or |
16152 | // 8 bytes for the scalar case. |
16153 | |
16154 | MVT LogicVT; |
16155 | MVT EltVT; |
16156 | |
16157 | if (VT.isVector()) { |
16158 | LogicVT = VT; |
16159 | EltVT = VT.getVectorElementType(); |
16160 | } else if (IsF128) { |
16161 | // SSE instructions are used for optimized f128 logical operations. |
16162 | LogicVT = MVT::f128; |
16163 | EltVT = VT; |
16164 | } else { |
16165 | // There are no scalar bitwise logical SSE/AVX instructions, so we |
16166 | // generate a 16-byte vector constant and logic op even for the scalar case. |
16167 | // Using a 16-byte mask allows folding the load of the mask with |
16168 | // the logic op, so it can save (~4 bytes) on code size. |
16169 | LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32; |
16170 | EltVT = VT; |
16171 | } |
16172 | |
16173 | unsigned EltBits = EltVT.getSizeInBits(); |
16174 | // For FABS, mask is 0x7f...; for FNEG, mask is 0x80... |
16175 | APInt MaskElt = |
16176 | IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignMask(EltBits); |
16177 | const fltSemantics &Sem = |
16178 | EltVT == MVT::f64 ? APFloat::IEEEdouble() : |
16179 | (IsF128 ? APFloat::IEEEquad() : APFloat::IEEEsingle()); |
16180 | SDValue Mask = DAG.getConstantFP(APFloat(Sem, MaskElt), dl, LogicVT); |
16181 | |
16182 | SDValue Op0 = Op.getOperand(0); |
16183 | bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS); |
16184 | unsigned LogicOp = |
16185 | IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR; |
16186 | SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0; |
16187 | |
16188 | if (VT.isVector() || IsF128) |
16189 | return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask); |
16190 | |
16191 | // For the scalar case extend to a 128-bit vector, perform the logic op, |
16192 | // and extract the scalar result back out. |
16193 | Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand); |
16194 | SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask); |
16195 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode, |
16196 | DAG.getIntPtrConstant(0, dl)); |
16197 | } |
16198 | |
16199 | static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) { |
16200 | SDValue Mag = Op.getOperand(0); |
16201 | SDValue Sign = Op.getOperand(1); |
16202 | SDLoc dl(Op); |
16203 | |
16204 | // If the sign operand is smaller, extend it first. |
16205 | MVT VT = Op.getSimpleValueType(); |
16206 | if (Sign.getSimpleValueType().bitsLT(VT)) |
16207 | Sign = DAG.getNode(ISD::FP_EXTEND, dl, VT, Sign); |
16208 | |
16209 | // And if it is bigger, shrink it first. |
16210 | if (Sign.getSimpleValueType().bitsGT(VT)) |
16211 | Sign = DAG.getNode(ISD::FP_ROUND, dl, VT, Sign, DAG.getIntPtrConstant(1, dl)); |
16212 | |
16213 | // At this point the operands and the result should have the same |
16214 | // type, and that won't be f80 since that is not custom lowered. |
16215 | bool IsF128 = (VT == MVT::f128); |
16216 | assert((VT == MVT::f64 || VT == MVT::f32 || VT == MVT::f128 ||(((VT == MVT::f64 || VT == MVT::f32 || VT == MVT::f128 || VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v4f32 || VT == MVT ::v8f32 || VT == MVT::v8f64 || VT == MVT::v16f32) && "Unexpected type in LowerFCOPYSIGN" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::f64 || VT == MVT::f32 || VT == MVT::f128 || VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v4f32 || VT == MVT::v8f32 || VT == MVT::v8f64 || VT == MVT::v16f32) && \"Unexpected type in LowerFCOPYSIGN\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16219, __PRETTY_FUNCTION__)) |
16217 | VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v4f32 ||(((VT == MVT::f64 || VT == MVT::f32 || VT == MVT::f128 || VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v4f32 || VT == MVT ::v8f32 || VT == MVT::v8f64 || VT == MVT::v16f32) && "Unexpected type in LowerFCOPYSIGN" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::f64 || VT == MVT::f32 || VT == MVT::f128 || VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v4f32 || VT == MVT::v8f32 || VT == MVT::v8f64 || VT == MVT::v16f32) && \"Unexpected type in LowerFCOPYSIGN\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16219, __PRETTY_FUNCTION__)) |
16218 | VT == MVT::v8f32 || VT == MVT::v8f64 || VT == MVT::v16f32) &&(((VT == MVT::f64 || VT == MVT::f32 || VT == MVT::f128 || VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v4f32 || VT == MVT ::v8f32 || VT == MVT::v8f64 || VT == MVT::v16f32) && "Unexpected type in LowerFCOPYSIGN" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::f64 || VT == MVT::f32 || VT == MVT::f128 || VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v4f32 || VT == MVT::v8f32 || VT == MVT::v8f64 || VT == MVT::v16f32) && \"Unexpected type in LowerFCOPYSIGN\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16219, __PRETTY_FUNCTION__)) |
16219 | "Unexpected type in LowerFCOPYSIGN")(((VT == MVT::f64 || VT == MVT::f32 || VT == MVT::f128 || VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v4f32 || VT == MVT ::v8f32 || VT == MVT::v8f64 || VT == MVT::v16f32) && "Unexpected type in LowerFCOPYSIGN" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::f64 || VT == MVT::f32 || VT == MVT::f128 || VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v4f32 || VT == MVT::v8f32 || VT == MVT::v8f64 || VT == MVT::v16f32) && \"Unexpected type in LowerFCOPYSIGN\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16219, __PRETTY_FUNCTION__)); |
16220 | |
16221 | MVT EltVT = VT.getScalarType(); |
16222 | const fltSemantics &Sem = |
16223 | EltVT == MVT::f64 ? APFloat::IEEEdouble() |
16224 | : (IsF128 ? APFloat::IEEEquad() : APFloat::IEEEsingle()); |
16225 | |
16226 | // Perform all scalar logic operations as 16-byte vectors because there are no |
16227 | // scalar FP logic instructions in SSE. |
16228 | // TODO: This isn't necessary. If we used scalar types, we might avoid some |
16229 | // unnecessary splats, but we might miss load folding opportunities. Should |
16230 | // this decision be based on OptimizeForSize? |
16231 | bool IsFakeVector = !VT.isVector() && !IsF128; |
16232 | MVT LogicVT = VT; |
16233 | if (IsFakeVector) |
16234 | LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32; |
16235 | |
16236 | // The mask constants are automatically splatted for vector types. |
16237 | unsigned EltSizeInBits = VT.getScalarSizeInBits(); |
16238 | SDValue SignMask = DAG.getConstantFP( |
16239 | APFloat(Sem, APInt::getSignMask(EltSizeInBits)), dl, LogicVT); |
16240 | SDValue MagMask = DAG.getConstantFP( |
16241 | APFloat(Sem, ~APInt::getSignMask(EltSizeInBits)), dl, LogicVT); |
16242 | |
16243 | // First, clear all bits but the sign bit from the second operand (sign). |
16244 | if (IsFakeVector) |
16245 | Sign = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Sign); |
16246 | SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Sign, SignMask); |
16247 | |
16248 | // Next, clear the sign bit from the first operand (magnitude). |
16249 | // TODO: If we had general constant folding for FP logic ops, this check |
16250 | // wouldn't be necessary. |
16251 | SDValue MagBits; |
16252 | if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Mag)) { |
16253 | APFloat APF = Op0CN->getValueAPF(); |
16254 | APF.clearSign(); |
16255 | MagBits = DAG.getConstantFP(APF, dl, LogicVT); |
16256 | } else { |
16257 | // If the magnitude operand wasn't a constant, we need to AND out the sign. |
16258 | if (IsFakeVector) |
16259 | Mag = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Mag); |
16260 | MagBits = DAG.getNode(X86ISD::FAND, dl, LogicVT, Mag, MagMask); |
16261 | } |
16262 | |
16263 | // OR the magnitude value with the sign bit. |
16264 | SDValue Or = DAG.getNode(X86ISD::FOR, dl, LogicVT, MagBits, SignBit); |
16265 | return !IsFakeVector ? Or : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Or, |
16266 | DAG.getIntPtrConstant(0, dl)); |
16267 | } |
16268 | |
16269 | static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) { |
16270 | SDValue N0 = Op.getOperand(0); |
16271 | SDLoc dl(Op); |
16272 | MVT VT = Op.getSimpleValueType(); |
16273 | |
16274 | MVT OpVT = N0.getSimpleValueType(); |
16275 | assert((OpVT == MVT::f32 || OpVT == MVT::f64) &&(((OpVT == MVT::f32 || OpVT == MVT::f64) && "Unexpected type for FGETSIGN" ) ? static_cast<void> (0) : __assert_fail ("(OpVT == MVT::f32 || OpVT == MVT::f64) && \"Unexpected type for FGETSIGN\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16276, __PRETTY_FUNCTION__)) |
16276 | "Unexpected type for FGETSIGN")(((OpVT == MVT::f32 || OpVT == MVT::f64) && "Unexpected type for FGETSIGN" ) ? static_cast<void> (0) : __assert_fail ("(OpVT == MVT::f32 || OpVT == MVT::f64) && \"Unexpected type for FGETSIGN\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16276, __PRETTY_FUNCTION__)); |
16277 | |
16278 | // Lower ISD::FGETSIGN to (AND (X86ISD::MOVMSK ...) 1). |
16279 | MVT VecVT = (OpVT == MVT::f32 ? MVT::v4f32 : MVT::v2f64); |
16280 | SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, N0); |
16281 | Res = DAG.getNode(X86ISD::MOVMSK, dl, MVT::i32, Res); |
16282 | Res = DAG.getZExtOrTrunc(Res, dl, VT); |
16283 | Res = DAG.getNode(ISD::AND, dl, VT, Res, DAG.getConstant(1, dl, VT)); |
16284 | return Res; |
16285 | } |
16286 | |
16287 | // Check whether an OR'd tree is PTEST-able. |
16288 | static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget &Subtarget, |
16289 | SelectionDAG &DAG) { |
16290 | assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.")((Op.getOpcode() == ISD::OR && "Only check OR'd tree." ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::OR && \"Only check OR'd tree.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16290, __PRETTY_FUNCTION__)); |
16291 | |
16292 | if (!Subtarget.hasSSE41()) |
16293 | return SDValue(); |
16294 | |
16295 | if (!Op->hasOneUse()) |
16296 | return SDValue(); |
16297 | |
16298 | SDNode *N = Op.getNode(); |
16299 | SDLoc DL(N); |
16300 | |
16301 | SmallVector<SDValue, 8> Opnds; |
16302 | DenseMap<SDValue, unsigned> VecInMap; |
16303 | SmallVector<SDValue, 8> VecIns; |
16304 | EVT VT = MVT::Other; |
16305 | |
16306 | // Recognize a special case where a vector is casted into wide integer to |
16307 | // test all 0s. |
16308 | Opnds.push_back(N->getOperand(0)); |
16309 | Opnds.push_back(N->getOperand(1)); |
16310 | |
16311 | for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) { |
16312 | SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot; |
16313 | // BFS traverse all OR'd operands. |
16314 | if (I->getOpcode() == ISD::OR) { |
16315 | Opnds.push_back(I->getOperand(0)); |
16316 | Opnds.push_back(I->getOperand(1)); |
16317 | // Re-evaluate the number of nodes to be traversed. |
16318 | e += 2; // 2 more nodes (LHS and RHS) are pushed. |
16319 | continue; |
16320 | } |
16321 | |
16322 | // Quit if a non-EXTRACT_VECTOR_ELT |
16323 | if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
16324 | return SDValue(); |
16325 | |
16326 | // Quit if without a constant index. |
16327 | SDValue Idx = I->getOperand(1); |
16328 | if (!isa<ConstantSDNode>(Idx)) |
16329 | return SDValue(); |
16330 | |
16331 | SDValue ExtractedFromVec = I->getOperand(0); |
16332 | DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec); |
16333 | if (M == VecInMap.end()) { |
16334 | VT = ExtractedFromVec.getValueType(); |
16335 | // Quit if not 128/256-bit vector. |
16336 | if (!VT.is128BitVector() && !VT.is256BitVector()) |
16337 | return SDValue(); |
16338 | // Quit if not the same type. |
16339 | if (VecInMap.begin() != VecInMap.end() && |
16340 | VT != VecInMap.begin()->first.getValueType()) |
16341 | return SDValue(); |
16342 | M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first; |
16343 | VecIns.push_back(ExtractedFromVec); |
16344 | } |
16345 | M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue(); |
16346 | } |
16347 | |
16348 | assert((VT.is128BitVector() || VT.is256BitVector()) &&(((VT.is128BitVector() || VT.is256BitVector()) && "Not extracted from 128-/256-bit vector." ) ? static_cast<void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector()) && \"Not extracted from 128-/256-bit vector.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16349, __PRETTY_FUNCTION__)) |
16349 | "Not extracted from 128-/256-bit vector.")(((VT.is128BitVector() || VT.is256BitVector()) && "Not extracted from 128-/256-bit vector." ) ? static_cast<void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector()) && \"Not extracted from 128-/256-bit vector.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16349, __PRETTY_FUNCTION__)); |
16350 | |
16351 | unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U; |
16352 | |
16353 | for (DenseMap<SDValue, unsigned>::const_iterator |
16354 | I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) { |
16355 | // Quit if not all elements are used. |
16356 | if (I->second != FullMask) |
16357 | return SDValue(); |
16358 | } |
16359 | |
16360 | MVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64; |
16361 | |
16362 | // Cast all vectors into TestVT for PTEST. |
16363 | for (unsigned i = 0, e = VecIns.size(); i < e; ++i) |
16364 | VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]); |
16365 | |
16366 | // If more than one full vector is evaluated, OR them first before PTEST. |
16367 | for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) { |
16368 | // Each iteration will OR 2 nodes and append the result until there is only |
16369 | // 1 node left, i.e. the final OR'd value of all vectors. |
16370 | SDValue LHS = VecIns[Slot]; |
16371 | SDValue RHS = VecIns[Slot + 1]; |
16372 | VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS)); |
16373 | } |
16374 | |
16375 | return DAG.getNode(X86ISD::PTEST, DL, MVT::i32, VecIns.back(), VecIns.back()); |
16376 | } |
16377 | |
16378 | /// \brief return true if \c Op has a use that doesn't just read flags. |
16379 | static bool hasNonFlagsUse(SDValue Op) { |
16380 | for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE; |
16381 | ++UI) { |
16382 | SDNode *User = *UI; |
16383 | unsigned UOpNo = UI.getOperandNo(); |
16384 | if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) { |
16385 | // Look pass truncate. |
16386 | UOpNo = User->use_begin().getOperandNo(); |
16387 | User = *User->use_begin(); |
16388 | } |
16389 | |
16390 | if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC && |
16391 | !(User->getOpcode() == ISD::SELECT && UOpNo == 0)) |
16392 | return true; |
16393 | } |
16394 | return false; |
16395 | } |
16396 | |
16397 | // Emit KTEST instruction for bit vectors on AVX-512 |
16398 | static SDValue EmitKTEST(SDValue Op, SelectionDAG &DAG, |
16399 | const X86Subtarget &Subtarget) { |
16400 | if (Op.getOpcode() == ISD::BITCAST) { |
16401 | auto hasKTEST = [&](MVT VT) { |
16402 | unsigned SizeInBits = VT.getSizeInBits(); |
16403 | return (Subtarget.hasDQI() && (SizeInBits == 8 || SizeInBits == 16)) || |
16404 | (Subtarget.hasBWI() && (SizeInBits == 32 || SizeInBits == 64)); |
16405 | }; |
16406 | SDValue Op0 = Op.getOperand(0); |
16407 | MVT Op0VT = Op0.getValueType().getSimpleVT(); |
16408 | if (Op0VT.isVector() && Op0VT.getVectorElementType() == MVT::i1 && |
16409 | hasKTEST(Op0VT)) |
16410 | return DAG.getNode(X86ISD::KTEST, SDLoc(Op), Op0VT, Op0, Op0); |
16411 | } |
16412 | return SDValue(); |
16413 | } |
16414 | |
16415 | /// Emit nodes that will be selected as "test Op0,Op0", or something |
16416 | /// equivalent. |
16417 | SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, const SDLoc &dl, |
16418 | SelectionDAG &DAG) const { |
16419 | if (Op.getValueType() == MVT::i1) { |
16420 | SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op); |
16421 | return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp, |
16422 | DAG.getConstant(0, dl, MVT::i8)); |
16423 | } |
16424 | // CF and OF aren't always set the way we want. Determine which |
16425 | // of these we need. |
16426 | bool NeedCF = false; |
16427 | bool NeedOF = false; |
16428 | switch (X86CC) { |
16429 | default: break; |
16430 | case X86::COND_A: case X86::COND_AE: |
16431 | case X86::COND_B: case X86::COND_BE: |
16432 | NeedCF = true; |
16433 | break; |
16434 | case X86::COND_G: case X86::COND_GE: |
16435 | case X86::COND_L: case X86::COND_LE: |
16436 | case X86::COND_O: case X86::COND_NO: { |
16437 | // Check if we really need to set the |
16438 | // Overflow flag. If NoSignedWrap is present |
16439 | // that is not actually needed. |
16440 | switch (Op->getOpcode()) { |
16441 | case ISD::ADD: |
16442 | case ISD::SUB: |
16443 | case ISD::MUL: |
16444 | case ISD::SHL: |
16445 | if (Op.getNode()->getFlags().hasNoSignedWrap()) |
16446 | break; |
16447 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
16448 | default: |
16449 | NeedOF = true; |
16450 | break; |
16451 | } |
16452 | break; |
16453 | } |
16454 | } |
16455 | // See if we can use the EFLAGS value from the operand instead of |
16456 | // doing a separate TEST. TEST always sets OF and CF to 0, so unless |
16457 | // we prove that the arithmetic won't overflow, we can't use OF or CF. |
16458 | if (Op.getResNo() != 0 || NeedOF || NeedCF) { |
16459 | // Emit KTEST for bit vectors |
16460 | if (auto Node = EmitKTEST(Op, DAG, Subtarget)) |
16461 | return Node; |
16462 | // Emit a CMP with 0, which is the TEST pattern. |
16463 | return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op, |
16464 | DAG.getConstant(0, dl, Op.getValueType())); |
16465 | } |
16466 | unsigned Opcode = 0; |
16467 | unsigned NumOperands = 0; |
16468 | |
16469 | // Truncate operations may prevent the merge of the SETCC instruction |
16470 | // and the arithmetic instruction before it. Attempt to truncate the operands |
16471 | // of the arithmetic instruction and use a reduced bit-width instruction. |
16472 | bool NeedTruncation = false; |
16473 | SDValue ArithOp = Op; |
16474 | if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) { |
16475 | SDValue Arith = Op->getOperand(0); |
16476 | // Both the trunc and the arithmetic op need to have one user each. |
16477 | if (Arith->hasOneUse()) |
16478 | switch (Arith.getOpcode()) { |
16479 | default: break; |
16480 | case ISD::ADD: |
16481 | case ISD::SUB: |
16482 | case ISD::AND: |
16483 | case ISD::OR: |
16484 | case ISD::XOR: { |
16485 | NeedTruncation = true; |
16486 | ArithOp = Arith; |
16487 | } |
16488 | } |
16489 | } |
16490 | |
16491 | // Sometimes flags can be set either with an AND or with an SRL/SHL |
16492 | // instruction. SRL/SHL variant should be preferred for masks longer than this |
16493 | // number of bits. |
16494 | const int ShiftToAndMaxMaskWidth = 32; |
16495 | const bool ZeroCheck = (X86CC == X86::COND_E || X86CC == X86::COND_NE); |
16496 | |
16497 | // NOTICE: In the code below we use ArithOp to hold the arithmetic operation |
16498 | // which may be the result of a CAST. We use the variable 'Op', which is the |
16499 | // non-casted variable when we check for possible users. |
16500 | switch (ArithOp.getOpcode()) { |
16501 | case ISD::ADD: |
16502 | // Due to an isel shortcoming, be conservative if this add is likely to be |
16503 | // selected as part of a load-modify-store instruction. When the root node |
16504 | // in a match is a store, isel doesn't know how to remap non-chain non-flag |
16505 | // uses of other nodes in the match, such as the ADD in this case. This |
16506 | // leads to the ADD being left around and reselected, with the result being |
16507 | // two adds in the output. Alas, even if none our users are stores, that |
16508 | // doesn't prove we're O.K. Ergo, if we have any parents that aren't |
16509 | // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require |
16510 | // climbing the DAG back to the root, and it doesn't seem to be worth the |
16511 | // effort. |
16512 | for (SDNode::use_iterator UI = Op.getNode()->use_begin(), |
16513 | UE = Op.getNode()->use_end(); UI != UE; ++UI) |
16514 | if (UI->getOpcode() != ISD::CopyToReg && |
16515 | UI->getOpcode() != ISD::SETCC && |
16516 | UI->getOpcode() != ISD::STORE) |
16517 | goto default_case; |
16518 | |
16519 | if (ConstantSDNode *C = |
16520 | dyn_cast<ConstantSDNode>(ArithOp.getOperand(1))) { |
16521 | // An add of one will be selected as an INC. |
16522 | if (C->isOne() && !Subtarget.slowIncDec()) { |
16523 | Opcode = X86ISD::INC; |
16524 | NumOperands = 1; |
16525 | break; |
16526 | } |
16527 | |
16528 | // An add of negative one (subtract of one) will be selected as a DEC. |
16529 | if (C->isAllOnesValue() && !Subtarget.slowIncDec()) { |
16530 | Opcode = X86ISD::DEC; |
16531 | NumOperands = 1; |
16532 | break; |
16533 | } |
16534 | } |
16535 | |
16536 | // Otherwise use a regular EFLAGS-setting add. |
16537 | Opcode = X86ISD::ADD; |
16538 | NumOperands = 2; |
16539 | break; |
16540 | case ISD::SHL: |
16541 | case ISD::SRL: |
16542 | // If we have a constant logical shift that's only used in a comparison |
16543 | // against zero turn it into an equivalent AND. This allows turning it into |
16544 | // a TEST instruction later. |
16545 | if (ZeroCheck && Op->hasOneUse() && |
16546 | isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) { |
16547 | EVT VT = Op.getValueType(); |
16548 | unsigned BitWidth = VT.getSizeInBits(); |
16549 | unsigned ShAmt = Op->getConstantOperandVal(1); |
16550 | if (ShAmt >= BitWidth) // Avoid undefined shifts. |
16551 | break; |
16552 | APInt Mask = ArithOp.getOpcode() == ISD::SRL |
16553 | ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt) |
16554 | : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt); |
16555 | if (!Mask.isSignedIntN(ShiftToAndMaxMaskWidth)) |
16556 | break; |
16557 | Op = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0), |
16558 | DAG.getConstant(Mask, dl, VT)); |
16559 | } |
16560 | break; |
16561 | |
16562 | case ISD::AND: |
16563 | // If the primary 'and' result isn't used, don't bother using X86ISD::AND, |
16564 | // because a TEST instruction will be better. However, AND should be |
16565 | // preferred if the instruction can be combined into ANDN. |
16566 | if (!hasNonFlagsUse(Op)) { |
16567 | SDValue Op0 = ArithOp->getOperand(0); |
16568 | SDValue Op1 = ArithOp->getOperand(1); |
16569 | EVT VT = ArithOp.getValueType(); |
16570 | bool isAndn = isBitwiseNot(Op0) || isBitwiseNot(Op1); |
16571 | bool isLegalAndnType = VT == MVT::i32 || VT == MVT::i64; |
16572 | bool isProperAndn = isAndn && isLegalAndnType && Subtarget.hasBMI(); |
16573 | |
16574 | // If we cannot select an ANDN instruction, check if we can replace |
16575 | // AND+IMM64 with a shift before giving up. This is possible for masks |
16576 | // like 0xFF000000 or 0x00FFFFFF and if we care only about the zero flag. |
16577 | if (!isProperAndn) { |
16578 | if (!ZeroCheck) |
16579 | break; |
16580 | |
16581 | assert(!isa<ConstantSDNode>(Op0) && "AND node isn't canonicalized")((!isa<ConstantSDNode>(Op0) && "AND node isn't canonicalized" ) ? static_cast<void> (0) : __assert_fail ("!isa<ConstantSDNode>(Op0) && \"AND node isn't canonicalized\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16581, __PRETTY_FUNCTION__)); |
16582 | auto *CN = dyn_cast<ConstantSDNode>(Op1); |
16583 | if (!CN) |
16584 | break; |
16585 | |
16586 | const APInt &Mask = CN->getAPIntValue(); |
16587 | if (Mask.isSignedIntN(ShiftToAndMaxMaskWidth)) |
16588 | break; // Prefer TEST instruction. |
16589 | |
16590 | unsigned BitWidth = Mask.getBitWidth(); |
16591 | unsigned LeadingOnes = Mask.countLeadingOnes(); |
16592 | unsigned TrailingZeros = Mask.countTrailingZeros(); |
16593 | |
16594 | if (LeadingOnes + TrailingZeros == BitWidth) { |
16595 | assert(TrailingZeros < VT.getSizeInBits() &&((TrailingZeros < VT.getSizeInBits() && "Shift amount should be less than the type width" ) ? static_cast<void> (0) : __assert_fail ("TrailingZeros < VT.getSizeInBits() && \"Shift amount should be less than the type width\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16596, __PRETTY_FUNCTION__)) |
16596 | "Shift amount should be less than the type width")((TrailingZeros < VT.getSizeInBits() && "Shift amount should be less than the type width" ) ? static_cast<void> (0) : __assert_fail ("TrailingZeros < VT.getSizeInBits() && \"Shift amount should be less than the type width\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16596, __PRETTY_FUNCTION__)); |
16597 | MVT ShTy = getScalarShiftAmountTy(DAG.getDataLayout(), VT); |
16598 | SDValue ShAmt = DAG.getConstant(TrailingZeros, dl, ShTy); |
16599 | Op = DAG.getNode(ISD::SRL, dl, VT, Op0, ShAmt); |
16600 | break; |
16601 | } |
16602 | |
16603 | unsigned LeadingZeros = Mask.countLeadingZeros(); |
16604 | unsigned TrailingOnes = Mask.countTrailingOnes(); |
16605 | |
16606 | if (LeadingZeros + TrailingOnes == BitWidth) { |
16607 | assert(LeadingZeros < VT.getSizeInBits() &&((LeadingZeros < VT.getSizeInBits() && "Shift amount should be less than the type width" ) ? static_cast<void> (0) : __assert_fail ("LeadingZeros < VT.getSizeInBits() && \"Shift amount should be less than the type width\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16608, __PRETTY_FUNCTION__)) |
16608 | "Shift amount should be less than the type width")((LeadingZeros < VT.getSizeInBits() && "Shift amount should be less than the type width" ) ? static_cast<void> (0) : __assert_fail ("LeadingZeros < VT.getSizeInBits() && \"Shift amount should be less than the type width\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16608, __PRETTY_FUNCTION__)); |
16609 | MVT ShTy = getScalarShiftAmountTy(DAG.getDataLayout(), VT); |
16610 | SDValue ShAmt = DAG.getConstant(LeadingZeros, dl, ShTy); |
16611 | Op = DAG.getNode(ISD::SHL, dl, VT, Op0, ShAmt); |
16612 | break; |
16613 | } |
16614 | |
16615 | break; |
16616 | } |
16617 | } |
16618 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
16619 | case ISD::SUB: |
16620 | case ISD::OR: |
16621 | case ISD::XOR: |
16622 | // Due to the ISEL shortcoming noted above, be conservative if this op is |
16623 | // likely to be selected as part of a load-modify-store instruction. |
16624 | for (SDNode::use_iterator UI = Op.getNode()->use_begin(), |
16625 | UE = Op.getNode()->use_end(); UI != UE; ++UI) |
16626 | if (UI->getOpcode() == ISD::STORE) |
16627 | goto default_case; |
16628 | |
16629 | // Otherwise use a regular EFLAGS-setting instruction. |
16630 | switch (ArithOp.getOpcode()) { |
16631 | default: llvm_unreachable("unexpected operator!")::llvm::llvm_unreachable_internal("unexpected operator!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16631); |
16632 | case ISD::SUB: Opcode = X86ISD::SUB; break; |
16633 | case ISD::XOR: Opcode = X86ISD::XOR; break; |
16634 | case ISD::AND: Opcode = X86ISD::AND; break; |
16635 | case ISD::OR: { |
16636 | if (!NeedTruncation && ZeroCheck) { |
16637 | if (SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG)) |
16638 | return EFLAGS; |
16639 | } |
16640 | Opcode = X86ISD::OR; |
16641 | break; |
16642 | } |
16643 | } |
16644 | |
16645 | NumOperands = 2; |
16646 | break; |
16647 | case X86ISD::ADD: |
16648 | case X86ISD::SUB: |
16649 | case X86ISD::INC: |
16650 | case X86ISD::DEC: |
16651 | case X86ISD::OR: |
16652 | case X86ISD::XOR: |
16653 | case X86ISD::AND: |
16654 | return SDValue(Op.getNode(), 1); |
16655 | default: |
16656 | default_case: |
16657 | break; |
16658 | } |
16659 | |
16660 | // If we found that truncation is beneficial, perform the truncation and |
16661 | // update 'Op'. |
16662 | if (NeedTruncation) { |
16663 | EVT VT = Op.getValueType(); |
16664 | SDValue WideVal = Op->getOperand(0); |
16665 | EVT WideVT = WideVal.getValueType(); |
16666 | unsigned ConvertedOp = 0; |
16667 | // Use a target machine opcode to prevent further DAGCombine |
16668 | // optimizations that may separate the arithmetic operations |
16669 | // from the setcc node. |
16670 | switch (WideVal.getOpcode()) { |
16671 | default: break; |
16672 | case ISD::ADD: ConvertedOp = X86ISD::ADD; break; |
16673 | case ISD::SUB: ConvertedOp = X86ISD::SUB; break; |
16674 | case ISD::AND: ConvertedOp = X86ISD::AND; break; |
16675 | case ISD::OR: ConvertedOp = X86ISD::OR; break; |
16676 | case ISD::XOR: ConvertedOp = X86ISD::XOR; break; |
16677 | } |
16678 | |
16679 | if (ConvertedOp) { |
16680 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
16681 | if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) { |
16682 | SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0)); |
16683 | SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1)); |
16684 | Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1); |
16685 | } |
16686 | } |
16687 | } |
16688 | |
16689 | if (Opcode == 0) { |
16690 | // Emit KTEST for bit vectors |
16691 | if (auto Node = EmitKTEST(Op, DAG, Subtarget)) |
16692 | return Node; |
16693 | |
16694 | // Emit a CMP with 0, which is the TEST pattern. |
16695 | return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op, |
16696 | DAG.getConstant(0, dl, Op.getValueType())); |
16697 | } |
16698 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32); |
16699 | SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands); |
16700 | |
16701 | SDValue New = DAG.getNode(Opcode, dl, VTs, Ops); |
16702 | DAG.ReplaceAllUsesWith(Op, New); |
16703 | return SDValue(New.getNode(), 1); |
16704 | } |
16705 | |
16706 | /// Emit nodes that will be selected as "cmp Op0,Op1", or something |
16707 | /// equivalent. |
16708 | SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, |
16709 | const SDLoc &dl, SelectionDAG &DAG) const { |
16710 | if (isNullConstant(Op1)) |
16711 | return EmitTest(Op0, X86CC, dl, DAG); |
16712 | |
16713 | assert(!(isa<ConstantSDNode>(Op1) && Op0.getValueType() == MVT::i1) &&((!(isa<ConstantSDNode>(Op1) && Op0.getValueType () == MVT::i1) && "Unexpected comparison operation for MVT::i1 operands" ) ? static_cast<void> (0) : __assert_fail ("!(isa<ConstantSDNode>(Op1) && Op0.getValueType() == MVT::i1) && \"Unexpected comparison operation for MVT::i1 operands\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16714, __PRETTY_FUNCTION__)) |
16714 | "Unexpected comparison operation for MVT::i1 operands")((!(isa<ConstantSDNode>(Op1) && Op0.getValueType () == MVT::i1) && "Unexpected comparison operation for MVT::i1 operands" ) ? static_cast<void> (0) : __assert_fail ("!(isa<ConstantSDNode>(Op1) && Op0.getValueType() == MVT::i1) && \"Unexpected comparison operation for MVT::i1 operands\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16714, __PRETTY_FUNCTION__)); |
16715 | |
16716 | if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 || |
16717 | Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) { |
16718 | // Only promote the compare up to I32 if it is a 16 bit operation |
16719 | // with an immediate. 16 bit immediates are to be avoided. |
16720 | if ((Op0.getValueType() == MVT::i16 && |
16721 | (isa<ConstantSDNode>(Op0) || isa<ConstantSDNode>(Op1))) && |
16722 | !DAG.getMachineFunction().getFunction()->optForMinSize() && |
16723 | !Subtarget.isAtom()) { |
16724 | unsigned ExtendOp = |
16725 | isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND; |
16726 | Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0); |
16727 | Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1); |
16728 | } |
16729 | // Use SUB instead of CMP to enable CSE between SUB and CMP. |
16730 | SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32); |
16731 | SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs, |
16732 | Op0, Op1); |
16733 | return SDValue(Sub.getNode(), 1); |
16734 | } |
16735 | return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1); |
16736 | } |
16737 | |
16738 | /// Convert a comparison if required by the subtarget. |
16739 | SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp, |
16740 | SelectionDAG &DAG) const { |
16741 | // If the subtarget does not support the FUCOMI instruction, floating-point |
16742 | // comparisons have to be converted. |
16743 | if (Subtarget.hasCMov() || |
16744 | Cmp.getOpcode() != X86ISD::CMP || |
16745 | !Cmp.getOperand(0).getValueType().isFloatingPoint() || |
16746 | !Cmp.getOperand(1).getValueType().isFloatingPoint()) |
16747 | return Cmp; |
16748 | |
16749 | // The instruction selector will select an FUCOM instruction instead of |
16750 | // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence |
16751 | // build an SDNode sequence that transfers the result from FPSW into EFLAGS: |
16752 | // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8)))) |
16753 | SDLoc dl(Cmp); |
16754 | SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp); |
16755 | SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW); |
16756 | SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW, |
16757 | DAG.getConstant(8, dl, MVT::i8)); |
16758 | SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl); |
16759 | |
16760 | // Some 64-bit targets lack SAHF support, but they do support FCOMI. |
16761 | assert(Subtarget.hasLAHFSAHF() && "Target doesn't support SAHF or FCOMI?")((Subtarget.hasLAHFSAHF() && "Target doesn't support SAHF or FCOMI?" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasLAHFSAHF() && \"Target doesn't support SAHF or FCOMI?\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16761, __PRETTY_FUNCTION__)); |
16762 | return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl); |
16763 | } |
16764 | |
16765 | /// Check if replacement of SQRT with RSQRT should be disabled. |
16766 | bool X86TargetLowering::isFsqrtCheap(SDValue Op, SelectionDAG &DAG) const { |
16767 | EVT VT = Op.getValueType(); |
16768 | |
16769 | // We never want to use both SQRT and RSQRT instructions for the same input. |
16770 | if (DAG.getNodeIfExists(X86ISD::FRSQRT, DAG.getVTList(VT), Op)) |
16771 | return false; |
16772 | |
16773 | if (VT.isVector()) |
16774 | return Subtarget.hasFastVectorFSQRT(); |
16775 | return Subtarget.hasFastScalarFSQRT(); |
16776 | } |
16777 | |
16778 | /// The minimum architected relative accuracy is 2^-12. We need one |
16779 | /// Newton-Raphson step to have a good float result (24 bits of precision). |
16780 | SDValue X86TargetLowering::getSqrtEstimate(SDValue Op, |
16781 | SelectionDAG &DAG, int Enabled, |
16782 | int &RefinementSteps, |
16783 | bool &UseOneConstNR, |
16784 | bool Reciprocal) const { |
16785 | EVT VT = Op.getValueType(); |
16786 | |
16787 | // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps. |
16788 | // TODO: Add support for AVX512 (v16f32). |
16789 | // It is likely not profitable to do this for f64 because a double-precision |
16790 | // rsqrt estimate with refinement on x86 prior to FMA requires at least 16 |
16791 | // instructions: convert to single, rsqrtss, convert back to double, refine |
16792 | // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA |
16793 | // along with FMA, this could be a throughput win. |
16794 | if ((VT == MVT::f32 && Subtarget.hasSSE1()) || |
16795 | (VT == MVT::v4f32 && Subtarget.hasSSE1()) || |
16796 | (VT == MVT::v8f32 && Subtarget.hasAVX())) { |
16797 | if (RefinementSteps == ReciprocalEstimate::Unspecified) |
16798 | RefinementSteps = 1; |
16799 | |
16800 | UseOneConstNR = false; |
16801 | return DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op); |
16802 | } |
16803 | return SDValue(); |
16804 | } |
16805 | |
16806 | /// The minimum architected relative accuracy is 2^-12. We need one |
16807 | /// Newton-Raphson step to have a good float result (24 bits of precision). |
16808 | SDValue X86TargetLowering::getRecipEstimate(SDValue Op, SelectionDAG &DAG, |
16809 | int Enabled, |
16810 | int &RefinementSteps) const { |
16811 | EVT VT = Op.getValueType(); |
16812 | |
16813 | // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps. |
16814 | // TODO: Add support for AVX512 (v16f32). |
16815 | // It is likely not profitable to do this for f64 because a double-precision |
16816 | // reciprocal estimate with refinement on x86 prior to FMA requires |
16817 | // 15 instructions: convert to single, rcpss, convert back to double, refine |
16818 | // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA |
16819 | // along with FMA, this could be a throughput win. |
16820 | |
16821 | if ((VT == MVT::f32 && Subtarget.hasSSE1()) || |
16822 | (VT == MVT::v4f32 && Subtarget.hasSSE1()) || |
16823 | (VT == MVT::v8f32 && Subtarget.hasAVX())) { |
16824 | // Enable estimate codegen with 1 refinement step for vector division. |
16825 | // Scalar division estimates are disabled because they break too much |
16826 | // real-world code. These defaults are intended to match GCC behavior. |
16827 | if (VT == MVT::f32 && Enabled == ReciprocalEstimate::Unspecified) |
16828 | return SDValue(); |
16829 | |
16830 | if (RefinementSteps == ReciprocalEstimate::Unspecified) |
16831 | RefinementSteps = 1; |
16832 | |
16833 | return DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op); |
16834 | } |
16835 | return SDValue(); |
16836 | } |
16837 | |
16838 | /// If we have at least two divisions that use the same divisor, convert to |
16839 | /// multiplication by a reciprocal. This may need to be adjusted for a given |
16840 | /// CPU if a division's cost is not at least twice the cost of a multiplication. |
16841 | /// This is because we still need one division to calculate the reciprocal and |
16842 | /// then we need two multiplies by that reciprocal as replacements for the |
16843 | /// original divisions. |
16844 | unsigned X86TargetLowering::combineRepeatedFPDivisors() const { |
16845 | return 2; |
16846 | } |
16847 | |
16848 | /// Helper for creating a X86ISD::SETCC node. |
16849 | static SDValue getSETCC(X86::CondCode Cond, SDValue EFLAGS, const SDLoc &dl, |
16850 | SelectionDAG &DAG) { |
16851 | return DAG.getNode(X86ISD::SETCC, dl, MVT::i8, |
16852 | DAG.getConstant(Cond, dl, MVT::i8), EFLAGS); |
16853 | } |
16854 | |
16855 | /// Create a BT (Bit Test) node - Test bit \p BitNo in \p Src and set condition |
16856 | /// according to equal/not-equal condition code \p CC. |
16857 | static SDValue getBitTestCondition(SDValue Src, SDValue BitNo, ISD::CondCode CC, |
16858 | const SDLoc &dl, SelectionDAG &DAG) { |
16859 | // If Src is i8, promote it to i32 with any_extend. There is no i8 BT |
16860 | // instruction. Since the shift amount is in-range-or-undefined, we know |
16861 | // that doing a bittest on the i32 value is ok. We extend to i32 because |
16862 | // the encoding for the i16 version is larger than the i32 version. |
16863 | // Also promote i16 to i32 for performance / code size reason. |
16864 | if (Src.getValueType() == MVT::i8 || Src.getValueType() == MVT::i16) |
16865 | Src = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Src); |
16866 | |
16867 | // See if we can use the 32-bit instruction instead of the 64-bit one for a |
16868 | // shorter encoding. Since the former takes the modulo 32 of BitNo and the |
16869 | // latter takes the modulo 64, this is only valid if the 5th bit of BitNo is |
16870 | // known to be zero. |
16871 | if (Src.getValueType() == MVT::i64 && |
16872 | DAG.MaskedValueIsZero(BitNo, APInt(BitNo.getValueSizeInBits(), 32))) |
16873 | Src = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Src); |
16874 | |
16875 | // If the operand types disagree, extend the shift amount to match. Since |
16876 | // BT ignores high bits (like shifts) we can use anyextend. |
16877 | if (Src.getValueType() != BitNo.getValueType()) |
16878 | BitNo = DAG.getNode(ISD::ANY_EXTEND, dl, Src.getValueType(), BitNo); |
16879 | |
16880 | SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, Src, BitNo); |
16881 | X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B; |
16882 | return getSETCC(Cond, BT, dl , DAG); |
16883 | } |
16884 | |
16885 | /// Result of 'and' is compared against zero. Change to a BT node if possible. |
16886 | static SDValue LowerAndToBT(SDValue And, ISD::CondCode CC, |
16887 | const SDLoc &dl, SelectionDAG &DAG) { |
16888 | SDValue Op0 = And.getOperand(0); |
16889 | SDValue Op1 = And.getOperand(1); |
16890 | if (Op0.getOpcode() == ISD::TRUNCATE) |
16891 | Op0 = Op0.getOperand(0); |
16892 | if (Op1.getOpcode() == ISD::TRUNCATE) |
16893 | Op1 = Op1.getOperand(0); |
16894 | |
16895 | SDValue LHS, RHS; |
16896 | if (Op1.getOpcode() == ISD::SHL) |
16897 | std::swap(Op0, Op1); |
16898 | if (Op0.getOpcode() == ISD::SHL) { |
16899 | if (isOneConstant(Op0.getOperand(0))) { |
16900 | // If we looked past a truncate, check that it's only truncating away |
16901 | // known zeros. |
16902 | unsigned BitWidth = Op0.getValueSizeInBits(); |
16903 | unsigned AndBitWidth = And.getValueSizeInBits(); |
16904 | if (BitWidth > AndBitWidth) { |
16905 | KnownBits Known; |
16906 | DAG.computeKnownBits(Op0, Known); |
16907 | if (Known.countMinLeadingZeros() < BitWidth - AndBitWidth) |
16908 | return SDValue(); |
16909 | } |
16910 | LHS = Op1; |
16911 | RHS = Op0.getOperand(1); |
16912 | } |
16913 | } else if (Op1.getOpcode() == ISD::Constant) { |
16914 | ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1); |
16915 | uint64_t AndRHSVal = AndRHS->getZExtValue(); |
16916 | SDValue AndLHS = Op0; |
16917 | |
16918 | if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) { |
16919 | LHS = AndLHS.getOperand(0); |
16920 | RHS = AndLHS.getOperand(1); |
16921 | } |
16922 | |
16923 | // Use BT if the immediate can't be encoded in a TEST instruction. |
16924 | if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) { |
16925 | LHS = AndLHS; |
16926 | RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType()); |
16927 | } |
16928 | } |
16929 | |
16930 | if (LHS.getNode()) |
16931 | return getBitTestCondition(LHS, RHS, CC, dl, DAG); |
16932 | |
16933 | return SDValue(); |
16934 | } |
16935 | |
16936 | // Convert (truncate (srl X, N) to i1) to (bt X, N) |
16937 | static SDValue LowerTruncateToBT(SDValue Op, ISD::CondCode CC, |
16938 | const SDLoc &dl, SelectionDAG &DAG) { |
16939 | |
16940 | assert(Op.getOpcode() == ISD::TRUNCATE && Op.getValueType() == MVT::i1 &&((Op.getOpcode() == ISD::TRUNCATE && Op.getValueType( ) == MVT::i1 && "Expected TRUNCATE to i1 node") ? static_cast <void> (0) : __assert_fail ("Op.getOpcode() == ISD::TRUNCATE && Op.getValueType() == MVT::i1 && \"Expected TRUNCATE to i1 node\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16941, __PRETTY_FUNCTION__)) |
16941 | "Expected TRUNCATE to i1 node")((Op.getOpcode() == ISD::TRUNCATE && Op.getValueType( ) == MVT::i1 && "Expected TRUNCATE to i1 node") ? static_cast <void> (0) : __assert_fail ("Op.getOpcode() == ISD::TRUNCATE && Op.getValueType() == MVT::i1 && \"Expected TRUNCATE to i1 node\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16941, __PRETTY_FUNCTION__)); |
16942 | |
16943 | if (Op.getOperand(0).getOpcode() != ISD::SRL) |
16944 | return SDValue(); |
16945 | |
16946 | SDValue ShiftRight = Op.getOperand(0); |
16947 | return getBitTestCondition(ShiftRight.getOperand(0), ShiftRight.getOperand(1), |
16948 | CC, dl, DAG); |
16949 | } |
16950 | |
16951 | /// Result of 'and' or 'trunc to i1' is compared against zero. |
16952 | /// Change to a BT node if possible. |
16953 | SDValue X86TargetLowering::LowerToBT(SDValue Op, ISD::CondCode CC, |
16954 | const SDLoc &dl, SelectionDAG &DAG) const { |
16955 | if (Op.getOpcode() == ISD::AND) |
16956 | return LowerAndToBT(Op, CC, dl, DAG); |
16957 | if (Op.getOpcode() == ISD::TRUNCATE && Op.getValueType() == MVT::i1) |
16958 | return LowerTruncateToBT(Op, CC, dl, DAG); |
16959 | return SDValue(); |
16960 | } |
16961 | |
16962 | /// Turns an ISD::CondCode into a value suitable for SSE floating-point mask |
16963 | /// CMPs. |
16964 | static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0, |
16965 | SDValue &Op1) { |
16966 | unsigned SSECC; |
16967 | bool Swap = false; |
16968 | |
16969 | // SSE Condition code mapping: |
16970 | // 0 - EQ |
16971 | // 1 - LT |
16972 | // 2 - LE |
16973 | // 3 - UNORD |
16974 | // 4 - NEQ |
16975 | // 5 - NLT |
16976 | // 6 - NLE |
16977 | // 7 - ORD |
16978 | switch (SetCCOpcode) { |
16979 | default: llvm_unreachable("Unexpected SETCC condition")::llvm::llvm_unreachable_internal("Unexpected SETCC condition" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 16979); |
16980 | case ISD::SETOEQ: |
16981 | case ISD::SETEQ: SSECC = 0; break; |
16982 | case ISD::SETOGT: |
16983 | case ISD::SETGT: Swap = true; LLVM_FALLTHROUGH[[clang::fallthrough]]; |
16984 | case ISD::SETLT: |
16985 | case ISD::SETOLT: SSECC = 1; break; |
16986 | case ISD::SETOGE: |
16987 | case ISD::SETGE: Swap = true; LLVM_FALLTHROUGH[[clang::fallthrough]]; |
16988 | case ISD::SETLE: |
16989 | case ISD::SETOLE: SSECC = 2; break; |
16990 | case ISD::SETUO: SSECC = 3; break; |
16991 | case ISD::SETUNE: |
16992 | case ISD::SETNE: SSECC = 4; break; |
16993 | case ISD::SETULE: Swap = true; LLVM_FALLTHROUGH[[clang::fallthrough]]; |
16994 | case ISD::SETUGE: SSECC = 5; break; |
16995 | case ISD::SETULT: Swap = true; LLVM_FALLTHROUGH[[clang::fallthrough]]; |
16996 | case ISD::SETUGT: SSECC = 6; break; |
16997 | case ISD::SETO: SSECC = 7; break; |
16998 | case ISD::SETUEQ: |
16999 | case ISD::SETONE: SSECC = 8; break; |
17000 | } |
17001 | if (Swap) |
17002 | std::swap(Op0, Op1); |
17003 | |
17004 | return SSECC; |
17005 | } |
17006 | |
17007 | /// Break a VSETCC 256-bit integer VSETCC into two new 128 ones and then |
17008 | /// concatenate the result back. |
17009 | static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) { |
17010 | MVT VT = Op.getSimpleValueType(); |
17011 | |
17012 | assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&((VT.is256BitVector() && Op.getOpcode() == ISD::SETCC && "Unsupported value type for operation") ? static_cast <void> (0) : __assert_fail ("VT.is256BitVector() && Op.getOpcode() == ISD::SETCC && \"Unsupported value type for operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17013, __PRETTY_FUNCTION__)) |
17013 | "Unsupported value type for operation")((VT.is256BitVector() && Op.getOpcode() == ISD::SETCC && "Unsupported value type for operation") ? static_cast <void> (0) : __assert_fail ("VT.is256BitVector() && Op.getOpcode() == ISD::SETCC && \"Unsupported value type for operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17013, __PRETTY_FUNCTION__)); |
17014 | |
17015 | unsigned NumElems = VT.getVectorNumElements(); |
17016 | SDLoc dl(Op); |
17017 | SDValue CC = Op.getOperand(2); |
17018 | |
17019 | // Extract the LHS vectors |
17020 | SDValue LHS = Op.getOperand(0); |
17021 | SDValue LHS1 = extract128BitVector(LHS, 0, DAG, dl); |
17022 | SDValue LHS2 = extract128BitVector(LHS, NumElems / 2, DAG, dl); |
17023 | |
17024 | // Extract the RHS vectors |
17025 | SDValue RHS = Op.getOperand(1); |
17026 | SDValue RHS1 = extract128BitVector(RHS, 0, DAG, dl); |
17027 | SDValue RHS2 = extract128BitVector(RHS, NumElems / 2, DAG, dl); |
17028 | |
17029 | // Issue the operation on the smaller types and concatenate the result back |
17030 | MVT EltVT = VT.getVectorElementType(); |
17031 | MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2); |
17032 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, |
17033 | DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC), |
17034 | DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC)); |
17035 | } |
17036 | |
17037 | static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) { |
17038 | SDValue Op0 = Op.getOperand(0); |
17039 | SDValue Op1 = Op.getOperand(1); |
17040 | SDValue CC = Op.getOperand(2); |
17041 | MVT VT = Op.getSimpleValueType(); |
17042 | SDLoc dl(Op); |
17043 | |
17044 | assert(Op0.getSimpleValueType().getVectorElementType() == MVT::i1 &&((Op0.getSimpleValueType().getVectorElementType() == MVT::i1 && "Unexpected type for boolean compare operation") ? static_cast <void> (0) : __assert_fail ("Op0.getSimpleValueType().getVectorElementType() == MVT::i1 && \"Unexpected type for boolean compare operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17045, __PRETTY_FUNCTION__)) |
17045 | "Unexpected type for boolean compare operation")((Op0.getSimpleValueType().getVectorElementType() == MVT::i1 && "Unexpected type for boolean compare operation") ? static_cast <void> (0) : __assert_fail ("Op0.getSimpleValueType().getVectorElementType() == MVT::i1 && \"Unexpected type for boolean compare operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17045, __PRETTY_FUNCTION__)); |
17046 | ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get(); |
17047 | SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0, |
17048 | DAG.getConstant(-1, dl, VT)); |
17049 | SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1, |
17050 | DAG.getConstant(-1, dl, VT)); |
17051 | switch (SetCCOpcode) { |
17052 | default: llvm_unreachable("Unexpected SETCC condition")::llvm::llvm_unreachable_internal("Unexpected SETCC condition" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17052); |
17053 | case ISD::SETEQ: |
17054 | // (x == y) -> ~(x ^ y) |
17055 | return DAG.getNode(ISD::XOR, dl, VT, |
17056 | DAG.getNode(ISD::XOR, dl, VT, Op0, Op1), |
17057 | DAG.getConstant(-1, dl, VT)); |
17058 | case ISD::SETNE: |
17059 | // (x != y) -> (x ^ y) |
17060 | return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1); |
17061 | case ISD::SETUGT: |
17062 | case ISD::SETGT: |
17063 | // (x > y) -> (x & ~y) |
17064 | return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1); |
17065 | case ISD::SETULT: |
17066 | case ISD::SETLT: |
17067 | // (x < y) -> (~x & y) |
17068 | return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1); |
17069 | case ISD::SETULE: |
17070 | case ISD::SETLE: |
17071 | // (x <= y) -> (~x | y) |
17072 | return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1); |
17073 | case ISD::SETUGE: |
17074 | case ISD::SETGE: |
17075 | // (x >=y) -> (x | ~y) |
17076 | return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1); |
17077 | } |
17078 | } |
17079 | |
17080 | static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) { |
17081 | |
17082 | SDValue Op0 = Op.getOperand(0); |
17083 | SDValue Op1 = Op.getOperand(1); |
17084 | SDValue CC = Op.getOperand(2); |
17085 | MVT VT = Op.getSimpleValueType(); |
17086 | SDLoc dl(Op); |
17087 | |
17088 | assert(VT.getVectorElementType() == MVT::i1 &&((VT.getVectorElementType() == MVT::i1 && "Cannot set masked compare for this operation" ) ? static_cast<void> (0) : __assert_fail ("VT.getVectorElementType() == MVT::i1 && \"Cannot set masked compare for this operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17089, __PRETTY_FUNCTION__)) |
17089 | "Cannot set masked compare for this operation")((VT.getVectorElementType() == MVT::i1 && "Cannot set masked compare for this operation" ) ? static_cast<void> (0) : __assert_fail ("VT.getVectorElementType() == MVT::i1 && \"Cannot set masked compare for this operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17089, __PRETTY_FUNCTION__)); |
17090 | |
17091 | ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get(); |
17092 | unsigned Opc = 0; |
17093 | bool Unsigned = false; |
17094 | bool Swap = false; |
17095 | unsigned SSECC; |
17096 | switch (SetCCOpcode) { |
17097 | default: llvm_unreachable("Unexpected SETCC condition")::llvm::llvm_unreachable_internal("Unexpected SETCC condition" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17097); |
17098 | case ISD::SETNE: SSECC = 4; break; |
17099 | case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break; |
17100 | case ISD::SETUGT: SSECC = 6; Unsigned = true; break; |
17101 | case ISD::SETLT: Swap = true; LLVM_FALLTHROUGH[[clang::fallthrough]]; |
17102 | case ISD::SETGT: Opc = X86ISD::PCMPGTM; break; |
17103 | case ISD::SETULT: SSECC = 1; Unsigned = true; break; |
17104 | case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT |
17105 | case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap |
17106 | case ISD::SETULE: Unsigned = true; LLVM_FALLTHROUGH[[clang::fallthrough]]; |
17107 | case ISD::SETLE: SSECC = 2; break; |
17108 | } |
17109 | |
17110 | if (Swap) |
17111 | std::swap(Op0, Op1); |
17112 | if (Opc) |
17113 | return DAG.getNode(Opc, dl, VT, Op0, Op1); |
17114 | Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM; |
17115 | return DAG.getNode(Opc, dl, VT, Op0, Op1, |
17116 | DAG.getConstant(SSECC, dl, MVT::i8)); |
17117 | } |
17118 | |
17119 | /// \brief Try to turn a VSETULT into a VSETULE by modifying its second |
17120 | /// operand \p Op1. If non-trivial (for example because it's not constant) |
17121 | /// return an empty value. |
17122 | static SDValue ChangeVSETULTtoVSETULE(const SDLoc &dl, SDValue Op1, |
17123 | SelectionDAG &DAG) { |
17124 | BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode()); |
17125 | if (!BV) |
17126 | return SDValue(); |
17127 | |
17128 | MVT VT = Op1.getSimpleValueType(); |
17129 | MVT EVT = VT.getVectorElementType(); |
17130 | unsigned n = VT.getVectorNumElements(); |
17131 | SmallVector<SDValue, 8> ULTOp1; |
17132 | |
17133 | for (unsigned i = 0; i < n; ++i) { |
17134 | ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i)); |
17135 | if (!Elt || Elt->isOpaque() || Elt->getSimpleValueType(0) != EVT) |
17136 | return SDValue(); |
17137 | |
17138 | // Avoid underflow. |
17139 | APInt Val = Elt->getAPIntValue(); |
17140 | if (Val == 0) |
17141 | return SDValue(); |
17142 | |
17143 | ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT)); |
17144 | } |
17145 | |
17146 | return DAG.getBuildVector(VT, dl, ULTOp1); |
17147 | } |
17148 | |
17149 | static SDValue LowerVSETCC(SDValue Op, const X86Subtarget &Subtarget, |
17150 | SelectionDAG &DAG) { |
17151 | SDValue Op0 = Op.getOperand(0); |
17152 | SDValue Op1 = Op.getOperand(1); |
17153 | SDValue CC = Op.getOperand(2); |
17154 | MVT VT = Op.getSimpleValueType(); |
17155 | ISD::CondCode Cond = cast<CondCodeSDNode>(CC)->get(); |
17156 | bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint(); |
17157 | SDLoc dl(Op); |
17158 | |
17159 | if (isFP) { |
17160 | #ifndef NDEBUG |
17161 | MVT EltVT = Op0.getSimpleValueType().getVectorElementType(); |
17162 | assert(EltVT == MVT::f32 || EltVT == MVT::f64)((EltVT == MVT::f32 || EltVT == MVT::f64) ? static_cast<void > (0) : __assert_fail ("EltVT == MVT::f32 || EltVT == MVT::f64" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17162, __PRETTY_FUNCTION__)); |
17163 | #endif |
17164 | |
17165 | unsigned Opc; |
17166 | if (Subtarget.hasAVX512() && VT.getVectorElementType() == MVT::i1) { |
17167 | assert(VT.getVectorNumElements() <= 16)((VT.getVectorNumElements() <= 16) ? static_cast<void> (0) : __assert_fail ("VT.getVectorNumElements() <= 16", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17167, __PRETTY_FUNCTION__)); |
17168 | Opc = X86ISD::CMPM; |
17169 | } else { |
17170 | Opc = X86ISD::CMPP; |
17171 | // The SSE/AVX packed FP comparison nodes are defined with a |
17172 | // floating-point vector result that matches the operand type. This allows |
17173 | // them to work with an SSE1 target (integer vector types are not legal). |
17174 | VT = Op0.getSimpleValueType(); |
17175 | } |
17176 | |
17177 | // In the two cases not handled by SSE compare predicates (SETUEQ/SETONE), |
17178 | // emit two comparisons and a logic op to tie them together. |
17179 | // TODO: This can be avoided if Intel (and only Intel as of 2016) AVX is |
17180 | // available. |
17181 | SDValue Cmp; |
17182 | unsigned SSECC = translateX86FSETCC(Cond, Op0, Op1); |
17183 | if (SSECC == 8) { |
17184 | // LLVM predicate is SETUEQ or SETONE. |
17185 | unsigned CC0, CC1; |
17186 | unsigned CombineOpc; |
17187 | if (Cond == ISD::SETUEQ) { |
17188 | CC0 = 3; // UNORD |
17189 | CC1 = 0; // EQ |
17190 | CombineOpc = Opc == X86ISD::CMPP ? static_cast<unsigned>(X86ISD::FOR) : |
17191 | static_cast<unsigned>(ISD::OR); |
17192 | } else { |
17193 | assert(Cond == ISD::SETONE)((Cond == ISD::SETONE) ? static_cast<void> (0) : __assert_fail ("Cond == ISD::SETONE", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17193, __PRETTY_FUNCTION__)); |
17194 | CC0 = 7; // ORD |
17195 | CC1 = 4; // NEQ |
17196 | CombineOpc = Opc == X86ISD::CMPP ? static_cast<unsigned>(X86ISD::FAND) : |
17197 | static_cast<unsigned>(ISD::AND); |
17198 | } |
17199 | |
17200 | SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1, |
17201 | DAG.getConstant(CC0, dl, MVT::i8)); |
17202 | SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1, |
17203 | DAG.getConstant(CC1, dl, MVT::i8)); |
17204 | Cmp = DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1); |
17205 | } else { |
17206 | // Handle all other FP comparisons here. |
17207 | Cmp = DAG.getNode(Opc, dl, VT, Op0, Op1, |
17208 | DAG.getConstant(SSECC, dl, MVT::i8)); |
17209 | } |
17210 | |
17211 | // If this is SSE/AVX CMPP, bitcast the result back to integer to match the |
17212 | // result type of SETCC. The bitcast is expected to be optimized away |
17213 | // during combining/isel. |
17214 | if (Opc == X86ISD::CMPP) |
17215 | Cmp = DAG.getBitcast(Op.getSimpleValueType(), Cmp); |
17216 | |
17217 | return Cmp; |
17218 | } |
17219 | |
17220 | MVT VTOp0 = Op0.getSimpleValueType(); |
17221 | assert(VTOp0 == Op1.getSimpleValueType() &&((VTOp0 == Op1.getSimpleValueType() && "Expected operands with same type!" ) ? static_cast<void> (0) : __assert_fail ("VTOp0 == Op1.getSimpleValueType() && \"Expected operands with same type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17222, __PRETTY_FUNCTION__)) |
17222 | "Expected operands with same type!")((VTOp0 == Op1.getSimpleValueType() && "Expected operands with same type!" ) ? static_cast<void> (0) : __assert_fail ("VTOp0 == Op1.getSimpleValueType() && \"Expected operands with same type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17222, __PRETTY_FUNCTION__)); |
17223 | assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&((VT.getVectorNumElements() == VTOp0.getVectorNumElements() && "Invalid number of packed elements for source and destination!" ) ? static_cast<void> (0) : __assert_fail ("VT.getVectorNumElements() == VTOp0.getVectorNumElements() && \"Invalid number of packed elements for source and destination!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17224, __PRETTY_FUNCTION__)) |
17224 | "Invalid number of packed elements for source and destination!")((VT.getVectorNumElements() == VTOp0.getVectorNumElements() && "Invalid number of packed elements for source and destination!" ) ? static_cast<void> (0) : __assert_fail ("VT.getVectorNumElements() == VTOp0.getVectorNumElements() && \"Invalid number of packed elements for source and destination!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17224, __PRETTY_FUNCTION__)); |
17225 | |
17226 | if (VT.is128BitVector() && VTOp0.is256BitVector()) { |
17227 | // On non-AVX512 targets, a vector of MVT::i1 is promoted by the type |
17228 | // legalizer to a wider vector type. In the case of 'vsetcc' nodes, the |
17229 | // legalizer firstly checks if the first operand in input to the setcc has |
17230 | // a legal type. If so, then it promotes the return type to that same type. |
17231 | // Otherwise, the return type is promoted to the 'next legal type' which, |
17232 | // for a vector of MVT::i1 is always a 128-bit integer vector type. |
17233 | // |
17234 | // We reach this code only if the following two conditions are met: |
17235 | // 1. Both return type and operand type have been promoted to wider types |
17236 | // by the type legalizer. |
17237 | // 2. The original operand type has been promoted to a 256-bit vector. |
17238 | // |
17239 | // Note that condition 2. only applies for AVX targets. |
17240 | SDValue NewOp = DAG.getSetCC(dl, VTOp0, Op0, Op1, Cond); |
17241 | return DAG.getZExtOrTrunc(NewOp, dl, VT); |
17242 | } |
17243 | |
17244 | // The non-AVX512 code below works under the assumption that source and |
17245 | // destination types are the same. |
17246 | assert((Subtarget.hasAVX512() || (VT == VTOp0)) &&(((Subtarget.hasAVX512() || (VT == VTOp0)) && "Value types for source and destination must be the same!" ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasAVX512() || (VT == VTOp0)) && \"Value types for source and destination must be the same!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17247, __PRETTY_FUNCTION__)) |
17247 | "Value types for source and destination must be the same!")(((Subtarget.hasAVX512() || (VT == VTOp0)) && "Value types for source and destination must be the same!" ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasAVX512() || (VT == VTOp0)) && \"Value types for source and destination must be the same!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17247, __PRETTY_FUNCTION__)); |
17248 | |
17249 | // Break 256-bit integer vector compare into smaller ones. |
17250 | if (VT.is256BitVector() && !Subtarget.hasInt256()) |
17251 | return Lower256IntVSETCC(Op, DAG); |
17252 | |
17253 | // Operands are boolean (vectors of i1) |
17254 | MVT OpVT = Op1.getSimpleValueType(); |
17255 | if (OpVT.getVectorElementType() == MVT::i1) |
17256 | return LowerBoolVSETCC_AVX512(Op, DAG); |
17257 | |
17258 | // The result is boolean, but operands are int/float |
17259 | if (VT.getVectorElementType() == MVT::i1) { |
17260 | // In AVX-512 architecture setcc returns mask with i1 elements, |
17261 | // But there is no compare instruction for i8 and i16 elements in KNL. |
17262 | // In this case use SSE compare |
17263 | bool UseAVX512Inst = |
17264 | (OpVT.is512BitVector() || |
17265 | OpVT.getScalarSizeInBits() >= 32 || |
17266 | (Subtarget.hasBWI() && Subtarget.hasVLX())); |
17267 | |
17268 | if (UseAVX512Inst) |
17269 | return LowerIntVSETCC_AVX512(Op, DAG); |
17270 | |
17271 | return DAG.getNode(ISD::TRUNCATE, dl, VT, |
17272 | DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC)); |
17273 | } |
17274 | |
17275 | // Lower using XOP integer comparisons. |
17276 | if ((VT == MVT::v16i8 || VT == MVT::v8i16 || |
17277 | VT == MVT::v4i32 || VT == MVT::v2i64) && Subtarget.hasXOP()) { |
17278 | // Translate compare code to XOP PCOM compare mode. |
17279 | unsigned CmpMode = 0; |
17280 | switch (Cond) { |
17281 | default: llvm_unreachable("Unexpected SETCC condition")::llvm::llvm_unreachable_internal("Unexpected SETCC condition" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17281); |
17282 | case ISD::SETULT: |
17283 | case ISD::SETLT: CmpMode = 0x00; break; |
17284 | case ISD::SETULE: |
17285 | case ISD::SETLE: CmpMode = 0x01; break; |
17286 | case ISD::SETUGT: |
17287 | case ISD::SETGT: CmpMode = 0x02; break; |
17288 | case ISD::SETUGE: |
17289 | case ISD::SETGE: CmpMode = 0x03; break; |
17290 | case ISD::SETEQ: CmpMode = 0x04; break; |
17291 | case ISD::SETNE: CmpMode = 0x05; break; |
17292 | } |
17293 | |
17294 | // Are we comparing unsigned or signed integers? |
17295 | unsigned Opc = |
17296 | ISD::isUnsignedIntSetCC(Cond) ? X86ISD::VPCOMU : X86ISD::VPCOM; |
17297 | |
17298 | return DAG.getNode(Opc, dl, VT, Op0, Op1, |
17299 | DAG.getConstant(CmpMode, dl, MVT::i8)); |
17300 | } |
17301 | |
17302 | // We are handling one of the integer comparisons here. Since SSE only has |
17303 | // GT and EQ comparisons for integer, swapping operands and multiple |
17304 | // operations may be required for some comparisons. |
17305 | unsigned Opc = (Cond == ISD::SETEQ || Cond == ISD::SETNE) ? X86ISD::PCMPEQ |
17306 | : X86ISD::PCMPGT; |
17307 | bool Swap = Cond == ISD::SETLT || Cond == ISD::SETULT || |
17308 | Cond == ISD::SETGE || Cond == ISD::SETUGE; |
17309 | bool Invert = Cond == ISD::SETNE || |
17310 | (Cond != ISD::SETEQ && ISD::isTrueWhenEqual(Cond)); |
17311 | |
17312 | // If both operands are known non-negative, then an unsigned compare is the |
17313 | // same as a signed compare and there's no need to flip signbits. |
17314 | // TODO: We could check for more general simplifications here since we're |
17315 | // computing known bits. |
17316 | bool FlipSigns = ISD::isUnsignedIntSetCC(Cond) && |
17317 | !(DAG.SignBitIsZero(Op0) && DAG.SignBitIsZero(Op1)); |
17318 | |
17319 | // Special case: Use min/max operations for SETULE/SETUGE |
17320 | MVT VET = VT.getVectorElementType(); |
17321 | bool HasMinMax = |
17322 | (Subtarget.hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32)) || |
17323 | (Subtarget.hasSSE2() && (VET == MVT::i8)); |
17324 | bool MinMax = false; |
17325 | if (HasMinMax) { |
17326 | switch (Cond) { |
17327 | default: break; |
17328 | case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break; |
17329 | case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break; |
17330 | } |
17331 | |
17332 | if (MinMax) |
17333 | Swap = Invert = FlipSigns = false; |
17334 | } |
17335 | |
17336 | bool HasSubus = Subtarget.hasSSE2() && (VET == MVT::i8 || VET == MVT::i16); |
17337 | bool Subus = false; |
17338 | if (!MinMax && HasSubus) { |
17339 | // As another special case, use PSUBUS[BW] when it's profitable. E.g. for |
17340 | // Op0 u<= Op1: |
17341 | // t = psubus Op0, Op1 |
17342 | // pcmpeq t, <0..0> |
17343 | switch (Cond) { |
17344 | default: break; |
17345 | case ISD::SETULT: { |
17346 | // If the comparison is against a constant we can turn this into a |
17347 | // setule. With psubus, setule does not require a swap. This is |
17348 | // beneficial because the constant in the register is no longer |
17349 | // destructed as the destination so it can be hoisted out of a loop. |
17350 | // Only do this pre-AVX since vpcmp* is no longer destructive. |
17351 | if (Subtarget.hasAVX()) |
17352 | break; |
17353 | if (SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG)) { |
17354 | Op1 = ULEOp1; |
17355 | Subus = true; Invert = false; Swap = false; |
17356 | } |
17357 | break; |
17358 | } |
17359 | // Psubus is better than flip-sign because it requires no inversion. |
17360 | case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break; |
17361 | case ISD::SETULE: Subus = true; Invert = false; Swap = false; break; |
17362 | } |
17363 | |
17364 | if (Subus) { |
17365 | Opc = X86ISD::SUBUS; |
17366 | FlipSigns = false; |
17367 | } |
17368 | } |
17369 | |
17370 | if (Swap) |
17371 | std::swap(Op0, Op1); |
17372 | |
17373 | // Check that the operation in question is available (most are plain SSE2, |
17374 | // but PCMPGTQ and PCMPEQQ have different requirements). |
17375 | if (VT == MVT::v2i64) { |
17376 | if (Opc == X86ISD::PCMPGT && !Subtarget.hasSSE42()) { |
17377 | assert(Subtarget.hasSSE2() && "Don't know how to lower!")((Subtarget.hasSSE2() && "Don't know how to lower!") ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSE2() && \"Don't know how to lower!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17377, __PRETTY_FUNCTION__)); |
17378 | |
17379 | // First cast everything to the right type. |
17380 | Op0 = DAG.getBitcast(MVT::v4i32, Op0); |
17381 | Op1 = DAG.getBitcast(MVT::v4i32, Op1); |
17382 | |
17383 | // Since SSE has no unsigned integer comparisons, we need to flip the sign |
17384 | // bits of the inputs before performing those operations. The lower |
17385 | // compare is always unsigned. |
17386 | SDValue SB; |
17387 | if (FlipSigns) { |
17388 | SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32); |
17389 | } else { |
17390 | SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32); |
17391 | SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32); |
17392 | SB = DAG.getBuildVector(MVT::v4i32, dl, {Sign, Zero, Sign, Zero}); |
17393 | } |
17394 | Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB); |
17395 | Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB); |
17396 | |
17397 | // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2)) |
17398 | SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1); |
17399 | SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1); |
17400 | |
17401 | // Create masks for only the low parts/high parts of the 64 bit integers. |
17402 | static const int MaskHi[] = { 1, 1, 3, 3 }; |
17403 | static const int MaskLo[] = { 0, 0, 2, 2 }; |
17404 | SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi); |
17405 | SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo); |
17406 | SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi); |
17407 | |
17408 | SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo); |
17409 | Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi); |
17410 | |
17411 | if (Invert) |
17412 | Result = DAG.getNOT(dl, Result, MVT::v4i32); |
17413 | |
17414 | return DAG.getBitcast(VT, Result); |
17415 | } |
17416 | |
17417 | if (Opc == X86ISD::PCMPEQ && !Subtarget.hasSSE41()) { |
17418 | // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with |
17419 | // pcmpeqd + pshufd + pand. |
17420 | assert(Subtarget.hasSSE2() && !FlipSigns && "Don't know how to lower!")((Subtarget.hasSSE2() && !FlipSigns && "Don't know how to lower!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSE2() && !FlipSigns && \"Don't know how to lower!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17420, __PRETTY_FUNCTION__)); |
17421 | |
17422 | // First cast everything to the right type. |
17423 | Op0 = DAG.getBitcast(MVT::v4i32, Op0); |
17424 | Op1 = DAG.getBitcast(MVT::v4i32, Op1); |
17425 | |
17426 | // Do the compare. |
17427 | SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1); |
17428 | |
17429 | // Make sure the lower and upper halves are both all-ones. |
17430 | static const int Mask[] = { 1, 0, 3, 2 }; |
17431 | SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask); |
17432 | Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf); |
17433 | |
17434 | if (Invert) |
17435 | Result = DAG.getNOT(dl, Result, MVT::v4i32); |
17436 | |
17437 | return DAG.getBitcast(VT, Result); |
17438 | } |
17439 | } |
17440 | |
17441 | // Since SSE has no unsigned integer comparisons, we need to flip the sign |
17442 | // bits of the inputs before performing those operations. |
17443 | if (FlipSigns) { |
17444 | MVT EltVT = VT.getVectorElementType(); |
17445 | SDValue SM = DAG.getConstant(APInt::getSignMask(EltVT.getSizeInBits()), dl, |
17446 | VT); |
17447 | Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SM); |
17448 | Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SM); |
17449 | } |
17450 | |
17451 | SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1); |
17452 | |
17453 | // If the logical-not of the result is required, perform that now. |
17454 | if (Invert) |
17455 | Result = DAG.getNOT(dl, Result, VT); |
17456 | |
17457 | if (MinMax) |
17458 | Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result); |
17459 | |
17460 | if (Subus) |
17461 | Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result, |
17462 | getZeroVector(VT, Subtarget, DAG, dl)); |
17463 | |
17464 | return Result; |
17465 | } |
17466 | |
17467 | SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const { |
17468 | |
17469 | MVT VT = Op.getSimpleValueType(); |
17470 | |
17471 | if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG); |
17472 | |
17473 | assert(VT == MVT::i8 && "SetCC type must be 8-bit integer")((VT == MVT::i8 && "SetCC type must be 8-bit integer" ) ? static_cast<void> (0) : __assert_fail ("VT == MVT::i8 && \"SetCC type must be 8-bit integer\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17473, __PRETTY_FUNCTION__)); |
17474 | SDValue Op0 = Op.getOperand(0); |
17475 | SDValue Op1 = Op.getOperand(1); |
17476 | SDLoc dl(Op); |
17477 | ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); |
17478 | |
17479 | // Optimize to BT if possible. |
17480 | // Lower (X & (1 << N)) == 0 to BT(X, N). |
17481 | // Lower ((X >>u N) & 1) != 0 to BT(X, N). |
17482 | // Lower ((X >>s N) & 1) != 0 to BT(X, N). |
17483 | // Lower (trunc (X >> N) to i1) to BT(X, N). |
17484 | if (Op0.hasOneUse() && isNullConstant(Op1) && |
17485 | (CC == ISD::SETEQ || CC == ISD::SETNE)) { |
17486 | if (SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG)) { |
17487 | if (VT == MVT::i1) |
17488 | return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC); |
17489 | return NewSetCC; |
17490 | } |
17491 | } |
17492 | |
17493 | // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of |
17494 | // these. |
17495 | if ((isOneConstant(Op1) || isNullConstant(Op1)) && |
17496 | (CC == ISD::SETEQ || CC == ISD::SETNE)) { |
17497 | |
17498 | // If the input is a setcc, then reuse the input setcc or use a new one with |
17499 | // the inverted condition. |
17500 | if (Op0.getOpcode() == X86ISD::SETCC) { |
17501 | X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0); |
17502 | bool Invert = (CC == ISD::SETNE) ^ isNullConstant(Op1); |
17503 | if (!Invert) |
17504 | return Op0; |
17505 | |
17506 | CCode = X86::GetOppositeBranchCondition(CCode); |
17507 | SDValue SetCC = getSETCC(CCode, Op0.getOperand(1), dl, DAG); |
17508 | if (VT == MVT::i1) |
17509 | return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC); |
17510 | return SetCC; |
17511 | } |
17512 | } |
17513 | if (Op0.getValueType() == MVT::i1 && (CC == ISD::SETEQ || CC == ISD::SETNE)) { |
17514 | if (isOneConstant(Op1)) { |
17515 | ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true); |
17516 | return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC); |
17517 | } |
17518 | if (!isNullConstant(Op1)) { |
17519 | SDValue Xor = DAG.getNode(ISD::XOR, dl, MVT::i1, Op0, Op1); |
17520 | return DAG.getSetCC(dl, VT, Xor, DAG.getConstant(0, dl, MVT::i1), CC); |
17521 | } |
17522 | } |
17523 | |
17524 | bool IsFP = Op1.getSimpleValueType().isFloatingPoint(); |
17525 | X86::CondCode X86CC = TranslateX86CC(CC, dl, IsFP, Op0, Op1, DAG); |
17526 | if (X86CC == X86::COND_INVALID) |
17527 | return SDValue(); |
17528 | |
17529 | SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG); |
17530 | EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG); |
17531 | SDValue SetCC = getSETCC(X86CC, EFLAGS, dl, DAG); |
17532 | if (VT == MVT::i1) |
17533 | return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC); |
17534 | return SetCC; |
17535 | } |
17536 | |
17537 | SDValue X86TargetLowering::LowerSETCCCARRY(SDValue Op, SelectionDAG &DAG) const { |
17538 | SDValue LHS = Op.getOperand(0); |
17539 | SDValue RHS = Op.getOperand(1); |
17540 | SDValue Carry = Op.getOperand(2); |
17541 | SDValue Cond = Op.getOperand(3); |
17542 | SDLoc DL(Op); |
17543 | |
17544 | assert(LHS.getSimpleValueType().isInteger() && "SETCCCARRY is integer only.")((LHS.getSimpleValueType().isInteger() && "SETCCCARRY is integer only." ) ? static_cast<void> (0) : __assert_fail ("LHS.getSimpleValueType().isInteger() && \"SETCCCARRY is integer only.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17544, __PRETTY_FUNCTION__)); |
17545 | X86::CondCode CC = TranslateIntegerX86CC(cast<CondCodeSDNode>(Cond)->get()); |
17546 | |
17547 | // Recreate the carry if needed. |
17548 | EVT CarryVT = Carry.getValueType(); |
17549 | APInt NegOne = APInt::getAllOnesValue(CarryVT.getScalarSizeInBits()); |
17550 | Carry = DAG.getNode(X86ISD::ADD, DL, DAG.getVTList(CarryVT, MVT::i32), |
17551 | Carry, DAG.getConstant(NegOne, DL, CarryVT)); |
17552 | |
17553 | SDVTList VTs = DAG.getVTList(LHS.getValueType(), MVT::i32); |
17554 | SDValue Cmp = DAG.getNode(X86ISD::SBB, DL, VTs, LHS, RHS, Carry.getValue(1)); |
17555 | SDValue SetCC = getSETCC(CC, Cmp.getValue(1), DL, DAG); |
17556 | if (Op.getSimpleValueType() == MVT::i1) |
17557 | return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, SetCC); |
17558 | return SetCC; |
17559 | } |
17560 | |
17561 | /// Return true if opcode is a X86 logical comparison. |
17562 | static bool isX86LogicalCmp(SDValue Op) { |
17563 | unsigned Opc = Op.getOpcode(); |
17564 | if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI || |
17565 | Opc == X86ISD::SAHF) |
17566 | return true; |
17567 | if (Op.getResNo() == 1 && |
17568 | (Opc == X86ISD::ADD || Opc == X86ISD::SUB || Opc == X86ISD::ADC || |
17569 | Opc == X86ISD::SBB || Opc == X86ISD::SMUL || Opc == X86ISD::UMUL || |
17570 | Opc == X86ISD::INC || Opc == X86ISD::DEC || Opc == X86ISD::OR || |
17571 | Opc == X86ISD::XOR || Opc == X86ISD::AND)) |
17572 | return true; |
17573 | |
17574 | if (Op.getResNo() == 2 && Opc == X86ISD::UMUL) |
17575 | return true; |
17576 | |
17577 | return false; |
17578 | } |
17579 | |
17580 | static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) { |
17581 | if (V.getOpcode() != ISD::TRUNCATE) |
17582 | return false; |
17583 | |
17584 | SDValue VOp0 = V.getOperand(0); |
17585 | unsigned InBits = VOp0.getValueSizeInBits(); |
17586 | unsigned Bits = V.getValueSizeInBits(); |
17587 | return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits)); |
17588 | } |
17589 | |
17590 | SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const { |
17591 | bool AddTest = true; |
17592 | SDValue Cond = Op.getOperand(0); |
17593 | SDValue Op1 = Op.getOperand(1); |
17594 | SDValue Op2 = Op.getOperand(2); |
17595 | SDLoc DL(Op); |
17596 | MVT VT = Op1.getSimpleValueType(); |
17597 | SDValue CC; |
17598 | |
17599 | // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops |
17600 | // are available or VBLENDV if AVX is available. |
17601 | // Otherwise FP cmovs get lowered into a less efficient branch sequence later. |
17602 | if (Cond.getOpcode() == ISD::SETCC && |
17603 | ((Subtarget.hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) || |
17604 | (Subtarget.hasSSE1() && VT == MVT::f32)) && |
17605 | VT == Cond.getOperand(0).getSimpleValueType() && Cond->hasOneUse()) { |
17606 | SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1); |
17607 | int SSECC = translateX86FSETCC( |
17608 | cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1); |
17609 | |
17610 | if (SSECC != 8) { |
17611 | if (Subtarget.hasAVX512()) { |
17612 | SDValue Cmp = DAG.getNode(X86ISD::FSETCCM, DL, MVT::v1i1, CondOp0, |
17613 | CondOp1, DAG.getConstant(SSECC, DL, MVT::i8)); |
17614 | return DAG.getNode(VT.isVector() ? X86ISD::SELECT : X86ISD::SELECTS, |
17615 | DL, VT, Cmp, Op1, Op2); |
17616 | } |
17617 | |
17618 | SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1, |
17619 | DAG.getConstant(SSECC, DL, MVT::i8)); |
17620 | |
17621 | // If we have AVX, we can use a variable vector select (VBLENDV) instead |
17622 | // of 3 logic instructions for size savings and potentially speed. |
17623 | // Unfortunately, there is no scalar form of VBLENDV. |
17624 | |
17625 | // If either operand is a constant, don't try this. We can expect to |
17626 | // optimize away at least one of the logic instructions later in that |
17627 | // case, so that sequence would be faster than a variable blend. |
17628 | |
17629 | // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly |
17630 | // uses XMM0 as the selection register. That may need just as many |
17631 | // instructions as the AND/ANDN/OR sequence due to register moves, so |
17632 | // don't bother. |
17633 | |
17634 | if (Subtarget.hasAVX() && |
17635 | !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) { |
17636 | |
17637 | // Convert to vectors, do a VSELECT, and convert back to scalar. |
17638 | // All of the conversions should be optimized away. |
17639 | |
17640 | MVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64; |
17641 | SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1); |
17642 | SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2); |
17643 | SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp); |
17644 | |
17645 | MVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64; |
17646 | VCmp = DAG.getBitcast(VCmpVT, VCmp); |
17647 | |
17648 | SDValue VSel = DAG.getSelect(DL, VecVT, VCmp, VOp1, VOp2); |
17649 | |
17650 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, |
17651 | VSel, DAG.getIntPtrConstant(0, DL)); |
17652 | } |
17653 | SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2); |
17654 | SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1); |
17655 | return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And); |
17656 | } |
17657 | } |
17658 | |
17659 | // AVX512 fallback is to lower selects of scalar floats to masked moves. |
17660 | if ((VT == MVT::f64 || VT == MVT::f32) && Subtarget.hasAVX512()) { |
17661 | SDValue Cmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v1i1, Cond); |
17662 | return DAG.getNode(X86ISD::SELECTS, DL, VT, Cmp, Op1, Op2); |
17663 | } |
17664 | |
17665 | if (VT.isVector() && VT.getVectorElementType() == MVT::i1) { |
17666 | SDValue Op1Scalar; |
17667 | if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode())) |
17668 | Op1Scalar = ConvertI1VectorToInteger(Op1, DAG); |
17669 | else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0)) |
17670 | Op1Scalar = Op1.getOperand(0); |
17671 | SDValue Op2Scalar; |
17672 | if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode())) |
17673 | Op2Scalar = ConvertI1VectorToInteger(Op2, DAG); |
17674 | else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0)) |
17675 | Op2Scalar = Op2.getOperand(0); |
17676 | if (Op1Scalar.getNode() && Op2Scalar.getNode()) { |
17677 | SDValue newSelect = DAG.getSelect(DL, Op1Scalar.getValueType(), Cond, |
17678 | Op1Scalar, Op2Scalar); |
17679 | if (newSelect.getValueSizeInBits() == VT.getSizeInBits()) |
17680 | return DAG.getBitcast(VT, newSelect); |
17681 | SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect); |
17682 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec, |
17683 | DAG.getIntPtrConstant(0, DL)); |
17684 | } |
17685 | } |
17686 | |
17687 | if (VT == MVT::v4i1 || VT == MVT::v2i1) { |
17688 | SDValue zeroConst = DAG.getIntPtrConstant(0, DL); |
17689 | Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1, |
17690 | DAG.getUNDEF(MVT::v8i1), Op1, zeroConst); |
17691 | Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1, |
17692 | DAG.getUNDEF(MVT::v8i1), Op2, zeroConst); |
17693 | SDValue newSelect = DAG.getSelect(DL, MVT::v8i1, Cond, Op1, Op2); |
17694 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst); |
17695 | } |
17696 | |
17697 | if (Cond.getOpcode() == ISD::SETCC) { |
17698 | if (SDValue NewCond = LowerSETCC(Cond, DAG)) { |
17699 | Cond = NewCond; |
17700 | // If the condition was updated, it's possible that the operands of the |
17701 | // select were also updated (for example, EmitTest has a RAUW). Refresh |
17702 | // the local references to the select operands in case they got stale. |
17703 | Op1 = Op.getOperand(1); |
17704 | Op2 = Op.getOperand(2); |
17705 | } |
17706 | } |
17707 | |
17708 | // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y |
17709 | // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y |
17710 | // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y |
17711 | // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y |
17712 | // (select (and (x , 0x1) == 0), y, (z ^ y) ) -> (-(and (x , 0x1)) & z ) ^ y |
17713 | // (select (and (x , 0x1) == 0), y, (z | y) ) -> (-(and (x , 0x1)) & z ) | y |
17714 | if (Cond.getOpcode() == X86ISD::SETCC && |
17715 | Cond.getOperand(1).getOpcode() == X86ISD::CMP && |
17716 | isNullConstant(Cond.getOperand(1).getOperand(1))) { |
17717 | SDValue Cmp = Cond.getOperand(1); |
17718 | unsigned CondCode = |
17719 | cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue(); |
17720 | |
17721 | if ((isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) && |
17722 | (CondCode == X86::COND_E || CondCode == X86::COND_NE)) { |
17723 | SDValue Y = isAllOnesConstant(Op2) ? Op1 : Op2; |
17724 | SDValue CmpOp0 = Cmp.getOperand(0); |
17725 | |
17726 | // Apply further optimizations for special cases |
17727 | // (select (x != 0), -1, 0) -> neg & sbb |
17728 | // (select (x == 0), 0, -1) -> neg & sbb |
17729 | if (isNullConstant(Y) && |
17730 | (isAllOnesConstant(Op1) == (CondCode == X86::COND_NE))) { |
17731 | SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32); |
17732 | SDValue Zero = DAG.getConstant(0, DL, CmpOp0.getValueType()); |
17733 | SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs, Zero, CmpOp0); |
17734 | SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(), |
17735 | DAG.getConstant(X86::COND_B, DL, MVT::i8), |
17736 | SDValue(Neg.getNode(), 1)); |
17737 | return Res; |
17738 | } |
17739 | |
17740 | Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, |
17741 | CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType())); |
17742 | Cmp = ConvertCmpIfNecessary(Cmp, DAG); |
17743 | |
17744 | SDValue Res = // Res = 0 or -1. |
17745 | DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(), |
17746 | DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp); |
17747 | |
17748 | if (isAllOnesConstant(Op1) != (CondCode == X86::COND_E)) |
17749 | Res = DAG.getNOT(DL, Res, Res.getValueType()); |
17750 | |
17751 | if (!isNullConstant(Op2)) |
17752 | Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y); |
17753 | return Res; |
17754 | } else if (!Subtarget.hasCMov() && CondCode == X86::COND_E && |
17755 | Cmp.getOperand(0).getOpcode() == ISD::AND && |
17756 | isOneConstant(Cmp.getOperand(0).getOperand(1))) { |
17757 | SDValue CmpOp0 = Cmp.getOperand(0); |
17758 | SDValue Src1, Src2; |
17759 | // true if Op2 is XOR or OR operator and one of its operands |
17760 | // is equal to Op1 |
17761 | // ( a , a op b) || ( b , a op b) |
17762 | auto isOrXorPattern = [&]() { |
17763 | if ((Op2.getOpcode() == ISD::XOR || Op2.getOpcode() == ISD::OR) && |
17764 | (Op2.getOperand(0) == Op1 || Op2.getOperand(1) == Op1)) { |
17765 | Src1 = |
17766 | Op2.getOperand(0) == Op1 ? Op2.getOperand(1) : Op2.getOperand(0); |
17767 | Src2 = Op1; |
17768 | return true; |
17769 | } |
17770 | return false; |
17771 | }; |
17772 | |
17773 | if (isOrXorPattern()) { |
17774 | SDValue Neg; |
17775 | unsigned int CmpSz = CmpOp0.getSimpleValueType().getSizeInBits(); |
17776 | // we need mask of all zeros or ones with same size of the other |
17777 | // operands. |
17778 | if (CmpSz > VT.getSizeInBits()) |
17779 | Neg = DAG.getNode(ISD::TRUNCATE, DL, VT, CmpOp0); |
17780 | else if (CmpSz < VT.getSizeInBits()) |
17781 | Neg = DAG.getNode(ISD::AND, DL, VT, |
17782 | DAG.getNode(ISD::ANY_EXTEND, DL, VT, CmpOp0.getOperand(0)), |
17783 | DAG.getConstant(1, DL, VT)); |
17784 | else |
17785 | Neg = CmpOp0; |
17786 | SDValue Mask = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), |
17787 | Neg); // -(and (x, 0x1)) |
17788 | SDValue And = DAG.getNode(ISD::AND, DL, VT, Mask, Src1); // Mask & z |
17789 | return DAG.getNode(Op2.getOpcode(), DL, VT, And, Src2); // And Op y |
17790 | } |
17791 | } |
17792 | } |
17793 | |
17794 | // Look past (and (setcc_carry (cmp ...)), 1). |
17795 | if (Cond.getOpcode() == ISD::AND && |
17796 | Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY && |
17797 | isOneConstant(Cond.getOperand(1))) |
17798 | Cond = Cond.getOperand(0); |
17799 | |
17800 | // If condition flag is set by a X86ISD::CMP, then use it as the condition |
17801 | // setting operand in place of the X86ISD::SETCC. |
17802 | unsigned CondOpcode = Cond.getOpcode(); |
17803 | if (CondOpcode == X86ISD::SETCC || |
17804 | CondOpcode == X86ISD::SETCC_CARRY) { |
17805 | CC = Cond.getOperand(0); |
17806 | |
17807 | SDValue Cmp = Cond.getOperand(1); |
17808 | unsigned Opc = Cmp.getOpcode(); |
17809 | MVT VT = Op.getSimpleValueType(); |
17810 | |
17811 | bool IllegalFPCMov = false; |
17812 | if (VT.isFloatingPoint() && !VT.isVector() && |
17813 | !isScalarFPTypeInSSEReg(VT)) // FPStack? |
17814 | IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue()); |
17815 | |
17816 | if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) || |
17817 | Opc == X86ISD::BT) { // FIXME |
17818 | Cond = Cmp; |
17819 | AddTest = false; |
17820 | } |
17821 | } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO || |
17822 | CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO || |
17823 | ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) && |
17824 | Cond.getOperand(0).getValueType() != MVT::i8)) { |
17825 | SDValue LHS = Cond.getOperand(0); |
17826 | SDValue RHS = Cond.getOperand(1); |
17827 | unsigned X86Opcode; |
17828 | unsigned X86Cond; |
17829 | SDVTList VTs; |
17830 | switch (CondOpcode) { |
17831 | case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break; |
17832 | case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break; |
17833 | case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break; |
17834 | case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break; |
17835 | case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break; |
17836 | case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break; |
17837 | default: llvm_unreachable("unexpected overflowing operator")::llvm::llvm_unreachable_internal("unexpected overflowing operator" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17837); |
17838 | } |
17839 | if (CondOpcode == ISD::UMULO) |
17840 | VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(), |
17841 | MVT::i32); |
17842 | else |
17843 | VTs = DAG.getVTList(LHS.getValueType(), MVT::i32); |
17844 | |
17845 | SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS); |
17846 | |
17847 | if (CondOpcode == ISD::UMULO) |
17848 | Cond = X86Op.getValue(2); |
17849 | else |
17850 | Cond = X86Op.getValue(1); |
17851 | |
17852 | CC = DAG.getConstant(X86Cond, DL, MVT::i8); |
17853 | AddTest = false; |
17854 | } |
17855 | |
17856 | if (AddTest) { |
17857 | // Look past the truncate if the high bits are known zero. |
17858 | if (isTruncWithZeroHighBitsInput(Cond, DAG)) |
17859 | Cond = Cond.getOperand(0); |
17860 | |
17861 | // We know the result of AND is compared against zero. Try to match |
17862 | // it to BT. |
17863 | if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) { |
17864 | if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG)) { |
17865 | CC = NewSetCC.getOperand(0); |
17866 | Cond = NewSetCC.getOperand(1); |
17867 | AddTest = false; |
17868 | } |
17869 | } |
17870 | } |
17871 | |
17872 | if (AddTest) { |
17873 | CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8); |
17874 | Cond = EmitTest(Cond, X86::COND_NE, DL, DAG); |
17875 | } |
17876 | |
17877 | // a < b ? -1 : 0 -> RES = ~setcc_carry |
17878 | // a < b ? 0 : -1 -> RES = setcc_carry |
17879 | // a >= b ? -1 : 0 -> RES = setcc_carry |
17880 | // a >= b ? 0 : -1 -> RES = ~setcc_carry |
17881 | if (Cond.getOpcode() == X86ISD::SUB) { |
17882 | Cond = ConvertCmpIfNecessary(Cond, DAG); |
17883 | unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue(); |
17884 | |
17885 | if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) && |
17886 | (isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) && |
17887 | (isNullConstant(Op1) || isNullConstant(Op2))) { |
17888 | SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(), |
17889 | DAG.getConstant(X86::COND_B, DL, MVT::i8), |
17890 | Cond); |
17891 | if (isAllOnesConstant(Op1) != (CondCode == X86::COND_B)) |
17892 | return DAG.getNOT(DL, Res, Res.getValueType()); |
17893 | return Res; |
17894 | } |
17895 | } |
17896 | |
17897 | // X86 doesn't have an i8 cmov. If both operands are the result of a truncate |
17898 | // widen the cmov and push the truncate through. This avoids introducing a new |
17899 | // branch during isel and doesn't add any extensions. |
17900 | if (Op.getValueType() == MVT::i8 && |
17901 | Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) { |
17902 | SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0); |
17903 | if (T1.getValueType() == T2.getValueType() && |
17904 | // Blacklist CopyFromReg to avoid partial register stalls. |
17905 | T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){ |
17906 | SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue); |
17907 | SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond); |
17908 | return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov); |
17909 | } |
17910 | } |
17911 | |
17912 | // X86ISD::CMOV means set the result (which is operand 1) to the RHS if |
17913 | // condition is true. |
17914 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue); |
17915 | SDValue Ops[] = { Op2, Op1, CC, Cond }; |
17916 | return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops); |
17917 | } |
17918 | |
17919 | static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, |
17920 | const X86Subtarget &Subtarget, |
17921 | SelectionDAG &DAG) { |
17922 | MVT VT = Op->getSimpleValueType(0); |
17923 | SDValue In = Op->getOperand(0); |
17924 | MVT InVT = In.getSimpleValueType(); |
17925 | MVT VTElt = VT.getVectorElementType(); |
17926 | MVT InVTElt = InVT.getVectorElementType(); |
17927 | SDLoc dl(Op); |
17928 | |
17929 | // SKX processor |
17930 | if ((InVTElt == MVT::i1) && |
17931 | (((Subtarget.hasBWI() && VTElt.getSizeInBits() <= 16)) || |
17932 | |
17933 | ((Subtarget.hasDQI() && VTElt.getSizeInBits() >= 32)))) |
17934 | |
17935 | return DAG.getNode(X86ISD::VSEXT, dl, VT, In); |
17936 | |
17937 | unsigned NumElts = VT.getVectorNumElements(); |
17938 | |
17939 | if (VT.is512BitVector() && InVTElt != MVT::i1 && |
17940 | (NumElts == 8 || NumElts == 16 || Subtarget.hasBWI())) { |
17941 | if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT) |
17942 | return getExtendInVec(In.getOpcode(), dl, VT, In.getOperand(0), DAG); |
17943 | return getExtendInVec(X86ISD::VSEXT, dl, VT, In, DAG); |
17944 | } |
17945 | |
17946 | if (InVTElt != MVT::i1) |
17947 | return SDValue(); |
17948 | |
17949 | MVT ExtVT = VT; |
17950 | if (!VT.is512BitVector() && !Subtarget.hasVLX()) |
17951 | ExtVT = MVT::getVectorVT(MVT::getIntegerVT(512/NumElts), NumElts); |
17952 | |
17953 | SDValue V; |
17954 | if (Subtarget.hasDQI()) { |
17955 | V = getExtendInVec(X86ISD::VSEXT, dl, ExtVT, In, DAG); |
17956 | assert(!VT.is512BitVector() && "Unexpected vector type")((!VT.is512BitVector() && "Unexpected vector type") ? static_cast<void> (0) : __assert_fail ("!VT.is512BitVector() && \"Unexpected vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17956, __PRETTY_FUNCTION__)); |
17957 | } else { |
17958 | SDValue NegOne = getOnesVector(ExtVT, DAG, dl); |
17959 | SDValue Zero = getZeroVector(ExtVT, Subtarget, DAG, dl); |
17960 | V = DAG.getSelect(dl, ExtVT, In, NegOne, Zero); |
17961 | if (ExtVT == VT) |
17962 | return V; |
17963 | } |
17964 | |
17965 | return DAG.getNode(X86ISD::VTRUNC, dl, VT, V); |
17966 | } |
17967 | |
17968 | // Lowering for SIGN_EXTEND_VECTOR_INREG and ZERO_EXTEND_VECTOR_INREG. |
17969 | // For sign extend this needs to handle all vector sizes and SSE4.1 and |
17970 | // non-SSE4.1 targets. For zero extend this should only handle inputs of |
17971 | // MVT::v64i8 when BWI is not supported, but AVX512 is. |
17972 | static SDValue LowerEXTEND_VECTOR_INREG(SDValue Op, |
17973 | const X86Subtarget &Subtarget, |
17974 | SelectionDAG &DAG) { |
17975 | SDValue In = Op->getOperand(0); |
17976 | MVT VT = Op->getSimpleValueType(0); |
17977 | MVT InVT = In.getSimpleValueType(); |
17978 | assert(VT.getSizeInBits() == InVT.getSizeInBits())((VT.getSizeInBits() == InVT.getSizeInBits()) ? static_cast< void> (0) : __assert_fail ("VT.getSizeInBits() == InVT.getSizeInBits()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17978, __PRETTY_FUNCTION__)); |
17979 | |
17980 | MVT SVT = VT.getVectorElementType(); |
17981 | MVT InSVT = InVT.getVectorElementType(); |
17982 | assert(SVT.getSizeInBits() > InSVT.getSizeInBits())((SVT.getSizeInBits() > InSVT.getSizeInBits()) ? static_cast <void> (0) : __assert_fail ("SVT.getSizeInBits() > InSVT.getSizeInBits()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 17982, __PRETTY_FUNCTION__)); |
17983 | |
17984 | if (SVT != MVT::i64 && SVT != MVT::i32 && SVT != MVT::i16) |
17985 | return SDValue(); |
17986 | if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8) |
17987 | return SDValue(); |
17988 | if (!(VT.is128BitVector() && Subtarget.hasSSE2()) && |
17989 | !(VT.is256BitVector() && Subtarget.hasInt256()) && |
17990 | !(VT.is512BitVector() && Subtarget.hasAVX512())) |
17991 | return SDValue(); |
17992 | |
17993 | SDLoc dl(Op); |
17994 | |
17995 | // For 256-bit vectors, we only need the lower (128-bit) half of the input. |
17996 | // For 512-bit vectors, we need 128-bits or 256-bits. |
17997 | if (VT.getSizeInBits() > 128) { |
17998 | // Input needs to be at least the same number of elements as output, and |
17999 | // at least 128-bits. |
18000 | int InSize = InSVT.getSizeInBits() * VT.getVectorNumElements(); |
18001 | In = extractSubVector(In, 0, DAG, dl, std::max(InSize, 128)); |
18002 | } |
18003 | |
18004 | assert((Op.getOpcode() != ISD::ZERO_EXTEND_VECTOR_INREG ||(((Op.getOpcode() != ISD::ZERO_EXTEND_VECTOR_INREG || InVT == MVT::v64i8) && "Zero extend only for v64i8 input!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() != ISD::ZERO_EXTEND_VECTOR_INREG || InVT == MVT::v64i8) && \"Zero extend only for v64i8 input!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18005, __PRETTY_FUNCTION__)) |
18005 | InVT == MVT::v64i8) && "Zero extend only for v64i8 input!")(((Op.getOpcode() != ISD::ZERO_EXTEND_VECTOR_INREG || InVT == MVT::v64i8) && "Zero extend only for v64i8 input!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() != ISD::ZERO_EXTEND_VECTOR_INREG || InVT == MVT::v64i8) && \"Zero extend only for v64i8 input!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18005, __PRETTY_FUNCTION__)); |
18006 | |
18007 | // SSE41 targets can use the pmovsx* instructions directly for 128-bit results, |
18008 | // so are legal and shouldn't occur here. AVX2/AVX512 pmovsx* instructions still |
18009 | // need to be handled here for 256/512-bit results. |
18010 | if (Subtarget.hasInt256()) { |
18011 | assert(VT.getSizeInBits() > 128 && "Unexpected 128-bit vector extension")((VT.getSizeInBits() > 128 && "Unexpected 128-bit vector extension" ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() > 128 && \"Unexpected 128-bit vector extension\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18011, __PRETTY_FUNCTION__)); |
18012 | unsigned ExtOpc = Op.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG ? |
18013 | X86ISD::VSEXT : X86ISD::VZEXT; |
18014 | return DAG.getNode(ExtOpc, dl, VT, In); |
18015 | } |
18016 | |
18017 | // We should only get here for sign extend. |
18018 | assert(Op.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG &&((Op.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG && "Unexpected opcode!" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG && \"Unexpected opcode!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18019, __PRETTY_FUNCTION__)) |
18019 | "Unexpected opcode!")((Op.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG && "Unexpected opcode!" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG && \"Unexpected opcode!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18019, __PRETTY_FUNCTION__)); |
18020 | |
18021 | // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI. |
18022 | SDValue Curr = In; |
18023 | MVT CurrVT = InVT; |
18024 | |
18025 | // As SRAI is only available on i16/i32 types, we expand only up to i32 |
18026 | // and handle i64 separately. |
18027 | while (CurrVT != VT && CurrVT.getVectorElementType() != MVT::i32) { |
18028 | Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr); |
18029 | MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2); |
18030 | CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2); |
18031 | Curr = DAG.getBitcast(CurrVT, Curr); |
18032 | } |
18033 | |
18034 | SDValue SignExt = Curr; |
18035 | if (CurrVT != InVT) { |
18036 | unsigned SignExtShift = |
18037 | CurrVT.getScalarSizeInBits() - InSVT.getSizeInBits(); |
18038 | SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr, |
18039 | DAG.getConstant(SignExtShift, dl, MVT::i8)); |
18040 | } |
18041 | |
18042 | if (CurrVT == VT) |
18043 | return SignExt; |
18044 | |
18045 | if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) { |
18046 | SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr, |
18047 | DAG.getConstant(31, dl, MVT::i8)); |
18048 | SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5}); |
18049 | return DAG.getBitcast(VT, Ext); |
18050 | } |
18051 | |
18052 | return SDValue(); |
18053 | } |
18054 | |
18055 | static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget &Subtarget, |
18056 | SelectionDAG &DAG) { |
18057 | MVT VT = Op->getSimpleValueType(0); |
18058 | SDValue In = Op->getOperand(0); |
18059 | MVT InVT = In.getSimpleValueType(); |
18060 | SDLoc dl(Op); |
18061 | |
18062 | if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1) |
18063 | return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG); |
18064 | |
18065 | if ((VT != MVT::v4i64 || InVT != MVT::v4i32) && |
18066 | (VT != MVT::v8i32 || InVT != MVT::v8i16) && |
18067 | (VT != MVT::v16i16 || InVT != MVT::v16i8)) |
18068 | return SDValue(); |
18069 | |
18070 | if (Subtarget.hasInt256()) |
18071 | return DAG.getNode(X86ISD::VSEXT, dl, VT, In); |
18072 | |
18073 | // Optimize vectors in AVX mode |
18074 | // Sign extend v8i16 to v8i32 and |
18075 | // v4i32 to v4i64 |
18076 | // |
18077 | // Divide input vector into two parts |
18078 | // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1} |
18079 | // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32 |
18080 | // concat the vectors to original VT |
18081 | |
18082 | unsigned NumElems = InVT.getVectorNumElements(); |
18083 | SDValue Undef = DAG.getUNDEF(InVT); |
18084 | |
18085 | SmallVector<int,8> ShufMask1(NumElems, -1); |
18086 | for (unsigned i = 0; i != NumElems/2; ++i) |
18087 | ShufMask1[i] = i; |
18088 | |
18089 | SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, ShufMask1); |
18090 | |
18091 | SmallVector<int,8> ShufMask2(NumElems, -1); |
18092 | for (unsigned i = 0; i != NumElems/2; ++i) |
18093 | ShufMask2[i] = i + NumElems/2; |
18094 | |
18095 | SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, ShufMask2); |
18096 | |
18097 | MVT HalfVT = MVT::getVectorVT(VT.getVectorElementType(), |
18098 | VT.getVectorNumElements() / 2); |
18099 | |
18100 | OpLo = DAG.getSignExtendVectorInReg(OpLo, dl, HalfVT); |
18101 | OpHi = DAG.getSignExtendVectorInReg(OpHi, dl, HalfVT); |
18102 | |
18103 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi); |
18104 | } |
18105 | |
18106 | // Lower truncating store. We need a special lowering to vXi1 vectors |
18107 | static SDValue LowerTruncatingStore(SDValue StOp, const X86Subtarget &Subtarget, |
18108 | SelectionDAG &DAG) { |
18109 | StoreSDNode *St = cast<StoreSDNode>(StOp.getNode()); |
18110 | SDLoc dl(St); |
18111 | EVT MemVT = St->getMemoryVT(); |
18112 | assert(St->isTruncatingStore() && "We only custom truncating store.")((St->isTruncatingStore() && "We only custom truncating store." ) ? static_cast<void> (0) : __assert_fail ("St->isTruncatingStore() && \"We only custom truncating store.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18112, __PRETTY_FUNCTION__)); |
18113 | assert(MemVT.isVector() && MemVT.getVectorElementType() == MVT::i1 &&((MemVT.isVector() && MemVT.getVectorElementType() == MVT::i1 && "Expected truncstore of i1 vector") ? static_cast <void> (0) : __assert_fail ("MemVT.isVector() && MemVT.getVectorElementType() == MVT::i1 && \"Expected truncstore of i1 vector\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18114, __PRETTY_FUNCTION__)) |
18114 | "Expected truncstore of i1 vector")((MemVT.isVector() && MemVT.getVectorElementType() == MVT::i1 && "Expected truncstore of i1 vector") ? static_cast <void> (0) : __assert_fail ("MemVT.isVector() && MemVT.getVectorElementType() == MVT::i1 && \"Expected truncstore of i1 vector\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18114, __PRETTY_FUNCTION__)); |
18115 | |
18116 | SDValue Op = St->getValue(); |
18117 | MVT OpVT = Op.getValueType().getSimpleVT(); |
18118 | unsigned NumElts = OpVT.getVectorNumElements(); |
18119 | if ((Subtarget.hasVLX() && Subtarget.hasBWI() && Subtarget.hasDQI()) || |
18120 | NumElts == 16) { |
18121 | // Truncate and store - everything is legal |
18122 | Op = DAG.getNode(ISD::TRUNCATE, dl, MemVT, Op); |
18123 | if (MemVT.getSizeInBits() < 8) |
18124 | Op = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, MVT::v8i1, |
18125 | DAG.getUNDEF(MVT::v8i1), Op, |
18126 | DAG.getIntPtrConstant(0, dl)); |
18127 | return DAG.getStore(St->getChain(), dl, Op, St->getBasePtr(), |
18128 | St->getMemOperand()); |
18129 | } |
18130 | |
18131 | // A subset, assume that we have only AVX-512F |
18132 | if (NumElts <= 8) { |
18133 | if (NumElts < 8) { |
18134 | // Extend to 8-elts vector |
18135 | MVT ExtVT = MVT::getVectorVT(OpVT.getScalarType(), 8); |
18136 | Op = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ExtVT, |
18137 | DAG.getUNDEF(ExtVT), Op, DAG.getIntPtrConstant(0, dl)); |
18138 | } |
18139 | Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i1, Op); |
18140 | return DAG.getStore(St->getChain(), dl, Op, St->getBasePtr(), |
18141 | St->getMemOperand()); |
18142 | } |
18143 | // v32i8 |
18144 | assert(OpVT == MVT::v32i8 && "Unexpected operand type")((OpVT == MVT::v32i8 && "Unexpected operand type") ? static_cast <void> (0) : __assert_fail ("OpVT == MVT::v32i8 && \"Unexpected operand type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18144, __PRETTY_FUNCTION__)); |
18145 | // Divide the vector into 2 parts and store each part separately |
18146 | SDValue Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v16i8, Op, |
18147 | DAG.getIntPtrConstant(0, dl)); |
18148 | Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::v16i1, Lo); |
18149 | SDValue BasePtr = St->getBasePtr(); |
18150 | SDValue StLo = DAG.getStore(St->getChain(), dl, Lo, BasePtr, |
18151 | St->getMemOperand()); |
18152 | SDValue Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v16i8, Op, |
18153 | DAG.getIntPtrConstant(16, dl)); |
18154 | Hi = DAG.getNode(ISD::TRUNCATE, dl, MVT::v16i1, Hi); |
18155 | |
18156 | SDValue BasePtrHi = |
18157 | DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr, |
18158 | DAG.getConstant(2, dl, BasePtr.getValueType())); |
18159 | |
18160 | SDValue StHi = DAG.getStore(St->getChain(), dl, Hi, |
18161 | BasePtrHi, St->getMemOperand()); |
18162 | return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, StLo, StHi); |
18163 | } |
18164 | |
18165 | static SDValue LowerExtended1BitVectorLoad(SDValue Op, |
18166 | const X86Subtarget &Subtarget, |
18167 | SelectionDAG &DAG) { |
18168 | |
18169 | LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode()); |
18170 | SDLoc dl(Ld); |
18171 | EVT MemVT = Ld->getMemoryVT(); |
18172 | assert(MemVT.isVector() && MemVT.getScalarType() == MVT::i1 &&((MemVT.isVector() && MemVT.getScalarType() == MVT::i1 && "Expected i1 vector load") ? static_cast<void> (0) : __assert_fail ("MemVT.isVector() && MemVT.getScalarType() == MVT::i1 && \"Expected i1 vector load\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18173, __PRETTY_FUNCTION__)) |
18173 | "Expected i1 vector load")((MemVT.isVector() && MemVT.getScalarType() == MVT::i1 && "Expected i1 vector load") ? static_cast<void> (0) : __assert_fail ("MemVT.isVector() && MemVT.getScalarType() == MVT::i1 && \"Expected i1 vector load\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18173, __PRETTY_FUNCTION__)); |
18174 | unsigned ExtOpcode = Ld->getExtensionType() == ISD::ZEXTLOAD ? |
18175 | ISD::ZERO_EXTEND : ISD::SIGN_EXTEND; |
18176 | MVT VT = Op.getValueType().getSimpleVT(); |
18177 | unsigned NumElts = VT.getVectorNumElements(); |
18178 | |
18179 | if ((Subtarget.hasBWI() && NumElts >= 32) || |
18180 | (Subtarget.hasDQI() && NumElts < 16) || |
18181 | NumElts == 16) { |
18182 | // Load and extend - everything is legal |
18183 | if (NumElts < 8) { |
18184 | SDValue Load = DAG.getLoad(MVT::v8i1, dl, Ld->getChain(), |
18185 | Ld->getBasePtr(), |
18186 | Ld->getMemOperand()); |
18187 | // Replace chain users with the new chain. |
18188 | assert(Load->getNumValues() == 2 && "Loads must carry a chain!")((Load->getNumValues() == 2 && "Loads must carry a chain!" ) ? static_cast<void> (0) : __assert_fail ("Load->getNumValues() == 2 && \"Loads must carry a chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18188, __PRETTY_FUNCTION__)); |
18189 | DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1)); |
18190 | MVT ExtVT = MVT::getVectorVT(VT.getScalarType(), 8); |
18191 | SDValue ExtVec = DAG.getNode(ExtOpcode, dl, ExtVT, Load); |
18192 | |
18193 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec, |
18194 | DAG.getIntPtrConstant(0, dl)); |
18195 | } |
18196 | SDValue Load = DAG.getLoad(MemVT, dl, Ld->getChain(), |
18197 | Ld->getBasePtr(), |
18198 | Ld->getMemOperand()); |
18199 | // Replace chain users with the new chain. |
18200 | assert(Load->getNumValues() == 2 && "Loads must carry a chain!")((Load->getNumValues() == 2 && "Loads must carry a chain!" ) ? static_cast<void> (0) : __assert_fail ("Load->getNumValues() == 2 && \"Loads must carry a chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18200, __PRETTY_FUNCTION__)); |
18201 | DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1)); |
18202 | |
18203 | // Finally, do a normal sign-extend to the desired register. |
18204 | return DAG.getNode(ExtOpcode, dl, Op.getValueType(), Load); |
18205 | } |
18206 | |
18207 | if (NumElts <= 8) { |
18208 | // A subset, assume that we have only AVX-512F |
18209 | unsigned NumBitsToLoad = 8; |
18210 | MVT TypeToLoad = MVT::getIntegerVT(NumBitsToLoad); |
18211 | SDValue Load = DAG.getLoad(TypeToLoad, dl, Ld->getChain(), |
18212 | Ld->getBasePtr(), |
18213 | Ld->getMemOperand()); |
18214 | // Replace chain users with the new chain. |
18215 | assert(Load->getNumValues() == 2 && "Loads must carry a chain!")((Load->getNumValues() == 2 && "Loads must carry a chain!" ) ? static_cast<void> (0) : __assert_fail ("Load->getNumValues() == 2 && \"Loads must carry a chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18215, __PRETTY_FUNCTION__)); |
18216 | DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1)); |
18217 | |
18218 | MVT MaskVT = MVT::getVectorVT(MVT::i1, NumBitsToLoad); |
18219 | SDValue BitVec = DAG.getBitcast(MaskVT, Load); |
18220 | |
18221 | if (NumElts == 8) |
18222 | return DAG.getNode(ExtOpcode, dl, VT, BitVec); |
18223 | |
18224 | // we should take care to v4i1 and v2i1 |
18225 | |
18226 | MVT ExtVT = MVT::getVectorVT(VT.getScalarType(), 8); |
18227 | SDValue ExtVec = DAG.getNode(ExtOpcode, dl, ExtVT, BitVec); |
18228 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec, |
18229 | DAG.getIntPtrConstant(0, dl)); |
18230 | } |
18231 | |
18232 | assert(VT == MVT::v32i8 && "Unexpected extload type")((VT == MVT::v32i8 && "Unexpected extload type") ? static_cast <void> (0) : __assert_fail ("VT == MVT::v32i8 && \"Unexpected extload type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18232, __PRETTY_FUNCTION__)); |
18233 | |
18234 | SmallVector<SDValue, 2> Chains; |
18235 | |
18236 | SDValue BasePtr = Ld->getBasePtr(); |
18237 | SDValue LoadLo = DAG.getLoad(MVT::v16i1, dl, Ld->getChain(), |
18238 | Ld->getBasePtr(), |
18239 | Ld->getMemOperand()); |
18240 | Chains.push_back(LoadLo.getValue(1)); |
18241 | |
18242 | SDValue BasePtrHi = |
18243 | DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr, |
18244 | DAG.getConstant(2, dl, BasePtr.getValueType())); |
18245 | |
18246 | SDValue LoadHi = DAG.getLoad(MVT::v16i1, dl, Ld->getChain(), |
18247 | BasePtrHi, |
18248 | Ld->getMemOperand()); |
18249 | Chains.push_back(LoadHi.getValue(1)); |
18250 | SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains); |
18251 | DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain); |
18252 | |
18253 | SDValue Lo = DAG.getNode(ExtOpcode, dl, MVT::v16i8, LoadLo); |
18254 | SDValue Hi = DAG.getNode(ExtOpcode, dl, MVT::v16i8, LoadHi); |
18255 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v32i8, Lo, Hi); |
18256 | } |
18257 | |
18258 | // Lower vector extended loads using a shuffle. If SSSE3 is not available we |
18259 | // may emit an illegal shuffle but the expansion is still better than scalar |
18260 | // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise |
18261 | // we'll emit a shuffle and a arithmetic shift. |
18262 | // FIXME: Is the expansion actually better than scalar code? It doesn't seem so. |
18263 | // TODO: It is possible to support ZExt by zeroing the undef values during |
18264 | // the shuffle phase or after the shuffle. |
18265 | static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget &Subtarget, |
18266 | SelectionDAG &DAG) { |
18267 | MVT RegVT = Op.getSimpleValueType(); |
18268 | assert(RegVT.isVector() && "We only custom lower vector sext loads.")((RegVT.isVector() && "We only custom lower vector sext loads." ) ? static_cast<void> (0) : __assert_fail ("RegVT.isVector() && \"We only custom lower vector sext loads.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18268, __PRETTY_FUNCTION__)); |
18269 | assert(RegVT.isInteger() &&((RegVT.isInteger() && "We only custom lower integer vector sext loads." ) ? static_cast<void> (0) : __assert_fail ("RegVT.isInteger() && \"We only custom lower integer vector sext loads.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18270, __PRETTY_FUNCTION__)) |
18270 | "We only custom lower integer vector sext loads.")((RegVT.isInteger() && "We only custom lower integer vector sext loads." ) ? static_cast<void> (0) : __assert_fail ("RegVT.isInteger() && \"We only custom lower integer vector sext loads.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18270, __PRETTY_FUNCTION__)); |
18271 | |
18272 | // Nothing useful we can do without SSE2 shuffles. |
18273 | assert(Subtarget.hasSSE2() && "We only custom lower sext loads with SSE2.")((Subtarget.hasSSE2() && "We only custom lower sext loads with SSE2." ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSE2() && \"We only custom lower sext loads with SSE2.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18273, __PRETTY_FUNCTION__)); |
18274 | |
18275 | LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode()); |
18276 | SDLoc dl(Ld); |
18277 | EVT MemVT = Ld->getMemoryVT(); |
18278 | if (MemVT.getScalarType() == MVT::i1) |
18279 | return LowerExtended1BitVectorLoad(Op, Subtarget, DAG); |
18280 | |
18281 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
18282 | unsigned RegSz = RegVT.getSizeInBits(); |
18283 | |
18284 | ISD::LoadExtType Ext = Ld->getExtensionType(); |
18285 | |
18286 | assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)(((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD) && "Only anyext and sext are currently implemented." ) ? static_cast<void> (0) : __assert_fail ("(Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD) && \"Only anyext and sext are currently implemented.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18287, __PRETTY_FUNCTION__)) |
18287 | && "Only anyext and sext are currently implemented.")(((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD) && "Only anyext and sext are currently implemented." ) ? static_cast<void> (0) : __assert_fail ("(Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD) && \"Only anyext and sext are currently implemented.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18287, __PRETTY_FUNCTION__)); |
18288 | assert(MemVT != RegVT && "Cannot extend to the same type")((MemVT != RegVT && "Cannot extend to the same type") ? static_cast<void> (0) : __assert_fail ("MemVT != RegVT && \"Cannot extend to the same type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18288, __PRETTY_FUNCTION__)); |
18289 | assert(MemVT.isVector() && "Must load a vector from memory")((MemVT.isVector() && "Must load a vector from memory" ) ? static_cast<void> (0) : __assert_fail ("MemVT.isVector() && \"Must load a vector from memory\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18289, __PRETTY_FUNCTION__)); |
18290 | |
18291 | unsigned NumElems = RegVT.getVectorNumElements(); |
18292 | unsigned MemSz = MemVT.getSizeInBits(); |
18293 | assert(RegSz > MemSz && "Register size must be greater than the mem size")((RegSz > MemSz && "Register size must be greater than the mem size" ) ? static_cast<void> (0) : __assert_fail ("RegSz > MemSz && \"Register size must be greater than the mem size\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18293, __PRETTY_FUNCTION__)); |
18294 | |
18295 | if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget.hasInt256()) { |
18296 | // The only way in which we have a legal 256-bit vector result but not the |
18297 | // integer 256-bit operations needed to directly lower a sextload is if we |
18298 | // have AVX1 but not AVX2. In that case, we can always emit a sextload to |
18299 | // a 128-bit vector and a normal sign_extend to 256-bits that should get |
18300 | // correctly legalized. We do this late to allow the canonical form of |
18301 | // sextload to persist throughout the rest of the DAG combiner -- it wants |
18302 | // to fold together any extensions it can, and so will fuse a sign_extend |
18303 | // of an sextload into a sextload targeting a wider value. |
18304 | SDValue Load; |
18305 | if (MemSz == 128) { |
18306 | // Just switch this to a normal load. |
18307 | assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "((TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, " "it must be a legal 128-bit vector " "type!") ? static_cast< void> (0) : __assert_fail ("TLI.isTypeLegal(MemVT) && \"If the memory type is a 128-bit type, \" \"it must be a legal 128-bit vector \" \"type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18309, __PRETTY_FUNCTION__)) |
18308 | "it must be a legal 128-bit vector "((TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, " "it must be a legal 128-bit vector " "type!") ? static_cast< void> (0) : __assert_fail ("TLI.isTypeLegal(MemVT) && \"If the memory type is a 128-bit type, \" \"it must be a legal 128-bit vector \" \"type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18309, __PRETTY_FUNCTION__)) |
18309 | "type!")((TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, " "it must be a legal 128-bit vector " "type!") ? static_cast< void> (0) : __assert_fail ("TLI.isTypeLegal(MemVT) && \"If the memory type is a 128-bit type, \" \"it must be a legal 128-bit vector \" \"type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18309, __PRETTY_FUNCTION__)); |
18310 | Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(), |
18311 | Ld->getPointerInfo(), Ld->getAlignment(), |
18312 | Ld->getMemOperand()->getFlags()); |
18313 | } else { |
18314 | assert(MemSz < 128 &&((MemSz < 128 && "Can't extend a type wider than 128 bits to a 256 bit vector!" ) ? static_cast<void> (0) : __assert_fail ("MemSz < 128 && \"Can't extend a type wider than 128 bits to a 256 bit vector!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18315, __PRETTY_FUNCTION__)) |
18315 | "Can't extend a type wider than 128 bits to a 256 bit vector!")((MemSz < 128 && "Can't extend a type wider than 128 bits to a 256 bit vector!" ) ? static_cast<void> (0) : __assert_fail ("MemSz < 128 && \"Can't extend a type wider than 128 bits to a 256 bit vector!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18315, __PRETTY_FUNCTION__)); |
18316 | // Do an sext load to a 128-bit vector type. We want to use the same |
18317 | // number of elements, but elements half as wide. This will end up being |
18318 | // recursively lowered by this routine, but will succeed as we definitely |
18319 | // have all the necessary features if we're using AVX1. |
18320 | EVT HalfEltVT = |
18321 | EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2); |
18322 | EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems); |
18323 | Load = |
18324 | DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(), |
18325 | Ld->getPointerInfo(), MemVT, Ld->getAlignment(), |
18326 | Ld->getMemOperand()->getFlags()); |
18327 | } |
18328 | |
18329 | // Replace chain users with the new chain. |
18330 | assert(Load->getNumValues() == 2 && "Loads must carry a chain!")((Load->getNumValues() == 2 && "Loads must carry a chain!" ) ? static_cast<void> (0) : __assert_fail ("Load->getNumValues() == 2 && \"Loads must carry a chain!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18330, __PRETTY_FUNCTION__)); |
18331 | DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1)); |
18332 | |
18333 | // Finally, do a normal sign-extend to the desired register. |
18334 | return DAG.getSExtOrTrunc(Load, dl, RegVT); |
18335 | } |
18336 | |
18337 | // All sizes must be a power of two. |
18338 | assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&((isPowerOf2_32(RegSz * MemSz * NumElems) && "Non-power-of-two elements are not custom lowered!" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(RegSz * MemSz * NumElems) && \"Non-power-of-two elements are not custom lowered!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18339, __PRETTY_FUNCTION__)) |
18339 | "Non-power-of-two elements are not custom lowered!")((isPowerOf2_32(RegSz * MemSz * NumElems) && "Non-power-of-two elements are not custom lowered!" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(RegSz * MemSz * NumElems) && \"Non-power-of-two elements are not custom lowered!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18339, __PRETTY_FUNCTION__)); |
18340 | |
18341 | // Attempt to load the original value using scalar loads. |
18342 | // Find the largest scalar type that divides the total loaded size. |
18343 | MVT SclrLoadTy = MVT::i8; |
18344 | for (MVT Tp : MVT::integer_valuetypes()) { |
18345 | if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) { |
18346 | SclrLoadTy = Tp; |
18347 | } |
18348 | } |
18349 | |
18350 | // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64. |
18351 | if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 && |
18352 | (64 <= MemSz)) |
18353 | SclrLoadTy = MVT::f64; |
18354 | |
18355 | // Calculate the number of scalar loads that we need to perform |
18356 | // in order to load our vector from memory. |
18357 | unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits(); |
18358 | |
18359 | assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&(((Ext != ISD::SEXTLOAD || NumLoads == 1) && "Can only lower sext loads with a single scalar load!" ) ? static_cast<void> (0) : __assert_fail ("(Ext != ISD::SEXTLOAD || NumLoads == 1) && \"Can only lower sext loads with a single scalar load!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18360, __PRETTY_FUNCTION__)) |
18360 | "Can only lower sext loads with a single scalar load!")(((Ext != ISD::SEXTLOAD || NumLoads == 1) && "Can only lower sext loads with a single scalar load!" ) ? static_cast<void> (0) : __assert_fail ("(Ext != ISD::SEXTLOAD || NumLoads == 1) && \"Can only lower sext loads with a single scalar load!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18360, __PRETTY_FUNCTION__)); |
18361 | |
18362 | unsigned loadRegZize = RegSz; |
18363 | if (Ext == ISD::SEXTLOAD && RegSz >= 256) |
18364 | loadRegZize = 128; |
18365 | |
18366 | // Represent our vector as a sequence of elements which are the |
18367 | // largest scalar that we can load. |
18368 | EVT LoadUnitVecVT = EVT::getVectorVT( |
18369 | *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits()); |
18370 | |
18371 | // Represent the data using the same element type that is stored in |
18372 | // memory. In practice, we ''widen'' MemVT. |
18373 | EVT WideVecVT = |
18374 | EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(), |
18375 | loadRegZize / MemVT.getScalarSizeInBits()); |
18376 | |
18377 | assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&((WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() && "Invalid vector type") ? static_cast<void> (0) : __assert_fail ("WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() && \"Invalid vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18378, __PRETTY_FUNCTION__)) |
18378 | "Invalid vector type")((WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() && "Invalid vector type") ? static_cast<void> (0) : __assert_fail ("WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() && \"Invalid vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18378, __PRETTY_FUNCTION__)); |
18379 | |
18380 | // We can't shuffle using an illegal type. |
18381 | assert(TLI.isTypeLegal(WideVecVT) &&((TLI.isTypeLegal(WideVecVT) && "We only lower types that form legal widened vector types" ) ? static_cast<void> (0) : __assert_fail ("TLI.isTypeLegal(WideVecVT) && \"We only lower types that form legal widened vector types\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18382, __PRETTY_FUNCTION__)) |
18382 | "We only lower types that form legal widened vector types")((TLI.isTypeLegal(WideVecVT) && "We only lower types that form legal widened vector types" ) ? static_cast<void> (0) : __assert_fail ("TLI.isTypeLegal(WideVecVT) && \"We only lower types that form legal widened vector types\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18382, __PRETTY_FUNCTION__)); |
18383 | |
18384 | SmallVector<SDValue, 8> Chains; |
18385 | SDValue Ptr = Ld->getBasePtr(); |
18386 | SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl, |
18387 | TLI.getPointerTy(DAG.getDataLayout())); |
18388 | SDValue Res = DAG.getUNDEF(LoadUnitVecVT); |
18389 | |
18390 | for (unsigned i = 0; i < NumLoads; ++i) { |
18391 | // Perform a single load. |
18392 | SDValue ScalarLoad = |
18393 | DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(), |
18394 | Ld->getAlignment(), Ld->getMemOperand()->getFlags()); |
18395 | Chains.push_back(ScalarLoad.getValue(1)); |
18396 | // Create the first element type using SCALAR_TO_VECTOR in order to avoid |
18397 | // another round of DAGCombining. |
18398 | if (i == 0) |
18399 | Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad); |
18400 | else |
18401 | Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res, |
18402 | ScalarLoad, DAG.getIntPtrConstant(i, dl)); |
18403 | |
18404 | Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment); |
18405 | } |
18406 | |
18407 | SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains); |
18408 | |
18409 | // Bitcast the loaded value to a vector of the original element type, in |
18410 | // the size of the target vector type. |
18411 | SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res); |
18412 | unsigned SizeRatio = RegSz / MemSz; |
18413 | |
18414 | if (Ext == ISD::SEXTLOAD) { |
18415 | // If we have SSE4.1, we can directly emit a VSEXT node. |
18416 | if (Subtarget.hasSSE41()) { |
18417 | SDValue Sext = getExtendInVec(X86ISD::VSEXT, dl, RegVT, SlicedVec, DAG); |
18418 | DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF); |
18419 | return Sext; |
18420 | } |
18421 | |
18422 | // Otherwise we'll use SIGN_EXTEND_VECTOR_INREG to sign extend the lowest |
18423 | // lanes. |
18424 | assert(TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) &&((TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) && "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!" ) ? static_cast<void> (0) : __assert_fail ("TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) && \"We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18425, __PRETTY_FUNCTION__)) |
18425 | "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!")((TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) && "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!" ) ? static_cast<void> (0) : __assert_fail ("TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) && \"We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18425, __PRETTY_FUNCTION__)); |
18426 | |
18427 | SDValue Shuff = DAG.getSignExtendVectorInReg(SlicedVec, dl, RegVT); |
18428 | DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF); |
18429 | return Shuff; |
18430 | } |
18431 | |
18432 | // Redistribute the loaded elements into the different locations. |
18433 | SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1); |
18434 | for (unsigned i = 0; i != NumElems; ++i) |
18435 | ShuffleVec[i * SizeRatio] = i; |
18436 | |
18437 | SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec, |
18438 | DAG.getUNDEF(WideVecVT), ShuffleVec); |
18439 | |
18440 | // Bitcast to the requested type. |
18441 | Shuff = DAG.getBitcast(RegVT, Shuff); |
18442 | DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF); |
18443 | return Shuff; |
18444 | } |
18445 | |
18446 | /// Return true if node is an ISD::AND or ISD::OR of two X86ISD::SETCC nodes |
18447 | /// each of which has no other use apart from the AND / OR. |
18448 | static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) { |
18449 | Opc = Op.getOpcode(); |
18450 | if (Opc != ISD::OR && Opc != ISD::AND) |
18451 | return false; |
18452 | return (Op.getOperand(0).getOpcode() == X86ISD::SETCC && |
18453 | Op.getOperand(0).hasOneUse() && |
18454 | Op.getOperand(1).getOpcode() == X86ISD::SETCC && |
18455 | Op.getOperand(1).hasOneUse()); |
18456 | } |
18457 | |
18458 | /// Return true if node is an ISD::XOR of a X86ISD::SETCC and 1 and that the |
18459 | /// SETCC node has a single use. |
18460 | static bool isXor1OfSetCC(SDValue Op) { |
18461 | if (Op.getOpcode() != ISD::XOR) |
18462 | return false; |
18463 | if (isOneConstant(Op.getOperand(1))) |
18464 | return Op.getOperand(0).getOpcode() == X86ISD::SETCC && |
18465 | Op.getOperand(0).hasOneUse(); |
18466 | return false; |
18467 | } |
18468 | |
18469 | SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const { |
18470 | bool addTest = true; |
18471 | SDValue Chain = Op.getOperand(0); |
18472 | SDValue Cond = Op.getOperand(1); |
18473 | SDValue Dest = Op.getOperand(2); |
18474 | SDLoc dl(Op); |
18475 | SDValue CC; |
18476 | bool Inverted = false; |
18477 | |
18478 | if (Cond.getOpcode() == ISD::SETCC) { |
18479 | // Check for setcc([su]{add,sub,mul}o == 0). |
18480 | if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ && |
18481 | isNullConstant(Cond.getOperand(1)) && |
18482 | Cond.getOperand(0).getResNo() == 1 && |
18483 | (Cond.getOperand(0).getOpcode() == ISD::SADDO || |
18484 | Cond.getOperand(0).getOpcode() == ISD::UADDO || |
18485 | Cond.getOperand(0).getOpcode() == ISD::SSUBO || |
18486 | Cond.getOperand(0).getOpcode() == ISD::USUBO || |
18487 | Cond.getOperand(0).getOpcode() == ISD::SMULO || |
18488 | Cond.getOperand(0).getOpcode() == ISD::UMULO)) { |
18489 | Inverted = true; |
18490 | Cond = Cond.getOperand(0); |
18491 | } else { |
18492 | if (SDValue NewCond = LowerSETCC(Cond, DAG)) |
18493 | Cond = NewCond; |
18494 | } |
18495 | } |
18496 | #if 0 |
18497 | // FIXME: LowerXALUO doesn't handle these!! |
18498 | else if (Cond.getOpcode() == X86ISD::ADD || |
18499 | Cond.getOpcode() == X86ISD::SUB || |
18500 | Cond.getOpcode() == X86ISD::SMUL || |
18501 | Cond.getOpcode() == X86ISD::UMUL) |
18502 | Cond = LowerXALUO(Cond, DAG); |
18503 | #endif |
18504 | |
18505 | // Look pass (and (setcc_carry (cmp ...)), 1). |
18506 | if (Cond.getOpcode() == ISD::AND && |
18507 | Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY && |
18508 | isOneConstant(Cond.getOperand(1))) |
18509 | Cond = Cond.getOperand(0); |
18510 | |
18511 | // If condition flag is set by a X86ISD::CMP, then use it as the condition |
18512 | // setting operand in place of the X86ISD::SETCC. |
18513 | unsigned CondOpcode = Cond.getOpcode(); |
18514 | if (CondOpcode == X86ISD::SETCC || |
18515 | CondOpcode == X86ISD::SETCC_CARRY) { |
18516 | CC = Cond.getOperand(0); |
18517 | |
18518 | SDValue Cmp = Cond.getOperand(1); |
18519 | unsigned Opc = Cmp.getOpcode(); |
18520 | // FIXME: WHY THE SPECIAL CASING OF LogicalCmp?? |
18521 | if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) { |
18522 | Cond = Cmp; |
18523 | addTest = false; |
18524 | } else { |
18525 | switch (cast<ConstantSDNode>(CC)->getZExtValue()) { |
18526 | default: break; |
18527 | case X86::COND_O: |
18528 | case X86::COND_B: |
18529 | // These can only come from an arithmetic instruction with overflow, |
18530 | // e.g. SADDO, UADDO. |
18531 | Cond = Cond.getOperand(1); |
18532 | addTest = false; |
18533 | break; |
18534 | } |
18535 | } |
18536 | } |
18537 | CondOpcode = Cond.getOpcode(); |
18538 | if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO || |
18539 | CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO || |
18540 | ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) && |
18541 | Cond.getOperand(0).getValueType() != MVT::i8)) { |
18542 | SDValue LHS = Cond.getOperand(0); |
18543 | SDValue RHS = Cond.getOperand(1); |
18544 | unsigned X86Opcode; |
18545 | unsigned X86Cond; |
18546 | SDVTList VTs; |
18547 | // Keep this in sync with LowerXALUO, otherwise we might create redundant |
18548 | // instructions that can't be removed afterwards (i.e. X86ISD::ADD and |
18549 | // X86ISD::INC). |
18550 | switch (CondOpcode) { |
18551 | case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break; |
18552 | case ISD::SADDO: |
18553 | if (isOneConstant(RHS)) { |
18554 | X86Opcode = X86ISD::INC; X86Cond = X86::COND_O; |
18555 | break; |
18556 | } |
18557 | X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break; |
18558 | case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break; |
18559 | case ISD::SSUBO: |
18560 | if (isOneConstant(RHS)) { |
18561 | X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O; |
18562 | break; |
18563 | } |
18564 | X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break; |
18565 | case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break; |
18566 | case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break; |
18567 | default: llvm_unreachable("unexpected overflowing operator")::llvm::llvm_unreachable_internal("unexpected overflowing operator" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18567); |
18568 | } |
18569 | if (Inverted) |
18570 | X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond); |
18571 | if (CondOpcode == ISD::UMULO) |
18572 | VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(), |
18573 | MVT::i32); |
18574 | else |
18575 | VTs = DAG.getVTList(LHS.getValueType(), MVT::i32); |
18576 | |
18577 | SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS); |
18578 | |
18579 | if (CondOpcode == ISD::UMULO) |
18580 | Cond = X86Op.getValue(2); |
18581 | else |
18582 | Cond = X86Op.getValue(1); |
18583 | |
18584 | CC = DAG.getConstant(X86Cond, dl, MVT::i8); |
18585 | addTest = false; |
18586 | } else { |
18587 | unsigned CondOpc; |
18588 | if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) { |
18589 | SDValue Cmp = Cond.getOperand(0).getOperand(1); |
18590 | if (CondOpc == ISD::OR) { |
18591 | // Also, recognize the pattern generated by an FCMP_UNE. We can emit |
18592 | // two branches instead of an explicit OR instruction with a |
18593 | // separate test. |
18594 | if (Cmp == Cond.getOperand(1).getOperand(1) && |
18595 | isX86LogicalCmp(Cmp)) { |
18596 | CC = Cond.getOperand(0).getOperand(0); |
18597 | Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(), |
18598 | Chain, Dest, CC, Cmp); |
18599 | CC = Cond.getOperand(1).getOperand(0); |
18600 | Cond = Cmp; |
18601 | addTest = false; |
18602 | } |
18603 | } else { // ISD::AND |
18604 | // Also, recognize the pattern generated by an FCMP_OEQ. We can emit |
18605 | // two branches instead of an explicit AND instruction with a |
18606 | // separate test. However, we only do this if this block doesn't |
18607 | // have a fall-through edge, because this requires an explicit |
18608 | // jmp when the condition is false. |
18609 | if (Cmp == Cond.getOperand(1).getOperand(1) && |
18610 | isX86LogicalCmp(Cmp) && |
18611 | Op.getNode()->hasOneUse()) { |
18612 | X86::CondCode CCode = |
18613 | (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0); |
18614 | CCode = X86::GetOppositeBranchCondition(CCode); |
18615 | CC = DAG.getConstant(CCode, dl, MVT::i8); |
18616 | SDNode *User = *Op.getNode()->use_begin(); |
18617 | // Look for an unconditional branch following this conditional branch. |
18618 | // We need this because we need to reverse the successors in order |
18619 | // to implement FCMP_OEQ. |
18620 | if (User->getOpcode() == ISD::BR) { |
18621 | SDValue FalseBB = User->getOperand(1); |
18622 | SDNode *NewBR = |
18623 | DAG.UpdateNodeOperands(User, User->getOperand(0), Dest); |
18624 | assert(NewBR == User)((NewBR == User) ? static_cast<void> (0) : __assert_fail ("NewBR == User", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18624, __PRETTY_FUNCTION__)); |
18625 | (void)NewBR; |
18626 | Dest = FalseBB; |
18627 | |
18628 | Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(), |
18629 | Chain, Dest, CC, Cmp); |
18630 | X86::CondCode CCode = |
18631 | (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0); |
18632 | CCode = X86::GetOppositeBranchCondition(CCode); |
18633 | CC = DAG.getConstant(CCode, dl, MVT::i8); |
18634 | Cond = Cmp; |
18635 | addTest = false; |
18636 | } |
18637 | } |
18638 | } |
18639 | } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) { |
18640 | // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition. |
18641 | // It should be transformed during dag combiner except when the condition |
18642 | // is set by a arithmetics with overflow node. |
18643 | X86::CondCode CCode = |
18644 | (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0); |
18645 | CCode = X86::GetOppositeBranchCondition(CCode); |
18646 | CC = DAG.getConstant(CCode, dl, MVT::i8); |
18647 | Cond = Cond.getOperand(0).getOperand(1); |
18648 | addTest = false; |
18649 | } else if (Cond.getOpcode() == ISD::SETCC && |
18650 | cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) { |
18651 | // For FCMP_OEQ, we can emit |
18652 | // two branches instead of an explicit AND instruction with a |
18653 | // separate test. However, we only do this if this block doesn't |
18654 | // have a fall-through edge, because this requires an explicit |
18655 | // jmp when the condition is false. |
18656 | if (Op.getNode()->hasOneUse()) { |
18657 | SDNode *User = *Op.getNode()->use_begin(); |
18658 | // Look for an unconditional branch following this conditional branch. |
18659 | // We need this because we need to reverse the successors in order |
18660 | // to implement FCMP_OEQ. |
18661 | if (User->getOpcode() == ISD::BR) { |
18662 | SDValue FalseBB = User->getOperand(1); |
18663 | SDNode *NewBR = |
18664 | DAG.UpdateNodeOperands(User, User->getOperand(0), Dest); |
18665 | assert(NewBR == User)((NewBR == User) ? static_cast<void> (0) : __assert_fail ("NewBR == User", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18665, __PRETTY_FUNCTION__)); |
18666 | (void)NewBR; |
18667 | Dest = FalseBB; |
18668 | |
18669 | SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32, |
18670 | Cond.getOperand(0), Cond.getOperand(1)); |
18671 | Cmp = ConvertCmpIfNecessary(Cmp, DAG); |
18672 | CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8); |
18673 | Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(), |
18674 | Chain, Dest, CC, Cmp); |
18675 | CC = DAG.getConstant(X86::COND_P, dl, MVT::i8); |
18676 | Cond = Cmp; |
18677 | addTest = false; |
18678 | } |
18679 | } |
18680 | } else if (Cond.getOpcode() == ISD::SETCC && |
18681 | cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) { |
18682 | // For FCMP_UNE, we can emit |
18683 | // two branches instead of an explicit AND instruction with a |
18684 | // separate test. However, we only do this if this block doesn't |
18685 | // have a fall-through edge, because this requires an explicit |
18686 | // jmp when the condition is false. |
18687 | if (Op.getNode()->hasOneUse()) { |
18688 | SDNode *User = *Op.getNode()->use_begin(); |
18689 | // Look for an unconditional branch following this conditional branch. |
18690 | // We need this because we need to reverse the successors in order |
18691 | // to implement FCMP_UNE. |
18692 | if (User->getOpcode() == ISD::BR) { |
18693 | SDValue FalseBB = User->getOperand(1); |
18694 | SDNode *NewBR = |
18695 | DAG.UpdateNodeOperands(User, User->getOperand(0), Dest); |
18696 | assert(NewBR == User)((NewBR == User) ? static_cast<void> (0) : __assert_fail ("NewBR == User", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18696, __PRETTY_FUNCTION__)); |
18697 | (void)NewBR; |
18698 | |
18699 | SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32, |
18700 | Cond.getOperand(0), Cond.getOperand(1)); |
18701 | Cmp = ConvertCmpIfNecessary(Cmp, DAG); |
18702 | CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8); |
18703 | Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(), |
18704 | Chain, Dest, CC, Cmp); |
18705 | CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8); |
18706 | Cond = Cmp; |
18707 | addTest = false; |
18708 | Dest = FalseBB; |
18709 | } |
18710 | } |
18711 | } |
18712 | } |
18713 | |
18714 | if (addTest) { |
18715 | // Look pass the truncate if the high bits are known zero. |
18716 | if (isTruncWithZeroHighBitsInput(Cond, DAG)) |
18717 | Cond = Cond.getOperand(0); |
18718 | |
18719 | // We know the result is compared against zero. Try to match it to BT. |
18720 | if (Cond.hasOneUse()) { |
18721 | if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG)) { |
18722 | CC = NewSetCC.getOperand(0); |
18723 | Cond = NewSetCC.getOperand(1); |
18724 | addTest = false; |
18725 | } |
18726 | } |
18727 | } |
18728 | |
18729 | if (addTest) { |
18730 | X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE; |
18731 | CC = DAG.getConstant(X86Cond, dl, MVT::i8); |
18732 | Cond = EmitTest(Cond, X86Cond, dl, DAG); |
18733 | } |
18734 | Cond = ConvertCmpIfNecessary(Cond, DAG); |
18735 | return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(), |
18736 | Chain, Dest, CC, Cond); |
18737 | } |
18738 | |
18739 | // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets. |
18740 | // Calls to _alloca are needed to probe the stack when allocating more than 4k |
18741 | // bytes in one go. Touching the stack at 4K increments is necessary to ensure |
18742 | // that the guard pages used by the OS virtual memory manager are allocated in |
18743 | // correct sequence. |
18744 | SDValue |
18745 | X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, |
18746 | SelectionDAG &DAG) const { |
18747 | MachineFunction &MF = DAG.getMachineFunction(); |
18748 | bool SplitStack = MF.shouldSplitStack(); |
18749 | bool EmitStackProbe = !getStackProbeSymbolName(MF).empty(); |
18750 | bool Lower = (Subtarget.isOSWindows() && !Subtarget.isTargetMachO()) || |
18751 | SplitStack || EmitStackProbe; |
18752 | SDLoc dl(Op); |
18753 | |
18754 | // Get the inputs. |
18755 | SDNode *Node = Op.getNode(); |
18756 | SDValue Chain = Op.getOperand(0); |
18757 | SDValue Size = Op.getOperand(1); |
18758 | unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue(); |
18759 | EVT VT = Node->getValueType(0); |
18760 | |
18761 | // Chain the dynamic stack allocation so that it doesn't modify the stack |
18762 | // pointer when other instructions are using the stack. |
18763 | Chain = DAG.getCALLSEQ_START(Chain, 0, 0, dl); |
18764 | |
18765 | bool Is64Bit = Subtarget.is64Bit(); |
18766 | MVT SPTy = getPointerTy(DAG.getDataLayout()); |
18767 | |
18768 | SDValue Result; |
18769 | if (!Lower) { |
18770 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
18771 | unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore(); |
18772 | assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"((SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and" " not tell us which reg is the stack pointer!") ? static_cast <void> (0) : __assert_fail ("SPReg && \"Target cannot require DYNAMIC_STACKALLOC expansion and\" \" not tell us which reg is the stack pointer!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18773, __PRETTY_FUNCTION__)) |
18773 | " not tell us which reg is the stack pointer!")((SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and" " not tell us which reg is the stack pointer!") ? static_cast <void> (0) : __assert_fail ("SPReg && \"Target cannot require DYNAMIC_STACKALLOC expansion and\" \" not tell us which reg is the stack pointer!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18773, __PRETTY_FUNCTION__)); |
18774 | |
18775 | SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT); |
18776 | Chain = SP.getValue(1); |
18777 | const TargetFrameLowering &TFI = *Subtarget.getFrameLowering(); |
18778 | unsigned StackAlign = TFI.getStackAlignment(); |
18779 | Result = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value |
18780 | if (Align > StackAlign) |
18781 | Result = DAG.getNode(ISD::AND, dl, VT, Result, |
18782 | DAG.getConstant(-(uint64_t)Align, dl, VT)); |
18783 | Chain = DAG.getCopyToReg(Chain, dl, SPReg, Result); // Output chain |
18784 | } else if (SplitStack) { |
18785 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
18786 | |
18787 | if (Is64Bit) { |
18788 | // The 64 bit implementation of segmented stacks needs to clobber both r10 |
18789 | // r11. This makes it impossible to use it along with nested parameters. |
18790 | const Function *F = MF.getFunction(); |
18791 | for (const auto &A : F->args()) { |
18792 | if (A.hasNestAttr()) |
18793 | report_fatal_error("Cannot use segmented stacks with functions that " |
18794 | "have nested arguments."); |
18795 | } |
18796 | } |
18797 | |
18798 | const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy); |
18799 | unsigned Vreg = MRI.createVirtualRegister(AddrRegClass); |
18800 | Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size); |
18801 | Result = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain, |
18802 | DAG.getRegister(Vreg, SPTy)); |
18803 | } else { |
18804 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); |
18805 | Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Size); |
18806 | MF.getInfo<X86MachineFunctionInfo>()->setHasWinAlloca(true); |
18807 | |
18808 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
18809 | unsigned SPReg = RegInfo->getStackRegister(); |
18810 | SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy); |
18811 | Chain = SP.getValue(1); |
18812 | |
18813 | if (Align) { |
18814 | SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0), |
18815 | DAG.getConstant(-(uint64_t)Align, dl, VT)); |
18816 | Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP); |
18817 | } |
18818 | |
18819 | Result = SP; |
18820 | } |
18821 | |
18822 | Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true), |
18823 | DAG.getIntPtrConstant(0, dl, true), SDValue(), dl); |
18824 | |
18825 | SDValue Ops[2] = {Result, Chain}; |
18826 | return DAG.getMergeValues(Ops, dl); |
18827 | } |
18828 | |
18829 | SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const { |
18830 | MachineFunction &MF = DAG.getMachineFunction(); |
18831 | auto PtrVT = getPointerTy(MF.getDataLayout()); |
18832 | X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); |
18833 | |
18834 | const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); |
18835 | SDLoc DL(Op); |
18836 | |
18837 | if (!Subtarget.is64Bit() || |
18838 | Subtarget.isCallingConvWin64(MF.getFunction()->getCallingConv())) { |
18839 | // vastart just stores the address of the VarArgsFrameIndex slot into the |
18840 | // memory location argument. |
18841 | SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); |
18842 | return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1), |
18843 | MachinePointerInfo(SV)); |
18844 | } |
18845 | |
18846 | // __va_list_tag: |
18847 | // gp_offset (0 - 6 * 8) |
18848 | // fp_offset (48 - 48 + 8 * 16) |
18849 | // overflow_arg_area (point to parameters coming in memory). |
18850 | // reg_save_area |
18851 | SmallVector<SDValue, 8> MemOps; |
18852 | SDValue FIN = Op.getOperand(1); |
18853 | // Store gp_offset |
18854 | SDValue Store = DAG.getStore( |
18855 | Op.getOperand(0), DL, |
18856 | DAG.getConstant(FuncInfo->getVarArgsGPOffset(), DL, MVT::i32), FIN, |
18857 | MachinePointerInfo(SV)); |
18858 | MemOps.push_back(Store); |
18859 | |
18860 | // Store fp_offset |
18861 | FIN = DAG.getMemBasePlusOffset(FIN, 4, DL); |
18862 | Store = DAG.getStore( |
18863 | Op.getOperand(0), DL, |
18864 | DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL, MVT::i32), FIN, |
18865 | MachinePointerInfo(SV, 4)); |
18866 | MemOps.push_back(Store); |
18867 | |
18868 | // Store ptr to overflow_arg_area |
18869 | FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL)); |
18870 | SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); |
18871 | Store = |
18872 | DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN, MachinePointerInfo(SV, 8)); |
18873 | MemOps.push_back(Store); |
18874 | |
18875 | // Store ptr to reg_save_area. |
18876 | FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant( |
18877 | Subtarget.isTarget64BitLP64() ? 8 : 4, DL)); |
18878 | SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT); |
18879 | Store = DAG.getStore( |
18880 | Op.getOperand(0), DL, RSFIN, FIN, |
18881 | MachinePointerInfo(SV, Subtarget.isTarget64BitLP64() ? 16 : 12)); |
18882 | MemOps.push_back(Store); |
18883 | return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps); |
18884 | } |
18885 | |
18886 | SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const { |
18887 | assert(Subtarget.is64Bit() &&((Subtarget.is64Bit() && "LowerVAARG only handles 64-bit va_arg!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.is64Bit() && \"LowerVAARG only handles 64-bit va_arg!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18888, __PRETTY_FUNCTION__)) |
18888 | "LowerVAARG only handles 64-bit va_arg!")((Subtarget.is64Bit() && "LowerVAARG only handles 64-bit va_arg!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.is64Bit() && \"LowerVAARG only handles 64-bit va_arg!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18888, __PRETTY_FUNCTION__)); |
18889 | assert(Op.getNumOperands() == 4)((Op.getNumOperands() == 4) ? static_cast<void> (0) : __assert_fail ("Op.getNumOperands() == 4", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18889, __PRETTY_FUNCTION__)); |
18890 | |
18891 | MachineFunction &MF = DAG.getMachineFunction(); |
18892 | if (Subtarget.isCallingConvWin64(MF.getFunction()->getCallingConv())) |
18893 | // The Win64 ABI uses char* instead of a structure. |
18894 | return DAG.expandVAArg(Op.getNode()); |
18895 | |
18896 | SDValue Chain = Op.getOperand(0); |
18897 | SDValue SrcPtr = Op.getOperand(1); |
18898 | const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); |
18899 | unsigned Align = Op.getConstantOperandVal(3); |
18900 | SDLoc dl(Op); |
18901 | |
18902 | EVT ArgVT = Op.getNode()->getValueType(0); |
18903 | Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext()); |
18904 | uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy); |
18905 | uint8_t ArgMode; |
18906 | |
18907 | // Decide which area this value should be read from. |
18908 | // TODO: Implement the AMD64 ABI in its entirety. This simple |
18909 | // selection mechanism works only for the basic types. |
18910 | if (ArgVT == MVT::f80) { |
18911 | llvm_unreachable("va_arg for f80 not yet implemented")::llvm::llvm_unreachable_internal("va_arg for f80 not yet implemented" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18911); |
18912 | } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) { |
18913 | ArgMode = 2; // Argument passed in XMM register. Use fp_offset. |
18914 | } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) { |
18915 | ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset. |
18916 | } else { |
18917 | llvm_unreachable("Unhandled argument type in LowerVAARG")::llvm::llvm_unreachable_internal("Unhandled argument type in LowerVAARG" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18917); |
18918 | } |
18919 | |
18920 | if (ArgMode == 2) { |
18921 | // Sanity Check: Make sure using fp_offset makes sense. |
18922 | assert(!Subtarget.useSoftFloat() &&((!Subtarget.useSoftFloat() && !(MF.getFunction()-> hasFnAttribute(Attribute::NoImplicitFloat)) && Subtarget .hasSSE1()) ? static_cast<void> (0) : __assert_fail ("!Subtarget.useSoftFloat() && !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) && Subtarget.hasSSE1()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18924, __PRETTY_FUNCTION__)) |
18923 | !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&((!Subtarget.useSoftFloat() && !(MF.getFunction()-> hasFnAttribute(Attribute::NoImplicitFloat)) && Subtarget .hasSSE1()) ? static_cast<void> (0) : __assert_fail ("!Subtarget.useSoftFloat() && !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) && Subtarget.hasSSE1()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18924, __PRETTY_FUNCTION__)) |
18924 | Subtarget.hasSSE1())((!Subtarget.useSoftFloat() && !(MF.getFunction()-> hasFnAttribute(Attribute::NoImplicitFloat)) && Subtarget .hasSSE1()) ? static_cast<void> (0) : __assert_fail ("!Subtarget.useSoftFloat() && !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) && Subtarget.hasSSE1()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18924, __PRETTY_FUNCTION__)); |
18925 | } |
18926 | |
18927 | // Insert VAARG_64 node into the DAG |
18928 | // VAARG_64 returns two values: Variable Argument Address, Chain |
18929 | SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32), |
18930 | DAG.getConstant(ArgMode, dl, MVT::i8), |
18931 | DAG.getConstant(Align, dl, MVT::i32)}; |
18932 | SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other); |
18933 | SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl, |
18934 | VTs, InstOps, MVT::i64, |
18935 | MachinePointerInfo(SV), |
18936 | /*Align=*/0, |
18937 | /*Volatile=*/false, |
18938 | /*ReadMem=*/true, |
18939 | /*WriteMem=*/true); |
18940 | Chain = VAARG.getValue(1); |
18941 | |
18942 | // Load the next argument and return it |
18943 | return DAG.getLoad(ArgVT, dl, Chain, VAARG, MachinePointerInfo()); |
18944 | } |
18945 | |
18946 | static SDValue LowerVACOPY(SDValue Op, const X86Subtarget &Subtarget, |
18947 | SelectionDAG &DAG) { |
18948 | // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows, |
18949 | // where a va_list is still an i8*. |
18950 | assert(Subtarget.is64Bit() && "This code only handles 64-bit va_copy!")((Subtarget.is64Bit() && "This code only handles 64-bit va_copy!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.is64Bit() && \"This code only handles 64-bit va_copy!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18950, __PRETTY_FUNCTION__)); |
18951 | if (Subtarget.isCallingConvWin64( |
18952 | DAG.getMachineFunction().getFunction()->getCallingConv())) |
18953 | // Probably a Win64 va_copy. |
18954 | return DAG.expandVACopy(Op.getNode()); |
18955 | |
18956 | SDValue Chain = Op.getOperand(0); |
18957 | SDValue DstPtr = Op.getOperand(1); |
18958 | SDValue SrcPtr = Op.getOperand(2); |
18959 | const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue(); |
18960 | const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue(); |
18961 | SDLoc DL(Op); |
18962 | |
18963 | return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr, |
18964 | DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false, |
18965 | false, false, |
18966 | MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV)); |
18967 | } |
18968 | |
18969 | /// Handle vector element shifts where the shift amount is a constant. |
18970 | /// Takes immediate version of shift as input. |
18971 | static SDValue getTargetVShiftByConstNode(unsigned Opc, const SDLoc &dl, MVT VT, |
18972 | SDValue SrcOp, uint64_t ShiftAmt, |
18973 | SelectionDAG &DAG) { |
18974 | MVT ElementType = VT.getVectorElementType(); |
18975 | |
18976 | // Bitcast the source vector to the output type, this is mainly necessary for |
18977 | // vXi8/vXi64 shifts. |
18978 | if (VT != SrcOp.getSimpleValueType()) |
18979 | SrcOp = DAG.getBitcast(VT, SrcOp); |
18980 | |
18981 | // Fold this packed shift into its first operand if ShiftAmt is 0. |
18982 | if (ShiftAmt == 0) |
18983 | return SrcOp; |
18984 | |
18985 | // Check for ShiftAmt >= element width |
18986 | if (ShiftAmt >= ElementType.getSizeInBits()) { |
18987 | if (Opc == X86ISD::VSRAI) |
18988 | ShiftAmt = ElementType.getSizeInBits() - 1; |
18989 | else |
18990 | return DAG.getConstant(0, dl, VT); |
18991 | } |
18992 | |
18993 | assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)(((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD ::VSRAI) && "Unknown target vector shift-by-constant node" ) ? static_cast<void> (0) : __assert_fail ("(Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI) && \"Unknown target vector shift-by-constant node\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18994, __PRETTY_FUNCTION__)) |
18994 | && "Unknown target vector shift-by-constant node")(((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD ::VSRAI) && "Unknown target vector shift-by-constant node" ) ? static_cast<void> (0) : __assert_fail ("(Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI) && \"Unknown target vector shift-by-constant node\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 18994, __PRETTY_FUNCTION__)); |
18995 | |
18996 | // Fold this packed vector shift into a build vector if SrcOp is a |
18997 | // vector of Constants or UNDEFs. |
18998 | if (ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) { |
18999 | SmallVector<SDValue, 8> Elts; |
19000 | unsigned NumElts = SrcOp->getNumOperands(); |
19001 | ConstantSDNode *ND; |
19002 | |
19003 | switch(Opc) { |
19004 | default: llvm_unreachable("Unknown opcode!")::llvm::llvm_unreachable_internal("Unknown opcode!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 19004); |
19005 | case X86ISD::VSHLI: |
19006 | for (unsigned i=0; i!=NumElts; ++i) { |
19007 | SDValue CurrentOp = SrcOp->getOperand(i); |
19008 | if (CurrentOp->isUndef()) { |
19009 | Elts.push_back(CurrentOp); |
19010 | continue; |
19011 | } |
19012 | ND = cast<ConstantSDNode>(CurrentOp); |
19013 | const APInt &C = ND->getAPIntValue(); |
19014 | Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType)); |
19015 | } |
19016 | break; |
19017 | case X86ISD::VSRLI: |
19018 | for (unsigned i=0; i!=NumElts; ++i) { |
19019 | SDValue CurrentOp = SrcOp->getOperand(i); |
19020 | if (CurrentOp->isUndef()) { |
19021 | Elts.push_back(CurrentOp); |
19022 | continue; |
19023 | } |
19024 | ND = cast<ConstantSDNode>(CurrentOp); |
19025 | const APInt &C = ND->getAPIntValue(); |
19026 | Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType)); |
19027 | } |
19028 | break; |
19029 | case X86ISD::VSRAI: |
19030 | for (unsigned i=0; i!=NumElts; ++i) { |
19031 | SDValue CurrentOp = SrcOp->getOperand(i); |
19032 | if (CurrentOp->isUndef()) { |
19033 | Elts.push_back(CurrentOp); |
19034 | continue; |
19035 | } |
19036 | ND = cast<ConstantSDNode>(CurrentOp); |
19037 | const APInt &C = ND->getAPIntValue(); |
19038 | Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType)); |
19039 | } |
19040 | break; |
19041 | } |
19042 | |
19043 | return DAG.getBuildVector(VT, dl, Elts); |
19044 | } |
19045 | |
19046 | return DAG.getNode(Opc, dl, VT, SrcOp, |
19047 | DAG.getConstant(ShiftAmt, dl, MVT::i8)); |
19048 | } |
19049 | |
19050 | /// Handle vector element shifts where the shift amount may or may not be a |
19051 | /// constant. Takes immediate version of shift as input. |
19052 | static SDValue getTargetVShiftNode(unsigned Opc, const SDLoc &dl, MVT VT, |
19053 | SDValue SrcOp, SDValue ShAmt, |
19054 | const X86Subtarget &Subtarget, |
19055 | SelectionDAG &DAG) { |
19056 | MVT SVT = ShAmt.getSimpleValueType(); |
19057 | assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!")(((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!" ) ? static_cast<void> (0) : __assert_fail ("(SVT == MVT::i32 || SVT == MVT::i64) && \"Unexpected value type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 19057, __PRETTY_FUNCTION__)); |
19058 | |
19059 | // Catch shift-by-constant. |
19060 | if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt)) |
19061 | return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp, |
19062 | CShAmt->getZExtValue(), DAG); |
19063 | |
19064 | // Change opcode to non-immediate version |
19065 | switch (Opc) { |
19066 | default: llvm_unreachable("Unknown target vector shift node")::llvm::llvm_unreachable_internal("Unknown target vector shift node" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 19066); |
19067 | case X86ISD::VSHLI: Opc = X86ISD::VSHL; break; |
19068 | case X86ISD::VSRLI: Opc = X86ISD::VSRL; break; |
19069 | case X86ISD::VSRAI: Opc = X86ISD::VSRA; break; |
19070 | } |
19071 | |
19072 | // Need to build a vector containing shift amount. |
19073 | // SSE/AVX packed shifts only use the lower 64-bit of the shift count. |
19074 | // +=================+============+=======================================+ |
19075 | // | ShAmt is | HasSSE4.1? | Construct ShAmt vector as | |
19076 | // +=================+============+=======================================+ |
19077 | // | i64 | Yes, No | Use ShAmt as lowest elt | |
19078 | // | i32 | Yes | zero-extend in-reg | |
19079 | // | (i32 zext(i16)) | Yes | zero-extend in-reg | |
19080 | // | i16/i32 | No | v4i32 build_vector(ShAmt, 0, ud, ud)) | |
19081 | // +=================+============+=======================================+ |
19082 | |
19083 | if (SVT == MVT::i64) |
19084 | ShAmt = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(ShAmt), MVT::v2i64, ShAmt); |
19085 | else if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND && |
19086 | ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) { |
19087 | ShAmt = ShAmt.getOperand(0); |
19088 | ShAmt = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(ShAmt), MVT::v8i16, ShAmt); |
19089 | ShAmt = DAG.getZeroExtendVectorInReg(ShAmt, SDLoc(ShAmt), MVT::v2i64); |
19090 | } else if (Subtarget.hasSSE41() && |
19091 | ShAmt.getOpcode() == ISD::EXTRACT_VECTOR_ELT) { |
19092 | ShAmt = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(ShAmt), MVT::v4i32, ShAmt); |
19093 | ShAmt = DAG.getZeroExtendVectorInReg(ShAmt, SDLoc(ShAmt), MVT::v2i64); |
19094 | } else { |
19095 | SmallVector<SDValue, 4> ShOps = {ShAmt, DAG.getConstant(0, dl, SVT), |
19096 | DAG.getUNDEF(SVT), DAG.getUNDEF(SVT)}; |
19097 | ShAmt = DAG.getBuildVector(MVT::v4i32, dl, ShOps); |
19098 | } |
19099 | |
19100 | // The return type has to be a 128-bit type with the same element |
19101 | // type as the input type. |
19102 | MVT EltVT = VT.getVectorElementType(); |
19103 | MVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits()); |
19104 | |
19105 | ShAmt = DAG.getBitcast(ShVT, ShAmt); |
19106 | return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt); |
19107 | } |
19108 | |
19109 | /// \brief Return Mask with the necessary casting or extending |
19110 | /// for \p Mask according to \p MaskVT when lowering masking intrinsics |
19111 | static SDValue getMaskNode(SDValue Mask, MVT MaskVT, |
19112 | const X86Subtarget &Subtarget, SelectionDAG &DAG, |
19113 | const SDLoc &dl) { |
19114 | |
19115 | if (isAllOnesConstant(Mask)) |
19116 | return DAG.getTargetConstant(1, dl, MaskVT); |
19117 | if (X86::isZeroNode(Mask)) |
19118 | return DAG.getTargetConstant(0, dl, MaskVT); |
19119 | |
19120 | if (MaskVT.bitsGT(Mask.getSimpleValueType())) { |
19121 | // Mask should be extended |
19122 | Mask = DAG.getNode(ISD::ANY_EXTEND, dl, |
19123 | MVT::getIntegerVT(MaskVT.getSizeInBits()), Mask); |
19124 | } |
19125 | |
19126 | if (Mask.getSimpleValueType() == MVT::i64 && Subtarget.is32Bit()) { |
19127 | if (MaskVT == MVT::v64i1) { |
19128 | assert(Subtarget.hasBWI() && "Expected AVX512BW target!")((Subtarget.hasBWI() && "Expected AVX512BW target!") ? static_cast<void> (0) : __assert_fail ("Subtarget.hasBWI() && \"Expected AVX512BW target!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 19128, __PRETTY_FUNCTION__)); |
19129 | // In case 32bit mode, bitcast i64 is illegal, extend/split it. |
19130 | SDValue Lo, Hi; |
19131 | Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Mask, |
19132 | DAG.getConstant(0, dl, MVT::i32)); |
19133 | Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Mask, |
19134 | DAG.getConstant(1, dl, MVT::i32)); |
19135 | |
19136 | Lo = DAG.getBitcast(MVT::v32i1, Lo); |
19137 | Hi = DAG.getBitcast(MVT::v32i1, Hi); |
19138 | |
19139 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v64i1, Lo, Hi); |
19140 | } else { |
19141 | // MaskVT require < 64bit. Truncate mask (should succeed in any case), |
19142 | // and bitcast. |
19143 | MVT TruncVT = MVT::getIntegerVT(MaskVT.getSizeInBits()); |
19144 | return DAG.getBitcast(MaskVT, |
19145 | DAG.getNode(ISD::TRUNCATE, dl, TruncVT, Mask)); |
19146 | } |
19147 | |
19148 | } else { |
19149 | MVT BitcastVT = MVT::getVectorVT(MVT::i1, |
19150 | Mask.getSimpleValueType().getSizeInBits()); |
19151 | // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements |
19152 | // are extracted by EXTRACT_SUBVECTOR. |
19153 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT, |
19154 | DAG.getBitcast(BitcastVT, Mask), |
19155 | DAG.getIntPtrConstant(0, dl)); |
19156 | } |
19157 | } |
19158 | |
19159 | /// \brief Return (and \p Op, \p Mask) for compare instructions or |
19160 | /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the |
19161 | /// necessary casting or extending for \p Mask when lowering masking intrinsics |
19162 | static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask, |
19163 | SDValue PreservedSrc, |
19164 | const X86Subtarget &Subtarget, |
19165 | SelectionDAG &DAG) { |
19166 | MVT VT = Op.getSimpleValueType(); |
19167 | MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements()); |
19168 | unsigned OpcodeSelect = ISD::VSELECT; |
19169 | SDLoc dl(Op); |
19170 | |
19171 | if (isAllOnesConstant(Mask)) |
19172 | return Op; |
19173 | |
19174 | SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); |
19175 | |
19176 | switch (Op.getOpcode()) { |
19177 | default: break; |
19178 | case X86ISD::PCMPEQM: |
19179 | case X86ISD::PCMPGTM: |
19180 | case X86ISD::CMPM: |
19181 | case X86ISD::CMPMU: |
19182 | return DAG.getNode(ISD::AND, dl, VT, Op, VMask); |
19183 | case X86ISD::VFPCLASS: |
19184 | case X86ISD::VFPCLASSS: |
19185 | return DAG.getNode(ISD::OR, dl, VT, Op, VMask); |
19186 | case X86ISD::VTRUNC: |
19187 | case X86ISD::VTRUNCS: |
19188 | case X86ISD::VTRUNCUS: |
19189 | case X86ISD::CVTPS2PH: |
19190 | // We can't use ISD::VSELECT here because it is not always "Legal" |
19191 | // for the destination type. For example vpmovqb require only AVX512 |
19192 | // and vselect that can operate on byte element type require BWI |
19193 | OpcodeSelect = X86ISD::SELECT; |
19194 | break; |
19195 | } |
19196 | if (PreservedSrc.isUndef()) |
19197 | PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl); |
19198 | return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc); |
19199 | } |
19200 | |
19201 | /// \brief Creates an SDNode for a predicated scalar operation. |
19202 | /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc). |
19203 | /// The mask is coming as MVT::i8 and it should be transformed |
19204 | /// to MVT::v1i1 while lowering masking intrinsics. |
19205 | /// The main difference between ScalarMaskingNode and VectorMaskingNode is using |
19206 | /// "X86select" instead of "vselect". We just can't create the "vselect" node |
19207 | /// for a scalar instruction. |
19208 | static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask, |
19209 | SDValue PreservedSrc, |
19210 | const X86Subtarget &Subtarget, |
19211 | SelectionDAG &DAG) { |
19212 | |
19213 | if (auto *MaskConst = dyn_cast<ConstantSDNode>(Mask)) |
19214 | if (MaskConst->getZExtValue() & 0x1) |
19215 | return Op; |
19216 | |
19217 | MVT VT = Op.getSimpleValueType(); |
19218 | SDLoc dl(Op); |
19219 | |
19220 | SDValue IMask = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i1, Mask); |
19221 | if (Op.getOpcode() == X86ISD::FSETCCM || |
19222 | Op.getOpcode() == X86ISD::FSETCCM_RND) |
19223 | return DAG.getNode(ISD::AND, dl, VT, Op, IMask); |
19224 | if (Op.getOpcode() == X86ISD::VFPCLASSS) |
19225 | return DAG.getNode(ISD::OR, dl, VT, Op, IMask); |
19226 | |
19227 | if (PreservedSrc.isUndef()) |
19228 | PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl); |
19229 | return DAG.getNode(X86ISD::SELECTS, dl, VT, IMask, Op, PreservedSrc); |
19230 | } |
19231 | |
19232 | static int getSEHRegistrationNodeSize(const Function *Fn) { |
19233 | if (!Fn->hasPersonalityFn()) |
19234 | report_fatal_error( |
19235 | "querying registration node size for function without personality"); |
19236 | // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See |
19237 | // WinEHStatePass for the full struct definition. |
19238 | switch (classifyEHPersonality(Fn->getPersonalityFn())) { |
19239 | case EHPersonality::MSVC_X86SEH: return 24; |
19240 | case EHPersonality::MSVC_CXX: return 16; |
19241 | default: break; |
19242 | } |
19243 | report_fatal_error( |
19244 | "can only recover FP for 32-bit MSVC EH personality functions"); |
19245 | } |
19246 | |
19247 | /// When the MSVC runtime transfers control to us, either to an outlined |
19248 | /// function or when returning to a parent frame after catching an exception, we |
19249 | /// recover the parent frame pointer by doing arithmetic on the incoming EBP. |
19250 | /// Here's the math: |
19251 | /// RegNodeBase = EntryEBP - RegNodeSize |
19252 | /// ParentFP = RegNodeBase - ParentFrameOffset |
19253 | /// Subtracting RegNodeSize takes us to the offset of the registration node, and |
19254 | /// subtracting the offset (negative on x86) takes us back to the parent FP. |
19255 | static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn, |
19256 | SDValue EntryEBP) { |
19257 | MachineFunction &MF = DAG.getMachineFunction(); |
19258 | SDLoc dl; |
19259 | |
19260 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
19261 | MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout()); |
19262 | |
19263 | // It's possible that the parent function no longer has a personality function |
19264 | // if the exceptional code was optimized away, in which case we just return |
19265 | // the incoming EBP. |
19266 | if (!Fn->hasPersonalityFn()) |
19267 | return EntryEBP; |
19268 | |
19269 | // Get an MCSymbol that will ultimately resolve to the frame offset of the EH |
19270 | // registration, or the .set_setframe offset. |
19271 | MCSymbol *OffsetSym = |
19272 | MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol( |
19273 | GlobalValue::dropLLVMManglingEscape(Fn->getName())); |
19274 | SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT); |
19275 | SDValue ParentFrameOffset = |
19276 | DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal); |
19277 | |
19278 | // Return EntryEBP + ParentFrameOffset for x64. This adjusts from RSP after |
19279 | // prologue to RBP in the parent function. |
19280 | const X86Subtarget &Subtarget = |
19281 | static_cast<const X86Subtarget &>(DAG.getSubtarget()); |
19282 | if (Subtarget.is64Bit()) |
19283 | return DAG.getNode(ISD::ADD, dl, PtrVT, EntryEBP, ParentFrameOffset); |
19284 | |
19285 | int RegNodeSize = getSEHRegistrationNodeSize(Fn); |
19286 | // RegNodeBase = EntryEBP - RegNodeSize |
19287 | // ParentFP = RegNodeBase - ParentFrameOffset |
19288 | SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP, |
19289 | DAG.getConstant(RegNodeSize, dl, PtrVT)); |
19290 | return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, ParentFrameOffset); |
19291 | } |
19292 | |
19293 | static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget &Subtarget, |
19294 | SelectionDAG &DAG) { |
19295 | // Helper to detect if the operand is CUR_DIRECTION rounding mode. |
19296 | auto isRoundModeCurDirection = [](SDValue Rnd) { |
19297 | if (!isa<ConstantSDNode>(Rnd)) |
19298 | return false; |
19299 | |
19300 | unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue(); |
19301 | return Round == X86::STATIC_ROUNDING::CUR_DIRECTION; |
19302 | }; |
19303 | |
19304 | SDLoc dl(Op); |
19305 | unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); |
19306 | MVT VT = Op.getSimpleValueType(); |
19307 | const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo); |
19308 | if (IntrData) { |
19309 | switch(IntrData->Type) { |
19310 | case INTR_TYPE_1OP: |
19311 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1)); |
19312 | case INTR_TYPE_2OP: |
19313 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1), |
19314 | Op.getOperand(2)); |
19315 | case INTR_TYPE_3OP: |
19316 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1), |
19317 | Op.getOperand(2), Op.getOperand(3)); |
19318 | case INTR_TYPE_4OP: |
19319 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1), |
19320 | Op.getOperand(2), Op.getOperand(3), Op.getOperand(4)); |
19321 | case INTR_TYPE_1OP_MASK_RM: { |
19322 | SDValue Src = Op.getOperand(1); |
19323 | SDValue PassThru = Op.getOperand(2); |
19324 | SDValue Mask = Op.getOperand(3); |
19325 | SDValue RoundingMode; |
19326 | // We always add rounding mode to the Node. |
19327 | // If the rounding mode is not specified, we add the |
19328 | // "current direction" mode. |
19329 | if (Op.getNumOperands() == 4) |
19330 | RoundingMode = |
19331 | DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32); |
19332 | else |
19333 | RoundingMode = Op.getOperand(4); |
19334 | assert(IntrData->Opc1 == 0 && "Unexpected second opcode!")((IntrData->Opc1 == 0 && "Unexpected second opcode!" ) ? static_cast<void> (0) : __assert_fail ("IntrData->Opc1 == 0 && \"Unexpected second opcode!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 19334, __PRETTY_FUNCTION__)); |
19335 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src, |
19336 | RoundingMode), |
19337 | Mask, PassThru, Subtarget, DAG); |
19338 | } |
19339 | case INTR_TYPE_1OP_MASK: { |
19340 | SDValue Src = Op.getOperand(1); |
19341 | SDValue PassThru = Op.getOperand(2); |
19342 | SDValue Mask = Op.getOperand(3); |
19343 | // We add rounding mode to the Node when |
19344 | // - RM Opcode is specified and |
19345 | // - RM is not "current direction". |
19346 | unsigned IntrWithRoundingModeOpcode = IntrData->Opc1; |
19347 | if (IntrWithRoundingModeOpcode != 0) { |
19348 | SDValue Rnd = Op.getOperand(4); |
19349 | if (!isRoundModeCurDirection(Rnd)) { |
19350 | return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode, |
19351 | dl, Op.getValueType(), |
19352 | Src, Rnd), |
19353 | Mask, PassThru, Subtarget, DAG); |
19354 | } |
19355 | } |
19356 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src), |
19357 | Mask, PassThru, Subtarget, DAG); |
19358 | } |
19359 | case INTR_TYPE_SCALAR_MASK: { |
19360 | SDValue Src1 = Op.getOperand(1); |
19361 | SDValue Src2 = Op.getOperand(2); |
19362 | SDValue passThru = Op.getOperand(3); |
19363 | SDValue Mask = Op.getOperand(4); |
19364 | unsigned IntrWithRoundingModeOpcode = IntrData->Opc1; |
19365 | if (IntrWithRoundingModeOpcode != 0) { |
19366 | SDValue Rnd = Op.getOperand(5); |
19367 | if (!isRoundModeCurDirection(Rnd)) |
19368 | return getScalarMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode, |
19369 | dl, VT, Src1, Src2, Rnd), |
19370 | Mask, passThru, Subtarget, DAG); |
19371 | } |
19372 | return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2), |
19373 | Mask, passThru, Subtarget, DAG); |
19374 | } |
19375 | case INTR_TYPE_SCALAR_MASK_RM: { |
19376 | SDValue Src1 = Op.getOperand(1); |
19377 | SDValue Src2 = Op.getOperand(2); |
19378 | SDValue Src0 = Op.getOperand(3); |
19379 | SDValue Mask = Op.getOperand(4); |
19380 | // There are 2 kinds of intrinsics in this group: |
19381 | // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands |
19382 | // (2) With rounding mode and sae - 7 operands. |
19383 | if (Op.getNumOperands() == 6) { |
19384 | SDValue Sae = Op.getOperand(5); |
19385 | return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2, |
19386 | Sae), |
19387 | Mask, Src0, Subtarget, DAG); |
19388 | } |
19389 | assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form")((Op.getNumOperands() == 7 && "Unexpected intrinsic form" ) ? static_cast<void> (0) : __assert_fail ("Op.getNumOperands() == 7 && \"Unexpected intrinsic form\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 19389, __PRETTY_FUNCTION__)); |
19390 | SDValue RoundingMode = Op.getOperand(5); |
19391 | SDValue Sae = Op.getOperand(6); |
19392 | return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2, |
19393 | RoundingMode, Sae), |
19394 | Mask, Src0, Subtarget, DAG); |
19395 | } |
19396 | case INTR_TYPE_2OP_MASK: |
19397 | case INTR_TYPE_2OP_IMM8_MASK: { |
19398 | SDValue Src1 = Op.getOperand(1); |
19399 | SDValue Src2 = Op.getOperand(2); |
19400 | SDValue PassThru = Op.getOperand(3); |
19401 | SDValue Mask = Op.getOperand(4); |
19402 | |
19403 | if (IntrData->Type == INTR_TYPE_2OP_IMM8_MASK) |
19404 | Src2 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src2); |
19405 | |
19406 | // We specify 2 possible opcodes for intrinsics with rounding modes. |
19407 | // First, we check if the intrinsic may have non-default rounding mode, |
19408 | // (IntrData->Opc1 != 0), then we check the rounding mode operand. |
19409 | unsigned IntrWithRoundingModeOpcode = IntrData->Opc1; |
19410 | if (IntrWithRoundingModeOpcode != 0) { |
19411 | SDValue Rnd = Op.getOperand(5); |
19412 | if (!isRoundModeCurDirection(Rnd)) { |
19413 | return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode, |
19414 | dl, Op.getValueType(), |
19415 | Src1, Src2, Rnd), |
19416 | Mask, PassThru, Subtarget, DAG); |
19417 | } |
19418 | } |
19419 | // TODO: Intrinsics should have fast-math-flags to propagate. |
19420 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2), |
19421 | Mask, PassThru, Subtarget, DAG); |
19422 | } |
19423 | case INTR_TYPE_2OP_MASK_RM: { |
19424 | SDValue Src1 = Op.getOperand(1); |
19425 | SDValue Src2 = Op.getOperand(2); |
19426 | SDValue PassThru = Op.getOperand(3); |
19427 | SDValue Mask = Op.getOperand(4); |
19428 | // We specify 2 possible modes for intrinsics, with/without rounding |
19429 | // modes. |
19430 | // First, we check if the intrinsic have rounding mode (6 operands), |
19431 | // if not, we set rounding mode to "current". |
19432 | SDValue Rnd; |
19433 | if (Op.getNumOperands() == 6) |
19434 | Rnd = Op.getOperand(5); |
19435 | else |
19436 | Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32); |
19437 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, |
19438 | Src1, Src2, Rnd), |
19439 | Mask, PassThru, Subtarget, DAG); |
19440 | } |
19441 | case INTR_TYPE_3OP_SCALAR_MASK_RM: { |
19442 | SDValue Src1 = Op.getOperand(1); |
19443 | SDValue Src2 = Op.getOperand(2); |
19444 | SDValue Src3 = Op.getOperand(3); |
19445 | SDValue PassThru = Op.getOperand(4); |
19446 | SDValue Mask = Op.getOperand(5); |
19447 | SDValue Sae = Op.getOperand(6); |
19448 | |
19449 | return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, |
19450 | Src2, Src3, Sae), |
19451 | Mask, PassThru, Subtarget, DAG); |
19452 | } |
19453 | case INTR_TYPE_3OP_MASK_RM: { |
19454 | SDValue Src1 = Op.getOperand(1); |
19455 | SDValue Src2 = Op.getOperand(2); |
19456 | SDValue Imm = Op.getOperand(3); |
19457 | SDValue PassThru = Op.getOperand(4); |
19458 | SDValue Mask = Op.getOperand(5); |
19459 | // We specify 2 possible modes for intrinsics, with/without rounding |
19460 | // modes. |
19461 | // First, we check if the intrinsic have rounding mode (7 operands), |
19462 | // if not, we set rounding mode to "current". |
19463 | SDValue Rnd; |
19464 | if (Op.getNumOperands() == 7) |
19465 | Rnd = Op.getOperand(6); |
19466 | else |
19467 | Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32); |
19468 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, |
19469 | Src1, Src2, Imm, Rnd), |
19470 | Mask, PassThru, Subtarget, DAG); |
19471 | } |
19472 | case INTR_TYPE_3OP_IMM8_MASK: |
19473 | case INTR_TYPE_3OP_MASK: { |
19474 | SDValue Src1 = Op.getOperand(1); |
19475 | SDValue Src2 = Op.getOperand(2); |
19476 | SDValue Src3 = Op.getOperand(3); |
19477 | SDValue PassThru = Op.getOperand(4); |
19478 | SDValue Mask = Op.getOperand(5); |
19479 | |
19480 | if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK) |
19481 | Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3); |
19482 | |
19483 | // We specify 2 possible opcodes for intrinsics with rounding modes. |
19484 | // First, we check if the intrinsic may have non-default rounding mode, |
19485 | // (IntrData->Opc1 != 0), then we check the rounding mode operand. |
19486 | unsigned IntrWithRoundingModeOpcode = IntrData->Opc1; |
19487 | if (IntrWithRoundingModeOpcode != 0) { |
19488 | SDValue Rnd = Op.getOperand(6); |
19489 | if (!isRoundModeCurDirection(Rnd)) { |
19490 | return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode, |
19491 | dl, Op.getValueType(), |
19492 | Src1, Src2, Src3, Rnd), |
19493 | Mask, PassThru, Subtarget, DAG); |
19494 | } |
19495 | } |
19496 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, |
19497 | Src1, Src2, Src3), |
19498 | Mask, PassThru, Subtarget, DAG); |
19499 | } |
19500 | case VPERM_2OP_MASK : { |
19501 | SDValue Src1 = Op.getOperand(1); |
19502 | SDValue Src2 = Op.getOperand(2); |
19503 | SDValue PassThru = Op.getOperand(3); |
19504 | SDValue Mask = Op.getOperand(4); |
19505 | |
19506 | // Swap Src1 and Src2 in the node creation |
19507 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src2, Src1), |
19508 | Mask, PassThru, Subtarget, DAG); |
19509 | } |
19510 | case VPERM_3OP_MASKZ: |
19511 | case VPERM_3OP_MASK:{ |
19512 | MVT VT = Op.getSimpleValueType(); |
19513 | // Src2 is the PassThru |
19514 | SDValue Src1 = Op.getOperand(1); |
19515 | // PassThru needs to be the same type as the destination in order |
19516 | // to pattern match correctly. |
19517 | SDValue Src2 = DAG.getBitcast(VT, Op.getOperand(2)); |
19518 | SDValue Src3 = Op.getOperand(3); |
19519 | SDValue Mask = Op.getOperand(4); |
19520 | SDValue PassThru = SDValue(); |
19521 | |
19522 | // set PassThru element |
19523 | if (IntrData->Type == VPERM_3OP_MASKZ) |
19524 | PassThru = getZeroVector(VT, Subtarget, DAG, dl); |
19525 | else |
19526 | PassThru = Src2; |
19527 | |
19528 | // Swap Src1 and Src2 in the node creation |
19529 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, |
19530 | dl, Op.getValueType(), |
19531 | Src2, Src1, Src3), |
19532 | Mask, PassThru, Subtarget, DAG); |
19533 | } |
19534 | case FMA_OP_MASK3: |
19535 | case FMA_OP_MASKZ: |
19536 | case FMA_OP_MASK: { |
19537 | SDValue Src1 = Op.getOperand(1); |
19538 | SDValue Src2 = Op.getOperand(2); |
19539 | SDValue Src3 = Op.getOperand(3); |
19540 | SDValue Mask = Op.getOperand(4); |
19541 | MVT VT = Op.getSimpleValueType(); |
19542 | SDValue PassThru = SDValue(); |
19543 | |
19544 | // set PassThru element |
19545 | if (IntrData->Type == FMA_OP_MASKZ) |
19546 | PassThru = getZeroVector(VT, Subtarget, DAG, dl); |
19547 | else if (IntrData->Type == FMA_OP_MASK3) |
19548 | PassThru = Src3; |
19549 | else |
19550 | PassThru = Src1; |
19551 | |
19552 | // We specify 2 possible opcodes for intrinsics with rounding modes. |
19553 | // First, we check if the intrinsic may have non-default rounding mode, |
19554 | // (IntrData->Opc1 != 0), then we check the rounding mode operand. |
19555 | unsigned IntrWithRoundingModeOpcode = IntrData->Opc1; |
19556 | if (IntrWithRoundingModeOpcode != 0) { |
19557 | SDValue Rnd = Op.getOperand(5); |
19558 | if (!isRoundModeCurDirection(Rnd)) |
19559 | return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode, |
19560 | dl, Op.getValueType(), |
19561 | Src1, Src2, Src3, Rnd), |
19562 | Mask, PassThru, Subtarget, DAG); |
19563 | } |
19564 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, |
19565 | dl, Op.getValueType(), |
19566 | Src1, Src2, Src3), |
19567 | Mask, PassThru, Subtarget, DAG); |
19568 | } |
19569 | case FMA_OP_SCALAR_MASK: |
19570 | case FMA_OP_SCALAR_MASK3: |
19571 | case FMA_OP_SCALAR_MASKZ: { |
19572 | SDValue Src1 = Op.getOperand(1); |
19573 | SDValue Src2 = Op.getOperand(2); |
19574 | SDValue Src3 = Op.getOperand(3); |
19575 | SDValue Mask = Op.getOperand(4); |
19576 | MVT VT = Op.getSimpleValueType(); |
19577 | SDValue PassThru = SDValue(); |
19578 | |
19579 | // set PassThru element |
19580 | if (IntrData->Type == FMA_OP_SCALAR_MASKZ) |
19581 | PassThru = getZeroVector(VT, Subtarget, DAG, dl); |
19582 | else if (IntrData->Type == FMA_OP_SCALAR_MASK3) |
19583 | PassThru = Src3; |
19584 | else |
19585 | PassThru = Src1; |
19586 | |
19587 | SDValue Rnd = Op.getOperand(5); |
19588 | return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, |
19589 | Op.getValueType(), Src1, Src2, |
19590 | Src3, Rnd), |
19591 | Mask, PassThru, Subtarget, DAG); |
19592 | } |
19593 | case TERLOG_OP_MASK: |
19594 | case TERLOG_OP_MASKZ: { |
19595 | SDValue Src1 = Op.getOperand(1); |
19596 | SDValue Src2 = Op.getOperand(2); |
19597 | SDValue Src3 = Op.getOperand(3); |
19598 | SDValue Src4 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(4)); |
19599 | SDValue Mask = Op.getOperand(5); |
19600 | MVT VT = Op.getSimpleValueType(); |
19601 | SDValue PassThru = Src1; |
19602 | // Set PassThru element. |
19603 | if (IntrData->Type == TERLOG_OP_MASKZ) |
19604 | PassThru = getZeroVector(VT, Subtarget, DAG, dl); |
19605 | |
19606 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, |
19607 | Src1, Src2, Src3, Src4), |
19608 | Mask, PassThru, Subtarget, DAG); |
19609 | } |
19610 | case CVTPD2PS: |
19611 | // ISD::FP_ROUND has a second argument that indicates if the truncation |
19612 | // does not change the value. Set it to 0 since it can change. |
19613 | return DAG.getNode(IntrData->Opc0, dl, VT, Op.getOperand(1), |
19614 | DAG.getIntPtrConstant(0, dl)); |
19615 | case CVTPD2PS_MASK: { |
19616 | SDValue Src = Op.getOperand(1); |
19617 | SDValue PassThru = Op.getOperand(2); |
19618 | SDValue Mask = Op.getOperand(3); |
19619 | // We add rounding mode to the Node when |
19620 | // - RM Opcode is specified and |
19621 | // - RM is not "current direction". |
19622 | unsigned IntrWithRoundingModeOpcode = IntrData->Opc1; |
19623 | if (IntrWithRoundingModeOpcode != 0) { |
19624 | SDValue Rnd = Op.getOperand(4); |
19625 | if (!isRoundModeCurDirection(Rnd)) { |
19626 | return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode, |
19627 | dl, Op.getValueType(), |
19628 | Src, Rnd), |
19629 | Mask, PassThru, Subtarget, DAG); |
19630 | } |
19631 | } |
19632 | assert(IntrData->Opc0 == ISD::FP_ROUND && "Unexpected opcode!")((IntrData->Opc0 == ISD::FP_ROUND && "Unexpected opcode!" ) ? static_cast<void> (0) : __assert_fail ("IntrData->Opc0 == ISD::FP_ROUND && \"Unexpected opcode!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 19632, __PRETTY_FUNCTION__)); |
19633 | // ISD::FP_ROUND has a second argument that indicates if the truncation |
19634 | // does not change the value. Set it to 0 since it can change. |
19635 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src, |
19636 | DAG.getIntPtrConstant(0, dl)), |
19637 | Mask, PassThru, Subtarget, DAG); |
19638 | } |
19639 | case FPCLASS: { |
19640 | // FPclass intrinsics with mask |
19641 | SDValue Src1 = Op.getOperand(1); |
19642 | MVT VT = Src1.getSimpleValueType(); |
19643 | MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements()); |
19644 | SDValue Imm = Op.getOperand(2); |
19645 | SDValue Mask = Op.getOperand(3); |
19646 | MVT BitcastVT = MVT::getVectorVT(MVT::i1, |
19647 | Mask.getSimpleValueType().getSizeInBits()); |
19648 | SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Imm); |
19649 | SDValue FPclassMask = getVectorMaskingNode(FPclass, Mask, |
19650 | DAG.getTargetConstant(0, dl, MaskVT), |
19651 | Subtarget, DAG); |
19652 | SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT, |
19653 | DAG.getUNDEF(BitcastVT), FPclassMask, |
19654 | DAG.getIntPtrConstant(0, dl)); |
19655 | return DAG.getBitcast(Op.getValueType(), Res); |
19656 | } |
19657 | case FPCLASSS: { |
19658 | SDValue Src1 = Op.getOperand(1); |
19659 | SDValue Imm = Op.getOperand(2); |
19660 | SDValue Mask = Op.getOperand(3); |
19661 | SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::v1i1, Src1, Imm); |
19662 | SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask, |
19663 | DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG); |
19664 | return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i8, FPclassMask, |
19665 | DAG.getIntPtrConstant(0, dl)); |
19666 | } |
19667 | case CMP_MASK: |
19668 | case CMP_MASK_CC: { |
19669 | // Comparison intrinsics with masks. |
19670 | // Example of transformation: |
19671 | // (i8 (int_x86_avx512_mask_pcmpeq_q_128 |
19672 | // (v2i64 %a), (v2i64 %b), (i8 %mask))) -> |
19673 | // (i8 (bitcast |
19674 | // (v8i1 (insert_subvector undef, |
19675 | // (v2i1 (and (PCMPEQM %a, %b), |
19676 | // (extract_subvector |
19677 | // (v8i1 (bitcast %mask)), 0))), 0)))) |
19678 | MVT VT = Op.getOperand(1).getSimpleValueType(); |
19679 | MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements()); |
19680 | SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3); |
19681 | MVT BitcastVT = MVT::getVectorVT(MVT::i1, |
19682 | Mask.getSimpleValueType().getSizeInBits()); |
19683 | SDValue Cmp; |
19684 | if (IntrData->Type == CMP_MASK_CC) { |
19685 | SDValue CC = Op.getOperand(3); |
19686 | CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC); |
19687 | // We specify 2 possible opcodes for intrinsics with rounding modes. |
19688 | // First, we check if the intrinsic may have non-default rounding mode, |
19689 | // (IntrData->Opc1 != 0), then we check the rounding mode operand. |
19690 | if (IntrData->Opc1 != 0) { |
19691 | SDValue Rnd = Op.getOperand(5); |
19692 | if (!isRoundModeCurDirection(Rnd)) |
19693 | Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1), |
19694 | Op.getOperand(2), CC, Rnd); |
19695 | } |
19696 | //default rounding mode |
19697 | if(!Cmp.getNode()) |
19698 | Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1), |
19699 | Op.getOperand(2), CC); |
19700 | |
19701 | } else { |
19702 | assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!")((IntrData->Type == CMP_MASK && "Unexpected intrinsic type!" ) ? static_cast<void> (0) : __assert_fail ("IntrData->Type == CMP_MASK && \"Unexpected intrinsic type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 19702, __PRETTY_FUNCTION__)); |
19703 | Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1), |
19704 | Op.getOperand(2)); |
19705 | } |
19706 | SDValue CmpMask = getVectorMaskingNode(Cmp, Mask, |
19707 | DAG.getTargetConstant(0, dl, |
19708 | MaskVT), |
19709 | Subtarget, DAG); |
19710 | SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT, |
19711 | DAG.getUNDEF(BitcastVT), CmpMask, |
19712 | DAG.getIntPtrConstant(0, dl)); |
19713 | return DAG.getBitcast(Op.getValueType(), Res); |
19714 | } |
19715 | case CMP_MASK_SCALAR_CC: { |
19716 | SDValue Src1 = Op.getOperand(1); |
19717 | SDValue Src2 = Op.getOperand(2); |
19718 | SDValue CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(3)); |
19719 | SDValue Mask = Op.getOperand(4); |
19720 | |
19721 | SDValue Cmp; |
19722 | if (IntrData->Opc1 != 0) { |
19723 | SDValue Rnd = Op.getOperand(5); |
19724 | if (!isRoundModeCurDirection(Rnd)) |
19725 | Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::v1i1, Src1, Src2, CC, Rnd); |
19726 | } |
19727 | //default rounding mode |
19728 | if(!Cmp.getNode()) |
19729 | Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::v1i1, Src1, Src2, CC); |
19730 | |
19731 | SDValue CmpMask = getScalarMaskingNode(Cmp, Mask, |
19732 | DAG.getTargetConstant(0, dl, |
19733 | MVT::i1), |
19734 | Subtarget, DAG); |
19735 | return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i8, CmpMask, |
19736 | DAG.getIntPtrConstant(0, dl)); |
19737 | } |
19738 | case COMI: { // Comparison intrinsics |
19739 | ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1; |
19740 | SDValue LHS = Op.getOperand(1); |
19741 | SDValue RHS = Op.getOperand(2); |
19742 | SDValue Comi = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS); |
19743 | SDValue InvComi = DAG.getNode(IntrData->Opc0, dl, MVT::i32, RHS, LHS); |
19744 | SDValue SetCC; |
19745 | switch (CC) { |
19746 | case ISD::SETEQ: { // (ZF = 0 and PF = 0) |
19747 | SetCC = getSETCC(X86::COND_E, Comi, dl, DAG); |
19748 | SDValue SetNP = getSETCC(X86::COND_NP, Comi, dl, DAG); |
19749 | SetCC = DAG.getNode(ISD::AND, dl, MVT::i8, SetCC, SetNP); |
19750 | break; |
19751 | } |
19752 | case ISD::SETNE: { // (ZF = 1 or PF = 1) |
19753 | SetCC = getSETCC(X86::COND_NE, Comi, dl, DAG); |
19754 | SDValue SetP = getSETCC(X86::COND_P, Comi, dl, DAG); |
19755 | SetCC = DAG.getNode(ISD::OR, dl, MVT::i8, SetCC, SetP); |
19756 | break; |
19757 | } |
19758 | case ISD::SETGT: // (CF = 0 and ZF = 0) |
19759 | SetCC = getSETCC(X86::COND_A, Comi, dl, DAG); |
19760 | break; |
19761 | case ISD::SETLT: { // The condition is opposite to GT. Swap the operands. |
19762 | SetCC = getSETCC(X86::COND_A, InvComi, dl, DAG); |
19763 | break; |
19764 | } |
19765 | case ISD::SETGE: // CF = 0 |
19766 | SetCC = getSETCC(X86::COND_AE, Comi, dl, DAG); |
19767 | break; |
19768 | case ISD::SETLE: // The condition is opposite to GE. Swap the operands. |
19769 | SetCC = getSETCC(X86::COND_AE, InvComi, dl, DAG); |
19770 | break; |
19771 | default: |
19772 | llvm_unreachable("Unexpected illegal condition!")::llvm::llvm_unreachable_internal("Unexpected illegal condition!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 19772); |
19773 | } |
19774 | return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC); |
19775 | } |
19776 | case COMI_RM: { // Comparison intrinsics with Sae |
19777 | SDValue LHS = Op.getOperand(1); |
19778 | SDValue RHS = Op.getOperand(2); |
19779 | unsigned CondVal = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue(); |
19780 | SDValue Sae = Op.getOperand(4); |
19781 | |
19782 | SDValue FCmp; |
19783 | if (isRoundModeCurDirection(Sae)) |
19784 | FCmp = DAG.getNode(X86ISD::FSETCCM, dl, MVT::v1i1, LHS, RHS, |
19785 | DAG.getConstant(CondVal, dl, MVT::i8)); |
19786 | else |
19787 | FCmp = DAG.getNode(X86ISD::FSETCCM_RND, dl, MVT::v1i1, LHS, RHS, |
19788 | DAG.getConstant(CondVal, dl, MVT::i8), Sae); |
19789 | return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i32, FCmp, |
19790 | DAG.getIntPtrConstant(0, dl)); |
19791 | } |
19792 | case VSHIFT: |
19793 | return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(), |
19794 | Op.getOperand(1), Op.getOperand(2), Subtarget, |
19795 | DAG); |
19796 | case COMPRESS_EXPAND_IN_REG: { |
19797 | SDValue Mask = Op.getOperand(3); |
19798 | SDValue DataToCompress = Op.getOperand(1); |
19799 | SDValue PassThru = Op.getOperand(2); |
19800 | if (isAllOnesConstant(Mask)) // return data as is |
19801 | return Op.getOperand(1); |
19802 | |
19803 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, |
19804 | DataToCompress), |
19805 | Mask, PassThru, Subtarget, DAG); |
19806 | } |
19807 | case BROADCASTM: { |
19808 | SDValue Mask = Op.getOperand(1); |
19809 | MVT MaskVT = MVT::getVectorVT(MVT::i1, |
19810 | Mask.getSimpleValueType().getSizeInBits()); |
19811 | Mask = DAG.getBitcast(MaskVT, Mask); |
19812 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Mask); |
19813 | } |
19814 | case KUNPCK: { |
19815 | MVT VT = Op.getSimpleValueType(); |
19816 | MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getSizeInBits()/2); |
19817 | |
19818 | SDValue Src1 = getMaskNode(Op.getOperand(1), MaskVT, Subtarget, DAG, dl); |
19819 | SDValue Src2 = getMaskNode(Op.getOperand(2), MaskVT, Subtarget, DAG, dl); |
19820 | // Arguments should be swapped. |
19821 | SDValue Res = DAG.getNode(IntrData->Opc0, dl, |
19822 | MVT::getVectorVT(MVT::i1, VT.getSizeInBits()), |
19823 | Src2, Src1); |
19824 | return DAG.getBitcast(VT, Res); |
19825 | } |
19826 | case MASK_BINOP: { |
19827 | MVT VT = Op.getSimpleValueType(); |
19828 | MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getSizeInBits()); |
19829 | |
19830 | SDValue Src1 = getMaskNode(Op.getOperand(1), MaskVT, Subtarget, DAG, dl); |
19831 | SDValue Src2 = getMaskNode(Op.getOperand(2), MaskVT, Subtarget, DAG, dl); |
19832 | SDValue Res = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Src2); |
19833 | return DAG.getBitcast(VT, Res); |
19834 | } |
19835 | case FIXUPIMMS: |
19836 | case FIXUPIMMS_MASKZ: |
19837 | case FIXUPIMM: |
19838 | case FIXUPIMM_MASKZ:{ |
19839 | SDValue Src1 = Op.getOperand(1); |
19840 | SDValue Src2 = Op.getOperand(2); |
19841 | SDValue Src3 = Op.getOperand(3); |
19842 | SDValue Imm = Op.getOperand(4); |
19843 | SDValue Mask = Op.getOperand(5); |
19844 | SDValue Passthru = (IntrData->Type == FIXUPIMM || IntrData->Type == FIXUPIMMS ) ? |
19845 | Src1 : getZeroVector(VT, Subtarget, DAG, dl); |
19846 | // We specify 2 possible modes for intrinsics, with/without rounding |
19847 | // modes. |
19848 | // First, we check if the intrinsic have rounding mode (7 operands), |
19849 | // if not, we set rounding mode to "current". |
19850 | SDValue Rnd; |
19851 | if (Op.getNumOperands() == 7) |
19852 | Rnd = Op.getOperand(6); |
19853 | else |
19854 | Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32); |
19855 | if (IntrData->Type == FIXUPIMM || IntrData->Type == FIXUPIMM_MASKZ) |
19856 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, |
19857 | Src1, Src2, Src3, Imm, Rnd), |
19858 | Mask, Passthru, Subtarget, DAG); |
19859 | else // Scalar - FIXUPIMMS, FIXUPIMMS_MASKZ |
19860 | return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, |
19861 | Src1, Src2, Src3, Imm, Rnd), |
19862 | Mask, Passthru, Subtarget, DAG); |
19863 | } |
19864 | case CONVERT_TO_MASK: { |
19865 | MVT SrcVT = Op.getOperand(1).getSimpleValueType(); |
19866 | MVT MaskVT = MVT::getVectorVT(MVT::i1, SrcVT.getVectorNumElements()); |
19867 | MVT BitcastVT = MVT::getVectorVT(MVT::i1, VT.getSizeInBits()); |
19868 | |
19869 | SDValue CvtMask = DAG.getNode(IntrData->Opc0, dl, MaskVT, |
19870 | Op.getOperand(1)); |
19871 | SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT, |
19872 | DAG.getUNDEF(BitcastVT), CvtMask, |
19873 | DAG.getIntPtrConstant(0, dl)); |
19874 | return DAG.getBitcast(Op.getValueType(), Res); |
19875 | } |
19876 | case BRCST_SUBVEC_TO_VEC: { |
19877 | SDValue Src = Op.getOperand(1); |
19878 | SDValue Passthru = Op.getOperand(2); |
19879 | SDValue Mask = Op.getOperand(3); |
19880 | EVT resVT = Passthru.getValueType(); |
19881 | SDValue subVec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, resVT, |
19882 | DAG.getUNDEF(resVT), Src, |
19883 | DAG.getIntPtrConstant(0, dl)); |
19884 | SDValue immVal; |
19885 | if (Src.getSimpleValueType().is256BitVector() && resVT.is512BitVector()) |
19886 | immVal = DAG.getConstant(0x44, dl, MVT::i8); |
19887 | else |
19888 | immVal = DAG.getConstant(0, dl, MVT::i8); |
19889 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, |
19890 | subVec, subVec, immVal), |
19891 | Mask, Passthru, Subtarget, DAG); |
19892 | } |
19893 | case BRCST32x2_TO_VEC: { |
19894 | SDValue Src = Op.getOperand(1); |
19895 | SDValue PassThru = Op.getOperand(2); |
19896 | SDValue Mask = Op.getOperand(3); |
19897 | |
19898 | assert((VT.getScalarType() == MVT::i32 ||(((VT.getScalarType() == MVT::i32 || VT.getScalarType() == MVT ::f32) && "Unexpected type!") ? static_cast<void> (0) : __assert_fail ("(VT.getScalarType() == MVT::i32 || VT.getScalarType() == MVT::f32) && \"Unexpected type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 19899, __PRETTY_FUNCTION__)) |
19899 | VT.getScalarType() == MVT::f32) && "Unexpected type!")(((VT.getScalarType() == MVT::i32 || VT.getScalarType() == MVT ::f32) && "Unexpected type!") ? static_cast<void> (0) : __assert_fail ("(VT.getScalarType() == MVT::i32 || VT.getScalarType() == MVT::f32) && \"Unexpected type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 19899, __PRETTY_FUNCTION__)); |
19900 | //bitcast Src to packed 64 |
19901 | MVT ScalarVT = VT.getScalarType() == MVT::i32 ? MVT::i64 : MVT::f64; |
19902 | MVT BitcastVT = MVT::getVectorVT(ScalarVT, Src.getValueSizeInBits()/64); |
19903 | Src = DAG.getBitcast(BitcastVT, Src); |
19904 | |
19905 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src), |
19906 | Mask, PassThru, Subtarget, DAG); |
19907 | } |
19908 | default: |
19909 | break; |
19910 | } |
19911 | } |
19912 | |
19913 | switch (IntNo) { |
19914 | default: return SDValue(); // Don't custom lower most intrinsics. |
19915 | |
19916 | case Intrinsic::x86_avx2_permd: |
19917 | case Intrinsic::x86_avx2_permps: |
19918 | // Operands intentionally swapped. Mask is last operand to intrinsic, |
19919 | // but second operand for node/instruction. |
19920 | return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(), |
19921 | Op.getOperand(2), Op.getOperand(1)); |
19922 | |
19923 | // ptest and testp intrinsics. The intrinsic these come from are designed to |
19924 | // return an integer value, not just an instruction so lower it to the ptest |
19925 | // or testp pattern and a setcc for the result. |
19926 | case Intrinsic::x86_sse41_ptestz: |
19927 | case Intrinsic::x86_sse41_ptestc: |
19928 | case Intrinsic::x86_sse41_ptestnzc: |
19929 | case Intrinsic::x86_avx_ptestz_256: |
19930 | case Intrinsic::x86_avx_ptestc_256: |
19931 | case Intrinsic::x86_avx_ptestnzc_256: |
19932 | case Intrinsic::x86_avx_vtestz_ps: |
19933 | case Intrinsic::x86_avx_vtestc_ps: |
19934 | case Intrinsic::x86_avx_vtestnzc_ps: |
19935 | case Intrinsic::x86_avx_vtestz_pd: |
19936 | case Intrinsic::x86_avx_vtestc_pd: |
19937 | case Intrinsic::x86_avx_vtestnzc_pd: |
19938 | case Intrinsic::x86_avx_vtestz_ps_256: |
19939 | case Intrinsic::x86_avx_vtestc_ps_256: |
19940 | case Intrinsic::x86_avx_vtestnzc_ps_256: |
19941 | case Intrinsic::x86_avx_vtestz_pd_256: |
19942 | case Intrinsic::x86_avx_vtestc_pd_256: |
19943 | case Intrinsic::x86_avx_vtestnzc_pd_256: { |
19944 | bool IsTestPacked = false; |
19945 | X86::CondCode X86CC; |
19946 | switch (IntNo) { |
19947 | default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.")::llvm::llvm_unreachable_internal("Bad fallthrough in Intrinsic lowering." , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 19947); |
19948 | case Intrinsic::x86_avx_vtestz_ps: |
19949 | case Intrinsic::x86_avx_vtestz_pd: |
19950 | case Intrinsic::x86_avx_vtestz_ps_256: |
19951 | case Intrinsic::x86_avx_vtestz_pd_256: |
19952 | IsTestPacked = true; |
19953 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
19954 | case Intrinsic::x86_sse41_ptestz: |
19955 | case Intrinsic::x86_avx_ptestz_256: |
19956 | // ZF = 1 |
19957 | X86CC = X86::COND_E; |
19958 | break; |
19959 | case Intrinsic::x86_avx_vtestc_ps: |
19960 | case Intrinsic::x86_avx_vtestc_pd: |
19961 | case Intrinsic::x86_avx_vtestc_ps_256: |
19962 | case Intrinsic::x86_avx_vtestc_pd_256: |
19963 | IsTestPacked = true; |
19964 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
19965 | case Intrinsic::x86_sse41_ptestc: |
19966 | case Intrinsic::x86_avx_ptestc_256: |
19967 | // CF = 1 |
19968 | X86CC = X86::COND_B; |
19969 | break; |
19970 | case Intrinsic::x86_avx_vtestnzc_ps: |
19971 | case Intrinsic::x86_avx_vtestnzc_pd: |
19972 | case Intrinsic::x86_avx_vtestnzc_ps_256: |
19973 | case Intrinsic::x86_avx_vtestnzc_pd_256: |
19974 | IsTestPacked = true; |
19975 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
19976 | case Intrinsic::x86_sse41_ptestnzc: |
19977 | case Intrinsic::x86_avx_ptestnzc_256: |
19978 | // ZF and CF = 0 |
19979 | X86CC = X86::COND_A; |
19980 | break; |
19981 | } |
19982 | |
19983 | SDValue LHS = Op.getOperand(1); |
19984 | SDValue RHS = Op.getOperand(2); |
19985 | unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST; |
19986 | SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS); |
19987 | SDValue SetCC = getSETCC(X86CC, Test, dl, DAG); |
19988 | return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC); |
19989 | } |
19990 | case Intrinsic::x86_avx512_kortestz_w: |
19991 | case Intrinsic::x86_avx512_kortestc_w: { |
19992 | X86::CondCode X86CC = |
19993 | (IntNo == Intrinsic::x86_avx512_kortestz_w) ? X86::COND_E : X86::COND_B; |
19994 | SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1)); |
19995 | SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2)); |
19996 | SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS); |
19997 | SDValue SetCC = getSETCC(X86CC, Test, dl, DAG); |
19998 | return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC); |
19999 | } |
20000 | |
20001 | case Intrinsic::x86_avx512_knot_w: { |
20002 | SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1)); |
20003 | SDValue RHS = DAG.getConstant(1, dl, MVT::v16i1); |
20004 | SDValue Res = DAG.getNode(ISD::XOR, dl, MVT::v16i1, LHS, RHS); |
20005 | return DAG.getBitcast(MVT::i16, Res); |
20006 | } |
20007 | |
20008 | case Intrinsic::x86_avx512_kandn_w: { |
20009 | SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1)); |
20010 | // Invert LHS for the not. |
20011 | LHS = DAG.getNode(ISD::XOR, dl, MVT::v16i1, LHS, |
20012 | DAG.getConstant(1, dl, MVT::v16i1)); |
20013 | SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2)); |
20014 | SDValue Res = DAG.getNode(ISD::AND, dl, MVT::v16i1, LHS, RHS); |
20015 | return DAG.getBitcast(MVT::i16, Res); |
20016 | } |
20017 | |
20018 | case Intrinsic::x86_avx512_kxnor_w: { |
20019 | SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1)); |
20020 | SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2)); |
20021 | SDValue Res = DAG.getNode(ISD::XOR, dl, MVT::v16i1, LHS, RHS); |
20022 | // Invert result for the not. |
20023 | Res = DAG.getNode(ISD::XOR, dl, MVT::v16i1, Res, |
20024 | DAG.getConstant(1, dl, MVT::v16i1)); |
20025 | return DAG.getBitcast(MVT::i16, Res); |
20026 | } |
20027 | |
20028 | case Intrinsic::x86_sse42_pcmpistria128: |
20029 | case Intrinsic::x86_sse42_pcmpestria128: |
20030 | case Intrinsic::x86_sse42_pcmpistric128: |
20031 | case Intrinsic::x86_sse42_pcmpestric128: |
20032 | case Intrinsic::x86_sse42_pcmpistrio128: |
20033 | case Intrinsic::x86_sse42_pcmpestrio128: |
20034 | case Intrinsic::x86_sse42_pcmpistris128: |
20035 | case Intrinsic::x86_sse42_pcmpestris128: |
20036 | case Intrinsic::x86_sse42_pcmpistriz128: |
20037 | case Intrinsic::x86_sse42_pcmpestriz128: { |
20038 | unsigned Opcode; |
20039 | X86::CondCode X86CC; |
20040 | switch (IntNo) { |
20041 | default: llvm_unreachable("Impossible intrinsic")::llvm::llvm_unreachable_internal("Impossible intrinsic", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20041); // Can't reach here. |
20042 | case Intrinsic::x86_sse42_pcmpistria128: |
20043 | Opcode = X86ISD::PCMPISTRI; |
20044 | X86CC = X86::COND_A; |
20045 | break; |
20046 | case Intrinsic::x86_sse42_pcmpestria128: |
20047 | Opcode = X86ISD::PCMPESTRI; |
20048 | X86CC = X86::COND_A; |
20049 | break; |
20050 | case Intrinsic::x86_sse42_pcmpistric128: |
20051 | Opcode = X86ISD::PCMPISTRI; |
20052 | X86CC = X86::COND_B; |
20053 | break; |
20054 | case Intrinsic::x86_sse42_pcmpestric128: |
20055 | Opcode = X86ISD::PCMPESTRI; |
20056 | X86CC = X86::COND_B; |
20057 | break; |
20058 | case Intrinsic::x86_sse42_pcmpistrio128: |
20059 | Opcode = X86ISD::PCMPISTRI; |
20060 | X86CC = X86::COND_O; |
20061 | break; |
20062 | case Intrinsic::x86_sse42_pcmpestrio128: |
20063 | Opcode = X86ISD::PCMPESTRI; |
20064 | X86CC = X86::COND_O; |
20065 | break; |
20066 | case Intrinsic::x86_sse42_pcmpistris128: |
20067 | Opcode = X86ISD::PCMPISTRI; |
20068 | X86CC = X86::COND_S; |
20069 | break; |
20070 | case Intrinsic::x86_sse42_pcmpestris128: |
20071 | Opcode = X86ISD::PCMPESTRI; |
20072 | X86CC = X86::COND_S; |
20073 | break; |
20074 | case Intrinsic::x86_sse42_pcmpistriz128: |
20075 | Opcode = X86ISD::PCMPISTRI; |
20076 | X86CC = X86::COND_E; |
20077 | break; |
20078 | case Intrinsic::x86_sse42_pcmpestriz128: |
20079 | Opcode = X86ISD::PCMPESTRI; |
20080 | X86CC = X86::COND_E; |
20081 | break; |
20082 | } |
20083 | SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end()); |
20084 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32); |
20085 | SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps); |
20086 | SDValue SetCC = getSETCC(X86CC, SDValue(PCMP.getNode(), 1), dl, DAG); |
20087 | return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC); |
20088 | } |
20089 | |
20090 | case Intrinsic::x86_sse42_pcmpistri128: |
20091 | case Intrinsic::x86_sse42_pcmpestri128: { |
20092 | unsigned Opcode; |
20093 | if (IntNo == Intrinsic::x86_sse42_pcmpistri128) |
20094 | Opcode = X86ISD::PCMPISTRI; |
20095 | else |
20096 | Opcode = X86ISD::PCMPESTRI; |
20097 | |
20098 | SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end()); |
20099 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32); |
20100 | return DAG.getNode(Opcode, dl, VTs, NewOps); |
20101 | } |
20102 | |
20103 | case Intrinsic::eh_sjlj_lsda: { |
20104 | MachineFunction &MF = DAG.getMachineFunction(); |
20105 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
20106 | MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout()); |
20107 | auto &Context = MF.getMMI().getContext(); |
20108 | MCSymbol *S = Context.getOrCreateSymbol(Twine("GCC_except_table") + |
20109 | Twine(MF.getFunctionNumber())); |
20110 | return DAG.getNode(X86ISD::Wrapper, dl, VT, DAG.getMCSymbol(S, PtrVT)); |
20111 | } |
20112 | |
20113 | case Intrinsic::x86_seh_lsda: { |
20114 | // Compute the symbol for the LSDA. We know it'll get emitted later. |
20115 | MachineFunction &MF = DAG.getMachineFunction(); |
20116 | SDValue Op1 = Op.getOperand(1); |
20117 | auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal()); |
20118 | MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol( |
20119 | GlobalValue::dropLLVMManglingEscape(Fn->getName())); |
20120 | |
20121 | // Generate a simple absolute symbol reference. This intrinsic is only |
20122 | // supported on 32-bit Windows, which isn't PIC. |
20123 | SDValue Result = DAG.getMCSymbol(LSDASym, VT); |
20124 | return DAG.getNode(X86ISD::Wrapper, dl, VT, Result); |
20125 | } |
20126 | |
20127 | case Intrinsic::x86_seh_recoverfp: { |
20128 | SDValue FnOp = Op.getOperand(1); |
20129 | SDValue IncomingFPOp = Op.getOperand(2); |
20130 | GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp); |
20131 | auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr); |
20132 | if (!Fn) |
20133 | report_fatal_error( |
20134 | "llvm.x86.seh.recoverfp must take a function as the first argument"); |
20135 | return recoverFramePointer(DAG, Fn, IncomingFPOp); |
20136 | } |
20137 | |
20138 | case Intrinsic::localaddress: { |
20139 | // Returns one of the stack, base, or frame pointer registers, depending on |
20140 | // which is used to reference local variables. |
20141 | MachineFunction &MF = DAG.getMachineFunction(); |
20142 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
20143 | unsigned Reg; |
20144 | if (RegInfo->hasBasePointer(MF)) |
20145 | Reg = RegInfo->getBaseRegister(); |
20146 | else // This function handles the SP or FP case. |
20147 | Reg = RegInfo->getPtrSizedFrameRegister(MF); |
20148 | return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT); |
20149 | } |
20150 | } |
20151 | } |
20152 | |
20153 | static SDValue getAVX2GatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG, |
20154 | SDValue Src, SDValue Mask, SDValue Base, |
20155 | SDValue Index, SDValue ScaleOp, SDValue Chain, |
20156 | const X86Subtarget &Subtarget) { |
20157 | SDLoc dl(Op); |
20158 | auto *C = cast<ConstantSDNode>(ScaleOp); |
20159 | SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8); |
20160 | EVT MaskVT = Mask.getValueType(); |
20161 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other); |
20162 | SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32); |
20163 | SDValue Segment = DAG.getRegister(0, MVT::i32); |
20164 | // If source is undef or we know it won't be used, use a zero vector |
20165 | // to break register dependency. |
20166 | // TODO: use undef instead and let ExecutionDepsFix deal with it? |
20167 | if (Src.isUndef() || ISD::isBuildVectorAllOnes(Mask.getNode())) |
20168 | Src = getZeroVector(Op.getSimpleValueType(), Subtarget, DAG, dl); |
20169 | SDValue Ops[] = {Src, Base, Scale, Index, Disp, Segment, Mask, Chain}; |
20170 | SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops); |
20171 | SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) }; |
20172 | return DAG.getMergeValues(RetOps, dl); |
20173 | } |
20174 | |
20175 | static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG, |
20176 | SDValue Src, SDValue Mask, SDValue Base, |
20177 | SDValue Index, SDValue ScaleOp, SDValue Chain, |
20178 | const X86Subtarget &Subtarget) { |
20179 | SDLoc dl(Op); |
20180 | auto *C = cast<ConstantSDNode>(ScaleOp); |
20181 | SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8); |
20182 | MVT MaskVT = MVT::getVectorVT(MVT::i1, |
20183 | Index.getSimpleValueType().getVectorNumElements()); |
20184 | |
20185 | SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); |
20186 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other); |
20187 | SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32); |
20188 | SDValue Segment = DAG.getRegister(0, MVT::i32); |
20189 | // If source is undef or we know it won't be used, use a zero vector |
20190 | // to break register dependency. |
20191 | // TODO: use undef instead and let ExecutionDepsFix deal with it? |
20192 | if (Src.isUndef() || ISD::isBuildVectorAllOnes(VMask.getNode())) |
20193 | Src = getZeroVector(Op.getSimpleValueType(), Subtarget, DAG, dl); |
20194 | SDValue Ops[] = {Src, VMask, Base, Scale, Index, Disp, Segment, Chain}; |
20195 | SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops); |
20196 | SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) }; |
20197 | return DAG.getMergeValues(RetOps, dl); |
20198 | } |
20199 | |
20200 | static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG, |
20201 | SDValue Src, SDValue Mask, SDValue Base, |
20202 | SDValue Index, SDValue ScaleOp, SDValue Chain, |
20203 | const X86Subtarget &Subtarget) { |
20204 | SDLoc dl(Op); |
20205 | auto *C = cast<ConstantSDNode>(ScaleOp); |
20206 | SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8); |
20207 | SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32); |
20208 | SDValue Segment = DAG.getRegister(0, MVT::i32); |
20209 | MVT MaskVT = MVT::getVectorVT(MVT::i1, |
20210 | Index.getSimpleValueType().getVectorNumElements()); |
20211 | |
20212 | SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); |
20213 | SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other); |
20214 | SDValue Ops[] = {Base, Scale, Index, Disp, Segment, VMask, Src, Chain}; |
20215 | SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops); |
20216 | return SDValue(Res, 1); |
20217 | } |
20218 | |
20219 | static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG, |
20220 | SDValue Mask, SDValue Base, SDValue Index, |
20221 | SDValue ScaleOp, SDValue Chain, |
20222 | const X86Subtarget &Subtarget) { |
20223 | SDLoc dl(Op); |
20224 | auto *C = cast<ConstantSDNode>(ScaleOp); |
20225 | SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8); |
20226 | SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32); |
20227 | SDValue Segment = DAG.getRegister(0, MVT::i32); |
20228 | MVT MaskVT = |
20229 | MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements()); |
20230 | SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); |
20231 | SDValue Ops[] = {VMask, Base, Scale, Index, Disp, Segment, Chain}; |
20232 | SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops); |
20233 | return SDValue(Res, 0); |
20234 | } |
20235 | |
20236 | /// Handles the lowering of builtin intrinsic that return the value |
20237 | /// of the extended control register. |
20238 | static void getExtendedControlRegister(SDNode *N, const SDLoc &DL, |
20239 | SelectionDAG &DAG, |
20240 | const X86Subtarget &Subtarget, |
20241 | SmallVectorImpl<SDValue> &Results) { |
20242 | assert(N->getNumOperands() == 3 && "Unexpected number of operands!")((N->getNumOperands() == 3 && "Unexpected number of operands!" ) ? static_cast<void> (0) : __assert_fail ("N->getNumOperands() == 3 && \"Unexpected number of operands!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20242, __PRETTY_FUNCTION__)); |
20243 | SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue); |
20244 | SDValue LO, HI; |
20245 | |
20246 | // The ECX register is used to select the index of the XCR register to |
20247 | // return. |
20248 | SDValue Chain = |
20249 | DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX, N->getOperand(2)); |
20250 | SDNode *N1 = DAG.getMachineNode(X86::XGETBV, DL, Tys, Chain); |
20251 | Chain = SDValue(N1, 0); |
20252 | |
20253 | // Reads the content of XCR and returns it in registers EDX:EAX. |
20254 | if (Subtarget.is64Bit()) { |
20255 | LO = DAG.getCopyFromReg(Chain, DL, X86::RAX, MVT::i64, SDValue(N1, 1)); |
20256 | HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64, |
20257 | LO.getValue(2)); |
20258 | } else { |
20259 | LO = DAG.getCopyFromReg(Chain, DL, X86::EAX, MVT::i32, SDValue(N1, 1)); |
20260 | HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32, |
20261 | LO.getValue(2)); |
20262 | } |
20263 | Chain = HI.getValue(1); |
20264 | |
20265 | if (Subtarget.is64Bit()) { |
20266 | // Merge the two 32-bit values into a 64-bit one.. |
20267 | SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI, |
20268 | DAG.getConstant(32, DL, MVT::i8)); |
20269 | Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp)); |
20270 | Results.push_back(Chain); |
20271 | return; |
20272 | } |
20273 | |
20274 | // Use a buildpair to merge the two 32-bit values into a 64-bit one. |
20275 | SDValue Ops[] = { LO, HI }; |
20276 | SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops); |
20277 | Results.push_back(Pair); |
20278 | Results.push_back(Chain); |
20279 | } |
20280 | |
20281 | /// Handles the lowering of builtin intrinsics that read performance monitor |
20282 | /// counters (x86_rdpmc). |
20283 | static void getReadPerformanceCounter(SDNode *N, const SDLoc &DL, |
20284 | SelectionDAG &DAG, |
20285 | const X86Subtarget &Subtarget, |
20286 | SmallVectorImpl<SDValue> &Results) { |
20287 | assert(N->getNumOperands() == 3 && "Unexpected number of operands!")((N->getNumOperands() == 3 && "Unexpected number of operands!" ) ? static_cast<void> (0) : __assert_fail ("N->getNumOperands() == 3 && \"Unexpected number of operands!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20287, __PRETTY_FUNCTION__)); |
20288 | SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue); |
20289 | SDValue LO, HI; |
20290 | |
20291 | // The ECX register is used to select the index of the performance counter |
20292 | // to read. |
20293 | SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX, |
20294 | N->getOperand(2)); |
20295 | SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain); |
20296 | |
20297 | // Reads the content of a 64-bit performance counter and returns it in the |
20298 | // registers EDX:EAX. |
20299 | if (Subtarget.is64Bit()) { |
20300 | LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1)); |
20301 | HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64, |
20302 | LO.getValue(2)); |
20303 | } else { |
20304 | LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1)); |
20305 | HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32, |
20306 | LO.getValue(2)); |
20307 | } |
20308 | Chain = HI.getValue(1); |
20309 | |
20310 | if (Subtarget.is64Bit()) { |
20311 | // The EAX register is loaded with the low-order 32 bits. The EDX register |
20312 | // is loaded with the supported high-order bits of the counter. |
20313 | SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI, |
20314 | DAG.getConstant(32, DL, MVT::i8)); |
20315 | Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp)); |
20316 | Results.push_back(Chain); |
20317 | return; |
20318 | } |
20319 | |
20320 | // Use a buildpair to merge the two 32-bit values into a 64-bit one. |
20321 | SDValue Ops[] = { LO, HI }; |
20322 | SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops); |
20323 | Results.push_back(Pair); |
20324 | Results.push_back(Chain); |
20325 | } |
20326 | |
20327 | /// Handles the lowering of builtin intrinsics that read the time stamp counter |
20328 | /// (x86_rdtsc and x86_rdtscp). This function is also used to custom lower |
20329 | /// READCYCLECOUNTER nodes. |
20330 | static void getReadTimeStampCounter(SDNode *N, const SDLoc &DL, unsigned Opcode, |
20331 | SelectionDAG &DAG, |
20332 | const X86Subtarget &Subtarget, |
20333 | SmallVectorImpl<SDValue> &Results) { |
20334 | SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue); |
20335 | SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0)); |
20336 | SDValue LO, HI; |
20337 | |
20338 | // The processor's time-stamp counter (a 64-bit MSR) is stored into the |
20339 | // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR |
20340 | // and the EAX register is loaded with the low-order 32 bits. |
20341 | if (Subtarget.is64Bit()) { |
20342 | LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1)); |
20343 | HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64, |
20344 | LO.getValue(2)); |
20345 | } else { |
20346 | LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1)); |
20347 | HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32, |
20348 | LO.getValue(2)); |
20349 | } |
20350 | SDValue Chain = HI.getValue(1); |
20351 | |
20352 | if (Opcode == X86ISD::RDTSCP_DAG) { |
20353 | assert(N->getNumOperands() == 3 && "Unexpected number of operands!")((N->getNumOperands() == 3 && "Unexpected number of operands!" ) ? static_cast<void> (0) : __assert_fail ("N->getNumOperands() == 3 && \"Unexpected number of operands!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20353, __PRETTY_FUNCTION__)); |
20354 | |
20355 | // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into |
20356 | // the ECX register. Add 'ecx' explicitly to the chain. |
20357 | SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32, |
20358 | HI.getValue(2)); |
20359 | // Explicitly store the content of ECX at the location passed in input |
20360 | // to the 'rdtscp' intrinsic. |
20361 | Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2), |
20362 | MachinePointerInfo()); |
20363 | } |
20364 | |
20365 | if (Subtarget.is64Bit()) { |
20366 | // The EDX register is loaded with the high-order 32 bits of the MSR, and |
20367 | // the EAX register is loaded with the low-order 32 bits. |
20368 | SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI, |
20369 | DAG.getConstant(32, DL, MVT::i8)); |
20370 | Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp)); |
20371 | Results.push_back(Chain); |
20372 | return; |
20373 | } |
20374 | |
20375 | // Use a buildpair to merge the two 32-bit values into a 64-bit one. |
20376 | SDValue Ops[] = { LO, HI }; |
20377 | SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops); |
20378 | Results.push_back(Pair); |
20379 | Results.push_back(Chain); |
20380 | } |
20381 | |
20382 | static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget &Subtarget, |
20383 | SelectionDAG &DAG) { |
20384 | SmallVector<SDValue, 2> Results; |
20385 | SDLoc DL(Op); |
20386 | getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget, |
20387 | Results); |
20388 | return DAG.getMergeValues(Results, DL); |
20389 | } |
20390 | |
20391 | static SDValue MarkEHRegistrationNode(SDValue Op, SelectionDAG &DAG) { |
20392 | MachineFunction &MF = DAG.getMachineFunction(); |
20393 | SDValue Chain = Op.getOperand(0); |
20394 | SDValue RegNode = Op.getOperand(2); |
20395 | WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo(); |
20396 | if (!EHInfo) |
20397 | report_fatal_error("EH registrations only live in functions using WinEH"); |
20398 | |
20399 | // Cast the operand to an alloca, and remember the frame index. |
20400 | auto *FINode = dyn_cast<FrameIndexSDNode>(RegNode); |
20401 | if (!FINode) |
20402 | report_fatal_error("llvm.x86.seh.ehregnode expects a static alloca"); |
20403 | EHInfo->EHRegNodeFrameIndex = FINode->getIndex(); |
20404 | |
20405 | // Return the chain operand without making any DAG nodes. |
20406 | return Chain; |
20407 | } |
20408 | |
20409 | static SDValue MarkEHGuard(SDValue Op, SelectionDAG &DAG) { |
20410 | MachineFunction &MF = DAG.getMachineFunction(); |
20411 | SDValue Chain = Op.getOperand(0); |
20412 | SDValue EHGuard = Op.getOperand(2); |
20413 | WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo(); |
20414 | if (!EHInfo) |
20415 | report_fatal_error("EHGuard only live in functions using WinEH"); |
20416 | |
20417 | // Cast the operand to an alloca, and remember the frame index. |
20418 | auto *FINode = dyn_cast<FrameIndexSDNode>(EHGuard); |
20419 | if (!FINode) |
20420 | report_fatal_error("llvm.x86.seh.ehguard expects a static alloca"); |
20421 | EHInfo->EHGuardFrameIndex = FINode->getIndex(); |
20422 | |
20423 | // Return the chain operand without making any DAG nodes. |
20424 | return Chain; |
20425 | } |
20426 | |
20427 | /// Emit Truncating Store with signed or unsigned saturation. |
20428 | static SDValue |
20429 | EmitTruncSStore(bool SignedSat, SDValue Chain, const SDLoc &Dl, SDValue Val, |
20430 | SDValue Ptr, EVT MemVT, MachineMemOperand *MMO, |
20431 | SelectionDAG &DAG) { |
20432 | |
20433 | SDVTList VTs = DAG.getVTList(MVT::Other); |
20434 | SDValue Undef = DAG.getUNDEF(Ptr.getValueType()); |
20435 | SDValue Ops[] = { Chain, Val, Ptr, Undef }; |
20436 | return SignedSat ? |
20437 | DAG.getTargetMemSDNode<TruncSStoreSDNode>(VTs, Ops, Dl, MemVT, MMO) : |
20438 | DAG.getTargetMemSDNode<TruncUSStoreSDNode>(VTs, Ops, Dl, MemVT, MMO); |
20439 | } |
20440 | |
20441 | /// Emit Masked Truncating Store with signed or unsigned saturation. |
20442 | static SDValue |
20443 | EmitMaskedTruncSStore(bool SignedSat, SDValue Chain, const SDLoc &Dl, |
20444 | SDValue Val, SDValue Ptr, SDValue Mask, EVT MemVT, |
20445 | MachineMemOperand *MMO, SelectionDAG &DAG) { |
20446 | |
20447 | SDVTList VTs = DAG.getVTList(MVT::Other); |
20448 | SDValue Ops[] = { Chain, Ptr, Mask, Val }; |
20449 | return SignedSat ? |
20450 | DAG.getTargetMemSDNode<MaskedTruncSStoreSDNode>(VTs, Ops, Dl, MemVT, MMO) : |
20451 | DAG.getTargetMemSDNode<MaskedTruncUSStoreSDNode>(VTs, Ops, Dl, MemVT, MMO); |
20452 | } |
20453 | |
20454 | static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget &Subtarget, |
20455 | SelectionDAG &DAG) { |
20456 | unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); |
20457 | |
20458 | const IntrinsicData *IntrData = getIntrinsicWithChain(IntNo); |
20459 | if (!IntrData) { |
20460 | switch (IntNo) { |
20461 | case llvm::Intrinsic::x86_seh_ehregnode: |
20462 | return MarkEHRegistrationNode(Op, DAG); |
20463 | case llvm::Intrinsic::x86_seh_ehguard: |
20464 | return MarkEHGuard(Op, DAG); |
20465 | case llvm::Intrinsic::x86_flags_read_u32: |
20466 | case llvm::Intrinsic::x86_flags_read_u64: |
20467 | case llvm::Intrinsic::x86_flags_write_u32: |
20468 | case llvm::Intrinsic::x86_flags_write_u64: { |
20469 | // We need a frame pointer because this will get lowered to a PUSH/POP |
20470 | // sequence. |
20471 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
20472 | MFI.setHasCopyImplyingStackAdjustment(true); |
20473 | // Don't do anything here, we will expand these intrinsics out later |
20474 | // during ExpandISelPseudos in EmitInstrWithCustomInserter. |
20475 | return SDValue(); |
20476 | } |
20477 | case Intrinsic::x86_lwpins32: |
20478 | case Intrinsic::x86_lwpins64: { |
20479 | SDLoc dl(Op); |
20480 | SDValue Chain = Op->getOperand(0); |
20481 | SDVTList VTs = DAG.getVTList(MVT::i32, MVT::Other); |
20482 | SDValue LwpIns = |
20483 | DAG.getNode(X86ISD::LWPINS, dl, VTs, Chain, Op->getOperand(2), |
20484 | Op->getOperand(3), Op->getOperand(4)); |
20485 | SDValue SetCC = getSETCC(X86::COND_B, LwpIns.getValue(0), dl, DAG); |
20486 | SDValue Result = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, SetCC); |
20487 | return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, |
20488 | LwpIns.getValue(1)); |
20489 | } |
20490 | } |
20491 | return SDValue(); |
20492 | } |
20493 | |
20494 | SDLoc dl(Op); |
20495 | switch(IntrData->Type) { |
20496 | default: llvm_unreachable("Unknown Intrinsic Type")::llvm::llvm_unreachable_internal("Unknown Intrinsic Type", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20496); |
20497 | case RDSEED: |
20498 | case RDRAND: { |
20499 | // Emit the node with the right value type. |
20500 | SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other); |
20501 | SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0)); |
20502 | |
20503 | // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1. |
20504 | // Otherwise return the value from Rand, which is always 0, casted to i32. |
20505 | SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)), |
20506 | DAG.getConstant(1, dl, Op->getValueType(1)), |
20507 | DAG.getConstant(X86::COND_B, dl, MVT::i32), |
20508 | SDValue(Result.getNode(), 1) }; |
20509 | SDValue isValid = DAG.getNode(X86ISD::CMOV, dl, |
20510 | DAG.getVTList(Op->getValueType(1), MVT::Glue), |
20511 | Ops); |
20512 | |
20513 | // Return { result, isValid, chain }. |
20514 | return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid, |
20515 | SDValue(Result.getNode(), 2)); |
20516 | } |
20517 | case GATHER_AVX2: { |
20518 | SDValue Chain = Op.getOperand(0); |
20519 | SDValue Src = Op.getOperand(2); |
20520 | SDValue Base = Op.getOperand(3); |
20521 | SDValue Index = Op.getOperand(4); |
20522 | SDValue Mask = Op.getOperand(5); |
20523 | SDValue Scale = Op.getOperand(6); |
20524 | return getAVX2GatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, |
20525 | Scale, Chain, Subtarget); |
20526 | } |
20527 | case GATHER: { |
20528 | //gather(v1, mask, index, base, scale); |
20529 | SDValue Chain = Op.getOperand(0); |
20530 | SDValue Src = Op.getOperand(2); |
20531 | SDValue Base = Op.getOperand(3); |
20532 | SDValue Index = Op.getOperand(4); |
20533 | SDValue Mask = Op.getOperand(5); |
20534 | SDValue Scale = Op.getOperand(6); |
20535 | return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, |
20536 | Chain, Subtarget); |
20537 | } |
20538 | case SCATTER: { |
20539 | //scatter(base, mask, index, v1, scale); |
20540 | SDValue Chain = Op.getOperand(0); |
20541 | SDValue Base = Op.getOperand(2); |
20542 | SDValue Mask = Op.getOperand(3); |
20543 | SDValue Index = Op.getOperand(4); |
20544 | SDValue Src = Op.getOperand(5); |
20545 | SDValue Scale = Op.getOperand(6); |
20546 | return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, |
20547 | Scale, Chain, Subtarget); |
20548 | } |
20549 | case PREFETCH: { |
20550 | SDValue Hint = Op.getOperand(6); |
20551 | unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue(); |
20552 | assert((HintVal == 2 || HintVal == 3) &&(((HintVal == 2 || HintVal == 3) && "Wrong prefetch hint in intrinsic: should be 2 or 3" ) ? static_cast<void> (0) : __assert_fail ("(HintVal == 2 || HintVal == 3) && \"Wrong prefetch hint in intrinsic: should be 2 or 3\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20553, __PRETTY_FUNCTION__)) |
20553 | "Wrong prefetch hint in intrinsic: should be 2 or 3")(((HintVal == 2 || HintVal == 3) && "Wrong prefetch hint in intrinsic: should be 2 or 3" ) ? static_cast<void> (0) : __assert_fail ("(HintVal == 2 || HintVal == 3) && \"Wrong prefetch hint in intrinsic: should be 2 or 3\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20553, __PRETTY_FUNCTION__)); |
20554 | unsigned Opcode = (HintVal == 2 ? IntrData->Opc1 : IntrData->Opc0); |
20555 | SDValue Chain = Op.getOperand(0); |
20556 | SDValue Mask = Op.getOperand(2); |
20557 | SDValue Index = Op.getOperand(3); |
20558 | SDValue Base = Op.getOperand(4); |
20559 | SDValue Scale = Op.getOperand(5); |
20560 | return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain, |
20561 | Subtarget); |
20562 | } |
20563 | // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP). |
20564 | case RDTSC: { |
20565 | SmallVector<SDValue, 2> Results; |
20566 | getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget, |
20567 | Results); |
20568 | return DAG.getMergeValues(Results, dl); |
20569 | } |
20570 | // Read Performance Monitoring Counters. |
20571 | case RDPMC: { |
20572 | SmallVector<SDValue, 2> Results; |
20573 | getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results); |
20574 | return DAG.getMergeValues(Results, dl); |
20575 | } |
20576 | // Get Extended Control Register. |
20577 | case XGETBV: { |
20578 | SmallVector<SDValue, 2> Results; |
20579 | getExtendedControlRegister(Op.getNode(), dl, DAG, Subtarget, Results); |
20580 | return DAG.getMergeValues(Results, dl); |
20581 | } |
20582 | // XTEST intrinsics. |
20583 | case XTEST: { |
20584 | SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other); |
20585 | SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0)); |
20586 | |
20587 | SDValue SetCC = getSETCC(X86::COND_NE, InTrans, dl, DAG); |
20588 | SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC); |
20589 | return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), |
20590 | Ret, SDValue(InTrans.getNode(), 1)); |
20591 | } |
20592 | // ADC/ADCX/SBB |
20593 | case ADX: { |
20594 | SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other); |
20595 | SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other); |
20596 | SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2), |
20597 | DAG.getConstant(-1, dl, MVT::i8)); |
20598 | SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3), |
20599 | Op.getOperand(4), GenCF.getValue(1)); |
20600 | SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0), |
20601 | Op.getOperand(5), MachinePointerInfo()); |
20602 | SDValue SetCC = getSETCC(X86::COND_B, Res.getValue(1), dl, DAG); |
20603 | SDValue Results[] = { SetCC, Store }; |
20604 | return DAG.getMergeValues(Results, dl); |
20605 | } |
20606 | case COMPRESS_TO_MEM: { |
20607 | SDValue Mask = Op.getOperand(4); |
20608 | SDValue DataToCompress = Op.getOperand(3); |
20609 | SDValue Addr = Op.getOperand(2); |
20610 | SDValue Chain = Op.getOperand(0); |
20611 | MVT VT = DataToCompress.getSimpleValueType(); |
20612 | |
20613 | MemIntrinsicSDNode *MemIntr = dyn_cast<MemIntrinsicSDNode>(Op); |
20614 | assert(MemIntr && "Expected MemIntrinsicSDNode!")((MemIntr && "Expected MemIntrinsicSDNode!") ? static_cast <void> (0) : __assert_fail ("MemIntr && \"Expected MemIntrinsicSDNode!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20614, __PRETTY_FUNCTION__)); |
20615 | |
20616 | if (isAllOnesConstant(Mask)) // return just a store |
20617 | return DAG.getStore(Chain, dl, DataToCompress, Addr, |
20618 | MemIntr->getMemOperand()); |
20619 | |
20620 | MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements()); |
20621 | SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); |
20622 | |
20623 | return DAG.getMaskedStore(Chain, dl, DataToCompress, Addr, VMask, VT, |
20624 | MemIntr->getMemOperand(), |
20625 | false /* truncating */, true /* compressing */); |
20626 | } |
20627 | case TRUNCATE_TO_MEM_VI8: |
20628 | case TRUNCATE_TO_MEM_VI16: |
20629 | case TRUNCATE_TO_MEM_VI32: { |
20630 | SDValue Mask = Op.getOperand(4); |
20631 | SDValue DataToTruncate = Op.getOperand(3); |
20632 | SDValue Addr = Op.getOperand(2); |
20633 | SDValue Chain = Op.getOperand(0); |
20634 | |
20635 | MemIntrinsicSDNode *MemIntr = dyn_cast<MemIntrinsicSDNode>(Op); |
20636 | assert(MemIntr && "Expected MemIntrinsicSDNode!")((MemIntr && "Expected MemIntrinsicSDNode!") ? static_cast <void> (0) : __assert_fail ("MemIntr && \"Expected MemIntrinsicSDNode!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20636, __PRETTY_FUNCTION__)); |
20637 | |
20638 | EVT MemVT = MemIntr->getMemoryVT(); |
20639 | |
20640 | uint16_t TruncationOp = IntrData->Opc0; |
20641 | switch (TruncationOp) { |
20642 | case X86ISD::VTRUNC: { |
20643 | if (isAllOnesConstant(Mask)) // return just a truncate store |
20644 | return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr, MemVT, |
20645 | MemIntr->getMemOperand()); |
20646 | |
20647 | MVT MaskVT = MVT::getVectorVT(MVT::i1, MemVT.getVectorNumElements()); |
20648 | SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); |
20649 | |
20650 | return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr, VMask, MemVT, |
20651 | MemIntr->getMemOperand(), true /* truncating */); |
20652 | } |
20653 | case X86ISD::VTRUNCUS: |
20654 | case X86ISD::VTRUNCS: { |
20655 | bool IsSigned = (TruncationOp == X86ISD::VTRUNCS); |
20656 | if (isAllOnesConstant(Mask)) |
20657 | return EmitTruncSStore(IsSigned, Chain, dl, DataToTruncate, Addr, MemVT, |
20658 | MemIntr->getMemOperand(), DAG); |
20659 | |
20660 | MVT MaskVT = MVT::getVectorVT(MVT::i1, MemVT.getVectorNumElements()); |
20661 | SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); |
20662 | |
20663 | return EmitMaskedTruncSStore(IsSigned, Chain, dl, DataToTruncate, Addr, |
20664 | VMask, MemVT, MemIntr->getMemOperand(), DAG); |
20665 | } |
20666 | default: |
20667 | llvm_unreachable("Unsupported truncstore intrinsic")::llvm::llvm_unreachable_internal("Unsupported truncstore intrinsic" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20667); |
20668 | } |
20669 | } |
20670 | |
20671 | case EXPAND_FROM_MEM: { |
20672 | SDValue Mask = Op.getOperand(4); |
20673 | SDValue PassThru = Op.getOperand(3); |
20674 | SDValue Addr = Op.getOperand(2); |
20675 | SDValue Chain = Op.getOperand(0); |
20676 | MVT VT = Op.getSimpleValueType(); |
20677 | |
20678 | MemIntrinsicSDNode *MemIntr = dyn_cast<MemIntrinsicSDNode>(Op); |
20679 | assert(MemIntr && "Expected MemIntrinsicSDNode!")((MemIntr && "Expected MemIntrinsicSDNode!") ? static_cast <void> (0) : __assert_fail ("MemIntr && \"Expected MemIntrinsicSDNode!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20679, __PRETTY_FUNCTION__)); |
20680 | |
20681 | if (isAllOnesConstant(Mask)) // Return a regular (unmasked) vector load. |
20682 | return DAG.getLoad(VT, dl, Chain, Addr, MemIntr->getMemOperand()); |
20683 | if (X86::isZeroNode(Mask)) |
20684 | return DAG.getUNDEF(VT); |
20685 | |
20686 | MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements()); |
20687 | SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); |
20688 | return DAG.getMaskedLoad(VT, dl, Chain, Addr, VMask, PassThru, VT, |
20689 | MemIntr->getMemOperand(), ISD::NON_EXTLOAD, |
20690 | true /* expanding */); |
20691 | } |
20692 | } |
20693 | } |
20694 | |
20695 | SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, |
20696 | SelectionDAG &DAG) const { |
20697 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
20698 | MFI.setReturnAddressIsTaken(true); |
20699 | |
20700 | if (verifyReturnAddressArgumentIsConstant(Op, DAG)) |
20701 | return SDValue(); |
20702 | |
20703 | unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); |
20704 | SDLoc dl(Op); |
20705 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
20706 | |
20707 | if (Depth > 0) { |
20708 | SDValue FrameAddr = LowerFRAMEADDR(Op, DAG); |
20709 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
20710 | SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT); |
20711 | return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), |
20712 | DAG.getNode(ISD::ADD, dl, PtrVT, FrameAddr, Offset), |
20713 | MachinePointerInfo()); |
20714 | } |
20715 | |
20716 | // Just load the return address. |
20717 | SDValue RetAddrFI = getReturnAddressFrameIndex(DAG); |
20718 | return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), RetAddrFI, |
20719 | MachinePointerInfo()); |
20720 | } |
20721 | |
20722 | SDValue X86TargetLowering::LowerADDROFRETURNADDR(SDValue Op, |
20723 | SelectionDAG &DAG) const { |
20724 | DAG.getMachineFunction().getFrameInfo().setReturnAddressIsTaken(true); |
20725 | return getReturnAddressFrameIndex(DAG); |
20726 | } |
20727 | |
20728 | SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { |
20729 | MachineFunction &MF = DAG.getMachineFunction(); |
20730 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
20731 | X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); |
20732 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
20733 | EVT VT = Op.getValueType(); |
20734 | |
20735 | MFI.setFrameAddressIsTaken(true); |
20736 | |
20737 | if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) { |
20738 | // Depth > 0 makes no sense on targets which use Windows unwind codes. It |
20739 | // is not possible to crawl up the stack without looking at the unwind codes |
20740 | // simultaneously. |
20741 | int FrameAddrIndex = FuncInfo->getFAIndex(); |
20742 | if (!FrameAddrIndex) { |
20743 | // Set up a frame object for the return address. |
20744 | unsigned SlotSize = RegInfo->getSlotSize(); |
20745 | FrameAddrIndex = MF.getFrameInfo().CreateFixedObject( |
20746 | SlotSize, /*Offset=*/0, /*IsImmutable=*/false); |
20747 | FuncInfo->setFAIndex(FrameAddrIndex); |
20748 | } |
20749 | return DAG.getFrameIndex(FrameAddrIndex, VT); |
20750 | } |
20751 | |
20752 | unsigned FrameReg = |
20753 | RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction()); |
20754 | SDLoc dl(Op); // FIXME probably not meaningful |
20755 | unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); |
20756 | assert(((FrameReg == X86::RBP && VT == MVT::i64) ||((((FrameReg == X86::RBP && VT == MVT::i64) || (FrameReg == X86::EBP && VT == MVT::i32)) && "Invalid Frame Register!" ) ? static_cast<void> (0) : __assert_fail ("((FrameReg == X86::RBP && VT == MVT::i64) || (FrameReg == X86::EBP && VT == MVT::i32)) && \"Invalid Frame Register!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20758, __PRETTY_FUNCTION__)) |
20757 | (FrameReg == X86::EBP && VT == MVT::i32)) &&((((FrameReg == X86::RBP && VT == MVT::i64) || (FrameReg == X86::EBP && VT == MVT::i32)) && "Invalid Frame Register!" ) ? static_cast<void> (0) : __assert_fail ("((FrameReg == X86::RBP && VT == MVT::i64) || (FrameReg == X86::EBP && VT == MVT::i32)) && \"Invalid Frame Register!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20758, __PRETTY_FUNCTION__)) |
20758 | "Invalid Frame Register!")((((FrameReg == X86::RBP && VT == MVT::i64) || (FrameReg == X86::EBP && VT == MVT::i32)) && "Invalid Frame Register!" ) ? static_cast<void> (0) : __assert_fail ("((FrameReg == X86::RBP && VT == MVT::i64) || (FrameReg == X86::EBP && VT == MVT::i32)) && \"Invalid Frame Register!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20758, __PRETTY_FUNCTION__)); |
20759 | SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT); |
20760 | while (Depth--) |
20761 | FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, |
20762 | MachinePointerInfo()); |
20763 | return FrameAddr; |
20764 | } |
20765 | |
20766 | // FIXME? Maybe this could be a TableGen attribute on some registers and |
20767 | // this table could be generated automatically from RegInfo. |
20768 | unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT, |
20769 | SelectionDAG &DAG) const { |
20770 | const TargetFrameLowering &TFI = *Subtarget.getFrameLowering(); |
20771 | const MachineFunction &MF = DAG.getMachineFunction(); |
20772 | |
20773 | unsigned Reg = StringSwitch<unsigned>(RegName) |
20774 | .Case("esp", X86::ESP) |
20775 | .Case("rsp", X86::RSP) |
20776 | .Case("ebp", X86::EBP) |
20777 | .Case("rbp", X86::RBP) |
20778 | .Default(0); |
20779 | |
20780 | if (Reg == X86::EBP || Reg == X86::RBP) { |
20781 | if (!TFI.hasFP(MF)) |
20782 | report_fatal_error("register " + StringRef(RegName) + |
20783 | " is allocatable: function has no frame pointer"); |
20784 | #ifndef NDEBUG |
20785 | else { |
20786 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
20787 | unsigned FrameReg = |
20788 | RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction()); |
20789 | assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&(((FrameReg == X86::EBP || FrameReg == X86::RBP) && "Invalid Frame Register!" ) ? static_cast<void> (0) : __assert_fail ("(FrameReg == X86::EBP || FrameReg == X86::RBP) && \"Invalid Frame Register!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20790, __PRETTY_FUNCTION__)) |
20790 | "Invalid Frame Register!")(((FrameReg == X86::EBP || FrameReg == X86::RBP) && "Invalid Frame Register!" ) ? static_cast<void> (0) : __assert_fail ("(FrameReg == X86::EBP || FrameReg == X86::RBP) && \"Invalid Frame Register!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20790, __PRETTY_FUNCTION__)); |
20791 | } |
20792 | #endif |
20793 | } |
20794 | |
20795 | if (Reg) |
20796 | return Reg; |
20797 | |
20798 | report_fatal_error("Invalid register name global variable"); |
20799 | } |
20800 | |
20801 | SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op, |
20802 | SelectionDAG &DAG) const { |
20803 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
20804 | return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op)); |
20805 | } |
20806 | |
20807 | unsigned X86TargetLowering::getExceptionPointerRegister( |
20808 | const Constant *PersonalityFn) const { |
20809 | if (classifyEHPersonality(PersonalityFn) == EHPersonality::CoreCLR) |
20810 | return Subtarget.isTarget64BitLP64() ? X86::RDX : X86::EDX; |
20811 | |
20812 | return Subtarget.isTarget64BitLP64() ? X86::RAX : X86::EAX; |
20813 | } |
20814 | |
20815 | unsigned X86TargetLowering::getExceptionSelectorRegister( |
20816 | const Constant *PersonalityFn) const { |
20817 | // Funclet personalities don't use selectors (the runtime does the selection). |
20818 | assert(!isFuncletEHPersonality(classifyEHPersonality(PersonalityFn)))((!isFuncletEHPersonality(classifyEHPersonality(PersonalityFn ))) ? static_cast<void> (0) : __assert_fail ("!isFuncletEHPersonality(classifyEHPersonality(PersonalityFn))" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20818, __PRETTY_FUNCTION__)); |
20819 | return Subtarget.isTarget64BitLP64() ? X86::RDX : X86::EDX; |
20820 | } |
20821 | |
20822 | bool X86TargetLowering::needsFixedCatchObjects() const { |
20823 | return Subtarget.isTargetWin64(); |
20824 | } |
20825 | |
20826 | SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const { |
20827 | SDValue Chain = Op.getOperand(0); |
20828 | SDValue Offset = Op.getOperand(1); |
20829 | SDValue Handler = Op.getOperand(2); |
20830 | SDLoc dl (Op); |
20831 | |
20832 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
20833 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
20834 | unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction()); |
20835 | assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||((((FrameReg == X86::RBP && PtrVT == MVT::i64) || (FrameReg == X86::EBP && PtrVT == MVT::i32)) && "Invalid Frame Register!" ) ? static_cast<void> (0) : __assert_fail ("((FrameReg == X86::RBP && PtrVT == MVT::i64) || (FrameReg == X86::EBP && PtrVT == MVT::i32)) && \"Invalid Frame Register!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20837, __PRETTY_FUNCTION__)) |
20836 | (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&((((FrameReg == X86::RBP && PtrVT == MVT::i64) || (FrameReg == X86::EBP && PtrVT == MVT::i32)) && "Invalid Frame Register!" ) ? static_cast<void> (0) : __assert_fail ("((FrameReg == X86::RBP && PtrVT == MVT::i64) || (FrameReg == X86::EBP && PtrVT == MVT::i32)) && \"Invalid Frame Register!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20837, __PRETTY_FUNCTION__)) |
20837 | "Invalid Frame Register!")((((FrameReg == X86::RBP && PtrVT == MVT::i64) || (FrameReg == X86::EBP && PtrVT == MVT::i32)) && "Invalid Frame Register!" ) ? static_cast<void> (0) : __assert_fail ("((FrameReg == X86::RBP && PtrVT == MVT::i64) || (FrameReg == X86::EBP && PtrVT == MVT::i32)) && \"Invalid Frame Register!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20837, __PRETTY_FUNCTION__)); |
20838 | SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT); |
20839 | unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX; |
20840 | |
20841 | SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame, |
20842 | DAG.getIntPtrConstant(RegInfo->getSlotSize(), |
20843 | dl)); |
20844 | StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset); |
20845 | Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo()); |
20846 | Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr); |
20847 | |
20848 | return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain, |
20849 | DAG.getRegister(StoreAddrReg, PtrVT)); |
20850 | } |
20851 | |
20852 | SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op, |
20853 | SelectionDAG &DAG) const { |
20854 | SDLoc DL(Op); |
20855 | // If the subtarget is not 64bit, we may need the global base reg |
20856 | // after isel expand pseudo, i.e., after CGBR pass ran. |
20857 | // Therefore, ask for the GlobalBaseReg now, so that the pass |
20858 | // inserts the code for us in case we need it. |
20859 | // Otherwise, we will end up in a situation where we will |
20860 | // reference a virtual register that is not defined! |
20861 | if (!Subtarget.is64Bit()) { |
20862 | const X86InstrInfo *TII = Subtarget.getInstrInfo(); |
20863 | (void)TII->getGlobalBaseReg(&DAG.getMachineFunction()); |
20864 | } |
20865 | return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL, |
20866 | DAG.getVTList(MVT::i32, MVT::Other), |
20867 | Op.getOperand(0), Op.getOperand(1)); |
20868 | } |
20869 | |
20870 | SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op, |
20871 | SelectionDAG &DAG) const { |
20872 | SDLoc DL(Op); |
20873 | return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other, |
20874 | Op.getOperand(0), Op.getOperand(1)); |
20875 | } |
20876 | |
20877 | SDValue X86TargetLowering::lowerEH_SJLJ_SETUP_DISPATCH(SDValue Op, |
20878 | SelectionDAG &DAG) const { |
20879 | SDLoc DL(Op); |
20880 | return DAG.getNode(X86ISD::EH_SJLJ_SETUP_DISPATCH, DL, MVT::Other, |
20881 | Op.getOperand(0)); |
20882 | } |
20883 | |
20884 | static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) { |
20885 | return Op.getOperand(0); |
20886 | } |
20887 | |
20888 | SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op, |
20889 | SelectionDAG &DAG) const { |
20890 | SDValue Root = Op.getOperand(0); |
20891 | SDValue Trmp = Op.getOperand(1); // trampoline |
20892 | SDValue FPtr = Op.getOperand(2); // nested function |
20893 | SDValue Nest = Op.getOperand(3); // 'nest' parameter value |
20894 | SDLoc dl (Op); |
20895 | |
20896 | const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue(); |
20897 | const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); |
20898 | |
20899 | if (Subtarget.is64Bit()) { |
20900 | SDValue OutChains[6]; |
20901 | |
20902 | // Large code-model. |
20903 | const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode. |
20904 | const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode. |
20905 | |
20906 | const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7; |
20907 | const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7; |
20908 | |
20909 | const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix |
20910 | |
20911 | // Load the pointer to the nested function into R11. |
20912 | unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11 |
20913 | SDValue Addr = Trmp; |
20914 | OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16), |
20915 | Addr, MachinePointerInfo(TrmpAddr)); |
20916 | |
20917 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp, |
20918 | DAG.getConstant(2, dl, MVT::i64)); |
20919 | OutChains[1] = |
20920 | DAG.getStore(Root, dl, FPtr, Addr, MachinePointerInfo(TrmpAddr, 2), |
20921 | /* Alignment = */ 2); |
20922 | |
20923 | // Load the 'nest' parameter value into R10. |
20924 | // R10 is specified in X86CallingConv.td |
20925 | OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10 |
20926 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp, |
20927 | DAG.getConstant(10, dl, MVT::i64)); |
20928 | OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16), |
20929 | Addr, MachinePointerInfo(TrmpAddr, 10)); |
20930 | |
20931 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp, |
20932 | DAG.getConstant(12, dl, MVT::i64)); |
20933 | OutChains[3] = |
20934 | DAG.getStore(Root, dl, Nest, Addr, MachinePointerInfo(TrmpAddr, 12), |
20935 | /* Alignment = */ 2); |
20936 | |
20937 | // Jump to the nested function. |
20938 | OpCode = (JMP64r << 8) | REX_WB; // jmpq *... |
20939 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp, |
20940 | DAG.getConstant(20, dl, MVT::i64)); |
20941 | OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16), |
20942 | Addr, MachinePointerInfo(TrmpAddr, 20)); |
20943 | |
20944 | unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11 |
20945 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp, |
20946 | DAG.getConstant(22, dl, MVT::i64)); |
20947 | OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8), |
20948 | Addr, MachinePointerInfo(TrmpAddr, 22)); |
20949 | |
20950 | return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains); |
20951 | } else { |
20952 | const Function *Func = |
20953 | cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue()); |
20954 | CallingConv::ID CC = Func->getCallingConv(); |
20955 | unsigned NestReg; |
20956 | |
20957 | switch (CC) { |
20958 | default: |
20959 | llvm_unreachable("Unsupported calling convention")::llvm::llvm_unreachable_internal("Unsupported calling convention" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 20959); |
20960 | case CallingConv::C: |
20961 | case CallingConv::X86_StdCall: { |
20962 | // Pass 'nest' parameter in ECX. |
20963 | // Must be kept in sync with X86CallingConv.td |
20964 | NestReg = X86::ECX; |
20965 | |
20966 | // Check that ECX wasn't needed by an 'inreg' parameter. |
20967 | FunctionType *FTy = Func->getFunctionType(); |
20968 | const AttributeList &Attrs = Func->getAttributes(); |
20969 | |
20970 | if (!Attrs.isEmpty() && !Func->isVarArg()) { |
20971 | unsigned InRegCount = 0; |
20972 | unsigned Idx = 1; |
20973 | |
20974 | for (FunctionType::param_iterator I = FTy->param_begin(), |
20975 | E = FTy->param_end(); I != E; ++I, ++Idx) |
20976 | if (Attrs.hasAttribute(Idx, Attribute::InReg)) { |
20977 | auto &DL = DAG.getDataLayout(); |
20978 | // FIXME: should only count parameters that are lowered to integers. |
20979 | InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32; |
20980 | } |
20981 | |
20982 | if (InRegCount > 2) { |
20983 | report_fatal_error("Nest register in use - reduce number of inreg" |
20984 | " parameters!"); |
20985 | } |
20986 | } |
20987 | break; |
20988 | } |
20989 | case CallingConv::X86_FastCall: |
20990 | case CallingConv::X86_ThisCall: |
20991 | case CallingConv::Fast: |
20992 | // Pass 'nest' parameter in EAX. |
20993 | // Must be kept in sync with X86CallingConv.td |
20994 | NestReg = X86::EAX; |
20995 | break; |
20996 | } |
20997 | |
20998 | SDValue OutChains[4]; |
20999 | SDValue Addr, Disp; |
21000 | |
21001 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, |
21002 | DAG.getConstant(10, dl, MVT::i32)); |
21003 | Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr); |
21004 | |
21005 | // This is storing the opcode for MOV32ri. |
21006 | const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte. |
21007 | const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7; |
21008 | OutChains[0] = |
21009 | DAG.getStore(Root, dl, DAG.getConstant(MOV32ri | N86Reg, dl, MVT::i8), |
21010 | Trmp, MachinePointerInfo(TrmpAddr)); |
21011 | |
21012 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, |
21013 | DAG.getConstant(1, dl, MVT::i32)); |
21014 | OutChains[1] = |
21015 | DAG.getStore(Root, dl, Nest, Addr, MachinePointerInfo(TrmpAddr, 1), |
21016 | /* Alignment = */ 1); |
21017 | |
21018 | const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode. |
21019 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, |
21020 | DAG.getConstant(5, dl, MVT::i32)); |
21021 | OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8), |
21022 | Addr, MachinePointerInfo(TrmpAddr, 5), |
21023 | /* Alignment = */ 1); |
21024 | |
21025 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, |
21026 | DAG.getConstant(6, dl, MVT::i32)); |
21027 | OutChains[3] = |
21028 | DAG.getStore(Root, dl, Disp, Addr, MachinePointerInfo(TrmpAddr, 6), |
21029 | /* Alignment = */ 1); |
21030 | |
21031 | return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains); |
21032 | } |
21033 | } |
21034 | |
21035 | SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, |
21036 | SelectionDAG &DAG) const { |
21037 | /* |
21038 | The rounding mode is in bits 11:10 of FPSR, and has the following |
21039 | settings: |
21040 | 00 Round to nearest |
21041 | 01 Round to -inf |
21042 | 10 Round to +inf |
21043 | 11 Round to 0 |
21044 | |
21045 | FLT_ROUNDS, on the other hand, expects the following: |
21046 | -1 Undefined |
21047 | 0 Round to 0 |
21048 | 1 Round to nearest |
21049 | 2 Round to +inf |
21050 | 3 Round to -inf |
21051 | |
21052 | To perform the conversion, we do: |
21053 | (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3) |
21054 | */ |
21055 | |
21056 | MachineFunction &MF = DAG.getMachineFunction(); |
21057 | const TargetFrameLowering &TFI = *Subtarget.getFrameLowering(); |
21058 | unsigned StackAlignment = TFI.getStackAlignment(); |
21059 | MVT VT = Op.getSimpleValueType(); |
21060 | SDLoc DL(Op); |
21061 | |
21062 | // Save FP Control Word to stack slot |
21063 | int SSFI = MF.getFrameInfo().CreateStackObject(2, StackAlignment, false); |
21064 | SDValue StackSlot = |
21065 | DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout())); |
21066 | |
21067 | MachineMemOperand *MMO = |
21068 | MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI), |
21069 | MachineMemOperand::MOStore, 2, 2); |
21070 | |
21071 | SDValue Ops[] = { DAG.getEntryNode(), StackSlot }; |
21072 | SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL, |
21073 | DAG.getVTList(MVT::Other), |
21074 | Ops, MVT::i16, MMO); |
21075 | |
21076 | // Load FP Control Word from stack slot |
21077 | SDValue CWD = |
21078 | DAG.getLoad(MVT::i16, DL, Chain, StackSlot, MachinePointerInfo()); |
21079 | |
21080 | // Transform as necessary |
21081 | SDValue CWD1 = |
21082 | DAG.getNode(ISD::SRL, DL, MVT::i16, |
21083 | DAG.getNode(ISD::AND, DL, MVT::i16, |
21084 | CWD, DAG.getConstant(0x800, DL, MVT::i16)), |
21085 | DAG.getConstant(11, DL, MVT::i8)); |
21086 | SDValue CWD2 = |
21087 | DAG.getNode(ISD::SRL, DL, MVT::i16, |
21088 | DAG.getNode(ISD::AND, DL, MVT::i16, |
21089 | CWD, DAG.getConstant(0x400, DL, MVT::i16)), |
21090 | DAG.getConstant(9, DL, MVT::i8)); |
21091 | |
21092 | SDValue RetVal = |
21093 | DAG.getNode(ISD::AND, DL, MVT::i16, |
21094 | DAG.getNode(ISD::ADD, DL, MVT::i16, |
21095 | DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2), |
21096 | DAG.getConstant(1, DL, MVT::i16)), |
21097 | DAG.getConstant(3, DL, MVT::i16)); |
21098 | |
21099 | return DAG.getNode((VT.getSizeInBits() < 16 ? |
21100 | ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal); |
21101 | } |
21102 | |
21103 | // Split an unary integer op into 2 half sized ops. |
21104 | static SDValue LowerVectorIntUnary(SDValue Op, SelectionDAG &DAG) { |
21105 | MVT VT = Op.getSimpleValueType(); |
21106 | unsigned NumElems = VT.getVectorNumElements(); |
21107 | unsigned SizeInBits = VT.getSizeInBits(); |
21108 | |
21109 | // Extract the Lo/Hi vectors |
21110 | SDLoc dl(Op); |
21111 | SDValue Src = Op.getOperand(0); |
21112 | SDValue Lo = extractSubVector(Src, 0, DAG, dl, SizeInBits / 2); |
21113 | SDValue Hi = extractSubVector(Src, NumElems / 2, DAG, dl, SizeInBits / 2); |
21114 | |
21115 | MVT EltVT = VT.getVectorElementType(); |
21116 | MVT NewVT = MVT::getVectorVT(EltVT, NumElems / 2); |
21117 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, |
21118 | DAG.getNode(Op.getOpcode(), dl, NewVT, Lo), |
21119 | DAG.getNode(Op.getOpcode(), dl, NewVT, Hi)); |
21120 | } |
21121 | |
21122 | // Decompose 256-bit ops into smaller 128-bit ops. |
21123 | static SDValue Lower256IntUnary(SDValue Op, SelectionDAG &DAG) { |
21124 | assert(Op.getSimpleValueType().is256BitVector() &&((Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType ().isInteger() && "Only handle AVX 256-bit vector integer operation" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType().isInteger() && \"Only handle AVX 256-bit vector integer operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21126, __PRETTY_FUNCTION__)) |
21125 | Op.getSimpleValueType().isInteger() &&((Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType ().isInteger() && "Only handle AVX 256-bit vector integer operation" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType().isInteger() && \"Only handle AVX 256-bit vector integer operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21126, __PRETTY_FUNCTION__)) |
21126 | "Only handle AVX 256-bit vector integer operation")((Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType ().isInteger() && "Only handle AVX 256-bit vector integer operation" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType().isInteger() && \"Only handle AVX 256-bit vector integer operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21126, __PRETTY_FUNCTION__)); |
21127 | return LowerVectorIntUnary(Op, DAG); |
21128 | } |
21129 | |
21130 | // Decompose 512-bit ops into smaller 256-bit ops. |
21131 | static SDValue Lower512IntUnary(SDValue Op, SelectionDAG &DAG) { |
21132 | assert(Op.getSimpleValueType().is512BitVector() &&((Op.getSimpleValueType().is512BitVector() && Op.getSimpleValueType ().isInteger() && "Only handle AVX 512-bit vector integer operation" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().is512BitVector() && Op.getSimpleValueType().isInteger() && \"Only handle AVX 512-bit vector integer operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21134, __PRETTY_FUNCTION__)) |
21133 | Op.getSimpleValueType().isInteger() &&((Op.getSimpleValueType().is512BitVector() && Op.getSimpleValueType ().isInteger() && "Only handle AVX 512-bit vector integer operation" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().is512BitVector() && Op.getSimpleValueType().isInteger() && \"Only handle AVX 512-bit vector integer operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21134, __PRETTY_FUNCTION__)) |
21134 | "Only handle AVX 512-bit vector integer operation")((Op.getSimpleValueType().is512BitVector() && Op.getSimpleValueType ().isInteger() && "Only handle AVX 512-bit vector integer operation" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().is512BitVector() && Op.getSimpleValueType().isInteger() && \"Only handle AVX 512-bit vector integer operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21134, __PRETTY_FUNCTION__)); |
21135 | return LowerVectorIntUnary(Op, DAG); |
21136 | } |
21137 | |
21138 | /// \brief Lower a vector CTLZ using native supported vector CTLZ instruction. |
21139 | // |
21140 | // i8/i16 vector implemented using dword LZCNT vector instruction |
21141 | // ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal, |
21142 | // split the vector, perform operation on it's Lo a Hi part and |
21143 | // concatenate the results. |
21144 | static SDValue LowerVectorCTLZ_AVX512CDI(SDValue Op, SelectionDAG &DAG) { |
21145 | assert(Op.getOpcode() == ISD::CTLZ)((Op.getOpcode() == ISD::CTLZ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::CTLZ", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21145, __PRETTY_FUNCTION__)); |
21146 | SDLoc dl(Op); |
21147 | MVT VT = Op.getSimpleValueType(); |
21148 | MVT EltVT = VT.getVectorElementType(); |
21149 | unsigned NumElems = VT.getVectorNumElements(); |
21150 | |
21151 | assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&(((EltVT == MVT::i8 || EltVT == MVT::i16) && "Unsupported element type" ) ? static_cast<void> (0) : __assert_fail ("(EltVT == MVT::i8 || EltVT == MVT::i16) && \"Unsupported element type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21152, __PRETTY_FUNCTION__)) |
21152 | "Unsupported element type")(((EltVT == MVT::i8 || EltVT == MVT::i16) && "Unsupported element type" ) ? static_cast<void> (0) : __assert_fail ("(EltVT == MVT::i8 || EltVT == MVT::i16) && \"Unsupported element type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21152, __PRETTY_FUNCTION__)); |
21153 | |
21154 | // Split vector, it's Lo and Hi parts will be handled in next iteration. |
21155 | if (16 < NumElems) |
21156 | return LowerVectorIntUnary(Op, DAG); |
21157 | |
21158 | MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems); |
21159 | assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&(((NewVT.is256BitVector() || NewVT.is512BitVector()) && "Unsupported value type for operation") ? static_cast<void > (0) : __assert_fail ("(NewVT.is256BitVector() || NewVT.is512BitVector()) && \"Unsupported value type for operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21160, __PRETTY_FUNCTION__)) |
21160 | "Unsupported value type for operation")(((NewVT.is256BitVector() || NewVT.is512BitVector()) && "Unsupported value type for operation") ? static_cast<void > (0) : __assert_fail ("(NewVT.is256BitVector() || NewVT.is512BitVector()) && \"Unsupported value type for operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21160, __PRETTY_FUNCTION__)); |
21161 | |
21162 | // Use native supported vector instruction vplzcntd. |
21163 | Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0)); |
21164 | SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op); |
21165 | SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode); |
21166 | SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT); |
21167 | |
21168 | return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta); |
21169 | } |
21170 | |
21171 | // Lower CTLZ using a PSHUFB lookup table implementation. |
21172 | static SDValue LowerVectorCTLZInRegLUT(SDValue Op, const SDLoc &DL, |
21173 | const X86Subtarget &Subtarget, |
21174 | SelectionDAG &DAG) { |
21175 | MVT VT = Op.getSimpleValueType(); |
21176 | int NumElts = VT.getVectorNumElements(); |
21177 | int NumBytes = NumElts * (VT.getScalarSizeInBits() / 8); |
21178 | MVT CurrVT = MVT::getVectorVT(MVT::i8, NumBytes); |
21179 | |
21180 | // Per-nibble leading zero PSHUFB lookup table. |
21181 | const int LUT[16] = {/* 0 */ 4, /* 1 */ 3, /* 2 */ 2, /* 3 */ 2, |
21182 | /* 4 */ 1, /* 5 */ 1, /* 6 */ 1, /* 7 */ 1, |
21183 | /* 8 */ 0, /* 9 */ 0, /* a */ 0, /* b */ 0, |
21184 | /* c */ 0, /* d */ 0, /* e */ 0, /* f */ 0}; |
21185 | |
21186 | SmallVector<SDValue, 64> LUTVec; |
21187 | for (int i = 0; i < NumBytes; ++i) |
21188 | LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8)); |
21189 | SDValue InRegLUT = DAG.getBuildVector(CurrVT, DL, LUTVec); |
21190 | |
21191 | // Begin by bitcasting the input to byte vector, then split those bytes |
21192 | // into lo/hi nibbles and use the PSHUFB LUT to perform CLTZ on each of them. |
21193 | // If the hi input nibble is zero then we add both results together, otherwise |
21194 | // we just take the hi result (by masking the lo result to zero before the |
21195 | // add). |
21196 | SDValue Op0 = DAG.getBitcast(CurrVT, Op.getOperand(0)); |
21197 | SDValue Zero = getZeroVector(CurrVT, Subtarget, DAG, DL); |
21198 | |
21199 | SDValue NibbleMask = DAG.getConstant(0xF, DL, CurrVT); |
21200 | SDValue NibbleShift = DAG.getConstant(0x4, DL, CurrVT); |
21201 | SDValue Lo = DAG.getNode(ISD::AND, DL, CurrVT, Op0, NibbleMask); |
21202 | SDValue Hi = DAG.getNode(ISD::SRL, DL, CurrVT, Op0, NibbleShift); |
21203 | SDValue HiZ = DAG.getSetCC(DL, CurrVT, Hi, Zero, ISD::SETEQ); |
21204 | |
21205 | Lo = DAG.getNode(X86ISD::PSHUFB, DL, CurrVT, InRegLUT, Lo); |
21206 | Hi = DAG.getNode(X86ISD::PSHUFB, DL, CurrVT, InRegLUT, Hi); |
21207 | Lo = DAG.getNode(ISD::AND, DL, CurrVT, Lo, HiZ); |
21208 | SDValue Res = DAG.getNode(ISD::ADD, DL, CurrVT, Lo, Hi); |
21209 | |
21210 | // Merge result back from vXi8 back to VT, working on the lo/hi halves |
21211 | // of the current vector width in the same way we did for the nibbles. |
21212 | // If the upper half of the input element is zero then add the halves' |
21213 | // leading zero counts together, otherwise just use the upper half's. |
21214 | // Double the width of the result until we are at target width. |
21215 | while (CurrVT != VT) { |
21216 | int CurrScalarSizeInBits = CurrVT.getScalarSizeInBits(); |
21217 | int CurrNumElts = CurrVT.getVectorNumElements(); |
21218 | MVT NextSVT = MVT::getIntegerVT(CurrScalarSizeInBits * 2); |
21219 | MVT NextVT = MVT::getVectorVT(NextSVT, CurrNumElts / 2); |
21220 | SDValue Shift = DAG.getConstant(CurrScalarSizeInBits, DL, NextVT); |
21221 | |
21222 | // Check if the upper half of the input element is zero. |
21223 | SDValue HiZ = DAG.getSetCC(DL, CurrVT, DAG.getBitcast(CurrVT, Op0), |
21224 | DAG.getBitcast(CurrVT, Zero), ISD::SETEQ); |
21225 | HiZ = DAG.getBitcast(NextVT, HiZ); |
21226 | |
21227 | // Move the upper/lower halves to the lower bits as we'll be extending to |
21228 | // NextVT. Mask the lower result to zero if HiZ is true and add the results |
21229 | // together. |
21230 | SDValue ResNext = Res = DAG.getBitcast(NextVT, Res); |
21231 | SDValue R0 = DAG.getNode(ISD::SRL, DL, NextVT, ResNext, Shift); |
21232 | SDValue R1 = DAG.getNode(ISD::SRL, DL, NextVT, HiZ, Shift); |
21233 | R1 = DAG.getNode(ISD::AND, DL, NextVT, ResNext, R1); |
21234 | Res = DAG.getNode(ISD::ADD, DL, NextVT, R0, R1); |
21235 | CurrVT = NextVT; |
21236 | } |
21237 | |
21238 | return Res; |
21239 | } |
21240 | |
21241 | static SDValue LowerVectorCTLZ(SDValue Op, const SDLoc &DL, |
21242 | const X86Subtarget &Subtarget, |
21243 | SelectionDAG &DAG) { |
21244 | MVT VT = Op.getSimpleValueType(); |
21245 | |
21246 | if (Subtarget.hasCDI()) |
21247 | return LowerVectorCTLZ_AVX512CDI(Op, DAG); |
21248 | |
21249 | // Decompose 256-bit ops into smaller 128-bit ops. |
21250 | if (VT.is256BitVector() && !Subtarget.hasInt256()) |
21251 | return Lower256IntUnary(Op, DAG); |
21252 | |
21253 | // Decompose 512-bit ops into smaller 256-bit ops. |
21254 | if (VT.is512BitVector() && !Subtarget.hasBWI()) |
21255 | return Lower512IntUnary(Op, DAG); |
21256 | |
21257 | assert(Subtarget.hasSSSE3() && "Expected SSSE3 support for PSHUFB")((Subtarget.hasSSSE3() && "Expected SSSE3 support for PSHUFB" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSSE3() && \"Expected SSSE3 support for PSHUFB\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21257, __PRETTY_FUNCTION__)); |
21258 | return LowerVectorCTLZInRegLUT(Op, DL, Subtarget, DAG); |
21259 | } |
21260 | |
21261 | static SDValue LowerCTLZ(SDValue Op, const X86Subtarget &Subtarget, |
21262 | SelectionDAG &DAG) { |
21263 | MVT VT = Op.getSimpleValueType(); |
21264 | MVT OpVT = VT; |
21265 | unsigned NumBits = VT.getSizeInBits(); |
21266 | SDLoc dl(Op); |
21267 | unsigned Opc = Op.getOpcode(); |
21268 | |
21269 | if (VT.isVector()) |
21270 | return LowerVectorCTLZ(Op, dl, Subtarget, DAG); |
21271 | |
21272 | Op = Op.getOperand(0); |
21273 | if (VT == MVT::i8) { |
21274 | // Zero extend to i32 since there is not an i8 bsr. |
21275 | OpVT = MVT::i32; |
21276 | Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op); |
21277 | } |
21278 | |
21279 | // Issue a bsr (scan bits in reverse) which also sets EFLAGS. |
21280 | SDVTList VTs = DAG.getVTList(OpVT, MVT::i32); |
21281 | Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op); |
21282 | |
21283 | if (Opc == ISD::CTLZ) { |
21284 | // If src is zero (i.e. bsr sets ZF), returns NumBits. |
21285 | SDValue Ops[] = { |
21286 | Op, |
21287 | DAG.getConstant(NumBits + NumBits - 1, dl, OpVT), |
21288 | DAG.getConstant(X86::COND_E, dl, MVT::i8), |
21289 | Op.getValue(1) |
21290 | }; |
21291 | Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops); |
21292 | } |
21293 | |
21294 | // Finally xor with NumBits-1. |
21295 | Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, |
21296 | DAG.getConstant(NumBits - 1, dl, OpVT)); |
21297 | |
21298 | if (VT == MVT::i8) |
21299 | Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op); |
21300 | return Op; |
21301 | } |
21302 | |
21303 | static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) { |
21304 | MVT VT = Op.getSimpleValueType(); |
21305 | unsigned NumBits = VT.getScalarSizeInBits(); |
21306 | SDLoc dl(Op); |
21307 | |
21308 | if (VT.isVector()) { |
21309 | SDValue N0 = Op.getOperand(0); |
21310 | SDValue Zero = DAG.getConstant(0, dl, VT); |
21311 | |
21312 | // lsb(x) = (x & -x) |
21313 | SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0, |
21314 | DAG.getNode(ISD::SUB, dl, VT, Zero, N0)); |
21315 | |
21316 | // cttz_undef(x) = (width - 1) - ctlz(lsb) |
21317 | if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF) { |
21318 | SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT); |
21319 | return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne, |
21320 | DAG.getNode(ISD::CTLZ, dl, VT, LSB)); |
21321 | } |
21322 | |
21323 | // cttz(x) = ctpop(lsb - 1) |
21324 | SDValue One = DAG.getConstant(1, dl, VT); |
21325 | return DAG.getNode(ISD::CTPOP, dl, VT, |
21326 | DAG.getNode(ISD::SUB, dl, VT, LSB, One)); |
21327 | } |
21328 | |
21329 | assert(Op.getOpcode() == ISD::CTTZ &&((Op.getOpcode() == ISD::CTTZ && "Only scalar CTTZ requires custom lowering" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::CTTZ && \"Only scalar CTTZ requires custom lowering\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21330, __PRETTY_FUNCTION__)) |
21330 | "Only scalar CTTZ requires custom lowering")((Op.getOpcode() == ISD::CTTZ && "Only scalar CTTZ requires custom lowering" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::CTTZ && \"Only scalar CTTZ requires custom lowering\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21330, __PRETTY_FUNCTION__)); |
21331 | |
21332 | // Issue a bsf (scan bits forward) which also sets EFLAGS. |
21333 | SDVTList VTs = DAG.getVTList(VT, MVT::i32); |
21334 | Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0)); |
21335 | |
21336 | // If src is zero (i.e. bsf sets ZF), returns NumBits. |
21337 | SDValue Ops[] = { |
21338 | Op, |
21339 | DAG.getConstant(NumBits, dl, VT), |
21340 | DAG.getConstant(X86::COND_E, dl, MVT::i8), |
21341 | Op.getValue(1) |
21342 | }; |
21343 | return DAG.getNode(X86ISD::CMOV, dl, VT, Ops); |
21344 | } |
21345 | |
21346 | /// Break a 256-bit integer operation into two new 128-bit ones and then |
21347 | /// concatenate the result back. |
21348 | static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) { |
21349 | MVT VT = Op.getSimpleValueType(); |
21350 | |
21351 | assert(VT.is256BitVector() && VT.isInteger() &&((VT.is256BitVector() && VT.isInteger() && "Unsupported value type for operation" ) ? static_cast<void> (0) : __assert_fail ("VT.is256BitVector() && VT.isInteger() && \"Unsupported value type for operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21352, __PRETTY_FUNCTION__)) |
21352 | "Unsupported value type for operation")((VT.is256BitVector() && VT.isInteger() && "Unsupported value type for operation" ) ? static_cast<void> (0) : __assert_fail ("VT.is256BitVector() && VT.isInteger() && \"Unsupported value type for operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21352, __PRETTY_FUNCTION__)); |
21353 | |
21354 | unsigned NumElems = VT.getVectorNumElements(); |
21355 | SDLoc dl(Op); |
21356 | |
21357 | // Extract the LHS vectors |
21358 | SDValue LHS = Op.getOperand(0); |
21359 | SDValue LHS1 = extract128BitVector(LHS, 0, DAG, dl); |
21360 | SDValue LHS2 = extract128BitVector(LHS, NumElems / 2, DAG, dl); |
21361 | |
21362 | // Extract the RHS vectors |
21363 | SDValue RHS = Op.getOperand(1); |
21364 | SDValue RHS1 = extract128BitVector(RHS, 0, DAG, dl); |
21365 | SDValue RHS2 = extract128BitVector(RHS, NumElems / 2, DAG, dl); |
21366 | |
21367 | MVT EltVT = VT.getVectorElementType(); |
21368 | MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2); |
21369 | |
21370 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, |
21371 | DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1), |
21372 | DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2)); |
21373 | } |
21374 | |
21375 | /// Break a 512-bit integer operation into two new 256-bit ones and then |
21376 | /// concatenate the result back. |
21377 | static SDValue Lower512IntArith(SDValue Op, SelectionDAG &DAG) { |
21378 | MVT VT = Op.getSimpleValueType(); |
21379 | |
21380 | assert(VT.is512BitVector() && VT.isInteger() &&((VT.is512BitVector() && VT.isInteger() && "Unsupported value type for operation" ) ? static_cast<void> (0) : __assert_fail ("VT.is512BitVector() && VT.isInteger() && \"Unsupported value type for operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21381, __PRETTY_FUNCTION__)) |
21381 | "Unsupported value type for operation")((VT.is512BitVector() && VT.isInteger() && "Unsupported value type for operation" ) ? static_cast<void> (0) : __assert_fail ("VT.is512BitVector() && VT.isInteger() && \"Unsupported value type for operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21381, __PRETTY_FUNCTION__)); |
21382 | |
21383 | unsigned NumElems = VT.getVectorNumElements(); |
21384 | SDLoc dl(Op); |
21385 | |
21386 | // Extract the LHS vectors |
21387 | SDValue LHS = Op.getOperand(0); |
21388 | SDValue LHS1 = extract256BitVector(LHS, 0, DAG, dl); |
21389 | SDValue LHS2 = extract256BitVector(LHS, NumElems / 2, DAG, dl); |
21390 | |
21391 | // Extract the RHS vectors |
21392 | SDValue RHS = Op.getOperand(1); |
21393 | SDValue RHS1 = extract256BitVector(RHS, 0, DAG, dl); |
21394 | SDValue RHS2 = extract256BitVector(RHS, NumElems / 2, DAG, dl); |
21395 | |
21396 | MVT EltVT = VT.getVectorElementType(); |
21397 | MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2); |
21398 | |
21399 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, |
21400 | DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1), |
21401 | DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2)); |
21402 | } |
21403 | |
21404 | static SDValue LowerADD_SUB(SDValue Op, SelectionDAG &DAG) { |
21405 | MVT VT = Op.getSimpleValueType(); |
21406 | if (VT.getScalarType() == MVT::i1) |
21407 | return DAG.getNode(ISD::XOR, SDLoc(Op), VT, |
21408 | Op.getOperand(0), Op.getOperand(1)); |
21409 | assert(Op.getSimpleValueType().is256BitVector() &&((Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType ().isInteger() && "Only handle AVX 256-bit vector integer operation" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType().isInteger() && \"Only handle AVX 256-bit vector integer operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21411, __PRETTY_FUNCTION__)) |
21410 | Op.getSimpleValueType().isInteger() &&((Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType ().isInteger() && "Only handle AVX 256-bit vector integer operation" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType().isInteger() && \"Only handle AVX 256-bit vector integer operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21411, __PRETTY_FUNCTION__)) |
21411 | "Only handle AVX 256-bit vector integer operation")((Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType ().isInteger() && "Only handle AVX 256-bit vector integer operation" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType().isInteger() && \"Only handle AVX 256-bit vector integer operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21411, __PRETTY_FUNCTION__)); |
21412 | return Lower256IntArith(Op, DAG); |
21413 | } |
21414 | |
21415 | static SDValue LowerABS(SDValue Op, SelectionDAG &DAG) { |
21416 | assert(Op.getSimpleValueType().is256BitVector() &&((Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType ().isInteger() && "Only handle AVX 256-bit vector integer operation" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType().isInteger() && \"Only handle AVX 256-bit vector integer operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21418, __PRETTY_FUNCTION__)) |
21417 | Op.getSimpleValueType().isInteger() &&((Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType ().isInteger() && "Only handle AVX 256-bit vector integer operation" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType().isInteger() && \"Only handle AVX 256-bit vector integer operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21418, __PRETTY_FUNCTION__)) |
21418 | "Only handle AVX 256-bit vector integer operation")((Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType ().isInteger() && "Only handle AVX 256-bit vector integer operation" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType().isInteger() && \"Only handle AVX 256-bit vector integer operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21418, __PRETTY_FUNCTION__)); |
21419 | return Lower256IntUnary(Op, DAG); |
21420 | } |
21421 | |
21422 | static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) { |
21423 | assert(Op.getSimpleValueType().is256BitVector() &&((Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType ().isInteger() && "Only handle AVX 256-bit vector integer operation" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType().isInteger() && \"Only handle AVX 256-bit vector integer operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21425, __PRETTY_FUNCTION__)) |
21424 | Op.getSimpleValueType().isInteger() &&((Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType ().isInteger() && "Only handle AVX 256-bit vector integer operation" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType().isInteger() && \"Only handle AVX 256-bit vector integer operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21425, __PRETTY_FUNCTION__)) |
21425 | "Only handle AVX 256-bit vector integer operation")((Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType ().isInteger() && "Only handle AVX 256-bit vector integer operation" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().is256BitVector() && Op.getSimpleValueType().isInteger() && \"Only handle AVX 256-bit vector integer operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21425, __PRETTY_FUNCTION__)); |
21426 | return Lower256IntArith(Op, DAG); |
21427 | } |
21428 | |
21429 | static SDValue LowerMUL(SDValue Op, const X86Subtarget &Subtarget, |
21430 | SelectionDAG &DAG) { |
21431 | SDLoc dl(Op); |
21432 | MVT VT = Op.getSimpleValueType(); |
21433 | |
21434 | if (VT.getScalarType() == MVT::i1) |
21435 | return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1)); |
21436 | |
21437 | // Decompose 256-bit ops into smaller 128-bit ops. |
21438 | if (VT.is256BitVector() && !Subtarget.hasInt256()) |
21439 | return Lower256IntArith(Op, DAG); |
21440 | |
21441 | SDValue A = Op.getOperand(0); |
21442 | SDValue B = Op.getOperand(1); |
21443 | |
21444 | // Lower v16i8/v32i8/v64i8 mul as sign-extension to v8i16/v16i16/v32i16 |
21445 | // vector pairs, multiply and truncate. |
21446 | if (VT == MVT::v16i8 || VT == MVT::v32i8 || VT == MVT::v64i8) { |
21447 | if (Subtarget.hasInt256()) { |
21448 | // For 512-bit vectors, split into 256-bit vectors to allow the |
21449 | // sign-extension to occur. |
21450 | if (VT == MVT::v64i8) |
21451 | return Lower512IntArith(Op, DAG); |
21452 | |
21453 | // For 256-bit vectors, split into 128-bit vectors to allow the |
21454 | // sign-extension to occur. We don't need this on AVX512BW as we can |
21455 | // safely sign-extend to v32i16. |
21456 | if (VT == MVT::v32i8 && !Subtarget.hasBWI()) |
21457 | return Lower256IntArith(Op, DAG); |
21458 | |
21459 | MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements()); |
21460 | return DAG.getNode( |
21461 | ISD::TRUNCATE, dl, VT, |
21462 | DAG.getNode(ISD::MUL, dl, ExVT, |
21463 | DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A), |
21464 | DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B))); |
21465 | } |
21466 | |
21467 | assert(VT == MVT::v16i8 &&((VT == MVT::v16i8 && "Pre-AVX2 support only supports v16i8 multiplication" ) ? static_cast<void> (0) : __assert_fail ("VT == MVT::v16i8 && \"Pre-AVX2 support only supports v16i8 multiplication\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21468, __PRETTY_FUNCTION__)) |
21468 | "Pre-AVX2 support only supports v16i8 multiplication")((VT == MVT::v16i8 && "Pre-AVX2 support only supports v16i8 multiplication" ) ? static_cast<void> (0) : __assert_fail ("VT == MVT::v16i8 && \"Pre-AVX2 support only supports v16i8 multiplication\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21468, __PRETTY_FUNCTION__)); |
21469 | MVT ExVT = MVT::v8i16; |
21470 | |
21471 | // Extract the lo parts and sign extend to i16 |
21472 | SDValue ALo, BLo; |
21473 | if (Subtarget.hasSSE41()) { |
21474 | ALo = DAG.getSignExtendVectorInReg(A, dl, ExVT); |
21475 | BLo = DAG.getSignExtendVectorInReg(B, dl, ExVT); |
21476 | } else { |
21477 | const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3, |
21478 | -1, 4, -1, 5, -1, 6, -1, 7}; |
21479 | ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask); |
21480 | BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask); |
21481 | ALo = DAG.getBitcast(ExVT, ALo); |
21482 | BLo = DAG.getBitcast(ExVT, BLo); |
21483 | ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT)); |
21484 | BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT)); |
21485 | } |
21486 | |
21487 | // Extract the hi parts and sign extend to i16 |
21488 | SDValue AHi, BHi; |
21489 | if (Subtarget.hasSSE41()) { |
21490 | const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15, |
21491 | -1, -1, -1, -1, -1, -1, -1, -1}; |
21492 | AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask); |
21493 | BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask); |
21494 | AHi = DAG.getSignExtendVectorInReg(AHi, dl, ExVT); |
21495 | BHi = DAG.getSignExtendVectorInReg(BHi, dl, ExVT); |
21496 | } else { |
21497 | const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11, |
21498 | -1, 12, -1, 13, -1, 14, -1, 15}; |
21499 | AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask); |
21500 | BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask); |
21501 | AHi = DAG.getBitcast(ExVT, AHi); |
21502 | BHi = DAG.getBitcast(ExVT, BHi); |
21503 | AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT)); |
21504 | BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT)); |
21505 | } |
21506 | |
21507 | // Multiply, mask the lower 8bits of the lo/hi results and pack |
21508 | SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo); |
21509 | SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi); |
21510 | RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT)); |
21511 | RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT)); |
21512 | return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi); |
21513 | } |
21514 | |
21515 | // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle. |
21516 | if (VT == MVT::v4i32) { |
21517 | assert(Subtarget.hasSSE2() && !Subtarget.hasSSE41() &&((Subtarget.hasSSE2() && !Subtarget.hasSSE41() && "Should not custom lower when pmuldq is available!") ? static_cast <void> (0) : __assert_fail ("Subtarget.hasSSE2() && !Subtarget.hasSSE41() && \"Should not custom lower when pmuldq is available!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21518, __PRETTY_FUNCTION__)) |
21518 | "Should not custom lower when pmuldq is available!")((Subtarget.hasSSE2() && !Subtarget.hasSSE41() && "Should not custom lower when pmuldq is available!") ? static_cast <void> (0) : __assert_fail ("Subtarget.hasSSE2() && !Subtarget.hasSSE41() && \"Should not custom lower when pmuldq is available!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21518, __PRETTY_FUNCTION__)); |
21519 | |
21520 | // Extract the odd parts. |
21521 | static const int UnpackMask[] = { 1, -1, 3, -1 }; |
21522 | SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask); |
21523 | SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask); |
21524 | |
21525 | // Multiply the even parts. |
21526 | SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B); |
21527 | // Now multiply odd parts. |
21528 | SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds); |
21529 | |
21530 | Evens = DAG.getBitcast(VT, Evens); |
21531 | Odds = DAG.getBitcast(VT, Odds); |
21532 | |
21533 | // Merge the two vectors back together with a shuffle. This expands into 2 |
21534 | // shuffles. |
21535 | static const int ShufMask[] = { 0, 4, 2, 6 }; |
21536 | return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask); |
21537 | } |
21538 | |
21539 | assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&(((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) && "Only know how to lower V2I64/V4I64/V8I64 multiply") ? static_cast <void> (0) : __assert_fail ("(VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) && \"Only know how to lower V2I64/V4I64/V8I64 multiply\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21540, __PRETTY_FUNCTION__)) |
21540 | "Only know how to lower V2I64/V4I64/V8I64 multiply")(((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) && "Only know how to lower V2I64/V4I64/V8I64 multiply") ? static_cast <void> (0) : __assert_fail ("(VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) && \"Only know how to lower V2I64/V4I64/V8I64 multiply\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21540, __PRETTY_FUNCTION__)); |
21541 | |
21542 | // 32-bit vector types used for MULDQ/MULUDQ. |
21543 | MVT MulVT = MVT::getVectorVT(MVT::i32, VT.getSizeInBits() / 32); |
21544 | |
21545 | // MULDQ returns the 64-bit result of the signed multiplication of the lower |
21546 | // 32-bits. We can lower with this if the sign bits stretch that far. |
21547 | if (Subtarget.hasSSE41() && DAG.ComputeNumSignBits(A) > 32 && |
21548 | DAG.ComputeNumSignBits(B) > 32) { |
21549 | return DAG.getNode(X86ISD::PMULDQ, dl, VT, DAG.getBitcast(MulVT, A), |
21550 | DAG.getBitcast(MulVT, B)); |
21551 | } |
21552 | |
21553 | // Ahi = psrlqi(a, 32); |
21554 | // Bhi = psrlqi(b, 32); |
21555 | // |
21556 | // AloBlo = pmuludq(a, b); |
21557 | // AloBhi = pmuludq(a, Bhi); |
21558 | // AhiBlo = pmuludq(Ahi, b); |
21559 | // |
21560 | // Hi = psllqi(AloBhi + AhiBlo, 32); |
21561 | // return AloBlo + Hi; |
21562 | APInt LowerBitsMask = APInt::getLowBitsSet(64, 32); |
21563 | bool ALoIsZero = DAG.MaskedValueIsZero(A, LowerBitsMask); |
21564 | bool BLoIsZero = DAG.MaskedValueIsZero(B, LowerBitsMask); |
21565 | |
21566 | APInt UpperBitsMask = APInt::getHighBitsSet(64, 32); |
21567 | bool AHiIsZero = DAG.MaskedValueIsZero(A, UpperBitsMask); |
21568 | bool BHiIsZero = DAG.MaskedValueIsZero(B, UpperBitsMask); |
21569 | |
21570 | // Bit cast to 32-bit vectors for MULUDQ. |
21571 | SDValue Alo = DAG.getBitcast(MulVT, A); |
21572 | SDValue Blo = DAG.getBitcast(MulVT, B); |
21573 | |
21574 | SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl); |
21575 | |
21576 | // Only multiply lo/hi halves that aren't known to be zero. |
21577 | SDValue AloBlo = Zero; |
21578 | if (!ALoIsZero && !BLoIsZero) |
21579 | AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Alo, Blo); |
21580 | |
21581 | SDValue AloBhi = Zero; |
21582 | if (!ALoIsZero && !BHiIsZero) { |
21583 | SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG); |
21584 | Bhi = DAG.getBitcast(MulVT, Bhi); |
21585 | AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Alo, Bhi); |
21586 | } |
21587 | |
21588 | SDValue AhiBlo = Zero; |
21589 | if (!AHiIsZero && !BLoIsZero) { |
21590 | SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG); |
21591 | Ahi = DAG.getBitcast(MulVT, Ahi); |
21592 | AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, Blo); |
21593 | } |
21594 | |
21595 | SDValue Hi = DAG.getNode(ISD::ADD, dl, VT, AloBhi, AhiBlo); |
21596 | Hi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Hi, 32, DAG); |
21597 | |
21598 | return DAG.getNode(ISD::ADD, dl, VT, AloBlo, Hi); |
21599 | } |
21600 | |
21601 | static SDValue LowerMULH(SDValue Op, const X86Subtarget &Subtarget, |
21602 | SelectionDAG &DAG) { |
21603 | SDLoc dl(Op); |
21604 | MVT VT = Op.getSimpleValueType(); |
21605 | |
21606 | // Decompose 256-bit ops into smaller 128-bit ops. |
21607 | if (VT.is256BitVector() && !Subtarget.hasInt256()) |
21608 | return Lower256IntArith(Op, DAG); |
21609 | |
21610 | // Only i8 vectors should need custom lowering after this. |
21611 | assert((VT == MVT::v16i8 || (VT == MVT::v32i8 && Subtarget.hasInt256())) &&(((VT == MVT::v16i8 || (VT == MVT::v32i8 && Subtarget .hasInt256())) && "Unsupported vector type") ? static_cast <void> (0) : __assert_fail ("(VT == MVT::v16i8 || (VT == MVT::v32i8 && Subtarget.hasInt256())) && \"Unsupported vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21612, __PRETTY_FUNCTION__)) |
21612 | "Unsupported vector type")(((VT == MVT::v16i8 || (VT == MVT::v32i8 && Subtarget .hasInt256())) && "Unsupported vector type") ? static_cast <void> (0) : __assert_fail ("(VT == MVT::v16i8 || (VT == MVT::v32i8 && Subtarget.hasInt256())) && \"Unsupported vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21612, __PRETTY_FUNCTION__)); |
21613 | |
21614 | // Lower v16i8/v32i8 as extension to v8i16/v16i16 vector pairs, multiply, |
21615 | // logical shift down the upper half and pack back to i8. |
21616 | SDValue A = Op.getOperand(0); |
21617 | SDValue B = Op.getOperand(1); |
21618 | |
21619 | // With SSE41 we can use sign/zero extend, but for pre-SSE41 we unpack |
21620 | // and then ashr/lshr the upper bits down to the lower bits before multiply. |
21621 | unsigned Opcode = Op.getOpcode(); |
21622 | unsigned ExShift = (ISD::MULHU == Opcode ? ISD::SRL : ISD::SRA); |
21623 | unsigned ExSSE41 = (ISD::MULHU == Opcode ? X86ISD::VZEXT : X86ISD::VSEXT); |
21624 | |
21625 | // AVX2 implementations - extend xmm subvectors to ymm. |
21626 | if (Subtarget.hasInt256()) { |
21627 | SDValue Lo = DAG.getIntPtrConstant(0, dl); |
21628 | SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl); |
21629 | |
21630 | if (VT == MVT::v32i8) { |
21631 | SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v16i8, A, Lo); |
21632 | SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v16i8, B, Lo); |
21633 | SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v16i8, A, Hi); |
21634 | SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v16i8, B, Hi); |
21635 | ALo = DAG.getNode(ExSSE41, dl, MVT::v16i16, ALo); |
21636 | BLo = DAG.getNode(ExSSE41, dl, MVT::v16i16, BLo); |
21637 | AHi = DAG.getNode(ExSSE41, dl, MVT::v16i16, AHi); |
21638 | BHi = DAG.getNode(ExSSE41, dl, MVT::v16i16, BHi); |
21639 | Lo = DAG.getNode(ISD::SRL, dl, MVT::v16i16, |
21640 | DAG.getNode(ISD::MUL, dl, MVT::v16i16, ALo, BLo), |
21641 | DAG.getConstant(8, dl, MVT::v16i16)); |
21642 | Hi = DAG.getNode(ISD::SRL, dl, MVT::v16i16, |
21643 | DAG.getNode(ISD::MUL, dl, MVT::v16i16, AHi, BHi), |
21644 | DAG.getConstant(8, dl, MVT::v16i16)); |
21645 | // The ymm variant of PACKUS treats the 128-bit lanes separately, so before |
21646 | // using PACKUS we need to permute the inputs to the correct lo/hi xmm lane. |
21647 | const int LoMask[] = {0, 1, 2, 3, 4, 5, 6, 7, |
21648 | 16, 17, 18, 19, 20, 21, 22, 23}; |
21649 | const int HiMask[] = {8, 9, 10, 11, 12, 13, 14, 15, |
21650 | 24, 25, 26, 27, 28, 29, 30, 31}; |
21651 | return DAG.getNode(X86ISD::PACKUS, dl, VT, |
21652 | DAG.getVectorShuffle(MVT::v16i16, dl, Lo, Hi, LoMask), |
21653 | DAG.getVectorShuffle(MVT::v16i16, dl, Lo, Hi, HiMask)); |
21654 | } |
21655 | |
21656 | SDValue ExA = getExtendInVec(ExSSE41, dl, MVT::v16i16, A, DAG); |
21657 | SDValue ExB = getExtendInVec(ExSSE41, dl, MVT::v16i16, B, DAG); |
21658 | SDValue Mul = DAG.getNode(ISD::MUL, dl, MVT::v16i16, ExA, ExB); |
21659 | SDValue MulH = DAG.getNode(ISD::SRL, dl, MVT::v16i16, Mul, |
21660 | DAG.getConstant(8, dl, MVT::v16i16)); |
21661 | Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v8i16, MulH, Lo); |
21662 | Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v8i16, MulH, Hi); |
21663 | return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi); |
21664 | } |
21665 | |
21666 | assert(VT == MVT::v16i8 &&((VT == MVT::v16i8 && "Pre-AVX2 support only supports v16i8 multiplication" ) ? static_cast<void> (0) : __assert_fail ("VT == MVT::v16i8 && \"Pre-AVX2 support only supports v16i8 multiplication\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21667, __PRETTY_FUNCTION__)) |
21667 | "Pre-AVX2 support only supports v16i8 multiplication")((VT == MVT::v16i8 && "Pre-AVX2 support only supports v16i8 multiplication" ) ? static_cast<void> (0) : __assert_fail ("VT == MVT::v16i8 && \"Pre-AVX2 support only supports v16i8 multiplication\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21667, __PRETTY_FUNCTION__)); |
21668 | MVT ExVT = MVT::v8i16; |
21669 | |
21670 | // Extract the lo parts and zero/sign extend to i16. |
21671 | SDValue ALo, BLo; |
21672 | if (Subtarget.hasSSE41()) { |
21673 | ALo = getExtendInVec(ExSSE41, dl, ExVT, A, DAG); |
21674 | BLo = getExtendInVec(ExSSE41, dl, ExVT, B, DAG); |
21675 | } else { |
21676 | const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3, |
21677 | -1, 4, -1, 5, -1, 6, -1, 7}; |
21678 | ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask); |
21679 | BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask); |
21680 | ALo = DAG.getBitcast(ExVT, ALo); |
21681 | BLo = DAG.getBitcast(ExVT, BLo); |
21682 | ALo = DAG.getNode(ExShift, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT)); |
21683 | BLo = DAG.getNode(ExShift, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT)); |
21684 | } |
21685 | |
21686 | // Extract the hi parts and zero/sign extend to i16. |
21687 | SDValue AHi, BHi; |
21688 | if (Subtarget.hasSSE41()) { |
21689 | const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15, |
21690 | -1, -1, -1, -1, -1, -1, -1, -1}; |
21691 | AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask); |
21692 | BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask); |
21693 | AHi = getExtendInVec(ExSSE41, dl, ExVT, AHi, DAG); |
21694 | BHi = getExtendInVec(ExSSE41, dl, ExVT, BHi, DAG); |
21695 | } else { |
21696 | const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11, |
21697 | -1, 12, -1, 13, -1, 14, -1, 15}; |
21698 | AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask); |
21699 | BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask); |
21700 | AHi = DAG.getBitcast(ExVT, AHi); |
21701 | BHi = DAG.getBitcast(ExVT, BHi); |
21702 | AHi = DAG.getNode(ExShift, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT)); |
21703 | BHi = DAG.getNode(ExShift, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT)); |
21704 | } |
21705 | |
21706 | // Multiply, lshr the upper 8bits to the lower 8bits of the lo/hi results and |
21707 | // pack back to v16i8. |
21708 | SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo); |
21709 | SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi); |
21710 | RLo = DAG.getNode(ISD::SRL, dl, ExVT, RLo, DAG.getConstant(8, dl, ExVT)); |
21711 | RHi = DAG.getNode(ISD::SRL, dl, ExVT, RHi, DAG.getConstant(8, dl, ExVT)); |
21712 | return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi); |
21713 | } |
21714 | |
21715 | SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const { |
21716 | assert(Subtarget.isTargetWin64() && "Unexpected target")((Subtarget.isTargetWin64() && "Unexpected target") ? static_cast<void> (0) : __assert_fail ("Subtarget.isTargetWin64() && \"Unexpected target\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21716, __PRETTY_FUNCTION__)); |
21717 | EVT VT = Op.getValueType(); |
21718 | assert(VT.isInteger() && VT.getSizeInBits() == 128 &&((VT.isInteger() && VT.getSizeInBits() == 128 && "Unexpected return type for lowering") ? static_cast<void > (0) : __assert_fail ("VT.isInteger() && VT.getSizeInBits() == 128 && \"Unexpected return type for lowering\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21719, __PRETTY_FUNCTION__)) |
21719 | "Unexpected return type for lowering")((VT.isInteger() && VT.getSizeInBits() == 128 && "Unexpected return type for lowering") ? static_cast<void > (0) : __assert_fail ("VT.isInteger() && VT.getSizeInBits() == 128 && \"Unexpected return type for lowering\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21719, __PRETTY_FUNCTION__)); |
21720 | |
21721 | RTLIB::Libcall LC; |
21722 | bool isSigned; |
21723 | switch (Op->getOpcode()) { |
21724 | default: llvm_unreachable("Unexpected request for libcall!")::llvm::llvm_unreachable_internal("Unexpected request for libcall!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21724); |
21725 | case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break; |
21726 | case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break; |
21727 | case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break; |
21728 | case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break; |
21729 | case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break; |
21730 | case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break; |
21731 | } |
21732 | |
21733 | SDLoc dl(Op); |
21734 | SDValue InChain = DAG.getEntryNode(); |
21735 | |
21736 | TargetLowering::ArgListTy Args; |
21737 | TargetLowering::ArgListEntry Entry; |
21738 | for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) { |
21739 | EVT ArgVT = Op->getOperand(i).getValueType(); |
21740 | assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&((ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 && "Unexpected argument type for lowering") ? static_cast<void > (0) : __assert_fail ("ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 && \"Unexpected argument type for lowering\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21741, __PRETTY_FUNCTION__)) |
21741 | "Unexpected argument type for lowering")((ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 && "Unexpected argument type for lowering") ? static_cast<void > (0) : __assert_fail ("ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 && \"Unexpected argument type for lowering\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21741, __PRETTY_FUNCTION__)); |
21742 | SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16); |
21743 | Entry.Node = StackPtr; |
21744 | InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, |
21745 | MachinePointerInfo(), /* Alignment = */ 16); |
21746 | Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext()); |
21747 | Entry.Ty = PointerType::get(ArgTy,0); |
21748 | Entry.IsSExt = false; |
21749 | Entry.IsZExt = false; |
21750 | Args.push_back(Entry); |
21751 | } |
21752 | |
21753 | SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC), |
21754 | getPointerTy(DAG.getDataLayout())); |
21755 | |
21756 | TargetLowering::CallLoweringInfo CLI(DAG); |
21757 | CLI.setDebugLoc(dl) |
21758 | .setChain(InChain) |
21759 | .setLibCallee( |
21760 | getLibcallCallingConv(LC), |
21761 | static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()), Callee, |
21762 | std::move(Args)) |
21763 | .setInRegister() |
21764 | .setSExtResult(isSigned) |
21765 | .setZExtResult(!isSigned); |
21766 | |
21767 | std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI); |
21768 | return DAG.getBitcast(VT, CallInfo.first); |
21769 | } |
21770 | |
21771 | static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget &Subtarget, |
21772 | SelectionDAG &DAG) { |
21773 | SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1); |
21774 | MVT VT = Op0.getSimpleValueType(); |
21775 | SDLoc dl(Op); |
21776 | |
21777 | // Decompose 256-bit ops into smaller 128-bit ops. |
21778 | if (VT.is256BitVector() && !Subtarget.hasInt256()) { |
21779 | unsigned Opcode = Op.getOpcode(); |
21780 | unsigned NumElems = VT.getVectorNumElements(); |
21781 | MVT HalfVT = MVT::getVectorVT(VT.getScalarType(), NumElems / 2); |
21782 | SDValue Lo0 = extract128BitVector(Op0, 0, DAG, dl); |
21783 | SDValue Lo1 = extract128BitVector(Op1, 0, DAG, dl); |
21784 | SDValue Hi0 = extract128BitVector(Op0, NumElems / 2, DAG, dl); |
21785 | SDValue Hi1 = extract128BitVector(Op1, NumElems / 2, DAG, dl); |
21786 | SDValue Lo = DAG.getNode(Opcode, dl, DAG.getVTList(HalfVT, HalfVT), Lo0, Lo1); |
21787 | SDValue Hi = DAG.getNode(Opcode, dl, DAG.getVTList(HalfVT, HalfVT), Hi0, Hi1); |
21788 | SDValue Ops[] = { |
21789 | DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo.getValue(0), Hi.getValue(0)), |
21790 | DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo.getValue(1), Hi.getValue(1)) |
21791 | }; |
21792 | return DAG.getMergeValues(Ops, dl); |
21793 | } |
21794 | |
21795 | assert((VT == MVT::v4i32 && Subtarget.hasSSE2()) ||(((VT == MVT::v4i32 && Subtarget.hasSSE2()) || (VT == MVT::v8i32 && Subtarget.hasInt256())) ? static_cast< void> (0) : __assert_fail ("(VT == MVT::v4i32 && Subtarget.hasSSE2()) || (VT == MVT::v8i32 && Subtarget.hasInt256())" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21796, __PRETTY_FUNCTION__)) |
21796 | (VT == MVT::v8i32 && Subtarget.hasInt256()))(((VT == MVT::v4i32 && Subtarget.hasSSE2()) || (VT == MVT::v8i32 && Subtarget.hasInt256())) ? static_cast< void> (0) : __assert_fail ("(VT == MVT::v4i32 && Subtarget.hasSSE2()) || (VT == MVT::v8i32 && Subtarget.hasInt256())" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21796, __PRETTY_FUNCTION__)); |
21797 | |
21798 | // PMULxD operations multiply each even value (starting at 0) of LHS with |
21799 | // the related value of RHS and produce a widen result. |
21800 | // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h> |
21801 | // => <2 x i64> <ae|cg> |
21802 | // |
21803 | // In other word, to have all the results, we need to perform two PMULxD: |
21804 | // 1. one with the even values. |
21805 | // 2. one with the odd values. |
21806 | // To achieve #2, with need to place the odd values at an even position. |
21807 | // |
21808 | // Place the odd value at an even position (basically, shift all values 1 |
21809 | // step to the left): |
21810 | const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1}; |
21811 | // <a|b|c|d> => <b|undef|d|undef> |
21812 | SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, |
21813 | makeArrayRef(&Mask[0], VT.getVectorNumElements())); |
21814 | // <e|f|g|h> => <f|undef|h|undef> |
21815 | SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, |
21816 | makeArrayRef(&Mask[0], VT.getVectorNumElements())); |
21817 | |
21818 | // Emit two multiplies, one for the lower 2 ints and one for the higher 2 |
21819 | // ints. |
21820 | MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64; |
21821 | bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI; |
21822 | unsigned Opcode = |
21823 | (!IsSigned || !Subtarget.hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ; |
21824 | // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h> |
21825 | // => <2 x i64> <ae|cg> |
21826 | SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1)); |
21827 | // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef> |
21828 | // => <2 x i64> <bf|dh> |
21829 | SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1)); |
21830 | |
21831 | // Shuffle it back into the right order. |
21832 | SDValue Highs, Lows; |
21833 | if (VT == MVT::v8i32) { |
21834 | const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15}; |
21835 | Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask); |
21836 | const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14}; |
21837 | Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask); |
21838 | } else { |
21839 | const int HighMask[] = {1, 5, 3, 7}; |
21840 | Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask); |
21841 | const int LowMask[] = {0, 4, 2, 6}; |
21842 | Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask); |
21843 | } |
21844 | |
21845 | // If we have a signed multiply but no PMULDQ fix up the high parts of a |
21846 | // unsigned multiply. |
21847 | if (IsSigned && !Subtarget.hasSSE41()) { |
21848 | SDValue ShAmt = DAG.getConstant( |
21849 | 31, dl, |
21850 | DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout())); |
21851 | SDValue T1 = DAG.getNode(ISD::AND, dl, VT, |
21852 | DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1); |
21853 | SDValue T2 = DAG.getNode(ISD::AND, dl, VT, |
21854 | DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0); |
21855 | |
21856 | SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2); |
21857 | Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup); |
21858 | } |
21859 | |
21860 | // The first result of MUL_LOHI is actually the low value, followed by the |
21861 | // high value. |
21862 | SDValue Ops[] = {Lows, Highs}; |
21863 | return DAG.getMergeValues(Ops, dl); |
21864 | } |
21865 | |
21866 | // Return true if the required (according to Opcode) shift-imm form is natively |
21867 | // supported by the Subtarget |
21868 | static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget &Subtarget, |
21869 | unsigned Opcode) { |
21870 | if (VT.getScalarSizeInBits() < 16) |
21871 | return false; |
21872 | |
21873 | if (VT.is512BitVector() && Subtarget.hasAVX512() && |
21874 | (VT.getScalarSizeInBits() > 16 || Subtarget.hasBWI())) |
21875 | return true; |
21876 | |
21877 | bool LShift = (VT.is128BitVector() && Subtarget.hasSSE2()) || |
21878 | (VT.is256BitVector() && Subtarget.hasInt256()); |
21879 | |
21880 | bool AShift = LShift && (Subtarget.hasAVX512() || |
21881 | (VT != MVT::v2i64 && VT != MVT::v4i64)); |
21882 | return (Opcode == ISD::SRA) ? AShift : LShift; |
21883 | } |
21884 | |
21885 | // The shift amount is a variable, but it is the same for all vector lanes. |
21886 | // These instructions are defined together with shift-immediate. |
21887 | static |
21888 | bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget &Subtarget, |
21889 | unsigned Opcode) { |
21890 | return SupportedVectorShiftWithImm(VT, Subtarget, Opcode); |
21891 | } |
21892 | |
21893 | // Return true if the required (according to Opcode) variable-shift form is |
21894 | // natively supported by the Subtarget |
21895 | static bool SupportedVectorVarShift(MVT VT, const X86Subtarget &Subtarget, |
21896 | unsigned Opcode) { |
21897 | |
21898 | if (!Subtarget.hasInt256() || VT.getScalarSizeInBits() < 16) |
21899 | return false; |
21900 | |
21901 | // vXi16 supported only on AVX-512, BWI |
21902 | if (VT.getScalarSizeInBits() == 16 && !Subtarget.hasBWI()) |
21903 | return false; |
21904 | |
21905 | if (Subtarget.hasAVX512()) |
21906 | return true; |
21907 | |
21908 | bool LShift = VT.is128BitVector() || VT.is256BitVector(); |
21909 | bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64; |
21910 | return (Opcode == ISD::SRA) ? AShift : LShift; |
21911 | } |
21912 | |
21913 | static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG, |
21914 | const X86Subtarget &Subtarget) { |
21915 | MVT VT = Op.getSimpleValueType(); |
21916 | SDLoc dl(Op); |
21917 | SDValue R = Op.getOperand(0); |
21918 | SDValue Amt = Op.getOperand(1); |
21919 | |
21920 | unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI : |
21921 | (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI; |
21922 | |
21923 | auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) { |
21924 | assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type")(((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v2i64 || VT == MVT::v4i64) && \"Unexpected SRA type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21924, __PRETTY_FUNCTION__)); |
21925 | MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2); |
21926 | SDValue Ex = DAG.getBitcast(ExVT, R); |
21927 | |
21928 | // ashr(R, 63) === cmp_slt(R, 0) |
21929 | if (ShiftAmt == 63 && Subtarget.hasSSE42()) { |
21930 | assert((VT != MVT::v4i64 || Subtarget.hasInt256()) &&(((VT != MVT::v4i64 || Subtarget.hasInt256()) && "Unsupported PCMPGT op" ) ? static_cast<void> (0) : __assert_fail ("(VT != MVT::v4i64 || Subtarget.hasInt256()) && \"Unsupported PCMPGT op\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21931, __PRETTY_FUNCTION__)) |
21931 | "Unsupported PCMPGT op")(((VT != MVT::v4i64 || Subtarget.hasInt256()) && "Unsupported PCMPGT op" ) ? static_cast<void> (0) : __assert_fail ("(VT != MVT::v4i64 || Subtarget.hasInt256()) && \"Unsupported PCMPGT op\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21931, __PRETTY_FUNCTION__)); |
21932 | return DAG.getNode(X86ISD::PCMPGT, dl, VT, |
21933 | getZeroVector(VT, Subtarget, DAG, dl), R); |
21934 | } |
21935 | |
21936 | if (ShiftAmt >= 32) { |
21937 | // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32. |
21938 | SDValue Upper = |
21939 | getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG); |
21940 | SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, |
21941 | ShiftAmt - 32, DAG); |
21942 | if (VT == MVT::v2i64) |
21943 | Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3}); |
21944 | if (VT == MVT::v4i64) |
21945 | Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, |
21946 | {9, 1, 11, 3, 13, 5, 15, 7}); |
21947 | } else { |
21948 | // SRA upper i32, SHL whole i64 and select lower i32. |
21949 | SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, |
21950 | ShiftAmt, DAG); |
21951 | SDValue Lower = |
21952 | getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG); |
21953 | Lower = DAG.getBitcast(ExVT, Lower); |
21954 | if (VT == MVT::v2i64) |
21955 | Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3}); |
21956 | if (VT == MVT::v4i64) |
21957 | Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, |
21958 | {8, 1, 10, 3, 12, 5, 14, 7}); |
21959 | } |
21960 | return DAG.getBitcast(VT, Ex); |
21961 | }; |
21962 | |
21963 | // Optimize shl/srl/sra with constant shift amount. |
21964 | if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) { |
21965 | if (auto *ShiftConst = BVAmt->getConstantSplatNode()) { |
21966 | uint64_t ShiftAmt = ShiftConst->getZExtValue(); |
21967 | |
21968 | if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode())) |
21969 | return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG); |
21970 | |
21971 | // i64 SRA needs to be performed as partial shifts. |
21972 | if ((VT == MVT::v2i64 || (Subtarget.hasInt256() && VT == MVT::v4i64)) && |
21973 | Op.getOpcode() == ISD::SRA && !Subtarget.hasXOP()) |
21974 | return ArithmeticShiftRight64(ShiftAmt); |
21975 | |
21976 | if (VT == MVT::v16i8 || |
21977 | (Subtarget.hasInt256() && VT == MVT::v32i8) || |
21978 | VT == MVT::v64i8) { |
21979 | unsigned NumElts = VT.getVectorNumElements(); |
21980 | MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2); |
21981 | |
21982 | // Simple i8 add case |
21983 | if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1) |
21984 | return DAG.getNode(ISD::ADD, dl, VT, R, R); |
21985 | |
21986 | // ashr(R, 7) === cmp_slt(R, 0) |
21987 | if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) { |
21988 | SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl); |
21989 | if (VT.is512BitVector()) { |
21990 | assert(VT == MVT::v64i8 && "Unexpected element type!")((VT == MVT::v64i8 && "Unexpected element type!") ? static_cast <void> (0) : __assert_fail ("VT == MVT::v64i8 && \"Unexpected element type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 21990, __PRETTY_FUNCTION__)); |
21991 | SDValue CMP = DAG.getNode(X86ISD::PCMPGTM, dl, MVT::v64i1, Zeros, R); |
21992 | return DAG.getNode(ISD::SIGN_EXTEND, dl, VT, CMP); |
21993 | } |
21994 | return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R); |
21995 | } |
21996 | |
21997 | // XOP can shift v16i8 directly instead of as shift v8i16 + mask. |
21998 | if (VT == MVT::v16i8 && Subtarget.hasXOP()) |
21999 | return SDValue(); |
22000 | |
22001 | if (Op.getOpcode() == ISD::SHL) { |
22002 | // Make a large shift. |
22003 | SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT, |
22004 | R, ShiftAmt, DAG); |
22005 | SHL = DAG.getBitcast(VT, SHL); |
22006 | // Zero out the rightmost bits. |
22007 | return DAG.getNode(ISD::AND, dl, VT, SHL, |
22008 | DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, VT)); |
22009 | } |
22010 | if (Op.getOpcode() == ISD::SRL) { |
22011 | // Make a large shift. |
22012 | SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT, |
22013 | R, ShiftAmt, DAG); |
22014 | SRL = DAG.getBitcast(VT, SRL); |
22015 | // Zero out the leftmost bits. |
22016 | return DAG.getNode(ISD::AND, dl, VT, SRL, |
22017 | DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, VT)); |
22018 | } |
22019 | if (Op.getOpcode() == ISD::SRA) { |
22020 | // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask) |
22021 | SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt); |
22022 | |
22023 | SDValue Mask = DAG.getConstant(128 >> ShiftAmt, dl, VT); |
22024 | Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask); |
22025 | Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask); |
22026 | return Res; |
22027 | } |
22028 | llvm_unreachable("Unknown shift opcode.")::llvm::llvm_unreachable_internal("Unknown shift opcode.", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22028); |
22029 | } |
22030 | } |
22031 | } |
22032 | |
22033 | // Special case in 32-bit mode, where i64 is expanded into high and low parts. |
22034 | // TODO: Replace constant extraction with getTargetConstantBitsFromNode. |
22035 | if (!Subtarget.is64Bit() && !Subtarget.hasXOP() && |
22036 | (VT == MVT::v2i64 || (Subtarget.hasInt256() && VT == MVT::v4i64) || |
22037 | (Subtarget.hasAVX512() && VT == MVT::v8i64))) { |
22038 | |
22039 | // AVX1 targets maybe extracting a 128-bit vector from a 256-bit constant. |
22040 | unsigned SubVectorScale = 1; |
22041 | if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR) { |
22042 | SubVectorScale = |
22043 | Amt.getOperand(0).getValueSizeInBits() / Amt.getValueSizeInBits(); |
22044 | Amt = Amt.getOperand(0); |
22045 | } |
22046 | |
22047 | // Peek through any splat that was introduced for i64 shift vectorization. |
22048 | int SplatIndex = -1; |
22049 | if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode())) |
22050 | if (SVN->isSplat()) { |
22051 | SplatIndex = SVN->getSplatIndex(); |
22052 | Amt = Amt.getOperand(0); |
22053 | assert(SplatIndex < (int)VT.getVectorNumElements() &&((SplatIndex < (int)VT.getVectorNumElements() && "Splat shuffle referencing second operand" ) ? static_cast<void> (0) : __assert_fail ("SplatIndex < (int)VT.getVectorNumElements() && \"Splat shuffle referencing second operand\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22054, __PRETTY_FUNCTION__)) |
22054 | "Splat shuffle referencing second operand")((SplatIndex < (int)VT.getVectorNumElements() && "Splat shuffle referencing second operand" ) ? static_cast<void> (0) : __assert_fail ("SplatIndex < (int)VT.getVectorNumElements() && \"Splat shuffle referencing second operand\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22054, __PRETTY_FUNCTION__)); |
22055 | } |
22056 | |
22057 | if (Amt.getOpcode() != ISD::BITCAST || |
22058 | Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR) |
22059 | return SDValue(); |
22060 | |
22061 | Amt = Amt.getOperand(0); |
22062 | unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() / |
22063 | (SubVectorScale * VT.getVectorNumElements()); |
22064 | unsigned RatioInLog2 = Log2_32_Ceil(Ratio); |
22065 | uint64_t ShiftAmt = 0; |
22066 | unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio); |
22067 | for (unsigned i = 0; i != Ratio; ++i) { |
22068 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp)); |
22069 | if (!C) |
22070 | return SDValue(); |
22071 | // 6 == Log2(64) |
22072 | ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2))); |
22073 | } |
22074 | |
22075 | // Check remaining shift amounts (if not a splat). |
22076 | if (SplatIndex < 0) { |
22077 | for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) { |
22078 | uint64_t ShAmt = 0; |
22079 | for (unsigned j = 0; j != Ratio; ++j) { |
22080 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j)); |
22081 | if (!C) |
22082 | return SDValue(); |
22083 | // 6 == Log2(64) |
22084 | ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2))); |
22085 | } |
22086 | if (ShAmt != ShiftAmt) |
22087 | return SDValue(); |
22088 | } |
22089 | } |
22090 | |
22091 | if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode())) |
22092 | return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG); |
22093 | |
22094 | if (Op.getOpcode() == ISD::SRA) |
22095 | return ArithmeticShiftRight64(ShiftAmt); |
22096 | } |
22097 | |
22098 | return SDValue(); |
22099 | } |
22100 | |
22101 | static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG, |
22102 | const X86Subtarget &Subtarget) { |
22103 | MVT VT = Op.getSimpleValueType(); |
22104 | SDLoc dl(Op); |
22105 | SDValue R = Op.getOperand(0); |
22106 | SDValue Amt = Op.getOperand(1); |
22107 | |
22108 | unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI : |
22109 | (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI; |
22110 | |
22111 | unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL : |
22112 | (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA; |
22113 | |
22114 | if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) { |
22115 | SDValue BaseShAmt; |
22116 | MVT EltVT = VT.getVectorElementType(); |
22117 | |
22118 | if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) { |
22119 | // Check if this build_vector node is doing a splat. |
22120 | // If so, then set BaseShAmt equal to the splat value. |
22121 | BaseShAmt = BV->getSplatValue(); |
22122 | if (BaseShAmt && BaseShAmt.isUndef()) |
22123 | BaseShAmt = SDValue(); |
22124 | } else { |
22125 | if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR) |
22126 | Amt = Amt.getOperand(0); |
22127 | |
22128 | ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt); |
22129 | if (SVN && SVN->isSplat()) { |
22130 | unsigned SplatIdx = (unsigned)SVN->getSplatIndex(); |
22131 | SDValue InVec = Amt.getOperand(0); |
22132 | if (InVec.getOpcode() == ISD::BUILD_VECTOR) { |
22133 | assert((SplatIdx < InVec.getSimpleValueType().getVectorNumElements()) &&(((SplatIdx < InVec.getSimpleValueType().getVectorNumElements ()) && "Unexpected shuffle index found!") ? static_cast <void> (0) : __assert_fail ("(SplatIdx < InVec.getSimpleValueType().getVectorNumElements()) && \"Unexpected shuffle index found!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22134, __PRETTY_FUNCTION__)) |
22134 | "Unexpected shuffle index found!")(((SplatIdx < InVec.getSimpleValueType().getVectorNumElements ()) && "Unexpected shuffle index found!") ? static_cast <void> (0) : __assert_fail ("(SplatIdx < InVec.getSimpleValueType().getVectorNumElements()) && \"Unexpected shuffle index found!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22134, __PRETTY_FUNCTION__)); |
22135 | BaseShAmt = InVec.getOperand(SplatIdx); |
22136 | } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) { |
22137 | if (ConstantSDNode *C = |
22138 | dyn_cast<ConstantSDNode>(InVec.getOperand(2))) { |
22139 | if (C->getZExtValue() == SplatIdx) |
22140 | BaseShAmt = InVec.getOperand(1); |
22141 | } |
22142 | } |
22143 | |
22144 | if (!BaseShAmt) |
22145 | // Avoid introducing an extract element from a shuffle. |
22146 | BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec, |
22147 | DAG.getIntPtrConstant(SplatIdx, dl)); |
22148 | } |
22149 | } |
22150 | |
22151 | if (BaseShAmt.getNode()) { |
22152 | assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!")((EltVT.bitsLE(MVT::i64) && "Unexpected element type!" ) ? static_cast<void> (0) : __assert_fail ("EltVT.bitsLE(MVT::i64) && \"Unexpected element type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22152, __PRETTY_FUNCTION__)); |
22153 | if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32)) |
22154 | BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt); |
22155 | else if (EltVT.bitsLT(MVT::i32)) |
22156 | BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt); |
22157 | |
22158 | return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, Subtarget, DAG); |
22159 | } |
22160 | } |
22161 | |
22162 | // Special case in 32-bit mode, where i64 is expanded into high and low parts. |
22163 | if (!Subtarget.is64Bit() && VT == MVT::v2i64 && |
22164 | Amt.getOpcode() == ISD::BITCAST && |
22165 | Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) { |
22166 | Amt = Amt.getOperand(0); |
22167 | unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() / |
22168 | VT.getVectorNumElements(); |
22169 | std::vector<SDValue> Vals(Ratio); |
22170 | for (unsigned i = 0; i != Ratio; ++i) |
22171 | Vals[i] = Amt.getOperand(i); |
22172 | for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) { |
22173 | for (unsigned j = 0; j != Ratio; ++j) |
22174 | if (Vals[j] != Amt.getOperand(i + j)) |
22175 | return SDValue(); |
22176 | } |
22177 | |
22178 | if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) |
22179 | return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1)); |
22180 | } |
22181 | return SDValue(); |
22182 | } |
22183 | |
22184 | static SDValue LowerShift(SDValue Op, const X86Subtarget &Subtarget, |
22185 | SelectionDAG &DAG) { |
22186 | MVT VT = Op.getSimpleValueType(); |
22187 | SDLoc dl(Op); |
22188 | SDValue R = Op.getOperand(0); |
22189 | SDValue Amt = Op.getOperand(1); |
22190 | bool ConstantAmt = ISD::isBuildVectorOfConstantSDNodes(Amt.getNode()); |
22191 | |
22192 | assert(VT.isVector() && "Custom lowering only for vector shifts!")((VT.isVector() && "Custom lowering only for vector shifts!" ) ? static_cast<void> (0) : __assert_fail ("VT.isVector() && \"Custom lowering only for vector shifts!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22192, __PRETTY_FUNCTION__)); |
22193 | assert(Subtarget.hasSSE2() && "Only custom lower when we have SSE2!")((Subtarget.hasSSE2() && "Only custom lower when we have SSE2!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSE2() && \"Only custom lower when we have SSE2!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22193, __PRETTY_FUNCTION__)); |
22194 | |
22195 | if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget)) |
22196 | return V; |
22197 | |
22198 | if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget)) |
22199 | return V; |
22200 | |
22201 | if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode())) |
22202 | return Op; |
22203 | |
22204 | // XOP has 128-bit variable logical/arithmetic shifts. |
22205 | // +ve/-ve Amt = shift left/right. |
22206 | if (Subtarget.hasXOP() && |
22207 | (VT == MVT::v2i64 || VT == MVT::v4i32 || |
22208 | VT == MVT::v8i16 || VT == MVT::v16i8)) { |
22209 | if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) { |
22210 | SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl); |
22211 | Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt); |
22212 | } |
22213 | if (Op.getOpcode() == ISD::SHL || Op.getOpcode() == ISD::SRL) |
22214 | return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt); |
22215 | if (Op.getOpcode() == ISD::SRA) |
22216 | return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt); |
22217 | } |
22218 | |
22219 | // 2i64 vector logical shifts can efficiently avoid scalarization - do the |
22220 | // shifts per-lane and then shuffle the partial results back together. |
22221 | if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) { |
22222 | // Splat the shift amounts so the scalar shifts above will catch it. |
22223 | SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0}); |
22224 | SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1}); |
22225 | SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0); |
22226 | SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1); |
22227 | return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3}); |
22228 | } |
22229 | |
22230 | // i64 vector arithmetic shift can be emulated with the transform: |
22231 | // M = lshr(SIGN_MASK, Amt) |
22232 | // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M) |
22233 | if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget.hasInt256())) && |
22234 | Op.getOpcode() == ISD::SRA) { |
22235 | SDValue S = DAG.getConstant(APInt::getSignMask(64), dl, VT); |
22236 | SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt); |
22237 | R = DAG.getNode(ISD::SRL, dl, VT, R, Amt); |
22238 | R = DAG.getNode(ISD::XOR, dl, VT, R, M); |
22239 | R = DAG.getNode(ISD::SUB, dl, VT, R, M); |
22240 | return R; |
22241 | } |
22242 | |
22243 | // If possible, lower this packed shift into a vector multiply instead of |
22244 | // expanding it into a sequence of scalar shifts. |
22245 | // Do this only if the vector shift count is a constant build_vector. |
22246 | if (ConstantAmt && Op.getOpcode() == ISD::SHL && |
22247 | (VT == MVT::v8i16 || VT == MVT::v4i32 || |
22248 | (Subtarget.hasInt256() && VT == MVT::v16i16))) { |
22249 | SmallVector<SDValue, 8> Elts; |
22250 | MVT SVT = VT.getVectorElementType(); |
22251 | unsigned SVTBits = SVT.getSizeInBits(); |
22252 | APInt One(SVTBits, 1); |
22253 | unsigned NumElems = VT.getVectorNumElements(); |
22254 | |
22255 | for (unsigned i=0; i !=NumElems; ++i) { |
22256 | SDValue Op = Amt->getOperand(i); |
22257 | if (Op->isUndef()) { |
22258 | Elts.push_back(Op); |
22259 | continue; |
22260 | } |
22261 | |
22262 | ConstantSDNode *ND = cast<ConstantSDNode>(Op); |
22263 | APInt C(SVTBits, ND->getAPIntValue().getZExtValue()); |
22264 | uint64_t ShAmt = C.getZExtValue(); |
22265 | if (ShAmt >= SVTBits) { |
22266 | Elts.push_back(DAG.getUNDEF(SVT)); |
22267 | continue; |
22268 | } |
22269 | Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT)); |
22270 | } |
22271 | SDValue BV = DAG.getBuildVector(VT, dl, Elts); |
22272 | return DAG.getNode(ISD::MUL, dl, VT, R, BV); |
22273 | } |
22274 | |
22275 | // Lower SHL with variable shift amount. |
22276 | if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) { |
22277 | Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT)); |
22278 | |
22279 | Op = DAG.getNode(ISD::ADD, dl, VT, Op, |
22280 | DAG.getConstant(0x3f800000U, dl, VT)); |
22281 | Op = DAG.getBitcast(MVT::v4f32, Op); |
22282 | Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op); |
22283 | return DAG.getNode(ISD::MUL, dl, VT, Op, R); |
22284 | } |
22285 | |
22286 | // If possible, lower this shift as a sequence of two shifts by |
22287 | // constant plus a MOVSS/MOVSD/PBLEND instead of scalarizing it. |
22288 | // Example: |
22289 | // (v4i32 (srl A, (build_vector < X, Y, Y, Y>))) |
22290 | // |
22291 | // Could be rewritten as: |
22292 | // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>))) |
22293 | // |
22294 | // The advantage is that the two shifts from the example would be |
22295 | // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing |
22296 | // the vector shift into four scalar shifts plus four pairs of vector |
22297 | // insert/extract. |
22298 | if (ConstantAmt && (VT == MVT::v8i16 || VT == MVT::v4i32)) { |
22299 | unsigned TargetOpcode = X86ISD::MOVSS; |
22300 | bool CanBeSimplified; |
22301 | // The splat value for the first packed shift (the 'X' from the example). |
22302 | SDValue Amt1 = Amt->getOperand(0); |
22303 | // The splat value for the second packed shift (the 'Y' from the example). |
22304 | SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) : Amt->getOperand(2); |
22305 | |
22306 | // See if it is possible to replace this node with a sequence of |
22307 | // two shifts followed by a MOVSS/MOVSD/PBLEND. |
22308 | if (VT == MVT::v4i32) { |
22309 | // Check if it is legal to use a MOVSS. |
22310 | CanBeSimplified = Amt2 == Amt->getOperand(2) && |
22311 | Amt2 == Amt->getOperand(3); |
22312 | if (!CanBeSimplified) { |
22313 | // Otherwise, check if we can still simplify this node using a MOVSD. |
22314 | CanBeSimplified = Amt1 == Amt->getOperand(1) && |
22315 | Amt->getOperand(2) == Amt->getOperand(3); |
22316 | TargetOpcode = X86ISD::MOVSD; |
22317 | Amt2 = Amt->getOperand(2); |
22318 | } |
22319 | } else { |
22320 | // Do similar checks for the case where the machine value type |
22321 | // is MVT::v8i16. |
22322 | CanBeSimplified = Amt1 == Amt->getOperand(1); |
22323 | for (unsigned i=3; i != 8 && CanBeSimplified; ++i) |
22324 | CanBeSimplified = Amt2 == Amt->getOperand(i); |
22325 | |
22326 | if (!CanBeSimplified) { |
22327 | TargetOpcode = X86ISD::MOVSD; |
22328 | CanBeSimplified = true; |
22329 | Amt2 = Amt->getOperand(4); |
22330 | for (unsigned i=0; i != 4 && CanBeSimplified; ++i) |
22331 | CanBeSimplified = Amt1 == Amt->getOperand(i); |
22332 | for (unsigned j=4; j != 8 && CanBeSimplified; ++j) |
22333 | CanBeSimplified = Amt2 == Amt->getOperand(j); |
22334 | } |
22335 | } |
22336 | |
22337 | if (CanBeSimplified && isa<ConstantSDNode>(Amt1) && |
22338 | isa<ConstantSDNode>(Amt2)) { |
22339 | // Replace this node with two shifts followed by a MOVSS/MOVSD/PBLEND. |
22340 | MVT CastVT = MVT::v4i32; |
22341 | SDValue Splat1 = |
22342 | DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT); |
22343 | SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1); |
22344 | SDValue Splat2 = |
22345 | DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT); |
22346 | SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2); |
22347 | SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1); |
22348 | SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2); |
22349 | if (TargetOpcode == X86ISD::MOVSD) |
22350 | return DAG.getBitcast(VT, DAG.getVectorShuffle(CastVT, dl, BitCast1, |
22351 | BitCast2, {0, 1, 6, 7})); |
22352 | return DAG.getBitcast(VT, DAG.getVectorShuffle(CastVT, dl, BitCast1, |
22353 | BitCast2, {0, 5, 6, 7})); |
22354 | } |
22355 | } |
22356 | |
22357 | // v4i32 Non Uniform Shifts. |
22358 | // If the shift amount is constant we can shift each lane using the SSE2 |
22359 | // immediate shifts, else we need to zero-extend each lane to the lower i64 |
22360 | // and shift using the SSE2 variable shifts. |
22361 | // The separate results can then be blended together. |
22362 | if (VT == MVT::v4i32) { |
22363 | unsigned Opc = Op.getOpcode(); |
22364 | SDValue Amt0, Amt1, Amt2, Amt3; |
22365 | if (ConstantAmt) { |
22366 | Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0}); |
22367 | Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1}); |
22368 | Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2}); |
22369 | Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3}); |
22370 | } else { |
22371 | // ISD::SHL is handled above but we include it here for completeness. |
22372 | switch (Opc) { |
22373 | default: |
22374 | llvm_unreachable("Unknown target vector shift node")::llvm::llvm_unreachable_internal("Unknown target vector shift node" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22374); |
22375 | case ISD::SHL: |
22376 | Opc = X86ISD::VSHL; |
22377 | break; |
22378 | case ISD::SRL: |
22379 | Opc = X86ISD::VSRL; |
22380 | break; |
22381 | case ISD::SRA: |
22382 | Opc = X86ISD::VSRA; |
22383 | break; |
22384 | } |
22385 | // The SSE2 shifts use the lower i64 as the same shift amount for |
22386 | // all lanes and the upper i64 is ignored. These shuffle masks |
22387 | // optimally zero-extend each lanes on SSE2/SSE41/AVX targets. |
22388 | SDValue Z = getZeroVector(VT, Subtarget, DAG, dl); |
22389 | Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1}); |
22390 | Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1}); |
22391 | Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1}); |
22392 | Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1}); |
22393 | } |
22394 | |
22395 | SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0); |
22396 | SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1); |
22397 | SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2); |
22398 | SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3); |
22399 | SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1}); |
22400 | SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7}); |
22401 | return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7}); |
22402 | } |
22403 | |
22404 | // It's worth extending once and using the vXi16/vXi32 shifts for smaller |
22405 | // types, but without AVX512 the extra overheads to get from vXi8 to vXi32 |
22406 | // make the existing SSE solution better. |
22407 | if ((Subtarget.hasInt256() && VT == MVT::v8i16) || |
22408 | (Subtarget.hasAVX512() && VT == MVT::v16i16) || |
22409 | (Subtarget.hasAVX512() && VT == MVT::v16i8) || |
22410 | (Subtarget.hasBWI() && VT == MVT::v32i8)) { |
22411 | MVT EvtSVT = (VT == MVT::v32i8 ? MVT::i16 : MVT::i32); |
22412 | MVT ExtVT = MVT::getVectorVT(EvtSVT, VT.getVectorNumElements()); |
22413 | unsigned ExtOpc = |
22414 | Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; |
22415 | R = DAG.getNode(ExtOpc, dl, ExtVT, R); |
22416 | Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt); |
22417 | return DAG.getNode(ISD::TRUNCATE, dl, VT, |
22418 | DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt)); |
22419 | } |
22420 | |
22421 | if (VT == MVT::v16i8 || |
22422 | (VT == MVT::v32i8 && Subtarget.hasInt256() && !Subtarget.hasXOP()) || |
22423 | (VT == MVT::v64i8 && Subtarget.hasBWI())) { |
22424 | MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2); |
22425 | unsigned ShiftOpcode = Op->getOpcode(); |
22426 | |
22427 | auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) { |
22428 | if (VT.is512BitVector()) { |
22429 | // On AVX512BW targets we make use of the fact that VSELECT lowers |
22430 | // to a masked blend which selects bytes based just on the sign bit |
22431 | // extracted to a mask. |
22432 | MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements()); |
22433 | V0 = DAG.getBitcast(VT, V0); |
22434 | V1 = DAG.getBitcast(VT, V1); |
22435 | Sel = DAG.getBitcast(VT, Sel); |
22436 | Sel = DAG.getNode(X86ISD::CVT2MASK, dl, MaskVT, Sel); |
22437 | return DAG.getBitcast(SelVT, DAG.getSelect(dl, VT, Sel, V0, V1)); |
22438 | } else if (Subtarget.hasSSE41()) { |
22439 | // On SSE41 targets we make use of the fact that VSELECT lowers |
22440 | // to PBLENDVB which selects bytes based just on the sign bit. |
22441 | V0 = DAG.getBitcast(VT, V0); |
22442 | V1 = DAG.getBitcast(VT, V1); |
22443 | Sel = DAG.getBitcast(VT, Sel); |
22444 | return DAG.getBitcast(SelVT, DAG.getSelect(dl, VT, Sel, V0, V1)); |
22445 | } |
22446 | // On pre-SSE41 targets we test for the sign bit by comparing to |
22447 | // zero - a negative value will set all bits of the lanes to true |
22448 | // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering. |
22449 | SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl); |
22450 | SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel); |
22451 | return DAG.getSelect(dl, SelVT, C, V0, V1); |
22452 | }; |
22453 | |
22454 | // Turn 'a' into a mask suitable for VSELECT: a = a << 5; |
22455 | // We can safely do this using i16 shifts as we're only interested in |
22456 | // the 3 lower bits of each byte. |
22457 | Amt = DAG.getBitcast(ExtVT, Amt); |
22458 | Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT)); |
22459 | Amt = DAG.getBitcast(VT, Amt); |
22460 | |
22461 | if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) { |
22462 | // r = VSELECT(r, shift(r, 4), a); |
22463 | SDValue M = |
22464 | DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT)); |
22465 | R = SignBitSelect(VT, Amt, M, R); |
22466 | |
22467 | // a += a |
22468 | Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt); |
22469 | |
22470 | // r = VSELECT(r, shift(r, 2), a); |
22471 | M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT)); |
22472 | R = SignBitSelect(VT, Amt, M, R); |
22473 | |
22474 | // a += a |
22475 | Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt); |
22476 | |
22477 | // return VSELECT(r, shift(r, 1), a); |
22478 | M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT)); |
22479 | R = SignBitSelect(VT, Amt, M, R); |
22480 | return R; |
22481 | } |
22482 | |
22483 | if (Op->getOpcode() == ISD::SRA) { |
22484 | // For SRA we need to unpack each byte to the higher byte of a i16 vector |
22485 | // so we can correctly sign extend. We don't care what happens to the |
22486 | // lower byte. |
22487 | SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt); |
22488 | SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt); |
22489 | SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R); |
22490 | SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R); |
22491 | ALo = DAG.getBitcast(ExtVT, ALo); |
22492 | AHi = DAG.getBitcast(ExtVT, AHi); |
22493 | RLo = DAG.getBitcast(ExtVT, RLo); |
22494 | RHi = DAG.getBitcast(ExtVT, RHi); |
22495 | |
22496 | // r = VSELECT(r, shift(r, 4), a); |
22497 | SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo, |
22498 | DAG.getConstant(4, dl, ExtVT)); |
22499 | SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi, |
22500 | DAG.getConstant(4, dl, ExtVT)); |
22501 | RLo = SignBitSelect(ExtVT, ALo, MLo, RLo); |
22502 | RHi = SignBitSelect(ExtVT, AHi, MHi, RHi); |
22503 | |
22504 | // a += a |
22505 | ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo); |
22506 | AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi); |
22507 | |
22508 | // r = VSELECT(r, shift(r, 2), a); |
22509 | MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo, |
22510 | DAG.getConstant(2, dl, ExtVT)); |
22511 | MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi, |
22512 | DAG.getConstant(2, dl, ExtVT)); |
22513 | RLo = SignBitSelect(ExtVT, ALo, MLo, RLo); |
22514 | RHi = SignBitSelect(ExtVT, AHi, MHi, RHi); |
22515 | |
22516 | // a += a |
22517 | ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo); |
22518 | AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi); |
22519 | |
22520 | // r = VSELECT(r, shift(r, 1), a); |
22521 | MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo, |
22522 | DAG.getConstant(1, dl, ExtVT)); |
22523 | MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi, |
22524 | DAG.getConstant(1, dl, ExtVT)); |
22525 | RLo = SignBitSelect(ExtVT, ALo, MLo, RLo); |
22526 | RHi = SignBitSelect(ExtVT, AHi, MHi, RHi); |
22527 | |
22528 | // Logical shift the result back to the lower byte, leaving a zero upper |
22529 | // byte |
22530 | // meaning that we can safely pack with PACKUSWB. |
22531 | RLo = |
22532 | DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT)); |
22533 | RHi = |
22534 | DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT)); |
22535 | return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi); |
22536 | } |
22537 | } |
22538 | |
22539 | if (Subtarget.hasInt256() && !Subtarget.hasXOP() && VT == MVT::v16i16) { |
22540 | MVT ExtVT = MVT::v8i32; |
22541 | SDValue Z = getZeroVector(VT, Subtarget, DAG, dl); |
22542 | SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z); |
22543 | SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z); |
22544 | SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Z, R); |
22545 | SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Z, R); |
22546 | ALo = DAG.getBitcast(ExtVT, ALo); |
22547 | AHi = DAG.getBitcast(ExtVT, AHi); |
22548 | RLo = DAG.getBitcast(ExtVT, RLo); |
22549 | RHi = DAG.getBitcast(ExtVT, RHi); |
22550 | SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo); |
22551 | SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi); |
22552 | Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT)); |
22553 | Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT)); |
22554 | return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi); |
22555 | } |
22556 | |
22557 | if (VT == MVT::v8i16) { |
22558 | unsigned ShiftOpcode = Op->getOpcode(); |
22559 | |
22560 | // If we have a constant shift amount, the non-SSE41 path is best as |
22561 | // avoiding bitcasts make it easier to constant fold and reduce to PBLENDW. |
22562 | bool UseSSE41 = Subtarget.hasSSE41() && |
22563 | !ISD::isBuildVectorOfConstantSDNodes(Amt.getNode()); |
22564 | |
22565 | auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) { |
22566 | // On SSE41 targets we make use of the fact that VSELECT lowers |
22567 | // to PBLENDVB which selects bytes based just on the sign bit. |
22568 | if (UseSSE41) { |
22569 | MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2); |
22570 | V0 = DAG.getBitcast(ExtVT, V0); |
22571 | V1 = DAG.getBitcast(ExtVT, V1); |
22572 | Sel = DAG.getBitcast(ExtVT, Sel); |
22573 | return DAG.getBitcast(VT, DAG.getSelect(dl, ExtVT, Sel, V0, V1)); |
22574 | } |
22575 | // On pre-SSE41 targets we splat the sign bit - a negative value will |
22576 | // set all bits of the lanes to true and VSELECT uses that in |
22577 | // its OR(AND(V0,C),AND(V1,~C)) lowering. |
22578 | SDValue C = |
22579 | DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT)); |
22580 | return DAG.getSelect(dl, VT, C, V0, V1); |
22581 | }; |
22582 | |
22583 | // Turn 'a' into a mask suitable for VSELECT: a = a << 12; |
22584 | if (UseSSE41) { |
22585 | // On SSE41 targets we need to replicate the shift mask in both |
22586 | // bytes for PBLENDVB. |
22587 | Amt = DAG.getNode( |
22588 | ISD::OR, dl, VT, |
22589 | DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)), |
22590 | DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT))); |
22591 | } else { |
22592 | Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)); |
22593 | } |
22594 | |
22595 | // r = VSELECT(r, shift(r, 8), a); |
22596 | SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT)); |
22597 | R = SignBitSelect(Amt, M, R); |
22598 | |
22599 | // a += a |
22600 | Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt); |
22601 | |
22602 | // r = VSELECT(r, shift(r, 4), a); |
22603 | M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT)); |
22604 | R = SignBitSelect(Amt, M, R); |
22605 | |
22606 | // a += a |
22607 | Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt); |
22608 | |
22609 | // r = VSELECT(r, shift(r, 2), a); |
22610 | M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT)); |
22611 | R = SignBitSelect(Amt, M, R); |
22612 | |
22613 | // a += a |
22614 | Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt); |
22615 | |
22616 | // return VSELECT(r, shift(r, 1), a); |
22617 | M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT)); |
22618 | R = SignBitSelect(Amt, M, R); |
22619 | return R; |
22620 | } |
22621 | |
22622 | // Decompose 256-bit shifts into smaller 128-bit shifts. |
22623 | if (VT.is256BitVector()) |
22624 | return Lower256IntArith(Op, DAG); |
22625 | |
22626 | return SDValue(); |
22627 | } |
22628 | |
22629 | static SDValue LowerRotate(SDValue Op, const X86Subtarget &Subtarget, |
22630 | SelectionDAG &DAG) { |
22631 | MVT VT = Op.getSimpleValueType(); |
22632 | SDLoc DL(Op); |
22633 | SDValue R = Op.getOperand(0); |
22634 | SDValue Amt = Op.getOperand(1); |
22635 | |
22636 | assert(VT.isVector() && "Custom lowering only for vector rotates!")((VT.isVector() && "Custom lowering only for vector rotates!" ) ? static_cast<void> (0) : __assert_fail ("VT.isVector() && \"Custom lowering only for vector rotates!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22636, __PRETTY_FUNCTION__)); |
22637 | assert(Subtarget.hasXOP() && "XOP support required for vector rotates!")((Subtarget.hasXOP() && "XOP support required for vector rotates!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasXOP() && \"XOP support required for vector rotates!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22637, __PRETTY_FUNCTION__)); |
22638 | assert((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported")(((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported" ) ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() == ISD::ROTL) && \"Only ROTL supported\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22638, __PRETTY_FUNCTION__)); |
22639 | |
22640 | // XOP has 128-bit vector variable + immediate rotates. |
22641 | // +ve/-ve Amt = rotate left/right. |
22642 | |
22643 | // Split 256-bit integers. |
22644 | if (VT.is256BitVector()) |
22645 | return Lower256IntArith(Op, DAG); |
22646 | |
22647 | assert(VT.is128BitVector() && "Only rotate 128-bit vectors!")((VT.is128BitVector() && "Only rotate 128-bit vectors!" ) ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"Only rotate 128-bit vectors!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22647, __PRETTY_FUNCTION__)); |
22648 | |
22649 | // Attempt to rotate by immediate. |
22650 | if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) { |
22651 | if (auto *RotateConst = BVAmt->getConstantSplatNode()) { |
22652 | uint64_t RotateAmt = RotateConst->getAPIntValue().getZExtValue(); |
22653 | assert(RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range")((RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range" ) ? static_cast<void> (0) : __assert_fail ("RotateAmt < VT.getScalarSizeInBits() && \"Rotation out of range\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22653, __PRETTY_FUNCTION__)); |
22654 | return DAG.getNode(X86ISD::VPROTI, DL, VT, R, |
22655 | DAG.getConstant(RotateAmt, DL, MVT::i8)); |
22656 | } |
22657 | } |
22658 | |
22659 | // Use general rotate by variable (per-element). |
22660 | return DAG.getNode(X86ISD::VPROT, DL, VT, R, Amt); |
22661 | } |
22662 | |
22663 | static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) { |
22664 | // Lower the "add/sub/mul with overflow" instruction into a regular ins plus |
22665 | // a "setcc" instruction that checks the overflow flag. The "brcond" lowering |
22666 | // looks for this combo and may remove the "setcc" instruction if the "setcc" |
22667 | // has only one use. |
22668 | SDNode *N = Op.getNode(); |
22669 | SDValue LHS = N->getOperand(0); |
22670 | SDValue RHS = N->getOperand(1); |
22671 | unsigned BaseOp = 0; |
22672 | X86::CondCode Cond; |
22673 | SDLoc DL(Op); |
22674 | switch (Op.getOpcode()) { |
22675 | default: llvm_unreachable("Unknown ovf instruction!")::llvm::llvm_unreachable_internal("Unknown ovf instruction!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22675); |
22676 | case ISD::SADDO: |
22677 | // A subtract of one will be selected as a INC. Note that INC doesn't |
22678 | // set CF, so we can't do this for UADDO. |
22679 | if (isOneConstant(RHS)) { |
22680 | BaseOp = X86ISD::INC; |
22681 | Cond = X86::COND_O; |
22682 | break; |
22683 | } |
22684 | BaseOp = X86ISD::ADD; |
22685 | Cond = X86::COND_O; |
22686 | break; |
22687 | case ISD::UADDO: |
22688 | BaseOp = X86ISD::ADD; |
22689 | Cond = X86::COND_B; |
22690 | break; |
22691 | case ISD::SSUBO: |
22692 | // A subtract of one will be selected as a DEC. Note that DEC doesn't |
22693 | // set CF, so we can't do this for USUBO. |
22694 | if (isOneConstant(RHS)) { |
22695 | BaseOp = X86ISD::DEC; |
22696 | Cond = X86::COND_O; |
22697 | break; |
22698 | } |
22699 | BaseOp = X86ISD::SUB; |
22700 | Cond = X86::COND_O; |
22701 | break; |
22702 | case ISD::USUBO: |
22703 | BaseOp = X86ISD::SUB; |
22704 | Cond = X86::COND_B; |
22705 | break; |
22706 | case ISD::SMULO: |
22707 | BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL; |
22708 | Cond = X86::COND_O; |
22709 | break; |
22710 | case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs |
22711 | if (N->getValueType(0) == MVT::i8) { |
22712 | BaseOp = X86ISD::UMUL8; |
22713 | Cond = X86::COND_O; |
22714 | break; |
22715 | } |
22716 | SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0), |
22717 | MVT::i32); |
22718 | SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS); |
22719 | |
22720 | SDValue SetCC = getSETCC(X86::COND_O, SDValue(Sum.getNode(), 2), DL, DAG); |
22721 | |
22722 | if (N->getValueType(1) == MVT::i1) |
22723 | SetCC = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, SetCC); |
22724 | |
22725 | return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC); |
22726 | } |
22727 | } |
22728 | |
22729 | // Also sets EFLAGS. |
22730 | SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32); |
22731 | SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS); |
22732 | |
22733 | SDValue SetCC = getSETCC(Cond, SDValue(Sum.getNode(), 1), DL, DAG); |
22734 | |
22735 | if (N->getValueType(1) == MVT::i1) |
22736 | SetCC = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, SetCC); |
22737 | |
22738 | return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC); |
22739 | } |
22740 | |
22741 | /// Returns true if the operand type is exactly twice the native width, and |
22742 | /// the corresponding cmpxchg8b or cmpxchg16b instruction is available. |
22743 | /// Used to know whether to use cmpxchg8/16b when expanding atomic operations |
22744 | /// (otherwise we leave them alone to become __sync_fetch_and_... calls). |
22745 | bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const { |
22746 | unsigned OpWidth = MemType->getPrimitiveSizeInBits(); |
22747 | |
22748 | if (OpWidth == 64) |
22749 | return !Subtarget.is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b |
22750 | else if (OpWidth == 128) |
22751 | return Subtarget.hasCmpxchg16b(); |
22752 | else |
22753 | return false; |
22754 | } |
22755 | |
22756 | bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const { |
22757 | return needsCmpXchgNb(SI->getValueOperand()->getType()); |
22758 | } |
22759 | |
22760 | // Note: this turns large loads into lock cmpxchg8b/16b. |
22761 | // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b. |
22762 | TargetLowering::AtomicExpansionKind |
22763 | X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const { |
22764 | auto PTy = cast<PointerType>(LI->getPointerOperandType()); |
22765 | return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg |
22766 | : AtomicExpansionKind::None; |
22767 | } |
22768 | |
22769 | TargetLowering::AtomicExpansionKind |
22770 | X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const { |
22771 | unsigned NativeWidth = Subtarget.is64Bit() ? 64 : 32; |
22772 | Type *MemType = AI->getType(); |
22773 | |
22774 | // If the operand is too big, we must see if cmpxchg8/16b is available |
22775 | // and default to library calls otherwise. |
22776 | if (MemType->getPrimitiveSizeInBits() > NativeWidth) { |
22777 | return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg |
22778 | : AtomicExpansionKind::None; |
22779 | } |
22780 | |
22781 | AtomicRMWInst::BinOp Op = AI->getOperation(); |
22782 | switch (Op) { |
22783 | default: |
22784 | llvm_unreachable("Unknown atomic operation")::llvm::llvm_unreachable_internal("Unknown atomic operation", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22784); |
22785 | case AtomicRMWInst::Xchg: |
22786 | case AtomicRMWInst::Add: |
22787 | case AtomicRMWInst::Sub: |
22788 | // It's better to use xadd, xsub or xchg for these in all cases. |
22789 | return AtomicExpansionKind::None; |
22790 | case AtomicRMWInst::Or: |
22791 | case AtomicRMWInst::And: |
22792 | case AtomicRMWInst::Xor: |
22793 | // If the atomicrmw's result isn't actually used, we can just add a "lock" |
22794 | // prefix to a normal instruction for these operations. |
22795 | return !AI->use_empty() ? AtomicExpansionKind::CmpXChg |
22796 | : AtomicExpansionKind::None; |
22797 | case AtomicRMWInst::Nand: |
22798 | case AtomicRMWInst::Max: |
22799 | case AtomicRMWInst::Min: |
22800 | case AtomicRMWInst::UMax: |
22801 | case AtomicRMWInst::UMin: |
22802 | // These always require a non-trivial set of data operations on x86. We must |
22803 | // use a cmpxchg loop. |
22804 | return AtomicExpansionKind::CmpXChg; |
22805 | } |
22806 | } |
22807 | |
22808 | LoadInst * |
22809 | X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const { |
22810 | unsigned NativeWidth = Subtarget.is64Bit() ? 64 : 32; |
22811 | Type *MemType = AI->getType(); |
22812 | // Accesses larger than the native width are turned into cmpxchg/libcalls, so |
22813 | // there is no benefit in turning such RMWs into loads, and it is actually |
22814 | // harmful as it introduces a mfence. |
22815 | if (MemType->getPrimitiveSizeInBits() > NativeWidth) |
22816 | return nullptr; |
22817 | |
22818 | auto Builder = IRBuilder<>(AI); |
22819 | Module *M = Builder.GetInsertBlock()->getParent()->getParent(); |
22820 | auto SynchScope = AI->getSynchScope(); |
22821 | // We must restrict the ordering to avoid generating loads with Release or |
22822 | // ReleaseAcquire orderings. |
22823 | auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering()); |
22824 | auto Ptr = AI->getPointerOperand(); |
22825 | |
22826 | // Before the load we need a fence. Here is an example lifted from |
22827 | // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence |
22828 | // is required: |
22829 | // Thread 0: |
22830 | // x.store(1, relaxed); |
22831 | // r1 = y.fetch_add(0, release); |
22832 | // Thread 1: |
22833 | // y.fetch_add(42, acquire); |
22834 | // r2 = x.load(relaxed); |
22835 | // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is |
22836 | // lowered to just a load without a fence. A mfence flushes the store buffer, |
22837 | // making the optimization clearly correct. |
22838 | // FIXME: it is required if isReleaseOrStronger(Order) but it is not clear |
22839 | // otherwise, we might be able to be more aggressive on relaxed idempotent |
22840 | // rmw. In practice, they do not look useful, so we don't try to be |
22841 | // especially clever. |
22842 | if (SynchScope == SingleThread) |
22843 | // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at |
22844 | // the IR level, so we must wrap it in an intrinsic. |
22845 | return nullptr; |
22846 | |
22847 | if (!Subtarget.hasMFence()) |
22848 | // FIXME: it might make sense to use a locked operation here but on a |
22849 | // different cache-line to prevent cache-line bouncing. In practice it |
22850 | // is probably a small win, and x86 processors without mfence are rare |
22851 | // enough that we do not bother. |
22852 | return nullptr; |
22853 | |
22854 | Function *MFence = |
22855 | llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence); |
22856 | Builder.CreateCall(MFence, {}); |
22857 | |
22858 | // Finally we can emit the atomic load. |
22859 | LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr, |
22860 | AI->getType()->getPrimitiveSizeInBits()); |
22861 | Loaded->setAtomic(Order, SynchScope); |
22862 | AI->replaceAllUsesWith(Loaded); |
22863 | AI->eraseFromParent(); |
22864 | return Loaded; |
22865 | } |
22866 | |
22867 | static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget &Subtarget, |
22868 | SelectionDAG &DAG) { |
22869 | SDLoc dl(Op); |
22870 | AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>( |
22871 | cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()); |
22872 | SynchronizationScope FenceScope = static_cast<SynchronizationScope>( |
22873 | cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue()); |
22874 | |
22875 | // The only fence that needs an instruction is a sequentially-consistent |
22876 | // cross-thread fence. |
22877 | if (FenceOrdering == AtomicOrdering::SequentiallyConsistent && |
22878 | FenceScope == CrossThread) { |
22879 | if (Subtarget.hasMFence()) |
22880 | return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0)); |
22881 | |
22882 | SDValue Chain = Op.getOperand(0); |
22883 | SDValue Zero = DAG.getConstant(0, dl, MVT::i32); |
22884 | SDValue Ops[] = { |
22885 | DAG.getRegister(X86::ESP, MVT::i32), // Base |
22886 | DAG.getTargetConstant(1, dl, MVT::i8), // Scale |
22887 | DAG.getRegister(0, MVT::i32), // Index |
22888 | DAG.getTargetConstant(0, dl, MVT::i32), // Disp |
22889 | DAG.getRegister(0, MVT::i32), // Segment. |
22890 | Zero, |
22891 | Chain |
22892 | }; |
22893 | SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops); |
22894 | return SDValue(Res, 0); |
22895 | } |
22896 | |
22897 | // MEMBARRIER is a compiler barrier; it codegens to a no-op. |
22898 | return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0)); |
22899 | } |
22900 | |
22901 | static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget &Subtarget, |
22902 | SelectionDAG &DAG) { |
22903 | MVT T = Op.getSimpleValueType(); |
22904 | SDLoc DL(Op); |
22905 | unsigned Reg = 0; |
22906 | unsigned size = 0; |
22907 | switch(T.SimpleTy) { |
22908 | default: llvm_unreachable("Invalid value type!")::llvm::llvm_unreachable_internal("Invalid value type!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22908); |
22909 | case MVT::i8: Reg = X86::AL; size = 1; break; |
22910 | case MVT::i16: Reg = X86::AX; size = 2; break; |
22911 | case MVT::i32: Reg = X86::EAX; size = 4; break; |
22912 | case MVT::i64: |
22913 | assert(Subtarget.is64Bit() && "Node not type legal!")((Subtarget.is64Bit() && "Node not type legal!") ? static_cast <void> (0) : __assert_fail ("Subtarget.is64Bit() && \"Node not type legal!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22913, __PRETTY_FUNCTION__)); |
22914 | Reg = X86::RAX; size = 8; |
22915 | break; |
22916 | } |
22917 | SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg, |
22918 | Op.getOperand(2), SDValue()); |
22919 | SDValue Ops[] = { cpIn.getValue(0), |
22920 | Op.getOperand(1), |
22921 | Op.getOperand(3), |
22922 | DAG.getTargetConstant(size, DL, MVT::i8), |
22923 | cpIn.getValue(1) }; |
22924 | SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue); |
22925 | MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand(); |
22926 | SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys, |
22927 | Ops, T, MMO); |
22928 | |
22929 | SDValue cpOut = |
22930 | DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1)); |
22931 | SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS, |
22932 | MVT::i32, cpOut.getValue(2)); |
22933 | SDValue Success = getSETCC(X86::COND_E, EFLAGS, DL, DAG); |
22934 | |
22935 | DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut); |
22936 | DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success); |
22937 | DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1)); |
22938 | return SDValue(); |
22939 | } |
22940 | |
22941 | static SDValue LowerBITCAST(SDValue Op, const X86Subtarget &Subtarget, |
22942 | SelectionDAG &DAG) { |
22943 | MVT SrcVT = Op.getOperand(0).getSimpleValueType(); |
22944 | MVT DstVT = Op.getSimpleValueType(); |
22945 | |
22946 | if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8 || |
22947 | SrcVT == MVT::i64) { |
22948 | assert(Subtarget.hasSSE2() && "Requires at least SSE2!")((Subtarget.hasSSE2() && "Requires at least SSE2!") ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSE2() && \"Requires at least SSE2!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22948, __PRETTY_FUNCTION__)); |
22949 | if (DstVT != MVT::f64) |
22950 | // This conversion needs to be expanded. |
22951 | return SDValue(); |
22952 | |
22953 | SDValue Op0 = Op->getOperand(0); |
22954 | SmallVector<SDValue, 16> Elts; |
22955 | SDLoc dl(Op); |
22956 | unsigned NumElts; |
22957 | MVT SVT; |
22958 | if (SrcVT.isVector()) { |
22959 | NumElts = SrcVT.getVectorNumElements(); |
22960 | SVT = SrcVT.getVectorElementType(); |
22961 | |
22962 | // Widen the vector in input in the case of MVT::v2i32. |
22963 | // Example: from MVT::v2i32 to MVT::v4i32. |
22964 | for (unsigned i = 0, e = NumElts; i != e; ++i) |
22965 | Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, Op0, |
22966 | DAG.getIntPtrConstant(i, dl))); |
22967 | } else { |
22968 | assert(SrcVT == MVT::i64 && !Subtarget.is64Bit() &&((SrcVT == MVT::i64 && !Subtarget.is64Bit() && "Unexpected source type in LowerBITCAST") ? static_cast<void > (0) : __assert_fail ("SrcVT == MVT::i64 && !Subtarget.is64Bit() && \"Unexpected source type in LowerBITCAST\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22969, __PRETTY_FUNCTION__)) |
22969 | "Unexpected source type in LowerBITCAST")((SrcVT == MVT::i64 && !Subtarget.is64Bit() && "Unexpected source type in LowerBITCAST") ? static_cast<void > (0) : __assert_fail ("SrcVT == MVT::i64 && !Subtarget.is64Bit() && \"Unexpected source type in LowerBITCAST\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22969, __PRETTY_FUNCTION__)); |
22970 | Elts.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op0, |
22971 | DAG.getIntPtrConstant(0, dl))); |
22972 | Elts.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op0, |
22973 | DAG.getIntPtrConstant(1, dl))); |
22974 | NumElts = 2; |
22975 | SVT = MVT::i32; |
22976 | } |
22977 | // Explicitly mark the extra elements as Undef. |
22978 | Elts.append(NumElts, DAG.getUNDEF(SVT)); |
22979 | |
22980 | EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2); |
22981 | SDValue BV = DAG.getBuildVector(NewVT, dl, Elts); |
22982 | SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV); |
22983 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64, |
22984 | DAG.getIntPtrConstant(0, dl)); |
22985 | } |
22986 | |
22987 | assert(Subtarget.is64Bit() && !Subtarget.hasSSE2() &&((Subtarget.is64Bit() && !Subtarget.hasSSE2() && Subtarget.hasMMX() && "Unexpected custom BITCAST") ? static_cast<void> (0) : __assert_fail ("Subtarget.is64Bit() && !Subtarget.hasSSE2() && Subtarget.hasMMX() && \"Unexpected custom BITCAST\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22988, __PRETTY_FUNCTION__)) |
22988 | Subtarget.hasMMX() && "Unexpected custom BITCAST")((Subtarget.is64Bit() && !Subtarget.hasSSE2() && Subtarget.hasMMX() && "Unexpected custom BITCAST") ? static_cast<void> (0) : __assert_fail ("Subtarget.is64Bit() && !Subtarget.hasSSE2() && Subtarget.hasMMX() && \"Unexpected custom BITCAST\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22988, __PRETTY_FUNCTION__)); |
22989 | assert((DstVT == MVT::i64 ||(((DstVT == MVT::i64 || (DstVT.isVector() && DstVT.getSizeInBits ()==64)) && "Unexpected custom BITCAST") ? static_cast <void> (0) : __assert_fail ("(DstVT == MVT::i64 || (DstVT.isVector() && DstVT.getSizeInBits()==64)) && \"Unexpected custom BITCAST\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22991, __PRETTY_FUNCTION__)) |
22990 | (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&(((DstVT == MVT::i64 || (DstVT.isVector() && DstVT.getSizeInBits ()==64)) && "Unexpected custom BITCAST") ? static_cast <void> (0) : __assert_fail ("(DstVT == MVT::i64 || (DstVT.isVector() && DstVT.getSizeInBits()==64)) && \"Unexpected custom BITCAST\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22991, __PRETTY_FUNCTION__)) |
22991 | "Unexpected custom BITCAST")(((DstVT == MVT::i64 || (DstVT.isVector() && DstVT.getSizeInBits ()==64)) && "Unexpected custom BITCAST") ? static_cast <void> (0) : __assert_fail ("(DstVT == MVT::i64 || (DstVT.isVector() && DstVT.getSizeInBits()==64)) && \"Unexpected custom BITCAST\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 22991, __PRETTY_FUNCTION__)); |
22992 | // i64 <=> MMX conversions are Legal. |
22993 | if (SrcVT==MVT::i64 && DstVT.isVector()) |
22994 | return Op; |
22995 | if (DstVT==MVT::i64 && SrcVT.isVector()) |
22996 | return Op; |
22997 | // MMX <=> MMX conversions are Legal. |
22998 | if (SrcVT.isVector() && DstVT.isVector()) |
22999 | return Op; |
23000 | // All other conversions need to be expanded. |
23001 | return SDValue(); |
23002 | } |
23003 | |
23004 | /// Compute the horizontal sum of bytes in V for the elements of VT. |
23005 | /// |
23006 | /// Requires V to be a byte vector and VT to be an integer vector type with |
23007 | /// wider elements than V's type. The width of the elements of VT determines |
23008 | /// how many bytes of V are summed horizontally to produce each element of the |
23009 | /// result. |
23010 | static SDValue LowerHorizontalByteSum(SDValue V, MVT VT, |
23011 | const X86Subtarget &Subtarget, |
23012 | SelectionDAG &DAG) { |
23013 | SDLoc DL(V); |
23014 | MVT ByteVecVT = V.getSimpleValueType(); |
23015 | MVT EltVT = VT.getVectorElementType(); |
23016 | assert(ByteVecVT.getVectorElementType() == MVT::i8 &&((ByteVecVT.getVectorElementType() == MVT::i8 && "Expected value to have byte element type." ) ? static_cast<void> (0) : __assert_fail ("ByteVecVT.getVectorElementType() == MVT::i8 && \"Expected value to have byte element type.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23017, __PRETTY_FUNCTION__)) |
23017 | "Expected value to have byte element type.")((ByteVecVT.getVectorElementType() == MVT::i8 && "Expected value to have byte element type." ) ? static_cast<void> (0) : __assert_fail ("ByteVecVT.getVectorElementType() == MVT::i8 && \"Expected value to have byte element type.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23017, __PRETTY_FUNCTION__)); |
23018 | assert(EltVT != MVT::i8 &&((EltVT != MVT::i8 && "Horizontal byte sum only makes sense for wider elements!" ) ? static_cast<void> (0) : __assert_fail ("EltVT != MVT::i8 && \"Horizontal byte sum only makes sense for wider elements!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23019, __PRETTY_FUNCTION__)) |
23019 | "Horizontal byte sum only makes sense for wider elements!")((EltVT != MVT::i8 && "Horizontal byte sum only makes sense for wider elements!" ) ? static_cast<void> (0) : __assert_fail ("EltVT != MVT::i8 && \"Horizontal byte sum only makes sense for wider elements!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23019, __PRETTY_FUNCTION__)); |
23020 | unsigned VecSize = VT.getSizeInBits(); |
23021 | assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!")((ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!" ) ? static_cast<void> (0) : __assert_fail ("ByteVecVT.getSizeInBits() == VecSize && \"Cannot change vector size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23021, __PRETTY_FUNCTION__)); |
23022 | |
23023 | // PSADBW instruction horizontally add all bytes and leave the result in i64 |
23024 | // chunks, thus directly computes the pop count for v2i64 and v4i64. |
23025 | if (EltVT == MVT::i64) { |
23026 | SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL); |
23027 | MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64); |
23028 | V = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT, V, Zeros); |
23029 | return DAG.getBitcast(VT, V); |
23030 | } |
23031 | |
23032 | if (EltVT == MVT::i32) { |
23033 | // We unpack the low half and high half into i32s interleaved with zeros so |
23034 | // that we can use PSADBW to horizontally sum them. The most useful part of |
23035 | // this is that it lines up the results of two PSADBW instructions to be |
23036 | // two v2i64 vectors which concatenated are the 4 population counts. We can |
23037 | // then use PACKUSWB to shrink and concatenate them into a v4i32 again. |
23038 | SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL); |
23039 | SDValue V32 = DAG.getBitcast(VT, V); |
23040 | SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V32, Zeros); |
23041 | SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V32, Zeros); |
23042 | |
23043 | // Do the horizontal sums into two v2i64s. |
23044 | Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL); |
23045 | MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64); |
23046 | Low = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT, |
23047 | DAG.getBitcast(ByteVecVT, Low), Zeros); |
23048 | High = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT, |
23049 | DAG.getBitcast(ByteVecVT, High), Zeros); |
23050 | |
23051 | // Merge them together. |
23052 | MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16); |
23053 | V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT, |
23054 | DAG.getBitcast(ShortVecVT, Low), |
23055 | DAG.getBitcast(ShortVecVT, High)); |
23056 | |
23057 | return DAG.getBitcast(VT, V); |
23058 | } |
23059 | |
23060 | // The only element type left is i16. |
23061 | assert(EltVT == MVT::i16 && "Unknown how to handle type")((EltVT == MVT::i16 && "Unknown how to handle type") ? static_cast<void> (0) : __assert_fail ("EltVT == MVT::i16 && \"Unknown how to handle type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23061, __PRETTY_FUNCTION__)); |
23062 | |
23063 | // To obtain pop count for each i16 element starting from the pop count for |
23064 | // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s |
23065 | // right by 8. It is important to shift as i16s as i8 vector shift isn't |
23066 | // directly supported. |
23067 | SDValue ShifterV = DAG.getConstant(8, DL, VT); |
23068 | SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), ShifterV); |
23069 | V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl), |
23070 | DAG.getBitcast(ByteVecVT, V)); |
23071 | return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), ShifterV); |
23072 | } |
23073 | |
23074 | static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, const SDLoc &DL, |
23075 | const X86Subtarget &Subtarget, |
23076 | SelectionDAG &DAG) { |
23077 | MVT VT = Op.getSimpleValueType(); |
23078 | MVT EltVT = VT.getVectorElementType(); |
23079 | unsigned VecSize = VT.getSizeInBits(); |
23080 | |
23081 | // Implement a lookup table in register by using an algorithm based on: |
23082 | // http://wm.ite.pl/articles/sse-popcount.html |
23083 | // |
23084 | // The general idea is that every lower byte nibble in the input vector is an |
23085 | // index into a in-register pre-computed pop count table. We then split up the |
23086 | // input vector in two new ones: (1) a vector with only the shifted-right |
23087 | // higher nibbles for each byte and (2) a vector with the lower nibbles (and |
23088 | // masked out higher ones) for each byte. PSHUFB is used separately with both |
23089 | // to index the in-register table. Next, both are added and the result is a |
23090 | // i8 vector where each element contains the pop count for input byte. |
23091 | // |
23092 | // To obtain the pop count for elements != i8, we follow up with the same |
23093 | // approach and use additional tricks as described below. |
23094 | // |
23095 | const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2, |
23096 | /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3, |
23097 | /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3, |
23098 | /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4}; |
23099 | |
23100 | int NumByteElts = VecSize / 8; |
23101 | MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts); |
23102 | SDValue In = DAG.getBitcast(ByteVecVT, Op); |
23103 | SmallVector<SDValue, 64> LUTVec; |
23104 | for (int i = 0; i < NumByteElts; ++i) |
23105 | LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8)); |
23106 | SDValue InRegLUT = DAG.getBuildVector(ByteVecVT, DL, LUTVec); |
23107 | SDValue M0F = DAG.getConstant(0x0F, DL, ByteVecVT); |
23108 | |
23109 | // High nibbles |
23110 | SDValue FourV = DAG.getConstant(4, DL, ByteVecVT); |
23111 | SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV); |
23112 | |
23113 | // Low nibbles |
23114 | SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F); |
23115 | |
23116 | // The input vector is used as the shuffle mask that index elements into the |
23117 | // LUT. After counting low and high nibbles, add the vector to obtain the |
23118 | // final pop count per i8 element. |
23119 | SDValue HighPopCnt = |
23120 | DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles); |
23121 | SDValue LowPopCnt = |
23122 | DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles); |
23123 | SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt); |
23124 | |
23125 | if (EltVT == MVT::i8) |
23126 | return PopCnt; |
23127 | |
23128 | return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG); |
23129 | } |
23130 | |
23131 | static SDValue LowerVectorCTPOPBitmath(SDValue Op, const SDLoc &DL, |
23132 | const X86Subtarget &Subtarget, |
23133 | SelectionDAG &DAG) { |
23134 | MVT VT = Op.getSimpleValueType(); |
23135 | assert(VT.is128BitVector() &&((VT.is128BitVector() && "Only 128-bit vector bitmath lowering supported." ) ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"Only 128-bit vector bitmath lowering supported.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23136, __PRETTY_FUNCTION__)) |
23136 | "Only 128-bit vector bitmath lowering supported.")((VT.is128BitVector() && "Only 128-bit vector bitmath lowering supported." ) ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"Only 128-bit vector bitmath lowering supported.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23136, __PRETTY_FUNCTION__)); |
23137 | |
23138 | int VecSize = VT.getSizeInBits(); |
23139 | MVT EltVT = VT.getVectorElementType(); |
23140 | int Len = EltVT.getSizeInBits(); |
23141 | |
23142 | // This is the vectorized version of the "best" algorithm from |
23143 | // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel |
23144 | // with a minor tweak to use a series of adds + shifts instead of vector |
23145 | // multiplications. Implemented for all integer vector types. We only use |
23146 | // this when we don't have SSSE3 which allows a LUT-based lowering that is |
23147 | // much faster, even faster than using native popcnt instructions. |
23148 | |
23149 | auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) { |
23150 | MVT VT = V.getSimpleValueType(); |
23151 | SDValue ShifterV = DAG.getConstant(Shifter, DL, VT); |
23152 | return DAG.getNode(OpCode, DL, VT, V, ShifterV); |
23153 | }; |
23154 | auto GetMask = [&](SDValue V, APInt Mask) { |
23155 | MVT VT = V.getSimpleValueType(); |
23156 | SDValue MaskV = DAG.getConstant(Mask, DL, VT); |
23157 | return DAG.getNode(ISD::AND, DL, VT, V, MaskV); |
23158 | }; |
23159 | |
23160 | // We don't want to incur the implicit masks required to SRL vNi8 vectors on |
23161 | // x86, so set the SRL type to have elements at least i16 wide. This is |
23162 | // correct because all of our SRLs are followed immediately by a mask anyways |
23163 | // that handles any bits that sneak into the high bits of the byte elements. |
23164 | MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16); |
23165 | |
23166 | SDValue V = Op; |
23167 | |
23168 | // v = v - ((v >> 1) & 0x55555555...) |
23169 | SDValue Srl = |
23170 | DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1)); |
23171 | SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55))); |
23172 | V = DAG.getNode(ISD::SUB, DL, VT, V, And); |
23173 | |
23174 | // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...) |
23175 | SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33))); |
23176 | Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2)); |
23177 | SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33))); |
23178 | V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS); |
23179 | |
23180 | // v = (v + (v >> 4)) & 0x0F0F0F0F... |
23181 | Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4)); |
23182 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl); |
23183 | V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F))); |
23184 | |
23185 | // At this point, V contains the byte-wise population count, and we are |
23186 | // merely doing a horizontal sum if necessary to get the wider element |
23187 | // counts. |
23188 | if (EltVT == MVT::i8) |
23189 | return V; |
23190 | |
23191 | return LowerHorizontalByteSum( |
23192 | DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget, |
23193 | DAG); |
23194 | } |
23195 | |
23196 | // Please ensure that any codegen change from LowerVectorCTPOP is reflected in |
23197 | // updated cost models in X86TTIImpl::getIntrinsicInstrCost. |
23198 | static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget &Subtarget, |
23199 | SelectionDAG &DAG) { |
23200 | MVT VT = Op.getSimpleValueType(); |
23201 | assert((VT.is512BitVector() || VT.is256BitVector() || VT.is128BitVector()) &&(((VT.is512BitVector() || VT.is256BitVector() || VT.is128BitVector ()) && "Unknown CTPOP type to handle") ? static_cast< void> (0) : __assert_fail ("(VT.is512BitVector() || VT.is256BitVector() || VT.is128BitVector()) && \"Unknown CTPOP type to handle\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23202, __PRETTY_FUNCTION__)) |
23202 | "Unknown CTPOP type to handle")(((VT.is512BitVector() || VT.is256BitVector() || VT.is128BitVector ()) && "Unknown CTPOP type to handle") ? static_cast< void> (0) : __assert_fail ("(VT.is512BitVector() || VT.is256BitVector() || VT.is128BitVector()) && \"Unknown CTPOP type to handle\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23202, __PRETTY_FUNCTION__)); |
23203 | SDLoc DL(Op.getNode()); |
23204 | SDValue Op0 = Op.getOperand(0); |
23205 | |
23206 | if (!Subtarget.hasSSSE3()) { |
23207 | // We can't use the fast LUT approach, so fall back on vectorized bitmath. |
23208 | assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!")((VT.is128BitVector() && "Only 128-bit vectors supported in SSE!" ) ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"Only 128-bit vectors supported in SSE!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23208, __PRETTY_FUNCTION__)); |
23209 | return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG); |
23210 | } |
23211 | |
23212 | // Decompose 256-bit ops into smaller 128-bit ops. |
23213 | if (VT.is256BitVector() && !Subtarget.hasInt256()) |
23214 | return Lower256IntUnary(Op, DAG); |
23215 | |
23216 | // Decompose 512-bit ops into smaller 256-bit ops. |
23217 | if (VT.is512BitVector() && !Subtarget.hasBWI()) |
23218 | return Lower512IntUnary(Op, DAG); |
23219 | |
23220 | return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG); |
23221 | } |
23222 | |
23223 | static SDValue LowerCTPOP(SDValue Op, const X86Subtarget &Subtarget, |
23224 | SelectionDAG &DAG) { |
23225 | assert(Op.getSimpleValueType().isVector() &&((Op.getSimpleValueType().isVector() && "We only do custom lowering for vector population count." ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().isVector() && \"We only do custom lowering for vector population count.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23226, __PRETTY_FUNCTION__)) |
23226 | "We only do custom lowering for vector population count.")((Op.getSimpleValueType().isVector() && "We only do custom lowering for vector population count." ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType().isVector() && \"We only do custom lowering for vector population count.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23226, __PRETTY_FUNCTION__)); |
23227 | return LowerVectorCTPOP(Op, Subtarget, DAG); |
23228 | } |
23229 | |
23230 | static SDValue LowerBITREVERSE_XOP(SDValue Op, SelectionDAG &DAG) { |
23231 | MVT VT = Op.getSimpleValueType(); |
23232 | SDValue In = Op.getOperand(0); |
23233 | SDLoc DL(Op); |
23234 | |
23235 | // For scalars, its still beneficial to transfer to/from the SIMD unit to |
23236 | // perform the BITREVERSE. |
23237 | if (!VT.isVector()) { |
23238 | MVT VecVT = MVT::getVectorVT(VT, 128 / VT.getSizeInBits()); |
23239 | SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, In); |
23240 | Res = DAG.getNode(ISD::BITREVERSE, DL, VecVT, Res); |
23241 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Res, |
23242 | DAG.getIntPtrConstant(0, DL)); |
23243 | } |
23244 | |
23245 | int NumElts = VT.getVectorNumElements(); |
23246 | int ScalarSizeInBytes = VT.getScalarSizeInBits() / 8; |
23247 | |
23248 | // Decompose 256-bit ops into smaller 128-bit ops. |
23249 | if (VT.is256BitVector()) |
23250 | return Lower256IntUnary(Op, DAG); |
23251 | |
23252 | assert(VT.is128BitVector() &&((VT.is128BitVector() && "Only 128-bit vector bitreverse lowering supported." ) ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"Only 128-bit vector bitreverse lowering supported.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23253, __PRETTY_FUNCTION__)) |
23253 | "Only 128-bit vector bitreverse lowering supported.")((VT.is128BitVector() && "Only 128-bit vector bitreverse lowering supported." ) ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"Only 128-bit vector bitreverse lowering supported.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23253, __PRETTY_FUNCTION__)); |
23254 | |
23255 | // VPPERM reverses the bits of a byte with the permute Op (2 << 5), and we |
23256 | // perform the BSWAP in the shuffle. |
23257 | // Its best to shuffle using the second operand as this will implicitly allow |
23258 | // memory folding for multiple vectors. |
23259 | SmallVector<SDValue, 16> MaskElts; |
23260 | for (int i = 0; i != NumElts; ++i) { |
23261 | for (int j = ScalarSizeInBytes - 1; j >= 0; --j) { |
23262 | int SourceByte = 16 + (i * ScalarSizeInBytes) + j; |
23263 | int PermuteByte = SourceByte | (2 << 5); |
23264 | MaskElts.push_back(DAG.getConstant(PermuteByte, DL, MVT::i8)); |
23265 | } |
23266 | } |
23267 | |
23268 | SDValue Mask = DAG.getBuildVector(MVT::v16i8, DL, MaskElts); |
23269 | SDValue Res = DAG.getBitcast(MVT::v16i8, In); |
23270 | Res = DAG.getNode(X86ISD::VPPERM, DL, MVT::v16i8, DAG.getUNDEF(MVT::v16i8), |
23271 | Res, Mask); |
23272 | return DAG.getBitcast(VT, Res); |
23273 | } |
23274 | |
23275 | static SDValue LowerBITREVERSE(SDValue Op, const X86Subtarget &Subtarget, |
23276 | SelectionDAG &DAG) { |
23277 | if (Subtarget.hasXOP()) |
23278 | return LowerBITREVERSE_XOP(Op, DAG); |
23279 | |
23280 | assert(Subtarget.hasSSSE3() && "SSSE3 required for BITREVERSE")((Subtarget.hasSSSE3() && "SSSE3 required for BITREVERSE" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSSE3() && \"SSSE3 required for BITREVERSE\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23280, __PRETTY_FUNCTION__)); |
23281 | |
23282 | MVT VT = Op.getSimpleValueType(); |
23283 | SDValue In = Op.getOperand(0); |
23284 | SDLoc DL(Op); |
23285 | |
23286 | unsigned NumElts = VT.getVectorNumElements(); |
23287 | assert(VT.getScalarType() == MVT::i8 &&((VT.getScalarType() == MVT::i8 && "Only byte vector BITREVERSE supported" ) ? static_cast<void> (0) : __assert_fail ("VT.getScalarType() == MVT::i8 && \"Only byte vector BITREVERSE supported\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23288, __PRETTY_FUNCTION__)) |
23288 | "Only byte vector BITREVERSE supported")((VT.getScalarType() == MVT::i8 && "Only byte vector BITREVERSE supported" ) ? static_cast<void> (0) : __assert_fail ("VT.getScalarType() == MVT::i8 && \"Only byte vector BITREVERSE supported\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23288, __PRETTY_FUNCTION__)); |
23289 | |
23290 | // Decompose 256-bit ops into smaller 128-bit ops on pre-AVX2. |
23291 | if (VT.is256BitVector() && !Subtarget.hasInt256()) |
23292 | return Lower256IntUnary(Op, DAG); |
23293 | |
23294 | // Perform BITREVERSE using PSHUFB lookups. Each byte is split into |
23295 | // two nibbles and a PSHUFB lookup to find the bitreverse of each |
23296 | // 0-15 value (moved to the other nibble). |
23297 | SDValue NibbleMask = DAG.getConstant(0xF, DL, VT); |
23298 | SDValue Lo = DAG.getNode(ISD::AND, DL, VT, In, NibbleMask); |
23299 | SDValue Hi = DAG.getNode(ISD::SRL, DL, VT, In, DAG.getConstant(4, DL, VT)); |
23300 | |
23301 | const int LoLUT[16] = { |
23302 | /* 0 */ 0x00, /* 1 */ 0x80, /* 2 */ 0x40, /* 3 */ 0xC0, |
23303 | /* 4 */ 0x20, /* 5 */ 0xA0, /* 6 */ 0x60, /* 7 */ 0xE0, |
23304 | /* 8 */ 0x10, /* 9 */ 0x90, /* a */ 0x50, /* b */ 0xD0, |
23305 | /* c */ 0x30, /* d */ 0xB0, /* e */ 0x70, /* f */ 0xF0}; |
23306 | const int HiLUT[16] = { |
23307 | /* 0 */ 0x00, /* 1 */ 0x08, /* 2 */ 0x04, /* 3 */ 0x0C, |
23308 | /* 4 */ 0x02, /* 5 */ 0x0A, /* 6 */ 0x06, /* 7 */ 0x0E, |
23309 | /* 8 */ 0x01, /* 9 */ 0x09, /* a */ 0x05, /* b */ 0x0D, |
23310 | /* c */ 0x03, /* d */ 0x0B, /* e */ 0x07, /* f */ 0x0F}; |
23311 | |
23312 | SmallVector<SDValue, 16> LoMaskElts, HiMaskElts; |
23313 | for (unsigned i = 0; i < NumElts; ++i) { |
23314 | LoMaskElts.push_back(DAG.getConstant(LoLUT[i % 16], DL, MVT::i8)); |
23315 | HiMaskElts.push_back(DAG.getConstant(HiLUT[i % 16], DL, MVT::i8)); |
23316 | } |
23317 | |
23318 | SDValue LoMask = DAG.getBuildVector(VT, DL, LoMaskElts); |
23319 | SDValue HiMask = DAG.getBuildVector(VT, DL, HiMaskElts); |
23320 | Lo = DAG.getNode(X86ISD::PSHUFB, DL, VT, LoMask, Lo); |
23321 | Hi = DAG.getNode(X86ISD::PSHUFB, DL, VT, HiMask, Hi); |
23322 | return DAG.getNode(ISD::OR, DL, VT, Lo, Hi); |
23323 | } |
23324 | |
23325 | static SDValue lowerAtomicArithWithLOCK(SDValue N, SelectionDAG &DAG) { |
23326 | unsigned NewOpc = 0; |
23327 | switch (N->getOpcode()) { |
23328 | case ISD::ATOMIC_LOAD_ADD: |
23329 | NewOpc = X86ISD::LADD; |
23330 | break; |
23331 | case ISD::ATOMIC_LOAD_SUB: |
23332 | NewOpc = X86ISD::LSUB; |
23333 | break; |
23334 | case ISD::ATOMIC_LOAD_OR: |
23335 | NewOpc = X86ISD::LOR; |
23336 | break; |
23337 | case ISD::ATOMIC_LOAD_XOR: |
23338 | NewOpc = X86ISD::LXOR; |
23339 | break; |
23340 | case ISD::ATOMIC_LOAD_AND: |
23341 | NewOpc = X86ISD::LAND; |
23342 | break; |
23343 | default: |
23344 | llvm_unreachable("Unknown ATOMIC_LOAD_ opcode")::llvm::llvm_unreachable_internal("Unknown ATOMIC_LOAD_ opcode" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23344); |
23345 | } |
23346 | |
23347 | MachineMemOperand *MMO = cast<MemSDNode>(N)->getMemOperand(); |
23348 | return DAG.getMemIntrinsicNode( |
23349 | NewOpc, SDLoc(N), DAG.getVTList(MVT::i32, MVT::Other), |
23350 | {N->getOperand(0), N->getOperand(1), N->getOperand(2)}, |
23351 | /*MemVT=*/N->getSimpleValueType(0), MMO); |
23352 | } |
23353 | |
23354 | /// Lower atomic_load_ops into LOCK-prefixed operations. |
23355 | static SDValue lowerAtomicArith(SDValue N, SelectionDAG &DAG, |
23356 | const X86Subtarget &Subtarget) { |
23357 | SDValue Chain = N->getOperand(0); |
23358 | SDValue LHS = N->getOperand(1); |
23359 | SDValue RHS = N->getOperand(2); |
23360 | unsigned Opc = N->getOpcode(); |
23361 | MVT VT = N->getSimpleValueType(0); |
23362 | SDLoc DL(N); |
23363 | |
23364 | // We can lower atomic_load_add into LXADD. However, any other atomicrmw op |
23365 | // can only be lowered when the result is unused. They should have already |
23366 | // been transformed into a cmpxchg loop in AtomicExpand. |
23367 | if (N->hasAnyUseOfValue(0)) { |
23368 | // Handle (atomic_load_sub p, v) as (atomic_load_add p, -v), to be able to |
23369 | // select LXADD if LOCK_SUB can't be selected. |
23370 | if (Opc == ISD::ATOMIC_LOAD_SUB) { |
23371 | AtomicSDNode *AN = cast<AtomicSDNode>(N.getNode()); |
23372 | RHS = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), RHS); |
23373 | return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, DL, VT, Chain, LHS, |
23374 | RHS, AN->getMemOperand()); |
23375 | } |
23376 | assert(Opc == ISD::ATOMIC_LOAD_ADD &&((Opc == ISD::ATOMIC_LOAD_ADD && "Used AtomicRMW ops other than Add should have been expanded!" ) ? static_cast<void> (0) : __assert_fail ("Opc == ISD::ATOMIC_LOAD_ADD && \"Used AtomicRMW ops other than Add should have been expanded!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23377, __PRETTY_FUNCTION__)) |
23377 | "Used AtomicRMW ops other than Add should have been expanded!")((Opc == ISD::ATOMIC_LOAD_ADD && "Used AtomicRMW ops other than Add should have been expanded!" ) ? static_cast<void> (0) : __assert_fail ("Opc == ISD::ATOMIC_LOAD_ADD && \"Used AtomicRMW ops other than Add should have been expanded!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23377, __PRETTY_FUNCTION__)); |
23378 | return N; |
23379 | } |
23380 | |
23381 | SDValue LockOp = lowerAtomicArithWithLOCK(N, DAG); |
23382 | // RAUW the chain, but don't worry about the result, as it's unused. |
23383 | assert(!N->hasAnyUseOfValue(0))((!N->hasAnyUseOfValue(0)) ? static_cast<void> (0) : __assert_fail ("!N->hasAnyUseOfValue(0)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23383, __PRETTY_FUNCTION__)); |
23384 | DAG.ReplaceAllUsesOfValueWith(N.getValue(1), LockOp.getValue(1)); |
23385 | return SDValue(); |
23386 | } |
23387 | |
23388 | static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) { |
23389 | SDNode *Node = Op.getNode(); |
23390 | SDLoc dl(Node); |
23391 | EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT(); |
23392 | |
23393 | // Convert seq_cst store -> xchg |
23394 | // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b) |
23395 | // FIXME: On 32-bit, store -> fist or movq would be more efficient |
23396 | // (The only way to get a 16-byte store is cmpxchg16b) |
23397 | // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment. |
23398 | if (cast<AtomicSDNode>(Node)->getOrdering() == |
23399 | AtomicOrdering::SequentiallyConsistent || |
23400 | !DAG.getTargetLoweringInfo().isTypeLegal(VT)) { |
23401 | SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl, |
23402 | cast<AtomicSDNode>(Node)->getMemoryVT(), |
23403 | Node->getOperand(0), |
23404 | Node->getOperand(1), Node->getOperand(2), |
23405 | cast<AtomicSDNode>(Node)->getMemOperand()); |
23406 | return Swap.getValue(1); |
23407 | } |
23408 | // Other atomic stores have a simple pattern. |
23409 | return Op; |
23410 | } |
23411 | |
23412 | static SDValue LowerADDSUBCARRY(SDValue Op, SelectionDAG &DAG) { |
23413 | SDNode *N = Op.getNode(); |
23414 | MVT VT = N->getSimpleValueType(0); |
23415 | |
23416 | // Let legalize expand this if it isn't a legal type yet. |
23417 | if (!DAG.getTargetLoweringInfo().isTypeLegal(VT)) |
23418 | return SDValue(); |
23419 | |
23420 | SDVTList VTs = DAG.getVTList(VT, MVT::i32); |
23421 | SDLoc DL(N); |
23422 | |
23423 | // Set the carry flag. |
23424 | SDValue Carry = Op.getOperand(2); |
23425 | EVT CarryVT = Carry.getValueType(); |
23426 | APInt NegOne = APInt::getAllOnesValue(CarryVT.getScalarSizeInBits()); |
23427 | Carry = DAG.getNode(X86ISD::ADD, DL, DAG.getVTList(CarryVT, MVT::i32), |
23428 | Carry, DAG.getConstant(NegOne, DL, CarryVT)); |
23429 | |
23430 | unsigned Opc = Op.getOpcode() == ISD::ADDCARRY ? X86ISD::ADC : X86ISD::SBB; |
23431 | SDValue Sum = DAG.getNode(Opc, DL, VTs, Op.getOperand(0), |
23432 | Op.getOperand(1), Carry.getValue(1)); |
23433 | |
23434 | SDValue SetCC = getSETCC(X86::COND_B, Sum.getValue(1), DL, DAG); |
23435 | if (N->getValueType(1) == MVT::i1) |
23436 | SetCC = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, SetCC); |
23437 | |
23438 | return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC); |
23439 | } |
23440 | |
23441 | static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget &Subtarget, |
23442 | SelectionDAG &DAG) { |
23443 | assert(Subtarget.isTargetDarwin() && Subtarget.is64Bit())((Subtarget.isTargetDarwin() && Subtarget.is64Bit()) ? static_cast<void> (0) : __assert_fail ("Subtarget.isTargetDarwin() && Subtarget.is64Bit()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23443, __PRETTY_FUNCTION__)); |
23444 | |
23445 | // For MacOSX, we want to call an alternative entry point: __sincos_stret, |
23446 | // which returns the values as { float, float } (in XMM0) or |
23447 | // { double, double } (which is returned in XMM0, XMM1). |
23448 | SDLoc dl(Op); |
23449 | SDValue Arg = Op.getOperand(0); |
23450 | EVT ArgVT = Arg.getValueType(); |
23451 | Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext()); |
23452 | |
23453 | TargetLowering::ArgListTy Args; |
23454 | TargetLowering::ArgListEntry Entry; |
23455 | |
23456 | Entry.Node = Arg; |
23457 | Entry.Ty = ArgTy; |
23458 | Entry.IsSExt = false; |
23459 | Entry.IsZExt = false; |
23460 | Args.push_back(Entry); |
23461 | |
23462 | bool isF64 = ArgVT == MVT::f64; |
23463 | // Only optimize x86_64 for now. i386 is a bit messy. For f32, |
23464 | // the small struct {f32, f32} is returned in (eax, edx). For f64, |
23465 | // the results are returned via SRet in memory. |
23466 | const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret"; |
23467 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
23468 | SDValue Callee = |
23469 | DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout())); |
23470 | |
23471 | Type *RetTy = isF64 ? (Type *)StructType::get(ArgTy, ArgTy) |
23472 | : (Type *)VectorType::get(ArgTy, 4); |
23473 | |
23474 | TargetLowering::CallLoweringInfo CLI(DAG); |
23475 | CLI.setDebugLoc(dl) |
23476 | .setChain(DAG.getEntryNode()) |
23477 | .setLibCallee(CallingConv::C, RetTy, Callee, std::move(Args)); |
23478 | |
23479 | std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI); |
23480 | |
23481 | if (isF64) |
23482 | // Returned in xmm0 and xmm1. |
23483 | return CallResult.first; |
23484 | |
23485 | // Returned in bits 0:31 and 32:64 xmm0. |
23486 | SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT, |
23487 | CallResult.first, DAG.getIntPtrConstant(0, dl)); |
23488 | SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT, |
23489 | CallResult.first, DAG.getIntPtrConstant(1, dl)); |
23490 | SDVTList Tys = DAG.getVTList(ArgVT, ArgVT); |
23491 | return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal); |
23492 | } |
23493 | |
23494 | /// Widen a vector input to a vector of NVT. The |
23495 | /// input vector must have the same element type as NVT. |
23496 | static SDValue ExtendToType(SDValue InOp, MVT NVT, SelectionDAG &DAG, |
23497 | bool FillWithZeroes = false) { |
23498 | // Check if InOp already has the right width. |
23499 | MVT InVT = InOp.getSimpleValueType(); |
23500 | if (InVT == NVT) |
23501 | return InOp; |
23502 | |
23503 | if (InOp.isUndef()) |
23504 | return DAG.getUNDEF(NVT); |
23505 | |
23506 | assert(InVT.getVectorElementType() == NVT.getVectorElementType() &&((InVT.getVectorElementType() == NVT.getVectorElementType() && "input and widen element type must match") ? static_cast< void> (0) : __assert_fail ("InVT.getVectorElementType() == NVT.getVectorElementType() && \"input and widen element type must match\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23507, __PRETTY_FUNCTION__)) |
23507 | "input and widen element type must match")((InVT.getVectorElementType() == NVT.getVectorElementType() && "input and widen element type must match") ? static_cast< void> (0) : __assert_fail ("InVT.getVectorElementType() == NVT.getVectorElementType() && \"input and widen element type must match\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23507, __PRETTY_FUNCTION__)); |
23508 | |
23509 | unsigned InNumElts = InVT.getVectorNumElements(); |
23510 | unsigned WidenNumElts = NVT.getVectorNumElements(); |
23511 | assert(WidenNumElts > InNumElts && WidenNumElts % InNumElts == 0 &&((WidenNumElts > InNumElts && WidenNumElts % InNumElts == 0 && "Unexpected request for vector widening") ? static_cast <void> (0) : __assert_fail ("WidenNumElts > InNumElts && WidenNumElts % InNumElts == 0 && \"Unexpected request for vector widening\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23512, __PRETTY_FUNCTION__)) |
23512 | "Unexpected request for vector widening")((WidenNumElts > InNumElts && WidenNumElts % InNumElts == 0 && "Unexpected request for vector widening") ? static_cast <void> (0) : __assert_fail ("WidenNumElts > InNumElts && WidenNumElts % InNumElts == 0 && \"Unexpected request for vector widening\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23512, __PRETTY_FUNCTION__)); |
23513 | |
23514 | SDLoc dl(InOp); |
23515 | if (InOp.getOpcode() == ISD::CONCAT_VECTORS && |
23516 | InOp.getNumOperands() == 2) { |
23517 | SDValue N1 = InOp.getOperand(1); |
23518 | if ((ISD::isBuildVectorAllZeros(N1.getNode()) && FillWithZeroes) || |
23519 | N1.isUndef()) { |
23520 | InOp = InOp.getOperand(0); |
23521 | InVT = InOp.getSimpleValueType(); |
23522 | InNumElts = InVT.getVectorNumElements(); |
23523 | } |
23524 | } |
23525 | if (ISD::isBuildVectorOfConstantSDNodes(InOp.getNode()) || |
23526 | ISD::isBuildVectorOfConstantFPSDNodes(InOp.getNode())) { |
23527 | SmallVector<SDValue, 16> Ops; |
23528 | for (unsigned i = 0; i < InNumElts; ++i) |
23529 | Ops.push_back(InOp.getOperand(i)); |
23530 | |
23531 | EVT EltVT = InOp.getOperand(0).getValueType(); |
23532 | |
23533 | SDValue FillVal = FillWithZeroes ? DAG.getConstant(0, dl, EltVT) : |
23534 | DAG.getUNDEF(EltVT); |
23535 | for (unsigned i = 0; i < WidenNumElts - InNumElts; ++i) |
23536 | Ops.push_back(FillVal); |
23537 | return DAG.getBuildVector(NVT, dl, Ops); |
23538 | } |
23539 | SDValue FillVal = FillWithZeroes ? DAG.getConstant(0, dl, NVT) : |
23540 | DAG.getUNDEF(NVT); |
23541 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NVT, FillVal, |
23542 | InOp, DAG.getIntPtrConstant(0, dl)); |
23543 | } |
23544 | |
23545 | static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget &Subtarget, |
23546 | SelectionDAG &DAG) { |
23547 | assert(Subtarget.hasAVX512() &&((Subtarget.hasAVX512() && "MGATHER/MSCATTER are supported on AVX-512 arch only" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX512() && \"MGATHER/MSCATTER are supported on AVX-512 arch only\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23548, __PRETTY_FUNCTION__)) |
23548 | "MGATHER/MSCATTER are supported on AVX-512 arch only")((Subtarget.hasAVX512() && "MGATHER/MSCATTER are supported on AVX-512 arch only" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX512() && \"MGATHER/MSCATTER are supported on AVX-512 arch only\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23548, __PRETTY_FUNCTION__)); |
23549 | |
23550 | // X86 scatter kills mask register, so its type should be added to |
23551 | // the list of return values. |
23552 | // If the "scatter" has 2 return values, it is already handled. |
23553 | if (Op.getNode()->getNumValues() == 2) |
23554 | return Op; |
23555 | |
23556 | MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode()); |
23557 | SDValue Src = N->getValue(); |
23558 | MVT VT = Src.getSimpleValueType(); |
23559 | assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op")((VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op" ) ? static_cast<void> (0) : __assert_fail ("VT.getScalarSizeInBits() >= 32 && \"Unsupported scatter op\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23559, __PRETTY_FUNCTION__)); |
23560 | SDLoc dl(Op); |
23561 | |
23562 | SDValue NewScatter; |
23563 | SDValue Index = N->getIndex(); |
23564 | SDValue Mask = N->getMask(); |
23565 | SDValue Chain = N->getChain(); |
23566 | SDValue BasePtr = N->getBasePtr(); |
23567 | MVT MemVT = N->getMemoryVT().getSimpleVT(); |
23568 | MVT IndexVT = Index.getSimpleValueType(); |
23569 | MVT MaskVT = Mask.getSimpleValueType(); |
23570 | |
23571 | if (MemVT.getScalarSizeInBits() < VT.getScalarSizeInBits()) { |
23572 | // The v2i32 value was promoted to v2i64. |
23573 | // Now we "redo" the type legalizer's work and widen the original |
23574 | // v2i32 value to v4i32. The original v2i32 is retrieved from v2i64 |
23575 | // with a shuffle. |
23576 | assert((MemVT == MVT::v2i32 && VT == MVT::v2i64) &&(((MemVT == MVT::v2i32 && VT == MVT::v2i64) && "Unexpected memory type") ? static_cast<void> (0) : __assert_fail ("(MemVT == MVT::v2i32 && VT == MVT::v2i64) && \"Unexpected memory type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23577, __PRETTY_FUNCTION__)) |
23577 | "Unexpected memory type")(((MemVT == MVT::v2i32 && VT == MVT::v2i64) && "Unexpected memory type") ? static_cast<void> (0) : __assert_fail ("(MemVT == MVT::v2i32 && VT == MVT::v2i64) && \"Unexpected memory type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23577, __PRETTY_FUNCTION__)); |
23578 | int ShuffleMask[] = {0, 2, -1, -1}; |
23579 | Src = DAG.getVectorShuffle(MVT::v4i32, dl, DAG.getBitcast(MVT::v4i32, Src), |
23580 | DAG.getUNDEF(MVT::v4i32), ShuffleMask); |
23581 | // Now we have 4 elements instead of 2. |
23582 | // Expand the index. |
23583 | MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), 4); |
23584 | Index = ExtendToType(Index, NewIndexVT, DAG); |
23585 | |
23586 | // Expand the mask with zeroes |
23587 | // Mask may be <2 x i64> or <2 x i1> at this moment |
23588 | assert((MaskVT == MVT::v2i1 || MaskVT == MVT::v2i64) &&(((MaskVT == MVT::v2i1 || MaskVT == MVT::v2i64) && "Unexpected mask type" ) ? static_cast<void> (0) : __assert_fail ("(MaskVT == MVT::v2i1 || MaskVT == MVT::v2i64) && \"Unexpected mask type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23589, __PRETTY_FUNCTION__)) |
23589 | "Unexpected mask type")(((MaskVT == MVT::v2i1 || MaskVT == MVT::v2i64) && "Unexpected mask type" ) ? static_cast<void> (0) : __assert_fail ("(MaskVT == MVT::v2i1 || MaskVT == MVT::v2i64) && \"Unexpected mask type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23589, __PRETTY_FUNCTION__)); |
23590 | MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), 4); |
23591 | Mask = ExtendToType(Mask, ExtMaskVT, DAG, true); |
23592 | VT = MVT::v4i32; |
23593 | } |
23594 | |
23595 | unsigned NumElts = VT.getVectorNumElements(); |
23596 | if (!Subtarget.hasVLX() && !VT.is512BitVector() && |
23597 | !Index.getSimpleValueType().is512BitVector()) { |
23598 | // AVX512F supports only 512-bit vectors. Or data or index should |
23599 | // be 512 bit wide. If now the both index and data are 256-bit, but |
23600 | // the vector contains 8 elements, we just sign-extend the index |
23601 | if (IndexVT == MVT::v8i32) |
23602 | // Just extend index |
23603 | Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index); |
23604 | else { |
23605 | // The minimal number of elts in scatter is 8 |
23606 | NumElts = 8; |
23607 | // Index |
23608 | MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), NumElts); |
23609 | // Use original index here, do not modify the index twice |
23610 | Index = ExtendToType(N->getIndex(), NewIndexVT, DAG); |
23611 | if (IndexVT.getScalarType() == MVT::i32) |
23612 | Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index); |
23613 | |
23614 | // Mask |
23615 | // At this point we have promoted mask operand |
23616 | assert(MaskVT.getScalarSizeInBits() >= 32 && "unexpected mask type")((MaskVT.getScalarSizeInBits() >= 32 && "unexpected mask type" ) ? static_cast<void> (0) : __assert_fail ("MaskVT.getScalarSizeInBits() >= 32 && \"unexpected mask type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23616, __PRETTY_FUNCTION__)); |
23617 | MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), NumElts); |
23618 | // Use the original mask here, do not modify the mask twice |
23619 | Mask = ExtendToType(N->getMask(), ExtMaskVT, DAG, true); |
23620 | |
23621 | // The value that should be stored |
23622 | MVT NewVT = MVT::getVectorVT(VT.getScalarType(), NumElts); |
23623 | Src = ExtendToType(Src, NewVT, DAG); |
23624 | } |
23625 | } |
23626 | // If the mask is "wide" at this point - truncate it to i1 vector |
23627 | MVT BitMaskVT = MVT::getVectorVT(MVT::i1, NumElts); |
23628 | Mask = DAG.getNode(ISD::TRUNCATE, dl, BitMaskVT, Mask); |
23629 | |
23630 | // The mask is killed by scatter, add it to the values |
23631 | SDVTList VTs = DAG.getVTList(BitMaskVT, MVT::Other); |
23632 | SDValue Ops[] = {Chain, Src, Mask, BasePtr, Index}; |
23633 | NewScatter = DAG.getMaskedScatter(VTs, N->getMemoryVT(), dl, Ops, |
23634 | N->getMemOperand()); |
23635 | DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1)); |
23636 | return SDValue(NewScatter.getNode(), 1); |
23637 | } |
23638 | |
23639 | static SDValue LowerMLOAD(SDValue Op, const X86Subtarget &Subtarget, |
23640 | SelectionDAG &DAG) { |
23641 | |
23642 | MaskedLoadSDNode *N = cast<MaskedLoadSDNode>(Op.getNode()); |
23643 | MVT VT = Op.getSimpleValueType(); |
23644 | MVT ScalarVT = VT.getScalarType(); |
23645 | SDValue Mask = N->getMask(); |
23646 | SDLoc dl(Op); |
23647 | |
23648 | assert((!N->isExpandingLoad() || Subtarget.hasAVX512()) &&(((!N->isExpandingLoad() || Subtarget.hasAVX512()) && "Expanding masked load is supported on AVX-512 target only!" ) ? static_cast<void> (0) : __assert_fail ("(!N->isExpandingLoad() || Subtarget.hasAVX512()) && \"Expanding masked load is supported on AVX-512 target only!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23649, __PRETTY_FUNCTION__)) |
23649 | "Expanding masked load is supported on AVX-512 target only!")(((!N->isExpandingLoad() || Subtarget.hasAVX512()) && "Expanding masked load is supported on AVX-512 target only!" ) ? static_cast<void> (0) : __assert_fail ("(!N->isExpandingLoad() || Subtarget.hasAVX512()) && \"Expanding masked load is supported on AVX-512 target only!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23649, __PRETTY_FUNCTION__)); |
23650 | |
23651 | assert((!N->isExpandingLoad() || ScalarVT.getSizeInBits() >= 32) &&(((!N->isExpandingLoad() || ScalarVT.getSizeInBits() >= 32) && "Expanding masked load is supported for 32 and 64-bit types only!" ) ? static_cast<void> (0) : __assert_fail ("(!N->isExpandingLoad() || ScalarVT.getSizeInBits() >= 32) && \"Expanding masked load is supported for 32 and 64-bit types only!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23652, __PRETTY_FUNCTION__)) |
23652 | "Expanding masked load is supported for 32 and 64-bit types only!")(((!N->isExpandingLoad() || ScalarVT.getSizeInBits() >= 32) && "Expanding masked load is supported for 32 and 64-bit types only!" ) ? static_cast<void> (0) : __assert_fail ("(!N->isExpandingLoad() || ScalarVT.getSizeInBits() >= 32) && \"Expanding masked load is supported for 32 and 64-bit types only!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23652, __PRETTY_FUNCTION__)); |
23653 | |
23654 | // 4x32, 4x64 and 2x64 vectors of non-expanding loads are legal regardless of |
23655 | // VLX. These types for exp-loads are handled here. |
23656 | if (!N->isExpandingLoad() && VT.getVectorNumElements() <= 4) |
23657 | return Op; |
23658 | |
23659 | assert(Subtarget.hasAVX512() && !Subtarget.hasVLX() && !VT.is512BitVector() &&((Subtarget.hasAVX512() && !Subtarget.hasVLX() && !VT.is512BitVector() && "Cannot lower masked load op." ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX512() && !Subtarget.hasVLX() && !VT.is512BitVector() && \"Cannot lower masked load op.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23660, __PRETTY_FUNCTION__)) |
23660 | "Cannot lower masked load op.")((Subtarget.hasAVX512() && !Subtarget.hasVLX() && !VT.is512BitVector() && "Cannot lower masked load op." ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX512() && !Subtarget.hasVLX() && !VT.is512BitVector() && \"Cannot lower masked load op.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23660, __PRETTY_FUNCTION__)); |
23661 | |
23662 | assert((ScalarVT.getSizeInBits() >= 32 ||(((ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && "Unsupported masked load op." ) ? static_cast<void> (0) : __assert_fail ("(ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && \"Unsupported masked load op.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23665, __PRETTY_FUNCTION__)) |
23663 | (Subtarget.hasBWI() &&(((ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && "Unsupported masked load op." ) ? static_cast<void> (0) : __assert_fail ("(ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && \"Unsupported masked load op.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23665, __PRETTY_FUNCTION__)) |
23664 | (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) &&(((ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && "Unsupported masked load op." ) ? static_cast<void> (0) : __assert_fail ("(ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && \"Unsupported masked load op.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23665, __PRETTY_FUNCTION__)) |
23665 | "Unsupported masked load op.")(((ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && "Unsupported masked load op." ) ? static_cast<void> (0) : __assert_fail ("(ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && \"Unsupported masked load op.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23665, __PRETTY_FUNCTION__)); |
23666 | |
23667 | // This operation is legal for targets with VLX, but without |
23668 | // VLX the vector should be widened to 512 bit |
23669 | unsigned NumEltsInWideVec = 512 / VT.getScalarSizeInBits(); |
23670 | MVT WideDataVT = MVT::getVectorVT(ScalarVT, NumEltsInWideVec); |
23671 | SDValue Src0 = N->getSrc0(); |
23672 | Src0 = ExtendToType(Src0, WideDataVT, DAG); |
23673 | |
23674 | // Mask element has to be i1. |
23675 | MVT MaskEltTy = Mask.getSimpleValueType().getScalarType(); |
23676 | assert((MaskEltTy == MVT::i1 || VT.getVectorNumElements() <= 4) &&(((MaskEltTy == MVT::i1 || VT.getVectorNumElements() <= 4) && "We handle 4x32, 4x64 and 2x64 vectors only in this case" ) ? static_cast<void> (0) : __assert_fail ("(MaskEltTy == MVT::i1 || VT.getVectorNumElements() <= 4) && \"We handle 4x32, 4x64 and 2x64 vectors only in this case\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23677, __PRETTY_FUNCTION__)) |
23677 | "We handle 4x32, 4x64 and 2x64 vectors only in this case")(((MaskEltTy == MVT::i1 || VT.getVectorNumElements() <= 4) && "We handle 4x32, 4x64 and 2x64 vectors only in this case" ) ? static_cast<void> (0) : __assert_fail ("(MaskEltTy == MVT::i1 || VT.getVectorNumElements() <= 4) && \"We handle 4x32, 4x64 and 2x64 vectors only in this case\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23677, __PRETTY_FUNCTION__)); |
23678 | |
23679 | MVT WideMaskVT = MVT::getVectorVT(MaskEltTy, NumEltsInWideVec); |
23680 | |
23681 | Mask = ExtendToType(Mask, WideMaskVT, DAG, true); |
23682 | if (MaskEltTy != MVT::i1) |
23683 | Mask = DAG.getNode(ISD::TRUNCATE, dl, |
23684 | MVT::getVectorVT(MVT::i1, NumEltsInWideVec), Mask); |
23685 | SDValue NewLoad = DAG.getMaskedLoad(WideDataVT, dl, N->getChain(), |
23686 | N->getBasePtr(), Mask, Src0, |
23687 | N->getMemoryVT(), N->getMemOperand(), |
23688 | N->getExtensionType(), |
23689 | N->isExpandingLoad()); |
23690 | |
23691 | SDValue Exract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, |
23692 | NewLoad.getValue(0), |
23693 | DAG.getIntPtrConstant(0, dl)); |
23694 | SDValue RetOps[] = {Exract, NewLoad.getValue(1)}; |
23695 | return DAG.getMergeValues(RetOps, dl); |
23696 | } |
23697 | |
23698 | static SDValue LowerMSTORE(SDValue Op, const X86Subtarget &Subtarget, |
23699 | SelectionDAG &DAG) { |
23700 | MaskedStoreSDNode *N = cast<MaskedStoreSDNode>(Op.getNode()); |
23701 | SDValue DataToStore = N->getValue(); |
23702 | MVT VT = DataToStore.getSimpleValueType(); |
23703 | MVT ScalarVT = VT.getScalarType(); |
23704 | SDValue Mask = N->getMask(); |
23705 | SDLoc dl(Op); |
23706 | |
23707 | assert((!N->isCompressingStore() || Subtarget.hasAVX512()) &&(((!N->isCompressingStore() || Subtarget.hasAVX512()) && "Expanding masked load is supported on AVX-512 target only!" ) ? static_cast<void> (0) : __assert_fail ("(!N->isCompressingStore() || Subtarget.hasAVX512()) && \"Expanding masked load is supported on AVX-512 target only!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23708, __PRETTY_FUNCTION__)) |
23708 | "Expanding masked load is supported on AVX-512 target only!")(((!N->isCompressingStore() || Subtarget.hasAVX512()) && "Expanding masked load is supported on AVX-512 target only!" ) ? static_cast<void> (0) : __assert_fail ("(!N->isCompressingStore() || Subtarget.hasAVX512()) && \"Expanding masked load is supported on AVX-512 target only!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23708, __PRETTY_FUNCTION__)); |
23709 | |
23710 | assert((!N->isCompressingStore() || ScalarVT.getSizeInBits() >= 32) &&(((!N->isCompressingStore() || ScalarVT.getSizeInBits() >= 32) && "Expanding masked load is supported for 32 and 64-bit types only!" ) ? static_cast<void> (0) : __assert_fail ("(!N->isCompressingStore() || ScalarVT.getSizeInBits() >= 32) && \"Expanding masked load is supported for 32 and 64-bit types only!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23711, __PRETTY_FUNCTION__)) |
23711 | "Expanding masked load is supported for 32 and 64-bit types only!")(((!N->isCompressingStore() || ScalarVT.getSizeInBits() >= 32) && "Expanding masked load is supported for 32 and 64-bit types only!" ) ? static_cast<void> (0) : __assert_fail ("(!N->isCompressingStore() || ScalarVT.getSizeInBits() >= 32) && \"Expanding masked load is supported for 32 and 64-bit types only!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23711, __PRETTY_FUNCTION__)); |
23712 | |
23713 | // 4x32 and 2x64 vectors of non-compressing stores are legal regardless to VLX. |
23714 | if (!N->isCompressingStore() && VT.getVectorNumElements() <= 4) |
23715 | return Op; |
23716 | |
23717 | assert(Subtarget.hasAVX512() && !Subtarget.hasVLX() && !VT.is512BitVector() &&((Subtarget.hasAVX512() && !Subtarget.hasVLX() && !VT.is512BitVector() && "Cannot lower masked store op." ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX512() && !Subtarget.hasVLX() && !VT.is512BitVector() && \"Cannot lower masked store op.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23718, __PRETTY_FUNCTION__)) |
23718 | "Cannot lower masked store op.")((Subtarget.hasAVX512() && !Subtarget.hasVLX() && !VT.is512BitVector() && "Cannot lower masked store op." ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX512() && !Subtarget.hasVLX() && !VT.is512BitVector() && \"Cannot lower masked store op.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23718, __PRETTY_FUNCTION__)); |
23719 | |
23720 | assert((ScalarVT.getSizeInBits() >= 32 ||(((ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && "Unsupported masked store op." ) ? static_cast<void> (0) : __assert_fail ("(ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && \"Unsupported masked store op.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23723, __PRETTY_FUNCTION__)) |
23721 | (Subtarget.hasBWI() &&(((ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && "Unsupported masked store op." ) ? static_cast<void> (0) : __assert_fail ("(ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && \"Unsupported masked store op.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23723, __PRETTY_FUNCTION__)) |
23722 | (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) &&(((ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && "Unsupported masked store op." ) ? static_cast<void> (0) : __assert_fail ("(ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && \"Unsupported masked store op.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23723, __PRETTY_FUNCTION__)) |
23723 | "Unsupported masked store op.")(((ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && "Unsupported masked store op." ) ? static_cast<void> (0) : __assert_fail ("(ScalarVT.getSizeInBits() >= 32 || (Subtarget.hasBWI() && (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) && \"Unsupported masked store op.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23723, __PRETTY_FUNCTION__)); |
23724 | |
23725 | // This operation is legal for targets with VLX, but without |
23726 | // VLX the vector should be widened to 512 bit |
23727 | unsigned NumEltsInWideVec = 512/VT.getScalarSizeInBits(); |
23728 | MVT WideDataVT = MVT::getVectorVT(ScalarVT, NumEltsInWideVec); |
23729 | |
23730 | // Mask element has to be i1. |
23731 | MVT MaskEltTy = Mask.getSimpleValueType().getScalarType(); |
23732 | assert((MaskEltTy == MVT::i1 || VT.getVectorNumElements() <= 4) &&(((MaskEltTy == MVT::i1 || VT.getVectorNumElements() <= 4) && "We handle 4x32, 4x64 and 2x64 vectors only in this case" ) ? static_cast<void> (0) : __assert_fail ("(MaskEltTy == MVT::i1 || VT.getVectorNumElements() <= 4) && \"We handle 4x32, 4x64 and 2x64 vectors only in this case\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23733, __PRETTY_FUNCTION__)) |
23733 | "We handle 4x32, 4x64 and 2x64 vectors only in this case")(((MaskEltTy == MVT::i1 || VT.getVectorNumElements() <= 4) && "We handle 4x32, 4x64 and 2x64 vectors only in this case" ) ? static_cast<void> (0) : __assert_fail ("(MaskEltTy == MVT::i1 || VT.getVectorNumElements() <= 4) && \"We handle 4x32, 4x64 and 2x64 vectors only in this case\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23733, __PRETTY_FUNCTION__)); |
23734 | |
23735 | MVT WideMaskVT = MVT::getVectorVT(MaskEltTy, NumEltsInWideVec); |
23736 | |
23737 | DataToStore = ExtendToType(DataToStore, WideDataVT, DAG); |
23738 | Mask = ExtendToType(Mask, WideMaskVT, DAG, true); |
23739 | if (MaskEltTy != MVT::i1) |
23740 | Mask = DAG.getNode(ISD::TRUNCATE, dl, |
23741 | MVT::getVectorVT(MVT::i1, NumEltsInWideVec), Mask); |
23742 | return DAG.getMaskedStore(N->getChain(), dl, DataToStore, N->getBasePtr(), |
23743 | Mask, N->getMemoryVT(), N->getMemOperand(), |
23744 | N->isTruncatingStore(), N->isCompressingStore()); |
23745 | } |
23746 | |
23747 | static SDValue LowerMGATHER(SDValue Op, const X86Subtarget &Subtarget, |
23748 | SelectionDAG &DAG) { |
23749 | assert(Subtarget.hasAVX512() &&((Subtarget.hasAVX512() && "MGATHER/MSCATTER are supported on AVX-512 arch only" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX512() && \"MGATHER/MSCATTER are supported on AVX-512 arch only\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23750, __PRETTY_FUNCTION__)) |
23750 | "MGATHER/MSCATTER are supported on AVX-512 arch only")((Subtarget.hasAVX512() && "MGATHER/MSCATTER are supported on AVX-512 arch only" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX512() && \"MGATHER/MSCATTER are supported on AVX-512 arch only\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23750, __PRETTY_FUNCTION__)); |
23751 | |
23752 | MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode()); |
23753 | SDLoc dl(Op); |
23754 | MVT VT = Op.getSimpleValueType(); |
23755 | SDValue Index = N->getIndex(); |
23756 | SDValue Mask = N->getMask(); |
23757 | SDValue Src0 = N->getValue(); |
23758 | MVT IndexVT = Index.getSimpleValueType(); |
23759 | MVT MaskVT = Mask.getSimpleValueType(); |
23760 | |
23761 | unsigned NumElts = VT.getVectorNumElements(); |
23762 | assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op")((VT.getScalarSizeInBits() >= 32 && "Unsupported gather op" ) ? static_cast<void> (0) : __assert_fail ("VT.getScalarSizeInBits() >= 32 && \"Unsupported gather op\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23762, __PRETTY_FUNCTION__)); |
23763 | |
23764 | if (!Subtarget.hasVLX() && !VT.is512BitVector() && |
23765 | !Index.getSimpleValueType().is512BitVector()) { |
23766 | // AVX512F supports only 512-bit vectors. Or data or index should |
23767 | // be 512 bit wide. If now the both index and data are 256-bit, but |
23768 | // the vector contains 8 elements, we just sign-extend the index |
23769 | if (NumElts == 8) { |
23770 | Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index); |
23771 | SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2), |
23772 | N->getOperand(3), Index }; |
23773 | DAG.UpdateNodeOperands(N, Ops); |
23774 | return Op; |
23775 | } |
23776 | |
23777 | // Minimal number of elements in Gather |
23778 | NumElts = 8; |
23779 | // Index |
23780 | MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), NumElts); |
23781 | Index = ExtendToType(Index, NewIndexVT, DAG); |
23782 | if (IndexVT.getScalarType() == MVT::i32) |
23783 | Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index); |
23784 | |
23785 | // Mask |
23786 | MVT MaskBitVT = MVT::getVectorVT(MVT::i1, NumElts); |
23787 | // At this point we have promoted mask operand |
23788 | assert(MaskVT.getScalarSizeInBits() >= 32 && "unexpected mask type")((MaskVT.getScalarSizeInBits() >= 32 && "unexpected mask type" ) ? static_cast<void> (0) : __assert_fail ("MaskVT.getScalarSizeInBits() >= 32 && \"unexpected mask type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23788, __PRETTY_FUNCTION__)); |
23789 | MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), NumElts); |
23790 | Mask = ExtendToType(Mask, ExtMaskVT, DAG, true); |
23791 | Mask = DAG.getNode(ISD::TRUNCATE, dl, MaskBitVT, Mask); |
23792 | |
23793 | // The pass-through value |
23794 | MVT NewVT = MVT::getVectorVT(VT.getScalarType(), NumElts); |
23795 | Src0 = ExtendToType(Src0, NewVT, DAG); |
23796 | |
23797 | SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index }; |
23798 | SDValue NewGather = DAG.getMaskedGather(DAG.getVTList(NewVT, MVT::Other), |
23799 | N->getMemoryVT(), dl, Ops, |
23800 | N->getMemOperand()); |
23801 | SDValue Exract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, |
23802 | NewGather.getValue(0), |
23803 | DAG.getIntPtrConstant(0, dl)); |
23804 | SDValue RetOps[] = {Exract, NewGather.getValue(1)}; |
23805 | return DAG.getMergeValues(RetOps, dl); |
23806 | } |
23807 | if (N->getMemoryVT() == MVT::v2i32 && Subtarget.hasVLX()) { |
23808 | // There is a special case when the return type is v2i32 is illegal and |
23809 | // the type legaizer extended it to v2i64. Without this conversion we end up |
23810 | // with VPGATHERQQ (reading q-words from the memory) instead of VPGATHERQD. |
23811 | // In order to avoid this situation, we'll build an X86 specific Gather node |
23812 | // with index v2i64 and value type v4i32. |
23813 | assert(VT == MVT::v2i64 && Src0.getValueType() == MVT::v2i64 &&((VT == MVT::v2i64 && Src0.getValueType() == MVT::v2i64 && "Unexpected type in masked gather") ? static_cast <void> (0) : __assert_fail ("VT == MVT::v2i64 && Src0.getValueType() == MVT::v2i64 && \"Unexpected type in masked gather\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23814, __PRETTY_FUNCTION__)) |
23814 | "Unexpected type in masked gather")((VT == MVT::v2i64 && Src0.getValueType() == MVT::v2i64 && "Unexpected type in masked gather") ? static_cast <void> (0) : __assert_fail ("VT == MVT::v2i64 && Src0.getValueType() == MVT::v2i64 && \"Unexpected type in masked gather\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23814, __PRETTY_FUNCTION__)); |
23815 | Src0 = DAG.getVectorShuffle(MVT::v4i32, dl, |
23816 | DAG.getBitcast(MVT::v4i32, Src0), |
23817 | DAG.getUNDEF(MVT::v4i32), { 0, 2, -1, -1 }); |
23818 | // The mask should match the destination type. Extending mask with zeroes |
23819 | // is not necessary since instruction itself reads only two values from |
23820 | // memory. |
23821 | Mask = ExtendToType(Mask, MVT::v4i1, DAG, false); |
23822 | SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index }; |
23823 | SDValue NewGather = DAG.getTargetMemSDNode<X86MaskedGatherSDNode>( |
23824 | DAG.getVTList(MVT::v4i32, MVT::Other), Ops, dl, N->getMemoryVT(), |
23825 | N->getMemOperand()); |
23826 | |
23827 | SDValue Sext = getExtendInVec(X86ISD::VSEXT, dl, MVT::v2i64, |
23828 | NewGather.getValue(0), DAG); |
23829 | SDValue RetOps[] = { Sext, NewGather.getValue(1) }; |
23830 | return DAG.getMergeValues(RetOps, dl); |
23831 | } |
23832 | if (N->getMemoryVT() == MVT::v2f32 && Subtarget.hasVLX()) { |
23833 | // This transformation is for optimization only. |
23834 | // The type legalizer extended mask and index to 4 elements vector |
23835 | // in order to match requirements of the common gather node - same |
23836 | // vector width of index and value. X86 Gather node allows mismatch |
23837 | // of vector width in order to select more optimal instruction at the |
23838 | // end. |
23839 | assert(VT == MVT::v4f32 && Src0.getValueType() == MVT::v4f32 &&((VT == MVT::v4f32 && Src0.getValueType() == MVT::v4f32 && "Unexpected type in masked gather") ? static_cast <void> (0) : __assert_fail ("VT == MVT::v4f32 && Src0.getValueType() == MVT::v4f32 && \"Unexpected type in masked gather\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23840, __PRETTY_FUNCTION__)) |
23840 | "Unexpected type in masked gather")((VT == MVT::v4f32 && Src0.getValueType() == MVT::v4f32 && "Unexpected type in masked gather") ? static_cast <void> (0) : __assert_fail ("VT == MVT::v4f32 && Src0.getValueType() == MVT::v4f32 && \"Unexpected type in masked gather\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23840, __PRETTY_FUNCTION__)); |
23841 | if (Mask.getOpcode() == ISD::CONCAT_VECTORS && |
23842 | ISD::isBuildVectorAllZeros(Mask.getOperand(1).getNode()) && |
23843 | Index.getOpcode() == ISD::CONCAT_VECTORS && |
23844 | Index.getOperand(1).isUndef()) { |
23845 | Mask = ExtendToType(Mask.getOperand(0), MVT::v4i1, DAG, false); |
23846 | Index = Index.getOperand(0); |
23847 | } else |
23848 | return Op; |
23849 | SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index }; |
23850 | SDValue NewGather = DAG.getTargetMemSDNode<X86MaskedGatherSDNode>( |
23851 | DAG.getVTList(MVT::v4f32, MVT::Other), Ops, dl, N->getMemoryVT(), |
23852 | N->getMemOperand()); |
23853 | |
23854 | SDValue RetOps[] = { NewGather.getValue(0), NewGather.getValue(1) }; |
23855 | return DAG.getMergeValues(RetOps, dl); |
23856 | |
23857 | } |
23858 | return Op; |
23859 | } |
23860 | |
23861 | SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op, |
23862 | SelectionDAG &DAG) const { |
23863 | // TODO: Eventually, the lowering of these nodes should be informed by or |
23864 | // deferred to the GC strategy for the function in which they appear. For |
23865 | // now, however, they must be lowered to something. Since they are logically |
23866 | // no-ops in the case of a null GC strategy (or a GC strategy which does not |
23867 | // require special handling for these nodes), lower them as literal NOOPs for |
23868 | // the time being. |
23869 | SmallVector<SDValue, 2> Ops; |
23870 | |
23871 | Ops.push_back(Op.getOperand(0)); |
23872 | if (Op->getGluedNode()) |
23873 | Ops.push_back(Op->getOperand(Op->getNumOperands() - 1)); |
23874 | |
23875 | SDLoc OpDL(Op); |
23876 | SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue); |
23877 | SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0); |
23878 | |
23879 | return NOOP; |
23880 | } |
23881 | |
23882 | SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op, |
23883 | SelectionDAG &DAG) const { |
23884 | // TODO: Eventually, the lowering of these nodes should be informed by or |
23885 | // deferred to the GC strategy for the function in which they appear. For |
23886 | // now, however, they must be lowered to something. Since they are logically |
23887 | // no-ops in the case of a null GC strategy (or a GC strategy which does not |
23888 | // require special handling for these nodes), lower them as literal NOOPs for |
23889 | // the time being. |
23890 | SmallVector<SDValue, 2> Ops; |
23891 | |
23892 | Ops.push_back(Op.getOperand(0)); |
23893 | if (Op->getGluedNode()) |
23894 | Ops.push_back(Op->getOperand(Op->getNumOperands() - 1)); |
23895 | |
23896 | SDLoc OpDL(Op); |
23897 | SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue); |
23898 | SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0); |
23899 | |
23900 | return NOOP; |
23901 | } |
23902 | |
23903 | /// Provide custom lowering hooks for some operations. |
23904 | SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { |
23905 | switch (Op.getOpcode()) { |
23906 | default: llvm_unreachable("Should not custom lower this!")::llvm::llvm_unreachable_internal("Should not custom lower this!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 23906); |
23907 | case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG); |
23908 | case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: |
23909 | return LowerCMP_SWAP(Op, Subtarget, DAG); |
23910 | case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG); |
23911 | case ISD::ATOMIC_LOAD_ADD: |
23912 | case ISD::ATOMIC_LOAD_SUB: |
23913 | case ISD::ATOMIC_LOAD_OR: |
23914 | case ISD::ATOMIC_LOAD_XOR: |
23915 | case ISD::ATOMIC_LOAD_AND: return lowerAtomicArith(Op, DAG, Subtarget); |
23916 | case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op, DAG); |
23917 | case ISD::BITREVERSE: return LowerBITREVERSE(Op, Subtarget, DAG); |
23918 | case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG); |
23919 | case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG); |
23920 | case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG); |
23921 | case ISD::VSELECT: return LowerVSELECT(Op, DAG); |
23922 | case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG); |
23923 | case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG); |
23924 | case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG); |
23925 | case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG); |
23926 | case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, Subtarget,DAG); |
23927 | case ISD::ConstantPool: return LowerConstantPool(Op, DAG); |
23928 | case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG); |
23929 | case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); |
23930 | case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG); |
23931 | case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); |
23932 | case ISD::SHL_PARTS: |
23933 | case ISD::SRA_PARTS: |
23934 | case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG); |
23935 | case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG); |
23936 | case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG); |
23937 | case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG); |
23938 | case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG); |
23939 | case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG); |
23940 | case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG); |
23941 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
23942 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
23943 | return LowerEXTEND_VECTOR_INREG(Op, Subtarget, DAG); |
23944 | case ISD::FP_TO_SINT: |
23945 | case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG); |
23946 | case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG); |
23947 | case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG); |
23948 | case ISD::FABS: |
23949 | case ISD::FNEG: return LowerFABSorFNEG(Op, DAG); |
23950 | case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG); |
23951 | case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG); |
23952 | case ISD::SETCC: return LowerSETCC(Op, DAG); |
23953 | case ISD::SETCCCARRY: return LowerSETCCCARRY(Op, DAG); |
23954 | case ISD::SELECT: return LowerSELECT(Op, DAG); |
23955 | case ISD::BRCOND: return LowerBRCOND(Op, DAG); |
23956 | case ISD::JumpTable: return LowerJumpTable(Op, DAG); |
23957 | case ISD::VASTART: return LowerVASTART(Op, DAG); |
23958 | case ISD::VAARG: return LowerVAARG(Op, DAG); |
23959 | case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG); |
23960 | case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG); |
23961 | case ISD::INTRINSIC_VOID: |
23962 | case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG); |
23963 | case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG); |
23964 | case ISD::ADDROFRETURNADDR: return LowerADDROFRETURNADDR(Op, DAG); |
23965 | case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); |
23966 | case ISD::FRAME_TO_ARGS_OFFSET: |
23967 | return LowerFRAME_TO_ARGS_OFFSET(Op, DAG); |
23968 | case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG); |
23969 | case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG); |
23970 | case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG); |
23971 | case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG); |
23972 | case ISD::EH_SJLJ_SETUP_DISPATCH: |
23973 | return lowerEH_SJLJ_SETUP_DISPATCH(Op, DAG); |
23974 | case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG); |
23975 | case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG); |
23976 | case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG); |
23977 | case ISD::CTLZ: |
23978 | case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ(Op, Subtarget, DAG); |
23979 | case ISD::CTTZ: |
23980 | case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG); |
23981 | case ISD::MUL: return LowerMUL(Op, Subtarget, DAG); |
23982 | case ISD::MULHS: |
23983 | case ISD::MULHU: return LowerMULH(Op, Subtarget, DAG); |
23984 | case ISD::UMUL_LOHI: |
23985 | case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG); |
23986 | case ISD::ROTL: return LowerRotate(Op, Subtarget, DAG); |
23987 | case ISD::SRA: |
23988 | case ISD::SRL: |
23989 | case ISD::SHL: return LowerShift(Op, Subtarget, DAG); |
23990 | case ISD::SADDO: |
23991 | case ISD::UADDO: |
23992 | case ISD::SSUBO: |
23993 | case ISD::USUBO: |
23994 | case ISD::SMULO: |
23995 | case ISD::UMULO: return LowerXALUO(Op, DAG); |
23996 | case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG); |
23997 | case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG); |
23998 | case ISD::ADDCARRY: |
23999 | case ISD::SUBCARRY: return LowerADDSUBCARRY(Op, DAG); |
24000 | case ISD::ADD: |
24001 | case ISD::SUB: return LowerADD_SUB(Op, DAG); |
24002 | case ISD::SMAX: |
24003 | case ISD::SMIN: |
24004 | case ISD::UMAX: |
24005 | case ISD::UMIN: return LowerMINMAX(Op, DAG); |
24006 | case ISD::ABS: return LowerABS(Op, DAG); |
24007 | case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG); |
24008 | case ISD::MLOAD: return LowerMLOAD(Op, Subtarget, DAG); |
24009 | case ISD::MSTORE: return LowerMSTORE(Op, Subtarget, DAG); |
24010 | case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG); |
24011 | case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG); |
24012 | case ISD::GC_TRANSITION_START: |
24013 | return LowerGC_TRANSITION_START(Op, DAG); |
24014 | case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG); |
24015 | case ISD::STORE: return LowerTruncatingStore(Op, Subtarget, DAG); |
24016 | } |
24017 | } |
24018 | |
24019 | /// Places new result values for the node in Results (their number |
24020 | /// and types must exactly match those of the original return values of |
24021 | /// the node), or leaves Results empty, which indicates that the node is not |
24022 | /// to be custom lowered after all. |
24023 | void X86TargetLowering::LowerOperationWrapper(SDNode *N, |
24024 | SmallVectorImpl<SDValue> &Results, |
24025 | SelectionDAG &DAG) const { |
24026 | SDValue Res = LowerOperation(SDValue(N, 0), DAG); |
24027 | |
24028 | if (!Res.getNode()) |
24029 | return; |
24030 | |
24031 | assert((N->getNumValues() <= Res->getNumValues()) &&(((N->getNumValues() <= Res->getNumValues()) && "Lowering returned the wrong number of results!") ? static_cast <void> (0) : __assert_fail ("(N->getNumValues() <= Res->getNumValues()) && \"Lowering returned the wrong number of results!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24032, __PRETTY_FUNCTION__)) |
24032 | "Lowering returned the wrong number of results!")(((N->getNumValues() <= Res->getNumValues()) && "Lowering returned the wrong number of results!") ? static_cast <void> (0) : __assert_fail ("(N->getNumValues() <= Res->getNumValues()) && \"Lowering returned the wrong number of results!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24032, __PRETTY_FUNCTION__)); |
24033 | |
24034 | // Places new result values base on N result number. |
24035 | // In some cases (LowerSINT_TO_FP for example) Res has more result values |
24036 | // than original node, chain should be dropped(last value). |
24037 | for (unsigned I = 0, E = N->getNumValues(); I != E; ++I) |
24038 | Results.push_back(Res.getValue(I)); |
24039 | } |
24040 | |
24041 | /// Replace a node with an illegal result type with a new node built out of |
24042 | /// custom code. |
24043 | void X86TargetLowering::ReplaceNodeResults(SDNode *N, |
24044 | SmallVectorImpl<SDValue>&Results, |
24045 | SelectionDAG &DAG) const { |
24046 | SDLoc dl(N); |
24047 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
24048 | switch (N->getOpcode()) { |
24049 | default: |
24050 | llvm_unreachable("Do not know how to custom type legalize this operation!")::llvm::llvm_unreachable_internal("Do not know how to custom type legalize this operation!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24050); |
24051 | case X86ISD::AVG: { |
24052 | // Legalize types for X86ISD::AVG by expanding vectors. |
24053 | assert(Subtarget.hasSSE2() && "Requires at least SSE2!")((Subtarget.hasSSE2() && "Requires at least SSE2!") ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSE2() && \"Requires at least SSE2!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24053, __PRETTY_FUNCTION__)); |
24054 | |
24055 | auto InVT = N->getValueType(0); |
24056 | auto InVTSize = InVT.getSizeInBits(); |
24057 | const unsigned RegSize = |
24058 | (InVTSize > 128) ? ((InVTSize > 256) ? 512 : 256) : 128; |
24059 | assert((Subtarget.hasBWI() || RegSize < 512) &&(((Subtarget.hasBWI() || RegSize < 512) && "512-bit vector requires AVX512BW" ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasBWI() || RegSize < 512) && \"512-bit vector requires AVX512BW\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24060, __PRETTY_FUNCTION__)) |
24060 | "512-bit vector requires AVX512BW")(((Subtarget.hasBWI() || RegSize < 512) && "512-bit vector requires AVX512BW" ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasBWI() || RegSize < 512) && \"512-bit vector requires AVX512BW\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24060, __PRETTY_FUNCTION__)); |
24061 | assert((Subtarget.hasAVX2() || RegSize < 256) &&(((Subtarget.hasAVX2() || RegSize < 256) && "256-bit vector requires AVX2" ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasAVX2() || RegSize < 256) && \"256-bit vector requires AVX2\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24062, __PRETTY_FUNCTION__)) |
24062 | "256-bit vector requires AVX2")(((Subtarget.hasAVX2() || RegSize < 256) && "256-bit vector requires AVX2" ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.hasAVX2() || RegSize < 256) && \"256-bit vector requires AVX2\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24062, __PRETTY_FUNCTION__)); |
24063 | |
24064 | auto ElemVT = InVT.getVectorElementType(); |
24065 | auto RegVT = EVT::getVectorVT(*DAG.getContext(), ElemVT, |
24066 | RegSize / ElemVT.getSizeInBits()); |
24067 | assert(RegSize % InVT.getSizeInBits() == 0)((RegSize % InVT.getSizeInBits() == 0) ? static_cast<void> (0) : __assert_fail ("RegSize % InVT.getSizeInBits() == 0", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24067, __PRETTY_FUNCTION__)); |
24068 | unsigned NumConcat = RegSize / InVT.getSizeInBits(); |
24069 | |
24070 | SmallVector<SDValue, 16> Ops(NumConcat, DAG.getUNDEF(InVT)); |
24071 | Ops[0] = N->getOperand(0); |
24072 | SDValue InVec0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, RegVT, Ops); |
24073 | Ops[0] = N->getOperand(1); |
24074 | SDValue InVec1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, RegVT, Ops); |
24075 | |
24076 | SDValue Res = DAG.getNode(X86ISD::AVG, dl, RegVT, InVec0, InVec1); |
24077 | Results.push_back(DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InVT, Res, |
24078 | DAG.getIntPtrConstant(0, dl))); |
24079 | return; |
24080 | } |
24081 | // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32. |
24082 | case X86ISD::FMINC: |
24083 | case X86ISD::FMIN: |
24084 | case X86ISD::FMAXC: |
24085 | case X86ISD::FMAX: { |
24086 | EVT VT = N->getValueType(0); |
24087 | assert(VT == MVT::v2f32 && "Unexpected type (!= v2f32) on FMIN/FMAX.")((VT == MVT::v2f32 && "Unexpected type (!= v2f32) on FMIN/FMAX." ) ? static_cast<void> (0) : __assert_fail ("VT == MVT::v2f32 && \"Unexpected type (!= v2f32) on FMIN/FMAX.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24087, __PRETTY_FUNCTION__)); |
24088 | SDValue UNDEF = DAG.getUNDEF(VT); |
24089 | SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32, |
24090 | N->getOperand(0), UNDEF); |
24091 | SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32, |
24092 | N->getOperand(1), UNDEF); |
24093 | Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS)); |
24094 | return; |
24095 | } |
24096 | case ISD::SDIV: |
24097 | case ISD::UDIV: |
24098 | case ISD::SREM: |
24099 | case ISD::UREM: |
24100 | case ISD::SDIVREM: |
24101 | case ISD::UDIVREM: { |
24102 | SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG); |
24103 | Results.push_back(V); |
24104 | return; |
24105 | } |
24106 | case ISD::FP_TO_SINT: |
24107 | case ISD::FP_TO_UINT: { |
24108 | bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT; |
24109 | |
24110 | if (N->getValueType(0) == MVT::v2i32) { |
24111 | assert((IsSigned || Subtarget.hasAVX512()) &&(((IsSigned || Subtarget.hasAVX512()) && "Can only handle signed conversion without AVX512" ) ? static_cast<void> (0) : __assert_fail ("(IsSigned || Subtarget.hasAVX512()) && \"Can only handle signed conversion without AVX512\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24112, __PRETTY_FUNCTION__)) |
24112 | "Can only handle signed conversion without AVX512")(((IsSigned || Subtarget.hasAVX512()) && "Can only handle signed conversion without AVX512" ) ? static_cast<void> (0) : __assert_fail ("(IsSigned || Subtarget.hasAVX512()) && \"Can only handle signed conversion without AVX512\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24112, __PRETTY_FUNCTION__)); |
24113 | assert(Subtarget.hasSSE2() && "Requires at least SSE2!")((Subtarget.hasSSE2() && "Requires at least SSE2!") ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSE2() && \"Requires at least SSE2!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24113, __PRETTY_FUNCTION__)); |
24114 | SDValue Src = N->getOperand(0); |
24115 | if (Src.getValueType() == MVT::v2f64) { |
24116 | SDValue Idx = DAG.getIntPtrConstant(0, dl); |
24117 | SDValue Res = DAG.getNode(IsSigned ? X86ISD::CVTTP2SI |
24118 | : X86ISD::CVTTP2UI, |
24119 | dl, MVT::v4i32, Src); |
24120 | Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v2i32, Res, Idx); |
24121 | Results.push_back(Res); |
24122 | return; |
24123 | } |
24124 | if (Src.getValueType() == MVT::v2f32) { |
24125 | SDValue Idx = DAG.getIntPtrConstant(0, dl); |
24126 | SDValue Res = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32, Src, |
24127 | DAG.getUNDEF(MVT::v2f32)); |
24128 | Res = DAG.getNode(IsSigned ? ISD::FP_TO_SINT |
24129 | : ISD::FP_TO_UINT, dl, MVT::v4i32, Res); |
24130 | Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v2i32, Res, Idx); |
24131 | Results.push_back(Res); |
24132 | return; |
24133 | } |
24134 | |
24135 | // The FP_TO_INTHelper below only handles f32/f64/f80 scalar inputs, |
24136 | // so early out here. |
24137 | return; |
24138 | } |
24139 | |
24140 | std::pair<SDValue,SDValue> Vals = |
24141 | FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true); |
24142 | SDValue FIST = Vals.first, StackSlot = Vals.second; |
24143 | if (FIST.getNode()) { |
24144 | EVT VT = N->getValueType(0); |
24145 | // Return a load from the stack slot. |
24146 | if (StackSlot.getNode()) |
24147 | Results.push_back( |
24148 | DAG.getLoad(VT, dl, FIST, StackSlot, MachinePointerInfo())); |
24149 | else |
24150 | Results.push_back(FIST); |
24151 | } |
24152 | return; |
24153 | } |
24154 | case ISD::SINT_TO_FP: { |
24155 | assert(Subtarget.hasDQI() && Subtarget.hasVLX() && "Requires AVX512DQVL!")((Subtarget.hasDQI() && Subtarget.hasVLX() && "Requires AVX512DQVL!") ? static_cast<void> (0) : __assert_fail ("Subtarget.hasDQI() && Subtarget.hasVLX() && \"Requires AVX512DQVL!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24155, __PRETTY_FUNCTION__)); |
24156 | SDValue Src = N->getOperand(0); |
24157 | if (N->getValueType(0) != MVT::v2f32 || Src.getValueType() != MVT::v2i64) |
24158 | return; |
24159 | Results.push_back(DAG.getNode(X86ISD::CVTSI2P, dl, MVT::v4f32, Src)); |
24160 | return; |
24161 | } |
24162 | case ISD::UINT_TO_FP: { |
24163 | assert(Subtarget.hasSSE2() && "Requires at least SSE2!")((Subtarget.hasSSE2() && "Requires at least SSE2!") ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSE2() && \"Requires at least SSE2!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24163, __PRETTY_FUNCTION__)); |
24164 | EVT VT = N->getValueType(0); |
24165 | if (VT != MVT::v2f32) |
24166 | return; |
24167 | SDValue Src = N->getOperand(0); |
24168 | EVT SrcVT = Src.getValueType(); |
24169 | if (Subtarget.hasDQI() && Subtarget.hasVLX() && SrcVT == MVT::v2i64) { |
24170 | Results.push_back(DAG.getNode(X86ISD::CVTUI2P, dl, MVT::v4f32, Src)); |
24171 | return; |
24172 | } |
24173 | if (SrcVT != MVT::v2i32) |
24174 | return; |
24175 | SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64, Src); |
24176 | SDValue VBias = |
24177 | DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl, MVT::v2f64); |
24178 | SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn, |
24179 | DAG.getBitcast(MVT::v2i64, VBias)); |
24180 | Or = DAG.getBitcast(MVT::v2f64, Or); |
24181 | // TODO: Are there any fast-math-flags to propagate here? |
24182 | SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias); |
24183 | Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub)); |
24184 | return; |
24185 | } |
24186 | case ISD::FP_ROUND: { |
24187 | if (!TLI.isTypeLegal(N->getOperand(0).getValueType())) |
24188 | return; |
24189 | SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0)); |
24190 | Results.push_back(V); |
24191 | return; |
24192 | } |
24193 | case ISD::FP_EXTEND: { |
24194 | // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND. |
24195 | // No other ValueType for FP_EXTEND should reach this point. |
24196 | assert(N->getValueType(0) == MVT::v2f32 &&((N->getValueType(0) == MVT::v2f32 && "Do not know how to legalize this Node" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::v2f32 && \"Do not know how to legalize this Node\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24197, __PRETTY_FUNCTION__)) |
24197 | "Do not know how to legalize this Node")((N->getValueType(0) == MVT::v2f32 && "Do not know how to legalize this Node" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::v2f32 && \"Do not know how to legalize this Node\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24197, __PRETTY_FUNCTION__)); |
24198 | return; |
24199 | } |
24200 | case ISD::INTRINSIC_W_CHAIN: { |
24201 | unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); |
24202 | switch (IntNo) { |
24203 | default : llvm_unreachable("Do not know how to custom type "::llvm::llvm_unreachable_internal("Do not know how to custom type " "legalize this intrinsic operation!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24204) |
24204 | "legalize this intrinsic operation!")::llvm::llvm_unreachable_internal("Do not know how to custom type " "legalize this intrinsic operation!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24204); |
24205 | case Intrinsic::x86_rdtsc: |
24206 | return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget, |
24207 | Results); |
24208 | case Intrinsic::x86_rdtscp: |
24209 | return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget, |
24210 | Results); |
24211 | case Intrinsic::x86_rdpmc: |
24212 | return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results); |
24213 | |
24214 | case Intrinsic::x86_xgetbv: |
24215 | return getExtendedControlRegister(N, dl, DAG, Subtarget, Results); |
24216 | } |
24217 | } |
24218 | case ISD::INTRINSIC_WO_CHAIN: { |
24219 | if (SDValue V = LowerINTRINSIC_WO_CHAIN(SDValue(N, 0), Subtarget, DAG)) |
24220 | Results.push_back(V); |
24221 | return; |
24222 | } |
24223 | case ISD::READCYCLECOUNTER: { |
24224 | return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget, |
24225 | Results); |
24226 | } |
24227 | case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: { |
24228 | EVT T = N->getValueType(0); |
24229 | assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair")(((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair" ) ? static_cast<void> (0) : __assert_fail ("(T == MVT::i64 || T == MVT::i128) && \"can only expand cmpxchg pair\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24229, __PRETTY_FUNCTION__)); |
24230 | bool Regs64bit = T == MVT::i128; |
24231 | MVT HalfT = Regs64bit ? MVT::i64 : MVT::i32; |
24232 | SDValue cpInL, cpInH; |
24233 | cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2), |
24234 | DAG.getConstant(0, dl, HalfT)); |
24235 | cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2), |
24236 | DAG.getConstant(1, dl, HalfT)); |
24237 | cpInL = DAG.getCopyToReg(N->getOperand(0), dl, |
24238 | Regs64bit ? X86::RAX : X86::EAX, |
24239 | cpInL, SDValue()); |
24240 | cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, |
24241 | Regs64bit ? X86::RDX : X86::EDX, |
24242 | cpInH, cpInL.getValue(1)); |
24243 | SDValue swapInL, swapInH; |
24244 | swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3), |
24245 | DAG.getConstant(0, dl, HalfT)); |
24246 | swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3), |
24247 | DAG.getConstant(1, dl, HalfT)); |
24248 | swapInH = |
24249 | DAG.getCopyToReg(cpInH.getValue(0), dl, Regs64bit ? X86::RCX : X86::ECX, |
24250 | swapInH, cpInH.getValue(1)); |
24251 | // If the current function needs the base pointer, RBX, |
24252 | // we shouldn't use cmpxchg directly. |
24253 | // Indeed the lowering of that instruction will clobber |
24254 | // that register and since RBX will be a reserved register |
24255 | // the register allocator will not make sure its value will |
24256 | // be properly saved and restored around this live-range. |
24257 | const X86RegisterInfo *TRI = Subtarget.getRegisterInfo(); |
24258 | SDValue Result; |
24259 | SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue); |
24260 | unsigned BasePtr = TRI->getBaseRegister(); |
24261 | MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand(); |
24262 | if (TRI->hasBasePointer(DAG.getMachineFunction()) && |
24263 | (BasePtr == X86::RBX || BasePtr == X86::EBX)) { |
24264 | // ISel prefers the LCMPXCHG64 variant. |
24265 | // If that assert breaks, that means it is not the case anymore, |
24266 | // and we need to teach LCMPXCHG8_SAVE_EBX_DAG how to save RBX, |
24267 | // not just EBX. This is a matter of accepting i64 input for that |
24268 | // pseudo, and restoring into the register of the right wide |
24269 | // in expand pseudo. Everything else should just work. |
24270 | assert(((Regs64bit == (BasePtr == X86::RBX)) || BasePtr == X86::EBX) &&((((Regs64bit == (BasePtr == X86::RBX)) || BasePtr == X86::EBX ) && "Saving only half of the RBX") ? static_cast< void> (0) : __assert_fail ("((Regs64bit == (BasePtr == X86::RBX)) || BasePtr == X86::EBX) && \"Saving only half of the RBX\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24271, __PRETTY_FUNCTION__)) |
24271 | "Saving only half of the RBX")((((Regs64bit == (BasePtr == X86::RBX)) || BasePtr == X86::EBX ) && "Saving only half of the RBX") ? static_cast< void> (0) : __assert_fail ("((Regs64bit == (BasePtr == X86::RBX)) || BasePtr == X86::EBX) && \"Saving only half of the RBX\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24271, __PRETTY_FUNCTION__)); |
24272 | unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_SAVE_RBX_DAG |
24273 | : X86ISD::LCMPXCHG8_SAVE_EBX_DAG; |
24274 | SDValue RBXSave = DAG.getCopyFromReg(swapInH.getValue(0), dl, |
24275 | Regs64bit ? X86::RBX : X86::EBX, |
24276 | HalfT, swapInH.getValue(1)); |
24277 | SDValue Ops[] = {/*Chain*/ RBXSave.getValue(1), N->getOperand(1), swapInL, |
24278 | RBXSave, |
24279 | /*Glue*/ RBXSave.getValue(2)}; |
24280 | Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO); |
24281 | } else { |
24282 | unsigned Opcode = |
24283 | Regs64bit ? X86ISD::LCMPXCHG16_DAG : X86ISD::LCMPXCHG8_DAG; |
24284 | swapInL = DAG.getCopyToReg(swapInH.getValue(0), dl, |
24285 | Regs64bit ? X86::RBX : X86::EBX, swapInL, |
24286 | swapInH.getValue(1)); |
24287 | SDValue Ops[] = {swapInL.getValue(0), N->getOperand(1), |
24288 | swapInL.getValue(1)}; |
24289 | Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO); |
24290 | } |
24291 | SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, |
24292 | Regs64bit ? X86::RAX : X86::EAX, |
24293 | HalfT, Result.getValue(1)); |
24294 | SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, |
24295 | Regs64bit ? X86::RDX : X86::EDX, |
24296 | HalfT, cpOutL.getValue(2)); |
24297 | SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)}; |
24298 | |
24299 | SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS, |
24300 | MVT::i32, cpOutH.getValue(2)); |
24301 | SDValue Success = getSETCC(X86::COND_E, EFLAGS, dl, DAG); |
24302 | Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1)); |
24303 | |
24304 | Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF)); |
24305 | Results.push_back(Success); |
24306 | Results.push_back(EFLAGS.getValue(1)); |
24307 | return; |
24308 | } |
24309 | case ISD::ATOMIC_SWAP: |
24310 | case ISD::ATOMIC_LOAD_ADD: |
24311 | case ISD::ATOMIC_LOAD_SUB: |
24312 | case ISD::ATOMIC_LOAD_AND: |
24313 | case ISD::ATOMIC_LOAD_OR: |
24314 | case ISD::ATOMIC_LOAD_XOR: |
24315 | case ISD::ATOMIC_LOAD_NAND: |
24316 | case ISD::ATOMIC_LOAD_MIN: |
24317 | case ISD::ATOMIC_LOAD_MAX: |
24318 | case ISD::ATOMIC_LOAD_UMIN: |
24319 | case ISD::ATOMIC_LOAD_UMAX: |
24320 | case ISD::ATOMIC_LOAD: { |
24321 | // Delegate to generic TypeLegalization. Situations we can really handle |
24322 | // should have already been dealt with by AtomicExpandPass.cpp. |
24323 | break; |
24324 | } |
24325 | case ISD::BITCAST: { |
24326 | assert(Subtarget.hasSSE2() && "Requires at least SSE2!")((Subtarget.hasSSE2() && "Requires at least SSE2!") ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSE2() && \"Requires at least SSE2!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24326, __PRETTY_FUNCTION__)); |
24327 | EVT DstVT = N->getValueType(0); |
24328 | EVT SrcVT = N->getOperand(0)->getValueType(0); |
24329 | |
24330 | if (SrcVT != MVT::f64 || |
24331 | (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8)) |
24332 | return; |
24333 | |
24334 | unsigned NumElts = DstVT.getVectorNumElements(); |
24335 | EVT SVT = DstVT.getVectorElementType(); |
24336 | EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2); |
24337 | SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, |
24338 | MVT::v2f64, N->getOperand(0)); |
24339 | SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded); |
24340 | |
24341 | if (ExperimentalVectorWideningLegalization) { |
24342 | // If we are legalizing vectors by widening, we already have the desired |
24343 | // legal vector type, just return it. |
24344 | Results.push_back(ToVecInt); |
24345 | return; |
24346 | } |
24347 | |
24348 | SmallVector<SDValue, 8> Elts; |
24349 | for (unsigned i = 0, e = NumElts; i != e; ++i) |
24350 | Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, |
24351 | ToVecInt, DAG.getIntPtrConstant(i, dl))); |
24352 | |
24353 | Results.push_back(DAG.getBuildVector(DstVT, dl, Elts)); |
24354 | } |
24355 | } |
24356 | } |
24357 | |
24358 | const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const { |
24359 | switch ((X86ISD::NodeType)Opcode) { |
24360 | case X86ISD::FIRST_NUMBER: break; |
24361 | case X86ISD::BSF: return "X86ISD::BSF"; |
24362 | case X86ISD::BSR: return "X86ISD::BSR"; |
24363 | case X86ISD::SHLD: return "X86ISD::SHLD"; |
24364 | case X86ISD::SHRD: return "X86ISD::SHRD"; |
24365 | case X86ISD::FAND: return "X86ISD::FAND"; |
24366 | case X86ISD::FANDN: return "X86ISD::FANDN"; |
24367 | case X86ISD::FOR: return "X86ISD::FOR"; |
24368 | case X86ISD::FXOR: return "X86ISD::FXOR"; |
24369 | case X86ISD::FILD: return "X86ISD::FILD"; |
24370 | case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG"; |
24371 | case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM"; |
24372 | case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM"; |
24373 | case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM"; |
24374 | case X86ISD::FLD: return "X86ISD::FLD"; |
24375 | case X86ISD::FST: return "X86ISD::FST"; |
24376 | case X86ISD::CALL: return "X86ISD::CALL"; |
24377 | case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG"; |
24378 | case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG"; |
24379 | case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG"; |
24380 | case X86ISD::BT: return "X86ISD::BT"; |
24381 | case X86ISD::CMP: return "X86ISD::CMP"; |
24382 | case X86ISD::COMI: return "X86ISD::COMI"; |
24383 | case X86ISD::UCOMI: return "X86ISD::UCOMI"; |
24384 | case X86ISD::CMPM: return "X86ISD::CMPM"; |
24385 | case X86ISD::CMPMU: return "X86ISD::CMPMU"; |
24386 | case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND"; |
24387 | case X86ISD::SETCC: return "X86ISD::SETCC"; |
24388 | case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY"; |
24389 | case X86ISD::FSETCC: return "X86ISD::FSETCC"; |
24390 | case X86ISD::FSETCCM: return "X86ISD::FSETCCM"; |
24391 | case X86ISD::FSETCCM_RND: return "X86ISD::FSETCCM_RND"; |
24392 | case X86ISD::CMOV: return "X86ISD::CMOV"; |
24393 | case X86ISD::BRCOND: return "X86ISD::BRCOND"; |
24394 | case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG"; |
24395 | case X86ISD::IRET: return "X86ISD::IRET"; |
24396 | case X86ISD::REP_STOS: return "X86ISD::REP_STOS"; |
24397 | case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS"; |
24398 | case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg"; |
24399 | case X86ISD::Wrapper: return "X86ISD::Wrapper"; |
24400 | case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP"; |
24401 | case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q"; |
24402 | case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W"; |
24403 | case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D"; |
24404 | case X86ISD::PEXTRB: return "X86ISD::PEXTRB"; |
24405 | case X86ISD::PEXTRW: return "X86ISD::PEXTRW"; |
24406 | case X86ISD::INSERTPS: return "X86ISD::INSERTPS"; |
24407 | case X86ISD::PINSRB: return "X86ISD::PINSRB"; |
24408 | case X86ISD::PINSRW: return "X86ISD::PINSRW"; |
24409 | case X86ISD::PSHUFB: return "X86ISD::PSHUFB"; |
24410 | case X86ISD::ANDNP: return "X86ISD::ANDNP"; |
24411 | case X86ISD::BLENDI: return "X86ISD::BLENDI"; |
24412 | case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND"; |
24413 | case X86ISD::ADDUS: return "X86ISD::ADDUS"; |
24414 | case X86ISD::SUBUS: return "X86ISD::SUBUS"; |
24415 | case X86ISD::HADD: return "X86ISD::HADD"; |
24416 | case X86ISD::HSUB: return "X86ISD::HSUB"; |
24417 | case X86ISD::FHADD: return "X86ISD::FHADD"; |
24418 | case X86ISD::FHSUB: return "X86ISD::FHSUB"; |
24419 | case X86ISD::CONFLICT: return "X86ISD::CONFLICT"; |
24420 | case X86ISD::FMAX: return "X86ISD::FMAX"; |
24421 | case X86ISD::FMAXS: return "X86ISD::FMAXS"; |
24422 | case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND"; |
24423 | case X86ISD::FMAXS_RND: return "X86ISD::FMAX_RND"; |
24424 | case X86ISD::FMIN: return "X86ISD::FMIN"; |
24425 | case X86ISD::FMINS: return "X86ISD::FMINS"; |
24426 | case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND"; |
24427 | case X86ISD::FMINS_RND: return "X86ISD::FMINS_RND"; |
24428 | case X86ISD::FMAXC: return "X86ISD::FMAXC"; |
24429 | case X86ISD::FMINC: return "X86ISD::FMINC"; |
24430 | case X86ISD::FRSQRT: return "X86ISD::FRSQRT"; |
24431 | case X86ISD::FRSQRTS: return "X86ISD::FRSQRTS"; |
24432 | case X86ISD::FRCP: return "X86ISD::FRCP"; |
24433 | case X86ISD::FRCPS: return "X86ISD::FRCPS"; |
24434 | case X86ISD::EXTRQI: return "X86ISD::EXTRQI"; |
24435 | case X86ISD::INSERTQI: return "X86ISD::INSERTQI"; |
24436 | case X86ISD::TLSADDR: return "X86ISD::TLSADDR"; |
24437 | case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR"; |
24438 | case X86ISD::TLSCALL: return "X86ISD::TLSCALL"; |
24439 | case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP"; |
24440 | case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP"; |
24441 | case X86ISD::EH_SJLJ_SETUP_DISPATCH: |
24442 | return "X86ISD::EH_SJLJ_SETUP_DISPATCH"; |
24443 | case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN"; |
24444 | case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN"; |
24445 | case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m"; |
24446 | case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r"; |
24447 | case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG"; |
24448 | case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG"; |
24449 | case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG"; |
24450 | case X86ISD::LCMPXCHG8_SAVE_EBX_DAG: |
24451 | return "X86ISD::LCMPXCHG8_SAVE_EBX_DAG"; |
24452 | case X86ISD::LCMPXCHG16_SAVE_RBX_DAG: |
24453 | return "X86ISD::LCMPXCHG16_SAVE_RBX_DAG"; |
24454 | case X86ISD::LADD: return "X86ISD::LADD"; |
24455 | case X86ISD::LSUB: return "X86ISD::LSUB"; |
24456 | case X86ISD::LOR: return "X86ISD::LOR"; |
24457 | case X86ISD::LXOR: return "X86ISD::LXOR"; |
24458 | case X86ISD::LAND: return "X86ISD::LAND"; |
24459 | case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL"; |
24460 | case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD"; |
24461 | case X86ISD::VZEXT: return "X86ISD::VZEXT"; |
24462 | case X86ISD::VSEXT: return "X86ISD::VSEXT"; |
24463 | case X86ISD::VTRUNC: return "X86ISD::VTRUNC"; |
24464 | case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS"; |
24465 | case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS"; |
24466 | case X86ISD::VTRUNCSTORES: return "X86ISD::VTRUNCSTORES"; |
24467 | case X86ISD::VTRUNCSTOREUS: return "X86ISD::VTRUNCSTOREUS"; |
24468 | case X86ISD::VMTRUNCSTORES: return "X86ISD::VMTRUNCSTORES"; |
24469 | case X86ISD::VMTRUNCSTOREUS: return "X86ISD::VMTRUNCSTOREUS"; |
24470 | case X86ISD::VFPEXT: return "X86ISD::VFPEXT"; |
24471 | case X86ISD::VFPEXT_RND: return "X86ISD::VFPEXT_RND"; |
24472 | case X86ISD::VFPEXTS_RND: return "X86ISD::VFPEXTS_RND"; |
24473 | case X86ISD::VFPROUND: return "X86ISD::VFPROUND"; |
24474 | case X86ISD::VFPROUND_RND: return "X86ISD::VFPROUND_RND"; |
24475 | case X86ISD::VFPROUNDS_RND: return "X86ISD::VFPROUNDS_RND"; |
24476 | case X86ISD::CVT2MASK: return "X86ISD::CVT2MASK"; |
24477 | case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ"; |
24478 | case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ"; |
24479 | case X86ISD::VSHL: return "X86ISD::VSHL"; |
24480 | case X86ISD::VSRL: return "X86ISD::VSRL"; |
24481 | case X86ISD::VSRA: return "X86ISD::VSRA"; |
24482 | case X86ISD::VSHLI: return "X86ISD::VSHLI"; |
24483 | case X86ISD::VSRLI: return "X86ISD::VSRLI"; |
24484 | case X86ISD::VSRAI: return "X86ISD::VSRAI"; |
24485 | case X86ISD::VSRAV: return "X86ISD::VSRAV"; |
24486 | case X86ISD::VROTLI: return "X86ISD::VROTLI"; |
24487 | case X86ISD::VROTRI: return "X86ISD::VROTRI"; |
24488 | case X86ISD::VPPERM: return "X86ISD::VPPERM"; |
24489 | case X86ISD::CMPP: return "X86ISD::CMPP"; |
24490 | case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ"; |
24491 | case X86ISD::PCMPGT: return "X86ISD::PCMPGT"; |
24492 | case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM"; |
24493 | case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM"; |
24494 | case X86ISD::ADD: return "X86ISD::ADD"; |
24495 | case X86ISD::SUB: return "X86ISD::SUB"; |
24496 | case X86ISD::ADC: return "X86ISD::ADC"; |
24497 | case X86ISD::SBB: return "X86ISD::SBB"; |
24498 | case X86ISD::SMUL: return "X86ISD::SMUL"; |
24499 | case X86ISD::UMUL: return "X86ISD::UMUL"; |
24500 | case X86ISD::SMUL8: return "X86ISD::SMUL8"; |
24501 | case X86ISD::UMUL8: return "X86ISD::UMUL8"; |
24502 | case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG"; |
24503 | case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG"; |
24504 | case X86ISD::INC: return "X86ISD::INC"; |
24505 | case X86ISD::DEC: return "X86ISD::DEC"; |
24506 | case X86ISD::OR: return "X86ISD::OR"; |
24507 | case X86ISD::XOR: return "X86ISD::XOR"; |
24508 | case X86ISD::AND: return "X86ISD::AND"; |
24509 | case X86ISD::BEXTR: return "X86ISD::BEXTR"; |
24510 | case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM"; |
24511 | case X86ISD::MOVMSK: return "X86ISD::MOVMSK"; |
24512 | case X86ISD::PTEST: return "X86ISD::PTEST"; |
24513 | case X86ISD::TESTP: return "X86ISD::TESTP"; |
24514 | case X86ISD::TESTM: return "X86ISD::TESTM"; |
24515 | case X86ISD::TESTNM: return "X86ISD::TESTNM"; |
24516 | case X86ISD::KORTEST: return "X86ISD::KORTEST"; |
24517 | case X86ISD::KTEST: return "X86ISD::KTEST"; |
24518 | case X86ISD::KSHIFTL: return "X86ISD::KSHIFTL"; |
24519 | case X86ISD::KSHIFTR: return "X86ISD::KSHIFTR"; |
24520 | case X86ISD::PACKSS: return "X86ISD::PACKSS"; |
24521 | case X86ISD::PACKUS: return "X86ISD::PACKUS"; |
24522 | case X86ISD::PALIGNR: return "X86ISD::PALIGNR"; |
24523 | case X86ISD::VALIGN: return "X86ISD::VALIGN"; |
24524 | case X86ISD::PSHUFD: return "X86ISD::PSHUFD"; |
24525 | case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW"; |
24526 | case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW"; |
24527 | case X86ISD::SHUFP: return "X86ISD::SHUFP"; |
24528 | case X86ISD::SHUF128: return "X86ISD::SHUF128"; |
24529 | case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS"; |
24530 | case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD"; |
24531 | case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS"; |
24532 | case X86ISD::MOVLPS: return "X86ISD::MOVLPS"; |
24533 | case X86ISD::MOVLPD: return "X86ISD::MOVLPD"; |
24534 | case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP"; |
24535 | case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP"; |
24536 | case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP"; |
24537 | case X86ISD::MOVSD: return "X86ISD::MOVSD"; |
24538 | case X86ISD::MOVSS: return "X86ISD::MOVSS"; |
24539 | case X86ISD::UNPCKL: return "X86ISD::UNPCKL"; |
24540 | case X86ISD::UNPCKH: return "X86ISD::UNPCKH"; |
24541 | case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST"; |
24542 | case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM"; |
24543 | case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST"; |
24544 | case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT"; |
24545 | case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV"; |
24546 | case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI"; |
24547 | case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128"; |
24548 | case X86ISD::VPERMV: return "X86ISD::VPERMV"; |
24549 | case X86ISD::VPERMV3: return "X86ISD::VPERMV3"; |
24550 | case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3"; |
24551 | case X86ISD::VPERMI: return "X86ISD::VPERMI"; |
24552 | case X86ISD::VPTERNLOG: return "X86ISD::VPTERNLOG"; |
24553 | case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM"; |
24554 | case X86ISD::VFIXUPIMMS: return "X86ISD::VFIXUPIMMS"; |
24555 | case X86ISD::VRANGE: return "X86ISD::VRANGE"; |
24556 | case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ"; |
24557 | case X86ISD::PMULDQ: return "X86ISD::PMULDQ"; |
24558 | case X86ISD::PSADBW: return "X86ISD::PSADBW"; |
24559 | case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW"; |
24560 | case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS"; |
24561 | case X86ISD::VAARG_64: return "X86ISD::VAARG_64"; |
24562 | case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA"; |
24563 | case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER"; |
24564 | case X86ISD::MFENCE: return "X86ISD::MFENCE"; |
24565 | case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA"; |
24566 | case X86ISD::SAHF: return "X86ISD::SAHF"; |
24567 | case X86ISD::RDRAND: return "X86ISD::RDRAND"; |
24568 | case X86ISD::RDSEED: return "X86ISD::RDSEED"; |
24569 | case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW"; |
24570 | case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD"; |
24571 | case X86ISD::VPROT: return "X86ISD::VPROT"; |
24572 | case X86ISD::VPROTI: return "X86ISD::VPROTI"; |
24573 | case X86ISD::VPSHA: return "X86ISD::VPSHA"; |
24574 | case X86ISD::VPSHL: return "X86ISD::VPSHL"; |
24575 | case X86ISD::VPCOM: return "X86ISD::VPCOM"; |
24576 | case X86ISD::VPCOMU: return "X86ISD::VPCOMU"; |
24577 | case X86ISD::VPERMIL2: return "X86ISD::VPERMIL2"; |
24578 | case X86ISD::FMADD: return "X86ISD::FMADD"; |
24579 | case X86ISD::FMSUB: return "X86ISD::FMSUB"; |
24580 | case X86ISD::FNMADD: return "X86ISD::FNMADD"; |
24581 | case X86ISD::FNMSUB: return "X86ISD::FNMSUB"; |
24582 | case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB"; |
24583 | case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD"; |
24584 | case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND"; |
24585 | case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND"; |
24586 | case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND"; |
24587 | case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND"; |
24588 | case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND"; |
24589 | case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND"; |
24590 | case X86ISD::FMADDS1_RND: return "X86ISD::FMADDS1_RND"; |
24591 | case X86ISD::FNMADDS1_RND: return "X86ISD::FNMADDS1_RND"; |
24592 | case X86ISD::FMSUBS1_RND: return "X86ISD::FMSUBS1_RND"; |
24593 | case X86ISD::FNMSUBS1_RND: return "X86ISD::FNMSUBS1_RND"; |
24594 | case X86ISD::FMADDS3_RND: return "X86ISD::FMADDS3_RND"; |
24595 | case X86ISD::FNMADDS3_RND: return "X86ISD::FNMADDS3_RND"; |
24596 | case X86ISD::FMSUBS3_RND: return "X86ISD::FMSUBS3_RND"; |
24597 | case X86ISD::FNMSUBS3_RND: return "X86ISD::FNMSUBS3_RND"; |
24598 | case X86ISD::VPMADD52H: return "X86ISD::VPMADD52H"; |
24599 | case X86ISD::VPMADD52L: return "X86ISD::VPMADD52L"; |
24600 | case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE"; |
24601 | case X86ISD::VRNDSCALES: return "X86ISD::VRNDSCALES"; |
24602 | case X86ISD::VREDUCE: return "X86ISD::VREDUCE"; |
24603 | case X86ISD::VREDUCES: return "X86ISD::VREDUCES"; |
24604 | case X86ISD::VGETMANT: return "X86ISD::VGETMANT"; |
24605 | case X86ISD::VGETMANTS: return "X86ISD::VGETMANTS"; |
24606 | case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI"; |
24607 | case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI"; |
24608 | case X86ISD::XTEST: return "X86ISD::XTEST"; |
24609 | case X86ISD::COMPRESS: return "X86ISD::COMPRESS"; |
24610 | case X86ISD::EXPAND: return "X86ISD::EXPAND"; |
24611 | case X86ISD::SELECT: return "X86ISD::SELECT"; |
24612 | case X86ISD::SELECTS: return "X86ISD::SELECTS"; |
24613 | case X86ISD::ADDSUB: return "X86ISD::ADDSUB"; |
24614 | case X86ISD::RCP28: return "X86ISD::RCP28"; |
24615 | case X86ISD::RCP28S: return "X86ISD::RCP28S"; |
24616 | case X86ISD::EXP2: return "X86ISD::EXP2"; |
24617 | case X86ISD::RSQRT28: return "X86ISD::RSQRT28"; |
24618 | case X86ISD::RSQRT28S: return "X86ISD::RSQRT28S"; |
24619 | case X86ISD::FADD_RND: return "X86ISD::FADD_RND"; |
24620 | case X86ISD::FADDS_RND: return "X86ISD::FADDS_RND"; |
24621 | case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND"; |
24622 | case X86ISD::FSUBS_RND: return "X86ISD::FSUBS_RND"; |
24623 | case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND"; |
24624 | case X86ISD::FMULS_RND: return "X86ISD::FMULS_RND"; |
24625 | case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND"; |
24626 | case X86ISD::FDIVS_RND: return "X86ISD::FDIVS_RND"; |
24627 | case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND"; |
24628 | case X86ISD::FSQRTS_RND: return "X86ISD::FSQRTS_RND"; |
24629 | case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND"; |
24630 | case X86ISD::FGETEXPS_RND: return "X86ISD::FGETEXPS_RND"; |
24631 | case X86ISD::SCALEF: return "X86ISD::SCALEF"; |
24632 | case X86ISD::SCALEFS: return "X86ISD::SCALEFS"; |
24633 | case X86ISD::ADDS: return "X86ISD::ADDS"; |
24634 | case X86ISD::SUBS: return "X86ISD::SUBS"; |
24635 | case X86ISD::AVG: return "X86ISD::AVG"; |
24636 | case X86ISD::MULHRS: return "X86ISD::MULHRS"; |
24637 | case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND"; |
24638 | case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND"; |
24639 | case X86ISD::CVTTP2SI: return "X86ISD::CVTTP2SI"; |
24640 | case X86ISD::CVTTP2UI: return "X86ISD::CVTTP2UI"; |
24641 | case X86ISD::CVTTP2SI_RND: return "X86ISD::CVTTP2SI_RND"; |
24642 | case X86ISD::CVTTP2UI_RND: return "X86ISD::CVTTP2UI_RND"; |
24643 | case X86ISD::CVTTS2SI_RND: return "X86ISD::CVTTS2SI_RND"; |
24644 | case X86ISD::CVTTS2UI_RND: return "X86ISD::CVTTS2UI_RND"; |
24645 | case X86ISD::CVTSI2P: return "X86ISD::CVTSI2P"; |
24646 | case X86ISD::CVTUI2P: return "X86ISD::CVTUI2P"; |
24647 | case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS"; |
24648 | case X86ISD::VFPCLASSS: return "X86ISD::VFPCLASSS"; |
24649 | case X86ISD::MULTISHIFT: return "X86ISD::MULTISHIFT"; |
24650 | case X86ISD::SCALAR_SINT_TO_FP_RND: return "X86ISD::SCALAR_SINT_TO_FP_RND"; |
24651 | case X86ISD::SCALAR_UINT_TO_FP_RND: return "X86ISD::SCALAR_UINT_TO_FP_RND"; |
24652 | case X86ISD::CVTPS2PH: return "X86ISD::CVTPS2PH"; |
24653 | case X86ISD::CVTPH2PS: return "X86ISD::CVTPH2PS"; |
24654 | case X86ISD::CVTP2SI: return "X86ISD::CVTP2SI"; |
24655 | case X86ISD::CVTP2UI: return "X86ISD::CVTP2UI"; |
24656 | case X86ISD::CVTP2SI_RND: return "X86ISD::CVTP2SI_RND"; |
24657 | case X86ISD::CVTP2UI_RND: return "X86ISD::CVTP2UI_RND"; |
24658 | case X86ISD::CVTS2SI_RND: return "X86ISD::CVTS2SI_RND"; |
24659 | case X86ISD::CVTS2UI_RND: return "X86ISD::CVTS2UI_RND"; |
24660 | case X86ISD::LWPINS: return "X86ISD::LWPINS"; |
24661 | case X86ISD::MGATHER: return "X86ISD::MGATHER"; |
24662 | } |
24663 | return nullptr; |
24664 | } |
24665 | |
24666 | /// Return true if the addressing mode represented by AM is legal for this |
24667 | /// target, for a load/store of the specified type. |
24668 | bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL, |
24669 | const AddrMode &AM, Type *Ty, |
24670 | unsigned AS) const { |
24671 | // X86 supports extremely general addressing modes. |
24672 | CodeModel::Model M = getTargetMachine().getCodeModel(); |
24673 | |
24674 | // X86 allows a sign-extended 32-bit immediate field as a displacement. |
24675 | if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr)) |
24676 | return false; |
24677 | |
24678 | if (AM.BaseGV) { |
24679 | unsigned GVFlags = Subtarget.classifyGlobalReference(AM.BaseGV); |
24680 | |
24681 | // If a reference to this global requires an extra load, we can't fold it. |
24682 | if (isGlobalStubReference(GVFlags)) |
24683 | return false; |
24684 | |
24685 | // If BaseGV requires a register for the PIC base, we cannot also have a |
24686 | // BaseReg specified. |
24687 | if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags)) |
24688 | return false; |
24689 | |
24690 | // If lower 4G is not available, then we must use rip-relative addressing. |
24691 | if ((M != CodeModel::Small || isPositionIndependent()) && |
24692 | Subtarget.is64Bit() && (AM.BaseOffs || AM.Scale > 1)) |
24693 | return false; |
24694 | } |
24695 | |
24696 | switch (AM.Scale) { |
24697 | case 0: |
24698 | case 1: |
24699 | case 2: |
24700 | case 4: |
24701 | case 8: |
24702 | // These scales always work. |
24703 | break; |
24704 | case 3: |
24705 | case 5: |
24706 | case 9: |
24707 | // These scales are formed with basereg+scalereg. Only accept if there is |
24708 | // no basereg yet. |
24709 | if (AM.HasBaseReg) |
24710 | return false; |
24711 | break; |
24712 | default: // Other stuff never works. |
24713 | return false; |
24714 | } |
24715 | |
24716 | return true; |
24717 | } |
24718 | |
24719 | bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const { |
24720 | unsigned Bits = Ty->getScalarSizeInBits(); |
24721 | |
24722 | // 8-bit shifts are always expensive, but versions with a scalar amount aren't |
24723 | // particularly cheaper than those without. |
24724 | if (Bits == 8) |
24725 | return false; |
24726 | |
24727 | // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make |
24728 | // variable shifts just as cheap as scalar ones. |
24729 | if (Subtarget.hasInt256() && (Bits == 32 || Bits == 64)) |
24730 | return false; |
24731 | |
24732 | // Otherwise, it's significantly cheaper to shift by a scalar amount than by a |
24733 | // fully general vector. |
24734 | return true; |
24735 | } |
24736 | |
24737 | bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const { |
24738 | if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy()) |
24739 | return false; |
24740 | unsigned NumBits1 = Ty1->getPrimitiveSizeInBits(); |
24741 | unsigned NumBits2 = Ty2->getPrimitiveSizeInBits(); |
24742 | return NumBits1 > NumBits2; |
24743 | } |
24744 | |
24745 | bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const { |
24746 | if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy()) |
24747 | return false; |
24748 | |
24749 | if (!isTypeLegal(EVT::getEVT(Ty1))) |
24750 | return false; |
24751 | |
24752 | assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop")((Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop" ) ? static_cast<void> (0) : __assert_fail ("Ty1->getPrimitiveSizeInBits() <= 64 && \"i128 is probably not a noop\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24752, __PRETTY_FUNCTION__)); |
24753 | |
24754 | // Assuming the caller doesn't have a zeroext or signext return parameter, |
24755 | // truncation all the way down to i1 is valid. |
24756 | return true; |
24757 | } |
24758 | |
24759 | bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const { |
24760 | return isInt<32>(Imm); |
24761 | } |
24762 | |
24763 | bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const { |
24764 | // Can also use sub to handle negated immediates. |
24765 | return isInt<32>(Imm); |
24766 | } |
24767 | |
24768 | bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const { |
24769 | if (!VT1.isInteger() || !VT2.isInteger()) |
24770 | return false; |
24771 | unsigned NumBits1 = VT1.getSizeInBits(); |
24772 | unsigned NumBits2 = VT2.getSizeInBits(); |
24773 | return NumBits1 > NumBits2; |
24774 | } |
24775 | |
24776 | bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const { |
24777 | // x86-64 implicitly zero-extends 32-bit results in 64-bit registers. |
24778 | return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget.is64Bit(); |
24779 | } |
24780 | |
24781 | bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const { |
24782 | // x86-64 implicitly zero-extends 32-bit results in 64-bit registers. |
24783 | return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget.is64Bit(); |
24784 | } |
24785 | |
24786 | bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const { |
24787 | EVT VT1 = Val.getValueType(); |
24788 | if (isZExtFree(VT1, VT2)) |
24789 | return true; |
24790 | |
24791 | if (Val.getOpcode() != ISD::LOAD) |
24792 | return false; |
24793 | |
24794 | if (!VT1.isSimple() || !VT1.isInteger() || |
24795 | !VT2.isSimple() || !VT2.isInteger()) |
24796 | return false; |
24797 | |
24798 | switch (VT1.getSimpleVT().SimpleTy) { |
24799 | default: break; |
24800 | case MVT::i8: |
24801 | case MVT::i16: |
24802 | case MVT::i32: |
24803 | // X86 has 8, 16, and 32-bit zero-extending loads. |
24804 | return true; |
24805 | } |
24806 | |
24807 | return false; |
24808 | } |
24809 | |
24810 | bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; } |
24811 | |
24812 | bool |
24813 | X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const { |
24814 | if (!Subtarget.hasAnyFMA()) |
24815 | return false; |
24816 | |
24817 | VT = VT.getScalarType(); |
24818 | |
24819 | if (!VT.isSimple()) |
24820 | return false; |
24821 | |
24822 | switch (VT.getSimpleVT().SimpleTy) { |
24823 | case MVT::f32: |
24824 | case MVT::f64: |
24825 | return true; |
24826 | default: |
24827 | break; |
24828 | } |
24829 | |
24830 | return false; |
24831 | } |
24832 | |
24833 | bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const { |
24834 | // i16 instructions are longer (0x66 prefix) and potentially slower. |
24835 | return !(VT1 == MVT::i32 && VT2 == MVT::i16); |
24836 | } |
24837 | |
24838 | /// Targets can use this to indicate that they only support *some* |
24839 | /// VECTOR_SHUFFLE operations, those with specific masks. |
24840 | /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values |
24841 | /// are assumed to be legal. |
24842 | bool |
24843 | X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M, |
24844 | EVT VT) const { |
24845 | if (!VT.isSimple()) |
24846 | return false; |
24847 | |
24848 | // Not for i1 vectors |
24849 | if (VT.getSimpleVT().getScalarType() == MVT::i1) |
24850 | return false; |
24851 | |
24852 | // Very little shuffling can be done for 64-bit vectors right now. |
24853 | if (VT.getSimpleVT().getSizeInBits() == 64) |
24854 | return false; |
24855 | |
24856 | // We only care that the types being shuffled are legal. The lowering can |
24857 | // handle any possible shuffle mask that results. |
24858 | return isTypeLegal(VT.getSimpleVT()); |
24859 | } |
24860 | |
24861 | bool |
24862 | X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask, |
24863 | EVT VT) const { |
24864 | // Just delegate to the generic legality, clear masks aren't special. |
24865 | return isShuffleMaskLegal(Mask, VT); |
24866 | } |
24867 | |
24868 | //===----------------------------------------------------------------------===// |
24869 | // X86 Scheduler Hooks |
24870 | //===----------------------------------------------------------------------===// |
24871 | |
24872 | /// Utility function to emit xbegin specifying the start of an RTM region. |
24873 | static MachineBasicBlock *emitXBegin(MachineInstr &MI, MachineBasicBlock *MBB, |
24874 | const TargetInstrInfo *TII) { |
24875 | DebugLoc DL = MI.getDebugLoc(); |
24876 | |
24877 | const BasicBlock *BB = MBB->getBasicBlock(); |
24878 | MachineFunction::iterator I = ++MBB->getIterator(); |
24879 | |
24880 | // For the v = xbegin(), we generate |
24881 | // |
24882 | // thisMBB: |
24883 | // xbegin sinkMBB |
24884 | // |
24885 | // mainMBB: |
24886 | // s0 = -1 |
24887 | // |
24888 | // fallBB: |
24889 | // eax = # XABORT_DEF |
24890 | // s1 = eax |
24891 | // |
24892 | // sinkMBB: |
24893 | // v = phi(s0/mainBB, s1/fallBB) |
24894 | |
24895 | MachineBasicBlock *thisMBB = MBB; |
24896 | MachineFunction *MF = MBB->getParent(); |
24897 | MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB); |
24898 | MachineBasicBlock *fallMBB = MF->CreateMachineBasicBlock(BB); |
24899 | MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB); |
24900 | MF->insert(I, mainMBB); |
24901 | MF->insert(I, fallMBB); |
24902 | MF->insert(I, sinkMBB); |
24903 | |
24904 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
24905 | sinkMBB->splice(sinkMBB->begin(), MBB, |
24906 | std::next(MachineBasicBlock::iterator(MI)), MBB->end()); |
24907 | sinkMBB->transferSuccessorsAndUpdatePHIs(MBB); |
24908 | |
24909 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
24910 | unsigned DstReg = MI.getOperand(0).getReg(); |
24911 | const TargetRegisterClass *RC = MRI.getRegClass(DstReg); |
24912 | unsigned mainDstReg = MRI.createVirtualRegister(RC); |
24913 | unsigned fallDstReg = MRI.createVirtualRegister(RC); |
24914 | |
24915 | // thisMBB: |
24916 | // xbegin fallMBB |
24917 | // # fallthrough to mainMBB |
24918 | // # abortion to fallMBB |
24919 | BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(fallMBB); |
24920 | thisMBB->addSuccessor(mainMBB); |
24921 | thisMBB->addSuccessor(fallMBB); |
24922 | |
24923 | // mainMBB: |
24924 | // mainDstReg := -1 |
24925 | BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), mainDstReg).addImm(-1); |
24926 | BuildMI(mainMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB); |
24927 | mainMBB->addSuccessor(sinkMBB); |
24928 | |
24929 | // fallMBB: |
24930 | // ; pseudo instruction to model hardware's definition from XABORT |
24931 | // EAX := XABORT_DEF |
24932 | // fallDstReg := EAX |
24933 | BuildMI(fallMBB, DL, TII->get(X86::XABORT_DEF)); |
24934 | BuildMI(fallMBB, DL, TII->get(TargetOpcode::COPY), fallDstReg) |
24935 | .addReg(X86::EAX); |
24936 | fallMBB->addSuccessor(sinkMBB); |
24937 | |
24938 | // sinkMBB: |
24939 | // DstReg := phi(mainDstReg/mainBB, fallDstReg/fallBB) |
24940 | BuildMI(*sinkMBB, sinkMBB->begin(), DL, TII->get(X86::PHI), DstReg) |
24941 | .addReg(mainDstReg).addMBB(mainMBB) |
24942 | .addReg(fallDstReg).addMBB(fallMBB); |
24943 | |
24944 | MI.eraseFromParent(); |
24945 | return sinkMBB; |
24946 | } |
24947 | |
24948 | // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8 |
24949 | // or XMM0_V32I8 in AVX all of this code can be replaced with that |
24950 | // in the .td file. |
24951 | static MachineBasicBlock *emitPCMPSTRM(MachineInstr &MI, MachineBasicBlock *BB, |
24952 | const TargetInstrInfo *TII) { |
24953 | unsigned Opc; |
24954 | switch (MI.getOpcode()) { |
24955 | default: llvm_unreachable("illegal opcode!")::llvm::llvm_unreachable_internal("illegal opcode!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24955); |
24956 | case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break; |
24957 | case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break; |
24958 | case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break; |
24959 | case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break; |
24960 | case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break; |
24961 | case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break; |
24962 | case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break; |
24963 | case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break; |
24964 | } |
24965 | |
24966 | DebugLoc dl = MI.getDebugLoc(); |
24967 | MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc)); |
24968 | |
24969 | unsigned NumArgs = MI.getNumOperands(); |
24970 | for (unsigned i = 1; i < NumArgs; ++i) { |
24971 | MachineOperand &Op = MI.getOperand(i); |
24972 | if (!(Op.isReg() && Op.isImplicit())) |
24973 | MIB.add(Op); |
24974 | } |
24975 | if (MI.hasOneMemOperand()) |
24976 | MIB->setMemRefs(MI.memoperands_begin(), MI.memoperands_end()); |
24977 | |
24978 | BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), MI.getOperand(0).getReg()) |
24979 | .addReg(X86::XMM0); |
24980 | |
24981 | MI.eraseFromParent(); |
24982 | return BB; |
24983 | } |
24984 | |
24985 | // FIXME: Custom handling because TableGen doesn't support multiple implicit |
24986 | // defs in an instruction pattern |
24987 | static MachineBasicBlock *emitPCMPSTRI(MachineInstr &MI, MachineBasicBlock *BB, |
24988 | const TargetInstrInfo *TII) { |
24989 | unsigned Opc; |
24990 | switch (MI.getOpcode()) { |
24991 | default: llvm_unreachable("illegal opcode!")::llvm::llvm_unreachable_internal("illegal opcode!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 24991); |
24992 | case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break; |
24993 | case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break; |
24994 | case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break; |
24995 | case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break; |
24996 | case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break; |
24997 | case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break; |
24998 | case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break; |
24999 | case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break; |
25000 | } |
25001 | |
25002 | DebugLoc dl = MI.getDebugLoc(); |
25003 | MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc)); |
25004 | |
25005 | unsigned NumArgs = MI.getNumOperands(); // remove the results |
25006 | for (unsigned i = 1; i < NumArgs; ++i) { |
25007 | MachineOperand &Op = MI.getOperand(i); |
25008 | if (!(Op.isReg() && Op.isImplicit())) |
25009 | MIB.add(Op); |
25010 | } |
25011 | if (MI.hasOneMemOperand()) |
25012 | MIB->setMemRefs(MI.memoperands_begin(), MI.memoperands_end()); |
25013 | |
25014 | BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), MI.getOperand(0).getReg()) |
25015 | .addReg(X86::ECX); |
25016 | |
25017 | MI.eraseFromParent(); |
25018 | return BB; |
25019 | } |
25020 | |
25021 | static MachineBasicBlock *emitWRPKRU(MachineInstr &MI, MachineBasicBlock *BB, |
25022 | const X86Subtarget &Subtarget) { |
25023 | DebugLoc dl = MI.getDebugLoc(); |
25024 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
25025 | |
25026 | // insert input VAL into EAX |
25027 | BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX) |
25028 | .addReg(MI.getOperand(0).getReg()); |
25029 | // insert zero to ECX |
25030 | BuildMI(*BB, MI, dl, TII->get(X86::MOV32r0), X86::ECX); |
25031 | |
25032 | // insert zero to EDX |
25033 | BuildMI(*BB, MI, dl, TII->get(X86::MOV32r0), X86::EDX); |
25034 | |
25035 | // insert WRPKRU instruction |
25036 | BuildMI(*BB, MI, dl, TII->get(X86::WRPKRUr)); |
25037 | |
25038 | MI.eraseFromParent(); // The pseudo is gone now. |
25039 | return BB; |
25040 | } |
25041 | |
25042 | static MachineBasicBlock *emitRDPKRU(MachineInstr &MI, MachineBasicBlock *BB, |
25043 | const X86Subtarget &Subtarget) { |
25044 | DebugLoc dl = MI.getDebugLoc(); |
25045 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
25046 | |
25047 | // insert zero to ECX |
25048 | BuildMI(*BB, MI, dl, TII->get(X86::MOV32r0), X86::ECX); |
25049 | |
25050 | // insert RDPKRU instruction |
25051 | BuildMI(*BB, MI, dl, TII->get(X86::RDPKRUr)); |
25052 | BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), MI.getOperand(0).getReg()) |
25053 | .addReg(X86::EAX); |
25054 | |
25055 | MI.eraseFromParent(); // The pseudo is gone now. |
25056 | return BB; |
25057 | } |
25058 | |
25059 | static MachineBasicBlock *emitMonitor(MachineInstr &MI, MachineBasicBlock *BB, |
25060 | const X86Subtarget &Subtarget, |
25061 | unsigned Opc) { |
25062 | DebugLoc dl = MI.getDebugLoc(); |
25063 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
25064 | // Address into RAX/EAX, other two args into ECX, EDX. |
25065 | unsigned MemOpc = Subtarget.is64Bit() ? X86::LEA64r : X86::LEA32r; |
25066 | unsigned MemReg = Subtarget.is64Bit() ? X86::RAX : X86::EAX; |
25067 | MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg); |
25068 | for (int i = 0; i < X86::AddrNumOperands; ++i) |
25069 | MIB.add(MI.getOperand(i)); |
25070 | |
25071 | unsigned ValOps = X86::AddrNumOperands; |
25072 | BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX) |
25073 | .addReg(MI.getOperand(ValOps).getReg()); |
25074 | BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX) |
25075 | .addReg(MI.getOperand(ValOps + 1).getReg()); |
25076 | |
25077 | // The instruction doesn't actually take any operands though. |
25078 | BuildMI(*BB, MI, dl, TII->get(Opc)); |
25079 | |
25080 | MI.eraseFromParent(); // The pseudo is gone now. |
25081 | return BB; |
25082 | } |
25083 | |
25084 | static MachineBasicBlock *emitClzero(MachineInstr *MI, MachineBasicBlock *BB, |
25085 | const X86Subtarget &Subtarget) { |
25086 | DebugLoc dl = MI->getDebugLoc(); |
25087 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
25088 | // Address into RAX/EAX |
25089 | unsigned MemOpc = Subtarget.is64Bit() ? X86::LEA64r : X86::LEA32r; |
25090 | unsigned MemReg = Subtarget.is64Bit() ? X86::RAX : X86::EAX; |
25091 | MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg); |
25092 | for (int i = 0; i < X86::AddrNumOperands; ++i) |
25093 | MIB.add(MI->getOperand(i)); |
25094 | |
25095 | // The instruction doesn't actually take any operands though. |
25096 | BuildMI(*BB, MI, dl, TII->get(X86::CLZEROr)); |
25097 | |
25098 | MI->eraseFromParent(); // The pseudo is gone now. |
25099 | return BB; |
25100 | } |
25101 | |
25102 | |
25103 | |
25104 | MachineBasicBlock * |
25105 | X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr &MI, |
25106 | MachineBasicBlock *MBB) const { |
25107 | // Emit va_arg instruction on X86-64. |
25108 | |
25109 | // Operands to this pseudo-instruction: |
25110 | // 0 ) Output : destination address (reg) |
25111 | // 1-5) Input : va_list address (addr, i64mem) |
25112 | // 6 ) ArgSize : Size (in bytes) of vararg type |
25113 | // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset |
25114 | // 8 ) Align : Alignment of type |
25115 | // 9 ) EFLAGS (implicit-def) |
25116 | |
25117 | assert(MI.getNumOperands() == 10 && "VAARG_64 should have 10 operands!")((MI.getNumOperands() == 10 && "VAARG_64 should have 10 operands!" ) ? static_cast<void> (0) : __assert_fail ("MI.getNumOperands() == 10 && \"VAARG_64 should have 10 operands!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 25117, __PRETTY_FUNCTION__)); |
25118 | static_assert(X86::AddrNumOperands == 5, |
25119 | "VAARG_64 assumes 5 address operands"); |
25120 | |
25121 | unsigned DestReg = MI.getOperand(0).getReg(); |
25122 | MachineOperand &Base = MI.getOperand(1); |
25123 | MachineOperand &Scale = MI.getOperand(2); |
25124 | MachineOperand &Index = MI.getOperand(3); |
25125 | MachineOperand &Disp = MI.getOperand(4); |
25126 | MachineOperand &Segment = MI.getOperand(5); |
25127 | unsigned ArgSize = MI.getOperand(6).getImm(); |
25128 | unsigned ArgMode = MI.getOperand(7).getImm(); |
25129 | unsigned Align = MI.getOperand(8).getImm(); |
25130 | |
25131 | // Memory Reference |
25132 | assert(MI.hasOneMemOperand() && "Expected VAARG_64 to have one memoperand")((MI.hasOneMemOperand() && "Expected VAARG_64 to have one memoperand" ) ? static_cast<void> (0) : __assert_fail ("MI.hasOneMemOperand() && \"Expected VAARG_64 to have one memoperand\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 25132, __PRETTY_FUNCTION__)); |
25133 | MachineInstr::mmo_iterator MMOBegin = MI.memoperands_begin(); |
25134 | MachineInstr::mmo_iterator MMOEnd = MI.memoperands_end(); |
25135 | |
25136 | // Machine Information |
25137 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
25138 | MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); |
25139 | const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64); |
25140 | const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32); |
25141 | DebugLoc DL = MI.getDebugLoc(); |
25142 | |
25143 | // struct va_list { |
25144 | // i32 gp_offset |
25145 | // i32 fp_offset |
25146 | // i64 overflow_area (address) |
25147 | // i64 reg_save_area (address) |
25148 | // } |
25149 | // sizeof(va_list) = 24 |
25150 | // alignment(va_list) = 8 |
25151 | |
25152 | unsigned TotalNumIntRegs = 6; |
25153 | unsigned TotalNumXMMRegs = 8; |
25154 | bool UseGPOffset = (ArgMode == 1); |
25155 | bool UseFPOffset = (ArgMode == 2); |
25156 | unsigned MaxOffset = TotalNumIntRegs * 8 + |
25157 | (UseFPOffset ? TotalNumXMMRegs * 16 : 0); |
25158 | |
25159 | /* Align ArgSize to a multiple of 8 */ |
25160 | unsigned ArgSizeA8 = (ArgSize + 7) & ~7; |
25161 | bool NeedsAlign = (Align > 8); |
25162 | |
25163 | MachineBasicBlock *thisMBB = MBB; |
25164 | MachineBasicBlock *overflowMBB; |
25165 | MachineBasicBlock *offsetMBB; |
25166 | MachineBasicBlock *endMBB; |
25167 | |
25168 | unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB |
25169 | unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB |
25170 | unsigned OffsetReg = 0; |
25171 | |
25172 | if (!UseGPOffset && !UseFPOffset) { |
25173 | // If we only pull from the overflow region, we don't create a branch. |
25174 | // We don't need to alter control flow. |
25175 | OffsetDestReg = 0; // unused |
25176 | OverflowDestReg = DestReg; |
25177 | |
25178 | offsetMBB = nullptr; |
25179 | overflowMBB = thisMBB; |
25180 | endMBB = thisMBB; |
25181 | } else { |
25182 | // First emit code to check if gp_offset (or fp_offset) is below the bound. |
25183 | // If so, pull the argument from reg_save_area. (branch to offsetMBB) |
25184 | // If not, pull from overflow_area. (branch to overflowMBB) |
25185 | // |
25186 | // thisMBB |
25187 | // | . |
25188 | // | . |
25189 | // offsetMBB overflowMBB |
25190 | // | . |
25191 | // | . |
25192 | // endMBB |
25193 | |
25194 | // Registers for the PHI in endMBB |
25195 | OffsetDestReg = MRI.createVirtualRegister(AddrRegClass); |
25196 | OverflowDestReg = MRI.createVirtualRegister(AddrRegClass); |
25197 | |
25198 | const BasicBlock *LLVM_BB = MBB->getBasicBlock(); |
25199 | MachineFunction *MF = MBB->getParent(); |
25200 | overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB); |
25201 | offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB); |
25202 | endMBB = MF->CreateMachineBasicBlock(LLVM_BB); |
25203 | |
25204 | MachineFunction::iterator MBBIter = ++MBB->getIterator(); |
25205 | |
25206 | // Insert the new basic blocks |
25207 | MF->insert(MBBIter, offsetMBB); |
25208 | MF->insert(MBBIter, overflowMBB); |
25209 | MF->insert(MBBIter, endMBB); |
25210 | |
25211 | // Transfer the remainder of MBB and its successor edges to endMBB. |
25212 | endMBB->splice(endMBB->begin(), thisMBB, |
25213 | std::next(MachineBasicBlock::iterator(MI)), thisMBB->end()); |
25214 | endMBB->transferSuccessorsAndUpdatePHIs(thisMBB); |
25215 | |
25216 | // Make offsetMBB and overflowMBB successors of thisMBB |
25217 | thisMBB->addSuccessor(offsetMBB); |
25218 | thisMBB->addSuccessor(overflowMBB); |
25219 | |
25220 | // endMBB is a successor of both offsetMBB and overflowMBB |
25221 | offsetMBB->addSuccessor(endMBB); |
25222 | overflowMBB->addSuccessor(endMBB); |
25223 | |
25224 | // Load the offset value into a register |
25225 | OffsetReg = MRI.createVirtualRegister(OffsetRegClass); |
25226 | BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg) |
25227 | .add(Base) |
25228 | .add(Scale) |
25229 | .add(Index) |
25230 | .addDisp(Disp, UseFPOffset ? 4 : 0) |
25231 | .add(Segment) |
25232 | .setMemRefs(MMOBegin, MMOEnd); |
25233 | |
25234 | // Check if there is enough room left to pull this argument. |
25235 | BuildMI(thisMBB, DL, TII->get(X86::CMP32ri)) |
25236 | .addReg(OffsetReg) |
25237 | .addImm(MaxOffset + 8 - ArgSizeA8); |
25238 | |
25239 | // Branch to "overflowMBB" if offset >= max |
25240 | // Fall through to "offsetMBB" otherwise |
25241 | BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE))) |
25242 | .addMBB(overflowMBB); |
25243 | } |
25244 | |
25245 | // In offsetMBB, emit code to use the reg_save_area. |
25246 | if (offsetMBB) { |
25247 | assert(OffsetReg != 0)((OffsetReg != 0) ? static_cast<void> (0) : __assert_fail ("OffsetReg != 0", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 25247, __PRETTY_FUNCTION__)); |
25248 | |
25249 | // Read the reg_save_area address. |
25250 | unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass); |
25251 | BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg) |
25252 | .add(Base) |
25253 | .add(Scale) |
25254 | .add(Index) |
25255 | .addDisp(Disp, 16) |
25256 | .add(Segment) |
25257 | .setMemRefs(MMOBegin, MMOEnd); |
25258 | |
25259 | // Zero-extend the offset |
25260 | unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass); |
25261 | BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64) |
25262 | .addImm(0) |
25263 | .addReg(OffsetReg) |
25264 | .addImm(X86::sub_32bit); |
25265 | |
25266 | // Add the offset to the reg_save_area to get the final address. |
25267 | BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg) |
25268 | .addReg(OffsetReg64) |
25269 | .addReg(RegSaveReg); |
25270 | |
25271 | // Compute the offset for the next argument |
25272 | unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass); |
25273 | BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg) |
25274 | .addReg(OffsetReg) |
25275 | .addImm(UseFPOffset ? 16 : 8); |
25276 | |
25277 | // Store it back into the va_list. |
25278 | BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr)) |
25279 | .add(Base) |
25280 | .add(Scale) |
25281 | .add(Index) |
25282 | .addDisp(Disp, UseFPOffset ? 4 : 0) |
25283 | .add(Segment) |
25284 | .addReg(NextOffsetReg) |
25285 | .setMemRefs(MMOBegin, MMOEnd); |
25286 | |
25287 | // Jump to endMBB |
25288 | BuildMI(offsetMBB, DL, TII->get(X86::JMP_1)) |
25289 | .addMBB(endMBB); |
25290 | } |
25291 | |
25292 | // |
25293 | // Emit code to use overflow area |
25294 | // |
25295 | |
25296 | // Load the overflow_area address into a register. |
25297 | unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass); |
25298 | BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg) |
25299 | .add(Base) |
25300 | .add(Scale) |
25301 | .add(Index) |
25302 | .addDisp(Disp, 8) |
25303 | .add(Segment) |
25304 | .setMemRefs(MMOBegin, MMOEnd); |
25305 | |
25306 | // If we need to align it, do so. Otherwise, just copy the address |
25307 | // to OverflowDestReg. |
25308 | if (NeedsAlign) { |
25309 | // Align the overflow address |
25310 | assert(isPowerOf2_32(Align) && "Alignment must be a power of 2")((isPowerOf2_32(Align) && "Alignment must be a power of 2" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(Align) && \"Alignment must be a power of 2\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 25310, __PRETTY_FUNCTION__)); |
25311 | unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass); |
25312 | |
25313 | // aligned_addr = (addr + (align-1)) & ~(align-1) |
25314 | BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg) |
25315 | .addReg(OverflowAddrReg) |
25316 | .addImm(Align-1); |
25317 | |
25318 | BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg) |
25319 | .addReg(TmpReg) |
25320 | .addImm(~(uint64_t)(Align-1)); |
25321 | } else { |
25322 | BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg) |
25323 | .addReg(OverflowAddrReg); |
25324 | } |
25325 | |
25326 | // Compute the next overflow address after this argument. |
25327 | // (the overflow address should be kept 8-byte aligned) |
25328 | unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass); |
25329 | BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg) |
25330 | .addReg(OverflowDestReg) |
25331 | .addImm(ArgSizeA8); |
25332 | |
25333 | // Store the new overflow address. |
25334 | BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr)) |
25335 | .add(Base) |
25336 | .add(Scale) |
25337 | .add(Index) |
25338 | .addDisp(Disp, 8) |
25339 | .add(Segment) |
25340 | .addReg(NextAddrReg) |
25341 | .setMemRefs(MMOBegin, MMOEnd); |
25342 | |
25343 | // If we branched, emit the PHI to the front of endMBB. |
25344 | if (offsetMBB) { |
25345 | BuildMI(*endMBB, endMBB->begin(), DL, |
25346 | TII->get(X86::PHI), DestReg) |
25347 | .addReg(OffsetDestReg).addMBB(offsetMBB) |
25348 | .addReg(OverflowDestReg).addMBB(overflowMBB); |
25349 | } |
25350 | |
25351 | // Erase the pseudo instruction |
25352 | MI.eraseFromParent(); |
25353 | |
25354 | return endMBB; |
25355 | } |
25356 | |
25357 | MachineBasicBlock *X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter( |
25358 | MachineInstr &MI, MachineBasicBlock *MBB) const { |
25359 | // Emit code to save XMM registers to the stack. The ABI says that the |
25360 | // number of registers to save is given in %al, so it's theoretically |
25361 | // possible to do an indirect jump trick to avoid saving all of them, |
25362 | // however this code takes a simpler approach and just executes all |
25363 | // of the stores if %al is non-zero. It's less code, and it's probably |
25364 | // easier on the hardware branch predictor, and stores aren't all that |
25365 | // expensive anyway. |
25366 | |
25367 | // Create the new basic blocks. One block contains all the XMM stores, |
25368 | // and one block is the final destination regardless of whether any |
25369 | // stores were performed. |
25370 | const BasicBlock *LLVM_BB = MBB->getBasicBlock(); |
25371 | MachineFunction *F = MBB->getParent(); |
25372 | MachineFunction::iterator MBBIter = ++MBB->getIterator(); |
25373 | MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB); |
25374 | MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB); |
25375 | F->insert(MBBIter, XMMSaveMBB); |
25376 | F->insert(MBBIter, EndMBB); |
25377 | |
25378 | // Transfer the remainder of MBB and its successor edges to EndMBB. |
25379 | EndMBB->splice(EndMBB->begin(), MBB, |
25380 | std::next(MachineBasicBlock::iterator(MI)), MBB->end()); |
25381 | EndMBB->transferSuccessorsAndUpdatePHIs(MBB); |
25382 | |
25383 | // The original block will now fall through to the XMM save block. |
25384 | MBB->addSuccessor(XMMSaveMBB); |
25385 | // The XMMSaveMBB will fall through to the end block. |
25386 | XMMSaveMBB->addSuccessor(EndMBB); |
25387 | |
25388 | // Now add the instructions. |
25389 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
25390 | DebugLoc DL = MI.getDebugLoc(); |
25391 | |
25392 | unsigned CountReg = MI.getOperand(0).getReg(); |
25393 | int64_t RegSaveFrameIndex = MI.getOperand(1).getImm(); |
25394 | int64_t VarArgsFPOffset = MI.getOperand(2).getImm(); |
25395 | |
25396 | if (!Subtarget.isCallingConvWin64(F->getFunction()->getCallingConv())) { |
25397 | // If %al is 0, branch around the XMM save block. |
25398 | BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg); |
25399 | BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB); |
25400 | MBB->addSuccessor(EndMBB); |
25401 | } |
25402 | |
25403 | // Make sure the last operand is EFLAGS, which gets clobbered by the branch |
25404 | // that was just emitted, but clearly shouldn't be "saved". |
25405 | assert((MI.getNumOperands() <= 3 ||(((MI.getNumOperands() <= 3 || !MI.getOperand(MI.getNumOperands () - 1).isReg() || MI.getOperand(MI.getNumOperands() - 1).getReg () == X86::EFLAGS) && "Expected last argument to be EFLAGS" ) ? static_cast<void> (0) : __assert_fail ("(MI.getNumOperands() <= 3 || !MI.getOperand(MI.getNumOperands() - 1).isReg() || MI.getOperand(MI.getNumOperands() - 1).getReg() == X86::EFLAGS) && \"Expected last argument to be EFLAGS\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 25408, __PRETTY_FUNCTION__)) |
25406 | !MI.getOperand(MI.getNumOperands() - 1).isReg() ||(((MI.getNumOperands() <= 3 || !MI.getOperand(MI.getNumOperands () - 1).isReg() || MI.getOperand(MI.getNumOperands() - 1).getReg () == X86::EFLAGS) && "Expected last argument to be EFLAGS" ) ? static_cast<void> (0) : __assert_fail ("(MI.getNumOperands() <= 3 || !MI.getOperand(MI.getNumOperands() - 1).isReg() || MI.getOperand(MI.getNumOperands() - 1).getReg() == X86::EFLAGS) && \"Expected last argument to be EFLAGS\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 25408, __PRETTY_FUNCTION__)) |
25407 | MI.getOperand(MI.getNumOperands() - 1).getReg() == X86::EFLAGS) &&(((MI.getNumOperands() <= 3 || !MI.getOperand(MI.getNumOperands () - 1).isReg() || MI.getOperand(MI.getNumOperands() - 1).getReg () == X86::EFLAGS) && "Expected last argument to be EFLAGS" ) ? static_cast<void> (0) : __assert_fail ("(MI.getNumOperands() <= 3 || !MI.getOperand(MI.getNumOperands() - 1).isReg() || MI.getOperand(MI.getNumOperands() - 1).getReg() == X86::EFLAGS) && \"Expected last argument to be EFLAGS\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 25408, __PRETTY_FUNCTION__)) |
25408 | "Expected last argument to be EFLAGS")(((MI.getNumOperands() <= 3 || !MI.getOperand(MI.getNumOperands () - 1).isReg() || MI.getOperand(MI.getNumOperands() - 1).getReg () == X86::EFLAGS) && "Expected last argument to be EFLAGS" ) ? static_cast<void> (0) : __assert_fail ("(MI.getNumOperands() <= 3 || !MI.getOperand(MI.getNumOperands() - 1).isReg() || MI.getOperand(MI.getNumOperands() - 1).getReg() == X86::EFLAGS) && \"Expected last argument to be EFLAGS\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 25408, __PRETTY_FUNCTION__)); |
25409 | unsigned MOVOpc = Subtarget.hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr; |
25410 | // In the XMM save block, save all the XMM argument registers. |
25411 | for (int i = 3, e = MI.getNumOperands() - 1; i != e; ++i) { |
25412 | int64_t Offset = (i - 3) * 16 + VarArgsFPOffset; |
25413 | MachineMemOperand *MMO = F->getMachineMemOperand( |
25414 | MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset), |
25415 | MachineMemOperand::MOStore, |
25416 | /*Size=*/16, /*Align=*/16); |
25417 | BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc)) |
25418 | .addFrameIndex(RegSaveFrameIndex) |
25419 | .addImm(/*Scale=*/1) |
25420 | .addReg(/*IndexReg=*/0) |
25421 | .addImm(/*Disp=*/Offset) |
25422 | .addReg(/*Segment=*/0) |
25423 | .addReg(MI.getOperand(i).getReg()) |
25424 | .addMemOperand(MMO); |
25425 | } |
25426 | |
25427 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
25428 | |
25429 | return EndMBB; |
25430 | } |
25431 | |
25432 | // The EFLAGS operand of SelectItr might be missing a kill marker |
25433 | // because there were multiple uses of EFLAGS, and ISel didn't know |
25434 | // which to mark. Figure out whether SelectItr should have had a |
25435 | // kill marker, and set it if it should. Returns the correct kill |
25436 | // marker value. |
25437 | static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr, |
25438 | MachineBasicBlock* BB, |
25439 | const TargetRegisterInfo* TRI) { |
25440 | // Scan forward through BB for a use/def of EFLAGS. |
25441 | MachineBasicBlock::iterator miI(std::next(SelectItr)); |
25442 | for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) { |
25443 | const MachineInstr& mi = *miI; |
25444 | if (mi.readsRegister(X86::EFLAGS)) |
25445 | return false; |
25446 | if (mi.definesRegister(X86::EFLAGS)) |
25447 | break; // Should have kill-flag - update below. |
25448 | } |
25449 | |
25450 | // If we hit the end of the block, check whether EFLAGS is live into a |
25451 | // successor. |
25452 | if (miI == BB->end()) { |
25453 | for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(), |
25454 | sEnd = BB->succ_end(); |
25455 | sItr != sEnd; ++sItr) { |
25456 | MachineBasicBlock* succ = *sItr; |
25457 | if (succ->isLiveIn(X86::EFLAGS)) |
25458 | return false; |
25459 | } |
25460 | } |
25461 | |
25462 | // We found a def, or hit the end of the basic block and EFLAGS wasn't live |
25463 | // out. SelectMI should have a kill flag on EFLAGS. |
25464 | SelectItr->addRegisterKilled(X86::EFLAGS, TRI); |
25465 | return true; |
25466 | } |
25467 | |
25468 | // Return true if it is OK for this CMOV pseudo-opcode to be cascaded |
25469 | // together with other CMOV pseudo-opcodes into a single basic-block with |
25470 | // conditional jump around it. |
25471 | static bool isCMOVPseudo(MachineInstr &MI) { |
25472 | switch (MI.getOpcode()) { |
25473 | case X86::CMOV_FR32: |
25474 | case X86::CMOV_FR64: |
25475 | case X86::CMOV_GR8: |
25476 | case X86::CMOV_GR16: |
25477 | case X86::CMOV_GR32: |
25478 | case X86::CMOV_RFP32: |
25479 | case X86::CMOV_RFP64: |
25480 | case X86::CMOV_RFP80: |
25481 | case X86::CMOV_V2F64: |
25482 | case X86::CMOV_V2I64: |
25483 | case X86::CMOV_V4F32: |
25484 | case X86::CMOV_V4F64: |
25485 | case X86::CMOV_V4I64: |
25486 | case X86::CMOV_V16F32: |
25487 | case X86::CMOV_V8F32: |
25488 | case X86::CMOV_V8F64: |
25489 | case X86::CMOV_V8I64: |
25490 | case X86::CMOV_V8I1: |
25491 | case X86::CMOV_V16I1: |
25492 | case X86::CMOV_V32I1: |
25493 | case X86::CMOV_V64I1: |
25494 | return true; |
25495 | |
25496 | default: |
25497 | return false; |
25498 | } |
25499 | } |
25500 | |
25501 | MachineBasicBlock * |
25502 | X86TargetLowering::EmitLoweredSelect(MachineInstr &MI, |
25503 | MachineBasicBlock *BB) const { |
25504 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
25505 | DebugLoc DL = MI.getDebugLoc(); |
25506 | |
25507 | // To "insert" a SELECT_CC instruction, we actually have to insert the |
25508 | // diamond control-flow pattern. The incoming instruction knows the |
25509 | // destination vreg to set, the condition code register to branch on, the |
25510 | // true/false values to select between, and a branch opcode to use. |
25511 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
25512 | MachineFunction::iterator It = ++BB->getIterator(); |
25513 | |
25514 | // thisMBB: |
25515 | // ... |
25516 | // TrueVal = ... |
25517 | // cmpTY ccX, r1, r2 |
25518 | // bCC copy1MBB |
25519 | // fallthrough --> copy0MBB |
25520 | MachineBasicBlock *thisMBB = BB; |
25521 | MachineFunction *F = BB->getParent(); |
25522 | |
25523 | // This code lowers all pseudo-CMOV instructions. Generally it lowers these |
25524 | // as described above, by inserting a BB, and then making a PHI at the join |
25525 | // point to select the true and false operands of the CMOV in the PHI. |
25526 | // |
25527 | // The code also handles two different cases of multiple CMOV opcodes |
25528 | // in a row. |
25529 | // |
25530 | // Case 1: |
25531 | // In this case, there are multiple CMOVs in a row, all which are based on |
25532 | // the same condition setting (or the exact opposite condition setting). |
25533 | // In this case we can lower all the CMOVs using a single inserted BB, and |
25534 | // then make a number of PHIs at the join point to model the CMOVs. The only |
25535 | // trickiness here, is that in a case like: |
25536 | // |
25537 | // t2 = CMOV cond1 t1, f1 |
25538 | // t3 = CMOV cond1 t2, f2 |
25539 | // |
25540 | // when rewriting this into PHIs, we have to perform some renaming on the |
25541 | // temps since you cannot have a PHI operand refer to a PHI result earlier |
25542 | // in the same block. The "simple" but wrong lowering would be: |
25543 | // |
25544 | // t2 = PHI t1(BB1), f1(BB2) |
25545 | // t3 = PHI t2(BB1), f2(BB2) |
25546 | // |
25547 | // but clearly t2 is not defined in BB1, so that is incorrect. The proper |
25548 | // renaming is to note that on the path through BB1, t2 is really just a |
25549 | // copy of t1, and do that renaming, properly generating: |
25550 | // |
25551 | // t2 = PHI t1(BB1), f1(BB2) |
25552 | // t3 = PHI t1(BB1), f2(BB2) |
25553 | // |
25554 | // Case 2, we lower cascaded CMOVs such as |
25555 | // |
25556 | // (CMOV (CMOV F, T, cc1), T, cc2) |
25557 | // |
25558 | // to two successive branches. For that, we look for another CMOV as the |
25559 | // following instruction. |
25560 | // |
25561 | // Without this, we would add a PHI between the two jumps, which ends up |
25562 | // creating a few copies all around. For instance, for |
25563 | // |
25564 | // (sitofp (zext (fcmp une))) |
25565 | // |
25566 | // we would generate: |
25567 | // |
25568 | // ucomiss %xmm1, %xmm0 |
25569 | // movss <1.0f>, %xmm0 |
25570 | // movaps %xmm0, %xmm1 |
25571 | // jne .LBB5_2 |
25572 | // xorps %xmm1, %xmm1 |
25573 | // .LBB5_2: |
25574 | // jp .LBB5_4 |
25575 | // movaps %xmm1, %xmm0 |
25576 | // .LBB5_4: |
25577 | // retq |
25578 | // |
25579 | // because this custom-inserter would have generated: |
25580 | // |
25581 | // A |
25582 | // | \ |
25583 | // | B |
25584 | // | / |
25585 | // C |
25586 | // | \ |
25587 | // | D |
25588 | // | / |
25589 | // E |
25590 | // |
25591 | // A: X = ...; Y = ... |
25592 | // B: empty |
25593 | // C: Z = PHI [X, A], [Y, B] |
25594 | // D: empty |
25595 | // E: PHI [X, C], [Z, D] |
25596 | // |
25597 | // If we lower both CMOVs in a single step, we can instead generate: |
25598 | // |
25599 | // A |
25600 | // | \ |
25601 | // | C |
25602 | // | /| |
25603 | // |/ | |
25604 | // | | |
25605 | // | D |
25606 | // | / |
25607 | // E |
25608 | // |
25609 | // A: X = ...; Y = ... |
25610 | // D: empty |
25611 | // E: PHI [X, A], [X, C], [Y, D] |
25612 | // |
25613 | // Which, in our sitofp/fcmp example, gives us something like: |
25614 | // |
25615 | // ucomiss %xmm1, %xmm0 |
25616 | // movss <1.0f>, %xmm0 |
25617 | // jne .LBB5_4 |
25618 | // jp .LBB5_4 |
25619 | // xorps %xmm0, %xmm0 |
25620 | // .LBB5_4: |
25621 | // retq |
25622 | // |
25623 | MachineInstr *CascadedCMOV = nullptr; |
25624 | MachineInstr *LastCMOV = &MI; |
25625 | X86::CondCode CC = X86::CondCode(MI.getOperand(3).getImm()); |
25626 | X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC); |
25627 | MachineBasicBlock::iterator NextMIIt = |
25628 | std::next(MachineBasicBlock::iterator(MI)); |
25629 | |
25630 | // Check for case 1, where there are multiple CMOVs with the same condition |
25631 | // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the |
25632 | // number of jumps the most. |
25633 | |
25634 | if (isCMOVPseudo(MI)) { |
25635 | // See if we have a string of CMOVS with the same condition. |
25636 | while (NextMIIt != BB->end() && isCMOVPseudo(*NextMIIt) && |
25637 | (NextMIIt->getOperand(3).getImm() == CC || |
25638 | NextMIIt->getOperand(3).getImm() == OppCC)) { |
25639 | LastCMOV = &*NextMIIt; |
25640 | ++NextMIIt; |
25641 | } |
25642 | } |
25643 | |
25644 | // This checks for case 2, but only do this if we didn't already find |
25645 | // case 1, as indicated by LastCMOV == MI. |
25646 | if (LastCMOV == &MI && NextMIIt != BB->end() && |
25647 | NextMIIt->getOpcode() == MI.getOpcode() && |
25648 | NextMIIt->getOperand(2).getReg() == MI.getOperand(2).getReg() && |
25649 | NextMIIt->getOperand(1).getReg() == MI.getOperand(0).getReg() && |
25650 | NextMIIt->getOperand(1).isKill()) { |
25651 | CascadedCMOV = &*NextMIIt; |
25652 | } |
25653 | |
25654 | MachineBasicBlock *jcc1MBB = nullptr; |
25655 | |
25656 | // If we have a cascaded CMOV, we lower it to two successive branches to |
25657 | // the same block. EFLAGS is used by both, so mark it as live in the second. |
25658 | if (CascadedCMOV) { |
25659 | jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB); |
25660 | F->insert(It, jcc1MBB); |
25661 | jcc1MBB->addLiveIn(X86::EFLAGS); |
25662 | } |
25663 | |
25664 | MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); |
25665 | MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); |
25666 | F->insert(It, copy0MBB); |
25667 | F->insert(It, sinkMBB); |
25668 | |
25669 | // If the EFLAGS register isn't dead in the terminator, then claim that it's |
25670 | // live into the sink and copy blocks. |
25671 | const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); |
25672 | |
25673 | MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV; |
25674 | if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) && |
25675 | !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) { |
25676 | copy0MBB->addLiveIn(X86::EFLAGS); |
25677 | sinkMBB->addLiveIn(X86::EFLAGS); |
25678 | } |
25679 | |
25680 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
25681 | sinkMBB->splice(sinkMBB->begin(), BB, |
25682 | std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end()); |
25683 | sinkMBB->transferSuccessorsAndUpdatePHIs(BB); |
25684 | |
25685 | // Add the true and fallthrough blocks as its successors. |
25686 | if (CascadedCMOV) { |
25687 | // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV. |
25688 | BB->addSuccessor(jcc1MBB); |
25689 | |
25690 | // In that case, jcc1MBB will itself fallthrough the copy0MBB, and |
25691 | // jump to the sinkMBB. |
25692 | jcc1MBB->addSuccessor(copy0MBB); |
25693 | jcc1MBB->addSuccessor(sinkMBB); |
25694 | } else { |
25695 | BB->addSuccessor(copy0MBB); |
25696 | } |
25697 | |
25698 | // The true block target of the first (or only) branch is always sinkMBB. |
25699 | BB->addSuccessor(sinkMBB); |
25700 | |
25701 | // Create the conditional branch instruction. |
25702 | unsigned Opc = X86::GetCondBranchFromCond(CC); |
25703 | BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB); |
25704 | |
25705 | if (CascadedCMOV) { |
25706 | unsigned Opc2 = X86::GetCondBranchFromCond( |
25707 | (X86::CondCode)CascadedCMOV->getOperand(3).getImm()); |
25708 | BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB); |
25709 | } |
25710 | |
25711 | // copy0MBB: |
25712 | // %FalseValue = ... |
25713 | // # fallthrough to sinkMBB |
25714 | copy0MBB->addSuccessor(sinkMBB); |
25715 | |
25716 | // sinkMBB: |
25717 | // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] |
25718 | // ... |
25719 | MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI); |
25720 | MachineBasicBlock::iterator MIItEnd = |
25721 | std::next(MachineBasicBlock::iterator(LastCMOV)); |
25722 | MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin(); |
25723 | DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable; |
25724 | MachineInstrBuilder MIB; |
25725 | |
25726 | // As we are creating the PHIs, we have to be careful if there is more than |
25727 | // one. Later CMOVs may reference the results of earlier CMOVs, but later |
25728 | // PHIs have to reference the individual true/false inputs from earlier PHIs. |
25729 | // That also means that PHI construction must work forward from earlier to |
25730 | // later, and that the code must maintain a mapping from earlier PHI's |
25731 | // destination registers, and the registers that went into the PHI. |
25732 | |
25733 | for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) { |
25734 | unsigned DestReg = MIIt->getOperand(0).getReg(); |
25735 | unsigned Op1Reg = MIIt->getOperand(1).getReg(); |
25736 | unsigned Op2Reg = MIIt->getOperand(2).getReg(); |
25737 | |
25738 | // If this CMOV we are generating is the opposite condition from |
25739 | // the jump we generated, then we have to swap the operands for the |
25740 | // PHI that is going to be generated. |
25741 | if (MIIt->getOperand(3).getImm() == OppCC) |
25742 | std::swap(Op1Reg, Op2Reg); |
25743 | |
25744 | if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end()) |
25745 | Op1Reg = RegRewriteTable[Op1Reg].first; |
25746 | |
25747 | if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end()) |
25748 | Op2Reg = RegRewriteTable[Op2Reg].second; |
25749 | |
25750 | MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL, |
25751 | TII->get(X86::PHI), DestReg) |
25752 | .addReg(Op1Reg).addMBB(copy0MBB) |
25753 | .addReg(Op2Reg).addMBB(thisMBB); |
25754 | |
25755 | // Add this PHI to the rewrite table. |
25756 | RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg); |
25757 | } |
25758 | |
25759 | // If we have a cascaded CMOV, the second Jcc provides the same incoming |
25760 | // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes). |
25761 | if (CascadedCMOV) { |
25762 | MIB.addReg(MI.getOperand(2).getReg()).addMBB(jcc1MBB); |
25763 | // Copy the PHI result to the register defined by the second CMOV. |
25764 | BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())), |
25765 | DL, TII->get(TargetOpcode::COPY), |
25766 | CascadedCMOV->getOperand(0).getReg()) |
25767 | .addReg(MI.getOperand(0).getReg()); |
25768 | CascadedCMOV->eraseFromParent(); |
25769 | } |
25770 | |
25771 | // Now remove the CMOV(s). |
25772 | for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ) |
25773 | (MIIt++)->eraseFromParent(); |
25774 | |
25775 | return sinkMBB; |
25776 | } |
25777 | |
25778 | MachineBasicBlock * |
25779 | X86TargetLowering::EmitLoweredAtomicFP(MachineInstr &MI, |
25780 | MachineBasicBlock *BB) const { |
25781 | // Combine the following atomic floating-point modification pattern: |
25782 | // a.store(reg OP a.load(acquire), release) |
25783 | // Transform them into: |
25784 | // OPss (%gpr), %xmm |
25785 | // movss %xmm, (%gpr) |
25786 | // Or sd equivalent for 64-bit operations. |
25787 | unsigned MOp, FOp; |
25788 | switch (MI.getOpcode()) { |
25789 | default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP")::llvm::llvm_unreachable_internal("unexpected instr type for EmitLoweredAtomicFP" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 25789); |
25790 | case X86::RELEASE_FADD32mr: |
25791 | FOp = X86::ADDSSrm; |
25792 | MOp = X86::MOVSSmr; |
25793 | break; |
25794 | case X86::RELEASE_FADD64mr: |
25795 | FOp = X86::ADDSDrm; |
25796 | MOp = X86::MOVSDmr; |
25797 | break; |
25798 | } |
25799 | const X86InstrInfo *TII = Subtarget.getInstrInfo(); |
25800 | DebugLoc DL = MI.getDebugLoc(); |
25801 | MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); |
25802 | unsigned ValOpIdx = X86::AddrNumOperands; |
25803 | unsigned VSrc = MI.getOperand(ValOpIdx).getReg(); |
25804 | MachineInstrBuilder MIB = |
25805 | BuildMI(*BB, MI, DL, TII->get(FOp), |
25806 | MRI.createVirtualRegister(MRI.getRegClass(VSrc))) |
25807 | .addReg(VSrc); |
25808 | for (int i = 0; i < X86::AddrNumOperands; ++i) { |
25809 | MachineOperand &Operand = MI.getOperand(i); |
25810 | // Clear any kill flags on register operands as we'll create a second |
25811 | // instruction using the same address operands. |
25812 | if (Operand.isReg()) |
25813 | Operand.setIsKill(false); |
25814 | MIB.add(Operand); |
25815 | } |
25816 | MachineInstr *FOpMI = MIB; |
25817 | MIB = BuildMI(*BB, MI, DL, TII->get(MOp)); |
25818 | for (int i = 0; i < X86::AddrNumOperands; ++i) |
25819 | MIB.add(MI.getOperand(i)); |
25820 | MIB.addReg(FOpMI->getOperand(0).getReg(), RegState::Kill); |
25821 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
25822 | return BB; |
25823 | } |
25824 | |
25825 | MachineBasicBlock * |
25826 | X86TargetLowering::EmitLoweredSegAlloca(MachineInstr &MI, |
25827 | MachineBasicBlock *BB) const { |
25828 | MachineFunction *MF = BB->getParent(); |
25829 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
25830 | DebugLoc DL = MI.getDebugLoc(); |
25831 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
25832 | |
25833 | assert(MF->shouldSplitStack())((MF->shouldSplitStack()) ? static_cast<void> (0) : __assert_fail ("MF->shouldSplitStack()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 25833, __PRETTY_FUNCTION__)); |
25834 | |
25835 | const bool Is64Bit = Subtarget.is64Bit(); |
25836 | const bool IsLP64 = Subtarget.isTarget64BitLP64(); |
25837 | |
25838 | const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS; |
25839 | const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30; |
25840 | |
25841 | // BB: |
25842 | // ... [Till the alloca] |
25843 | // If stacklet is not large enough, jump to mallocMBB |
25844 | // |
25845 | // bumpMBB: |
25846 | // Allocate by subtracting from RSP |
25847 | // Jump to continueMBB |
25848 | // |
25849 | // mallocMBB: |
25850 | // Allocate by call to runtime |
25851 | // |
25852 | // continueMBB: |
25853 | // ... |
25854 | // [rest of original BB] |
25855 | // |
25856 | |
25857 | MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB); |
25858 | MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB); |
25859 | MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB); |
25860 | |
25861 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
25862 | const TargetRegisterClass *AddrRegClass = |
25863 | getRegClassFor(getPointerTy(MF->getDataLayout())); |
25864 | |
25865 | unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass), |
25866 | bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass), |
25867 | tmpSPVReg = MRI.createVirtualRegister(AddrRegClass), |
25868 | SPLimitVReg = MRI.createVirtualRegister(AddrRegClass), |
25869 | sizeVReg = MI.getOperand(1).getReg(), |
25870 | physSPReg = |
25871 | IsLP64 || Subtarget.isTargetNaCl64() ? X86::RSP : X86::ESP; |
25872 | |
25873 | MachineFunction::iterator MBBIter = ++BB->getIterator(); |
25874 | |
25875 | MF->insert(MBBIter, bumpMBB); |
25876 | MF->insert(MBBIter, mallocMBB); |
25877 | MF->insert(MBBIter, continueMBB); |
25878 | |
25879 | continueMBB->splice(continueMBB->begin(), BB, |
25880 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); |
25881 | continueMBB->transferSuccessorsAndUpdatePHIs(BB); |
25882 | |
25883 | // Add code to the main basic block to check if the stack limit has been hit, |
25884 | // and if so, jump to mallocMBB otherwise to bumpMBB. |
25885 | BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg); |
25886 | BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg) |
25887 | .addReg(tmpSPVReg).addReg(sizeVReg); |
25888 | BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr)) |
25889 | .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg) |
25890 | .addReg(SPLimitVReg); |
25891 | BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB); |
25892 | |
25893 | // bumpMBB simply decreases the stack pointer, since we know the current |
25894 | // stacklet has enough space. |
25895 | BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg) |
25896 | .addReg(SPLimitVReg); |
25897 | BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg) |
25898 | .addReg(SPLimitVReg); |
25899 | BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB); |
25900 | |
25901 | // Calls into a routine in libgcc to allocate more space from the heap. |
25902 | const uint32_t *RegMask = |
25903 | Subtarget.getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C); |
25904 | if (IsLP64) { |
25905 | BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI) |
25906 | .addReg(sizeVReg); |
25907 | BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32)) |
25908 | .addExternalSymbol("__morestack_allocate_stack_space") |
25909 | .addRegMask(RegMask) |
25910 | .addReg(X86::RDI, RegState::Implicit) |
25911 | .addReg(X86::RAX, RegState::ImplicitDefine); |
25912 | } else if (Is64Bit) { |
25913 | BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI) |
25914 | .addReg(sizeVReg); |
25915 | BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32)) |
25916 | .addExternalSymbol("__morestack_allocate_stack_space") |
25917 | .addRegMask(RegMask) |
25918 | .addReg(X86::EDI, RegState::Implicit) |
25919 | .addReg(X86::EAX, RegState::ImplicitDefine); |
25920 | } else { |
25921 | BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg) |
25922 | .addImm(12); |
25923 | BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg); |
25924 | BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32)) |
25925 | .addExternalSymbol("__morestack_allocate_stack_space") |
25926 | .addRegMask(RegMask) |
25927 | .addReg(X86::EAX, RegState::ImplicitDefine); |
25928 | } |
25929 | |
25930 | if (!Is64Bit) |
25931 | BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg) |
25932 | .addImm(16); |
25933 | |
25934 | BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg) |
25935 | .addReg(IsLP64 ? X86::RAX : X86::EAX); |
25936 | BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB); |
25937 | |
25938 | // Set up the CFG correctly. |
25939 | BB->addSuccessor(bumpMBB); |
25940 | BB->addSuccessor(mallocMBB); |
25941 | mallocMBB->addSuccessor(continueMBB); |
25942 | bumpMBB->addSuccessor(continueMBB); |
25943 | |
25944 | // Take care of the PHI nodes. |
25945 | BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI), |
25946 | MI.getOperand(0).getReg()) |
25947 | .addReg(mallocPtrVReg) |
25948 | .addMBB(mallocMBB) |
25949 | .addReg(bumpSPPtrVReg) |
25950 | .addMBB(bumpMBB); |
25951 | |
25952 | // Delete the original pseudo instruction. |
25953 | MI.eraseFromParent(); |
25954 | |
25955 | // And we're done. |
25956 | return continueMBB; |
25957 | } |
25958 | |
25959 | MachineBasicBlock * |
25960 | X86TargetLowering::EmitLoweredCatchRet(MachineInstr &MI, |
25961 | MachineBasicBlock *BB) const { |
25962 | MachineFunction *MF = BB->getParent(); |
25963 | const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); |
25964 | MachineBasicBlock *TargetMBB = MI.getOperand(0).getMBB(); |
25965 | DebugLoc DL = MI.getDebugLoc(); |
25966 | |
25967 | assert(!isAsynchronousEHPersonality(((!isAsynchronousEHPersonality( classifyEHPersonality(MF-> getFunction()->getPersonalityFn())) && "SEH does not use catchret!" ) ? static_cast<void> (0) : __assert_fail ("!isAsynchronousEHPersonality( classifyEHPersonality(MF->getFunction()->getPersonalityFn())) && \"SEH does not use catchret!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 25969, __PRETTY_FUNCTION__)) |
25968 | classifyEHPersonality(MF->getFunction()->getPersonalityFn())) &&((!isAsynchronousEHPersonality( classifyEHPersonality(MF-> getFunction()->getPersonalityFn())) && "SEH does not use catchret!" ) ? static_cast<void> (0) : __assert_fail ("!isAsynchronousEHPersonality( classifyEHPersonality(MF->getFunction()->getPersonalityFn())) && \"SEH does not use catchret!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 25969, __PRETTY_FUNCTION__)) |
25969 | "SEH does not use catchret!")((!isAsynchronousEHPersonality( classifyEHPersonality(MF-> getFunction()->getPersonalityFn())) && "SEH does not use catchret!" ) ? static_cast<void> (0) : __assert_fail ("!isAsynchronousEHPersonality( classifyEHPersonality(MF->getFunction()->getPersonalityFn())) && \"SEH does not use catchret!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 25969, __PRETTY_FUNCTION__)); |
25970 | |
25971 | // Only 32-bit EH needs to worry about manually restoring stack pointers. |
25972 | if (!Subtarget.is32Bit()) |
25973 | return BB; |
25974 | |
25975 | // C++ EH creates a new target block to hold the restore code, and wires up |
25976 | // the new block to the return destination with a normal JMP_4. |
25977 | MachineBasicBlock *RestoreMBB = |
25978 | MF->CreateMachineBasicBlock(BB->getBasicBlock()); |
25979 | assert(BB->succ_size() == 1)((BB->succ_size() == 1) ? static_cast<void> (0) : __assert_fail ("BB->succ_size() == 1", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 25979, __PRETTY_FUNCTION__)); |
25980 | MF->insert(std::next(BB->getIterator()), RestoreMBB); |
25981 | RestoreMBB->transferSuccessorsAndUpdatePHIs(BB); |
25982 | BB->addSuccessor(RestoreMBB); |
25983 | MI.getOperand(0).setMBB(RestoreMBB); |
25984 | |
25985 | auto RestoreMBBI = RestoreMBB->begin(); |
25986 | BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::EH_RESTORE)); |
25987 | BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::JMP_4)).addMBB(TargetMBB); |
25988 | return BB; |
25989 | } |
25990 | |
25991 | MachineBasicBlock * |
25992 | X86TargetLowering::EmitLoweredCatchPad(MachineInstr &MI, |
25993 | MachineBasicBlock *BB) const { |
25994 | MachineFunction *MF = BB->getParent(); |
25995 | const Constant *PerFn = MF->getFunction()->getPersonalityFn(); |
25996 | bool IsSEH = isAsynchronousEHPersonality(classifyEHPersonality(PerFn)); |
25997 | // Only 32-bit SEH requires special handling for catchpad. |
25998 | if (IsSEH && Subtarget.is32Bit()) { |
25999 | const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); |
26000 | DebugLoc DL = MI.getDebugLoc(); |
26001 | BuildMI(*BB, MI, DL, TII.get(X86::EH_RESTORE)); |
26002 | } |
26003 | MI.eraseFromParent(); |
26004 | return BB; |
26005 | } |
26006 | |
26007 | MachineBasicBlock * |
26008 | X86TargetLowering::EmitLoweredTLSAddr(MachineInstr &MI, |
26009 | MachineBasicBlock *BB) const { |
26010 | // So, here we replace TLSADDR with the sequence: |
26011 | // adjust_stackdown -> TLSADDR -> adjust_stackup. |
26012 | // We need this because TLSADDR is lowered into calls |
26013 | // inside MC, therefore without the two markers shrink-wrapping |
26014 | // may push the prologue/epilogue pass them. |
26015 | const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); |
26016 | DebugLoc DL = MI.getDebugLoc(); |
26017 | MachineFunction &MF = *BB->getParent(); |
26018 | |
26019 | // Emit CALLSEQ_START right before the instruction. |
26020 | unsigned AdjStackDown = TII.getCallFrameSetupOpcode(); |
26021 | MachineInstrBuilder CallseqStart = |
26022 | BuildMI(MF, DL, TII.get(AdjStackDown)).addImm(0).addImm(0).addImm(0); |
26023 | BB->insert(MachineBasicBlock::iterator(MI), CallseqStart); |
26024 | |
26025 | // Emit CALLSEQ_END right after the instruction. |
26026 | // We don't call erase from parent because we want to keep the |
26027 | // original instruction around. |
26028 | unsigned AdjStackUp = TII.getCallFrameDestroyOpcode(); |
26029 | MachineInstrBuilder CallseqEnd = |
26030 | BuildMI(MF, DL, TII.get(AdjStackUp)).addImm(0).addImm(0); |
26031 | BB->insertAfter(MachineBasicBlock::iterator(MI), CallseqEnd); |
26032 | |
26033 | return BB; |
26034 | } |
26035 | |
26036 | MachineBasicBlock * |
26037 | X86TargetLowering::EmitLoweredTLSCall(MachineInstr &MI, |
26038 | MachineBasicBlock *BB) const { |
26039 | // This is pretty easy. We're taking the value that we received from |
26040 | // our load from the relocation, sticking it in either RDI (x86-64) |
26041 | // or EAX and doing an indirect call. The return value will then |
26042 | // be in the normal return register. |
26043 | MachineFunction *F = BB->getParent(); |
26044 | const X86InstrInfo *TII = Subtarget.getInstrInfo(); |
26045 | DebugLoc DL = MI.getDebugLoc(); |
26046 | |
26047 | assert(Subtarget.isTargetDarwin() && "Darwin only instr emitted?")((Subtarget.isTargetDarwin() && "Darwin only instr emitted?" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.isTargetDarwin() && \"Darwin only instr emitted?\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26047, __PRETTY_FUNCTION__)); |
26048 | assert(MI.getOperand(3).isGlobal() && "This should be a global")((MI.getOperand(3).isGlobal() && "This should be a global" ) ? static_cast<void> (0) : __assert_fail ("MI.getOperand(3).isGlobal() && \"This should be a global\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26048, __PRETTY_FUNCTION__)); |
26049 | |
26050 | // Get a register mask for the lowered call. |
26051 | // FIXME: The 32-bit calls have non-standard calling conventions. Use a |
26052 | // proper register mask. |
26053 | const uint32_t *RegMask = |
26054 | Subtarget.is64Bit() ? |
26055 | Subtarget.getRegisterInfo()->getDarwinTLSCallPreservedMask() : |
26056 | Subtarget.getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C); |
26057 | if (Subtarget.is64Bit()) { |
26058 | MachineInstrBuilder MIB = |
26059 | BuildMI(*BB, MI, DL, TII->get(X86::MOV64rm), X86::RDI) |
26060 | .addReg(X86::RIP) |
26061 | .addImm(0) |
26062 | .addReg(0) |
26063 | .addGlobalAddress(MI.getOperand(3).getGlobal(), 0, |
26064 | MI.getOperand(3).getTargetFlags()) |
26065 | .addReg(0); |
26066 | MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m)); |
26067 | addDirectMem(MIB, X86::RDI); |
26068 | MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask); |
26069 | } else if (!isPositionIndependent()) { |
26070 | MachineInstrBuilder MIB = |
26071 | BuildMI(*BB, MI, DL, TII->get(X86::MOV32rm), X86::EAX) |
26072 | .addReg(0) |
26073 | .addImm(0) |
26074 | .addReg(0) |
26075 | .addGlobalAddress(MI.getOperand(3).getGlobal(), 0, |
26076 | MI.getOperand(3).getTargetFlags()) |
26077 | .addReg(0); |
26078 | MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m)); |
26079 | addDirectMem(MIB, X86::EAX); |
26080 | MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask); |
26081 | } else { |
26082 | MachineInstrBuilder MIB = |
26083 | BuildMI(*BB, MI, DL, TII->get(X86::MOV32rm), X86::EAX) |
26084 | .addReg(TII->getGlobalBaseReg(F)) |
26085 | .addImm(0) |
26086 | .addReg(0) |
26087 | .addGlobalAddress(MI.getOperand(3).getGlobal(), 0, |
26088 | MI.getOperand(3).getTargetFlags()) |
26089 | .addReg(0); |
26090 | MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m)); |
26091 | addDirectMem(MIB, X86::EAX); |
26092 | MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask); |
26093 | } |
26094 | |
26095 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
26096 | return BB; |
26097 | } |
26098 | |
26099 | MachineBasicBlock * |
26100 | X86TargetLowering::emitEHSjLjSetJmp(MachineInstr &MI, |
26101 | MachineBasicBlock *MBB) const { |
26102 | DebugLoc DL = MI.getDebugLoc(); |
26103 | MachineFunction *MF = MBB->getParent(); |
26104 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
26105 | const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); |
26106 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
26107 | |
26108 | const BasicBlock *BB = MBB->getBasicBlock(); |
26109 | MachineFunction::iterator I = ++MBB->getIterator(); |
26110 | |
26111 | // Memory Reference |
26112 | MachineInstr::mmo_iterator MMOBegin = MI.memoperands_begin(); |
26113 | MachineInstr::mmo_iterator MMOEnd = MI.memoperands_end(); |
26114 | |
26115 | unsigned DstReg; |
26116 | unsigned MemOpndSlot = 0; |
26117 | |
26118 | unsigned CurOp = 0; |
26119 | |
26120 | DstReg = MI.getOperand(CurOp++).getReg(); |
26121 | const TargetRegisterClass *RC = MRI.getRegClass(DstReg); |
26122 | assert(TRI->isTypeLegalForClass(*RC, MVT::i32) && "Invalid destination!")((TRI->isTypeLegalForClass(*RC, MVT::i32) && "Invalid destination!" ) ? static_cast<void> (0) : __assert_fail ("TRI->isTypeLegalForClass(*RC, MVT::i32) && \"Invalid destination!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26122, __PRETTY_FUNCTION__)); |
26123 | (void)TRI; |
26124 | unsigned mainDstReg = MRI.createVirtualRegister(RC); |
26125 | unsigned restoreDstReg = MRI.createVirtualRegister(RC); |
26126 | |
26127 | MemOpndSlot = CurOp; |
26128 | |
26129 | MVT PVT = getPointerTy(MF->getDataLayout()); |
26130 | assert((PVT == MVT::i64 || PVT == MVT::i32) &&(((PVT == MVT::i64 || PVT == MVT::i32) && "Invalid Pointer Size!" ) ? static_cast<void> (0) : __assert_fail ("(PVT == MVT::i64 || PVT == MVT::i32) && \"Invalid Pointer Size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26131, __PRETTY_FUNCTION__)) |
26131 | "Invalid Pointer Size!")(((PVT == MVT::i64 || PVT == MVT::i32) && "Invalid Pointer Size!" ) ? static_cast<void> (0) : __assert_fail ("(PVT == MVT::i64 || PVT == MVT::i32) && \"Invalid Pointer Size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26131, __PRETTY_FUNCTION__)); |
26132 | |
26133 | // For v = setjmp(buf), we generate |
26134 | // |
26135 | // thisMBB: |
26136 | // buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB |
26137 | // SjLjSetup restoreMBB |
26138 | // |
26139 | // mainMBB: |
26140 | // v_main = 0 |
26141 | // |
26142 | // sinkMBB: |
26143 | // v = phi(main, restore) |
26144 | // |
26145 | // restoreMBB: |
26146 | // if base pointer being used, load it from frame |
26147 | // v_restore = 1 |
26148 | |
26149 | MachineBasicBlock *thisMBB = MBB; |
26150 | MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB); |
26151 | MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB); |
26152 | MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB); |
26153 | MF->insert(I, mainMBB); |
26154 | MF->insert(I, sinkMBB); |
26155 | MF->push_back(restoreMBB); |
26156 | restoreMBB->setHasAddressTaken(); |
26157 | |
26158 | MachineInstrBuilder MIB; |
26159 | |
26160 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
26161 | sinkMBB->splice(sinkMBB->begin(), MBB, |
26162 | std::next(MachineBasicBlock::iterator(MI)), MBB->end()); |
26163 | sinkMBB->transferSuccessorsAndUpdatePHIs(MBB); |
26164 | |
26165 | // thisMBB: |
26166 | unsigned PtrStoreOpc = 0; |
26167 | unsigned LabelReg = 0; |
26168 | const int64_t LabelOffset = 1 * PVT.getStoreSize(); |
26169 | bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) && |
26170 | !isPositionIndependent(); |
26171 | |
26172 | // Prepare IP either in reg or imm. |
26173 | if (!UseImmLabel) { |
26174 | PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr; |
26175 | const TargetRegisterClass *PtrRC = getRegClassFor(PVT); |
26176 | LabelReg = MRI.createVirtualRegister(PtrRC); |
26177 | if (Subtarget.is64Bit()) { |
26178 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg) |
26179 | .addReg(X86::RIP) |
26180 | .addImm(0) |
26181 | .addReg(0) |
26182 | .addMBB(restoreMBB) |
26183 | .addReg(0); |
26184 | } else { |
26185 | const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII); |
26186 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg) |
26187 | .addReg(XII->getGlobalBaseReg(MF)) |
26188 | .addImm(0) |
26189 | .addReg(0) |
26190 | .addMBB(restoreMBB, Subtarget.classifyBlockAddressReference()) |
26191 | .addReg(0); |
26192 | } |
26193 | } else |
26194 | PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi; |
26195 | // Store IP |
26196 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc)); |
26197 | for (unsigned i = 0; i < X86::AddrNumOperands; ++i) { |
26198 | if (i == X86::AddrDisp) |
26199 | MIB.addDisp(MI.getOperand(MemOpndSlot + i), LabelOffset); |
26200 | else |
26201 | MIB.add(MI.getOperand(MemOpndSlot + i)); |
26202 | } |
26203 | if (!UseImmLabel) |
26204 | MIB.addReg(LabelReg); |
26205 | else |
26206 | MIB.addMBB(restoreMBB); |
26207 | MIB.setMemRefs(MMOBegin, MMOEnd); |
26208 | // Setup |
26209 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup)) |
26210 | .addMBB(restoreMBB); |
26211 | |
26212 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
26213 | MIB.addRegMask(RegInfo->getNoPreservedMask()); |
26214 | thisMBB->addSuccessor(mainMBB); |
26215 | thisMBB->addSuccessor(restoreMBB); |
26216 | |
26217 | // mainMBB: |
26218 | // EAX = 0 |
26219 | BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg); |
26220 | mainMBB->addSuccessor(sinkMBB); |
26221 | |
26222 | // sinkMBB: |
26223 | BuildMI(*sinkMBB, sinkMBB->begin(), DL, |
26224 | TII->get(X86::PHI), DstReg) |
26225 | .addReg(mainDstReg).addMBB(mainMBB) |
26226 | .addReg(restoreDstReg).addMBB(restoreMBB); |
26227 | |
26228 | // restoreMBB: |
26229 | if (RegInfo->hasBasePointer(*MF)) { |
26230 | const bool Uses64BitFramePtr = |
26231 | Subtarget.isTarget64BitLP64() || Subtarget.isTargetNaCl64(); |
26232 | X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>(); |
26233 | X86FI->setRestoreBasePointer(MF); |
26234 | unsigned FramePtr = RegInfo->getFrameRegister(*MF); |
26235 | unsigned BasePtr = RegInfo->getBaseRegister(); |
26236 | unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm; |
26237 | addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr), |
26238 | FramePtr, true, X86FI->getRestoreBasePointerOffset()) |
26239 | .setMIFlag(MachineInstr::FrameSetup); |
26240 | } |
26241 | BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1); |
26242 | BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB); |
26243 | restoreMBB->addSuccessor(sinkMBB); |
26244 | |
26245 | MI.eraseFromParent(); |
26246 | return sinkMBB; |
26247 | } |
26248 | |
26249 | MachineBasicBlock * |
26250 | X86TargetLowering::emitEHSjLjLongJmp(MachineInstr &MI, |
26251 | MachineBasicBlock *MBB) const { |
26252 | DebugLoc DL = MI.getDebugLoc(); |
26253 | MachineFunction *MF = MBB->getParent(); |
26254 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
26255 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
26256 | |
26257 | // Memory Reference |
26258 | MachineInstr::mmo_iterator MMOBegin = MI.memoperands_begin(); |
26259 | MachineInstr::mmo_iterator MMOEnd = MI.memoperands_end(); |
26260 | |
26261 | MVT PVT = getPointerTy(MF->getDataLayout()); |
26262 | assert((PVT == MVT::i64 || PVT == MVT::i32) &&(((PVT == MVT::i64 || PVT == MVT::i32) && "Invalid Pointer Size!" ) ? static_cast<void> (0) : __assert_fail ("(PVT == MVT::i64 || PVT == MVT::i32) && \"Invalid Pointer Size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26263, __PRETTY_FUNCTION__)) |
26263 | "Invalid Pointer Size!")(((PVT == MVT::i64 || PVT == MVT::i32) && "Invalid Pointer Size!" ) ? static_cast<void> (0) : __assert_fail ("(PVT == MVT::i64 || PVT == MVT::i32) && \"Invalid Pointer Size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26263, __PRETTY_FUNCTION__)); |
26264 | |
26265 | const TargetRegisterClass *RC = |
26266 | (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass; |
26267 | unsigned Tmp = MRI.createVirtualRegister(RC); |
26268 | // Since FP is only updated here but NOT referenced, it's treated as GPR. |
26269 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
26270 | unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP; |
26271 | unsigned SP = RegInfo->getStackRegister(); |
26272 | |
26273 | MachineInstrBuilder MIB; |
26274 | |
26275 | const int64_t LabelOffset = 1 * PVT.getStoreSize(); |
26276 | const int64_t SPOffset = 2 * PVT.getStoreSize(); |
26277 | |
26278 | unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm; |
26279 | unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r; |
26280 | |
26281 | // Reload FP |
26282 | MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP); |
26283 | for (unsigned i = 0; i < X86::AddrNumOperands; ++i) |
26284 | MIB.add(MI.getOperand(i)); |
26285 | MIB.setMemRefs(MMOBegin, MMOEnd); |
26286 | // Reload IP |
26287 | MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp); |
26288 | for (unsigned i = 0; i < X86::AddrNumOperands; ++i) { |
26289 | if (i == X86::AddrDisp) |
26290 | MIB.addDisp(MI.getOperand(i), LabelOffset); |
26291 | else |
26292 | MIB.add(MI.getOperand(i)); |
26293 | } |
26294 | MIB.setMemRefs(MMOBegin, MMOEnd); |
26295 | // Reload SP |
26296 | MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP); |
26297 | for (unsigned i = 0; i < X86::AddrNumOperands; ++i) { |
26298 | if (i == X86::AddrDisp) |
26299 | MIB.addDisp(MI.getOperand(i), SPOffset); |
26300 | else |
26301 | MIB.add(MI.getOperand(i)); |
26302 | } |
26303 | MIB.setMemRefs(MMOBegin, MMOEnd); |
26304 | // Jump |
26305 | BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp); |
26306 | |
26307 | MI.eraseFromParent(); |
26308 | return MBB; |
26309 | } |
26310 | |
26311 | void X86TargetLowering::SetupEntryBlockForSjLj(MachineInstr &MI, |
26312 | MachineBasicBlock *MBB, |
26313 | MachineBasicBlock *DispatchBB, |
26314 | int FI) const { |
26315 | DebugLoc DL = MI.getDebugLoc(); |
26316 | MachineFunction *MF = MBB->getParent(); |
26317 | MachineRegisterInfo *MRI = &MF->getRegInfo(); |
26318 | const X86InstrInfo *TII = Subtarget.getInstrInfo(); |
26319 | |
26320 | MVT PVT = getPointerTy(MF->getDataLayout()); |
26321 | assert((PVT == MVT::i64 || PVT == MVT::i32) && "Invalid Pointer Size!")(((PVT == MVT::i64 || PVT == MVT::i32) && "Invalid Pointer Size!" ) ? static_cast<void> (0) : __assert_fail ("(PVT == MVT::i64 || PVT == MVT::i32) && \"Invalid Pointer Size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26321, __PRETTY_FUNCTION__)); |
26322 | |
26323 | unsigned Op = 0; |
26324 | unsigned VR = 0; |
26325 | |
26326 | bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) && |
26327 | !isPositionIndependent(); |
26328 | |
26329 | if (UseImmLabel) { |
26330 | Op = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi; |
26331 | } else { |
26332 | const TargetRegisterClass *TRC = |
26333 | (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass; |
26334 | VR = MRI->createVirtualRegister(TRC); |
26335 | Op = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr; |
26336 | |
26337 | if (Subtarget.is64Bit()) |
26338 | BuildMI(*MBB, MI, DL, TII->get(X86::LEA64r), VR) |
26339 | .addReg(X86::RIP) |
26340 | .addImm(1) |
26341 | .addReg(0) |
26342 | .addMBB(DispatchBB) |
26343 | .addReg(0); |
26344 | else |
26345 | BuildMI(*MBB, MI, DL, TII->get(X86::LEA32r), VR) |
26346 | .addReg(0) /* TII->getGlobalBaseReg(MF) */ |
26347 | .addImm(1) |
26348 | .addReg(0) |
26349 | .addMBB(DispatchBB, Subtarget.classifyBlockAddressReference()) |
26350 | .addReg(0); |
26351 | } |
26352 | |
26353 | MachineInstrBuilder MIB = BuildMI(*MBB, MI, DL, TII->get(Op)); |
26354 | addFrameReference(MIB, FI, 36); |
26355 | if (UseImmLabel) |
26356 | MIB.addMBB(DispatchBB); |
26357 | else |
26358 | MIB.addReg(VR); |
26359 | } |
26360 | |
26361 | MachineBasicBlock * |
26362 | X86TargetLowering::EmitSjLjDispatchBlock(MachineInstr &MI, |
26363 | MachineBasicBlock *BB) const { |
26364 | DebugLoc DL = MI.getDebugLoc(); |
26365 | MachineFunction *MF = BB->getParent(); |
26366 | MachineFrameInfo &MFI = MF->getFrameInfo(); |
26367 | MachineRegisterInfo *MRI = &MF->getRegInfo(); |
26368 | const X86InstrInfo *TII = Subtarget.getInstrInfo(); |
26369 | int FI = MFI.getFunctionContextIndex(); |
26370 | |
26371 | // Get a mapping of the call site numbers to all of the landing pads they're |
26372 | // associated with. |
26373 | DenseMap<unsigned, SmallVector<MachineBasicBlock *, 2>> CallSiteNumToLPad; |
26374 | unsigned MaxCSNum = 0; |
26375 | for (auto &MBB : *MF) { |
26376 | if (!MBB.isEHPad()) |
26377 | continue; |
26378 | |
26379 | MCSymbol *Sym = nullptr; |
26380 | for (const auto &MI : MBB) { |
26381 | if (MI.isDebugValue()) |
26382 | continue; |
26383 | |
26384 | assert(MI.isEHLabel() && "expected EH_LABEL")((MI.isEHLabel() && "expected EH_LABEL") ? static_cast <void> (0) : __assert_fail ("MI.isEHLabel() && \"expected EH_LABEL\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26384, __PRETTY_FUNCTION__)); |
26385 | Sym = MI.getOperand(0).getMCSymbol(); |
26386 | break; |
26387 | } |
26388 | |
26389 | if (!MF->hasCallSiteLandingPad(Sym)) |
26390 | continue; |
26391 | |
26392 | for (unsigned CSI : MF->getCallSiteLandingPad(Sym)) { |
26393 | CallSiteNumToLPad[CSI].push_back(&MBB); |
26394 | MaxCSNum = std::max(MaxCSNum, CSI); |
26395 | } |
26396 | } |
26397 | |
26398 | // Get an ordered list of the machine basic blocks for the jump table. |
26399 | std::vector<MachineBasicBlock *> LPadList; |
26400 | SmallPtrSet<MachineBasicBlock *, 32> InvokeBBs; |
26401 | LPadList.reserve(CallSiteNumToLPad.size()); |
26402 | |
26403 | for (unsigned CSI = 1; CSI <= MaxCSNum; ++CSI) { |
26404 | for (auto &LP : CallSiteNumToLPad[CSI]) { |
26405 | LPadList.push_back(LP); |
26406 | InvokeBBs.insert(LP->pred_begin(), LP->pred_end()); |
26407 | } |
26408 | } |
26409 | |
26410 | assert(!LPadList.empty() &&((!LPadList.empty() && "No landing pad destinations for the dispatch jump table!" ) ? static_cast<void> (0) : __assert_fail ("!LPadList.empty() && \"No landing pad destinations for the dispatch jump table!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26411, __PRETTY_FUNCTION__)) |
26411 | "No landing pad destinations for the dispatch jump table!")((!LPadList.empty() && "No landing pad destinations for the dispatch jump table!" ) ? static_cast<void> (0) : __assert_fail ("!LPadList.empty() && \"No landing pad destinations for the dispatch jump table!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26411, __PRETTY_FUNCTION__)); |
26412 | |
26413 | // Create the MBBs for the dispatch code. |
26414 | |
26415 | // Shove the dispatch's address into the return slot in the function context. |
26416 | MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock(); |
26417 | DispatchBB->setIsEHPad(true); |
26418 | |
26419 | MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock(); |
26420 | BuildMI(TrapBB, DL, TII->get(X86::TRAP)); |
26421 | DispatchBB->addSuccessor(TrapBB); |
26422 | |
26423 | MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock(); |
26424 | DispatchBB->addSuccessor(DispContBB); |
26425 | |
26426 | // Insert MBBs. |
26427 | MF->push_back(DispatchBB); |
26428 | MF->push_back(DispContBB); |
26429 | MF->push_back(TrapBB); |
26430 | |
26431 | // Insert code into the entry block that creates and registers the function |
26432 | // context. |
26433 | SetupEntryBlockForSjLj(MI, BB, DispatchBB, FI); |
26434 | |
26435 | // Create the jump table and associated information |
26436 | MachineJumpTableInfo *JTI = |
26437 | MF->getOrCreateJumpTableInfo(getJumpTableEncoding()); |
26438 | unsigned MJTI = JTI->createJumpTableIndex(LPadList); |
26439 | |
26440 | const X86RegisterInfo &RI = TII->getRegisterInfo(); |
26441 | // Add a register mask with no preserved registers. This results in all |
26442 | // registers being marked as clobbered. |
26443 | if (RI.hasBasePointer(*MF)) { |
26444 | const bool FPIs64Bit = |
26445 | Subtarget.isTarget64BitLP64() || Subtarget.isTargetNaCl64(); |
26446 | X86MachineFunctionInfo *MFI = MF->getInfo<X86MachineFunctionInfo>(); |
26447 | MFI->setRestoreBasePointer(MF); |
26448 | |
26449 | unsigned FP = RI.getFrameRegister(*MF); |
26450 | unsigned BP = RI.getBaseRegister(); |
26451 | unsigned Op = FPIs64Bit ? X86::MOV64rm : X86::MOV32rm; |
26452 | addRegOffset(BuildMI(DispatchBB, DL, TII->get(Op), BP), FP, true, |
26453 | MFI->getRestoreBasePointerOffset()) |
26454 | .addRegMask(RI.getNoPreservedMask()); |
26455 | } else { |
26456 | BuildMI(DispatchBB, DL, TII->get(X86::NOOP)) |
26457 | .addRegMask(RI.getNoPreservedMask()); |
26458 | } |
26459 | |
26460 | unsigned IReg = MRI->createVirtualRegister(&X86::GR32RegClass); |
26461 | addFrameReference(BuildMI(DispatchBB, DL, TII->get(X86::MOV32rm), IReg), FI, |
26462 | 4); |
26463 | BuildMI(DispatchBB, DL, TII->get(X86::CMP32ri)) |
26464 | .addReg(IReg) |
26465 | .addImm(LPadList.size()); |
26466 | BuildMI(DispatchBB, DL, TII->get(X86::JA_1)).addMBB(TrapBB); |
26467 | |
26468 | unsigned JReg = MRI->createVirtualRegister(&X86::GR32RegClass); |
26469 | BuildMI(DispContBB, DL, TII->get(X86::SUB32ri), JReg) |
26470 | .addReg(IReg) |
26471 | .addImm(1); |
26472 | BuildMI(DispContBB, DL, |
26473 | TII->get(Subtarget.is64Bit() ? X86::JMP64m : X86::JMP32m)) |
26474 | .addReg(0) |
26475 | .addImm(Subtarget.is64Bit() ? 8 : 4) |
26476 | .addReg(JReg) |
26477 | .addJumpTableIndex(MJTI) |
26478 | .addReg(0); |
26479 | |
26480 | // Add the jump table entries as successors to the MBB. |
26481 | SmallPtrSet<MachineBasicBlock *, 8> SeenMBBs; |
26482 | for (auto &LP : LPadList) |
26483 | if (SeenMBBs.insert(LP).second) |
26484 | DispContBB->addSuccessor(LP); |
26485 | |
26486 | // N.B. the order the invoke BBs are processed in doesn't matter here. |
26487 | SmallVector<MachineBasicBlock *, 64> MBBLPads; |
26488 | const MCPhysReg *SavedRegs = MF->getRegInfo().getCalleeSavedRegs(); |
26489 | for (MachineBasicBlock *MBB : InvokeBBs) { |
26490 | // Remove the landing pad successor from the invoke block and replace it |
26491 | // with the new dispatch block. |
26492 | // Keep a copy of Successors since it's modified inside the loop. |
26493 | SmallVector<MachineBasicBlock *, 8> Successors(MBB->succ_rbegin(), |
26494 | MBB->succ_rend()); |
26495 | // FIXME: Avoid quadratic complexity. |
26496 | for (auto MBBS : Successors) { |
26497 | if (MBBS->isEHPad()) { |
26498 | MBB->removeSuccessor(MBBS); |
26499 | MBBLPads.push_back(MBBS); |
26500 | } |
26501 | } |
26502 | |
26503 | MBB->addSuccessor(DispatchBB); |
26504 | |
26505 | // Find the invoke call and mark all of the callee-saved registers as |
26506 | // 'implicit defined' so that they're spilled. This prevents code from |
26507 | // moving instructions to before the EH block, where they will never be |
26508 | // executed. |
26509 | for (auto &II : reverse(*MBB)) { |
26510 | if (!II.isCall()) |
26511 | continue; |
26512 | |
26513 | DenseMap<unsigned, bool> DefRegs; |
26514 | for (auto &MOp : II.operands()) |
26515 | if (MOp.isReg()) |
26516 | DefRegs[MOp.getReg()] = true; |
26517 | |
26518 | MachineInstrBuilder MIB(*MF, &II); |
26519 | for (unsigned RI = 0; SavedRegs[RI]; ++RI) { |
26520 | unsigned Reg = SavedRegs[RI]; |
26521 | if (!DefRegs[Reg]) |
26522 | MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead); |
26523 | } |
26524 | |
26525 | break; |
26526 | } |
26527 | } |
26528 | |
26529 | // Mark all former landing pads as non-landing pads. The dispatch is the only |
26530 | // landing pad now. |
26531 | for (auto &LP : MBBLPads) |
26532 | LP->setIsEHPad(false); |
26533 | |
26534 | // The instruction is gone now. |
26535 | MI.eraseFromParent(); |
26536 | return BB; |
26537 | } |
26538 | |
26539 | MachineBasicBlock * |
26540 | X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI, |
26541 | MachineBasicBlock *BB) const { |
26542 | MachineFunction *MF = BB->getParent(); |
26543 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
26544 | DebugLoc DL = MI.getDebugLoc(); |
26545 | |
26546 | switch (MI.getOpcode()) { |
26547 | default: llvm_unreachable("Unexpected instr type to insert")::llvm::llvm_unreachable_internal("Unexpected instr type to insert" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26547); |
26548 | case X86::TAILJMPd64: |
26549 | case X86::TAILJMPr64: |
26550 | case X86::TAILJMPm64: |
26551 | case X86::TAILJMPr64_REX: |
26552 | case X86::TAILJMPm64_REX: |
26553 | llvm_unreachable("TAILJMP64 would not be touched here.")::llvm::llvm_unreachable_internal("TAILJMP64 would not be touched here." , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26553); |
26554 | case X86::TCRETURNdi64: |
26555 | case X86::TCRETURNri64: |
26556 | case X86::TCRETURNmi64: |
26557 | return BB; |
26558 | case X86::TLS_addr32: |
26559 | case X86::TLS_addr64: |
26560 | case X86::TLS_base_addr32: |
26561 | case X86::TLS_base_addr64: |
26562 | return EmitLoweredTLSAddr(MI, BB); |
26563 | case X86::CATCHRET: |
26564 | return EmitLoweredCatchRet(MI, BB); |
26565 | case X86::CATCHPAD: |
26566 | return EmitLoweredCatchPad(MI, BB); |
26567 | case X86::SEG_ALLOCA_32: |
26568 | case X86::SEG_ALLOCA_64: |
26569 | return EmitLoweredSegAlloca(MI, BB); |
26570 | case X86::TLSCall_32: |
26571 | case X86::TLSCall_64: |
26572 | return EmitLoweredTLSCall(MI, BB); |
26573 | case X86::CMOV_FR32: |
26574 | case X86::CMOV_FR64: |
26575 | case X86::CMOV_FR128: |
26576 | case X86::CMOV_GR8: |
26577 | case X86::CMOV_GR16: |
26578 | case X86::CMOV_GR32: |
26579 | case X86::CMOV_RFP32: |
26580 | case X86::CMOV_RFP64: |
26581 | case X86::CMOV_RFP80: |
26582 | case X86::CMOV_V2F64: |
26583 | case X86::CMOV_V2I64: |
26584 | case X86::CMOV_V4F32: |
26585 | case X86::CMOV_V4F64: |
26586 | case X86::CMOV_V4I64: |
26587 | case X86::CMOV_V16F32: |
26588 | case X86::CMOV_V8F32: |
26589 | case X86::CMOV_V8F64: |
26590 | case X86::CMOV_V8I64: |
26591 | case X86::CMOV_V8I1: |
26592 | case X86::CMOV_V16I1: |
26593 | case X86::CMOV_V32I1: |
26594 | case X86::CMOV_V64I1: |
26595 | return EmitLoweredSelect(MI, BB); |
26596 | |
26597 | case X86::RDFLAGS32: |
26598 | case X86::RDFLAGS64: { |
26599 | unsigned PushF = |
26600 | MI.getOpcode() == X86::RDFLAGS32 ? X86::PUSHF32 : X86::PUSHF64; |
26601 | unsigned Pop = MI.getOpcode() == X86::RDFLAGS32 ? X86::POP32r : X86::POP64r; |
26602 | MachineInstr *Push = BuildMI(*BB, MI, DL, TII->get(PushF)); |
26603 | // Permit reads of the FLAGS register without it being defined. |
26604 | // This intrinsic exists to read external processor state in flags, such as |
26605 | // the trap flag, interrupt flag, and direction flag, none of which are |
26606 | // modeled by the backend. |
26607 | Push->getOperand(2).setIsUndef(); |
26608 | BuildMI(*BB, MI, DL, TII->get(Pop), MI.getOperand(0).getReg()); |
26609 | |
26610 | MI.eraseFromParent(); // The pseudo is gone now. |
26611 | return BB; |
26612 | } |
26613 | |
26614 | case X86::WRFLAGS32: |
26615 | case X86::WRFLAGS64: { |
26616 | unsigned Push = |
26617 | MI.getOpcode() == X86::WRFLAGS32 ? X86::PUSH32r : X86::PUSH64r; |
26618 | unsigned PopF = |
26619 | MI.getOpcode() == X86::WRFLAGS32 ? X86::POPF32 : X86::POPF64; |
26620 | BuildMI(*BB, MI, DL, TII->get(Push)).addReg(MI.getOperand(0).getReg()); |
26621 | BuildMI(*BB, MI, DL, TII->get(PopF)); |
26622 | |
26623 | MI.eraseFromParent(); // The pseudo is gone now. |
26624 | return BB; |
26625 | } |
26626 | |
26627 | case X86::RELEASE_FADD32mr: |
26628 | case X86::RELEASE_FADD64mr: |
26629 | return EmitLoweredAtomicFP(MI, BB); |
26630 | |
26631 | case X86::FP32_TO_INT16_IN_MEM: |
26632 | case X86::FP32_TO_INT32_IN_MEM: |
26633 | case X86::FP32_TO_INT64_IN_MEM: |
26634 | case X86::FP64_TO_INT16_IN_MEM: |
26635 | case X86::FP64_TO_INT32_IN_MEM: |
26636 | case X86::FP64_TO_INT64_IN_MEM: |
26637 | case X86::FP80_TO_INT16_IN_MEM: |
26638 | case X86::FP80_TO_INT32_IN_MEM: |
26639 | case X86::FP80_TO_INT64_IN_MEM: { |
26640 | // Change the floating point control register to use "round towards zero" |
26641 | // mode when truncating to an integer value. |
26642 | int CWFrameIdx = MF->getFrameInfo().CreateStackObject(2, 2, false); |
26643 | addFrameReference(BuildMI(*BB, MI, DL, |
26644 | TII->get(X86::FNSTCW16m)), CWFrameIdx); |
26645 | |
26646 | // Load the old value of the high byte of the control word... |
26647 | unsigned OldCW = |
26648 | MF->getRegInfo().createVirtualRegister(&X86::GR16RegClass); |
26649 | addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW), |
26650 | CWFrameIdx); |
26651 | |
26652 | // Set the high part to be round to zero... |
26653 | addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx) |
26654 | .addImm(0xC7F); |
26655 | |
26656 | // Reload the modified control word now... |
26657 | addFrameReference(BuildMI(*BB, MI, DL, |
26658 | TII->get(X86::FLDCW16m)), CWFrameIdx); |
26659 | |
26660 | // Restore the memory image of control word to original value |
26661 | addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx) |
26662 | .addReg(OldCW); |
26663 | |
26664 | // Get the X86 opcode to use. |
26665 | unsigned Opc; |
26666 | switch (MI.getOpcode()) { |
26667 | default: llvm_unreachable("illegal opcode!")::llvm::llvm_unreachable_internal("illegal opcode!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26667); |
26668 | case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break; |
26669 | case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break; |
26670 | case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break; |
26671 | case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break; |
26672 | case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break; |
26673 | case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break; |
26674 | case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break; |
26675 | case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break; |
26676 | case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break; |
26677 | } |
26678 | |
26679 | X86AddressMode AM = getAddressFromInstr(&MI, 0); |
26680 | addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM) |
26681 | .addReg(MI.getOperand(X86::AddrNumOperands).getReg()); |
26682 | |
26683 | // Reload the original control word now. |
26684 | addFrameReference(BuildMI(*BB, MI, DL, |
26685 | TII->get(X86::FLDCW16m)), CWFrameIdx); |
26686 | |
26687 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
26688 | return BB; |
26689 | } |
26690 | // String/text processing lowering. |
26691 | case X86::PCMPISTRM128REG: |
26692 | case X86::VPCMPISTRM128REG: |
26693 | case X86::PCMPISTRM128MEM: |
26694 | case X86::VPCMPISTRM128MEM: |
26695 | case X86::PCMPESTRM128REG: |
26696 | case X86::VPCMPESTRM128REG: |
26697 | case X86::PCMPESTRM128MEM: |
26698 | case X86::VPCMPESTRM128MEM: |
26699 | assert(Subtarget.hasSSE42() &&((Subtarget.hasSSE42() && "Target must have SSE4.2 or AVX features enabled" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSE42() && \"Target must have SSE4.2 or AVX features enabled\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26700, __PRETTY_FUNCTION__)) |
26700 | "Target must have SSE4.2 or AVX features enabled")((Subtarget.hasSSE42() && "Target must have SSE4.2 or AVX features enabled" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSE42() && \"Target must have SSE4.2 or AVX features enabled\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26700, __PRETTY_FUNCTION__)); |
26701 | return emitPCMPSTRM(MI, BB, Subtarget.getInstrInfo()); |
26702 | |
26703 | // String/text processing lowering. |
26704 | case X86::PCMPISTRIREG: |
26705 | case X86::VPCMPISTRIREG: |
26706 | case X86::PCMPISTRIMEM: |
26707 | case X86::VPCMPISTRIMEM: |
26708 | case X86::PCMPESTRIREG: |
26709 | case X86::VPCMPESTRIREG: |
26710 | case X86::PCMPESTRIMEM: |
26711 | case X86::VPCMPESTRIMEM: |
26712 | assert(Subtarget.hasSSE42() &&((Subtarget.hasSSE42() && "Target must have SSE4.2 or AVX features enabled" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSE42() && \"Target must have SSE4.2 or AVX features enabled\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26713, __PRETTY_FUNCTION__)) |
26713 | "Target must have SSE4.2 or AVX features enabled")((Subtarget.hasSSE42() && "Target must have SSE4.2 or AVX features enabled" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasSSE42() && \"Target must have SSE4.2 or AVX features enabled\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26713, __PRETTY_FUNCTION__)); |
26714 | return emitPCMPSTRI(MI, BB, Subtarget.getInstrInfo()); |
26715 | |
26716 | // Thread synchronization. |
26717 | case X86::MONITOR: |
26718 | return emitMonitor(MI, BB, Subtarget, X86::MONITORrrr); |
26719 | case X86::MONITORX: |
26720 | return emitMonitor(MI, BB, Subtarget, X86::MONITORXrrr); |
26721 | |
26722 | // Cache line zero |
26723 | case X86::CLZERO: |
26724 | return emitClzero(&MI, BB, Subtarget); |
26725 | |
26726 | // PKU feature |
26727 | case X86::WRPKRU: |
26728 | return emitWRPKRU(MI, BB, Subtarget); |
26729 | case X86::RDPKRU: |
26730 | return emitRDPKRU(MI, BB, Subtarget); |
26731 | // xbegin |
26732 | case X86::XBEGIN: |
26733 | return emitXBegin(MI, BB, Subtarget.getInstrInfo()); |
26734 | |
26735 | case X86::VASTART_SAVE_XMM_REGS: |
26736 | return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB); |
26737 | |
26738 | case X86::VAARG_64: |
26739 | return EmitVAARG64WithCustomInserter(MI, BB); |
26740 | |
26741 | case X86::EH_SjLj_SetJmp32: |
26742 | case X86::EH_SjLj_SetJmp64: |
26743 | return emitEHSjLjSetJmp(MI, BB); |
26744 | |
26745 | case X86::EH_SjLj_LongJmp32: |
26746 | case X86::EH_SjLj_LongJmp64: |
26747 | return emitEHSjLjLongJmp(MI, BB); |
26748 | |
26749 | case X86::Int_eh_sjlj_setup_dispatch: |
26750 | return EmitSjLjDispatchBlock(MI, BB); |
26751 | |
26752 | case TargetOpcode::STATEPOINT: |
26753 | // As an implementation detail, STATEPOINT shares the STACKMAP format at |
26754 | // this point in the process. We diverge later. |
26755 | return emitPatchPoint(MI, BB); |
26756 | |
26757 | case TargetOpcode::STACKMAP: |
26758 | case TargetOpcode::PATCHPOINT: |
26759 | return emitPatchPoint(MI, BB); |
26760 | |
26761 | case TargetOpcode::PATCHABLE_EVENT_CALL: |
26762 | // Do nothing here, handle in xray instrumentation pass. |
26763 | return BB; |
26764 | |
26765 | case X86::LCMPXCHG8B: { |
26766 | const X86RegisterInfo *TRI = Subtarget.getRegisterInfo(); |
26767 | // In addition to 4 E[ABCD] registers implied by encoding, CMPXCHG8B |
26768 | // requires a memory operand. If it happens that current architecture is |
26769 | // i686 and for current function we need a base pointer |
26770 | // - which is ESI for i686 - register allocator would not be able to |
26771 | // allocate registers for an address in form of X(%reg, %reg, Y) |
26772 | // - there never would be enough unreserved registers during regalloc |
26773 | // (without the need for base ptr the only option would be X(%edi, %esi, Y). |
26774 | // We are giving a hand to register allocator by precomputing the address in |
26775 | // a new vreg using LEA. |
26776 | |
26777 | // If it is not i686 or there is no base pointer - nothing to do here. |
26778 | if (!Subtarget.is32Bit() || !TRI->hasBasePointer(*MF)) |
26779 | return BB; |
26780 | |
26781 | // Even though this code does not necessarily needs the base pointer to |
26782 | // be ESI, we check for that. The reason: if this assert fails, there are |
26783 | // some changes happened in the compiler base pointer handling, which most |
26784 | // probably have to be addressed somehow here. |
26785 | assert(TRI->getBaseRegister() == X86::ESI &&((TRI->getBaseRegister() == X86::ESI && "LCMPXCHG8B custom insertion for i686 is written with X86::ESI as a " "base pointer in mind") ? static_cast<void> (0) : __assert_fail ("TRI->getBaseRegister() == X86::ESI && \"LCMPXCHG8B custom insertion for i686 is written with X86::ESI as a \" \"base pointer in mind\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26787, __PRETTY_FUNCTION__)) |
26786 | "LCMPXCHG8B custom insertion for i686 is written with X86::ESI as a "((TRI->getBaseRegister() == X86::ESI && "LCMPXCHG8B custom insertion for i686 is written with X86::ESI as a " "base pointer in mind") ? static_cast<void> (0) : __assert_fail ("TRI->getBaseRegister() == X86::ESI && \"LCMPXCHG8B custom insertion for i686 is written with X86::ESI as a \" \"base pointer in mind\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26787, __PRETTY_FUNCTION__)) |
26787 | "base pointer in mind")((TRI->getBaseRegister() == X86::ESI && "LCMPXCHG8B custom insertion for i686 is written with X86::ESI as a " "base pointer in mind") ? static_cast<void> (0) : __assert_fail ("TRI->getBaseRegister() == X86::ESI && \"LCMPXCHG8B custom insertion for i686 is written with X86::ESI as a \" \"base pointer in mind\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26787, __PRETTY_FUNCTION__)); |
26788 | |
26789 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
26790 | MVT SPTy = getPointerTy(MF->getDataLayout()); |
26791 | const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy); |
26792 | unsigned computedAddrVReg = MRI.createVirtualRegister(AddrRegClass); |
26793 | |
26794 | X86AddressMode AM = getAddressFromInstr(&MI, 0); |
26795 | // Regalloc does not need any help when the memory operand of CMPXCHG8B |
26796 | // does not use index register. |
26797 | if (AM.IndexReg == X86::NoRegister) |
26798 | return BB; |
26799 | |
26800 | // After X86TargetLowering::ReplaceNodeResults CMPXCHG8B is glued to its |
26801 | // four operand definitions that are E[ABCD] registers. We skip them and |
26802 | // then insert the LEA. |
26803 | MachineBasicBlock::iterator MBBI(MI); |
26804 | while (MBBI->definesRegister(X86::EAX) || MBBI->definesRegister(X86::EBX) || |
26805 | MBBI->definesRegister(X86::ECX) || MBBI->definesRegister(X86::EDX)) |
26806 | --MBBI; |
26807 | addFullAddress( |
26808 | BuildMI(*BB, *MBBI, DL, TII->get(X86::LEA32r), computedAddrVReg), AM); |
26809 | |
26810 | setDirectAddressInInstr(&MI, 0, computedAddrVReg); |
26811 | |
26812 | return BB; |
26813 | } |
26814 | case X86::LCMPXCHG16B: |
26815 | return BB; |
26816 | case X86::LCMPXCHG8B_SAVE_EBX: |
26817 | case X86::LCMPXCHG16B_SAVE_RBX: { |
26818 | unsigned BasePtr = |
26819 | MI.getOpcode() == X86::LCMPXCHG8B_SAVE_EBX ? X86::EBX : X86::RBX; |
26820 | if (!BB->isLiveIn(BasePtr)) |
26821 | BB->addLiveIn(BasePtr); |
26822 | return BB; |
26823 | } |
26824 | } |
26825 | } |
26826 | |
26827 | //===----------------------------------------------------------------------===// |
26828 | // X86 Optimization Hooks |
26829 | //===----------------------------------------------------------------------===// |
26830 | |
26831 | void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op, |
26832 | KnownBits &Known, |
26833 | const APInt &DemandedElts, |
26834 | const SelectionDAG &DAG, |
26835 | unsigned Depth) const { |
26836 | unsigned BitWidth = Known.getBitWidth(); |
26837 | unsigned Opc = Op.getOpcode(); |
26838 | EVT VT = Op.getValueType(); |
26839 | assert((Opc >= ISD::BUILTIN_OP_END ||(((Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID ) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26844, __PRETTY_FUNCTION__)) |
26840 | Opc == ISD::INTRINSIC_WO_CHAIN ||(((Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID ) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26844, __PRETTY_FUNCTION__)) |
26841 | Opc == ISD::INTRINSIC_W_CHAIN ||(((Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID ) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26844, __PRETTY_FUNCTION__)) |
26842 | Opc == ISD::INTRINSIC_VOID) &&(((Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID ) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26844, __PRETTY_FUNCTION__)) |
26843 | "Should use MaskedValueIsZero if you don't know whether Op"(((Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID ) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26844, __PRETTY_FUNCTION__)) |
26844 | " is a target node!")(((Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID ) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26844, __PRETTY_FUNCTION__)); |
26845 | |
26846 | Known.resetAll(); |
26847 | switch (Opc) { |
26848 | default: break; |
26849 | case X86ISD::ADD: |
26850 | case X86ISD::SUB: |
26851 | case X86ISD::ADC: |
26852 | case X86ISD::SBB: |
26853 | case X86ISD::SMUL: |
26854 | case X86ISD::UMUL: |
26855 | case X86ISD::INC: |
26856 | case X86ISD::DEC: |
26857 | case X86ISD::OR: |
26858 | case X86ISD::XOR: |
26859 | case X86ISD::AND: |
26860 | // These nodes' second result is a boolean. |
26861 | if (Op.getResNo() == 0) |
26862 | break; |
26863 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
26864 | case X86ISD::SETCC: |
26865 | Known.Zero.setBitsFrom(1); |
26866 | break; |
26867 | case X86ISD::MOVMSK: { |
26868 | unsigned NumLoBits = Op.getOperand(0).getValueType().getVectorNumElements(); |
26869 | Known.Zero.setBitsFrom(NumLoBits); |
26870 | break; |
26871 | } |
26872 | case X86ISD::VSHLI: |
26873 | case X86ISD::VSRLI: { |
26874 | if (auto *ShiftImm = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { |
26875 | if (ShiftImm->getAPIntValue().uge(VT.getScalarSizeInBits())) { |
26876 | Known.setAllZero(); |
26877 | break; |
26878 | } |
26879 | |
26880 | DAG.computeKnownBits(Op.getOperand(0), Known, Depth + 1); |
26881 | unsigned ShAmt = ShiftImm->getZExtValue(); |
26882 | if (Opc == X86ISD::VSHLI) { |
26883 | Known.Zero <<= ShAmt; |
26884 | Known.One <<= ShAmt; |
26885 | // Low bits are known zero. |
26886 | Known.Zero.setLowBits(ShAmt); |
26887 | } else { |
26888 | Known.Zero.lshrInPlace(ShAmt); |
26889 | Known.One.lshrInPlace(ShAmt); |
26890 | // High bits are known zero. |
26891 | Known.Zero.setHighBits(ShAmt); |
26892 | } |
26893 | } |
26894 | break; |
26895 | } |
26896 | case X86ISD::VZEXT: { |
26897 | SDValue N0 = Op.getOperand(0); |
26898 | unsigned NumElts = VT.getVectorNumElements(); |
26899 | |
26900 | EVT SrcVT = N0.getValueType(); |
26901 | unsigned InNumElts = SrcVT.getVectorNumElements(); |
26902 | unsigned InBitWidth = SrcVT.getScalarSizeInBits(); |
26903 | assert(InNumElts >= NumElts && "Illegal VZEXT input")((InNumElts >= NumElts && "Illegal VZEXT input") ? static_cast<void> (0) : __assert_fail ("InNumElts >= NumElts && \"Illegal VZEXT input\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 26903, __PRETTY_FUNCTION__)); |
26904 | |
26905 | Known = KnownBits(InBitWidth); |
26906 | APInt DemandedSrcElts = APInt::getLowBitsSet(InNumElts, NumElts); |
26907 | DAG.computeKnownBits(N0, Known, DemandedSrcElts, Depth + 1); |
26908 | Known = Known.zext(BitWidth); |
26909 | Known.Zero.setBitsFrom(InBitWidth); |
26910 | break; |
26911 | } |
26912 | } |
26913 | } |
26914 | |
26915 | unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode( |
26916 | SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG, |
26917 | unsigned Depth) const { |
26918 | unsigned VTBits = Op.getScalarValueSizeInBits(); |
26919 | unsigned Opcode = Op.getOpcode(); |
26920 | switch (Opcode) { |
26921 | case X86ISD::SETCC_CARRY: |
26922 | // SETCC_CARRY sets the dest to ~0 for true or 0 for false. |
26923 | return VTBits; |
26924 | |
26925 | case X86ISD::VSEXT: { |
26926 | SDValue Src = Op.getOperand(0); |
26927 | unsigned Tmp = DAG.ComputeNumSignBits(Src, Depth + 1); |
26928 | Tmp += VTBits - Src.getScalarValueSizeInBits(); |
26929 | return Tmp; |
26930 | } |
26931 | |
26932 | case X86ISD::VSHLI: { |
26933 | SDValue Src = Op.getOperand(0); |
26934 | unsigned Tmp = DAG.ComputeNumSignBits(Src, Depth + 1); |
26935 | APInt ShiftVal = cast<ConstantSDNode>(Op.getOperand(1))->getAPIntValue(); |
26936 | if (ShiftVal.uge(VTBits)) |
26937 | return VTBits; // Shifted all bits out --> zero. |
26938 | if (ShiftVal.uge(Tmp)) |
26939 | return 1; // Shifted all sign bits out --> unknown. |
26940 | return Tmp - ShiftVal.getZExtValue(); |
26941 | } |
26942 | |
26943 | case X86ISD::VSRAI: { |
26944 | SDValue Src = Op.getOperand(0); |
26945 | unsigned Tmp = DAG.ComputeNumSignBits(Src, Depth + 1); |
26946 | APInt ShiftVal = cast<ConstantSDNode>(Op.getOperand(1))->getAPIntValue(); |
26947 | ShiftVal += Tmp; |
26948 | return ShiftVal.uge(VTBits) ? VTBits : ShiftVal.getZExtValue(); |
26949 | } |
26950 | |
26951 | case X86ISD::PCMPGT: |
26952 | case X86ISD::PCMPEQ: |
26953 | case X86ISD::CMPP: |
26954 | case X86ISD::VPCOM: |
26955 | case X86ISD::VPCOMU: |
26956 | // Vector compares return zero/all-bits result values. |
26957 | return VTBits; |
26958 | } |
26959 | |
26960 | // Fallback case. |
26961 | return 1; |
26962 | } |
26963 | |
26964 | /// Returns true (and the GlobalValue and the offset) if the node is a |
26965 | /// GlobalAddress + offset. |
26966 | bool X86TargetLowering::isGAPlusOffset(SDNode *N, |
26967 | const GlobalValue* &GA, |
26968 | int64_t &Offset) const { |
26969 | if (N->getOpcode() == X86ISD::Wrapper) { |
26970 | if (isa<GlobalAddressSDNode>(N->getOperand(0))) { |
26971 | GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal(); |
26972 | Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset(); |
26973 | return true; |
26974 | } |
26975 | } |
26976 | return TargetLowering::isGAPlusOffset(N, GA, Offset); |
26977 | } |
26978 | |
26979 | // Attempt to match a combined shuffle mask against supported unary shuffle |
26980 | // instructions. |
26981 | // TODO: Investigate sharing more of this with shuffle lowering. |
26982 | static bool matchUnaryVectorShuffle(MVT MaskVT, ArrayRef<int> Mask, |
26983 | bool AllowFloatDomain, bool AllowIntDomain, |
26984 | SDValue &V1, SDLoc &DL, SelectionDAG &DAG, |
26985 | const X86Subtarget &Subtarget, |
26986 | unsigned &Shuffle, MVT &SrcVT, MVT &DstVT) { |
26987 | unsigned NumMaskElts = Mask.size(); |
26988 | unsigned MaskEltSize = MaskVT.getScalarSizeInBits(); |
26989 | |
26990 | // Match against a ZERO_EXTEND_VECTOR_INREG/VZEXT instruction. |
26991 | // TODO: Add 512-bit vector support (split AVX512F and AVX512BW). |
26992 | if (AllowIntDomain && ((MaskVT.is128BitVector() && Subtarget.hasSSE41()) || |
26993 | (MaskVT.is256BitVector() && Subtarget.hasInt256()))) { |
26994 | unsigned MaxScale = 64 / MaskEltSize; |
26995 | for (unsigned Scale = 2; Scale <= MaxScale; Scale *= 2) { |
26996 | bool Match = true; |
26997 | unsigned NumDstElts = NumMaskElts / Scale; |
26998 | for (unsigned i = 0; i != NumDstElts && Match; ++i) { |
26999 | Match &= isUndefOrEqual(Mask[i * Scale], (int)i); |
27000 | Match &= isUndefOrZeroInRange(Mask, (i * Scale) + 1, Scale - 1); |
27001 | } |
27002 | if (Match) { |
27003 | unsigned SrcSize = std::max(128u, NumDstElts * MaskEltSize); |
27004 | SrcVT = MVT::getVectorVT(MaskVT.getScalarType(), SrcSize / MaskEltSize); |
27005 | if (SrcVT != MaskVT) |
27006 | V1 = extractSubVector(V1, 0, DAG, DL, SrcSize); |
27007 | DstVT = MVT::getIntegerVT(Scale * MaskEltSize); |
27008 | DstVT = MVT::getVectorVT(DstVT, NumDstElts); |
27009 | Shuffle = SrcVT != MaskVT ? unsigned(X86ISD::VZEXT) |
27010 | : unsigned(ISD::ZERO_EXTEND_VECTOR_INREG); |
27011 | return true; |
27012 | } |
27013 | } |
27014 | } |
27015 | |
27016 | // Match against a VZEXT_MOVL instruction, SSE1 only supports 32-bits (MOVSS). |
27017 | if (((MaskEltSize == 32) || (MaskEltSize == 64 && Subtarget.hasSSE2())) && |
27018 | isUndefOrEqual(Mask[0], 0) && |
27019 | isUndefOrZeroInRange(Mask, 1, NumMaskElts - 1)) { |
27020 | Shuffle = X86ISD::VZEXT_MOVL; |
27021 | SrcVT = DstVT = !Subtarget.hasSSE2() ? MVT::v4f32 : MaskVT; |
27022 | return true; |
27023 | } |
27024 | |
27025 | // Check if we have SSE3 which will let us use MOVDDUP etc. The |
27026 | // instructions are no slower than UNPCKLPD but has the option to |
27027 | // fold the input operand into even an unaligned memory load. |
27028 | if (MaskVT.is128BitVector() && Subtarget.hasSSE3() && AllowFloatDomain) { |
27029 | if (isTargetShuffleEquivalent(Mask, {0, 0})) { |
27030 | Shuffle = X86ISD::MOVDDUP; |
27031 | SrcVT = DstVT = MVT::v2f64; |
27032 | return true; |
27033 | } |
27034 | if (isTargetShuffleEquivalent(Mask, {0, 0, 2, 2})) { |
27035 | Shuffle = X86ISD::MOVSLDUP; |
27036 | SrcVT = DstVT = MVT::v4f32; |
27037 | return true; |
27038 | } |
27039 | if (isTargetShuffleEquivalent(Mask, {1, 1, 3, 3})) { |
27040 | Shuffle = X86ISD::MOVSHDUP; |
27041 | SrcVT = DstVT = MVT::v4f32; |
27042 | return true; |
27043 | } |
27044 | } |
27045 | |
27046 | if (MaskVT.is256BitVector() && AllowFloatDomain) { |
27047 | assert(Subtarget.hasAVX() && "AVX required for 256-bit vector shuffles")((Subtarget.hasAVX() && "AVX required for 256-bit vector shuffles" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX() && \"AVX required for 256-bit vector shuffles\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27047, __PRETTY_FUNCTION__)); |
27048 | if (isTargetShuffleEquivalent(Mask, {0, 0, 2, 2})) { |
27049 | Shuffle = X86ISD::MOVDDUP; |
27050 | SrcVT = DstVT = MVT::v4f64; |
27051 | return true; |
27052 | } |
27053 | if (isTargetShuffleEquivalent(Mask, {0, 0, 2, 2, 4, 4, 6, 6})) { |
27054 | Shuffle = X86ISD::MOVSLDUP; |
27055 | SrcVT = DstVT = MVT::v8f32; |
27056 | return true; |
27057 | } |
27058 | if (isTargetShuffleEquivalent(Mask, {1, 1, 3, 3, 5, 5, 7, 7})) { |
27059 | Shuffle = X86ISD::MOVSHDUP; |
27060 | SrcVT = DstVT = MVT::v8f32; |
27061 | return true; |
27062 | } |
27063 | } |
27064 | |
27065 | if (MaskVT.is512BitVector() && AllowFloatDomain) { |
27066 | assert(Subtarget.hasAVX512() &&((Subtarget.hasAVX512() && "AVX512 required for 512-bit vector shuffles" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX512() && \"AVX512 required for 512-bit vector shuffles\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27067, __PRETTY_FUNCTION__)) |
27067 | "AVX512 required for 512-bit vector shuffles")((Subtarget.hasAVX512() && "AVX512 required for 512-bit vector shuffles" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasAVX512() && \"AVX512 required for 512-bit vector shuffles\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27067, __PRETTY_FUNCTION__)); |
27068 | if (isTargetShuffleEquivalent(Mask, {0, 0, 2, 2, 4, 4, 6, 6})) { |
27069 | Shuffle = X86ISD::MOVDDUP; |
27070 | SrcVT = DstVT = MVT::v8f64; |
27071 | return true; |
27072 | } |
27073 | if (isTargetShuffleEquivalent( |
27074 | Mask, {0, 0, 2, 2, 4, 4, 6, 6, 8, 8, 10, 10, 12, 12, 14, 14})) { |
27075 | Shuffle = X86ISD::MOVSLDUP; |
27076 | SrcVT = DstVT = MVT::v16f32; |
27077 | return true; |
27078 | } |
27079 | if (isTargetShuffleEquivalent( |
27080 | Mask, {1, 1, 3, 3, 5, 5, 7, 7, 9, 9, 11, 11, 13, 13, 15, 15})) { |
27081 | Shuffle = X86ISD::MOVSHDUP; |
27082 | SrcVT = DstVT = MVT::v16f32; |
27083 | return true; |
27084 | } |
27085 | } |
27086 | |
27087 | // Attempt to match against broadcast-from-vector. |
27088 | if (Subtarget.hasAVX2()) { |
27089 | SmallVector<int, 64> BroadcastMask(NumMaskElts, 0); |
27090 | if (isTargetShuffleEquivalent(Mask, BroadcastMask)) { |
27091 | SrcVT = DstVT = MaskVT; |
27092 | Shuffle = X86ISD::VBROADCAST; |
27093 | return true; |
27094 | } |
27095 | } |
27096 | |
27097 | return false; |
27098 | } |
27099 | |
27100 | // Attempt to match a combined shuffle mask against supported unary immediate |
27101 | // permute instructions. |
27102 | // TODO: Investigate sharing more of this with shuffle lowering. |
27103 | static bool matchUnaryPermuteVectorShuffle(MVT MaskVT, ArrayRef<int> Mask, |
27104 | bool AllowFloatDomain, |
27105 | bool AllowIntDomain, |
27106 | const X86Subtarget &Subtarget, |
27107 | unsigned &Shuffle, MVT &ShuffleVT, |
27108 | unsigned &PermuteImm) { |
27109 | unsigned NumMaskElts = Mask.size(); |
27110 | |
27111 | bool ContainsZeros = false; |
27112 | APInt Zeroable(NumMaskElts, false); |
27113 | for (unsigned i = 0; i != NumMaskElts; ++i) { |
27114 | int M = Mask[i]; |
27115 | if (isUndefOrZero(M)) |
27116 | Zeroable.setBit(i); |
27117 | ContainsZeros |= (M == SM_SentinelZero); |
27118 | } |
27119 | |
27120 | // Attempt to match against byte/bit shifts. |
27121 | // FIXME: Add 512-bit support. |
27122 | if (AllowIntDomain && ((MaskVT.is128BitVector() && Subtarget.hasSSE2()) || |
27123 | (MaskVT.is256BitVector() && Subtarget.hasAVX2()))) { |
27124 | int ShiftAmt = matchVectorShuffleAsShift(ShuffleVT, Shuffle, |
27125 | MaskVT.getScalarSizeInBits(), Mask, |
27126 | 0, Zeroable, Subtarget); |
27127 | if (0 < ShiftAmt) { |
27128 | PermuteImm = (unsigned)ShiftAmt; |
27129 | return true; |
27130 | } |
27131 | } |
27132 | |
27133 | // Ensure we don't contain any zero elements. |
27134 | if (ContainsZeros) |
27135 | return false; |
27136 | |
27137 | assert(llvm::all_of(Mask, [&](int M) {((llvm::all_of(Mask, [&](int M) { return SM_SentinelUndef <= M && M < (int)NumMaskElts; }) && "Expected unary shuffle" ) ? static_cast<void> (0) : __assert_fail ("llvm::all_of(Mask, [&](int M) { return SM_SentinelUndef <= M && M < (int)NumMaskElts; }) && \"Expected unary shuffle\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27139, __PRETTY_FUNCTION__)) |
27138 | return SM_SentinelUndef <= M && M < (int)NumMaskElts;((llvm::all_of(Mask, [&](int M) { return SM_SentinelUndef <= M && M < (int)NumMaskElts; }) && "Expected unary shuffle" ) ? static_cast<void> (0) : __assert_fail ("llvm::all_of(Mask, [&](int M) { return SM_SentinelUndef <= M && M < (int)NumMaskElts; }) && \"Expected unary shuffle\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27139, __PRETTY_FUNCTION__)) |
27139 | }) && "Expected unary shuffle")((llvm::all_of(Mask, [&](int M) { return SM_SentinelUndef <= M && M < (int)NumMaskElts; }) && "Expected unary shuffle" ) ? static_cast<void> (0) : __assert_fail ("llvm::all_of(Mask, [&](int M) { return SM_SentinelUndef <= M && M < (int)NumMaskElts; }) && \"Expected unary shuffle\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27139, __PRETTY_FUNCTION__)); |
27140 | |
27141 | unsigned InputSizeInBits = MaskVT.getSizeInBits(); |
27142 | unsigned MaskScalarSizeInBits = InputSizeInBits / Mask.size(); |
27143 | MVT MaskEltVT = MVT::getIntegerVT(MaskScalarSizeInBits); |
27144 | |
27145 | // Handle PSHUFLW/PSHUFHW repeated patterns. |
27146 | if (MaskScalarSizeInBits == 16) { |
27147 | SmallVector<int, 4> RepeatedMask; |
27148 | if (is128BitLaneRepeatedShuffleMask(MaskEltVT, Mask, RepeatedMask)) { |
27149 | ArrayRef<int> LoMask(Mask.data() + 0, 4); |
27150 | ArrayRef<int> HiMask(Mask.data() + 4, 4); |
27151 | |
27152 | // PSHUFLW: permute lower 4 elements only. |
27153 | if (isUndefOrInRange(LoMask, 0, 4) && |
27154 | isSequentialOrUndefInRange(HiMask, 0, 4, 4)) { |
27155 | Shuffle = X86ISD::PSHUFLW; |
27156 | ShuffleVT = MVT::getVectorVT(MVT::i16, InputSizeInBits / 16); |
27157 | PermuteImm = getV4X86ShuffleImm(LoMask); |
27158 | return true; |
27159 | } |
27160 | |
27161 | // PSHUFHW: permute upper 4 elements only. |
27162 | if (isUndefOrInRange(HiMask, 4, 8) && |
27163 | isSequentialOrUndefInRange(LoMask, 0, 4, 0)) { |
27164 | // Offset the HiMask so that we can create the shuffle immediate. |
27165 | int OffsetHiMask[4]; |
27166 | for (int i = 0; i != 4; ++i) |
27167 | OffsetHiMask[i] = (HiMask[i] < 0 ? HiMask[i] : HiMask[i] - 4); |
27168 | |
27169 | Shuffle = X86ISD::PSHUFHW; |
27170 | ShuffleVT = MVT::getVectorVT(MVT::i16, InputSizeInBits / 16); |
27171 | PermuteImm = getV4X86ShuffleImm(OffsetHiMask); |
27172 | return true; |
27173 | } |
27174 | |
27175 | return false; |
27176 | } |
27177 | return false; |
27178 | } |
27179 | |
27180 | // We only support permutation of 32/64 bit elements after this. |
27181 | if (MaskScalarSizeInBits != 32 && MaskScalarSizeInBits != 64) |
27182 | return false; |
27183 | |
27184 | // AVX introduced the VPERMILPD/VPERMILPS float permutes, before then we |
27185 | // had to use 2-input SHUFPD/SHUFPS shuffles (not handled here). |
27186 | if ((AllowFloatDomain && !AllowIntDomain) && !Subtarget.hasAVX()) |
27187 | return false; |
27188 | |
27189 | // Pre-AVX2 we must use float shuffles on 256-bit vectors. |
27190 | if (MaskVT.is256BitVector() && !Subtarget.hasAVX2()) { |
27191 | AllowFloatDomain = true; |
27192 | AllowIntDomain = false; |
Value stored to 'AllowIntDomain' is never read | |
27193 | } |
27194 | |
27195 | // Check for lane crossing permutes. |
27196 | if (is128BitLaneCrossingShuffleMask(MaskEltVT, Mask)) { |
27197 | // PERMPD/PERMQ permutes within a 256-bit vector (AVX2+). |
27198 | if (Subtarget.hasAVX2() && MaskVT.is256BitVector() && Mask.size() == 4) { |
27199 | Shuffle = X86ISD::VPERMI; |
27200 | ShuffleVT = (AllowFloatDomain ? MVT::v4f64 : MVT::v4i64); |
27201 | PermuteImm = getV4X86ShuffleImm(Mask); |
27202 | return true; |
27203 | } |
27204 | if (Subtarget.hasAVX512() && MaskVT.is512BitVector() && Mask.size() == 8) { |
27205 | SmallVector<int, 4> RepeatedMask; |
27206 | if (is256BitLaneRepeatedShuffleMask(MVT::v8f64, Mask, RepeatedMask)) { |
27207 | Shuffle = X86ISD::VPERMI; |
27208 | ShuffleVT = (AllowFloatDomain ? MVT::v8f64 : MVT::v8i64); |
27209 | PermuteImm = getV4X86ShuffleImm(RepeatedMask); |
27210 | return true; |
27211 | } |
27212 | } |
27213 | return false; |
27214 | } |
27215 | |
27216 | // VPERMILPD can permute with a non-repeating shuffle. |
27217 | if (AllowFloatDomain && MaskScalarSizeInBits == 64) { |
27218 | Shuffle = X86ISD::VPERMILPI; |
27219 | ShuffleVT = MVT::getVectorVT(MVT::f64, Mask.size()); |
27220 | PermuteImm = 0; |
27221 | for (int i = 0, e = Mask.size(); i != e; ++i) { |
27222 | int M = Mask[i]; |
27223 | if (M == SM_SentinelUndef) |
27224 | continue; |
27225 | assert(((M / 2) == (i / 2)) && "Out of range shuffle mask index")((((M / 2) == (i / 2)) && "Out of range shuffle mask index" ) ? static_cast<void> (0) : __assert_fail ("((M / 2) == (i / 2)) && \"Out of range shuffle mask index\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27225, __PRETTY_FUNCTION__)); |
27226 | PermuteImm |= (M & 1) << i; |
27227 | } |
27228 | return true; |
27229 | } |
27230 | |
27231 | // We need a repeating shuffle mask for VPERMILPS/PSHUFD. |
27232 | SmallVector<int, 4> RepeatedMask; |
27233 | if (!is128BitLaneRepeatedShuffleMask(MaskEltVT, Mask, RepeatedMask)) |
27234 | return false; |
27235 | |
27236 | // Narrow the repeated mask for 32-bit element permutes. |
27237 | SmallVector<int, 4> WordMask = RepeatedMask; |
27238 | if (MaskScalarSizeInBits == 64) |
27239 | scaleShuffleMask(2, RepeatedMask, WordMask); |
27240 | |
27241 | Shuffle = (AllowFloatDomain ? X86ISD::VPERMILPI : X86ISD::PSHUFD); |
27242 | ShuffleVT = (AllowFloatDomain ? MVT::f32 : MVT::i32); |
27243 | ShuffleVT = MVT::getVectorVT(ShuffleVT, InputSizeInBits / 32); |
27244 | PermuteImm = getV4X86ShuffleImm(WordMask); |
27245 | return true; |
27246 | } |
27247 | |
27248 | // Attempt to match a combined unary shuffle mask against supported binary |
27249 | // shuffle instructions. |
27250 | // TODO: Investigate sharing more of this with shuffle lowering. |
27251 | static bool matchBinaryVectorShuffle(MVT MaskVT, ArrayRef<int> Mask, |
27252 | bool AllowFloatDomain, bool AllowIntDomain, |
27253 | SDValue &V1, SDValue &V2, SDLoc &DL, |
27254 | SelectionDAG &DAG, |
27255 | const X86Subtarget &Subtarget, |
27256 | unsigned &Shuffle, MVT &ShuffleVT, |
27257 | bool IsUnary) { |
27258 | unsigned EltSizeInBits = MaskVT.getScalarSizeInBits(); |
27259 | |
27260 | if (MaskVT.is128BitVector()) { |
27261 | if (isTargetShuffleEquivalent(Mask, {0, 0}) && AllowFloatDomain) { |
27262 | V2 = V1; |
27263 | Shuffle = X86ISD::MOVLHPS; |
27264 | ShuffleVT = MVT::v4f32; |
27265 | return true; |
27266 | } |
27267 | if (isTargetShuffleEquivalent(Mask, {1, 1}) && AllowFloatDomain) { |
27268 | V2 = V1; |
27269 | Shuffle = X86ISD::MOVHLPS; |
27270 | ShuffleVT = MVT::v4f32; |
27271 | return true; |
27272 | } |
27273 | if (isTargetShuffleEquivalent(Mask, {0, 3}) && Subtarget.hasSSE2() && |
27274 | (AllowFloatDomain || !Subtarget.hasSSE41())) { |
27275 | std::swap(V1, V2); |
27276 | Shuffle = X86ISD::MOVSD; |
27277 | ShuffleVT = MaskVT; |
27278 | return true; |
27279 | } |
27280 | if (isTargetShuffleEquivalent(Mask, {4, 1, 2, 3}) && |
27281 | (AllowFloatDomain || !Subtarget.hasSSE41())) { |
27282 | Shuffle = X86ISD::MOVSS; |
27283 | ShuffleVT = MaskVT; |
27284 | return true; |
27285 | } |
27286 | } |
27287 | |
27288 | // Attempt to match against either a unary or binary UNPCKL/UNPCKH shuffle. |
27289 | if ((MaskVT == MVT::v4f32 && Subtarget.hasSSE1()) || |
27290 | (MaskVT.is128BitVector() && Subtarget.hasSSE2()) || |
27291 | (MaskVT.is256BitVector() && 32 <= EltSizeInBits && Subtarget.hasAVX()) || |
27292 | (MaskVT.is256BitVector() && Subtarget.hasAVX2()) || |
27293 | (MaskVT.is512BitVector() && Subtarget.hasAVX512())) { |
27294 | if (matchVectorShuffleWithUNPCK(MaskVT, V1, V2, Shuffle, IsUnary, Mask, DL, |
27295 | DAG, Subtarget)) { |
27296 | ShuffleVT = MaskVT; |
27297 | if (ShuffleVT.is256BitVector() && !Subtarget.hasAVX2()) |
27298 | ShuffleVT = (32 == EltSizeInBits ? MVT::v8f32 : MVT::v4f64); |
27299 | return true; |
27300 | } |
27301 | } |
27302 | |
27303 | return false; |
27304 | } |
27305 | |
27306 | static bool matchBinaryPermuteVectorShuffle(MVT MaskVT, ArrayRef<int> Mask, |
27307 | bool AllowFloatDomain, |
27308 | bool AllowIntDomain, |
27309 | SDValue &V1, SDValue &V2, SDLoc &DL, |
27310 | SelectionDAG &DAG, |
27311 | const X86Subtarget &Subtarget, |
27312 | unsigned &Shuffle, MVT &ShuffleVT, |
27313 | unsigned &PermuteImm) { |
27314 | unsigned NumMaskElts = Mask.size(); |
27315 | unsigned EltSizeInBits = MaskVT.getScalarSizeInBits(); |
27316 | |
27317 | // Attempt to match against PALIGNR byte rotate. |
27318 | if (AllowIntDomain && ((MaskVT.is128BitVector() && Subtarget.hasSSSE3()) || |
27319 | (MaskVT.is256BitVector() && Subtarget.hasAVX2()))) { |
27320 | int ByteRotation = matchVectorShuffleAsByteRotate(MaskVT, V1, V2, Mask); |
27321 | if (0 < ByteRotation) { |
27322 | Shuffle = X86ISD::PALIGNR; |
27323 | ShuffleVT = MVT::getVectorVT(MVT::i8, MaskVT.getSizeInBits() / 8); |
27324 | PermuteImm = ByteRotation; |
27325 | return true; |
27326 | } |
27327 | } |
27328 | |
27329 | // Attempt to combine to X86ISD::BLENDI. |
27330 | if ((NumMaskElts <= 8 && ((Subtarget.hasSSE41() && MaskVT.is128BitVector()) || |
27331 | (Subtarget.hasAVX() && MaskVT.is256BitVector()))) || |
27332 | (MaskVT == MVT::v16i16 && Subtarget.hasAVX2())) { |
27333 | uint64_t BlendMask = 0; |
27334 | bool ForceV1Zero = false, ForceV2Zero = false; |
27335 | SmallVector<int, 8> TargetMask(Mask.begin(), Mask.end()); |
27336 | if (matchVectorShuffleAsBlend(V1, V2, TargetMask, ForceV1Zero, ForceV2Zero, |
27337 | BlendMask)) { |
27338 | if (MaskVT == MVT::v16i16) { |
27339 | // We can only use v16i16 PBLENDW if the lanes are repeated. |
27340 | SmallVector<int, 8> RepeatedMask; |
27341 | if (isRepeatedTargetShuffleMask(128, MaskVT, TargetMask, |
27342 | RepeatedMask)) { |
27343 | assert(RepeatedMask.size() == 8 &&((RepeatedMask.size() == 8 && "Repeated mask size doesn't match!" ) ? static_cast<void> (0) : __assert_fail ("RepeatedMask.size() == 8 && \"Repeated mask size doesn't match!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27344, __PRETTY_FUNCTION__)) |
27344 | "Repeated mask size doesn't match!")((RepeatedMask.size() == 8 && "Repeated mask size doesn't match!" ) ? static_cast<void> (0) : __assert_fail ("RepeatedMask.size() == 8 && \"Repeated mask size doesn't match!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27344, __PRETTY_FUNCTION__)); |
27345 | PermuteImm = 0; |
27346 | for (int i = 0; i < 8; ++i) |
27347 | if (RepeatedMask[i] >= 8) |
27348 | PermuteImm |= 1 << i; |
27349 | V1 = ForceV1Zero ? getZeroVector(MaskVT, Subtarget, DAG, DL) : V1; |
27350 | V2 = ForceV2Zero ? getZeroVector(MaskVT, Subtarget, DAG, DL) : V2; |
27351 | Shuffle = X86ISD::BLENDI; |
27352 | ShuffleVT = MaskVT; |
27353 | return true; |
27354 | } |
27355 | } else { |
27356 | // Determine a type compatible with X86ISD::BLENDI. |
27357 | ShuffleVT = MaskVT; |
27358 | if (Subtarget.hasAVX2()) { |
27359 | if (ShuffleVT == MVT::v4i64) |
27360 | ShuffleVT = MVT::v8i32; |
27361 | else if (ShuffleVT == MVT::v2i64) |
27362 | ShuffleVT = MVT::v4i32; |
27363 | } else { |
27364 | if (ShuffleVT == MVT::v2i64 || ShuffleVT == MVT::v4i32) |
27365 | ShuffleVT = MVT::v8i16; |
27366 | else if (ShuffleVT == MVT::v4i64) |
27367 | ShuffleVT = MVT::v4f64; |
27368 | else if (ShuffleVT == MVT::v8i32) |
27369 | ShuffleVT = MVT::v8f32; |
27370 | } |
27371 | |
27372 | if (!ShuffleVT.isFloatingPoint()) { |
27373 | int Scale = EltSizeInBits / ShuffleVT.getScalarSizeInBits(); |
27374 | BlendMask = |
27375 | scaleVectorShuffleBlendMask(BlendMask, NumMaskElts, Scale); |
27376 | ShuffleVT = MVT::getIntegerVT(EltSizeInBits / Scale); |
27377 | ShuffleVT = MVT::getVectorVT(ShuffleVT, NumMaskElts * Scale); |
27378 | } |
27379 | |
27380 | V1 = ForceV1Zero ? getZeroVector(MaskVT, Subtarget, DAG, DL) : V1; |
27381 | V2 = ForceV2Zero ? getZeroVector(MaskVT, Subtarget, DAG, DL) : V2; |
27382 | PermuteImm = (unsigned)BlendMask; |
27383 | Shuffle = X86ISD::BLENDI; |
27384 | return true; |
27385 | } |
27386 | } |
27387 | } |
27388 | |
27389 | // Attempt to combine to INSERTPS. |
27390 | if (AllowFloatDomain && EltSizeInBits == 32 && Subtarget.hasSSE41() && |
27391 | MaskVT.is128BitVector()) { |
27392 | APInt Zeroable(4, 0); |
27393 | for (unsigned i = 0; i != NumMaskElts; ++i) |
27394 | if (Mask[i] < 0) |
27395 | Zeroable.setBit(i); |
27396 | |
27397 | if (Zeroable.getBoolValue() && |
27398 | matchVectorShuffleAsInsertPS(V1, V2, PermuteImm, Zeroable, Mask, DAG)) { |
27399 | Shuffle = X86ISD::INSERTPS; |
27400 | ShuffleVT = MVT::v4f32; |
27401 | return true; |
27402 | } |
27403 | } |
27404 | |
27405 | // Attempt to combine to SHUFPD. |
27406 | if (AllowFloatDomain && EltSizeInBits == 64 && |
27407 | ((MaskVT.is128BitVector() && Subtarget.hasSSE2()) || |
27408 | (MaskVT.is256BitVector() && Subtarget.hasAVX()) || |
27409 | (MaskVT.is512BitVector() && Subtarget.hasAVX512()))) { |
27410 | if (matchVectorShuffleWithSHUFPD(MaskVT, V1, V2, PermuteImm, Mask)) { |
27411 | Shuffle = X86ISD::SHUFP; |
27412 | ShuffleVT = MVT::getVectorVT(MVT::f64, MaskVT.getSizeInBits() / 64); |
27413 | return true; |
27414 | } |
27415 | } |
27416 | |
27417 | // Attempt to combine to SHUFPS. |
27418 | if (AllowFloatDomain && EltSizeInBits == 32 && |
27419 | ((MaskVT.is128BitVector() && Subtarget.hasSSE1()) || |
27420 | (MaskVT.is256BitVector() && Subtarget.hasAVX()) || |
27421 | (MaskVT.is512BitVector() && Subtarget.hasAVX512()))) { |
27422 | SmallVector<int, 4> RepeatedMask; |
27423 | if (isRepeatedTargetShuffleMask(128, MaskVT, Mask, RepeatedMask)) { |
27424 | // Match each half of the repeated mask, to determine if its just |
27425 | // referencing one of the vectors, is zeroable or entirely undef. |
27426 | auto MatchHalf = [&](unsigned Offset, int &S0, int &S1) { |
27427 | int M0 = RepeatedMask[Offset]; |
27428 | int M1 = RepeatedMask[Offset + 1]; |
27429 | |
27430 | if (isUndefInRange(RepeatedMask, Offset, 2)) { |
27431 | return DAG.getUNDEF(MaskVT); |
27432 | } else if (isUndefOrZeroInRange(RepeatedMask, Offset, 2)) { |
27433 | S0 = (SM_SentinelUndef == M0 ? -1 : 0); |
27434 | S1 = (SM_SentinelUndef == M1 ? -1 : 1); |
27435 | return getZeroVector(MaskVT, Subtarget, DAG, DL); |
27436 | } else if (isUndefOrInRange(M0, 0, 4) && isUndefOrInRange(M1, 0, 4)) { |
27437 | S0 = (SM_SentinelUndef == M0 ? -1 : M0 & 3); |
27438 | S1 = (SM_SentinelUndef == M1 ? -1 : M1 & 3); |
27439 | return V1; |
27440 | } else if (isUndefOrInRange(M0, 4, 8) && isUndefOrInRange(M1, 4, 8)) { |
27441 | S0 = (SM_SentinelUndef == M0 ? -1 : M0 & 3); |
27442 | S1 = (SM_SentinelUndef == M1 ? -1 : M1 & 3); |
27443 | return V2; |
27444 | } |
27445 | |
27446 | return SDValue(); |
27447 | }; |
27448 | |
27449 | int ShufMask[4] = {-1, -1, -1, -1}; |
27450 | SDValue Lo = MatchHalf(0, ShufMask[0], ShufMask[1]); |
27451 | SDValue Hi = MatchHalf(2, ShufMask[2], ShufMask[3]); |
27452 | |
27453 | if (Lo && Hi) { |
27454 | V1 = Lo; |
27455 | V2 = Hi; |
27456 | Shuffle = X86ISD::SHUFP; |
27457 | ShuffleVT = MVT::getVectorVT(MVT::f32, MaskVT.getSizeInBits() / 32); |
27458 | PermuteImm = getV4X86ShuffleImm(ShufMask); |
27459 | return true; |
27460 | } |
27461 | } |
27462 | } |
27463 | |
27464 | return false; |
27465 | } |
27466 | |
27467 | /// \brief Combine an arbitrary chain of shuffles into a single instruction if |
27468 | /// possible. |
27469 | /// |
27470 | /// This is the leaf of the recursive combine below. When we have found some |
27471 | /// chain of single-use x86 shuffle instructions and accumulated the combined |
27472 | /// shuffle mask represented by them, this will try to pattern match that mask |
27473 | /// into either a single instruction if there is a special purpose instruction |
27474 | /// for this operation, or into a PSHUFB instruction which is a fully general |
27475 | /// instruction but should only be used to replace chains over a certain depth. |
27476 | static bool combineX86ShuffleChain(ArrayRef<SDValue> Inputs, SDValue Root, |
27477 | ArrayRef<int> BaseMask, int Depth, |
27478 | bool HasVariableMask, SelectionDAG &DAG, |
27479 | TargetLowering::DAGCombinerInfo &DCI, |
27480 | const X86Subtarget &Subtarget) { |
27481 | assert(!BaseMask.empty() && "Cannot combine an empty shuffle mask!")((!BaseMask.empty() && "Cannot combine an empty shuffle mask!" ) ? static_cast<void> (0) : __assert_fail ("!BaseMask.empty() && \"Cannot combine an empty shuffle mask!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27481, __PRETTY_FUNCTION__)); |
27482 | assert((Inputs.size() == 1 || Inputs.size() == 2) &&(((Inputs.size() == 1 || Inputs.size() == 2) && "Unexpected number of shuffle inputs!" ) ? static_cast<void> (0) : __assert_fail ("(Inputs.size() == 1 || Inputs.size() == 2) && \"Unexpected number of shuffle inputs!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27483, __PRETTY_FUNCTION__)) |
27483 | "Unexpected number of shuffle inputs!")(((Inputs.size() == 1 || Inputs.size() == 2) && "Unexpected number of shuffle inputs!" ) ? static_cast<void> (0) : __assert_fail ("(Inputs.size() == 1 || Inputs.size() == 2) && \"Unexpected number of shuffle inputs!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27483, __PRETTY_FUNCTION__)); |
27484 | |
27485 | // Find the inputs that enter the chain. Note that multiple uses are OK |
27486 | // here, we're not going to remove the operands we find. |
27487 | bool UnaryShuffle = (Inputs.size() == 1); |
27488 | SDValue V1 = peekThroughBitcasts(Inputs[0]); |
27489 | SDValue V2 = (UnaryShuffle ? DAG.getUNDEF(V1.getValueType()) |
27490 | : peekThroughBitcasts(Inputs[1])); |
27491 | |
27492 | MVT VT1 = V1.getSimpleValueType(); |
27493 | MVT VT2 = V2.getSimpleValueType(); |
27494 | MVT RootVT = Root.getSimpleValueType(); |
27495 | assert(VT1.getSizeInBits() == RootVT.getSizeInBits() &&((VT1.getSizeInBits() == RootVT.getSizeInBits() && VT2 .getSizeInBits() == RootVT.getSizeInBits() && "Vector size mismatch" ) ? static_cast<void> (0) : __assert_fail ("VT1.getSizeInBits() == RootVT.getSizeInBits() && VT2.getSizeInBits() == RootVT.getSizeInBits() && \"Vector size mismatch\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27497, __PRETTY_FUNCTION__)) |
27496 | VT2.getSizeInBits() == RootVT.getSizeInBits() &&((VT1.getSizeInBits() == RootVT.getSizeInBits() && VT2 .getSizeInBits() == RootVT.getSizeInBits() && "Vector size mismatch" ) ? static_cast<void> (0) : __assert_fail ("VT1.getSizeInBits() == RootVT.getSizeInBits() && VT2.getSizeInBits() == RootVT.getSizeInBits() && \"Vector size mismatch\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27497, __PRETTY_FUNCTION__)) |
27497 | "Vector size mismatch")((VT1.getSizeInBits() == RootVT.getSizeInBits() && VT2 .getSizeInBits() == RootVT.getSizeInBits() && "Vector size mismatch" ) ? static_cast<void> (0) : __assert_fail ("VT1.getSizeInBits() == RootVT.getSizeInBits() && VT2.getSizeInBits() == RootVT.getSizeInBits() && \"Vector size mismatch\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27497, __PRETTY_FUNCTION__)); |
27498 | |
27499 | SDLoc DL(Root); |
27500 | SDValue Res; |
27501 | |
27502 | unsigned NumBaseMaskElts = BaseMask.size(); |
27503 | if (NumBaseMaskElts == 1) { |
27504 | assert(BaseMask[0] == 0 && "Invalid shuffle index found!")((BaseMask[0] == 0 && "Invalid shuffle index found!") ? static_cast<void> (0) : __assert_fail ("BaseMask[0] == 0 && \"Invalid shuffle index found!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27504, __PRETTY_FUNCTION__)); |
27505 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, V1), |
27506 | /*AddTo*/ true); |
27507 | return true; |
27508 | } |
27509 | |
27510 | unsigned RootSizeInBits = RootVT.getSizeInBits(); |
27511 | unsigned NumRootElts = RootVT.getVectorNumElements(); |
27512 | unsigned BaseMaskEltSizeInBits = RootSizeInBits / NumBaseMaskElts; |
27513 | bool FloatDomain = VT1.isFloatingPoint() || VT2.isFloatingPoint() || |
27514 | (RootVT.is256BitVector() && !Subtarget.hasAVX2()); |
27515 | |
27516 | // Don't combine if we are a AVX512/EVEX target and the mask element size |
27517 | // is different from the root element size - this would prevent writemasks |
27518 | // from being reused. |
27519 | // TODO - this currently prevents all lane shuffles from occurring. |
27520 | // TODO - check for writemasks usage instead of always preventing combining. |
27521 | // TODO - attempt to narrow Mask back to writemask size. |
27522 | bool IsEVEXShuffle = |
27523 | RootSizeInBits == 512 || (Subtarget.hasVLX() && RootSizeInBits >= 128); |
27524 | if (IsEVEXShuffle && (RootVT.getScalarSizeInBits() != BaseMaskEltSizeInBits)) |
27525 | return false; |
27526 | |
27527 | // TODO - handle 128/256-bit lane shuffles of 512-bit vectors. |
27528 | |
27529 | // Handle 128-bit lane shuffles of 256-bit vectors. |
27530 | // TODO - this should support binary shuffles. |
27531 | if (UnaryShuffle && RootVT.is256BitVector() && NumBaseMaskElts == 2 && |
27532 | !isSequentialOrUndefOrZeroInRange(BaseMask, 0, 2, 0)) { |
27533 | if (Depth == 1 && Root.getOpcode() == X86ISD::VPERM2X128) |
27534 | return false; // Nothing to do! |
27535 | MVT ShuffleVT = (FloatDomain ? MVT::v4f64 : MVT::v4i64); |
27536 | unsigned PermMask = 0; |
27537 | PermMask |= ((BaseMask[0] < 0 ? 0x8 : (BaseMask[0] & 1)) << 0); |
27538 | PermMask |= ((BaseMask[1] < 0 ? 0x8 : (BaseMask[1] & 1)) << 4); |
27539 | |
27540 | Res = DAG.getBitcast(ShuffleVT, V1); |
27541 | DCI.AddToWorklist(Res.getNode()); |
27542 | Res = DAG.getNode(X86ISD::VPERM2X128, DL, ShuffleVT, Res, |
27543 | DAG.getUNDEF(ShuffleVT), |
27544 | DAG.getConstant(PermMask, DL, MVT::i8)); |
27545 | DCI.AddToWorklist(Res.getNode()); |
27546 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res), |
27547 | /*AddTo*/ true); |
27548 | return true; |
27549 | } |
27550 | |
27551 | // For masks that have been widened to 128-bit elements or more, |
27552 | // narrow back down to 64-bit elements. |
27553 | SmallVector<int, 64> Mask; |
27554 | if (BaseMaskEltSizeInBits > 64) { |
27555 | assert((BaseMaskEltSizeInBits % 64) == 0 && "Illegal mask size")(((BaseMaskEltSizeInBits % 64) == 0 && "Illegal mask size" ) ? static_cast<void> (0) : __assert_fail ("(BaseMaskEltSizeInBits % 64) == 0 && \"Illegal mask size\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27555, __PRETTY_FUNCTION__)); |
27556 | int MaskScale = BaseMaskEltSizeInBits / 64; |
27557 | scaleShuffleMask(MaskScale, BaseMask, Mask); |
27558 | } else { |
27559 | Mask = SmallVector<int, 64>(BaseMask.begin(), BaseMask.end()); |
27560 | } |
27561 | |
27562 | unsigned NumMaskElts = Mask.size(); |
27563 | unsigned MaskEltSizeInBits = RootSizeInBits / NumMaskElts; |
27564 | |
27565 | // Determine the effective mask value type. |
27566 | FloatDomain &= (32 <= MaskEltSizeInBits); |
27567 | MVT MaskVT = FloatDomain ? MVT::getFloatingPointVT(MaskEltSizeInBits) |
27568 | : MVT::getIntegerVT(MaskEltSizeInBits); |
27569 | MaskVT = MVT::getVectorVT(MaskVT, NumMaskElts); |
27570 | |
27571 | // Only allow legal mask types. |
27572 | if (!DAG.getTargetLoweringInfo().isTypeLegal(MaskVT)) |
27573 | return false; |
27574 | |
27575 | // Attempt to match the mask against known shuffle patterns. |
27576 | MVT ShuffleSrcVT, ShuffleVT; |
27577 | unsigned Shuffle, PermuteImm; |
27578 | |
27579 | // Which shuffle domains are permitted? |
27580 | // Permit domain crossing at higher combine depths. |
27581 | bool AllowFloatDomain = FloatDomain || (Depth > 3); |
27582 | bool AllowIntDomain = !FloatDomain || (Depth > 3); |
27583 | |
27584 | if (UnaryShuffle) { |
27585 | // If we are shuffling a X86ISD::VZEXT_LOAD then we can use the load |
27586 | // directly if we don't shuffle the lower element and we shuffle the upper |
27587 | // (zero) elements within themselves. |
27588 | if (V1.getOpcode() == X86ISD::VZEXT_LOAD && |
27589 | (V1.getScalarValueSizeInBits() % MaskEltSizeInBits) == 0) { |
27590 | unsigned Scale = V1.getScalarValueSizeInBits() / MaskEltSizeInBits; |
27591 | ArrayRef<int> HiMask(Mask.data() + Scale, NumMaskElts - Scale); |
27592 | if (isSequentialOrUndefInRange(Mask, 0, Scale, 0) && |
27593 | isUndefOrZeroOrInRange(HiMask, Scale, NumMaskElts)) { |
27594 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, V1), |
27595 | /*AddTo*/ true); |
27596 | return true; |
27597 | } |
27598 | } |
27599 | |
27600 | if (matchUnaryVectorShuffle(MaskVT, Mask, AllowFloatDomain, AllowIntDomain, |
27601 | V1, DL, DAG, Subtarget, Shuffle, ShuffleSrcVT, |
27602 | ShuffleVT)) { |
27603 | if (Depth == 1 && Root.getOpcode() == Shuffle) |
27604 | return false; // Nothing to do! |
27605 | if (IsEVEXShuffle && (NumRootElts != ShuffleVT.getVectorNumElements())) |
27606 | return false; // AVX512 Writemask clash. |
27607 | Res = DAG.getBitcast(ShuffleSrcVT, V1); |
27608 | DCI.AddToWorklist(Res.getNode()); |
27609 | Res = DAG.getNode(Shuffle, DL, ShuffleVT, Res); |
27610 | DCI.AddToWorklist(Res.getNode()); |
27611 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res), |
27612 | /*AddTo*/ true); |
27613 | return true; |
27614 | } |
27615 | |
27616 | if (matchUnaryPermuteVectorShuffle(MaskVT, Mask, AllowFloatDomain, |
27617 | AllowIntDomain, Subtarget, Shuffle, |
27618 | ShuffleVT, PermuteImm)) { |
27619 | if (Depth == 1 && Root.getOpcode() == Shuffle) |
27620 | return false; // Nothing to do! |
27621 | if (IsEVEXShuffle && (NumRootElts != ShuffleVT.getVectorNumElements())) |
27622 | return false; // AVX512 Writemask clash. |
27623 | Res = DAG.getBitcast(ShuffleVT, V1); |
27624 | DCI.AddToWorklist(Res.getNode()); |
27625 | Res = DAG.getNode(Shuffle, DL, ShuffleVT, Res, |
27626 | DAG.getConstant(PermuteImm, DL, MVT::i8)); |
27627 | DCI.AddToWorklist(Res.getNode()); |
27628 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res), |
27629 | /*AddTo*/ true); |
27630 | return true; |
27631 | } |
27632 | } |
27633 | |
27634 | if (matchBinaryVectorShuffle(MaskVT, Mask, AllowFloatDomain, AllowIntDomain, |
27635 | V1, V2, DL, DAG, Subtarget, Shuffle, ShuffleVT, |
27636 | UnaryShuffle)) { |
27637 | if (Depth == 1 && Root.getOpcode() == Shuffle) |
27638 | return false; // Nothing to do! |
27639 | if (IsEVEXShuffle && (NumRootElts != ShuffleVT.getVectorNumElements())) |
27640 | return false; // AVX512 Writemask clash. |
27641 | V1 = DAG.getBitcast(ShuffleVT, V1); |
27642 | DCI.AddToWorklist(V1.getNode()); |
27643 | V2 = DAG.getBitcast(ShuffleVT, V2); |
27644 | DCI.AddToWorklist(V2.getNode()); |
27645 | Res = DAG.getNode(Shuffle, DL, ShuffleVT, V1, V2); |
27646 | DCI.AddToWorklist(Res.getNode()); |
27647 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res), |
27648 | /*AddTo*/ true); |
27649 | return true; |
27650 | } |
27651 | |
27652 | if (matchBinaryPermuteVectorShuffle(MaskVT, Mask, AllowFloatDomain, |
27653 | AllowIntDomain, V1, V2, DL, DAG, |
27654 | Subtarget, Shuffle, ShuffleVT, |
27655 | PermuteImm)) { |
27656 | if (Depth == 1 && Root.getOpcode() == Shuffle) |
27657 | return false; // Nothing to do! |
27658 | if (IsEVEXShuffle && (NumRootElts != ShuffleVT.getVectorNumElements())) |
27659 | return false; // AVX512 Writemask clash. |
27660 | V1 = DAG.getBitcast(ShuffleVT, V1); |
27661 | DCI.AddToWorklist(V1.getNode()); |
27662 | V2 = DAG.getBitcast(ShuffleVT, V2); |
27663 | DCI.AddToWorklist(V2.getNode()); |
27664 | Res = DAG.getNode(Shuffle, DL, ShuffleVT, V1, V2, |
27665 | DAG.getConstant(PermuteImm, DL, MVT::i8)); |
27666 | DCI.AddToWorklist(Res.getNode()); |
27667 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res), |
27668 | /*AddTo*/ true); |
27669 | return true; |
27670 | } |
27671 | |
27672 | // Don't try to re-form single instruction chains under any circumstances now |
27673 | // that we've done encoding canonicalization for them. |
27674 | if (Depth < 2) |
27675 | return false; |
27676 | |
27677 | bool MaskContainsZeros = |
27678 | any_of(Mask, [](int M) { return M == SM_SentinelZero; }); |
27679 | |
27680 | if (is128BitLaneCrossingShuffleMask(MaskVT, Mask)) { |
27681 | // If we have a single input lane-crossing shuffle then lower to VPERMV. |
27682 | if (UnaryShuffle && (Depth >= 3 || HasVariableMask) && !MaskContainsZeros && |
27683 | ((Subtarget.hasAVX2() && |
27684 | (MaskVT == MVT::v8f32 || MaskVT == MVT::v8i32)) || |
27685 | (Subtarget.hasAVX512() && |
27686 | (MaskVT == MVT::v8f64 || MaskVT == MVT::v8i64 || |
27687 | MaskVT == MVT::v16f32 || MaskVT == MVT::v16i32)) || |
27688 | (Subtarget.hasBWI() && MaskVT == MVT::v32i16) || |
27689 | (Subtarget.hasBWI() && Subtarget.hasVLX() && MaskVT == MVT::v16i16) || |
27690 | (Subtarget.hasVBMI() && MaskVT == MVT::v64i8) || |
27691 | (Subtarget.hasVBMI() && Subtarget.hasVLX() && MaskVT == MVT::v32i8))) { |
27692 | MVT VPermMaskSVT = MVT::getIntegerVT(MaskEltSizeInBits); |
27693 | MVT VPermMaskVT = MVT::getVectorVT(VPermMaskSVT, NumMaskElts); |
27694 | SDValue VPermMask = getConstVector(Mask, VPermMaskVT, DAG, DL, true); |
27695 | DCI.AddToWorklist(VPermMask.getNode()); |
27696 | Res = DAG.getBitcast(MaskVT, V1); |
27697 | DCI.AddToWorklist(Res.getNode()); |
27698 | Res = DAG.getNode(X86ISD::VPERMV, DL, MaskVT, VPermMask, Res); |
27699 | DCI.AddToWorklist(Res.getNode()); |
27700 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res), |
27701 | /*AddTo*/ true); |
27702 | return true; |
27703 | } |
27704 | |
27705 | // Lower a unary+zero lane-crossing shuffle as VPERMV3 with a zero |
27706 | // vector as the second source. |
27707 | if (UnaryShuffle && (Depth >= 3 || HasVariableMask) && |
27708 | ((Subtarget.hasAVX512() && |
27709 | (MaskVT == MVT::v8f64 || MaskVT == MVT::v8i64 || |
27710 | MaskVT == MVT::v16f32 || MaskVT == MVT::v16i32)) || |
27711 | (Subtarget.hasVLX() && |
27712 | (MaskVT == MVT::v4f64 || MaskVT == MVT::v4i64 || |
27713 | MaskVT == MVT::v8f32 || MaskVT == MVT::v8i32)) || |
27714 | (Subtarget.hasBWI() && MaskVT == MVT::v32i16) || |
27715 | (Subtarget.hasBWI() && Subtarget.hasVLX() && MaskVT == MVT::v16i16) || |
27716 | (Subtarget.hasVBMI() && MaskVT == MVT::v64i8) || |
27717 | (Subtarget.hasVBMI() && Subtarget.hasVLX() && MaskVT == MVT::v32i8))) { |
27718 | // Adjust shuffle mask - replace SM_SentinelZero with second source index. |
27719 | for (unsigned i = 0; i != NumMaskElts; ++i) |
27720 | if (Mask[i] == SM_SentinelZero) |
27721 | Mask[i] = NumMaskElts + i; |
27722 | |
27723 | MVT VPermMaskSVT = MVT::getIntegerVT(MaskEltSizeInBits); |
27724 | MVT VPermMaskVT = MVT::getVectorVT(VPermMaskSVT, NumMaskElts); |
27725 | SDValue VPermMask = getConstVector(Mask, VPermMaskVT, DAG, DL, true); |
27726 | DCI.AddToWorklist(VPermMask.getNode()); |
27727 | Res = DAG.getBitcast(MaskVT, V1); |
27728 | DCI.AddToWorklist(Res.getNode()); |
27729 | SDValue Zero = getZeroVector(MaskVT, Subtarget, DAG, DL); |
27730 | DCI.AddToWorklist(Zero.getNode()); |
27731 | Res = DAG.getNode(X86ISD::VPERMV3, DL, MaskVT, Res, VPermMask, Zero); |
27732 | DCI.AddToWorklist(Res.getNode()); |
27733 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res), |
27734 | /*AddTo*/ true); |
27735 | return true; |
27736 | } |
27737 | |
27738 | // If we have a dual input lane-crossing shuffle then lower to VPERMV3. |
27739 | if ((Depth >= 3 || HasVariableMask) && !MaskContainsZeros && |
27740 | ((Subtarget.hasAVX512() && |
27741 | (MaskVT == MVT::v8f64 || MaskVT == MVT::v8i64 || |
27742 | MaskVT == MVT::v16f32 || MaskVT == MVT::v16i32)) || |
27743 | (Subtarget.hasVLX() && |
27744 | (MaskVT == MVT::v4f64 || MaskVT == MVT::v4i64 || |
27745 | MaskVT == MVT::v8f32 || MaskVT == MVT::v8i32)) || |
27746 | (Subtarget.hasBWI() && MaskVT == MVT::v32i16) || |
27747 | (Subtarget.hasBWI() && Subtarget.hasVLX() && MaskVT == MVT::v16i16) || |
27748 | (Subtarget.hasVBMI() && MaskVT == MVT::v64i8) || |
27749 | (Subtarget.hasVBMI() && Subtarget.hasVLX() && MaskVT == MVT::v32i8))) { |
27750 | MVT VPermMaskSVT = MVT::getIntegerVT(MaskEltSizeInBits); |
27751 | MVT VPermMaskVT = MVT::getVectorVT(VPermMaskSVT, NumMaskElts); |
27752 | SDValue VPermMask = getConstVector(Mask, VPermMaskVT, DAG, DL, true); |
27753 | DCI.AddToWorklist(VPermMask.getNode()); |
27754 | V1 = DAG.getBitcast(MaskVT, V1); |
27755 | DCI.AddToWorklist(V1.getNode()); |
27756 | V2 = DAG.getBitcast(MaskVT, V2); |
27757 | DCI.AddToWorklist(V2.getNode()); |
27758 | Res = DAG.getNode(X86ISD::VPERMV3, DL, MaskVT, V1, VPermMask, V2); |
27759 | DCI.AddToWorklist(Res.getNode()); |
27760 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res), |
27761 | /*AddTo*/ true); |
27762 | return true; |
27763 | } |
27764 | return false; |
27765 | } |
27766 | |
27767 | // See if we can combine a single input shuffle with zeros to a bit-mask, |
27768 | // which is much simpler than any shuffle. |
27769 | if (UnaryShuffle && MaskContainsZeros && (Depth >= 3 || HasVariableMask) && |
27770 | isSequentialOrUndefOrZeroInRange(Mask, 0, NumMaskElts, 0) && |
27771 | DAG.getTargetLoweringInfo().isTypeLegal(MaskVT)) { |
27772 | APInt Zero = APInt::getNullValue(MaskEltSizeInBits); |
27773 | APInt AllOnes = APInt::getAllOnesValue(MaskEltSizeInBits); |
27774 | APInt UndefElts(NumMaskElts, 0); |
27775 | SmallVector<APInt, 64> EltBits(NumMaskElts, Zero); |
27776 | for (unsigned i = 0; i != NumMaskElts; ++i) { |
27777 | int M = Mask[i]; |
27778 | if (M == SM_SentinelUndef) { |
27779 | UndefElts.setBit(i); |
27780 | continue; |
27781 | } |
27782 | if (M == SM_SentinelZero) |
27783 | continue; |
27784 | EltBits[i] = AllOnes; |
27785 | } |
27786 | SDValue BitMask = getConstVector(EltBits, UndefElts, MaskVT, DAG, DL); |
27787 | DCI.AddToWorklist(BitMask.getNode()); |
27788 | Res = DAG.getBitcast(MaskVT, V1); |
27789 | DCI.AddToWorklist(Res.getNode()); |
27790 | unsigned AndOpcode = |
27791 | FloatDomain ? unsigned(X86ISD::FAND) : unsigned(ISD::AND); |
27792 | Res = DAG.getNode(AndOpcode, DL, MaskVT, Res, BitMask); |
27793 | DCI.AddToWorklist(Res.getNode()); |
27794 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res), |
27795 | /*AddTo*/ true); |
27796 | return true; |
27797 | } |
27798 | |
27799 | // If we have a single input shuffle with different shuffle patterns in the |
27800 | // the 128-bit lanes use the variable mask to VPERMILPS. |
27801 | // TODO Combine other mask types at higher depths. |
27802 | if (UnaryShuffle && HasVariableMask && !MaskContainsZeros && |
27803 | ((MaskVT == MVT::v8f32 && Subtarget.hasAVX()) || |
27804 | (MaskVT == MVT::v16f32 && Subtarget.hasAVX512()))) { |
27805 | SmallVector<SDValue, 16> VPermIdx; |
27806 | for (int M : Mask) { |
27807 | SDValue Idx = |
27808 | M < 0 ? DAG.getUNDEF(MVT::i32) : DAG.getConstant(M % 4, DL, MVT::i32); |
27809 | VPermIdx.push_back(Idx); |
27810 | } |
27811 | MVT VPermMaskVT = MVT::getVectorVT(MVT::i32, NumMaskElts); |
27812 | SDValue VPermMask = DAG.getBuildVector(VPermMaskVT, DL, VPermIdx); |
27813 | DCI.AddToWorklist(VPermMask.getNode()); |
27814 | Res = DAG.getBitcast(MaskVT, V1); |
27815 | DCI.AddToWorklist(Res.getNode()); |
27816 | Res = DAG.getNode(X86ISD::VPERMILPV, DL, MaskVT, Res, VPermMask); |
27817 | DCI.AddToWorklist(Res.getNode()); |
27818 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res), |
27819 | /*AddTo*/ true); |
27820 | return true; |
27821 | } |
27822 | |
27823 | // With XOP, binary shuffles of 128/256-bit floating point vectors can combine |
27824 | // to VPERMIL2PD/VPERMIL2PS. |
27825 | if ((Depth >= 3 || HasVariableMask) && Subtarget.hasXOP() && |
27826 | (MaskVT == MVT::v2f64 || MaskVT == MVT::v4f64 || MaskVT == MVT::v4f32 || |
27827 | MaskVT == MVT::v8f32)) { |
27828 | // VPERMIL2 Operation. |
27829 | // Bits[3] - Match Bit. |
27830 | // Bits[2:1] - (Per Lane) PD Shuffle Mask. |
27831 | // Bits[2:0] - (Per Lane) PS Shuffle Mask. |
27832 | unsigned NumLanes = MaskVT.getSizeInBits() / 128; |
27833 | unsigned NumEltsPerLane = NumMaskElts / NumLanes; |
27834 | SmallVector<int, 8> VPerm2Idx; |
27835 | MVT MaskIdxSVT = MVT::getIntegerVT(MaskVT.getScalarSizeInBits()); |
27836 | MVT MaskIdxVT = MVT::getVectorVT(MaskIdxSVT, NumMaskElts); |
27837 | unsigned M2ZImm = 0; |
27838 | for (int M : Mask) { |
27839 | if (M == SM_SentinelUndef) { |
27840 | VPerm2Idx.push_back(-1); |
27841 | continue; |
27842 | } |
27843 | if (M == SM_SentinelZero) { |
27844 | M2ZImm = 2; |
27845 | VPerm2Idx.push_back(8); |
27846 | continue; |
27847 | } |
27848 | int Index = (M % NumEltsPerLane) + ((M / NumMaskElts) * NumEltsPerLane); |
27849 | Index = (MaskVT.getScalarSizeInBits() == 64 ? Index << 1 : Index); |
27850 | VPerm2Idx.push_back(Index); |
27851 | } |
27852 | V1 = DAG.getBitcast(MaskVT, V1); |
27853 | DCI.AddToWorklist(V1.getNode()); |
27854 | V2 = DAG.getBitcast(MaskVT, V2); |
27855 | DCI.AddToWorklist(V2.getNode()); |
27856 | SDValue VPerm2MaskOp = getConstVector(VPerm2Idx, MaskIdxVT, DAG, DL, true); |
27857 | DCI.AddToWorklist(VPerm2MaskOp.getNode()); |
27858 | Res = DAG.getNode(X86ISD::VPERMIL2, DL, MaskVT, V1, V2, VPerm2MaskOp, |
27859 | DAG.getConstant(M2ZImm, DL, MVT::i8)); |
27860 | DCI.AddToWorklist(Res.getNode()); |
27861 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res), |
27862 | /*AddTo*/ true); |
27863 | return true; |
27864 | } |
27865 | |
27866 | // If we have 3 or more shuffle instructions or a chain involving a variable |
27867 | // mask, we can replace them with a single PSHUFB instruction profitably. |
27868 | // Intel's manuals suggest only using PSHUFB if doing so replacing 5 |
27869 | // instructions, but in practice PSHUFB tends to be *very* fast so we're |
27870 | // more aggressive. |
27871 | if (UnaryShuffle && (Depth >= 3 || HasVariableMask) && |
27872 | ((RootVT.is128BitVector() && Subtarget.hasSSSE3()) || |
27873 | (RootVT.is256BitVector() && Subtarget.hasAVX2()) || |
27874 | (RootVT.is512BitVector() && Subtarget.hasBWI()))) { |
27875 | SmallVector<SDValue, 16> PSHUFBMask; |
27876 | int NumBytes = RootVT.getSizeInBits() / 8; |
27877 | int Ratio = NumBytes / NumMaskElts; |
27878 | for (int i = 0; i < NumBytes; ++i) { |
27879 | int M = Mask[i / Ratio]; |
27880 | if (M == SM_SentinelUndef) { |
27881 | PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8)); |
27882 | continue; |
27883 | } |
27884 | if (M == SM_SentinelZero) { |
27885 | PSHUFBMask.push_back(DAG.getConstant(255, DL, MVT::i8)); |
27886 | continue; |
27887 | } |
27888 | M = Ratio * M + i % Ratio; |
27889 | assert ((M / 16) == (i / 16) && "Lane crossing detected")(((M / 16) == (i / 16) && "Lane crossing detected") ? static_cast<void> (0) : __assert_fail ("(M / 16) == (i / 16) && \"Lane crossing detected\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27889, __PRETTY_FUNCTION__)); |
27890 | PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8)); |
27891 | } |
27892 | MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes); |
27893 | Res = DAG.getBitcast(ByteVT, V1); |
27894 | DCI.AddToWorklist(Res.getNode()); |
27895 | SDValue PSHUFBMaskOp = DAG.getBuildVector(ByteVT, DL, PSHUFBMask); |
27896 | DCI.AddToWorklist(PSHUFBMaskOp.getNode()); |
27897 | Res = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Res, PSHUFBMaskOp); |
27898 | DCI.AddToWorklist(Res.getNode()); |
27899 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res), |
27900 | /*AddTo*/ true); |
27901 | return true; |
27902 | } |
27903 | |
27904 | // With XOP, if we have a 128-bit binary input shuffle we can always combine |
27905 | // to VPPERM. We match the depth requirement of PSHUFB - VPPERM is never |
27906 | // slower than PSHUFB on targets that support both. |
27907 | if ((Depth >= 3 || HasVariableMask) && RootVT.is128BitVector() && |
27908 | Subtarget.hasXOP()) { |
27909 | // VPPERM Mask Operation |
27910 | // Bits[4:0] - Byte Index (0 - 31) |
27911 | // Bits[7:5] - Permute Operation (0 - Source byte, 4 - ZERO) |
27912 | SmallVector<SDValue, 16> VPPERMMask; |
27913 | int NumBytes = 16; |
27914 | int Ratio = NumBytes / NumMaskElts; |
27915 | for (int i = 0; i < NumBytes; ++i) { |
27916 | int M = Mask[i / Ratio]; |
27917 | if (M == SM_SentinelUndef) { |
27918 | VPPERMMask.push_back(DAG.getUNDEF(MVT::i8)); |
27919 | continue; |
27920 | } |
27921 | if (M == SM_SentinelZero) { |
27922 | VPPERMMask.push_back(DAG.getConstant(128, DL, MVT::i8)); |
27923 | continue; |
27924 | } |
27925 | M = Ratio * M + i % Ratio; |
27926 | VPPERMMask.push_back(DAG.getConstant(M, DL, MVT::i8)); |
27927 | } |
27928 | MVT ByteVT = MVT::v16i8; |
27929 | V1 = DAG.getBitcast(ByteVT, V1); |
27930 | DCI.AddToWorklist(V1.getNode()); |
27931 | V2 = DAG.getBitcast(ByteVT, V2); |
27932 | DCI.AddToWorklist(V2.getNode()); |
27933 | SDValue VPPERMMaskOp = DAG.getBuildVector(ByteVT, DL, VPPERMMask); |
27934 | DCI.AddToWorklist(VPPERMMaskOp.getNode()); |
27935 | Res = DAG.getNode(X86ISD::VPPERM, DL, ByteVT, V1, V2, VPPERMMaskOp); |
27936 | DCI.AddToWorklist(Res.getNode()); |
27937 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res), |
27938 | /*AddTo*/ true); |
27939 | return true; |
27940 | } |
27941 | |
27942 | // Failed to find any combines. |
27943 | return false; |
27944 | } |
27945 | |
27946 | // Attempt to constant fold all of the constant source ops. |
27947 | // Returns true if the entire shuffle is folded to a constant. |
27948 | // TODO: Extend this to merge multiple constant Ops and update the mask. |
27949 | static bool combineX86ShufflesConstants(const SmallVectorImpl<SDValue> &Ops, |
27950 | ArrayRef<int> Mask, SDValue Root, |
27951 | bool HasVariableMask, SelectionDAG &DAG, |
27952 | TargetLowering::DAGCombinerInfo &DCI, |
27953 | const X86Subtarget &Subtarget) { |
27954 | MVT VT = Root.getSimpleValueType(); |
27955 | |
27956 | unsigned SizeInBits = VT.getSizeInBits(); |
27957 | unsigned NumMaskElts = Mask.size(); |
27958 | unsigned MaskSizeInBits = SizeInBits / NumMaskElts; |
27959 | unsigned NumOps = Ops.size(); |
27960 | |
27961 | // Extract constant bits from each source op. |
27962 | bool OneUseConstantOp = false; |
27963 | SmallVector<APInt, 16> UndefEltsOps(NumOps); |
27964 | SmallVector<SmallVector<APInt, 16>, 16> RawBitsOps(NumOps); |
27965 | for (unsigned i = 0; i != NumOps; ++i) { |
27966 | SDValue SrcOp = Ops[i]; |
27967 | OneUseConstantOp |= SrcOp.hasOneUse(); |
27968 | if (!getTargetConstantBitsFromNode(SrcOp, MaskSizeInBits, UndefEltsOps[i], |
27969 | RawBitsOps[i])) |
27970 | return false; |
27971 | } |
27972 | |
27973 | // Only fold if at least one of the constants is only used once or |
27974 | // the combined shuffle has included a variable mask shuffle, this |
27975 | // is to avoid constant pool bloat. |
27976 | if (!OneUseConstantOp && !HasVariableMask) |
27977 | return false; |
27978 | |
27979 | // Shuffle the constant bits according to the mask. |
27980 | APInt UndefElts(NumMaskElts, 0); |
27981 | APInt ZeroElts(NumMaskElts, 0); |
27982 | APInt ConstantElts(NumMaskElts, 0); |
27983 | SmallVector<APInt, 8> ConstantBitData(NumMaskElts, |
27984 | APInt::getNullValue(MaskSizeInBits)); |
27985 | for (unsigned i = 0; i != NumMaskElts; ++i) { |
27986 | int M = Mask[i]; |
27987 | if (M == SM_SentinelUndef) { |
27988 | UndefElts.setBit(i); |
27989 | continue; |
27990 | } else if (M == SM_SentinelZero) { |
27991 | ZeroElts.setBit(i); |
27992 | continue; |
27993 | } |
27994 | assert(0 <= M && M < (int)(NumMaskElts * NumOps))((0 <= M && M < (int)(NumMaskElts * NumOps)) ? static_cast <void> (0) : __assert_fail ("0 <= M && M < (int)(NumMaskElts * NumOps)" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 27994, __PRETTY_FUNCTION__)); |
27995 | |
27996 | unsigned SrcOpIdx = (unsigned)M / NumMaskElts; |
27997 | unsigned SrcMaskIdx = (unsigned)M % NumMaskElts; |
27998 | |
27999 | auto &SrcUndefElts = UndefEltsOps[SrcOpIdx]; |
28000 | if (SrcUndefElts[SrcMaskIdx]) { |
28001 | UndefElts.setBit(i); |
28002 | continue; |
28003 | } |
28004 | |
28005 | auto &SrcEltBits = RawBitsOps[SrcOpIdx]; |
28006 | APInt &Bits = SrcEltBits[SrcMaskIdx]; |
28007 | if (!Bits) { |
28008 | ZeroElts.setBit(i); |
28009 | continue; |
28010 | } |
28011 | |
28012 | ConstantElts.setBit(i); |
28013 | ConstantBitData[i] = Bits; |
28014 | } |
28015 | assert((UndefElts | ZeroElts | ConstantElts).isAllOnesValue())(((UndefElts | ZeroElts | ConstantElts).isAllOnesValue()) ? static_cast <void> (0) : __assert_fail ("(UndefElts | ZeroElts | ConstantElts).isAllOnesValue()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28015, __PRETTY_FUNCTION__)); |
28016 | |
28017 | // Create the constant data. |
28018 | MVT MaskSVT; |
28019 | if (VT.isFloatingPoint() && (MaskSizeInBits == 32 || MaskSizeInBits == 64)) |
28020 | MaskSVT = MVT::getFloatingPointVT(MaskSizeInBits); |
28021 | else |
28022 | MaskSVT = MVT::getIntegerVT(MaskSizeInBits); |
28023 | |
28024 | MVT MaskVT = MVT::getVectorVT(MaskSVT, NumMaskElts); |
28025 | |
28026 | SDLoc DL(Root); |
28027 | SDValue CstOp = getConstVector(ConstantBitData, UndefElts, MaskVT, DAG, DL); |
28028 | DCI.AddToWorklist(CstOp.getNode()); |
28029 | DCI.CombineTo(Root.getNode(), DAG.getBitcast(VT, CstOp)); |
28030 | return true; |
28031 | } |
28032 | |
28033 | /// \brief Fully generic combining of x86 shuffle instructions. |
28034 | /// |
28035 | /// This should be the last combine run over the x86 shuffle instructions. Once |
28036 | /// they have been fully optimized, this will recursively consider all chains |
28037 | /// of single-use shuffle instructions, build a generic model of the cumulative |
28038 | /// shuffle operation, and check for simpler instructions which implement this |
28039 | /// operation. We use this primarily for two purposes: |
28040 | /// |
28041 | /// 1) Collapse generic shuffles to specialized single instructions when |
28042 | /// equivalent. In most cases, this is just an encoding size win, but |
28043 | /// sometimes we will collapse multiple generic shuffles into a single |
28044 | /// special-purpose shuffle. |
28045 | /// 2) Look for sequences of shuffle instructions with 3 or more total |
28046 | /// instructions, and replace them with the slightly more expensive SSSE3 |
28047 | /// PSHUFB instruction if available. We do this as the last combining step |
28048 | /// to ensure we avoid using PSHUFB if we can implement the shuffle with |
28049 | /// a suitable short sequence of other instructions. The PSHUFB will either |
28050 | /// use a register or have to read from memory and so is slightly (but only |
28051 | /// slightly) more expensive than the other shuffle instructions. |
28052 | /// |
28053 | /// Because this is inherently a quadratic operation (for each shuffle in |
28054 | /// a chain, we recurse up the chain), the depth is limited to 8 instructions. |
28055 | /// This should never be an issue in practice as the shuffle lowering doesn't |
28056 | /// produce sequences of more than 8 instructions. |
28057 | /// |
28058 | /// FIXME: We will currently miss some cases where the redundant shuffling |
28059 | /// would simplify under the threshold for PSHUFB formation because of |
28060 | /// combine-ordering. To fix this, we should do the redundant instruction |
28061 | /// combining in this recursive walk. |
28062 | static bool combineX86ShufflesRecursively(ArrayRef<SDValue> SrcOps, |
28063 | int SrcOpIndex, SDValue Root, |
28064 | ArrayRef<int> RootMask, |
28065 | ArrayRef<const SDNode*> SrcNodes, |
28066 | int Depth, bool HasVariableMask, |
28067 | SelectionDAG &DAG, |
28068 | TargetLowering::DAGCombinerInfo &DCI, |
28069 | const X86Subtarget &Subtarget) { |
28070 | // Bound the depth of our recursive combine because this is ultimately |
28071 | // quadratic in nature. |
28072 | if (Depth > 8) |
28073 | return false; |
28074 | |
28075 | // Directly rip through bitcasts to find the underlying operand. |
28076 | SDValue Op = SrcOps[SrcOpIndex]; |
28077 | Op = peekThroughOneUseBitcasts(Op); |
28078 | |
28079 | MVT VT = Op.getSimpleValueType(); |
28080 | if (!VT.isVector()) |
28081 | return false; // Bail if we hit a non-vector. |
28082 | |
28083 | assert(Root.getSimpleValueType().isVector() &&((Root.getSimpleValueType().isVector() && "Shuffles operate on vector types!" ) ? static_cast<void> (0) : __assert_fail ("Root.getSimpleValueType().isVector() && \"Shuffles operate on vector types!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28084, __PRETTY_FUNCTION__)) |
28084 | "Shuffles operate on vector types!")((Root.getSimpleValueType().isVector() && "Shuffles operate on vector types!" ) ? static_cast<void> (0) : __assert_fail ("Root.getSimpleValueType().isVector() && \"Shuffles operate on vector types!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28084, __PRETTY_FUNCTION__)); |
28085 | assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&((VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits () && "Can only combine shuffles of the same vector register size." ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() && \"Can only combine shuffles of the same vector register size.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28086, __PRETTY_FUNCTION__)) |
28086 | "Can only combine shuffles of the same vector register size.")((VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits () && "Can only combine shuffles of the same vector register size." ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() && \"Can only combine shuffles of the same vector register size.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28086, __PRETTY_FUNCTION__)); |
28087 | |
28088 | // Extract target shuffle mask and resolve sentinels and inputs. |
28089 | SmallVector<int, 64> OpMask; |
28090 | SmallVector<SDValue, 2> OpInputs; |
28091 | if (!resolveTargetShuffleInputs(Op, OpInputs, OpMask, DAG)) |
28092 | return false; |
28093 | |
28094 | assert(OpInputs.size() <= 2 && "Too many shuffle inputs")((OpInputs.size() <= 2 && "Too many shuffle inputs" ) ? static_cast<void> (0) : __assert_fail ("OpInputs.size() <= 2 && \"Too many shuffle inputs\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28094, __PRETTY_FUNCTION__)); |
28095 | SDValue Input0 = (OpInputs.size() > 0 ? OpInputs[0] : SDValue()); |
28096 | SDValue Input1 = (OpInputs.size() > 1 ? OpInputs[1] : SDValue()); |
28097 | |
28098 | // Add the inputs to the Ops list, avoiding duplicates. |
28099 | SmallVector<SDValue, 16> Ops(SrcOps.begin(), SrcOps.end()); |
28100 | |
28101 | int InputIdx0 = -1, InputIdx1 = -1; |
28102 | for (int i = 0, e = Ops.size(); i < e; ++i) { |
28103 | SDValue BC = peekThroughBitcasts(Ops[i]); |
28104 | if (Input0 && BC == peekThroughBitcasts(Input0)) |
28105 | InputIdx0 = i; |
28106 | if (Input1 && BC == peekThroughBitcasts(Input1)) |
28107 | InputIdx1 = i; |
28108 | } |
28109 | |
28110 | if (Input0 && InputIdx0 < 0) { |
28111 | InputIdx0 = SrcOpIndex; |
28112 | Ops[SrcOpIndex] = Input0; |
28113 | } |
28114 | if (Input1 && InputIdx1 < 0) { |
28115 | InputIdx1 = Ops.size(); |
28116 | Ops.push_back(Input1); |
28117 | } |
28118 | |
28119 | assert(((RootMask.size() > OpMask.size() &&((((RootMask.size() > OpMask.size() && RootMask.size () % OpMask.size() == 0) || (OpMask.size() > RootMask.size () && OpMask.size() % RootMask.size() == 0) || OpMask .size() == RootMask.size()) && "The smaller number of elements must divide the larger." ) ? static_cast<void> (0) : __assert_fail ("((RootMask.size() > OpMask.size() && RootMask.size() % OpMask.size() == 0) || (OpMask.size() > RootMask.size() && OpMask.size() % RootMask.size() == 0) || OpMask.size() == RootMask.size()) && \"The smaller number of elements must divide the larger.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28124, __PRETTY_FUNCTION__)) |
28120 | RootMask.size() % OpMask.size() == 0) ||((((RootMask.size() > OpMask.size() && RootMask.size () % OpMask.size() == 0) || (OpMask.size() > RootMask.size () && OpMask.size() % RootMask.size() == 0) || OpMask .size() == RootMask.size()) && "The smaller number of elements must divide the larger." ) ? static_cast<void> (0) : __assert_fail ("((RootMask.size() > OpMask.size() && RootMask.size() % OpMask.size() == 0) || (OpMask.size() > RootMask.size() && OpMask.size() % RootMask.size() == 0) || OpMask.size() == RootMask.size()) && \"The smaller number of elements must divide the larger.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28124, __PRETTY_FUNCTION__)) |
28121 | (OpMask.size() > RootMask.size() &&((((RootMask.size() > OpMask.size() && RootMask.size () % OpMask.size() == 0) || (OpMask.size() > RootMask.size () && OpMask.size() % RootMask.size() == 0) || OpMask .size() == RootMask.size()) && "The smaller number of elements must divide the larger." ) ? static_cast<void> (0) : __assert_fail ("((RootMask.size() > OpMask.size() && RootMask.size() % OpMask.size() == 0) || (OpMask.size() > RootMask.size() && OpMask.size() % RootMask.size() == 0) || OpMask.size() == RootMask.size()) && \"The smaller number of elements must divide the larger.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28124, __PRETTY_FUNCTION__)) |
28122 | OpMask.size() % RootMask.size() == 0) ||((((RootMask.size() > OpMask.size() && RootMask.size () % OpMask.size() == 0) || (OpMask.size() > RootMask.size () && OpMask.size() % RootMask.size() == 0) || OpMask .size() == RootMask.size()) && "The smaller number of elements must divide the larger." ) ? static_cast<void> (0) : __assert_fail ("((RootMask.size() > OpMask.size() && RootMask.size() % OpMask.size() == 0) || (OpMask.size() > RootMask.size() && OpMask.size() % RootMask.size() == 0) || OpMask.size() == RootMask.size()) && \"The smaller number of elements must divide the larger.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28124, __PRETTY_FUNCTION__)) |
28123 | OpMask.size() == RootMask.size()) &&((((RootMask.size() > OpMask.size() && RootMask.size () % OpMask.size() == 0) || (OpMask.size() > RootMask.size () && OpMask.size() % RootMask.size() == 0) || OpMask .size() == RootMask.size()) && "The smaller number of elements must divide the larger." ) ? static_cast<void> (0) : __assert_fail ("((RootMask.size() > OpMask.size() && RootMask.size() % OpMask.size() == 0) || (OpMask.size() > RootMask.size() && OpMask.size() % RootMask.size() == 0) || OpMask.size() == RootMask.size()) && \"The smaller number of elements must divide the larger.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28124, __PRETTY_FUNCTION__)) |
28124 | "The smaller number of elements must divide the larger.")((((RootMask.size() > OpMask.size() && RootMask.size () % OpMask.size() == 0) || (OpMask.size() > RootMask.size () && OpMask.size() % RootMask.size() == 0) || OpMask .size() == RootMask.size()) && "The smaller number of elements must divide the larger." ) ? static_cast<void> (0) : __assert_fail ("((RootMask.size() > OpMask.size() && RootMask.size() % OpMask.size() == 0) || (OpMask.size() > RootMask.size() && OpMask.size() % RootMask.size() == 0) || OpMask.size() == RootMask.size()) && \"The smaller number of elements must divide the larger.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28124, __PRETTY_FUNCTION__)); |
28125 | |
28126 | // This function can be performance-critical, so we rely on the power-of-2 |
28127 | // knowledge that we have about the mask sizes to replace div/rem ops with |
28128 | // bit-masks and shifts. |
28129 | assert(isPowerOf2_32(RootMask.size()) && "Non-power-of-2 shuffle mask sizes")((isPowerOf2_32(RootMask.size()) && "Non-power-of-2 shuffle mask sizes" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(RootMask.size()) && \"Non-power-of-2 shuffle mask sizes\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28129, __PRETTY_FUNCTION__)); |
28130 | assert(isPowerOf2_32(OpMask.size()) && "Non-power-of-2 shuffle mask sizes")((isPowerOf2_32(OpMask.size()) && "Non-power-of-2 shuffle mask sizes" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(OpMask.size()) && \"Non-power-of-2 shuffle mask sizes\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28130, __PRETTY_FUNCTION__)); |
28131 | unsigned RootMaskSizeLog2 = countTrailingZeros(RootMask.size()); |
28132 | unsigned OpMaskSizeLog2 = countTrailingZeros(OpMask.size()); |
28133 | |
28134 | unsigned MaskWidth = std::max<unsigned>(OpMask.size(), RootMask.size()); |
28135 | unsigned RootRatio = std::max<unsigned>(1, OpMask.size() >> RootMaskSizeLog2); |
28136 | unsigned OpRatio = std::max<unsigned>(1, RootMask.size() >> OpMaskSizeLog2); |
28137 | assert((RootRatio == 1 || OpRatio == 1) &&(((RootRatio == 1 || OpRatio == 1) && "Must not have a ratio for both incoming and op masks!" ) ? static_cast<void> (0) : __assert_fail ("(RootRatio == 1 || OpRatio == 1) && \"Must not have a ratio for both incoming and op masks!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28138, __PRETTY_FUNCTION__)) |
28138 | "Must not have a ratio for both incoming and op masks!")(((RootRatio == 1 || OpRatio == 1) && "Must not have a ratio for both incoming and op masks!" ) ? static_cast<void> (0) : __assert_fail ("(RootRatio == 1 || OpRatio == 1) && \"Must not have a ratio for both incoming and op masks!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28138, __PRETTY_FUNCTION__)); |
28139 | |
28140 | assert(isPowerOf2_32(MaskWidth) && "Non-power-of-2 shuffle mask sizes")((isPowerOf2_32(MaskWidth) && "Non-power-of-2 shuffle mask sizes" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(MaskWidth) && \"Non-power-of-2 shuffle mask sizes\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28140, __PRETTY_FUNCTION__)); |
28141 | assert(isPowerOf2_32(RootRatio) && "Non-power-of-2 shuffle mask sizes")((isPowerOf2_32(RootRatio) && "Non-power-of-2 shuffle mask sizes" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(RootRatio) && \"Non-power-of-2 shuffle mask sizes\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28141, __PRETTY_FUNCTION__)); |
28142 | assert(isPowerOf2_32(OpRatio) && "Non-power-of-2 shuffle mask sizes")((isPowerOf2_32(OpRatio) && "Non-power-of-2 shuffle mask sizes" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(OpRatio) && \"Non-power-of-2 shuffle mask sizes\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28142, __PRETTY_FUNCTION__)); |
28143 | unsigned RootRatioLog2 = countTrailingZeros(RootRatio); |
28144 | unsigned OpRatioLog2 = countTrailingZeros(OpRatio); |
28145 | |
28146 | SmallVector<int, 64> Mask(MaskWidth, SM_SentinelUndef); |
28147 | |
28148 | // Merge this shuffle operation's mask into our accumulated mask. Note that |
28149 | // this shuffle's mask will be the first applied to the input, followed by the |
28150 | // root mask to get us all the way to the root value arrangement. The reason |
28151 | // for this order is that we are recursing up the operation chain. |
28152 | for (unsigned i = 0; i < MaskWidth; ++i) { |
28153 | unsigned RootIdx = i >> RootRatioLog2; |
28154 | if (RootMask[RootIdx] < 0) { |
28155 | // This is a zero or undef lane, we're done. |
28156 | Mask[i] = RootMask[RootIdx]; |
28157 | continue; |
28158 | } |
28159 | |
28160 | unsigned RootMaskedIdx = |
28161 | RootRatio == 1 |
28162 | ? RootMask[RootIdx] |
28163 | : (RootMask[RootIdx] << RootRatioLog2) + (i & (RootRatio - 1)); |
28164 | |
28165 | // Just insert the scaled root mask value if it references an input other |
28166 | // than the SrcOp we're currently inserting. |
28167 | if ((RootMaskedIdx < (SrcOpIndex * MaskWidth)) || |
28168 | (((SrcOpIndex + 1) * MaskWidth) <= RootMaskedIdx)) { |
28169 | Mask[i] = RootMaskedIdx; |
28170 | continue; |
28171 | } |
28172 | |
28173 | RootMaskedIdx = RootMaskedIdx & (MaskWidth - 1); |
28174 | unsigned OpIdx = RootMaskedIdx >> OpRatioLog2; |
28175 | if (OpMask[OpIdx] < 0) { |
28176 | // The incoming lanes are zero or undef, it doesn't matter which ones we |
28177 | // are using. |
28178 | Mask[i] = OpMask[OpIdx]; |
28179 | continue; |
28180 | } |
28181 | |
28182 | // Ok, we have non-zero lanes, map them through to one of the Op's inputs. |
28183 | unsigned OpMaskedIdx = |
28184 | OpRatio == 1 |
28185 | ? OpMask[OpIdx] |
28186 | : (OpMask[OpIdx] << OpRatioLog2) + (RootMaskedIdx & (OpRatio - 1)); |
28187 | |
28188 | OpMaskedIdx = OpMaskedIdx & (MaskWidth - 1); |
28189 | if (OpMask[OpIdx] < (int)OpMask.size()) { |
28190 | assert(0 <= InputIdx0 && "Unknown target shuffle input")((0 <= InputIdx0 && "Unknown target shuffle input" ) ? static_cast<void> (0) : __assert_fail ("0 <= InputIdx0 && \"Unknown target shuffle input\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28190, __PRETTY_FUNCTION__)); |
28191 | OpMaskedIdx += InputIdx0 * MaskWidth; |
28192 | } else { |
28193 | assert(0 <= InputIdx1 && "Unknown target shuffle input")((0 <= InputIdx1 && "Unknown target shuffle input" ) ? static_cast<void> (0) : __assert_fail ("0 <= InputIdx1 && \"Unknown target shuffle input\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28193, __PRETTY_FUNCTION__)); |
28194 | OpMaskedIdx += InputIdx1 * MaskWidth; |
28195 | } |
28196 | |
28197 | Mask[i] = OpMaskedIdx; |
28198 | } |
28199 | |
28200 | // Handle the all undef/zero cases early. |
28201 | if (all_of(Mask, [](int Idx) { return Idx == SM_SentinelUndef; })) { |
28202 | DCI.CombineTo(Root.getNode(), DAG.getUNDEF(Root.getValueType())); |
28203 | return true; |
28204 | } |
28205 | if (all_of(Mask, [](int Idx) { return Idx < 0; })) { |
28206 | // TODO - should we handle the mixed zero/undef case as well? Just returning |
28207 | // a zero mask will lose information on undef elements possibly reducing |
28208 | // future combine possibilities. |
28209 | DCI.CombineTo(Root.getNode(), getZeroVector(Root.getSimpleValueType(), |
28210 | Subtarget, DAG, SDLoc(Root))); |
28211 | return true; |
28212 | } |
28213 | |
28214 | // Remove unused shuffle source ops. |
28215 | resolveTargetShuffleInputsAndMask(Ops, Mask); |
28216 | assert(!Ops.empty() && "Shuffle with no inputs detected")((!Ops.empty() && "Shuffle with no inputs detected") ? static_cast<void> (0) : __assert_fail ("!Ops.empty() && \"Shuffle with no inputs detected\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28216, __PRETTY_FUNCTION__)); |
28217 | |
28218 | HasVariableMask |= isTargetShuffleVariableMask(Op.getOpcode()); |
28219 | |
28220 | // Update the list of shuffle nodes that have been combined so far. |
28221 | SmallVector<const SDNode *, 16> CombinedNodes(SrcNodes.begin(), |
28222 | SrcNodes.end()); |
28223 | CombinedNodes.push_back(Op.getNode()); |
28224 | |
28225 | // See if we can recurse into each shuffle source op (if it's a target |
28226 | // shuffle). The source op should only be combined if it either has a |
28227 | // single use (i.e. current Op) or all its users have already been combined. |
28228 | for (int i = 0, e = Ops.size(); i < e; ++i) |
28229 | if (Ops[i].getNode()->hasOneUse() || |
28230 | SDNode::areOnlyUsersOf(CombinedNodes, Ops[i].getNode())) |
28231 | if (combineX86ShufflesRecursively(Ops, i, Root, Mask, CombinedNodes, |
28232 | Depth + 1, HasVariableMask, DAG, DCI, |
28233 | Subtarget)) |
28234 | return true; |
28235 | |
28236 | // Attempt to constant fold all of the constant source ops. |
28237 | if (combineX86ShufflesConstants(Ops, Mask, Root, HasVariableMask, DAG, DCI, |
28238 | Subtarget)) |
28239 | return true; |
28240 | |
28241 | // We can only combine unary and binary shuffle mask cases. |
28242 | if (Ops.size() > 2) |
28243 | return false; |
28244 | |
28245 | // Minor canonicalization of the accumulated shuffle mask to make it easier |
28246 | // to match below. All this does is detect masks with sequential pairs of |
28247 | // elements, and shrink them to the half-width mask. It does this in a loop |
28248 | // so it will reduce the size of the mask to the minimal width mask which |
28249 | // performs an equivalent shuffle. |
28250 | SmallVector<int, 64> WidenedMask; |
28251 | while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) { |
28252 | Mask = std::move(WidenedMask); |
28253 | } |
28254 | |
28255 | // Canonicalization of binary shuffle masks to improve pattern matching by |
28256 | // commuting the inputs. |
28257 | if (Ops.size() == 2 && canonicalizeShuffleMaskWithCommute(Mask)) { |
28258 | ShuffleVectorSDNode::commuteMask(Mask); |
28259 | std::swap(Ops[0], Ops[1]); |
28260 | } |
28261 | |
28262 | return combineX86ShuffleChain(Ops, Root, Mask, Depth, HasVariableMask, DAG, |
28263 | DCI, Subtarget); |
28264 | } |
28265 | |
28266 | /// \brief Get the PSHUF-style mask from PSHUF node. |
28267 | /// |
28268 | /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4 |
28269 | /// PSHUF-style masks that can be reused with such instructions. |
28270 | static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) { |
28271 | MVT VT = N.getSimpleValueType(); |
28272 | SmallVector<int, 4> Mask; |
28273 | SmallVector<SDValue, 2> Ops; |
28274 | bool IsUnary; |
28275 | bool HaveMask = |
28276 | getTargetShuffleMask(N.getNode(), VT, false, Ops, Mask, IsUnary); |
28277 | (void)HaveMask; |
28278 | assert(HaveMask)((HaveMask) ? static_cast<void> (0) : __assert_fail ("HaveMask" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28278, __PRETTY_FUNCTION__)); |
28279 | |
28280 | // If we have more than 128-bits, only the low 128-bits of shuffle mask |
28281 | // matter. Check that the upper masks are repeats and remove them. |
28282 | if (VT.getSizeInBits() > 128) { |
28283 | int LaneElts = 128 / VT.getScalarSizeInBits(); |
28284 | #ifndef NDEBUG |
28285 | for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i) |
28286 | for (int j = 0; j < LaneElts; ++j) |
28287 | assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&((Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) && "Mask doesn't repeat in high 128-bit lanes!") ? static_cast< void> (0) : __assert_fail ("Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) && \"Mask doesn't repeat in high 128-bit lanes!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28288, __PRETTY_FUNCTION__)) |
28288 | "Mask doesn't repeat in high 128-bit lanes!")((Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) && "Mask doesn't repeat in high 128-bit lanes!") ? static_cast< void> (0) : __assert_fail ("Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) && \"Mask doesn't repeat in high 128-bit lanes!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28288, __PRETTY_FUNCTION__)); |
28289 | #endif |
28290 | Mask.resize(LaneElts); |
28291 | } |
28292 | |
28293 | switch (N.getOpcode()) { |
28294 | case X86ISD::PSHUFD: |
28295 | return Mask; |
28296 | case X86ISD::PSHUFLW: |
28297 | Mask.resize(4); |
28298 | return Mask; |
28299 | case X86ISD::PSHUFHW: |
28300 | Mask.erase(Mask.begin(), Mask.begin() + 4); |
28301 | for (int &M : Mask) |
28302 | M -= 4; |
28303 | return Mask; |
28304 | default: |
28305 | llvm_unreachable("No valid shuffle instruction found!")::llvm::llvm_unreachable_internal("No valid shuffle instruction found!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28305); |
28306 | } |
28307 | } |
28308 | |
28309 | /// \brief Search for a combinable shuffle across a chain ending in pshufd. |
28310 | /// |
28311 | /// We walk up the chain and look for a combinable shuffle, skipping over |
28312 | /// shuffles that we could hoist this shuffle's transformation past without |
28313 | /// altering anything. |
28314 | static SDValue |
28315 | combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask, |
28316 | SelectionDAG &DAG) { |
28317 | assert(N.getOpcode() == X86ISD::PSHUFD &&((N.getOpcode() == X86ISD::PSHUFD && "Called with something other than an x86 128-bit half shuffle!" ) ? static_cast<void> (0) : __assert_fail ("N.getOpcode() == X86ISD::PSHUFD && \"Called with something other than an x86 128-bit half shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28318, __PRETTY_FUNCTION__)) |
28318 | "Called with something other than an x86 128-bit half shuffle!")((N.getOpcode() == X86ISD::PSHUFD && "Called with something other than an x86 128-bit half shuffle!" ) ? static_cast<void> (0) : __assert_fail ("N.getOpcode() == X86ISD::PSHUFD && \"Called with something other than an x86 128-bit half shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28318, __PRETTY_FUNCTION__)); |
28319 | SDLoc DL(N); |
28320 | |
28321 | // Walk up a single-use chain looking for a combinable shuffle. Keep a stack |
28322 | // of the shuffles in the chain so that we can form a fresh chain to replace |
28323 | // this one. |
28324 | SmallVector<SDValue, 8> Chain; |
28325 | SDValue V = N.getOperand(0); |
28326 | for (; V.hasOneUse(); V = V.getOperand(0)) { |
28327 | switch (V.getOpcode()) { |
28328 | default: |
28329 | return SDValue(); // Nothing combined! |
28330 | |
28331 | case ISD::BITCAST: |
28332 | // Skip bitcasts as we always know the type for the target specific |
28333 | // instructions. |
28334 | continue; |
28335 | |
28336 | case X86ISD::PSHUFD: |
28337 | // Found another dword shuffle. |
28338 | break; |
28339 | |
28340 | case X86ISD::PSHUFLW: |
28341 | // Check that the low words (being shuffled) are the identity in the |
28342 | // dword shuffle, and the high words are self-contained. |
28343 | if (Mask[0] != 0 || Mask[1] != 1 || |
28344 | !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4)) |
28345 | return SDValue(); |
28346 | |
28347 | Chain.push_back(V); |
28348 | continue; |
28349 | |
28350 | case X86ISD::PSHUFHW: |
28351 | // Check that the high words (being shuffled) are the identity in the |
28352 | // dword shuffle, and the low words are self-contained. |
28353 | if (Mask[2] != 2 || Mask[3] != 3 || |
28354 | !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2)) |
28355 | return SDValue(); |
28356 | |
28357 | Chain.push_back(V); |
28358 | continue; |
28359 | |
28360 | case X86ISD::UNPCKL: |
28361 | case X86ISD::UNPCKH: |
28362 | // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword |
28363 | // shuffle into a preceding word shuffle. |
28364 | if (V.getSimpleValueType().getVectorElementType() != MVT::i8 && |
28365 | V.getSimpleValueType().getVectorElementType() != MVT::i16) |
28366 | return SDValue(); |
28367 | |
28368 | // Search for a half-shuffle which we can combine with. |
28369 | unsigned CombineOp = |
28370 | V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW; |
28371 | if (V.getOperand(0) != V.getOperand(1) || |
28372 | !V->isOnlyUserOf(V.getOperand(0).getNode())) |
28373 | return SDValue(); |
28374 | Chain.push_back(V); |
28375 | V = V.getOperand(0); |
28376 | do { |
28377 | switch (V.getOpcode()) { |
28378 | default: |
28379 | return SDValue(); // Nothing to combine. |
28380 | |
28381 | case X86ISD::PSHUFLW: |
28382 | case X86ISD::PSHUFHW: |
28383 | if (V.getOpcode() == CombineOp) |
28384 | break; |
28385 | |
28386 | Chain.push_back(V); |
28387 | |
28388 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
28389 | case ISD::BITCAST: |
28390 | V = V.getOperand(0); |
28391 | continue; |
28392 | } |
28393 | break; |
28394 | } while (V.hasOneUse()); |
28395 | break; |
28396 | } |
28397 | // Break out of the loop if we break out of the switch. |
28398 | break; |
28399 | } |
28400 | |
28401 | if (!V.hasOneUse()) |
28402 | // We fell out of the loop without finding a viable combining instruction. |
28403 | return SDValue(); |
28404 | |
28405 | // Merge this node's mask and our incoming mask. |
28406 | SmallVector<int, 4> VMask = getPSHUFShuffleMask(V); |
28407 | for (int &M : Mask) |
28408 | M = VMask[M]; |
28409 | V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0), |
28410 | getV4X86ShuffleImm8ForMask(Mask, DL, DAG)); |
28411 | |
28412 | // Rebuild the chain around this new shuffle. |
28413 | while (!Chain.empty()) { |
28414 | SDValue W = Chain.pop_back_val(); |
28415 | |
28416 | if (V.getValueType() != W.getOperand(0).getValueType()) |
28417 | V = DAG.getBitcast(W.getOperand(0).getValueType(), V); |
28418 | |
28419 | switch (W.getOpcode()) { |
28420 | default: |
28421 | llvm_unreachable("Only PSHUF and UNPCK instructions get here!")::llvm::llvm_unreachable_internal("Only PSHUF and UNPCK instructions get here!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28421); |
28422 | |
28423 | case X86ISD::UNPCKL: |
28424 | case X86ISD::UNPCKH: |
28425 | V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V); |
28426 | break; |
28427 | |
28428 | case X86ISD::PSHUFD: |
28429 | case X86ISD::PSHUFLW: |
28430 | case X86ISD::PSHUFHW: |
28431 | V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1)); |
28432 | break; |
28433 | } |
28434 | } |
28435 | if (V.getValueType() != N.getValueType()) |
28436 | V = DAG.getBitcast(N.getValueType(), V); |
28437 | |
28438 | // Return the new chain to replace N. |
28439 | return V; |
28440 | } |
28441 | |
28442 | /// \brief Search for a combinable shuffle across a chain ending in pshuflw or |
28443 | /// pshufhw. |
28444 | /// |
28445 | /// We walk up the chain, skipping shuffles of the other half and looking |
28446 | /// through shuffles which switch halves trying to find a shuffle of the same |
28447 | /// pair of dwords. |
28448 | static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask, |
28449 | SelectionDAG &DAG, |
28450 | TargetLowering::DAGCombinerInfo &DCI) { |
28451 | assert((((N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD ::PSHUFHW) && "Called with something other than an x86 128-bit half shuffle!" ) ? static_cast<void> (0) : __assert_fail ("(N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) && \"Called with something other than an x86 128-bit half shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28453, __PRETTY_FUNCTION__)) |
28452 | (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&(((N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD ::PSHUFHW) && "Called with something other than an x86 128-bit half shuffle!" ) ? static_cast<void> (0) : __assert_fail ("(N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) && \"Called with something other than an x86 128-bit half shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28453, __PRETTY_FUNCTION__)) |
28453 | "Called with something other than an x86 128-bit half shuffle!")(((N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD ::PSHUFHW) && "Called with something other than an x86 128-bit half shuffle!" ) ? static_cast<void> (0) : __assert_fail ("(N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) && \"Called with something other than an x86 128-bit half shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28453, __PRETTY_FUNCTION__)); |
28454 | SDLoc DL(N); |
28455 | unsigned CombineOpcode = N.getOpcode(); |
28456 | |
28457 | // Walk up a single-use chain looking for a combinable shuffle. |
28458 | SDValue V = N.getOperand(0); |
28459 | for (; V.hasOneUse(); V = V.getOperand(0)) { |
28460 | switch (V.getOpcode()) { |
28461 | default: |
28462 | return false; // Nothing combined! |
28463 | |
28464 | case ISD::BITCAST: |
28465 | // Skip bitcasts as we always know the type for the target specific |
28466 | // instructions. |
28467 | continue; |
28468 | |
28469 | case X86ISD::PSHUFLW: |
28470 | case X86ISD::PSHUFHW: |
28471 | if (V.getOpcode() == CombineOpcode) |
28472 | break; |
28473 | |
28474 | // Other-half shuffles are no-ops. |
28475 | continue; |
28476 | } |
28477 | // Break out of the loop if we break out of the switch. |
28478 | break; |
28479 | } |
28480 | |
28481 | if (!V.hasOneUse()) |
28482 | // We fell out of the loop without finding a viable combining instruction. |
28483 | return false; |
28484 | |
28485 | // Combine away the bottom node as its shuffle will be accumulated into |
28486 | // a preceding shuffle. |
28487 | DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true); |
28488 | |
28489 | // Record the old value. |
28490 | SDValue Old = V; |
28491 | |
28492 | // Merge this node's mask and our incoming mask (adjusted to account for all |
28493 | // the pshufd instructions encountered). |
28494 | SmallVector<int, 4> VMask = getPSHUFShuffleMask(V); |
28495 | for (int &M : Mask) |
28496 | M = VMask[M]; |
28497 | V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0), |
28498 | getV4X86ShuffleImm8ForMask(Mask, DL, DAG)); |
28499 | |
28500 | // Check that the shuffles didn't cancel each other out. If not, we need to |
28501 | // combine to the new one. |
28502 | if (Old != V) |
28503 | // Replace the combinable shuffle with the combined one, updating all users |
28504 | // so that we re-evaluate the chain here. |
28505 | DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true); |
28506 | |
28507 | return true; |
28508 | } |
28509 | |
28510 | /// \brief Try to combine x86 target specific shuffles. |
28511 | static SDValue combineTargetShuffle(SDValue N, SelectionDAG &DAG, |
28512 | TargetLowering::DAGCombinerInfo &DCI, |
28513 | const X86Subtarget &Subtarget) { |
28514 | SDLoc DL(N); |
28515 | MVT VT = N.getSimpleValueType(); |
28516 | SmallVector<int, 4> Mask; |
28517 | |
28518 | unsigned Opcode = N.getOpcode(); |
28519 | switch (Opcode) { |
28520 | case X86ISD::PSHUFD: |
28521 | case X86ISD::PSHUFLW: |
28522 | case X86ISD::PSHUFHW: |
28523 | Mask = getPSHUFShuffleMask(N); |
28524 | assert(Mask.size() == 4)((Mask.size() == 4) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 4", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28524, __PRETTY_FUNCTION__)); |
28525 | break; |
28526 | case X86ISD::UNPCKL: { |
28527 | auto Op0 = N.getOperand(0); |
28528 | auto Op1 = N.getOperand(1); |
28529 | unsigned Opcode0 = Op0.getOpcode(); |
28530 | unsigned Opcode1 = Op1.getOpcode(); |
28531 | |
28532 | // Combine X86ISD::UNPCKL with 2 X86ISD::FHADD inputs into a single |
28533 | // X86ISD::FHADD. This is generated by UINT_TO_FP v2f64 scalarization. |
28534 | // TODO: Add other horizontal operations as required. |
28535 | if (VT == MVT::v2f64 && Opcode0 == Opcode1 && Opcode0 == X86ISD::FHADD) |
28536 | return DAG.getNode(Opcode0, DL, VT, Op0.getOperand(0), Op1.getOperand(0)); |
28537 | |
28538 | // Combine X86ISD::UNPCKL and ISD::VECTOR_SHUFFLE into X86ISD::UNPCKH, in |
28539 | // which X86ISD::UNPCKL has a ISD::UNDEF operand, and ISD::VECTOR_SHUFFLE |
28540 | // moves upper half elements into the lower half part. For example: |
28541 | // |
28542 | // t2: v16i8 = vector_shuffle<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u> t1, |
28543 | // undef:v16i8 |
28544 | // t3: v16i8 = X86ISD::UNPCKL undef:v16i8, t2 |
28545 | // |
28546 | // will be combined to: |
28547 | // |
28548 | // t3: v16i8 = X86ISD::UNPCKH undef:v16i8, t1 |
28549 | |
28550 | // This is only for 128-bit vectors. From SSE4.1 onward this combine may not |
28551 | // happen due to advanced instructions. |
28552 | if (!VT.is128BitVector()) |
28553 | return SDValue(); |
28554 | |
28555 | if (Op0.isUndef() && Opcode1 == ISD::VECTOR_SHUFFLE) { |
28556 | ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op1.getNode())->getMask(); |
28557 | |
28558 | unsigned NumElts = VT.getVectorNumElements(); |
28559 | SmallVector<int, 8> ExpectedMask(NumElts, -1); |
28560 | std::iota(ExpectedMask.begin(), ExpectedMask.begin() + NumElts / 2, |
28561 | NumElts / 2); |
28562 | |
28563 | auto ShufOp = Op1.getOperand(0); |
28564 | if (isShuffleEquivalent(Op1, ShufOp, Mask, ExpectedMask)) |
28565 | return DAG.getNode(X86ISD::UNPCKH, DL, VT, N.getOperand(0), ShufOp); |
28566 | } |
28567 | return SDValue(); |
28568 | } |
28569 | case X86ISD::BLENDI: { |
28570 | SDValue V0 = N->getOperand(0); |
28571 | SDValue V1 = N->getOperand(1); |
28572 | assert(VT == V0.getSimpleValueType() && VT == V1.getSimpleValueType() &&((VT == V0.getSimpleValueType() && VT == V1.getSimpleValueType () && "Unexpected input vector types") ? static_cast< void> (0) : __assert_fail ("VT == V0.getSimpleValueType() && VT == V1.getSimpleValueType() && \"Unexpected input vector types\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28573, __PRETTY_FUNCTION__)) |
28573 | "Unexpected input vector types")((VT == V0.getSimpleValueType() && VT == V1.getSimpleValueType () && "Unexpected input vector types") ? static_cast< void> (0) : __assert_fail ("VT == V0.getSimpleValueType() && VT == V1.getSimpleValueType() && \"Unexpected input vector types\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28573, __PRETTY_FUNCTION__)); |
28574 | |
28575 | // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector |
28576 | // operands and changing the mask to 1. This saves us a bunch of |
28577 | // pattern-matching possibilities related to scalar math ops in SSE/AVX. |
28578 | // x86InstrInfo knows how to commute this back after instruction selection |
28579 | // if it would help register allocation. |
28580 | |
28581 | // TODO: If optimizing for size or a processor that doesn't suffer from |
28582 | // partial register update stalls, this should be transformed into a MOVSD |
28583 | // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD. |
28584 | |
28585 | if (VT == MVT::v2f64) |
28586 | if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2))) |
28587 | if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) { |
28588 | SDValue NewMask = DAG.getConstant(1, DL, MVT::i8); |
28589 | return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask); |
28590 | } |
28591 | |
28592 | return SDValue(); |
28593 | } |
28594 | case X86ISD::MOVSD: |
28595 | case X86ISD::MOVSS: { |
28596 | SDValue V0 = peekThroughBitcasts(N->getOperand(0)); |
28597 | SDValue V1 = peekThroughBitcasts(N->getOperand(1)); |
28598 | bool isZero0 = ISD::isBuildVectorAllZeros(V0.getNode()); |
28599 | bool isZero1 = ISD::isBuildVectorAllZeros(V1.getNode()); |
28600 | if (isZero0 && isZero1) |
28601 | return SDValue(); |
28602 | |
28603 | // We often lower to MOVSD/MOVSS from integer as well as native float |
28604 | // types; remove unnecessary domain-crossing bitcasts if we can to make it |
28605 | // easier to combine shuffles later on. We've already accounted for the |
28606 | // domain switching cost when we decided to lower with it. |
28607 | bool isFloat = VT.isFloatingPoint(); |
28608 | bool isFloat0 = V0.getSimpleValueType().isFloatingPoint(); |
28609 | bool isFloat1 = V1.getSimpleValueType().isFloatingPoint(); |
28610 | if ((isFloat != isFloat0 || isZero0) && (isFloat != isFloat1 || isZero1)) { |
28611 | MVT NewVT = isFloat ? (X86ISD::MOVSD == Opcode ? MVT::v2i64 : MVT::v4i32) |
28612 | : (X86ISD::MOVSD == Opcode ? MVT::v2f64 : MVT::v4f32); |
28613 | V0 = DAG.getBitcast(NewVT, V0); |
28614 | V1 = DAG.getBitcast(NewVT, V1); |
28615 | return DAG.getBitcast(VT, DAG.getNode(Opcode, DL, NewVT, V0, V1)); |
28616 | } |
28617 | |
28618 | return SDValue(); |
28619 | } |
28620 | case X86ISD::INSERTPS: { |
28621 | assert(VT == MVT::v4f32 && "INSERTPS ValueType must be MVT::v4f32")((VT == MVT::v4f32 && "INSERTPS ValueType must be MVT::v4f32" ) ? static_cast<void> (0) : __assert_fail ("VT == MVT::v4f32 && \"INSERTPS ValueType must be MVT::v4f32\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28621, __PRETTY_FUNCTION__)); |
28622 | SDValue Op0 = N.getOperand(0); |
28623 | SDValue Op1 = N.getOperand(1); |
28624 | SDValue Op2 = N.getOperand(2); |
28625 | unsigned InsertPSMask = cast<ConstantSDNode>(Op2)->getZExtValue(); |
28626 | unsigned SrcIdx = (InsertPSMask >> 6) & 0x3; |
28627 | unsigned DstIdx = (InsertPSMask >> 4) & 0x3; |
28628 | unsigned ZeroMask = InsertPSMask & 0xF; |
28629 | |
28630 | // If we zero out all elements from Op0 then we don't need to reference it. |
28631 | if (((ZeroMask | (1u << DstIdx)) == 0xF) && !Op0.isUndef()) |
28632 | return DAG.getNode(X86ISD::INSERTPS, DL, VT, DAG.getUNDEF(VT), Op1, |
28633 | DAG.getConstant(InsertPSMask, DL, MVT::i8)); |
28634 | |
28635 | // If we zero out the element from Op1 then we don't need to reference it. |
28636 | if ((ZeroMask & (1u << DstIdx)) && !Op1.isUndef()) |
28637 | return DAG.getNode(X86ISD::INSERTPS, DL, VT, Op0, DAG.getUNDEF(VT), |
28638 | DAG.getConstant(InsertPSMask, DL, MVT::i8)); |
28639 | |
28640 | // Attempt to merge insertps Op1 with an inner target shuffle node. |
28641 | SmallVector<int, 8> TargetMask1; |
28642 | SmallVector<SDValue, 2> Ops1; |
28643 | if (setTargetShuffleZeroElements(Op1, TargetMask1, Ops1)) { |
28644 | int M = TargetMask1[SrcIdx]; |
28645 | if (isUndefOrZero(M)) { |
28646 | // Zero/UNDEF insertion - zero out element and remove dependency. |
28647 | InsertPSMask |= (1u << DstIdx); |
28648 | return DAG.getNode(X86ISD::INSERTPS, DL, VT, Op0, DAG.getUNDEF(VT), |
28649 | DAG.getConstant(InsertPSMask, DL, MVT::i8)); |
28650 | } |
28651 | // Update insertps mask srcidx and reference the source input directly. |
28652 | assert(0 <= M && M < 8 && "Shuffle index out of range")((0 <= M && M < 8 && "Shuffle index out of range" ) ? static_cast<void> (0) : __assert_fail ("0 <= M && M < 8 && \"Shuffle index out of range\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28652, __PRETTY_FUNCTION__)); |
28653 | InsertPSMask = (InsertPSMask & 0x3f) | ((M & 0x3) << 6); |
28654 | Op1 = Ops1[M < 4 ? 0 : 1]; |
28655 | return DAG.getNode(X86ISD::INSERTPS, DL, VT, Op0, Op1, |
28656 | DAG.getConstant(InsertPSMask, DL, MVT::i8)); |
28657 | } |
28658 | |
28659 | // Attempt to merge insertps Op0 with an inner target shuffle node. |
28660 | SmallVector<int, 8> TargetMask0; |
28661 | SmallVector<SDValue, 2> Ops0; |
28662 | if (!setTargetShuffleZeroElements(Op0, TargetMask0, Ops0)) |
28663 | return SDValue(); |
28664 | |
28665 | bool Updated = false; |
28666 | bool UseInput00 = false; |
28667 | bool UseInput01 = false; |
28668 | for (int i = 0; i != 4; ++i) { |
28669 | int M = TargetMask0[i]; |
28670 | if ((InsertPSMask & (1u << i)) || (i == (int)DstIdx)) { |
28671 | // No change if element is already zero or the inserted element. |
28672 | continue; |
28673 | } else if (isUndefOrZero(M)) { |
28674 | // If the target mask is undef/zero then we must zero the element. |
28675 | InsertPSMask |= (1u << i); |
28676 | Updated = true; |
28677 | continue; |
28678 | } |
28679 | |
28680 | // The input vector element must be inline. |
28681 | if (M != i && M != (i + 4)) |
28682 | return SDValue(); |
28683 | |
28684 | // Determine which inputs of the target shuffle we're using. |
28685 | UseInput00 |= (0 <= M && M < 4); |
28686 | UseInput01 |= (4 <= M); |
28687 | } |
28688 | |
28689 | // If we're not using both inputs of the target shuffle then use the |
28690 | // referenced input directly. |
28691 | if (UseInput00 && !UseInput01) { |
28692 | Updated = true; |
28693 | Op0 = Ops0[0]; |
28694 | } else if (!UseInput00 && UseInput01) { |
28695 | Updated = true; |
28696 | Op0 = Ops0[1]; |
28697 | } |
28698 | |
28699 | if (Updated) |
28700 | return DAG.getNode(X86ISD::INSERTPS, DL, VT, Op0, Op1, |
28701 | DAG.getConstant(InsertPSMask, DL, MVT::i8)); |
28702 | |
28703 | return SDValue(); |
28704 | } |
28705 | default: |
28706 | return SDValue(); |
28707 | } |
28708 | |
28709 | // Nuke no-op shuffles that show up after combining. |
28710 | if (isNoopShuffleMask(Mask)) |
28711 | return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true); |
28712 | |
28713 | // Look for simplifications involving one or two shuffle instructions. |
28714 | SDValue V = N.getOperand(0); |
28715 | switch (N.getOpcode()) { |
28716 | default: |
28717 | break; |
28718 | case X86ISD::PSHUFLW: |
28719 | case X86ISD::PSHUFHW: |
28720 | assert(VT.getVectorElementType() == MVT::i16 && "Bad word shuffle type!")((VT.getVectorElementType() == MVT::i16 && "Bad word shuffle type!" ) ? static_cast<void> (0) : __assert_fail ("VT.getVectorElementType() == MVT::i16 && \"Bad word shuffle type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 28720, __PRETTY_FUNCTION__)); |
28721 | |
28722 | if (combineRedundantHalfShuffle(N, Mask, DAG, DCI)) |
28723 | return SDValue(); // We combined away this shuffle, so we're done. |
28724 | |
28725 | // See if this reduces to a PSHUFD which is no more expensive and can |
28726 | // combine with more operations. Note that it has to at least flip the |
28727 | // dwords as otherwise it would have been removed as a no-op. |
28728 | if (makeArrayRef(Mask).equals({2, 3, 0, 1})) { |
28729 | int DMask[] = {0, 1, 2, 3}; |
28730 | int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2; |
28731 | DMask[DOffset + 0] = DOffset + 1; |
28732 | DMask[DOffset + 1] = DOffset + 0; |
28733 | MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2); |
28734 | V = DAG.getBitcast(DVT, V); |
28735 | DCI.AddToWorklist(V.getNode()); |
28736 | V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V, |
28737 | getV4X86ShuffleImm8ForMask(DMask, DL, DAG)); |
28738 | DCI.AddToWorklist(V.getNode()); |
28739 | return DAG.getBitcast(VT, V); |
28740 | } |
28741 | |
28742 | // Look for shuffle patterns which can be implemented as a single unpack. |
28743 | // FIXME: This doesn't handle the location of the PSHUFD generically, and |
28744 | // only works when we have a PSHUFD followed by two half-shuffles. |
28745 | if (Mask[0] == Mask[1] && Mask[2] == Mask[3] && |
28746 | (V.getOpcode() == X86ISD::PSHUFLW || |
28747 | V.getOpcode() == X86ISD::PSHUFHW) && |
28748 | V.getOpcode() != N.getOpcode() && |
28749 | V.hasOneUse()) { |
28750 | SDValue D = peekThroughOneUseBitcasts(V.getOperand(0)); |
28751 | if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) { |
28752 | SmallVector<int, 4> VMask = getPSHUFShuffleMask(V); |
28753 | SmallVector<int, 4> DMask = getPSHUFShuffleMask(D); |
28754 | int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4; |
28755 | int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4; |
28756 | int WordMask[8]; |
28757 | for (int i = 0; i < 4; ++i) { |
28758 | WordMask[i + NOffset] = Mask[i] + NOffset; |
28759 | WordMask[i + VOffset] = VMask[i] + VOffset; |
28760 | } |
28761 | // Map the word mask through the DWord mask. |
28762 | int MappedMask[8]; |
28763 | for (int i = 0; i < 8; ++i) |
28764 | MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2; |
28765 | if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) || |
28766 | makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) { |
28767 | // We can replace all three shuffles with an unpack. |
28768 | V = DAG.getBitcast(VT, D.getOperand(0)); |
28769 | DCI.AddToWorklist(V.getNode()); |
28770 | return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL |
28771 | : X86ISD::UNPCKH, |
28772 | DL, VT, V, V); |
28773 | } |
28774 | } |
28775 | } |
28776 | |
28777 | break; |
28778 | |
28779 | case X86ISD::PSHUFD: |
28780 | if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG)) |
28781 | return NewN; |
28782 | |
28783 | break; |
28784 | } |
28785 | |
28786 | return SDValue(); |
28787 | } |
28788 | |
28789 | /// Returns true iff the shuffle node \p N can be replaced with ADDSUB |
28790 | /// operation. If true is returned then the operands of ADDSUB operation |
28791 | /// are written to the parameters \p Opnd0 and \p Opnd1. |
28792 | /// |
28793 | /// We combine shuffle to ADDSUB directly on the abstract vector shuffle nodes |
28794 | /// so it is easier to generically match. We also insert dummy vector shuffle |
28795 | /// nodes for the operands which explicitly discard the lanes which are unused |
28796 | /// by this operation to try to flow through the rest of the combiner |
28797 | /// the fact that they're unused. |
28798 | static bool isAddSub(SDNode *N, const X86Subtarget &Subtarget, |
28799 | SDValue &Opnd0, SDValue &Opnd1) { |
28800 | |
28801 | EVT VT = N->getValueType(0); |
28802 | if ((!Subtarget.hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) && |
28803 | (!Subtarget.hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)) && |
28804 | (!Subtarget.hasAVX512() || (VT != MVT::v16f32 && VT != MVT::v8f64))) |
28805 | return false; |
28806 | |
28807 | // We only handle target-independent shuffles. |
28808 | // FIXME: It would be easy and harmless to use the target shuffle mask |
28809 | // extraction tool to support more. |
28810 | if (N->getOpcode() != ISD::VECTOR_SHUFFLE) |
28811 | return false; |
28812 | |
28813 | ArrayRef<int> OrigMask = cast<ShuffleVectorSDNode>(N)->getMask(); |
28814 | SmallVector<int, 16> Mask(OrigMask.begin(), OrigMask.end()); |
28815 | |
28816 | SDValue V1 = N->getOperand(0); |
28817 | SDValue V2 = N->getOperand(1); |
28818 | |
28819 | // We require the first shuffle operand to be the FSUB node, and the second to |
28820 | // be the FADD node. |
28821 | if (V1.getOpcode() == ISD::FADD && V2.getOpcode() == ISD::FSUB) { |
28822 | ShuffleVectorSDNode::commuteMask(Mask); |
28823 | std::swap(V1, V2); |
28824 | } else if (V1.getOpcode() != ISD::FSUB || V2.getOpcode() != ISD::FADD) |
28825 | return false; |
28826 | |
28827 | // If there are other uses of these operations we can't fold them. |
28828 | if (!V1->hasOneUse() || !V2->hasOneUse()) |
28829 | return false; |
28830 | |
28831 | // Ensure that both operations have the same operands. Note that we can |
28832 | // commute the FADD operands. |
28833 | SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1); |
28834 | if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) && |
28835 | (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS)) |
28836 | return false; |
28837 | |
28838 | // We're looking for blends between FADD and FSUB nodes. We insist on these |
28839 | // nodes being lined up in a specific expected pattern. |
28840 | if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) || |
28841 | isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) || |
28842 | isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15}) || |
28843 | isShuffleEquivalent(V1, V2, Mask, {0, 17, 2, 19, 4, 21, 6, 23, |
28844 | 8, 25, 10, 27, 12, 29, 14, 31}))) |
28845 | return false; |
28846 | |
28847 | Opnd0 = LHS; |
28848 | Opnd1 = RHS; |
28849 | return true; |
28850 | } |
28851 | |
28852 | /// \brief Try to combine a shuffle into a target-specific add-sub or |
28853 | /// mul-add-sub node. |
28854 | static SDValue combineShuffleToAddSubOrFMAddSub(SDNode *N, |
28855 | const X86Subtarget &Subtarget, |
28856 | SelectionDAG &DAG) { |
28857 | SDValue Opnd0, Opnd1; |
28858 | if (!isAddSub(N, Subtarget, Opnd0, Opnd1)) |
28859 | return SDValue(); |
28860 | |
28861 | EVT VT = N->getValueType(0); |
28862 | SDLoc DL(N); |
28863 | |
28864 | // Try to generate X86ISD::FMADDSUB node here. |
28865 | SDValue Opnd2; |
28866 | if (isFMAddSub(Subtarget, DAG, Opnd0, Opnd1, Opnd2)) |
28867 | return DAG.getNode(X86ISD::FMADDSUB, DL, VT, Opnd0, Opnd1, Opnd2); |
28868 | |
28869 | // Do not generate X86ISD::ADDSUB node for 512-bit types even though |
28870 | // the ADDSUB idiom has been successfully recognized. There are no known |
28871 | // X86 targets with 512-bit ADDSUB instructions! |
28872 | if (VT.is512BitVector()) |
28873 | return SDValue(); |
28874 | |
28875 | return DAG.getNode(X86ISD::ADDSUB, DL, VT, Opnd0, Opnd1); |
28876 | } |
28877 | |
28878 | // We are looking for a shuffle where both sources are concatenated with undef |
28879 | // and have a width that is half of the output's width. AVX2 has VPERMD/Q, so |
28880 | // if we can express this as a single-source shuffle, that's preferable. |
28881 | static SDValue combineShuffleOfConcatUndef(SDNode *N, SelectionDAG &DAG, |
28882 | const X86Subtarget &Subtarget) { |
28883 | if (!Subtarget.hasAVX2() || !isa<ShuffleVectorSDNode>(N)) |
28884 | return SDValue(); |
28885 | |
28886 | EVT VT = N->getValueType(0); |
28887 | |
28888 | // We only care about shuffles of 128/256-bit vectors of 32/64-bit values. |
28889 | if (!VT.is128BitVector() && !VT.is256BitVector()) |
28890 | return SDValue(); |
28891 | |
28892 | if (VT.getVectorElementType() != MVT::i32 && |
28893 | VT.getVectorElementType() != MVT::i64 && |
28894 | VT.getVectorElementType() != MVT::f32 && |
28895 | VT.getVectorElementType() != MVT::f64) |
28896 | return SDValue(); |
28897 | |
28898 | SDValue N0 = N->getOperand(0); |
28899 | SDValue N1 = N->getOperand(1); |
28900 | |
28901 | // Check that both sources are concats with undef. |
28902 | if (N0.getOpcode() != ISD::CONCAT_VECTORS || |
28903 | N1.getOpcode() != ISD::CONCAT_VECTORS || N0.getNumOperands() != 2 || |
28904 | N1.getNumOperands() != 2 || !N0.getOperand(1).isUndef() || |
28905 | !N1.getOperand(1).isUndef()) |
28906 | return SDValue(); |
28907 | |
28908 | // Construct the new shuffle mask. Elements from the first source retain their |
28909 | // index, but elements from the second source no longer need to skip an undef. |
28910 | SmallVector<int, 8> Mask; |
28911 | int NumElts = VT.getVectorNumElements(); |
28912 | |
28913 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N); |
28914 | for (int Elt : SVOp->getMask()) |
28915 | Mask.push_back(Elt < NumElts ? Elt : (Elt - NumElts / 2)); |
28916 | |
28917 | SDLoc DL(N); |
28918 | SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, N0.getOperand(0), |
28919 | N1.getOperand(0)); |
28920 | return DAG.getVectorShuffle(VT, DL, Concat, DAG.getUNDEF(VT), Mask); |
28921 | } |
28922 | |
28923 | static SDValue combineShuffle(SDNode *N, SelectionDAG &DAG, |
28924 | TargetLowering::DAGCombinerInfo &DCI, |
28925 | const X86Subtarget &Subtarget) { |
28926 | SDLoc dl(N); |
28927 | EVT VT = N->getValueType(0); |
28928 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
28929 | // If we have legalized the vector types, look for blends of FADD and FSUB |
28930 | // nodes that we can fuse into an ADDSUB node. |
28931 | if (TLI.isTypeLegal(VT)) |
28932 | if (SDValue AddSub = combineShuffleToAddSubOrFMAddSub(N, Subtarget, DAG)) |
28933 | return AddSub; |
28934 | |
28935 | // During Type Legalization, when promoting illegal vector types, |
28936 | // the backend might introduce new shuffle dag nodes and bitcasts. |
28937 | // |
28938 | // This code performs the following transformation: |
28939 | // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) -> |
28940 | // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>) |
28941 | // |
28942 | // We do this only if both the bitcast and the BINOP dag nodes have |
28943 | // one use. Also, perform this transformation only if the new binary |
28944 | // operation is legal. This is to avoid introducing dag nodes that |
28945 | // potentially need to be further expanded (or custom lowered) into a |
28946 | // less optimal sequence of dag nodes. |
28947 | if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() && |
28948 | N->getOpcode() == ISD::VECTOR_SHUFFLE && |
28949 | N->getOperand(0).getOpcode() == ISD::BITCAST && |
28950 | N->getOperand(1).isUndef() && N->getOperand(0).hasOneUse()) { |
28951 | SDValue N0 = N->getOperand(0); |
28952 | SDValue N1 = N->getOperand(1); |
28953 | |
28954 | SDValue BC0 = N0.getOperand(0); |
28955 | EVT SVT = BC0.getValueType(); |
28956 | unsigned Opcode = BC0.getOpcode(); |
28957 | unsigned NumElts = VT.getVectorNumElements(); |
28958 | |
28959 | if (BC0.hasOneUse() && SVT.isVector() && |
28960 | SVT.getVectorNumElements() * 2 == NumElts && |
28961 | TLI.isOperationLegal(Opcode, VT)) { |
28962 | bool CanFold = false; |
28963 | switch (Opcode) { |
28964 | default : break; |
28965 | case ISD::ADD: |
28966 | case ISD::SUB: |
28967 | case ISD::MUL: |
28968 | // isOperationLegal lies for integer ops on floating point types. |
28969 | CanFold = VT.isInteger(); |
28970 | break; |
28971 | case ISD::FADD: |
28972 | case ISD::FSUB: |
28973 | case ISD::FMUL: |
28974 | // isOperationLegal lies for floating point ops on integer types. |
28975 | CanFold = VT.isFloatingPoint(); |
28976 | break; |
28977 | } |
28978 | |
28979 | unsigned SVTNumElts = SVT.getVectorNumElements(); |
28980 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N); |
28981 | for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i) |
28982 | CanFold = SVOp->getMaskElt(i) == (int)(i * 2); |
28983 | for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i) |
28984 | CanFold = SVOp->getMaskElt(i) < 0; |
28985 | |
28986 | if (CanFold) { |
28987 | SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0)); |
28988 | SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1)); |
28989 | SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01); |
28990 | return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, SVOp->getMask()); |
28991 | } |
28992 | } |
28993 | } |
28994 | |
28995 | // Combine a vector_shuffle that is equal to build_vector load1, load2, load3, |
28996 | // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are |
28997 | // consecutive, non-overlapping, and in the right order. |
28998 | SmallVector<SDValue, 16> Elts; |
28999 | for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) { |
29000 | if (SDValue Elt = getShuffleScalarElt(N, i, DAG, 0)) { |
29001 | Elts.push_back(Elt); |
29002 | continue; |
29003 | } |
29004 | Elts.clear(); |
29005 | break; |
29006 | } |
29007 | |
29008 | if (Elts.size() == VT.getVectorNumElements()) |
29009 | if (SDValue LD = |
29010 | EltsFromConsecutiveLoads(VT, Elts, dl, DAG, Subtarget, true)) |
29011 | return LD; |
29012 | |
29013 | // For AVX2, we sometimes want to combine |
29014 | // (vector_shuffle <mask> (concat_vectors t1, undef) |
29015 | // (concat_vectors t2, undef)) |
29016 | // Into: |
29017 | // (vector_shuffle <mask> (concat_vectors t1, t2), undef) |
29018 | // Since the latter can be efficiently lowered with VPERMD/VPERMQ |
29019 | if (SDValue ShufConcat = combineShuffleOfConcatUndef(N, DAG, Subtarget)) |
29020 | return ShufConcat; |
29021 | |
29022 | if (isTargetShuffle(N->getOpcode())) { |
29023 | SDValue Op(N, 0); |
29024 | if (SDValue Shuffle = combineTargetShuffle(Op, DAG, DCI, Subtarget)) |
29025 | return Shuffle; |
29026 | |
29027 | // Try recursively combining arbitrary sequences of x86 shuffle |
29028 | // instructions into higher-order shuffles. We do this after combining |
29029 | // specific PSHUF instruction sequences into their minimal form so that we |
29030 | // can evaluate how many specialized shuffle instructions are involved in |
29031 | // a particular chain. |
29032 | SmallVector<int, 1> NonceMask; // Just a placeholder. |
29033 | NonceMask.push_back(0); |
29034 | if (combineX86ShufflesRecursively({Op}, 0, Op, NonceMask, {}, |
29035 | /*Depth*/ 1, /*HasVarMask*/ false, DAG, |
29036 | DCI, Subtarget)) |
29037 | return SDValue(); // This routine will use CombineTo to replace N. |
29038 | } |
29039 | |
29040 | return SDValue(); |
29041 | } |
29042 | |
29043 | /// Check if a vector extract from a target-specific shuffle of a load can be |
29044 | /// folded into a single element load. |
29045 | /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but |
29046 | /// shuffles have been custom lowered so we need to handle those here. |
29047 | static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG, |
29048 | TargetLowering::DAGCombinerInfo &DCI) { |
29049 | if (DCI.isBeforeLegalizeOps()) |
29050 | return SDValue(); |
29051 | |
29052 | SDValue InVec = N->getOperand(0); |
29053 | SDValue EltNo = N->getOperand(1); |
29054 | EVT EltVT = N->getValueType(0); |
29055 | |
29056 | if (!isa<ConstantSDNode>(EltNo)) |
29057 | return SDValue(); |
29058 | |
29059 | EVT OriginalVT = InVec.getValueType(); |
29060 | |
29061 | // Peek through bitcasts, don't duplicate a load with other uses. |
29062 | InVec = peekThroughOneUseBitcasts(InVec); |
29063 | |
29064 | EVT CurrentVT = InVec.getValueType(); |
29065 | if (!CurrentVT.isVector() || |
29066 | CurrentVT.getVectorNumElements() != OriginalVT.getVectorNumElements()) |
29067 | return SDValue(); |
29068 | |
29069 | if (!isTargetShuffle(InVec.getOpcode())) |
29070 | return SDValue(); |
29071 | |
29072 | // Don't duplicate a load with other uses. |
29073 | if (!InVec.hasOneUse()) |
29074 | return SDValue(); |
29075 | |
29076 | SmallVector<int, 16> ShuffleMask; |
29077 | SmallVector<SDValue, 2> ShuffleOps; |
29078 | bool UnaryShuffle; |
29079 | if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(), true, |
29080 | ShuffleOps, ShuffleMask, UnaryShuffle)) |
29081 | return SDValue(); |
29082 | |
29083 | // Select the input vector, guarding against out of range extract vector. |
29084 | unsigned NumElems = CurrentVT.getVectorNumElements(); |
29085 | int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue(); |
29086 | int Idx = (Elt > (int)NumElems) ? SM_SentinelUndef : ShuffleMask[Elt]; |
29087 | |
29088 | if (Idx == SM_SentinelZero) |
29089 | return EltVT.isInteger() ? DAG.getConstant(0, SDLoc(N), EltVT) |
29090 | : DAG.getConstantFP(+0.0, SDLoc(N), EltVT); |
29091 | if (Idx == SM_SentinelUndef) |
29092 | return DAG.getUNDEF(EltVT); |
29093 | |
29094 | assert(0 <= Idx && Idx < (int)(2 * NumElems) && "Shuffle index out of range")((0 <= Idx && Idx < (int)(2 * NumElems) && "Shuffle index out of range") ? static_cast<void> (0) : __assert_fail ("0 <= Idx && Idx < (int)(2 * NumElems) && \"Shuffle index out of range\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 29094, __PRETTY_FUNCTION__)); |
29095 | SDValue LdNode = (Idx < (int)NumElems) ? ShuffleOps[0] |
29096 | : ShuffleOps[1]; |
29097 | |
29098 | // If inputs to shuffle are the same for both ops, then allow 2 uses |
29099 | unsigned AllowedUses = |
29100 | (ShuffleOps.size() > 1 && ShuffleOps[0] == ShuffleOps[1]) ? 2 : 1; |
29101 | |
29102 | if (LdNode.getOpcode() == ISD::BITCAST) { |
29103 | // Don't duplicate a load with other uses. |
29104 | if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0)) |
29105 | return SDValue(); |
29106 | |
29107 | AllowedUses = 1; // only allow 1 load use if we have a bitcast |
29108 | LdNode = LdNode.getOperand(0); |
29109 | } |
29110 | |
29111 | if (!ISD::isNormalLoad(LdNode.getNode())) |
29112 | return SDValue(); |
29113 | |
29114 | LoadSDNode *LN0 = cast<LoadSDNode>(LdNode); |
29115 | |
29116 | if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile()) |
29117 | return SDValue(); |
29118 | |
29119 | // If there's a bitcast before the shuffle, check if the load type and |
29120 | // alignment is valid. |
29121 | unsigned Align = LN0->getAlignment(); |
29122 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
29123 | unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment( |
29124 | EltVT.getTypeForEVT(*DAG.getContext())); |
29125 | |
29126 | if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT)) |
29127 | return SDValue(); |
29128 | |
29129 | // All checks match so transform back to vector_shuffle so that DAG combiner |
29130 | // can finish the job |
29131 | SDLoc dl(N); |
29132 | |
29133 | // Create shuffle node taking into account the case that its a unary shuffle |
29134 | SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT) : ShuffleOps[1]; |
29135 | Shuffle = DAG.getVectorShuffle(CurrentVT, dl, ShuffleOps[0], Shuffle, |
29136 | ShuffleMask); |
29137 | Shuffle = DAG.getBitcast(OriginalVT, Shuffle); |
29138 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle, |
29139 | EltNo); |
29140 | } |
29141 | |
29142 | // Try to match patterns such as |
29143 | // (i16 bitcast (v16i1 x)) |
29144 | // -> |
29145 | // (i16 movmsk (16i8 sext (v16i1 x))) |
29146 | // before the illegal vector is scalarized on subtargets that don't have legal |
29147 | // vxi1 types. |
29148 | static SDValue combineBitcastvxi1(SelectionDAG &DAG, SDValue BitCast, |
29149 | const X86Subtarget &Subtarget) { |
29150 | EVT VT = BitCast.getValueType(); |
29151 | SDValue N0 = BitCast.getOperand(0); |
29152 | EVT VecVT = N0->getValueType(0); |
29153 | |
29154 | if (!VT.isScalarInteger() || !VecVT.isSimple()) |
29155 | return SDValue(); |
29156 | |
29157 | // With AVX512 vxi1 types are legal and we prefer using k-regs. |
29158 | // MOVMSK is supported in SSE2 or later. |
29159 | if (Subtarget.hasAVX512() || !Subtarget.hasSSE2()) |
29160 | return SDValue(); |
29161 | |
29162 | // There are MOVMSK flavors for types v16i8, v32i8, v4f32, v8f32, v4f64 and |
29163 | // v8f64. So all legal 128-bit and 256-bit vectors are covered except for |
29164 | // v8i16 and v16i16. |
29165 | // For these two cases, we can shuffle the upper element bytes to a |
29166 | // consecutive sequence at the start of the vector and treat the results as |
29167 | // v16i8 or v32i8, and for v61i8 this is the prefferable solution. However, |
29168 | // for v16i16 this is not the case, because the shuffle is expensive, so we |
29169 | // avoid sign-exteding to this type entirely. |
29170 | // For example, t0 := (v8i16 sext(v8i1 x)) needs to be shuffled as: |
29171 | // (v16i8 shuffle <0,2,4,6,8,10,12,14,u,u,...,u> (v16i8 bitcast t0), undef) |
29172 | MVT SExtVT; |
29173 | MVT FPCastVT = MVT::INVALID_SIMPLE_VALUE_TYPE; |
29174 | switch (VecVT.getSimpleVT().SimpleTy) { |
29175 | default: |
29176 | return SDValue(); |
29177 | case MVT::v2i1: |
29178 | SExtVT = MVT::v2i64; |
29179 | FPCastVT = MVT::v2f64; |
29180 | break; |
29181 | case MVT::v4i1: |
29182 | SExtVT = MVT::v4i32; |
29183 | FPCastVT = MVT::v4f32; |
29184 | // For cases such as (i4 bitcast (v4i1 setcc v4i64 v1, v2)) |
29185 | // sign-extend to a 256-bit operation to avoid truncation. |
29186 | if (N0->getOpcode() == ISD::SETCC && |
29187 | N0->getOperand(0)->getValueType(0).is256BitVector() && |
29188 | Subtarget.hasInt256()) { |
29189 | SExtVT = MVT::v4i64; |
29190 | FPCastVT = MVT::v4f64; |
29191 | } |
29192 | break; |
29193 | case MVT::v8i1: |
29194 | SExtVT = MVT::v8i16; |
29195 | // For cases such as (i8 bitcast (v8i1 setcc v8i32 v1, v2)), |
29196 | // sign-extend to a 256-bit operation to match the compare. |
29197 | // If the setcc operand is 128-bit, prefer sign-extending to 128-bit over |
29198 | // 256-bit because the shuffle is cheaper than sign extending the result of |
29199 | // the compare. |
29200 | if (N0->getOpcode() == ISD::SETCC && |
29201 | N0->getOperand(0)->getValueType(0).is256BitVector() && |
29202 | Subtarget.hasInt256()) { |
29203 | SExtVT = MVT::v8i32; |
29204 | FPCastVT = MVT::v8f32; |
29205 | } |
29206 | break; |
29207 | case MVT::v16i1: |
29208 | SExtVT = MVT::v16i8; |
29209 | // For the case (i16 bitcast (v16i1 setcc v16i16 v1, v2)), |
29210 | // it is not profitable to sign-extend to 256-bit because this will |
29211 | // require an extra cross-lane shuffle which is more exprensive than |
29212 | // truncating the result of the compare to 128-bits. |
29213 | break; |
29214 | case MVT::v32i1: |
29215 | // TODO: Handle pre-AVX2 cases by splitting to two v16i1's. |
29216 | if (!Subtarget.hasInt256()) |
29217 | return SDValue(); |
29218 | SExtVT = MVT::v32i8; |
29219 | break; |
29220 | }; |
29221 | |
29222 | SDLoc DL(BitCast); |
29223 | SDValue V = DAG.getSExtOrTrunc(N0, DL, SExtVT); |
29224 | if (SExtVT == MVT::v8i16) { |
29225 | V = DAG.getBitcast(MVT::v16i8, V); |
29226 | V = DAG.getVectorShuffle( |
29227 | MVT::v16i8, DL, V, DAG.getUNDEF(MVT::v16i8), |
29228 | {0, 2, 4, 6, 8, 10, 12, 14, -1, -1, -1, -1, -1, -1, -1, -1}); |
29229 | } else |
29230 | assert(SExtVT.getScalarType() != MVT::i16 &&((SExtVT.getScalarType() != MVT::i16 && "Vectors of i16 must be shuffled" ) ? static_cast<void> (0) : __assert_fail ("SExtVT.getScalarType() != MVT::i16 && \"Vectors of i16 must be shuffled\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 29231, __PRETTY_FUNCTION__)) |
29231 | "Vectors of i16 must be shuffled")((SExtVT.getScalarType() != MVT::i16 && "Vectors of i16 must be shuffled" ) ? static_cast<void> (0) : __assert_fail ("SExtVT.getScalarType() != MVT::i16 && \"Vectors of i16 must be shuffled\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 29231, __PRETTY_FUNCTION__)); |
29232 | if (FPCastVT != MVT::INVALID_SIMPLE_VALUE_TYPE) |
29233 | V = DAG.getBitcast(FPCastVT, V); |
29234 | V = DAG.getNode(X86ISD::MOVMSK, DL, MVT::i32, V); |
29235 | return DAG.getZExtOrTrunc(V, DL, VT); |
29236 | } |
29237 | |
29238 | static SDValue combineBitcast(SDNode *N, SelectionDAG &DAG, |
29239 | TargetLowering::DAGCombinerInfo &DCI, |
29240 | const X86Subtarget &Subtarget) { |
29241 | SDValue N0 = N->getOperand(0); |
29242 | EVT VT = N->getValueType(0); |
29243 | EVT SrcVT = N0.getValueType(); |
29244 | |
29245 | // Try to match patterns such as |
29246 | // (i16 bitcast (v16i1 x)) |
29247 | // -> |
29248 | // (i16 movmsk (16i8 sext (v16i1 x))) |
29249 | // before the setcc result is scalarized on subtargets that don't have legal |
29250 | // vxi1 types. |
29251 | if (DCI.isBeforeLegalize()) |
29252 | if (SDValue V = combineBitcastvxi1(DAG, SDValue(N, 0), Subtarget)) |
29253 | return V; |
29254 | // Since MMX types are special and don't usually play with other vector types, |
29255 | // it's better to handle them early to be sure we emit efficient code by |
29256 | // avoiding store-load conversions. |
29257 | |
29258 | // Detect bitcasts between i32 to x86mmx low word. |
29259 | if (VT == MVT::x86mmx && N0.getOpcode() == ISD::BUILD_VECTOR && |
29260 | SrcVT == MVT::v2i32 && isNullConstant(N0.getOperand(1))) { |
29261 | SDValue N00 = N0->getOperand(0); |
29262 | if (N00.getValueType() == MVT::i32) |
29263 | return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00); |
29264 | } |
29265 | |
29266 | // Detect bitcasts between element or subvector extraction to x86mmx. |
29267 | if (VT == MVT::x86mmx && |
29268 | (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT || |
29269 | N0.getOpcode() == ISD::EXTRACT_SUBVECTOR) && |
29270 | isNullConstant(N0.getOperand(1))) { |
29271 | SDValue N00 = N0->getOperand(0); |
29272 | if (N00.getValueType().is128BitVector()) |
29273 | return DAG.getNode(X86ISD::MOVDQ2Q, SDLoc(N00), VT, |
29274 | DAG.getBitcast(MVT::v2i64, N00)); |
29275 | } |
29276 | |
29277 | // Detect bitcasts from FP_TO_SINT to x86mmx. |
29278 | if (VT == MVT::x86mmx && SrcVT == MVT::v2i32 && |
29279 | N0.getOpcode() == ISD::FP_TO_SINT) { |
29280 | SDLoc DL(N0); |
29281 | SDValue Res = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4i32, N0, |
29282 | DAG.getUNDEF(MVT::v2i32)); |
29283 | return DAG.getNode(X86ISD::MOVDQ2Q, DL, VT, |
29284 | DAG.getBitcast(MVT::v2i64, Res)); |
29285 | } |
29286 | |
29287 | // Convert a bitcasted integer logic operation that has one bitcasted |
29288 | // floating-point operand into a floating-point logic operation. This may |
29289 | // create a load of a constant, but that is cheaper than materializing the |
29290 | // constant in an integer register and transferring it to an SSE register or |
29291 | // transferring the SSE operand to integer register and back. |
29292 | unsigned FPOpcode; |
29293 | switch (N0.getOpcode()) { |
29294 | case ISD::AND: FPOpcode = X86ISD::FAND; break; |
29295 | case ISD::OR: FPOpcode = X86ISD::FOR; break; |
29296 | case ISD::XOR: FPOpcode = X86ISD::FXOR; break; |
29297 | default: return SDValue(); |
29298 | } |
29299 | |
29300 | if (!((Subtarget.hasSSE1() && VT == MVT::f32) || |
29301 | (Subtarget.hasSSE2() && VT == MVT::f64))) |
29302 | return SDValue(); |
29303 | |
29304 | SDValue LogicOp0 = N0.getOperand(0); |
29305 | SDValue LogicOp1 = N0.getOperand(1); |
29306 | SDLoc DL0(N0); |
29307 | |
29308 | // bitcast(logic(bitcast(X), Y)) --> logic'(X, bitcast(Y)) |
29309 | if (N0.hasOneUse() && LogicOp0.getOpcode() == ISD::BITCAST && |
29310 | LogicOp0.hasOneUse() && LogicOp0.getOperand(0).getValueType() == VT && |
29311 | !isa<ConstantSDNode>(LogicOp0.getOperand(0))) { |
29312 | SDValue CastedOp1 = DAG.getBitcast(VT, LogicOp1); |
29313 | return DAG.getNode(FPOpcode, DL0, VT, LogicOp0.getOperand(0), CastedOp1); |
29314 | } |
29315 | // bitcast(logic(X, bitcast(Y))) --> logic'(bitcast(X), Y) |
29316 | if (N0.hasOneUse() && LogicOp1.getOpcode() == ISD::BITCAST && |
29317 | LogicOp1.hasOneUse() && LogicOp1.getOperand(0).getValueType() == VT && |
29318 | !isa<ConstantSDNode>(LogicOp1.getOperand(0))) { |
29319 | SDValue CastedOp0 = DAG.getBitcast(VT, LogicOp0); |
29320 | return DAG.getNode(FPOpcode, DL0, VT, LogicOp1.getOperand(0), CastedOp0); |
29321 | } |
29322 | |
29323 | return SDValue(); |
29324 | } |
29325 | |
29326 | // Match a binop + shuffle pyramid that represents a horizontal reduction over |
29327 | // the elements of a vector. |
29328 | // Returns the vector that is being reduced on, or SDValue() if a reduction |
29329 | // was not matched. |
29330 | static SDValue matchBinOpReduction(SDNode *Extract, ISD::NodeType BinOp) { |
29331 | // The pattern must end in an extract from index 0. |
29332 | if ((Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT) || |
29333 | !isNullConstant(Extract->getOperand(1))) |
29334 | return SDValue(); |
29335 | |
29336 | unsigned Stages = |
29337 | Log2_32(Extract->getOperand(0).getValueType().getVectorNumElements()); |
29338 | |
29339 | SDValue Op = Extract->getOperand(0); |
29340 | // At each stage, we're looking for something that looks like: |
29341 | // %s = shufflevector <8 x i32> %op, <8 x i32> undef, |
29342 | // <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, |
29343 | // i32 undef, i32 undef, i32 undef, i32 undef> |
29344 | // %a = binop <8 x i32> %op, %s |
29345 | // Where the mask changes according to the stage. E.g. for a 3-stage pyramid, |
29346 | // we expect something like: |
29347 | // <4,5,6,7,u,u,u,u> |
29348 | // <2,3,u,u,u,u,u,u> |
29349 | // <1,u,u,u,u,u,u,u> |
29350 | for (unsigned i = 0; i < Stages; ++i) { |
29351 | if (Op.getOpcode() != BinOp) |
29352 | return SDValue(); |
29353 | |
29354 | ShuffleVectorSDNode *Shuffle = |
29355 | dyn_cast<ShuffleVectorSDNode>(Op.getOperand(0).getNode()); |
29356 | if (Shuffle) { |
29357 | Op = Op.getOperand(1); |
29358 | } else { |
29359 | Shuffle = dyn_cast<ShuffleVectorSDNode>(Op.getOperand(1).getNode()); |
29360 | Op = Op.getOperand(0); |
29361 | } |
29362 | |
29363 | // The first operand of the shuffle should be the same as the other operand |
29364 | // of the add. |
29365 | if (!Shuffle || (Shuffle->getOperand(0) != Op)) |
29366 | return SDValue(); |
29367 | |
29368 | // Verify the shuffle has the expected (at this stage of the pyramid) mask. |
29369 | for (int Index = 0, MaskEnd = 1 << i; Index < MaskEnd; ++Index) |
29370 | if (Shuffle->getMaskElt(Index) != MaskEnd + Index) |
29371 | return SDValue(); |
29372 | } |
29373 | |
29374 | return Op; |
29375 | } |
29376 | |
29377 | // Given a select, detect the following pattern: |
29378 | // 1: %2 = zext <N x i8> %0 to <N x i32> |
29379 | // 2: %3 = zext <N x i8> %1 to <N x i32> |
29380 | // 3: %4 = sub nsw <N x i32> %2, %3 |
29381 | // 4: %5 = icmp sgt <N x i32> %4, [0 x N] or [-1 x N] |
29382 | // 5: %6 = sub nsw <N x i32> zeroinitializer, %4 |
29383 | // 6: %7 = select <N x i1> %5, <N x i32> %4, <N x i32> %6 |
29384 | // This is useful as it is the input into a SAD pattern. |
29385 | static bool detectZextAbsDiff(const SDValue &Select, SDValue &Op0, |
29386 | SDValue &Op1) { |
29387 | // Check the condition of the select instruction is greater-than. |
29388 | SDValue SetCC = Select->getOperand(0); |
29389 | if (SetCC.getOpcode() != ISD::SETCC) |
29390 | return false; |
29391 | ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get(); |
29392 | if (CC != ISD::SETGT && CC != ISD::SETLT) |
29393 | return false; |
29394 | |
29395 | SDValue SelectOp1 = Select->getOperand(1); |
29396 | SDValue SelectOp2 = Select->getOperand(2); |
29397 | |
29398 | // The following instructions assume SelectOp1 is the subtraction operand |
29399 | // and SelectOp2 is the negation operand. |
29400 | // In the case of SETLT this is the other way around. |
29401 | if (CC == ISD::SETLT) |
29402 | std::swap(SelectOp1, SelectOp2); |
29403 | |
29404 | // The second operand of the select should be the negation of the first |
29405 | // operand, which is implemented as 0 - SelectOp1. |
29406 | if (!(SelectOp2.getOpcode() == ISD::SUB && |
29407 | ISD::isBuildVectorAllZeros(SelectOp2.getOperand(0).getNode()) && |
29408 | SelectOp2.getOperand(1) == SelectOp1)) |
29409 | return false; |
29410 | |
29411 | // The first operand of SetCC is the first operand of the select, which is the |
29412 | // difference between the two input vectors. |
29413 | if (SetCC.getOperand(0) != SelectOp1) |
29414 | return false; |
29415 | |
29416 | // In SetLT case, The second operand of the comparison can be either 1 or 0. |
29417 | APInt SplatVal; |
29418 | if ((CC == ISD::SETLT) && |
29419 | !((ISD::isConstantSplatVector(SetCC.getOperand(1).getNode(), SplatVal) && |
29420 | SplatVal == 1) || |
29421 | (ISD::isBuildVectorAllZeros(SetCC.getOperand(1).getNode())))) |
29422 | return false; |
29423 | |
29424 | // In SetGT case, The second operand of the comparison can be either -1 or 0. |
29425 | if ((CC == ISD::SETGT) && |
29426 | !(ISD::isBuildVectorAllZeros(SetCC.getOperand(1).getNode()) || |
29427 | ISD::isBuildVectorAllOnes(SetCC.getOperand(1).getNode()))) |
29428 | return false; |
29429 | |
29430 | // The first operand of the select is the difference between the two input |
29431 | // vectors. |
29432 | if (SelectOp1.getOpcode() != ISD::SUB) |
29433 | return false; |
29434 | |
29435 | Op0 = SelectOp1.getOperand(0); |
29436 | Op1 = SelectOp1.getOperand(1); |
29437 | |
29438 | // Check if the operands of the sub are zero-extended from vectors of i8. |
29439 | if (Op0.getOpcode() != ISD::ZERO_EXTEND || |
29440 | Op0.getOperand(0).getValueType().getVectorElementType() != MVT::i8 || |
29441 | Op1.getOpcode() != ISD::ZERO_EXTEND || |
29442 | Op1.getOperand(0).getValueType().getVectorElementType() != MVT::i8) |
29443 | return false; |
29444 | |
29445 | return true; |
29446 | } |
29447 | |
29448 | // Given two zexts of <k x i8> to <k x i32>, create a PSADBW of the inputs |
29449 | // to these zexts. |
29450 | static SDValue createPSADBW(SelectionDAG &DAG, const SDValue &Zext0, |
29451 | const SDValue &Zext1, const SDLoc &DL) { |
29452 | |
29453 | // Find the appropriate width for the PSADBW. |
29454 | EVT InVT = Zext0.getOperand(0).getValueType(); |
29455 | unsigned RegSize = std::max(128u, InVT.getSizeInBits()); |
29456 | |
29457 | // "Zero-extend" the i8 vectors. This is not a per-element zext, rather we |
29458 | // fill in the missing vector elements with 0. |
29459 | unsigned NumConcat = RegSize / InVT.getSizeInBits(); |
29460 | SmallVector<SDValue, 16> Ops(NumConcat, DAG.getConstant(0, DL, InVT)); |
29461 | Ops[0] = Zext0.getOperand(0); |
29462 | MVT ExtendedVT = MVT::getVectorVT(MVT::i8, RegSize / 8); |
29463 | SDValue SadOp0 = DAG.getNode(ISD::CONCAT_VECTORS, DL, ExtendedVT, Ops); |
29464 | Ops[0] = Zext1.getOperand(0); |
29465 | SDValue SadOp1 = DAG.getNode(ISD::CONCAT_VECTORS, DL, ExtendedVT, Ops); |
29466 | |
29467 | // Actually build the SAD |
29468 | MVT SadVT = MVT::getVectorVT(MVT::i64, RegSize / 64); |
29469 | return DAG.getNode(X86ISD::PSADBW, DL, SadVT, SadOp0, SadOp1); |
29470 | } |
29471 | |
29472 | // Attempt to replace an all_of/any_of style horizontal reduction with a MOVMSK. |
29473 | static SDValue combineHorizontalPredicateResult(SDNode *Extract, |
29474 | SelectionDAG &DAG, |
29475 | const X86Subtarget &Subtarget) { |
29476 | // Bail without SSE2 or with AVX512VL (which uses predicate registers). |
29477 | if (!Subtarget.hasSSE2() || Subtarget.hasVLX()) |
29478 | return SDValue(); |
29479 | |
29480 | EVT ExtractVT = Extract->getValueType(0); |
29481 | unsigned BitWidth = ExtractVT.getSizeInBits(); |
29482 | if (ExtractVT != MVT::i64 && ExtractVT != MVT::i32 && ExtractVT != MVT::i16 && |
29483 | ExtractVT != MVT::i8) |
29484 | return SDValue(); |
29485 | |
29486 | // Check for OR(any_of) and AND(all_of) horizontal reduction patterns. |
29487 | for (ISD::NodeType Op : {ISD::OR, ISD::AND}) { |
29488 | SDValue Match = matchBinOpReduction(Extract, Op); |
29489 | if (!Match) |
29490 | continue; |
29491 | |
29492 | // EXTRACT_VECTOR_ELT can require implicit extension of the vector element |
29493 | // which we can't support here for now. |
29494 | if (Match.getScalarValueSizeInBits() != BitWidth) |
29495 | continue; |
29496 | |
29497 | // We require AVX2 for PMOVMSKB for v16i16/v32i8; |
29498 | unsigned MatchSizeInBits = Match.getValueSizeInBits(); |
29499 | if (!(MatchSizeInBits == 128 || |
29500 | (MatchSizeInBits == 256 && |
29501 | ((Subtarget.hasAVX() && BitWidth >= 32) || Subtarget.hasAVX2())))) |
29502 | return SDValue(); |
29503 | |
29504 | // Don't bother performing this for 2-element vectors. |
29505 | if (Match.getValueType().getVectorNumElements() <= 2) |
29506 | return SDValue(); |
29507 | |
29508 | // Check that we are extracting a reduction of all sign bits. |
29509 | if (DAG.ComputeNumSignBits(Match) != BitWidth) |
29510 | return SDValue(); |
29511 | |
29512 | // For 32/64 bit comparisons use MOVMSKPS/MOVMSKPD, else PMOVMSKB. |
29513 | MVT MaskVT; |
29514 | if (64 == BitWidth || 32 == BitWidth) |
29515 | MaskVT = MVT::getVectorVT(MVT::getFloatingPointVT(BitWidth), |
29516 | MatchSizeInBits / BitWidth); |
29517 | else |
29518 | MaskVT = MVT::getVectorVT(MVT::i8, MatchSizeInBits / 8); |
29519 | |
29520 | APInt CompareBits; |
29521 | ISD::CondCode CondCode; |
29522 | if (Op == ISD::OR) { |
29523 | // any_of -> MOVMSK != 0 |
29524 | CompareBits = APInt::getNullValue(32); |
29525 | CondCode = ISD::CondCode::SETNE; |
29526 | } else { |
29527 | // all_of -> MOVMSK == ((1 << NumElts) - 1) |
29528 | CompareBits = APInt::getLowBitsSet(32, MaskVT.getVectorNumElements()); |
29529 | CondCode = ISD::CondCode::SETEQ; |
29530 | } |
29531 | |
29532 | // Perform the select as i32/i64 and then truncate to avoid partial register |
29533 | // stalls. |
29534 | unsigned ResWidth = std::max(BitWidth, 32u); |
29535 | EVT ResVT = EVT::getIntegerVT(*DAG.getContext(), ResWidth); |
29536 | SDLoc DL(Extract); |
29537 | SDValue Zero = DAG.getConstant(0, DL, ResVT); |
29538 | SDValue Ones = DAG.getAllOnesConstant(DL, ResVT); |
29539 | SDValue Res = DAG.getBitcast(MaskVT, Match); |
29540 | Res = DAG.getNode(X86ISD::MOVMSK, DL, MVT::i32, Res); |
29541 | Res = DAG.getSelectCC(DL, Res, DAG.getConstant(CompareBits, DL, MVT::i32), |
29542 | Ones, Zero, CondCode); |
29543 | return DAG.getSExtOrTrunc(Res, DL, ExtractVT); |
29544 | } |
29545 | |
29546 | return SDValue(); |
29547 | } |
29548 | |
29549 | static SDValue combineBasicSADPattern(SDNode *Extract, SelectionDAG &DAG, |
29550 | const X86Subtarget &Subtarget) { |
29551 | // PSADBW is only supported on SSE2 and up. |
29552 | if (!Subtarget.hasSSE2()) |
29553 | return SDValue(); |
29554 | |
29555 | // Verify the type we're extracting from is any integer type above i16. |
29556 | EVT VT = Extract->getOperand(0).getValueType(); |
29557 | if (!VT.isSimple() || !(VT.getVectorElementType().getSizeInBits() > 16)) |
29558 | return SDValue(); |
29559 | |
29560 | unsigned RegSize = 128; |
29561 | if (Subtarget.hasBWI()) |
29562 | RegSize = 512; |
29563 | else if (Subtarget.hasAVX2()) |
29564 | RegSize = 256; |
29565 | |
29566 | // We handle upto v16i* for SSE2 / v32i* for AVX2 / v64i* for AVX512. |
29567 | // TODO: We should be able to handle larger vectors by splitting them before |
29568 | // feeding them into several SADs, and then reducing over those. |
29569 | if (RegSize / VT.getVectorNumElements() < 8) |
29570 | return SDValue(); |
29571 | |
29572 | // Match shuffle + add pyramid. |
29573 | SDValue Root = matchBinOpReduction(Extract, ISD::ADD); |
29574 | |
29575 | // The operand is expected to be zero extended from i8 |
29576 | // (verified in detectZextAbsDiff). |
29577 | // In order to convert to i64 and above, additional any/zero/sign |
29578 | // extend is expected. |
29579 | // The zero extend from 32 bit has no mathematical effect on the result. |
29580 | // Also the sign extend is basically zero extend |
29581 | // (extends the sign bit which is zero). |
29582 | // So it is correct to skip the sign/zero extend instruction. |
29583 | if (Root && (Root.getOpcode() == ISD::SIGN_EXTEND || |
29584 | Root.getOpcode() == ISD::ZERO_EXTEND || |
29585 | Root.getOpcode() == ISD::ANY_EXTEND)) |
29586 | Root = Root.getOperand(0); |
29587 | |
29588 | // If there was a match, we want Root to be a select that is the root of an |
29589 | // abs-diff pattern. |
29590 | if (!Root || (Root.getOpcode() != ISD::VSELECT)) |
29591 | return SDValue(); |
29592 | |
29593 | // Check whether we have an abs-diff pattern feeding into the select. |
29594 | SDValue Zext0, Zext1; |
29595 | if (!detectZextAbsDiff(Root, Zext0, Zext1)) |
29596 | return SDValue(); |
29597 | |
29598 | // Create the SAD instruction. |
29599 | SDLoc DL(Extract); |
29600 | SDValue SAD = createPSADBW(DAG, Zext0, Zext1, DL); |
29601 | |
29602 | // If the original vector was wider than 8 elements, sum over the results |
29603 | // in the SAD vector. |
29604 | unsigned Stages = Log2_32(VT.getVectorNumElements()); |
29605 | MVT SadVT = SAD.getSimpleValueType(); |
29606 | if (Stages > 3) { |
29607 | unsigned SadElems = SadVT.getVectorNumElements(); |
29608 | |
29609 | for(unsigned i = Stages - 3; i > 0; --i) { |
29610 | SmallVector<int, 16> Mask(SadElems, -1); |
29611 | for(unsigned j = 0, MaskEnd = 1 << (i - 1); j < MaskEnd; ++j) |
29612 | Mask[j] = MaskEnd + j; |
29613 | |
29614 | SDValue Shuffle = |
29615 | DAG.getVectorShuffle(SadVT, DL, SAD, DAG.getUNDEF(SadVT), Mask); |
29616 | SAD = DAG.getNode(ISD::ADD, DL, SadVT, SAD, Shuffle); |
29617 | } |
29618 | } |
29619 | |
29620 | MVT Type = Extract->getSimpleValueType(0); |
29621 | unsigned TypeSizeInBits = Type.getSizeInBits(); |
29622 | // Return the lowest TypeSizeInBits bits. |
29623 | MVT ResVT = MVT::getVectorVT(Type, SadVT.getSizeInBits() / TypeSizeInBits); |
29624 | SAD = DAG.getNode(ISD::BITCAST, DL, ResVT, SAD); |
29625 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, Type, SAD, |
29626 | Extract->getOperand(1)); |
29627 | } |
29628 | |
29629 | // Attempt to peek through a target shuffle and extract the scalar from the |
29630 | // source. |
29631 | static SDValue combineExtractWithShuffle(SDNode *N, SelectionDAG &DAG, |
29632 | TargetLowering::DAGCombinerInfo &DCI, |
29633 | const X86Subtarget &Subtarget) { |
29634 | if (DCI.isBeforeLegalizeOps()) |
29635 | return SDValue(); |
29636 | |
29637 | SDValue Src = N->getOperand(0); |
29638 | SDValue Idx = N->getOperand(1); |
29639 | |
29640 | EVT VT = N->getValueType(0); |
29641 | EVT SrcVT = Src.getValueType(); |
29642 | EVT SrcSVT = SrcVT.getVectorElementType(); |
29643 | unsigned NumSrcElts = SrcVT.getVectorNumElements(); |
29644 | |
29645 | // Don't attempt this for boolean mask vectors or unknown extraction indices. |
29646 | if (SrcSVT == MVT::i1 || !isa<ConstantSDNode>(Idx)) |
29647 | return SDValue(); |
29648 | |
29649 | // Resolve the target shuffle inputs and mask. |
29650 | SmallVector<int, 16> Mask; |
29651 | SmallVector<SDValue, 2> Ops; |
29652 | if (!resolveTargetShuffleInputs(peekThroughBitcasts(Src), Ops, Mask, DAG)) |
29653 | return SDValue(); |
29654 | |
29655 | // Attempt to narrow/widen the shuffle mask to the correct size. |
29656 | if (Mask.size() != NumSrcElts) { |
29657 | if ((NumSrcElts % Mask.size()) == 0) { |
29658 | SmallVector<int, 16> ScaledMask; |
29659 | int Scale = NumSrcElts / Mask.size(); |
29660 | scaleShuffleMask(Scale, Mask, ScaledMask); |
29661 | Mask = std::move(ScaledMask); |
29662 | } else if ((Mask.size() % NumSrcElts) == 0) { |
29663 | SmallVector<int, 16> WidenedMask; |
29664 | while (Mask.size() > NumSrcElts && |
29665 | canWidenShuffleElements(Mask, WidenedMask)) |
29666 | Mask = std::move(WidenedMask); |
29667 | // TODO - investigate support for wider shuffle masks with known upper |
29668 | // undef/zero elements for implicit zero-extension. |
29669 | } |
29670 | } |
29671 | |
29672 | // Check if narrowing/widening failed. |
29673 | if (Mask.size() != NumSrcElts) |
29674 | return SDValue(); |
29675 | |
29676 | int SrcIdx = Mask[N->getConstantOperandVal(1)]; |
29677 | SDLoc dl(N); |
29678 | |
29679 | // If the shuffle source element is undef/zero then we can just accept it. |
29680 | if (SrcIdx == SM_SentinelUndef) |
29681 | return DAG.getUNDEF(VT); |
29682 | |
29683 | if (SrcIdx == SM_SentinelZero) |
29684 | return VT.isFloatingPoint() ? DAG.getConstantFP(0.0, dl, VT) |
29685 | : DAG.getConstant(0, dl, VT); |
29686 | |
29687 | SDValue SrcOp = Ops[SrcIdx / Mask.size()]; |
29688 | SrcOp = DAG.getBitcast(SrcVT, SrcOp); |
29689 | SrcIdx = SrcIdx % Mask.size(); |
29690 | |
29691 | // We can only extract other elements from 128-bit vectors and in certain |
29692 | // circumstances, depending on SSE-level. |
29693 | // TODO: Investigate using extract_subvector for larger vectors. |
29694 | // TODO: Investigate float/double extraction if it will be just stored. |
29695 | if ((SrcVT == MVT::v4i32 || SrcVT == MVT::v2i64) && |
29696 | ((SrcIdx == 0 && Subtarget.hasSSE2()) || Subtarget.hasSSE41())) { |
29697 | assert(SrcSVT == VT && "Unexpected extraction type")((SrcSVT == VT && "Unexpected extraction type") ? static_cast <void> (0) : __assert_fail ("SrcSVT == VT && \"Unexpected extraction type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 29697, __PRETTY_FUNCTION__)); |
29698 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcSVT, SrcOp, |
29699 | DAG.getIntPtrConstant(SrcIdx, dl)); |
29700 | } |
29701 | |
29702 | if ((SrcVT == MVT::v8i16 && Subtarget.hasSSE2()) || |
29703 | (SrcVT == MVT::v16i8 && Subtarget.hasSSE41())) { |
29704 | assert(VT.getSizeInBits() >= SrcSVT.getSizeInBits() &&((VT.getSizeInBits() >= SrcSVT.getSizeInBits() && "Unexpected extraction type" ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() >= SrcSVT.getSizeInBits() && \"Unexpected extraction type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 29705, __PRETTY_FUNCTION__)) |
29705 | "Unexpected extraction type")((VT.getSizeInBits() >= SrcSVT.getSizeInBits() && "Unexpected extraction type" ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() >= SrcSVT.getSizeInBits() && \"Unexpected extraction type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 29705, __PRETTY_FUNCTION__)); |
29706 | unsigned OpCode = (SrcVT == MVT::v8i16 ? X86ISD::PEXTRW : X86ISD::PEXTRB); |
29707 | SDValue ExtOp = DAG.getNode(OpCode, dl, MVT::i32, SrcOp, |
29708 | DAG.getIntPtrConstant(SrcIdx, dl)); |
29709 | SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, ExtOp, |
29710 | DAG.getValueType(SrcSVT)); |
29711 | return DAG.getZExtOrTrunc(Assert, dl, VT); |
29712 | } |
29713 | |
29714 | return SDValue(); |
29715 | } |
29716 | |
29717 | /// Detect vector gather/scatter index generation and convert it from being a |
29718 | /// bunch of shuffles and extracts into a somewhat faster sequence. |
29719 | /// For i686, the best sequence is apparently storing the value and loading |
29720 | /// scalars back, while for x64 we should use 64-bit extracts and shifts. |
29721 | static SDValue combineExtractVectorElt(SDNode *N, SelectionDAG &DAG, |
29722 | TargetLowering::DAGCombinerInfo &DCI, |
29723 | const X86Subtarget &Subtarget) { |
29724 | if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI)) |
29725 | return NewOp; |
29726 | |
29727 | if (SDValue NewOp = combineExtractWithShuffle(N, DAG, DCI, Subtarget)) |
29728 | return NewOp; |
29729 | |
29730 | SDValue InputVector = N->getOperand(0); |
29731 | SDValue EltIdx = N->getOperand(1); |
29732 | |
29733 | EVT SrcVT = InputVector.getValueType(); |
29734 | EVT VT = N->getValueType(0); |
29735 | SDLoc dl(InputVector); |
29736 | |
29737 | // Detect mmx extraction of all bits as a i64. It works better as a bitcast. |
29738 | if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() && |
29739 | VT == MVT::i64 && SrcVT == MVT::v1i64 && isNullConstant(EltIdx)) { |
29740 | SDValue MMXSrc = InputVector.getOperand(0); |
29741 | |
29742 | // The bitcast source is a direct mmx result. |
29743 | if (MMXSrc.getValueType() == MVT::x86mmx) |
29744 | return DAG.getBitcast(VT, InputVector); |
29745 | } |
29746 | |
29747 | // Detect mmx to i32 conversion through a v2i32 elt extract. |
29748 | if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() && |
29749 | VT == MVT::i32 && SrcVT == MVT::v2i32 && isNullConstant(EltIdx)) { |
29750 | SDValue MMXSrc = InputVector.getOperand(0); |
29751 | |
29752 | // The bitcast source is a direct mmx result. |
29753 | if (MMXSrc.getValueType() == MVT::x86mmx) |
29754 | return DAG.getNode(X86ISD::MMX_MOVD2W, dl, MVT::i32, MMXSrc); |
29755 | } |
29756 | |
29757 | if (VT == MVT::i1 && InputVector.getOpcode() == ISD::BITCAST && |
29758 | isa<ConstantSDNode>(EltIdx) && |
29759 | isa<ConstantSDNode>(InputVector.getOperand(0))) { |
29760 | uint64_t ExtractedElt = N->getConstantOperandVal(1); |
29761 | uint64_t InputValue = InputVector.getConstantOperandVal(0); |
29762 | uint64_t Res = (InputValue >> ExtractedElt) & 1; |
29763 | return DAG.getConstant(Res, dl, MVT::i1); |
29764 | } |
29765 | |
29766 | // Check whether this extract is the root of a sum of absolute differences |
29767 | // pattern. This has to be done here because we really want it to happen |
29768 | // pre-legalization, |
29769 | if (SDValue SAD = combineBasicSADPattern(N, DAG, Subtarget)) |
29770 | return SAD; |
29771 | |
29772 | // Attempt to replace an all_of/any_of horizontal reduction with a MOVMSK. |
29773 | if (SDValue Cmp = combineHorizontalPredicateResult(N, DAG, Subtarget)) |
29774 | return Cmp; |
29775 | |
29776 | // Only operate on vectors of 4 elements, where the alternative shuffling |
29777 | // gets to be more expensive. |
29778 | if (SrcVT != MVT::v4i32) |
29779 | return SDValue(); |
29780 | |
29781 | // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a |
29782 | // single use which is a sign-extend or zero-extend, and all elements are |
29783 | // used. |
29784 | SmallVector<SDNode *, 4> Uses; |
29785 | unsigned ExtractedElements = 0; |
29786 | for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(), |
29787 | UE = InputVector.getNode()->use_end(); UI != UE; ++UI) { |
29788 | if (UI.getUse().getResNo() != InputVector.getResNo()) |
29789 | return SDValue(); |
29790 | |
29791 | SDNode *Extract = *UI; |
29792 | if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
29793 | return SDValue(); |
29794 | |
29795 | if (Extract->getValueType(0) != MVT::i32) |
29796 | return SDValue(); |
29797 | if (!Extract->hasOneUse()) |
29798 | return SDValue(); |
29799 | if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND && |
29800 | Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND) |
29801 | return SDValue(); |
29802 | if (!isa<ConstantSDNode>(Extract->getOperand(1))) |
29803 | return SDValue(); |
29804 | |
29805 | // Record which element was extracted. |
29806 | ExtractedElements |= 1 << Extract->getConstantOperandVal(1); |
29807 | Uses.push_back(Extract); |
29808 | } |
29809 | |
29810 | // If not all the elements were used, this may not be worthwhile. |
29811 | if (ExtractedElements != 15) |
29812 | return SDValue(); |
29813 | |
29814 | // Ok, we've now decided to do the transformation. |
29815 | // If 64-bit shifts are legal, use the extract-shift sequence, |
29816 | // otherwise bounce the vector off the cache. |
29817 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
29818 | SDValue Vals[4]; |
29819 | |
29820 | if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) { |
29821 | SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector); |
29822 | auto &DL = DAG.getDataLayout(); |
29823 | EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL); |
29824 | SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst, |
29825 | DAG.getConstant(0, dl, VecIdxTy)); |
29826 | SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst, |
29827 | DAG.getConstant(1, dl, VecIdxTy)); |
29828 | |
29829 | SDValue ShAmt = DAG.getConstant( |
29830 | 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL)); |
29831 | Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf); |
29832 | Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, |
29833 | DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt)); |
29834 | Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf); |
29835 | Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, |
29836 | DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt)); |
29837 | } else { |
29838 | // Store the value to a temporary stack slot. |
29839 | SDValue StackPtr = DAG.CreateStackTemporary(SrcVT); |
29840 | SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, |
29841 | MachinePointerInfo()); |
29842 | |
29843 | EVT ElementType = SrcVT.getVectorElementType(); |
29844 | unsigned EltSize = ElementType.getSizeInBits() / 8; |
29845 | |
29846 | // Replace each use (extract) with a load of the appropriate element. |
29847 | for (unsigned i = 0; i < 4; ++i) { |
29848 | uint64_t Offset = EltSize * i; |
29849 | auto PtrVT = TLI.getPointerTy(DAG.getDataLayout()); |
29850 | SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT); |
29851 | |
29852 | SDValue ScalarAddr = |
29853 | DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal); |
29854 | |
29855 | // Load the scalar. |
29856 | Vals[i] = |
29857 | DAG.getLoad(ElementType, dl, Ch, ScalarAddr, MachinePointerInfo()); |
29858 | } |
29859 | } |
29860 | |
29861 | // Replace the extracts |
29862 | for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(), |
29863 | UE = Uses.end(); UI != UE; ++UI) { |
29864 | SDNode *Extract = *UI; |
29865 | |
29866 | uint64_t IdxVal = Extract->getConstantOperandVal(1); |
29867 | DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]); |
29868 | } |
29869 | |
29870 | // The replacement was made in place; don't return anything. |
29871 | return SDValue(); |
29872 | } |
29873 | |
29874 | // TODO - merge with combineExtractVectorElt once it can handle the implicit |
29875 | // zero-extension of X86ISD::PINSRW/X86ISD::PINSRB in: |
29876 | // XFormVExtractWithShuffleIntoLoad, combineHorizontalPredicateResult and |
29877 | // combineBasicSADPattern. |
29878 | static SDValue combineExtractVectorElt_SSE(SDNode *N, SelectionDAG &DAG, |
29879 | TargetLowering::DAGCombinerInfo &DCI, |
29880 | const X86Subtarget &Subtarget) { |
29881 | return combineExtractWithShuffle(N, DAG, DCI, Subtarget); |
29882 | } |
29883 | |
29884 | /// If a vector select has an operand that is -1 or 0, try to simplify the |
29885 | /// select to a bitwise logic operation. |
29886 | static SDValue |
29887 | combineVSelectWithAllOnesOrZeros(SDNode *N, SelectionDAG &DAG, |
29888 | TargetLowering::DAGCombinerInfo &DCI, |
29889 | const X86Subtarget &Subtarget) { |
29890 | SDValue Cond = N->getOperand(0); |
29891 | SDValue LHS = N->getOperand(1); |
29892 | SDValue RHS = N->getOperand(2); |
29893 | EVT VT = LHS.getValueType(); |
29894 | EVT CondVT = Cond.getValueType(); |
29895 | SDLoc DL(N); |
29896 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
29897 | |
29898 | if (N->getOpcode() != ISD::VSELECT) |
29899 | return SDValue(); |
29900 | |
29901 | assert(CondVT.isVector() && "Vector select expects a vector selector!")((CondVT.isVector() && "Vector select expects a vector selector!" ) ? static_cast<void> (0) : __assert_fail ("CondVT.isVector() && \"Vector select expects a vector selector!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 29901, __PRETTY_FUNCTION__)); |
29902 | |
29903 | bool FValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode()); |
29904 | // Check if the first operand is all zeros and Cond type is vXi1. |
29905 | // This situation only applies to avx512. |
29906 | if (FValIsAllZeros && Subtarget.hasAVX512() && Cond.hasOneUse() && |
29907 | CondVT.getVectorElementType() == MVT::i1) { |
29908 | // Invert the cond to not(cond) : xor(op,allones)=not(op) |
29909 | SDValue CondNew = DAG.getNode(ISD::XOR, DL, CondVT, Cond, |
29910 | DAG.getAllOnesConstant(DL, CondVT)); |
29911 | // Vselect cond, op1, op2 = Vselect not(cond), op2, op1 |
29912 | return DAG.getSelect(DL, VT, CondNew, RHS, LHS); |
29913 | } |
29914 | |
29915 | // To use the condition operand as a bitwise mask, it must have elements that |
29916 | // are the same size as the select elements. Ie, the condition operand must |
29917 | // have already been promoted from the IR select condition type <N x i1>. |
29918 | // Don't check if the types themselves are equal because that excludes |
29919 | // vector floating-point selects. |
29920 | if (CondVT.getScalarSizeInBits() != VT.getScalarSizeInBits()) |
29921 | return SDValue(); |
29922 | |
29923 | bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode()); |
29924 | FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode()); |
29925 | |
29926 | // Try to invert the condition if true value is not all 1s and false value is |
29927 | // not all 0s. |
29928 | if (!TValIsAllOnes && !FValIsAllZeros && |
29929 | // Check if the selector will be produced by CMPP*/PCMP*. |
29930 | Cond.getOpcode() == ISD::SETCC && |
29931 | // Check if SETCC has already been promoted. |
29932 | TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) == |
29933 | CondVT) { |
29934 | bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode()); |
29935 | bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode()); |
29936 | |
29937 | if (TValIsAllZeros || FValIsAllOnes) { |
29938 | SDValue CC = Cond.getOperand(2); |
29939 | ISD::CondCode NewCC = |
29940 | ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(), |
29941 | Cond.getOperand(0).getValueType().isInteger()); |
29942 | Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), |
29943 | NewCC); |
29944 | std::swap(LHS, RHS); |
29945 | TValIsAllOnes = FValIsAllOnes; |
29946 | FValIsAllZeros = TValIsAllZeros; |
29947 | } |
29948 | } |
29949 | |
29950 | // vselect Cond, 111..., 000... -> Cond |
29951 | if (TValIsAllOnes && FValIsAllZeros) |
29952 | return DAG.getBitcast(VT, Cond); |
29953 | |
29954 | if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(CondVT)) |
29955 | return SDValue(); |
29956 | |
29957 | // vselect Cond, 111..., X -> or Cond, X |
29958 | if (TValIsAllOnes) { |
29959 | SDValue CastRHS = DAG.getBitcast(CondVT, RHS); |
29960 | SDValue Or = DAG.getNode(ISD::OR, DL, CondVT, Cond, CastRHS); |
29961 | return DAG.getBitcast(VT, Or); |
29962 | } |
29963 | |
29964 | // vselect Cond, X, 000... -> and Cond, X |
29965 | if (FValIsAllZeros) { |
29966 | SDValue CastLHS = DAG.getBitcast(CondVT, LHS); |
29967 | SDValue And = DAG.getNode(ISD::AND, DL, CondVT, Cond, CastLHS); |
29968 | return DAG.getBitcast(VT, And); |
29969 | } |
29970 | |
29971 | return SDValue(); |
29972 | } |
29973 | |
29974 | static SDValue combineSelectOfTwoConstants(SDNode *N, SelectionDAG &DAG) { |
29975 | SDValue Cond = N->getOperand(0); |
29976 | SDValue LHS = N->getOperand(1); |
29977 | SDValue RHS = N->getOperand(2); |
29978 | SDLoc DL(N); |
29979 | |
29980 | auto *TrueC = dyn_cast<ConstantSDNode>(LHS); |
29981 | auto *FalseC = dyn_cast<ConstantSDNode>(RHS); |
29982 | if (!TrueC || !FalseC) |
29983 | return SDValue(); |
29984 | |
29985 | // Don't do this for crazy integer types. |
29986 | if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) |
29987 | return SDValue(); |
29988 | |
29989 | // If this is efficiently invertible, canonicalize the LHSC/RHSC values |
29990 | // so that TrueC (the true value) is larger than FalseC. |
29991 | bool NeedsCondInvert = false; |
29992 | if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) && |
29993 | // Efficiently invertible. |
29994 | (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible. |
29995 | (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible. |
29996 | isa<ConstantSDNode>(Cond.getOperand(1))))) { |
29997 | NeedsCondInvert = true; |
29998 | std::swap(TrueC, FalseC); |
29999 | } |
30000 | |
30001 | // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0. |
30002 | if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) { |
30003 | if (NeedsCondInvert) // Invert the condition if needed. |
30004 | Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond, |
30005 | DAG.getConstant(1, DL, Cond.getValueType())); |
30006 | |
30007 | // Zero extend the condition if needed. |
30008 | Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond); |
30009 | |
30010 | unsigned ShAmt = TrueC->getAPIntValue().logBase2(); |
30011 | return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond, |
30012 | DAG.getConstant(ShAmt, DL, MVT::i8)); |
30013 | } |
30014 | |
30015 | // Optimize cases that will turn into an LEA instruction. This requires |
30016 | // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9). |
30017 | if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) { |
30018 | uint64_t Diff = TrueC->getZExtValue() - FalseC->getZExtValue(); |
30019 | if (N->getValueType(0) == MVT::i32) |
30020 | Diff = (unsigned)Diff; |
30021 | |
30022 | bool IsFastMultiplier = false; |
30023 | if (Diff < 10) { |
30024 | switch ((unsigned char)Diff) { |
30025 | default: |
30026 | break; |
30027 | case 1: // result = add base, cond |
30028 | case 2: // result = lea base( , cond*2) |
30029 | case 3: // result = lea base(cond, cond*2) |
30030 | case 4: // result = lea base( , cond*4) |
30031 | case 5: // result = lea base(cond, cond*4) |
30032 | case 8: // result = lea base( , cond*8) |
30033 | case 9: // result = lea base(cond, cond*8) |
30034 | IsFastMultiplier = true; |
30035 | break; |
30036 | } |
30037 | } |
30038 | |
30039 | if (IsFastMultiplier) { |
30040 | APInt Diff = TrueC->getAPIntValue() - FalseC->getAPIntValue(); |
30041 | if (NeedsCondInvert) // Invert the condition if needed. |
30042 | Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond, |
30043 | DAG.getConstant(1, DL, Cond.getValueType())); |
30044 | |
30045 | // Zero extend the condition if needed. |
30046 | Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0), Cond); |
30047 | // Scale the condition by the difference. |
30048 | if (Diff != 1) |
30049 | Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond, |
30050 | DAG.getConstant(Diff, DL, Cond.getValueType())); |
30051 | |
30052 | // Add the base if non-zero. |
30053 | if (FalseC->getAPIntValue() != 0) |
30054 | Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond, |
30055 | SDValue(FalseC, 0)); |
30056 | return Cond; |
30057 | } |
30058 | } |
30059 | |
30060 | return SDValue(); |
30061 | } |
30062 | |
30063 | // If this is a bitcasted op that can be represented as another type, push the |
30064 | // the bitcast to the inputs. This allows more opportunities for pattern |
30065 | // matching masked instructions. This is called when we know that the operation |
30066 | // is used as one of the inputs of a vselect. |
30067 | static bool combineBitcastForMaskedOp(SDValue OrigOp, SelectionDAG &DAG, |
30068 | TargetLowering::DAGCombinerInfo &DCI) { |
30069 | // Make sure we have a bitcast. |
30070 | if (OrigOp.getOpcode() != ISD::BITCAST) |
30071 | return false; |
30072 | |
30073 | SDValue Op = OrigOp.getOperand(0); |
30074 | |
30075 | // If the operation is used by anything other than the bitcast, we shouldn't |
30076 | // do this combine as that would replicate the operation. |
30077 | if (!Op.hasOneUse()) |
30078 | return false; |
30079 | |
30080 | MVT VT = OrigOp.getSimpleValueType(); |
30081 | MVT EltVT = VT.getVectorElementType(); |
30082 | SDLoc DL(Op.getNode()); |
30083 | |
30084 | auto BitcastAndCombineShuffle = [&](unsigned Opcode, SDValue Op0, SDValue Op1, |
30085 | SDValue Op2) { |
30086 | Op0 = DAG.getBitcast(VT, Op0); |
30087 | DCI.AddToWorklist(Op0.getNode()); |
30088 | Op1 = DAG.getBitcast(VT, Op1); |
30089 | DCI.AddToWorklist(Op1.getNode()); |
30090 | DCI.CombineTo(OrigOp.getNode(), |
30091 | DAG.getNode(Opcode, DL, VT, Op0, Op1, Op2)); |
30092 | return true; |
30093 | }; |
30094 | |
30095 | unsigned Opcode = Op.getOpcode(); |
30096 | switch (Opcode) { |
30097 | case X86ISD::PALIGNR: |
30098 | // PALIGNR can be converted to VALIGND/Q for 128-bit vectors. |
30099 | if (!VT.is128BitVector()) |
30100 | return false; |
30101 | Opcode = X86ISD::VALIGN; |
30102 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
30103 | case X86ISD::VALIGN: { |
30104 | if (EltVT != MVT::i32 && EltVT != MVT::i64) |
30105 | return false; |
30106 | uint64_t Imm = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue(); |
30107 | MVT OpEltVT = Op.getSimpleValueType().getVectorElementType(); |
30108 | unsigned ShiftAmt = Imm * OpEltVT.getSizeInBits(); |
30109 | unsigned EltSize = EltVT.getSizeInBits(); |
30110 | // Make sure we can represent the same shift with the new VT. |
30111 | if ((ShiftAmt % EltSize) != 0) |
30112 | return false; |
30113 | Imm = ShiftAmt / EltSize; |
30114 | return BitcastAndCombineShuffle(Opcode, Op.getOperand(0), Op.getOperand(1), |
30115 | DAG.getConstant(Imm, DL, MVT::i8)); |
30116 | } |
30117 | case X86ISD::SHUF128: { |
30118 | if (EltVT.getSizeInBits() != 32 && EltVT.getSizeInBits() != 64) |
30119 | return false; |
30120 | // Only change element size, not type. |
30121 | if (VT.isInteger() != Op.getSimpleValueType().isInteger()) |
30122 | return false; |
30123 | return BitcastAndCombineShuffle(Opcode, Op.getOperand(0), Op.getOperand(1), |
30124 | Op.getOperand(2)); |
30125 | } |
30126 | case ISD::INSERT_SUBVECTOR: { |
30127 | unsigned EltSize = EltVT.getSizeInBits(); |
30128 | if (EltSize != 32 && EltSize != 64) |
30129 | return false; |
30130 | MVT OpEltVT = Op.getSimpleValueType().getVectorElementType(); |
30131 | // Only change element size, not type. |
30132 | if (EltVT.isInteger() != OpEltVT.isInteger()) |
30133 | return false; |
30134 | uint64_t Imm = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue(); |
30135 | Imm = (Imm * OpEltVT.getSizeInBits()) / EltSize; |
30136 | SDValue Op0 = DAG.getBitcast(VT, Op.getOperand(0)); |
30137 | DCI.AddToWorklist(Op0.getNode()); |
30138 | // Op1 needs to be bitcasted to a smaller vector with the same element type. |
30139 | SDValue Op1 = Op.getOperand(1); |
30140 | MVT Op1VT = MVT::getVectorVT(EltVT, |
30141 | Op1.getSimpleValueType().getSizeInBits() / EltSize); |
30142 | Op1 = DAG.getBitcast(Op1VT, Op1); |
30143 | DCI.AddToWorklist(Op1.getNode()); |
30144 | DCI.CombineTo(OrigOp.getNode(), |
30145 | DAG.getNode(Opcode, DL, VT, Op0, Op1, |
30146 | DAG.getIntPtrConstant(Imm, DL))); |
30147 | return true; |
30148 | } |
30149 | case ISD::EXTRACT_SUBVECTOR: { |
30150 | unsigned EltSize = EltVT.getSizeInBits(); |
30151 | if (EltSize != 32 && EltSize != 64) |
30152 | return false; |
30153 | MVT OpEltVT = Op.getSimpleValueType().getVectorElementType(); |
30154 | // Only change element size, not type. |
30155 | if (EltVT.isInteger() != OpEltVT.isInteger()) |
30156 | return false; |
30157 | uint64_t Imm = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); |
30158 | Imm = (Imm * OpEltVT.getSizeInBits()) / EltSize; |
30159 | // Op0 needs to be bitcasted to a larger vector with the same element type. |
30160 | SDValue Op0 = Op.getOperand(0); |
30161 | MVT Op0VT = MVT::getVectorVT(EltVT, |
30162 | Op0.getSimpleValueType().getSizeInBits() / EltSize); |
30163 | Op0 = DAG.getBitcast(Op0VT, Op0); |
30164 | DCI.AddToWorklist(Op0.getNode()); |
30165 | DCI.CombineTo(OrigOp.getNode(), |
30166 | DAG.getNode(Opcode, DL, VT, Op0, |
30167 | DAG.getIntPtrConstant(Imm, DL))); |
30168 | return true; |
30169 | } |
30170 | case X86ISD::SUBV_BROADCAST: { |
30171 | unsigned EltSize = EltVT.getSizeInBits(); |
30172 | if (EltSize != 32 && EltSize != 64) |
30173 | return false; |
30174 | // Only change element size, not type. |
30175 | if (VT.isInteger() != Op.getSimpleValueType().isInteger()) |
30176 | return false; |
30177 | SDValue Op0 = Op.getOperand(0); |
30178 | MVT Op0VT = MVT::getVectorVT(EltVT, |
30179 | Op0.getSimpleValueType().getSizeInBits() / EltSize); |
30180 | Op0 = DAG.getBitcast(Op0VT, Op.getOperand(0)); |
30181 | DCI.AddToWorklist(Op0.getNode()); |
30182 | DCI.CombineTo(OrigOp.getNode(), |
30183 | DAG.getNode(Opcode, DL, VT, Op0)); |
30184 | return true; |
30185 | } |
30186 | } |
30187 | |
30188 | return false; |
30189 | } |
30190 | |
30191 | /// Do target-specific dag combines on SELECT and VSELECT nodes. |
30192 | static SDValue combineSelect(SDNode *N, SelectionDAG &DAG, |
30193 | TargetLowering::DAGCombinerInfo &DCI, |
30194 | const X86Subtarget &Subtarget) { |
30195 | SDLoc DL(N); |
30196 | SDValue Cond = N->getOperand(0); |
30197 | // Get the LHS/RHS of the select. |
30198 | SDValue LHS = N->getOperand(1); |
30199 | SDValue RHS = N->getOperand(2); |
30200 | EVT VT = LHS.getValueType(); |
30201 | EVT CondVT = Cond.getValueType(); |
30202 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
30203 | |
30204 | // If we have SSE[12] support, try to form min/max nodes. SSE min/max |
30205 | // instructions match the semantics of the common C idiom x<y?x:y but not |
30206 | // x<=y?x:y, because of how they handle negative zero (which can be |
30207 | // ignored in unsafe-math mode). |
30208 | // We also try to create v2f32 min/max nodes, which we later widen to v4f32. |
30209 | if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() && |
30210 | VT != MVT::f80 && VT != MVT::f128 && |
30211 | (TLI.isTypeLegal(VT) || VT == MVT::v2f32) && |
30212 | (Subtarget.hasSSE2() || |
30213 | (Subtarget.hasSSE1() && VT.getScalarType() == MVT::f32))) { |
30214 | ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get(); |
30215 | |
30216 | unsigned Opcode = 0; |
30217 | // Check for x CC y ? x : y. |
30218 | if (DAG.isEqualTo(LHS, Cond.getOperand(0)) && |
30219 | DAG.isEqualTo(RHS, Cond.getOperand(1))) { |
30220 | switch (CC) { |
30221 | default: break; |
30222 | case ISD::SETULT: |
30223 | // Converting this to a min would handle NaNs incorrectly, and swapping |
30224 | // the operands would cause it to handle comparisons between positive |
30225 | // and negative zero incorrectly. |
30226 | if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) { |
30227 | if (!DAG.getTarget().Options.UnsafeFPMath && |
30228 | !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) |
30229 | break; |
30230 | std::swap(LHS, RHS); |
30231 | } |
30232 | Opcode = X86ISD::FMIN; |
30233 | break; |
30234 | case ISD::SETOLE: |
30235 | // Converting this to a min would handle comparisons between positive |
30236 | // and negative zero incorrectly. |
30237 | if (!DAG.getTarget().Options.UnsafeFPMath && |
30238 | !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) |
30239 | break; |
30240 | Opcode = X86ISD::FMIN; |
30241 | break; |
30242 | case ISD::SETULE: |
30243 | // Converting this to a min would handle both negative zeros and NaNs |
30244 | // incorrectly, but we can swap the operands to fix both. |
30245 | std::swap(LHS, RHS); |
30246 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
30247 | case ISD::SETOLT: |
30248 | case ISD::SETLT: |
30249 | case ISD::SETLE: |
30250 | Opcode = X86ISD::FMIN; |
30251 | break; |
30252 | |
30253 | case ISD::SETOGE: |
30254 | // Converting this to a max would handle comparisons between positive |
30255 | // and negative zero incorrectly. |
30256 | if (!DAG.getTarget().Options.UnsafeFPMath && |
30257 | !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) |
30258 | break; |
30259 | Opcode = X86ISD::FMAX; |
30260 | break; |
30261 | case ISD::SETUGT: |
30262 | // Converting this to a max would handle NaNs incorrectly, and swapping |
30263 | // the operands would cause it to handle comparisons between positive |
30264 | // and negative zero incorrectly. |
30265 | if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) { |
30266 | if (!DAG.getTarget().Options.UnsafeFPMath && |
30267 | !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) |
30268 | break; |
30269 | std::swap(LHS, RHS); |
30270 | } |
30271 | Opcode = X86ISD::FMAX; |
30272 | break; |
30273 | case ISD::SETUGE: |
30274 | // Converting this to a max would handle both negative zeros and NaNs |
30275 | // incorrectly, but we can swap the operands to fix both. |
30276 | std::swap(LHS, RHS); |
30277 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
30278 | case ISD::SETOGT: |
30279 | case ISD::SETGT: |
30280 | case ISD::SETGE: |
30281 | Opcode = X86ISD::FMAX; |
30282 | break; |
30283 | } |
30284 | // Check for x CC y ? y : x -- a min/max with reversed arms. |
30285 | } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) && |
30286 | DAG.isEqualTo(RHS, Cond.getOperand(0))) { |
30287 | switch (CC) { |
30288 | default: break; |
30289 | case ISD::SETOGE: |
30290 | // Converting this to a min would handle comparisons between positive |
30291 | // and negative zero incorrectly, and swapping the operands would |
30292 | // cause it to handle NaNs incorrectly. |
30293 | if (!DAG.getTarget().Options.UnsafeFPMath && |
30294 | !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) { |
30295 | if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) |
30296 | break; |
30297 | std::swap(LHS, RHS); |
30298 | } |
30299 | Opcode = X86ISD::FMIN; |
30300 | break; |
30301 | case ISD::SETUGT: |
30302 | // Converting this to a min would handle NaNs incorrectly. |
30303 | if (!DAG.getTarget().Options.UnsafeFPMath && |
30304 | (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) |
30305 | break; |
30306 | Opcode = X86ISD::FMIN; |
30307 | break; |
30308 | case ISD::SETUGE: |
30309 | // Converting this to a min would handle both negative zeros and NaNs |
30310 | // incorrectly, but we can swap the operands to fix both. |
30311 | std::swap(LHS, RHS); |
30312 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
30313 | case ISD::SETOGT: |
30314 | case ISD::SETGT: |
30315 | case ISD::SETGE: |
30316 | Opcode = X86ISD::FMIN; |
30317 | break; |
30318 | |
30319 | case ISD::SETULT: |
30320 | // Converting this to a max would handle NaNs incorrectly. |
30321 | if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) |
30322 | break; |
30323 | Opcode = X86ISD::FMAX; |
30324 | break; |
30325 | case ISD::SETOLE: |
30326 | // Converting this to a max would handle comparisons between positive |
30327 | // and negative zero incorrectly, and swapping the operands would |
30328 | // cause it to handle NaNs incorrectly. |
30329 | if (!DAG.getTarget().Options.UnsafeFPMath && |
30330 | !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) { |
30331 | if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) |
30332 | break; |
30333 | std::swap(LHS, RHS); |
30334 | } |
30335 | Opcode = X86ISD::FMAX; |
30336 | break; |
30337 | case ISD::SETULE: |
30338 | // Converting this to a max would handle both negative zeros and NaNs |
30339 | // incorrectly, but we can swap the operands to fix both. |
30340 | std::swap(LHS, RHS); |
30341 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
30342 | case ISD::SETOLT: |
30343 | case ISD::SETLT: |
30344 | case ISD::SETLE: |
30345 | Opcode = X86ISD::FMAX; |
30346 | break; |
30347 | } |
30348 | } |
30349 | |
30350 | if (Opcode) |
30351 | return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS); |
30352 | } |
30353 | |
30354 | // v16i8 (select v16i1, v16i8, v16i8) does not have a proper |
30355 | // lowering on KNL. In this case we convert it to |
30356 | // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction. |
30357 | // The same situation for all 128 and 256-bit vectors of i8 and i16. |
30358 | // Since SKX these selects have a proper lowering. |
30359 | if (Subtarget.hasAVX512() && CondVT.isVector() && |
30360 | CondVT.getVectorElementType() == MVT::i1 && |
30361 | (VT.is128BitVector() || VT.is256BitVector()) && |
30362 | (VT.getVectorElementType() == MVT::i8 || |
30363 | VT.getVectorElementType() == MVT::i16) && |
30364 | !(Subtarget.hasBWI() && Subtarget.hasVLX())) { |
30365 | Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, VT, Cond); |
30366 | DCI.AddToWorklist(Cond.getNode()); |
30367 | return DAG.getNode(N->getOpcode(), DL, VT, Cond, LHS, RHS); |
30368 | } |
30369 | |
30370 | if (SDValue V = combineSelectOfTwoConstants(N, DAG)) |
30371 | return V; |
30372 | |
30373 | // Canonicalize max and min: |
30374 | // (x > y) ? x : y -> (x >= y) ? x : y |
30375 | // (x < y) ? x : y -> (x <= y) ? x : y |
30376 | // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates |
30377 | // the need for an extra compare |
30378 | // against zero. e.g. |
30379 | // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0 |
30380 | // subl %esi, %edi |
30381 | // testl %edi, %edi |
30382 | // movl $0, %eax |
30383 | // cmovgl %edi, %eax |
30384 | // => |
30385 | // xorl %eax, %eax |
30386 | // subl %esi, $edi |
30387 | // cmovsl %eax, %edi |
30388 | if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC && |
30389 | DAG.isEqualTo(LHS, Cond.getOperand(0)) && |
30390 | DAG.isEqualTo(RHS, Cond.getOperand(1))) { |
30391 | ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get(); |
30392 | switch (CC) { |
30393 | default: break; |
30394 | case ISD::SETLT: |
30395 | case ISD::SETGT: { |
30396 | ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE; |
30397 | Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(), |
30398 | Cond.getOperand(0), Cond.getOperand(1), NewCC); |
30399 | return DAG.getSelect(DL, VT, Cond, LHS, RHS); |
30400 | } |
30401 | } |
30402 | } |
30403 | |
30404 | // Early exit check |
30405 | if (!TLI.isTypeLegal(VT)) |
30406 | return SDValue(); |
30407 | |
30408 | // Match VSELECTs into subs with unsigned saturation. |
30409 | if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC && |
30410 | // psubus is available in SSE2 and AVX2 for i8 and i16 vectors. |
30411 | ((Subtarget.hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) || |
30412 | (Subtarget.hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) { |
30413 | ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get(); |
30414 | |
30415 | // Check if one of the arms of the VSELECT is a zero vector. If it's on the |
30416 | // left side invert the predicate to simplify logic below. |
30417 | SDValue Other; |
30418 | if (ISD::isBuildVectorAllZeros(LHS.getNode())) { |
30419 | Other = RHS; |
30420 | CC = ISD::getSetCCInverse(CC, true); |
30421 | } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) { |
30422 | Other = LHS; |
30423 | } |
30424 | |
30425 | if (Other.getNode() && Other->getNumOperands() == 2 && |
30426 | DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) { |
30427 | SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1); |
30428 | SDValue CondRHS = Cond->getOperand(1); |
30429 | |
30430 | // Look for a general sub with unsigned saturation first. |
30431 | // x >= y ? x-y : 0 --> subus x, y |
30432 | // x > y ? x-y : 0 --> subus x, y |
30433 | if ((CC == ISD::SETUGE || CC == ISD::SETUGT) && |
30434 | Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS)) |
30435 | return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS); |
30436 | |
30437 | if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS)) |
30438 | if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) { |
30439 | if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS)) |
30440 | if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode()) |
30441 | // If the RHS is a constant we have to reverse the const |
30442 | // canonicalization. |
30443 | // x > C-1 ? x+-C : 0 --> subus x, C |
30444 | if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD && |
30445 | CondRHSConst->getAPIntValue() == |
30446 | (-OpRHSConst->getAPIntValue() - 1)) |
30447 | return DAG.getNode( |
30448 | X86ISD::SUBUS, DL, VT, OpLHS, |
30449 | DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT)); |
30450 | |
30451 | // Another special case: If C was a sign bit, the sub has been |
30452 | // canonicalized into a xor. |
30453 | // FIXME: Would it be better to use computeKnownBits to determine |
30454 | // whether it's safe to decanonicalize the xor? |
30455 | // x s< 0 ? x^C : 0 --> subus x, C |
30456 | if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR && |
30457 | ISD::isBuildVectorAllZeros(CondRHS.getNode()) && |
30458 | OpRHSConst->getAPIntValue().isSignMask()) |
30459 | // Note that we have to rebuild the RHS constant here to ensure we |
30460 | // don't rely on particular values of undef lanes. |
30461 | return DAG.getNode( |
30462 | X86ISD::SUBUS, DL, VT, OpLHS, |
30463 | DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT)); |
30464 | } |
30465 | } |
30466 | } |
30467 | |
30468 | if (SDValue V = combineVSelectWithAllOnesOrZeros(N, DAG, DCI, Subtarget)) |
30469 | return V; |
30470 | |
30471 | // If this is a *dynamic* select (non-constant condition) and we can match |
30472 | // this node with one of the variable blend instructions, restructure the |
30473 | // condition so that blends can use the high (sign) bit of each element and |
30474 | // use SimplifyDemandedBits to simplify the condition operand. |
30475 | if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() && |
30476 | !DCI.isBeforeLegalize() && |
30477 | !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) { |
30478 | unsigned BitWidth = Cond.getScalarValueSizeInBits(); |
30479 | |
30480 | // Don't optimize vector selects that map to mask-registers. |
30481 | if (BitWidth == 1) |
30482 | return SDValue(); |
30483 | |
30484 | // We can only handle the cases where VSELECT is directly legal on the |
30485 | // subtarget. We custom lower VSELECT nodes with constant conditions and |
30486 | // this makes it hard to see whether a dynamic VSELECT will correctly |
30487 | // lower, so we both check the operation's status and explicitly handle the |
30488 | // cases where a *dynamic* blend will fail even though a constant-condition |
30489 | // blend could be custom lowered. |
30490 | // FIXME: We should find a better way to handle this class of problems. |
30491 | // Potentially, we should combine constant-condition vselect nodes |
30492 | // pre-legalization into shuffles and not mark as many types as custom |
30493 | // lowered. |
30494 | if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT)) |
30495 | return SDValue(); |
30496 | // FIXME: We don't support i16-element blends currently. We could and |
30497 | // should support them by making *all* the bits in the condition be set |
30498 | // rather than just the high bit and using an i8-element blend. |
30499 | if (VT.getVectorElementType() == MVT::i16) |
30500 | return SDValue(); |
30501 | // Dynamic blending was only available from SSE4.1 onward. |
30502 | if (VT.is128BitVector() && !Subtarget.hasSSE41()) |
30503 | return SDValue(); |
30504 | // Byte blends are only available in AVX2 |
30505 | if (VT == MVT::v32i8 && !Subtarget.hasAVX2()) |
30506 | return SDValue(); |
30507 | |
30508 | assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size")((BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size" ) ? static_cast<void> (0) : __assert_fail ("BitWidth >= 8 && BitWidth <= 64 && \"Invalid mask size\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 30508, __PRETTY_FUNCTION__)); |
30509 | APInt DemandedMask(APInt::getSignMask(BitWidth)); |
30510 | KnownBits Known; |
30511 | TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(), |
30512 | DCI.isBeforeLegalizeOps()); |
30513 | if (TLI.ShrinkDemandedConstant(Cond, DemandedMask, TLO) || |
30514 | TLI.SimplifyDemandedBits(Cond, DemandedMask, Known, TLO)) { |
30515 | // If we changed the computation somewhere in the DAG, this change will |
30516 | // affect all users of Cond. Make sure it is fine and update all the nodes |
30517 | // so that we do not use the generic VSELECT anymore. Otherwise, we may |
30518 | // perform wrong optimizations as we messed with the actual expectation |
30519 | // for the vector boolean values. |
30520 | if (Cond != TLO.Old) { |
30521 | // Check all uses of the condition operand to check whether it will be |
30522 | // consumed by non-BLEND instructions. Those may require that all bits |
30523 | // are set properly. |
30524 | for (SDNode *U : Cond->uses()) { |
30525 | // TODO: Add other opcodes eventually lowered into BLEND. |
30526 | if (U->getOpcode() != ISD::VSELECT) |
30527 | return SDValue(); |
30528 | } |
30529 | |
30530 | // Update all users of the condition before committing the change, so |
30531 | // that the VSELECT optimizations that expect the correct vector boolean |
30532 | // value will not be triggered. |
30533 | for (SDNode *U : Cond->uses()) { |
30534 | SDValue SB = DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(U), |
30535 | U->getValueType(0), Cond, U->getOperand(1), |
30536 | U->getOperand(2)); |
30537 | DAG.ReplaceAllUsesOfValueWith(SDValue(U, 0), SB); |
30538 | } |
30539 | DCI.CommitTargetLoweringOpt(TLO); |
30540 | return SDValue(); |
30541 | } |
30542 | // Only Cond (rather than other nodes in the computation chain) was |
30543 | // changed. Change the condition just for N to keep the opportunity to |
30544 | // optimize all other users their own way. |
30545 | SDValue SB = DAG.getNode(X86ISD::SHRUNKBLEND, DL, VT, TLO.New, LHS, RHS); |
30546 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), SB); |
30547 | return SDValue(); |
30548 | } |
30549 | } |
30550 | |
30551 | // Look for vselects with LHS/RHS being bitcasted from an operation that |
30552 | // can be executed on another type. Push the bitcast to the inputs of |
30553 | // the operation. This exposes opportunities for using masking instructions. |
30554 | if (N->getOpcode() == ISD::VSELECT && DCI.isAfterLegalizeVectorOps() && |
30555 | CondVT.getVectorElementType() == MVT::i1) { |
30556 | if (combineBitcastForMaskedOp(LHS, DAG, DCI)) |
30557 | return SDValue(N, 0); |
30558 | if (combineBitcastForMaskedOp(RHS, DAG, DCI)) |
30559 | return SDValue(N, 0); |
30560 | } |
30561 | |
30562 | return SDValue(); |
30563 | } |
30564 | |
30565 | /// Combine: |
30566 | /// (brcond/cmov/setcc .., (cmp (atomic_load_add x, 1), 0), COND_S) |
30567 | /// to: |
30568 | /// (brcond/cmov/setcc .., (LADD x, 1), COND_LE) |
30569 | /// i.e., reusing the EFLAGS produced by the LOCKed instruction. |
30570 | /// Note that this is only legal for some op/cc combinations. |
30571 | static SDValue combineSetCCAtomicArith(SDValue Cmp, X86::CondCode &CC, |
30572 | SelectionDAG &DAG) { |
30573 | // This combine only operates on CMP-like nodes. |
30574 | if (!(Cmp.getOpcode() == X86ISD::CMP || |
30575 | (Cmp.getOpcode() == X86ISD::SUB && !Cmp->hasAnyUseOfValue(0)))) |
30576 | return SDValue(); |
30577 | |
30578 | // Can't replace the cmp if it has more uses than the one we're looking at. |
30579 | // FIXME: We would like to be able to handle this, but would need to make sure |
30580 | // all uses were updated. |
30581 | if (!Cmp.hasOneUse()) |
30582 | return SDValue(); |
30583 | |
30584 | // This only applies to variations of the common case: |
30585 | // (icmp slt x, 0) -> (icmp sle (add x, 1), 0) |
30586 | // (icmp sge x, 0) -> (icmp sgt (add x, 1), 0) |
30587 | // (icmp sle x, 0) -> (icmp slt (sub x, 1), 0) |
30588 | // (icmp sgt x, 0) -> (icmp sge (sub x, 1), 0) |
30589 | // Using the proper condcodes (see below), overflow is checked for. |
30590 | |
30591 | // FIXME: We can generalize both constraints: |
30592 | // - XOR/OR/AND (if they were made to survive AtomicExpand) |
30593 | // - LHS != 1 |
30594 | // if the result is compared. |
30595 | |
30596 | SDValue CmpLHS = Cmp.getOperand(0); |
30597 | SDValue CmpRHS = Cmp.getOperand(1); |
30598 | |
30599 | if (!CmpLHS.hasOneUse()) |
30600 | return SDValue(); |
30601 | |
30602 | auto *CmpRHSC = dyn_cast<ConstantSDNode>(CmpRHS); |
30603 | if (!CmpRHSC || CmpRHSC->getZExtValue() != 0) |
30604 | return SDValue(); |
30605 | |
30606 | const unsigned Opc = CmpLHS.getOpcode(); |
30607 | |
30608 | if (Opc != ISD::ATOMIC_LOAD_ADD && Opc != ISD::ATOMIC_LOAD_SUB) |
30609 | return SDValue(); |
30610 | |
30611 | SDValue OpRHS = CmpLHS.getOperand(2); |
30612 | auto *OpRHSC = dyn_cast<ConstantSDNode>(OpRHS); |
30613 | if (!OpRHSC) |
30614 | return SDValue(); |
30615 | |
30616 | APInt Addend = OpRHSC->getAPIntValue(); |
30617 | if (Opc == ISD::ATOMIC_LOAD_SUB) |
30618 | Addend = -Addend; |
30619 | |
30620 | if (CC == X86::COND_S && Addend == 1) |
30621 | CC = X86::COND_LE; |
30622 | else if (CC == X86::COND_NS && Addend == 1) |
30623 | CC = X86::COND_G; |
30624 | else if (CC == X86::COND_G && Addend == -1) |
30625 | CC = X86::COND_GE; |
30626 | else if (CC == X86::COND_LE && Addend == -1) |
30627 | CC = X86::COND_L; |
30628 | else |
30629 | return SDValue(); |
30630 | |
30631 | SDValue LockOp = lowerAtomicArithWithLOCK(CmpLHS, DAG); |
30632 | DAG.ReplaceAllUsesOfValueWith(CmpLHS.getValue(0), |
30633 | DAG.getUNDEF(CmpLHS.getValueType())); |
30634 | DAG.ReplaceAllUsesOfValueWith(CmpLHS.getValue(1), LockOp.getValue(1)); |
30635 | return LockOp; |
30636 | } |
30637 | |
30638 | // Check whether a boolean test is testing a boolean value generated by |
30639 | // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition |
30640 | // code. |
30641 | // |
30642 | // Simplify the following patterns: |
30643 | // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or |
30644 | // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ) |
30645 | // to (Op EFLAGS Cond) |
30646 | // |
30647 | // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or |
30648 | // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ) |
30649 | // to (Op EFLAGS !Cond) |
30650 | // |
30651 | // where Op could be BRCOND or CMOV. |
30652 | // |
30653 | static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) { |
30654 | // This combine only operates on CMP-like nodes. |
30655 | if (!(Cmp.getOpcode() == X86ISD::CMP || |
30656 | (Cmp.getOpcode() == X86ISD::SUB && !Cmp->hasAnyUseOfValue(0)))) |
30657 | return SDValue(); |
30658 | |
30659 | // Quit if not used as a boolean value. |
30660 | if (CC != X86::COND_E && CC != X86::COND_NE) |
30661 | return SDValue(); |
30662 | |
30663 | // Check CMP operands. One of them should be 0 or 1 and the other should be |
30664 | // an SetCC or extended from it. |
30665 | SDValue Op1 = Cmp.getOperand(0); |
30666 | SDValue Op2 = Cmp.getOperand(1); |
30667 | |
30668 | SDValue SetCC; |
30669 | const ConstantSDNode* C = nullptr; |
30670 | bool needOppositeCond = (CC == X86::COND_E); |
30671 | bool checkAgainstTrue = false; // Is it a comparison against 1? |
30672 | |
30673 | if ((C = dyn_cast<ConstantSDNode>(Op1))) |
30674 | SetCC = Op2; |
30675 | else if ((C = dyn_cast<ConstantSDNode>(Op2))) |
30676 | SetCC = Op1; |
30677 | else // Quit if all operands are not constants. |
30678 | return SDValue(); |
30679 | |
30680 | if (C->getZExtValue() == 1) { |
30681 | needOppositeCond = !needOppositeCond; |
30682 | checkAgainstTrue = true; |
30683 | } else if (C->getZExtValue() != 0) |
30684 | // Quit if the constant is neither 0 or 1. |
30685 | return SDValue(); |
30686 | |
30687 | bool truncatedToBoolWithAnd = false; |
30688 | // Skip (zext $x), (trunc $x), or (and $x, 1) node. |
30689 | while (SetCC.getOpcode() == ISD::ZERO_EXTEND || |
30690 | SetCC.getOpcode() == ISD::TRUNCATE || |
30691 | SetCC.getOpcode() == ISD::AND) { |
30692 | if (SetCC.getOpcode() == ISD::AND) { |
30693 | int OpIdx = -1; |
30694 | if (isOneConstant(SetCC.getOperand(0))) |
30695 | OpIdx = 1; |
30696 | if (isOneConstant(SetCC.getOperand(1))) |
30697 | OpIdx = 0; |
30698 | if (OpIdx < 0) |
30699 | break; |
30700 | SetCC = SetCC.getOperand(OpIdx); |
30701 | truncatedToBoolWithAnd = true; |
30702 | } else |
30703 | SetCC = SetCC.getOperand(0); |
30704 | } |
30705 | |
30706 | switch (SetCC.getOpcode()) { |
30707 | case X86ISD::SETCC_CARRY: |
30708 | // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to |
30709 | // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1, |
30710 | // i.e. it's a comparison against true but the result of SETCC_CARRY is not |
30711 | // truncated to i1 using 'and'. |
30712 | if (checkAgainstTrue && !truncatedToBoolWithAnd) |
30713 | break; |
30714 | assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&((X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B && "Invalid use of SETCC_CARRY!") ? static_cast<void > (0) : __assert_fail ("X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B && \"Invalid use of SETCC_CARRY!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 30715, __PRETTY_FUNCTION__)) |
30715 | "Invalid use of SETCC_CARRY!")((X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B && "Invalid use of SETCC_CARRY!") ? static_cast<void > (0) : __assert_fail ("X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B && \"Invalid use of SETCC_CARRY!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 30715, __PRETTY_FUNCTION__)); |
30716 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
30717 | case X86ISD::SETCC: |
30718 | // Set the condition code or opposite one if necessary. |
30719 | CC = X86::CondCode(SetCC.getConstantOperandVal(0)); |
30720 | if (needOppositeCond) |
30721 | CC = X86::GetOppositeBranchCondition(CC); |
30722 | return SetCC.getOperand(1); |
30723 | case X86ISD::CMOV: { |
30724 | // Check whether false/true value has canonical one, i.e. 0 or 1. |
30725 | ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0)); |
30726 | ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1)); |
30727 | // Quit if true value is not a constant. |
30728 | if (!TVal) |
30729 | return SDValue(); |
30730 | // Quit if false value is not a constant. |
30731 | if (!FVal) { |
30732 | SDValue Op = SetCC.getOperand(0); |
30733 | // Skip 'zext' or 'trunc' node. |
30734 | if (Op.getOpcode() == ISD::ZERO_EXTEND || |
30735 | Op.getOpcode() == ISD::TRUNCATE) |
30736 | Op = Op.getOperand(0); |
30737 | // A special case for rdrand/rdseed, where 0 is set if false cond is |
30738 | // found. |
30739 | if ((Op.getOpcode() != X86ISD::RDRAND && |
30740 | Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0) |
30741 | return SDValue(); |
30742 | } |
30743 | // Quit if false value is not the constant 0 or 1. |
30744 | bool FValIsFalse = true; |
30745 | if (FVal && FVal->getZExtValue() != 0) { |
30746 | if (FVal->getZExtValue() != 1) |
30747 | return SDValue(); |
30748 | // If FVal is 1, opposite cond is needed. |
30749 | needOppositeCond = !needOppositeCond; |
30750 | FValIsFalse = false; |
30751 | } |
30752 | // Quit if TVal is not the constant opposite of FVal. |
30753 | if (FValIsFalse && TVal->getZExtValue() != 1) |
30754 | return SDValue(); |
30755 | if (!FValIsFalse && TVal->getZExtValue() != 0) |
30756 | return SDValue(); |
30757 | CC = X86::CondCode(SetCC.getConstantOperandVal(2)); |
30758 | if (needOppositeCond) |
30759 | CC = X86::GetOppositeBranchCondition(CC); |
30760 | return SetCC.getOperand(3); |
30761 | } |
30762 | } |
30763 | |
30764 | return SDValue(); |
30765 | } |
30766 | |
30767 | /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS. |
30768 | /// Match: |
30769 | /// (X86or (X86setcc) (X86setcc)) |
30770 | /// (X86cmp (and (X86setcc) (X86setcc)), 0) |
30771 | static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0, |
30772 | X86::CondCode &CC1, SDValue &Flags, |
30773 | bool &isAnd) { |
30774 | if (Cond->getOpcode() == X86ISD::CMP) { |
30775 | if (!isNullConstant(Cond->getOperand(1))) |
30776 | return false; |
30777 | |
30778 | Cond = Cond->getOperand(0); |
30779 | } |
30780 | |
30781 | isAnd = false; |
30782 | |
30783 | SDValue SetCC0, SetCC1; |
30784 | switch (Cond->getOpcode()) { |
30785 | default: return false; |
30786 | case ISD::AND: |
30787 | case X86ISD::AND: |
30788 | isAnd = true; |
30789 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
30790 | case ISD::OR: |
30791 | case X86ISD::OR: |
30792 | SetCC0 = Cond->getOperand(0); |
30793 | SetCC1 = Cond->getOperand(1); |
30794 | break; |
30795 | }; |
30796 | |
30797 | // Make sure we have SETCC nodes, using the same flags value. |
30798 | if (SetCC0.getOpcode() != X86ISD::SETCC || |
30799 | SetCC1.getOpcode() != X86ISD::SETCC || |
30800 | SetCC0->getOperand(1) != SetCC1->getOperand(1)) |
30801 | return false; |
30802 | |
30803 | CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0); |
30804 | CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0); |
30805 | Flags = SetCC0->getOperand(1); |
30806 | return true; |
30807 | } |
30808 | |
30809 | /// Optimize an EFLAGS definition used according to the condition code \p CC |
30810 | /// into a simpler EFLAGS value, potentially returning a new \p CC and replacing |
30811 | /// uses of chain values. |
30812 | static SDValue combineSetCCEFLAGS(SDValue EFLAGS, X86::CondCode &CC, |
30813 | SelectionDAG &DAG) { |
30814 | if (SDValue R = checkBoolTestSetCCCombine(EFLAGS, CC)) |
30815 | return R; |
30816 | return combineSetCCAtomicArith(EFLAGS, CC, DAG); |
30817 | } |
30818 | |
30819 | /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL] |
30820 | static SDValue combineCMov(SDNode *N, SelectionDAG &DAG, |
30821 | TargetLowering::DAGCombinerInfo &DCI, |
30822 | const X86Subtarget &Subtarget) { |
30823 | SDLoc DL(N); |
30824 | |
30825 | // If the flag operand isn't dead, don't touch this CMOV. |
30826 | if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty()) |
30827 | return SDValue(); |
30828 | |
30829 | SDValue FalseOp = N->getOperand(0); |
30830 | SDValue TrueOp = N->getOperand(1); |
30831 | X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2); |
30832 | SDValue Cond = N->getOperand(3); |
30833 | |
30834 | if (CC == X86::COND_E || CC == X86::COND_NE) { |
30835 | switch (Cond.getOpcode()) { |
30836 | default: break; |
30837 | case X86ISD::BSR: |
30838 | case X86ISD::BSF: |
30839 | // If operand of BSR / BSF are proven never zero, then ZF cannot be set. |
30840 | if (DAG.isKnownNeverZero(Cond.getOperand(0))) |
30841 | return (CC == X86::COND_E) ? FalseOp : TrueOp; |
30842 | } |
30843 | } |
30844 | |
30845 | // Try to simplify the EFLAGS and condition code operands. |
30846 | // We can't always do this as FCMOV only supports a subset of X86 cond. |
30847 | if (SDValue Flags = combineSetCCEFLAGS(Cond, CC, DAG)) { |
30848 | if (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC)) { |
30849 | SDValue Ops[] = {FalseOp, TrueOp, DAG.getConstant(CC, DL, MVT::i8), |
30850 | Flags}; |
30851 | return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops); |
30852 | } |
30853 | } |
30854 | |
30855 | // If this is a select between two integer constants, try to do some |
30856 | // optimizations. Note that the operands are ordered the opposite of SELECT |
30857 | // operands. |
30858 | if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) { |
30859 | if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) { |
30860 | // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is |
30861 | // larger than FalseC (the false value). |
30862 | if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) { |
30863 | CC = X86::GetOppositeBranchCondition(CC); |
30864 | std::swap(TrueC, FalseC); |
30865 | std::swap(TrueOp, FalseOp); |
30866 | } |
30867 | |
30868 | // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0. |
30869 | // This is efficient for any integer data type (including i8/i16) and |
30870 | // shift amount. |
30871 | if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) { |
30872 | Cond = getSETCC(CC, Cond, DL, DAG); |
30873 | |
30874 | // Zero extend the condition if needed. |
30875 | Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond); |
30876 | |
30877 | unsigned ShAmt = TrueC->getAPIntValue().logBase2(); |
30878 | Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond, |
30879 | DAG.getConstant(ShAmt, DL, MVT::i8)); |
30880 | if (N->getNumValues() == 2) // Dead flag value? |
30881 | return DCI.CombineTo(N, Cond, SDValue()); |
30882 | return Cond; |
30883 | } |
30884 | |
30885 | // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient |
30886 | // for any integer data type, including i8/i16. |
30887 | if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) { |
30888 | Cond = getSETCC(CC, Cond, DL, DAG); |
30889 | |
30890 | // Zero extend the condition if needed. |
30891 | Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, |
30892 | FalseC->getValueType(0), Cond); |
30893 | Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond, |
30894 | SDValue(FalseC, 0)); |
30895 | |
30896 | if (N->getNumValues() == 2) // Dead flag value? |
30897 | return DCI.CombineTo(N, Cond, SDValue()); |
30898 | return Cond; |
30899 | } |
30900 | |
30901 | // Optimize cases that will turn into an LEA instruction. This requires |
30902 | // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9). |
30903 | if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) { |
30904 | uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue(); |
30905 | if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff; |
30906 | |
30907 | bool isFastMultiplier = false; |
30908 | if (Diff < 10) { |
30909 | switch ((unsigned char)Diff) { |
30910 | default: break; |
30911 | case 1: // result = add base, cond |
30912 | case 2: // result = lea base( , cond*2) |
30913 | case 3: // result = lea base(cond, cond*2) |
30914 | case 4: // result = lea base( , cond*4) |
30915 | case 5: // result = lea base(cond, cond*4) |
30916 | case 8: // result = lea base( , cond*8) |
30917 | case 9: // result = lea base(cond, cond*8) |
30918 | isFastMultiplier = true; |
30919 | break; |
30920 | } |
30921 | } |
30922 | |
30923 | if (isFastMultiplier) { |
30924 | APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue(); |
30925 | Cond = getSETCC(CC, Cond, DL ,DAG); |
30926 | // Zero extend the condition if needed. |
30927 | Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0), |
30928 | Cond); |
30929 | // Scale the condition by the difference. |
30930 | if (Diff != 1) |
30931 | Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond, |
30932 | DAG.getConstant(Diff, DL, Cond.getValueType())); |
30933 | |
30934 | // Add the base if non-zero. |
30935 | if (FalseC->getAPIntValue() != 0) |
30936 | Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond, |
30937 | SDValue(FalseC, 0)); |
30938 | if (N->getNumValues() == 2) // Dead flag value? |
30939 | return DCI.CombineTo(N, Cond, SDValue()); |
30940 | return Cond; |
30941 | } |
30942 | } |
30943 | } |
30944 | } |
30945 | |
30946 | // Handle these cases: |
30947 | // (select (x != c), e, c) -> select (x != c), e, x), |
30948 | // (select (x == c), c, e) -> select (x == c), x, e) |
30949 | // where the c is an integer constant, and the "select" is the combination |
30950 | // of CMOV and CMP. |
30951 | // |
30952 | // The rationale for this change is that the conditional-move from a constant |
30953 | // needs two instructions, however, conditional-move from a register needs |
30954 | // only one instruction. |
30955 | // |
30956 | // CAVEAT: By replacing a constant with a symbolic value, it may obscure |
30957 | // some instruction-combining opportunities. This opt needs to be |
30958 | // postponed as late as possible. |
30959 | // |
30960 | if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) { |
30961 | // the DCI.xxxx conditions are provided to postpone the optimization as |
30962 | // late as possible. |
30963 | |
30964 | ConstantSDNode *CmpAgainst = nullptr; |
30965 | if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) && |
30966 | (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) && |
30967 | !isa<ConstantSDNode>(Cond.getOperand(0))) { |
30968 | |
30969 | if (CC == X86::COND_NE && |
30970 | CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) { |
30971 | CC = X86::GetOppositeBranchCondition(CC); |
30972 | std::swap(TrueOp, FalseOp); |
30973 | } |
30974 | |
30975 | if (CC == X86::COND_E && |
30976 | CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) { |
30977 | SDValue Ops[] = { FalseOp, Cond.getOperand(0), |
30978 | DAG.getConstant(CC, DL, MVT::i8), Cond }; |
30979 | return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops); |
30980 | } |
30981 | } |
30982 | } |
30983 | |
30984 | // Fold and/or of setcc's to double CMOV: |
30985 | // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2) |
30986 | // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2) |
30987 | // |
30988 | // This combine lets us generate: |
30989 | // cmovcc1 (jcc1 if we don't have CMOV) |
30990 | // cmovcc2 (same) |
30991 | // instead of: |
30992 | // setcc1 |
30993 | // setcc2 |
30994 | // and/or |
30995 | // cmovne (jne if we don't have CMOV) |
30996 | // When we can't use the CMOV instruction, it might increase branch |
30997 | // mispredicts. |
30998 | // When we can use CMOV, or when there is no mispredict, this improves |
30999 | // throughput and reduces register pressure. |
31000 | // |
31001 | if (CC == X86::COND_NE) { |
31002 | SDValue Flags; |
31003 | X86::CondCode CC0, CC1; |
31004 | bool isAndSetCC; |
31005 | if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) { |
31006 | if (isAndSetCC) { |
31007 | std::swap(FalseOp, TrueOp); |
31008 | CC0 = X86::GetOppositeBranchCondition(CC0); |
31009 | CC1 = X86::GetOppositeBranchCondition(CC1); |
31010 | } |
31011 | |
31012 | SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8), |
31013 | Flags}; |
31014 | SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps); |
31015 | SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags}; |
31016 | SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops); |
31017 | DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1)); |
31018 | return CMOV; |
31019 | } |
31020 | } |
31021 | |
31022 | return SDValue(); |
31023 | } |
31024 | |
31025 | /// Different mul shrinking modes. |
31026 | enum ShrinkMode { MULS8, MULU8, MULS16, MULU16 }; |
31027 | |
31028 | static bool canReduceVMulWidth(SDNode *N, SelectionDAG &DAG, ShrinkMode &Mode) { |
31029 | EVT VT = N->getOperand(0).getValueType(); |
31030 | if (VT.getScalarSizeInBits() != 32) |
31031 | return false; |
31032 | |
31033 | assert(N->getNumOperands() == 2 && "NumOperands of Mul are 2")((N->getNumOperands() == 2 && "NumOperands of Mul are 2" ) ? static_cast<void> (0) : __assert_fail ("N->getNumOperands() == 2 && \"NumOperands of Mul are 2\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31033, __PRETTY_FUNCTION__)); |
31034 | unsigned SignBits[2] = {1, 1}; |
31035 | bool IsPositive[2] = {false, false}; |
31036 | for (unsigned i = 0; i < 2; i++) { |
31037 | SDValue Opd = N->getOperand(i); |
31038 | |
31039 | // DAG.ComputeNumSignBits return 1 for ISD::ANY_EXTEND, so we need to |
31040 | // compute signbits for it separately. |
31041 | if (Opd.getOpcode() == ISD::ANY_EXTEND) { |
31042 | // For anyextend, it is safe to assume an appropriate number of leading |
31043 | // sign/zero bits. |
31044 | if (Opd.getOperand(0).getValueType().getVectorElementType() == MVT::i8) |
31045 | SignBits[i] = 25; |
31046 | else if (Opd.getOperand(0).getValueType().getVectorElementType() == |
31047 | MVT::i16) |
31048 | SignBits[i] = 17; |
31049 | else |
31050 | return false; |
31051 | IsPositive[i] = true; |
31052 | } else if (Opd.getOpcode() == ISD::BUILD_VECTOR) { |
31053 | // All the operands of BUILD_VECTOR need to be int constant. |
31054 | // Find the smallest value range which all the operands belong to. |
31055 | SignBits[i] = 32; |
31056 | IsPositive[i] = true; |
31057 | for (const SDValue &SubOp : Opd.getNode()->op_values()) { |
31058 | if (SubOp.isUndef()) |
31059 | continue; |
31060 | auto *CN = dyn_cast<ConstantSDNode>(SubOp); |
31061 | if (!CN) |
31062 | return false; |
31063 | APInt IntVal = CN->getAPIntValue(); |
31064 | if (IntVal.isNegative()) |
31065 | IsPositive[i] = false; |
31066 | SignBits[i] = std::min(SignBits[i], IntVal.getNumSignBits()); |
31067 | } |
31068 | } else { |
31069 | SignBits[i] = DAG.ComputeNumSignBits(Opd); |
31070 | if (Opd.getOpcode() == ISD::ZERO_EXTEND) |
31071 | IsPositive[i] = true; |
31072 | } |
31073 | } |
31074 | |
31075 | bool AllPositive = IsPositive[0] && IsPositive[1]; |
31076 | unsigned MinSignBits = std::min(SignBits[0], SignBits[1]); |
31077 | // When ranges are from -128 ~ 127, use MULS8 mode. |
31078 | if (MinSignBits >= 25) |
31079 | Mode = MULS8; |
31080 | // When ranges are from 0 ~ 255, use MULU8 mode. |
31081 | else if (AllPositive && MinSignBits >= 24) |
31082 | Mode = MULU8; |
31083 | // When ranges are from -32768 ~ 32767, use MULS16 mode. |
31084 | else if (MinSignBits >= 17) |
31085 | Mode = MULS16; |
31086 | // When ranges are from 0 ~ 65535, use MULU16 mode. |
31087 | else if (AllPositive && MinSignBits >= 16) |
31088 | Mode = MULU16; |
31089 | else |
31090 | return false; |
31091 | return true; |
31092 | } |
31093 | |
31094 | /// When the operands of vector mul are extended from smaller size values, |
31095 | /// like i8 and i16, the type of mul may be shrinked to generate more |
31096 | /// efficient code. Two typical patterns are handled: |
31097 | /// Pattern1: |
31098 | /// %2 = sext/zext <N x i8> %1 to <N x i32> |
31099 | /// %4 = sext/zext <N x i8> %3 to <N x i32> |
31100 | // or %4 = build_vector <N x i32> %C1, ..., %CN (%C1..%CN are constants) |
31101 | /// %5 = mul <N x i32> %2, %4 |
31102 | /// |
31103 | /// Pattern2: |
31104 | /// %2 = zext/sext <N x i16> %1 to <N x i32> |
31105 | /// %4 = zext/sext <N x i16> %3 to <N x i32> |
31106 | /// or %4 = build_vector <N x i32> %C1, ..., %CN (%C1..%CN are constants) |
31107 | /// %5 = mul <N x i32> %2, %4 |
31108 | /// |
31109 | /// There are four mul shrinking modes: |
31110 | /// If %2 == sext32(trunc8(%2)), i.e., the scalar value range of %2 is |
31111 | /// -128 to 128, and the scalar value range of %4 is also -128 to 128, |
31112 | /// generate pmullw+sext32 for it (MULS8 mode). |
31113 | /// If %2 == zext32(trunc8(%2)), i.e., the scalar value range of %2 is |
31114 | /// 0 to 255, and the scalar value range of %4 is also 0 to 255, |
31115 | /// generate pmullw+zext32 for it (MULU8 mode). |
31116 | /// If %2 == sext32(trunc16(%2)), i.e., the scalar value range of %2 is |
31117 | /// -32768 to 32767, and the scalar value range of %4 is also -32768 to 32767, |
31118 | /// generate pmullw+pmulhw for it (MULS16 mode). |
31119 | /// If %2 == zext32(trunc16(%2)), i.e., the scalar value range of %2 is |
31120 | /// 0 to 65535, and the scalar value range of %4 is also 0 to 65535, |
31121 | /// generate pmullw+pmulhuw for it (MULU16 mode). |
31122 | static SDValue reduceVMULWidth(SDNode *N, SelectionDAG &DAG, |
31123 | const X86Subtarget &Subtarget) { |
31124 | // Check for legality |
31125 | // pmullw/pmulhw are not supported by SSE. |
31126 | if (!Subtarget.hasSSE2()) |
31127 | return SDValue(); |
31128 | |
31129 | // Check for profitability |
31130 | // pmulld is supported since SSE41. It is better to use pmulld |
31131 | // instead of pmullw+pmulhw, except for subtargets where pmulld is slower than |
31132 | // the expansion. |
31133 | bool OptForMinSize = DAG.getMachineFunction().getFunction()->optForMinSize(); |
31134 | if (Subtarget.hasSSE41() && (OptForMinSize || !Subtarget.isPMULLDSlow())) |
31135 | return SDValue(); |
31136 | |
31137 | ShrinkMode Mode; |
31138 | if (!canReduceVMulWidth(N, DAG, Mode)) |
31139 | return SDValue(); |
31140 | |
31141 | SDLoc DL(N); |
31142 | SDValue N0 = N->getOperand(0); |
31143 | SDValue N1 = N->getOperand(1); |
31144 | EVT VT = N->getOperand(0).getValueType(); |
31145 | unsigned RegSize = 128; |
31146 | MVT OpsVT = MVT::getVectorVT(MVT::i16, RegSize / 16); |
31147 | EVT ReducedVT = |
31148 | EVT::getVectorVT(*DAG.getContext(), MVT::i16, VT.getVectorNumElements()); |
31149 | // Shrink the operands of mul. |
31150 | SDValue NewN0 = DAG.getNode(ISD::TRUNCATE, DL, ReducedVT, N0); |
31151 | SDValue NewN1 = DAG.getNode(ISD::TRUNCATE, DL, ReducedVT, N1); |
31152 | |
31153 | if (VT.getVectorNumElements() >= OpsVT.getVectorNumElements()) { |
31154 | // Generate the lower part of mul: pmullw. For MULU8/MULS8, only the |
31155 | // lower part is needed. |
31156 | SDValue MulLo = DAG.getNode(ISD::MUL, DL, ReducedVT, NewN0, NewN1); |
31157 | if (Mode == MULU8 || Mode == MULS8) { |
31158 | return DAG.getNode((Mode == MULU8) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND, |
31159 | DL, VT, MulLo); |
31160 | } else { |
31161 | MVT ResVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2); |
31162 | // Generate the higher part of mul: pmulhw/pmulhuw. For MULU16/MULS16, |
31163 | // the higher part is also needed. |
31164 | SDValue MulHi = DAG.getNode(Mode == MULS16 ? ISD::MULHS : ISD::MULHU, DL, |
31165 | ReducedVT, NewN0, NewN1); |
31166 | |
31167 | // Repack the lower part and higher part result of mul into a wider |
31168 | // result. |
31169 | // Generate shuffle functioning as punpcklwd. |
31170 | SmallVector<int, 16> ShuffleMask(VT.getVectorNumElements()); |
31171 | for (unsigned i = 0; i < VT.getVectorNumElements() / 2; i++) { |
31172 | ShuffleMask[2 * i] = i; |
31173 | ShuffleMask[2 * i + 1] = i + VT.getVectorNumElements(); |
31174 | } |
31175 | SDValue ResLo = |
31176 | DAG.getVectorShuffle(ReducedVT, DL, MulLo, MulHi, ShuffleMask); |
31177 | ResLo = DAG.getNode(ISD::BITCAST, DL, ResVT, ResLo); |
31178 | // Generate shuffle functioning as punpckhwd. |
31179 | for (unsigned i = 0; i < VT.getVectorNumElements() / 2; i++) { |
31180 | ShuffleMask[2 * i] = i + VT.getVectorNumElements() / 2; |
31181 | ShuffleMask[2 * i + 1] = i + VT.getVectorNumElements() * 3 / 2; |
31182 | } |
31183 | SDValue ResHi = |
31184 | DAG.getVectorShuffle(ReducedVT, DL, MulLo, MulHi, ShuffleMask); |
31185 | ResHi = DAG.getNode(ISD::BITCAST, DL, ResVT, ResHi); |
31186 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, ResLo, ResHi); |
31187 | } |
31188 | } else { |
31189 | // When VT.getVectorNumElements() < OpsVT.getVectorNumElements(), we want |
31190 | // to legalize the mul explicitly because implicit legalization for type |
31191 | // <4 x i16> to <4 x i32> sometimes involves unnecessary unpack |
31192 | // instructions which will not exist when we explicitly legalize it by |
31193 | // extending <4 x i16> to <8 x i16> (concatenating the <4 x i16> val with |
31194 | // <4 x i16> undef). |
31195 | // |
31196 | // Legalize the operands of mul. |
31197 | // FIXME: We may be able to handle non-concatenated vectors by insertion. |
31198 | unsigned ReducedSizeInBits = ReducedVT.getSizeInBits(); |
31199 | if ((RegSize % ReducedSizeInBits) != 0) |
31200 | return SDValue(); |
31201 | |
31202 | SmallVector<SDValue, 16> Ops(RegSize / ReducedSizeInBits, |
31203 | DAG.getUNDEF(ReducedVT)); |
31204 | Ops[0] = NewN0; |
31205 | NewN0 = DAG.getNode(ISD::CONCAT_VECTORS, DL, OpsVT, Ops); |
31206 | Ops[0] = NewN1; |
31207 | NewN1 = DAG.getNode(ISD::CONCAT_VECTORS, DL, OpsVT, Ops); |
31208 | |
31209 | if (Mode == MULU8 || Mode == MULS8) { |
31210 | // Generate lower part of mul: pmullw. For MULU8/MULS8, only the lower |
31211 | // part is needed. |
31212 | SDValue Mul = DAG.getNode(ISD::MUL, DL, OpsVT, NewN0, NewN1); |
31213 | |
31214 | // convert the type of mul result to VT. |
31215 | MVT ResVT = MVT::getVectorVT(MVT::i32, RegSize / 32); |
31216 | SDValue Res = DAG.getNode(Mode == MULU8 ? ISD::ZERO_EXTEND_VECTOR_INREG |
31217 | : ISD::SIGN_EXTEND_VECTOR_INREG, |
31218 | DL, ResVT, Mul); |
31219 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Res, |
31220 | DAG.getIntPtrConstant(0, DL)); |
31221 | } else { |
31222 | // Generate the lower and higher part of mul: pmulhw/pmulhuw. For |
31223 | // MULU16/MULS16, both parts are needed. |
31224 | SDValue MulLo = DAG.getNode(ISD::MUL, DL, OpsVT, NewN0, NewN1); |
31225 | SDValue MulHi = DAG.getNode(Mode == MULS16 ? ISD::MULHS : ISD::MULHU, DL, |
31226 | OpsVT, NewN0, NewN1); |
31227 | |
31228 | // Repack the lower part and higher part result of mul into a wider |
31229 | // result. Make sure the type of mul result is VT. |
31230 | MVT ResVT = MVT::getVectorVT(MVT::i32, RegSize / 32); |
31231 | SDValue Res = DAG.getNode(X86ISD::UNPCKL, DL, OpsVT, MulLo, MulHi); |
31232 | Res = DAG.getNode(ISD::BITCAST, DL, ResVT, Res); |
31233 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Res, |
31234 | DAG.getIntPtrConstant(0, DL)); |
31235 | } |
31236 | } |
31237 | } |
31238 | |
31239 | static SDValue combineMulSpecial(uint64_t MulAmt, SDNode *N, SelectionDAG &DAG, |
31240 | EVT VT, SDLoc DL) { |
31241 | |
31242 | auto combineMulShlAddOrSub = [&](int Mult, int Shift, bool isAdd) { |
31243 | SDValue Result = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0), |
31244 | DAG.getConstant(Mult, DL, VT)); |
31245 | Result = DAG.getNode(ISD::SHL, DL, VT, Result, |
31246 | DAG.getConstant(Shift, DL, MVT::i8)); |
31247 | Result = DAG.getNode(isAdd ? ISD::ADD : ISD::SUB, DL, VT, Result, |
31248 | N->getOperand(0)); |
31249 | return Result; |
31250 | }; |
31251 | |
31252 | auto combineMulMulAddOrSub = [&](bool isAdd) { |
31253 | SDValue Result = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0), |
31254 | DAG.getConstant(9, DL, VT)); |
31255 | Result = DAG.getNode(ISD::MUL, DL, VT, Result, DAG.getConstant(3, DL, VT)); |
31256 | Result = DAG.getNode(isAdd ? ISD::ADD : ISD::SUB, DL, VT, Result, |
31257 | N->getOperand(0)); |
31258 | return Result; |
31259 | }; |
31260 | |
31261 | switch (MulAmt) { |
31262 | default: |
31263 | break; |
31264 | case 11: |
31265 | // mul x, 11 => add ((shl (mul x, 5), 1), x) |
31266 | return combineMulShlAddOrSub(5, 1, /*isAdd*/ true); |
31267 | case 21: |
31268 | // mul x, 21 => add ((shl (mul x, 5), 2), x) |
31269 | return combineMulShlAddOrSub(5, 2, /*isAdd*/ true); |
31270 | case 22: |
31271 | // mul x, 22 => add (add ((shl (mul x, 5), 2), x), x) |
31272 | return DAG.getNode(ISD::ADD, DL, VT, N->getOperand(0), |
31273 | combineMulShlAddOrSub(5, 2, /*isAdd*/ true)); |
31274 | case 19: |
31275 | // mul x, 19 => sub ((shl (mul x, 5), 2), x) |
31276 | return combineMulShlAddOrSub(5, 2, /*isAdd*/ false); |
31277 | case 13: |
31278 | // mul x, 13 => add ((shl (mul x, 3), 2), x) |
31279 | return combineMulShlAddOrSub(3, 2, /*isAdd*/ true); |
31280 | case 23: |
31281 | // mul x, 13 => sub ((shl (mul x, 3), 3), x) |
31282 | return combineMulShlAddOrSub(3, 3, /*isAdd*/ false); |
31283 | case 14: |
31284 | // mul x, 14 => add (add ((shl (mul x, 3), 2), x), x) |
31285 | return DAG.getNode(ISD::ADD, DL, VT, N->getOperand(0), |
31286 | combineMulShlAddOrSub(3, 2, /*isAdd*/ true)); |
31287 | case 26: |
31288 | // mul x, 26 => sub ((mul (mul x, 9), 3), x) |
31289 | return combineMulMulAddOrSub(/*isAdd*/ false); |
31290 | case 28: |
31291 | // mul x, 28 => add ((mul (mul x, 9), 3), x) |
31292 | return combineMulMulAddOrSub(/*isAdd*/ true); |
31293 | case 29: |
31294 | // mul x, 29 => add (add ((mul (mul x, 9), 3), x), x) |
31295 | return DAG.getNode(ISD::ADD, DL, VT, N->getOperand(0), |
31296 | combineMulMulAddOrSub(/*isAdd*/ true)); |
31297 | case 30: |
31298 | // mul x, 30 => sub (sub ((shl x, 5), x), x) |
31299 | return DAG.getNode( |
31300 | ISD::SUB, DL, VT, |
31301 | DAG.getNode(ISD::SUB, DL, VT, |
31302 | DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0), |
31303 | DAG.getConstant(5, DL, MVT::i8)), |
31304 | N->getOperand(0)), |
31305 | N->getOperand(0)); |
31306 | } |
31307 | return SDValue(); |
31308 | } |
31309 | |
31310 | /// Optimize a single multiply with constant into two operations in order to |
31311 | /// implement it with two cheaper instructions, e.g. LEA + SHL, LEA + LEA. |
31312 | static SDValue combineMul(SDNode *N, SelectionDAG &DAG, |
31313 | TargetLowering::DAGCombinerInfo &DCI, |
31314 | const X86Subtarget &Subtarget) { |
31315 | EVT VT = N->getValueType(0); |
31316 | if (DCI.isBeforeLegalize() && VT.isVector()) |
31317 | return reduceVMULWidth(N, DAG, Subtarget); |
31318 | |
31319 | if (!MulConstantOptimization) |
31320 | return SDValue(); |
31321 | // An imul is usually smaller than the alternative sequence. |
31322 | if (DAG.getMachineFunction().getFunction()->optForMinSize()) |
31323 | return SDValue(); |
31324 | |
31325 | if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer()) |
31326 | return SDValue(); |
31327 | |
31328 | if (VT != MVT::i64 && VT != MVT::i32) |
31329 | return SDValue(); |
31330 | |
31331 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)); |
31332 | if (!C) |
31333 | return SDValue(); |
31334 | uint64_t MulAmt = C->getZExtValue(); |
31335 | if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9) |
31336 | return SDValue(); |
31337 | |
31338 | uint64_t MulAmt1 = 0; |
31339 | uint64_t MulAmt2 = 0; |
31340 | if ((MulAmt % 9) == 0) { |
31341 | MulAmt1 = 9; |
31342 | MulAmt2 = MulAmt / 9; |
31343 | } else if ((MulAmt % 5) == 0) { |
31344 | MulAmt1 = 5; |
31345 | MulAmt2 = MulAmt / 5; |
31346 | } else if ((MulAmt % 3) == 0) { |
31347 | MulAmt1 = 3; |
31348 | MulAmt2 = MulAmt / 3; |
31349 | } |
31350 | |
31351 | SDLoc DL(N); |
31352 | SDValue NewMul; |
31353 | if (MulAmt2 && |
31354 | (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){ |
31355 | |
31356 | if (isPowerOf2_64(MulAmt2) && |
31357 | !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD)) |
31358 | // If second multiplifer is pow2, issue it first. We want the multiply by |
31359 | // 3, 5, or 9 to be folded into the addressing mode unless the lone use |
31360 | // is an add. |
31361 | std::swap(MulAmt1, MulAmt2); |
31362 | |
31363 | if (isPowerOf2_64(MulAmt1)) |
31364 | NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0), |
31365 | DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8)); |
31366 | else |
31367 | NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0), |
31368 | DAG.getConstant(MulAmt1, DL, VT)); |
31369 | |
31370 | if (isPowerOf2_64(MulAmt2)) |
31371 | NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul, |
31372 | DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8)); |
31373 | else |
31374 | NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul, |
31375 | DAG.getConstant(MulAmt2, DL, VT)); |
31376 | } else if (!Subtarget.slowLEA()) |
31377 | NewMul = combineMulSpecial(MulAmt, N, DAG, VT, DL); |
31378 | |
31379 | if (!NewMul) { |
31380 | assert(MulAmt != 0 &&((MulAmt != 0 && MulAmt != (VT == MVT::i64 ? (18446744073709551615UL ) : (4294967295U)) && "Both cases that could cause potential overflows should have " "already been handled.") ? static_cast<void> (0) : __assert_fail ("MulAmt != 0 && MulAmt != (VT == MVT::i64 ? UINT64_MAX : UINT32_MAX) && \"Both cases that could cause potential overflows should have \" \"already been handled.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31383, __PRETTY_FUNCTION__)) |
31381 | MulAmt != (VT == MVT::i64 ? UINT64_MAX : UINT32_MAX) &&((MulAmt != 0 && MulAmt != (VT == MVT::i64 ? (18446744073709551615UL ) : (4294967295U)) && "Both cases that could cause potential overflows should have " "already been handled.") ? static_cast<void> (0) : __assert_fail ("MulAmt != 0 && MulAmt != (VT == MVT::i64 ? UINT64_MAX : UINT32_MAX) && \"Both cases that could cause potential overflows should have \" \"already been handled.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31383, __PRETTY_FUNCTION__)) |
31382 | "Both cases that could cause potential overflows should have "((MulAmt != 0 && MulAmt != (VT == MVT::i64 ? (18446744073709551615UL ) : (4294967295U)) && "Both cases that could cause potential overflows should have " "already been handled.") ? static_cast<void> (0) : __assert_fail ("MulAmt != 0 && MulAmt != (VT == MVT::i64 ? UINT64_MAX : UINT32_MAX) && \"Both cases that could cause potential overflows should have \" \"already been handled.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31383, __PRETTY_FUNCTION__)) |
31383 | "already been handled.")((MulAmt != 0 && MulAmt != (VT == MVT::i64 ? (18446744073709551615UL ) : (4294967295U)) && "Both cases that could cause potential overflows should have " "already been handled.") ? static_cast<void> (0) : __assert_fail ("MulAmt != 0 && MulAmt != (VT == MVT::i64 ? UINT64_MAX : UINT32_MAX) && \"Both cases that could cause potential overflows should have \" \"already been handled.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31383, __PRETTY_FUNCTION__)); |
31384 | int64_t SignMulAmt = C->getSExtValue(); |
31385 | if ((SignMulAmt != INT64_MIN(-9223372036854775807L -1)) && (SignMulAmt != INT64_MAX(9223372036854775807L)) && |
31386 | (SignMulAmt != -INT64_MAX(9223372036854775807L))) { |
31387 | int NumSign = SignMulAmt > 0 ? 1 : -1; |
31388 | bool IsPowerOf2_64PlusOne = isPowerOf2_64(NumSign * SignMulAmt - 1); |
31389 | bool IsPowerOf2_64MinusOne = isPowerOf2_64(NumSign * SignMulAmt + 1); |
31390 | if (IsPowerOf2_64PlusOne) { |
31391 | // (mul x, 2^N + 1) => (add (shl x, N), x) |
31392 | NewMul = DAG.getNode( |
31393 | ISD::ADD, DL, VT, N->getOperand(0), |
31394 | DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0), |
31395 | DAG.getConstant(Log2_64(NumSign * SignMulAmt - 1), DL, |
31396 | MVT::i8))); |
31397 | } else if (IsPowerOf2_64MinusOne) { |
31398 | // (mul x, 2^N - 1) => (sub (shl x, N), x) |
31399 | NewMul = DAG.getNode( |
31400 | ISD::SUB, DL, VT, |
31401 | DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0), |
31402 | DAG.getConstant(Log2_64(NumSign * SignMulAmt + 1), DL, |
31403 | MVT::i8)), |
31404 | N->getOperand(0)); |
31405 | } |
31406 | // To negate, subtract the number from zero |
31407 | if ((IsPowerOf2_64PlusOne || IsPowerOf2_64MinusOne) && NumSign == -1) |
31408 | NewMul = |
31409 | DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), NewMul); |
31410 | } |
31411 | } |
31412 | |
31413 | if (NewMul) |
31414 | // Do not add new nodes to DAG combiner worklist. |
31415 | DCI.CombineTo(N, NewMul, false); |
31416 | |
31417 | return SDValue(); |
31418 | } |
31419 | |
31420 | static SDValue combineShiftLeft(SDNode *N, SelectionDAG &DAG) { |
31421 | SDValue N0 = N->getOperand(0); |
31422 | SDValue N1 = N->getOperand(1); |
31423 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); |
31424 | EVT VT = N0.getValueType(); |
31425 | |
31426 | // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2)) |
31427 | // since the result of setcc_c is all zero's or all ones. |
31428 | if (VT.isInteger() && !VT.isVector() && |
31429 | N1C && N0.getOpcode() == ISD::AND && |
31430 | N0.getOperand(1).getOpcode() == ISD::Constant) { |
31431 | SDValue N00 = N0.getOperand(0); |
31432 | APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); |
31433 | Mask <<= N1C->getAPIntValue(); |
31434 | bool MaskOK = false; |
31435 | // We can handle cases concerning bit-widening nodes containing setcc_c if |
31436 | // we carefully interrogate the mask to make sure we are semantics |
31437 | // preserving. |
31438 | // The transform is not safe if the result of C1 << C2 exceeds the bitwidth |
31439 | // of the underlying setcc_c operation if the setcc_c was zero extended. |
31440 | // Consider the following example: |
31441 | // zext(setcc_c) -> i32 0x0000FFFF |
31442 | // c1 -> i32 0x0000FFFF |
31443 | // c2 -> i32 0x00000001 |
31444 | // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE |
31445 | // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE |
31446 | if (N00.getOpcode() == X86ISD::SETCC_CARRY) { |
31447 | MaskOK = true; |
31448 | } else if (N00.getOpcode() == ISD::SIGN_EXTEND && |
31449 | N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) { |
31450 | MaskOK = true; |
31451 | } else if ((N00.getOpcode() == ISD::ZERO_EXTEND || |
31452 | N00.getOpcode() == ISD::ANY_EXTEND) && |
31453 | N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) { |
31454 | MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits()); |
31455 | } |
31456 | if (MaskOK && Mask != 0) { |
31457 | SDLoc DL(N); |
31458 | return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT)); |
31459 | } |
31460 | } |
31461 | |
31462 | // Hardware support for vector shifts is sparse which makes us scalarize the |
31463 | // vector operations in many cases. Also, on sandybridge ADD is faster than |
31464 | // shl. |
31465 | // (shl V, 1) -> add V,V |
31466 | if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1)) |
31467 | if (auto *N1SplatC = N1BV->getConstantSplatNode()) { |
31468 | assert(N0.getValueType().isVector() && "Invalid vector shift type")((N0.getValueType().isVector() && "Invalid vector shift type" ) ? static_cast<void> (0) : __assert_fail ("N0.getValueType().isVector() && \"Invalid vector shift type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31468, __PRETTY_FUNCTION__)); |
31469 | // We shift all of the values by one. In many cases we do not have |
31470 | // hardware support for this operation. This is better expressed as an ADD |
31471 | // of two values. |
31472 | if (N1SplatC->getAPIntValue() == 1) |
31473 | return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0); |
31474 | } |
31475 | |
31476 | return SDValue(); |
31477 | } |
31478 | |
31479 | static SDValue combineShiftRightAlgebraic(SDNode *N, SelectionDAG &DAG) { |
31480 | SDValue N0 = N->getOperand(0); |
31481 | SDValue N1 = N->getOperand(1); |
31482 | EVT VT = N0.getValueType(); |
31483 | unsigned Size = VT.getSizeInBits(); |
31484 | |
31485 | // fold (ashr (shl, a, [56,48,32,24,16]), SarConst) |
31486 | // into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or |
31487 | // into (lshr, (sext (a), SarConst - [56,48,32,24,16])) |
31488 | // depending on sign of (SarConst - [56,48,32,24,16]) |
31489 | |
31490 | // sexts in X86 are MOVs. The MOVs have the same code size |
31491 | // as above SHIFTs (only SHIFT on 1 has lower code size). |
31492 | // However the MOVs have 2 advantages to a SHIFT: |
31493 | // 1. MOVs can write to a register that differs from source |
31494 | // 2. MOVs accept memory operands |
31495 | |
31496 | if (!VT.isInteger() || VT.isVector() || N1.getOpcode() != ISD::Constant || |
31497 | N0.getOpcode() != ISD::SHL || !N0.hasOneUse() || |
31498 | N0.getOperand(1).getOpcode() != ISD::Constant) |
31499 | return SDValue(); |
31500 | |
31501 | SDValue N00 = N0.getOperand(0); |
31502 | SDValue N01 = N0.getOperand(1); |
31503 | APInt ShlConst = (cast<ConstantSDNode>(N01))->getAPIntValue(); |
31504 | APInt SarConst = (cast<ConstantSDNode>(N1))->getAPIntValue(); |
31505 | EVT CVT = N1.getValueType(); |
31506 | |
31507 | if (SarConst.isNegative()) |
31508 | return SDValue(); |
31509 | |
31510 | for (MVT SVT : MVT::integer_valuetypes()) { |
31511 | unsigned ShiftSize = SVT.getSizeInBits(); |
31512 | // skipping types without corresponding sext/zext and |
31513 | // ShlConst that is not one of [56,48,32,24,16] |
31514 | if (ShiftSize < 8 || ShiftSize > 64 || ShlConst != Size - ShiftSize) |
31515 | continue; |
31516 | SDLoc DL(N); |
31517 | SDValue NN = |
31518 | DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, N00, DAG.getValueType(SVT)); |
31519 | SarConst = SarConst - (Size - ShiftSize); |
31520 | if (SarConst == 0) |
31521 | return NN; |
31522 | else if (SarConst.isNegative()) |
31523 | return DAG.getNode(ISD::SHL, DL, VT, NN, |
31524 | DAG.getConstant(-SarConst, DL, CVT)); |
31525 | else |
31526 | return DAG.getNode(ISD::SRA, DL, VT, NN, |
31527 | DAG.getConstant(SarConst, DL, CVT)); |
31528 | } |
31529 | return SDValue(); |
31530 | } |
31531 | |
31532 | /// \brief Returns a vector of 0s if the node in input is a vector logical |
31533 | /// shift by a constant amount which is known to be bigger than or equal |
31534 | /// to the vector element size in bits. |
31535 | static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG, |
31536 | const X86Subtarget &Subtarget) { |
31537 | EVT VT = N->getValueType(0); |
31538 | |
31539 | if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 && |
31540 | (!Subtarget.hasInt256() || |
31541 | (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16))) |
31542 | return SDValue(); |
31543 | |
31544 | SDValue Amt = N->getOperand(1); |
31545 | SDLoc DL(N); |
31546 | if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt)) |
31547 | if (auto *AmtSplat = AmtBV->getConstantSplatNode()) { |
31548 | const APInt &ShiftAmt = AmtSplat->getAPIntValue(); |
31549 | unsigned MaxAmount = |
31550 | VT.getSimpleVT().getScalarSizeInBits(); |
31551 | |
31552 | // SSE2/AVX2 logical shifts always return a vector of 0s |
31553 | // if the shift amount is bigger than or equal to |
31554 | // the element size. The constant shift amount will be |
31555 | // encoded as a 8-bit immediate. |
31556 | if (ShiftAmt.trunc(8).uge(MaxAmount)) |
31557 | return getZeroVector(VT.getSimpleVT(), Subtarget, DAG, DL); |
31558 | } |
31559 | |
31560 | return SDValue(); |
31561 | } |
31562 | |
31563 | static SDValue combineShift(SDNode* N, SelectionDAG &DAG, |
31564 | TargetLowering::DAGCombinerInfo &DCI, |
31565 | const X86Subtarget &Subtarget) { |
31566 | if (N->getOpcode() == ISD::SHL) |
31567 | if (SDValue V = combineShiftLeft(N, DAG)) |
31568 | return V; |
31569 | |
31570 | if (N->getOpcode() == ISD::SRA) |
31571 | if (SDValue V = combineShiftRightAlgebraic(N, DAG)) |
31572 | return V; |
31573 | |
31574 | // Try to fold this logical shift into a zero vector. |
31575 | if (N->getOpcode() != ISD::SRA) |
31576 | if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget)) |
31577 | return V; |
31578 | |
31579 | return SDValue(); |
31580 | } |
31581 | |
31582 | static SDValue combineVectorShiftImm(SDNode *N, SelectionDAG &DAG, |
31583 | TargetLowering::DAGCombinerInfo &DCI, |
31584 | const X86Subtarget &Subtarget) { |
31585 | unsigned Opcode = N->getOpcode(); |
31586 | assert((X86ISD::VSHLI == Opcode || X86ISD::VSRAI == Opcode ||(((X86ISD::VSHLI == Opcode || X86ISD::VSRAI == Opcode || X86ISD ::VSRLI == Opcode) && "Unexpected shift opcode") ? static_cast <void> (0) : __assert_fail ("(X86ISD::VSHLI == Opcode || X86ISD::VSRAI == Opcode || X86ISD::VSRLI == Opcode) && \"Unexpected shift opcode\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31588, __PRETTY_FUNCTION__)) |
31587 | X86ISD::VSRLI == Opcode) &&(((X86ISD::VSHLI == Opcode || X86ISD::VSRAI == Opcode || X86ISD ::VSRLI == Opcode) && "Unexpected shift opcode") ? static_cast <void> (0) : __assert_fail ("(X86ISD::VSHLI == Opcode || X86ISD::VSRAI == Opcode || X86ISD::VSRLI == Opcode) && \"Unexpected shift opcode\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31588, __PRETTY_FUNCTION__)) |
31588 | "Unexpected shift opcode")(((X86ISD::VSHLI == Opcode || X86ISD::VSRAI == Opcode || X86ISD ::VSRLI == Opcode) && "Unexpected shift opcode") ? static_cast <void> (0) : __assert_fail ("(X86ISD::VSHLI == Opcode || X86ISD::VSRAI == Opcode || X86ISD::VSRLI == Opcode) && \"Unexpected shift opcode\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31588, __PRETTY_FUNCTION__)); |
31589 | bool LogicalShift = X86ISD::VSHLI == Opcode || X86ISD::VSRLI == Opcode; |
31590 | EVT VT = N->getValueType(0); |
31591 | SDValue N0 = N->getOperand(0); |
31592 | SDValue N1 = N->getOperand(1); |
31593 | unsigned NumBitsPerElt = VT.getScalarSizeInBits(); |
31594 | assert(VT == N0.getValueType() && (NumBitsPerElt % 8) == 0 &&((VT == N0.getValueType() && (NumBitsPerElt % 8) == 0 && "Unexpected value type") ? static_cast<void> (0) : __assert_fail ("VT == N0.getValueType() && (NumBitsPerElt % 8) == 0 && \"Unexpected value type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31595, __PRETTY_FUNCTION__)) |
31595 | "Unexpected value type")((VT == N0.getValueType() && (NumBitsPerElt % 8) == 0 && "Unexpected value type") ? static_cast<void> (0) : __assert_fail ("VT == N0.getValueType() && (NumBitsPerElt % 8) == 0 && \"Unexpected value type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31595, __PRETTY_FUNCTION__)); |
31596 | |
31597 | // Out of range logical bit shifts are guaranteed to be zero. |
31598 | // Out of range arithmetic bit shifts splat the sign bit. |
31599 | APInt ShiftVal = cast<ConstantSDNode>(N1)->getAPIntValue(); |
31600 | if (ShiftVal.zextOrTrunc(8).uge(NumBitsPerElt)) { |
31601 | if (LogicalShift) |
31602 | return getZeroVector(VT.getSimpleVT(), Subtarget, DAG, SDLoc(N)); |
31603 | else |
31604 | ShiftVal = NumBitsPerElt - 1; |
31605 | } |
31606 | |
31607 | // Shift N0 by zero -> N0. |
31608 | if (!ShiftVal) |
31609 | return N0; |
31610 | |
31611 | // Shift zero -> zero. |
31612 | if (ISD::isBuildVectorAllZeros(N0.getNode())) |
31613 | return getZeroVector(VT.getSimpleVT(), Subtarget, DAG, SDLoc(N)); |
31614 | |
31615 | // fold (VSRLI (VSRAI X, Y), 31) -> (VSRLI X, 31). |
31616 | // This VSRLI only looks at the sign bit, which is unmodified by VSRAI. |
31617 | // TODO - support other sra opcodes as needed. |
31618 | if (Opcode == X86ISD::VSRLI && (ShiftVal + 1) == NumBitsPerElt && |
31619 | N0.getOpcode() == X86ISD::VSRAI) |
31620 | return DAG.getNode(X86ISD::VSRLI, SDLoc(N), VT, N0.getOperand(0), N1); |
31621 | |
31622 | // We can decode 'whole byte' logical bit shifts as shuffles. |
31623 | if (LogicalShift && (ShiftVal.getZExtValue() % 8) == 0) { |
31624 | SDValue Op(N, 0); |
31625 | SmallVector<int, 1> NonceMask; // Just a placeholder. |
31626 | NonceMask.push_back(0); |
31627 | if (combineX86ShufflesRecursively({Op}, 0, Op, NonceMask, {}, |
31628 | /*Depth*/ 1, /*HasVarMask*/ false, DAG, |
31629 | DCI, Subtarget)) |
31630 | return SDValue(); // This routine will use CombineTo to replace N. |
31631 | } |
31632 | |
31633 | // Constant Folding. |
31634 | APInt UndefElts; |
31635 | SmallVector<APInt, 32> EltBits; |
31636 | if (N->isOnlyUserOf(N0.getNode()) && |
31637 | getTargetConstantBitsFromNode(N0, NumBitsPerElt, UndefElts, EltBits)) { |
31638 | assert(EltBits.size() == VT.getVectorNumElements() &&((EltBits.size() == VT.getVectorNumElements() && "Unexpected shift value type" ) ? static_cast<void> (0) : __assert_fail ("EltBits.size() == VT.getVectorNumElements() && \"Unexpected shift value type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31639, __PRETTY_FUNCTION__)) |
31639 | "Unexpected shift value type")((EltBits.size() == VT.getVectorNumElements() && "Unexpected shift value type" ) ? static_cast<void> (0) : __assert_fail ("EltBits.size() == VT.getVectorNumElements() && \"Unexpected shift value type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31639, __PRETTY_FUNCTION__)); |
31640 | unsigned ShiftImm = ShiftVal.getZExtValue(); |
31641 | for (APInt &Elt : EltBits) { |
31642 | if (X86ISD::VSHLI == Opcode) |
31643 | Elt <<= ShiftImm; |
31644 | else if (X86ISD::VSRAI == Opcode) |
31645 | Elt.ashrInPlace(ShiftImm); |
31646 | else |
31647 | Elt.lshrInPlace(ShiftImm); |
31648 | } |
31649 | return getConstVector(EltBits, UndefElts, VT.getSimpleVT(), DAG, SDLoc(N)); |
31650 | } |
31651 | |
31652 | return SDValue(); |
31653 | } |
31654 | |
31655 | static SDValue combineVectorInsert(SDNode *N, SelectionDAG &DAG, |
31656 | TargetLowering::DAGCombinerInfo &DCI, |
31657 | const X86Subtarget &Subtarget) { |
31658 | assert(((((N->getOpcode() == X86ISD::PINSRB && N->getValueType (0) == MVT::v16i8) || (N->getOpcode() == X86ISD::PINSRW && N->getValueType(0) == MVT::v8i16)) && "Unexpected vector insertion" ) ? static_cast<void> (0) : __assert_fail ("((N->getOpcode() == X86ISD::PINSRB && N->getValueType(0) == MVT::v16i8) || (N->getOpcode() == X86ISD::PINSRW && N->getValueType(0) == MVT::v8i16)) && \"Unexpected vector insertion\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31662, __PRETTY_FUNCTION__)) |
31659 | ((N->getOpcode() == X86ISD::PINSRB && N->getValueType(0) == MVT::v16i8) ||((((N->getOpcode() == X86ISD::PINSRB && N->getValueType (0) == MVT::v16i8) || (N->getOpcode() == X86ISD::PINSRW && N->getValueType(0) == MVT::v8i16)) && "Unexpected vector insertion" ) ? static_cast<void> (0) : __assert_fail ("((N->getOpcode() == X86ISD::PINSRB && N->getValueType(0) == MVT::v16i8) || (N->getOpcode() == X86ISD::PINSRW && N->getValueType(0) == MVT::v8i16)) && \"Unexpected vector insertion\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31662, __PRETTY_FUNCTION__)) |
31660 | (N->getOpcode() == X86ISD::PINSRW &&((((N->getOpcode() == X86ISD::PINSRB && N->getValueType (0) == MVT::v16i8) || (N->getOpcode() == X86ISD::PINSRW && N->getValueType(0) == MVT::v8i16)) && "Unexpected vector insertion" ) ? static_cast<void> (0) : __assert_fail ("((N->getOpcode() == X86ISD::PINSRB && N->getValueType(0) == MVT::v16i8) || (N->getOpcode() == X86ISD::PINSRW && N->getValueType(0) == MVT::v8i16)) && \"Unexpected vector insertion\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31662, __PRETTY_FUNCTION__)) |
31661 | N->getValueType(0) == MVT::v8i16)) &&((((N->getOpcode() == X86ISD::PINSRB && N->getValueType (0) == MVT::v16i8) || (N->getOpcode() == X86ISD::PINSRW && N->getValueType(0) == MVT::v8i16)) && "Unexpected vector insertion" ) ? static_cast<void> (0) : __assert_fail ("((N->getOpcode() == X86ISD::PINSRB && N->getValueType(0) == MVT::v16i8) || (N->getOpcode() == X86ISD::PINSRW && N->getValueType(0) == MVT::v8i16)) && \"Unexpected vector insertion\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31662, __PRETTY_FUNCTION__)) |
31662 | "Unexpected vector insertion")((((N->getOpcode() == X86ISD::PINSRB && N->getValueType (0) == MVT::v16i8) || (N->getOpcode() == X86ISD::PINSRW && N->getValueType(0) == MVT::v8i16)) && "Unexpected vector insertion" ) ? static_cast<void> (0) : __assert_fail ("((N->getOpcode() == X86ISD::PINSRB && N->getValueType(0) == MVT::v16i8) || (N->getOpcode() == X86ISD::PINSRW && N->getValueType(0) == MVT::v8i16)) && \"Unexpected vector insertion\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31662, __PRETTY_FUNCTION__)); |
31663 | |
31664 | // Attempt to combine PINSRB/PINSRW patterns to a shuffle. |
31665 | SDValue Op(N, 0); |
31666 | SmallVector<int, 1> NonceMask; // Just a placeholder. |
31667 | NonceMask.push_back(0); |
31668 | combineX86ShufflesRecursively({Op}, 0, Op, NonceMask, {}, |
31669 | /*Depth*/ 1, /*HasVarMask*/ false, DAG, |
31670 | DCI, Subtarget); |
31671 | return SDValue(); |
31672 | } |
31673 | |
31674 | /// Recognize the distinctive (AND (setcc ...) (setcc ..)) where both setccs |
31675 | /// reference the same FP CMP, and rewrite for CMPEQSS and friends. Likewise for |
31676 | /// OR -> CMPNEQSS. |
31677 | static SDValue combineCompareEqual(SDNode *N, SelectionDAG &DAG, |
31678 | TargetLowering::DAGCombinerInfo &DCI, |
31679 | const X86Subtarget &Subtarget) { |
31680 | unsigned opcode; |
31681 | |
31682 | // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but |
31683 | // we're requiring SSE2 for both. |
31684 | if (Subtarget.hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) { |
31685 | SDValue N0 = N->getOperand(0); |
31686 | SDValue N1 = N->getOperand(1); |
31687 | SDValue CMP0 = N0->getOperand(1); |
31688 | SDValue CMP1 = N1->getOperand(1); |
31689 | SDLoc DL(N); |
31690 | |
31691 | // The SETCCs should both refer to the same CMP. |
31692 | if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1) |
31693 | return SDValue(); |
31694 | |
31695 | SDValue CMP00 = CMP0->getOperand(0); |
31696 | SDValue CMP01 = CMP0->getOperand(1); |
31697 | EVT VT = CMP00.getValueType(); |
31698 | |
31699 | if (VT == MVT::f32 || VT == MVT::f64) { |
31700 | bool ExpectingFlags = false; |
31701 | // Check for any users that want flags: |
31702 | for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); |
31703 | !ExpectingFlags && UI != UE; ++UI) |
31704 | switch (UI->getOpcode()) { |
31705 | default: |
31706 | case ISD::BR_CC: |
31707 | case ISD::BRCOND: |
31708 | case ISD::SELECT: |
31709 | ExpectingFlags = true; |
31710 | break; |
31711 | case ISD::CopyToReg: |
31712 | case ISD::SIGN_EXTEND: |
31713 | case ISD::ZERO_EXTEND: |
31714 | case ISD::ANY_EXTEND: |
31715 | break; |
31716 | } |
31717 | |
31718 | if (!ExpectingFlags) { |
31719 | enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0); |
31720 | enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0); |
31721 | |
31722 | if (cc1 == X86::COND_E || cc1 == X86::COND_NE) { |
31723 | X86::CondCode tmp = cc0; |
31724 | cc0 = cc1; |
31725 | cc1 = tmp; |
31726 | } |
31727 | |
31728 | if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) || |
31729 | (cc0 == X86::COND_NE && cc1 == X86::COND_P)) { |
31730 | // FIXME: need symbolic constants for these magic numbers. |
31731 | // See X86ATTInstPrinter.cpp:printSSECC(). |
31732 | unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4; |
31733 | if (Subtarget.hasAVX512()) { |
31734 | SDValue FSetCC = |
31735 | DAG.getNode(X86ISD::FSETCCM, DL, MVT::v1i1, CMP00, CMP01, |
31736 | DAG.getConstant(x86cc, DL, MVT::i8)); |
31737 | return DAG.getNode(X86ISD::VEXTRACT, DL, N->getSimpleValueType(0), |
31738 | FSetCC, DAG.getIntPtrConstant(0, DL)); |
31739 | } |
31740 | SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL, |
31741 | CMP00.getValueType(), CMP00, CMP01, |
31742 | DAG.getConstant(x86cc, DL, |
31743 | MVT::i8)); |
31744 | |
31745 | bool is64BitFP = (CMP00.getValueType() == MVT::f64); |
31746 | MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32; |
31747 | |
31748 | if (is64BitFP && !Subtarget.is64Bit()) { |
31749 | // On a 32-bit target, we cannot bitcast the 64-bit float to a |
31750 | // 64-bit integer, since that's not a legal type. Since |
31751 | // OnesOrZeroesF is all ones of all zeroes, we don't need all the |
31752 | // bits, but can do this little dance to extract the lowest 32 bits |
31753 | // and work with those going forward. |
31754 | SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, |
31755 | OnesOrZeroesF); |
31756 | SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64); |
31757 | OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, |
31758 | Vector32, DAG.getIntPtrConstant(0, DL)); |
31759 | IntVT = MVT::i32; |
31760 | } |
31761 | |
31762 | SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF); |
31763 | SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI, |
31764 | DAG.getConstant(1, DL, IntVT)); |
31765 | SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, |
31766 | ANDed); |
31767 | return OneBitOfTruth; |
31768 | } |
31769 | } |
31770 | } |
31771 | } |
31772 | return SDValue(); |
31773 | } |
31774 | |
31775 | /// Try to fold: (and (xor X, -1), Y) -> (andnp X, Y). |
31776 | static SDValue combineANDXORWithAllOnesIntoANDNP(SDNode *N, SelectionDAG &DAG) { |
31777 | assert(N->getOpcode() == ISD::AND)((N->getOpcode() == ISD::AND) ? static_cast<void> (0 ) : __assert_fail ("N->getOpcode() == ISD::AND", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31777, __PRETTY_FUNCTION__)); |
31778 | |
31779 | EVT VT = N->getValueType(0); |
31780 | SDValue N0 = N->getOperand(0); |
31781 | SDValue N1 = N->getOperand(1); |
31782 | SDLoc DL(N); |
31783 | |
31784 | if (VT != MVT::v2i64 && VT != MVT::v4i64 && VT != MVT::v8i64) |
31785 | return SDValue(); |
31786 | |
31787 | if (N0.getOpcode() == ISD::XOR && |
31788 | ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode())) |
31789 | return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1); |
31790 | |
31791 | if (N1.getOpcode() == ISD::XOR && |
31792 | ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode())) |
31793 | return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0); |
31794 | |
31795 | return SDValue(); |
31796 | } |
31797 | |
31798 | // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized |
31799 | // register. In most cases we actually compare or select YMM-sized registers |
31800 | // and mixing the two types creates horrible code. This method optimizes |
31801 | // some of the transition sequences. |
31802 | static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG, |
31803 | TargetLowering::DAGCombinerInfo &DCI, |
31804 | const X86Subtarget &Subtarget) { |
31805 | EVT VT = N->getValueType(0); |
31806 | if (!VT.is256BitVector()) |
31807 | return SDValue(); |
31808 | |
31809 | assert((N->getOpcode() == ISD::ANY_EXTEND ||(((N->getOpcode() == ISD::ANY_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND || N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node") ? static_cast<void> (0) : __assert_fail ("(N->getOpcode() == ISD::ANY_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND || N->getOpcode() == ISD::SIGN_EXTEND) && \"Invalid Node\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31811, __PRETTY_FUNCTION__)) |
31810 | N->getOpcode() == ISD::ZERO_EXTEND ||(((N->getOpcode() == ISD::ANY_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND || N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node") ? static_cast<void> (0) : __assert_fail ("(N->getOpcode() == ISD::ANY_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND || N->getOpcode() == ISD::SIGN_EXTEND) && \"Invalid Node\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31811, __PRETTY_FUNCTION__)) |
31811 | N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node")(((N->getOpcode() == ISD::ANY_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND || N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node") ? static_cast<void> (0) : __assert_fail ("(N->getOpcode() == ISD::ANY_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND || N->getOpcode() == ISD::SIGN_EXTEND) && \"Invalid Node\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31811, __PRETTY_FUNCTION__)); |
31812 | |
31813 | SDValue Narrow = N->getOperand(0); |
31814 | EVT NarrowVT = Narrow->getValueType(0); |
31815 | if (!NarrowVT.is128BitVector()) |
31816 | return SDValue(); |
31817 | |
31818 | if (Narrow->getOpcode() != ISD::XOR && |
31819 | Narrow->getOpcode() != ISD::AND && |
31820 | Narrow->getOpcode() != ISD::OR) |
31821 | return SDValue(); |
31822 | |
31823 | SDValue N0 = Narrow->getOperand(0); |
31824 | SDValue N1 = Narrow->getOperand(1); |
31825 | SDLoc DL(Narrow); |
31826 | |
31827 | // The Left side has to be a trunc. |
31828 | if (N0.getOpcode() != ISD::TRUNCATE) |
31829 | return SDValue(); |
31830 | |
31831 | // The type of the truncated inputs. |
31832 | EVT WideVT = N0->getOperand(0)->getValueType(0); |
31833 | if (WideVT != VT) |
31834 | return SDValue(); |
31835 | |
31836 | // The right side has to be a 'trunc' or a constant vector. |
31837 | bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE; |
31838 | ConstantSDNode *RHSConstSplat = nullptr; |
31839 | if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1)) |
31840 | RHSConstSplat = RHSBV->getConstantSplatNode(); |
31841 | if (!RHSTrunc && !RHSConstSplat) |
31842 | return SDValue(); |
31843 | |
31844 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
31845 | |
31846 | if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT)) |
31847 | return SDValue(); |
31848 | |
31849 | // Set N0 and N1 to hold the inputs to the new wide operation. |
31850 | N0 = N0->getOperand(0); |
31851 | if (RHSConstSplat) { |
31852 | N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getVectorElementType(), |
31853 | SDValue(RHSConstSplat, 0)); |
31854 | N1 = DAG.getSplatBuildVector(WideVT, DL, N1); |
31855 | } else if (RHSTrunc) { |
31856 | N1 = N1->getOperand(0); |
31857 | } |
31858 | |
31859 | // Generate the wide operation. |
31860 | SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1); |
31861 | unsigned Opcode = N->getOpcode(); |
31862 | switch (Opcode) { |
31863 | case ISD::ANY_EXTEND: |
31864 | return Op; |
31865 | case ISD::ZERO_EXTEND: { |
31866 | unsigned InBits = NarrowVT.getScalarSizeInBits(); |
31867 | APInt Mask = APInt::getAllOnesValue(InBits); |
31868 | Mask = Mask.zext(VT.getScalarSizeInBits()); |
31869 | return DAG.getNode(ISD::AND, DL, VT, |
31870 | Op, DAG.getConstant(Mask, DL, VT)); |
31871 | } |
31872 | case ISD::SIGN_EXTEND: |
31873 | return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, |
31874 | Op, DAG.getValueType(NarrowVT)); |
31875 | default: |
31876 | llvm_unreachable("Unexpected opcode")::llvm::llvm_unreachable_internal("Unexpected opcode", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31876); |
31877 | } |
31878 | } |
31879 | |
31880 | /// If both input operands of a logic op are being cast from floating point |
31881 | /// types, try to convert this into a floating point logic node to avoid |
31882 | /// unnecessary moves from SSE to integer registers. |
31883 | static SDValue convertIntLogicToFPLogic(SDNode *N, SelectionDAG &DAG, |
31884 | const X86Subtarget &Subtarget) { |
31885 | unsigned FPOpcode = ISD::DELETED_NODE; |
31886 | if (N->getOpcode() == ISD::AND) |
31887 | FPOpcode = X86ISD::FAND; |
31888 | else if (N->getOpcode() == ISD::OR) |
31889 | FPOpcode = X86ISD::FOR; |
31890 | else if (N->getOpcode() == ISD::XOR) |
31891 | FPOpcode = X86ISD::FXOR; |
31892 | |
31893 | assert(FPOpcode != ISD::DELETED_NODE &&((FPOpcode != ISD::DELETED_NODE && "Unexpected input node for FP logic conversion" ) ? static_cast<void> (0) : __assert_fail ("FPOpcode != ISD::DELETED_NODE && \"Unexpected input node for FP logic conversion\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31894, __PRETTY_FUNCTION__)) |
31894 | "Unexpected input node for FP logic conversion")((FPOpcode != ISD::DELETED_NODE && "Unexpected input node for FP logic conversion" ) ? static_cast<void> (0) : __assert_fail ("FPOpcode != ISD::DELETED_NODE && \"Unexpected input node for FP logic conversion\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 31894, __PRETTY_FUNCTION__)); |
31895 | |
31896 | EVT VT = N->getValueType(0); |
31897 | SDValue N0 = N->getOperand(0); |
31898 | SDValue N1 = N->getOperand(1); |
31899 | SDLoc DL(N); |
31900 | if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST && |
31901 | ((Subtarget.hasSSE1() && VT == MVT::i32) || |
31902 | (Subtarget.hasSSE2() && VT == MVT::i64))) { |
31903 | SDValue N00 = N0.getOperand(0); |
31904 | SDValue N10 = N1.getOperand(0); |
31905 | EVT N00Type = N00.getValueType(); |
31906 | EVT N10Type = N10.getValueType(); |
31907 | if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) { |
31908 | SDValue FPLogic = DAG.getNode(FPOpcode, DL, N00Type, N00, N10); |
31909 | return DAG.getBitcast(VT, FPLogic); |
31910 | } |
31911 | } |
31912 | return SDValue(); |
31913 | } |
31914 | |
31915 | /// If this is a zero/all-bits result that is bitwise-anded with a low bits |
31916 | /// mask. (Mask == 1 for the x86 lowering of a SETCC + ZEXT), replace the 'and' |
31917 | /// with a shift-right to eliminate loading the vector constant mask value. |
31918 | static SDValue combineAndMaskToShift(SDNode *N, SelectionDAG &DAG, |
31919 | const X86Subtarget &Subtarget) { |
31920 | SDValue Op0 = peekThroughBitcasts(N->getOperand(0)); |
31921 | SDValue Op1 = peekThroughBitcasts(N->getOperand(1)); |
31922 | EVT VT0 = Op0.getValueType(); |
31923 | EVT VT1 = Op1.getValueType(); |
31924 | |
31925 | if (VT0 != VT1 || !VT0.isSimple() || !VT0.isInteger()) |
31926 | return SDValue(); |
31927 | |
31928 | APInt SplatVal; |
31929 | if (!ISD::isConstantSplatVector(Op1.getNode(), SplatVal) || |
31930 | !SplatVal.isMask()) |
31931 | return SDValue(); |
31932 | |
31933 | if (!SupportedVectorShiftWithImm(VT0.getSimpleVT(), Subtarget, ISD::SRL)) |
31934 | return SDValue(); |
31935 | |
31936 | unsigned EltBitWidth = VT0.getScalarSizeInBits(); |
31937 | if (EltBitWidth != DAG.ComputeNumSignBits(Op0)) |
31938 | return SDValue(); |
31939 | |
31940 | SDLoc DL(N); |
31941 | unsigned ShiftVal = SplatVal.countTrailingOnes(); |
31942 | SDValue ShAmt = DAG.getConstant(EltBitWidth - ShiftVal, DL, MVT::i8); |
31943 | SDValue Shift = DAG.getNode(X86ISD::VSRLI, DL, VT0, Op0, ShAmt); |
31944 | return DAG.getBitcast(N->getValueType(0), Shift); |
31945 | } |
31946 | |
31947 | static SDValue combineAnd(SDNode *N, SelectionDAG &DAG, |
31948 | TargetLowering::DAGCombinerInfo &DCI, |
31949 | const X86Subtarget &Subtarget) { |
31950 | if (DCI.isBeforeLegalizeOps()) |
31951 | return SDValue(); |
31952 | |
31953 | if (SDValue R = combineCompareEqual(N, DAG, DCI, Subtarget)) |
31954 | return R; |
31955 | |
31956 | if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget)) |
31957 | return FPLogic; |
31958 | |
31959 | if (SDValue R = combineANDXORWithAllOnesIntoANDNP(N, DAG)) |
31960 | return R; |
31961 | |
31962 | if (SDValue ShiftRight = combineAndMaskToShift(N, DAG, Subtarget)) |
31963 | return ShiftRight; |
31964 | |
31965 | EVT VT = N->getValueType(0); |
31966 | SDValue N0 = N->getOperand(0); |
31967 | SDValue N1 = N->getOperand(1); |
31968 | SDLoc DL(N); |
31969 | |
31970 | // Attempt to recursively combine a bitmask AND with shuffles. |
31971 | if (VT.isVector() && (VT.getScalarSizeInBits() % 8) == 0) { |
31972 | SDValue Op(N, 0); |
31973 | SmallVector<int, 1> NonceMask; // Just a placeholder. |
31974 | NonceMask.push_back(0); |
31975 | if (combineX86ShufflesRecursively({Op}, 0, Op, NonceMask, {}, |
31976 | /*Depth*/ 1, /*HasVarMask*/ false, DAG, |
31977 | DCI, Subtarget)) |
31978 | return SDValue(); // This routine will use CombineTo to replace N. |
31979 | } |
31980 | |
31981 | // Create BEXTR instructions |
31982 | // BEXTR is ((X >> imm) & (2**size-1)) |
31983 | if (VT != MVT::i32 && VT != MVT::i64) |
31984 | return SDValue(); |
31985 | |
31986 | if (!Subtarget.hasBMI() && !Subtarget.hasTBM()) |
31987 | return SDValue(); |
31988 | if (N0.getOpcode() != ISD::SRA && N0.getOpcode() != ISD::SRL) |
31989 | return SDValue(); |
31990 | |
31991 | ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1); |
31992 | ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1)); |
31993 | if (MaskNode && ShiftNode) { |
31994 | uint64_t Mask = MaskNode->getZExtValue(); |
31995 | uint64_t Shift = ShiftNode->getZExtValue(); |
31996 | if (isMask_64(Mask)) { |
31997 | uint64_t MaskSize = countPopulation(Mask); |
31998 | if (Shift + MaskSize <= VT.getSizeInBits()) |
31999 | return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0), |
32000 | DAG.getConstant(Shift | (MaskSize << 8), DL, |
32001 | VT)); |
32002 | } |
32003 | } |
32004 | return SDValue(); |
32005 | } |
32006 | |
32007 | // Try to fold: |
32008 | // (or (and (m, y), (pandn m, x))) |
32009 | // into: |
32010 | // (vselect m, x, y) |
32011 | // As a special case, try to fold: |
32012 | // (or (and (m, (sub 0, x)), (pandn m, x))) |
32013 | // into: |
32014 | // (sub (xor X, M), M) |
32015 | static SDValue combineLogicBlendIntoPBLENDV(SDNode *N, SelectionDAG &DAG, |
32016 | const X86Subtarget &Subtarget) { |
32017 | assert(N->getOpcode() == ISD::OR && "Unexpected Opcode")((N->getOpcode() == ISD::OR && "Unexpected Opcode" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::OR && \"Unexpected Opcode\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 32017, __PRETTY_FUNCTION__)); |
32018 | |
32019 | SDValue N0 = N->getOperand(0); |
32020 | SDValue N1 = N->getOperand(1); |
32021 | EVT VT = N->getValueType(0); |
32022 | |
32023 | if (!((VT.is128BitVector() && Subtarget.hasSSE2()) || |
32024 | (VT.is256BitVector() && Subtarget.hasInt256()))) |
32025 | return SDValue(); |
32026 | |
32027 | // Canonicalize AND to LHS. |
32028 | if (N1.getOpcode() == ISD::AND) |
32029 | std::swap(N0, N1); |
32030 | |
32031 | // TODO: Attempt to match against AND(XOR(-1,X),Y) as well, waiting for |
32032 | // ANDNP combine allows other combines to happen that prevent matching. |
32033 | if (N0.getOpcode() != ISD::AND || N1.getOpcode() != X86ISD::ANDNP) |
32034 | return SDValue(); |
32035 | |
32036 | SDValue Mask = N1.getOperand(0); |
32037 | SDValue X = N1.getOperand(1); |
32038 | SDValue Y; |
32039 | if (N0.getOperand(0) == Mask) |
32040 | Y = N0.getOperand(1); |
32041 | if (N0.getOperand(1) == Mask) |
32042 | Y = N0.getOperand(0); |
32043 | |
32044 | // Check to see if the mask appeared in both the AND and ANDNP. |
32045 | if (!Y.getNode()) |
32046 | return SDValue(); |
32047 | |
32048 | // Validate that X, Y, and Mask are bitcasts, and see through them. |
32049 | Mask = peekThroughBitcasts(Mask); |
32050 | X = peekThroughBitcasts(X); |
32051 | Y = peekThroughBitcasts(Y); |
32052 | |
32053 | EVT MaskVT = Mask.getValueType(); |
32054 | unsigned EltBits = MaskVT.getScalarSizeInBits(); |
32055 | |
32056 | // TODO: Attempt to handle floating point cases as well? |
32057 | if (!MaskVT.isInteger() || DAG.ComputeNumSignBits(Mask) != EltBits) |
32058 | return SDValue(); |
32059 | |
32060 | SDLoc DL(N); |
32061 | |
32062 | // Try to match: |
32063 | // (or (and (M, (sub 0, X)), (pandn M, X))) |
32064 | // which is a special case of vselect: |
32065 | // (vselect M, (sub 0, X), X) |
32066 | // Per: |
32067 | // http://graphics.stanford.edu/~seander/bithacks.html#ConditionalNegate |
32068 | // We know that, if fNegate is 0 or 1: |
32069 | // (fNegate ? -v : v) == ((v ^ -fNegate) + fNegate) |
32070 | // |
32071 | // Here, we have a mask, M (all 1s or 0), and, similarly, we know that: |
32072 | // ((M & 1) ? -X : X) == ((X ^ -(M & 1)) + (M & 1)) |
32073 | // ( M ? -X : X) == ((X ^ M ) + (M & 1)) |
32074 | // This lets us transform our vselect to: |
32075 | // (add (xor X, M), (and M, 1)) |
32076 | // And further to: |
32077 | // (sub (xor X, M), M) |
32078 | if (X.getValueType() == MaskVT && Y.getValueType() == MaskVT && |
32079 | DAG.getTargetLoweringInfo().isOperationLegal(ISD::SUB, MaskVT)) { |
32080 | auto IsNegV = [](SDNode *N, SDValue V) { |
32081 | return N->getOpcode() == ISD::SUB && N->getOperand(1) == V && |
32082 | ISD::isBuildVectorAllZeros(N->getOperand(0).getNode()); |
32083 | }; |
32084 | SDValue V; |
32085 | if (IsNegV(Y.getNode(), X)) |
32086 | V = X; |
32087 | else if (IsNegV(X.getNode(), Y)) |
32088 | V = Y; |
32089 | |
32090 | if (V) { |
32091 | SDValue SubOp1 = DAG.getNode(ISD::XOR, DL, MaskVT, V, Mask); |
32092 | SDValue SubOp2 = Mask; |
32093 | |
32094 | // If the negate was on the false side of the select, then |
32095 | // the operands of the SUB need to be swapped. PR 27251. |
32096 | // This is because the pattern being matched above is |
32097 | // (vselect M, (sub (0, X), X) -> (sub (xor X, M), M) |
32098 | // but if the pattern matched was |
32099 | // (vselect M, X, (sub (0, X))), that is really negation of the pattern |
32100 | // above, -(vselect M, (sub 0, X), X), and therefore the replacement |
32101 | // pattern also needs to be a negation of the replacement pattern above. |
32102 | // And -(sub X, Y) is just sub (Y, X), so swapping the operands of the |
32103 | // sub accomplishes the negation of the replacement pattern. |
32104 | if (V == Y) |
32105 | std::swap(SubOp1, SubOp2); |
32106 | |
32107 | SDValue Res = DAG.getNode(ISD::SUB, DL, MaskVT, SubOp1, SubOp2); |
32108 | return DAG.getBitcast(VT, Res); |
32109 | } |
32110 | } |
32111 | |
32112 | // PBLENDVB is only available on SSE 4.1. |
32113 | if (!Subtarget.hasSSE41()) |
32114 | return SDValue(); |
32115 | |
32116 | MVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8; |
32117 | |
32118 | X = DAG.getBitcast(BlendVT, X); |
32119 | Y = DAG.getBitcast(BlendVT, Y); |
32120 | Mask = DAG.getBitcast(BlendVT, Mask); |
32121 | Mask = DAG.getSelect(DL, BlendVT, Mask, Y, X); |
32122 | return DAG.getBitcast(VT, Mask); |
32123 | } |
32124 | |
32125 | // Helper function for combineOrCmpEqZeroToCtlzSrl |
32126 | // Transforms: |
32127 | // seteq(cmp x, 0) |
32128 | // into: |
32129 | // srl(ctlz x), log2(bitsize(x)) |
32130 | // Input pattern is checked by caller. |
32131 | static SDValue lowerX86CmpEqZeroToCtlzSrl(SDValue Op, EVT ExtTy, |
32132 | SelectionDAG &DAG) { |
32133 | SDValue Cmp = Op.getOperand(1); |
32134 | EVT VT = Cmp.getOperand(0).getValueType(); |
32135 | unsigned Log2b = Log2_32(VT.getSizeInBits()); |
32136 | SDLoc dl(Op); |
32137 | SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Cmp->getOperand(0)); |
32138 | // The result of the shift is true or false, and on X86, the 32-bit |
32139 | // encoding of shr and lzcnt is more desirable. |
32140 | SDValue Trunc = DAG.getZExtOrTrunc(Clz, dl, MVT::i32); |
32141 | SDValue Scc = DAG.getNode(ISD::SRL, dl, MVT::i32, Trunc, |
32142 | DAG.getConstant(Log2b, dl, VT)); |
32143 | return DAG.getZExtOrTrunc(Scc, dl, ExtTy); |
32144 | } |
32145 | |
32146 | // Try to transform: |
32147 | // zext(or(setcc(eq, (cmp x, 0)), setcc(eq, (cmp y, 0)))) |
32148 | // into: |
32149 | // srl(or(ctlz(x), ctlz(y)), log2(bitsize(x)) |
32150 | // Will also attempt to match more generic cases, eg: |
32151 | // zext(or(or(setcc(eq, cmp 0), setcc(eq, cmp 0)), setcc(eq, cmp 0))) |
32152 | // Only applies if the target supports the FastLZCNT feature. |
32153 | static SDValue combineOrCmpEqZeroToCtlzSrl(SDNode *N, SelectionDAG &DAG, |
32154 | TargetLowering::DAGCombinerInfo &DCI, |
32155 | const X86Subtarget &Subtarget) { |
32156 | if (DCI.isBeforeLegalize() || !Subtarget.getTargetLowering()->isCtlzFast()) |
32157 | return SDValue(); |
32158 | |
32159 | auto isORCandidate = [](SDValue N) { |
32160 | return (N->getOpcode() == ISD::OR && N->hasOneUse()); |
32161 | }; |
32162 | |
32163 | // Check the zero extend is extending to 32-bit or more. The code generated by |
32164 | // srl(ctlz) for 16-bit or less variants of the pattern would require extra |
32165 | // instructions to clear the upper bits. |
32166 | if (!N->hasOneUse() || !N->getSimpleValueType(0).bitsGE(MVT::i32) || |
32167 | !isORCandidate(N->getOperand(0))) |
32168 | return SDValue(); |
32169 | |
32170 | // Check the node matches: setcc(eq, cmp 0) |
32171 | auto isSetCCCandidate = [](SDValue N) { |
32172 | return N->getOpcode() == X86ISD::SETCC && N->hasOneUse() && |
32173 | X86::CondCode(N->getConstantOperandVal(0)) == X86::COND_E && |
32174 | N->getOperand(1).getOpcode() == X86ISD::CMP && |
32175 | isNullConstant(N->getOperand(1).getOperand(1)) && |
32176 | N->getOperand(1).getValueType().bitsGE(MVT::i32); |
32177 | }; |
32178 | |
32179 | SDNode *OR = N->getOperand(0).getNode(); |
32180 | SDValue LHS = OR->getOperand(0); |
32181 | SDValue RHS = OR->getOperand(1); |
32182 | |
32183 | // Save nodes matching or(or, setcc(eq, cmp 0)). |
32184 | SmallVector<SDNode *, 2> ORNodes; |
32185 | while (((isORCandidate(LHS) && isSetCCCandidate(RHS)) || |
32186 | (isORCandidate(RHS) && isSetCCCandidate(LHS)))) { |
32187 | ORNodes.push_back(OR); |
32188 | OR = (LHS->getOpcode() == ISD::OR) ? LHS.getNode() : RHS.getNode(); |
32189 | LHS = OR->getOperand(0); |
32190 | RHS = OR->getOperand(1); |
32191 | } |
32192 | |
32193 | // The last OR node should match or(setcc(eq, cmp 0), setcc(eq, cmp 0)). |
32194 | if (!(isSetCCCandidate(LHS) && isSetCCCandidate(RHS)) || |
32195 | !isORCandidate(SDValue(OR, 0))) |
32196 | return SDValue(); |
32197 | |
32198 | // We have a or(setcc(eq, cmp 0), setcc(eq, cmp 0)) pattern, try to lower it |
32199 | // to |
32200 | // or(srl(ctlz),srl(ctlz)). |
32201 | // The dag combiner can then fold it into: |
32202 | // srl(or(ctlz, ctlz)). |
32203 | EVT VT = OR->getValueType(0); |
32204 | SDValue NewLHS = lowerX86CmpEqZeroToCtlzSrl(LHS, VT, DAG); |
32205 | SDValue Ret, NewRHS; |
32206 | if (NewLHS && (NewRHS = lowerX86CmpEqZeroToCtlzSrl(RHS, VT, DAG))) |
32207 | Ret = DAG.getNode(ISD::OR, SDLoc(OR), VT, NewLHS, NewRHS); |
32208 | |
32209 | if (!Ret) |
32210 | return SDValue(); |
32211 | |
32212 | // Try to lower nodes matching the or(or, setcc(eq, cmp 0)) pattern. |
32213 | while (ORNodes.size() > 0) { |
32214 | OR = ORNodes.pop_back_val(); |
32215 | LHS = OR->getOperand(0); |
32216 | RHS = OR->getOperand(1); |
32217 | // Swap rhs with lhs to match or(setcc(eq, cmp, 0), or). |
32218 | if (RHS->getOpcode() == ISD::OR) |
32219 | std::swap(LHS, RHS); |
32220 | EVT VT = OR->getValueType(0); |
32221 | SDValue NewRHS = lowerX86CmpEqZeroToCtlzSrl(RHS, VT, DAG); |
32222 | if (!NewRHS) |
32223 | return SDValue(); |
32224 | Ret = DAG.getNode(ISD::OR, SDLoc(OR), VT, Ret, NewRHS); |
32225 | } |
32226 | |
32227 | if (Ret) |
32228 | Ret = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0), Ret); |
32229 | |
32230 | return Ret; |
32231 | } |
32232 | |
32233 | static SDValue combineOr(SDNode *N, SelectionDAG &DAG, |
32234 | TargetLowering::DAGCombinerInfo &DCI, |
32235 | const X86Subtarget &Subtarget) { |
32236 | if (DCI.isBeforeLegalizeOps()) |
32237 | return SDValue(); |
32238 | |
32239 | if (SDValue R = combineCompareEqual(N, DAG, DCI, Subtarget)) |
32240 | return R; |
32241 | |
32242 | if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget)) |
32243 | return FPLogic; |
32244 | |
32245 | if (SDValue R = combineLogicBlendIntoPBLENDV(N, DAG, Subtarget)) |
32246 | return R; |
32247 | |
32248 | SDValue N0 = N->getOperand(0); |
32249 | SDValue N1 = N->getOperand(1); |
32250 | EVT VT = N->getValueType(0); |
32251 | |
32252 | if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64) |
32253 | return SDValue(); |
32254 | |
32255 | // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c) |
32256 | bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize(); |
32257 | |
32258 | // SHLD/SHRD instructions have lower register pressure, but on some |
32259 | // platforms they have higher latency than the equivalent |
32260 | // series of shifts/or that would otherwise be generated. |
32261 | // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions |
32262 | // have higher latencies and we are not optimizing for size. |
32263 | if (!OptForSize && Subtarget.isSHLDSlow()) |
32264 | return SDValue(); |
32265 | |
32266 | if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL) |
32267 | std::swap(N0, N1); |
32268 | if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL) |
32269 | return SDValue(); |
32270 | if (!N0.hasOneUse() || !N1.hasOneUse()) |
32271 | return SDValue(); |
32272 | |
32273 | SDValue ShAmt0 = N0.getOperand(1); |
32274 | if (ShAmt0.getValueType() != MVT::i8) |
32275 | return SDValue(); |
32276 | SDValue ShAmt1 = N1.getOperand(1); |
32277 | if (ShAmt1.getValueType() != MVT::i8) |
32278 | return SDValue(); |
32279 | if (ShAmt0.getOpcode() == ISD::TRUNCATE) |
32280 | ShAmt0 = ShAmt0.getOperand(0); |
32281 | if (ShAmt1.getOpcode() == ISD::TRUNCATE) |
32282 | ShAmt1 = ShAmt1.getOperand(0); |
32283 | |
32284 | SDLoc DL(N); |
32285 | unsigned Opc = X86ISD::SHLD; |
32286 | SDValue Op0 = N0.getOperand(0); |
32287 | SDValue Op1 = N1.getOperand(0); |
32288 | if (ShAmt0.getOpcode() == ISD::SUB || |
32289 | ShAmt0.getOpcode() == ISD::XOR) { |
32290 | Opc = X86ISD::SHRD; |
32291 | std::swap(Op0, Op1); |
32292 | std::swap(ShAmt0, ShAmt1); |
32293 | } |
32294 | |
32295 | // OR( SHL( X, C ), SRL( Y, 32 - C ) ) -> SHLD( X, Y, C ) |
32296 | // OR( SRL( X, C ), SHL( Y, 32 - C ) ) -> SHRD( X, Y, C ) |
32297 | // OR( SHL( X, C ), SRL( SRL( Y, 1 ), XOR( C, 31 ) ) ) -> SHLD( X, Y, C ) |
32298 | // OR( SRL( X, C ), SHL( SHL( Y, 1 ), XOR( C, 31 ) ) ) -> SHRD( X, Y, C ) |
32299 | unsigned Bits = VT.getSizeInBits(); |
32300 | if (ShAmt1.getOpcode() == ISD::SUB) { |
32301 | SDValue Sum = ShAmt1.getOperand(0); |
32302 | if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) { |
32303 | SDValue ShAmt1Op1 = ShAmt1.getOperand(1); |
32304 | if (ShAmt1Op1.getOpcode() == ISD::TRUNCATE) |
32305 | ShAmt1Op1 = ShAmt1Op1.getOperand(0); |
32306 | if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0) |
32307 | return DAG.getNode(Opc, DL, VT, |
32308 | Op0, Op1, |
32309 | DAG.getNode(ISD::TRUNCATE, DL, |
32310 | MVT::i8, ShAmt0)); |
32311 | } |
32312 | } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) { |
32313 | ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0); |
32314 | if (ShAmt0C && (ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue()) == Bits) |
32315 | return DAG.getNode(Opc, DL, VT, |
32316 | N0.getOperand(0), N1.getOperand(0), |
32317 | DAG.getNode(ISD::TRUNCATE, DL, |
32318 | MVT::i8, ShAmt0)); |
32319 | } else if (ShAmt1.getOpcode() == ISD::XOR) { |
32320 | SDValue Mask = ShAmt1.getOperand(1); |
32321 | if (ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask)) { |
32322 | unsigned InnerShift = (X86ISD::SHLD == Opc ? ISD::SRL : ISD::SHL); |
32323 | SDValue ShAmt1Op0 = ShAmt1.getOperand(0); |
32324 | if (ShAmt1Op0.getOpcode() == ISD::TRUNCATE) |
32325 | ShAmt1Op0 = ShAmt1Op0.getOperand(0); |
32326 | if (MaskC->getSExtValue() == (Bits - 1) && ShAmt1Op0 == ShAmt0) { |
32327 | if (Op1.getOpcode() == InnerShift && |
32328 | isa<ConstantSDNode>(Op1.getOperand(1)) && |
32329 | Op1.getConstantOperandVal(1) == 1) { |
32330 | return DAG.getNode(Opc, DL, VT, Op0, Op1.getOperand(0), |
32331 | DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ShAmt0)); |
32332 | } |
32333 | // Test for ADD( Y, Y ) as an equivalent to SHL( Y, 1 ). |
32334 | if (InnerShift == ISD::SHL && Op1.getOpcode() == ISD::ADD && |
32335 | Op1.getOperand(0) == Op1.getOperand(1)) { |
32336 | return DAG.getNode(Opc, DL, VT, Op0, Op1.getOperand(0), |
32337 | DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ShAmt0)); |
32338 | } |
32339 | } |
32340 | } |
32341 | } |
32342 | |
32343 | return SDValue(); |
32344 | } |
32345 | |
32346 | /// Generate NEG and CMOV for integer abs. |
32347 | static SDValue combineIntegerAbs(SDNode *N, SelectionDAG &DAG) { |
32348 | EVT VT = N->getValueType(0); |
32349 | |
32350 | // Since X86 does not have CMOV for 8-bit integer, we don't convert |
32351 | // 8-bit integer abs to NEG and CMOV. |
32352 | if (VT.isInteger() && VT.getSizeInBits() == 8) |
32353 | return SDValue(); |
32354 | |
32355 | SDValue N0 = N->getOperand(0); |
32356 | SDValue N1 = N->getOperand(1); |
32357 | SDLoc DL(N); |
32358 | |
32359 | // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1) |
32360 | // and change it to SUB and CMOV. |
32361 | if (VT.isInteger() && N->getOpcode() == ISD::XOR && |
32362 | N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 && |
32363 | N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0)) { |
32364 | auto *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); |
32365 | if (Y1C && Y1C->getAPIntValue() == VT.getSizeInBits() - 1) { |
32366 | // Generate SUB & CMOV. |
32367 | SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32), |
32368 | DAG.getConstant(0, DL, VT), N0.getOperand(0)); |
32369 | SDValue Ops[] = {N0.getOperand(0), Neg, |
32370 | DAG.getConstant(X86::COND_GE, DL, MVT::i8), |
32371 | SDValue(Neg.getNode(), 1)}; |
32372 | return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops); |
32373 | } |
32374 | } |
32375 | return SDValue(); |
32376 | } |
32377 | |
32378 | /// Try to turn tests against the signbit in the form of: |
32379 | /// XOR(TRUNCATE(SRL(X, size(X)-1)), 1) |
32380 | /// into: |
32381 | /// SETGT(X, -1) |
32382 | static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) { |
32383 | // This is only worth doing if the output type is i8 or i1. |
32384 | EVT ResultType = N->getValueType(0); |
32385 | if (ResultType != MVT::i8 && ResultType != MVT::i1) |
32386 | return SDValue(); |
32387 | |
32388 | SDValue N0 = N->getOperand(0); |
32389 | SDValue N1 = N->getOperand(1); |
32390 | |
32391 | // We should be performing an xor against a truncated shift. |
32392 | if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse()) |
32393 | return SDValue(); |
32394 | |
32395 | // Make sure we are performing an xor against one. |
32396 | if (!isOneConstant(N1)) |
32397 | return SDValue(); |
32398 | |
32399 | // SetCC on x86 zero extends so only act on this if it's a logical shift. |
32400 | SDValue Shift = N0.getOperand(0); |
32401 | if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse()) |
32402 | return SDValue(); |
32403 | |
32404 | // Make sure we are truncating from one of i16, i32 or i64. |
32405 | EVT ShiftTy = Shift.getValueType(); |
32406 | if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64) |
32407 | return SDValue(); |
32408 | |
32409 | // Make sure the shift amount extracts the sign bit. |
32410 | if (!isa<ConstantSDNode>(Shift.getOperand(1)) || |
32411 | Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1) |
32412 | return SDValue(); |
32413 | |
32414 | // Create a greater-than comparison against -1. |
32415 | // N.B. Using SETGE against 0 works but we want a canonical looking |
32416 | // comparison, using SETGT matches up with what TranslateX86CC. |
32417 | SDLoc DL(N); |
32418 | SDValue ShiftOp = Shift.getOperand(0); |
32419 | EVT ShiftOpTy = ShiftOp.getValueType(); |
32420 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
32421 | EVT SetCCResultType = TLI.getSetCCResultType(DAG.getDataLayout(), |
32422 | *DAG.getContext(), ResultType); |
32423 | SDValue Cond = DAG.getSetCC(DL, SetCCResultType, ShiftOp, |
32424 | DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT); |
32425 | if (SetCCResultType != ResultType) |
32426 | Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, ResultType, Cond); |
32427 | return Cond; |
32428 | } |
32429 | |
32430 | /// Turn vector tests of the signbit in the form of: |
32431 | /// xor (sra X, elt_size(X)-1), -1 |
32432 | /// into: |
32433 | /// pcmpgt X, -1 |
32434 | /// |
32435 | /// This should be called before type legalization because the pattern may not |
32436 | /// persist after that. |
32437 | static SDValue foldVectorXorShiftIntoCmp(SDNode *N, SelectionDAG &DAG, |
32438 | const X86Subtarget &Subtarget) { |
32439 | EVT VT = N->getValueType(0); |
32440 | if (!VT.isSimple()) |
32441 | return SDValue(); |
32442 | |
32443 | switch (VT.getSimpleVT().SimpleTy) { |
32444 | default: return SDValue(); |
32445 | case MVT::v16i8: |
32446 | case MVT::v8i16: |
32447 | case MVT::v4i32: if (!Subtarget.hasSSE2()) return SDValue(); break; |
32448 | case MVT::v2i64: if (!Subtarget.hasSSE42()) return SDValue(); break; |
32449 | case MVT::v32i8: |
32450 | case MVT::v16i16: |
32451 | case MVT::v8i32: |
32452 | case MVT::v4i64: if (!Subtarget.hasAVX2()) return SDValue(); break; |
32453 | } |
32454 | |
32455 | // There must be a shift right algebraic before the xor, and the xor must be a |
32456 | // 'not' operation. |
32457 | SDValue Shift = N->getOperand(0); |
32458 | SDValue Ones = N->getOperand(1); |
32459 | if (Shift.getOpcode() != ISD::SRA || !Shift.hasOneUse() || |
32460 | !ISD::isBuildVectorAllOnes(Ones.getNode())) |
32461 | return SDValue(); |
32462 | |
32463 | // The shift should be smearing the sign bit across each vector element. |
32464 | auto *ShiftBV = dyn_cast<BuildVectorSDNode>(Shift.getOperand(1)); |
32465 | if (!ShiftBV) |
32466 | return SDValue(); |
32467 | |
32468 | EVT ShiftEltTy = Shift.getValueType().getVectorElementType(); |
32469 | auto *ShiftAmt = ShiftBV->getConstantSplatNode(); |
32470 | if (!ShiftAmt || ShiftAmt->getZExtValue() != ShiftEltTy.getSizeInBits() - 1) |
32471 | return SDValue(); |
32472 | |
32473 | // Create a greater-than comparison against -1. We don't use the more obvious |
32474 | // greater-than-or-equal-to-zero because SSE/AVX don't have that instruction. |
32475 | return DAG.getNode(X86ISD::PCMPGT, SDLoc(N), VT, Shift.getOperand(0), Ones); |
32476 | } |
32477 | |
32478 | /// Check if truncation with saturation form type \p SrcVT to \p DstVT |
32479 | /// is valid for the given \p Subtarget. |
32480 | static bool isSATValidOnAVX512Subtarget(EVT SrcVT, EVT DstVT, |
32481 | const X86Subtarget &Subtarget) { |
32482 | if (!Subtarget.hasAVX512()) |
32483 | return false; |
32484 | |
32485 | // FIXME: Scalar type may be supported if we move it to vector register. |
32486 | if (!SrcVT.isVector() || !SrcVT.isSimple() || SrcVT.getSizeInBits() > 512) |
32487 | return false; |
32488 | |
32489 | EVT SrcElVT = SrcVT.getScalarType(); |
32490 | EVT DstElVT = DstVT.getScalarType(); |
32491 | if (SrcElVT.getSizeInBits() < 16 || SrcElVT.getSizeInBits() > 64) |
32492 | return false; |
32493 | if (DstElVT.getSizeInBits() < 8 || DstElVT.getSizeInBits() > 32) |
32494 | return false; |
32495 | if (SrcVT.is512BitVector() || Subtarget.hasVLX()) |
32496 | return SrcElVT.getSizeInBits() >= 32 || Subtarget.hasBWI(); |
32497 | return false; |
32498 | } |
32499 | |
32500 | /// Detect a pattern of truncation with saturation: |
32501 | /// (truncate (umin (x, unsigned_max_of_dest_type)) to dest_type). |
32502 | /// Return the source value to be truncated or SDValue() if the pattern was not |
32503 | /// matched. |
32504 | static SDValue detectUSatPattern(SDValue In, EVT VT) { |
32505 | if (In.getOpcode() != ISD::UMIN) |
32506 | return SDValue(); |
32507 | |
32508 | //Saturation with truncation. We truncate from InVT to VT. |
32509 | assert(In.getScalarValueSizeInBits() > VT.getScalarSizeInBits() &&((In.getScalarValueSizeInBits() > VT.getScalarSizeInBits() && "Unexpected types for truncate operation") ? static_cast <void> (0) : __assert_fail ("In.getScalarValueSizeInBits() > VT.getScalarSizeInBits() && \"Unexpected types for truncate operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 32510, __PRETTY_FUNCTION__)) |
32510 | "Unexpected types for truncate operation")((In.getScalarValueSizeInBits() > VT.getScalarSizeInBits() && "Unexpected types for truncate operation") ? static_cast <void> (0) : __assert_fail ("In.getScalarValueSizeInBits() > VT.getScalarSizeInBits() && \"Unexpected types for truncate operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 32510, __PRETTY_FUNCTION__)); |
32511 | |
32512 | APInt C; |
32513 | if (ISD::isConstantSplatVector(In.getOperand(1).getNode(), C)) { |
32514 | // C should be equal to UINT32_MAX / UINT16_MAX / UINT8_MAX according |
32515 | // the element size of the destination type. |
32516 | return C.isMask(VT.getScalarSizeInBits()) ? In.getOperand(0) : |
32517 | SDValue(); |
32518 | } |
32519 | return SDValue(); |
32520 | } |
32521 | |
32522 | /// Detect a pattern of truncation with saturation: |
32523 | /// (truncate (umin (x, unsigned_max_of_dest_type)) to dest_type). |
32524 | /// The types should allow to use VPMOVUS* instruction on AVX512. |
32525 | /// Return the source value to be truncated or SDValue() if the pattern was not |
32526 | /// matched. |
32527 | static SDValue detectAVX512USatPattern(SDValue In, EVT VT, |
32528 | const X86Subtarget &Subtarget) { |
32529 | if (!isSATValidOnAVX512Subtarget(In.getValueType(), VT, Subtarget)) |
32530 | return SDValue(); |
32531 | return detectUSatPattern(In, VT); |
32532 | } |
32533 | |
32534 | static SDValue |
32535 | combineTruncateWithUSat(SDValue In, EVT VT, SDLoc &DL, SelectionDAG &DAG, |
32536 | const X86Subtarget &Subtarget) { |
32537 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
32538 | if (!TLI.isTypeLegal(In.getValueType()) || !TLI.isTypeLegal(VT)) |
32539 | return SDValue(); |
32540 | if (auto USatVal = detectUSatPattern(In, VT)) |
32541 | if (isSATValidOnAVX512Subtarget(In.getValueType(), VT, Subtarget)) |
32542 | return DAG.getNode(X86ISD::VTRUNCUS, DL, VT, USatVal); |
32543 | return SDValue(); |
32544 | } |
32545 | |
32546 | /// This function detects the AVG pattern between vectors of unsigned i8/i16, |
32547 | /// which is c = (a + b + 1) / 2, and replace this operation with the efficient |
32548 | /// X86ISD::AVG instruction. |
32549 | static SDValue detectAVGPattern(SDValue In, EVT VT, SelectionDAG &DAG, |
32550 | const X86Subtarget &Subtarget, |
32551 | const SDLoc &DL) { |
32552 | if (!VT.isVector() || !VT.isSimple()) |
32553 | return SDValue(); |
32554 | EVT InVT = In.getValueType(); |
32555 | unsigned NumElems = VT.getVectorNumElements(); |
32556 | |
32557 | EVT ScalarVT = VT.getVectorElementType(); |
32558 | if (!((ScalarVT == MVT::i8 || ScalarVT == MVT::i16) && |
32559 | isPowerOf2_32(NumElems))) |
32560 | return SDValue(); |
32561 | |
32562 | // InScalarVT is the intermediate type in AVG pattern and it should be greater |
32563 | // than the original input type (i8/i16). |
32564 | EVT InScalarVT = InVT.getVectorElementType(); |
32565 | if (InScalarVT.getSizeInBits() <= ScalarVT.getSizeInBits()) |
32566 | return SDValue(); |
32567 | |
32568 | if (!Subtarget.hasSSE2()) |
32569 | return SDValue(); |
32570 | if (Subtarget.hasBWI()) { |
32571 | if (VT.getSizeInBits() > 512) |
32572 | return SDValue(); |
32573 | } else if (Subtarget.hasAVX2()) { |
32574 | if (VT.getSizeInBits() > 256) |
32575 | return SDValue(); |
32576 | } else { |
32577 | if (VT.getSizeInBits() > 128) |
32578 | return SDValue(); |
32579 | } |
32580 | |
32581 | // Detect the following pattern: |
32582 | // |
32583 | // %1 = zext <N x i8> %a to <N x i32> |
32584 | // %2 = zext <N x i8> %b to <N x i32> |
32585 | // %3 = add nuw nsw <N x i32> %1, <i32 1 x N> |
32586 | // %4 = add nuw nsw <N x i32> %3, %2 |
32587 | // %5 = lshr <N x i32> %N, <i32 1 x N> |
32588 | // %6 = trunc <N x i32> %5 to <N x i8> |
32589 | // |
32590 | // In AVX512, the last instruction can also be a trunc store. |
32591 | |
32592 | if (In.getOpcode() != ISD::SRL) |
32593 | return SDValue(); |
32594 | |
32595 | // A lambda checking the given SDValue is a constant vector and each element |
32596 | // is in the range [Min, Max]. |
32597 | auto IsConstVectorInRange = [](SDValue V, unsigned Min, unsigned Max) { |
32598 | BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(V); |
32599 | if (!BV || !BV->isConstant()) |
32600 | return false; |
32601 | for (SDValue Op : V->ops()) { |
32602 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); |
32603 | if (!C) |
32604 | return false; |
32605 | uint64_t Val = C->getZExtValue(); |
32606 | if (Val < Min || Val > Max) |
32607 | return false; |
32608 | } |
32609 | return true; |
32610 | }; |
32611 | |
32612 | // Check if each element of the vector is left-shifted by one. |
32613 | auto LHS = In.getOperand(0); |
32614 | auto RHS = In.getOperand(1); |
32615 | if (!IsConstVectorInRange(RHS, 1, 1)) |
32616 | return SDValue(); |
32617 | if (LHS.getOpcode() != ISD::ADD) |
32618 | return SDValue(); |
32619 | |
32620 | // Detect a pattern of a + b + 1 where the order doesn't matter. |
32621 | SDValue Operands[3]; |
32622 | Operands[0] = LHS.getOperand(0); |
32623 | Operands[1] = LHS.getOperand(1); |
32624 | |
32625 | // Take care of the case when one of the operands is a constant vector whose |
32626 | // element is in the range [1, 256]. |
32627 | if (IsConstVectorInRange(Operands[1], 1, ScalarVT == MVT::i8 ? 256 : 65536) && |
32628 | Operands[0].getOpcode() == ISD::ZERO_EXTEND && |
32629 | Operands[0].getOperand(0).getValueType() == VT) { |
32630 | // The pattern is detected. Subtract one from the constant vector, then |
32631 | // demote it and emit X86ISD::AVG instruction. |
32632 | SDValue VecOnes = DAG.getConstant(1, DL, InVT); |
32633 | Operands[1] = DAG.getNode(ISD::SUB, DL, InVT, Operands[1], VecOnes); |
32634 | Operands[1] = DAG.getNode(ISD::TRUNCATE, DL, VT, Operands[1]); |
32635 | return DAG.getNode(X86ISD::AVG, DL, VT, Operands[0].getOperand(0), |
32636 | Operands[1]); |
32637 | } |
32638 | |
32639 | if (Operands[0].getOpcode() == ISD::ADD) |
32640 | std::swap(Operands[0], Operands[1]); |
32641 | else if (Operands[1].getOpcode() != ISD::ADD) |
32642 | return SDValue(); |
32643 | Operands[2] = Operands[1].getOperand(0); |
32644 | Operands[1] = Operands[1].getOperand(1); |
32645 | |
32646 | // Now we have three operands of two additions. Check that one of them is a |
32647 | // constant vector with ones, and the other two are promoted from i8/i16. |
32648 | for (int i = 0; i < 3; ++i) { |
32649 | if (!IsConstVectorInRange(Operands[i], 1, 1)) |
32650 | continue; |
32651 | std::swap(Operands[i], Operands[2]); |
32652 | |
32653 | // Check if Operands[0] and Operands[1] are results of type promotion. |
32654 | for (int j = 0; j < 2; ++j) |
32655 | if (Operands[j].getOpcode() != ISD::ZERO_EXTEND || |
32656 | Operands[j].getOperand(0).getValueType() != VT) |
32657 | return SDValue(); |
32658 | |
32659 | // The pattern is detected, emit X86ISD::AVG instruction. |
32660 | return DAG.getNode(X86ISD::AVG, DL, VT, Operands[0].getOperand(0), |
32661 | Operands[1].getOperand(0)); |
32662 | } |
32663 | |
32664 | return SDValue(); |
32665 | } |
32666 | |
32667 | static SDValue combineLoad(SDNode *N, SelectionDAG &DAG, |
32668 | TargetLowering::DAGCombinerInfo &DCI, |
32669 | const X86Subtarget &Subtarget) { |
32670 | LoadSDNode *Ld = cast<LoadSDNode>(N); |
32671 | EVT RegVT = Ld->getValueType(0); |
32672 | EVT MemVT = Ld->getMemoryVT(); |
32673 | SDLoc dl(Ld); |
32674 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
32675 | |
32676 | // For chips with slow 32-byte unaligned loads, break the 32-byte operation |
32677 | // into two 16-byte operations. Also split non-temporal aligned loads on |
32678 | // pre-AVX2 targets as 32-byte loads will lower to regular temporal loads. |
32679 | ISD::LoadExtType Ext = Ld->getExtensionType(); |
32680 | bool Fast; |
32681 | unsigned AddressSpace = Ld->getAddressSpace(); |
32682 | unsigned Alignment = Ld->getAlignment(); |
32683 | if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() && |
32684 | Ext == ISD::NON_EXTLOAD && |
32685 | ((Ld->isNonTemporal() && !Subtarget.hasInt256() && Alignment >= 16) || |
32686 | (TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT, |
32687 | AddressSpace, Alignment, &Fast) && !Fast))) { |
32688 | unsigned NumElems = RegVT.getVectorNumElements(); |
32689 | if (NumElems < 2) |
32690 | return SDValue(); |
32691 | |
32692 | SDValue Ptr = Ld->getBasePtr(); |
32693 | |
32694 | EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(), |
32695 | NumElems/2); |
32696 | SDValue Load1 = |
32697 | DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(), |
32698 | Alignment, Ld->getMemOperand()->getFlags()); |
32699 | |
32700 | Ptr = DAG.getMemBasePlusOffset(Ptr, 16, dl); |
32701 | SDValue Load2 = |
32702 | DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(), |
32703 | std::min(16U, Alignment), Ld->getMemOperand()->getFlags()); |
32704 | SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, |
32705 | Load1.getValue(1), |
32706 | Load2.getValue(1)); |
32707 | |
32708 | SDValue NewVec = DAG.getUNDEF(RegVT); |
32709 | NewVec = insert128BitVector(NewVec, Load1, 0, DAG, dl); |
32710 | NewVec = insert128BitVector(NewVec, Load2, NumElems / 2, DAG, dl); |
32711 | return DCI.CombineTo(N, NewVec, TF, true); |
32712 | } |
32713 | |
32714 | return SDValue(); |
32715 | } |
32716 | |
32717 | /// If V is a build vector of boolean constants and exactly one of those |
32718 | /// constants is true, return the operand index of that true element. |
32719 | /// Otherwise, return -1. |
32720 | static int getOneTrueElt(SDValue V) { |
32721 | // This needs to be a build vector of booleans. |
32722 | // TODO: Checking for the i1 type matches the IR definition for the mask, |
32723 | // but the mask check could be loosened to i8 or other types. That might |
32724 | // also require checking more than 'allOnesValue'; eg, the x86 HW |
32725 | // instructions only require that the MSB is set for each mask element. |
32726 | // The ISD::MSTORE comments/definition do not specify how the mask operand |
32727 | // is formatted. |
32728 | auto *BV = dyn_cast<BuildVectorSDNode>(V); |
32729 | if (!BV || BV->getValueType(0).getVectorElementType() != MVT::i1) |
32730 | return -1; |
32731 | |
32732 | int TrueIndex = -1; |
32733 | unsigned NumElts = BV->getValueType(0).getVectorNumElements(); |
32734 | for (unsigned i = 0; i < NumElts; ++i) { |
32735 | const SDValue &Op = BV->getOperand(i); |
32736 | if (Op.isUndef()) |
32737 | continue; |
32738 | auto *ConstNode = dyn_cast<ConstantSDNode>(Op); |
32739 | if (!ConstNode) |
32740 | return -1; |
32741 | if (ConstNode->getAPIntValue().isAllOnesValue()) { |
32742 | // If we already found a one, this is too many. |
32743 | if (TrueIndex >= 0) |
32744 | return -1; |
32745 | TrueIndex = i; |
32746 | } |
32747 | } |
32748 | return TrueIndex; |
32749 | } |
32750 | |
32751 | /// Given a masked memory load/store operation, return true if it has one mask |
32752 | /// bit set. If it has one mask bit set, then also return the memory address of |
32753 | /// the scalar element to load/store, the vector index to insert/extract that |
32754 | /// scalar element, and the alignment for the scalar memory access. |
32755 | static bool getParamsForOneTrueMaskedElt(MaskedLoadStoreSDNode *MaskedOp, |
32756 | SelectionDAG &DAG, SDValue &Addr, |
32757 | SDValue &Index, unsigned &Alignment) { |
32758 | int TrueMaskElt = getOneTrueElt(MaskedOp->getMask()); |
32759 | if (TrueMaskElt < 0) |
32760 | return false; |
32761 | |
32762 | // Get the address of the one scalar element that is specified by the mask |
32763 | // using the appropriate offset from the base pointer. |
32764 | EVT EltVT = MaskedOp->getMemoryVT().getVectorElementType(); |
32765 | Addr = MaskedOp->getBasePtr(); |
32766 | if (TrueMaskElt != 0) { |
32767 | unsigned Offset = TrueMaskElt * EltVT.getStoreSize(); |
32768 | Addr = DAG.getMemBasePlusOffset(Addr, Offset, SDLoc(MaskedOp)); |
32769 | } |
32770 | |
32771 | Index = DAG.getIntPtrConstant(TrueMaskElt, SDLoc(MaskedOp)); |
32772 | Alignment = MinAlign(MaskedOp->getAlignment(), EltVT.getStoreSize()); |
32773 | return true; |
32774 | } |
32775 | |
32776 | /// If exactly one element of the mask is set for a non-extending masked load, |
32777 | /// it is a scalar load and vector insert. |
32778 | /// Note: It is expected that the degenerate cases of an all-zeros or all-ones |
32779 | /// mask have already been optimized in IR, so we don't bother with those here. |
32780 | static SDValue |
32781 | reduceMaskedLoadToScalarLoad(MaskedLoadSDNode *ML, SelectionDAG &DAG, |
32782 | TargetLowering::DAGCombinerInfo &DCI) { |
32783 | // TODO: This is not x86-specific, so it could be lifted to DAGCombiner. |
32784 | // However, some target hooks may need to be added to know when the transform |
32785 | // is profitable. Endianness would also have to be considered. |
32786 | |
32787 | SDValue Addr, VecIndex; |
32788 | unsigned Alignment; |
32789 | if (!getParamsForOneTrueMaskedElt(ML, DAG, Addr, VecIndex, Alignment)) |
32790 | return SDValue(); |
32791 | |
32792 | // Load the one scalar element that is specified by the mask using the |
32793 | // appropriate offset from the base pointer. |
32794 | SDLoc DL(ML); |
32795 | EVT VT = ML->getValueType(0); |
32796 | EVT EltVT = VT.getVectorElementType(); |
32797 | SDValue Load = |
32798 | DAG.getLoad(EltVT, DL, ML->getChain(), Addr, ML->getPointerInfo(), |
32799 | Alignment, ML->getMemOperand()->getFlags()); |
32800 | |
32801 | // Insert the loaded element into the appropriate place in the vector. |
32802 | SDValue Insert = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, ML->getSrc0(), |
32803 | Load, VecIndex); |
32804 | return DCI.CombineTo(ML, Insert, Load.getValue(1), true); |
32805 | } |
32806 | |
32807 | static SDValue |
32808 | combineMaskedLoadConstantMask(MaskedLoadSDNode *ML, SelectionDAG &DAG, |
32809 | TargetLowering::DAGCombinerInfo &DCI) { |
32810 | if (!ISD::isBuildVectorOfConstantSDNodes(ML->getMask().getNode())) |
32811 | return SDValue(); |
32812 | |
32813 | SDLoc DL(ML); |
32814 | EVT VT = ML->getValueType(0); |
32815 | |
32816 | // If we are loading the first and last elements of a vector, it is safe and |
32817 | // always faster to load the whole vector. Replace the masked load with a |
32818 | // vector load and select. |
32819 | unsigned NumElts = VT.getVectorNumElements(); |
32820 | BuildVectorSDNode *MaskBV = cast<BuildVectorSDNode>(ML->getMask()); |
32821 | bool LoadFirstElt = !isNullConstant(MaskBV->getOperand(0)); |
32822 | bool LoadLastElt = !isNullConstant(MaskBV->getOperand(NumElts - 1)); |
32823 | if (LoadFirstElt && LoadLastElt) { |
32824 | SDValue VecLd = DAG.getLoad(VT, DL, ML->getChain(), ML->getBasePtr(), |
32825 | ML->getMemOperand()); |
32826 | SDValue Blend = DAG.getSelect(DL, VT, ML->getMask(), VecLd, ML->getSrc0()); |
32827 | return DCI.CombineTo(ML, Blend, VecLd.getValue(1), true); |
32828 | } |
32829 | |
32830 | // Convert a masked load with a constant mask into a masked load and a select. |
32831 | // This allows the select operation to use a faster kind of select instruction |
32832 | // (for example, vblendvps -> vblendps). |
32833 | |
32834 | // Don't try this if the pass-through operand is already undefined. That would |
32835 | // cause an infinite loop because that's what we're about to create. |
32836 | if (ML->getSrc0().isUndef()) |
32837 | return SDValue(); |
32838 | |
32839 | // The new masked load has an undef pass-through operand. The select uses the |
32840 | // original pass-through operand. |
32841 | SDValue NewML = DAG.getMaskedLoad(VT, DL, ML->getChain(), ML->getBasePtr(), |
32842 | ML->getMask(), DAG.getUNDEF(VT), |
32843 | ML->getMemoryVT(), ML->getMemOperand(), |
32844 | ML->getExtensionType()); |
32845 | SDValue Blend = DAG.getSelect(DL, VT, ML->getMask(), NewML, ML->getSrc0()); |
32846 | |
32847 | return DCI.CombineTo(ML, Blend, NewML.getValue(1), true); |
32848 | } |
32849 | |
32850 | static SDValue combineMaskedLoad(SDNode *N, SelectionDAG &DAG, |
32851 | TargetLowering::DAGCombinerInfo &DCI, |
32852 | const X86Subtarget &Subtarget) { |
32853 | MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N); |
32854 | |
32855 | // TODO: Expanding load with constant mask may be optimized as well. |
32856 | if (Mld->isExpandingLoad()) |
32857 | return SDValue(); |
32858 | |
32859 | if (Mld->getExtensionType() == ISD::NON_EXTLOAD) { |
32860 | if (SDValue ScalarLoad = reduceMaskedLoadToScalarLoad(Mld, DAG, DCI)) |
32861 | return ScalarLoad; |
32862 | // TODO: Do some AVX512 subsets benefit from this transform? |
32863 | if (!Subtarget.hasAVX512()) |
32864 | if (SDValue Blend = combineMaskedLoadConstantMask(Mld, DAG, DCI)) |
32865 | return Blend; |
32866 | } |
32867 | |
32868 | if (Mld->getExtensionType() != ISD::SEXTLOAD) |
32869 | return SDValue(); |
32870 | |
32871 | // Resolve extending loads. |
32872 | EVT VT = Mld->getValueType(0); |
32873 | unsigned NumElems = VT.getVectorNumElements(); |
32874 | EVT LdVT = Mld->getMemoryVT(); |
32875 | SDLoc dl(Mld); |
32876 | |
32877 | assert(LdVT != VT && "Cannot extend to the same type")((LdVT != VT && "Cannot extend to the same type") ? static_cast <void> (0) : __assert_fail ("LdVT != VT && \"Cannot extend to the same type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 32877, __PRETTY_FUNCTION__)); |
32878 | unsigned ToSz = VT.getScalarSizeInBits(); |
32879 | unsigned FromSz = LdVT.getScalarSizeInBits(); |
32880 | // From/To sizes and ElemCount must be pow of two. |
32881 | assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&((isPowerOf2_32(NumElems * FromSz * ToSz) && "Unexpected size for extending masked load" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(NumElems * FromSz * ToSz) && \"Unexpected size for extending masked load\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 32882, __PRETTY_FUNCTION__)) |
32882 | "Unexpected size for extending masked load")((isPowerOf2_32(NumElems * FromSz * ToSz) && "Unexpected size for extending masked load" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(NumElems * FromSz * ToSz) && \"Unexpected size for extending masked load\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 32882, __PRETTY_FUNCTION__)); |
32883 | |
32884 | unsigned SizeRatio = ToSz / FromSz; |
32885 | assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits())((SizeRatio * NumElems * FromSz == VT.getSizeInBits()) ? static_cast <void> (0) : __assert_fail ("SizeRatio * NumElems * FromSz == VT.getSizeInBits()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 32885, __PRETTY_FUNCTION__)); |
32886 | |
32887 | // Create a type on which we perform the shuffle. |
32888 | EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), |
32889 | LdVT.getScalarType(), NumElems*SizeRatio); |
32890 | assert(WideVecVT.getSizeInBits() == VT.getSizeInBits())((WideVecVT.getSizeInBits() == VT.getSizeInBits()) ? static_cast <void> (0) : __assert_fail ("WideVecVT.getSizeInBits() == VT.getSizeInBits()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 32890, __PRETTY_FUNCTION__)); |
32891 | |
32892 | // Convert Src0 value. |
32893 | SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0()); |
32894 | if (!Mld->getSrc0().isUndef()) { |
32895 | SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1); |
32896 | for (unsigned i = 0; i != NumElems; ++i) |
32897 | ShuffleVec[i] = i * SizeRatio; |
32898 | |
32899 | // Can't shuffle using an illegal type. |
32900 | assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&((DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) && "WideVecVT should be legal") ? static_cast<void> (0) : __assert_fail ("DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) && \"WideVecVT should be legal\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 32901, __PRETTY_FUNCTION__)) |
32901 | "WideVecVT should be legal")((DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) && "WideVecVT should be legal") ? static_cast<void> (0) : __assert_fail ("DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) && \"WideVecVT should be legal\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 32901, __PRETTY_FUNCTION__)); |
32902 | WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0, |
32903 | DAG.getUNDEF(WideVecVT), ShuffleVec); |
32904 | } |
32905 | // Prepare the new mask. |
32906 | SDValue NewMask; |
32907 | SDValue Mask = Mld->getMask(); |
32908 | if (Mask.getValueType() == VT) { |
32909 | // Mask and original value have the same type. |
32910 | NewMask = DAG.getBitcast(WideVecVT, Mask); |
32911 | SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1); |
32912 | for (unsigned i = 0; i != NumElems; ++i) |
32913 | ShuffleVec[i] = i * SizeRatio; |
32914 | for (unsigned i = NumElems; i != NumElems * SizeRatio; ++i) |
32915 | ShuffleVec[i] = NumElems * SizeRatio; |
32916 | NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask, |
32917 | DAG.getConstant(0, dl, WideVecVT), |
32918 | ShuffleVec); |
32919 | } else { |
32920 | assert(Mask.getValueType().getVectorElementType() == MVT::i1)((Mask.getValueType().getVectorElementType() == MVT::i1) ? static_cast <void> (0) : __assert_fail ("Mask.getValueType().getVectorElementType() == MVT::i1" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 32920, __PRETTY_FUNCTION__)); |
32921 | unsigned WidenNumElts = NumElems*SizeRatio; |
32922 | unsigned MaskNumElts = VT.getVectorNumElements(); |
32923 | EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, |
32924 | WidenNumElts); |
32925 | |
32926 | unsigned NumConcat = WidenNumElts / MaskNumElts; |
32927 | SmallVector<SDValue, 16> Ops(NumConcat); |
32928 | SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType()); |
32929 | Ops[0] = Mask; |
32930 | for (unsigned i = 1; i != NumConcat; ++i) |
32931 | Ops[i] = ZeroVal; |
32932 | |
32933 | NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops); |
32934 | } |
32935 | |
32936 | SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(), |
32937 | Mld->getBasePtr(), NewMask, WideSrc0, |
32938 | Mld->getMemoryVT(), Mld->getMemOperand(), |
32939 | ISD::NON_EXTLOAD); |
32940 | SDValue NewVec = getExtendInVec(X86ISD::VSEXT, dl, VT, WideLd, DAG); |
32941 | return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true); |
32942 | } |
32943 | |
32944 | /// If exactly one element of the mask is set for a non-truncating masked store, |
32945 | /// it is a vector extract and scalar store. |
32946 | /// Note: It is expected that the degenerate cases of an all-zeros or all-ones |
32947 | /// mask have already been optimized in IR, so we don't bother with those here. |
32948 | static SDValue reduceMaskedStoreToScalarStore(MaskedStoreSDNode *MS, |
32949 | SelectionDAG &DAG) { |
32950 | // TODO: This is not x86-specific, so it could be lifted to DAGCombiner. |
32951 | // However, some target hooks may need to be added to know when the transform |
32952 | // is profitable. Endianness would also have to be considered. |
32953 | |
32954 | SDValue Addr, VecIndex; |
32955 | unsigned Alignment; |
32956 | if (!getParamsForOneTrueMaskedElt(MS, DAG, Addr, VecIndex, Alignment)) |
32957 | return SDValue(); |
32958 | |
32959 | // Extract the one scalar element that is actually being stored. |
32960 | SDLoc DL(MS); |
32961 | EVT VT = MS->getValue().getValueType(); |
32962 | EVT EltVT = VT.getVectorElementType(); |
32963 | SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, |
32964 | MS->getValue(), VecIndex); |
32965 | |
32966 | // Store that element at the appropriate offset from the base pointer. |
32967 | return DAG.getStore(MS->getChain(), DL, Extract, Addr, MS->getPointerInfo(), |
32968 | Alignment, MS->getMemOperand()->getFlags()); |
32969 | } |
32970 | |
32971 | static SDValue combineMaskedStore(SDNode *N, SelectionDAG &DAG, |
32972 | const X86Subtarget &Subtarget) { |
32973 | MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N); |
32974 | |
32975 | if (Mst->isCompressingStore()) |
32976 | return SDValue(); |
32977 | |
32978 | if (!Mst->isTruncatingStore()) |
32979 | return reduceMaskedStoreToScalarStore(Mst, DAG); |
32980 | |
32981 | // Resolve truncating stores. |
32982 | EVT VT = Mst->getValue().getValueType(); |
32983 | unsigned NumElems = VT.getVectorNumElements(); |
32984 | EVT StVT = Mst->getMemoryVT(); |
32985 | SDLoc dl(Mst); |
32986 | |
32987 | assert(StVT != VT && "Cannot truncate to the same type")((StVT != VT && "Cannot truncate to the same type") ? static_cast<void> (0) : __assert_fail ("StVT != VT && \"Cannot truncate to the same type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 32987, __PRETTY_FUNCTION__)); |
32988 | unsigned FromSz = VT.getScalarSizeInBits(); |
32989 | unsigned ToSz = StVT.getScalarSizeInBits(); |
32990 | |
32991 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
32992 | |
32993 | // The truncating store is legal in some cases. For example |
32994 | // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw |
32995 | // are designated for truncate store. |
32996 | // In this case we don't need any further transformations. |
32997 | if (TLI.isTruncStoreLegal(VT, StVT)) |
32998 | return SDValue(); |
32999 | |
33000 | // From/To sizes and ElemCount must be pow of two. |
33001 | assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&((isPowerOf2_32(NumElems * FromSz * ToSz) && "Unexpected size for truncating masked store" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(NumElems * FromSz * ToSz) && \"Unexpected size for truncating masked store\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33002, __PRETTY_FUNCTION__)) |
33002 | "Unexpected size for truncating masked store")((isPowerOf2_32(NumElems * FromSz * ToSz) && "Unexpected size for truncating masked store" ) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(NumElems * FromSz * ToSz) && \"Unexpected size for truncating masked store\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33002, __PRETTY_FUNCTION__)); |
33003 | // We are going to use the original vector elt for storing. |
33004 | // Accumulated smaller vector elements must be a multiple of the store size. |
33005 | assert (((NumElems * FromSz) % ToSz) == 0 &&((((NumElems * FromSz) % ToSz) == 0 && "Unexpected ratio for truncating masked store" ) ? static_cast<void> (0) : __assert_fail ("((NumElems * FromSz) % ToSz) == 0 && \"Unexpected ratio for truncating masked store\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33006, __PRETTY_FUNCTION__)) |
33006 | "Unexpected ratio for truncating masked store")((((NumElems * FromSz) % ToSz) == 0 && "Unexpected ratio for truncating masked store" ) ? static_cast<void> (0) : __assert_fail ("((NumElems * FromSz) % ToSz) == 0 && \"Unexpected ratio for truncating masked store\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33006, __PRETTY_FUNCTION__)); |
33007 | |
33008 | unsigned SizeRatio = FromSz / ToSz; |
33009 | assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits())((SizeRatio * NumElems * ToSz == VT.getSizeInBits()) ? static_cast <void> (0) : __assert_fail ("SizeRatio * NumElems * ToSz == VT.getSizeInBits()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33009, __PRETTY_FUNCTION__)); |
33010 | |
33011 | // Create a type on which we perform the shuffle. |
33012 | EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), |
33013 | StVT.getScalarType(), NumElems*SizeRatio); |
33014 | |
33015 | assert(WideVecVT.getSizeInBits() == VT.getSizeInBits())((WideVecVT.getSizeInBits() == VT.getSizeInBits()) ? static_cast <void> (0) : __assert_fail ("WideVecVT.getSizeInBits() == VT.getSizeInBits()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33015, __PRETTY_FUNCTION__)); |
33016 | |
33017 | SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue()); |
33018 | SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1); |
33019 | for (unsigned i = 0; i != NumElems; ++i) |
33020 | ShuffleVec[i] = i * SizeRatio; |
33021 | |
33022 | // Can't shuffle using an illegal type. |
33023 | assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&((DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) && "WideVecVT should be legal") ? static_cast<void> (0) : __assert_fail ("DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) && \"WideVecVT should be legal\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33024, __PRETTY_FUNCTION__)) |
33024 | "WideVecVT should be legal")((DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) && "WideVecVT should be legal") ? static_cast<void> (0) : __assert_fail ("DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) && \"WideVecVT should be legal\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33024, __PRETTY_FUNCTION__)); |
33025 | |
33026 | SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec, |
33027 | DAG.getUNDEF(WideVecVT), |
33028 | ShuffleVec); |
33029 | |
33030 | SDValue NewMask; |
33031 | SDValue Mask = Mst->getMask(); |
33032 | if (Mask.getValueType() == VT) { |
33033 | // Mask and original value have the same type. |
33034 | NewMask = DAG.getBitcast(WideVecVT, Mask); |
33035 | for (unsigned i = 0; i != NumElems; ++i) |
33036 | ShuffleVec[i] = i * SizeRatio; |
33037 | for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i) |
33038 | ShuffleVec[i] = NumElems*SizeRatio; |
33039 | NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask, |
33040 | DAG.getConstant(0, dl, WideVecVT), |
33041 | ShuffleVec); |
33042 | } else { |
33043 | assert(Mask.getValueType().getVectorElementType() == MVT::i1)((Mask.getValueType().getVectorElementType() == MVT::i1) ? static_cast <void> (0) : __assert_fail ("Mask.getValueType().getVectorElementType() == MVT::i1" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33043, __PRETTY_FUNCTION__)); |
33044 | unsigned WidenNumElts = NumElems*SizeRatio; |
33045 | unsigned MaskNumElts = VT.getVectorNumElements(); |
33046 | EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, |
33047 | WidenNumElts); |
33048 | |
33049 | unsigned NumConcat = WidenNumElts / MaskNumElts; |
33050 | SmallVector<SDValue, 16> Ops(NumConcat); |
33051 | SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType()); |
33052 | Ops[0] = Mask; |
33053 | for (unsigned i = 1; i != NumConcat; ++i) |
33054 | Ops[i] = ZeroVal; |
33055 | |
33056 | NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops); |
33057 | } |
33058 | |
33059 | return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, |
33060 | Mst->getBasePtr(), NewMask, StVT, |
33061 | Mst->getMemOperand(), false); |
33062 | } |
33063 | |
33064 | static SDValue combineStore(SDNode *N, SelectionDAG &DAG, |
33065 | const X86Subtarget &Subtarget) { |
33066 | StoreSDNode *St = cast<StoreSDNode>(N); |
33067 | EVT VT = St->getValue().getValueType(); |
33068 | EVT StVT = St->getMemoryVT(); |
33069 | SDLoc dl(St); |
33070 | SDValue StoredVal = St->getOperand(1); |
33071 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
33072 | |
33073 | // If we are saving a concatenation of two XMM registers and 32-byte stores |
33074 | // are slow, such as on Sandy Bridge, perform two 16-byte stores. |
33075 | bool Fast; |
33076 | unsigned AddressSpace = St->getAddressSpace(); |
33077 | unsigned Alignment = St->getAlignment(); |
33078 | if (VT.is256BitVector() && StVT == VT && |
33079 | TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT, |
33080 | AddressSpace, Alignment, &Fast) && |
33081 | !Fast) { |
33082 | unsigned NumElems = VT.getVectorNumElements(); |
33083 | if (NumElems < 2) |
33084 | return SDValue(); |
33085 | |
33086 | SDValue Value0 = extract128BitVector(StoredVal, 0, DAG, dl); |
33087 | SDValue Value1 = extract128BitVector(StoredVal, NumElems / 2, DAG, dl); |
33088 | |
33089 | SDValue Ptr0 = St->getBasePtr(); |
33090 | SDValue Ptr1 = DAG.getMemBasePlusOffset(Ptr0, 16, dl); |
33091 | |
33092 | SDValue Ch0 = |
33093 | DAG.getStore(St->getChain(), dl, Value0, Ptr0, St->getPointerInfo(), |
33094 | Alignment, St->getMemOperand()->getFlags()); |
33095 | SDValue Ch1 = |
33096 | DAG.getStore(St->getChain(), dl, Value1, Ptr1, St->getPointerInfo(), |
33097 | std::min(16U, Alignment), St->getMemOperand()->getFlags()); |
33098 | return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1); |
33099 | } |
33100 | |
33101 | // Optimize trunc store (of multiple scalars) to shuffle and store. |
33102 | // First, pack all of the elements in one place. Next, store to memory |
33103 | // in fewer chunks. |
33104 | if (St->isTruncatingStore() && VT.isVector()) { |
33105 | // Check if we can detect an AVG pattern from the truncation. If yes, |
33106 | // replace the trunc store by a normal store with the result of X86ISD::AVG |
33107 | // instruction. |
33108 | if (SDValue Avg = detectAVGPattern(St->getValue(), St->getMemoryVT(), DAG, |
33109 | Subtarget, dl)) |
33110 | return DAG.getStore(St->getChain(), dl, Avg, St->getBasePtr(), |
33111 | St->getPointerInfo(), St->getAlignment(), |
33112 | St->getMemOperand()->getFlags()); |
33113 | |
33114 | if (SDValue Val = |
33115 | detectAVX512USatPattern(St->getValue(), St->getMemoryVT(), Subtarget)) |
33116 | return EmitTruncSStore(false /* Unsigned saturation */, St->getChain(), |
33117 | dl, Val, St->getBasePtr(), |
33118 | St->getMemoryVT(), St->getMemOperand(), DAG); |
33119 | |
33120 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
33121 | unsigned NumElems = VT.getVectorNumElements(); |
33122 | assert(StVT != VT && "Cannot truncate to the same type")((StVT != VT && "Cannot truncate to the same type") ? static_cast<void> (0) : __assert_fail ("StVT != VT && \"Cannot truncate to the same type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33122, __PRETTY_FUNCTION__)); |
33123 | unsigned FromSz = VT.getScalarSizeInBits(); |
33124 | unsigned ToSz = StVT.getScalarSizeInBits(); |
33125 | |
33126 | // The truncating store is legal in some cases. For example |
33127 | // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw |
33128 | // are designated for truncate store. |
33129 | // In this case we don't need any further transformations. |
33130 | if (TLI.isTruncStoreLegalOrCustom(VT, StVT)) |
33131 | return SDValue(); |
33132 | |
33133 | // From, To sizes and ElemCount must be pow of two |
33134 | if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue(); |
33135 | // We are going to use the original vector elt for storing. |
33136 | // Accumulated smaller vector elements must be a multiple of the store size. |
33137 | if (0 != (NumElems * FromSz) % ToSz) return SDValue(); |
33138 | |
33139 | unsigned SizeRatio = FromSz / ToSz; |
33140 | |
33141 | assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits())((SizeRatio * NumElems * ToSz == VT.getSizeInBits()) ? static_cast <void> (0) : __assert_fail ("SizeRatio * NumElems * ToSz == VT.getSizeInBits()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33141, __PRETTY_FUNCTION__)); |
33142 | |
33143 | // Create a type on which we perform the shuffle |
33144 | EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), |
33145 | StVT.getScalarType(), NumElems*SizeRatio); |
33146 | |
33147 | assert(WideVecVT.getSizeInBits() == VT.getSizeInBits())((WideVecVT.getSizeInBits() == VT.getSizeInBits()) ? static_cast <void> (0) : __assert_fail ("WideVecVT.getSizeInBits() == VT.getSizeInBits()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33147, __PRETTY_FUNCTION__)); |
33148 | |
33149 | SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue()); |
33150 | SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1); |
33151 | for (unsigned i = 0; i != NumElems; ++i) |
33152 | ShuffleVec[i] = i * SizeRatio; |
33153 | |
33154 | // Can't shuffle using an illegal type. |
33155 | if (!TLI.isTypeLegal(WideVecVT)) |
33156 | return SDValue(); |
33157 | |
33158 | SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec, |
33159 | DAG.getUNDEF(WideVecVT), |
33160 | ShuffleVec); |
33161 | // At this point all of the data is stored at the bottom of the |
33162 | // register. We now need to save it to mem. |
33163 | |
33164 | // Find the largest store unit |
33165 | MVT StoreType = MVT::i8; |
33166 | for (MVT Tp : MVT::integer_valuetypes()) { |
33167 | if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz) |
33168 | StoreType = Tp; |
33169 | } |
33170 | |
33171 | // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64. |
33172 | if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 && |
33173 | (64 <= NumElems * ToSz)) |
33174 | StoreType = MVT::f64; |
33175 | |
33176 | // Bitcast the original vector into a vector of store-size units |
33177 | EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(), |
33178 | StoreType, VT.getSizeInBits()/StoreType.getSizeInBits()); |
33179 | assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits())((StoreVecVT.getSizeInBits() == VT.getSizeInBits()) ? static_cast <void> (0) : __assert_fail ("StoreVecVT.getSizeInBits() == VT.getSizeInBits()" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33179, __PRETTY_FUNCTION__)); |
33180 | SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff); |
33181 | SmallVector<SDValue, 8> Chains; |
33182 | SDValue Ptr = St->getBasePtr(); |
33183 | |
33184 | // Perform one or more big stores into memory. |
33185 | for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) { |
33186 | SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, |
33187 | StoreType, ShuffWide, |
33188 | DAG.getIntPtrConstant(i, dl)); |
33189 | SDValue Ch = |
33190 | DAG.getStore(St->getChain(), dl, SubVec, Ptr, St->getPointerInfo(), |
33191 | St->getAlignment(), St->getMemOperand()->getFlags()); |
33192 | Ptr = DAG.getMemBasePlusOffset(Ptr, StoreType.getStoreSize(), dl); |
33193 | Chains.push_back(Ch); |
33194 | } |
33195 | |
33196 | return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains); |
33197 | } |
33198 | |
33199 | // Turn load->store of MMX types into GPR load/stores. This avoids clobbering |
33200 | // the FP state in cases where an emms may be missing. |
33201 | // A preferable solution to the general problem is to figure out the right |
33202 | // places to insert EMMS. This qualifies as a quick hack. |
33203 | |
33204 | // Similarly, turn load->store of i64 into double load/stores in 32-bit mode. |
33205 | if (VT.getSizeInBits() != 64) |
33206 | return SDValue(); |
33207 | |
33208 | const Function *F = DAG.getMachineFunction().getFunction(); |
33209 | bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat); |
33210 | bool F64IsLegal = |
33211 | !Subtarget.useSoftFloat() && !NoImplicitFloatOps && Subtarget.hasSSE2(); |
33212 | if ((VT.isVector() || |
33213 | (VT == MVT::i64 && F64IsLegal && !Subtarget.is64Bit())) && |
33214 | isa<LoadSDNode>(St->getValue()) && |
33215 | !cast<LoadSDNode>(St->getValue())->isVolatile() && |
33216 | St->getChain().hasOneUse() && !St->isVolatile()) { |
33217 | SDNode* LdVal = St->getValue().getNode(); |
33218 | LoadSDNode *Ld = nullptr; |
33219 | int TokenFactorIndex = -1; |
33220 | SmallVector<SDValue, 8> Ops; |
33221 | SDNode* ChainVal = St->getChain().getNode(); |
33222 | // Must be a store of a load. We currently handle two cases: the load |
33223 | // is a direct child, and it's under an intervening TokenFactor. It is |
33224 | // possible to dig deeper under nested TokenFactors. |
33225 | if (ChainVal == LdVal) |
33226 | Ld = cast<LoadSDNode>(St->getChain()); |
33227 | else if (St->getValue().hasOneUse() && |
33228 | ChainVal->getOpcode() == ISD::TokenFactor) { |
33229 | for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) { |
33230 | if (ChainVal->getOperand(i).getNode() == LdVal) { |
33231 | TokenFactorIndex = i; |
33232 | Ld = cast<LoadSDNode>(St->getValue()); |
33233 | } else |
33234 | Ops.push_back(ChainVal->getOperand(i)); |
33235 | } |
33236 | } |
33237 | |
33238 | if (!Ld || !ISD::isNormalLoad(Ld)) |
33239 | return SDValue(); |
33240 | |
33241 | // If this is not the MMX case, i.e. we are just turning i64 load/store |
33242 | // into f64 load/store, avoid the transformation if there are multiple |
33243 | // uses of the loaded value. |
33244 | if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0)) |
33245 | return SDValue(); |
33246 | |
33247 | SDLoc LdDL(Ld); |
33248 | SDLoc StDL(N); |
33249 | // If we are a 64-bit capable x86, lower to a single movq load/store pair. |
33250 | // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store |
33251 | // pair instead. |
33252 | if (Subtarget.is64Bit() || F64IsLegal) { |
33253 | MVT LdVT = Subtarget.is64Bit() ? MVT::i64 : MVT::f64; |
33254 | SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(), |
33255 | Ld->getPointerInfo(), Ld->getAlignment(), |
33256 | Ld->getMemOperand()->getFlags()); |
33257 | SDValue NewChain = NewLd.getValue(1); |
33258 | if (TokenFactorIndex >= 0) { |
33259 | Ops.push_back(NewChain); |
33260 | NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops); |
33261 | } |
33262 | return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(), |
33263 | St->getPointerInfo(), St->getAlignment(), |
33264 | St->getMemOperand()->getFlags()); |
33265 | } |
33266 | |
33267 | // Otherwise, lower to two pairs of 32-bit loads / stores. |
33268 | SDValue LoAddr = Ld->getBasePtr(); |
33269 | SDValue HiAddr = DAG.getMemBasePlusOffset(LoAddr, 4, LdDL); |
33270 | |
33271 | SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr, |
33272 | Ld->getPointerInfo(), Ld->getAlignment(), |
33273 | Ld->getMemOperand()->getFlags()); |
33274 | SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr, |
33275 | Ld->getPointerInfo().getWithOffset(4), |
33276 | MinAlign(Ld->getAlignment(), 4), |
33277 | Ld->getMemOperand()->getFlags()); |
33278 | |
33279 | SDValue NewChain = LoLd.getValue(1); |
33280 | if (TokenFactorIndex >= 0) { |
33281 | Ops.push_back(LoLd); |
33282 | Ops.push_back(HiLd); |
33283 | NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops); |
33284 | } |
33285 | |
33286 | LoAddr = St->getBasePtr(); |
33287 | HiAddr = DAG.getMemBasePlusOffset(LoAddr, 4, StDL); |
33288 | |
33289 | SDValue LoSt = |
33290 | DAG.getStore(NewChain, StDL, LoLd, LoAddr, St->getPointerInfo(), |
33291 | St->getAlignment(), St->getMemOperand()->getFlags()); |
33292 | SDValue HiSt = DAG.getStore( |
33293 | NewChain, StDL, HiLd, HiAddr, St->getPointerInfo().getWithOffset(4), |
33294 | MinAlign(St->getAlignment(), 4), St->getMemOperand()->getFlags()); |
33295 | return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt); |
33296 | } |
33297 | |
33298 | // This is similar to the above case, but here we handle a scalar 64-bit |
33299 | // integer store that is extracted from a vector on a 32-bit target. |
33300 | // If we have SSE2, then we can treat it like a floating-point double |
33301 | // to get past legalization. The execution dependencies fixup pass will |
33302 | // choose the optimal machine instruction for the store if this really is |
33303 | // an integer or v2f32 rather than an f64. |
33304 | if (VT == MVT::i64 && F64IsLegal && !Subtarget.is64Bit() && |
33305 | St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) { |
33306 | SDValue OldExtract = St->getOperand(1); |
33307 | SDValue ExtOp0 = OldExtract.getOperand(0); |
33308 | unsigned VecSize = ExtOp0.getValueSizeInBits(); |
33309 | EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64); |
33310 | SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0); |
33311 | SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, |
33312 | BitCast, OldExtract.getOperand(1)); |
33313 | return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(), |
33314 | St->getPointerInfo(), St->getAlignment(), |
33315 | St->getMemOperand()->getFlags()); |
33316 | } |
33317 | |
33318 | return SDValue(); |
33319 | } |
33320 | |
33321 | /// Return 'true' if this vector operation is "horizontal" |
33322 | /// and return the operands for the horizontal operation in LHS and RHS. A |
33323 | /// horizontal operation performs the binary operation on successive elements |
33324 | /// of its first operand, then on successive elements of its second operand, |
33325 | /// returning the resulting values in a vector. For example, if |
33326 | /// A = < float a0, float a1, float a2, float a3 > |
33327 | /// and |
33328 | /// B = < float b0, float b1, float b2, float b3 > |
33329 | /// then the result of doing a horizontal operation on A and B is |
33330 | /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >. |
33331 | /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form |
33332 | /// A horizontal-op B, for some already available A and B, and if so then LHS is |
33333 | /// set to A, RHS to B, and the routine returns 'true'. |
33334 | /// Note that the binary operation should have the property that if one of the |
33335 | /// operands is UNDEF then the result is UNDEF. |
33336 | static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) { |
33337 | // Look for the following pattern: if |
33338 | // A = < float a0, float a1, float a2, float a3 > |
33339 | // B = < float b0, float b1, float b2, float b3 > |
33340 | // and |
33341 | // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6> |
33342 | // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7> |
33343 | // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 > |
33344 | // which is A horizontal-op B. |
33345 | |
33346 | // At least one of the operands should be a vector shuffle. |
33347 | if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE && |
33348 | RHS.getOpcode() != ISD::VECTOR_SHUFFLE) |
33349 | return false; |
33350 | |
33351 | MVT VT = LHS.getSimpleValueType(); |
33352 | |
33353 | assert((VT.is128BitVector() || VT.is256BitVector()) &&(((VT.is128BitVector() || VT.is256BitVector()) && "Unsupported vector type for horizontal add/sub" ) ? static_cast<void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector()) && \"Unsupported vector type for horizontal add/sub\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33354, __PRETTY_FUNCTION__)) |
33354 | "Unsupported vector type for horizontal add/sub")(((VT.is128BitVector() || VT.is256BitVector()) && "Unsupported vector type for horizontal add/sub" ) ? static_cast<void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector()) && \"Unsupported vector type for horizontal add/sub\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33354, __PRETTY_FUNCTION__)); |
33355 | |
33356 | // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to |
33357 | // operate independently on 128-bit lanes. |
33358 | unsigned NumElts = VT.getVectorNumElements(); |
33359 | unsigned NumLanes = VT.getSizeInBits()/128; |
33360 | unsigned NumLaneElts = NumElts / NumLanes; |
33361 | assert((NumLaneElts % 2 == 0) &&(((NumLaneElts % 2 == 0) && "Vector type should have an even number of elements in each lane" ) ? static_cast<void> (0) : __assert_fail ("(NumLaneElts % 2 == 0) && \"Vector type should have an even number of elements in each lane\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33362, __PRETTY_FUNCTION__)) |
33362 | "Vector type should have an even number of elements in each lane")(((NumLaneElts % 2 == 0) && "Vector type should have an even number of elements in each lane" ) ? static_cast<void> (0) : __assert_fail ("(NumLaneElts % 2 == 0) && \"Vector type should have an even number of elements in each lane\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33362, __PRETTY_FUNCTION__)); |
33363 | unsigned HalfLaneElts = NumLaneElts/2; |
33364 | |
33365 | // View LHS in the form |
33366 | // LHS = VECTOR_SHUFFLE A, B, LMask |
33367 | // If LHS is not a shuffle then pretend it is the shuffle |
33368 | // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1> |
33369 | // NOTE: in what follows a default initialized SDValue represents an UNDEF of |
33370 | // type VT. |
33371 | SDValue A, B; |
33372 | SmallVector<int, 16> LMask(NumElts); |
33373 | if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) { |
33374 | if (!LHS.getOperand(0).isUndef()) |
33375 | A = LHS.getOperand(0); |
33376 | if (!LHS.getOperand(1).isUndef()) |
33377 | B = LHS.getOperand(1); |
33378 | ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask(); |
33379 | std::copy(Mask.begin(), Mask.end(), LMask.begin()); |
33380 | } else { |
33381 | if (!LHS.isUndef()) |
33382 | A = LHS; |
33383 | for (unsigned i = 0; i != NumElts; ++i) |
33384 | LMask[i] = i; |
33385 | } |
33386 | |
33387 | // Likewise, view RHS in the form |
33388 | // RHS = VECTOR_SHUFFLE C, D, RMask |
33389 | SDValue C, D; |
33390 | SmallVector<int, 16> RMask(NumElts); |
33391 | if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) { |
33392 | if (!RHS.getOperand(0).isUndef()) |
33393 | C = RHS.getOperand(0); |
33394 | if (!RHS.getOperand(1).isUndef()) |
33395 | D = RHS.getOperand(1); |
33396 | ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask(); |
33397 | std::copy(Mask.begin(), Mask.end(), RMask.begin()); |
33398 | } else { |
33399 | if (!RHS.isUndef()) |
33400 | C = RHS; |
33401 | for (unsigned i = 0; i != NumElts; ++i) |
33402 | RMask[i] = i; |
33403 | } |
33404 | |
33405 | // Check that the shuffles are both shuffling the same vectors. |
33406 | if (!(A == C && B == D) && !(A == D && B == C)) |
33407 | return false; |
33408 | |
33409 | // If everything is UNDEF then bail out: it would be better to fold to UNDEF. |
33410 | if (!A.getNode() && !B.getNode()) |
33411 | return false; |
33412 | |
33413 | // If A and B occur in reverse order in RHS, then "swap" them (which means |
33414 | // rewriting the mask). |
33415 | if (A != C) |
33416 | ShuffleVectorSDNode::commuteMask(RMask); |
33417 | |
33418 | // At this point LHS and RHS are equivalent to |
33419 | // LHS = VECTOR_SHUFFLE A, B, LMask |
33420 | // RHS = VECTOR_SHUFFLE A, B, RMask |
33421 | // Check that the masks correspond to performing a horizontal operation. |
33422 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { |
33423 | for (unsigned i = 0; i != NumLaneElts; ++i) { |
33424 | int LIdx = LMask[i+l], RIdx = RMask[i+l]; |
33425 | |
33426 | // Ignore any UNDEF components. |
33427 | if (LIdx < 0 || RIdx < 0 || |
33428 | (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) || |
33429 | (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts))) |
33430 | continue; |
33431 | |
33432 | // Check that successive elements are being operated on. If not, this is |
33433 | // not a horizontal operation. |
33434 | unsigned Src = (i/HalfLaneElts); // each lane is split between srcs |
33435 | int Index = 2*(i%HalfLaneElts) + NumElts*Src + l; |
33436 | if (!(LIdx == Index && RIdx == Index + 1) && |
33437 | !(IsCommutative && LIdx == Index + 1 && RIdx == Index)) |
33438 | return false; |
33439 | } |
33440 | } |
33441 | |
33442 | LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it. |
33443 | RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it. |
33444 | return true; |
33445 | } |
33446 | |
33447 | /// Do target-specific dag combines on floating-point adds/subs. |
33448 | static SDValue combineFaddFsub(SDNode *N, SelectionDAG &DAG, |
33449 | const X86Subtarget &Subtarget) { |
33450 | EVT VT = N->getValueType(0); |
33451 | SDValue LHS = N->getOperand(0); |
33452 | SDValue RHS = N->getOperand(1); |
33453 | bool IsFadd = N->getOpcode() == ISD::FADD; |
33454 | assert((IsFadd || N->getOpcode() == ISD::FSUB) && "Wrong opcode")(((IsFadd || N->getOpcode() == ISD::FSUB) && "Wrong opcode" ) ? static_cast<void> (0) : __assert_fail ("(IsFadd || N->getOpcode() == ISD::FSUB) && \"Wrong opcode\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33454, __PRETTY_FUNCTION__)); |
33455 | |
33456 | // Try to synthesize horizontal add/sub from adds/subs of shuffles. |
33457 | if (((Subtarget.hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) || |
33458 | (Subtarget.hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) && |
33459 | isHorizontalBinOp(LHS, RHS, IsFadd)) { |
33460 | auto NewOpcode = IsFadd ? X86ISD::FHADD : X86ISD::FHSUB; |
33461 | return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS); |
33462 | } |
33463 | return SDValue(); |
33464 | } |
33465 | |
33466 | /// Attempt to pre-truncate inputs to arithmetic ops if it will simplify |
33467 | /// the codegen. |
33468 | /// e.g. TRUNC( BINOP( X, Y ) ) --> BINOP( TRUNC( X ), TRUNC( Y ) ) |
33469 | static SDValue combineTruncatedArithmetic(SDNode *N, SelectionDAG &DAG, |
33470 | const X86Subtarget &Subtarget, |
33471 | SDLoc &DL) { |
33472 | assert(N->getOpcode() == ISD::TRUNCATE && "Wrong opcode")((N->getOpcode() == ISD::TRUNCATE && "Wrong opcode" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::TRUNCATE && \"Wrong opcode\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33472, __PRETTY_FUNCTION__)); |
33473 | SDValue Src = N->getOperand(0); |
33474 | unsigned Opcode = Src.getOpcode(); |
33475 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
33476 | |
33477 | EVT VT = N->getValueType(0); |
33478 | EVT SrcVT = Src.getValueType(); |
33479 | |
33480 | auto IsRepeatedOpOrFreeTruncation = [VT](SDValue Op0, SDValue Op1) { |
33481 | unsigned TruncSizeInBits = VT.getScalarSizeInBits(); |
33482 | |
33483 | // Repeated operand, so we are only trading one output truncation for |
33484 | // one input truncation. |
33485 | if (Op0 == Op1) |
33486 | return true; |
33487 | |
33488 | // See if either operand has been extended from a smaller/equal size to |
33489 | // the truncation size, allowing a truncation to combine with the extend. |
33490 | unsigned Opcode0 = Op0.getOpcode(); |
33491 | if ((Opcode0 == ISD::ANY_EXTEND || Opcode0 == ISD::SIGN_EXTEND || |
33492 | Opcode0 == ISD::ZERO_EXTEND) && |
33493 | Op0.getOperand(0).getScalarValueSizeInBits() <= TruncSizeInBits) |
33494 | return true; |
33495 | |
33496 | unsigned Opcode1 = Op1.getOpcode(); |
33497 | if ((Opcode1 == ISD::ANY_EXTEND || Opcode1 == ISD::SIGN_EXTEND || |
33498 | Opcode1 == ISD::ZERO_EXTEND) && |
33499 | Op1.getOperand(0).getScalarValueSizeInBits() <= TruncSizeInBits) |
33500 | return true; |
33501 | |
33502 | // See if either operand is a single use constant which can be constant |
33503 | // folded. |
33504 | SDValue BC0 = peekThroughOneUseBitcasts(Op0); |
33505 | SDValue BC1 = peekThroughOneUseBitcasts(Op1); |
33506 | return ISD::isBuildVectorOfConstantSDNodes(BC0.getNode()) || |
33507 | ISD::isBuildVectorOfConstantSDNodes(BC1.getNode()); |
33508 | }; |
33509 | |
33510 | auto TruncateArithmetic = [&](SDValue N0, SDValue N1) { |
33511 | SDValue Trunc0 = DAG.getNode(ISD::TRUNCATE, DL, VT, N0); |
33512 | SDValue Trunc1 = DAG.getNode(ISD::TRUNCATE, DL, VT, N1); |
33513 | return DAG.getNode(Opcode, DL, VT, Trunc0, Trunc1); |
33514 | }; |
33515 | |
33516 | // Don't combine if the operation has other uses. |
33517 | if (!N->isOnlyUserOf(Src.getNode())) |
33518 | return SDValue(); |
33519 | |
33520 | // Only support vector truncation for now. |
33521 | // TODO: i64 scalar math would benefit as well. |
33522 | if (!VT.isVector()) |
33523 | return SDValue(); |
33524 | |
33525 | // In most cases its only worth pre-truncating if we're only facing the cost |
33526 | // of one truncation. |
33527 | // i.e. if one of the inputs will constant fold or the input is repeated. |
33528 | switch (Opcode) { |
33529 | case ISD::AND: |
33530 | case ISD::XOR: |
33531 | case ISD::OR: { |
33532 | SDValue Op0 = Src.getOperand(0); |
33533 | SDValue Op1 = Src.getOperand(1); |
33534 | if (TLI.isOperationLegalOrPromote(Opcode, VT) && |
33535 | IsRepeatedOpOrFreeTruncation(Op0, Op1)) |
33536 | return TruncateArithmetic(Op0, Op1); |
33537 | break; |
33538 | } |
33539 | |
33540 | case ISD::MUL: |
33541 | // X86 is rubbish at scalar and vector i64 multiplies (until AVX512DQ) - its |
33542 | // better to truncate if we have the chance. |
33543 | if (SrcVT.getScalarType() == MVT::i64 && TLI.isOperationLegal(Opcode, VT) && |
33544 | !TLI.isOperationLegal(Opcode, SrcVT)) |
33545 | return TruncateArithmetic(Src.getOperand(0), Src.getOperand(1)); |
33546 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
33547 | case ISD::ADD: { |
33548 | SDValue Op0 = Src.getOperand(0); |
33549 | SDValue Op1 = Src.getOperand(1); |
33550 | if (TLI.isOperationLegal(Opcode, VT) && |
33551 | IsRepeatedOpOrFreeTruncation(Op0, Op1)) |
33552 | return TruncateArithmetic(Op0, Op1); |
33553 | break; |
33554 | } |
33555 | } |
33556 | |
33557 | return SDValue(); |
33558 | } |
33559 | |
33560 | /// Truncate a group of v4i32 into v16i8/v8i16 using X86ISD::PACKUS. |
33561 | static SDValue |
33562 | combineVectorTruncationWithPACKUS(SDNode *N, SelectionDAG &DAG, |
33563 | SmallVector<SDValue, 8> &Regs) { |
33564 | assert(Regs.size() > 0 && (Regs[0].getValueType() == MVT::v4i32 ||((Regs.size() > 0 && (Regs[0].getValueType() == MVT ::v4i32 || Regs[0].getValueType() == MVT::v2i64)) ? static_cast <void> (0) : __assert_fail ("Regs.size() > 0 && (Regs[0].getValueType() == MVT::v4i32 || Regs[0].getValueType() == MVT::v2i64)" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33565, __PRETTY_FUNCTION__)) |
33565 | Regs[0].getValueType() == MVT::v2i64))((Regs.size() > 0 && (Regs[0].getValueType() == MVT ::v4i32 || Regs[0].getValueType() == MVT::v2i64)) ? static_cast <void> (0) : __assert_fail ("Regs.size() > 0 && (Regs[0].getValueType() == MVT::v4i32 || Regs[0].getValueType() == MVT::v2i64)" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33565, __PRETTY_FUNCTION__)); |
33566 | EVT OutVT = N->getValueType(0); |
33567 | EVT OutSVT = OutVT.getVectorElementType(); |
33568 | EVT InVT = Regs[0].getValueType(); |
33569 | EVT InSVT = InVT.getVectorElementType(); |
33570 | SDLoc DL(N); |
33571 | |
33572 | // First, use mask to unset all bits that won't appear in the result. |
33573 | assert((OutSVT == MVT::i8 || OutSVT == MVT::i16) &&(((OutSVT == MVT::i8 || OutSVT == MVT::i16) && "OutSVT can only be either i8 or i16." ) ? static_cast<void> (0) : __assert_fail ("(OutSVT == MVT::i8 || OutSVT == MVT::i16) && \"OutSVT can only be either i8 or i16.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33574, __PRETTY_FUNCTION__)) |
33574 | "OutSVT can only be either i8 or i16.")(((OutSVT == MVT::i8 || OutSVT == MVT::i16) && "OutSVT can only be either i8 or i16." ) ? static_cast<void> (0) : __assert_fail ("(OutSVT == MVT::i8 || OutSVT == MVT::i16) && \"OutSVT can only be either i8 or i16.\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33574, __PRETTY_FUNCTION__)); |
33575 | APInt Mask = |
33576 | APInt::getLowBitsSet(InSVT.getSizeInBits(), OutSVT.getSizeInBits()); |
33577 | SDValue MaskVal = DAG.getConstant(Mask, DL, InVT); |
33578 | for (auto &Reg : Regs) |
33579 | Reg = DAG.getNode(ISD::AND, DL, InVT, MaskVal, Reg); |
33580 | |
33581 | MVT UnpackedVT, PackedVT; |
33582 | if (OutSVT == MVT::i8) { |
33583 | UnpackedVT = MVT::v8i16; |
33584 | PackedVT = MVT::v16i8; |
33585 | } else { |
33586 | UnpackedVT = MVT::v4i32; |
33587 | PackedVT = MVT::v8i16; |
33588 | } |
33589 | |
33590 | // In each iteration, truncate the type by a half size. |
33591 | auto RegNum = Regs.size(); |
33592 | for (unsigned j = 1, e = InSVT.getSizeInBits() / OutSVT.getSizeInBits(); |
33593 | j < e; j *= 2, RegNum /= 2) { |
33594 | for (unsigned i = 0; i < RegNum; i++) |
33595 | Regs[i] = DAG.getBitcast(UnpackedVT, Regs[i]); |
33596 | for (unsigned i = 0; i < RegNum / 2; i++) |
33597 | Regs[i] = DAG.getNode(X86ISD::PACKUS, DL, PackedVT, Regs[i * 2], |
33598 | Regs[i * 2 + 1]); |
33599 | } |
33600 | |
33601 | // If the type of the result is v8i8, we need do one more X86ISD::PACKUS, and |
33602 | // then extract a subvector as the result since v8i8 is not a legal type. |
33603 | if (OutVT == MVT::v8i8) { |
33604 | Regs[0] = DAG.getNode(X86ISD::PACKUS, DL, PackedVT, Regs[0], Regs[0]); |
33605 | Regs[0] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OutVT, Regs[0], |
33606 | DAG.getIntPtrConstant(0, DL)); |
33607 | return Regs[0]; |
33608 | } else if (RegNum > 1) { |
33609 | Regs.resize(RegNum); |
33610 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Regs); |
33611 | } else |
33612 | return Regs[0]; |
33613 | } |
33614 | |
33615 | /// Truncate a group of v4i32 into v8i16 using X86ISD::PACKSS. |
33616 | static SDValue |
33617 | combineVectorTruncationWithPACKSS(SDNode *N, const X86Subtarget &Subtarget, |
33618 | SelectionDAG &DAG, |
33619 | SmallVector<SDValue, 8> &Regs) { |
33620 | assert(Regs.size() > 0 && Regs[0].getValueType() == MVT::v4i32)((Regs.size() > 0 && Regs[0].getValueType() == MVT ::v4i32) ? static_cast<void> (0) : __assert_fail ("Regs.size() > 0 && Regs[0].getValueType() == MVT::v4i32" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33620, __PRETTY_FUNCTION__)); |
33621 | EVT OutVT = N->getValueType(0); |
33622 | SDLoc DL(N); |
33623 | |
33624 | // Shift left by 16 bits, then arithmetic-shift right by 16 bits. |
33625 | SDValue ShAmt = DAG.getConstant(16, DL, MVT::i32); |
33626 | for (auto &Reg : Regs) { |
33627 | Reg = getTargetVShiftNode(X86ISD::VSHLI, DL, MVT::v4i32, Reg, ShAmt, |
33628 | Subtarget, DAG); |
33629 | Reg = getTargetVShiftNode(X86ISD::VSRAI, DL, MVT::v4i32, Reg, ShAmt, |
33630 | Subtarget, DAG); |
33631 | } |
33632 | |
33633 | for (unsigned i = 0, e = Regs.size() / 2; i < e; i++) |
33634 | Regs[i] = DAG.getNode(X86ISD::PACKSS, DL, MVT::v8i16, Regs[i * 2], |
33635 | Regs[i * 2 + 1]); |
33636 | |
33637 | if (Regs.size() > 2) { |
33638 | Regs.resize(Regs.size() / 2); |
33639 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Regs); |
33640 | } else |
33641 | return Regs[0]; |
33642 | } |
33643 | |
33644 | /// This function transforms truncation from vXi32/vXi64 to vXi8/vXi16 into |
33645 | /// X86ISD::PACKUS/X86ISD::PACKSS operations. We do it here because after type |
33646 | /// legalization the truncation will be translated into a BUILD_VECTOR with each |
33647 | /// element that is extracted from a vector and then truncated, and it is |
33648 | /// difficult to do this optimization based on them. |
33649 | static SDValue combineVectorTruncation(SDNode *N, SelectionDAG &DAG, |
33650 | const X86Subtarget &Subtarget) { |
33651 | EVT OutVT = N->getValueType(0); |
33652 | if (!OutVT.isVector()) |
33653 | return SDValue(); |
33654 | |
33655 | SDValue In = N->getOperand(0); |
33656 | if (!In.getValueType().isSimple()) |
33657 | return SDValue(); |
33658 | |
33659 | EVT InVT = In.getValueType(); |
33660 | unsigned NumElems = OutVT.getVectorNumElements(); |
33661 | |
33662 | // TODO: On AVX2, the behavior of X86ISD::PACKUS is different from that on |
33663 | // SSE2, and we need to take care of it specially. |
33664 | // AVX512 provides vpmovdb. |
33665 | if (!Subtarget.hasSSE2() || Subtarget.hasAVX2()) |
33666 | return SDValue(); |
33667 | |
33668 | EVT OutSVT = OutVT.getVectorElementType(); |
33669 | EVT InSVT = InVT.getVectorElementType(); |
33670 | if (!((InSVT == MVT::i32 || InSVT == MVT::i64) && |
33671 | (OutSVT == MVT::i8 || OutSVT == MVT::i16) && isPowerOf2_32(NumElems) && |
33672 | NumElems >= 8)) |
33673 | return SDValue(); |
33674 | |
33675 | // SSSE3's pshufb results in less instructions in the cases below. |
33676 | if (Subtarget.hasSSSE3() && NumElems == 8 && |
33677 | ((OutSVT == MVT::i8 && InSVT != MVT::i64) || |
33678 | (InSVT == MVT::i32 && OutSVT == MVT::i16))) |
33679 | return SDValue(); |
33680 | |
33681 | SDLoc DL(N); |
33682 | |
33683 | // Split a long vector into vectors of legal type. |
33684 | unsigned RegNum = InVT.getSizeInBits() / 128; |
33685 | SmallVector<SDValue, 8> SubVec(RegNum); |
33686 | unsigned NumSubRegElts = 128 / InSVT.getSizeInBits(); |
33687 | EVT SubRegVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubRegElts); |
33688 | |
33689 | for (unsigned i = 0; i < RegNum; i++) |
33690 | SubVec[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubRegVT, In, |
33691 | DAG.getIntPtrConstant(i * NumSubRegElts, DL)); |
33692 | |
33693 | // SSE2 provides PACKUS for only 2 x v8i16 -> v16i8 and SSE4.1 provides PACKUS |
33694 | // for 2 x v4i32 -> v8i16. For SSSE3 and below, we need to use PACKSS to |
33695 | // truncate 2 x v4i32 to v8i16. |
33696 | if (Subtarget.hasSSE41() || OutSVT == MVT::i8) |
33697 | return combineVectorTruncationWithPACKUS(N, DAG, SubVec); |
33698 | else if (InSVT == MVT::i32) |
33699 | return combineVectorTruncationWithPACKSS(N, Subtarget, DAG, SubVec); |
33700 | else |
33701 | return SDValue(); |
33702 | } |
33703 | |
33704 | /// This function transforms vector truncation of 'all or none' bits values. |
33705 | /// vXi16/vXi32/vXi64 to vXi8/vXi16/vXi32 into X86ISD::PACKSS operations. |
33706 | static SDValue combineVectorSignBitsTruncation(SDNode *N, SDLoc &DL, |
33707 | SelectionDAG &DAG, |
33708 | const X86Subtarget &Subtarget) { |
33709 | // Requires SSE2 but AVX512 has fast truncate. |
33710 | if (!Subtarget.hasSSE2() || Subtarget.hasAVX512()) |
33711 | return SDValue(); |
33712 | |
33713 | if (!N->getValueType(0).isVector() || !N->getValueType(0).isSimple()) |
33714 | return SDValue(); |
33715 | |
33716 | SDValue In = N->getOperand(0); |
33717 | if (!In.getValueType().isSimple()) |
33718 | return SDValue(); |
33719 | |
33720 | MVT VT = N->getValueType(0).getSimpleVT(); |
33721 | MVT SVT = VT.getScalarType(); |
33722 | |
33723 | MVT InVT = In.getValueType().getSimpleVT(); |
33724 | MVT InSVT = InVT.getScalarType(); |
33725 | |
33726 | // Use PACKSS if the input is a splatted sign bit. |
33727 | // e.g. Comparison result, sext_in_reg, etc. |
33728 | unsigned NumSignBits = DAG.ComputeNumSignBits(In); |
33729 | if (NumSignBits != InSVT.getSizeInBits()) |
33730 | return SDValue(); |
33731 | |
33732 | // Check we have a truncation suited for PACKSS. |
33733 | if (!VT.is128BitVector() && !VT.is256BitVector()) |
33734 | return SDValue(); |
33735 | if (SVT != MVT::i8 && SVT != MVT::i16 && SVT != MVT::i32) |
33736 | return SDValue(); |
33737 | if (InSVT != MVT::i16 && InSVT != MVT::i32 && InSVT != MVT::i64) |
33738 | return SDValue(); |
33739 | |
33740 | return truncateVectorCompareWithPACKSS(VT, In, DL, DAG, Subtarget); |
33741 | } |
33742 | |
33743 | static SDValue combineTruncate(SDNode *N, SelectionDAG &DAG, |
33744 | const X86Subtarget &Subtarget) { |
33745 | EVT VT = N->getValueType(0); |
33746 | SDValue Src = N->getOperand(0); |
33747 | SDLoc DL(N); |
33748 | |
33749 | // Attempt to pre-truncate inputs to arithmetic ops instead. |
33750 | if (SDValue V = combineTruncatedArithmetic(N, DAG, Subtarget, DL)) |
33751 | return V; |
33752 | |
33753 | // Try to detect AVG pattern first. |
33754 | if (SDValue Avg = detectAVGPattern(Src, VT, DAG, Subtarget, DL)) |
33755 | return Avg; |
33756 | |
33757 | // Try to combine truncation with unsigned saturation. |
33758 | if (SDValue Val = combineTruncateWithUSat(Src, VT, DL, DAG, Subtarget)) |
33759 | return Val; |
33760 | |
33761 | // The bitcast source is a direct mmx result. |
33762 | // Detect bitcasts between i32 to x86mmx |
33763 | if (Src.getOpcode() == ISD::BITCAST && VT == MVT::i32) { |
33764 | SDValue BCSrc = Src.getOperand(0); |
33765 | if (BCSrc.getValueType() == MVT::x86mmx) |
33766 | return DAG.getNode(X86ISD::MMX_MOVD2W, DL, MVT::i32, BCSrc); |
33767 | } |
33768 | |
33769 | // Try to truncate extended sign bits with PACKSS. |
33770 | if (SDValue V = combineVectorSignBitsTruncation(N, DL, DAG, Subtarget)) |
33771 | return V; |
33772 | |
33773 | return combineVectorTruncation(N, DAG, Subtarget); |
33774 | } |
33775 | |
33776 | /// Returns the negated value if the node \p N flips sign of FP value. |
33777 | /// |
33778 | /// FP-negation node may have different forms: FNEG(x) or FXOR (x, 0x80000000). |
33779 | /// AVX512F does not have FXOR, so FNEG is lowered as |
33780 | /// (bitcast (xor (bitcast x), (bitcast ConstantFP(0x80000000)))). |
33781 | /// In this case we go though all bitcasts. |
33782 | static SDValue isFNEG(SDNode *N) { |
33783 | if (N->getOpcode() == ISD::FNEG) |
33784 | return N->getOperand(0); |
33785 | |
33786 | SDValue Op = peekThroughBitcasts(SDValue(N, 0)); |
33787 | if (Op.getOpcode() != X86ISD::FXOR && Op.getOpcode() != ISD::XOR) |
33788 | return SDValue(); |
33789 | |
33790 | SDValue Op1 = peekThroughBitcasts(Op.getOperand(1)); |
33791 | if (!Op1.getValueType().isFloatingPoint()) |
33792 | return SDValue(); |
33793 | |
33794 | SDValue Op0 = peekThroughBitcasts(Op.getOperand(0)); |
33795 | |
33796 | unsigned EltBits = Op1.getScalarValueSizeInBits(); |
33797 | auto isSignMask = [&](const ConstantFP *C) { |
33798 | return C->getValueAPF().bitcastToAPInt() == APInt::getSignMask(EltBits); |
33799 | }; |
33800 | |
33801 | // There is more than one way to represent the same constant on |
33802 | // the different X86 targets. The type of the node may also depend on size. |
33803 | // - load scalar value and broadcast |
33804 | // - BUILD_VECTOR node |
33805 | // - load from a constant pool. |
33806 | // We check all variants here. |
33807 | if (Op1.getOpcode() == X86ISD::VBROADCAST) { |
33808 | if (auto *C = getTargetConstantFromNode(Op1.getOperand(0))) |
33809 | if (isSignMask(cast<ConstantFP>(C))) |
33810 | return Op0; |
33811 | |
33812 | } else if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1)) { |
33813 | if (ConstantFPSDNode *CN = BV->getConstantFPSplatNode()) |
33814 | if (isSignMask(CN->getConstantFPValue())) |
33815 | return Op0; |
33816 | |
33817 | } else if (auto *C = getTargetConstantFromNode(Op1)) { |
33818 | if (C->getType()->isVectorTy()) { |
33819 | if (auto *SplatV = C->getSplatValue()) |
33820 | if (isSignMask(cast<ConstantFP>(SplatV))) |
33821 | return Op0; |
33822 | } else if (auto *FPConst = dyn_cast<ConstantFP>(C)) |
33823 | if (isSignMask(FPConst)) |
33824 | return Op0; |
33825 | } |
33826 | return SDValue(); |
33827 | } |
33828 | |
33829 | /// Do target-specific dag combines on floating point negations. |
33830 | static SDValue combineFneg(SDNode *N, SelectionDAG &DAG, |
33831 | const X86Subtarget &Subtarget) { |
33832 | EVT OrigVT = N->getValueType(0); |
33833 | SDValue Arg = isFNEG(N); |
33834 | assert(Arg.getNode() && "N is expected to be an FNEG node")((Arg.getNode() && "N is expected to be an FNEG node" ) ? static_cast<void> (0) : __assert_fail ("Arg.getNode() && \"N is expected to be an FNEG node\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33834, __PRETTY_FUNCTION__)); |
33835 | |
33836 | EVT VT = Arg.getValueType(); |
33837 | EVT SVT = VT.getScalarType(); |
33838 | SDLoc DL(N); |
33839 | |
33840 | // Let legalize expand this if it isn't a legal type yet. |
33841 | if (!DAG.getTargetLoweringInfo().isTypeLegal(VT)) |
33842 | return SDValue(); |
33843 | |
33844 | // If we're negating a FMUL node on a target with FMA, then we can avoid the |
33845 | // use of a constant by performing (-0 - A*B) instead. |
33846 | // FIXME: Check rounding control flags as well once it becomes available. |
33847 | if (Arg.getOpcode() == ISD::FMUL && (SVT == MVT::f32 || SVT == MVT::f64) && |
33848 | Arg->getFlags().hasNoSignedZeros() && Subtarget.hasAnyFMA()) { |
33849 | SDValue Zero = DAG.getConstantFP(0.0, DL, VT); |
33850 | SDValue NewNode = DAG.getNode(X86ISD::FNMSUB, DL, VT, Arg.getOperand(0), |
33851 | Arg.getOperand(1), Zero); |
33852 | return DAG.getBitcast(OrigVT, NewNode); |
33853 | } |
33854 | |
33855 | // If we're negating an FMA node, then we can adjust the |
33856 | // instruction to include the extra negation. |
33857 | unsigned NewOpcode = 0; |
33858 | if (Arg.hasOneUse()) { |
33859 | switch (Arg.getOpcode()) { |
33860 | case X86ISD::FMADD: NewOpcode = X86ISD::FNMSUB; break; |
33861 | case X86ISD::FMSUB: NewOpcode = X86ISD::FNMADD; break; |
33862 | case X86ISD::FNMADD: NewOpcode = X86ISD::FMSUB; break; |
33863 | case X86ISD::FNMSUB: NewOpcode = X86ISD::FMADD; break; |
33864 | case X86ISD::FMADD_RND: NewOpcode = X86ISD::FNMSUB_RND; break; |
33865 | case X86ISD::FMSUB_RND: NewOpcode = X86ISD::FNMADD_RND; break; |
33866 | case X86ISD::FNMADD_RND: NewOpcode = X86ISD::FMSUB_RND; break; |
33867 | case X86ISD::FNMSUB_RND: NewOpcode = X86ISD::FMADD_RND; break; |
33868 | // We can't handle scalar intrinsic node here because it would only |
33869 | // invert one element and not the whole vector. But we could try to handle |
33870 | // a negation of the lower element only. |
33871 | } |
33872 | } |
33873 | if (NewOpcode) |
33874 | return DAG.getBitcast(OrigVT, DAG.getNode(NewOpcode, DL, VT, |
33875 | Arg.getNode()->ops())); |
33876 | |
33877 | return SDValue(); |
33878 | } |
33879 | |
33880 | static SDValue lowerX86FPLogicOp(SDNode *N, SelectionDAG &DAG, |
33881 | const X86Subtarget &Subtarget) { |
33882 | MVT VT = N->getSimpleValueType(0); |
33883 | // If we have integer vector types available, use the integer opcodes. |
33884 | if (VT.isVector() && Subtarget.hasSSE2()) { |
33885 | SDLoc dl(N); |
33886 | |
33887 | MVT IntVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64); |
33888 | |
33889 | SDValue Op0 = DAG.getBitcast(IntVT, N->getOperand(0)); |
33890 | SDValue Op1 = DAG.getBitcast(IntVT, N->getOperand(1)); |
33891 | unsigned IntOpcode; |
33892 | switch (N->getOpcode()) { |
33893 | default: llvm_unreachable("Unexpected FP logic op")::llvm::llvm_unreachable_internal("Unexpected FP logic op", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 33893); |
33894 | case X86ISD::FOR: IntOpcode = ISD::OR; break; |
33895 | case X86ISD::FXOR: IntOpcode = ISD::XOR; break; |
33896 | case X86ISD::FAND: IntOpcode = ISD::AND; break; |
33897 | case X86ISD::FANDN: IntOpcode = X86ISD::ANDNP; break; |
33898 | } |
33899 | SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1); |
33900 | return DAG.getBitcast(VT, IntOp); |
33901 | } |
33902 | return SDValue(); |
33903 | } |
33904 | |
33905 | static SDValue combineXor(SDNode *N, SelectionDAG &DAG, |
33906 | TargetLowering::DAGCombinerInfo &DCI, |
33907 | const X86Subtarget &Subtarget) { |
33908 | if (SDValue Cmp = foldVectorXorShiftIntoCmp(N, DAG, Subtarget)) |
33909 | return Cmp; |
33910 | |
33911 | if (DCI.isBeforeLegalizeOps()) |
33912 | return SDValue(); |
33913 | |
33914 | if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG)) |
33915 | return RV; |
33916 | |
33917 | if (Subtarget.hasCMov()) |
33918 | if (SDValue RV = combineIntegerAbs(N, DAG)) |
33919 | return RV; |
33920 | |
33921 | if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget)) |
33922 | return FPLogic; |
33923 | |
33924 | if (isFNEG(N)) |
33925 | return combineFneg(N, DAG, Subtarget); |
33926 | return SDValue(); |
33927 | } |
33928 | |
33929 | |
33930 | static bool isNullFPScalarOrVectorConst(SDValue V) { |
33931 | return isNullFPConstant(V) || ISD::isBuildVectorAllZeros(V.getNode()); |
33932 | } |
33933 | |
33934 | /// If a value is a scalar FP zero or a vector FP zero (potentially including |
33935 | /// undefined elements), return a zero constant that may be used to fold away |
33936 | /// that value. In the case of a vector, the returned constant will not contain |
33937 | /// undefined elements even if the input parameter does. This makes it suitable |
33938 | /// to be used as a replacement operand with operations (eg, bitwise-and) where |
33939 | /// an undef should not propagate. |
33940 | static SDValue getNullFPConstForNullVal(SDValue V, SelectionDAG &DAG, |
33941 | const X86Subtarget &Subtarget) { |
33942 | if (!isNullFPScalarOrVectorConst(V)) |
33943 | return SDValue(); |
33944 | |
33945 | if (V.getValueType().isVector()) |
33946 | return getZeroVector(V.getSimpleValueType(), Subtarget, DAG, SDLoc(V)); |
33947 | |
33948 | return V; |
33949 | } |
33950 | |
33951 | static SDValue combineFAndFNotToFAndn(SDNode *N, SelectionDAG &DAG, |
33952 | const X86Subtarget &Subtarget) { |
33953 | SDValue N0 = N->getOperand(0); |
33954 | SDValue N1 = N->getOperand(1); |
33955 | EVT VT = N->getValueType(0); |
33956 | SDLoc DL(N); |
33957 | |
33958 | // Vector types are handled in combineANDXORWithAllOnesIntoANDNP(). |
33959 | if (!((VT == MVT::f32 && Subtarget.hasSSE1()) || |
33960 | (VT == MVT::f64 && Subtarget.hasSSE2()))) |
33961 | return SDValue(); |
33962 | |
33963 | auto isAllOnesConstantFP = [](SDValue V) { |
33964 | auto *C = dyn_cast<ConstantFPSDNode>(V); |
33965 | return C && C->getConstantFPValue()->isAllOnesValue(); |
33966 | }; |
33967 | |
33968 | // fand (fxor X, -1), Y --> fandn X, Y |
33969 | if (N0.getOpcode() == X86ISD::FXOR && isAllOnesConstantFP(N0.getOperand(1))) |
33970 | return DAG.getNode(X86ISD::FANDN, DL, VT, N0.getOperand(0), N1); |
33971 | |
33972 | // fand X, (fxor Y, -1) --> fandn Y, X |
33973 | if (N1.getOpcode() == X86ISD::FXOR && isAllOnesConstantFP(N1.getOperand(1))) |
33974 | return DAG.getNode(X86ISD::FANDN, DL, VT, N1.getOperand(0), N0); |
33975 | |
33976 | return SDValue(); |
33977 | } |
33978 | |
33979 | /// Do target-specific dag combines on X86ISD::FAND nodes. |
33980 | static SDValue combineFAnd(SDNode *N, SelectionDAG &DAG, |
33981 | const X86Subtarget &Subtarget) { |
33982 | // FAND(0.0, x) -> 0.0 |
33983 | if (SDValue V = getNullFPConstForNullVal(N->getOperand(0), DAG, Subtarget)) |
33984 | return V; |
33985 | |
33986 | // FAND(x, 0.0) -> 0.0 |
33987 | if (SDValue V = getNullFPConstForNullVal(N->getOperand(1), DAG, Subtarget)) |
33988 | return V; |
33989 | |
33990 | if (SDValue V = combineFAndFNotToFAndn(N, DAG, Subtarget)) |
33991 | return V; |
33992 | |
33993 | return lowerX86FPLogicOp(N, DAG, Subtarget); |
33994 | } |
33995 | |
33996 | /// Do target-specific dag combines on X86ISD::FANDN nodes. |
33997 | static SDValue combineFAndn(SDNode *N, SelectionDAG &DAG, |
33998 | const X86Subtarget &Subtarget) { |
33999 | // FANDN(0.0, x) -> x |
34000 | if (isNullFPScalarOrVectorConst(N->getOperand(0))) |
34001 | return N->getOperand(1); |
34002 | |
34003 | // FANDN(x, 0.0) -> 0.0 |
34004 | if (SDValue V = getNullFPConstForNullVal(N->getOperand(1), DAG, Subtarget)) |
34005 | return V; |
34006 | |
34007 | return lowerX86FPLogicOp(N, DAG, Subtarget); |
34008 | } |
34009 | |
34010 | /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes. |
34011 | static SDValue combineFOr(SDNode *N, SelectionDAG &DAG, |
34012 | const X86Subtarget &Subtarget) { |
34013 | assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR)((N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD ::FXOR) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 34013, __PRETTY_FUNCTION__)); |
34014 | |
34015 | // F[X]OR(0.0, x) -> x |
34016 | if (isNullFPScalarOrVectorConst(N->getOperand(0))) |
34017 | return N->getOperand(1); |
34018 | |
34019 | // F[X]OR(x, 0.0) -> x |
34020 | if (isNullFPScalarOrVectorConst(N->getOperand(1))) |
34021 | return N->getOperand(0); |
34022 | |
34023 | if (isFNEG(N)) |
34024 | if (SDValue NewVal = combineFneg(N, DAG, Subtarget)) |
34025 | return NewVal; |
34026 | |
34027 | return lowerX86FPLogicOp(N, DAG, Subtarget); |
34028 | } |
34029 | |
34030 | /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes. |
34031 | static SDValue combineFMinFMax(SDNode *N, SelectionDAG &DAG) { |
34032 | assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX)((N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD ::FMAX) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 34032, __PRETTY_FUNCTION__)); |
34033 | |
34034 | // Only perform optimizations if UnsafeMath is used. |
34035 | if (!DAG.getTarget().Options.UnsafeFPMath) |
34036 | return SDValue(); |
34037 | |
34038 | // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes |
34039 | // into FMINC and FMAXC, which are Commutative operations. |
34040 | unsigned NewOp = 0; |
34041 | switch (N->getOpcode()) { |
34042 | default: llvm_unreachable("unknown opcode")::llvm::llvm_unreachable_internal("unknown opcode", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 34042); |
34043 | case X86ISD::FMIN: NewOp = X86ISD::FMINC; break; |
34044 | case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break; |
34045 | } |
34046 | |
34047 | return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0), |
34048 | N->getOperand(0), N->getOperand(1)); |
34049 | } |
34050 | |
34051 | static SDValue combineFMinNumFMaxNum(SDNode *N, SelectionDAG &DAG, |
34052 | const X86Subtarget &Subtarget) { |
34053 | if (Subtarget.useSoftFloat()) |
34054 | return SDValue(); |
34055 | |
34056 | // TODO: Check for global or instruction-level "nnan". In that case, we |
34057 | // should be able to lower to FMAX/FMIN alone. |
34058 | // TODO: If an operand is already known to be a NaN or not a NaN, this |
34059 | // should be an optional swap and FMAX/FMIN. |
34060 | |
34061 | EVT VT = N->getValueType(0); |
34062 | if (!((Subtarget.hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) || |
34063 | (Subtarget.hasSSE2() && (VT == MVT::f64 || VT == MVT::v2f64)) || |
34064 | (Subtarget.hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64)))) |
34065 | return SDValue(); |
34066 | |
34067 | // This takes at least 3 instructions, so favor a library call when operating |
34068 | // on a scalar and minimizing code size. |
34069 | if (!VT.isVector() && DAG.getMachineFunction().getFunction()->optForMinSize()) |
34070 | return SDValue(); |
34071 | |
34072 | SDValue Op0 = N->getOperand(0); |
34073 | SDValue Op1 = N->getOperand(1); |
34074 | SDLoc DL(N); |
34075 | EVT SetCCType = DAG.getTargetLoweringInfo().getSetCCResultType( |
34076 | DAG.getDataLayout(), *DAG.getContext(), VT); |
34077 | |
34078 | // There are 4 possibilities involving NaN inputs, and these are the required |
34079 | // outputs: |
34080 | // Op1 |
34081 | // Num NaN |
34082 | // ---------------- |
34083 | // Num | Max | Op0 | |
34084 | // Op0 ---------------- |
34085 | // NaN | Op1 | NaN | |
34086 | // ---------------- |
34087 | // |
34088 | // The SSE FP max/min instructions were not designed for this case, but rather |
34089 | // to implement: |
34090 | // Min = Op1 < Op0 ? Op1 : Op0 |
34091 | // Max = Op1 > Op0 ? Op1 : Op0 |
34092 | // |
34093 | // So they always return Op0 if either input is a NaN. However, we can still |
34094 | // use those instructions for fmaxnum by selecting away a NaN input. |
34095 | |
34096 | // If either operand is NaN, the 2nd source operand (Op0) is passed through. |
34097 | auto MinMaxOp = N->getOpcode() == ISD::FMAXNUM ? X86ISD::FMAX : X86ISD::FMIN; |
34098 | SDValue MinOrMax = DAG.getNode(MinMaxOp, DL, VT, Op1, Op0); |
34099 | SDValue IsOp0Nan = DAG.getSetCC(DL, SetCCType , Op0, Op0, ISD::SETUO); |
34100 | |
34101 | // If Op0 is a NaN, select Op1. Otherwise, select the max. If both operands |
34102 | // are NaN, the NaN value of Op1 is the result. |
34103 | return DAG.getSelect(DL, VT, IsOp0Nan, Op1, MinOrMax); |
34104 | } |
34105 | |
34106 | /// Do target-specific dag combines on X86ISD::ANDNP nodes. |
34107 | static SDValue combineAndnp(SDNode *N, SelectionDAG &DAG, |
34108 | TargetLowering::DAGCombinerInfo &DCI, |
34109 | const X86Subtarget &Subtarget) { |
34110 | // ANDNP(0, x) -> x |
34111 | if (ISD::isBuildVectorAllZeros(N->getOperand(0).getNode())) |
34112 | return N->getOperand(1); |
34113 | |
34114 | // ANDNP(x, 0) -> 0 |
34115 | if (ISD::isBuildVectorAllZeros(N->getOperand(1).getNode())) |
34116 | return getZeroVector(N->getSimpleValueType(0), Subtarget, DAG, SDLoc(N)); |
34117 | |
34118 | EVT VT = N->getValueType(0); |
34119 | |
34120 | // Attempt to recursively combine a bitmask ANDNP with shuffles. |
34121 | if (VT.isVector() && (VT.getScalarSizeInBits() % 8) == 0) { |
34122 | SDValue Op(N, 0); |
34123 | SmallVector<int, 1> NonceMask; // Just a placeholder. |
34124 | NonceMask.push_back(0); |
34125 | if (combineX86ShufflesRecursively({Op}, 0, Op, NonceMask, {}, |
34126 | /*Depth*/ 1, /*HasVarMask*/ false, DAG, |
34127 | DCI, Subtarget)) |
34128 | return SDValue(); // This routine will use CombineTo to replace N. |
34129 | } |
34130 | |
34131 | return SDValue(); |
34132 | } |
34133 | |
34134 | static SDValue combineBT(SDNode *N, SelectionDAG &DAG, |
34135 | TargetLowering::DAGCombinerInfo &DCI) { |
34136 | // BT ignores high bits in the bit index operand. |
34137 | SDValue Op1 = N->getOperand(1); |
34138 | if (Op1.hasOneUse()) { |
34139 | unsigned BitWidth = Op1.getValueSizeInBits(); |
34140 | APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth)); |
34141 | KnownBits Known; |
34142 | TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(), |
34143 | !DCI.isBeforeLegalizeOps()); |
34144 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
34145 | if (TLI.ShrinkDemandedConstant(Op1, DemandedMask, TLO) || |
34146 | TLI.SimplifyDemandedBits(Op1, DemandedMask, Known, TLO)) |
34147 | DCI.CommitTargetLoweringOpt(TLO); |
34148 | } |
34149 | return SDValue(); |
34150 | } |
34151 | |
34152 | static SDValue combineSignExtendInReg(SDNode *N, SelectionDAG &DAG, |
34153 | const X86Subtarget &Subtarget) { |
34154 | EVT VT = N->getValueType(0); |
34155 | if (!VT.isVector()) |
34156 | return SDValue(); |
34157 | |
34158 | SDValue N0 = N->getOperand(0); |
34159 | SDValue N1 = N->getOperand(1); |
34160 | EVT ExtraVT = cast<VTSDNode>(N1)->getVT(); |
34161 | SDLoc dl(N); |
34162 | |
34163 | // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the |
34164 | // both SSE and AVX2 since there is no sign-extended shift right |
34165 | // operation on a vector with 64-bit elements. |
34166 | //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) -> |
34167 | // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT))) |
34168 | if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND || |
34169 | N0.getOpcode() == ISD::SIGN_EXTEND)) { |
34170 | SDValue N00 = N0.getOperand(0); |
34171 | |
34172 | // EXTLOAD has a better solution on AVX2, |
34173 | // it may be replaced with X86ISD::VSEXT node. |
34174 | if (N00.getOpcode() == ISD::LOAD && Subtarget.hasInt256()) |
34175 | if (!ISD::isNormalLoad(N00.getNode())) |
34176 | return SDValue(); |
34177 | |
34178 | if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) { |
34179 | SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32, |
34180 | N00, N1); |
34181 | return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp); |
34182 | } |
34183 | } |
34184 | return SDValue(); |
34185 | } |
34186 | |
34187 | /// sext(add_nsw(x, C)) --> add(sext(x), C_sext) |
34188 | /// zext(add_nuw(x, C)) --> add(zext(x), C_zext) |
34189 | /// Promoting a sign/zero extension ahead of a no overflow 'add' exposes |
34190 | /// opportunities to combine math ops, use an LEA, or use a complex addressing |
34191 | /// mode. This can eliminate extend, add, and shift instructions. |
34192 | static SDValue promoteExtBeforeAdd(SDNode *Ext, SelectionDAG &DAG, |
34193 | const X86Subtarget &Subtarget) { |
34194 | if (Ext->getOpcode() != ISD::SIGN_EXTEND && |
34195 | Ext->getOpcode() != ISD::ZERO_EXTEND) |
34196 | return SDValue(); |
34197 | |
34198 | // TODO: This should be valid for other integer types. |
34199 | EVT VT = Ext->getValueType(0); |
34200 | if (VT != MVT::i64) |
34201 | return SDValue(); |
34202 | |
34203 | SDValue Add = Ext->getOperand(0); |
34204 | if (Add.getOpcode() != ISD::ADD) |
34205 | return SDValue(); |
34206 | |
34207 | bool Sext = Ext->getOpcode() == ISD::SIGN_EXTEND; |
34208 | bool NSW = Add->getFlags().hasNoSignedWrap(); |
34209 | bool NUW = Add->getFlags().hasNoUnsignedWrap(); |
34210 | |
34211 | // We need an 'add nsw' feeding into the 'sext' or 'add nuw' feeding |
34212 | // into the 'zext' |
34213 | if ((Sext && !NSW) || (!Sext && !NUW)) |
34214 | return SDValue(); |
34215 | |
34216 | // Having a constant operand to the 'add' ensures that we are not increasing |
34217 | // the instruction count because the constant is extended for free below. |
34218 | // A constant operand can also become the displacement field of an LEA. |
34219 | auto *AddOp1 = dyn_cast<ConstantSDNode>(Add.getOperand(1)); |
34220 | if (!AddOp1) |
34221 | return SDValue(); |
34222 | |
34223 | // Don't make the 'add' bigger if there's no hope of combining it with some |
34224 | // other 'add' or 'shl' instruction. |
34225 | // TODO: It may be profitable to generate simpler LEA instructions in place |
34226 | // of single 'add' instructions, but the cost model for selecting an LEA |
34227 | // currently has a high threshold. |
34228 | bool HasLEAPotential = false; |
34229 | for (auto *User : Ext->uses()) { |
34230 | if (User->getOpcode() == ISD::ADD || User->getOpcode() == ISD::SHL) { |
34231 | HasLEAPotential = true; |
34232 | break; |
34233 | } |
34234 | } |
34235 | if (!HasLEAPotential) |
34236 | return SDValue(); |
34237 | |
34238 | // Everything looks good, so pull the '{s|z}ext' ahead of the 'add'. |
34239 | int64_t AddConstant = Sext ? AddOp1->getSExtValue() : AddOp1->getZExtValue(); |
34240 | SDValue AddOp0 = Add.getOperand(0); |
34241 | SDValue NewExt = DAG.getNode(Ext->getOpcode(), SDLoc(Ext), VT, AddOp0); |
34242 | SDValue NewConstant = DAG.getConstant(AddConstant, SDLoc(Add), VT); |
34243 | |
34244 | // The wider add is guaranteed to not wrap because both operands are |
34245 | // sign-extended. |
34246 | SDNodeFlags Flags; |
34247 | Flags.setNoSignedWrap(NSW); |
34248 | Flags.setNoUnsignedWrap(NUW); |
34249 | return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewExt, NewConstant, Flags); |
34250 | } |
34251 | |
34252 | /// (i8,i32 {s/z}ext ({s/u}divrem (i8 x, i8 y)) -> |
34253 | /// (i8,i32 ({s/u}divrem_sext_hreg (i8 x, i8 y) |
34254 | /// This exposes the {s/z}ext to the sdivrem lowering, so that it directly |
34255 | /// extends from AH (which we otherwise need to do contortions to access). |
34256 | static SDValue getDivRem8(SDNode *N, SelectionDAG &DAG) { |
34257 | SDValue N0 = N->getOperand(0); |
34258 | auto OpcodeN = N->getOpcode(); |
34259 | auto OpcodeN0 = N0.getOpcode(); |
34260 | if (!((OpcodeN == ISD::SIGN_EXTEND && OpcodeN0 == ISD::SDIVREM) || |
34261 | (OpcodeN == ISD::ZERO_EXTEND && OpcodeN0 == ISD::UDIVREM))) |
34262 | return SDValue(); |
34263 | |
34264 | EVT VT = N->getValueType(0); |
34265 | EVT InVT = N0.getValueType(); |
34266 | if (N0.getResNo() != 1 || InVT != MVT::i8 || VT != MVT::i32) |
34267 | return SDValue(); |
34268 | |
34269 | SDVTList NodeTys = DAG.getVTList(MVT::i8, VT); |
34270 | auto DivRemOpcode = OpcodeN0 == ISD::SDIVREM ? X86ISD::SDIVREM8_SEXT_HREG |
34271 | : X86ISD::UDIVREM8_ZEXT_HREG; |
34272 | SDValue R = DAG.getNode(DivRemOpcode, SDLoc(N), NodeTys, N0.getOperand(0), |
34273 | N0.getOperand(1)); |
34274 | DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0)); |
34275 | return R.getValue(1); |
34276 | } |
34277 | |
34278 | /// Convert a SEXT or ZEXT of a vector to a SIGN_EXTEND_VECTOR_INREG or |
34279 | /// ZERO_EXTEND_VECTOR_INREG, this requires the splitting (or concatenating |
34280 | /// with UNDEFs) of the input to vectors of the same size as the target type |
34281 | /// which then extends the lowest elements. |
34282 | static SDValue combineToExtendVectorInReg(SDNode *N, SelectionDAG &DAG, |
34283 | TargetLowering::DAGCombinerInfo &DCI, |
34284 | const X86Subtarget &Subtarget) { |
34285 | unsigned Opcode = N->getOpcode(); |
34286 | if (Opcode != ISD::SIGN_EXTEND && Opcode != ISD::ZERO_EXTEND) |
34287 | return SDValue(); |
34288 | if (!DCI.isBeforeLegalizeOps()) |
34289 | return SDValue(); |
34290 | if (!Subtarget.hasSSE2()) |
34291 | return SDValue(); |
34292 | |
34293 | SDValue N0 = N->getOperand(0); |
34294 | EVT VT = N->getValueType(0); |
34295 | EVT SVT = VT.getScalarType(); |
34296 | EVT InVT = N0.getValueType(); |
34297 | EVT InSVT = InVT.getScalarType(); |
34298 | |
34299 | // Input type must be a vector and we must be extending legal integer types. |
34300 | if (!VT.isVector()) |
34301 | return SDValue(); |
34302 | if (SVT != MVT::i64 && SVT != MVT::i32 && SVT != MVT::i16) |
34303 | return SDValue(); |
34304 | if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8) |
34305 | return SDValue(); |
34306 | |
34307 | // On AVX2+ targets, if the input/output types are both legal then we will be |
34308 | // able to use SIGN_EXTEND/ZERO_EXTEND directly. |
34309 | if (Subtarget.hasInt256() && DAG.getTargetLoweringInfo().isTypeLegal(VT) && |
34310 | DAG.getTargetLoweringInfo().isTypeLegal(InVT)) |
34311 | return SDValue(); |
34312 | |
34313 | SDLoc DL(N); |
34314 | |
34315 | auto ExtendVecSize = [&DAG](const SDLoc &DL, SDValue N, unsigned Size) { |
34316 | EVT InVT = N.getValueType(); |
34317 | EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(), |
34318 | Size / InVT.getScalarSizeInBits()); |
34319 | SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(), |
34320 | DAG.getUNDEF(InVT)); |
34321 | Opnds[0] = N; |
34322 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds); |
34323 | }; |
34324 | |
34325 | // If target-size is less than 128-bits, extend to a type that would extend |
34326 | // to 128 bits, extend that and extract the original target vector. |
34327 | if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits())) { |
34328 | unsigned Scale = 128 / VT.getSizeInBits(); |
34329 | EVT ExVT = |
34330 | EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits()); |
34331 | SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits()); |
34332 | SDValue SExt = DAG.getNode(Opcode, DL, ExVT, Ex); |
34333 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt, |
34334 | DAG.getIntPtrConstant(0, DL)); |
34335 | } |
34336 | |
34337 | // If target-size is 128-bits (or 256-bits on AVX2 target), then convert to |
34338 | // ISD::*_EXTEND_VECTOR_INREG which ensures lowering to X86ISD::V*EXT. |
34339 | // Also use this if we don't have SSE41 to allow the legalizer do its job. |
34340 | if (!Subtarget.hasSSE41() || VT.is128BitVector() || |
34341 | (VT.is256BitVector() && Subtarget.hasInt256()) || |
34342 | (VT.is512BitVector() && Subtarget.hasAVX512())) { |
34343 | SDValue ExOp = ExtendVecSize(DL, N0, VT.getSizeInBits()); |
34344 | return Opcode == ISD::SIGN_EXTEND |
34345 | ? DAG.getSignExtendVectorInReg(ExOp, DL, VT) |
34346 | : DAG.getZeroExtendVectorInReg(ExOp, DL, VT); |
34347 | } |
34348 | |
34349 | auto SplitAndExtendInReg = [&](unsigned SplitSize) { |
34350 | unsigned NumVecs = VT.getSizeInBits() / SplitSize; |
34351 | unsigned NumSubElts = SplitSize / SVT.getSizeInBits(); |
34352 | EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts); |
34353 | EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts); |
34354 | |
34355 | SmallVector<SDValue, 8> Opnds; |
34356 | for (unsigned i = 0, Offset = 0; i != NumVecs; ++i, Offset += NumSubElts) { |
34357 | SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0, |
34358 | DAG.getIntPtrConstant(Offset, DL)); |
34359 | SrcVec = ExtendVecSize(DL, SrcVec, SplitSize); |
34360 | SrcVec = Opcode == ISD::SIGN_EXTEND |
34361 | ? DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT) |
34362 | : DAG.getZeroExtendVectorInReg(SrcVec, DL, SubVT); |
34363 | Opnds.push_back(SrcVec); |
34364 | } |
34365 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds); |
34366 | }; |
34367 | |
34368 | // On pre-AVX2 targets, split into 128-bit nodes of |
34369 | // ISD::*_EXTEND_VECTOR_INREG. |
34370 | if (!Subtarget.hasInt256() && !(VT.getSizeInBits() % 128)) |
34371 | return SplitAndExtendInReg(128); |
34372 | |
34373 | // On pre-AVX512 targets, split into 256-bit nodes of |
34374 | // ISD::*_EXTEND_VECTOR_INREG. |
34375 | if (!Subtarget.hasAVX512() && !(VT.getSizeInBits() % 256)) |
34376 | return SplitAndExtendInReg(256); |
34377 | |
34378 | return SDValue(); |
34379 | } |
34380 | |
34381 | static SDValue combineSext(SDNode *N, SelectionDAG &DAG, |
34382 | TargetLowering::DAGCombinerInfo &DCI, |
34383 | const X86Subtarget &Subtarget) { |
34384 | SDValue N0 = N->getOperand(0); |
34385 | EVT VT = N->getValueType(0); |
34386 | EVT InVT = N0.getValueType(); |
34387 | SDLoc DL(N); |
34388 | |
34389 | if (SDValue DivRem8 = getDivRem8(N, DAG)) |
34390 | return DivRem8; |
34391 | |
34392 | if (!DCI.isBeforeLegalizeOps()) { |
34393 | if (InVT == MVT::i1) { |
34394 | SDValue Zero = DAG.getConstant(0, DL, VT); |
34395 | SDValue AllOnes = DAG.getAllOnesConstant(DL, VT); |
34396 | return DAG.getSelect(DL, VT, N0, AllOnes, Zero); |
34397 | } |
34398 | return SDValue(); |
34399 | } |
34400 | |
34401 | if (InVT == MVT::i1 && N0.getOpcode() == ISD::XOR && |
34402 | isAllOnesConstant(N0.getOperand(1)) && N0.hasOneUse()) { |
34403 | // Invert and sign-extend a boolean is the same as zero-extend and subtract |
34404 | // 1 because 0 becomes -1 and 1 becomes 0. The subtract is efficiently |
34405 | // lowered with an LEA or a DEC. This is the same as: select Bool, 0, -1. |
34406 | // sext (xor Bool, -1) --> sub (zext Bool), 1 |
34407 | SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)); |
34408 | return DAG.getNode(ISD::SUB, DL, VT, Zext, DAG.getConstant(1, DL, VT)); |
34409 | } |
34410 | |
34411 | if (SDValue V = combineToExtendVectorInReg(N, DAG, DCI, Subtarget)) |
34412 | return V; |
34413 | |
34414 | if (Subtarget.hasAVX() && VT.is256BitVector()) |
34415 | if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget)) |
34416 | return R; |
34417 | |
34418 | if (SDValue NewAdd = promoteExtBeforeAdd(N, DAG, Subtarget)) |
34419 | return NewAdd; |
34420 | |
34421 | return SDValue(); |
34422 | } |
34423 | |
34424 | static SDValue combineFMA(SDNode *N, SelectionDAG &DAG, |
34425 | const X86Subtarget &Subtarget) { |
34426 | SDLoc dl(N); |
34427 | EVT VT = N->getValueType(0); |
34428 | |
34429 | // Let legalize expand this if it isn't a legal type yet. |
34430 | if (!DAG.getTargetLoweringInfo().isTypeLegal(VT)) |
34431 | return SDValue(); |
34432 | |
34433 | EVT ScalarVT = VT.getScalarType(); |
34434 | if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) || !Subtarget.hasAnyFMA()) |
34435 | return SDValue(); |
34436 | |
34437 | SDValue A = N->getOperand(0); |
34438 | SDValue B = N->getOperand(1); |
34439 | SDValue C = N->getOperand(2); |
34440 | |
34441 | auto invertIfNegative = [](SDValue &V) { |
34442 | if (SDValue NegVal = isFNEG(V.getNode())) { |
34443 | V = NegVal; |
34444 | return true; |
34445 | } |
34446 | return false; |
34447 | }; |
34448 | |
34449 | // Do not convert the passthru input of scalar intrinsics. |
34450 | // FIXME: We could allow negations of the lower element only. |
34451 | bool NegA = N->getOpcode() != X86ISD::FMADDS1_RND && invertIfNegative(A); |
34452 | bool NegB = invertIfNegative(B); |
34453 | bool NegC = N->getOpcode() != X86ISD::FMADDS3_RND && invertIfNegative(C); |
34454 | |
34455 | // Negative multiplication when NegA xor NegB |
34456 | bool NegMul = (NegA != NegB); |
34457 | |
34458 | unsigned NewOpcode; |
34459 | if (!NegMul) |
34460 | NewOpcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB; |
34461 | else |
34462 | NewOpcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB; |
34463 | |
34464 | |
34465 | if (N->getOpcode() == X86ISD::FMADD_RND) { |
34466 | switch (NewOpcode) { |
34467 | case X86ISD::FMADD: NewOpcode = X86ISD::FMADD_RND; break; |
34468 | case X86ISD::FMSUB: NewOpcode = X86ISD::FMSUB_RND; break; |
34469 | case X86ISD::FNMADD: NewOpcode = X86ISD::FNMADD_RND; break; |
34470 | case X86ISD::FNMSUB: NewOpcode = X86ISD::FNMSUB_RND; break; |
34471 | } |
34472 | } else if (N->getOpcode() == X86ISD::FMADDS1_RND) { |
34473 | switch (NewOpcode) { |
34474 | case X86ISD::FMADD: NewOpcode = X86ISD::FMADDS1_RND; break; |
34475 | case X86ISD::FMSUB: NewOpcode = X86ISD::FMSUBS1_RND; break; |
34476 | case X86ISD::FNMADD: NewOpcode = X86ISD::FNMADDS1_RND; break; |
34477 | case X86ISD::FNMSUB: NewOpcode = X86ISD::FNMSUBS1_RND; break; |
34478 | } |
34479 | } else if (N->getOpcode() == X86ISD::FMADDS3_RND) { |
34480 | switch (NewOpcode) { |
34481 | case X86ISD::FMADD: NewOpcode = X86ISD::FMADDS3_RND; break; |
34482 | case X86ISD::FMSUB: NewOpcode = X86ISD::FMSUBS3_RND; break; |
34483 | case X86ISD::FNMADD: NewOpcode = X86ISD::FNMADDS3_RND; break; |
34484 | case X86ISD::FNMSUB: NewOpcode = X86ISD::FNMSUBS3_RND; break; |
34485 | } |
34486 | } else { |
34487 | assert((N->getOpcode() == X86ISD::FMADD || N->getOpcode() == ISD::FMA) &&(((N->getOpcode() == X86ISD::FMADD || N->getOpcode() == ISD::FMA) && "Unexpected opcode!") ? static_cast< void> (0) : __assert_fail ("(N->getOpcode() == X86ISD::FMADD || N->getOpcode() == ISD::FMA) && \"Unexpected opcode!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 34488, __PRETTY_FUNCTION__)) |
34488 | "Unexpected opcode!")(((N->getOpcode() == X86ISD::FMADD || N->getOpcode() == ISD::FMA) && "Unexpected opcode!") ? static_cast< void> (0) : __assert_fail ("(N->getOpcode() == X86ISD::FMADD || N->getOpcode() == ISD::FMA) && \"Unexpected opcode!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 34488, __PRETTY_FUNCTION__)); |
34489 | return DAG.getNode(NewOpcode, dl, VT, A, B, C); |
34490 | } |
34491 | |
34492 | return DAG.getNode(NewOpcode, dl, VT, A, B, C, N->getOperand(3)); |
34493 | } |
34494 | |
34495 | static SDValue combineZext(SDNode *N, SelectionDAG &DAG, |
34496 | TargetLowering::DAGCombinerInfo &DCI, |
34497 | const X86Subtarget &Subtarget) { |
34498 | // (i32 zext (and (i8 x86isd::setcc_carry), 1)) -> |
34499 | // (and (i32 x86isd::setcc_carry), 1) |
34500 | // This eliminates the zext. This transformation is necessary because |
34501 | // ISD::SETCC is always legalized to i8. |
34502 | SDLoc dl(N); |
34503 | SDValue N0 = N->getOperand(0); |
34504 | EVT VT = N->getValueType(0); |
34505 | |
34506 | if (N0.getOpcode() == ISD::AND && |
34507 | N0.hasOneUse() && |
34508 | N0.getOperand(0).hasOneUse()) { |
34509 | SDValue N00 = N0.getOperand(0); |
34510 | if (N00.getOpcode() == X86ISD::SETCC_CARRY) { |
34511 | if (!isOneConstant(N0.getOperand(1))) |
34512 | return SDValue(); |
34513 | return DAG.getNode(ISD::AND, dl, VT, |
34514 | DAG.getNode(X86ISD::SETCC_CARRY, dl, VT, |
34515 | N00.getOperand(0), N00.getOperand(1)), |
34516 | DAG.getConstant(1, dl, VT)); |
34517 | } |
34518 | } |
34519 | |
34520 | if (N0.getOpcode() == ISD::TRUNCATE && |
34521 | N0.hasOneUse() && |
34522 | N0.getOperand(0).hasOneUse()) { |
34523 | SDValue N00 = N0.getOperand(0); |
34524 | if (N00.getOpcode() == X86ISD::SETCC_CARRY) { |
34525 | return DAG.getNode(ISD::AND, dl, VT, |
34526 | DAG.getNode(X86ISD::SETCC_CARRY, dl, VT, |
34527 | N00.getOperand(0), N00.getOperand(1)), |
34528 | DAG.getConstant(1, dl, VT)); |
34529 | } |
34530 | } |
34531 | |
34532 | if (SDValue V = combineToExtendVectorInReg(N, DAG, DCI, Subtarget)) |
34533 | return V; |
34534 | |
34535 | if (VT.is256BitVector()) |
34536 | if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget)) |
34537 | return R; |
34538 | |
34539 | if (SDValue DivRem8 = getDivRem8(N, DAG)) |
34540 | return DivRem8; |
34541 | |
34542 | if (SDValue NewAdd = promoteExtBeforeAdd(N, DAG, Subtarget)) |
34543 | return NewAdd; |
34544 | |
34545 | if (SDValue R = combineOrCmpEqZeroToCtlzSrl(N, DAG, DCI, Subtarget)) |
34546 | return R; |
34547 | |
34548 | return SDValue(); |
34549 | } |
34550 | |
34551 | /// Try to map a 128-bit or larger integer comparison to vector instructions |
34552 | /// before type legalization splits it up into chunks. |
34553 | static SDValue combineVectorSizedSetCCEquality(SDNode *SetCC, SelectionDAG &DAG, |
34554 | const X86Subtarget &Subtarget) { |
34555 | ISD::CondCode CC = cast<CondCodeSDNode>(SetCC->getOperand(2))->get(); |
34556 | assert((CC == ISD::SETNE || CC == ISD::SETEQ) && "Bad comparison predicate")(((CC == ISD::SETNE || CC == ISD::SETEQ) && "Bad comparison predicate" ) ? static_cast<void> (0) : __assert_fail ("(CC == ISD::SETNE || CC == ISD::SETEQ) && \"Bad comparison predicate\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 34556, __PRETTY_FUNCTION__)); |
34557 | |
34558 | // We're looking for an oversized integer equality comparison, but ignore a |
34559 | // comparison with zero because that gets special treatment in EmitTest(). |
34560 | SDValue X = SetCC->getOperand(0); |
34561 | SDValue Y = SetCC->getOperand(1); |
34562 | EVT OpVT = X.getValueType(); |
34563 | unsigned OpSize = OpVT.getSizeInBits(); |
34564 | if (!OpVT.isScalarInteger() || OpSize < 128 || isNullConstant(Y)) |
34565 | return SDValue(); |
34566 | |
34567 | // TODO: Use PXOR + PTEST for SSE4.1 or later? |
34568 | // TODO: Add support for AVX-512. |
34569 | EVT VT = SetCC->getValueType(0); |
34570 | SDLoc DL(SetCC); |
34571 | if ((OpSize == 128 && Subtarget.hasSSE2()) || |
34572 | (OpSize == 256 && Subtarget.hasAVX2())) { |
34573 | EVT VecVT = OpSize == 128 ? MVT::v16i8 : MVT::v32i8; |
34574 | SDValue VecX = DAG.getBitcast(VecVT, X); |
34575 | SDValue VecY = DAG.getBitcast(VecVT, Y); |
34576 | |
34577 | // If all bytes match (bitmask is 0x(FFFF)FFFF), that's equality. |
34578 | // setcc i128 X, Y, eq --> setcc (pmovmskb (pcmpeqb X, Y)), 0xFFFF, eq |
34579 | // setcc i128 X, Y, ne --> setcc (pmovmskb (pcmpeqb X, Y)), 0xFFFF, ne |
34580 | // setcc i256 X, Y, eq --> setcc (vpmovmskb (vpcmpeqb X, Y)), 0xFFFFFFFF, eq |
34581 | // setcc i256 X, Y, ne --> setcc (vpmovmskb (vpcmpeqb X, Y)), 0xFFFFFFFF, ne |
34582 | SDValue Cmp = DAG.getNode(X86ISD::PCMPEQ, DL, VecVT, VecX, VecY); |
34583 | SDValue MovMsk = DAG.getNode(X86ISD::MOVMSK, DL, MVT::i32, Cmp); |
34584 | SDValue FFFFs = DAG.getConstant(OpSize == 128 ? 0xFFFF : 0xFFFFFFFF, DL, |
34585 | MVT::i32); |
34586 | return DAG.getSetCC(DL, VT, MovMsk, FFFFs, CC); |
34587 | } |
34588 | |
34589 | return SDValue(); |
34590 | } |
34591 | |
34592 | static SDValue combineSetCC(SDNode *N, SelectionDAG &DAG, |
34593 | const X86Subtarget &Subtarget) { |
34594 | ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get(); |
34595 | SDValue LHS = N->getOperand(0); |
34596 | SDValue RHS = N->getOperand(1); |
34597 | EVT VT = N->getValueType(0); |
34598 | SDLoc DL(N); |
34599 | |
34600 | if (CC == ISD::SETNE || CC == ISD::SETEQ) { |
34601 | EVT OpVT = LHS.getValueType(); |
34602 | // 0-x == y --> x+y == 0 |
34603 | // 0-x != y --> x+y != 0 |
34604 | if (LHS.getOpcode() == ISD::SUB && isNullConstant(LHS.getOperand(0)) && |
34605 | LHS.hasOneUse()) { |
34606 | SDValue Add = DAG.getNode(ISD::ADD, DL, OpVT, RHS, LHS.getOperand(1)); |
34607 | return DAG.getSetCC(DL, VT, Add, DAG.getConstant(0, DL, OpVT), CC); |
34608 | } |
34609 | // x == 0-y --> x+y == 0 |
34610 | // x != 0-y --> x+y != 0 |
34611 | if (RHS.getOpcode() == ISD::SUB && isNullConstant(RHS.getOperand(0)) && |
34612 | RHS.hasOneUse()) { |
34613 | SDValue Add = DAG.getNode(ISD::ADD, DL, OpVT, LHS, RHS.getOperand(1)); |
34614 | return DAG.getSetCC(DL, VT, Add, DAG.getConstant(0, DL, OpVT), CC); |
34615 | } |
34616 | |
34617 | if (SDValue V = combineVectorSizedSetCCEquality(N, DAG, Subtarget)) |
34618 | return V; |
34619 | } |
34620 | |
34621 | if (VT.getScalarType() == MVT::i1 && |
34622 | (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) { |
34623 | bool IsSEXT0 = |
34624 | (LHS.getOpcode() == ISD::SIGN_EXTEND) && |
34625 | (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1); |
34626 | bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode()); |
34627 | |
34628 | if (!IsSEXT0 || !IsVZero1) { |
34629 | // Swap the operands and update the condition code. |
34630 | std::swap(LHS, RHS); |
34631 | CC = ISD::getSetCCSwappedOperands(CC); |
34632 | |
34633 | IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) && |
34634 | (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1); |
34635 | IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode()); |
34636 | } |
34637 | |
34638 | if (IsSEXT0 && IsVZero1) { |
34639 | assert(VT == LHS.getOperand(0).getValueType() &&((VT == LHS.getOperand(0).getValueType() && "Uexpected operand type" ) ? static_cast<void> (0) : __assert_fail ("VT == LHS.getOperand(0).getValueType() && \"Uexpected operand type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 34640, __PRETTY_FUNCTION__)) |
34640 | "Uexpected operand type")((VT == LHS.getOperand(0).getValueType() && "Uexpected operand type" ) ? static_cast<void> (0) : __assert_fail ("VT == LHS.getOperand(0).getValueType() && \"Uexpected operand type\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 34640, __PRETTY_FUNCTION__)); |
34641 | if (CC == ISD::SETGT) |
34642 | return DAG.getConstant(0, DL, VT); |
34643 | if (CC == ISD::SETLE) |
34644 | return DAG.getConstant(1, DL, VT); |
34645 | if (CC == ISD::SETEQ || CC == ISD::SETGE) |
34646 | return DAG.getNOT(DL, LHS.getOperand(0), VT); |
34647 | |
34648 | assert((CC == ISD::SETNE || CC == ISD::SETLT) &&(((CC == ISD::SETNE || CC == ISD::SETLT) && "Unexpected condition code!" ) ? static_cast<void> (0) : __assert_fail ("(CC == ISD::SETNE || CC == ISD::SETLT) && \"Unexpected condition code!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 34649, __PRETTY_FUNCTION__)) |
34649 | "Unexpected condition code!")(((CC == ISD::SETNE || CC == ISD::SETLT) && "Unexpected condition code!" ) ? static_cast<void> (0) : __assert_fail ("(CC == ISD::SETNE || CC == ISD::SETLT) && \"Unexpected condition code!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 34649, __PRETTY_FUNCTION__)); |
34650 | return LHS.getOperand(0); |
34651 | } |
34652 | } |
34653 | |
34654 | // For an SSE1-only target, lower a comparison of v4f32 to X86ISD::CMPP early |
34655 | // to avoid scalarization via legalization because v4i32 is not a legal type. |
34656 | if (Subtarget.hasSSE1() && !Subtarget.hasSSE2() && VT == MVT::v4i32 && |
34657 | LHS.getValueType() == MVT::v4f32) |
34658 | return LowerVSETCC(SDValue(N, 0), Subtarget, DAG); |
34659 | |
34660 | return SDValue(); |
34661 | } |
34662 | |
34663 | static SDValue combineGatherScatter(SDNode *N, SelectionDAG &DAG) { |
34664 | SDLoc DL(N); |
34665 | // Gather and Scatter instructions use k-registers for masks. The type of |
34666 | // the masks is v*i1. So the mask will be truncated anyway. |
34667 | // The SIGN_EXTEND_INREG my be dropped. |
34668 | SDValue Mask = N->getOperand(2); |
34669 | if (Mask.getOpcode() == ISD::SIGN_EXTEND_INREG) { |
34670 | SmallVector<SDValue, 5> NewOps(N->op_begin(), N->op_end()); |
34671 | NewOps[2] = Mask.getOperand(0); |
34672 | DAG.UpdateNodeOperands(N, NewOps); |
34673 | } |
34674 | return SDValue(); |
34675 | } |
34676 | |
34677 | // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT |
34678 | static SDValue combineX86SetCC(SDNode *N, SelectionDAG &DAG, |
34679 | const X86Subtarget &Subtarget) { |
34680 | SDLoc DL(N); |
34681 | X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0)); |
34682 | SDValue EFLAGS = N->getOperand(1); |
34683 | |
34684 | // Try to simplify the EFLAGS and condition code operands. |
34685 | if (SDValue Flags = combineSetCCEFLAGS(EFLAGS, CC, DAG)) |
34686 | return getSETCC(CC, Flags, DL, DAG); |
34687 | |
34688 | return SDValue(); |
34689 | } |
34690 | |
34691 | /// Optimize branch condition evaluation. |
34692 | static SDValue combineBrCond(SDNode *N, SelectionDAG &DAG, |
34693 | const X86Subtarget &Subtarget) { |
34694 | SDLoc DL(N); |
34695 | SDValue EFLAGS = N->getOperand(3); |
34696 | X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2)); |
34697 | |
34698 | // Try to simplify the EFLAGS and condition code operands. |
34699 | // Make sure to not keep references to operands, as combineSetCCEFLAGS can |
34700 | // RAUW them under us. |
34701 | if (SDValue Flags = combineSetCCEFLAGS(EFLAGS, CC, DAG)) { |
34702 | SDValue Cond = DAG.getConstant(CC, DL, MVT::i8); |
34703 | return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), N->getOperand(0), |
34704 | N->getOperand(1), Cond, Flags); |
34705 | } |
34706 | |
34707 | return SDValue(); |
34708 | } |
34709 | |
34710 | static SDValue combineVectorCompareAndMaskUnaryOp(SDNode *N, |
34711 | SelectionDAG &DAG) { |
34712 | // Take advantage of vector comparisons producing 0 or -1 in each lane to |
34713 | // optimize away operation when it's from a constant. |
34714 | // |
34715 | // The general transformation is: |
34716 | // UNARYOP(AND(VECTOR_CMP(x,y), constant)) --> |
34717 | // AND(VECTOR_CMP(x,y), constant2) |
34718 | // constant2 = UNARYOP(constant) |
34719 | |
34720 | // Early exit if this isn't a vector operation, the operand of the |
34721 | // unary operation isn't a bitwise AND, or if the sizes of the operations |
34722 | // aren't the same. |
34723 | EVT VT = N->getValueType(0); |
34724 | if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND || |
34725 | N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC || |
34726 | VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits()) |
34727 | return SDValue(); |
34728 | |
34729 | // Now check that the other operand of the AND is a constant. We could |
34730 | // make the transformation for non-constant splats as well, but it's unclear |
34731 | // that would be a benefit as it would not eliminate any operations, just |
34732 | // perform one more step in scalar code before moving to the vector unit. |
34733 | if (BuildVectorSDNode *BV = |
34734 | dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) { |
34735 | // Bail out if the vector isn't a constant. |
34736 | if (!BV->isConstant()) |
34737 | return SDValue(); |
34738 | |
34739 | // Everything checks out. Build up the new and improved node. |
34740 | SDLoc DL(N); |
34741 | EVT IntVT = BV->getValueType(0); |
34742 | // Create a new constant of the appropriate type for the transformed |
34743 | // DAG. |
34744 | SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0)); |
34745 | // The AND node needs bitcasts to/from an integer vector type around it. |
34746 | SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst); |
34747 | SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT, |
34748 | N->getOperand(0)->getOperand(0), MaskConst); |
34749 | SDValue Res = DAG.getBitcast(VT, NewAnd); |
34750 | return Res; |
34751 | } |
34752 | |
34753 | return SDValue(); |
34754 | } |
34755 | |
34756 | static SDValue combineUIntToFP(SDNode *N, SelectionDAG &DAG, |
34757 | const X86Subtarget &Subtarget) { |
34758 | SDValue Op0 = N->getOperand(0); |
34759 | EVT VT = N->getValueType(0); |
34760 | EVT InVT = Op0.getValueType(); |
34761 | EVT InSVT = InVT.getScalarType(); |
34762 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
34763 | |
34764 | // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32)) |
34765 | // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32)) |
34766 | if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) { |
34767 | SDLoc dl(N); |
34768 | EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, |
34769 | InVT.getVectorNumElements()); |
34770 | SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0); |
34771 | |
34772 | if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT)) |
34773 | return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P); |
34774 | |
34775 | return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P); |
34776 | } |
34777 | |
34778 | // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't |
34779 | // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform |
34780 | // the optimization here. |
34781 | if (DAG.SignBitIsZero(Op0)) |
34782 | return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, Op0); |
34783 | |
34784 | return SDValue(); |
34785 | } |
34786 | |
34787 | static SDValue combineSIntToFP(SDNode *N, SelectionDAG &DAG, |
34788 | const X86Subtarget &Subtarget) { |
34789 | // First try to optimize away the conversion entirely when it's |
34790 | // conditionally from a constant. Vectors only. |
34791 | if (SDValue Res = combineVectorCompareAndMaskUnaryOp(N, DAG)) |
34792 | return Res; |
34793 | |
34794 | // Now move on to more general possibilities. |
34795 | SDValue Op0 = N->getOperand(0); |
34796 | EVT VT = N->getValueType(0); |
34797 | EVT InVT = Op0.getValueType(); |
34798 | EVT InSVT = InVT.getScalarType(); |
34799 | |
34800 | // SINT_TO_FP(vXi1) -> SINT_TO_FP(SEXT(vXi1 to vXi32)) |
34801 | // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32)) |
34802 | // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32)) |
34803 | if (InVT.isVector() && |
34804 | (InSVT == MVT::i8 || InSVT == MVT::i16 || |
34805 | (InSVT == MVT::i1 && !DAG.getTargetLoweringInfo().isTypeLegal(InVT)))) { |
34806 | SDLoc dl(N); |
34807 | EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, |
34808 | InVT.getVectorNumElements()); |
34809 | SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0); |
34810 | return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P); |
34811 | } |
34812 | |
34813 | // Without AVX512DQ we only support i64 to float scalar conversion. For both |
34814 | // vectors and scalars, see if we know that the upper bits are all the sign |
34815 | // bit, in which case we can truncate the input to i32 and convert from that. |
34816 | if (InVT.getScalarSizeInBits() > 32 && !Subtarget.hasDQI()) { |
34817 | unsigned BitWidth = InVT.getScalarSizeInBits(); |
34818 | unsigned NumSignBits = DAG.ComputeNumSignBits(Op0); |
34819 | if (NumSignBits >= (BitWidth - 31)) { |
34820 | EVT TruncVT = EVT::getIntegerVT(*DAG.getContext(), 32); |
34821 | if (InVT.isVector()) |
34822 | TruncVT = EVT::getVectorVT(*DAG.getContext(), TruncVT, |
34823 | InVT.getVectorNumElements()); |
34824 | SDLoc dl(N); |
34825 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, TruncVT, Op0); |
34826 | return DAG.getNode(ISD::SINT_TO_FP, dl, VT, Trunc); |
34827 | } |
34828 | } |
34829 | |
34830 | // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have |
34831 | // a 32-bit target where SSE doesn't support i64->FP operations. |
34832 | if (!Subtarget.useSoftFloat() && Op0.getOpcode() == ISD::LOAD) { |
34833 | LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode()); |
34834 | EVT LdVT = Ld->getValueType(0); |
34835 | |
34836 | // This transformation is not supported if the result type is f16 or f128. |
34837 | if (VT == MVT::f16 || VT == MVT::f128) |
34838 | return SDValue(); |
34839 | |
34840 | if (!Ld->isVolatile() && !VT.isVector() && |
34841 | ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() && |
34842 | !Subtarget.is64Bit() && LdVT == MVT::i64) { |
34843 | SDValue FILDChain = Subtarget.getTargetLowering()->BuildFILD( |
34844 | SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG); |
34845 | DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1)); |
34846 | return FILDChain; |
34847 | } |
34848 | } |
34849 | return SDValue(); |
34850 | } |
34851 | |
34852 | // Optimize RES, EFLAGS = X86ISD::ADD LHS, RHS |
34853 | static SDValue combineX86ADD(SDNode *N, SelectionDAG &DAG, |
34854 | X86TargetLowering::DAGCombinerInfo &DCI) { |
34855 | // When legalizing carry, we create carries via add X, -1 |
34856 | // If that comes from an actual carry, via setcc, we use the |
34857 | // carry directly. |
34858 | if (isAllOnesConstant(N->getOperand(1)) && N->hasAnyUseOfValue(1)) { |
34859 | SDValue Carry = N->getOperand(0); |
34860 | while (Carry.getOpcode() == ISD::TRUNCATE || |
34861 | Carry.getOpcode() == ISD::ZERO_EXTEND || |
34862 | Carry.getOpcode() == ISD::SIGN_EXTEND || |
34863 | Carry.getOpcode() == ISD::ANY_EXTEND || |
34864 | (Carry.getOpcode() == ISD::AND && |
34865 | isOneConstant(Carry.getOperand(1)))) |
34866 | Carry = Carry.getOperand(0); |
34867 | |
34868 | if (Carry.getOpcode() == X86ISD::SETCC || |
34869 | Carry.getOpcode() == X86ISD::SETCC_CARRY) { |
34870 | if (Carry.getConstantOperandVal(0) == X86::COND_B) |
34871 | return DCI.CombineTo(N, SDValue(N, 0), Carry.getOperand(1)); |
34872 | } |
34873 | } |
34874 | |
34875 | return SDValue(); |
34876 | } |
34877 | |
34878 | // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS |
34879 | static SDValue combineADC(SDNode *N, SelectionDAG &DAG, |
34880 | X86TargetLowering::DAGCombinerInfo &DCI) { |
34881 | // If the LHS and RHS of the ADC node are zero, then it can't overflow and |
34882 | // the result is either zero or one (depending on the input carry bit). |
34883 | // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1. |
34884 | if (X86::isZeroNode(N->getOperand(0)) && |
34885 | X86::isZeroNode(N->getOperand(1)) && |
34886 | // We don't have a good way to replace an EFLAGS use, so only do this when |
34887 | // dead right now. |
34888 | SDValue(N, 1).use_empty()) { |
34889 | SDLoc DL(N); |
34890 | EVT VT = N->getValueType(0); |
34891 | SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1)); |
34892 | SDValue Res1 = DAG.getNode(ISD::AND, DL, VT, |
34893 | DAG.getNode(X86ISD::SETCC_CARRY, DL, VT, |
34894 | DAG.getConstant(X86::COND_B, DL, |
34895 | MVT::i8), |
34896 | N->getOperand(2)), |
34897 | DAG.getConstant(1, DL, VT)); |
34898 | return DCI.CombineTo(N, Res1, CarryOut); |
34899 | } |
34900 | |
34901 | return SDValue(); |
34902 | } |
34903 | |
34904 | /// Materialize "setb reg" as "sbb reg,reg", since it produces an all-ones bit |
34905 | /// which is more useful than 0/1 in some cases. |
34906 | static SDValue materializeSBB(SDNode *N, SDValue EFLAGS, SelectionDAG &DAG) { |
34907 | SDLoc DL(N); |
34908 | // "Condition code B" is also known as "the carry flag" (CF). |
34909 | SDValue CF = DAG.getConstant(X86::COND_B, DL, MVT::i8); |
34910 | SDValue SBB = DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8, CF, EFLAGS); |
34911 | MVT VT = N->getSimpleValueType(0); |
34912 | if (VT == MVT::i8) |
34913 | return DAG.getNode(ISD::AND, DL, VT, SBB, DAG.getConstant(1, DL, VT)); |
34914 | |
34915 | assert(VT == MVT::i1 && "Unexpected type for SETCC node")((VT == MVT::i1 && "Unexpected type for SETCC node") ? static_cast<void> (0) : __assert_fail ("VT == MVT::i1 && \"Unexpected type for SETCC node\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 34915, __PRETTY_FUNCTION__)); |
34916 | return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, SBB); |
34917 | } |
34918 | |
34919 | /// If this is an add or subtract where one operand is produced by a cmp+setcc, |
34920 | /// then try to convert it to an ADC or SBB. This replaces TEST+SET+{ADD/SUB} |
34921 | /// with CMP+{ADC, SBB}. |
34922 | static SDValue combineAddOrSubToADCOrSBB(SDNode *N, SelectionDAG &DAG) { |
34923 | bool IsSub = N->getOpcode() == ISD::SUB; |
34924 | SDValue X = N->getOperand(0); |
34925 | SDValue Y = N->getOperand(1); |
34926 | |
34927 | // If this is an add, canonicalize a zext operand to the RHS. |
34928 | // TODO: Incomplete? What if both sides are zexts? |
34929 | if (!IsSub && X.getOpcode() == ISD::ZERO_EXTEND && |
34930 | Y.getOpcode() != ISD::ZERO_EXTEND) |
34931 | std::swap(X, Y); |
34932 | |
34933 | // Look through a one-use zext. |
34934 | bool PeekedThroughZext = false; |
34935 | if (Y.getOpcode() == ISD::ZERO_EXTEND && Y.hasOneUse()) { |
34936 | Y = Y.getOperand(0); |
34937 | PeekedThroughZext = true; |
34938 | } |
34939 | |
34940 | // If this is an add, canonicalize a setcc operand to the RHS. |
34941 | // TODO: Incomplete? What if both sides are setcc? |
34942 | // TODO: Should we allow peeking through a zext of the other operand? |
34943 | if (!IsSub && !PeekedThroughZext && X.getOpcode() == X86ISD::SETCC && |
34944 | Y.getOpcode() != X86ISD::SETCC) |
34945 | std::swap(X, Y); |
34946 | |
34947 | if (Y.getOpcode() != X86ISD::SETCC || !Y.hasOneUse()) |
34948 | return SDValue(); |
34949 | |
34950 | SDLoc DL(N); |
34951 | EVT VT = N->getValueType(0); |
34952 | X86::CondCode CC = (X86::CondCode)Y.getConstantOperandVal(0); |
34953 | |
34954 | if (CC == X86::COND_B) { |
34955 | // X + SETB Z --> X + (mask SBB Z, Z) |
34956 | // X - SETB Z --> X - (mask SBB Z, Z) |
34957 | // TODO: Produce ADC/SBB here directly and avoid SETCC_CARRY? |
34958 | SDValue SBB = materializeSBB(Y.getNode(), Y.getOperand(1), DAG); |
34959 | if (SBB.getValueSizeInBits() != VT.getSizeInBits()) |
34960 | SBB = DAG.getZExtOrTrunc(SBB, DL, VT); |
34961 | return DAG.getNode(IsSub ? ISD::SUB : ISD::ADD, DL, VT, X, SBB); |
34962 | } |
34963 | |
34964 | if (CC == X86::COND_A) { |
34965 | SDValue EFLAGS = Y->getOperand(1); |
34966 | // Try to convert COND_A into COND_B in an attempt to facilitate |
34967 | // materializing "setb reg". |
34968 | // |
34969 | // Do not flip "e > c", where "c" is a constant, because Cmp instruction |
34970 | // cannot take an immediate as its first operand. |
34971 | // |
34972 | if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() && |
34973 | EFLAGS.getValueType().isInteger() && |
34974 | !isa<ConstantSDNode>(EFLAGS.getOperand(1))) { |
34975 | SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS), |
34976 | EFLAGS.getNode()->getVTList(), |
34977 | EFLAGS.getOperand(1), EFLAGS.getOperand(0)); |
34978 | SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo()); |
34979 | SDValue SBB = materializeSBB(Y.getNode(), NewEFLAGS, DAG); |
34980 | if (SBB.getValueSizeInBits() != VT.getSizeInBits()) |
34981 | SBB = DAG.getZExtOrTrunc(SBB, DL, VT); |
34982 | return DAG.getNode(IsSub ? ISD::SUB : ISD::ADD, DL, VT, X, SBB); |
34983 | } |
34984 | } |
34985 | |
34986 | if (CC != X86::COND_E && CC != X86::COND_NE) |
34987 | return SDValue(); |
34988 | |
34989 | SDValue Cmp = Y.getOperand(1); |
34990 | if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() || |
34991 | !X86::isZeroNode(Cmp.getOperand(1)) || |
34992 | !Cmp.getOperand(0).getValueType().isInteger()) |
34993 | return SDValue(); |
34994 | |
34995 | SDValue Z = Cmp.getOperand(0); |
34996 | EVT ZVT = Z.getValueType(); |
34997 | |
34998 | // If X is -1 or 0, then we have an opportunity to avoid constants required in |
34999 | // the general case below. |
35000 | if (auto *ConstantX = dyn_cast<ConstantSDNode>(X)) { |
35001 | // 'neg' sets the carry flag when Z != 0, so create 0 or -1 using 'sbb' with |
35002 | // fake operands: |
35003 | // 0 - (Z != 0) --> sbb %eax, %eax, (neg Z) |
35004 | // -1 + (Z == 0) --> sbb %eax, %eax, (neg Z) |
35005 | if ((IsSub && CC == X86::COND_NE && ConstantX->isNullValue()) || |
35006 | (!IsSub && CC == X86::COND_E && ConstantX->isAllOnesValue())) { |
35007 | SDValue Zero = DAG.getConstant(0, DL, ZVT); |
35008 | SDVTList X86SubVTs = DAG.getVTList(ZVT, MVT::i32); |
35009 | SDValue Neg = DAG.getNode(X86ISD::SUB, DL, X86SubVTs, Zero, Z); |
35010 | return DAG.getNode(X86ISD::SETCC_CARRY, DL, VT, |
35011 | DAG.getConstant(X86::COND_B, DL, MVT::i8), |
35012 | SDValue(Neg.getNode(), 1)); |
35013 | } |
35014 | |
35015 | // cmp with 1 sets the carry flag when Z == 0, so create 0 or -1 using 'sbb' |
35016 | // with fake operands: |
35017 | // 0 - (Z == 0) --> sbb %eax, %eax, (cmp Z, 1) |
35018 | // -1 + (Z != 0) --> sbb %eax, %eax, (cmp Z, 1) |
35019 | if ((IsSub && CC == X86::COND_E && ConstantX->isNullValue()) || |
35020 | (!IsSub && CC == X86::COND_NE && ConstantX->isAllOnesValue())) { |
35021 | SDValue One = DAG.getConstant(1, DL, ZVT); |
35022 | SDValue Cmp1 = DAG.getNode(X86ISD::CMP, DL, MVT::i32, Z, One); |
35023 | return DAG.getNode(X86ISD::SETCC_CARRY, DL, VT, |
35024 | DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp1); |
35025 | } |
35026 | } |
35027 | |
35028 | // (cmp Z, 1) sets the carry flag if Z is 0. |
35029 | SDValue One = DAG.getConstant(1, DL, ZVT); |
35030 | SDValue Cmp1 = DAG.getNode(X86ISD::CMP, DL, MVT::i32, Z, One); |
35031 | |
35032 | // Add the flags type for ADC/SBB nodes. |
35033 | SDVTList VTs = DAG.getVTList(VT, MVT::i32); |
35034 | |
35035 | // X - (Z != 0) --> sub X, (zext(setne Z, 0)) --> adc X, -1, (cmp Z, 1) |
35036 | // X + (Z != 0) --> add X, (zext(setne Z, 0)) --> sbb X, -1, (cmp Z, 1) |
35037 | if (CC == X86::COND_NE) |
35038 | return DAG.getNode(IsSub ? X86ISD::ADC : X86ISD::SBB, DL, VTs, X, |
35039 | DAG.getConstant(-1ULL, DL, VT), Cmp1); |
35040 | |
35041 | // X - (Z == 0) --> sub X, (zext(sete Z, 0)) --> sbb X, 0, (cmp Z, 1) |
35042 | // X + (Z == 0) --> add X, (zext(sete Z, 0)) --> adc X, 0, (cmp Z, 1) |
35043 | return DAG.getNode(IsSub ? X86ISD::SBB : X86ISD::ADC, DL, VTs, X, |
35044 | DAG.getConstant(0, DL, VT), Cmp1); |
35045 | } |
35046 | |
35047 | static SDValue combineLoopMAddPattern(SDNode *N, SelectionDAG &DAG, |
35048 | const X86Subtarget &Subtarget) { |
35049 | SDValue MulOp = N->getOperand(0); |
35050 | SDValue Phi = N->getOperand(1); |
35051 | |
35052 | if (MulOp.getOpcode() != ISD::MUL) |
35053 | std::swap(MulOp, Phi); |
35054 | if (MulOp.getOpcode() != ISD::MUL) |
35055 | return SDValue(); |
35056 | |
35057 | ShrinkMode Mode; |
35058 | if (!canReduceVMulWidth(MulOp.getNode(), DAG, Mode) || Mode == MULU16) |
35059 | return SDValue(); |
35060 | |
35061 | EVT VT = N->getValueType(0); |
35062 | |
35063 | unsigned RegSize = 128; |
35064 | if (Subtarget.hasBWI()) |
35065 | RegSize = 512; |
35066 | else if (Subtarget.hasAVX2()) |
35067 | RegSize = 256; |
35068 | unsigned VectorSize = VT.getVectorNumElements() * 16; |
35069 | // If the vector size is less than 128, or greater than the supported RegSize, |
35070 | // do not use PMADD. |
35071 | if (VectorSize < 128 || VectorSize > RegSize) |
35072 | return SDValue(); |
35073 | |
35074 | SDLoc DL(N); |
35075 | EVT ReducedVT = EVT::getVectorVT(*DAG.getContext(), MVT::i16, |
35076 | VT.getVectorNumElements()); |
35077 | EVT MAddVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, |
35078 | VT.getVectorNumElements() / 2); |
35079 | |
35080 | // Shrink the operands of mul. |
35081 | SDValue N0 = DAG.getNode(ISD::TRUNCATE, DL, ReducedVT, MulOp->getOperand(0)); |
35082 | SDValue N1 = DAG.getNode(ISD::TRUNCATE, DL, ReducedVT, MulOp->getOperand(1)); |
35083 | |
35084 | // Madd vector size is half of the original vector size |
35085 | SDValue Madd = DAG.getNode(X86ISD::VPMADDWD, DL, MAddVT, N0, N1); |
35086 | // Fill the rest of the output with 0 |
35087 | SDValue Zero = getZeroVector(Madd.getSimpleValueType(), Subtarget, DAG, DL); |
35088 | SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Madd, Zero); |
35089 | return DAG.getNode(ISD::ADD, DL, VT, Concat, Phi); |
35090 | } |
35091 | |
35092 | static SDValue combineLoopSADPattern(SDNode *N, SelectionDAG &DAG, |
35093 | const X86Subtarget &Subtarget) { |
35094 | SDLoc DL(N); |
35095 | EVT VT = N->getValueType(0); |
35096 | SDValue Op0 = N->getOperand(0); |
35097 | SDValue Op1 = N->getOperand(1); |
35098 | |
35099 | // TODO: There's nothing special about i32, any integer type above i16 should |
35100 | // work just as well. |
35101 | if (!VT.isVector() || !VT.isSimple() || |
35102 | !(VT.getVectorElementType() == MVT::i32)) |
35103 | return SDValue(); |
35104 | |
35105 | unsigned RegSize = 128; |
35106 | if (Subtarget.hasBWI()) |
35107 | RegSize = 512; |
35108 | else if (Subtarget.hasAVX2()) |
35109 | RegSize = 256; |
35110 | |
35111 | // We only handle v16i32 for SSE2 / v32i32 for AVX2 / v64i32 for AVX512. |
35112 | // TODO: We should be able to handle larger vectors by splitting them before |
35113 | // feeding them into several SADs, and then reducing over those. |
35114 | if (VT.getSizeInBits() / 4 > RegSize) |
35115 | return SDValue(); |
35116 | |
35117 | // We know N is a reduction add, which means one of its operands is a phi. |
35118 | // To match SAD, we need the other operand to be a vector select. |
35119 | SDValue SelectOp, Phi; |
35120 | if (Op0.getOpcode() == ISD::VSELECT) { |
35121 | SelectOp = Op0; |
35122 | Phi = Op1; |
35123 | } else if (Op1.getOpcode() == ISD::VSELECT) { |
35124 | SelectOp = Op1; |
35125 | Phi = Op0; |
35126 | } else |
35127 | return SDValue(); |
35128 | |
35129 | // Check whether we have an abs-diff pattern feeding into the select. |
35130 | if(!detectZextAbsDiff(SelectOp, Op0, Op1)) |
35131 | return SDValue(); |
35132 | |
35133 | // SAD pattern detected. Now build a SAD instruction and an addition for |
35134 | // reduction. Note that the number of elements of the result of SAD is less |
35135 | // than the number of elements of its input. Therefore, we could only update |
35136 | // part of elements in the reduction vector. |
35137 | SDValue Sad = createPSADBW(DAG, Op0, Op1, DL); |
35138 | |
35139 | // The output of PSADBW is a vector of i64. |
35140 | // We need to turn the vector of i64 into a vector of i32. |
35141 | // If the reduction vector is at least as wide as the psadbw result, just |
35142 | // bitcast. If it's narrower, truncate - the high i32 of each i64 is zero |
35143 | // anyway. |
35144 | MVT ResVT = MVT::getVectorVT(MVT::i32, Sad.getValueSizeInBits() / 32); |
35145 | if (VT.getSizeInBits() >= ResVT.getSizeInBits()) |
35146 | Sad = DAG.getNode(ISD::BITCAST, DL, ResVT, Sad); |
35147 | else |
35148 | Sad = DAG.getNode(ISD::TRUNCATE, DL, VT, Sad); |
35149 | |
35150 | if (VT.getSizeInBits() > ResVT.getSizeInBits()) { |
35151 | // Update part of elements of the reduction vector. This is done by first |
35152 | // extracting a sub-vector from it, updating this sub-vector, and inserting |
35153 | // it back. |
35154 | SDValue SubPhi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ResVT, Phi, |
35155 | DAG.getIntPtrConstant(0, DL)); |
35156 | SDValue Res = DAG.getNode(ISD::ADD, DL, ResVT, Sad, SubPhi); |
35157 | return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, Phi, Res, |
35158 | DAG.getIntPtrConstant(0, DL)); |
35159 | } else |
35160 | return DAG.getNode(ISD::ADD, DL, VT, Sad, Phi); |
35161 | } |
35162 | |
35163 | /// Convert vector increment or decrement to sub/add with an all-ones constant: |
35164 | /// add X, <1, 1...> --> sub X, <-1, -1...> |
35165 | /// sub X, <1, 1...> --> add X, <-1, -1...> |
35166 | /// The all-ones vector constant can be materialized using a pcmpeq instruction |
35167 | /// that is commonly recognized as an idiom (has no register dependency), so |
35168 | /// that's better/smaller than loading a splat 1 constant. |
35169 | static SDValue combineIncDecVector(SDNode *N, SelectionDAG &DAG) { |
35170 | assert((N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) &&(((N->getOpcode() == ISD::ADD || N->getOpcode() == ISD:: SUB) && "Unexpected opcode for increment/decrement transform" ) ? static_cast<void> (0) : __assert_fail ("(N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) && \"Unexpected opcode for increment/decrement transform\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 35171, __PRETTY_FUNCTION__)) |
35171 | "Unexpected opcode for increment/decrement transform")(((N->getOpcode() == ISD::ADD || N->getOpcode() == ISD:: SUB) && "Unexpected opcode for increment/decrement transform" ) ? static_cast<void> (0) : __assert_fail ("(N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) && \"Unexpected opcode for increment/decrement transform\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 35171, __PRETTY_FUNCTION__)); |
35172 | |
35173 | // Pseudo-legality check: getOnesVector() expects one of these types, so bail |
35174 | // out and wait for legalization if we have an unsupported vector length. |
35175 | EVT VT = N->getValueType(0); |
35176 | if (!VT.is128BitVector() && !VT.is256BitVector() && !VT.is512BitVector()) |
35177 | return SDValue(); |
35178 | |
35179 | SDNode *N1 = N->getOperand(1).getNode(); |
35180 | APInt SplatVal; |
35181 | if (!ISD::isConstantSplatVector(N1, SplatVal) || !SplatVal.isOneValue()) |
35182 | return SDValue(); |
35183 | |
35184 | SDValue AllOnesVec = getOnesVector(VT, DAG, SDLoc(N)); |
35185 | unsigned NewOpcode = N->getOpcode() == ISD::ADD ? ISD::SUB : ISD::ADD; |
35186 | return DAG.getNode(NewOpcode, SDLoc(N), VT, N->getOperand(0), AllOnesVec); |
35187 | } |
35188 | |
35189 | static SDValue combineAdd(SDNode *N, SelectionDAG &DAG, |
35190 | const X86Subtarget &Subtarget) { |
35191 | const SDNodeFlags Flags = N->getFlags(); |
35192 | if (Flags.hasVectorReduction()) { |
35193 | if (SDValue Sad = combineLoopSADPattern(N, DAG, Subtarget)) |
35194 | return Sad; |
35195 | if (SDValue MAdd = combineLoopMAddPattern(N, DAG, Subtarget)) |
35196 | return MAdd; |
35197 | } |
35198 | EVT VT = N->getValueType(0); |
35199 | SDValue Op0 = N->getOperand(0); |
35200 | SDValue Op1 = N->getOperand(1); |
35201 | |
35202 | // Try to synthesize horizontal adds from adds of shuffles. |
35203 | if (((Subtarget.hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) || |
35204 | (Subtarget.hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) && |
35205 | isHorizontalBinOp(Op0, Op1, true)) |
35206 | return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1); |
35207 | |
35208 | if (SDValue V = combineIncDecVector(N, DAG)) |
35209 | return V; |
35210 | |
35211 | return combineAddOrSubToADCOrSBB(N, DAG); |
35212 | } |
35213 | |
35214 | static SDValue combineSub(SDNode *N, SelectionDAG &DAG, |
35215 | const X86Subtarget &Subtarget) { |
35216 | SDValue Op0 = N->getOperand(0); |
35217 | SDValue Op1 = N->getOperand(1); |
35218 | |
35219 | // X86 can't encode an immediate LHS of a sub. See if we can push the |
35220 | // negation into a preceding instruction. |
35221 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) { |
35222 | // If the RHS of the sub is a XOR with one use and a constant, invert the |
35223 | // immediate. Then add one to the LHS of the sub so we can turn |
35224 | // X-Y -> X+~Y+1, saving one register. |
35225 | if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR && |
35226 | isa<ConstantSDNode>(Op1.getOperand(1))) { |
35227 | APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue(); |
35228 | EVT VT = Op0.getValueType(); |
35229 | SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT, |
35230 | Op1.getOperand(0), |
35231 | DAG.getConstant(~XorC, SDLoc(Op1), VT)); |
35232 | return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor, |
35233 | DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT)); |
35234 | } |
35235 | } |
35236 | |
35237 | // Try to synthesize horizontal subs from subs of shuffles. |
35238 | EVT VT = N->getValueType(0); |
35239 | if (((Subtarget.hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) || |
35240 | (Subtarget.hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) && |
35241 | isHorizontalBinOp(Op0, Op1, false)) |
35242 | return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1); |
35243 | |
35244 | if (SDValue V = combineIncDecVector(N, DAG)) |
35245 | return V; |
35246 | |
35247 | return combineAddOrSubToADCOrSBB(N, DAG); |
35248 | } |
35249 | |
35250 | static SDValue combineVSZext(SDNode *N, SelectionDAG &DAG, |
35251 | TargetLowering::DAGCombinerInfo &DCI, |
35252 | const X86Subtarget &Subtarget) { |
35253 | if (DCI.isBeforeLegalize()) |
35254 | return SDValue(); |
35255 | |
35256 | SDLoc DL(N); |
35257 | unsigned Opcode = N->getOpcode(); |
35258 | MVT VT = N->getSimpleValueType(0); |
35259 | MVT SVT = VT.getVectorElementType(); |
35260 | unsigned NumElts = VT.getVectorNumElements(); |
35261 | unsigned EltSizeInBits = SVT.getSizeInBits(); |
35262 | |
35263 | SDValue Op = N->getOperand(0); |
35264 | MVT OpVT = Op.getSimpleValueType(); |
35265 | MVT OpEltVT = OpVT.getVectorElementType(); |
35266 | unsigned OpEltSizeInBits = OpEltVT.getSizeInBits(); |
35267 | unsigned InputBits = OpEltSizeInBits * NumElts; |
35268 | |
35269 | // Perform any constant folding. |
35270 | // FIXME: Reduce constant pool usage and don't fold when OptSize is enabled. |
35271 | APInt UndefElts; |
35272 | SmallVector<APInt, 64> EltBits; |
35273 | if (getTargetConstantBitsFromNode(Op, OpEltSizeInBits, UndefElts, EltBits)) { |
35274 | APInt Undefs(NumElts, 0); |
35275 | SmallVector<APInt, 4> Vals(NumElts, APInt(EltSizeInBits, 0)); |
35276 | bool IsZEXT = |
35277 | (Opcode == X86ISD::VZEXT) || (Opcode == ISD::ZERO_EXTEND_VECTOR_INREG); |
35278 | for (unsigned i = 0; i != NumElts; ++i) { |
35279 | if (UndefElts[i]) { |
35280 | Undefs.setBit(i); |
35281 | continue; |
35282 | } |
35283 | Vals[i] = IsZEXT ? EltBits[i].zextOrTrunc(EltSizeInBits) |
35284 | : EltBits[i].sextOrTrunc(EltSizeInBits); |
35285 | } |
35286 | return getConstVector(Vals, Undefs, VT, DAG, DL); |
35287 | } |
35288 | |
35289 | // (vzext (bitcast (vzext (x)) -> (vzext x) |
35290 | // TODO: (vsext (bitcast (vsext (x)) -> (vsext x) |
35291 | SDValue V = peekThroughBitcasts(Op); |
35292 | if (Opcode == X86ISD::VZEXT && V != Op && V.getOpcode() == X86ISD::VZEXT) { |
35293 | MVT InnerVT = V.getSimpleValueType(); |
35294 | MVT InnerEltVT = InnerVT.getVectorElementType(); |
35295 | |
35296 | // If the element sizes match exactly, we can just do one larger vzext. This |
35297 | // is always an exact type match as vzext operates on integer types. |
35298 | if (OpEltVT == InnerEltVT) { |
35299 | assert(OpVT == InnerVT && "Types must match for vzext!")((OpVT == InnerVT && "Types must match for vzext!") ? static_cast<void> (0) : __assert_fail ("OpVT == InnerVT && \"Types must match for vzext!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 35299, __PRETTY_FUNCTION__)); |
35300 | return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0)); |
35301 | } |
35302 | |
35303 | // The only other way we can combine them is if only a single element of the |
35304 | // inner vzext is used in the input to the outer vzext. |
35305 | if (InnerEltVT.getSizeInBits() < InputBits) |
35306 | return SDValue(); |
35307 | |
35308 | // In this case, the inner vzext is completely dead because we're going to |
35309 | // only look at bits inside of the low element. Just do the outer vzext on |
35310 | // a bitcast of the input to the inner. |
35311 | return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V)); |
35312 | } |
35313 | |
35314 | // Check if we can bypass extracting and re-inserting an element of an input |
35315 | // vector. Essentially: |
35316 | // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x) |
35317 | // TODO: Add X86ISD::VSEXT support |
35318 | if (Opcode == X86ISD::VZEXT && |
35319 | V.getOpcode() == ISD::SCALAR_TO_VECTOR && |
35320 | V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT && |
35321 | V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) { |
35322 | SDValue ExtractedV = V.getOperand(0); |
35323 | SDValue OrigV = ExtractedV.getOperand(0); |
35324 | if (isNullConstant(ExtractedV.getOperand(1))) { |
35325 | MVT OrigVT = OrigV.getSimpleValueType(); |
35326 | // Extract a subvector if necessary... |
35327 | if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) { |
35328 | int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits(); |
35329 | OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(), |
35330 | OrigVT.getVectorNumElements() / Ratio); |
35331 | OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV, |
35332 | DAG.getIntPtrConstant(0, DL)); |
35333 | } |
35334 | Op = DAG.getBitcast(OpVT, OrigV); |
35335 | return DAG.getNode(X86ISD::VZEXT, DL, VT, Op); |
35336 | } |
35337 | } |
35338 | |
35339 | return SDValue(); |
35340 | } |
35341 | |
35342 | /// Canonicalize (LSUB p, 1) -> (LADD p, -1). |
35343 | static SDValue combineLockSub(SDNode *N, SelectionDAG &DAG, |
35344 | const X86Subtarget &Subtarget) { |
35345 | SDValue Chain = N->getOperand(0); |
35346 | SDValue LHS = N->getOperand(1); |
35347 | SDValue RHS = N->getOperand(2); |
35348 | MVT VT = RHS.getSimpleValueType(); |
35349 | SDLoc DL(N); |
35350 | |
35351 | auto *C = dyn_cast<ConstantSDNode>(RHS); |
35352 | if (!C || C->getZExtValue() != 1) |
35353 | return SDValue(); |
35354 | |
35355 | RHS = DAG.getConstant(-1, DL, VT); |
35356 | MachineMemOperand *MMO = cast<MemSDNode>(N)->getMemOperand(); |
35357 | return DAG.getMemIntrinsicNode(X86ISD::LADD, DL, |
35358 | DAG.getVTList(MVT::i32, MVT::Other), |
35359 | {Chain, LHS, RHS}, VT, MMO); |
35360 | } |
35361 | |
35362 | // TEST (AND a, b) ,(AND a, b) -> TEST a, b |
35363 | static SDValue combineTestM(SDNode *N, SelectionDAG &DAG) { |
35364 | SDValue Op0 = N->getOperand(0); |
35365 | SDValue Op1 = N->getOperand(1); |
35366 | |
35367 | if (Op0 != Op1 || Op1->getOpcode() != ISD::AND) |
35368 | return SDValue(); |
35369 | |
35370 | EVT VT = N->getValueType(0); |
35371 | SDLoc DL(N); |
35372 | |
35373 | return DAG.getNode(X86ISD::TESTM, DL, VT, |
35374 | Op0->getOperand(0), Op0->getOperand(1)); |
35375 | } |
35376 | |
35377 | static SDValue combineVectorCompare(SDNode *N, SelectionDAG &DAG, |
35378 | const X86Subtarget &Subtarget) { |
35379 | MVT VT = N->getSimpleValueType(0); |
35380 | SDLoc DL(N); |
35381 | |
35382 | if (N->getOperand(0) == N->getOperand(1)) { |
35383 | if (N->getOpcode() == X86ISD::PCMPEQ) |
35384 | return getOnesVector(VT, DAG, DL); |
35385 | if (N->getOpcode() == X86ISD::PCMPGT) |
35386 | return getZeroVector(VT, Subtarget, DAG, DL); |
35387 | } |
35388 | |
35389 | return SDValue(); |
35390 | } |
35391 | |
35392 | static SDValue combineInsertSubvector(SDNode *N, SelectionDAG &DAG, |
35393 | TargetLowering::DAGCombinerInfo &DCI, |
35394 | const X86Subtarget &Subtarget) { |
35395 | if (DCI.isBeforeLegalizeOps()) |
35396 | return SDValue(); |
35397 | |
35398 | SDLoc dl(N); |
35399 | SDValue Vec = N->getOperand(0); |
35400 | SDValue SubVec = N->getOperand(1); |
35401 | SDValue Idx = N->getOperand(2); |
35402 | |
35403 | unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); |
35404 | MVT OpVT = N->getSimpleValueType(0); |
35405 | MVT SubVecVT = SubVec.getSimpleValueType(); |
35406 | |
35407 | // If this is an insert of an extract, combine to a shuffle. Don't do this |
35408 | // if the insert or extract can be represented with a subvector operation. |
35409 | if (SubVec.getOpcode() == ISD::EXTRACT_SUBVECTOR && |
35410 | SubVec.getOperand(0).getSimpleValueType() == OpVT && |
35411 | (IdxVal != 0 || !Vec.isUndef())) { |
35412 | int ExtIdxVal = cast<ConstantSDNode>(SubVec.getOperand(1))->getZExtValue(); |
35413 | if (ExtIdxVal != 0) { |
35414 | int VecNumElts = OpVT.getVectorNumElements(); |
35415 | int SubVecNumElts = SubVecVT.getVectorNumElements(); |
35416 | SmallVector<int, 64> Mask(VecNumElts); |
35417 | // First create an identity shuffle mask. |
35418 | for (int i = 0; i != VecNumElts; ++i) |
35419 | Mask[i] = i; |
35420 | // Now insert the extracted portion. |
35421 | for (int i = 0; i != SubVecNumElts; ++i) |
35422 | Mask[i + IdxVal] = i + ExtIdxVal + VecNumElts; |
35423 | |
35424 | return DAG.getVectorShuffle(OpVT, dl, Vec, SubVec.getOperand(0), Mask); |
35425 | } |
35426 | } |
35427 | |
35428 | // Fold two 16-byte or 32-byte subvector loads into one 32-byte or 64-byte |
35429 | // load: |
35430 | // (insert_subvector (insert_subvector undef, (load16 addr), 0), |
35431 | // (load16 addr + 16), Elts/2) |
35432 | // --> load32 addr |
35433 | // or: |
35434 | // (insert_subvector (insert_subvector undef, (load32 addr), 0), |
35435 | // (load32 addr + 32), Elts/2) |
35436 | // --> load64 addr |
35437 | // or a 16-byte or 32-byte broadcast: |
35438 | // (insert_subvector (insert_subvector undef, (load16 addr), 0), |
35439 | // (load16 addr), Elts/2) |
35440 | // --> X86SubVBroadcast(load16 addr) |
35441 | // or: |
35442 | // (insert_subvector (insert_subvector undef, (load32 addr), 0), |
35443 | // (load32 addr), Elts/2) |
35444 | // --> X86SubVBroadcast(load32 addr) |
35445 | if ((IdxVal == OpVT.getVectorNumElements() / 2) && |
35446 | Vec.getOpcode() == ISD::INSERT_SUBVECTOR && |
35447 | OpVT.getSizeInBits() == SubVecVT.getSizeInBits() * 2) { |
35448 | auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2)); |
35449 | if (Idx2 && Idx2->getZExtValue() == 0) { |
35450 | SDValue SubVec2 = Vec.getOperand(1); |
35451 | // If needed, look through bitcasts to get to the load. |
35452 | if (auto *FirstLd = dyn_cast<LoadSDNode>(peekThroughBitcasts(SubVec2))) { |
35453 | bool Fast; |
35454 | unsigned Alignment = FirstLd->getAlignment(); |
35455 | unsigned AS = FirstLd->getAddressSpace(); |
35456 | const X86TargetLowering *TLI = Subtarget.getTargetLowering(); |
35457 | if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), |
35458 | OpVT, AS, Alignment, &Fast) && Fast) { |
35459 | SDValue Ops[] = {SubVec2, SubVec}; |
35460 | if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, |
35461 | Subtarget, false)) |
35462 | return Ld; |
35463 | } |
35464 | } |
35465 | // If lower/upper loads are the same and the only users of the load, then |
35466 | // lower to a VBROADCASTF128/VBROADCASTI128/etc. |
35467 | if (auto *Ld = dyn_cast<LoadSDNode>(peekThroughOneUseBitcasts(SubVec2))) { |
35468 | if (SubVec2 == SubVec && ISD::isNormalLoad(Ld) && |
35469 | SDNode::areOnlyUsersOf({N, Vec.getNode()}, SubVec2.getNode())) { |
35470 | return DAG.getNode(X86ISD::SUBV_BROADCAST, dl, OpVT, SubVec); |
35471 | } |
35472 | } |
35473 | // If this is subv_broadcast insert into both halves, use a larger |
35474 | // subv_broadcast. |
35475 | if (SubVec.getOpcode() == X86ISD::SUBV_BROADCAST && SubVec == SubVec2) { |
35476 | return DAG.getNode(X86ISD::SUBV_BROADCAST, dl, OpVT, |
35477 | SubVec.getOperand(0)); |
35478 | } |
35479 | } |
35480 | } |
35481 | |
35482 | return SDValue(); |
35483 | } |
35484 | |
35485 | |
35486 | SDValue X86TargetLowering::PerformDAGCombine(SDNode *N, |
35487 | DAGCombinerInfo &DCI) const { |
35488 | SelectionDAG &DAG = DCI.DAG; |
35489 | switch (N->getOpcode()) { |
35490 | default: break; |
35491 | case ISD::EXTRACT_VECTOR_ELT: |
35492 | return combineExtractVectorElt(N, DAG, DCI, Subtarget); |
35493 | case X86ISD::PEXTRW: |
35494 | case X86ISD::PEXTRB: |
35495 | return combineExtractVectorElt_SSE(N, DAG, DCI, Subtarget); |
35496 | case ISD::INSERT_SUBVECTOR: |
35497 | return combineInsertSubvector(N, DAG, DCI, Subtarget); |
35498 | case ISD::VSELECT: |
35499 | case ISD::SELECT: |
35500 | case X86ISD::SHRUNKBLEND: return combineSelect(N, DAG, DCI, Subtarget); |
35501 | case ISD::BITCAST: return combineBitcast(N, DAG, DCI, Subtarget); |
35502 | case X86ISD::CMOV: return combineCMov(N, DAG, DCI, Subtarget); |
35503 | case ISD::ADD: return combineAdd(N, DAG, Subtarget); |
35504 | case ISD::SUB: return combineSub(N, DAG, Subtarget); |
35505 | case X86ISD::ADD: return combineX86ADD(N, DAG, DCI); |
35506 | case X86ISD::ADC: return combineADC(N, DAG, DCI); |
35507 | case ISD::MUL: return combineMul(N, DAG, DCI, Subtarget); |
35508 | case ISD::SHL: |
35509 | case ISD::SRA: |
35510 | case ISD::SRL: return combineShift(N, DAG, DCI, Subtarget); |
35511 | case ISD::AND: return combineAnd(N, DAG, DCI, Subtarget); |
35512 | case ISD::OR: return combineOr(N, DAG, DCI, Subtarget); |
35513 | case ISD::XOR: return combineXor(N, DAG, DCI, Subtarget); |
35514 | case ISD::LOAD: return combineLoad(N, DAG, DCI, Subtarget); |
35515 | case ISD::MLOAD: return combineMaskedLoad(N, DAG, DCI, Subtarget); |
35516 | case ISD::STORE: return combineStore(N, DAG, Subtarget); |
35517 | case ISD::MSTORE: return combineMaskedStore(N, DAG, Subtarget); |
35518 | case ISD::SINT_TO_FP: return combineSIntToFP(N, DAG, Subtarget); |
35519 | case ISD::UINT_TO_FP: return combineUIntToFP(N, DAG, Subtarget); |
35520 | case ISD::FADD: |
35521 | case ISD::FSUB: return combineFaddFsub(N, DAG, Subtarget); |
35522 | case ISD::FNEG: return combineFneg(N, DAG, Subtarget); |
35523 | case ISD::TRUNCATE: return combineTruncate(N, DAG, Subtarget); |
35524 | case X86ISD::ANDNP: return combineAndnp(N, DAG, DCI, Subtarget); |
35525 | case X86ISD::FAND: return combineFAnd(N, DAG, Subtarget); |
35526 | case X86ISD::FANDN: return combineFAndn(N, DAG, Subtarget); |
35527 | case X86ISD::FXOR: |
35528 | case X86ISD::FOR: return combineFOr(N, DAG, Subtarget); |
35529 | case X86ISD::FMIN: |
35530 | case X86ISD::FMAX: return combineFMinFMax(N, DAG); |
35531 | case ISD::FMINNUM: |
35532 | case ISD::FMAXNUM: return combineFMinNumFMaxNum(N, DAG, Subtarget); |
35533 | case X86ISD::BT: return combineBT(N, DAG, DCI); |
35534 | case ISD::ANY_EXTEND: |
35535 | case ISD::ZERO_EXTEND: return combineZext(N, DAG, DCI, Subtarget); |
35536 | case ISD::SIGN_EXTEND: return combineSext(N, DAG, DCI, Subtarget); |
35537 | case ISD::SIGN_EXTEND_INREG: return combineSignExtendInReg(N, DAG, Subtarget); |
35538 | case ISD::SETCC: return combineSetCC(N, DAG, Subtarget); |
35539 | case X86ISD::SETCC: return combineX86SetCC(N, DAG, Subtarget); |
35540 | case X86ISD::BRCOND: return combineBrCond(N, DAG, Subtarget); |
35541 | case X86ISD::VSHLI: |
35542 | case X86ISD::VSRAI: |
35543 | case X86ISD::VSRLI: |
35544 | return combineVectorShiftImm(N, DAG, DCI, Subtarget); |
35545 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
35546 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
35547 | case X86ISD::VSEXT: |
35548 | case X86ISD::VZEXT: return combineVSZext(N, DAG, DCI, Subtarget); |
35549 | case X86ISD::PINSRB: |
35550 | case X86ISD::PINSRW: return combineVectorInsert(N, DAG, DCI, Subtarget); |
35551 | case X86ISD::SHUFP: // Handle all target specific shuffles |
35552 | case X86ISD::INSERTPS: |
35553 | case X86ISD::PALIGNR: |
35554 | case X86ISD::VSHLDQ: |
35555 | case X86ISD::VSRLDQ: |
35556 | case X86ISD::BLENDI: |
35557 | case X86ISD::UNPCKH: |
35558 | case X86ISD::UNPCKL: |
35559 | case X86ISD::MOVHLPS: |
35560 | case X86ISD::MOVLHPS: |
35561 | case X86ISD::PSHUFB: |
35562 | case X86ISD::PSHUFD: |
35563 | case X86ISD::PSHUFHW: |
35564 | case X86ISD::PSHUFLW: |
35565 | case X86ISD::MOVSHDUP: |
35566 | case X86ISD::MOVSLDUP: |
35567 | case X86ISD::MOVDDUP: |
35568 | case X86ISD::MOVSS: |
35569 | case X86ISD::MOVSD: |
35570 | case X86ISD::VPPERM: |
35571 | case X86ISD::VPERMI: |
35572 | case X86ISD::VPERMV: |
35573 | case X86ISD::VPERMV3: |
35574 | case X86ISD::VPERMIV3: |
35575 | case X86ISD::VPERMIL2: |
35576 | case X86ISD::VPERMILPI: |
35577 | case X86ISD::VPERMILPV: |
35578 | case X86ISD::VPERM2X128: |
35579 | case X86ISD::VZEXT_MOVL: |
35580 | case ISD::VECTOR_SHUFFLE: return combineShuffle(N, DAG, DCI,Subtarget); |
35581 | case X86ISD::FMADD: |
35582 | case X86ISD::FMADD_RND: |
35583 | case X86ISD::FMADDS1_RND: |
35584 | case X86ISD::FMADDS3_RND: |
35585 | case ISD::FMA: return combineFMA(N, DAG, Subtarget); |
35586 | case ISD::MGATHER: |
35587 | case ISD::MSCATTER: return combineGatherScatter(N, DAG); |
35588 | case X86ISD::LSUB: return combineLockSub(N, DAG, Subtarget); |
35589 | case X86ISD::TESTM: return combineTestM(N, DAG); |
35590 | case X86ISD::PCMPEQ: |
35591 | case X86ISD::PCMPGT: return combineVectorCompare(N, DAG, Subtarget); |
35592 | } |
35593 | |
35594 | return SDValue(); |
35595 | } |
35596 | |
35597 | /// Return true if the target has native support for the specified value type |
35598 | /// and it is 'desirable' to use the type for the given node type. e.g. On x86 |
35599 | /// i16 is legal, but undesirable since i16 instruction encodings are longer and |
35600 | /// some i16 instructions are slow. |
35601 | bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const { |
35602 | if (!isTypeLegal(VT)) |
35603 | return false; |
35604 | if (VT != MVT::i16) |
35605 | return true; |
35606 | |
35607 | switch (Opc) { |
35608 | default: |
35609 | return true; |
35610 | case ISD::LOAD: |
35611 | case ISD::SIGN_EXTEND: |
35612 | case ISD::ZERO_EXTEND: |
35613 | case ISD::ANY_EXTEND: |
35614 | case ISD::SHL: |
35615 | case ISD::SRL: |
35616 | case ISD::SUB: |
35617 | case ISD::ADD: |
35618 | case ISD::MUL: |
35619 | case ISD::AND: |
35620 | case ISD::OR: |
35621 | case ISD::XOR: |
35622 | return false; |
35623 | } |
35624 | } |
35625 | |
35626 | /// This function checks if any of the users of EFLAGS copies the EFLAGS. We |
35627 | /// know that the code that lowers COPY of EFLAGS has to use the stack, and if |
35628 | /// we don't adjust the stack we clobber the first frame index. |
35629 | /// See X86InstrInfo::copyPhysReg. |
35630 | static bool hasCopyImplyingStackAdjustment(const MachineFunction &MF) { |
35631 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
35632 | return any_of(MRI.reg_instructions(X86::EFLAGS), |
35633 | [](const MachineInstr &RI) { return RI.isCopy(); }); |
35634 | } |
35635 | |
35636 | void X86TargetLowering::finalizeLowering(MachineFunction &MF) const { |
35637 | if (hasCopyImplyingStackAdjustment(MF)) { |
35638 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
35639 | MFI.setHasCopyImplyingStackAdjustment(true); |
35640 | } |
35641 | |
35642 | TargetLoweringBase::finalizeLowering(MF); |
35643 | } |
35644 | |
35645 | /// This method query the target whether it is beneficial for dag combiner to |
35646 | /// promote the specified node. If true, it should return the desired promotion |
35647 | /// type by reference. |
35648 | bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const { |
35649 | EVT VT = Op.getValueType(); |
35650 | if (VT != MVT::i16) |
35651 | return false; |
35652 | |
35653 | bool Promote = false; |
35654 | bool Commute = false; |
35655 | switch (Op.getOpcode()) { |
35656 | default: break; |
35657 | case ISD::SIGN_EXTEND: |
35658 | case ISD::ZERO_EXTEND: |
35659 | case ISD::ANY_EXTEND: |
35660 | Promote = true; |
35661 | break; |
35662 | case ISD::SHL: |
35663 | case ISD::SRL: { |
35664 | SDValue N0 = Op.getOperand(0); |
35665 | // Look out for (store (shl (load), x)). |
35666 | if (MayFoldLoad(N0) && MayFoldIntoStore(Op)) |
35667 | return false; |
35668 | Promote = true; |
35669 | break; |
35670 | } |
35671 | case ISD::ADD: |
35672 | case ISD::MUL: |
35673 | case ISD::AND: |
35674 | case ISD::OR: |
35675 | case ISD::XOR: |
35676 | Commute = true; |
35677 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
35678 | case ISD::SUB: { |
35679 | SDValue N0 = Op.getOperand(0); |
35680 | SDValue N1 = Op.getOperand(1); |
35681 | if (!Commute && MayFoldLoad(N1)) |
35682 | return false; |
35683 | // Avoid disabling potential load folding opportunities. |
35684 | if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op))) |
35685 | return false; |
35686 | if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op))) |
35687 | return false; |
35688 | Promote = true; |
35689 | } |
35690 | } |
35691 | |
35692 | PVT = MVT::i32; |
35693 | return Promote; |
35694 | } |
35695 | |
35696 | //===----------------------------------------------------------------------===// |
35697 | // X86 Inline Assembly Support |
35698 | //===----------------------------------------------------------------------===// |
35699 | |
35700 | // Helper to match a string separated by whitespace. |
35701 | static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) { |
35702 | S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace. |
35703 | |
35704 | for (StringRef Piece : Pieces) { |
35705 | if (!S.startswith(Piece)) // Check if the piece matches. |
35706 | return false; |
35707 | |
35708 | S = S.substr(Piece.size()); |
35709 | StringRef::size_type Pos = S.find_first_not_of(" \t"); |
35710 | if (Pos == 0) // We matched a prefix. |
35711 | return false; |
35712 | |
35713 | S = S.substr(Pos); |
35714 | } |
35715 | |
35716 | return S.empty(); |
35717 | } |
35718 | |
35719 | static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) { |
35720 | |
35721 | if (AsmPieces.size() == 3 || AsmPieces.size() == 4) { |
35722 | if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") && |
35723 | std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") && |
35724 | std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) { |
35725 | |
35726 | if (AsmPieces.size() == 3) |
35727 | return true; |
35728 | else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}")) |
35729 | return true; |
35730 | } |
35731 | } |
35732 | return false; |
35733 | } |
35734 | |
35735 | bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const { |
35736 | InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue()); |
35737 | |
35738 | const std::string &AsmStr = IA->getAsmString(); |
35739 | |
35740 | IntegerType *Ty = dyn_cast<IntegerType>(CI->getType()); |
35741 | if (!Ty || Ty->getBitWidth() % 16 != 0) |
35742 | return false; |
35743 | |
35744 | // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a" |
35745 | SmallVector<StringRef, 4> AsmPieces; |
35746 | SplitString(AsmStr, AsmPieces, ";\n"); |
35747 | |
35748 | switch (AsmPieces.size()) { |
35749 | default: return false; |
35750 | case 1: |
35751 | // FIXME: this should verify that we are targeting a 486 or better. If not, |
35752 | // we will turn this bswap into something that will be lowered to logical |
35753 | // ops instead of emitting the bswap asm. For now, we don't support 486 or |
35754 | // lower so don't worry about this. |
35755 | // bswap $0 |
35756 | if (matchAsm(AsmPieces[0], {"bswap", "$0"}) || |
35757 | matchAsm(AsmPieces[0], {"bswapl", "$0"}) || |
35758 | matchAsm(AsmPieces[0], {"bswapq", "$0"}) || |
35759 | matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) || |
35760 | matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) || |
35761 | matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) { |
35762 | // No need to check constraints, nothing other than the equivalent of |
35763 | // "=r,0" would be valid here. |
35764 | return IntrinsicLowering::LowerToByteSwap(CI); |
35765 | } |
35766 | |
35767 | // rorw $$8, ${0:w} --> llvm.bswap.i16 |
35768 | if (CI->getType()->isIntegerTy(16) && |
35769 | IA->getConstraintString().compare(0, 5, "=r,0,") == 0 && |
35770 | (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) || |
35771 | matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) { |
35772 | AsmPieces.clear(); |
35773 | StringRef ConstraintsStr = IA->getConstraintString(); |
35774 | SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ","); |
35775 | array_pod_sort(AsmPieces.begin(), AsmPieces.end()); |
35776 | if (clobbersFlagRegisters(AsmPieces)) |
35777 | return IntrinsicLowering::LowerToByteSwap(CI); |
35778 | } |
35779 | break; |
35780 | case 3: |
35781 | if (CI->getType()->isIntegerTy(32) && |
35782 | IA->getConstraintString().compare(0, 5, "=r,0,") == 0 && |
35783 | matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) && |
35784 | matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) && |
35785 | matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) { |
35786 | AsmPieces.clear(); |
35787 | StringRef ConstraintsStr = IA->getConstraintString(); |
35788 | SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ","); |
35789 | array_pod_sort(AsmPieces.begin(), AsmPieces.end()); |
35790 | if (clobbersFlagRegisters(AsmPieces)) |
35791 | return IntrinsicLowering::LowerToByteSwap(CI); |
35792 | } |
35793 | |
35794 | if (CI->getType()->isIntegerTy(64)) { |
35795 | InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints(); |
35796 | if (Constraints.size() >= 2 && |
35797 | Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" && |
35798 | Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") { |
35799 | // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64 |
35800 | if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) && |
35801 | matchAsm(AsmPieces[1], {"bswap", "%edx"}) && |
35802 | matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"})) |
35803 | return IntrinsicLowering::LowerToByteSwap(CI); |
35804 | } |
35805 | } |
35806 | break; |
35807 | } |
35808 | return false; |
35809 | } |
35810 | |
35811 | /// Given a constraint letter, return the type of constraint for this target. |
35812 | X86TargetLowering::ConstraintType |
35813 | X86TargetLowering::getConstraintType(StringRef Constraint) const { |
35814 | if (Constraint.size() == 1) { |
35815 | switch (Constraint[0]) { |
35816 | case 'R': |
35817 | case 'q': |
35818 | case 'Q': |
35819 | case 'f': |
35820 | case 't': |
35821 | case 'u': |
35822 | case 'y': |
35823 | case 'x': |
35824 | case 'v': |
35825 | case 'Y': |
35826 | case 'l': |
35827 | return C_RegisterClass; |
35828 | case 'k': // AVX512 masking registers. |
35829 | case 'a': |
35830 | case 'b': |
35831 | case 'c': |
35832 | case 'd': |
35833 | case 'S': |
35834 | case 'D': |
35835 | case 'A': |
35836 | return C_Register; |
35837 | case 'I': |
35838 | case 'J': |
35839 | case 'K': |
35840 | case 'L': |
35841 | case 'M': |
35842 | case 'N': |
35843 | case 'G': |
35844 | case 'C': |
35845 | case 'e': |
35846 | case 'Z': |
35847 | return C_Other; |
35848 | default: |
35849 | break; |
35850 | } |
35851 | } |
35852 | else if (Constraint.size() == 2) { |
35853 | switch (Constraint[0]) { |
35854 | default: |
35855 | break; |
35856 | case 'Y': |
35857 | switch (Constraint[1]) { |
35858 | default: |
35859 | break; |
35860 | case 'k': |
35861 | return C_Register; |
35862 | } |
35863 | } |
35864 | } |
35865 | return TargetLowering::getConstraintType(Constraint); |
35866 | } |
35867 | |
35868 | /// Examine constraint type and operand type and determine a weight value. |
35869 | /// This object must already have been set up with the operand type |
35870 | /// and the current alternative constraint selected. |
35871 | TargetLowering::ConstraintWeight |
35872 | X86TargetLowering::getSingleConstraintMatchWeight( |
35873 | AsmOperandInfo &info, const char *constraint) const { |
35874 | ConstraintWeight weight = CW_Invalid; |
35875 | Value *CallOperandVal = info.CallOperandVal; |
35876 | // If we don't have a value, we can't do a match, |
35877 | // but allow it at the lowest weight. |
35878 | if (!CallOperandVal) |
35879 | return CW_Default; |
35880 | Type *type = CallOperandVal->getType(); |
35881 | // Look at the constraint type. |
35882 | switch (*constraint) { |
35883 | default: |
35884 | weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); |
35885 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
35886 | case 'R': |
35887 | case 'q': |
35888 | case 'Q': |
35889 | case 'a': |
35890 | case 'b': |
35891 | case 'c': |
35892 | case 'd': |
35893 | case 'S': |
35894 | case 'D': |
35895 | case 'A': |
35896 | if (CallOperandVal->getType()->isIntegerTy()) |
35897 | weight = CW_SpecificReg; |
35898 | break; |
35899 | case 'f': |
35900 | case 't': |
35901 | case 'u': |
35902 | if (type->isFloatingPointTy()) |
35903 | weight = CW_SpecificReg; |
35904 | break; |
35905 | case 'y': |
35906 | if (type->isX86_MMXTy() && Subtarget.hasMMX()) |
35907 | weight = CW_SpecificReg; |
35908 | break; |
35909 | case 'Y': |
35910 | // Other "Y<x>" (e.g. "Yk") constraints should be implemented below. |
35911 | if (constraint[1] == 'k') { |
35912 | // Support for 'Yk' (similarly to the 'k' variant below). |
35913 | weight = CW_SpecificReg; |
35914 | break; |
35915 | } |
35916 | // Else fall through (handle "Y" constraint). |
35917 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
35918 | case 'v': |
35919 | if ((type->getPrimitiveSizeInBits() == 512) && Subtarget.hasAVX512()) |
35920 | weight = CW_Register; |
35921 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
35922 | case 'x': |
35923 | if (((type->getPrimitiveSizeInBits() == 128) && Subtarget.hasSSE1()) || |
35924 | ((type->getPrimitiveSizeInBits() == 256) && Subtarget.hasFp256())) |
35925 | weight = CW_Register; |
35926 | break; |
35927 | case 'k': |
35928 | // Enable conditional vector operations using %k<#> registers. |
35929 | weight = CW_SpecificReg; |
35930 | break; |
35931 | case 'I': |
35932 | if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) { |
35933 | if (C->getZExtValue() <= 31) |
35934 | weight = CW_Constant; |
35935 | } |
35936 | break; |
35937 | case 'J': |
35938 | if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) { |
35939 | if (C->getZExtValue() <= 63) |
35940 | weight = CW_Constant; |
35941 | } |
35942 | break; |
35943 | case 'K': |
35944 | if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) { |
35945 | if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f)) |
35946 | weight = CW_Constant; |
35947 | } |
35948 | break; |
35949 | case 'L': |
35950 | if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) { |
35951 | if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff)) |
35952 | weight = CW_Constant; |
35953 | } |
35954 | break; |
35955 | case 'M': |
35956 | if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) { |
35957 | if (C->getZExtValue() <= 3) |
35958 | weight = CW_Constant; |
35959 | } |
35960 | break; |
35961 | case 'N': |
35962 | if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) { |
35963 | if (C->getZExtValue() <= 0xff) |
35964 | weight = CW_Constant; |
35965 | } |
35966 | break; |
35967 | case 'G': |
35968 | case 'C': |
35969 | if (isa<ConstantFP>(CallOperandVal)) { |
35970 | weight = CW_Constant; |
35971 | } |
35972 | break; |
35973 | case 'e': |
35974 | if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) { |
35975 | if ((C->getSExtValue() >= -0x80000000LL) && |
35976 | (C->getSExtValue() <= 0x7fffffffLL)) |
35977 | weight = CW_Constant; |
35978 | } |
35979 | break; |
35980 | case 'Z': |
35981 | if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) { |
35982 | if (C->getZExtValue() <= 0xffffffff) |
35983 | weight = CW_Constant; |
35984 | } |
35985 | break; |
35986 | } |
35987 | return weight; |
35988 | } |
35989 | |
35990 | /// Try to replace an X constraint, which matches anything, with another that |
35991 | /// has more specific requirements based on the type of the corresponding |
35992 | /// operand. |
35993 | const char *X86TargetLowering:: |
35994 | LowerXConstraint(EVT ConstraintVT) const { |
35995 | // FP X constraints get lowered to SSE1/2 registers if available, otherwise |
35996 | // 'f' like normal targets. |
35997 | if (ConstraintVT.isFloatingPoint()) { |
35998 | if (Subtarget.hasSSE2()) |
35999 | return "Y"; |
36000 | if (Subtarget.hasSSE1()) |
36001 | return "x"; |
36002 | } |
36003 | |
36004 | return TargetLowering::LowerXConstraint(ConstraintVT); |
36005 | } |
36006 | |
36007 | /// Lower the specified operand into the Ops vector. |
36008 | /// If it is invalid, don't add anything to Ops. |
36009 | void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op, |
36010 | std::string &Constraint, |
36011 | std::vector<SDValue>&Ops, |
36012 | SelectionDAG &DAG) const { |
36013 | SDValue Result; |
36014 | |
36015 | // Only support length 1 constraints for now. |
36016 | if (Constraint.length() > 1) return; |
36017 | |
36018 | char ConstraintLetter = Constraint[0]; |
36019 | switch (ConstraintLetter) { |
36020 | default: break; |
36021 | case 'I': |
36022 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { |
36023 | if (C->getZExtValue() <= 31) { |
36024 | Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op), |
36025 | Op.getValueType()); |
36026 | break; |
36027 | } |
36028 | } |
36029 | return; |
36030 | case 'J': |
36031 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { |
36032 | if (C->getZExtValue() <= 63) { |
36033 | Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op), |
36034 | Op.getValueType()); |
36035 | break; |
36036 | } |
36037 | } |
36038 | return; |
36039 | case 'K': |
36040 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { |
36041 | if (isInt<8>(C->getSExtValue())) { |
36042 | Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op), |
36043 | Op.getValueType()); |
36044 | break; |
36045 | } |
36046 | } |
36047 | return; |
36048 | case 'L': |
36049 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { |
36050 | if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff || |
36051 | (Subtarget.is64Bit() && C->getZExtValue() == 0xffffffff)) { |
36052 | Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), |
36053 | Op.getValueType()); |
36054 | break; |
36055 | } |
36056 | } |
36057 | return; |
36058 | case 'M': |
36059 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { |
36060 | if (C->getZExtValue() <= 3) { |
36061 | Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op), |
36062 | Op.getValueType()); |
36063 | break; |
36064 | } |
36065 | } |
36066 | return; |
36067 | case 'N': |
36068 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { |
36069 | if (C->getZExtValue() <= 255) { |
36070 | Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op), |
36071 | Op.getValueType()); |
36072 | break; |
36073 | } |
36074 | } |
36075 | return; |
36076 | case 'O': |
36077 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { |
36078 | if (C->getZExtValue() <= 127) { |
36079 | Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op), |
36080 | Op.getValueType()); |
36081 | break; |
36082 | } |
36083 | } |
36084 | return; |
36085 | case 'e': { |
36086 | // 32-bit signed value |
36087 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { |
36088 | if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()), |
36089 | C->getSExtValue())) { |
36090 | // Widen to 64 bits here to get it sign extended. |
36091 | Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64); |
36092 | break; |
36093 | } |
36094 | // FIXME gcc accepts some relocatable values here too, but only in certain |
36095 | // memory models; it's complicated. |
36096 | } |
36097 | return; |
36098 | } |
36099 | case 'Z': { |
36100 | // 32-bit unsigned value |
36101 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { |
36102 | if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()), |
36103 | C->getZExtValue())) { |
36104 | Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op), |
36105 | Op.getValueType()); |
36106 | break; |
36107 | } |
36108 | } |
36109 | // FIXME gcc accepts some relocatable values here too, but only in certain |
36110 | // memory models; it's complicated. |
36111 | return; |
36112 | } |
36113 | case 'i': { |
36114 | // Literal immediates are always ok. |
36115 | if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) { |
36116 | // Widen to 64 bits here to get it sign extended. |
36117 | Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64); |
36118 | break; |
36119 | } |
36120 | |
36121 | // In any sort of PIC mode addresses need to be computed at runtime by |
36122 | // adding in a register or some sort of table lookup. These can't |
36123 | // be used as immediates. |
36124 | if (Subtarget.isPICStyleGOT() || Subtarget.isPICStyleStubPIC()) |
36125 | return; |
36126 | |
36127 | // If we are in non-pic codegen mode, we allow the address of a global (with |
36128 | // an optional displacement) to be used with 'i'. |
36129 | GlobalAddressSDNode *GA = nullptr; |
36130 | int64_t Offset = 0; |
36131 | |
36132 | // Match either (GA), (GA+C), (GA+C1+C2), etc. |
36133 | while (1) { |
36134 | if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) { |
36135 | Offset += GA->getOffset(); |
36136 | break; |
36137 | } else if (Op.getOpcode() == ISD::ADD) { |
36138 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { |
36139 | Offset += C->getZExtValue(); |
36140 | Op = Op.getOperand(0); |
36141 | continue; |
36142 | } |
36143 | } else if (Op.getOpcode() == ISD::SUB) { |
36144 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { |
36145 | Offset += -C->getZExtValue(); |
36146 | Op = Op.getOperand(0); |
36147 | continue; |
36148 | } |
36149 | } |
36150 | |
36151 | // Otherwise, this isn't something we can handle, reject it. |
36152 | return; |
36153 | } |
36154 | |
36155 | const GlobalValue *GV = GA->getGlobal(); |
36156 | // If we require an extra load to get this address, as in PIC mode, we |
36157 | // can't accept it. |
36158 | if (isGlobalStubReference(Subtarget.classifyGlobalReference(GV))) |
36159 | return; |
36160 | |
36161 | Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op), |
36162 | GA->getValueType(0), Offset); |
36163 | break; |
36164 | } |
36165 | } |
36166 | |
36167 | if (Result.getNode()) { |
36168 | Ops.push_back(Result); |
36169 | return; |
36170 | } |
36171 | return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); |
36172 | } |
36173 | |
36174 | /// Check if \p RC is a general purpose register class. |
36175 | /// I.e., GR* or one of their variant. |
36176 | static bool isGRClass(const TargetRegisterClass &RC) { |
36177 | return RC.hasSuperClassEq(&X86::GR8RegClass) || |
36178 | RC.hasSuperClassEq(&X86::GR16RegClass) || |
36179 | RC.hasSuperClassEq(&X86::GR32RegClass) || |
36180 | RC.hasSuperClassEq(&X86::GR64RegClass) || |
36181 | RC.hasSuperClassEq(&X86::LOW32_ADDR_ACCESS_RBPRegClass); |
36182 | } |
36183 | |
36184 | /// Check if \p RC is a vector register class. |
36185 | /// I.e., FR* / VR* or one of their variant. |
36186 | static bool isFRClass(const TargetRegisterClass &RC) { |
36187 | return RC.hasSuperClassEq(&X86::FR32XRegClass) || |
36188 | RC.hasSuperClassEq(&X86::FR64XRegClass) || |
36189 | RC.hasSuperClassEq(&X86::VR128XRegClass) || |
36190 | RC.hasSuperClassEq(&X86::VR256XRegClass) || |
36191 | RC.hasSuperClassEq(&X86::VR512RegClass); |
36192 | } |
36193 | |
36194 | std::pair<unsigned, const TargetRegisterClass *> |
36195 | X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, |
36196 | StringRef Constraint, |
36197 | MVT VT) const { |
36198 | // First, see if this is a constraint that directly corresponds to an LLVM |
36199 | // register class. |
36200 | if (Constraint.size() == 1) { |
36201 | // GCC Constraint Letters |
36202 | switch (Constraint[0]) { |
36203 | default: break; |
36204 | // TODO: Slight differences here in allocation order and leaving |
36205 | // RIP in the class. Do they matter any more here than they do |
36206 | // in the normal allocation? |
36207 | case 'k': |
36208 | if (Subtarget.hasAVX512()) { |
36209 | // Only supported in AVX512 or later. |
36210 | switch (VT.SimpleTy) { |
36211 | default: break; |
36212 | case MVT::i32: |
36213 | return std::make_pair(0U, &X86::VK32RegClass); |
36214 | case MVT::i16: |
36215 | return std::make_pair(0U, &X86::VK16RegClass); |
36216 | case MVT::i8: |
36217 | return std::make_pair(0U, &X86::VK8RegClass); |
36218 | case MVT::i1: |
36219 | return std::make_pair(0U, &X86::VK1RegClass); |
36220 | case MVT::i64: |
36221 | return std::make_pair(0U, &X86::VK64RegClass); |
36222 | } |
36223 | } |
36224 | break; |
36225 | case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode. |
36226 | if (Subtarget.is64Bit()) { |
36227 | if (VT == MVT::i32 || VT == MVT::f32) |
36228 | return std::make_pair(0U, &X86::GR32RegClass); |
36229 | if (VT == MVT::i16) |
36230 | return std::make_pair(0U, &X86::GR16RegClass); |
36231 | if (VT == MVT::i8 || VT == MVT::i1) |
36232 | return std::make_pair(0U, &X86::GR8RegClass); |
36233 | if (VT == MVT::i64 || VT == MVT::f64) |
36234 | return std::make_pair(0U, &X86::GR64RegClass); |
36235 | break; |
36236 | } |
36237 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
36238 | // 32-bit fallthrough |
36239 | case 'Q': // Q_REGS |
36240 | if (VT == MVT::i32 || VT == MVT::f32) |
36241 | return std::make_pair(0U, &X86::GR32_ABCDRegClass); |
36242 | if (VT == MVT::i16) |
36243 | return std::make_pair(0U, &X86::GR16_ABCDRegClass); |
36244 | if (VT == MVT::i8 || VT == MVT::i1) |
36245 | return std::make_pair(0U, &X86::GR8_ABCD_LRegClass); |
36246 | if (VT == MVT::i64) |
36247 | return std::make_pair(0U, &X86::GR64_ABCDRegClass); |
36248 | break; |
36249 | case 'r': // GENERAL_REGS |
36250 | case 'l': // INDEX_REGS |
36251 | if (VT == MVT::i8 || VT == MVT::i1) |
36252 | return std::make_pair(0U, &X86::GR8RegClass); |
36253 | if (VT == MVT::i16) |
36254 | return std::make_pair(0U, &X86::GR16RegClass); |
36255 | if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget.is64Bit()) |
36256 | return std::make_pair(0U, &X86::GR32RegClass); |
36257 | return std::make_pair(0U, &X86::GR64RegClass); |
36258 | case 'R': // LEGACY_REGS |
36259 | if (VT == MVT::i8 || VT == MVT::i1) |
36260 | return std::make_pair(0U, &X86::GR8_NOREXRegClass); |
36261 | if (VT == MVT::i16) |
36262 | return std::make_pair(0U, &X86::GR16_NOREXRegClass); |
36263 | if (VT == MVT::i32 || !Subtarget.is64Bit()) |
36264 | return std::make_pair(0U, &X86::GR32_NOREXRegClass); |
36265 | return std::make_pair(0U, &X86::GR64_NOREXRegClass); |
36266 | case 'f': // FP Stack registers. |
36267 | // If SSE is enabled for this VT, use f80 to ensure the isel moves the |
36268 | // value to the correct fpstack register class. |
36269 | if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT)) |
36270 | return std::make_pair(0U, &X86::RFP32RegClass); |
36271 | if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT)) |
36272 | return std::make_pair(0U, &X86::RFP64RegClass); |
36273 | return std::make_pair(0U, &X86::RFP80RegClass); |
36274 | case 'y': // MMX_REGS if MMX allowed. |
36275 | if (!Subtarget.hasMMX()) break; |
36276 | return std::make_pair(0U, &X86::VR64RegClass); |
36277 | case 'Y': // SSE_REGS if SSE2 allowed |
36278 | if (!Subtarget.hasSSE2()) break; |
36279 | LLVM_FALLTHROUGH[[clang::fallthrough]]; |
36280 | case 'v': |
36281 | case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed |
36282 | if (!Subtarget.hasSSE1()) break; |
36283 | bool VConstraint = (Constraint[0] == 'v'); |
36284 | |
36285 | switch (VT.SimpleTy) { |
36286 | default: break; |
36287 | // Scalar SSE types. |
36288 | case MVT::f32: |
36289 | case MVT::i32: |
36290 | if (VConstraint && Subtarget.hasAVX512() && Subtarget.hasVLX()) |
36291 | return std::make_pair(0U, &X86::FR32XRegClass); |
36292 | return std::make_pair(0U, &X86::FR32RegClass); |
36293 | case MVT::f64: |
36294 | case MVT::i64: |
36295 | if (VConstraint && Subtarget.hasVLX()) |
36296 | return std::make_pair(0U, &X86::FR64XRegClass); |
36297 | return std::make_pair(0U, &X86::FR64RegClass); |
36298 | // TODO: Handle f128 and i128 in FR128RegClass after it is tested well. |
36299 | // Vector types. |
36300 | case MVT::v16i8: |
36301 | case MVT::v8i16: |
36302 | case MVT::v4i32: |
36303 | case MVT::v2i64: |
36304 | case MVT::v4f32: |
36305 | case MVT::v2f64: |
36306 | if (VConstraint && Subtarget.hasVLX()) |
36307 | return std::make_pair(0U, &X86::VR128XRegClass); |
36308 | return std::make_pair(0U, &X86::VR128RegClass); |
36309 | // AVX types. |
36310 | case MVT::v32i8: |
36311 | case MVT::v16i16: |
36312 | case MVT::v8i32: |
36313 | case MVT::v4i64: |
36314 | case MVT::v8f32: |
36315 | case MVT::v4f64: |
36316 | if (VConstraint && Subtarget.hasVLX()) |
36317 | return std::make_pair(0U, &X86::VR256XRegClass); |
36318 | return std::make_pair(0U, &X86::VR256RegClass); |
36319 | case MVT::v8f64: |
36320 | case MVT::v16f32: |
36321 | case MVT::v16i32: |
36322 | case MVT::v8i64: |
36323 | return std::make_pair(0U, &X86::VR512RegClass); |
36324 | } |
36325 | break; |
36326 | } |
36327 | } else if (Constraint.size() == 2 && Constraint[0] == 'Y') { |
36328 | switch (Constraint[1]) { |
36329 | default: |
36330 | break; |
36331 | case 'k': |
36332 | // This register class doesn't allocate k0 for masked vector operation. |
36333 | if (Subtarget.hasAVX512()) { // Only supported in AVX512. |
36334 | switch (VT.SimpleTy) { |
36335 | default: break; |
36336 | case MVT::i32: |
36337 | return std::make_pair(0U, &X86::VK32WMRegClass); |
36338 | case MVT::i16: |
36339 | return std::make_pair(0U, &X86::VK16WMRegClass); |
36340 | case MVT::i8: |
36341 | return std::make_pair(0U, &X86::VK8WMRegClass); |
36342 | case MVT::i1: |
36343 | return std::make_pair(0U, &X86::VK1WMRegClass); |
36344 | case MVT::i64: |
36345 | return std::make_pair(0U, &X86::VK64WMRegClass); |
36346 | } |
36347 | } |
36348 | break; |
36349 | } |
36350 | } |
36351 | |
36352 | // Use the default implementation in TargetLowering to convert the register |
36353 | // constraint into a member of a register class. |
36354 | std::pair<unsigned, const TargetRegisterClass*> Res; |
36355 | Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); |
36356 | |
36357 | // Not found as a standard register? |
36358 | if (!Res.second) { |
36359 | // Map st(0) -> st(7) -> ST0 |
36360 | if (Constraint.size() == 7 && Constraint[0] == '{' && |
36361 | tolower(Constraint[1]) == 's' && |
36362 | tolower(Constraint[2]) == 't' && |
36363 | Constraint[3] == '(' && |
36364 | (Constraint[4] >= '0' && Constraint[4] <= '7') && |
36365 | Constraint[5] == ')' && |
36366 | Constraint[6] == '}') { |
36367 | |
36368 | Res.first = X86::FP0+Constraint[4]-'0'; |
36369 | Res.second = &X86::RFP80RegClass; |
36370 | return Res; |
36371 | } |
36372 | |
36373 | // GCC allows "st(0)" to be called just plain "st". |
36374 | if (StringRef("{st}").equals_lower(Constraint)) { |
36375 | Res.first = X86::FP0; |
36376 | Res.second = &X86::RFP80RegClass; |
36377 | return Res; |
36378 | } |
36379 | |
36380 | // flags -> EFLAGS |
36381 | if (StringRef("{flags}").equals_lower(Constraint)) { |
36382 | Res.first = X86::EFLAGS; |
36383 | Res.second = &X86::CCRRegClass; |
36384 | return Res; |
36385 | } |
36386 | |
36387 | // 'A' means [ER]AX + [ER]DX. |
36388 | if (Constraint == "A") { |
36389 | if (Subtarget.is64Bit()) { |
36390 | Res.first = X86::RAX; |
36391 | Res.second = &X86::GR64_ADRegClass; |
36392 | } else { |
36393 | assert((Subtarget.is32Bit() || Subtarget.is16Bit()) &&(((Subtarget.is32Bit() || Subtarget.is16Bit()) && "Expecting 64, 32 or 16 bit subtarget" ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.is32Bit() || Subtarget.is16Bit()) && \"Expecting 64, 32 or 16 bit subtarget\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 36394, __PRETTY_FUNCTION__)) |
36394 | "Expecting 64, 32 or 16 bit subtarget")(((Subtarget.is32Bit() || Subtarget.is16Bit()) && "Expecting 64, 32 or 16 bit subtarget" ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.is32Bit() || Subtarget.is16Bit()) && \"Expecting 64, 32 or 16 bit subtarget\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 36394, __PRETTY_FUNCTION__)); |
36395 | Res.first = X86::EAX; |
36396 | Res.second = &X86::GR32_ADRegClass; |
36397 | } |
36398 | return Res; |
36399 | } |
36400 | return Res; |
36401 | } |
36402 | |
36403 | // Otherwise, check to see if this is a register class of the wrong value |
36404 | // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to |
36405 | // turn into {ax},{dx}. |
36406 | // MVT::Other is used to specify clobber names. |
36407 | if (TRI->isTypeLegalForClass(*Res.second, VT) || VT == MVT::Other) |
36408 | return Res; // Correct type already, nothing to do. |
36409 | |
36410 | // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should |
36411 | // return "eax". This should even work for things like getting 64bit integer |
36412 | // registers when given an f64 type. |
36413 | const TargetRegisterClass *Class = Res.second; |
36414 | // The generic code will match the first register class that contains the |
36415 | // given register. Thus, based on the ordering of the tablegened file, |
36416 | // the "plain" GR classes might not come first. |
36417 | // Therefore, use a helper method. |
36418 | if (isGRClass(*Class)) { |
36419 | unsigned Size = VT.getSizeInBits(); |
36420 | if (Size == 1) Size = 8; |
36421 | unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, Size); |
36422 | if (DestReg > 0) { |
36423 | Res.first = DestReg; |
36424 | Res.second = Size == 8 ? &X86::GR8RegClass |
36425 | : Size == 16 ? &X86::GR16RegClass |
36426 | : Size == 32 ? &X86::GR32RegClass |
36427 | : &X86::GR64RegClass; |
36428 | assert(Res.second->contains(Res.first) && "Register in register class")((Res.second->contains(Res.first) && "Register in register class" ) ? static_cast<void> (0) : __assert_fail ("Res.second->contains(Res.first) && \"Register in register class\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 36428, __PRETTY_FUNCTION__)); |
36429 | } else { |
36430 | // No register found/type mismatch. |
36431 | Res.first = 0; |
36432 | Res.second = nullptr; |
36433 | } |
36434 | } else if (isFRClass(*Class)) { |
36435 | // Handle references to XMM physical registers that got mapped into the |
36436 | // wrong class. This can happen with constraints like {xmm0} where the |
36437 | // target independent register mapper will just pick the first match it can |
36438 | // find, ignoring the required type. |
36439 | |
36440 | // TODO: Handle f128 and i128 in FR128RegClass after it is tested well. |
36441 | if (VT == MVT::f32 || VT == MVT::i32) |
36442 | Res.second = &X86::FR32RegClass; |
36443 | else if (VT == MVT::f64 || VT == MVT::i64) |
36444 | Res.second = &X86::FR64RegClass; |
36445 | else if (TRI->isTypeLegalForClass(X86::VR128RegClass, VT)) |
36446 | Res.second = &X86::VR128RegClass; |
36447 | else if (TRI->isTypeLegalForClass(X86::VR256RegClass, VT)) |
36448 | Res.second = &X86::VR256RegClass; |
36449 | else if (TRI->isTypeLegalForClass(X86::VR512RegClass, VT)) |
36450 | Res.second = &X86::VR512RegClass; |
36451 | else { |
36452 | // Type mismatch and not a clobber: Return an error; |
36453 | Res.first = 0; |
36454 | Res.second = nullptr; |
36455 | } |
36456 | } |
36457 | |
36458 | return Res; |
36459 | } |
36460 | |
36461 | int X86TargetLowering::getScalingFactorCost(const DataLayout &DL, |
36462 | const AddrMode &AM, Type *Ty, |
36463 | unsigned AS) const { |
36464 | // Scaling factors are not free at all. |
36465 | // An indexed folded instruction, i.e., inst (reg1, reg2, scale), |
36466 | // will take 2 allocations in the out of order engine instead of 1 |
36467 | // for plain addressing mode, i.e. inst (reg1). |
36468 | // E.g., |
36469 | // vaddps (%rsi,%drx), %ymm0, %ymm1 |
36470 | // Requires two allocations (one for the load, one for the computation) |
36471 | // whereas: |
36472 | // vaddps (%rsi), %ymm0, %ymm1 |
36473 | // Requires just 1 allocation, i.e., freeing allocations for other operations |
36474 | // and having less micro operations to execute. |
36475 | // |
36476 | // For some X86 architectures, this is even worse because for instance for |
36477 | // stores, the complex addressing mode forces the instruction to use the |
36478 | // "load" ports instead of the dedicated "store" port. |
36479 | // E.g., on Haswell: |
36480 | // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3. |
36481 | // vmovaps %ymm1, (%r8) can use port 2, 3, or 7. |
36482 | if (isLegalAddressingMode(DL, AM, Ty, AS)) |
36483 | // Scale represents reg2 * scale, thus account for 1 |
36484 | // as soon as we use a second register. |
36485 | return AM.Scale != 0; |
36486 | return -1; |
36487 | } |
36488 | |
36489 | bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeList Attr) const { |
36490 | // Integer division on x86 is expensive. However, when aggressively optimizing |
36491 | // for code size, we prefer to use a div instruction, as it is usually smaller |
36492 | // than the alternative sequence. |
36493 | // The exception to this is vector division. Since x86 doesn't have vector |
36494 | // integer division, leaving the division as-is is a loss even in terms of |
36495 | // size, because it will have to be scalarized, while the alternative code |
36496 | // sequence can be performed in vector form. |
36497 | bool OptSize = |
36498 | Attr.hasAttribute(AttributeList::FunctionIndex, Attribute::MinSize); |
36499 | return OptSize && !VT.isVector(); |
36500 | } |
36501 | |
36502 | void X86TargetLowering::initializeSplitCSR(MachineBasicBlock *Entry) const { |
36503 | if (!Subtarget.is64Bit()) |
36504 | return; |
36505 | |
36506 | // Update IsSplitCSR in X86MachineFunctionInfo. |
36507 | X86MachineFunctionInfo *AFI = |
36508 | Entry->getParent()->getInfo<X86MachineFunctionInfo>(); |
36509 | AFI->setIsSplitCSR(true); |
36510 | } |
36511 | |
36512 | void X86TargetLowering::insertCopiesSplitCSR( |
36513 | MachineBasicBlock *Entry, |
36514 | const SmallVectorImpl<MachineBasicBlock *> &Exits) const { |
36515 | const X86RegisterInfo *TRI = Subtarget.getRegisterInfo(); |
36516 | const MCPhysReg *IStart = TRI->getCalleeSavedRegsViaCopy(Entry->getParent()); |
36517 | if (!IStart) |
36518 | return; |
36519 | |
36520 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
36521 | MachineRegisterInfo *MRI = &Entry->getParent()->getRegInfo(); |
36522 | MachineBasicBlock::iterator MBBI = Entry->begin(); |
36523 | for (const MCPhysReg *I = IStart; *I; ++I) { |
36524 | const TargetRegisterClass *RC = nullptr; |
36525 | if (X86::GR64RegClass.contains(*I)) |
36526 | RC = &X86::GR64RegClass; |
36527 | else |
36528 | llvm_unreachable("Unexpected register class in CSRsViaCopy!")::llvm::llvm_unreachable_internal("Unexpected register class in CSRsViaCopy!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 36528); |
36529 | |
36530 | unsigned NewVR = MRI->createVirtualRegister(RC); |
36531 | // Create copy from CSR to a virtual register. |
36532 | // FIXME: this currently does not emit CFI pseudo-instructions, it works |
36533 | // fine for CXX_FAST_TLS since the C++-style TLS access functions should be |
36534 | // nounwind. If we want to generalize this later, we may need to emit |
36535 | // CFI pseudo-instructions. |
36536 | assert(Entry->getParent()->getFunction()->hasFnAttribute(((Entry->getParent()->getFunction()->hasFnAttribute( Attribute::NoUnwind) && "Function should be nounwind in insertCopiesSplitCSR!" ) ? static_cast<void> (0) : __assert_fail ("Entry->getParent()->getFunction()->hasFnAttribute( Attribute::NoUnwind) && \"Function should be nounwind in insertCopiesSplitCSR!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 36538, __PRETTY_FUNCTION__)) |
36537 | Attribute::NoUnwind) &&((Entry->getParent()->getFunction()->hasFnAttribute( Attribute::NoUnwind) && "Function should be nounwind in insertCopiesSplitCSR!" ) ? static_cast<void> (0) : __assert_fail ("Entry->getParent()->getFunction()->hasFnAttribute( Attribute::NoUnwind) && \"Function should be nounwind in insertCopiesSplitCSR!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 36538, __PRETTY_FUNCTION__)) |
36538 | "Function should be nounwind in insertCopiesSplitCSR!")((Entry->getParent()->getFunction()->hasFnAttribute( Attribute::NoUnwind) && "Function should be nounwind in insertCopiesSplitCSR!" ) ? static_cast<void> (0) : __assert_fail ("Entry->getParent()->getFunction()->hasFnAttribute( Attribute::NoUnwind) && \"Function should be nounwind in insertCopiesSplitCSR!\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/X86/X86ISelLowering.cpp" , 36538, __PRETTY_FUNCTION__)); |
36539 | Entry->addLiveIn(*I); |
36540 | BuildMI(*Entry, MBBI, DebugLoc(), TII->get(TargetOpcode::COPY), NewVR) |
36541 | .addReg(*I); |
36542 | |
36543 | // Insert the copy-back instructions right before the terminator. |
36544 | for (auto *Exit : Exits) |
36545 | BuildMI(*Exit, Exit->getFirstTerminator(), DebugLoc(), |
36546 | TII->get(TargetOpcode::COPY), *I) |
36547 | .addReg(NewVR); |
36548 | } |
36549 | } |
36550 | |
36551 | bool X86TargetLowering::supportSwiftError() const { |
36552 | return Subtarget.is64Bit(); |
36553 | } |
36554 | |
36555 | /// Returns the name of the symbol used to emit stack probes or the empty |
36556 | /// string if not applicable. |
36557 | StringRef X86TargetLowering::getStackProbeSymbolName(MachineFunction &MF) const { |
36558 | // If the function specifically requests stack probes, emit them. |
36559 | if (MF.getFunction()->hasFnAttribute("probe-stack")) |
36560 | return MF.getFunction()->getFnAttribute("probe-stack").getValueAsString(); |
36561 | |
36562 | // Generally, if we aren't on Windows, the platform ABI does not include |
36563 | // support for stack probes, so don't emit them. |
36564 | if (!Subtarget.isOSWindows() || Subtarget.isTargetMachO()) |
36565 | return ""; |
36566 | |
36567 | // We need a stack probe to conform to the Windows ABI. Choose the right |
36568 | // symbol. |
36569 | if (Subtarget.is64Bit()) |
36570 | return Subtarget.isTargetCygMing() ? "___chkstk_ms" : "__chkstk"; |
36571 | return Subtarget.isTargetCygMing() ? "_alloca" : "_chkstk"; |
36572 | } |