File: | lib/Target/X86/X86ISelLowering.cpp |
Location: | line 6075, column 33 |
Description: | Called C++ object pointer is null |
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 "X86MachineFunctionInfo.h" | |||
21 | #include "X86TargetMachine.h" | |||
22 | #include "X86TargetObjectFile.h" | |||
23 | #include "llvm/ADT/SmallBitVector.h" | |||
24 | #include "llvm/ADT/SmallSet.h" | |||
25 | #include "llvm/ADT/Statistic.h" | |||
26 | #include "llvm/ADT/StringExtras.h" | |||
27 | #include "llvm/ADT/StringSwitch.h" | |||
28 | #include "llvm/ADT/VariadicFunction.h" | |||
29 | #include "llvm/CodeGen/IntrinsicLowering.h" | |||
30 | #include "llvm/CodeGen/MachineFrameInfo.h" | |||
31 | #include "llvm/CodeGen/MachineFunction.h" | |||
32 | #include "llvm/CodeGen/MachineInstrBuilder.h" | |||
33 | #include "llvm/CodeGen/MachineJumpTableInfo.h" | |||
34 | #include "llvm/CodeGen/MachineModuleInfo.h" | |||
35 | #include "llvm/CodeGen/MachineRegisterInfo.h" | |||
36 | #include "llvm/IR/CallSite.h" | |||
37 | #include "llvm/IR/CallingConv.h" | |||
38 | #include "llvm/IR/Constants.h" | |||
39 | #include "llvm/IR/DerivedTypes.h" | |||
40 | #include "llvm/IR/Function.h" | |||
41 | #include "llvm/IR/GlobalAlias.h" | |||
42 | #include "llvm/IR/GlobalVariable.h" | |||
43 | #include "llvm/IR/Instructions.h" | |||
44 | #include "llvm/IR/Intrinsics.h" | |||
45 | #include "llvm/MC/MCAsmInfo.h" | |||
46 | #include "llvm/MC/MCContext.h" | |||
47 | #include "llvm/MC/MCExpr.h" | |||
48 | #include "llvm/MC/MCSymbol.h" | |||
49 | #include "llvm/Support/CommandLine.h" | |||
50 | #include "llvm/Support/Debug.h" | |||
51 | #include "llvm/Support/ErrorHandling.h" | |||
52 | #include "llvm/Support/MathExtras.h" | |||
53 | #include "llvm/Target/TargetOptions.h" | |||
54 | #include "X86IntrinsicsInfo.h" | |||
55 | #include <bitset> | |||
56 | #include <numeric> | |||
57 | #include <cctype> | |||
58 | using namespace llvm; | |||
59 | ||||
60 | #define DEBUG_TYPE"x86-isel" "x86-isel" | |||
61 | ||||
62 | STATISTIC(NumTailCalls, "Number of tail calls")static llvm::Statistic NumTailCalls = { "x86-isel", "Number of tail calls" , 0, 0 }; | |||
63 | ||||
64 | static cl::opt<bool> ExperimentalVectorWideningLegalization( | |||
65 | "x86-experimental-vector-widening-legalization", cl::init(false), | |||
66 | cl::desc("Enable an experimental vector type legalization through widening " | |||
67 | "rather than promotion."), | |||
68 | cl::Hidden); | |||
69 | ||||
70 | static cl::opt<bool> ExperimentalVectorShuffleLowering( | |||
71 | "x86-experimental-vector-shuffle-lowering", cl::init(true), | |||
72 | cl::desc("Enable an experimental vector shuffle lowering code path."), | |||
73 | cl::Hidden); | |||
74 | ||||
75 | static cl::opt<bool> ExperimentalVectorShuffleLegality( | |||
76 | "x86-experimental-vector-shuffle-legality", cl::init(false), | |||
77 | cl::desc("Enable experimental shuffle legality based on the experimental " | |||
78 | "shuffle lowering. Should only be used with the experimental " | |||
79 | "shuffle lowering."), | |||
80 | cl::Hidden); | |||
81 | ||||
82 | static cl::opt<int> ReciprocalEstimateRefinementSteps( | |||
83 | "x86-recip-refinement-steps", cl::init(1), | |||
84 | cl::desc("Specify the number of Newton-Raphson iterations applied to the " | |||
85 | "result of the hardware reciprocal estimate instruction."), | |||
86 | cl::NotHidden); | |||
87 | ||||
88 | // Forward declarations. | |||
89 | static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1, | |||
90 | SDValue V2); | |||
91 | ||||
92 | static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal, | |||
93 | SelectionDAG &DAG, SDLoc dl, | |||
94 | unsigned vectorWidth) { | |||
95 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 96, __PRETTY_FUNCTION__)) | |||
96 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 96, __PRETTY_FUNCTION__)); | |||
97 | EVT VT = Vec.getValueType(); | |||
98 | EVT ElVT = VT.getVectorElementType(); | |||
99 | unsigned Factor = VT.getSizeInBits()/vectorWidth; | |||
100 | EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT, | |||
101 | VT.getVectorNumElements()/Factor); | |||
102 | ||||
103 | // Extract from UNDEF is UNDEF. | |||
104 | if (Vec.getOpcode() == ISD::UNDEF) | |||
105 | return DAG.getUNDEF(ResultVT); | |||
106 | ||||
107 | // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR | |||
108 | unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits(); | |||
109 | ||||
110 | // This is the index of the first element of the vectorWidth-bit chunk | |||
111 | // we want. | |||
112 | unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth) | |||
113 | * ElemsPerChunk); | |||
114 | ||||
115 | // If the input is a buildvector just emit a smaller one. | |||
116 | if (Vec.getOpcode() == ISD::BUILD_VECTOR) | |||
117 | return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT, | |||
118 | makeArrayRef(Vec->op_begin() + NormalizedIdxVal, | |||
119 | ElemsPerChunk)); | |||
120 | ||||
121 | SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal); | |||
122 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx); | |||
123 | } | |||
124 | ||||
125 | /// Generate a DAG to grab 128-bits from a vector > 128 bits. This | |||
126 | /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128 | |||
127 | /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4 | |||
128 | /// instructions or a simple subregister reference. Idx is an index in the | |||
129 | /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes | |||
130 | /// lowering EXTRACT_VECTOR_ELT operations easier. | |||
131 | static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal, | |||
132 | SelectionDAG &DAG, SDLoc dl) { | |||
133 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 134, __PRETTY_FUNCTION__)) | |||
134 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 134, __PRETTY_FUNCTION__)); | |||
135 | return ExtractSubVector(Vec, IdxVal, DAG, dl, 128); | |||
136 | } | |||
137 | ||||
138 | /// Generate a DAG to grab 256-bits from a 512-bit vector. | |||
139 | static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal, | |||
140 | SelectionDAG &DAG, SDLoc dl) { | |||
141 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 141, __PRETTY_FUNCTION__)); | |||
142 | return ExtractSubVector(Vec, IdxVal, DAG, dl, 256); | |||
143 | } | |||
144 | ||||
145 | static SDValue InsertSubVector(SDValue Result, SDValue Vec, | |||
146 | unsigned IdxVal, SelectionDAG &DAG, | |||
147 | SDLoc dl, unsigned vectorWidth) { | |||
148 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 149, __PRETTY_FUNCTION__)) | |||
149 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 149, __PRETTY_FUNCTION__)); | |||
150 | // Inserting UNDEF is Result | |||
151 | if (Vec.getOpcode() == ISD::UNDEF) | |||
152 | return Result; | |||
153 | EVT VT = Vec.getValueType(); | |||
154 | EVT ElVT = VT.getVectorElementType(); | |||
155 | EVT ResultVT = Result.getValueType(); | |||
156 | ||||
157 | // Insert the relevant vectorWidth bits. | |||
158 | unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits(); | |||
159 | ||||
160 | // This is the index of the first element of the vectorWidth-bit chunk | |||
161 | // we want. | |||
162 | unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth) | |||
163 | * ElemsPerChunk); | |||
164 | ||||
165 | SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal); | |||
166 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx); | |||
167 | } | |||
168 | ||||
169 | /// Generate a DAG to put 128-bits into a vector > 128 bits. This | |||
170 | /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or | |||
171 | /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a | |||
172 | /// simple superregister reference. Idx is an index in the 128 bits | |||
173 | /// we want. It need not be aligned to a 128-bit boundary. That makes | |||
174 | /// lowering INSERT_VECTOR_ELT operations easier. | |||
175 | static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal, | |||
176 | SelectionDAG &DAG,SDLoc dl) { | |||
177 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 177, __PRETTY_FUNCTION__)); | |||
178 | return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128); | |||
179 | } | |||
180 | ||||
181 | static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal, | |||
182 | SelectionDAG &DAG, SDLoc dl) { | |||
183 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 183, __PRETTY_FUNCTION__)); | |||
184 | return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256); | |||
185 | } | |||
186 | ||||
187 | /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128 | |||
188 | /// instructions. This is used because creating CONCAT_VECTOR nodes of | |||
189 | /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower | |||
190 | /// large BUILD_VECTORS. | |||
191 | static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT, | |||
192 | unsigned NumElems, SelectionDAG &DAG, | |||
193 | SDLoc dl) { | |||
194 | SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl); | |||
195 | return Insert128BitVector(V, V2, NumElems/2, DAG, dl); | |||
196 | } | |||
197 | ||||
198 | static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT, | |||
199 | unsigned NumElems, SelectionDAG &DAG, | |||
200 | SDLoc dl) { | |||
201 | SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl); | |||
202 | return Insert256BitVector(V, V2, NumElems/2, DAG, dl); | |||
203 | } | |||
204 | ||||
205 | X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM) | |||
206 | : TargetLowering(TM) { | |||
207 | Subtarget = &TM.getSubtarget<X86Subtarget>(); | |||
208 | X86ScalarSSEf64 = Subtarget->hasSSE2(); | |||
209 | X86ScalarSSEf32 = Subtarget->hasSSE1(); | |||
210 | TD = getDataLayout(); | |||
211 | ||||
212 | // Set up the TargetLowering object. | |||
213 | static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }; | |||
214 | ||||
215 | // X86 is weird. It always uses i8 for shift amounts and setcc results. | |||
216 | setBooleanContents(ZeroOrOneBooleanContent); | |||
217 | // X86-SSE is even stranger. It uses -1 or 0 for vector masks. | |||
218 | setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); | |||
219 | ||||
220 | // For 64-bit, since we have so many registers, use the ILP scheduler. | |||
221 | // For 32-bit, use the register pressure specific scheduling. | |||
222 | // For Atom, always use ILP scheduling. | |||
223 | if (Subtarget->isAtom()) | |||
224 | setSchedulingPreference(Sched::ILP); | |||
225 | else if (Subtarget->is64Bit()) | |||
226 | setSchedulingPreference(Sched::ILP); | |||
227 | else | |||
228 | setSchedulingPreference(Sched::RegPressure); | |||
229 | const X86RegisterInfo *RegInfo = | |||
230 | TM.getSubtarget<X86Subtarget>().getRegisterInfo(); | |||
231 | setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister()); | |||
232 | ||||
233 | // Bypass expensive divides on Atom when compiling with O2. | |||
234 | if (TM.getOptLevel() >= CodeGenOpt::Default) { | |||
235 | if (Subtarget->hasSlowDivide32()) | |||
236 | addBypassSlowDiv(32, 8); | |||
237 | if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit()) | |||
238 | addBypassSlowDiv(64, 16); | |||
239 | } | |||
240 | ||||
241 | if (Subtarget->isTargetKnownWindowsMSVC()) { | |||
242 | // Setup Windows compiler runtime calls. | |||
243 | setLibcallName(RTLIB::SDIV_I64, "_alldiv"); | |||
244 | setLibcallName(RTLIB::UDIV_I64, "_aulldiv"); | |||
245 | setLibcallName(RTLIB::SREM_I64, "_allrem"); | |||
246 | setLibcallName(RTLIB::UREM_I64, "_aullrem"); | |||
247 | setLibcallName(RTLIB::MUL_I64, "_allmul"); | |||
248 | setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall); | |||
249 | setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall); | |||
250 | setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall); | |||
251 | setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall); | |||
252 | setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall); | |||
253 | ||||
254 | // The _ftol2 runtime function has an unusual calling conv, which | |||
255 | // is modeled by a special pseudo-instruction. | |||
256 | setLibcallName(RTLIB::FPTOUINT_F64_I64, nullptr); | |||
257 | setLibcallName(RTLIB::FPTOUINT_F32_I64, nullptr); | |||
258 | setLibcallName(RTLIB::FPTOUINT_F64_I32, nullptr); | |||
259 | setLibcallName(RTLIB::FPTOUINT_F32_I32, nullptr); | |||
260 | } | |||
261 | ||||
262 | if (Subtarget->isTargetDarwin()) { | |||
263 | // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp. | |||
264 | setUseUnderscoreSetJmp(false); | |||
265 | setUseUnderscoreLongJmp(false); | |||
266 | } else if (Subtarget->isTargetWindowsGNU()) { | |||
267 | // MS runtime is weird: it exports _setjmp, but longjmp! | |||
268 | setUseUnderscoreSetJmp(true); | |||
269 | setUseUnderscoreLongJmp(false); | |||
270 | } else { | |||
271 | setUseUnderscoreSetJmp(true); | |||
272 | setUseUnderscoreLongJmp(true); | |||
273 | } | |||
274 | ||||
275 | // Set up the register classes. | |||
276 | addRegisterClass(MVT::i8, &X86::GR8RegClass); | |||
277 | addRegisterClass(MVT::i16, &X86::GR16RegClass); | |||
278 | addRegisterClass(MVT::i32, &X86::GR32RegClass); | |||
279 | if (Subtarget->is64Bit()) | |||
280 | addRegisterClass(MVT::i64, &X86::GR64RegClass); | |||
281 | ||||
282 | for (MVT VT : MVT::integer_valuetypes()) | |||
283 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote); | |||
284 | ||||
285 | // We don't accept any truncstore of integer registers. | |||
286 | setTruncStoreAction(MVT::i64, MVT::i32, Expand); | |||
287 | setTruncStoreAction(MVT::i64, MVT::i16, Expand); | |||
288 | setTruncStoreAction(MVT::i64, MVT::i8 , Expand); | |||
289 | setTruncStoreAction(MVT::i32, MVT::i16, Expand); | |||
290 | setTruncStoreAction(MVT::i32, MVT::i8 , Expand); | |||
291 | setTruncStoreAction(MVT::i16, MVT::i8, Expand); | |||
292 | ||||
293 | setTruncStoreAction(MVT::f64, MVT::f32, Expand); | |||
294 | ||||
295 | // SETOEQ and SETUNE require checking two conditions. | |||
296 | setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand); | |||
297 | setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand); | |||
298 | setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand); | |||
299 | setCondCodeAction(ISD::SETUNE, MVT::f32, Expand); | |||
300 | setCondCodeAction(ISD::SETUNE, MVT::f64, Expand); | |||
301 | setCondCodeAction(ISD::SETUNE, MVT::f80, Expand); | |||
302 | ||||
303 | // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this | |||
304 | // operation. | |||
305 | setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote); | |||
306 | setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote); | |||
307 | setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote); | |||
308 | ||||
309 | if (Subtarget->is64Bit()) { | |||
310 | setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote); | |||
311 | setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom); | |||
312 | } else if (!TM.Options.UseSoftFloat) { | |||
313 | // We have an algorithm for SSE2->double, and we turn this into a | |||
314 | // 64-bit FILD followed by conditional FADD for other targets. | |||
315 | setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom); | |||
316 | // We have an algorithm for SSE2, and we turn this into a 64-bit | |||
317 | // FILD for other targets. | |||
318 | setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom); | |||
319 | } | |||
320 | ||||
321 | // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have | |||
322 | // this operation. | |||
323 | setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote); | |||
324 | setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote); | |||
325 | ||||
326 | if (!TM.Options.UseSoftFloat) { | |||
327 | // SSE has no i16 to fp conversion, only i32 | |||
328 | if (X86ScalarSSEf32) { | |||
329 | setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote); | |||
330 | // f32 and f64 cases are Legal, f80 case is not | |||
331 | setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom); | |||
332 | } else { | |||
333 | setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom); | |||
334 | setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom); | |||
335 | } | |||
336 | } else { | |||
337 | setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote); | |||
338 | setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote); | |||
339 | } | |||
340 | ||||
341 | // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64 | |||
342 | // are Legal, f80 is custom lowered. | |||
343 | setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom); | |||
344 | setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom); | |||
345 | ||||
346 | // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have | |||
347 | // this operation. | |||
348 | setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote); | |||
349 | setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote); | |||
350 | ||||
351 | if (X86ScalarSSEf32) { | |||
352 | setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote); | |||
353 | // f32 and f64 cases are Legal, f80 case is not | |||
354 | setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom); | |||
355 | } else { | |||
356 | setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom); | |||
357 | setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom); | |||
358 | } | |||
359 | ||||
360 | // Handle FP_TO_UINT by promoting the destination to a larger signed | |||
361 | // conversion. | |||
362 | setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote); | |||
363 | setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote); | |||
364 | setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote); | |||
365 | ||||
366 | if (Subtarget->is64Bit()) { | |||
367 | setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand); | |||
368 | setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote); | |||
369 | } else if (!TM.Options.UseSoftFloat) { | |||
370 | // Since AVX is a superset of SSE3, only check for SSE here. | |||
371 | if (Subtarget->hasSSE1() && !Subtarget->hasSSE3()) | |||
372 | // Expand FP_TO_UINT into a select. | |||
373 | // FIXME: We would like to use a Custom expander here eventually to do | |||
374 | // the optimal thing for SSE vs. the default expansion in the legalizer. | |||
375 | setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand); | |||
376 | else | |||
377 | // With SSE3 we can use fisttpll to convert to a signed i64; without | |||
378 | // SSE, we're stuck with a fistpll. | |||
379 | setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom); | |||
380 | } | |||
381 | ||||
382 | if (isTargetFTOL()) { | |||
383 | // Use the _ftol2 runtime function, which has a pseudo-instruction | |||
384 | // to handle its weird calling convention. | |||
385 | setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom); | |||
386 | } | |||
387 | ||||
388 | // TODO: when we have SSE, these could be more efficient, by using movd/movq. | |||
389 | if (!X86ScalarSSEf64) { | |||
390 | setOperationAction(ISD::BITCAST , MVT::f32 , Expand); | |||
391 | setOperationAction(ISD::BITCAST , MVT::i32 , Expand); | |||
392 | if (Subtarget->is64Bit()) { | |||
393 | setOperationAction(ISD::BITCAST , MVT::f64 , Expand); | |||
394 | // Without SSE, i64->f64 goes through memory. | |||
395 | setOperationAction(ISD::BITCAST , MVT::i64 , Expand); | |||
396 | } | |||
397 | } | |||
398 | ||||
399 | // Scalar integer divide and remainder are lowered to use operations that | |||
400 | // produce two results, to match the available instructions. This exposes | |||
401 | // the two-result form to trivial CSE, which is able to combine x/y and x%y | |||
402 | // into a single instruction. | |||
403 | // | |||
404 | // Scalar integer multiply-high is also lowered to use two-result | |||
405 | // operations, to match the available instructions. However, plain multiply | |||
406 | // (low) operations are left as Legal, as there are single-result | |||
407 | // instructions for this in x86. Using the two-result multiply instructions | |||
408 | // when both high and low results are needed must be arranged by dagcombine. | |||
409 | for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) { | |||
410 | MVT VT = IntVTs[i]; | |||
411 | setOperationAction(ISD::MULHS, VT, Expand); | |||
412 | setOperationAction(ISD::MULHU, VT, Expand); | |||
413 | setOperationAction(ISD::SDIV, VT, Expand); | |||
414 | setOperationAction(ISD::UDIV, VT, Expand); | |||
415 | setOperationAction(ISD::SREM, VT, Expand); | |||
416 | setOperationAction(ISD::UREM, VT, Expand); | |||
417 | ||||
418 | // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences. | |||
419 | setOperationAction(ISD::ADDC, VT, Custom); | |||
420 | setOperationAction(ISD::ADDE, VT, Custom); | |||
421 | setOperationAction(ISD::SUBC, VT, Custom); | |||
422 | setOperationAction(ISD::SUBE, VT, Custom); | |||
423 | } | |||
424 | ||||
425 | setOperationAction(ISD::BR_JT , MVT::Other, Expand); | |||
426 | setOperationAction(ISD::BRCOND , MVT::Other, Custom); | |||
427 | setOperationAction(ISD::BR_CC , MVT::f32, Expand); | |||
428 | setOperationAction(ISD::BR_CC , MVT::f64, Expand); | |||
429 | setOperationAction(ISD::BR_CC , MVT::f80, Expand); | |||
430 | setOperationAction(ISD::BR_CC , MVT::i8, Expand); | |||
431 | setOperationAction(ISD::BR_CC , MVT::i16, Expand); | |||
432 | setOperationAction(ISD::BR_CC , MVT::i32, Expand); | |||
433 | setOperationAction(ISD::BR_CC , MVT::i64, Expand); | |||
434 | setOperationAction(ISD::SELECT_CC , MVT::f32, Expand); | |||
435 | setOperationAction(ISD::SELECT_CC , MVT::f64, Expand); | |||
436 | setOperationAction(ISD::SELECT_CC , MVT::f80, Expand); | |||
437 | setOperationAction(ISD::SELECT_CC , MVT::i8, Expand); | |||
438 | setOperationAction(ISD::SELECT_CC , MVT::i16, Expand); | |||
439 | setOperationAction(ISD::SELECT_CC , MVT::i32, Expand); | |||
440 | setOperationAction(ISD::SELECT_CC , MVT::i64, Expand); | |||
441 | if (Subtarget->is64Bit()) | |||
442 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal); | |||
443 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal); | |||
444 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal); | |||
445 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand); | |||
446 | setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand); | |||
447 | setOperationAction(ISD::FREM , MVT::f32 , Expand); | |||
448 | setOperationAction(ISD::FREM , MVT::f64 , Expand); | |||
449 | setOperationAction(ISD::FREM , MVT::f80 , Expand); | |||
450 | setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom); | |||
451 | ||||
452 | // Promote the i8 variants and force them on up to i32 which has a shorter | |||
453 | // encoding. | |||
454 | setOperationAction(ISD::CTTZ , MVT::i8 , Promote); | |||
455 | AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32); | |||
456 | setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote); | |||
457 | AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32); | |||
458 | if (Subtarget->hasBMI()) { | |||
459 | setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand); | |||
460 | setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand); | |||
461 | if (Subtarget->is64Bit()) | |||
462 | setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand); | |||
463 | } else { | |||
464 | setOperationAction(ISD::CTTZ , MVT::i16 , Custom); | |||
465 | setOperationAction(ISD::CTTZ , MVT::i32 , Custom); | |||
466 | if (Subtarget->is64Bit()) | |||
467 | setOperationAction(ISD::CTTZ , MVT::i64 , Custom); | |||
468 | } | |||
469 | ||||
470 | if (Subtarget->hasLZCNT()) { | |||
471 | // When promoting the i8 variants, force them to i32 for a shorter | |||
472 | // encoding. | |||
473 | setOperationAction(ISD::CTLZ , MVT::i8 , Promote); | |||
474 | AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32); | |||
475 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote); | |||
476 | AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32); | |||
477 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand); | |||
478 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand); | |||
479 | if (Subtarget->is64Bit()) | |||
480 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand); | |||
481 | } else { | |||
482 | setOperationAction(ISD::CTLZ , MVT::i8 , Custom); | |||
483 | setOperationAction(ISD::CTLZ , MVT::i16 , Custom); | |||
484 | setOperationAction(ISD::CTLZ , MVT::i32 , Custom); | |||
485 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom); | |||
486 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom); | |||
487 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom); | |||
488 | if (Subtarget->is64Bit()) { | |||
489 | setOperationAction(ISD::CTLZ , MVT::i64 , Custom); | |||
490 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom); | |||
491 | } | |||
492 | } | |||
493 | ||||
494 | // Special handling for half-precision floating point conversions. | |||
495 | // If we don't have F16C support, then lower half float conversions | |||
496 | // into library calls. | |||
497 | if (TM.Options.UseSoftFloat || !Subtarget->hasF16C()) { | |||
498 | setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand); | |||
499 | setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand); | |||
500 | } | |||
501 | ||||
502 | // There's never any support for operations beyond MVT::f32. | |||
503 | setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand); | |||
504 | setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand); | |||
505 | setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand); | |||
506 | setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand); | |||
507 | ||||
508 | setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand); | |||
509 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand); | |||
510 | setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand); | |||
511 | setTruncStoreAction(MVT::f32, MVT::f16, Expand); | |||
512 | setTruncStoreAction(MVT::f64, MVT::f16, Expand); | |||
513 | setTruncStoreAction(MVT::f80, MVT::f16, Expand); | |||
514 | ||||
515 | if (Subtarget->hasPOPCNT()) { | |||
516 | setOperationAction(ISD::CTPOP , MVT::i8 , Promote); | |||
517 | } else { | |||
518 | setOperationAction(ISD::CTPOP , MVT::i8 , Expand); | |||
519 | setOperationAction(ISD::CTPOP , MVT::i16 , Expand); | |||
520 | setOperationAction(ISD::CTPOP , MVT::i32 , Expand); | |||
521 | if (Subtarget->is64Bit()) | |||
522 | setOperationAction(ISD::CTPOP , MVT::i64 , Expand); | |||
523 | } | |||
524 | ||||
525 | setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom); | |||
526 | ||||
527 | if (!Subtarget->hasMOVBE()) | |||
528 | setOperationAction(ISD::BSWAP , MVT::i16 , Expand); | |||
529 | ||||
530 | // These should be promoted to a larger select which is supported. | |||
531 | setOperationAction(ISD::SELECT , MVT::i1 , Promote); | |||
532 | // X86 wants to expand cmov itself. | |||
533 | setOperationAction(ISD::SELECT , MVT::i8 , Custom); | |||
534 | setOperationAction(ISD::SELECT , MVT::i16 , Custom); | |||
535 | setOperationAction(ISD::SELECT , MVT::i32 , Custom); | |||
536 | setOperationAction(ISD::SELECT , MVT::f32 , Custom); | |||
537 | setOperationAction(ISD::SELECT , MVT::f64 , Custom); | |||
538 | setOperationAction(ISD::SELECT , MVT::f80 , Custom); | |||
539 | setOperationAction(ISD::SETCC , MVT::i8 , Custom); | |||
540 | setOperationAction(ISD::SETCC , MVT::i16 , Custom); | |||
541 | setOperationAction(ISD::SETCC , MVT::i32 , Custom); | |||
542 | setOperationAction(ISD::SETCC , MVT::f32 , Custom); | |||
543 | setOperationAction(ISD::SETCC , MVT::f64 , Custom); | |||
544 | setOperationAction(ISD::SETCC , MVT::f80 , Custom); | |||
545 | if (Subtarget->is64Bit()) { | |||
546 | setOperationAction(ISD::SELECT , MVT::i64 , Custom); | |||
547 | setOperationAction(ISD::SETCC , MVT::i64 , Custom); | |||
548 | } | |||
549 | setOperationAction(ISD::EH_RETURN , MVT::Other, Custom); | |||
550 | // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support | |||
551 | // SjLj exception handling but a light-weight setjmp/longjmp replacement to | |||
552 | // support continuation, user-level threading, and etc.. As a result, no | |||
553 | // other SjLj exception interfaces are implemented and please don't build | |||
554 | // your own exception handling based on them. | |||
555 | // LLVM/Clang supports zero-cost DWARF exception handling. | |||
556 | setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom); | |||
557 | setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom); | |||
558 | ||||
559 | // Darwin ABI issue. | |||
560 | setOperationAction(ISD::ConstantPool , MVT::i32 , Custom); | |||
561 | setOperationAction(ISD::JumpTable , MVT::i32 , Custom); | |||
562 | setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom); | |||
563 | setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom); | |||
564 | if (Subtarget->is64Bit()) | |||
565 | setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom); | |||
566 | setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom); | |||
567 | setOperationAction(ISD::BlockAddress , MVT::i32 , Custom); | |||
568 | if (Subtarget->is64Bit()) { | |||
569 | setOperationAction(ISD::ConstantPool , MVT::i64 , Custom); | |||
570 | setOperationAction(ISD::JumpTable , MVT::i64 , Custom); | |||
571 | setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom); | |||
572 | setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom); | |||
573 | setOperationAction(ISD::BlockAddress , MVT::i64 , Custom); | |||
574 | } | |||
575 | // 64-bit addm sub, shl, sra, srl (iff 32-bit x86) | |||
576 | setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom); | |||
577 | setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom); | |||
578 | setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom); | |||
579 | if (Subtarget->is64Bit()) { | |||
580 | setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom); | |||
581 | setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom); | |||
582 | setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom); | |||
583 | } | |||
584 | ||||
585 | if (Subtarget->hasSSE1()) | |||
586 | setOperationAction(ISD::PREFETCH , MVT::Other, Legal); | |||
587 | ||||
588 | setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom); | |||
589 | ||||
590 | // Expand certain atomics | |||
591 | for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) { | |||
592 | MVT VT = IntVTs[i]; | |||
593 | setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom); | |||
594 | setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom); | |||
595 | setOperationAction(ISD::ATOMIC_STORE, VT, Custom); | |||
596 | } | |||
597 | ||||
598 | if (Subtarget->hasCmpxchg16b()) { | |||
599 | setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom); | |||
600 | } | |||
601 | ||||
602 | // FIXME - use subtarget debug flags | |||
603 | if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() && | |||
604 | !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) { | |||
605 | setOperationAction(ISD::EH_LABEL, MVT::Other, Expand); | |||
606 | } | |||
607 | ||||
608 | if (Subtarget->is64Bit()) { | |||
609 | setExceptionPointerRegister(X86::RAX); | |||
610 | setExceptionSelectorRegister(X86::RDX); | |||
611 | } else { | |||
612 | setExceptionPointerRegister(X86::EAX); | |||
613 | setExceptionSelectorRegister(X86::EDX); | |||
614 | } | |||
615 | setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom); | |||
616 | setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom); | |||
617 | ||||
618 | setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom); | |||
619 | setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom); | |||
620 | ||||
621 | setOperationAction(ISD::TRAP, MVT::Other, Legal); | |||
622 | setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal); | |||
623 | ||||
624 | // VASTART needs to be custom lowered to use the VarArgsFrameIndex | |||
625 | setOperationAction(ISD::VASTART , MVT::Other, Custom); | |||
626 | setOperationAction(ISD::VAEND , MVT::Other, Expand); | |||
627 | if (Subtarget->is64Bit() && !Subtarget->isTargetWin64()) { | |||
628 | // TargetInfo::X86_64ABIBuiltinVaList | |||
629 | setOperationAction(ISD::VAARG , MVT::Other, Custom); | |||
630 | setOperationAction(ISD::VACOPY , MVT::Other, Custom); | |||
631 | } else { | |||
632 | // TargetInfo::CharPtrBuiltinVaList | |||
633 | setOperationAction(ISD::VAARG , MVT::Other, Expand); | |||
634 | setOperationAction(ISD::VACOPY , MVT::Other, Expand); | |||
635 | } | |||
636 | ||||
637 | setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); | |||
638 | setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); | |||
639 | ||||
640 | setOperationAction(ISD::DYNAMIC_STACKALLOC, getPointerTy(), Custom); | |||
641 | ||||
642 | if (!TM.Options.UseSoftFloat && X86ScalarSSEf64) { | |||
643 | // f32 and f64 use SSE. | |||
644 | // Set up the FP register classes. | |||
645 | addRegisterClass(MVT::f32, &X86::FR32RegClass); | |||
646 | addRegisterClass(MVT::f64, &X86::FR64RegClass); | |||
647 | ||||
648 | // Use ANDPD to simulate FABS. | |||
649 | setOperationAction(ISD::FABS , MVT::f64, Custom); | |||
650 | setOperationAction(ISD::FABS , MVT::f32, Custom); | |||
651 | ||||
652 | // Use XORP to simulate FNEG. | |||
653 | setOperationAction(ISD::FNEG , MVT::f64, Custom); | |||
654 | setOperationAction(ISD::FNEG , MVT::f32, Custom); | |||
655 | ||||
656 | // Use ANDPD and ORPD to simulate FCOPYSIGN. | |||
657 | setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); | |||
658 | setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); | |||
659 | ||||
660 | // Lower this to FGETSIGNx86 plus an AND. | |||
661 | setOperationAction(ISD::FGETSIGN, MVT::i64, Custom); | |||
662 | setOperationAction(ISD::FGETSIGN, MVT::i32, Custom); | |||
663 | ||||
664 | // We don't support sin/cos/fmod | |||
665 | setOperationAction(ISD::FSIN , MVT::f64, Expand); | |||
666 | setOperationAction(ISD::FCOS , MVT::f64, Expand); | |||
667 | setOperationAction(ISD::FSINCOS, MVT::f64, Expand); | |||
668 | setOperationAction(ISD::FSIN , MVT::f32, Expand); | |||
669 | setOperationAction(ISD::FCOS , MVT::f32, Expand); | |||
670 | setOperationAction(ISD::FSINCOS, MVT::f32, Expand); | |||
671 | ||||
672 | // Expand FP immediates into loads from the stack, except for the special | |||
673 | // cases we handle. | |||
674 | addLegalFPImmediate(APFloat(+0.0)); // xorpd | |||
675 | addLegalFPImmediate(APFloat(+0.0f)); // xorps | |||
676 | } else if (!TM.Options.UseSoftFloat && X86ScalarSSEf32) { | |||
677 | // Use SSE for f32, x87 for f64. | |||
678 | // Set up the FP register classes. | |||
679 | addRegisterClass(MVT::f32, &X86::FR32RegClass); | |||
680 | addRegisterClass(MVT::f64, &X86::RFP64RegClass); | |||
681 | ||||
682 | // Use ANDPS to simulate FABS. | |||
683 | setOperationAction(ISD::FABS , MVT::f32, Custom); | |||
684 | ||||
685 | // Use XORP to simulate FNEG. | |||
686 | setOperationAction(ISD::FNEG , MVT::f32, Custom); | |||
687 | ||||
688 | setOperationAction(ISD::UNDEF, MVT::f64, Expand); | |||
689 | ||||
690 | // Use ANDPS and ORPS to simulate FCOPYSIGN. | |||
691 | setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand); | |||
692 | setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); | |||
693 | ||||
694 | // We don't support sin/cos/fmod | |||
695 | setOperationAction(ISD::FSIN , MVT::f32, Expand); | |||
696 | setOperationAction(ISD::FCOS , MVT::f32, Expand); | |||
697 | setOperationAction(ISD::FSINCOS, MVT::f32, Expand); | |||
698 | ||||
699 | // Special cases we handle for FP constants. | |||
700 | addLegalFPImmediate(APFloat(+0.0f)); // xorps | |||
701 | addLegalFPImmediate(APFloat(+0.0)); // FLD0 | |||
702 | addLegalFPImmediate(APFloat(+1.0)); // FLD1 | |||
703 | addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS | |||
704 | addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS | |||
705 | ||||
706 | if (!TM.Options.UnsafeFPMath) { | |||
707 | setOperationAction(ISD::FSIN , MVT::f64, Expand); | |||
708 | setOperationAction(ISD::FCOS , MVT::f64, Expand); | |||
709 | setOperationAction(ISD::FSINCOS, MVT::f64, Expand); | |||
710 | } | |||
711 | } else if (!TM.Options.UseSoftFloat) { | |||
712 | // f32 and f64 in x87. | |||
713 | // Set up the FP register classes. | |||
714 | addRegisterClass(MVT::f64, &X86::RFP64RegClass); | |||
715 | addRegisterClass(MVT::f32, &X86::RFP32RegClass); | |||
716 | ||||
717 | setOperationAction(ISD::UNDEF, MVT::f64, Expand); | |||
718 | setOperationAction(ISD::UNDEF, MVT::f32, Expand); | |||
719 | setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand); | |||
720 | setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand); | |||
721 | ||||
722 | if (!TM.Options.UnsafeFPMath) { | |||
723 | setOperationAction(ISD::FSIN , MVT::f64, Expand); | |||
724 | setOperationAction(ISD::FSIN , MVT::f32, Expand); | |||
725 | setOperationAction(ISD::FCOS , MVT::f64, Expand); | |||
726 | setOperationAction(ISD::FCOS , MVT::f32, Expand); | |||
727 | setOperationAction(ISD::FSINCOS, MVT::f64, Expand); | |||
728 | setOperationAction(ISD::FSINCOS, MVT::f32, Expand); | |||
729 | } | |||
730 | addLegalFPImmediate(APFloat(+0.0)); // FLD0 | |||
731 | addLegalFPImmediate(APFloat(+1.0)); // FLD1 | |||
732 | addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS | |||
733 | addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS | |||
734 | addLegalFPImmediate(APFloat(+0.0f)); // FLD0 | |||
735 | addLegalFPImmediate(APFloat(+1.0f)); // FLD1 | |||
736 | addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS | |||
737 | addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS | |||
738 | } | |||
739 | ||||
740 | // We don't support FMA. | |||
741 | setOperationAction(ISD::FMA, MVT::f64, Expand); | |||
742 | setOperationAction(ISD::FMA, MVT::f32, Expand); | |||
743 | ||||
744 | // Long double always uses X87. | |||
745 | if (!TM.Options.UseSoftFloat) { | |||
746 | addRegisterClass(MVT::f80, &X86::RFP80RegClass); | |||
747 | setOperationAction(ISD::UNDEF, MVT::f80, Expand); | |||
748 | setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand); | |||
749 | { | |||
750 | APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended); | |||
751 | addLegalFPImmediate(TmpFlt); // FLD0 | |||
752 | TmpFlt.changeSign(); | |||
753 | addLegalFPImmediate(TmpFlt); // FLD0/FCHS | |||
754 | ||||
755 | bool ignored; | |||
756 | APFloat TmpFlt2(+1.0); | |||
757 | TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven, | |||
758 | &ignored); | |||
759 | addLegalFPImmediate(TmpFlt2); // FLD1 | |||
760 | TmpFlt2.changeSign(); | |||
761 | addLegalFPImmediate(TmpFlt2); // FLD1/FCHS | |||
762 | } | |||
763 | ||||
764 | if (!TM.Options.UnsafeFPMath) { | |||
765 | setOperationAction(ISD::FSIN , MVT::f80, Expand); | |||
766 | setOperationAction(ISD::FCOS , MVT::f80, Expand); | |||
767 | setOperationAction(ISD::FSINCOS, MVT::f80, Expand); | |||
768 | } | |||
769 | ||||
770 | setOperationAction(ISD::FFLOOR, MVT::f80, Expand); | |||
771 | setOperationAction(ISD::FCEIL, MVT::f80, Expand); | |||
772 | setOperationAction(ISD::FTRUNC, MVT::f80, Expand); | |||
773 | setOperationAction(ISD::FRINT, MVT::f80, Expand); | |||
774 | setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand); | |||
775 | setOperationAction(ISD::FMA, MVT::f80, Expand); | |||
776 | } | |||
777 | ||||
778 | // Always use a library call for pow. | |||
779 | setOperationAction(ISD::FPOW , MVT::f32 , Expand); | |||
780 | setOperationAction(ISD::FPOW , MVT::f64 , Expand); | |||
781 | setOperationAction(ISD::FPOW , MVT::f80 , Expand); | |||
782 | ||||
783 | setOperationAction(ISD::FLOG, MVT::f80, Expand); | |||
784 | setOperationAction(ISD::FLOG2, MVT::f80, Expand); | |||
785 | setOperationAction(ISD::FLOG10, MVT::f80, Expand); | |||
786 | setOperationAction(ISD::FEXP, MVT::f80, Expand); | |||
787 | setOperationAction(ISD::FEXP2, MVT::f80, Expand); | |||
788 | setOperationAction(ISD::FMINNUM, MVT::f80, Expand); | |||
789 | setOperationAction(ISD::FMAXNUM, MVT::f80, Expand); | |||
790 | ||||
791 | // First set operation action for all vector types to either promote | |||
792 | // (for widening) or expand (for scalarization). Then we will selectively | |||
793 | // turn on ones that can be effectively codegen'd. | |||
794 | for (MVT VT : MVT::vector_valuetypes()) { | |||
795 | setOperationAction(ISD::ADD , VT, Expand); | |||
796 | setOperationAction(ISD::SUB , VT, Expand); | |||
797 | setOperationAction(ISD::FADD, VT, Expand); | |||
798 | setOperationAction(ISD::FNEG, VT, Expand); | |||
799 | setOperationAction(ISD::FSUB, VT, Expand); | |||
800 | setOperationAction(ISD::MUL , VT, Expand); | |||
801 | setOperationAction(ISD::FMUL, VT, Expand); | |||
802 | setOperationAction(ISD::SDIV, VT, Expand); | |||
803 | setOperationAction(ISD::UDIV, VT, Expand); | |||
804 | setOperationAction(ISD::FDIV, VT, Expand); | |||
805 | setOperationAction(ISD::SREM, VT, Expand); | |||
806 | setOperationAction(ISD::UREM, VT, Expand); | |||
807 | setOperationAction(ISD::LOAD, VT, Expand); | |||
808 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand); | |||
809 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand); | |||
810 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand); | |||
811 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand); | |||
812 | setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand); | |||
813 | setOperationAction(ISD::FABS, VT, Expand); | |||
814 | setOperationAction(ISD::FSIN, VT, Expand); | |||
815 | setOperationAction(ISD::FSINCOS, VT, Expand); | |||
816 | setOperationAction(ISD::FCOS, VT, Expand); | |||
817 | setOperationAction(ISD::FSINCOS, VT, Expand); | |||
818 | setOperationAction(ISD::FREM, VT, Expand); | |||
819 | setOperationAction(ISD::FMA, VT, Expand); | |||
820 | setOperationAction(ISD::FPOWI, VT, Expand); | |||
821 | setOperationAction(ISD::FSQRT, VT, Expand); | |||
822 | setOperationAction(ISD::FCOPYSIGN, VT, Expand); | |||
823 | setOperationAction(ISD::FFLOOR, VT, Expand); | |||
824 | setOperationAction(ISD::FCEIL, VT, Expand); | |||
825 | setOperationAction(ISD::FTRUNC, VT, Expand); | |||
826 | setOperationAction(ISD::FRINT, VT, Expand); | |||
827 | setOperationAction(ISD::FNEARBYINT, VT, Expand); | |||
828 | setOperationAction(ISD::SMUL_LOHI, VT, Expand); | |||
829 | setOperationAction(ISD::MULHS, VT, Expand); | |||
830 | setOperationAction(ISD::UMUL_LOHI, VT, Expand); | |||
831 | setOperationAction(ISD::MULHU, VT, Expand); | |||
832 | setOperationAction(ISD::SDIVREM, VT, Expand); | |||
833 | setOperationAction(ISD::UDIVREM, VT, Expand); | |||
834 | setOperationAction(ISD::FPOW, VT, Expand); | |||
835 | setOperationAction(ISD::CTPOP, VT, Expand); | |||
836 | setOperationAction(ISD::CTTZ, VT, Expand); | |||
837 | setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); | |||
838 | setOperationAction(ISD::CTLZ, VT, Expand); | |||
839 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); | |||
840 | setOperationAction(ISD::SHL, VT, Expand); | |||
841 | setOperationAction(ISD::SRA, VT, Expand); | |||
842 | setOperationAction(ISD::SRL, VT, Expand); | |||
843 | setOperationAction(ISD::ROTL, VT, Expand); | |||
844 | setOperationAction(ISD::ROTR, VT, Expand); | |||
845 | setOperationAction(ISD::BSWAP, VT, Expand); | |||
846 | setOperationAction(ISD::SETCC, VT, Expand); | |||
847 | setOperationAction(ISD::FLOG, VT, Expand); | |||
848 | setOperationAction(ISD::FLOG2, VT, Expand); | |||
849 | setOperationAction(ISD::FLOG10, VT, Expand); | |||
850 | setOperationAction(ISD::FEXP, VT, Expand); | |||
851 | setOperationAction(ISD::FEXP2, VT, Expand); | |||
852 | setOperationAction(ISD::FP_TO_UINT, VT, Expand); | |||
853 | setOperationAction(ISD::FP_TO_SINT, VT, Expand); | |||
854 | setOperationAction(ISD::UINT_TO_FP, VT, Expand); | |||
855 | setOperationAction(ISD::SINT_TO_FP, VT, Expand); | |||
856 | setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand); | |||
857 | setOperationAction(ISD::TRUNCATE, VT, Expand); | |||
858 | setOperationAction(ISD::SIGN_EXTEND, VT, Expand); | |||
859 | setOperationAction(ISD::ZERO_EXTEND, VT, Expand); | |||
860 | setOperationAction(ISD::ANY_EXTEND, VT, Expand); | |||
861 | setOperationAction(ISD::VSELECT, VT, Expand); | |||
862 | setOperationAction(ISD::SELECT_CC, VT, Expand); | |||
863 | for (MVT InnerVT : MVT::vector_valuetypes()) { | |||
864 | setTruncStoreAction(InnerVT, VT, Expand); | |||
865 | ||||
866 | setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand); | |||
867 | setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand); | |||
868 | ||||
869 | // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like | |||
870 | // types, we have to deal with them whether we ask for Expansion or not. | |||
871 | // Setting Expand causes its own optimisation problems though, so leave | |||
872 | // them legal. | |||
873 | if (VT.getVectorElementType() == MVT::i1) | |||
874 | setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand); | |||
875 | } | |||
876 | } | |||
877 | ||||
878 | // FIXME: In order to prevent SSE instructions being expanded to MMX ones | |||
879 | // with -msoft-float, disable use of MMX as well. | |||
880 | if (!TM.Options.UseSoftFloat && Subtarget->hasMMX()) { | |||
881 | addRegisterClass(MVT::x86mmx, &X86::VR64RegClass); | |||
882 | // No operations on x86mmx supported, everything uses intrinsics. | |||
883 | } | |||
884 | ||||
885 | // MMX-sized vectors (other than x86mmx) are expected to be expanded | |||
886 | // into smaller operations. | |||
887 | setOperationAction(ISD::MULHS, MVT::v8i8, Expand); | |||
888 | setOperationAction(ISD::MULHS, MVT::v4i16, Expand); | |||
889 | setOperationAction(ISD::MULHS, MVT::v2i32, Expand); | |||
890 | setOperationAction(ISD::MULHS, MVT::v1i64, Expand); | |||
891 | setOperationAction(ISD::AND, MVT::v8i8, Expand); | |||
892 | setOperationAction(ISD::AND, MVT::v4i16, Expand); | |||
893 | setOperationAction(ISD::AND, MVT::v2i32, Expand); | |||
894 | setOperationAction(ISD::AND, MVT::v1i64, Expand); | |||
895 | setOperationAction(ISD::OR, MVT::v8i8, Expand); | |||
896 | setOperationAction(ISD::OR, MVT::v4i16, Expand); | |||
897 | setOperationAction(ISD::OR, MVT::v2i32, Expand); | |||
898 | setOperationAction(ISD::OR, MVT::v1i64, Expand); | |||
899 | setOperationAction(ISD::XOR, MVT::v8i8, Expand); | |||
900 | setOperationAction(ISD::XOR, MVT::v4i16, Expand); | |||
901 | setOperationAction(ISD::XOR, MVT::v2i32, Expand); | |||
902 | setOperationAction(ISD::XOR, MVT::v1i64, Expand); | |||
903 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand); | |||
904 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand); | |||
905 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand); | |||
906 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand); | |||
907 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand); | |||
908 | setOperationAction(ISD::SELECT, MVT::v8i8, Expand); | |||
909 | setOperationAction(ISD::SELECT, MVT::v4i16, Expand); | |||
910 | setOperationAction(ISD::SELECT, MVT::v2i32, Expand); | |||
911 | setOperationAction(ISD::SELECT, MVT::v1i64, Expand); | |||
912 | setOperationAction(ISD::BITCAST, MVT::v8i8, Expand); | |||
913 | setOperationAction(ISD::BITCAST, MVT::v4i16, Expand); | |||
914 | setOperationAction(ISD::BITCAST, MVT::v2i32, Expand); | |||
915 | setOperationAction(ISD::BITCAST, MVT::v1i64, Expand); | |||
916 | ||||
917 | if (!TM.Options.UseSoftFloat && Subtarget->hasSSE1()) { | |||
918 | addRegisterClass(MVT::v4f32, &X86::VR128RegClass); | |||
919 | ||||
920 | setOperationAction(ISD::FADD, MVT::v4f32, Legal); | |||
921 | setOperationAction(ISD::FSUB, MVT::v4f32, Legal); | |||
922 | setOperationAction(ISD::FMUL, MVT::v4f32, Legal); | |||
923 | setOperationAction(ISD::FDIV, MVT::v4f32, Legal); | |||
924 | setOperationAction(ISD::FSQRT, MVT::v4f32, Legal); | |||
925 | setOperationAction(ISD::FNEG, MVT::v4f32, Custom); | |||
926 | setOperationAction(ISD::FABS, MVT::v4f32, Custom); | |||
927 | setOperationAction(ISD::LOAD, MVT::v4f32, Legal); | |||
928 | setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom); | |||
929 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom); | |||
930 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom); | |||
931 | setOperationAction(ISD::SELECT, MVT::v4f32, Custom); | |||
932 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom); | |||
933 | } | |||
934 | ||||
935 | if (!TM.Options.UseSoftFloat && Subtarget->hasSSE2()) { | |||
936 | addRegisterClass(MVT::v2f64, &X86::VR128RegClass); | |||
937 | ||||
938 | // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM | |||
939 | // registers cannot be used even for integer operations. | |||
940 | addRegisterClass(MVT::v16i8, &X86::VR128RegClass); | |||
941 | addRegisterClass(MVT::v8i16, &X86::VR128RegClass); | |||
942 | addRegisterClass(MVT::v4i32, &X86::VR128RegClass); | |||
943 | addRegisterClass(MVT::v2i64, &X86::VR128RegClass); | |||
944 | ||||
945 | setOperationAction(ISD::ADD, MVT::v16i8, Legal); | |||
946 | setOperationAction(ISD::ADD, MVT::v8i16, Legal); | |||
947 | setOperationAction(ISD::ADD, MVT::v4i32, Legal); | |||
948 | setOperationAction(ISD::ADD, MVT::v2i64, Legal); | |||
949 | setOperationAction(ISD::MUL, MVT::v4i32, Custom); | |||
950 | setOperationAction(ISD::MUL, MVT::v2i64, Custom); | |||
951 | setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom); | |||
952 | setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom); | |||
953 | setOperationAction(ISD::MULHU, MVT::v8i16, Legal); | |||
954 | setOperationAction(ISD::MULHS, MVT::v8i16, Legal); | |||
955 | setOperationAction(ISD::SUB, MVT::v16i8, Legal); | |||
956 | setOperationAction(ISD::SUB, MVT::v8i16, Legal); | |||
957 | setOperationAction(ISD::SUB, MVT::v4i32, Legal); | |||
958 | setOperationAction(ISD::SUB, MVT::v2i64, Legal); | |||
959 | setOperationAction(ISD::MUL, MVT::v8i16, Legal); | |||
960 | setOperationAction(ISD::FADD, MVT::v2f64, Legal); | |||
961 | setOperationAction(ISD::FSUB, MVT::v2f64, Legal); | |||
962 | setOperationAction(ISD::FMUL, MVT::v2f64, Legal); | |||
963 | setOperationAction(ISD::FDIV, MVT::v2f64, Legal); | |||
964 | setOperationAction(ISD::FSQRT, MVT::v2f64, Legal); | |||
965 | setOperationAction(ISD::FNEG, MVT::v2f64, Custom); | |||
966 | setOperationAction(ISD::FABS, MVT::v2f64, Custom); | |||
967 | ||||
968 | setOperationAction(ISD::SETCC, MVT::v2i64, Custom); | |||
969 | setOperationAction(ISD::SETCC, MVT::v16i8, Custom); | |||
970 | setOperationAction(ISD::SETCC, MVT::v8i16, Custom); | |||
971 | setOperationAction(ISD::SETCC, MVT::v4i32, Custom); | |||
972 | ||||
973 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom); | |||
974 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom); | |||
975 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom); | |||
976 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom); | |||
977 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom); | |||
978 | ||||
979 | // Only provide customized ctpop vector bit twiddling for vector types we | |||
980 | // know to perform better than using the popcnt instructions on each vector | |||
981 | // element. If popcnt isn't supported, always provide the custom version. | |||
982 | if (!Subtarget->hasPOPCNT()) { | |||
983 | setOperationAction(ISD::CTPOP, MVT::v4i32, Custom); | |||
984 | setOperationAction(ISD::CTPOP, MVT::v2i64, Custom); | |||
985 | } | |||
986 | ||||
987 | // Custom lower build_vector, vector_shuffle, and extract_vector_elt. | |||
988 | for (int i = MVT::v16i8; i != MVT::v2i64; ++i) { | |||
989 | MVT VT = (MVT::SimpleValueType)i; | |||
990 | // Do not attempt to custom lower non-power-of-2 vectors | |||
991 | if (!isPowerOf2_32(VT.getVectorNumElements())) | |||
992 | continue; | |||
993 | // Do not attempt to custom lower non-128-bit vectors | |||
994 | if (!VT.is128BitVector()) | |||
995 | continue; | |||
996 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); | |||
997 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); | |||
998 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); | |||
999 | } | |||
1000 | ||||
1001 | // We support custom legalizing of sext and anyext loads for specific | |||
1002 | // memory vector types which we can load as a scalar (or sequence of | |||
1003 | // scalars) and extend in-register to a legal 128-bit vector type. For sext | |||
1004 | // loads these must work with a single scalar load. | |||
1005 | for (MVT VT : MVT::integer_vector_valuetypes()) { | |||
1006 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom); | |||
1007 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom); | |||
1008 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom); | |||
1009 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom); | |||
1010 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom); | |||
1011 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom); | |||
1012 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom); | |||
1013 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom); | |||
1014 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom); | |||
1015 | } | |||
1016 | ||||
1017 | setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom); | |||
1018 | setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom); | |||
1019 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom); | |||
1020 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom); | |||
1021 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom); | |||
1022 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom); | |||
1023 | ||||
1024 | if (Subtarget->is64Bit()) { | |||
1025 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom); | |||
1026 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom); | |||
1027 | } | |||
1028 | ||||
1029 | // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64. | |||
1030 | for (int i = MVT::v16i8; i != MVT::v2i64; ++i) { | |||
1031 | MVT VT = (MVT::SimpleValueType)i; | |||
1032 | ||||
1033 | // Do not attempt to promote non-128-bit vectors | |||
1034 | if (!VT.is128BitVector()) | |||
1035 | continue; | |||
1036 | ||||
1037 | setOperationAction(ISD::AND, VT, Promote); | |||
1038 | AddPromotedToType (ISD::AND, VT, MVT::v2i64); | |||
1039 | setOperationAction(ISD::OR, VT, Promote); | |||
1040 | AddPromotedToType (ISD::OR, VT, MVT::v2i64); | |||
1041 | setOperationAction(ISD::XOR, VT, Promote); | |||
1042 | AddPromotedToType (ISD::XOR, VT, MVT::v2i64); | |||
1043 | setOperationAction(ISD::LOAD, VT, Promote); | |||
1044 | AddPromotedToType (ISD::LOAD, VT, MVT::v2i64); | |||
1045 | setOperationAction(ISD::SELECT, VT, Promote); | |||
1046 | AddPromotedToType (ISD::SELECT, VT, MVT::v2i64); | |||
1047 | } | |||
1048 | ||||
1049 | // Custom lower v2i64 and v2f64 selects. | |||
1050 | setOperationAction(ISD::LOAD, MVT::v2f64, Legal); | |||
1051 | setOperationAction(ISD::LOAD, MVT::v2i64, Legal); | |||
1052 | setOperationAction(ISD::SELECT, MVT::v2f64, Custom); | |||
1053 | setOperationAction(ISD::SELECT, MVT::v2i64, Custom); | |||
1054 | ||||
1055 | setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal); | |||
1056 | setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal); | |||
1057 | ||||
1058 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom); | |||
1059 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom); | |||
1060 | // As there is no 64-bit GPR available, we need build a special custom | |||
1061 | // sequence to convert from v2i32 to v2f32. | |||
1062 | if (!Subtarget->is64Bit()) | |||
1063 | setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom); | |||
1064 | ||||
1065 | setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom); | |||
1066 | setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom); | |||
1067 | ||||
1068 | for (MVT VT : MVT::fp_vector_valuetypes()) | |||
1069 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal); | |||
1070 | ||||
1071 | setOperationAction(ISD::BITCAST, MVT::v2i32, Custom); | |||
1072 | setOperationAction(ISD::BITCAST, MVT::v4i16, Custom); | |||
1073 | setOperationAction(ISD::BITCAST, MVT::v8i8, Custom); | |||
1074 | } | |||
1075 | ||||
1076 | if (!TM.Options.UseSoftFloat && Subtarget->hasSSE41()) { | |||
1077 | setOperationAction(ISD::FFLOOR, MVT::f32, Legal); | |||
1078 | setOperationAction(ISD::FCEIL, MVT::f32, Legal); | |||
1079 | setOperationAction(ISD::FTRUNC, MVT::f32, Legal); | |||
1080 | setOperationAction(ISD::FRINT, MVT::f32, Legal); | |||
1081 | setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal); | |||
1082 | setOperationAction(ISD::FFLOOR, MVT::f64, Legal); | |||
1083 | setOperationAction(ISD::FCEIL, MVT::f64, Legal); | |||
1084 | setOperationAction(ISD::FTRUNC, MVT::f64, Legal); | |||
1085 | setOperationAction(ISD::FRINT, MVT::f64, Legal); | |||
1086 | setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal); | |||
1087 | ||||
1088 | setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal); | |||
1089 | setOperationAction(ISD::FCEIL, MVT::v4f32, Legal); | |||
1090 | setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal); | |||
1091 | setOperationAction(ISD::FRINT, MVT::v4f32, Legal); | |||
1092 | setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal); | |||
1093 | setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal); | |||
1094 | setOperationAction(ISD::FCEIL, MVT::v2f64, Legal); | |||
1095 | setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal); | |||
1096 | setOperationAction(ISD::FRINT, MVT::v2f64, Legal); | |||
1097 | setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal); | |||
1098 | ||||
1099 | // FIXME: Do we need to handle scalar-to-vector here? | |||
1100 | setOperationAction(ISD::MUL, MVT::v4i32, Legal); | |||
1101 | ||||
1102 | setOperationAction(ISD::VSELECT, MVT::v2f64, Custom); | |||
1103 | setOperationAction(ISD::VSELECT, MVT::v2i64, Custom); | |||
1104 | setOperationAction(ISD::VSELECT, MVT::v4i32, Custom); | |||
1105 | setOperationAction(ISD::VSELECT, MVT::v4f32, Custom); | |||
1106 | setOperationAction(ISD::VSELECT, MVT::v8i16, Custom); | |||
1107 | // There is no BLENDI for byte vectors. We don't need to custom lower | |||
1108 | // some vselects for now. | |||
1109 | setOperationAction(ISD::VSELECT, MVT::v16i8, Legal); | |||
1110 | ||||
1111 | // SSE41 brings specific instructions for doing vector sign extend even in | |||
1112 | // cases where we don't have SRA. | |||
1113 | for (MVT VT : MVT::integer_vector_valuetypes()) { | |||
1114 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom); | |||
1115 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom); | |||
1116 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom); | |||
1117 | } | |||
1118 | ||||
1119 | // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X | |||
1120 | setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal); | |||
1121 | setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal); | |||
1122 | setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal); | |||
1123 | setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal); | |||
1124 | setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal); | |||
1125 | setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal); | |||
1126 | ||||
1127 | setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal); | |||
1128 | setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal); | |||
1129 | setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal); | |||
1130 | setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal); | |||
1131 | setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal); | |||
1132 | setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal); | |||
1133 | ||||
1134 | // i8 and i16 vectors are custom because the source register and source | |||
1135 | // source memory operand types are not the same width. f32 vectors are | |||
1136 | // custom since the immediate controlling the insert encodes additional | |||
1137 | // information. | |||
1138 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom); | |||
1139 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom); | |||
1140 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom); | |||
1141 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom); | |||
1142 | ||||
1143 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom); | |||
1144 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom); | |||
1145 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom); | |||
1146 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom); | |||
1147 | ||||
1148 | // FIXME: these should be Legal, but that's only for the case where | |||
1149 | // the index is constant. For now custom expand to deal with that. | |||
1150 | if (Subtarget->is64Bit()) { | |||
1151 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom); | |||
1152 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom); | |||
1153 | } | |||
1154 | } | |||
1155 | ||||
1156 | if (Subtarget->hasSSE2()) { | |||
1157 | setOperationAction(ISD::SRL, MVT::v8i16, Custom); | |||
1158 | setOperationAction(ISD::SRL, MVT::v16i8, Custom); | |||
1159 | ||||
1160 | setOperationAction(ISD::SHL, MVT::v8i16, Custom); | |||
1161 | setOperationAction(ISD::SHL, MVT::v16i8, Custom); | |||
1162 | ||||
1163 | setOperationAction(ISD::SRA, MVT::v8i16, Custom); | |||
1164 | setOperationAction(ISD::SRA, MVT::v16i8, Custom); | |||
1165 | ||||
1166 | // In the customized shift lowering, the legal cases in AVX2 will be | |||
1167 | // recognized. | |||
1168 | setOperationAction(ISD::SRL, MVT::v2i64, Custom); | |||
1169 | setOperationAction(ISD::SRL, MVT::v4i32, Custom); | |||
1170 | ||||
1171 | setOperationAction(ISD::SHL, MVT::v2i64, Custom); | |||
1172 | setOperationAction(ISD::SHL, MVT::v4i32, Custom); | |||
1173 | ||||
1174 | setOperationAction(ISD::SRA, MVT::v4i32, Custom); | |||
1175 | } | |||
1176 | ||||
1177 | if (!TM.Options.UseSoftFloat && Subtarget->hasFp256()) { | |||
1178 | addRegisterClass(MVT::v32i8, &X86::VR256RegClass); | |||
1179 | addRegisterClass(MVT::v16i16, &X86::VR256RegClass); | |||
1180 | addRegisterClass(MVT::v8i32, &X86::VR256RegClass); | |||
1181 | addRegisterClass(MVT::v8f32, &X86::VR256RegClass); | |||
1182 | addRegisterClass(MVT::v4i64, &X86::VR256RegClass); | |||
1183 | addRegisterClass(MVT::v4f64, &X86::VR256RegClass); | |||
1184 | ||||
1185 | setOperationAction(ISD::LOAD, MVT::v8f32, Legal); | |||
1186 | setOperationAction(ISD::LOAD, MVT::v4f64, Legal); | |||
1187 | setOperationAction(ISD::LOAD, MVT::v4i64, Legal); | |||
1188 | ||||
1189 | setOperationAction(ISD::FADD, MVT::v8f32, Legal); | |||
1190 | setOperationAction(ISD::FSUB, MVT::v8f32, Legal); | |||
1191 | setOperationAction(ISD::FMUL, MVT::v8f32, Legal); | |||
1192 | setOperationAction(ISD::FDIV, MVT::v8f32, Legal); | |||
1193 | setOperationAction(ISD::FSQRT, MVT::v8f32, Legal); | |||
1194 | setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal); | |||
1195 | setOperationAction(ISD::FCEIL, MVT::v8f32, Legal); | |||
1196 | setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal); | |||
1197 | setOperationAction(ISD::FRINT, MVT::v8f32, Legal); | |||
1198 | setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal); | |||
1199 | setOperationAction(ISD::FNEG, MVT::v8f32, Custom); | |||
1200 | setOperationAction(ISD::FABS, MVT::v8f32, Custom); | |||
1201 | ||||
1202 | setOperationAction(ISD::FADD, MVT::v4f64, Legal); | |||
1203 | setOperationAction(ISD::FSUB, MVT::v4f64, Legal); | |||
1204 | setOperationAction(ISD::FMUL, MVT::v4f64, Legal); | |||
1205 | setOperationAction(ISD::FDIV, MVT::v4f64, Legal); | |||
1206 | setOperationAction(ISD::FSQRT, MVT::v4f64, Legal); | |||
1207 | setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal); | |||
1208 | setOperationAction(ISD::FCEIL, MVT::v4f64, Legal); | |||
1209 | setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal); | |||
1210 | setOperationAction(ISD::FRINT, MVT::v4f64, Legal); | |||
1211 | setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal); | |||
1212 | setOperationAction(ISD::FNEG, MVT::v4f64, Custom); | |||
1213 | setOperationAction(ISD::FABS, MVT::v4f64, Custom); | |||
1214 | ||||
1215 | // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted | |||
1216 | // even though v8i16 is a legal type. | |||
1217 | setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote); | |||
1218 | setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote); | |||
1219 | setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal); | |||
1220 | ||||
1221 | setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote); | |||
1222 | setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal); | |||
1223 | setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal); | |||
1224 | ||||
1225 | setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom); | |||
1226 | setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom); | |||
1227 | ||||
1228 | for (MVT VT : MVT::fp_vector_valuetypes()) | |||
1229 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal); | |||
1230 | ||||
1231 | setOperationAction(ISD::SRL, MVT::v16i16, Custom); | |||
1232 | setOperationAction(ISD::SRL, MVT::v32i8, Custom); | |||
1233 | ||||
1234 | setOperationAction(ISD::SHL, MVT::v16i16, Custom); | |||
1235 | setOperationAction(ISD::SHL, MVT::v32i8, Custom); | |||
1236 | ||||
1237 | setOperationAction(ISD::SRA, MVT::v16i16, Custom); | |||
1238 | setOperationAction(ISD::SRA, MVT::v32i8, Custom); | |||
1239 | ||||
1240 | setOperationAction(ISD::SETCC, MVT::v32i8, Custom); | |||
1241 | setOperationAction(ISD::SETCC, MVT::v16i16, Custom); | |||
1242 | setOperationAction(ISD::SETCC, MVT::v8i32, Custom); | |||
1243 | setOperationAction(ISD::SETCC, MVT::v4i64, Custom); | |||
1244 | ||||
1245 | setOperationAction(ISD::SELECT, MVT::v4f64, Custom); | |||
1246 | setOperationAction(ISD::SELECT, MVT::v4i64, Custom); | |||
1247 | setOperationAction(ISD::SELECT, MVT::v8f32, Custom); | |||
1248 | ||||
1249 | setOperationAction(ISD::VSELECT, MVT::v4f64, Custom); | |||
1250 | setOperationAction(ISD::VSELECT, MVT::v4i64, Custom); | |||
1251 | setOperationAction(ISD::VSELECT, MVT::v8i32, Custom); | |||
1252 | setOperationAction(ISD::VSELECT, MVT::v8f32, Custom); | |||
1253 | ||||
1254 | setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom); | |||
1255 | setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom); | |||
1256 | setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom); | |||
1257 | setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom); | |||
1258 | setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom); | |||
1259 | setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom); | |||
1260 | setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom); | |||
1261 | setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom); | |||
1262 | setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom); | |||
1263 | setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom); | |||
1264 | setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom); | |||
1265 | setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom); | |||
1266 | ||||
1267 | if (Subtarget->hasFMA() || Subtarget->hasFMA4()) { | |||
1268 | setOperationAction(ISD::FMA, MVT::v8f32, Legal); | |||
1269 | setOperationAction(ISD::FMA, MVT::v4f64, Legal); | |||
1270 | setOperationAction(ISD::FMA, MVT::v4f32, Legal); | |||
1271 | setOperationAction(ISD::FMA, MVT::v2f64, Legal); | |||
1272 | setOperationAction(ISD::FMA, MVT::f32, Legal); | |||
1273 | setOperationAction(ISD::FMA, MVT::f64, Legal); | |||
1274 | } | |||
1275 | ||||
1276 | if (Subtarget->hasInt256()) { | |||
1277 | setOperationAction(ISD::ADD, MVT::v4i64, Legal); | |||
1278 | setOperationAction(ISD::ADD, MVT::v8i32, Legal); | |||
1279 | setOperationAction(ISD::ADD, MVT::v16i16, Legal); | |||
1280 | setOperationAction(ISD::ADD, MVT::v32i8, Legal); | |||
1281 | ||||
1282 | setOperationAction(ISD::SUB, MVT::v4i64, Legal); | |||
1283 | setOperationAction(ISD::SUB, MVT::v8i32, Legal); | |||
1284 | setOperationAction(ISD::SUB, MVT::v16i16, Legal); | |||
1285 | setOperationAction(ISD::SUB, MVT::v32i8, Legal); | |||
1286 | ||||
1287 | setOperationAction(ISD::MUL, MVT::v4i64, Custom); | |||
1288 | setOperationAction(ISD::MUL, MVT::v8i32, Legal); | |||
1289 | setOperationAction(ISD::MUL, MVT::v16i16, Legal); | |||
1290 | // Don't lower v32i8 because there is no 128-bit byte mul | |||
1291 | ||||
1292 | setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom); | |||
1293 | setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom); | |||
1294 | setOperationAction(ISD::MULHU, MVT::v16i16, Legal); | |||
1295 | setOperationAction(ISD::MULHS, MVT::v16i16, Legal); | |||
1296 | ||||
1297 | setOperationAction(ISD::VSELECT, MVT::v16i16, Custom); | |||
1298 | setOperationAction(ISD::VSELECT, MVT::v32i8, Legal); | |||
1299 | ||||
1300 | // The custom lowering for UINT_TO_FP for v8i32 becomes interesting | |||
1301 | // when we have a 256bit-wide blend with immediate. | |||
1302 | setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom); | |||
1303 | ||||
1304 | // Only provide customized ctpop vector bit twiddling for vector types we | |||
1305 | // know to perform better than using the popcnt instructions on each | |||
1306 | // vector element. If popcnt isn't supported, always provide the custom | |||
1307 | // version. | |||
1308 | if (!Subtarget->hasPOPCNT()) | |||
1309 | setOperationAction(ISD::CTPOP, MVT::v4i64, Custom); | |||
1310 | ||||
1311 | // Custom CTPOP always performs better on natively supported v8i32 | |||
1312 | setOperationAction(ISD::CTPOP, MVT::v8i32, Custom); | |||
1313 | ||||
1314 | // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X | |||
1315 | setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal); | |||
1316 | setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal); | |||
1317 | setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal); | |||
1318 | setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal); | |||
1319 | setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal); | |||
1320 | setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal); | |||
1321 | ||||
1322 | setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal); | |||
1323 | setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal); | |||
1324 | setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal); | |||
1325 | setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal); | |||
1326 | setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal); | |||
1327 | setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal); | |||
1328 | } else { | |||
1329 | setOperationAction(ISD::ADD, MVT::v4i64, Custom); | |||
1330 | setOperationAction(ISD::ADD, MVT::v8i32, Custom); | |||
1331 | setOperationAction(ISD::ADD, MVT::v16i16, Custom); | |||
1332 | setOperationAction(ISD::ADD, MVT::v32i8, Custom); | |||
1333 | ||||
1334 | setOperationAction(ISD::SUB, MVT::v4i64, Custom); | |||
1335 | setOperationAction(ISD::SUB, MVT::v8i32, Custom); | |||
1336 | setOperationAction(ISD::SUB, MVT::v16i16, Custom); | |||
1337 | setOperationAction(ISD::SUB, MVT::v32i8, Custom); | |||
1338 | ||||
1339 | setOperationAction(ISD::MUL, MVT::v4i64, Custom); | |||
1340 | setOperationAction(ISD::MUL, MVT::v8i32, Custom); | |||
1341 | setOperationAction(ISD::MUL, MVT::v16i16, Custom); | |||
1342 | // Don't lower v32i8 because there is no 128-bit byte mul | |||
1343 | } | |||
1344 | ||||
1345 | // In the customized shift lowering, the legal cases in AVX2 will be | |||
1346 | // recognized. | |||
1347 | setOperationAction(ISD::SRL, MVT::v4i64, Custom); | |||
1348 | setOperationAction(ISD::SRL, MVT::v8i32, Custom); | |||
1349 | ||||
1350 | setOperationAction(ISD::SHL, MVT::v4i64, Custom); | |||
1351 | setOperationAction(ISD::SHL, MVT::v8i32, Custom); | |||
1352 | ||||
1353 | setOperationAction(ISD::SRA, MVT::v8i32, Custom); | |||
1354 | ||||
1355 | // Custom lower several nodes for 256-bit types. | |||
1356 | for (MVT VT : MVT::vector_valuetypes()) { | |||
1357 | if (VT.getScalarSizeInBits() >= 32) { | |||
1358 | setOperationAction(ISD::MLOAD, VT, Legal); | |||
1359 | setOperationAction(ISD::MSTORE, VT, Legal); | |||
1360 | } | |||
1361 | // Extract subvector is special because the value type | |||
1362 | // (result) is 128-bit but the source is 256-bit wide. | |||
1363 | if (VT.is128BitVector()) { | |||
1364 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom); | |||
1365 | } | |||
1366 | // Do not attempt to custom lower other non-256-bit vectors | |||
1367 | if (!VT.is256BitVector()) | |||
1368 | continue; | |||
1369 | ||||
1370 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); | |||
1371 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); | |||
1372 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); | |||
1373 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); | |||
1374 | setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom); | |||
1375 | setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom); | |||
1376 | setOperationAction(ISD::CONCAT_VECTORS, VT, Custom); | |||
1377 | } | |||
1378 | ||||
1379 | // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64. | |||
1380 | for (int i = MVT::v32i8; i != MVT::v4i64; ++i) { | |||
1381 | MVT VT = (MVT::SimpleValueType)i; | |||
1382 | ||||
1383 | // Do not attempt to promote non-256-bit vectors | |||
1384 | if (!VT.is256BitVector()) | |||
1385 | continue; | |||
1386 | ||||
1387 | setOperationAction(ISD::AND, VT, Promote); | |||
1388 | AddPromotedToType (ISD::AND, VT, MVT::v4i64); | |||
1389 | setOperationAction(ISD::OR, VT, Promote); | |||
1390 | AddPromotedToType (ISD::OR, VT, MVT::v4i64); | |||
1391 | setOperationAction(ISD::XOR, VT, Promote); | |||
1392 | AddPromotedToType (ISD::XOR, VT, MVT::v4i64); | |||
1393 | setOperationAction(ISD::LOAD, VT, Promote); | |||
1394 | AddPromotedToType (ISD::LOAD, VT, MVT::v4i64); | |||
1395 | setOperationAction(ISD::SELECT, VT, Promote); | |||
1396 | AddPromotedToType (ISD::SELECT, VT, MVT::v4i64); | |||
1397 | } | |||
1398 | } | |||
1399 | ||||
1400 | if (!TM.Options.UseSoftFloat && Subtarget->hasAVX512()) { | |||
1401 | addRegisterClass(MVT::v16i32, &X86::VR512RegClass); | |||
1402 | addRegisterClass(MVT::v16f32, &X86::VR512RegClass); | |||
1403 | addRegisterClass(MVT::v8i64, &X86::VR512RegClass); | |||
1404 | addRegisterClass(MVT::v8f64, &X86::VR512RegClass); | |||
1405 | ||||
1406 | addRegisterClass(MVT::i1, &X86::VK1RegClass); | |||
1407 | addRegisterClass(MVT::v8i1, &X86::VK8RegClass); | |||
1408 | addRegisterClass(MVT::v16i1, &X86::VK16RegClass); | |||
1409 | ||||
1410 | for (MVT VT : MVT::fp_vector_valuetypes()) | |||
1411 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal); | |||
1412 | ||||
1413 | setOperationAction(ISD::BR_CC, MVT::i1, Expand); | |||
1414 | setOperationAction(ISD::SETCC, MVT::i1, Custom); | |||
1415 | setOperationAction(ISD::XOR, MVT::i1, Legal); | |||
1416 | setOperationAction(ISD::OR, MVT::i1, Legal); | |||
1417 | setOperationAction(ISD::AND, MVT::i1, Legal); | |||
1418 | setOperationAction(ISD::LOAD, MVT::v16f32, Legal); | |||
1419 | setOperationAction(ISD::LOAD, MVT::v8f64, Legal); | |||
1420 | setOperationAction(ISD::LOAD, MVT::v8i64, Legal); | |||
1421 | setOperationAction(ISD::LOAD, MVT::v16i32, Legal); | |||
1422 | setOperationAction(ISD::LOAD, MVT::v16i1, Legal); | |||
1423 | ||||
1424 | setOperationAction(ISD::FADD, MVT::v16f32, Legal); | |||
1425 | setOperationAction(ISD::FSUB, MVT::v16f32, Legal); | |||
1426 | setOperationAction(ISD::FMUL, MVT::v16f32, Legal); | |||
1427 | setOperationAction(ISD::FDIV, MVT::v16f32, Legal); | |||
1428 | setOperationAction(ISD::FSQRT, MVT::v16f32, Legal); | |||
1429 | setOperationAction(ISD::FNEG, MVT::v16f32, Custom); | |||
1430 | ||||
1431 | setOperationAction(ISD::FADD, MVT::v8f64, Legal); | |||
1432 | setOperationAction(ISD::FSUB, MVT::v8f64, Legal); | |||
1433 | setOperationAction(ISD::FMUL, MVT::v8f64, Legal); | |||
1434 | setOperationAction(ISD::FDIV, MVT::v8f64, Legal); | |||
1435 | setOperationAction(ISD::FSQRT, MVT::v8f64, Legal); | |||
1436 | setOperationAction(ISD::FNEG, MVT::v8f64, Custom); | |||
1437 | setOperationAction(ISD::FMA, MVT::v8f64, Legal); | |||
1438 | setOperationAction(ISD::FMA, MVT::v16f32, Legal); | |||
1439 | ||||
1440 | setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal); | |||
1441 | setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal); | |||
1442 | setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal); | |||
1443 | setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal); | |||
1444 | if (Subtarget->is64Bit()) { | |||
1445 | setOperationAction(ISD::FP_TO_UINT, MVT::i64, Legal); | |||
1446 | setOperationAction(ISD::FP_TO_SINT, MVT::i64, Legal); | |||
1447 | setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal); | |||
1448 | setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal); | |||
1449 | } | |||
1450 | setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal); | |||
1451 | setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal); | |||
1452 | setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal); | |||
1453 | setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal); | |||
1454 | setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal); | |||
1455 | setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom); | |||
1456 | setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom); | |||
1457 | setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote); | |||
1458 | setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote); | |||
1459 | setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal); | |||
1460 | setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal); | |||
1461 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal); | |||
1462 | setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal); | |||
1463 | setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal); | |||
1464 | ||||
1465 | setOperationAction(ISD::TRUNCATE, MVT::i1, Custom); | |||
1466 | setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom); | |||
1467 | setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom); | |||
1468 | setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom); | |||
1469 | setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom); | |||
1470 | setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom); | |||
1471 | setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom); | |||
1472 | setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom); | |||
1473 | setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom); | |||
1474 | setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom); | |||
1475 | setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom); | |||
1476 | setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom); | |||
1477 | setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom); | |||
1478 | ||||
1479 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom); | |||
1480 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom); | |||
1481 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom); | |||
1482 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom); | |||
1483 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom); | |||
1484 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal); | |||
1485 | ||||
1486 | setOperationAction(ISD::SETCC, MVT::v16i1, Custom); | |||
1487 | setOperationAction(ISD::SETCC, MVT::v8i1, Custom); | |||
1488 | ||||
1489 | setOperationAction(ISD::MUL, MVT::v8i64, Custom); | |||
1490 | ||||
1491 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom); | |||
1492 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom); | |||
1493 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom); | |||
1494 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom); | |||
1495 | setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom); | |||
1496 | setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom); | |||
1497 | setOperationAction(ISD::SELECT, MVT::v8f64, Custom); | |||
1498 | setOperationAction(ISD::SELECT, MVT::v8i64, Custom); | |||
1499 | setOperationAction(ISD::SELECT, MVT::v16f32, Custom); | |||
1500 | ||||
1501 | setOperationAction(ISD::ADD, MVT::v8i64, Legal); | |||
1502 | setOperationAction(ISD::ADD, MVT::v16i32, Legal); | |||
1503 | ||||
1504 | setOperationAction(ISD::SUB, MVT::v8i64, Legal); | |||
1505 | setOperationAction(ISD::SUB, MVT::v16i32, Legal); | |||
1506 | ||||
1507 | setOperationAction(ISD::MUL, MVT::v16i32, Legal); | |||
1508 | ||||
1509 | setOperationAction(ISD::SRL, MVT::v8i64, Custom); | |||
1510 | setOperationAction(ISD::SRL, MVT::v16i32, Custom); | |||
1511 | ||||
1512 | setOperationAction(ISD::SHL, MVT::v8i64, Custom); | |||
1513 | setOperationAction(ISD::SHL, MVT::v16i32, Custom); | |||
1514 | ||||
1515 | setOperationAction(ISD::SRA, MVT::v8i64, Custom); | |||
1516 | setOperationAction(ISD::SRA, MVT::v16i32, Custom); | |||
1517 | ||||
1518 | setOperationAction(ISD::AND, MVT::v8i64, Legal); | |||
1519 | setOperationAction(ISD::OR, MVT::v8i64, Legal); | |||
1520 | setOperationAction(ISD::XOR, MVT::v8i64, Legal); | |||
1521 | setOperationAction(ISD::AND, MVT::v16i32, Legal); | |||
1522 | setOperationAction(ISD::OR, MVT::v16i32, Legal); | |||
1523 | setOperationAction(ISD::XOR, MVT::v16i32, Legal); | |||
1524 | ||||
1525 | if (Subtarget->hasCDI()) { | |||
1526 | setOperationAction(ISD::CTLZ, MVT::v8i64, Legal); | |||
1527 | setOperationAction(ISD::CTLZ, MVT::v16i32, Legal); | |||
1528 | } | |||
1529 | ||||
1530 | // Custom lower several nodes. | |||
1531 | for (MVT VT : MVT::vector_valuetypes()) { | |||
1532 | unsigned EltSize = VT.getVectorElementType().getSizeInBits(); | |||
1533 | // Extract subvector is special because the value type | |||
1534 | // (result) is 256/128-bit but the source is 512-bit wide. | |||
1535 | if (VT.is128BitVector() || VT.is256BitVector()) { | |||
1536 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom); | |||
1537 | } | |||
1538 | if (VT.getVectorElementType() == MVT::i1) | |||
1539 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal); | |||
1540 | ||||
1541 | // Do not attempt to custom lower other non-512-bit vectors | |||
1542 | if (!VT.is512BitVector()) | |||
1543 | continue; | |||
1544 | ||||
1545 | if ( EltSize >= 32) { | |||
1546 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); | |||
1547 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); | |||
1548 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); | |||
1549 | setOperationAction(ISD::VSELECT, VT, Legal); | |||
1550 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); | |||
1551 | setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom); | |||
1552 | setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom); | |||
1553 | setOperationAction(ISD::MLOAD, VT, Legal); | |||
1554 | setOperationAction(ISD::MSTORE, VT, Legal); | |||
1555 | } | |||
1556 | } | |||
1557 | for (int i = MVT::v32i8; i != MVT::v8i64; ++i) { | |||
1558 | MVT VT = (MVT::SimpleValueType)i; | |||
1559 | ||||
1560 | // Do not attempt to promote non-512-bit vectors. | |||
1561 | if (!VT.is512BitVector()) | |||
1562 | continue; | |||
1563 | ||||
1564 | setOperationAction(ISD::SELECT, VT, Promote); | |||
1565 | AddPromotedToType (ISD::SELECT, VT, MVT::v8i64); | |||
1566 | } | |||
1567 | }// has AVX-512 | |||
1568 | ||||
1569 | if (!TM.Options.UseSoftFloat && Subtarget->hasBWI()) { | |||
1570 | addRegisterClass(MVT::v32i16, &X86::VR512RegClass); | |||
1571 | addRegisterClass(MVT::v64i8, &X86::VR512RegClass); | |||
1572 | ||||
1573 | addRegisterClass(MVT::v32i1, &X86::VK32RegClass); | |||
1574 | addRegisterClass(MVT::v64i1, &X86::VK64RegClass); | |||
1575 | ||||
1576 | setOperationAction(ISD::LOAD, MVT::v32i16, Legal); | |||
1577 | setOperationAction(ISD::LOAD, MVT::v64i8, Legal); | |||
1578 | setOperationAction(ISD::SETCC, MVT::v32i1, Custom); | |||
1579 | setOperationAction(ISD::SETCC, MVT::v64i1, Custom); | |||
1580 | setOperationAction(ISD::ADD, MVT::v32i16, Legal); | |||
1581 | setOperationAction(ISD::ADD, MVT::v64i8, Legal); | |||
1582 | setOperationAction(ISD::SUB, MVT::v32i16, Legal); | |||
1583 | setOperationAction(ISD::SUB, MVT::v64i8, Legal); | |||
1584 | setOperationAction(ISD::MUL, MVT::v32i16, Legal); | |||
1585 | ||||
1586 | for (int i = MVT::v32i8; i != MVT::v8i64; ++i) { | |||
1587 | const MVT VT = (MVT::SimpleValueType)i; | |||
1588 | ||||
1589 | const unsigned EltSize = VT.getVectorElementType().getSizeInBits(); | |||
1590 | ||||
1591 | // Do not attempt to promote non-512-bit vectors. | |||
1592 | if (!VT.is512BitVector()) | |||
1593 | continue; | |||
1594 | ||||
1595 | if (EltSize < 32) { | |||
1596 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); | |||
1597 | setOperationAction(ISD::VSELECT, VT, Legal); | |||
1598 | } | |||
1599 | } | |||
1600 | } | |||
1601 | ||||
1602 | if (!TM.Options.UseSoftFloat && Subtarget->hasVLX()) { | |||
1603 | addRegisterClass(MVT::v4i1, &X86::VK4RegClass); | |||
1604 | addRegisterClass(MVT::v2i1, &X86::VK2RegClass); | |||
1605 | ||||
1606 | setOperationAction(ISD::SETCC, MVT::v4i1, Custom); | |||
1607 | setOperationAction(ISD::SETCC, MVT::v2i1, Custom); | |||
1608 | setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Legal); | |||
1609 | ||||
1610 | setOperationAction(ISD::AND, MVT::v8i32, Legal); | |||
1611 | setOperationAction(ISD::OR, MVT::v8i32, Legal); | |||
1612 | setOperationAction(ISD::XOR, MVT::v8i32, Legal); | |||
1613 | setOperationAction(ISD::AND, MVT::v4i32, Legal); | |||
1614 | setOperationAction(ISD::OR, MVT::v4i32, Legal); | |||
1615 | setOperationAction(ISD::XOR, MVT::v4i32, Legal); | |||
1616 | } | |||
1617 | ||||
1618 | // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion | |||
1619 | // of this type with custom code. | |||
1620 | for (MVT VT : MVT::vector_valuetypes()) | |||
1621 | setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Custom); | |||
1622 | ||||
1623 | // We want to custom lower some of our intrinsics. | |||
1624 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); | |||
1625 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom); | |||
1626 | setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom); | |||
1627 | if (!Subtarget->is64Bit()) | |||
1628 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom); | |||
1629 | ||||
1630 | // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't | |||
1631 | // handle type legalization for these operations here. | |||
1632 | // | |||
1633 | // FIXME: We really should do custom legalization for addition and | |||
1634 | // subtraction on x86-32 once PR3203 is fixed. We really can't do much better | |||
1635 | // than generic legalization for 64-bit multiplication-with-overflow, though. | |||
1636 | for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) { | |||
1637 | // Add/Sub/Mul with overflow operations are custom lowered. | |||
1638 | MVT VT = IntVTs[i]; | |||
1639 | setOperationAction(ISD::SADDO, VT, Custom); | |||
1640 | setOperationAction(ISD::UADDO, VT, Custom); | |||
1641 | setOperationAction(ISD::SSUBO, VT, Custom); | |||
1642 | setOperationAction(ISD::USUBO, VT, Custom); | |||
1643 | setOperationAction(ISD::SMULO, VT, Custom); | |||
1644 | setOperationAction(ISD::UMULO, VT, Custom); | |||
1645 | } | |||
1646 | ||||
1647 | ||||
1648 | if (!Subtarget->is64Bit()) { | |||
1649 | // These libcalls are not available in 32-bit. | |||
1650 | setLibcallName(RTLIB::SHL_I128, nullptr); | |||
1651 | setLibcallName(RTLIB::SRL_I128, nullptr); | |||
1652 | setLibcallName(RTLIB::SRA_I128, nullptr); | |||
1653 | } | |||
1654 | ||||
1655 | // Combine sin / cos into one node or libcall if possible. | |||
1656 | if (Subtarget->hasSinCos()) { | |||
1657 | setLibcallName(RTLIB::SINCOS_F32, "sincosf"); | |||
1658 | setLibcallName(RTLIB::SINCOS_F64, "sincos"); | |||
1659 | if (Subtarget->isTargetDarwin()) { | |||
1660 | // For MacOSX, we don't want the normal expansion of a libcall to sincos. | |||
1661 | // We want to issue a libcall to __sincos_stret to avoid memory traffic. | |||
1662 | setOperationAction(ISD::FSINCOS, MVT::f64, Custom); | |||
1663 | setOperationAction(ISD::FSINCOS, MVT::f32, Custom); | |||
1664 | } | |||
1665 | } | |||
1666 | ||||
1667 | if (Subtarget->isTargetWin64()) { | |||
1668 | setOperationAction(ISD::SDIV, MVT::i128, Custom); | |||
1669 | setOperationAction(ISD::UDIV, MVT::i128, Custom); | |||
1670 | setOperationAction(ISD::SREM, MVT::i128, Custom); | |||
1671 | setOperationAction(ISD::UREM, MVT::i128, Custom); | |||
1672 | setOperationAction(ISD::SDIVREM, MVT::i128, Custom); | |||
1673 | setOperationAction(ISD::UDIVREM, MVT::i128, Custom); | |||
1674 | } | |||
1675 | ||||
1676 | // We have target-specific dag combine patterns for the following nodes: | |||
1677 | setTargetDAGCombine(ISD::VECTOR_SHUFFLE); | |||
1678 | setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT); | |||
1679 | setTargetDAGCombine(ISD::VSELECT); | |||
1680 | setTargetDAGCombine(ISD::SELECT); | |||
1681 | setTargetDAGCombine(ISD::SHL); | |||
1682 | setTargetDAGCombine(ISD::SRA); | |||
1683 | setTargetDAGCombine(ISD::SRL); | |||
1684 | setTargetDAGCombine(ISD::OR); | |||
1685 | setTargetDAGCombine(ISD::AND); | |||
1686 | setTargetDAGCombine(ISD::ADD); | |||
1687 | setTargetDAGCombine(ISD::FADD); | |||
1688 | setTargetDAGCombine(ISD::FSUB); | |||
1689 | setTargetDAGCombine(ISD::FMA); | |||
1690 | setTargetDAGCombine(ISD::SUB); | |||
1691 | setTargetDAGCombine(ISD::LOAD); | |||
1692 | setTargetDAGCombine(ISD::MLOAD); | |||
1693 | setTargetDAGCombine(ISD::STORE); | |||
1694 | setTargetDAGCombine(ISD::MSTORE); | |||
1695 | setTargetDAGCombine(ISD::ZERO_EXTEND); | |||
1696 | setTargetDAGCombine(ISD::ANY_EXTEND); | |||
1697 | setTargetDAGCombine(ISD::SIGN_EXTEND); | |||
1698 | setTargetDAGCombine(ISD::SIGN_EXTEND_INREG); | |||
1699 | setTargetDAGCombine(ISD::TRUNCATE); | |||
1700 | setTargetDAGCombine(ISD::SINT_TO_FP); | |||
1701 | setTargetDAGCombine(ISD::SETCC); | |||
1702 | setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN); | |||
1703 | setTargetDAGCombine(ISD::BUILD_VECTOR); | |||
1704 | setTargetDAGCombine(ISD::MUL); | |||
1705 | setTargetDAGCombine(ISD::XOR); | |||
1706 | ||||
1707 | computeRegisterProperties(); | |||
1708 | ||||
1709 | // On Darwin, -Os means optimize for size without hurting performance, | |||
1710 | // do not reduce the limit. | |||
1711 | MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores | |||
1712 | MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8; | |||
1713 | MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores | |||
1714 | MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4; | |||
1715 | MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores | |||
1716 | MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4; | |||
1717 | setPrefLoopAlignment(4); // 2^4 bytes. | |||
1718 | ||||
1719 | // Predictable cmov don't hurt on atom because it's in-order. | |||
1720 | PredictableSelectIsExpensive = !Subtarget->isAtom(); | |||
1721 | EnableExtLdPromotion = true; | |||
1722 | setPrefFunctionAlignment(4); // 2^4 bytes. | |||
1723 | ||||
1724 | verifyIntrinsicTables(); | |||
1725 | } | |||
1726 | ||||
1727 | // This has so far only been implemented for 64-bit MachO. | |||
1728 | bool X86TargetLowering::useLoadStackGuardNode() const { | |||
1729 | return Subtarget->isTargetMachO() && Subtarget->is64Bit(); | |||
1730 | } | |||
1731 | ||||
1732 | TargetLoweringBase::LegalizeTypeAction | |||
1733 | X86TargetLowering::getPreferredVectorAction(EVT VT) const { | |||
1734 | if (ExperimentalVectorWideningLegalization && | |||
1735 | VT.getVectorNumElements() != 1 && | |||
1736 | VT.getVectorElementType().getSimpleVT() != MVT::i1) | |||
1737 | return TypeWidenVector; | |||
1738 | ||||
1739 | return TargetLoweringBase::getPreferredVectorAction(VT); | |||
1740 | } | |||
1741 | ||||
1742 | EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const { | |||
1743 | if (!VT.isVector()) | |||
1744 | return Subtarget->hasAVX512() ? MVT::i1: MVT::i8; | |||
1745 | ||||
1746 | const unsigned NumElts = VT.getVectorNumElements(); | |||
1747 | const EVT EltVT = VT.getVectorElementType(); | |||
1748 | if (VT.is512BitVector()) { | |||
1749 | if (Subtarget->hasAVX512()) | |||
1750 | if (EltVT == MVT::i32 || EltVT == MVT::i64 || | |||
1751 | EltVT == MVT::f32 || EltVT == MVT::f64) | |||
1752 | switch(NumElts) { | |||
1753 | case 8: return MVT::v8i1; | |||
1754 | case 16: return MVT::v16i1; | |||
1755 | } | |||
1756 | if (Subtarget->hasBWI()) | |||
1757 | if (EltVT == MVT::i8 || EltVT == MVT::i16) | |||
1758 | switch(NumElts) { | |||
1759 | case 32: return MVT::v32i1; | |||
1760 | case 64: return MVT::v64i1; | |||
1761 | } | |||
1762 | } | |||
1763 | ||||
1764 | if (VT.is256BitVector() || VT.is128BitVector()) { | |||
1765 | if (Subtarget->hasVLX()) | |||
1766 | if (EltVT == MVT::i32 || EltVT == MVT::i64 || | |||
1767 | EltVT == MVT::f32 || EltVT == MVT::f64) | |||
1768 | switch(NumElts) { | |||
1769 | case 2: return MVT::v2i1; | |||
1770 | case 4: return MVT::v4i1; | |||
1771 | case 8: return MVT::v8i1; | |||
1772 | } | |||
1773 | if (Subtarget->hasBWI() && Subtarget->hasVLX()) | |||
1774 | if (EltVT == MVT::i8 || EltVT == MVT::i16) | |||
1775 | switch(NumElts) { | |||
1776 | case 8: return MVT::v8i1; | |||
1777 | case 16: return MVT::v16i1; | |||
1778 | case 32: return MVT::v32i1; | |||
1779 | } | |||
1780 | } | |||
1781 | ||||
1782 | return VT.changeVectorElementTypeToInteger(); | |||
1783 | } | |||
1784 | ||||
1785 | /// Helper for getByValTypeAlignment to determine | |||
1786 | /// the desired ByVal argument alignment. | |||
1787 | static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) { | |||
1788 | if (MaxAlign == 16) | |||
1789 | return; | |||
1790 | if (VectorType *VTy = dyn_cast<VectorType>(Ty)) { | |||
1791 | if (VTy->getBitWidth() == 128) | |||
1792 | MaxAlign = 16; | |||
1793 | } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { | |||
1794 | unsigned EltAlign = 0; | |||
1795 | getMaxByValAlign(ATy->getElementType(), EltAlign); | |||
1796 | if (EltAlign > MaxAlign) | |||
1797 | MaxAlign = EltAlign; | |||
1798 | } else if (StructType *STy = dyn_cast<StructType>(Ty)) { | |||
1799 | for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { | |||
1800 | unsigned EltAlign = 0; | |||
1801 | getMaxByValAlign(STy->getElementType(i), EltAlign); | |||
1802 | if (EltAlign > MaxAlign) | |||
1803 | MaxAlign = EltAlign; | |||
1804 | if (MaxAlign == 16) | |||
1805 | break; | |||
1806 | } | |||
1807 | } | |||
1808 | } | |||
1809 | ||||
1810 | /// Return the desired alignment for ByVal aggregate | |||
1811 | /// function arguments in the caller parameter area. For X86, aggregates | |||
1812 | /// that contain SSE vectors are placed at 16-byte boundaries while the rest | |||
1813 | /// are at 4-byte boundaries. | |||
1814 | unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const { | |||
1815 | if (Subtarget->is64Bit()) { | |||
1816 | // Max of 8 and alignment of type. | |||
1817 | unsigned TyAlign = TD->getABITypeAlignment(Ty); | |||
1818 | if (TyAlign > 8) | |||
1819 | return TyAlign; | |||
1820 | return 8; | |||
1821 | } | |||
1822 | ||||
1823 | unsigned Align = 4; | |||
1824 | if (Subtarget->hasSSE1()) | |||
1825 | getMaxByValAlign(Ty, Align); | |||
1826 | return Align; | |||
1827 | } | |||
1828 | ||||
1829 | /// Returns the target specific optimal type for load | |||
1830 | /// and store operations as a result of memset, memcpy, and memmove | |||
1831 | /// lowering. If DstAlign is zero that means it's safe to destination | |||
1832 | /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it | |||
1833 | /// means there isn't a need to check it against alignment requirement, | |||
1834 | /// probably because the source does not need to be loaded. If 'IsMemset' is | |||
1835 | /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that | |||
1836 | /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy | |||
1837 | /// source is constant so it does not need to be loaded. | |||
1838 | /// It returns EVT::Other if the type should be determined using generic | |||
1839 | /// target-independent logic. | |||
1840 | EVT | |||
1841 | X86TargetLowering::getOptimalMemOpType(uint64_t Size, | |||
1842 | unsigned DstAlign, unsigned SrcAlign, | |||
1843 | bool IsMemset, bool ZeroMemset, | |||
1844 | bool MemcpyStrSrc, | |||
1845 | MachineFunction &MF) const { | |||
1846 | const Function *F = MF.getFunction(); | |||
1847 | if ((!IsMemset || ZeroMemset) && | |||
1848 | !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex, | |||
1849 | Attribute::NoImplicitFloat)) { | |||
1850 | if (Size >= 16 && | |||
1851 | (Subtarget->isUnalignedMemAccessFast() || | |||
1852 | ((DstAlign == 0 || DstAlign >= 16) && | |||
1853 | (SrcAlign == 0 || SrcAlign >= 16)))) { | |||
1854 | if (Size >= 32) { | |||
1855 | if (Subtarget->hasInt256()) | |||
1856 | return MVT::v8i32; | |||
1857 | if (Subtarget->hasFp256()) | |||
1858 | return MVT::v8f32; | |||
1859 | } | |||
1860 | if (Subtarget->hasSSE2()) | |||
1861 | return MVT::v4i32; | |||
1862 | if (Subtarget->hasSSE1()) | |||
1863 | return MVT::v4f32; | |||
1864 | } else if (!MemcpyStrSrc && Size >= 8 && | |||
1865 | !Subtarget->is64Bit() && | |||
1866 | Subtarget->hasSSE2()) { | |||
1867 | // Do not use f64 to lower memcpy if source is string constant. It's | |||
1868 | // better to use i32 to avoid the loads. | |||
1869 | return MVT::f64; | |||
1870 | } | |||
1871 | } | |||
1872 | if (Subtarget->is64Bit() && Size >= 8) | |||
1873 | return MVT::i64; | |||
1874 | return MVT::i32; | |||
1875 | } | |||
1876 | ||||
1877 | bool X86TargetLowering::isSafeMemOpType(MVT VT) const { | |||
1878 | if (VT == MVT::f32) | |||
1879 | return X86ScalarSSEf32; | |||
1880 | else if (VT == MVT::f64) | |||
1881 | return X86ScalarSSEf64; | |||
1882 | return true; | |||
1883 | } | |||
1884 | ||||
1885 | bool | |||
1886 | X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT, | |||
1887 | unsigned, | |||
1888 | unsigned, | |||
1889 | bool *Fast) const { | |||
1890 | if (Fast) | |||
1891 | *Fast = Subtarget->isUnalignedMemAccessFast(); | |||
1892 | return true; | |||
1893 | } | |||
1894 | ||||
1895 | /// Return the entry encoding for a jump table in the | |||
1896 | /// current function. The returned value is a member of the | |||
1897 | /// MachineJumpTableInfo::JTEntryKind enum. | |||
1898 | unsigned X86TargetLowering::getJumpTableEncoding() const { | |||
1899 | // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF | |||
1900 | // symbol. | |||
1901 | if (getTargetMachine().getRelocationModel() == Reloc::PIC_ && | |||
1902 | Subtarget->isPICStyleGOT()) | |||
1903 | return MachineJumpTableInfo::EK_Custom32; | |||
1904 | ||||
1905 | // Otherwise, use the normal jump table encoding heuristics. | |||
1906 | return TargetLowering::getJumpTableEncoding(); | |||
1907 | } | |||
1908 | ||||
1909 | const MCExpr * | |||
1910 | X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI, | |||
1911 | const MachineBasicBlock *MBB, | |||
1912 | unsigned uid,MCContext &Ctx) const{ | |||
1913 | assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&((MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ && Subtarget->isPICStyleGOT()) ? static_cast <void> (0) : __assert_fail ("MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ && Subtarget->isPICStyleGOT()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 1914, __PRETTY_FUNCTION__)) | |||
1914 | Subtarget->isPICStyleGOT())((MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ && Subtarget->isPICStyleGOT()) ? static_cast <void> (0) : __assert_fail ("MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ && Subtarget->isPICStyleGOT()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 1914, __PRETTY_FUNCTION__)); | |||
1915 | // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF | |||
1916 | // entries. | |||
1917 | return MCSymbolRefExpr::Create(MBB->getSymbol(), | |||
1918 | MCSymbolRefExpr::VK_GOTOFF, Ctx); | |||
1919 | } | |||
1920 | ||||
1921 | /// Returns relocation base for the given PIC jumptable. | |||
1922 | SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table, | |||
1923 | SelectionDAG &DAG) const { | |||
1924 | if (!Subtarget->is64Bit()) | |||
1925 | // This doesn't have SDLoc associated with it, but is not really the | |||
1926 | // same as a Register. | |||
1927 | return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy()); | |||
1928 | return Table; | |||
1929 | } | |||
1930 | ||||
1931 | /// This returns the relocation base for the given PIC jumptable, | |||
1932 | /// the same as getPICJumpTableRelocBase, but as an MCExpr. | |||
1933 | const MCExpr *X86TargetLowering:: | |||
1934 | getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI, | |||
1935 | MCContext &Ctx) const { | |||
1936 | // X86-64 uses RIP relative addressing based on the jump table label. | |||
1937 | if (Subtarget->isPICStyleRIPRel()) | |||
1938 | return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx); | |||
1939 | ||||
1940 | // Otherwise, the reference is relative to the PIC base. | |||
1941 | return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx); | |||
1942 | } | |||
1943 | ||||
1944 | // FIXME: Why this routine is here? Move to RegInfo! | |||
1945 | std::pair<const TargetRegisterClass*, uint8_t> | |||
1946 | X86TargetLowering::findRepresentativeClass(MVT VT) const{ | |||
1947 | const TargetRegisterClass *RRC = nullptr; | |||
1948 | uint8_t Cost = 1; | |||
1949 | switch (VT.SimpleTy) { | |||
1950 | default: | |||
1951 | return TargetLowering::findRepresentativeClass(VT); | |||
1952 | case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64: | |||
1953 | RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass; | |||
1954 | break; | |||
1955 | case MVT::x86mmx: | |||
1956 | RRC = &X86::VR64RegClass; | |||
1957 | break; | |||
1958 | case MVT::f32: case MVT::f64: | |||
1959 | case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64: | |||
1960 | case MVT::v4f32: case MVT::v2f64: | |||
1961 | case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32: | |||
1962 | case MVT::v4f64: | |||
1963 | RRC = &X86::VR128RegClass; | |||
1964 | break; | |||
1965 | } | |||
1966 | return std::make_pair(RRC, Cost); | |||
1967 | } | |||
1968 | ||||
1969 | bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace, | |||
1970 | unsigned &Offset) const { | |||
1971 | if (!Subtarget->isTargetLinux()) | |||
1972 | return false; | |||
1973 | ||||
1974 | if (Subtarget->is64Bit()) { | |||
1975 | // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs: | |||
1976 | Offset = 0x28; | |||
1977 | if (getTargetMachine().getCodeModel() == CodeModel::Kernel) | |||
1978 | AddressSpace = 256; | |||
1979 | else | |||
1980 | AddressSpace = 257; | |||
1981 | } else { | |||
1982 | // %gs:0x14 on i386 | |||
1983 | Offset = 0x14; | |||
1984 | AddressSpace = 256; | |||
1985 | } | |||
1986 | return true; | |||
1987 | } | |||
1988 | ||||
1989 | bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS, | |||
1990 | unsigned DestAS) const { | |||
1991 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 1991, __PRETTY_FUNCTION__)); | |||
1992 | ||||
1993 | return SrcAS < 256 && DestAS < 256; | |||
1994 | } | |||
1995 | ||||
1996 | //===----------------------------------------------------------------------===// | |||
1997 | // Return Value Calling Convention Implementation | |||
1998 | //===----------------------------------------------------------------------===// | |||
1999 | ||||
2000 | #include "X86GenCallingConv.inc" | |||
2001 | ||||
2002 | bool | |||
2003 | X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, | |||
2004 | MachineFunction &MF, bool isVarArg, | |||
2005 | const SmallVectorImpl<ISD::OutputArg> &Outs, | |||
2006 | LLVMContext &Context) const { | |||
2007 | SmallVector<CCValAssign, 16> RVLocs; | |||
2008 | CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context); | |||
2009 | return CCInfo.CheckReturn(Outs, RetCC_X86); | |||
2010 | } | |||
2011 | ||||
2012 | const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const { | |||
2013 | static const MCPhysReg ScratchRegs[] = { X86::R11, 0 }; | |||
2014 | return ScratchRegs; | |||
2015 | } | |||
2016 | ||||
2017 | SDValue | |||
2018 | X86TargetLowering::LowerReturn(SDValue Chain, | |||
2019 | CallingConv::ID CallConv, bool isVarArg, | |||
2020 | const SmallVectorImpl<ISD::OutputArg> &Outs, | |||
2021 | const SmallVectorImpl<SDValue> &OutVals, | |||
2022 | SDLoc dl, SelectionDAG &DAG) const { | |||
2023 | MachineFunction &MF = DAG.getMachineFunction(); | |||
2024 | X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); | |||
2025 | ||||
2026 | SmallVector<CCValAssign, 16> RVLocs; | |||
2027 | CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext()); | |||
2028 | CCInfo.AnalyzeReturn(Outs, RetCC_X86); | |||
2029 | ||||
2030 | SDValue Flag; | |||
2031 | SmallVector<SDValue, 6> RetOps; | |||
2032 | RetOps.push_back(Chain); // Operand #0 = Chain (updated below) | |||
2033 | // Operand #1 = Bytes To Pop | |||
2034 | RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), | |||
2035 | MVT::i16)); | |||
2036 | ||||
2037 | // Copy the result values into the output registers. | |||
2038 | for (unsigned i = 0; i != RVLocs.size(); ++i) { | |||
2039 | CCValAssign &VA = RVLocs[i]; | |||
2040 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2040, __PRETTY_FUNCTION__)); | |||
2041 | SDValue ValToCopy = OutVals[i]; | |||
2042 | EVT ValVT = ValToCopy.getValueType(); | |||
2043 | ||||
2044 | // Promote values to the appropriate types. | |||
2045 | if (VA.getLocInfo() == CCValAssign::SExt) | |||
2046 | ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy); | |||
2047 | else if (VA.getLocInfo() == CCValAssign::ZExt) | |||
2048 | ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy); | |||
2049 | else if (VA.getLocInfo() == CCValAssign::AExt) | |||
2050 | ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy); | |||
2051 | else if (VA.getLocInfo() == CCValAssign::BCvt) | |||
2052 | ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy); | |||
2053 | ||||
2054 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2055, __PRETTY_FUNCTION__)) | |||
2055 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2055, __PRETTY_FUNCTION__)); | |||
2056 | ||||
2057 | // If this is x86-64, and we disabled SSE, we can't return FP values, | |||
2058 | // or SSE or MMX vectors. | |||
2059 | if ((ValVT == MVT::f32 || ValVT == MVT::f64 || | |||
2060 | VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) && | |||
2061 | (Subtarget->is64Bit() && !Subtarget->hasSSE1())) { | |||
2062 | report_fatal_error("SSE register return with SSE disabled"); | |||
2063 | } | |||
2064 | // Likewise we can't return F64 values with SSE1 only. gcc does so, but | |||
2065 | // llvm-gcc has never done it right and no one has noticed, so this | |||
2066 | // should be OK for now. | |||
2067 | if (ValVT == MVT::f64 && | |||
2068 | (Subtarget->is64Bit() && !Subtarget->hasSSE2())) | |||
2069 | report_fatal_error("SSE2 register return with SSE2 disabled"); | |||
2070 | ||||
2071 | // Returns in ST0/ST1 are handled specially: these are pushed as operands to | |||
2072 | // the RET instruction and handled by the FP Stackifier. | |||
2073 | if (VA.getLocReg() == X86::FP0 || | |||
2074 | VA.getLocReg() == X86::FP1) { | |||
2075 | // If this is a copy from an xmm register to ST(0), use an FPExtend to | |||
2076 | // change the value to the FP stack register class. | |||
2077 | if (isScalarFPTypeInSSEReg(VA.getValVT())) | |||
2078 | ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy); | |||
2079 | RetOps.push_back(ValToCopy); | |||
2080 | // Don't emit a copytoreg. | |||
2081 | continue; | |||
2082 | } | |||
2083 | ||||
2084 | // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64 | |||
2085 | // which is returned in RAX / RDX. | |||
2086 | if (Subtarget->is64Bit()) { | |||
2087 | if (ValVT == MVT::x86mmx) { | |||
2088 | if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) { | |||
2089 | ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy); | |||
2090 | ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, | |||
2091 | ValToCopy); | |||
2092 | // If we don't have SSE2 available, convert to v4f32 so the generated | |||
2093 | // register is legal. | |||
2094 | if (!Subtarget->hasSSE2()) | |||
2095 | ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy); | |||
2096 | } | |||
2097 | } | |||
2098 | } | |||
2099 | ||||
2100 | Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag); | |||
2101 | Flag = Chain.getValue(1); | |||
2102 | RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); | |||
2103 | } | |||
2104 | ||||
2105 | // The x86-64 ABIs require that for returning structs by value we copy | |||
2106 | // the sret argument into %rax/%eax (depending on ABI) for the return. | |||
2107 | // Win32 requires us to put the sret argument to %eax as well. | |||
2108 | // We saved the argument into a virtual register in the entry block, | |||
2109 | // so now we copy the value out and into %rax/%eax. | |||
2110 | if (DAG.getMachineFunction().getFunction()->hasStructRetAttr() && | |||
2111 | (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC())) { | |||
2112 | MachineFunction &MF = DAG.getMachineFunction(); | |||
2113 | X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); | |||
2114 | unsigned Reg = FuncInfo->getSRetReturnReg(); | |||
2115 | assert(Reg &&((Reg && "SRetReturnReg should have been set in LowerFormalArguments()." ) ? static_cast<void> (0) : __assert_fail ("Reg && \"SRetReturnReg should have been set in LowerFormalArguments().\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2116, __PRETTY_FUNCTION__)) | |||
2116 | "SRetReturnReg should have been set in LowerFormalArguments().")((Reg && "SRetReturnReg should have been set in LowerFormalArguments()." ) ? static_cast<void> (0) : __assert_fail ("Reg && \"SRetReturnReg should have been set in LowerFormalArguments().\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2116, __PRETTY_FUNCTION__)); | |||
2117 | SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy()); | |||
2118 | ||||
2119 | unsigned RetValReg | |||
2120 | = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ? | |||
2121 | X86::RAX : X86::EAX; | |||
2122 | Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag); | |||
2123 | Flag = Chain.getValue(1); | |||
2124 | ||||
2125 | // RAX/EAX now acts like a return value. | |||
2126 | RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy())); | |||
2127 | } | |||
2128 | ||||
2129 | RetOps[0] = Chain; // Update chain. | |||
2130 | ||||
2131 | // Add the flag if we have it. | |||
2132 | if (Flag.getNode()) | |||
2133 | RetOps.push_back(Flag); | |||
2134 | ||||
2135 | return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps); | |||
2136 | } | |||
2137 | ||||
2138 | bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const { | |||
2139 | if (N->getNumValues() != 1) | |||
2140 | return false; | |||
2141 | if (!N->hasNUsesOfValue(1, 0)) | |||
2142 | return false; | |||
2143 | ||||
2144 | SDValue TCChain = Chain; | |||
2145 | SDNode *Copy = *N->use_begin(); | |||
2146 | if (Copy->getOpcode() == ISD::CopyToReg) { | |||
2147 | // If the copy has a glue operand, we conservatively assume it isn't safe to | |||
2148 | // perform a tail call. | |||
2149 | if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue) | |||
2150 | return false; | |||
2151 | TCChain = Copy->getOperand(0); | |||
2152 | } else if (Copy->getOpcode() != ISD::FP_EXTEND) | |||
2153 | return false; | |||
2154 | ||||
2155 | bool HasRet = false; | |||
2156 | for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end(); | |||
2157 | UI != UE; ++UI) { | |||
2158 | if (UI->getOpcode() != X86ISD::RET_FLAG) | |||
2159 | return false; | |||
2160 | // If we are returning more than one value, we can definitely | |||
2161 | // not make a tail call see PR19530 | |||
2162 | if (UI->getNumOperands() > 4) | |||
2163 | return false; | |||
2164 | if (UI->getNumOperands() == 4 && | |||
2165 | UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue) | |||
2166 | return false; | |||
2167 | HasRet = true; | |||
2168 | } | |||
2169 | ||||
2170 | if (!HasRet) | |||
2171 | return false; | |||
2172 | ||||
2173 | Chain = TCChain; | |||
2174 | return true; | |||
2175 | } | |||
2176 | ||||
2177 | EVT | |||
2178 | X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT, | |||
2179 | ISD::NodeType ExtendKind) const { | |||
2180 | MVT ReturnMVT; | |||
2181 | // TODO: Is this also valid on 32-bit? | |||
2182 | if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND) | |||
2183 | ReturnMVT = MVT::i8; | |||
2184 | else | |||
2185 | ReturnMVT = MVT::i32; | |||
2186 | ||||
2187 | EVT MinVT = getRegisterType(Context, ReturnMVT); | |||
2188 | return VT.bitsLT(MinVT) ? MinVT : VT; | |||
2189 | } | |||
2190 | ||||
2191 | /// Lower the result values of a call into the | |||
2192 | /// appropriate copies out of appropriate physical registers. | |||
2193 | /// | |||
2194 | SDValue | |||
2195 | X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, | |||
2196 | CallingConv::ID CallConv, bool isVarArg, | |||
2197 | const SmallVectorImpl<ISD::InputArg> &Ins, | |||
2198 | SDLoc dl, SelectionDAG &DAG, | |||
2199 | SmallVectorImpl<SDValue> &InVals) const { | |||
2200 | ||||
2201 | // Assign locations to each value returned by this call. | |||
2202 | SmallVector<CCValAssign, 16> RVLocs; | |||
2203 | bool Is64Bit = Subtarget->is64Bit(); | |||
2204 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, | |||
2205 | *DAG.getContext()); | |||
2206 | CCInfo.AnalyzeCallResult(Ins, RetCC_X86); | |||
2207 | ||||
2208 | // Copy all of the result registers out of their specified physreg. | |||
2209 | for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) { | |||
2210 | CCValAssign &VA = RVLocs[i]; | |||
2211 | EVT CopyVT = VA.getValVT(); | |||
2212 | ||||
2213 | // If this is x86-64, and we disabled SSE, we can't return FP values | |||
2214 | if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) && | |||
2215 | ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) { | |||
2216 | report_fatal_error("SSE register return with SSE disabled"); | |||
2217 | } | |||
2218 | ||||
2219 | // If we prefer to use the value in xmm registers, copy it out as f80 and | |||
2220 | // use a truncate to move it from fp stack reg to xmm reg. | |||
2221 | if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) && | |||
2222 | isScalarFPTypeInSSEReg(VA.getValVT())) | |||
2223 | CopyVT = MVT::f80; | |||
2224 | ||||
2225 | Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), | |||
2226 | CopyVT, InFlag).getValue(1); | |||
2227 | SDValue Val = Chain.getValue(0); | |||
2228 | ||||
2229 | if (CopyVT != VA.getValVT()) | |||
2230 | Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val, | |||
2231 | // This truncation won't change the value. | |||
2232 | DAG.getIntPtrConstant(1)); | |||
2233 | ||||
2234 | InFlag = Chain.getValue(2); | |||
2235 | InVals.push_back(Val); | |||
2236 | } | |||
2237 | ||||
2238 | return Chain; | |||
2239 | } | |||
2240 | ||||
2241 | //===----------------------------------------------------------------------===// | |||
2242 | // C & StdCall & Fast Calling Convention implementation | |||
2243 | //===----------------------------------------------------------------------===// | |||
2244 | // StdCall calling convention seems to be standard for many Windows' API | |||
2245 | // routines and around. It differs from C calling convention just a little: | |||
2246 | // callee should clean up the stack, not caller. Symbols should be also | |||
2247 | // decorated in some fancy way :) It doesn't support any vector arguments. | |||
2248 | // For info on fast calling convention see Fast Calling Convention (tail call) | |||
2249 | // implementation LowerX86_32FastCCCallTo. | |||
2250 | ||||
2251 | /// CallIsStructReturn - Determines whether a call uses struct return | |||
2252 | /// semantics. | |||
2253 | enum StructReturnType { | |||
2254 | NotStructReturn, | |||
2255 | RegStructReturn, | |||
2256 | StackStructReturn | |||
2257 | }; | |||
2258 | static StructReturnType | |||
2259 | callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) { | |||
2260 | if (Outs.empty()) | |||
2261 | return NotStructReturn; | |||
2262 | ||||
2263 | const ISD::ArgFlagsTy &Flags = Outs[0].Flags; | |||
2264 | if (!Flags.isSRet()) | |||
2265 | return NotStructReturn; | |||
2266 | if (Flags.isInReg()) | |||
2267 | return RegStructReturn; | |||
2268 | return StackStructReturn; | |||
2269 | } | |||
2270 | ||||
2271 | /// Determines whether a function uses struct return semantics. | |||
2272 | static StructReturnType | |||
2273 | argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) { | |||
2274 | if (Ins.empty()) | |||
2275 | return NotStructReturn; | |||
2276 | ||||
2277 | const ISD::ArgFlagsTy &Flags = Ins[0].Flags; | |||
2278 | if (!Flags.isSRet()) | |||
2279 | return NotStructReturn; | |||
2280 | if (Flags.isInReg()) | |||
2281 | return RegStructReturn; | |||
2282 | return StackStructReturn; | |||
2283 | } | |||
2284 | ||||
2285 | /// Make a copy of an aggregate at address specified by "Src" to address | |||
2286 | /// "Dst" with size and alignment information specified by the specific | |||
2287 | /// parameter attribute. The copy will be passed as a byval function parameter. | |||
2288 | static SDValue | |||
2289 | CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain, | |||
2290 | ISD::ArgFlagsTy Flags, SelectionDAG &DAG, | |||
2291 | SDLoc dl) { | |||
2292 | SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32); | |||
2293 | ||||
2294 | return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(), | |||
2295 | /*isVolatile*/false, /*AlwaysInline=*/true, | |||
2296 | MachinePointerInfo(), MachinePointerInfo()); | |||
2297 | } | |||
2298 | ||||
2299 | /// Return true if the calling convention is one that | |||
2300 | /// supports tail call optimization. | |||
2301 | static bool IsTailCallConvention(CallingConv::ID CC) { | |||
2302 | return (CC == CallingConv::Fast || CC == CallingConv::GHC || | |||
2303 | CC == CallingConv::HiPE); | |||
2304 | } | |||
2305 | ||||
2306 | /// \brief Return true if the calling convention is a C calling convention. | |||
2307 | static bool IsCCallConvention(CallingConv::ID CC) { | |||
2308 | return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 || | |||
2309 | CC == CallingConv::X86_64_SysV); | |||
2310 | } | |||
2311 | ||||
2312 | bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const { | |||
2313 | if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls) | |||
2314 | return false; | |||
2315 | ||||
2316 | CallSite CS(CI); | |||
2317 | CallingConv::ID CalleeCC = CS.getCallingConv(); | |||
2318 | if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC)) | |||
2319 | return false; | |||
2320 | ||||
2321 | return true; | |||
2322 | } | |||
2323 | ||||
2324 | /// Return true if the function is being made into | |||
2325 | /// a tailcall target by changing its ABI. | |||
2326 | static bool FuncIsMadeTailCallSafe(CallingConv::ID CC, | |||
2327 | bool GuaranteedTailCallOpt) { | |||
2328 | return GuaranteedTailCallOpt && IsTailCallConvention(CC); | |||
2329 | } | |||
2330 | ||||
2331 | SDValue | |||
2332 | X86TargetLowering::LowerMemArgument(SDValue Chain, | |||
2333 | CallingConv::ID CallConv, | |||
2334 | const SmallVectorImpl<ISD::InputArg> &Ins, | |||
2335 | SDLoc dl, SelectionDAG &DAG, | |||
2336 | const CCValAssign &VA, | |||
2337 | MachineFrameInfo *MFI, | |||
2338 | unsigned i) const { | |||
2339 | // Create the nodes corresponding to a load from this parameter slot. | |||
2340 | ISD::ArgFlagsTy Flags = Ins[i].Flags; | |||
2341 | bool AlwaysUseMutable = FuncIsMadeTailCallSafe( | |||
2342 | CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt); | |||
2343 | bool isImmutable = !AlwaysUseMutable && !Flags.isByVal(); | |||
2344 | EVT ValVT; | |||
2345 | ||||
2346 | // If value is passed by pointer we have address passed instead of the value | |||
2347 | // itself. | |||
2348 | if (VA.getLocInfo() == CCValAssign::Indirect) | |||
2349 | ValVT = VA.getLocVT(); | |||
2350 | else | |||
2351 | ValVT = VA.getValVT(); | |||
2352 | ||||
2353 | // FIXME: For now, all byval parameter objects are marked mutable. This can be | |||
2354 | // changed with more analysis. | |||
2355 | // In case of tail call optimization mark all arguments mutable. Since they | |||
2356 | // could be overwritten by lowering of arguments in case of a tail call. | |||
2357 | if (Flags.isByVal()) { | |||
2358 | unsigned Bytes = Flags.getByValSize(); | |||
2359 | if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects. | |||
2360 | int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable); | |||
2361 | return DAG.getFrameIndex(FI, getPointerTy()); | |||
2362 | } else { | |||
2363 | int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8, | |||
2364 | VA.getLocMemOffset(), isImmutable); | |||
2365 | SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); | |||
2366 | return DAG.getLoad(ValVT, dl, Chain, FIN, | |||
2367 | MachinePointerInfo::getFixedStack(FI), | |||
2368 | false, false, false, 0); | |||
2369 | } | |||
2370 | } | |||
2371 | ||||
2372 | // FIXME: Get this from tablegen. | |||
2373 | static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv, | |||
2374 | const X86Subtarget *Subtarget) { | |||
2375 | assert(Subtarget->is64Bit())((Subtarget->is64Bit()) ? static_cast<void> (0) : __assert_fail ("Subtarget->is64Bit()", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2375, __PRETTY_FUNCTION__)); | |||
2376 | ||||
2377 | if (Subtarget->isCallingConvWin64(CallConv)) { | |||
2378 | static const MCPhysReg GPR64ArgRegsWin64[] = { | |||
2379 | X86::RCX, X86::RDX, X86::R8, X86::R9 | |||
2380 | }; | |||
2381 | return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64)); | |||
2382 | } | |||
2383 | ||||
2384 | static const MCPhysReg GPR64ArgRegs64Bit[] = { | |||
2385 | X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9 | |||
2386 | }; | |||
2387 | return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit)); | |||
2388 | } | |||
2389 | ||||
2390 | // FIXME: Get this from tablegen. | |||
2391 | static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF, | |||
2392 | CallingConv::ID CallConv, | |||
2393 | const X86Subtarget *Subtarget) { | |||
2394 | assert(Subtarget->is64Bit())((Subtarget->is64Bit()) ? static_cast<void> (0) : __assert_fail ("Subtarget->is64Bit()", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2394, __PRETTY_FUNCTION__)); | |||
2395 | if (Subtarget->isCallingConvWin64(CallConv)) { | |||
2396 | // The XMM registers which might contain var arg parameters are shadowed | |||
2397 | // in their paired GPR. So we only need to save the GPR to their home | |||
2398 | // slots. | |||
2399 | // TODO: __vectorcall will change this. | |||
2400 | return None; | |||
2401 | } | |||
2402 | ||||
2403 | const Function *Fn = MF.getFunction(); | |||
2404 | bool NoImplicitFloatOps = Fn->getAttributes(). | |||
2405 | hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat); | |||
2406 | assert(!(MF.getTarget().Options.UseSoftFloat && NoImplicitFloatOps) &&((!(MF.getTarget().Options.UseSoftFloat && NoImplicitFloatOps ) && "SSE register cannot be used when SSE is disabled!" ) ? static_cast<void> (0) : __assert_fail ("!(MF.getTarget().Options.UseSoftFloat && NoImplicitFloatOps) && \"SSE register cannot be used when SSE is disabled!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2407, __PRETTY_FUNCTION__)) | |||
2407 | "SSE register cannot be used when SSE is disabled!")((!(MF.getTarget().Options.UseSoftFloat && NoImplicitFloatOps ) && "SSE register cannot be used when SSE is disabled!" ) ? static_cast<void> (0) : __assert_fail ("!(MF.getTarget().Options.UseSoftFloat && NoImplicitFloatOps) && \"SSE register cannot be used when SSE is disabled!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2407, __PRETTY_FUNCTION__)); | |||
2408 | if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps || | |||
2409 | !Subtarget->hasSSE1()) | |||
2410 | // Kernel mode asks for SSE to be disabled, so there are no XMM argument | |||
2411 | // registers. | |||
2412 | return None; | |||
2413 | ||||
2414 | static const MCPhysReg XMMArgRegs64Bit[] = { | |||
2415 | X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, | |||
2416 | X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7 | |||
2417 | }; | |||
2418 | return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit)); | |||
2419 | } | |||
2420 | ||||
2421 | SDValue | |||
2422 | X86TargetLowering::LowerFormalArguments(SDValue Chain, | |||
2423 | CallingConv::ID CallConv, | |||
2424 | bool isVarArg, | |||
2425 | const SmallVectorImpl<ISD::InputArg> &Ins, | |||
2426 | SDLoc dl, | |||
2427 | SelectionDAG &DAG, | |||
2428 | SmallVectorImpl<SDValue> &InVals) | |||
2429 | const { | |||
2430 | MachineFunction &MF = DAG.getMachineFunction(); | |||
2431 | X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); | |||
2432 | ||||
2433 | const Function* Fn = MF.getFunction(); | |||
2434 | if (Fn->hasExternalLinkage() && | |||
2435 | Subtarget->isTargetCygMing() && | |||
2436 | Fn->getName() == "main") | |||
2437 | FuncInfo->setForceFramePointer(true); | |||
2438 | ||||
2439 | MachineFrameInfo *MFI = MF.getFrameInfo(); | |||
2440 | bool Is64Bit = Subtarget->is64Bit(); | |||
2441 | bool IsWin64 = Subtarget->isCallingConvWin64(CallConv); | |||
2442 | ||||
2443 | assert(!(isVarArg && IsTailCallConvention(CallConv)) &&((!(isVarArg && IsTailCallConvention(CallConv)) && "Var args not supported with calling convention fastcc, ghc or hipe" ) ? static_cast<void> (0) : __assert_fail ("!(isVarArg && IsTailCallConvention(CallConv)) && \"Var args not supported with calling convention fastcc, ghc or hipe\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2444, __PRETTY_FUNCTION__)) | |||
2444 | "Var args not supported with calling convention fastcc, ghc or hipe")((!(isVarArg && IsTailCallConvention(CallConv)) && "Var args not supported with calling convention fastcc, ghc or hipe" ) ? static_cast<void> (0) : __assert_fail ("!(isVarArg && IsTailCallConvention(CallConv)) && \"Var args not supported with calling convention fastcc, ghc or hipe\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2444, __PRETTY_FUNCTION__)); | |||
2445 | ||||
2446 | // Assign locations to all of the incoming arguments. | |||
2447 | SmallVector<CCValAssign, 16> ArgLocs; | |||
2448 | CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext()); | |||
2449 | ||||
2450 | // Allocate shadow area for Win64 | |||
2451 | if (IsWin64) | |||
2452 | CCInfo.AllocateStack(32, 8); | |||
2453 | ||||
2454 | CCInfo.AnalyzeFormalArguments(Ins, CC_X86); | |||
2455 | ||||
2456 | unsigned LastVal = ~0U; | |||
2457 | SDValue ArgValue; | |||
2458 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { | |||
2459 | CCValAssign &VA = ArgLocs[i]; | |||
2460 | // TODO: If an arg is passed in two places (e.g. reg and stack), skip later | |||
2461 | // places. | |||
2462 | assert(VA.getValNo() != LastVal &&((VA.getValNo() != LastVal && "Don't support value assigned to multiple locs yet" ) ? static_cast<void> (0) : __assert_fail ("VA.getValNo() != LastVal && \"Don't support value assigned to multiple locs yet\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2463, __PRETTY_FUNCTION__)) | |||
2463 | "Don't support value assigned to multiple locs yet")((VA.getValNo() != LastVal && "Don't support value assigned to multiple locs yet" ) ? static_cast<void> (0) : __assert_fail ("VA.getValNo() != LastVal && \"Don't support value assigned to multiple locs yet\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2463, __PRETTY_FUNCTION__)); | |||
2464 | (void)LastVal; | |||
2465 | LastVal = VA.getValNo(); | |||
2466 | ||||
2467 | if (VA.isRegLoc()) { | |||
2468 | EVT RegVT = VA.getLocVT(); | |||
2469 | const TargetRegisterClass *RC; | |||
2470 | if (RegVT == MVT::i32) | |||
2471 | RC = &X86::GR32RegClass; | |||
2472 | else if (Is64Bit && RegVT == MVT::i64) | |||
2473 | RC = &X86::GR64RegClass; | |||
2474 | else if (RegVT == MVT::f32) | |||
2475 | RC = &X86::FR32RegClass; | |||
2476 | else if (RegVT == MVT::f64) | |||
2477 | RC = &X86::FR64RegClass; | |||
2478 | else if (RegVT.is512BitVector()) | |||
2479 | RC = &X86::VR512RegClass; | |||
2480 | else if (RegVT.is256BitVector()) | |||
2481 | RC = &X86::VR256RegClass; | |||
2482 | else if (RegVT.is128BitVector()) | |||
2483 | RC = &X86::VR128RegClass; | |||
2484 | else if (RegVT == MVT::x86mmx) | |||
2485 | RC = &X86::VR64RegClass; | |||
2486 | else if (RegVT == MVT::i1) | |||
2487 | RC = &X86::VK1RegClass; | |||
2488 | else if (RegVT == MVT::v8i1) | |||
2489 | RC = &X86::VK8RegClass; | |||
2490 | else if (RegVT == MVT::v16i1) | |||
2491 | RC = &X86::VK16RegClass; | |||
2492 | else if (RegVT == MVT::v32i1) | |||
2493 | RC = &X86::VK32RegClass; | |||
2494 | else if (RegVT == MVT::v64i1) | |||
2495 | RC = &X86::VK64RegClass; | |||
2496 | else | |||
2497 | llvm_unreachable("Unknown argument type!")::llvm::llvm_unreachable_internal("Unknown argument type!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2497); | |||
2498 | ||||
2499 | unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); | |||
2500 | ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT); | |||
2501 | ||||
2502 | // If this is an 8 or 16-bit value, it is really passed promoted to 32 | |||
2503 | // bits. Insert an assert[sz]ext to capture this, then truncate to the | |||
2504 | // right size. | |||
2505 | if (VA.getLocInfo() == CCValAssign::SExt) | |||
2506 | ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue, | |||
2507 | DAG.getValueType(VA.getValVT())); | |||
2508 | else if (VA.getLocInfo() == CCValAssign::ZExt) | |||
2509 | ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue, | |||
2510 | DAG.getValueType(VA.getValVT())); | |||
2511 | else if (VA.getLocInfo() == CCValAssign::BCvt) | |||
2512 | ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue); | |||
2513 | ||||
2514 | if (VA.isExtInLoc()) { | |||
2515 | // Handle MMX values passed in XMM regs. | |||
2516 | if (RegVT.isVector()) | |||
2517 | ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue); | |||
2518 | else | |||
2519 | ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); | |||
2520 | } | |||
2521 | } else { | |||
2522 | assert(VA.isMemLoc())((VA.isMemLoc()) ? static_cast<void> (0) : __assert_fail ("VA.isMemLoc()", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2522, __PRETTY_FUNCTION__)); | |||
2523 | ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i); | |||
2524 | } | |||
2525 | ||||
2526 | // If value is passed via pointer - do a load. | |||
2527 | if (VA.getLocInfo() == CCValAssign::Indirect) | |||
2528 | ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, | |||
2529 | MachinePointerInfo(), false, false, false, 0); | |||
2530 | ||||
2531 | InVals.push_back(ArgValue); | |||
2532 | } | |||
2533 | ||||
2534 | if (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) { | |||
2535 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { | |||
2536 | // The x86-64 ABIs require that for returning structs by value we copy | |||
2537 | // the sret argument into %rax/%eax (depending on ABI) for the return. | |||
2538 | // Win32 requires us to put the sret argument to %eax as well. | |||
2539 | // Save the argument into a virtual register so that we can access it | |||
2540 | // from the return points. | |||
2541 | if (Ins[i].Flags.isSRet()) { | |||
2542 | unsigned Reg = FuncInfo->getSRetReturnReg(); | |||
2543 | if (!Reg) { | |||
2544 | MVT PtrTy = getPointerTy(); | |||
2545 | Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy)); | |||
2546 | FuncInfo->setSRetReturnReg(Reg); | |||
2547 | } | |||
2548 | SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]); | |||
2549 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain); | |||
2550 | break; | |||
2551 | } | |||
2552 | } | |||
2553 | } | |||
2554 | ||||
2555 | unsigned StackSize = CCInfo.getNextStackOffset(); | |||
2556 | // Align stack specially for tail calls. | |||
2557 | if (FuncIsMadeTailCallSafe(CallConv, | |||
2558 | MF.getTarget().Options.GuaranteedTailCallOpt)) | |||
2559 | StackSize = GetAlignedArgumentStackSize(StackSize, DAG); | |||
2560 | ||||
2561 | // If the function takes variable number of arguments, make a frame index for | |||
2562 | // the start of the first vararg value... for expansion of llvm.va_start. We | |||
2563 | // can skip this if there are no va_start calls. | |||
2564 | if (MFI->hasVAStart() && | |||
2565 | (Is64Bit || (CallConv != CallingConv::X86_FastCall && | |||
2566 | CallConv != CallingConv::X86_ThisCall))) { | |||
2567 | FuncInfo->setVarArgsFrameIndex( | |||
2568 | MFI->CreateFixedObject(1, StackSize, true)); | |||
2569 | } | |||
2570 | ||||
2571 | // Figure out if XMM registers are in use. | |||
2572 | assert(!(MF.getTarget().Options.UseSoftFloat &&((!(MF.getTarget().Options.UseSoftFloat && Fn->getAttributes ().hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat )) && "SSE register cannot be used when SSE is disabled!" ) ? static_cast<void> (0) : __assert_fail ("!(MF.getTarget().Options.UseSoftFloat && Fn->getAttributes().hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat)) && \"SSE register cannot be used when SSE is disabled!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2575, __PRETTY_FUNCTION__)) | |||
2573 | Fn->getAttributes().hasAttribute(AttributeSet::FunctionIndex,((!(MF.getTarget().Options.UseSoftFloat && Fn->getAttributes ().hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat )) && "SSE register cannot be used when SSE is disabled!" ) ? static_cast<void> (0) : __assert_fail ("!(MF.getTarget().Options.UseSoftFloat && Fn->getAttributes().hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat)) && \"SSE register cannot be used when SSE is disabled!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2575, __PRETTY_FUNCTION__)) | |||
2574 | Attribute::NoImplicitFloat)) &&((!(MF.getTarget().Options.UseSoftFloat && Fn->getAttributes ().hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat )) && "SSE register cannot be used when SSE is disabled!" ) ? static_cast<void> (0) : __assert_fail ("!(MF.getTarget().Options.UseSoftFloat && Fn->getAttributes().hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat)) && \"SSE register cannot be used when SSE is disabled!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2575, __PRETTY_FUNCTION__)) | |||
2575 | "SSE register cannot be used when SSE is disabled!")((!(MF.getTarget().Options.UseSoftFloat && Fn->getAttributes ().hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat )) && "SSE register cannot be used when SSE is disabled!" ) ? static_cast<void> (0) : __assert_fail ("!(MF.getTarget().Options.UseSoftFloat && Fn->getAttributes().hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat)) && \"SSE register cannot be used when SSE is disabled!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2575, __PRETTY_FUNCTION__)); | |||
2576 | ||||
2577 | // 64-bit calling conventions support varargs and register parameters, so we | |||
2578 | // have to do extra work to spill them in the prologue. | |||
2579 | if (Is64Bit && isVarArg && MFI->hasVAStart()) { | |||
2580 | // Find the first unallocated argument registers. | |||
2581 | ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget); | |||
2582 | ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget); | |||
2583 | unsigned NumIntRegs = | |||
2584 | CCInfo.getFirstUnallocated(ArgGPRs.data(), ArgGPRs.size()); | |||
2585 | unsigned NumXMMRegs = | |||
2586 | CCInfo.getFirstUnallocated(ArgXMMs.data(), ArgXMMs.size()); | |||
2587 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2588, __PRETTY_FUNCTION__)) | |||
2588 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2588, __PRETTY_FUNCTION__)); | |||
2589 | ||||
2590 | // Gather all the live in physical registers. | |||
2591 | SmallVector<SDValue, 6> LiveGPRs; | |||
2592 | SmallVector<SDValue, 8> LiveXMMRegs; | |||
2593 | SDValue ALVal; | |||
2594 | for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) { | |||
2595 | unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass); | |||
2596 | LiveGPRs.push_back( | |||
2597 | DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64)); | |||
2598 | } | |||
2599 | if (!ArgXMMs.empty()) { | |||
2600 | unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass); | |||
2601 | ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8); | |||
2602 | for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) { | |||
2603 | unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass); | |||
2604 | LiveXMMRegs.push_back( | |||
2605 | DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32)); | |||
2606 | } | |||
2607 | } | |||
2608 | ||||
2609 | if (IsWin64) { | |||
2610 | const TargetFrameLowering &TFI = *MF.getSubtarget().getFrameLowering(); | |||
2611 | // Get to the caller-allocated home save location. Add 8 to account | |||
2612 | // for the return address. | |||
2613 | int HomeOffset = TFI.getOffsetOfLocalArea() + 8; | |||
2614 | FuncInfo->setRegSaveFrameIndex( | |||
2615 | MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false)); | |||
2616 | // Fixup to set vararg frame on shadow area (4 x i64). | |||
2617 | if (NumIntRegs < 4) | |||
2618 | FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex()); | |||
2619 | } else { | |||
2620 | // For X86-64, if there are vararg parameters that are passed via | |||
2621 | // registers, then we must store them to their spots on the stack so | |||
2622 | // they may be loaded by deferencing the result of va_next. | |||
2623 | FuncInfo->setVarArgsGPOffset(NumIntRegs * 8); | |||
2624 | FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16); | |||
2625 | FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject( | |||
2626 | ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false)); | |||
2627 | } | |||
2628 | ||||
2629 | // Store the integer parameter registers. | |||
2630 | SmallVector<SDValue, 8> MemOps; | |||
2631 | SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), | |||
2632 | getPointerTy()); | |||
2633 | unsigned Offset = FuncInfo->getVarArgsGPOffset(); | |||
2634 | for (SDValue Val : LiveGPRs) { | |||
2635 | SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN, | |||
2636 | DAG.getIntPtrConstant(Offset)); | |||
2637 | SDValue Store = | |||
2638 | DAG.getStore(Val.getValue(1), dl, Val, FIN, | |||
2639 | MachinePointerInfo::getFixedStack( | |||
2640 | FuncInfo->getRegSaveFrameIndex(), Offset), | |||
2641 | false, false, 0); | |||
2642 | MemOps.push_back(Store); | |||
2643 | Offset += 8; | |||
2644 | } | |||
2645 | ||||
2646 | if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) { | |||
2647 | // Now store the XMM (fp + vector) parameter registers. | |||
2648 | SmallVector<SDValue, 12> SaveXMMOps; | |||
2649 | SaveXMMOps.push_back(Chain); | |||
2650 | SaveXMMOps.push_back(ALVal); | |||
2651 | SaveXMMOps.push_back(DAG.getIntPtrConstant( | |||
2652 | FuncInfo->getRegSaveFrameIndex())); | |||
2653 | SaveXMMOps.push_back(DAG.getIntPtrConstant( | |||
2654 | FuncInfo->getVarArgsFPOffset())); | |||
2655 | SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(), | |||
2656 | LiveXMMRegs.end()); | |||
2657 | MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl, | |||
2658 | MVT::Other, SaveXMMOps)); | |||
2659 | } | |||
2660 | ||||
2661 | if (!MemOps.empty()) | |||
2662 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps); | |||
2663 | } | |||
2664 | ||||
2665 | if (isVarArg && MFI->hasMustTailInVarArgFunc()) { | |||
2666 | // Find the largest legal vector type. | |||
2667 | MVT VecVT = MVT::Other; | |||
2668 | // FIXME: Only some x86_32 calling conventions support AVX512. | |||
2669 | if (Subtarget->hasAVX512() && | |||
2670 | (Is64Bit || (CallConv == CallingConv::X86_VectorCall || | |||
2671 | CallConv == CallingConv::Intel_OCL_BI))) | |||
2672 | VecVT = MVT::v16f32; | |||
2673 | else if (Subtarget->hasAVX()) | |||
2674 | VecVT = MVT::v8f32; | |||
2675 | else if (Subtarget->hasSSE2()) | |||
2676 | VecVT = MVT::v4f32; | |||
2677 | ||||
2678 | // We forward some GPRs and some vector types. | |||
2679 | SmallVector<MVT, 2> RegParmTypes; | |||
2680 | MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32; | |||
2681 | RegParmTypes.push_back(IntVT); | |||
2682 | if (VecVT != MVT::Other) | |||
2683 | RegParmTypes.push_back(VecVT); | |||
2684 | ||||
2685 | // Compute the set of forwarded registers. The rest are scratch. | |||
2686 | SmallVectorImpl<ForwardedRegister> &Forwards = | |||
2687 | FuncInfo->getForwardedMustTailRegParms(); | |||
2688 | CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86); | |||
2689 | ||||
2690 | // Conservatively forward AL on x86_64, since it might be used for varargs. | |||
2691 | if (Is64Bit && !CCInfo.isAllocated(X86::AL)) { | |||
2692 | unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass); | |||
2693 | Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8)); | |||
2694 | } | |||
2695 | ||||
2696 | // Copy all forwards from physical to virtual registers. | |||
2697 | for (ForwardedRegister &F : Forwards) { | |||
2698 | // FIXME: Can we use a less constrained schedule? | |||
2699 | SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT); | |||
2700 | F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT)); | |||
2701 | Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal); | |||
2702 | } | |||
2703 | } | |||
2704 | ||||
2705 | // Some CCs need callee pop. | |||
2706 | if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, | |||
2707 | MF.getTarget().Options.GuaranteedTailCallOpt)) { | |||
2708 | FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything. | |||
2709 | } else { | |||
2710 | FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing. | |||
2711 | // If this is an sret function, the return should pop the hidden pointer. | |||
2712 | if (!Is64Bit && !IsTailCallConvention(CallConv) && | |||
2713 | !Subtarget->getTargetTriple().isOSMSVCRT() && | |||
2714 | argsAreStructReturn(Ins) == StackStructReturn) | |||
2715 | FuncInfo->setBytesToPopOnReturn(4); | |||
2716 | } | |||
2717 | ||||
2718 | if (!Is64Bit) { | |||
2719 | // RegSaveFrameIndex is X86-64 only. | |||
2720 | FuncInfo->setRegSaveFrameIndex(0xAAAAAAA); | |||
2721 | if (CallConv == CallingConv::X86_FastCall || | |||
2722 | CallConv == CallingConv::X86_ThisCall) | |||
2723 | // fastcc functions can't have varargs. | |||
2724 | FuncInfo->setVarArgsFrameIndex(0xAAAAAAA); | |||
2725 | } | |||
2726 | ||||
2727 | FuncInfo->setArgumentStackSize(StackSize); | |||
2728 | ||||
2729 | return Chain; | |||
2730 | } | |||
2731 | ||||
2732 | SDValue | |||
2733 | X86TargetLowering::LowerMemOpCallTo(SDValue Chain, | |||
2734 | SDValue StackPtr, SDValue Arg, | |||
2735 | SDLoc dl, SelectionDAG &DAG, | |||
2736 | const CCValAssign &VA, | |||
2737 | ISD::ArgFlagsTy Flags) const { | |||
2738 | unsigned LocMemOffset = VA.getLocMemOffset(); | |||
2739 | SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset); | |||
2740 | PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff); | |||
2741 | if (Flags.isByVal()) | |||
2742 | return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl); | |||
2743 | ||||
2744 | return DAG.getStore(Chain, dl, Arg, PtrOff, | |||
2745 | MachinePointerInfo::getStack(LocMemOffset), | |||
2746 | false, false, 0); | |||
2747 | } | |||
2748 | ||||
2749 | /// Emit a load of return address if tail call | |||
2750 | /// optimization is performed and it is required. | |||
2751 | SDValue | |||
2752 | X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG, | |||
2753 | SDValue &OutRetAddr, SDValue Chain, | |||
2754 | bool IsTailCall, bool Is64Bit, | |||
2755 | int FPDiff, SDLoc dl) const { | |||
2756 | // Adjust the Return address stack slot. | |||
2757 | EVT VT = getPointerTy(); | |||
2758 | OutRetAddr = getReturnAddressFrameIndex(DAG); | |||
2759 | ||||
2760 | // Load the "old" Return address. | |||
2761 | OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(), | |||
2762 | false, false, false, 0); | |||
2763 | return SDValue(OutRetAddr.getNode(), 1); | |||
2764 | } | |||
2765 | ||||
2766 | /// Emit a store of the return address if tail call | |||
2767 | /// optimization is performed and it is required (FPDiff!=0). | |||
2768 | static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF, | |||
2769 | SDValue Chain, SDValue RetAddrFrIdx, | |||
2770 | EVT PtrVT, unsigned SlotSize, | |||
2771 | int FPDiff, SDLoc dl) { | |||
2772 | // Store the return address to the appropriate stack slot. | |||
2773 | if (!FPDiff) return Chain; | |||
2774 | // Calculate the new stack slot for the return address. | |||
2775 | int NewReturnAddrFI = | |||
2776 | MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize, | |||
2777 | false); | |||
2778 | SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT); | |||
2779 | Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx, | |||
2780 | MachinePointerInfo::getFixedStack(NewReturnAddrFI), | |||
2781 | false, false, 0); | |||
2782 | return Chain; | |||
2783 | } | |||
2784 | ||||
2785 | SDValue | |||
2786 | X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, | |||
2787 | SmallVectorImpl<SDValue> &InVals) const { | |||
2788 | SelectionDAG &DAG = CLI.DAG; | |||
2789 | SDLoc &dl = CLI.DL; | |||
2790 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; | |||
2791 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; | |||
2792 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; | |||
2793 | SDValue Chain = CLI.Chain; | |||
2794 | SDValue Callee = CLI.Callee; | |||
2795 | CallingConv::ID CallConv = CLI.CallConv; | |||
2796 | bool &isTailCall = CLI.IsTailCall; | |||
2797 | bool isVarArg = CLI.IsVarArg; | |||
2798 | ||||
2799 | MachineFunction &MF = DAG.getMachineFunction(); | |||
2800 | bool Is64Bit = Subtarget->is64Bit(); | |||
2801 | bool IsWin64 = Subtarget->isCallingConvWin64(CallConv); | |||
2802 | StructReturnType SR = callIsStructReturn(Outs); | |||
2803 | bool IsSibcall = false; | |||
2804 | X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>(); | |||
2805 | ||||
2806 | if (MF.getTarget().Options.DisableTailCalls) | |||
2807 | isTailCall = false; | |||
2808 | ||||
2809 | bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall(); | |||
2810 | if (IsMustTail) { | |||
2811 | // Force this to be a tail call. The verifier rules are enough to ensure | |||
2812 | // that we can lower this successfully without moving the return address | |||
2813 | // around. | |||
2814 | isTailCall = true; | |||
2815 | } else if (isTailCall) { | |||
2816 | // Check if it's really possible to do a tail call. | |||
2817 | isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, | |||
2818 | isVarArg, SR != NotStructReturn, | |||
2819 | MF.getFunction()->hasStructRetAttr(), CLI.RetTy, | |||
2820 | Outs, OutVals, Ins, DAG); | |||
2821 | ||||
2822 | // Sibcalls are automatically detected tailcalls which do not require | |||
2823 | // ABI changes. | |||
2824 | if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall) | |||
2825 | IsSibcall = true; | |||
2826 | ||||
2827 | if (isTailCall) | |||
2828 | ++NumTailCalls; | |||
2829 | } | |||
2830 | ||||
2831 | assert(!(isVarArg && IsTailCallConvention(CallConv)) &&((!(isVarArg && IsTailCallConvention(CallConv)) && "Var args not supported with calling convention fastcc, ghc or hipe" ) ? static_cast<void> (0) : __assert_fail ("!(isVarArg && IsTailCallConvention(CallConv)) && \"Var args not supported with calling convention fastcc, ghc or hipe\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2832, __PRETTY_FUNCTION__)) | |||
2832 | "Var args not supported with calling convention fastcc, ghc or hipe")((!(isVarArg && IsTailCallConvention(CallConv)) && "Var args not supported with calling convention fastcc, ghc or hipe" ) ? static_cast<void> (0) : __assert_fail ("!(isVarArg && IsTailCallConvention(CallConv)) && \"Var args not supported with calling convention fastcc, ghc or hipe\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2832, __PRETTY_FUNCTION__)); | |||
2833 | ||||
2834 | // Analyze operands of the call, assigning locations to each operand. | |||
2835 | SmallVector<CCValAssign, 16> ArgLocs; | |||
2836 | CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext()); | |||
2837 | ||||
2838 | // Allocate shadow area for Win64 | |||
2839 | if (IsWin64) | |||
2840 | CCInfo.AllocateStack(32, 8); | |||
2841 | ||||
2842 | CCInfo.AnalyzeCallOperands(Outs, CC_X86); | |||
2843 | ||||
2844 | // Get a count of how many bytes are to be pushed on the stack. | |||
2845 | unsigned NumBytes = CCInfo.getNextStackOffset(); | |||
2846 | if (IsSibcall) | |||
2847 | // This is a sibcall. The memory operands are available in caller's | |||
2848 | // own caller's stack. | |||
2849 | NumBytes = 0; | |||
2850 | else if (MF.getTarget().Options.GuaranteedTailCallOpt && | |||
2851 | IsTailCallConvention(CallConv)) | |||
2852 | NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG); | |||
2853 | ||||
2854 | int FPDiff = 0; | |||
2855 | if (isTailCall && !IsSibcall && !IsMustTail) { | |||
2856 | // Lower arguments at fp - stackoffset + fpdiff. | |||
2857 | unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn(); | |||
2858 | ||||
2859 | FPDiff = NumBytesCallerPushed - NumBytes; | |||
2860 | ||||
2861 | // Set the delta of movement of the returnaddr stackslot. | |||
2862 | // But only set if delta is greater than previous delta. | |||
2863 | if (FPDiff < X86Info->getTCReturnAddrDelta()) | |||
2864 | X86Info->setTCReturnAddrDelta(FPDiff); | |||
2865 | } | |||
2866 | ||||
2867 | unsigned NumBytesToPush = NumBytes; | |||
2868 | unsigned NumBytesToPop = NumBytes; | |||
2869 | ||||
2870 | // If we have an inalloca argument, all stack space has already been allocated | |||
2871 | // for us and be right at the top of the stack. We don't support multiple | |||
2872 | // arguments passed in memory when using inalloca. | |||
2873 | if (!Outs.empty() && Outs.back().Flags.isInAlloca()) { | |||
2874 | NumBytesToPush = 0; | |||
2875 | if (!ArgLocs.back().isMemLoc()) | |||
2876 | report_fatal_error("cannot use inalloca attribute on a register " | |||
2877 | "parameter"); | |||
2878 | if (ArgLocs.back().getLocMemOffset() != 0) | |||
2879 | report_fatal_error("any parameter with the inalloca attribute must be " | |||
2880 | "the only memory argument"); | |||
2881 | } | |||
2882 | ||||
2883 | if (!IsSibcall) | |||
2884 | Chain = DAG.getCALLSEQ_START( | |||
2885 | Chain, DAG.getIntPtrConstant(NumBytesToPush, true), dl); | |||
2886 | ||||
2887 | SDValue RetAddrFrIdx; | |||
2888 | // Load return address for tail calls. | |||
2889 | if (isTailCall && FPDiff) | |||
2890 | Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, | |||
2891 | Is64Bit, FPDiff, dl); | |||
2892 | ||||
2893 | SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; | |||
2894 | SmallVector<SDValue, 8> MemOpChains; | |||
2895 | SDValue StackPtr; | |||
2896 | ||||
2897 | // Walk the register/memloc assignments, inserting copies/loads. In the case | |||
2898 | // of tail call optimization arguments are handle later. | |||
2899 | const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>( | |||
2900 | DAG.getSubtarget().getRegisterInfo()); | |||
2901 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { | |||
2902 | // Skip inalloca arguments, they have already been written. | |||
2903 | ISD::ArgFlagsTy Flags = Outs[i].Flags; | |||
2904 | if (Flags.isInAlloca()) | |||
2905 | continue; | |||
2906 | ||||
2907 | CCValAssign &VA = ArgLocs[i]; | |||
2908 | EVT RegVT = VA.getLocVT(); | |||
2909 | SDValue Arg = OutVals[i]; | |||
2910 | bool isByVal = Flags.isByVal(); | |||
2911 | ||||
2912 | // Promote the value if needed. | |||
2913 | switch (VA.getLocInfo()) { | |||
2914 | default: llvm_unreachable("Unknown loc info!")::llvm::llvm_unreachable_internal("Unknown loc info!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2914); | |||
2915 | case CCValAssign::Full: break; | |||
2916 | case CCValAssign::SExt: | |||
2917 | Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg); | |||
2918 | break; | |||
2919 | case CCValAssign::ZExt: | |||
2920 | Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg); | |||
2921 | break; | |||
2922 | case CCValAssign::AExt: | |||
2923 | if (RegVT.is128BitVector()) { | |||
2924 | // Special case: passing MMX values in XMM registers. | |||
2925 | Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg); | |||
2926 | Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg); | |||
2927 | Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg); | |||
2928 | } else | |||
2929 | Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg); | |||
2930 | break; | |||
2931 | case CCValAssign::BCvt: | |||
2932 | Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg); | |||
2933 | break; | |||
2934 | case CCValAssign::Indirect: { | |||
2935 | // Store the argument. | |||
2936 | SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT()); | |||
2937 | int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex(); | |||
2938 | Chain = DAG.getStore(Chain, dl, Arg, SpillSlot, | |||
2939 | MachinePointerInfo::getFixedStack(FI), | |||
2940 | false, false, 0); | |||
2941 | Arg = SpillSlot; | |||
2942 | break; | |||
2943 | } | |||
2944 | } | |||
2945 | ||||
2946 | if (VA.isRegLoc()) { | |||
2947 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); | |||
2948 | if (isVarArg && IsWin64) { | |||
2949 | // Win64 ABI requires argument XMM reg to be copied to the corresponding | |||
2950 | // shadow reg if callee is a varargs function. | |||
2951 | unsigned ShadowReg = 0; | |||
2952 | switch (VA.getLocReg()) { | |||
2953 | case X86::XMM0: ShadowReg = X86::RCX; break; | |||
2954 | case X86::XMM1: ShadowReg = X86::RDX; break; | |||
2955 | case X86::XMM2: ShadowReg = X86::R8; break; | |||
2956 | case X86::XMM3: ShadowReg = X86::R9; break; | |||
2957 | } | |||
2958 | if (ShadowReg) | |||
2959 | RegsToPass.push_back(std::make_pair(ShadowReg, Arg)); | |||
2960 | } | |||
2961 | } else if (!IsSibcall && (!isTailCall || isByVal)) { | |||
2962 | assert(VA.isMemLoc())((VA.isMemLoc()) ? static_cast<void> (0) : __assert_fail ("VA.isMemLoc()", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 2962, __PRETTY_FUNCTION__)); | |||
2963 | if (!StackPtr.getNode()) | |||
2964 | StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(), | |||
2965 | getPointerTy()); | |||
2966 | MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg, | |||
2967 | dl, DAG, VA, Flags)); | |||
2968 | } | |||
2969 | } | |||
2970 | ||||
2971 | if (!MemOpChains.empty()) | |||
2972 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains); | |||
2973 | ||||
2974 | if (Subtarget->isPICStyleGOT()) { | |||
2975 | // ELF / PIC requires GOT in the EBX register before function calls via PLT | |||
2976 | // GOT pointer. | |||
2977 | if (!isTailCall) { | |||
2978 | RegsToPass.push_back(std::make_pair(unsigned(X86::EBX), | |||
2979 | DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy()))); | |||
2980 | } else { | |||
2981 | // If we are tail calling and generating PIC/GOT style code load the | |||
2982 | // address of the callee into ECX. The value in ecx is used as target of | |||
2983 | // the tail jump. This is done to circumvent the ebx/callee-saved problem | |||
2984 | // for tail calls on PIC/GOT architectures. Normally we would just put the | |||
2985 | // address of GOT into ebx and then call target@PLT. But for tail calls | |||
2986 | // ebx would be restored (since ebx is callee saved) before jumping to the | |||
2987 | // target@PLT. | |||
2988 | ||||
2989 | // Note: The actual moving to ECX is done further down. | |||
2990 | GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee); | |||
2991 | if (G && !G->getGlobal()->hasHiddenVisibility() && | |||
2992 | !G->getGlobal()->hasProtectedVisibility()) | |||
2993 | Callee = LowerGlobalAddress(Callee, DAG); | |||
2994 | else if (isa<ExternalSymbolSDNode>(Callee)) | |||
2995 | Callee = LowerExternalSymbol(Callee, DAG); | |||
2996 | } | |||
2997 | } | |||
2998 | ||||
2999 | if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) { | |||
3000 | // From AMD64 ABI document: | |||
3001 | // For calls that may call functions that use varargs or stdargs | |||
3002 | // (prototype-less calls or calls to functions containing ellipsis (...) in | |||
3003 | // the declaration) %al is used as hidden argument to specify the number | |||
3004 | // of SSE registers used. The contents of %al do not need to match exactly | |||
3005 | // the number of registers, but must be an ubound on the number of SSE | |||
3006 | // registers used and is in the range 0 - 8 inclusive. | |||
3007 | ||||
3008 | // Count the number of XMM registers allocated. | |||
3009 | static const MCPhysReg XMMArgRegs[] = { | |||
3010 | X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, | |||
3011 | X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7 | |||
3012 | }; | |||
3013 | unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8); | |||
3014 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 3015, __PRETTY_FUNCTION__)) | |||
3015 | && "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 3015, __PRETTY_FUNCTION__)); | |||
3016 | ||||
3017 | RegsToPass.push_back(std::make_pair(unsigned(X86::AL), | |||
3018 | DAG.getConstant(NumXMMRegs, MVT::i8))); | |||
3019 | } | |||
3020 | ||||
3021 | if (isVarArg && IsMustTail) { | |||
3022 | const auto &Forwards = X86Info->getForwardedMustTailRegParms(); | |||
3023 | for (const auto &F : Forwards) { | |||
3024 | SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT); | |||
3025 | RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val)); | |||
3026 | } | |||
3027 | } | |||
3028 | ||||
3029 | // For tail calls lower the arguments to the 'real' stack slots. Sibcalls | |||
3030 | // don't need this because the eligibility check rejects calls that require | |||
3031 | // shuffling arguments passed in memory. | |||
3032 | if (!IsSibcall && isTailCall) { | |||
3033 | // Force all the incoming stack arguments to be loaded from the stack | |||
3034 | // before any new outgoing arguments are stored to the stack, because the | |||
3035 | // outgoing stack slots may alias the incoming argument stack slots, and | |||
3036 | // the alias isn't otherwise explicit. This is slightly more conservative | |||
3037 | // than necessary, because it means that each store effectively depends | |||
3038 | // on every argument instead of just those arguments it would clobber. | |||
3039 | SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain); | |||
3040 | ||||
3041 | SmallVector<SDValue, 8> MemOpChains2; | |||
3042 | SDValue FIN; | |||
3043 | int FI = 0; | |||
3044 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { | |||
3045 | CCValAssign &VA = ArgLocs[i]; | |||
3046 | if (VA.isRegLoc()) | |||
3047 | continue; | |||
3048 | assert(VA.isMemLoc())((VA.isMemLoc()) ? static_cast<void> (0) : __assert_fail ("VA.isMemLoc()", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 3048, __PRETTY_FUNCTION__)); | |||
3049 | SDValue Arg = OutVals[i]; | |||
3050 | ISD::ArgFlagsTy Flags = Outs[i].Flags; | |||
3051 | // Skip inalloca arguments. They don't require any work. | |||
3052 | if (Flags.isInAlloca()) | |||
3053 | continue; | |||
3054 | // Create frame index. | |||
3055 | int32_t Offset = VA.getLocMemOffset()+FPDiff; | |||
3056 | uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8; | |||
3057 | FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true); | |||
3058 | FIN = DAG.getFrameIndex(FI, getPointerTy()); | |||
3059 | ||||
3060 | if (Flags.isByVal()) { | |||
3061 | // Copy relative to framepointer. | |||
3062 | SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset()); | |||
3063 | if (!StackPtr.getNode()) | |||
3064 | StackPtr = DAG.getCopyFromReg(Chain, dl, | |||
3065 | RegInfo->getStackRegister(), | |||
3066 | getPointerTy()); | |||
3067 | Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source); | |||
3068 | ||||
3069 | MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, | |||
3070 | ArgChain, | |||
3071 | Flags, DAG, dl)); | |||
3072 | } else { | |||
3073 | // Store relative to framepointer. | |||
3074 | MemOpChains2.push_back( | |||
3075 | DAG.getStore(ArgChain, dl, Arg, FIN, | |||
3076 | MachinePointerInfo::getFixedStack(FI), | |||
3077 | false, false, 0)); | |||
3078 | } | |||
3079 | } | |||
3080 | ||||
3081 | if (!MemOpChains2.empty()) | |||
3082 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2); | |||
3083 | ||||
3084 | // Store the return address to the appropriate stack slot. | |||
3085 | Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, | |||
3086 | getPointerTy(), RegInfo->getSlotSize(), | |||
3087 | FPDiff, dl); | |||
3088 | } | |||
3089 | ||||
3090 | // Build a sequence of copy-to-reg nodes chained together with token chain | |||
3091 | // and flag operands which copy the outgoing args into registers. | |||
3092 | SDValue InFlag; | |||
3093 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { | |||
3094 | Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, | |||
3095 | RegsToPass[i].second, InFlag); | |||
3096 | InFlag = Chain.getValue(1); | |||
3097 | } | |||
3098 | ||||
3099 | if (DAG.getTarget().getCodeModel() == CodeModel::Large) { | |||
3100 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 3100, __PRETTY_FUNCTION__)); | |||
3101 | // In the 64-bit large code model, we have to make all calls | |||
3102 | // through a register, since the call instruction's 32-bit | |||
3103 | // pc-relative offset may not be large enough to hold the whole | |||
3104 | // address. | |||
3105 | } else if (Callee->getOpcode() == ISD::GlobalAddress) { | |||
3106 | // If the callee is a GlobalAddress node (quite common, every direct call | |||
3107 | // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack | |||
3108 | // it. | |||
3109 | GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee); | |||
3110 | ||||
3111 | // We should use extra load for direct calls to dllimported functions in | |||
3112 | // non-JIT mode. | |||
3113 | const GlobalValue *GV = G->getGlobal(); | |||
3114 | if (!GV->hasDLLImportStorageClass()) { | |||
3115 | unsigned char OpFlags = 0; | |||
3116 | bool ExtraLoad = false; | |||
3117 | unsigned WrapperKind = ISD::DELETED_NODE; | |||
3118 | ||||
3119 | // On ELF targets, in both X86-64 and X86-32 mode, direct calls to | |||
3120 | // external symbols most go through the PLT in PIC mode. If the symbol | |||
3121 | // has hidden or protected visibility, or if it is static or local, then | |||
3122 | // we don't need to use the PLT - we can directly call it. | |||
3123 | if (Subtarget->isTargetELF() && | |||
3124 | DAG.getTarget().getRelocationModel() == Reloc::PIC_ && | |||
3125 | GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) { | |||
3126 | OpFlags = X86II::MO_PLT; | |||
3127 | } else if (Subtarget->isPICStyleStubAny() && | |||
3128 | (GV->isDeclaration() || GV->isWeakForLinker()) && | |||
3129 | (!Subtarget->getTargetTriple().isMacOSX() || | |||
3130 | Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) { | |||
3131 | // PC-relative references to external symbols should go through $stub, | |||
3132 | // unless we're building with the leopard linker or later, which | |||
3133 | // automatically synthesizes these stubs. | |||
3134 | OpFlags = X86II::MO_DARWIN_STUB; | |||
3135 | } else if (Subtarget->isPICStyleRIPRel() && | |||
3136 | isa<Function>(GV) && | |||
3137 | cast<Function>(GV)->getAttributes(). | |||
3138 | hasAttribute(AttributeSet::FunctionIndex, | |||
3139 | Attribute::NonLazyBind)) { | |||
3140 | // If the function is marked as non-lazy, generate an indirect call | |||
3141 | // which loads from the GOT directly. This avoids runtime overhead | |||
3142 | // at the cost of eager binding (and one extra byte of encoding). | |||
3143 | OpFlags = X86II::MO_GOTPCREL; | |||
3144 | WrapperKind = X86ISD::WrapperRIP; | |||
3145 | ExtraLoad = true; | |||
3146 | } | |||
3147 | ||||
3148 | Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), | |||
3149 | G->getOffset(), OpFlags); | |||
3150 | ||||
3151 | // Add a wrapper if needed. | |||
3152 | if (WrapperKind != ISD::DELETED_NODE) | |||
3153 | Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee); | |||
3154 | // Add extra indirection if needed. | |||
3155 | if (ExtraLoad) | |||
3156 | Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee, | |||
3157 | MachinePointerInfo::getGOT(), | |||
3158 | false, false, false, 0); | |||
3159 | } | |||
3160 | } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { | |||
3161 | unsigned char OpFlags = 0; | |||
3162 | ||||
3163 | // On ELF targets, in either X86-64 or X86-32 mode, direct calls to | |||
3164 | // external symbols should go through the PLT. | |||
3165 | if (Subtarget->isTargetELF() && | |||
3166 | DAG.getTarget().getRelocationModel() == Reloc::PIC_) { | |||
3167 | OpFlags = X86II::MO_PLT; | |||
3168 | } else if (Subtarget->isPICStyleStubAny() && | |||
3169 | (!Subtarget->getTargetTriple().isMacOSX() || | |||
3170 | Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) { | |||
3171 | // PC-relative references to external symbols should go through $stub, | |||
3172 | // unless we're building with the leopard linker or later, which | |||
3173 | // automatically synthesizes these stubs. | |||
3174 | OpFlags = X86II::MO_DARWIN_STUB; | |||
3175 | } | |||
3176 | ||||
3177 | Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(), | |||
3178 | OpFlags); | |||
3179 | } else if (Subtarget->isTarget64BitILP32() && Callee->getValueType(0) == MVT::i32) { | |||
3180 | // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI | |||
3181 | Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee); | |||
3182 | } | |||
3183 | ||||
3184 | // Returns a chain & a flag for retval copy to use. | |||
3185 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
3186 | SmallVector<SDValue, 8> Ops; | |||
3187 | ||||
3188 | if (!IsSibcall && isTailCall) { | |||
3189 | Chain = DAG.getCALLSEQ_END(Chain, | |||
3190 | DAG.getIntPtrConstant(NumBytesToPop, true), | |||
3191 | DAG.getIntPtrConstant(0, true), InFlag, dl); | |||
3192 | InFlag = Chain.getValue(1); | |||
3193 | } | |||
3194 | ||||
3195 | Ops.push_back(Chain); | |||
3196 | Ops.push_back(Callee); | |||
3197 | ||||
3198 | if (isTailCall) | |||
3199 | Ops.push_back(DAG.getConstant(FPDiff, MVT::i32)); | |||
3200 | ||||
3201 | // Add argument registers to the end of the list so that they are known live | |||
3202 | // into the call. | |||
3203 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) | |||
3204 | Ops.push_back(DAG.getRegister(RegsToPass[i].first, | |||
3205 | RegsToPass[i].second.getValueType())); | |||
3206 | ||||
3207 | // Add a register mask operand representing the call-preserved registers. | |||
3208 | const TargetRegisterInfo *TRI = DAG.getSubtarget().getRegisterInfo(); | |||
3209 | const uint32_t *Mask = TRI->getCallPreservedMask(CallConv); | |||
3210 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 3210, __PRETTY_FUNCTION__)); | |||
3211 | Ops.push_back(DAG.getRegisterMask(Mask)); | |||
3212 | ||||
3213 | if (InFlag.getNode()) | |||
3214 | Ops.push_back(InFlag); | |||
3215 | ||||
3216 | if (isTailCall) { | |||
3217 | // We used to do: | |||
3218 | //// If this is the first return lowered for this function, add the regs | |||
3219 | //// to the liveout set for the function. | |||
3220 | // This isn't right, although it's probably harmless on x86; liveouts | |||
3221 | // should be computed from returns not tail calls. Consider a void | |||
3222 | // function making a tail call to a function returning int. | |||
3223 | return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops); | |||
3224 | } | |||
3225 | ||||
3226 | Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops); | |||
3227 | InFlag = Chain.getValue(1); | |||
3228 | ||||
3229 | // Create the CALLSEQ_END node. | |||
3230 | unsigned NumBytesForCalleeToPop; | |||
3231 | if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, | |||
3232 | DAG.getTarget().Options.GuaranteedTailCallOpt)) | |||
3233 | NumBytesForCalleeToPop = NumBytes; // Callee pops everything | |||
3234 | else if (!Is64Bit && !IsTailCallConvention(CallConv) && | |||
3235 | !Subtarget->getTargetTriple().isOSMSVCRT() && | |||
3236 | SR == StackStructReturn) | |||
3237 | // If this is a call to a struct-return function, the callee | |||
3238 | // pops the hidden struct pointer, so we have to push it back. | |||
3239 | // This is common for Darwin/X86, Linux & Mingw32 targets. | |||
3240 | // For MSVC Win32 targets, the caller pops the hidden struct pointer. | |||
3241 | NumBytesForCalleeToPop = 4; | |||
3242 | else | |||
3243 | NumBytesForCalleeToPop = 0; // Callee pops nothing. | |||
3244 | ||||
3245 | // Returns a flag for retval copy to use. | |||
3246 | if (!IsSibcall) { | |||
3247 | Chain = DAG.getCALLSEQ_END(Chain, | |||
3248 | DAG.getIntPtrConstant(NumBytesToPop, true), | |||
3249 | DAG.getIntPtrConstant(NumBytesForCalleeToPop, | |||
3250 | true), | |||
3251 | InFlag, dl); | |||
3252 | InFlag = Chain.getValue(1); | |||
3253 | } | |||
3254 | ||||
3255 | // Handle result values, copying them out of physregs into vregs that we | |||
3256 | // return. | |||
3257 | return LowerCallResult(Chain, InFlag, CallConv, isVarArg, | |||
3258 | Ins, dl, DAG, InVals); | |||
3259 | } | |||
3260 | ||||
3261 | //===----------------------------------------------------------------------===// | |||
3262 | // Fast Calling Convention (tail call) implementation | |||
3263 | //===----------------------------------------------------------------------===// | |||
3264 | ||||
3265 | // Like std call, callee cleans arguments, convention except that ECX is | |||
3266 | // reserved for storing the tail called function address. Only 2 registers are | |||
3267 | // free for argument passing (inreg). Tail call optimization is performed | |||
3268 | // provided: | |||
3269 | // * tailcallopt is enabled | |||
3270 | // * caller/callee are fastcc | |||
3271 | // On X86_64 architecture with GOT-style position independent code only local | |||
3272 | // (within module) calls are supported at the moment. | |||
3273 | // To keep the stack aligned according to platform abi the function | |||
3274 | // GetAlignedArgumentStackSize ensures that argument delta is always multiples | |||
3275 | // of stack alignment. (Dynamic linkers need this - darwin's dyld for example) | |||
3276 | // If a tail called function callee has more arguments than the caller the | |||
3277 | // caller needs to make sure that there is room to move the RETADDR to. This is | |||
3278 | // achieved by reserving an area the size of the argument delta right after the | |||
3279 | // original RETADDR, but before the saved framepointer or the spilled registers | |||
3280 | // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4) | |||
3281 | // stack layout: | |||
3282 | // arg1 | |||
3283 | // arg2 | |||
3284 | // RETADDR | |||
3285 | // [ new RETADDR | |||
3286 | // move area ] | |||
3287 | // (possible EBP) | |||
3288 | // ESI | |||
3289 | // EDI | |||
3290 | // local1 .. | |||
3291 | ||||
3292 | /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned | |||
3293 | /// for a 16 byte align requirement. | |||
3294 | unsigned | |||
3295 | X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize, | |||
3296 | SelectionDAG& DAG) const { | |||
3297 | MachineFunction &MF = DAG.getMachineFunction(); | |||
3298 | const TargetMachine &TM = MF.getTarget(); | |||
3299 | const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>( | |||
3300 | TM.getSubtargetImpl()->getRegisterInfo()); | |||
3301 | const TargetFrameLowering &TFI = *TM.getSubtargetImpl()->getFrameLowering(); | |||
3302 | unsigned StackAlignment = TFI.getStackAlignment(); | |||
3303 | uint64_t AlignMask = StackAlignment - 1; | |||
3304 | int64_t Offset = StackSize; | |||
3305 | unsigned SlotSize = RegInfo->getSlotSize(); | |||
3306 | if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) { | |||
3307 | // Number smaller than 12 so just add the difference. | |||
3308 | Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask)); | |||
3309 | } else { | |||
3310 | // Mask out lower bits, add stackalignment once plus the 12 bytes. | |||
3311 | Offset = ((~AlignMask) & Offset) + StackAlignment + | |||
3312 | (StackAlignment-SlotSize); | |||
3313 | } | |||
3314 | return Offset; | |||
3315 | } | |||
3316 | ||||
3317 | /// MatchingStackOffset - Return true if the given stack call argument is | |||
3318 | /// already available in the same position (relatively) of the caller's | |||
3319 | /// incoming argument stack. | |||
3320 | static | |||
3321 | bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags, | |||
3322 | MachineFrameInfo *MFI, const MachineRegisterInfo *MRI, | |||
3323 | const X86InstrInfo *TII) { | |||
3324 | unsigned Bytes = Arg.getValueType().getSizeInBits() / 8; | |||
3325 | int FI = INT_MAX2147483647; | |||
3326 | if (Arg.getOpcode() == ISD::CopyFromReg) { | |||
3327 | unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg(); | |||
3328 | if (!TargetRegisterInfo::isVirtualRegister(VR)) | |||
3329 | return false; | |||
3330 | MachineInstr *Def = MRI->getVRegDef(VR); | |||
3331 | if (!Def) | |||
3332 | return false; | |||
3333 | if (!Flags.isByVal()) { | |||
3334 | if (!TII->isLoadFromStackSlot(Def, FI)) | |||
3335 | return false; | |||
3336 | } else { | |||
3337 | unsigned Opcode = Def->getOpcode(); | |||
3338 | if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r || | |||
3339 | Opcode == X86::LEA64_32r) && | |||
3340 | Def->getOperand(1).isFI()) { | |||
3341 | FI = Def->getOperand(1).getIndex(); | |||
3342 | Bytes = Flags.getByValSize(); | |||
3343 | } else | |||
3344 | return false; | |||
3345 | } | |||
3346 | } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) { | |||
3347 | if (Flags.isByVal()) | |||
3348 | // ByVal argument is passed in as a pointer but it's now being | |||
3349 | // dereferenced. e.g. | |||
3350 | // define @foo(%struct.X* %A) { | |||
3351 | // tail call @bar(%struct.X* byval %A) | |||
3352 | // } | |||
3353 | return false; | |||
3354 | SDValue Ptr = Ld->getBasePtr(); | |||
3355 | FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr); | |||
3356 | if (!FINode) | |||
3357 | return false; | |||
3358 | FI = FINode->getIndex(); | |||
3359 | } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) { | |||
3360 | FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg); | |||
3361 | FI = FINode->getIndex(); | |||
3362 | Bytes = Flags.getByValSize(); | |||
3363 | } else | |||
3364 | return false; | |||
3365 | ||||
3366 | assert(FI != INT_MAX)((FI != 2147483647) ? static_cast<void> (0) : __assert_fail ("FI != 2147483647", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 3366, __PRETTY_FUNCTION__)); | |||
3367 | if (!MFI->isFixedObjectIndex(FI)) | |||
3368 | return false; | |||
3369 | return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI); | |||
3370 | } | |||
3371 | ||||
3372 | /// IsEligibleForTailCallOptimization - Check whether the call is eligible | |||
3373 | /// for tail call optimization. Targets which want to do tail call | |||
3374 | /// optimization should implement this function. | |||
3375 | bool | |||
3376 | X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee, | |||
3377 | CallingConv::ID CalleeCC, | |||
3378 | bool isVarArg, | |||
3379 | bool isCalleeStructRet, | |||
3380 | bool isCallerStructRet, | |||
3381 | Type *RetTy, | |||
3382 | const SmallVectorImpl<ISD::OutputArg> &Outs, | |||
3383 | const SmallVectorImpl<SDValue> &OutVals, | |||
3384 | const SmallVectorImpl<ISD::InputArg> &Ins, | |||
3385 | SelectionDAG &DAG) const { | |||
3386 | if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC)) | |||
3387 | return false; | |||
3388 | ||||
3389 | // If -tailcallopt is specified, make fastcc functions tail-callable. | |||
3390 | const MachineFunction &MF = DAG.getMachineFunction(); | |||
3391 | const Function *CallerF = MF.getFunction(); | |||
3392 | ||||
3393 | // If the function return type is x86_fp80 and the callee return type is not, | |||
3394 | // then the FP_EXTEND of the call result is not a nop. It's not safe to | |||
3395 | // perform a tailcall optimization here. | |||
3396 | if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty()) | |||
3397 | return false; | |||
3398 | ||||
3399 | CallingConv::ID CallerCC = CallerF->getCallingConv(); | |||
3400 | bool CCMatch = CallerCC == CalleeCC; | |||
3401 | bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC); | |||
3402 | bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC); | |||
3403 | ||||
3404 | if (DAG.getTarget().Options.GuaranteedTailCallOpt) { | |||
3405 | if (IsTailCallConvention(CalleeCC) && CCMatch) | |||
3406 | return true; | |||
3407 | return false; | |||
3408 | } | |||
3409 | ||||
3410 | // Look for obvious safe cases to perform tail call optimization that do not | |||
3411 | // require ABI changes. This is what gcc calls sibcall. | |||
3412 | ||||
3413 | // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to | |||
3414 | // emit a special epilogue. | |||
3415 | const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>( | |||
3416 | DAG.getSubtarget().getRegisterInfo()); | |||
3417 | if (RegInfo->needsStackRealignment(MF)) | |||
3418 | return false; | |||
3419 | ||||
3420 | // Also avoid sibcall optimization if either caller or callee uses struct | |||
3421 | // return semantics. | |||
3422 | if (isCalleeStructRet || isCallerStructRet) | |||
3423 | return false; | |||
3424 | ||||
3425 | // An stdcall/thiscall caller is expected to clean up its arguments; the | |||
3426 | // callee isn't going to do that. | |||
3427 | // FIXME: this is more restrictive than needed. We could produce a tailcall | |||
3428 | // when the stack adjustment matches. For example, with a thiscall that takes | |||
3429 | // only one argument. | |||
3430 | if (!CCMatch && (CallerCC == CallingConv::X86_StdCall || | |||
3431 | CallerCC == CallingConv::X86_ThisCall)) | |||
3432 | return false; | |||
3433 | ||||
3434 | // Do not sibcall optimize vararg calls unless all arguments are passed via | |||
3435 | // registers. | |||
3436 | if (isVarArg && !Outs.empty()) { | |||
3437 | ||||
3438 | // Optimizing for varargs on Win64 is unlikely to be safe without | |||
3439 | // additional testing. | |||
3440 | if (IsCalleeWin64 || IsCallerWin64) | |||
3441 | return false; | |||
3442 | ||||
3443 | SmallVector<CCValAssign, 16> ArgLocs; | |||
3444 | CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs, | |||
3445 | *DAG.getContext()); | |||
3446 | ||||
3447 | CCInfo.AnalyzeCallOperands(Outs, CC_X86); | |||
3448 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) | |||
3449 | if (!ArgLocs[i].isRegLoc()) | |||
3450 | return false; | |||
3451 | } | |||
3452 | ||||
3453 | // If the call result is in ST0 / ST1, it needs to be popped off the x87 | |||
3454 | // stack. Therefore, if it's not used by the call it is not safe to optimize | |||
3455 | // this into a sibcall. | |||
3456 | bool Unused = false; | |||
3457 | for (unsigned i = 0, e = Ins.size(); i != e; ++i) { | |||
3458 | if (!Ins[i].Used) { | |||
3459 | Unused = true; | |||
3460 | break; | |||
3461 | } | |||
3462 | } | |||
3463 | if (Unused) { | |||
3464 | SmallVector<CCValAssign, 16> RVLocs; | |||
3465 | CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs, | |||
3466 | *DAG.getContext()); | |||
3467 | CCInfo.AnalyzeCallResult(Ins, RetCC_X86); | |||
3468 | for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) { | |||
3469 | CCValAssign &VA = RVLocs[i]; | |||
3470 | if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) | |||
3471 | return false; | |||
3472 | } | |||
3473 | } | |||
3474 | ||||
3475 | // If the calling conventions do not match, then we'd better make sure the | |||
3476 | // results are returned in the same way as what the caller expects. | |||
3477 | if (!CCMatch) { | |||
3478 | SmallVector<CCValAssign, 16> RVLocs1; | |||
3479 | CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1, | |||
3480 | *DAG.getContext()); | |||
3481 | CCInfo1.AnalyzeCallResult(Ins, RetCC_X86); | |||
3482 | ||||
3483 | SmallVector<CCValAssign, 16> RVLocs2; | |||
3484 | CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2, | |||
3485 | *DAG.getContext()); | |||
3486 | CCInfo2.AnalyzeCallResult(Ins, RetCC_X86); | |||
3487 | ||||
3488 | if (RVLocs1.size() != RVLocs2.size()) | |||
3489 | return false; | |||
3490 | for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) { | |||
3491 | if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc()) | |||
3492 | return false; | |||
3493 | if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo()) | |||
3494 | return false; | |||
3495 | if (RVLocs1[i].isRegLoc()) { | |||
3496 | if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg()) | |||
3497 | return false; | |||
3498 | } else { | |||
3499 | if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset()) | |||
3500 | return false; | |||
3501 | } | |||
3502 | } | |||
3503 | } | |||
3504 | ||||
3505 | // If the callee takes no arguments then go on to check the results of the | |||
3506 | // call. | |||
3507 | if (!Outs.empty()) { | |||
3508 | // Check if stack adjustment is needed. For now, do not do this if any | |||
3509 | // argument is passed on the stack. | |||
3510 | SmallVector<CCValAssign, 16> ArgLocs; | |||
3511 | CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs, | |||
3512 | *DAG.getContext()); | |||
3513 | ||||
3514 | // Allocate shadow area for Win64 | |||
3515 | if (IsCalleeWin64) | |||
3516 | CCInfo.AllocateStack(32, 8); | |||
3517 | ||||
3518 | CCInfo.AnalyzeCallOperands(Outs, CC_X86); | |||
3519 | if (CCInfo.getNextStackOffset()) { | |||
3520 | MachineFunction &MF = DAG.getMachineFunction(); | |||
3521 | if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn()) | |||
3522 | return false; | |||
3523 | ||||
3524 | // Check if the arguments are already laid out in the right way as | |||
3525 | // the caller's fixed stack objects. | |||
3526 | MachineFrameInfo *MFI = MF.getFrameInfo(); | |||
3527 | const MachineRegisterInfo *MRI = &MF.getRegInfo(); | |||
3528 | const X86InstrInfo *TII = | |||
3529 | static_cast<const X86InstrInfo *>(DAG.getSubtarget().getInstrInfo()); | |||
3530 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { | |||
3531 | CCValAssign &VA = ArgLocs[i]; | |||
3532 | SDValue Arg = OutVals[i]; | |||
3533 | ISD::ArgFlagsTy Flags = Outs[i].Flags; | |||
3534 | if (VA.getLocInfo() == CCValAssign::Indirect) | |||
3535 | return false; | |||
3536 | if (!VA.isRegLoc()) { | |||
3537 | if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags, | |||
3538 | MFI, MRI, TII)) | |||
3539 | return false; | |||
3540 | } | |||
3541 | } | |||
3542 | } | |||
3543 | ||||
3544 | // If the tailcall address may be in a register, then make sure it's | |||
3545 | // possible to register allocate for it. In 32-bit, the call address can | |||
3546 | // only target EAX, EDX, or ECX since the tail call must be scheduled after | |||
3547 | // callee-saved registers are restored. These happen to be the same | |||
3548 | // registers used to pass 'inreg' arguments so watch out for those. | |||
3549 | if (!Subtarget->is64Bit() && | |||
3550 | ((!isa<GlobalAddressSDNode>(Callee) && | |||
3551 | !isa<ExternalSymbolSDNode>(Callee)) || | |||
3552 | DAG.getTarget().getRelocationModel() == Reloc::PIC_)) { | |||
3553 | unsigned NumInRegs = 0; | |||
3554 | // In PIC we need an extra register to formulate the address computation | |||
3555 | // for the callee. | |||
3556 | unsigned MaxInRegs = | |||
3557 | (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3; | |||
3558 | ||||
3559 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { | |||
3560 | CCValAssign &VA = ArgLocs[i]; | |||
3561 | if (!VA.isRegLoc()) | |||
3562 | continue; | |||
3563 | unsigned Reg = VA.getLocReg(); | |||
3564 | switch (Reg) { | |||
3565 | default: break; | |||
3566 | case X86::EAX: case X86::EDX: case X86::ECX: | |||
3567 | if (++NumInRegs == MaxInRegs) | |||
3568 | return false; | |||
3569 | break; | |||
3570 | } | |||
3571 | } | |||
3572 | } | |||
3573 | } | |||
3574 | ||||
3575 | return true; | |||
3576 | } | |||
3577 | ||||
3578 | FastISel * | |||
3579 | X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo, | |||
3580 | const TargetLibraryInfo *libInfo) const { | |||
3581 | return X86::createFastISel(funcInfo, libInfo); | |||
3582 | } | |||
3583 | ||||
3584 | //===----------------------------------------------------------------------===// | |||
3585 | // Other Lowering Hooks | |||
3586 | //===----------------------------------------------------------------------===// | |||
3587 | ||||
3588 | static bool MayFoldLoad(SDValue Op) { | |||
3589 | return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode()); | |||
3590 | } | |||
3591 | ||||
3592 | static bool MayFoldIntoStore(SDValue Op) { | |||
3593 | return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin()); | |||
3594 | } | |||
3595 | ||||
3596 | static bool isTargetShuffle(unsigned Opcode) { | |||
3597 | switch(Opcode) { | |||
3598 | default: return false; | |||
3599 | case X86ISD::BLENDI: | |||
3600 | case X86ISD::PSHUFB: | |||
3601 | case X86ISD::PSHUFD: | |||
3602 | case X86ISD::PSHUFHW: | |||
3603 | case X86ISD::PSHUFLW: | |||
3604 | case X86ISD::SHUFP: | |||
3605 | case X86ISD::PALIGNR: | |||
3606 | case X86ISD::MOVLHPS: | |||
3607 | case X86ISD::MOVLHPD: | |||
3608 | case X86ISD::MOVHLPS: | |||
3609 | case X86ISD::MOVLPS: | |||
3610 | case X86ISD::MOVLPD: | |||
3611 | case X86ISD::MOVSHDUP: | |||
3612 | case X86ISD::MOVSLDUP: | |||
3613 | case X86ISD::MOVDDUP: | |||
3614 | case X86ISD::MOVSS: | |||
3615 | case X86ISD::MOVSD: | |||
3616 | case X86ISD::UNPCKL: | |||
3617 | case X86ISD::UNPCKH: | |||
3618 | case X86ISD::VPERMILPI: | |||
3619 | case X86ISD::VPERM2X128: | |||
3620 | case X86ISD::VPERMI: | |||
3621 | return true; | |||
3622 | } | |||
3623 | } | |||
3624 | ||||
3625 | static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT, | |||
3626 | SDValue V1, SelectionDAG &DAG) { | |||
3627 | switch(Opc) { | |||
3628 | default: llvm_unreachable("Unknown x86 shuffle node")::llvm::llvm_unreachable_internal("Unknown x86 shuffle node", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 3628); | |||
3629 | case X86ISD::MOVSHDUP: | |||
3630 | case X86ISD::MOVSLDUP: | |||
3631 | case X86ISD::MOVDDUP: | |||
3632 | return DAG.getNode(Opc, dl, VT, V1); | |||
3633 | } | |||
3634 | } | |||
3635 | ||||
3636 | static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT, | |||
3637 | SDValue V1, unsigned TargetMask, | |||
3638 | SelectionDAG &DAG) { | |||
3639 | switch(Opc) { | |||
3640 | default: llvm_unreachable("Unknown x86 shuffle node")::llvm::llvm_unreachable_internal("Unknown x86 shuffle node", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 3640); | |||
3641 | case X86ISD::PSHUFD: | |||
3642 | case X86ISD::PSHUFHW: | |||
3643 | case X86ISD::PSHUFLW: | |||
3644 | case X86ISD::VPERMILPI: | |||
3645 | case X86ISD::VPERMI: | |||
3646 | return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8)); | |||
3647 | } | |||
3648 | } | |||
3649 | ||||
3650 | static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT, | |||
3651 | SDValue V1, SDValue V2, unsigned TargetMask, | |||
3652 | SelectionDAG &DAG) { | |||
3653 | switch(Opc) { | |||
3654 | default: llvm_unreachable("Unknown x86 shuffle node")::llvm::llvm_unreachable_internal("Unknown x86 shuffle node", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 3654); | |||
3655 | case X86ISD::PALIGNR: | |||
3656 | case X86ISD::VALIGN: | |||
3657 | case X86ISD::SHUFP: | |||
3658 | case X86ISD::VPERM2X128: | |||
3659 | return DAG.getNode(Opc, dl, VT, V1, V2, | |||
3660 | DAG.getConstant(TargetMask, MVT::i8)); | |||
3661 | } | |||
3662 | } | |||
3663 | ||||
3664 | static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT, | |||
3665 | SDValue V1, SDValue V2, SelectionDAG &DAG) { | |||
3666 | switch(Opc) { | |||
3667 | default: llvm_unreachable("Unknown x86 shuffle node")::llvm::llvm_unreachable_internal("Unknown x86 shuffle node", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 3667); | |||
3668 | case X86ISD::MOVLHPS: | |||
3669 | case X86ISD::MOVLHPD: | |||
3670 | case X86ISD::MOVHLPS: | |||
3671 | case X86ISD::MOVLPS: | |||
3672 | case X86ISD::MOVLPD: | |||
3673 | case X86ISD::MOVSS: | |||
3674 | case X86ISD::MOVSD: | |||
3675 | case X86ISD::UNPCKL: | |||
3676 | case X86ISD::UNPCKH: | |||
3677 | return DAG.getNode(Opc, dl, VT, V1, V2); | |||
3678 | } | |||
3679 | } | |||
3680 | ||||
3681 | SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const { | |||
3682 | MachineFunction &MF = DAG.getMachineFunction(); | |||
3683 | const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>( | |||
3684 | DAG.getSubtarget().getRegisterInfo()); | |||
3685 | X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); | |||
3686 | int ReturnAddrIndex = FuncInfo->getRAIndex(); | |||
3687 | ||||
3688 | if (ReturnAddrIndex == 0) { | |||
3689 | // Set up a frame object for the return address. | |||
3690 | unsigned SlotSize = RegInfo->getSlotSize(); | |||
3691 | ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, | |||
3692 | -(int64_t)SlotSize, | |||
3693 | false); | |||
3694 | FuncInfo->setRAIndex(ReturnAddrIndex); | |||
3695 | } | |||
3696 | ||||
3697 | return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy()); | |||
3698 | } | |||
3699 | ||||
3700 | bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M, | |||
3701 | bool hasSymbolicDisplacement) { | |||
3702 | // Offset should fit into 32 bit immediate field. | |||
3703 | if (!isInt<32>(Offset)) | |||
3704 | return false; | |||
3705 | ||||
3706 | // If we don't have a symbolic displacement - we don't have any extra | |||
3707 | // restrictions. | |||
3708 | if (!hasSymbolicDisplacement) | |||
3709 | return true; | |||
3710 | ||||
3711 | // FIXME: Some tweaks might be needed for medium code model. | |||
3712 | if (M != CodeModel::Small && M != CodeModel::Kernel) | |||
3713 | return false; | |||
3714 | ||||
3715 | // For small code model we assume that latest object is 16MB before end of 31 | |||
3716 | // bits boundary. We may also accept pretty large negative constants knowing | |||
3717 | // that all objects are in the positive half of address space. | |||
3718 | if (M == CodeModel::Small && Offset < 16*1024*1024) | |||
3719 | return true; | |||
3720 | ||||
3721 | // For kernel code model we know that all object resist in the negative half | |||
3722 | // of 32bits address space. We may not accept negative offsets, since they may | |||
3723 | // be just off and we may accept pretty large positive ones. | |||
3724 | if (M == CodeModel::Kernel && Offset >= 0) | |||
3725 | return true; | |||
3726 | ||||
3727 | return false; | |||
3728 | } | |||
3729 | ||||
3730 | /// isCalleePop - Determines whether the callee is required to pop its | |||
3731 | /// own arguments. Callee pop is necessary to support tail calls. | |||
3732 | bool X86::isCalleePop(CallingConv::ID CallingConv, | |||
3733 | bool is64Bit, bool IsVarArg, bool TailCallOpt) { | |||
3734 | switch (CallingConv) { | |||
3735 | default: | |||
3736 | return false; | |||
3737 | case CallingConv::X86_StdCall: | |||
3738 | case CallingConv::X86_FastCall: | |||
3739 | case CallingConv::X86_ThisCall: | |||
3740 | return !is64Bit; | |||
3741 | case CallingConv::Fast: | |||
3742 | case CallingConv::GHC: | |||
3743 | case CallingConv::HiPE: | |||
3744 | if (IsVarArg) | |||
3745 | return false; | |||
3746 | return TailCallOpt; | |||
3747 | } | |||
3748 | } | |||
3749 | ||||
3750 | /// \brief Return true if the condition is an unsigned comparison operation. | |||
3751 | static bool isX86CCUnsigned(unsigned X86CC) { | |||
3752 | switch (X86CC) { | |||
3753 | default: llvm_unreachable("Invalid integer condition!")::llvm::llvm_unreachable_internal("Invalid integer condition!" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 3753); | |||
3754 | case X86::COND_E: return true; | |||
3755 | case X86::COND_G: return false; | |||
3756 | case X86::COND_GE: return false; | |||
3757 | case X86::COND_L: return false; | |||
3758 | case X86::COND_LE: return false; | |||
3759 | case X86::COND_NE: return true; | |||
3760 | case X86::COND_B: return true; | |||
3761 | case X86::COND_A: return true; | |||
3762 | case X86::COND_BE: return true; | |||
3763 | case X86::COND_AE: return true; | |||
3764 | } | |||
3765 | llvm_unreachable("covered switch fell through?!")::llvm::llvm_unreachable_internal("covered switch fell through?!" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 3765); | |||
3766 | } | |||
3767 | ||||
3768 | /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86 | |||
3769 | /// specific condition code, returning the condition code and the LHS/RHS of the | |||
3770 | /// comparison to make. | |||
3771 | static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP, | |||
3772 | SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) { | |||
3773 | if (!isFP) { | |||
3774 | if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) { | |||
3775 | if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) { | |||
3776 | // X > -1 -> X == 0, jump !sign. | |||
3777 | RHS = DAG.getConstant(0, RHS.getValueType()); | |||
3778 | return X86::COND_NS; | |||
3779 | } | |||
3780 | if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) { | |||
3781 | // X < 0 -> X == 0, jump on sign. | |||
3782 | return X86::COND_S; | |||
3783 | } | |||
3784 | if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) { | |||
3785 | // X < 1 -> X <= 0 | |||
3786 | RHS = DAG.getConstant(0, RHS.getValueType()); | |||
3787 | return X86::COND_LE; | |||
3788 | } | |||
3789 | } | |||
3790 | ||||
3791 | switch (SetCCOpcode) { | |||
3792 | default: llvm_unreachable("Invalid integer condition!")::llvm::llvm_unreachable_internal("Invalid integer condition!" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 3792); | |||
3793 | case ISD::SETEQ: return X86::COND_E; | |||
3794 | case ISD::SETGT: return X86::COND_G; | |||
3795 | case ISD::SETGE: return X86::COND_GE; | |||
3796 | case ISD::SETLT: return X86::COND_L; | |||
3797 | case ISD::SETLE: return X86::COND_LE; | |||
3798 | case ISD::SETNE: return X86::COND_NE; | |||
3799 | case ISD::SETULT: return X86::COND_B; | |||
3800 | case ISD::SETUGT: return X86::COND_A; | |||
3801 | case ISD::SETULE: return X86::COND_BE; | |||
3802 | case ISD::SETUGE: return X86::COND_AE; | |||
3803 | } | |||
3804 | } | |||
3805 | ||||
3806 | // First determine if it is required or is profitable to flip the operands. | |||
3807 | ||||
3808 | // If LHS is a foldable load, but RHS is not, flip the condition. | |||
3809 | if (ISD::isNON_EXTLoad(LHS.getNode()) && | |||
3810 | !ISD::isNON_EXTLoad(RHS.getNode())) { | |||
3811 | SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode); | |||
3812 | std::swap(LHS, RHS); | |||
3813 | } | |||
3814 | ||||
3815 | switch (SetCCOpcode) { | |||
3816 | default: break; | |||
3817 | case ISD::SETOLT: | |||
3818 | case ISD::SETOLE: | |||
3819 | case ISD::SETUGT: | |||
3820 | case ISD::SETUGE: | |||
3821 | std::swap(LHS, RHS); | |||
3822 | break; | |||
3823 | } | |||
3824 | ||||
3825 | // On a floating point condition, the flags are set as follows: | |||
3826 | // ZF PF CF op | |||
3827 | // 0 | 0 | 0 | X > Y | |||
3828 | // 0 | 0 | 1 | X < Y | |||
3829 | // 1 | 0 | 0 | X == Y | |||
3830 | // 1 | 1 | 1 | unordered | |||
3831 | switch (SetCCOpcode) { | |||
3832 | default: llvm_unreachable("Condcode should be pre-legalized away")::llvm::llvm_unreachable_internal("Condcode should be pre-legalized away" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 3832); | |||
3833 | case ISD::SETUEQ: | |||
3834 | case ISD::SETEQ: return X86::COND_E; | |||
3835 | case ISD::SETOLT: // flipped | |||
3836 | case ISD::SETOGT: | |||
3837 | case ISD::SETGT: return X86::COND_A; | |||
3838 | case ISD::SETOLE: // flipped | |||
3839 | case ISD::SETOGE: | |||
3840 | case ISD::SETGE: return X86::COND_AE; | |||
3841 | case ISD::SETUGT: // flipped | |||
3842 | case ISD::SETULT: | |||
3843 | case ISD::SETLT: return X86::COND_B; | |||
3844 | case ISD::SETUGE: // flipped | |||
3845 | case ISD::SETULE: | |||
3846 | case ISD::SETLE: return X86::COND_BE; | |||
3847 | case ISD::SETONE: | |||
3848 | case ISD::SETNE: return X86::COND_NE; | |||
3849 | case ISD::SETUO: return X86::COND_P; | |||
3850 | case ISD::SETO: return X86::COND_NP; | |||
3851 | case ISD::SETOEQ: | |||
3852 | case ISD::SETUNE: return X86::COND_INVALID; | |||
3853 | } | |||
3854 | } | |||
3855 | ||||
3856 | /// hasFPCMov - is there a floating point cmov for the specific X86 condition | |||
3857 | /// code. Current x86 isa includes the following FP cmov instructions: | |||
3858 | /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu. | |||
3859 | static bool hasFPCMov(unsigned X86CC) { | |||
3860 | switch (X86CC) { | |||
3861 | default: | |||
3862 | return false; | |||
3863 | case X86::COND_B: | |||
3864 | case X86::COND_BE: | |||
3865 | case X86::COND_E: | |||
3866 | case X86::COND_P: | |||
3867 | case X86::COND_A: | |||
3868 | case X86::COND_AE: | |||
3869 | case X86::COND_NE: | |||
3870 | case X86::COND_NP: | |||
3871 | return true; | |||
3872 | } | |||
3873 | } | |||
3874 | ||||
3875 | /// isFPImmLegal - Returns true if the target can instruction select the | |||
3876 | /// specified FP immediate natively. If false, the legalizer will | |||
3877 | /// materialize the FP immediate as a load from a constant pool. | |||
3878 | bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const { | |||
3879 | for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) { | |||
3880 | if (Imm.bitwiseIsEqual(LegalFPImmediates[i])) | |||
3881 | return true; | |||
3882 | } | |||
3883 | return false; | |||
3884 | } | |||
3885 | ||||
3886 | bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load, | |||
3887 | ISD::LoadExtType ExtTy, | |||
3888 | EVT NewVT) const { | |||
3889 | // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF | |||
3890 | // relocation target a movq or addq instruction: don't let the load shrink. | |||
3891 | SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr(); | |||
3892 | if (BasePtr.getOpcode() == X86ISD::WrapperRIP) | |||
3893 | if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0))) | |||
3894 | return GA->getTargetFlags() != X86II::MO_GOTTPOFF; | |||
3895 | return true; | |||
3896 | } | |||
3897 | ||||
3898 | /// \brief Returns true if it is beneficial to convert a load of a constant | |||
3899 | /// to just the constant itself. | |||
3900 | bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm, | |||
3901 | Type *Ty) const { | |||
3902 | assert(Ty->isIntegerTy())((Ty->isIntegerTy()) ? static_cast<void> (0) : __assert_fail ("Ty->isIntegerTy()", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 3902, __PRETTY_FUNCTION__)); | |||
3903 | ||||
3904 | unsigned BitSize = Ty->getPrimitiveSizeInBits(); | |||
3905 | if (BitSize == 0 || BitSize > 64) | |||
3906 | return false; | |||
3907 | return true; | |||
3908 | } | |||
3909 | ||||
3910 | bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT, | |||
3911 | unsigned Index) const { | |||
3912 | if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT)) | |||
3913 | return false; | |||
3914 | ||||
3915 | return (Index == 0 || Index == ResVT.getVectorNumElements()); | |||
3916 | } | |||
3917 | ||||
3918 | bool X86TargetLowering::isCheapToSpeculateCttz() const { | |||
3919 | // Speculate cttz only if we can directly use TZCNT. | |||
3920 | return Subtarget->hasBMI(); | |||
3921 | } | |||
3922 | ||||
3923 | bool X86TargetLowering::isCheapToSpeculateCtlz() const { | |||
3924 | // Speculate ctlz only if we can directly use LZCNT. | |||
3925 | return Subtarget->hasLZCNT(); | |||
3926 | } | |||
3927 | ||||
3928 | /// isUndefOrInRange - Return true if Val is undef or if its value falls within | |||
3929 | /// the specified range (L, H]. | |||
3930 | static bool isUndefOrInRange(int Val, int Low, int Hi) { | |||
3931 | return (Val < 0) || (Val >= Low && Val < Hi); | |||
3932 | } | |||
3933 | ||||
3934 | /// isUndefOrEqual - Val is either less than zero (undef) or equal to the | |||
3935 | /// specified value. | |||
3936 | static bool isUndefOrEqual(int Val, int CmpVal) { | |||
3937 | return (Val < 0 || Val == CmpVal); | |||
3938 | } | |||
3939 | ||||
3940 | /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning | |||
3941 | /// from position Pos and ending in Pos+Size, falls within the specified | |||
3942 | /// sequential range (Low, Low+Size]. or is undef. | |||
3943 | static bool isSequentialOrUndefInRange(ArrayRef<int> Mask, | |||
3944 | unsigned Pos, unsigned Size, int Low) { | |||
3945 | for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low) | |||
3946 | if (!isUndefOrEqual(Mask[i], Low)) | |||
3947 | return false; | |||
3948 | return true; | |||
3949 | } | |||
3950 | ||||
3951 | /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that | |||
3952 | /// is suitable for input to PSHUFD. That is, it doesn't reference the other | |||
3953 | /// operand - by default will match for first operand. | |||
3954 | static bool isPSHUFDMask(ArrayRef<int> Mask, MVT VT, | |||
3955 | bool TestSecondOperand = false) { | |||
3956 | if (VT != MVT::v4f32 && VT != MVT::v4i32 && | |||
3957 | VT != MVT::v2f64 && VT != MVT::v2i64) | |||
3958 | return false; | |||
3959 | ||||
3960 | unsigned NumElems = VT.getVectorNumElements(); | |||
3961 | unsigned Lo = TestSecondOperand ? NumElems : 0; | |||
3962 | unsigned Hi = Lo + NumElems; | |||
3963 | ||||
3964 | for (unsigned i = 0; i < NumElems; ++i) | |||
3965 | if (!isUndefOrInRange(Mask[i], (int)Lo, (int)Hi)) | |||
3966 | return false; | |||
3967 | ||||
3968 | return true; | |||
3969 | } | |||
3970 | ||||
3971 | /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that | |||
3972 | /// is suitable for input to PSHUFHW. | |||
3973 | static bool isPSHUFHWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) { | |||
3974 | if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16)) | |||
3975 | return false; | |||
3976 | ||||
3977 | // Lower quadword copied in order or undef. | |||
3978 | if (!isSequentialOrUndefInRange(Mask, 0, 4, 0)) | |||
3979 | return false; | |||
3980 | ||||
3981 | // Upper quadword shuffled. | |||
3982 | for (unsigned i = 4; i != 8; ++i) | |||
3983 | if (!isUndefOrInRange(Mask[i], 4, 8)) | |||
3984 | return false; | |||
3985 | ||||
3986 | if (VT == MVT::v16i16) { | |||
3987 | // Lower quadword copied in order or undef. | |||
3988 | if (!isSequentialOrUndefInRange(Mask, 8, 4, 8)) | |||
3989 | return false; | |||
3990 | ||||
3991 | // Upper quadword shuffled. | |||
3992 | for (unsigned i = 12; i != 16; ++i) | |||
3993 | if (!isUndefOrInRange(Mask[i], 12, 16)) | |||
3994 | return false; | |||
3995 | } | |||
3996 | ||||
3997 | return true; | |||
3998 | } | |||
3999 | ||||
4000 | /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that | |||
4001 | /// is suitable for input to PSHUFLW. | |||
4002 | static bool isPSHUFLWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) { | |||
4003 | if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16)) | |||
4004 | return false; | |||
4005 | ||||
4006 | // Upper quadword copied in order. | |||
4007 | if (!isSequentialOrUndefInRange(Mask, 4, 4, 4)) | |||
4008 | return false; | |||
4009 | ||||
4010 | // Lower quadword shuffled. | |||
4011 | for (unsigned i = 0; i != 4; ++i) | |||
4012 | if (!isUndefOrInRange(Mask[i], 0, 4)) | |||
4013 | return false; | |||
4014 | ||||
4015 | if (VT == MVT::v16i16) { | |||
4016 | // Upper quadword copied in order. | |||
4017 | if (!isSequentialOrUndefInRange(Mask, 12, 4, 12)) | |||
4018 | return false; | |||
4019 | ||||
4020 | // Lower quadword shuffled. | |||
4021 | for (unsigned i = 8; i != 12; ++i) | |||
4022 | if (!isUndefOrInRange(Mask[i], 8, 12)) | |||
4023 | return false; | |||
4024 | } | |||
4025 | ||||
4026 | return true; | |||
4027 | } | |||
4028 | ||||
4029 | /// \brief Return true if the mask specifies a shuffle of elements that is | |||
4030 | /// suitable for input to intralane (palignr) or interlane (valign) vector | |||
4031 | /// right-shift. | |||
4032 | static bool isAlignrMask(ArrayRef<int> Mask, MVT VT, bool InterLane) { | |||
4033 | unsigned NumElts = VT.getVectorNumElements(); | |||
4034 | unsigned NumLanes = InterLane ? 1: VT.getSizeInBits()/128; | |||
4035 | unsigned NumLaneElts = NumElts/NumLanes; | |||
4036 | ||||
4037 | // Do not handle 64-bit element shuffles with palignr. | |||
4038 | if (NumLaneElts == 2) | |||
4039 | return false; | |||
4040 | ||||
4041 | for (unsigned l = 0; l != NumElts; l+=NumLaneElts) { | |||
4042 | unsigned i; | |||
4043 | for (i = 0; i != NumLaneElts; ++i) { | |||
4044 | if (Mask[i+l] >= 0) | |||
4045 | break; | |||
4046 | } | |||
4047 | ||||
4048 | // Lane is all undef, go to next lane | |||
4049 | if (i == NumLaneElts) | |||
4050 | continue; | |||
4051 | ||||
4052 | int Start = Mask[i+l]; | |||
4053 | ||||
4054 | // Make sure its in this lane in one of the sources | |||
4055 | if (!isUndefOrInRange(Start, l, l+NumLaneElts) && | |||
4056 | !isUndefOrInRange(Start, l+NumElts, l+NumElts+NumLaneElts)) | |||
4057 | return false; | |||
4058 | ||||
4059 | // If not lane 0, then we must match lane 0 | |||
4060 | if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Start, Mask[i]+l)) | |||
4061 | return false; | |||
4062 | ||||
4063 | // Correct second source to be contiguous with first source | |||
4064 | if (Start >= (int)NumElts) | |||
4065 | Start -= NumElts - NumLaneElts; | |||
4066 | ||||
4067 | // Make sure we're shifting in the right direction. | |||
4068 | if (Start <= (int)(i+l)) | |||
4069 | return false; | |||
4070 | ||||
4071 | Start -= i; | |||
4072 | ||||
4073 | // Check the rest of the elements to see if they are consecutive. | |||
4074 | for (++i; i != NumLaneElts; ++i) { | |||
4075 | int Idx = Mask[i+l]; | |||
4076 | ||||
4077 | // Make sure its in this lane | |||
4078 | if (!isUndefOrInRange(Idx, l, l+NumLaneElts) && | |||
4079 | !isUndefOrInRange(Idx, l+NumElts, l+NumElts+NumLaneElts)) | |||
4080 | return false; | |||
4081 | ||||
4082 | // If not lane 0, then we must match lane 0 | |||
4083 | if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Idx, Mask[i]+l)) | |||
4084 | return false; | |||
4085 | ||||
4086 | if (Idx >= (int)NumElts) | |||
4087 | Idx -= NumElts - NumLaneElts; | |||
4088 | ||||
4089 | if (!isUndefOrEqual(Idx, Start+i)) | |||
4090 | return false; | |||
4091 | ||||
4092 | } | |||
4093 | } | |||
4094 | ||||
4095 | return true; | |||
4096 | } | |||
4097 | ||||
4098 | /// \brief Return true if the node specifies a shuffle of elements that is | |||
4099 | /// suitable for input to PALIGNR. | |||
4100 | static bool isPALIGNRMask(ArrayRef<int> Mask, MVT VT, | |||
4101 | const X86Subtarget *Subtarget) { | |||
4102 | if ((VT.is128BitVector() && !Subtarget->hasSSSE3()) || | |||
4103 | (VT.is256BitVector() && !Subtarget->hasInt256()) || | |||
4104 | VT.is512BitVector()) | |||
4105 | // FIXME: Add AVX512BW. | |||
4106 | return false; | |||
4107 | ||||
4108 | return isAlignrMask(Mask, VT, false); | |||
4109 | } | |||
4110 | ||||
4111 | /// \brief Return true if the node specifies a shuffle of elements that is | |||
4112 | /// suitable for input to VALIGN. | |||
4113 | static bool isVALIGNMask(ArrayRef<int> Mask, MVT VT, | |||
4114 | const X86Subtarget *Subtarget) { | |||
4115 | // FIXME: Add AVX512VL. | |||
4116 | if (!VT.is512BitVector() || !Subtarget->hasAVX512()) | |||
4117 | return false; | |||
4118 | return isAlignrMask(Mask, VT, true); | |||
4119 | } | |||
4120 | ||||
4121 | /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming | |||
4122 | /// the two vector operands have swapped position. | |||
4123 | static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, | |||
4124 | unsigned NumElems) { | |||
4125 | for (unsigned i = 0; i != NumElems; ++i) { | |||
4126 | int idx = Mask[i]; | |||
4127 | if (idx < 0) | |||
4128 | continue; | |||
4129 | else if (idx < (int)NumElems) | |||
4130 | Mask[i] = idx + NumElems; | |||
4131 | else | |||
4132 | Mask[i] = idx - NumElems; | |||
4133 | } | |||
4134 | } | |||
4135 | ||||
4136 | /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand | |||
4137 | /// specifies a shuffle of elements that is suitable for input to 128/256-bit | |||
4138 | /// SHUFPS and SHUFPD. If Commuted is true, then it checks for sources to be | |||
4139 | /// reverse of what x86 shuffles want. | |||
4140 | static bool isSHUFPMask(ArrayRef<int> Mask, MVT VT, bool Commuted = false) { | |||
4141 | ||||
4142 | unsigned NumElems = VT.getVectorNumElements(); | |||
4143 | unsigned NumLanes = VT.getSizeInBits()/128; | |||
4144 | unsigned NumLaneElems = NumElems/NumLanes; | |||
4145 | ||||
4146 | if (NumLaneElems != 2 && NumLaneElems != 4) | |||
4147 | return false; | |||
4148 | ||||
4149 | unsigned EltSize = VT.getVectorElementType().getSizeInBits(); | |||
4150 | bool symetricMaskRequired = | |||
4151 | (VT.getSizeInBits() >= 256) && (EltSize == 32); | |||
4152 | ||||
4153 | // VSHUFPSY divides the resulting vector into 4 chunks. | |||
4154 | // The sources are also splitted into 4 chunks, and each destination | |||
4155 | // chunk must come from a different source chunk. | |||
4156 | // | |||
4157 | // SRC1 => X7 X6 X5 X4 X3 X2 X1 X0 | |||
4158 | // SRC2 => Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y9 | |||
4159 | // | |||
4160 | // DST => Y7..Y4, Y7..Y4, X7..X4, X7..X4, | |||
4161 | // Y3..Y0, Y3..Y0, X3..X0, X3..X0 | |||
4162 | // | |||
4163 | // VSHUFPDY divides the resulting vector into 4 chunks. | |||
4164 | // The sources are also splitted into 4 chunks, and each destination | |||
4165 | // chunk must come from a different source chunk. | |||
4166 | // | |||
4167 | // SRC1 => X3 X2 X1 X0 | |||
4168 | // SRC2 => Y3 Y2 Y1 Y0 | |||
4169 | // | |||
4170 | // DST => Y3..Y2, X3..X2, Y1..Y0, X1..X0 | |||
4171 | // | |||
4172 | SmallVector<int, 4> MaskVal(NumLaneElems, -1); | |||
4173 | unsigned HalfLaneElems = NumLaneElems/2; | |||
4174 | for (unsigned l = 0; l != NumElems; l += NumLaneElems) { | |||
4175 | for (unsigned i = 0; i != NumLaneElems; ++i) { | |||
4176 | int Idx = Mask[i+l]; | |||
4177 | unsigned RngStart = l + ((Commuted == (i<HalfLaneElems)) ? NumElems : 0); | |||
4178 | if (!isUndefOrInRange(Idx, RngStart, RngStart+NumLaneElems)) | |||
4179 | return false; | |||
4180 | // For VSHUFPSY, the mask of the second half must be the same as the | |||
4181 | // first but with the appropriate offsets. This works in the same way as | |||
4182 | // VPERMILPS works with masks. | |||
4183 | if (!symetricMaskRequired || Idx < 0) | |||
4184 | continue; | |||
4185 | if (MaskVal[i] < 0) { | |||
4186 | MaskVal[i] = Idx - l; | |||
4187 | continue; | |||
4188 | } | |||
4189 | if ((signed)(Idx - l) != MaskVal[i]) | |||
4190 | return false; | |||
4191 | } | |||
4192 | } | |||
4193 | ||||
4194 | return true; | |||
4195 | } | |||
4196 | ||||
4197 | /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand | |||
4198 | /// specifies a shuffle of elements that is suitable for input to MOVHLPS. | |||
4199 | static bool isMOVHLPSMask(ArrayRef<int> Mask, MVT VT) { | |||
4200 | if (!VT.is128BitVector()) | |||
4201 | return false; | |||
4202 | ||||
4203 | unsigned NumElems = VT.getVectorNumElements(); | |||
4204 | ||||
4205 | if (NumElems != 4) | |||
4206 | return false; | |||
4207 | ||||
4208 | // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3 | |||
4209 | return isUndefOrEqual(Mask[0], 6) && | |||
4210 | isUndefOrEqual(Mask[1], 7) && | |||
4211 | isUndefOrEqual(Mask[2], 2) && | |||
4212 | isUndefOrEqual(Mask[3], 3); | |||
4213 | } | |||
4214 | ||||
4215 | /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form | |||
4216 | /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef, | |||
4217 | /// <2, 3, 2, 3> | |||
4218 | static bool isMOVHLPS_v_undef_Mask(ArrayRef<int> Mask, MVT VT) { | |||
4219 | if (!VT.is128BitVector()) | |||
4220 | return false; | |||
4221 | ||||
4222 | unsigned NumElems = VT.getVectorNumElements(); | |||
4223 | ||||
4224 | if (NumElems != 4) | |||
4225 | return false; | |||
4226 | ||||
4227 | return isUndefOrEqual(Mask[0], 2) && | |||
4228 | isUndefOrEqual(Mask[1], 3) && | |||
4229 | isUndefOrEqual(Mask[2], 2) && | |||
4230 | isUndefOrEqual(Mask[3], 3); | |||
4231 | } | |||
4232 | ||||
4233 | /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand | |||
4234 | /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}. | |||
4235 | static bool isMOVLPMask(ArrayRef<int> Mask, MVT VT) { | |||
4236 | if (!VT.is128BitVector()) | |||
4237 | return false; | |||
4238 | ||||
4239 | unsigned NumElems = VT.getVectorNumElements(); | |||
4240 | ||||
4241 | if (NumElems != 2 && NumElems != 4) | |||
4242 | return false; | |||
4243 | ||||
4244 | for (unsigned i = 0, e = NumElems/2; i != e; ++i) | |||
4245 | if (!isUndefOrEqual(Mask[i], i + NumElems)) | |||
4246 | return false; | |||
4247 | ||||
4248 | for (unsigned i = NumElems/2, e = NumElems; i != e; ++i) | |||
4249 | if (!isUndefOrEqual(Mask[i], i)) | |||
4250 | return false; | |||
4251 | ||||
4252 | return true; | |||
4253 | } | |||
4254 | ||||
4255 | /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand | |||
4256 | /// specifies a shuffle of elements that is suitable for input to MOVLHPS. | |||
4257 | static bool isMOVLHPSMask(ArrayRef<int> Mask, MVT VT) { | |||
4258 | if (!VT.is128BitVector()) | |||
4259 | return false; | |||
4260 | ||||
4261 | unsigned NumElems = VT.getVectorNumElements(); | |||
4262 | ||||
4263 | if (NumElems != 2 && NumElems != 4) | |||
4264 | return false; | |||
4265 | ||||
4266 | for (unsigned i = 0, e = NumElems/2; i != e; ++i) | |||
4267 | if (!isUndefOrEqual(Mask[i], i)) | |||
4268 | return false; | |||
4269 | ||||
4270 | for (unsigned i = 0, e = NumElems/2; i != e; ++i) | |||
4271 | if (!isUndefOrEqual(Mask[i + e], i + NumElems)) | |||
4272 | return false; | |||
4273 | ||||
4274 | return true; | |||
4275 | } | |||
4276 | ||||
4277 | /// isINSERTPSMask - Return true if the specified VECTOR_SHUFFLE operand | |||
4278 | /// specifies a shuffle of elements that is suitable for input to INSERTPS. | |||
4279 | /// i. e: If all but one element come from the same vector. | |||
4280 | static bool isINSERTPSMask(ArrayRef<int> Mask, MVT VT) { | |||
4281 | // TODO: Deal with AVX's VINSERTPS | |||
4282 | if (!VT.is128BitVector() || (VT != MVT::v4f32 && VT != MVT::v4i32)) | |||
4283 | return false; | |||
4284 | ||||
4285 | unsigned CorrectPosV1 = 0; | |||
4286 | unsigned CorrectPosV2 = 0; | |||
4287 | for (int i = 0, e = (int)VT.getVectorNumElements(); i != e; ++i) { | |||
4288 | if (Mask[i] == -1) { | |||
4289 | ++CorrectPosV1; | |||
4290 | ++CorrectPosV2; | |||
4291 | continue; | |||
4292 | } | |||
4293 | ||||
4294 | if (Mask[i] == i) | |||
4295 | ++CorrectPosV1; | |||
4296 | else if (Mask[i] == i + 4) | |||
4297 | ++CorrectPosV2; | |||
4298 | } | |||
4299 | ||||
4300 | if (CorrectPosV1 == 3 || CorrectPosV2 == 3) | |||
4301 | // We have 3 elements (undefs count as elements from any vector) from one | |||
4302 | // vector, and one from another. | |||
4303 | return true; | |||
4304 | ||||
4305 | return false; | |||
4306 | } | |||
4307 | ||||
4308 | // | |||
4309 | // Some special combinations that can be optimized. | |||
4310 | // | |||
4311 | static | |||
4312 | SDValue Compact8x32ShuffleNode(ShuffleVectorSDNode *SVOp, | |||
4313 | SelectionDAG &DAG) { | |||
4314 | MVT VT = SVOp->getSimpleValueType(0); | |||
4315 | SDLoc dl(SVOp); | |||
4316 | ||||
4317 | if (VT != MVT::v8i32 && VT != MVT::v8f32) | |||
4318 | return SDValue(); | |||
4319 | ||||
4320 | ArrayRef<int> Mask = SVOp->getMask(); | |||
4321 | ||||
4322 | // These are the special masks that may be optimized. | |||
4323 | static const int MaskToOptimizeEven[] = {0, 8, 2, 10, 4, 12, 6, 14}; | |||
4324 | static const int MaskToOptimizeOdd[] = {1, 9, 3, 11, 5, 13, 7, 15}; | |||
4325 | bool MatchEvenMask = true; | |||
4326 | bool MatchOddMask = true; | |||
4327 | for (int i=0; i<8; ++i) { | |||
4328 | if (!isUndefOrEqual(Mask[i], MaskToOptimizeEven[i])) | |||
4329 | MatchEvenMask = false; | |||
4330 | if (!isUndefOrEqual(Mask[i], MaskToOptimizeOdd[i])) | |||
4331 | MatchOddMask = false; | |||
4332 | } | |||
4333 | ||||
4334 | if (!MatchEvenMask && !MatchOddMask) | |||
4335 | return SDValue(); | |||
4336 | ||||
4337 | SDValue UndefNode = DAG.getNode(ISD::UNDEF, dl, VT); | |||
4338 | ||||
4339 | SDValue Op0 = SVOp->getOperand(0); | |||
4340 | SDValue Op1 = SVOp->getOperand(1); | |||
4341 | ||||
4342 | if (MatchEvenMask) { | |||
4343 | // Shift the second operand right to 32 bits. | |||
4344 | static const int ShiftRightMask[] = {-1, 0, -1, 2, -1, 4, -1, 6 }; | |||
4345 | Op1 = DAG.getVectorShuffle(VT, dl, Op1, UndefNode, ShiftRightMask); | |||
4346 | } else { | |||
4347 | // Shift the first operand left to 32 bits. | |||
4348 | static const int ShiftLeftMask[] = {1, -1, 3, -1, 5, -1, 7, -1 }; | |||
4349 | Op0 = DAG.getVectorShuffle(VT, dl, Op0, UndefNode, ShiftLeftMask); | |||
4350 | } | |||
4351 | static const int BlendMask[] = {0, 9, 2, 11, 4, 13, 6, 15}; | |||
4352 | return DAG.getVectorShuffle(VT, dl, Op0, Op1, BlendMask); | |||
4353 | } | |||
4354 | ||||
4355 | /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand | |||
4356 | /// specifies a shuffle of elements that is suitable for input to UNPCKL. | |||
4357 | static bool isUNPCKLMask(ArrayRef<int> Mask, MVT VT, | |||
4358 | bool HasInt256, bool V2IsSplat = false) { | |||
4359 | ||||
4360 | assert(VT.getSizeInBits() >= 128 &&((VT.getSizeInBits() >= 128 && "Unsupported vector type for unpckl" ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() >= 128 && \"Unsupported vector type for unpckl\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4361, __PRETTY_FUNCTION__)) | |||
4361 | "Unsupported vector type for unpckl")((VT.getSizeInBits() >= 128 && "Unsupported vector type for unpckl" ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() >= 128 && \"Unsupported vector type for unpckl\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4361, __PRETTY_FUNCTION__)); | |||
4362 | ||||
4363 | unsigned NumElts = VT.getVectorNumElements(); | |||
4364 | if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 && | |||
4365 | (!HasInt256 || (NumElts != 16 && NumElts != 32))) | |||
4366 | return false; | |||
4367 | ||||
4368 | assert((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) &&(((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) && "Unsupported vector type for unpckh") ? static_cast<void> (0) : __assert_fail ("(!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) && \"Unsupported vector type for unpckh\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4369, __PRETTY_FUNCTION__)) | |||
4369 | "Unsupported vector type for unpckh")(((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) && "Unsupported vector type for unpckh") ? static_cast<void> (0) : __assert_fail ("(!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) && \"Unsupported vector type for unpckh\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4369, __PRETTY_FUNCTION__)); | |||
4370 | ||||
4371 | // AVX defines UNPCK* to operate independently on 128-bit lanes. | |||
4372 | unsigned NumLanes = VT.getSizeInBits()/128; | |||
4373 | unsigned NumLaneElts = NumElts/NumLanes; | |||
4374 | ||||
4375 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { | |||
4376 | for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) { | |||
4377 | int BitI = Mask[l+i]; | |||
4378 | int BitI1 = Mask[l+i+1]; | |||
4379 | if (!isUndefOrEqual(BitI, j)) | |||
4380 | return false; | |||
4381 | if (V2IsSplat) { | |||
4382 | if (!isUndefOrEqual(BitI1, NumElts)) | |||
4383 | return false; | |||
4384 | } else { | |||
4385 | if (!isUndefOrEqual(BitI1, j + NumElts)) | |||
4386 | return false; | |||
4387 | } | |||
4388 | } | |||
4389 | } | |||
4390 | ||||
4391 | return true; | |||
4392 | } | |||
4393 | ||||
4394 | /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand | |||
4395 | /// specifies a shuffle of elements that is suitable for input to UNPCKH. | |||
4396 | static bool isUNPCKHMask(ArrayRef<int> Mask, MVT VT, | |||
4397 | bool HasInt256, bool V2IsSplat = false) { | |||
4398 | assert(VT.getSizeInBits() >= 128 &&((VT.getSizeInBits() >= 128 && "Unsupported vector type for unpckh" ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() >= 128 && \"Unsupported vector type for unpckh\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4399, __PRETTY_FUNCTION__)) | |||
4399 | "Unsupported vector type for unpckh")((VT.getSizeInBits() >= 128 && "Unsupported vector type for unpckh" ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() >= 128 && \"Unsupported vector type for unpckh\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4399, __PRETTY_FUNCTION__)); | |||
4400 | ||||
4401 | unsigned NumElts = VT.getVectorNumElements(); | |||
4402 | if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 && | |||
4403 | (!HasInt256 || (NumElts != 16 && NumElts != 32))) | |||
4404 | return false; | |||
4405 | ||||
4406 | assert((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) &&(((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) && "Unsupported vector type for unpckh") ? static_cast<void> (0) : __assert_fail ("(!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) && \"Unsupported vector type for unpckh\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4407, __PRETTY_FUNCTION__)) | |||
4407 | "Unsupported vector type for unpckh")(((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) && "Unsupported vector type for unpckh") ? static_cast<void> (0) : __assert_fail ("(!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) && \"Unsupported vector type for unpckh\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4407, __PRETTY_FUNCTION__)); | |||
4408 | ||||
4409 | // AVX defines UNPCK* to operate independently on 128-bit lanes. | |||
4410 | unsigned NumLanes = VT.getSizeInBits()/128; | |||
4411 | unsigned NumLaneElts = NumElts/NumLanes; | |||
4412 | ||||
4413 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { | |||
4414 | for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) { | |||
4415 | int BitI = Mask[l+i]; | |||
4416 | int BitI1 = Mask[l+i+1]; | |||
4417 | if (!isUndefOrEqual(BitI, j)) | |||
4418 | return false; | |||
4419 | if (V2IsSplat) { | |||
4420 | if (isUndefOrEqual(BitI1, NumElts)) | |||
4421 | return false; | |||
4422 | } else { | |||
4423 | if (!isUndefOrEqual(BitI1, j+NumElts)) | |||
4424 | return false; | |||
4425 | } | |||
4426 | } | |||
4427 | } | |||
4428 | return true; | |||
4429 | } | |||
4430 | ||||
4431 | /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form | |||
4432 | /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef, | |||
4433 | /// <0, 0, 1, 1> | |||
4434 | static bool isUNPCKL_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) { | |||
4435 | unsigned NumElts = VT.getVectorNumElements(); | |||
4436 | bool Is256BitVec = VT.is256BitVector(); | |||
4437 | ||||
4438 | if (VT.is512BitVector()) | |||
4439 | return false; | |||
4440 | assert((VT.is128BitVector() || VT.is256BitVector()) &&(((VT.is128BitVector() || VT.is256BitVector()) && "Unsupported vector type for unpckh" ) ? static_cast<void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector()) && \"Unsupported vector type for unpckh\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4441, __PRETTY_FUNCTION__)) | |||
4441 | "Unsupported vector type for unpckh")(((VT.is128BitVector() || VT.is256BitVector()) && "Unsupported vector type for unpckh" ) ? static_cast<void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector()) && \"Unsupported vector type for unpckh\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4441, __PRETTY_FUNCTION__)); | |||
4442 | ||||
4443 | if (Is256BitVec && NumElts != 4 && NumElts != 8 && | |||
4444 | (!HasInt256 || (NumElts != 16 && NumElts != 32))) | |||
4445 | return false; | |||
4446 | ||||
4447 | // For 256-bit i64/f64, use MOVDDUPY instead, so reject the matching pattern | |||
4448 | // FIXME: Need a better way to get rid of this, there's no latency difference | |||
4449 | // between UNPCKLPD and MOVDDUP, the later should always be checked first and | |||
4450 | // the former later. We should also remove the "_undef" special mask. | |||
4451 | if (NumElts == 4 && Is256BitVec) | |||
4452 | return false; | |||
4453 | ||||
4454 | // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate | |||
4455 | // independently on 128-bit lanes. | |||
4456 | unsigned NumLanes = VT.getSizeInBits()/128; | |||
4457 | unsigned NumLaneElts = NumElts/NumLanes; | |||
4458 | ||||
4459 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { | |||
4460 | for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) { | |||
4461 | int BitI = Mask[l+i]; | |||
4462 | int BitI1 = Mask[l+i+1]; | |||
4463 | ||||
4464 | if (!isUndefOrEqual(BitI, j)) | |||
4465 | return false; | |||
4466 | if (!isUndefOrEqual(BitI1, j)) | |||
4467 | return false; | |||
4468 | } | |||
4469 | } | |||
4470 | ||||
4471 | return true; | |||
4472 | } | |||
4473 | ||||
4474 | /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form | |||
4475 | /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef, | |||
4476 | /// <2, 2, 3, 3> | |||
4477 | static bool isUNPCKH_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) { | |||
4478 | unsigned NumElts = VT.getVectorNumElements(); | |||
4479 | ||||
4480 | if (VT.is512BitVector()) | |||
4481 | return false; | |||
4482 | ||||
4483 | assert((VT.is128BitVector() || VT.is256BitVector()) &&(((VT.is128BitVector() || VT.is256BitVector()) && "Unsupported vector type for unpckh" ) ? static_cast<void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector()) && \"Unsupported vector type for unpckh\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4484, __PRETTY_FUNCTION__)) | |||
4484 | "Unsupported vector type for unpckh")(((VT.is128BitVector() || VT.is256BitVector()) && "Unsupported vector type for unpckh" ) ? static_cast<void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector()) && \"Unsupported vector type for unpckh\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4484, __PRETTY_FUNCTION__)); | |||
4485 | ||||
4486 | if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 && | |||
4487 | (!HasInt256 || (NumElts != 16 && NumElts != 32))) | |||
4488 | return false; | |||
4489 | ||||
4490 | // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate | |||
4491 | // independently on 128-bit lanes. | |||
4492 | unsigned NumLanes = VT.getSizeInBits()/128; | |||
4493 | unsigned NumLaneElts = NumElts/NumLanes; | |||
4494 | ||||
4495 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { | |||
4496 | for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) { | |||
4497 | int BitI = Mask[l+i]; | |||
4498 | int BitI1 = Mask[l+i+1]; | |||
4499 | if (!isUndefOrEqual(BitI, j)) | |||
4500 | return false; | |||
4501 | if (!isUndefOrEqual(BitI1, j)) | |||
4502 | return false; | |||
4503 | } | |||
4504 | } | |||
4505 | return true; | |||
4506 | } | |||
4507 | ||||
4508 | // Match for INSERTI64x4 INSERTF64x4 instructions (src0[0], src1[0]) or | |||
4509 | // (src1[0], src0[1]), manipulation with 256-bit sub-vectors | |||
4510 | static bool isINSERT64x4Mask(ArrayRef<int> Mask, MVT VT, unsigned int *Imm) { | |||
4511 | if (!VT.is512BitVector()) | |||
4512 | return false; | |||
4513 | ||||
4514 | unsigned NumElts = VT.getVectorNumElements(); | |||
4515 | unsigned HalfSize = NumElts/2; | |||
4516 | if (isSequentialOrUndefInRange(Mask, 0, HalfSize, 0)) { | |||
4517 | if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, NumElts)) { | |||
4518 | *Imm = 1; | |||
4519 | return true; | |||
4520 | } | |||
4521 | } | |||
4522 | if (isSequentialOrUndefInRange(Mask, 0, HalfSize, NumElts)) { | |||
4523 | if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, HalfSize)) { | |||
4524 | *Imm = 0; | |||
4525 | return true; | |||
4526 | } | |||
4527 | } | |||
4528 | return false; | |||
4529 | } | |||
4530 | ||||
4531 | /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand | |||
4532 | /// specifies a shuffle of elements that is suitable for input to MOVSS, | |||
4533 | /// MOVSD, and MOVD, i.e. setting the lowest element. | |||
4534 | static bool isMOVLMask(ArrayRef<int> Mask, EVT VT) { | |||
4535 | if (VT.getVectorElementType().getSizeInBits() < 32) | |||
4536 | return false; | |||
4537 | if (!VT.is128BitVector()) | |||
4538 | return false; | |||
4539 | ||||
4540 | unsigned NumElts = VT.getVectorNumElements(); | |||
4541 | ||||
4542 | if (!isUndefOrEqual(Mask[0], NumElts)) | |||
4543 | return false; | |||
4544 | ||||
4545 | for (unsigned i = 1; i != NumElts; ++i) | |||
4546 | if (!isUndefOrEqual(Mask[i], i)) | |||
4547 | return false; | |||
4548 | ||||
4549 | return true; | |||
4550 | } | |||
4551 | ||||
4552 | /// isVPERM2X128Mask - Match 256-bit shuffles where the elements are considered | |||
4553 | /// as permutations between 128-bit chunks or halves. As an example: this | |||
4554 | /// shuffle bellow: | |||
4555 | /// vector_shuffle <4, 5, 6, 7, 12, 13, 14, 15> | |||
4556 | /// The first half comes from the second half of V1 and the second half from the | |||
4557 | /// the second half of V2. | |||
4558 | static bool isVPERM2X128Mask(ArrayRef<int> Mask, MVT VT, bool HasFp256) { | |||
4559 | if (!HasFp256 || !VT.is256BitVector()) | |||
4560 | return false; | |||
4561 | ||||
4562 | // The shuffle result is divided into half A and half B. In total the two | |||
4563 | // sources have 4 halves, namely: C, D, E, F. The final values of A and | |||
4564 | // B must come from C, D, E or F. | |||
4565 | unsigned HalfSize = VT.getVectorNumElements()/2; | |||
4566 | bool MatchA = false, MatchB = false; | |||
4567 | ||||
4568 | // Check if A comes from one of C, D, E, F. | |||
4569 | for (unsigned Half = 0; Half != 4; ++Half) { | |||
4570 | if (isSequentialOrUndefInRange(Mask, 0, HalfSize, Half*HalfSize)) { | |||
4571 | MatchA = true; | |||
4572 | break; | |||
4573 | } | |||
4574 | } | |||
4575 | ||||
4576 | // Check if B comes from one of C, D, E, F. | |||
4577 | for (unsigned Half = 0; Half != 4; ++Half) { | |||
4578 | if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, Half*HalfSize)) { | |||
4579 | MatchB = true; | |||
4580 | break; | |||
4581 | } | |||
4582 | } | |||
4583 | ||||
4584 | return MatchA && MatchB; | |||
4585 | } | |||
4586 | ||||
4587 | /// getShuffleVPERM2X128Immediate - Return the appropriate immediate to shuffle | |||
4588 | /// the specified VECTOR_MASK mask with VPERM2F128/VPERM2I128 instructions. | |||
4589 | static unsigned getShuffleVPERM2X128Immediate(ShuffleVectorSDNode *SVOp) { | |||
4590 | MVT VT = SVOp->getSimpleValueType(0); | |||
4591 | ||||
4592 | unsigned HalfSize = VT.getVectorNumElements()/2; | |||
4593 | ||||
4594 | unsigned FstHalf = 0, SndHalf = 0; | |||
4595 | for (unsigned i = 0; i < HalfSize; ++i) { | |||
4596 | if (SVOp->getMaskElt(i) > 0) { | |||
4597 | FstHalf = SVOp->getMaskElt(i)/HalfSize; | |||
4598 | break; | |||
4599 | } | |||
4600 | } | |||
4601 | for (unsigned i = HalfSize; i < HalfSize*2; ++i) { | |||
4602 | if (SVOp->getMaskElt(i) > 0) { | |||
4603 | SndHalf = SVOp->getMaskElt(i)/HalfSize; | |||
4604 | break; | |||
4605 | } | |||
4606 | } | |||
4607 | ||||
4608 | return (FstHalf | (SndHalf << 4)); | |||
4609 | } | |||
4610 | ||||
4611 | // Symetric in-lane mask. Each lane has 4 elements (for imm8) | |||
4612 | static bool isPermImmMask(ArrayRef<int> Mask, MVT VT, unsigned& Imm8) { | |||
4613 | unsigned EltSize = VT.getVectorElementType().getSizeInBits(); | |||
4614 | if (EltSize < 32) | |||
4615 | return false; | |||
4616 | ||||
4617 | unsigned NumElts = VT.getVectorNumElements(); | |||
4618 | Imm8 = 0; | |||
4619 | if (VT.is128BitVector() || (VT.is256BitVector() && EltSize == 64)) { | |||
4620 | for (unsigned i = 0; i != NumElts; ++i) { | |||
4621 | if (Mask[i] < 0) | |||
4622 | continue; | |||
4623 | Imm8 |= Mask[i] << (i*2); | |||
4624 | } | |||
4625 | return true; | |||
4626 | } | |||
4627 | ||||
4628 | unsigned LaneSize = 4; | |||
4629 | SmallVector<int, 4> MaskVal(LaneSize, -1); | |||
4630 | ||||
4631 | for (unsigned l = 0; l != NumElts; l += LaneSize) { | |||
4632 | for (unsigned i = 0; i != LaneSize; ++i) { | |||
4633 | if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize)) | |||
4634 | return false; | |||
4635 | if (Mask[i+l] < 0) | |||
4636 | continue; | |||
4637 | if (MaskVal[i] < 0) { | |||
4638 | MaskVal[i] = Mask[i+l] - l; | |||
4639 | Imm8 |= MaskVal[i] << (i*2); | |||
4640 | continue; | |||
4641 | } | |||
4642 | if (Mask[i+l] != (signed)(MaskVal[i]+l)) | |||
4643 | return false; | |||
4644 | } | |||
4645 | } | |||
4646 | return true; | |||
4647 | } | |||
4648 | ||||
4649 | /// isVPERMILPMask - Return true if the specified VECTOR_SHUFFLE operand | |||
4650 | /// specifies a shuffle of elements that is suitable for input to VPERMILPD*. | |||
4651 | /// Note that VPERMIL mask matching is different depending whether theunderlying | |||
4652 | /// type is 32 or 64. In the VPERMILPS the high half of the mask should point | |||
4653 | /// to the same elements of the low, but to the higher half of the source. | |||
4654 | /// In VPERMILPD the two lanes could be shuffled independently of each other | |||
4655 | /// with the same restriction that lanes can't be crossed. Also handles PSHUFDY. | |||
4656 | static bool isVPERMILPMask(ArrayRef<int> Mask, MVT VT) { | |||
4657 | unsigned EltSize = VT.getVectorElementType().getSizeInBits(); | |||
4658 | if (VT.getSizeInBits() < 256 || EltSize < 32) | |||
4659 | return false; | |||
4660 | bool symetricMaskRequired = (EltSize == 32); | |||
4661 | unsigned NumElts = VT.getVectorNumElements(); | |||
4662 | ||||
4663 | unsigned NumLanes = VT.getSizeInBits()/128; | |||
4664 | unsigned LaneSize = NumElts/NumLanes; | |||
4665 | // 2 or 4 elements in one lane | |||
4666 | ||||
4667 | SmallVector<int, 4> ExpectedMaskVal(LaneSize, -1); | |||
4668 | for (unsigned l = 0; l != NumElts; l += LaneSize) { | |||
4669 | for (unsigned i = 0; i != LaneSize; ++i) { | |||
4670 | if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize)) | |||
4671 | return false; | |||
4672 | if (symetricMaskRequired) { | |||
4673 | if (ExpectedMaskVal[i] < 0 && Mask[i+l] >= 0) { | |||
4674 | ExpectedMaskVal[i] = Mask[i+l] - l; | |||
4675 | continue; | |||
4676 | } | |||
4677 | if (!isUndefOrEqual(Mask[i+l], ExpectedMaskVal[i]+l)) | |||
4678 | return false; | |||
4679 | } | |||
4680 | } | |||
4681 | } | |||
4682 | return true; | |||
4683 | } | |||
4684 | ||||
4685 | /// isCommutedMOVLMask - Returns true if the shuffle mask is except the reverse | |||
4686 | /// of what x86 movss want. X86 movs requires the lowest element to be lowest | |||
4687 | /// element of vector 2 and the other elements to come from vector 1 in order. | |||
4688 | static bool isCommutedMOVLMask(ArrayRef<int> Mask, MVT VT, | |||
4689 | bool V2IsSplat = false, bool V2IsUndef = false) { | |||
4690 | if (!VT.is128BitVector()) | |||
4691 | return false; | |||
4692 | ||||
4693 | unsigned NumOps = VT.getVectorNumElements(); | |||
4694 | if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16) | |||
4695 | return false; | |||
4696 | ||||
4697 | if (!isUndefOrEqual(Mask[0], 0)) | |||
4698 | return false; | |||
4699 | ||||
4700 | for (unsigned i = 1; i != NumOps; ++i) | |||
4701 | if (!(isUndefOrEqual(Mask[i], i+NumOps) || | |||
4702 | (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) || | |||
4703 | (V2IsSplat && isUndefOrEqual(Mask[i], NumOps)))) | |||
4704 | return false; | |||
4705 | ||||
4706 | return true; | |||
4707 | } | |||
4708 | ||||
4709 | /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand | |||
4710 | /// specifies a shuffle of elements that is suitable for input to MOVSHDUP. | |||
4711 | /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7> | |||
4712 | static bool isMOVSHDUPMask(ArrayRef<int> Mask, MVT VT, | |||
4713 | const X86Subtarget *Subtarget) { | |||
4714 | if (!Subtarget->hasSSE3()) | |||
4715 | return false; | |||
4716 | ||||
4717 | unsigned NumElems = VT.getVectorNumElements(); | |||
4718 | ||||
4719 | if ((VT.is128BitVector() && NumElems != 4) || | |||
4720 | (VT.is256BitVector() && NumElems != 8) || | |||
4721 | (VT.is512BitVector() && NumElems != 16)) | |||
4722 | return false; | |||
4723 | ||||
4724 | // "i+1" is the value the indexed mask element must have | |||
4725 | for (unsigned i = 0; i != NumElems; i += 2) | |||
4726 | if (!isUndefOrEqual(Mask[i], i+1) || | |||
4727 | !isUndefOrEqual(Mask[i+1], i+1)) | |||
4728 | return false; | |||
4729 | ||||
4730 | return true; | |||
4731 | } | |||
4732 | ||||
4733 | /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand | |||
4734 | /// specifies a shuffle of elements that is suitable for input to MOVSLDUP. | |||
4735 | /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6> | |||
4736 | static bool isMOVSLDUPMask(ArrayRef<int> Mask, MVT VT, | |||
4737 | const X86Subtarget *Subtarget) { | |||
4738 | if (!Subtarget->hasSSE3()) | |||
4739 | return false; | |||
4740 | ||||
4741 | unsigned NumElems = VT.getVectorNumElements(); | |||
4742 | ||||
4743 | if ((VT.is128BitVector() && NumElems != 4) || | |||
4744 | (VT.is256BitVector() && NumElems != 8) || | |||
4745 | (VT.is512BitVector() && NumElems != 16)) | |||
4746 | return false; | |||
4747 | ||||
4748 | // "i" is the value the indexed mask element must have | |||
4749 | for (unsigned i = 0; i != NumElems; i += 2) | |||
4750 | if (!isUndefOrEqual(Mask[i], i) || | |||
4751 | !isUndefOrEqual(Mask[i+1], i)) | |||
4752 | return false; | |||
4753 | ||||
4754 | return true; | |||
4755 | } | |||
4756 | ||||
4757 | /// isMOVDDUPYMask - Return true if the specified VECTOR_SHUFFLE operand | |||
4758 | /// specifies a shuffle of elements that is suitable for input to 256-bit | |||
4759 | /// version of MOVDDUP. | |||
4760 | static bool isMOVDDUPYMask(ArrayRef<int> Mask, MVT VT, bool HasFp256) { | |||
4761 | if (!HasFp256 || !VT.is256BitVector()) | |||
4762 | return false; | |||
4763 | ||||
4764 | unsigned NumElts = VT.getVectorNumElements(); | |||
4765 | if (NumElts != 4) | |||
4766 | return false; | |||
4767 | ||||
4768 | for (unsigned i = 0; i != NumElts/2; ++i) | |||
4769 | if (!isUndefOrEqual(Mask[i], 0)) | |||
4770 | return false; | |||
4771 | for (unsigned i = NumElts/2; i != NumElts; ++i) | |||
4772 | if (!isUndefOrEqual(Mask[i], NumElts/2)) | |||
4773 | return false; | |||
4774 | return true; | |||
4775 | } | |||
4776 | ||||
4777 | /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand | |||
4778 | /// specifies a shuffle of elements that is suitable for input to 128-bit | |||
4779 | /// version of MOVDDUP. | |||
4780 | static bool isMOVDDUPMask(ArrayRef<int> Mask, MVT VT) { | |||
4781 | if (!VT.is128BitVector()) | |||
4782 | return false; | |||
4783 | ||||
4784 | unsigned e = VT.getVectorNumElements() / 2; | |||
4785 | for (unsigned i = 0; i != e; ++i) | |||
4786 | if (!isUndefOrEqual(Mask[i], i)) | |||
4787 | return false; | |||
4788 | for (unsigned i = 0; i != e; ++i) | |||
4789 | if (!isUndefOrEqual(Mask[e+i], i)) | |||
4790 | return false; | |||
4791 | return true; | |||
4792 | } | |||
4793 | ||||
4794 | /// isVEXTRACTIndex - Return true if the specified | |||
4795 | /// EXTRACT_SUBVECTOR operand specifies a vector extract that is | |||
4796 | /// suitable for instruction that extract 128 or 256 bit vectors | |||
4797 | static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) { | |||
4798 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4798, __PRETTY_FUNCTION__)); | |||
4799 | if (!isa<ConstantSDNode>(N->getOperand(1).getNode())) | |||
4800 | return false; | |||
4801 | ||||
4802 | // The index should be aligned on a vecWidth-bit boundary. | |||
4803 | uint64_t Index = | |||
4804 | cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue(); | |||
4805 | ||||
4806 | MVT VT = N->getSimpleValueType(0); | |||
4807 | unsigned ElSize = VT.getVectorElementType().getSizeInBits(); | |||
4808 | bool Result = (Index * ElSize) % vecWidth == 0; | |||
4809 | ||||
4810 | return Result; | |||
4811 | } | |||
4812 | ||||
4813 | /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR | |||
4814 | /// operand specifies a subvector insert that is suitable for input to | |||
4815 | /// insertion of 128 or 256-bit subvectors | |||
4816 | static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) { | |||
4817 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4817, __PRETTY_FUNCTION__)); | |||
4818 | if (!isa<ConstantSDNode>(N->getOperand(2).getNode())) | |||
4819 | return false; | |||
4820 | // The index should be aligned on a vecWidth-bit boundary. | |||
4821 | uint64_t Index = | |||
4822 | cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue(); | |||
4823 | ||||
4824 | MVT VT = N->getSimpleValueType(0); | |||
4825 | unsigned ElSize = VT.getVectorElementType().getSizeInBits(); | |||
4826 | bool Result = (Index * ElSize) % vecWidth == 0; | |||
4827 | ||||
4828 | return Result; | |||
4829 | } | |||
4830 | ||||
4831 | bool X86::isVINSERT128Index(SDNode *N) { | |||
4832 | return isVINSERTIndex(N, 128); | |||
4833 | } | |||
4834 | ||||
4835 | bool X86::isVINSERT256Index(SDNode *N) { | |||
4836 | return isVINSERTIndex(N, 256); | |||
4837 | } | |||
4838 | ||||
4839 | bool X86::isVEXTRACT128Index(SDNode *N) { | |||
4840 | return isVEXTRACTIndex(N, 128); | |||
4841 | } | |||
4842 | ||||
4843 | bool X86::isVEXTRACT256Index(SDNode *N) { | |||
4844 | return isVEXTRACTIndex(N, 256); | |||
4845 | } | |||
4846 | ||||
4847 | /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle | |||
4848 | /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions. | |||
4849 | /// Handles 128-bit and 256-bit. | |||
4850 | static unsigned getShuffleSHUFImmediate(ShuffleVectorSDNode *N) { | |||
4851 | MVT VT = N->getSimpleValueType(0); | |||
4852 | ||||
4853 | assert((VT.getSizeInBits() >= 128) &&(((VT.getSizeInBits() >= 128) && "Unsupported vector type for PSHUF/SHUFP" ) ? static_cast<void> (0) : __assert_fail ("(VT.getSizeInBits() >= 128) && \"Unsupported vector type for PSHUF/SHUFP\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4854, __PRETTY_FUNCTION__)) | |||
4854 | "Unsupported vector type for PSHUF/SHUFP")(((VT.getSizeInBits() >= 128) && "Unsupported vector type for PSHUF/SHUFP" ) ? static_cast<void> (0) : __assert_fail ("(VT.getSizeInBits() >= 128) && \"Unsupported vector type for PSHUF/SHUFP\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4854, __PRETTY_FUNCTION__)); | |||
4855 | ||||
4856 | // Handle 128 and 256-bit vector lengths. AVX defines PSHUF/SHUFP to operate | |||
4857 | // independently on 128-bit lanes. | |||
4858 | unsigned NumElts = VT.getVectorNumElements(); | |||
4859 | unsigned NumLanes = VT.getSizeInBits()/128; | |||
4860 | unsigned NumLaneElts = NumElts/NumLanes; | |||
4861 | ||||
4862 | assert((NumLaneElts == 2 || NumLaneElts == 4 || NumLaneElts == 8) &&(((NumLaneElts == 2 || NumLaneElts == 4 || NumLaneElts == 8) && "Only supports 2, 4 or 8 elements per lane") ? static_cast< void> (0) : __assert_fail ("(NumLaneElts == 2 || NumLaneElts == 4 || NumLaneElts == 8) && \"Only supports 2, 4 or 8 elements per lane\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4863, __PRETTY_FUNCTION__)) | |||
4863 | "Only supports 2, 4 or 8 elements per lane")(((NumLaneElts == 2 || NumLaneElts == 4 || NumLaneElts == 8) && "Only supports 2, 4 or 8 elements per lane") ? static_cast< void> (0) : __assert_fail ("(NumLaneElts == 2 || NumLaneElts == 4 || NumLaneElts == 8) && \"Only supports 2, 4 or 8 elements per lane\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4863, __PRETTY_FUNCTION__)); | |||
4864 | ||||
4865 | unsigned Shift = (NumLaneElts >= 4) ? 1 : 0; | |||
4866 | unsigned Mask = 0; | |||
4867 | for (unsigned i = 0; i != NumElts; ++i) { | |||
4868 | int Elt = N->getMaskElt(i); | |||
4869 | if (Elt < 0) continue; | |||
4870 | Elt &= NumLaneElts - 1; | |||
4871 | unsigned ShAmt = (i << Shift) % 8; | |||
4872 | Mask |= Elt << ShAmt; | |||
4873 | } | |||
4874 | ||||
4875 | return Mask; | |||
4876 | } | |||
4877 | ||||
4878 | /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle | |||
4879 | /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction. | |||
4880 | static unsigned getShufflePSHUFHWImmediate(ShuffleVectorSDNode *N) { | |||
4881 | MVT VT = N->getSimpleValueType(0); | |||
4882 | ||||
4883 | assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&(((VT == MVT::v8i16 || VT == MVT::v16i16) && "Unsupported vector type for PSHUFHW" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v8i16 || VT == MVT::v16i16) && \"Unsupported vector type for PSHUFHW\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4884, __PRETTY_FUNCTION__)) | |||
4884 | "Unsupported vector type for PSHUFHW")(((VT == MVT::v8i16 || VT == MVT::v16i16) && "Unsupported vector type for PSHUFHW" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v8i16 || VT == MVT::v16i16) && \"Unsupported vector type for PSHUFHW\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4884, __PRETTY_FUNCTION__)); | |||
4885 | ||||
4886 | unsigned NumElts = VT.getVectorNumElements(); | |||
4887 | ||||
4888 | unsigned Mask = 0; | |||
4889 | for (unsigned l = 0; l != NumElts; l += 8) { | |||
4890 | // 8 nodes per lane, but we only care about the last 4. | |||
4891 | for (unsigned i = 0; i < 4; ++i) { | |||
4892 | int Elt = N->getMaskElt(l+i+4); | |||
4893 | if (Elt < 0) continue; | |||
4894 | Elt &= 0x3; // only 2-bits. | |||
4895 | Mask |= Elt << (i * 2); | |||
4896 | } | |||
4897 | } | |||
4898 | ||||
4899 | return Mask; | |||
4900 | } | |||
4901 | ||||
4902 | /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle | |||
4903 | /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction. | |||
4904 | static unsigned getShufflePSHUFLWImmediate(ShuffleVectorSDNode *N) { | |||
4905 | MVT VT = N->getSimpleValueType(0); | |||
4906 | ||||
4907 | assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&(((VT == MVT::v8i16 || VT == MVT::v16i16) && "Unsupported vector type for PSHUFHW" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v8i16 || VT == MVT::v16i16) && \"Unsupported vector type for PSHUFHW\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4908, __PRETTY_FUNCTION__)) | |||
4908 | "Unsupported vector type for PSHUFHW")(((VT == MVT::v8i16 || VT == MVT::v16i16) && "Unsupported vector type for PSHUFHW" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v8i16 || VT == MVT::v16i16) && \"Unsupported vector type for PSHUFHW\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4908, __PRETTY_FUNCTION__)); | |||
4909 | ||||
4910 | unsigned NumElts = VT.getVectorNumElements(); | |||
4911 | ||||
4912 | unsigned Mask = 0; | |||
4913 | for (unsigned l = 0; l != NumElts; l += 8) { | |||
4914 | // 8 nodes per lane, but we only care about the first 4. | |||
4915 | for (unsigned i = 0; i < 4; ++i) { | |||
4916 | int Elt = N->getMaskElt(l+i); | |||
4917 | if (Elt < 0) continue; | |||
4918 | Elt &= 0x3; // only 2-bits | |||
4919 | Mask |= Elt << (i * 2); | |||
4920 | } | |||
4921 | } | |||
4922 | ||||
4923 | return Mask; | |||
4924 | } | |||
4925 | ||||
4926 | /// \brief Return the appropriate immediate to shuffle the specified | |||
4927 | /// VECTOR_SHUFFLE mask with the PALIGNR (if InterLane is false) or with | |||
4928 | /// VALIGN (if Interlane is true) instructions. | |||
4929 | static unsigned getShuffleAlignrImmediate(ShuffleVectorSDNode *SVOp, | |||
4930 | bool InterLane) { | |||
4931 | MVT VT = SVOp->getSimpleValueType(0); | |||
4932 | unsigned EltSize = InterLane ? 1 : | |||
4933 | VT.getVectorElementType().getSizeInBits() >> 3; | |||
4934 | ||||
4935 | unsigned NumElts = VT.getVectorNumElements(); | |||
4936 | unsigned NumLanes = VT.is512BitVector() ? 1 : VT.getSizeInBits()/128; | |||
4937 | unsigned NumLaneElts = NumElts/NumLanes; | |||
4938 | ||||
4939 | int Val = 0; | |||
4940 | unsigned i; | |||
4941 | for (i = 0; i != NumElts; ++i) { | |||
4942 | Val = SVOp->getMaskElt(i); | |||
4943 | if (Val >= 0) | |||
4944 | break; | |||
4945 | } | |||
4946 | if (Val >= (int)NumElts) | |||
4947 | Val -= NumElts - NumLaneElts; | |||
4948 | ||||
4949 | assert(Val - i > 0 && "PALIGNR imm should be positive")((Val - i > 0 && "PALIGNR imm should be positive") ? static_cast<void> (0) : __assert_fail ("Val - i > 0 && \"PALIGNR imm should be positive\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4949, __PRETTY_FUNCTION__)); | |||
4950 | return (Val - i) * EltSize; | |||
4951 | } | |||
4952 | ||||
4953 | /// \brief Return the appropriate immediate to shuffle the specified | |||
4954 | /// VECTOR_SHUFFLE mask with the PALIGNR instruction. | |||
4955 | static unsigned getShufflePALIGNRImmediate(ShuffleVectorSDNode *SVOp) { | |||
4956 | return getShuffleAlignrImmediate(SVOp, false); | |||
4957 | } | |||
4958 | ||||
4959 | /// \brief Return the appropriate immediate to shuffle the specified | |||
4960 | /// VECTOR_SHUFFLE mask with the VALIGN instruction. | |||
4961 | static unsigned getShuffleVALIGNImmediate(ShuffleVectorSDNode *SVOp) { | |||
4962 | return getShuffleAlignrImmediate(SVOp, true); | |||
4963 | } | |||
4964 | ||||
4965 | ||||
4966 | static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) { | |||
4967 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4967, __PRETTY_FUNCTION__)); | |||
4968 | if (!isa<ConstantSDNode>(N->getOperand(1).getNode())) | |||
4969 | llvm_unreachable("Illegal extract subvector for VEXTRACT")::llvm::llvm_unreachable_internal("Illegal extract subvector for VEXTRACT" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4969); | |||
4970 | ||||
4971 | uint64_t Index = | |||
4972 | cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue(); | |||
4973 | ||||
4974 | MVT VecVT = N->getOperand(0).getSimpleValueType(); | |||
4975 | MVT ElVT = VecVT.getVectorElementType(); | |||
4976 | ||||
4977 | unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits(); | |||
4978 | return Index / NumElemsPerChunk; | |||
4979 | } | |||
4980 | ||||
4981 | static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) { | |||
4982 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4982, __PRETTY_FUNCTION__)); | |||
4983 | if (!isa<ConstantSDNode>(N->getOperand(2).getNode())) | |||
4984 | llvm_unreachable("Illegal insert subvector for VINSERT")::llvm::llvm_unreachable_internal("Illegal insert subvector for VINSERT" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 4984); | |||
4985 | ||||
4986 | uint64_t Index = | |||
4987 | cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue(); | |||
4988 | ||||
4989 | MVT VecVT = N->getSimpleValueType(0); | |||
4990 | MVT ElVT = VecVT.getVectorElementType(); | |||
4991 | ||||
4992 | unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits(); | |||
4993 | return Index / NumElemsPerChunk; | |||
4994 | } | |||
4995 | ||||
4996 | /// getExtractVEXTRACT128Immediate - Return the appropriate immediate | |||
4997 | /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128 | |||
4998 | /// and VINSERTI128 instructions. | |||
4999 | unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) { | |||
5000 | return getExtractVEXTRACTImmediate(N, 128); | |||
5001 | } | |||
5002 | ||||
5003 | /// getExtractVEXTRACT256Immediate - Return the appropriate immediate | |||
5004 | /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4 | |||
5005 | /// and VINSERTI64x4 instructions. | |||
5006 | unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) { | |||
5007 | return getExtractVEXTRACTImmediate(N, 256); | |||
5008 | } | |||
5009 | ||||
5010 | /// getInsertVINSERT128Immediate - Return the appropriate immediate | |||
5011 | /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128 | |||
5012 | /// and VINSERTI128 instructions. | |||
5013 | unsigned X86::getInsertVINSERT128Immediate(SDNode *N) { | |||
5014 | return getInsertVINSERTImmediate(N, 128); | |||
5015 | } | |||
5016 | ||||
5017 | /// getInsertVINSERT256Immediate - Return the appropriate immediate | |||
5018 | /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4 | |||
5019 | /// and VINSERTI64x4 instructions. | |||
5020 | unsigned X86::getInsertVINSERT256Immediate(SDNode *N) { | |||
5021 | return getInsertVINSERTImmediate(N, 256); | |||
5022 | } | |||
5023 | ||||
5024 | /// isZero - Returns true if Elt is a constant integer zero | |||
5025 | static bool isZero(SDValue V) { | |||
5026 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(V); | |||
5027 | return C && C->isNullValue(); | |||
5028 | } | |||
5029 | ||||
5030 | /// isZeroNode - Returns true if Elt is a constant zero or a floating point | |||
5031 | /// constant +0.0. | |||
5032 | bool X86::isZeroNode(SDValue Elt) { | |||
5033 | if (isZero(Elt)) | |||
5034 | return true; | |||
5035 | if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt)) | |||
5036 | return CFP->getValueAPF().isPosZero(); | |||
5037 | return false; | |||
5038 | } | |||
5039 | ||||
5040 | /// ShouldXformToMOVHLPS - Return true if the node should be transformed to | |||
5041 | /// match movhlps. The lower half elements should come from upper half of | |||
5042 | /// V1 (and in order), and the upper half elements should come from the upper | |||
5043 | /// half of V2 (and in order). | |||
5044 | static bool ShouldXformToMOVHLPS(ArrayRef<int> Mask, MVT VT) { | |||
5045 | if (!VT.is128BitVector()) | |||
5046 | return false; | |||
5047 | if (VT.getVectorNumElements() != 4) | |||
5048 | return false; | |||
5049 | for (unsigned i = 0, e = 2; i != e; ++i) | |||
5050 | if (!isUndefOrEqual(Mask[i], i+2)) | |||
5051 | return false; | |||
5052 | for (unsigned i = 2; i != 4; ++i) | |||
5053 | if (!isUndefOrEqual(Mask[i], i+4)) | |||
5054 | return false; | |||
5055 | return true; | |||
5056 | } | |||
5057 | ||||
5058 | /// isScalarLoadToVector - Returns true if the node is a scalar load that | |||
5059 | /// is promoted to a vector. It also returns the LoadSDNode by reference if | |||
5060 | /// required. | |||
5061 | static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = nullptr) { | |||
5062 | if (N->getOpcode() != ISD::SCALAR_TO_VECTOR) | |||
5063 | return false; | |||
5064 | N = N->getOperand(0).getNode(); | |||
5065 | if (!ISD::isNON_EXTLoad(N)) | |||
5066 | return false; | |||
5067 | if (LD) | |||
5068 | *LD = cast<LoadSDNode>(N); | |||
5069 | return true; | |||
5070 | } | |||
5071 | ||||
5072 | // Test whether the given value is a vector value which will be legalized | |||
5073 | // into a load. | |||
5074 | static bool WillBeConstantPoolLoad(SDNode *N) { | |||
5075 | if (N->getOpcode() != ISD::BUILD_VECTOR) | |||
5076 | return false; | |||
5077 | ||||
5078 | // Check for any non-constant elements. | |||
5079 | for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) | |||
5080 | switch (N->getOperand(i).getNode()->getOpcode()) { | |||
5081 | case ISD::UNDEF: | |||
5082 | case ISD::ConstantFP: | |||
5083 | case ISD::Constant: | |||
5084 | break; | |||
5085 | default: | |||
5086 | return false; | |||
5087 | } | |||
5088 | ||||
5089 | // Vectors of all-zeros and all-ones are materialized with special | |||
5090 | // instructions rather than being loaded. | |||
5091 | return !ISD::isBuildVectorAllZeros(N) && | |||
5092 | !ISD::isBuildVectorAllOnes(N); | |||
5093 | } | |||
5094 | ||||
5095 | /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to | |||
5096 | /// match movlp{s|d}. The lower half elements should come from lower half of | |||
5097 | /// V1 (and in order), and the upper half elements should come from the upper | |||
5098 | /// half of V2 (and in order). And since V1 will become the source of the | |||
5099 | /// MOVLP, it must be either a vector load or a scalar load to vector. | |||
5100 | static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, | |||
5101 | ArrayRef<int> Mask, MVT VT) { | |||
5102 | if (!VT.is128BitVector()) | |||
5103 | return false; | |||
5104 | ||||
5105 | if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1)) | |||
5106 | return false; | |||
5107 | // Is V2 is a vector load, don't do this transformation. We will try to use | |||
5108 | // load folding shufps op. | |||
5109 | if (ISD::isNON_EXTLoad(V2) || WillBeConstantPoolLoad(V2)) | |||
5110 | return false; | |||
5111 | ||||
5112 | unsigned NumElems = VT.getVectorNumElements(); | |||
5113 | ||||
5114 | if (NumElems != 2 && NumElems != 4) | |||
5115 | return false; | |||
5116 | for (unsigned i = 0, e = NumElems/2; i != e; ++i) | |||
5117 | if (!isUndefOrEqual(Mask[i], i)) | |||
5118 | return false; | |||
5119 | for (unsigned i = NumElems/2, e = NumElems; i != e; ++i) | |||
5120 | if (!isUndefOrEqual(Mask[i], i+NumElems)) | |||
5121 | return false; | |||
5122 | return true; | |||
5123 | } | |||
5124 | ||||
5125 | /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved | |||
5126 | /// to an zero vector. | |||
5127 | /// FIXME: move to dag combiner / method on ShuffleVectorSDNode | |||
5128 | static bool isZeroShuffle(ShuffleVectorSDNode *N) { | |||
5129 | SDValue V1 = N->getOperand(0); | |||
5130 | SDValue V2 = N->getOperand(1); | |||
5131 | unsigned NumElems = N->getValueType(0).getVectorNumElements(); | |||
5132 | for (unsigned i = 0; i != NumElems; ++i) { | |||
5133 | int Idx = N->getMaskElt(i); | |||
5134 | if (Idx >= (int)NumElems) { | |||
5135 | unsigned Opc = V2.getOpcode(); | |||
5136 | if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode())) | |||
5137 | continue; | |||
5138 | if (Opc != ISD::BUILD_VECTOR || | |||
5139 | !X86::isZeroNode(V2.getOperand(Idx-NumElems))) | |||
5140 | return false; | |||
5141 | } else if (Idx >= 0) { | |||
5142 | unsigned Opc = V1.getOpcode(); | |||
5143 | if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode())) | |||
5144 | continue; | |||
5145 | if (Opc != ISD::BUILD_VECTOR || | |||
5146 | !X86::isZeroNode(V1.getOperand(Idx))) | |||
5147 | return false; | |||
5148 | } | |||
5149 | } | |||
5150 | return true; | |||
5151 | } | |||
5152 | ||||
5153 | /// getZeroVector - Returns a vector of specified type with all zero elements. | |||
5154 | /// | |||
5155 | static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget, | |||
5156 | SelectionDAG &DAG, SDLoc dl) { | |||
5157 | assert(VT.isVector() && "Expected a vector type")((VT.isVector() && "Expected a vector type") ? static_cast <void> (0) : __assert_fail ("VT.isVector() && \"Expected a vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5157, __PRETTY_FUNCTION__)); | |||
5158 | ||||
5159 | // Always build SSE zero vectors as <4 x i32> bitcasted | |||
5160 | // to their dest type. This ensures they get CSE'd. | |||
5161 | SDValue Vec; | |||
5162 | if (VT.is128BitVector()) { // SSE | |||
5163 | if (Subtarget->hasSSE2()) { // SSE2 | |||
5164 | SDValue Cst = DAG.getConstant(0, MVT::i32); | |||
5165 | Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst); | |||
5166 | } else { // SSE1 | |||
5167 | SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32); | |||
5168 | Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst); | |||
5169 | } | |||
5170 | } else if (VT.is256BitVector()) { // AVX | |||
5171 | if (Subtarget->hasInt256()) { // AVX2 | |||
5172 | SDValue Cst = DAG.getConstant(0, MVT::i32); | |||
5173 | SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst }; | |||
5174 | Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops); | |||
5175 | } else { | |||
5176 | // 256-bit logic and arithmetic instructions in AVX are all | |||
5177 | // floating-point, no support for integer ops. Emit fp zeroed vectors. | |||
5178 | SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32); | |||
5179 | SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst }; | |||
5180 | Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops); | |||
5181 | } | |||
5182 | } else if (VT.is512BitVector()) { // AVX-512 | |||
5183 | SDValue Cst = DAG.getConstant(0, MVT::i32); | |||
5184 | SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst, | |||
5185 | Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst }; | |||
5186 | Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops); | |||
5187 | } else if (VT.getScalarType() == MVT::i1) { | |||
5188 | assert(VT.getVectorNumElements() <= 16 && "Unexpected vector type")((VT.getVectorNumElements() <= 16 && "Unexpected vector type" ) ? static_cast<void> (0) : __assert_fail ("VT.getVectorNumElements() <= 16 && \"Unexpected vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5188, __PRETTY_FUNCTION__)); | |||
5189 | SDValue Cst = DAG.getConstant(0, MVT::i1); | |||
5190 | SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst); | |||
5191 | return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops); | |||
5192 | } else | |||
5193 | llvm_unreachable("Unexpected vector type")::llvm::llvm_unreachable_internal("Unexpected vector type", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5193); | |||
5194 | ||||
5195 | return DAG.getNode(ISD::BITCAST, dl, VT, Vec); | |||
5196 | } | |||
5197 | ||||
5198 | /// getOnesVector - Returns a vector of specified type with all bits set. | |||
5199 | /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with | |||
5200 | /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately. | |||
5201 | /// Then bitcast to their original type, ensuring they get CSE'd. | |||
5202 | static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG, | |||
5203 | SDLoc dl) { | |||
5204 | assert(VT.isVector() && "Expected a vector type")((VT.isVector() && "Expected a vector type") ? static_cast <void> (0) : __assert_fail ("VT.isVector() && \"Expected a vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5204, __PRETTY_FUNCTION__)); | |||
5205 | ||||
5206 | SDValue Cst = DAG.getConstant(~0U, MVT::i32); | |||
5207 | SDValue Vec; | |||
5208 | if (VT.is256BitVector()) { | |||
5209 | if (HasInt256) { // AVX2 | |||
5210 | SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst }; | |||
5211 | Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops); | |||
5212 | } else { // AVX | |||
5213 | Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst); | |||
5214 | Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl); | |||
5215 | } | |||
5216 | } else if (VT.is128BitVector()) { | |||
5217 | Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst); | |||
5218 | } else | |||
5219 | llvm_unreachable("Unexpected vector type")::llvm::llvm_unreachable_internal("Unexpected vector type", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5219); | |||
5220 | ||||
5221 | return DAG.getNode(ISD::BITCAST, dl, VT, Vec); | |||
5222 | } | |||
5223 | ||||
5224 | /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements | |||
5225 | /// that point to V2 points to its first element. | |||
5226 | static void NormalizeMask(SmallVectorImpl<int> &Mask, unsigned NumElems) { | |||
5227 | for (unsigned i = 0; i != NumElems; ++i) { | |||
5228 | if (Mask[i] > (int)NumElems) { | |||
5229 | Mask[i] = NumElems; | |||
5230 | } | |||
5231 | } | |||
5232 | } | |||
5233 | ||||
5234 | /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd | |||
5235 | /// operation of specified width. | |||
5236 | static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1, | |||
5237 | SDValue V2) { | |||
5238 | unsigned NumElems = VT.getVectorNumElements(); | |||
5239 | SmallVector<int, 8> Mask; | |||
5240 | Mask.push_back(NumElems); | |||
5241 | for (unsigned i = 1; i != NumElems; ++i) | |||
5242 | Mask.push_back(i); | |||
5243 | return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]); | |||
5244 | } | |||
5245 | ||||
5246 | /// getUnpackl - Returns a vector_shuffle node for an unpackl operation. | |||
5247 | static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1, | |||
5248 | SDValue V2) { | |||
5249 | unsigned NumElems = VT.getVectorNumElements(); | |||
5250 | SmallVector<int, 8> Mask; | |||
5251 | for (unsigned i = 0, e = NumElems/2; i != e; ++i) { | |||
5252 | Mask.push_back(i); | |||
5253 | Mask.push_back(i + NumElems); | |||
5254 | } | |||
5255 | return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]); | |||
5256 | } | |||
5257 | ||||
5258 | /// getUnpackh - Returns a vector_shuffle node for an unpackh operation. | |||
5259 | static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1, | |||
5260 | SDValue V2) { | |||
5261 | unsigned NumElems = VT.getVectorNumElements(); | |||
5262 | SmallVector<int, 8> Mask; | |||
5263 | for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) { | |||
5264 | Mask.push_back(i + Half); | |||
5265 | Mask.push_back(i + NumElems + Half); | |||
5266 | } | |||
5267 | return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]); | |||
5268 | } | |||
5269 | ||||
5270 | // PromoteSplati8i16 - All i16 and i8 vector types can't be used directly by | |||
5271 | // a generic shuffle instruction because the target has no such instructions. | |||
5272 | // Generate shuffles which repeat i16 and i8 several times until they can be | |||
5273 | // represented by v4f32 and then be manipulated by target suported shuffles. | |||
5274 | static SDValue PromoteSplati8i16(SDValue V, SelectionDAG &DAG, int &EltNo) { | |||
5275 | MVT VT = V.getSimpleValueType(); | |||
5276 | int NumElems = VT.getVectorNumElements(); | |||
5277 | SDLoc dl(V); | |||
5278 | ||||
5279 | while (NumElems > 4) { | |||
5280 | if (EltNo < NumElems/2) { | |||
5281 | V = getUnpackl(DAG, dl, VT, V, V); | |||
5282 | } else { | |||
5283 | V = getUnpackh(DAG, dl, VT, V, V); | |||
5284 | EltNo -= NumElems/2; | |||
5285 | } | |||
5286 | NumElems >>= 1; | |||
5287 | } | |||
5288 | return V; | |||
5289 | } | |||
5290 | ||||
5291 | /// getLegalSplat - Generate a legal splat with supported x86 shuffles | |||
5292 | static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) { | |||
5293 | MVT VT = V.getSimpleValueType(); | |||
5294 | SDLoc dl(V); | |||
5295 | ||||
5296 | if (VT.is128BitVector()) { | |||
5297 | V = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V); | |||
5298 | int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo }; | |||
5299 | V = DAG.getVectorShuffle(MVT::v4f32, dl, V, DAG.getUNDEF(MVT::v4f32), | |||
5300 | &SplatMask[0]); | |||
5301 | } else if (VT.is256BitVector()) { | |||
5302 | // To use VPERMILPS to splat scalars, the second half of indicies must | |||
5303 | // refer to the higher part, which is a duplication of the lower one, | |||
5304 | // because VPERMILPS can only handle in-lane permutations. | |||
5305 | int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo, | |||
5306 | EltNo+4, EltNo+4, EltNo+4, EltNo+4 }; | |||
5307 | ||||
5308 | V = DAG.getNode(ISD::BITCAST, dl, MVT::v8f32, V); | |||
5309 | V = DAG.getVectorShuffle(MVT::v8f32, dl, V, DAG.getUNDEF(MVT::v8f32), | |||
5310 | &SplatMask[0]); | |||
5311 | } else | |||
5312 | llvm_unreachable("Vector size not supported")::llvm::llvm_unreachable_internal("Vector size not supported" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5312); | |||
5313 | ||||
5314 | return DAG.getNode(ISD::BITCAST, dl, VT, V); | |||
5315 | } | |||
5316 | ||||
5317 | /// PromoteSplat - Splat is promoted to target supported vector shuffles. | |||
5318 | static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) { | |||
5319 | MVT SrcVT = SV->getSimpleValueType(0); | |||
5320 | SDValue V1 = SV->getOperand(0); | |||
5321 | SDLoc dl(SV); | |||
5322 | ||||
5323 | int EltNo = SV->getSplatIndex(); | |||
5324 | int NumElems = SrcVT.getVectorNumElements(); | |||
5325 | bool Is256BitVec = SrcVT.is256BitVector(); | |||
5326 | ||||
5327 | assert(((SrcVT.is128BitVector() && NumElems > 4) || Is256BitVec) &&((((SrcVT.is128BitVector() && NumElems > 4) || Is256BitVec ) && "Unknown how to promote splat for type") ? static_cast <void> (0) : __assert_fail ("((SrcVT.is128BitVector() && NumElems > 4) || Is256BitVec) && \"Unknown how to promote splat for type\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5328, __PRETTY_FUNCTION__)) | |||
5328 | "Unknown how to promote splat for type")((((SrcVT.is128BitVector() && NumElems > 4) || Is256BitVec ) && "Unknown how to promote splat for type") ? static_cast <void> (0) : __assert_fail ("((SrcVT.is128BitVector() && NumElems > 4) || Is256BitVec) && \"Unknown how to promote splat for type\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5328, __PRETTY_FUNCTION__)); | |||
5329 | ||||
5330 | // Extract the 128-bit part containing the splat element and update | |||
5331 | // the splat element index when it refers to the higher register. | |||
5332 | if (Is256BitVec) { | |||
5333 | V1 = Extract128BitVector(V1, EltNo, DAG, dl); | |||
5334 | if (EltNo >= NumElems/2) | |||
5335 | EltNo -= NumElems/2; | |||
5336 | } | |||
5337 | ||||
5338 | // All i16 and i8 vector types can't be used directly by a generic shuffle | |||
5339 | // instruction because the target has no such instruction. Generate shuffles | |||
5340 | // which repeat i16 and i8 several times until they fit in i32, and then can | |||
5341 | // be manipulated by target suported shuffles. | |||
5342 | MVT EltVT = SrcVT.getVectorElementType(); | |||
5343 | if (EltVT == MVT::i8 || EltVT == MVT::i16) | |||
5344 | V1 = PromoteSplati8i16(V1, DAG, EltNo); | |||
5345 | ||||
5346 | // Recreate the 256-bit vector and place the same 128-bit vector | |||
5347 | // into the low and high part. This is necessary because we want | |||
5348 | // to use VPERM* to shuffle the vectors | |||
5349 | if (Is256BitVec) { | |||
5350 | V1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, SrcVT, V1, V1); | |||
5351 | } | |||
5352 | ||||
5353 | return getLegalSplat(DAG, V1, EltNo); | |||
5354 | } | |||
5355 | ||||
5356 | /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified | |||
5357 | /// vector of zero or undef vector. This produces a shuffle where the low | |||
5358 | /// element of V2 is swizzled into the zero/undef vector, landing at element | |||
5359 | /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3). | |||
5360 | static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx, | |||
5361 | bool IsZero, | |||
5362 | const X86Subtarget *Subtarget, | |||
5363 | SelectionDAG &DAG) { | |||
5364 | MVT VT = V2.getSimpleValueType(); | |||
5365 | SDValue V1 = IsZero | |||
5366 | ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT); | |||
5367 | unsigned NumElems = VT.getVectorNumElements(); | |||
5368 | SmallVector<int, 16> MaskVec; | |||
5369 | for (unsigned i = 0; i != NumElems; ++i) | |||
5370 | // If this is the insertion idx, put the low elt of V2 here. | |||
5371 | MaskVec.push_back(i == Idx ? NumElems : i); | |||
5372 | return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]); | |||
5373 | } | |||
5374 | ||||
5375 | /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the | |||
5376 | /// target specific opcode. Returns true if the Mask could be calculated. Sets | |||
5377 | /// IsUnary to true if only uses one source. Note that this will set IsUnary for | |||
5378 | /// shuffles which use a single input multiple times, and in those cases it will | |||
5379 | /// adjust the mask to only have indices within that single input. | |||
5380 | static bool getTargetShuffleMask(SDNode *N, MVT VT, | |||
5381 | SmallVectorImpl<int> &Mask, bool &IsUnary) { | |||
5382 | unsigned NumElems = VT.getVectorNumElements(); | |||
5383 | SDValue ImmN; | |||
5384 | ||||
5385 | IsUnary = false; | |||
5386 | bool IsFakeUnary = false; | |||
5387 | switch(N->getOpcode()) { | |||
5388 | case X86ISD::BLENDI: | |||
5389 | ImmN = N->getOperand(N->getNumOperands()-1); | |||
5390 | DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); | |||
5391 | break; | |||
5392 | case X86ISD::SHUFP: | |||
5393 | ImmN = N->getOperand(N->getNumOperands()-1); | |||
5394 | DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); | |||
5395 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
5396 | break; | |||
5397 | case X86ISD::UNPCKH: | |||
5398 | DecodeUNPCKHMask(VT, Mask); | |||
5399 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
5400 | break; | |||
5401 | case X86ISD::UNPCKL: | |||
5402 | DecodeUNPCKLMask(VT, Mask); | |||
5403 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
5404 | break; | |||
5405 | case X86ISD::MOVHLPS: | |||
5406 | DecodeMOVHLPSMask(NumElems, Mask); | |||
5407 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
5408 | break; | |||
5409 | case X86ISD::MOVLHPS: | |||
5410 | DecodeMOVLHPSMask(NumElems, Mask); | |||
5411 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
5412 | break; | |||
5413 | case X86ISD::PALIGNR: | |||
5414 | ImmN = N->getOperand(N->getNumOperands()-1); | |||
5415 | DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); | |||
5416 | break; | |||
5417 | case X86ISD::PSHUFD: | |||
5418 | case X86ISD::VPERMILPI: | |||
5419 | ImmN = N->getOperand(N->getNumOperands()-1); | |||
5420 | DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); | |||
5421 | IsUnary = true; | |||
5422 | break; | |||
5423 | case X86ISD::PSHUFHW: | |||
5424 | ImmN = N->getOperand(N->getNumOperands()-1); | |||
5425 | DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); | |||
5426 | IsUnary = true; | |||
5427 | break; | |||
5428 | case X86ISD::PSHUFLW: | |||
5429 | ImmN = N->getOperand(N->getNumOperands()-1); | |||
5430 | DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); | |||
5431 | IsUnary = true; | |||
5432 | break; | |||
5433 | case X86ISD::PSHUFB: { | |||
5434 | IsUnary = true; | |||
5435 | SDValue MaskNode = N->getOperand(1); | |||
5436 | while (MaskNode->getOpcode() == ISD::BITCAST) | |||
5437 | MaskNode = MaskNode->getOperand(0); | |||
5438 | ||||
5439 | if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) { | |||
5440 | // If we have a build-vector, then things are easy. | |||
5441 | EVT VT = MaskNode.getValueType(); | |||
5442 | assert(VT.isVector() &&((VT.isVector() && "Can't produce a non-vector with a build_vector!" ) ? static_cast<void> (0) : __assert_fail ("VT.isVector() && \"Can't produce a non-vector with a build_vector!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5443, __PRETTY_FUNCTION__)) | |||
5443 | "Can't produce a non-vector with a build_vector!")((VT.isVector() && "Can't produce a non-vector with a build_vector!" ) ? static_cast<void> (0) : __assert_fail ("VT.isVector() && \"Can't produce a non-vector with a build_vector!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5443, __PRETTY_FUNCTION__)); | |||
5444 | if (!VT.isInteger()) | |||
5445 | return false; | |||
5446 | ||||
5447 | int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8; | |||
5448 | ||||
5449 | SmallVector<uint64_t, 32> RawMask; | |||
5450 | for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) { | |||
5451 | SDValue Op = MaskNode->getOperand(i); | |||
5452 | if (Op->getOpcode() == ISD::UNDEF) { | |||
5453 | RawMask.push_back((uint64_t)SM_SentinelUndef); | |||
5454 | continue; | |||
5455 | } | |||
5456 | auto *CN = dyn_cast<ConstantSDNode>(Op.getNode()); | |||
5457 | if (!CN) | |||
5458 | return false; | |||
5459 | APInt MaskElement = CN->getAPIntValue(); | |||
5460 | ||||
5461 | // We now have to decode the element which could be any integer size and | |||
5462 | // extract each byte of it. | |||
5463 | for (int j = 0; j < NumBytesPerElement; ++j) { | |||
5464 | // Note that this is x86 and so always little endian: the low byte is | |||
5465 | // the first byte of the mask. | |||
5466 | RawMask.push_back(MaskElement.getLoBits(8).getZExtValue()); | |||
5467 | MaskElement = MaskElement.lshr(8); | |||
5468 | } | |||
5469 | } | |||
5470 | DecodePSHUFBMask(RawMask, Mask); | |||
5471 | break; | |||
5472 | } | |||
5473 | ||||
5474 | auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode); | |||
5475 | if (!MaskLoad) | |||
5476 | return false; | |||
5477 | ||||
5478 | SDValue Ptr = MaskLoad->getBasePtr(); | |||
5479 | if (Ptr->getOpcode() == X86ISD::Wrapper) | |||
5480 | Ptr = Ptr->getOperand(0); | |||
5481 | ||||
5482 | auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr); | |||
5483 | if (!MaskCP || MaskCP->isMachineConstantPoolEntry()) | |||
5484 | return false; | |||
5485 | ||||
5486 | if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) { | |||
5487 | DecodePSHUFBMask(C, Mask); | |||
5488 | break; | |||
5489 | } | |||
5490 | ||||
5491 | return false; | |||
5492 | } | |||
5493 | case X86ISD::VPERMI: | |||
5494 | ImmN = N->getOperand(N->getNumOperands()-1); | |||
5495 | DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); | |||
5496 | IsUnary = true; | |||
5497 | break; | |||
5498 | case X86ISD::MOVSS: | |||
5499 | case X86ISD::MOVSD: | |||
5500 | DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask); | |||
5501 | break; | |||
5502 | case X86ISD::VPERM2X128: | |||
5503 | ImmN = N->getOperand(N->getNumOperands()-1); | |||
5504 | DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); | |||
5505 | if (Mask.empty()) return false; | |||
5506 | break; | |||
5507 | case X86ISD::MOVSLDUP: | |||
5508 | DecodeMOVSLDUPMask(VT, Mask); | |||
5509 | IsUnary = true; | |||
5510 | break; | |||
5511 | case X86ISD::MOVSHDUP: | |||
5512 | DecodeMOVSHDUPMask(VT, Mask); | |||
5513 | IsUnary = true; | |||
5514 | break; | |||
5515 | case X86ISD::MOVDDUP: | |||
5516 | DecodeMOVDDUPMask(VT, Mask); | |||
5517 | IsUnary = true; | |||
5518 | break; | |||
5519 | case X86ISD::MOVLHPD: | |||
5520 | case X86ISD::MOVLPD: | |||
5521 | case X86ISD::MOVLPS: | |||
5522 | // Not yet implemented | |||
5523 | return false; | |||
5524 | default: llvm_unreachable("unknown target shuffle node")::llvm::llvm_unreachable_internal("unknown target shuffle node" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5524); | |||
5525 | } | |||
5526 | ||||
5527 | // If we have a fake unary shuffle, the shuffle mask is spread across two | |||
5528 | // inputs that are actually the same node. Re-map the mask to always point | |||
5529 | // into the first input. | |||
5530 | if (IsFakeUnary) | |||
5531 | for (int &M : Mask) | |||
5532 | if (M >= (int)Mask.size()) | |||
5533 | M -= Mask.size(); | |||
5534 | ||||
5535 | return true; | |||
5536 | } | |||
5537 | ||||
5538 | /// getShuffleScalarElt - Returns the scalar element that will make up the ith | |||
5539 | /// element of the result of the vector shuffle. | |||
5540 | static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG, | |||
5541 | unsigned Depth) { | |||
5542 | if (Depth == 6) | |||
5543 | return SDValue(); // Limit search depth. | |||
5544 | ||||
5545 | SDValue V = SDValue(N, 0); | |||
5546 | EVT VT = V.getValueType(); | |||
5547 | unsigned Opcode = V.getOpcode(); | |||
5548 | ||||
5549 | // Recurse into ISD::VECTOR_SHUFFLE node to find scalars. | |||
5550 | if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) { | |||
5551 | int Elt = SV->getMaskElt(Index); | |||
5552 | ||||
5553 | if (Elt < 0) | |||
5554 | return DAG.getUNDEF(VT.getVectorElementType()); | |||
5555 | ||||
5556 | unsigned NumElems = VT.getVectorNumElements(); | |||
5557 | SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0) | |||
5558 | : SV->getOperand(1); | |||
5559 | return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1); | |||
5560 | } | |||
5561 | ||||
5562 | // Recurse into target specific vector shuffles to find scalars. | |||
5563 | if (isTargetShuffle(Opcode)) { | |||
5564 | MVT ShufVT = V.getSimpleValueType(); | |||
5565 | unsigned NumElems = ShufVT.getVectorNumElements(); | |||
5566 | SmallVector<int, 16> ShuffleMask; | |||
5567 | bool IsUnary; | |||
5568 | ||||
5569 | if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary)) | |||
5570 | return SDValue(); | |||
5571 | ||||
5572 | int Elt = ShuffleMask[Index]; | |||
5573 | if (Elt < 0) | |||
5574 | return DAG.getUNDEF(ShufVT.getVectorElementType()); | |||
5575 | ||||
5576 | SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0) | |||
5577 | : N->getOperand(1); | |||
5578 | return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, | |||
5579 | Depth+1); | |||
5580 | } | |||
5581 | ||||
5582 | // Actual nodes that may contain scalar elements | |||
5583 | if (Opcode == ISD::BITCAST) { | |||
5584 | V = V.getOperand(0); | |||
5585 | EVT SrcVT = V.getValueType(); | |||
5586 | unsigned NumElems = VT.getVectorNumElements(); | |||
5587 | ||||
5588 | if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems) | |||
5589 | return SDValue(); | |||
5590 | } | |||
5591 | ||||
5592 | if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) | |||
5593 | return (Index == 0) ? V.getOperand(0) | |||
5594 | : DAG.getUNDEF(VT.getVectorElementType()); | |||
5595 | ||||
5596 | if (V.getOpcode() == ISD::BUILD_VECTOR) | |||
5597 | return V.getOperand(Index); | |||
5598 | ||||
5599 | return SDValue(); | |||
5600 | } | |||
5601 | ||||
5602 | /// getNumOfConsecutiveZeros - Return the number of elements of a vector | |||
5603 | /// shuffle operation which come from a consecutively from a zero. The | |||
5604 | /// search can start in two different directions, from left or right. | |||
5605 | /// We count undefs as zeros until PreferredNum is reached. | |||
5606 | static unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, | |||
5607 | unsigned NumElems, bool ZerosFromLeft, | |||
5608 | SelectionDAG &DAG, | |||
5609 | unsigned PreferredNum = -1U) { | |||
5610 | unsigned NumZeros = 0; | |||
5611 | for (unsigned i = 0; i != NumElems; ++i) { | |||
5612 | unsigned Index = ZerosFromLeft ? i : NumElems - i - 1; | |||
5613 | SDValue Elt = getShuffleScalarElt(SVOp, Index, DAG, 0); | |||
5614 | if (!Elt.getNode()) | |||
5615 | break; | |||
5616 | ||||
5617 | if (X86::isZeroNode(Elt)) | |||
5618 | ++NumZeros; | |||
5619 | else if (Elt.getOpcode() == ISD::UNDEF) // Undef as zero up to PreferredNum. | |||
5620 | NumZeros = std::min(NumZeros + 1, PreferredNum); | |||
5621 | else | |||
5622 | break; | |||
5623 | } | |||
5624 | ||||
5625 | return NumZeros; | |||
5626 | } | |||
5627 | ||||
5628 | /// isShuffleMaskConsecutive - Check if the shuffle mask indicies [MaskI, MaskE) | |||
5629 | /// correspond consecutively to elements from one of the vector operands, | |||
5630 | /// starting from its index OpIdx. Also tell OpNum which source vector operand. | |||
5631 | static | |||
5632 | bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, | |||
5633 | unsigned MaskI, unsigned MaskE, unsigned OpIdx, | |||
5634 | unsigned NumElems, unsigned &OpNum) { | |||
5635 | bool SeenV1 = false; | |||
5636 | bool SeenV2 = false; | |||
5637 | ||||
5638 | for (unsigned i = MaskI; i != MaskE; ++i, ++OpIdx) { | |||
5639 | int Idx = SVOp->getMaskElt(i); | |||
5640 | // Ignore undef indicies | |||
5641 | if (Idx < 0) | |||
5642 | continue; | |||
5643 | ||||
5644 | if (Idx < (int)NumElems) | |||
5645 | SeenV1 = true; | |||
5646 | else | |||
5647 | SeenV2 = true; | |||
5648 | ||||
5649 | // Only accept consecutive elements from the same vector | |||
5650 | if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2)) | |||
5651 | return false; | |||
5652 | } | |||
5653 | ||||
5654 | OpNum = SeenV1 ? 0 : 1; | |||
5655 | return true; | |||
5656 | } | |||
5657 | ||||
5658 | /// isVectorShiftRight - Returns true if the shuffle can be implemented as a | |||
5659 | /// logical left shift of a vector. | |||
5660 | static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG, | |||
5661 | bool &isLeft, SDValue &ShVal, unsigned &ShAmt) { | |||
5662 | unsigned NumElems = | |||
5663 | SVOp->getSimpleValueType(0).getVectorNumElements(); | |||
5664 | unsigned NumZeros = getNumOfConsecutiveZeros( | |||
5665 | SVOp, NumElems, false /* check zeros from right */, DAG, | |||
5666 | SVOp->getMaskElt(0)); | |||
5667 | unsigned OpSrc; | |||
5668 | ||||
5669 | if (!NumZeros) | |||
5670 | return false; | |||
5671 | ||||
5672 | // Considering the elements in the mask that are not consecutive zeros, | |||
5673 | // check if they consecutively come from only one of the source vectors. | |||
5674 | // | |||
5675 | // V1 = {X, A, B, C} 0 | |||
5676 | // \ \ \ / | |||
5677 | // vector_shuffle V1, V2 <1, 2, 3, X> | |||
5678 | // | |||
5679 | if (!isShuffleMaskConsecutive(SVOp, | |||
5680 | 0, // Mask Start Index | |||
5681 | NumElems-NumZeros, // Mask End Index(exclusive) | |||
5682 | NumZeros, // Where to start looking in the src vector | |||
5683 | NumElems, // Number of elements in vector | |||
5684 | OpSrc)) // Which source operand ? | |||
5685 | return false; | |||
5686 | ||||
5687 | isLeft = false; | |||
5688 | ShAmt = NumZeros; | |||
5689 | ShVal = SVOp->getOperand(OpSrc); | |||
5690 | return true; | |||
5691 | } | |||
5692 | ||||
5693 | /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a | |||
5694 | /// logical left shift of a vector. | |||
5695 | static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG, | |||
5696 | bool &isLeft, SDValue &ShVal, unsigned &ShAmt) { | |||
5697 | unsigned NumElems = | |||
5698 | SVOp->getSimpleValueType(0).getVectorNumElements(); | |||
5699 | unsigned NumZeros = getNumOfConsecutiveZeros( | |||
5700 | SVOp, NumElems, true /* check zeros from left */, DAG, | |||
5701 | NumElems - SVOp->getMaskElt(NumElems - 1) - 1); | |||
5702 | unsigned OpSrc; | |||
5703 | ||||
5704 | if (!NumZeros) | |||
5705 | return false; | |||
5706 | ||||
5707 | // Considering the elements in the mask that are not consecutive zeros, | |||
5708 | // check if they consecutively come from only one of the source vectors. | |||
5709 | // | |||
5710 | // 0 { A, B, X, X } = V2 | |||
5711 | // / \ / / | |||
5712 | // vector_shuffle V1, V2 <X, X, 4, 5> | |||
5713 | // | |||
5714 | if (!isShuffleMaskConsecutive(SVOp, | |||
5715 | NumZeros, // Mask Start Index | |||
5716 | NumElems, // Mask End Index(exclusive) | |||
5717 | 0, // Where to start looking in the src vector | |||
5718 | NumElems, // Number of elements in vector | |||
5719 | OpSrc)) // Which source operand ? | |||
5720 | return false; | |||
5721 | ||||
5722 | isLeft = true; | |||
5723 | ShAmt = NumZeros; | |||
5724 | ShVal = SVOp->getOperand(OpSrc); | |||
5725 | return true; | |||
5726 | } | |||
5727 | ||||
5728 | /// isVectorShift - Returns true if the shuffle can be implemented as a | |||
5729 | /// logical left or right shift of a vector. | |||
5730 | static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG, | |||
5731 | bool &isLeft, SDValue &ShVal, unsigned &ShAmt) { | |||
5732 | // Although the logic below support any bitwidth size, there are no | |||
5733 | // shift instructions which handle more than 128-bit vectors. | |||
5734 | if (!SVOp->getSimpleValueType(0).is128BitVector()) | |||
5735 | return false; | |||
5736 | ||||
5737 | if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) || | |||
5738 | isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt)) | |||
5739 | return true; | |||
5740 | ||||
5741 | return false; | |||
5742 | } | |||
5743 | ||||
5744 | /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8. | |||
5745 | /// | |||
5746 | static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros, | |||
5747 | unsigned NumNonZero, unsigned NumZero, | |||
5748 | SelectionDAG &DAG, | |||
5749 | const X86Subtarget* Subtarget, | |||
5750 | const TargetLowering &TLI) { | |||
5751 | if (NumNonZero > 8) | |||
5752 | return SDValue(); | |||
5753 | ||||
5754 | SDLoc dl(Op); | |||
5755 | SDValue V; | |||
5756 | bool First = true; | |||
5757 | for (unsigned i = 0; i < 16; ++i) { | |||
5758 | bool ThisIsNonZero = (NonZeros & (1 << i)) != 0; | |||
5759 | if (ThisIsNonZero && First) { | |||
5760 | if (NumZero) | |||
5761 | V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl); | |||
5762 | else | |||
5763 | V = DAG.getUNDEF(MVT::v8i16); | |||
5764 | First = false; | |||
5765 | } | |||
5766 | ||||
5767 | if ((i & 1) != 0) { | |||
5768 | SDValue ThisElt, LastElt; | |||
5769 | bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0; | |||
5770 | if (LastIsNonZero) { | |||
5771 | LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl, | |||
5772 | MVT::i16, Op.getOperand(i-1)); | |||
5773 | } | |||
5774 | if (ThisIsNonZero) { | |||
5775 | ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i)); | |||
5776 | ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16, | |||
5777 | ThisElt, DAG.getConstant(8, MVT::i8)); | |||
5778 | if (LastIsNonZero) | |||
5779 | ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt); | |||
5780 | } else | |||
5781 | ThisElt = LastElt; | |||
5782 | ||||
5783 | if (ThisElt.getNode()) | |||
5784 | V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt, | |||
5785 | DAG.getIntPtrConstant(i/2)); | |||
5786 | } | |||
5787 | } | |||
5788 | ||||
5789 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V); | |||
5790 | } | |||
5791 | ||||
5792 | /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16. | |||
5793 | /// | |||
5794 | static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros, | |||
5795 | unsigned NumNonZero, unsigned NumZero, | |||
5796 | SelectionDAG &DAG, | |||
5797 | const X86Subtarget* Subtarget, | |||
5798 | const TargetLowering &TLI) { | |||
5799 | if (NumNonZero > 4) | |||
5800 | return SDValue(); | |||
5801 | ||||
5802 | SDLoc dl(Op); | |||
5803 | SDValue V; | |||
5804 | bool First = true; | |||
5805 | for (unsigned i = 0; i < 8; ++i) { | |||
5806 | bool isNonZero = (NonZeros & (1 << i)) != 0; | |||
5807 | if (isNonZero) { | |||
5808 | if (First) { | |||
5809 | if (NumZero) | |||
5810 | V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl); | |||
5811 | else | |||
5812 | V = DAG.getUNDEF(MVT::v8i16); | |||
5813 | First = false; | |||
5814 | } | |||
5815 | V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, | |||
5816 | MVT::v8i16, V, Op.getOperand(i), | |||
5817 | DAG.getIntPtrConstant(i)); | |||
5818 | } | |||
5819 | } | |||
5820 | ||||
5821 | return V; | |||
5822 | } | |||
5823 | ||||
5824 | /// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32. | |||
5825 | static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG, | |||
5826 | const X86Subtarget *Subtarget, | |||
5827 | const TargetLowering &TLI) { | |||
5828 | // Find all zeroable elements. | |||
5829 | bool Zeroable[4]; | |||
5830 | for (int i=0; i < 4; ++i) { | |||
5831 | SDValue Elt = Op->getOperand(i); | |||
5832 | Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt)); | |||
5833 | } | |||
5834 | assert(std::count_if(&Zeroable[0], &Zeroable[4],((std::count_if(&Zeroable[0], &Zeroable[4], [](bool M ) { return !M; }) > 1 && "We expect at least two non-zero elements!" ) ? static_cast<void> (0) : __assert_fail ("std::count_if(&Zeroable[0], &Zeroable[4], [](bool M) { return !M; }) > 1 && \"We expect at least two non-zero elements!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5836, __PRETTY_FUNCTION__)) | |||
5835 | [](bool M) { return !M; }) > 1 &&((std::count_if(&Zeroable[0], &Zeroable[4], [](bool M ) { return !M; }) > 1 && "We expect at least two non-zero elements!" ) ? static_cast<void> (0) : __assert_fail ("std::count_if(&Zeroable[0], &Zeroable[4], [](bool M) { return !M; }) > 1 && \"We expect at least two non-zero elements!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5836, __PRETTY_FUNCTION__)) | |||
5836 | "We expect at least two non-zero elements!")((std::count_if(&Zeroable[0], &Zeroable[4], [](bool M ) { return !M; }) > 1 && "We expect at least two non-zero elements!" ) ? static_cast<void> (0) : __assert_fail ("std::count_if(&Zeroable[0], &Zeroable[4], [](bool M) { return !M; }) > 1 && \"We expect at least two non-zero elements!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5836, __PRETTY_FUNCTION__)); | |||
5837 | ||||
5838 | // We only know how to deal with build_vector nodes where elements are either | |||
5839 | // zeroable or extract_vector_elt with constant index. | |||
5840 | SDValue FirstNonZero; | |||
5841 | unsigned FirstNonZeroIdx; | |||
5842 | for (unsigned i=0; i < 4; ++i) { | |||
5843 | if (Zeroable[i]) | |||
5844 | continue; | |||
5845 | SDValue Elt = Op->getOperand(i); | |||
5846 | if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT || | |||
5847 | !isa<ConstantSDNode>(Elt.getOperand(1))) | |||
5848 | return SDValue(); | |||
5849 | // Make sure that this node is extracting from a 128-bit vector. | |||
5850 | MVT VT = Elt.getOperand(0).getSimpleValueType(); | |||
5851 | if (!VT.is128BitVector()) | |||
5852 | return SDValue(); | |||
5853 | if (!FirstNonZero.getNode()) { | |||
5854 | FirstNonZero = Elt; | |||
5855 | FirstNonZeroIdx = i; | |||
5856 | } | |||
5857 | } | |||
5858 | ||||
5859 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5859, __PRETTY_FUNCTION__)); | |||
5860 | SDValue V1 = FirstNonZero.getOperand(0); | |||
5861 | MVT VT = V1.getSimpleValueType(); | |||
5862 | ||||
5863 | // See if this build_vector can be lowered as a blend with zero. | |||
5864 | SDValue Elt; | |||
5865 | unsigned EltMaskIdx, EltIdx; | |||
5866 | int Mask[4]; | |||
5867 | for (EltIdx = 0; EltIdx < 4; ++EltIdx) { | |||
5868 | if (Zeroable[EltIdx]) { | |||
5869 | // The zero vector will be on the right hand side. | |||
5870 | Mask[EltIdx] = EltIdx+4; | |||
5871 | continue; | |||
5872 | } | |||
5873 | ||||
5874 | Elt = Op->getOperand(EltIdx); | |||
5875 | // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index. | |||
5876 | EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue(); | |||
5877 | if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx) | |||
5878 | break; | |||
5879 | Mask[EltIdx] = EltIdx; | |||
5880 | } | |||
5881 | ||||
5882 | if (EltIdx == 4) { | |||
5883 | // Let the shuffle legalizer deal with blend operations. | |||
5884 | SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op)); | |||
5885 | if (V1.getSimpleValueType() != VT) | |||
5886 | V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1); | |||
5887 | return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]); | |||
5888 | } | |||
5889 | ||||
5890 | // See if we can lower this build_vector to a INSERTPS. | |||
5891 | if (!Subtarget->hasSSE41()) | |||
5892 | return SDValue(); | |||
5893 | ||||
5894 | SDValue V2 = Elt.getOperand(0); | |||
5895 | if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx) | |||
5896 | V1 = SDValue(); | |||
5897 | ||||
5898 | bool CanFold = true; | |||
5899 | for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) { | |||
5900 | if (Zeroable[i]) | |||
5901 | continue; | |||
5902 | ||||
5903 | SDValue Current = Op->getOperand(i); | |||
5904 | SDValue SrcVector = Current->getOperand(0); | |||
5905 | if (!V1.getNode()) | |||
5906 | V1 = SrcVector; | |||
5907 | CanFold = SrcVector == V1 && | |||
5908 | cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i; | |||
5909 | } | |||
5910 | ||||
5911 | if (!CanFold) | |||
5912 | return SDValue(); | |||
5913 | ||||
5914 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5914, __PRETTY_FUNCTION__)); | |||
5915 | if (V1.getSimpleValueType() != MVT::v4f32) | |||
5916 | V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1); | |||
5917 | if (V2.getSimpleValueType() != MVT::v4f32) | |||
5918 | V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2); | |||
5919 | ||||
5920 | // Ok, we can emit an INSERTPS instruction. | |||
5921 | unsigned ZMask = 0; | |||
5922 | for (int i = 0; i < 4; ++i) | |||
5923 | if (Zeroable[i]) | |||
5924 | ZMask |= 1 << i; | |||
5925 | ||||
5926 | unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask; | |||
5927 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5927, __PRETTY_FUNCTION__)); | |||
5928 | SDValue Result = DAG.getNode(X86ISD::INSERTPS, SDLoc(Op), MVT::v4f32, V1, V2, | |||
5929 | DAG.getIntPtrConstant(InsertPSMask)); | |||
5930 | return DAG.getNode(ISD::BITCAST, SDLoc(Op), VT, Result); | |||
5931 | } | |||
5932 | ||||
5933 | /// Return a vector logical shift node. | |||
5934 | static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp, | |||
5935 | unsigned NumBits, SelectionDAG &DAG, | |||
5936 | const TargetLowering &TLI, SDLoc dl) { | |||
5937 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 5937, __PRETTY_FUNCTION__)); | |||
5938 | MVT ShVT = MVT::v2i64; | |||
5939 | unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ; | |||
5940 | SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp); | |||
5941 | MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(SrcOp.getValueType()); | |||
5942 | SDValue ShiftVal = DAG.getConstant(NumBits, ScalarShiftTy); | |||
5943 | return DAG.getNode(ISD::BITCAST, dl, VT, | |||
5944 | DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal)); | |||
5945 | } | |||
5946 | ||||
5947 | static SDValue | |||
5948 | LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) { | |||
5949 | ||||
5950 | // Check if the scalar load can be widened into a vector load. And if | |||
5951 | // the address is "base + cst" see if the cst can be "absorbed" into | |||
5952 | // the shuffle mask. | |||
5953 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) { | |||
5954 | SDValue Ptr = LD->getBasePtr(); | |||
5955 | if (!ISD::isNormalLoad(LD) || LD->isVolatile()) | |||
5956 | return SDValue(); | |||
5957 | EVT PVT = LD->getValueType(0); | |||
5958 | if (PVT != MVT::i32 && PVT != MVT::f32) | |||
5959 | return SDValue(); | |||
5960 | ||||
5961 | int FI = -1; | |||
5962 | int64_t Offset = 0; | |||
5963 | if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) { | |||
5964 | FI = FINode->getIndex(); | |||
5965 | Offset = 0; | |||
5966 | } else if (DAG.isBaseWithConstantOffset(Ptr) && | |||
5967 | isa<FrameIndexSDNode>(Ptr.getOperand(0))) { | |||
5968 | FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex(); | |||
5969 | Offset = Ptr.getConstantOperandVal(1); | |||
5970 | Ptr = Ptr.getOperand(0); | |||
5971 | } else { | |||
5972 | return SDValue(); | |||
5973 | } | |||
5974 | ||||
5975 | // FIXME: 256-bit vector instructions don't require a strict alignment, | |||
5976 | // improve this code to support it better. | |||
5977 | unsigned RequiredAlign = VT.getSizeInBits()/8; | |||
5978 | SDValue Chain = LD->getChain(); | |||
5979 | // Make sure the stack object alignment is at least 16 or 32. | |||
5980 | MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); | |||
5981 | if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) { | |||
5982 | if (MFI->isFixedObjectIndex(FI)) { | |||
5983 | // Can't change the alignment. FIXME: It's possible to compute | |||
5984 | // the exact stack offset and reference FI + adjust offset instead. | |||
5985 | // If someone *really* cares about this. That's the way to implement it. | |||
5986 | return SDValue(); | |||
5987 | } else { | |||
5988 | MFI->setObjectAlignment(FI, RequiredAlign); | |||
5989 | } | |||
5990 | } | |||
5991 | ||||
5992 | // (Offset % 16 or 32) must be multiple of 4. Then address is then | |||
5993 | // Ptr + (Offset & ~15). | |||
5994 | if (Offset < 0) | |||
5995 | return SDValue(); | |||
5996 | if ((Offset % RequiredAlign) & 3) | |||
5997 | return SDValue(); | |||
5998 | int64_t StartOffset = Offset & ~(RequiredAlign-1); | |||
5999 | if (StartOffset) | |||
6000 | Ptr = DAG.getNode(ISD::ADD, SDLoc(Ptr), Ptr.getValueType(), | |||
6001 | Ptr,DAG.getConstant(StartOffset, Ptr.getValueType())); | |||
6002 | ||||
6003 | int EltNo = (Offset - StartOffset) >> 2; | |||
6004 | unsigned NumElems = VT.getVectorNumElements(); | |||
6005 | ||||
6006 | EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems); | |||
6007 | SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr, | |||
6008 | LD->getPointerInfo().getWithOffset(StartOffset), | |||
6009 | false, false, false, 0); | |||
6010 | ||||
6011 | SmallVector<int, 8> Mask; | |||
6012 | for (unsigned i = 0; i != NumElems; ++i) | |||
6013 | Mask.push_back(EltNo); | |||
6014 | ||||
6015 | return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]); | |||
6016 | } | |||
6017 | ||||
6018 | return SDValue(); | |||
6019 | } | |||
6020 | ||||
6021 | /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a | |||
6022 | /// vector of type 'VT', see if the elements can be replaced by a single large | |||
6023 | /// load which has the same value as a build_vector whose operands are 'elts'. | |||
6024 | /// | |||
6025 | /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a | |||
6026 | /// | |||
6027 | /// FIXME: we'd also like to handle the case where the last elements are zero | |||
6028 | /// rather than undef via VZEXT_LOAD, but we do not detect that case today. | |||
6029 | /// There's even a handy isZeroNode for that purpose. | |||
6030 | static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts, | |||
6031 | SDLoc &DL, SelectionDAG &DAG, | |||
6032 | bool isAfterLegalize) { | |||
6033 | EVT EltVT = VT.getVectorElementType(); | |||
6034 | unsigned NumElems = Elts.size(); | |||
6035 | ||||
6036 | LoadSDNode *LDBase = nullptr; | |||
6037 | unsigned LastLoadedElt = -1U; | |||
6038 | ||||
6039 | // For each element in the initializer, see if we've found a load or an undef. | |||
6040 | // If we don't find an initial load element, or later load elements are | |||
6041 | // non-consecutive, bail out. | |||
6042 | for (unsigned i = 0; i < NumElems; ++i) { | |||
6043 | SDValue Elt = Elts[i]; | |||
6044 | ||||
6045 | if (!Elt.getNode() || | |||
6046 | (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode()))) | |||
6047 | return SDValue(); | |||
6048 | if (!LDBase) { | |||
6049 | if (Elt.getNode()->getOpcode() == ISD::UNDEF) | |||
6050 | return SDValue(); | |||
6051 | LDBase = cast<LoadSDNode>(Elt.getNode()); | |||
6052 | LastLoadedElt = i; | |||
6053 | continue; | |||
6054 | } | |||
6055 | if (Elt.getOpcode() == ISD::UNDEF) | |||
6056 | continue; | |||
6057 | ||||
6058 | LoadSDNode *LD = cast<LoadSDNode>(Elt); | |||
6059 | if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i)) | |||
6060 | return SDValue(); | |||
6061 | LastLoadedElt = i; | |||
6062 | } | |||
6063 | ||||
6064 | // If we have found an entire vector of loads and undefs, then return a large | |||
6065 | // load of the entire vector width starting at the base pointer. If we found | |||
6066 | // consecutive loads for the low half, generate a vzext_load node. | |||
6067 | if (LastLoadedElt == NumElems - 1) { | |||
6068 | ||||
6069 | if (isAfterLegalize && | |||
6070 | !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT)) | |||
6071 | return SDValue(); | |||
6072 | ||||
6073 | SDValue NewLd = SDValue(); | |||
6074 | ||||
6075 | NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(), | |||
| ||||
6076 | LDBase->getPointerInfo(), LDBase->isVolatile(), | |||
6077 | LDBase->isNonTemporal(), LDBase->isInvariant(), | |||
6078 | LDBase->getAlignment()); | |||
6079 | ||||
6080 | if (LDBase->hasAnyUseOfValue(1)) { | |||
6081 | SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, | |||
6082 | SDValue(LDBase, 1), | |||
6083 | SDValue(NewLd.getNode(), 1)); | |||
6084 | DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain); | |||
6085 | DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1), | |||
6086 | SDValue(NewLd.getNode(), 1)); | |||
6087 | } | |||
6088 | ||||
6089 | return NewLd; | |||
6090 | } | |||
6091 | ||||
6092 | //TODO: The code below fires only for for loading the low v2i32 / v2f32 | |||
6093 | //of a v4i32 / v4f32. It's probably worth generalizing. | |||
6094 | if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) && | |||
6095 | DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) { | |||
6096 | SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other); | |||
6097 | SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() }; | |||
6098 | SDValue ResNode = | |||
6099 | DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64, | |||
6100 | LDBase->getPointerInfo(), | |||
6101 | LDBase->getAlignment(), | |||
6102 | false/*isVolatile*/, true/*ReadMem*/, | |||
6103 | false/*WriteMem*/); | |||
6104 | ||||
6105 | // Make sure the newly-created LOAD is in the same position as LDBase in | |||
6106 | // terms of dependency. We create a TokenFactor for LDBase and ResNode, and | |||
6107 | // update uses of LDBase's output chain to use the TokenFactor. | |||
6108 | if (LDBase->hasAnyUseOfValue(1)) { | |||
6109 | SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, | |||
6110 | SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1)); | |||
6111 | DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain); | |||
6112 | DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1), | |||
6113 | SDValue(ResNode.getNode(), 1)); | |||
6114 | } | |||
6115 | ||||
6116 | return DAG.getNode(ISD::BITCAST, DL, VT, ResNode); | |||
6117 | } | |||
6118 | return SDValue(); | |||
6119 | } | |||
6120 | ||||
6121 | /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction | |||
6122 | /// to generate a splat value for the following cases: | |||
6123 | /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant. | |||
6124 | /// 2. A splat shuffle which uses a scalar_to_vector node which comes from | |||
6125 | /// a scalar load, or a constant. | |||
6126 | /// The VBROADCAST node is returned when a pattern is found, | |||
6127 | /// or SDValue() otherwise. | |||
6128 | static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget, | |||
6129 | SelectionDAG &DAG) { | |||
6130 | // VBROADCAST requires AVX. | |||
6131 | // TODO: Splats could be generated for non-AVX CPUs using SSE | |||
6132 | // instructions, but there's less potential gain for only 128-bit vectors. | |||
6133 | if (!Subtarget->hasAVX()) | |||
6134 | return SDValue(); | |||
6135 | ||||
6136 | MVT VT = Op.getSimpleValueType(); | |||
6137 | SDLoc dl(Op); | |||
6138 | ||||
6139 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6140, __PRETTY_FUNCTION__)) | |||
6140 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6140, __PRETTY_FUNCTION__)); | |||
6141 | ||||
6142 | SDValue Ld; | |||
6143 | bool ConstSplatVal; | |||
6144 | ||||
6145 | switch (Op.getOpcode()) { | |||
6146 | default: | |||
6147 | // Unknown pattern found. | |||
6148 | return SDValue(); | |||
6149 | ||||
6150 | case ISD::BUILD_VECTOR: { | |||
6151 | auto *BVOp = cast<BuildVectorSDNode>(Op.getNode()); | |||
6152 | BitVector UndefElements; | |||
6153 | SDValue Splat = BVOp->getSplatValue(&UndefElements); | |||
6154 | ||||
6155 | // We need a splat of a single value to use broadcast, and it doesn't | |||
6156 | // make any sense if the value is only in one element of the vector. | |||
6157 | if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1) | |||
6158 | return SDValue(); | |||
6159 | ||||
6160 | Ld = Splat; | |||
6161 | ConstSplatVal = (Ld.getOpcode() == ISD::Constant || | |||
6162 | Ld.getOpcode() == ISD::ConstantFP); | |||
6163 | ||||
6164 | // Make sure that all of the users of a non-constant load are from the | |||
6165 | // BUILD_VECTOR node. | |||
6166 | if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode())) | |||
6167 | return SDValue(); | |||
6168 | break; | |||
6169 | } | |||
6170 | ||||
6171 | case ISD::VECTOR_SHUFFLE: { | |||
6172 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
6173 | ||||
6174 | // Shuffles must have a splat mask where the first element is | |||
6175 | // broadcasted. | |||
6176 | if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0) | |||
6177 | return SDValue(); | |||
6178 | ||||
6179 | SDValue Sc = Op.getOperand(0); | |||
6180 | if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR && | |||
6181 | Sc.getOpcode() != ISD::BUILD_VECTOR) { | |||
6182 | ||||
6183 | if (!Subtarget->hasInt256()) | |||
6184 | return SDValue(); | |||
6185 | ||||
6186 | // Use the register form of the broadcast instruction available on AVX2. | |||
6187 | if (VT.getSizeInBits() >= 256) | |||
6188 | Sc = Extract128BitVector(Sc, 0, DAG, dl); | |||
6189 | return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc); | |||
6190 | } | |||
6191 | ||||
6192 | Ld = Sc.getOperand(0); | |||
6193 | ConstSplatVal = (Ld.getOpcode() == ISD::Constant || | |||
6194 | Ld.getOpcode() == ISD::ConstantFP); | |||
6195 | ||||
6196 | // The scalar_to_vector node and the suspected | |||
6197 | // load node must have exactly one user. | |||
6198 | // Constants may have multiple users. | |||
6199 | ||||
6200 | // AVX-512 has register version of the broadcast | |||
6201 | bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() && | |||
6202 | Ld.getValueType().getSizeInBits() >= 32; | |||
6203 | if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) && | |||
6204 | !hasRegVer)) | |||
6205 | return SDValue(); | |||
6206 | break; | |||
6207 | } | |||
6208 | } | |||
6209 | ||||
6210 | unsigned ScalarSize = Ld.getValueType().getSizeInBits(); | |||
6211 | bool IsGE256 = (VT.getSizeInBits() >= 256); | |||
6212 | ||||
6213 | // When optimizing for size, generate up to 5 extra bytes for a broadcast | |||
6214 | // instruction to save 8 or more bytes of constant pool data. | |||
6215 | // TODO: If multiple splats are generated to load the same constant, | |||
6216 | // it may be detrimental to overall size. There needs to be a way to detect | |||
6217 | // that condition to know if this is truly a size win. | |||
6218 | const Function *F = DAG.getMachineFunction().getFunction(); | |||
6219 | bool OptForSize = F->getAttributes(). | |||
6220 | hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize); | |||
6221 | ||||
6222 | // Handle broadcasting a single constant scalar from the constant pool | |||
6223 | // into a vector. | |||
6224 | // On Sandybridge (no AVX2), it is still better to load a constant vector | |||
6225 | // from the constant pool and not to broadcast it from a scalar. | |||
6226 | // But override that restriction when optimizing for size. | |||
6227 | // TODO: Check if splatting is recommended for other AVX-capable CPUs. | |||
6228 | if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) { | |||
6229 | EVT CVT = Ld.getValueType(); | |||
6230 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6230, __PRETTY_FUNCTION__)); | |||
6231 | ||||
6232 | // Splat f32, i32, v4f64, v4i64 in all cases with AVX2. | |||
6233 | // For size optimization, also splat v2f64 and v2i64, and for size opt | |||
6234 | // with AVX2, also splat i8 and i16. | |||
6235 | // With pattern matching, the VBROADCAST node may become a VMOVDDUP. | |||
6236 | if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) || | |||
6237 | (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) { | |||
6238 | const Constant *C = nullptr; | |||
6239 | if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld)) | |||
6240 | C = CI->getConstantIntValue(); | |||
6241 | else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld)) | |||
6242 | C = CF->getConstantFPValue(); | |||
6243 | ||||
6244 | assert(C && "Invalid constant type")((C && "Invalid constant type") ? static_cast<void > (0) : __assert_fail ("C && \"Invalid constant type\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6244, __PRETTY_FUNCTION__)); | |||
6245 | ||||
6246 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
6247 | SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy()); | |||
6248 | unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment(); | |||
6249 | Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP, | |||
6250 | MachinePointerInfo::getConstantPool(), | |||
6251 | false, false, false, Alignment); | |||
6252 | ||||
6253 | return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld); | |||
6254 | } | |||
6255 | } | |||
6256 | ||||
6257 | bool IsLoad = ISD::isNormalLoad(Ld.getNode()); | |||
6258 | ||||
6259 | // Handle AVX2 in-register broadcasts. | |||
6260 | if (!IsLoad && Subtarget->hasInt256() && | |||
6261 | (ScalarSize == 32 || (IsGE256 && ScalarSize == 64))) | |||
6262 | return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld); | |||
6263 | ||||
6264 | // The scalar source must be a normal load. | |||
6265 | if (!IsLoad) | |||
6266 | return SDValue(); | |||
6267 | ||||
6268 | if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) || | |||
6269 | (Subtarget->hasVLX() && ScalarSize == 64)) | |||
6270 | return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld); | |||
6271 | ||||
6272 | // The integer check is needed for the 64-bit into 128-bit so it doesn't match | |||
6273 | // double since there is no vbroadcastsd xmm | |||
6274 | if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) { | |||
6275 | if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64) | |||
6276 | return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld); | |||
6277 | } | |||
6278 | ||||
6279 | // Unsupported broadcast. | |||
6280 | return SDValue(); | |||
6281 | } | |||
6282 | ||||
6283 | /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real | |||
6284 | /// underlying vector and index. | |||
6285 | /// | |||
6286 | /// Modifies \p ExtractedFromVec to the real vector and returns the real | |||
6287 | /// index. | |||
6288 | static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec, | |||
6289 | SDValue ExtIdx) { | |||
6290 | int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue(); | |||
6291 | if (!isa<ShuffleVectorSDNode>(ExtractedFromVec)) | |||
6292 | return Idx; | |||
6293 | ||||
6294 | // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already | |||
6295 | // lowered this: | |||
6296 | // (extract_vector_elt (v8f32 %vreg1), Constant<6>) | |||
6297 | // to: | |||
6298 | // (extract_vector_elt (vector_shuffle<2,u,u,u> | |||
6299 | // (extract_subvector (v8f32 %vreg0), Constant<4>), | |||
6300 | // undef) | |||
6301 | // Constant<0>) | |||
6302 | // In this case the vector is the extract_subvector expression and the index | |||
6303 | // is 2, as specified by the shuffle. | |||
6304 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec); | |||
6305 | SDValue ShuffleVec = SVOp->getOperand(0); | |||
6306 | MVT ShuffleVecVT = ShuffleVec.getSimpleValueType(); | |||
6307 | assert(ShuffleVecVT.getVectorElementType() ==((ShuffleVecVT.getVectorElementType() == ExtractedFromVec.getSimpleValueType ().getVectorElementType()) ? static_cast<void> (0) : __assert_fail ("ShuffleVecVT.getVectorElementType() == ExtractedFromVec.getSimpleValueType().getVectorElementType()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6308, __PRETTY_FUNCTION__)) | |||
6308 | ExtractedFromVec.getSimpleValueType().getVectorElementType())((ShuffleVecVT.getVectorElementType() == ExtractedFromVec.getSimpleValueType ().getVectorElementType()) ? static_cast<void> (0) : __assert_fail ("ShuffleVecVT.getVectorElementType() == ExtractedFromVec.getSimpleValueType().getVectorElementType()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6308, __PRETTY_FUNCTION__)); | |||
6309 | ||||
6310 | int ShuffleIdx = SVOp->getMaskElt(Idx); | |||
6311 | if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) { | |||
6312 | ExtractedFromVec = ShuffleVec; | |||
6313 | return ShuffleIdx; | |||
6314 | } | |||
6315 | return Idx; | |||
6316 | } | |||
6317 | ||||
6318 | static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) { | |||
6319 | MVT VT = Op.getSimpleValueType(); | |||
6320 | ||||
6321 | // Skip if insert_vec_elt is not supported. | |||
6322 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
6323 | if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT)) | |||
6324 | return SDValue(); | |||
6325 | ||||
6326 | SDLoc DL(Op); | |||
6327 | unsigned NumElems = Op.getNumOperands(); | |||
6328 | ||||
6329 | SDValue VecIn1; | |||
6330 | SDValue VecIn2; | |||
6331 | SmallVector<unsigned, 4> InsertIndices; | |||
6332 | SmallVector<int, 8> Mask(NumElems, -1); | |||
6333 | ||||
6334 | for (unsigned i = 0; i != NumElems; ++i) { | |||
6335 | unsigned Opc = Op.getOperand(i).getOpcode(); | |||
6336 | ||||
6337 | if (Opc == ISD::UNDEF) | |||
6338 | continue; | |||
6339 | ||||
6340 | if (Opc != ISD::EXTRACT_VECTOR_ELT) { | |||
6341 | // Quit if more than 1 elements need inserting. | |||
6342 | if (InsertIndices.size() > 1) | |||
6343 | return SDValue(); | |||
6344 | ||||
6345 | InsertIndices.push_back(i); | |||
6346 | continue; | |||
6347 | } | |||
6348 | ||||
6349 | SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0); | |||
6350 | SDValue ExtIdx = Op.getOperand(i).getOperand(1); | |||
6351 | // Quit if non-constant index. | |||
6352 | if (!isa<ConstantSDNode>(ExtIdx)) | |||
6353 | return SDValue(); | |||
6354 | int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx); | |||
6355 | ||||
6356 | // Quit if extracted from vector of different type. | |||
6357 | if (ExtractedFromVec.getValueType() != VT) | |||
6358 | return SDValue(); | |||
6359 | ||||
6360 | if (!VecIn1.getNode()) | |||
6361 | VecIn1 = ExtractedFromVec; | |||
6362 | else if (VecIn1 != ExtractedFromVec) { | |||
6363 | if (!VecIn2.getNode()) | |||
6364 | VecIn2 = ExtractedFromVec; | |||
6365 | else if (VecIn2 != ExtractedFromVec) | |||
6366 | // Quit if more than 2 vectors to shuffle | |||
6367 | return SDValue(); | |||
6368 | } | |||
6369 | ||||
6370 | if (ExtractedFromVec == VecIn1) | |||
6371 | Mask[i] = Idx; | |||
6372 | else if (ExtractedFromVec == VecIn2) | |||
6373 | Mask[i] = Idx + NumElems; | |||
6374 | } | |||
6375 | ||||
6376 | if (!VecIn1.getNode()) | |||
6377 | return SDValue(); | |||
6378 | ||||
6379 | VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT); | |||
6380 | SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]); | |||
6381 | for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) { | |||
6382 | unsigned Idx = InsertIndices[i]; | |||
6383 | NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx), | |||
6384 | DAG.getIntPtrConstant(Idx)); | |||
6385 | } | |||
6386 | ||||
6387 | return NV; | |||
6388 | } | |||
6389 | ||||
6390 | // Lower BUILD_VECTOR operation for v8i1 and v16i1 types. | |||
6391 | SDValue | |||
6392 | X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const { | |||
6393 | ||||
6394 | MVT VT = Op.getSimpleValueType(); | |||
6395 | assert((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) &&(((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits () <= 16) && "Unexpected type in LowerBUILD_VECTORvXi1!" ) ? static_cast<void> (0) : __assert_fail ("(VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) && \"Unexpected type in LowerBUILD_VECTORvXi1!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6396, __PRETTY_FUNCTION__)) | |||
6396 | "Unexpected type in LowerBUILD_VECTORvXi1!")(((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits () <= 16) && "Unexpected type in LowerBUILD_VECTORvXi1!" ) ? static_cast<void> (0) : __assert_fail ("(VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) && \"Unexpected type in LowerBUILD_VECTORvXi1!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6396, __PRETTY_FUNCTION__)); | |||
6397 | ||||
6398 | SDLoc dl(Op); | |||
6399 | if (ISD::isBuildVectorAllZeros(Op.getNode())) { | |||
6400 | SDValue Cst = DAG.getTargetConstant(0, MVT::i1); | |||
6401 | SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst); | |||
6402 | return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops); | |||
6403 | } | |||
6404 | ||||
6405 | if (ISD::isBuildVectorAllOnes(Op.getNode())) { | |||
6406 | SDValue Cst = DAG.getTargetConstant(1, MVT::i1); | |||
6407 | SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst); | |||
6408 | return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops); | |||
6409 | } | |||
6410 | ||||
6411 | bool AllContants = true; | |||
6412 | uint64_t Immediate = 0; | |||
6413 | int NonConstIdx = -1; | |||
6414 | bool IsSplat = true; | |||
6415 | unsigned NumNonConsts = 0; | |||
6416 | unsigned NumConsts = 0; | |||
6417 | for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) { | |||
6418 | SDValue In = Op.getOperand(idx); | |||
6419 | if (In.getOpcode() == ISD::UNDEF) | |||
6420 | continue; | |||
6421 | if (!isa<ConstantSDNode>(In)) { | |||
6422 | AllContants = false; | |||
6423 | NonConstIdx = idx; | |||
6424 | NumNonConsts++; | |||
6425 | } else { | |||
6426 | NumConsts++; | |||
6427 | if (cast<ConstantSDNode>(In)->getZExtValue()) | |||
6428 | Immediate |= (1ULL << idx); | |||
6429 | } | |||
6430 | if (In != Op.getOperand(0)) | |||
6431 | IsSplat = false; | |||
6432 | } | |||
6433 | ||||
6434 | if (AllContants) { | |||
6435 | SDValue FullMask = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, | |||
6436 | DAG.getConstant(Immediate, MVT::i16)); | |||
6437 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, FullMask, | |||
6438 | DAG.getIntPtrConstant(0)); | |||
6439 | } | |||
6440 | ||||
6441 | if (NumNonConsts == 1 && NonConstIdx != 0) { | |||
6442 | SDValue DstVec; | |||
6443 | if (NumConsts) { | |||
6444 | SDValue VecAsImm = DAG.getConstant(Immediate, | |||
6445 | MVT::getIntegerVT(VT.getSizeInBits())); | |||
6446 | DstVec = DAG.getNode(ISD::BITCAST, dl, VT, VecAsImm); | |||
6447 | } | |||
6448 | else | |||
6449 | DstVec = DAG.getUNDEF(VT); | |||
6450 | return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec, | |||
6451 | Op.getOperand(NonConstIdx), | |||
6452 | DAG.getIntPtrConstant(NonConstIdx)); | |||
6453 | } | |||
6454 | if (!IsSplat && (NonConstIdx != 0)) | |||
6455 | llvm_unreachable("Unsupported BUILD_VECTOR operation")::llvm::llvm_unreachable_internal("Unsupported BUILD_VECTOR operation" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6455); | |||
6456 | MVT SelectVT = (VT == MVT::v16i1)? MVT::i16 : MVT::i8; | |||
6457 | SDValue Select; | |||
6458 | if (IsSplat) | |||
6459 | Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0), | |||
6460 | DAG.getConstant(-1, SelectVT), | |||
6461 | DAG.getConstant(0, SelectVT)); | |||
6462 | else | |||
6463 | Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0), | |||
6464 | DAG.getConstant((Immediate | 1), SelectVT), | |||
6465 | DAG.getConstant(Immediate, SelectVT)); | |||
6466 | return DAG.getNode(ISD::BITCAST, dl, VT, Select); | |||
6467 | } | |||
6468 | ||||
6469 | /// \brief Return true if \p N implements a horizontal binop and return the | |||
6470 | /// operands for the horizontal binop into V0 and V1. | |||
6471 | /// | |||
6472 | /// This is a helper function of PerformBUILD_VECTORCombine. | |||
6473 | /// This function checks that the build_vector \p N in input implements a | |||
6474 | /// horizontal operation. Parameter \p Opcode defines the kind of horizontal | |||
6475 | /// operation to match. | |||
6476 | /// For example, if \p Opcode is equal to ISD::ADD, then this function | |||
6477 | /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode | |||
6478 | /// is equal to ISD::SUB, then this function checks if this is a horizontal | |||
6479 | /// arithmetic sub. | |||
6480 | /// | |||
6481 | /// This function only analyzes elements of \p N whose indices are | |||
6482 | /// in range [BaseIdx, LastIdx). | |||
6483 | static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode, | |||
6484 | SelectionDAG &DAG, | |||
6485 | unsigned BaseIdx, unsigned LastIdx, | |||
6486 | SDValue &V0, SDValue &V1) { | |||
6487 | EVT VT = N->getValueType(0); | |||
6488 | ||||
6489 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6489, __PRETTY_FUNCTION__)); | |||
6490 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6491, __PRETTY_FUNCTION__)) | |||
6491 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6491, __PRETTY_FUNCTION__)); | |||
6492 | ||||
6493 | bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD); | |||
6494 | bool CanFold = true; | |||
6495 | unsigned ExpectedVExtractIdx = BaseIdx; | |||
6496 | unsigned NumElts = LastIdx - BaseIdx; | |||
6497 | V0 = DAG.getUNDEF(VT); | |||
6498 | V1 = DAG.getUNDEF(VT); | |||
6499 | ||||
6500 | // Check if N implements a horizontal binop. | |||
6501 | for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) { | |||
6502 | SDValue Op = N->getOperand(i + BaseIdx); | |||
6503 | ||||
6504 | // Skip UNDEFs. | |||
6505 | if (Op->getOpcode() == ISD::UNDEF) { | |||
6506 | // Update the expected vector extract index. | |||
6507 | if (i * 2 == NumElts) | |||
6508 | ExpectedVExtractIdx = BaseIdx; | |||
6509 | ExpectedVExtractIdx += 2; | |||
6510 | continue; | |||
6511 | } | |||
6512 | ||||
6513 | CanFold = Op->getOpcode() == Opcode && Op->hasOneUse(); | |||
6514 | ||||
6515 | if (!CanFold) | |||
6516 | break; | |||
6517 | ||||
6518 | SDValue Op0 = Op.getOperand(0); | |||
6519 | SDValue Op1 = Op.getOperand(1); | |||
6520 | ||||
6521 | // Try to match the following pattern: | |||
6522 | // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1)) | |||
6523 | CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT && | |||
6524 | Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT && | |||
6525 | Op0.getOperand(0) == Op1.getOperand(0) && | |||
6526 | isa<ConstantSDNode>(Op0.getOperand(1)) && | |||
6527 | isa<ConstantSDNode>(Op1.getOperand(1))); | |||
6528 | if (!CanFold) | |||
6529 | break; | |||
6530 | ||||
6531 | unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue(); | |||
6532 | unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue(); | |||
6533 | ||||
6534 | if (i * 2 < NumElts) { | |||
6535 | if (V0.getOpcode() == ISD::UNDEF) | |||
6536 | V0 = Op0.getOperand(0); | |||
6537 | } else { | |||
6538 | if (V1.getOpcode() == ISD::UNDEF) | |||
6539 | V1 = Op0.getOperand(0); | |||
6540 | if (i * 2 == NumElts) | |||
6541 | ExpectedVExtractIdx = BaseIdx; | |||
6542 | } | |||
6543 | ||||
6544 | SDValue Expected = (i * 2 < NumElts) ? V0 : V1; | |||
6545 | if (I0 == ExpectedVExtractIdx) | |||
6546 | CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected; | |||
6547 | else if (IsCommutable && I1 == ExpectedVExtractIdx) { | |||
6548 | // Try to match the following dag sequence: | |||
6549 | // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I)) | |||
6550 | CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected; | |||
6551 | } else | |||
6552 | CanFold = false; | |||
6553 | ||||
6554 | ExpectedVExtractIdx += 2; | |||
6555 | } | |||
6556 | ||||
6557 | return CanFold; | |||
6558 | } | |||
6559 | ||||
6560 | /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by | |||
6561 | /// a concat_vector. | |||
6562 | /// | |||
6563 | /// This is a helper function of PerformBUILD_VECTORCombine. | |||
6564 | /// This function expects two 256-bit vectors called V0 and V1. | |||
6565 | /// At first, each vector is split into two separate 128-bit vectors. | |||
6566 | /// Then, the resulting 128-bit vectors are used to implement two | |||
6567 | /// horizontal binary operations. | |||
6568 | /// | |||
6569 | /// The kind of horizontal binary operation is defined by \p X86Opcode. | |||
6570 | /// | |||
6571 | /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to | |||
6572 | /// the two new horizontal binop. | |||
6573 | /// When Mode is set, the first horizontal binop dag node would take as input | |||
6574 | /// the lower 128-bit of V0 and the upper 128-bit of V0. The second | |||
6575 | /// horizontal binop dag node would take as input the lower 128-bit of V1 | |||
6576 | /// and the upper 128-bit of V1. | |||
6577 | /// Example: | |||
6578 | /// HADD V0_LO, V0_HI | |||
6579 | /// HADD V1_LO, V1_HI | |||
6580 | /// | |||
6581 | /// Otherwise, the first horizontal binop dag node takes as input the lower | |||
6582 | /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop | |||
6583 | /// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1. | |||
6584 | /// Example: | |||
6585 | /// HADD V0_LO, V1_LO | |||
6586 | /// HADD V0_HI, V1_HI | |||
6587 | /// | |||
6588 | /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower | |||
6589 | /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to | |||
6590 | /// the upper 128-bits of the result. | |||
6591 | static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1, | |||
6592 | SDLoc DL, SelectionDAG &DAG, | |||
6593 | unsigned X86Opcode, bool Mode, | |||
6594 | bool isUndefLO, bool isUndefHI) { | |||
6595 | EVT VT = V0.getValueType(); | |||
6596 | assert(VT.is256BitVector() && VT == V1.getValueType() &&((VT.is256BitVector() && VT == V1.getValueType() && "Invalid nodes in input!") ? static_cast<void> (0) : __assert_fail ("VT.is256BitVector() && VT == V1.getValueType() && \"Invalid nodes in input!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6597, __PRETTY_FUNCTION__)) | |||
6597 | "Invalid nodes in input!")((VT.is256BitVector() && VT == V1.getValueType() && "Invalid nodes in input!") ? static_cast<void> (0) : __assert_fail ("VT.is256BitVector() && VT == V1.getValueType() && \"Invalid nodes in input!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6597, __PRETTY_FUNCTION__)); | |||
6598 | ||||
6599 | unsigned NumElts = VT.getVectorNumElements(); | |||
6600 | SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL); | |||
6601 | SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL); | |||
6602 | SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL); | |||
6603 | SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL); | |||
6604 | EVT NewVT = V0_LO.getValueType(); | |||
6605 | ||||
6606 | SDValue LO = DAG.getUNDEF(NewVT); | |||
6607 | SDValue HI = DAG.getUNDEF(NewVT); | |||
6608 | ||||
6609 | if (Mode) { | |||
6610 | // Don't emit a horizontal binop if the result is expected to be UNDEF. | |||
6611 | if (!isUndefLO && V0->getOpcode() != ISD::UNDEF) | |||
6612 | LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI); | |||
6613 | if (!isUndefHI && V1->getOpcode() != ISD::UNDEF) | |||
6614 | HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI); | |||
6615 | } else { | |||
6616 | // Don't emit a horizontal binop if the result is expected to be UNDEF. | |||
6617 | if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF || | |||
6618 | V1_LO->getOpcode() != ISD::UNDEF)) | |||
6619 | LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO); | |||
6620 | ||||
6621 | if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF || | |||
6622 | V1_HI->getOpcode() != ISD::UNDEF)) | |||
6623 | HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI); | |||
6624 | } | |||
6625 | ||||
6626 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI); | |||
6627 | } | |||
6628 | ||||
6629 | /// \brief Try to fold a build_vector that performs an 'addsub' into the | |||
6630 | /// sequence of 'vadd + vsub + blendi'. | |||
6631 | static SDValue matchAddSub(const BuildVectorSDNode *BV, SelectionDAG &DAG, | |||
6632 | const X86Subtarget *Subtarget) { | |||
6633 | SDLoc DL(BV); | |||
6634 | EVT VT = BV->getValueType(0); | |||
6635 | unsigned NumElts = VT.getVectorNumElements(); | |||
6636 | SDValue InVec0 = DAG.getUNDEF(VT); | |||
6637 | SDValue InVec1 = DAG.getUNDEF(VT); | |||
6638 | ||||
6639 | assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||(((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 || VT == MVT::v2f64) && "build_vector with an invalid type found!" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 || VT == MVT::v2f64) && \"build_vector with an invalid type found!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6640, __PRETTY_FUNCTION__)) | |||
6640 | VT == MVT::v2f64) && "build_vector with an invalid type found!")(((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 || VT == MVT::v2f64) && "build_vector with an invalid type found!" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 || VT == MVT::v2f64) && \"build_vector with an invalid type found!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6640, __PRETTY_FUNCTION__)); | |||
6641 | ||||
6642 | // Odd-numbered elements in the input build vector are obtained from | |||
6643 | // adding two integer/float elements. | |||
6644 | // Even-numbered elements in the input build vector are obtained from | |||
6645 | // subtracting two integer/float elements. | |||
6646 | unsigned ExpectedOpcode = ISD::FSUB; | |||
6647 | unsigned NextExpectedOpcode = ISD::FADD; | |||
6648 | bool AddFound = false; | |||
6649 | bool SubFound = false; | |||
6650 | ||||
6651 | for (unsigned i = 0, e = NumElts; i != e; i++) { | |||
6652 | SDValue Op = BV->getOperand(i); | |||
6653 | ||||
6654 | // Skip 'undef' values. | |||
6655 | unsigned Opcode = Op.getOpcode(); | |||
6656 | if (Opcode == ISD::UNDEF) { | |||
6657 | std::swap(ExpectedOpcode, NextExpectedOpcode); | |||
6658 | continue; | |||
6659 | } | |||
6660 | ||||
6661 | // Early exit if we found an unexpected opcode. | |||
6662 | if (Opcode != ExpectedOpcode) | |||
6663 | return SDValue(); | |||
6664 | ||||
6665 | SDValue Op0 = Op.getOperand(0); | |||
6666 | SDValue Op1 = Op.getOperand(1); | |||
6667 | ||||
6668 | // Try to match the following pattern: | |||
6669 | // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i)) | |||
6670 | // Early exit if we cannot match that sequence. | |||
6671 | if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT || | |||
6672 | Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT || | |||
6673 | !isa<ConstantSDNode>(Op0.getOperand(1)) || | |||
6674 | !isa<ConstantSDNode>(Op1.getOperand(1)) || | |||
6675 | Op0.getOperand(1) != Op1.getOperand(1)) | |||
6676 | return SDValue(); | |||
6677 | ||||
6678 | unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue(); | |||
6679 | if (I0 != i) | |||
6680 | return SDValue(); | |||
6681 | ||||
6682 | // We found a valid add/sub node. Update the information accordingly. | |||
6683 | if (i & 1) | |||
6684 | AddFound = true; | |||
6685 | else | |||
6686 | SubFound = true; | |||
6687 | ||||
6688 | // Update InVec0 and InVec1. | |||
6689 | if (InVec0.getOpcode() == ISD::UNDEF) | |||
6690 | InVec0 = Op0.getOperand(0); | |||
6691 | if (InVec1.getOpcode() == ISD::UNDEF) | |||
6692 | InVec1 = Op1.getOperand(0); | |||
6693 | ||||
6694 | // Make sure that operands in input to each add/sub node always | |||
6695 | // come from a same pair of vectors. | |||
6696 | if (InVec0 != Op0.getOperand(0)) { | |||
6697 | if (ExpectedOpcode == ISD::FSUB) | |||
6698 | return SDValue(); | |||
6699 | ||||
6700 | // FADD is commutable. Try to commute the operands | |||
6701 | // and then test again. | |||
6702 | std::swap(Op0, Op1); | |||
6703 | if (InVec0 != Op0.getOperand(0)) | |||
6704 | return SDValue(); | |||
6705 | } | |||
6706 | ||||
6707 | if (InVec1 != Op1.getOperand(0)) | |||
6708 | return SDValue(); | |||
6709 | ||||
6710 | // Update the pair of expected opcodes. | |||
6711 | std::swap(ExpectedOpcode, NextExpectedOpcode); | |||
6712 | } | |||
6713 | ||||
6714 | // Don't try to fold this build_vector into an ADDSUB if the inputs are undef. | |||
6715 | if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF && | |||
6716 | InVec1.getOpcode() != ISD::UNDEF) | |||
6717 | return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1); | |||
6718 | ||||
6719 | return SDValue(); | |||
6720 | } | |||
6721 | ||||
6722 | static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG, | |||
6723 | const X86Subtarget *Subtarget) { | |||
6724 | SDLoc DL(N); | |||
6725 | EVT VT = N->getValueType(0); | |||
6726 | unsigned NumElts = VT.getVectorNumElements(); | |||
6727 | BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N); | |||
6728 | SDValue InVec0, InVec1; | |||
6729 | ||||
6730 | // Try to match an ADDSUB. | |||
6731 | if ((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) || | |||
6732 | (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) { | |||
6733 | SDValue Value = matchAddSub(BV, DAG, Subtarget); | |||
6734 | if (Value.getNode()) | |||
6735 | return Value; | |||
6736 | } | |||
6737 | ||||
6738 | // Try to match horizontal ADD/SUB. | |||
6739 | unsigned NumUndefsLO = 0; | |||
6740 | unsigned NumUndefsHI = 0; | |||
6741 | unsigned Half = NumElts/2; | |||
6742 | ||||
6743 | // Count the number of UNDEF operands in the build_vector in input. | |||
6744 | for (unsigned i = 0, e = Half; i != e; ++i) | |||
6745 | if (BV->getOperand(i)->getOpcode() == ISD::UNDEF) | |||
6746 | NumUndefsLO++; | |||
6747 | ||||
6748 | for (unsigned i = Half, e = NumElts; i != e; ++i) | |||
6749 | if (BV->getOperand(i)->getOpcode() == ISD::UNDEF) | |||
6750 | NumUndefsHI++; | |||
6751 | ||||
6752 | // Early exit if this is either a build_vector of all UNDEFs or all the | |||
6753 | // operands but one are UNDEF. | |||
6754 | if (NumUndefsLO + NumUndefsHI + 1 >= NumElts) | |||
6755 | return SDValue(); | |||
6756 | ||||
6757 | if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) { | |||
6758 | // Try to match an SSE3 float HADD/HSUB. | |||
6759 | if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1)) | |||
6760 | return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1); | |||
6761 | ||||
6762 | if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1)) | |||
6763 | return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1); | |||
6764 | } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) { | |||
6765 | // Try to match an SSSE3 integer HADD/HSUB. | |||
6766 | if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1)) | |||
6767 | return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1); | |||
6768 | ||||
6769 | if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1)) | |||
6770 | return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1); | |||
6771 | } | |||
6772 | ||||
6773 | if (!Subtarget->hasAVX()) | |||
6774 | return SDValue(); | |||
6775 | ||||
6776 | if ((VT == MVT::v8f32 || VT == MVT::v4f64)) { | |||
6777 | // Try to match an AVX horizontal add/sub of packed single/double | |||
6778 | // precision floating point values from 256-bit vectors. | |||
6779 | SDValue InVec2, InVec3; | |||
6780 | if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) && | |||
6781 | isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) && | |||
6782 | ((InVec0.getOpcode() == ISD::UNDEF || | |||
6783 | InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) && | |||
6784 | ((InVec1.getOpcode() == ISD::UNDEF || | |||
6785 | InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3)) | |||
6786 | return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1); | |||
6787 | ||||
6788 | if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) && | |||
6789 | isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) && | |||
6790 | ((InVec0.getOpcode() == ISD::UNDEF || | |||
6791 | InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) && | |||
6792 | ((InVec1.getOpcode() == ISD::UNDEF || | |||
6793 | InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3)) | |||
6794 | return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1); | |||
6795 | } else if (VT == MVT::v8i32 || VT == MVT::v16i16) { | |||
6796 | // Try to match an AVX2 horizontal add/sub of signed integers. | |||
6797 | SDValue InVec2, InVec3; | |||
6798 | unsigned X86Opcode; | |||
6799 | bool CanFold = true; | |||
6800 | ||||
6801 | if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) && | |||
6802 | isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) && | |||
6803 | ((InVec0.getOpcode() == ISD::UNDEF || | |||
6804 | InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) && | |||
6805 | ((InVec1.getOpcode() == ISD::UNDEF || | |||
6806 | InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3)) | |||
6807 | X86Opcode = X86ISD::HADD; | |||
6808 | else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) && | |||
6809 | isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) && | |||
6810 | ((InVec0.getOpcode() == ISD::UNDEF || | |||
6811 | InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) && | |||
6812 | ((InVec1.getOpcode() == ISD::UNDEF || | |||
6813 | InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3)) | |||
6814 | X86Opcode = X86ISD::HSUB; | |||
6815 | else | |||
6816 | CanFold = false; | |||
6817 | ||||
6818 | if (CanFold) { | |||
6819 | // Fold this build_vector into a single horizontal add/sub. | |||
6820 | // Do this only if the target has AVX2. | |||
6821 | if (Subtarget->hasAVX2()) | |||
6822 | return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1); | |||
6823 | ||||
6824 | // Do not try to expand this build_vector into a pair of horizontal | |||
6825 | // add/sub if we can emit a pair of scalar add/sub. | |||
6826 | if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half) | |||
6827 | return SDValue(); | |||
6828 | ||||
6829 | // Convert this build_vector into a pair of horizontal binop followed by | |||
6830 | // a concat vector. | |||
6831 | bool isUndefLO = NumUndefsLO == Half; | |||
6832 | bool isUndefHI = NumUndefsHI == Half; | |||
6833 | return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false, | |||
6834 | isUndefLO, isUndefHI); | |||
6835 | } | |||
6836 | } | |||
6837 | ||||
6838 | if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 || | |||
6839 | VT == MVT::v16i16) && Subtarget->hasAVX()) { | |||
6840 | unsigned X86Opcode; | |||
6841 | if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1)) | |||
6842 | X86Opcode = X86ISD::HADD; | |||
6843 | else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1)) | |||
6844 | X86Opcode = X86ISD::HSUB; | |||
6845 | else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1)) | |||
6846 | X86Opcode = X86ISD::FHADD; | |||
6847 | else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1)) | |||
6848 | X86Opcode = X86ISD::FHSUB; | |||
6849 | else | |||
6850 | return SDValue(); | |||
6851 | ||||
6852 | // Don't try to expand this build_vector into a pair of horizontal add/sub | |||
6853 | // if we can simply emit a pair of scalar add/sub. | |||
6854 | if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half) | |||
6855 | return SDValue(); | |||
6856 | ||||
6857 | // Convert this build_vector into two horizontal add/sub followed by | |||
6858 | // a concat vector. | |||
6859 | bool isUndefLO = NumUndefsLO == Half; | |||
6860 | bool isUndefHI = NumUndefsHI == Half; | |||
6861 | return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true, | |||
6862 | isUndefLO, isUndefHI); | |||
6863 | } | |||
6864 | ||||
6865 | return SDValue(); | |||
6866 | } | |||
6867 | ||||
6868 | SDValue | |||
6869 | X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const { | |||
6870 | SDLoc dl(Op); | |||
6871 | ||||
6872 | MVT VT = Op.getSimpleValueType(); | |||
6873 | MVT ExtVT = VT.getVectorElementType(); | |||
6874 | unsigned NumElems = Op.getNumOperands(); | |||
6875 | ||||
6876 | // Generate vectors for predicate vectors. | |||
6877 | if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512()) | |||
6878 | return LowerBUILD_VECTORvXi1(Op, DAG); | |||
6879 | ||||
6880 | // Vectors containing all zeros can be matched by pxor and xorps later | |||
6881 | if (ISD::isBuildVectorAllZeros(Op.getNode())) { | |||
6882 | // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd | |||
6883 | // and 2) ensure that i64 scalars are eliminated on x86-32 hosts. | |||
6884 | if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32) | |||
6885 | return Op; | |||
6886 | ||||
6887 | return getZeroVector(VT, Subtarget, DAG, dl); | |||
6888 | } | |||
6889 | ||||
6890 | // Vectors containing all ones can be matched by pcmpeqd on 128-bit width | |||
6891 | // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use | |||
6892 | // vpcmpeqd on 256-bit vectors. | |||
6893 | if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) { | |||
6894 | if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256())) | |||
6895 | return Op; | |||
6896 | ||||
6897 | if (!VT.is512BitVector()) | |||
6898 | return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl); | |||
6899 | } | |||
6900 | ||||
6901 | SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG); | |||
6902 | if (Broadcast.getNode()) | |||
6903 | return Broadcast; | |||
6904 | ||||
6905 | unsigned EVTBits = ExtVT.getSizeInBits(); | |||
6906 | ||||
6907 | unsigned NumZero = 0; | |||
6908 | unsigned NumNonZero = 0; | |||
6909 | unsigned NonZeros = 0; | |||
6910 | bool IsAllConstants = true; | |||
6911 | SmallSet<SDValue, 8> Values; | |||
6912 | for (unsigned i = 0; i < NumElems; ++i) { | |||
6913 | SDValue Elt = Op.getOperand(i); | |||
6914 | if (Elt.getOpcode() == ISD::UNDEF) | |||
6915 | continue; | |||
6916 | Values.insert(Elt); | |||
6917 | if (Elt.getOpcode() != ISD::Constant && | |||
6918 | Elt.getOpcode() != ISD::ConstantFP) | |||
6919 | IsAllConstants = false; | |||
6920 | if (X86::isZeroNode(Elt)) | |||
6921 | NumZero++; | |||
6922 | else { | |||
6923 | NonZeros |= (1 << i); | |||
6924 | NumNonZero++; | |||
6925 | } | |||
6926 | } | |||
6927 | ||||
6928 | // All undef vector. Return an UNDEF. All zero vectors were handled above. | |||
6929 | if (NumNonZero == 0) | |||
6930 | return DAG.getUNDEF(VT); | |||
6931 | ||||
6932 | // Special case for single non-zero, non-undef, element. | |||
6933 | if (NumNonZero == 1) { | |||
6934 | unsigned Idx = countTrailingZeros(NonZeros); | |||
6935 | SDValue Item = Op.getOperand(Idx); | |||
6936 | ||||
6937 | // If this is an insertion of an i64 value on x86-32, and if the top bits of | |||
6938 | // the value are obviously zero, truncate the value to i32 and do the | |||
6939 | // insertion that way. Only do this if the value is non-constant or if the | |||
6940 | // value is a constant being inserted into element 0. It is cheaper to do | |||
6941 | // a constant pool load than it is to do a movd + shuffle. | |||
6942 | if (ExtVT == MVT::i64 && !Subtarget->is64Bit() && | |||
6943 | (!IsAllConstants || Idx == 0)) { | |||
6944 | if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) { | |||
6945 | // Handle SSE only. | |||
6946 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6946, __PRETTY_FUNCTION__)); | |||
6947 | EVT VecVT = MVT::v4i32; | |||
6948 | unsigned VecElts = 4; | |||
6949 | ||||
6950 | // Truncate the value (which may itself be a constant) to i32, and | |||
6951 | // convert it to a vector with movd (S2V+shuffle to zero extend). | |||
6952 | Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item); | |||
6953 | Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item); | |||
6954 | ||||
6955 | // If using the new shuffle lowering, just directly insert this. | |||
6956 | if (ExperimentalVectorShuffleLowering) | |||
6957 | return DAG.getNode( | |||
6958 | ISD::BITCAST, dl, VT, | |||
6959 | getShuffleVectorZeroOrUndef(Item, Idx * 2, true, Subtarget, DAG)); | |||
6960 | ||||
6961 | Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG); | |||
6962 | ||||
6963 | // Now we have our 32-bit value zero extended in the low element of | |||
6964 | // a vector. If Idx != 0, swizzle it into place. | |||
6965 | if (Idx != 0) { | |||
6966 | SmallVector<int, 4> Mask; | |||
6967 | Mask.push_back(Idx); | |||
6968 | for (unsigned i = 1; i != VecElts; ++i) | |||
6969 | Mask.push_back(i); | |||
6970 | Item = DAG.getVectorShuffle(VecVT, dl, Item, DAG.getUNDEF(VecVT), | |||
6971 | &Mask[0]); | |||
6972 | } | |||
6973 | return DAG.getNode(ISD::BITCAST, dl, VT, Item); | |||
6974 | } | |||
6975 | } | |||
6976 | ||||
6977 | // If we have a constant or non-constant insertion into the low element of | |||
6978 | // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into | |||
6979 | // the rest of the elements. This will be matched as movd/movq/movss/movsd | |||
6980 | // depending on what the source datatype is. | |||
6981 | if (Idx == 0) { | |||
6982 | if (NumZero == 0) | |||
6983 | return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item); | |||
6984 | ||||
6985 | if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 || | |||
6986 | (ExtVT == MVT::i64 && Subtarget->is64Bit())) { | |||
6987 | if (VT.is256BitVector() || VT.is512BitVector()) { | |||
6988 | SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl); | |||
6989 | return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec, | |||
6990 | Item, DAG.getIntPtrConstant(0)); | |||
6991 | } | |||
6992 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 6992, __PRETTY_FUNCTION__)); | |||
6993 | Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item); | |||
6994 | // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector. | |||
6995 | return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG); | |||
6996 | } | |||
6997 | ||||
6998 | if (ExtVT == MVT::i16 || ExtVT == MVT::i8) { | |||
6999 | Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item); | |||
7000 | Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item); | |||
7001 | if (VT.is256BitVector()) { | |||
7002 | SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl); | |||
7003 | Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl); | |||
7004 | } else { | |||
7005 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7005, __PRETTY_FUNCTION__)); | |||
7006 | Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG); | |||
7007 | } | |||
7008 | return DAG.getNode(ISD::BITCAST, dl, VT, Item); | |||
7009 | } | |||
7010 | } | |||
7011 | ||||
7012 | // Is it a vector logical left shift? | |||
7013 | if (NumElems == 2 && Idx == 1 && | |||
7014 | X86::isZeroNode(Op.getOperand(0)) && | |||
7015 | !X86::isZeroNode(Op.getOperand(1))) { | |||
7016 | unsigned NumBits = VT.getSizeInBits(); | |||
7017 | return getVShift(true, VT, | |||
7018 | DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, | |||
7019 | VT, Op.getOperand(1)), | |||
7020 | NumBits/2, DAG, *this, dl); | |||
7021 | } | |||
7022 | ||||
7023 | if (IsAllConstants) // Otherwise, it's better to do a constpool load. | |||
7024 | return SDValue(); | |||
7025 | ||||
7026 | // Otherwise, if this is a vector with i32 or f32 elements, and the element | |||
7027 | // is a non-constant being inserted into an element other than the low one, | |||
7028 | // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka | |||
7029 | // movd/movss) to move this into the low element, then shuffle it into | |||
7030 | // place. | |||
7031 | if (EVTBits == 32) { | |||
7032 | Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item); | |||
7033 | ||||
7034 | // If using the new shuffle lowering, just directly insert this. | |||
7035 | if (ExperimentalVectorShuffleLowering) | |||
7036 | return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG); | |||
7037 | ||||
7038 | // Turn it into a shuffle of zero and zero-extended scalar to vector. | |||
7039 | Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0, Subtarget, DAG); | |||
7040 | SmallVector<int, 8> MaskVec; | |||
7041 | for (unsigned i = 0; i != NumElems; ++i) | |||
7042 | MaskVec.push_back(i == Idx ? 0 : 1); | |||
7043 | return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]); | |||
7044 | } | |||
7045 | } | |||
7046 | ||||
7047 | // Splat is obviously ok. Let legalizer expand it to a shuffle. | |||
7048 | if (Values.size() == 1) { | |||
7049 | if (EVTBits == 32) { | |||
7050 | // Instead of a shuffle like this: | |||
7051 | // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0> | |||
7052 | // Check if it's possible to issue this instead. | |||
7053 | // shuffle (vload ptr)), undef, <1, 1, 1, 1> | |||
7054 | unsigned Idx = countTrailingZeros(NonZeros); | |||
7055 | SDValue Item = Op.getOperand(Idx); | |||
7056 | if (Op.getNode()->isOnlyUserOf(Item.getNode())) | |||
7057 | return LowerAsSplatVectorLoad(Item, VT, dl, DAG); | |||
7058 | } | |||
7059 | return SDValue(); | |||
7060 | } | |||
7061 | ||||
7062 | // A vector full of immediates; various special cases are already | |||
7063 | // handled, so this is best done with a single constant-pool load. | |||
7064 | if (IsAllConstants) | |||
7065 | return SDValue(); | |||
7066 | ||||
7067 | // For AVX-length vectors, see if we can use a vector load to get all of the | |||
7068 | // elements, otherwise build the individual 128-bit pieces and use | |||
7069 | // shuffles to put them in place. | |||
7070 | if (VT.is256BitVector() || VT.is512BitVector()) { | |||
7071 | SmallVector<SDValue, 64> V; | |||
7072 | for (unsigned i = 0; i != NumElems; ++i) | |||
7073 | V.push_back(Op.getOperand(i)); | |||
7074 | ||||
7075 | // Check for a build vector of consecutive loads. | |||
7076 | if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false)) | |||
7077 | return LD; | |||
7078 | ||||
7079 | EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2); | |||
7080 | ||||
7081 | // Build both the lower and upper subvector. | |||
7082 | SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, | |||
7083 | makeArrayRef(&V[0], NumElems/2)); | |||
7084 | SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, | |||
7085 | makeArrayRef(&V[NumElems / 2], NumElems/2)); | |||
7086 | ||||
7087 | // Recreate the wider vector with the lower and upper part. | |||
7088 | if (VT.is256BitVector()) | |||
7089 | return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl); | |||
7090 | return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl); | |||
7091 | } | |||
7092 | ||||
7093 | // Let legalizer expand 2-wide build_vectors. | |||
7094 | if (EVTBits == 64) { | |||
7095 | if (NumNonZero == 1) { | |||
7096 | // One half is zero or undef. | |||
7097 | unsigned Idx = countTrailingZeros(NonZeros); | |||
7098 | SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, | |||
7099 | Op.getOperand(Idx)); | |||
7100 | return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG); | |||
7101 | } | |||
7102 | return SDValue(); | |||
7103 | } | |||
7104 | ||||
7105 | // If element VT is < 32 bits, convert it to inserts into a zero vector. | |||
7106 | if (EVTBits == 8 && NumElems == 16) { | |||
7107 | SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG, | |||
7108 | Subtarget, *this); | |||
7109 | if (V.getNode()) return V; | |||
7110 | } | |||
7111 | ||||
7112 | if (EVTBits == 16 && NumElems == 8) { | |||
7113 | SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG, | |||
7114 | Subtarget, *this); | |||
7115 | if (V.getNode()) return V; | |||
7116 | } | |||
7117 | ||||
7118 | // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS | |||
7119 | if (EVTBits == 32 && NumElems == 4) { | |||
7120 | SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this); | |||
7121 | if (V.getNode()) | |||
7122 | return V; | |||
7123 | } | |||
7124 | ||||
7125 | // If element VT is == 32 bits, turn it into a number of shuffles. | |||
7126 | SmallVector<SDValue, 8> V(NumElems); | |||
7127 | if (NumElems == 4 && NumZero > 0) { | |||
7128 | for (unsigned i = 0; i < 4; ++i) { | |||
7129 | bool isZero = !(NonZeros & (1 << i)); | |||
7130 | if (isZero) | |||
7131 | V[i] = getZeroVector(VT, Subtarget, DAG, dl); | |||
7132 | else | |||
7133 | V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i)); | |||
7134 | } | |||
7135 | ||||
7136 | for (unsigned i = 0; i < 2; ++i) { | |||
7137 | switch ((NonZeros & (0x3 << i*2)) >> (i*2)) { | |||
7138 | default: break; | |||
7139 | case 0: | |||
7140 | V[i] = V[i*2]; // Must be a zero vector. | |||
7141 | break; | |||
7142 | case 1: | |||
7143 | V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]); | |||
7144 | break; | |||
7145 | case 2: | |||
7146 | V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]); | |||
7147 | break; | |||
7148 | case 3: | |||
7149 | V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]); | |||
7150 | break; | |||
7151 | } | |||
7152 | } | |||
7153 | ||||
7154 | bool Reverse1 = (NonZeros & 0x3) == 2; | |||
7155 | bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2; | |||
7156 | int MaskVec[] = { | |||
7157 | Reverse1 ? 1 : 0, | |||
7158 | Reverse1 ? 0 : 1, | |||
7159 | static_cast<int>(Reverse2 ? NumElems+1 : NumElems), | |||
7160 | static_cast<int>(Reverse2 ? NumElems : NumElems+1) | |||
7161 | }; | |||
7162 | return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]); | |||
7163 | } | |||
7164 | ||||
7165 | if (Values.size() > 1 && VT.is128BitVector()) { | |||
7166 | // Check for a build vector of consecutive loads. | |||
7167 | for (unsigned i = 0; i < NumElems; ++i) | |||
7168 | V[i] = Op.getOperand(i); | |||
7169 | ||||
7170 | // Check for elements which are consecutive loads. | |||
7171 | SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false); | |||
7172 | if (LD.getNode()) | |||
7173 | return LD; | |||
7174 | ||||
7175 | // Check for a build vector from mostly shuffle plus few inserting. | |||
7176 | SDValue Sh = buildFromShuffleMostly(Op, DAG); | |||
7177 | if (Sh.getNode()) | |||
7178 | return Sh; | |||
7179 | ||||
7180 | // For SSE 4.1, use insertps to put the high elements into the low element. | |||
7181 | if (getSubtarget()->hasSSE41()) { | |||
7182 | SDValue Result; | |||
7183 | if (Op.getOperand(0).getOpcode() != ISD::UNDEF) | |||
7184 | Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0)); | |||
7185 | else | |||
7186 | Result = DAG.getUNDEF(VT); | |||
7187 | ||||
7188 | for (unsigned i = 1; i < NumElems; ++i) { | |||
7189 | if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue; | |||
7190 | Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result, | |||
7191 | Op.getOperand(i), DAG.getIntPtrConstant(i)); | |||
7192 | } | |||
7193 | return Result; | |||
7194 | } | |||
7195 | ||||
7196 | // Otherwise, expand into a number of unpckl*, start by extending each of | |||
7197 | // our (non-undef) elements to the full vector width with the element in the | |||
7198 | // bottom slot of the vector (which generates no code for SSE). | |||
7199 | for (unsigned i = 0; i < NumElems; ++i) { | |||
7200 | if (Op.getOperand(i).getOpcode() != ISD::UNDEF) | |||
7201 | V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i)); | |||
7202 | else | |||
7203 | V[i] = DAG.getUNDEF(VT); | |||
7204 | } | |||
7205 | ||||
7206 | // Next, we iteratively mix elements, e.g. for v4f32: | |||
7207 | // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0> | |||
7208 | // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1> | |||
7209 | // Step 2: unpcklps X, Y ==> <3, 2, 1, 0> | |||
7210 | unsigned EltStride = NumElems >> 1; | |||
7211 | while (EltStride != 0) { | |||
7212 | for (unsigned i = 0; i < EltStride; ++i) { | |||
7213 | // If V[i+EltStride] is undef and this is the first round of mixing, | |||
7214 | // then it is safe to just drop this shuffle: V[i] is already in the | |||
7215 | // right place, the one element (since it's the first round) being | |||
7216 | // inserted as undef can be dropped. This isn't safe for successive | |||
7217 | // rounds because they will permute elements within both vectors. | |||
7218 | if (V[i+EltStride].getOpcode() == ISD::UNDEF && | |||
7219 | EltStride == NumElems/2) | |||
7220 | continue; | |||
7221 | ||||
7222 | V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]); | |||
7223 | } | |||
7224 | EltStride >>= 1; | |||
7225 | } | |||
7226 | return V[0]; | |||
7227 | } | |||
7228 | return SDValue(); | |||
7229 | } | |||
7230 | ||||
7231 | // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction | |||
7232 | // to create 256-bit vectors from two other 128-bit ones. | |||
7233 | static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) { | |||
7234 | SDLoc dl(Op); | |||
7235 | MVT ResVT = Op.getSimpleValueType(); | |||
7236 | ||||
7237 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7238, __PRETTY_FUNCTION__)) | |||
7238 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7238, __PRETTY_FUNCTION__)); | |||
7239 | ||||
7240 | SDValue V1 = Op.getOperand(0); | |||
7241 | SDValue V2 = Op.getOperand(1); | |||
7242 | unsigned NumElems = ResVT.getVectorNumElements(); | |||
7243 | if(ResVT.is256BitVector()) | |||
7244 | return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl); | |||
7245 | ||||
7246 | if (Op.getNumOperands() == 4) { | |||
7247 | MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(), | |||
7248 | ResVT.getVectorNumElements()/2); | |||
7249 | SDValue V3 = Op.getOperand(2); | |||
7250 | SDValue V4 = Op.getOperand(3); | |||
7251 | return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl), | |||
7252 | Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl); | |||
7253 | } | |||
7254 | return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl); | |||
7255 | } | |||
7256 | ||||
7257 | static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) { | |||
7258 | MVT LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) VT = Op.getSimpleValueType(); | |||
7259 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7261, __PRETTY_FUNCTION__)) | |||
7260 | (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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7261, __PRETTY_FUNCTION__)) | |||
7261 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7261, __PRETTY_FUNCTION__)); | |||
7262 | ||||
7263 | // AVX can use the vinsertf128 instruction to create 256-bit vectors | |||
7264 | // from two other 128-bit ones. | |||
7265 | ||||
7266 | // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors | |||
7267 | return LowerAVXCONCAT_VECTORS(Op, DAG); | |||
7268 | } | |||
7269 | ||||
7270 | ||||
7271 | //===----------------------------------------------------------------------===// | |||
7272 | // Vector shuffle lowering | |||
7273 | // | |||
7274 | // This is an experimental code path for lowering vector shuffles on x86. It is | |||
7275 | // designed to handle arbitrary vector shuffles and blends, gracefully | |||
7276 | // degrading performance as necessary. It works hard to recognize idiomatic | |||
7277 | // shuffles and lower them to optimal instruction patterns without leaving | |||
7278 | // a framework that allows reasonably efficient handling of all vector shuffle | |||
7279 | // patterns. | |||
7280 | //===----------------------------------------------------------------------===// | |||
7281 | ||||
7282 | /// \brief Tiny helper function to identify a no-op mask. | |||
7283 | /// | |||
7284 | /// This is a somewhat boring predicate function. It checks whether the mask | |||
7285 | /// array input, which is assumed to be a single-input shuffle mask of the kind | |||
7286 | /// used by the X86 shuffle instructions (not a fully general | |||
7287 | /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an | |||
7288 | /// in-place shuffle are 'no-op's. | |||
7289 | static bool isNoopShuffleMask(ArrayRef<int> Mask) { | |||
7290 | for (int i = 0, Size = Mask.size(); i < Size; ++i) | |||
7291 | if (Mask[i] != -1 && Mask[i] != i) | |||
7292 | return false; | |||
7293 | return true; | |||
7294 | } | |||
7295 | ||||
7296 | /// \brief Helper function to classify a mask as a single-input mask. | |||
7297 | /// | |||
7298 | /// This isn't a generic single-input test because in the vector shuffle | |||
7299 | /// lowering we canonicalize single inputs to be the first input operand. This | |||
7300 | /// means we can more quickly test for a single input by only checking whether | |||
7301 | /// an input from the second operand exists. We also assume that the size of | |||
7302 | /// mask corresponds to the size of the input vectors which isn't true in the | |||
7303 | /// fully general case. | |||
7304 | static bool isSingleInputShuffleMask(ArrayRef<int> Mask) { | |||
7305 | for (int M : Mask) | |||
7306 | if (M >= (int)Mask.size()) | |||
7307 | return false; | |||
7308 | return true; | |||
7309 | } | |||
7310 | ||||
7311 | /// \brief Test whether there are elements crossing 128-bit lanes in this | |||
7312 | /// shuffle mask. | |||
7313 | /// | |||
7314 | /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations | |||
7315 | /// and we routinely test for these. | |||
7316 | static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) { | |||
7317 | int LaneSize = 128 / VT.getScalarSizeInBits(); | |||
7318 | int Size = Mask.size(); | |||
7319 | for (int i = 0; i < Size; ++i) | |||
7320 | if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize) | |||
7321 | return true; | |||
7322 | return false; | |||
7323 | } | |||
7324 | ||||
7325 | /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane. | |||
7326 | /// | |||
7327 | /// This checks a shuffle mask to see if it is performing the same | |||
7328 | /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies | |||
7329 | /// that it is also not lane-crossing. It may however involve a blend from the | |||
7330 | /// same lane of a second vector. | |||
7331 | /// | |||
7332 | /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is | |||
7333 | /// non-trivial to compute in the face of undef lanes. The representation is | |||
7334 | /// *not* suitable for use with existing 128-bit shuffles as it will contain | |||
7335 | /// entries from both V1 and V2 inputs to the wider mask. | |||
7336 | static bool | |||
7337 | is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask, | |||
7338 | SmallVectorImpl<int> &RepeatedMask) { | |||
7339 | int LaneSize = 128 / VT.getScalarSizeInBits(); | |||
7340 | RepeatedMask.resize(LaneSize, -1); | |||
7341 | int Size = Mask.size(); | |||
7342 | for (int i = 0; i < Size; ++i) { | |||
7343 | if (Mask[i] < 0) | |||
7344 | continue; | |||
7345 | if ((Mask[i] % Size) / LaneSize != i / LaneSize) | |||
7346 | // This entry crosses lanes, so there is no way to model this shuffle. | |||
7347 | return false; | |||
7348 | ||||
7349 | // Ok, handle the in-lane shuffles by detecting if and when they repeat. | |||
7350 | if (RepeatedMask[i % LaneSize] == -1) | |||
7351 | // This is the first non-undef entry in this slot of a 128-bit lane. | |||
7352 | RepeatedMask[i % LaneSize] = | |||
7353 | Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size; | |||
7354 | else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i]) | |||
7355 | // Found a mismatch with the repeated mask. | |||
7356 | return false; | |||
7357 | } | |||
7358 | return true; | |||
7359 | } | |||
7360 | ||||
7361 | // Hide this symbol with an anonymous namespace instead of 'static' so that MSVC | |||
7362 | // 2013 will allow us to use it as a non-type template parameter. | |||
7363 | namespace { | |||
7364 | ||||
7365 | /// \brief Implementation of the \c isShuffleEquivalent variadic functor. | |||
7366 | /// | |||
7367 | /// See its documentation for details. | |||
7368 | bool isShuffleEquivalentImpl(ArrayRef<int> Mask, ArrayRef<const int *> Args) { | |||
7369 | if (Mask.size() != Args.size()) | |||
7370 | return false; | |||
7371 | for (int i = 0, e = Mask.size(); i < e; ++i) { | |||
7372 | assert(*Args[i] >= 0 && "Arguments must be positive integers!")((*Args[i] >= 0 && "Arguments must be positive integers!" ) ? static_cast<void> (0) : __assert_fail ("*Args[i] >= 0 && \"Arguments must be positive integers!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7372, __PRETTY_FUNCTION__)); | |||
7373 | if (Mask[i] != -1 && Mask[i] != *Args[i]) | |||
7374 | return false; | |||
7375 | } | |||
7376 | return true; | |||
7377 | } | |||
7378 | ||||
7379 | } // namespace | |||
7380 | ||||
7381 | /// \brief Checks whether a shuffle mask is equivalent to an explicit list of | |||
7382 | /// arguments. | |||
7383 | /// | |||
7384 | /// This is a fast way to test a shuffle mask against a fixed pattern: | |||
7385 | /// | |||
7386 | /// if (isShuffleEquivalent(Mask, 3, 2, 1, 0)) { ... } | |||
7387 | /// | |||
7388 | /// It returns true if the mask is exactly as wide as the argument list, and | |||
7389 | /// each element of the mask is either -1 (signifying undef) or the value given | |||
7390 | /// in the argument. | |||
7391 | static const VariadicFunction1< | |||
7392 | bool, ArrayRef<int>, int, isShuffleEquivalentImpl> isShuffleEquivalent = {}; | |||
7393 | ||||
7394 | /// \brief Get a 4-lane 8-bit shuffle immediate for a mask. | |||
7395 | /// | |||
7396 | /// This helper function produces an 8-bit shuffle immediate corresponding to | |||
7397 | /// the ubiquitous shuffle encoding scheme used in x86 instructions for | |||
7398 | /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for | |||
7399 | /// example. | |||
7400 | /// | |||
7401 | /// NB: We rely heavily on "undef" masks preserving the input lane. | |||
7402 | static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, | |||
7403 | SelectionDAG &DAG) { | |||
7404 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7404, __PRETTY_FUNCTION__)); | |||
7405 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7405, __PRETTY_FUNCTION__)); | |||
7406 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7406, __PRETTY_FUNCTION__)); | |||
7407 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7407, __PRETTY_FUNCTION__)); | |||
7408 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7408, __PRETTY_FUNCTION__)); | |||
7409 | ||||
7410 | unsigned Imm = 0; | |||
7411 | Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0; | |||
7412 | Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2; | |||
7413 | Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4; | |||
7414 | Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6; | |||
7415 | return DAG.getConstant(Imm, MVT::i8); | |||
7416 | } | |||
7417 | ||||
7418 | /// \brief Try to emit a blend instruction for a shuffle. | |||
7419 | /// | |||
7420 | /// This doesn't do any checks for the availability of instructions for blending | |||
7421 | /// these values. It relies on the availability of the X86ISD::BLENDI pattern to | |||
7422 | /// be matched in the backend with the type given. What it does check for is | |||
7423 | /// that the shuffle mask is in fact a blend. | |||
7424 | static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1, | |||
7425 | SDValue V2, ArrayRef<int> Mask, | |||
7426 | const X86Subtarget *Subtarget, | |||
7427 | SelectionDAG &DAG) { | |||
7428 | ||||
7429 | unsigned BlendMask = 0; | |||
7430 | for (int i = 0, Size = Mask.size(); i < Size; ++i) { | |||
7431 | if (Mask[i] >= Size) { | |||
7432 | if (Mask[i] != i + Size) | |||
7433 | return SDValue(); // Shuffled V2 input! | |||
7434 | BlendMask |= 1u << i; | |||
7435 | continue; | |||
7436 | } | |||
7437 | if (Mask[i] >= 0 && Mask[i] != i) | |||
7438 | return SDValue(); // Shuffled V1 input! | |||
7439 | } | |||
7440 | switch (VT.SimpleTy) { | |||
7441 | case MVT::v2f64: | |||
7442 | case MVT::v4f32: | |||
7443 | case MVT::v4f64: | |||
7444 | case MVT::v8f32: | |||
7445 | return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2, | |||
7446 | DAG.getConstant(BlendMask, MVT::i8)); | |||
7447 | ||||
7448 | case MVT::v4i64: | |||
7449 | case MVT::v8i32: | |||
7450 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7450, __PRETTY_FUNCTION__)); | |||
7451 | // FALLTHROUGH | |||
7452 | case MVT::v2i64: | |||
7453 | case MVT::v4i32: | |||
7454 | // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into | |||
7455 | // that instruction. | |||
7456 | if (Subtarget->hasAVX2()) { | |||
7457 | // Scale the blend by the number of 32-bit dwords per element. | |||
7458 | int Scale = VT.getScalarSizeInBits() / 32; | |||
7459 | BlendMask = 0; | |||
7460 | for (int i = 0, Size = Mask.size(); i < Size; ++i) | |||
7461 | if (Mask[i] >= Size) | |||
7462 | for (int j = 0; j < Scale; ++j) | |||
7463 | BlendMask |= 1u << (i * Scale + j); | |||
7464 | ||||
7465 | MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32; | |||
7466 | V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1); | |||
7467 | V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2); | |||
7468 | return DAG.getNode(ISD::BITCAST, DL, VT, | |||
7469 | DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2, | |||
7470 | DAG.getConstant(BlendMask, MVT::i8))); | |||
7471 | } | |||
7472 | // FALLTHROUGH | |||
7473 | case MVT::v8i16: { | |||
7474 | // For integer shuffles we need to expand the mask and cast the inputs to | |||
7475 | // v8i16s prior to blending. | |||
7476 | int Scale = 8 / VT.getVectorNumElements(); | |||
7477 | BlendMask = 0; | |||
7478 | for (int i = 0, Size = Mask.size(); i < Size; ++i) | |||
7479 | if (Mask[i] >= Size) | |||
7480 | for (int j = 0; j < Scale; ++j) | |||
7481 | BlendMask |= 1u << (i * Scale + j); | |||
7482 | ||||
7483 | V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1); | |||
7484 | V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2); | |||
7485 | return DAG.getNode(ISD::BITCAST, DL, VT, | |||
7486 | DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2, | |||
7487 | DAG.getConstant(BlendMask, MVT::i8))); | |||
7488 | } | |||
7489 | ||||
7490 | case MVT::v16i16: { | |||
7491 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7491, __PRETTY_FUNCTION__)); | |||
7492 | SmallVector<int, 8> RepeatedMask; | |||
7493 | if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) { | |||
7494 | // We can lower these with PBLENDW which is mirrored across 128-bit lanes. | |||
7495 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7495, __PRETTY_FUNCTION__)); | |||
7496 | BlendMask = 0; | |||
7497 | for (int i = 0; i < 8; ++i) | |||
7498 | if (RepeatedMask[i] >= 16) | |||
7499 | BlendMask |= 1u << i; | |||
7500 | return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2, | |||
7501 | DAG.getConstant(BlendMask, MVT::i8)); | |||
7502 | } | |||
7503 | } | |||
7504 | // FALLTHROUGH | |||
7505 | case MVT::v32i8: { | |||
7506 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7506, __PRETTY_FUNCTION__)); | |||
7507 | // Scale the blend by the number of bytes per element. | |||
7508 | int Scale = VT.getScalarSizeInBits() / 8; | |||
7509 | assert(Mask.size() * Scale == 32 && "Not a 256-bit vector!")((Mask.size() * Scale == 32 && "Not a 256-bit vector!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() * Scale == 32 && \"Not a 256-bit vector!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7509, __PRETTY_FUNCTION__)); | |||
7510 | ||||
7511 | // Compute the VSELECT mask. Note that VSELECT is really confusing in the | |||
7512 | // mix of LLVM's code generator and the x86 backend. We tell the code | |||
7513 | // generator that boolean values in the elements of an x86 vector register | |||
7514 | // are -1 for true and 0 for false. We then use the LLVM semantics of 'true' | |||
7515 | // mapping a select to operand #1, and 'false' mapping to operand #2. The | |||
7516 | // reality in x86 is that vector masks (pre-AVX-512) use only the high bit | |||
7517 | // of the element (the remaining are ignored) and 0 in that high bit would | |||
7518 | // mean operand #1 while 1 in the high bit would mean operand #2. So while | |||
7519 | // the LLVM model for boolean values in vector elements gets the relevant | |||
7520 | // bit set, it is set backwards and over constrained relative to x86's | |||
7521 | // actual model. | |||
7522 | SDValue VSELECTMask[32]; | |||
7523 | for (int i = 0, Size = Mask.size(); i < Size; ++i) | |||
7524 | for (int j = 0; j < Scale; ++j) | |||
7525 | VSELECTMask[Scale * i + j] = | |||
7526 | Mask[i] < 0 ? DAG.getUNDEF(MVT::i8) | |||
7527 | : DAG.getConstant(Mask[i] < Size ? -1 : 0, MVT::i8); | |||
7528 | ||||
7529 | V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1); | |||
7530 | V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V2); | |||
7531 | return DAG.getNode( | |||
7532 | ISD::BITCAST, DL, VT, | |||
7533 | DAG.getNode(ISD::VSELECT, DL, MVT::v32i8, | |||
7534 | DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, VSELECTMask), | |||
7535 | V1, V2)); | |||
7536 | } | |||
7537 | ||||
7538 | default: | |||
7539 | llvm_unreachable("Not a supported integer vector type!")::llvm::llvm_unreachable_internal("Not a supported integer vector type!" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7539); | |||
7540 | } | |||
7541 | } | |||
7542 | ||||
7543 | /// \brief Generic routine to lower a shuffle and blend as a decomposed set of | |||
7544 | /// unblended shuffles followed by an unshuffled blend. | |||
7545 | /// | |||
7546 | /// This matches the extremely common pattern for handling combined | |||
7547 | /// shuffle+blend operations on newer X86 ISAs where we have very fast blend | |||
7548 | /// operations. | |||
7549 | static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT, | |||
7550 | SDValue V1, | |||
7551 | SDValue V2, | |||
7552 | ArrayRef<int> Mask, | |||
7553 | SelectionDAG &DAG) { | |||
7554 | // Shuffle the input elements into the desired positions in V1 and V2 and | |||
7555 | // blend them together. | |||
7556 | SmallVector<int, 32> V1Mask(Mask.size(), -1); | |||
7557 | SmallVector<int, 32> V2Mask(Mask.size(), -1); | |||
7558 | SmallVector<int, 32> BlendMask(Mask.size(), -1); | |||
7559 | for (int i = 0, Size = Mask.size(); i < Size; ++i) | |||
7560 | if (Mask[i] >= 0 && Mask[i] < Size) { | |||
7561 | V1Mask[i] = Mask[i]; | |||
7562 | BlendMask[i] = i; | |||
7563 | } else if (Mask[i] >= Size) { | |||
7564 | V2Mask[i] = Mask[i] - Size; | |||
7565 | BlendMask[i] = i + Size; | |||
7566 | } | |||
7567 | ||||
7568 | V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask); | |||
7569 | V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask); | |||
7570 | return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask); | |||
7571 | } | |||
7572 | ||||
7573 | /// \brief Try to lower a vector shuffle as a byte rotation. | |||
7574 | /// | |||
7575 | /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary | |||
7576 | /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use | |||
7577 | /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will | |||
7578 | /// try to generically lower a vector shuffle through such an pattern. It | |||
7579 | /// does not check for the profitability of lowering either as PALIGNR or | |||
7580 | /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form. | |||
7581 | /// This matches shuffle vectors that look like: | |||
7582 | /// | |||
7583 | /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2] | |||
7584 | /// | |||
7585 | /// Essentially it concatenates V1 and V2, shifts right by some number of | |||
7586 | /// elements, and takes the low elements as the result. Note that while this is | |||
7587 | /// specified as a *right shift* because x86 is little-endian, it is a *left | |||
7588 | /// rotate* of the vector lanes. | |||
7589 | /// | |||
7590 | /// Note that this only handles 128-bit vector widths currently. | |||
7591 | static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1, | |||
7592 | SDValue V2, | |||
7593 | ArrayRef<int> Mask, | |||
7594 | const X86Subtarget *Subtarget, | |||
7595 | SelectionDAG &DAG) { | |||
7596 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7596, __PRETTY_FUNCTION__)); | |||
7597 | ||||
7598 | // We need to detect various ways of spelling a rotation: | |||
7599 | // [11, 12, 13, 14, 15, 0, 1, 2] | |||
7600 | // [-1, 12, 13, 14, -1, -1, 1, -1] | |||
7601 | // [-1, -1, -1, -1, -1, -1, 1, 2] | |||
7602 | // [ 3, 4, 5, 6, 7, 8, 9, 10] | |||
7603 | // [-1, 4, 5, 6, -1, -1, 9, -1] | |||
7604 | // [-1, 4, 5, 6, -1, -1, -1, -1] | |||
7605 | int Rotation = 0; | |||
7606 | SDValue Lo, Hi; | |||
7607 | for (int i = 0, Size = Mask.size(); i < Size; ++i) { | |||
7608 | if (Mask[i] == -1) | |||
7609 | continue; | |||
7610 | assert(Mask[i] >= 0 && "Only -1 is a valid negative mask element!")((Mask[i] >= 0 && "Only -1 is a valid negative mask element!" ) ? static_cast<void> (0) : __assert_fail ("Mask[i] >= 0 && \"Only -1 is a valid negative mask element!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7610, __PRETTY_FUNCTION__)); | |||
7611 | ||||
7612 | // Based on the mod-Size value of this mask element determine where | |||
7613 | // a rotated vector would have started. | |||
7614 | int StartIdx = i - (Mask[i] % Size); | |||
7615 | if (StartIdx == 0) | |||
7616 | // The identity rotation isn't interesting, stop. | |||
7617 | return SDValue(); | |||
7618 | ||||
7619 | // If we found the tail of a vector the rotation must be the missing | |||
7620 | // front. If we found the head of a vector, it must be how much of the head. | |||
7621 | int CandidateRotation = StartIdx < 0 ? -StartIdx : Size - StartIdx; | |||
7622 | ||||
7623 | if (Rotation == 0) | |||
7624 | Rotation = CandidateRotation; | |||
7625 | else if (Rotation != CandidateRotation) | |||
7626 | // The rotations don't match, so we can't match this mask. | |||
7627 | return SDValue(); | |||
7628 | ||||
7629 | // Compute which value this mask is pointing at. | |||
7630 | SDValue MaskV = Mask[i] < Size ? V1 : V2; | |||
7631 | ||||
7632 | // Compute which of the two target values this index should be assigned to. | |||
7633 | // This reflects whether the high elements are remaining or the low elements | |||
7634 | // are remaining. | |||
7635 | SDValue &TargetV = StartIdx < 0 ? Hi : Lo; | |||
7636 | ||||
7637 | // Either set up this value if we've not encountered it before, or check | |||
7638 | // that it remains consistent. | |||
7639 | if (!TargetV) | |||
7640 | TargetV = MaskV; | |||
7641 | else if (TargetV != MaskV) | |||
7642 | // This may be a rotation, but it pulls from the inputs in some | |||
7643 | // unsupported interleaving. | |||
7644 | return SDValue(); | |||
7645 | } | |||
7646 | ||||
7647 | // Check that we successfully analyzed the mask, and normalize the results. | |||
7648 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7648, __PRETTY_FUNCTION__)); | |||
7649 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7649, __PRETTY_FUNCTION__)); | |||
7650 | if (!Lo) | |||
7651 | Lo = Hi; | |||
7652 | else if (!Hi) | |||
7653 | Hi = Lo; | |||
7654 | ||||
7655 | assert(VT.getSizeInBits() == 128 &&((VT.getSizeInBits() == 128 && "Rotate-based lowering only supports 128-bit lowering!" ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() == 128 && \"Rotate-based lowering only supports 128-bit lowering!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7656, __PRETTY_FUNCTION__)) | |||
7656 | "Rotate-based lowering only supports 128-bit lowering!")((VT.getSizeInBits() == 128 && "Rotate-based lowering only supports 128-bit lowering!" ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() == 128 && \"Rotate-based lowering only supports 128-bit lowering!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7656, __PRETTY_FUNCTION__)); | |||
7657 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7658, __PRETTY_FUNCTION__)) | |||
7658 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7658, __PRETTY_FUNCTION__)); | |||
7659 | ||||
7660 | // The actual rotate instruction rotates bytes, so we need to scale the | |||
7661 | // rotation based on how many bytes are in the vector. | |||
7662 | int Scale = 16 / Mask.size(); | |||
7663 | ||||
7664 | // SSSE3 targets can use the palignr instruction | |||
7665 | if (Subtarget->hasSSSE3()) { | |||
7666 | // Cast the inputs to v16i8 to match PALIGNR. | |||
7667 | Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Lo); | |||
7668 | Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Hi); | |||
7669 | ||||
7670 | return DAG.getNode(ISD::BITCAST, DL, VT, | |||
7671 | DAG.getNode(X86ISD::PALIGNR, DL, MVT::v16i8, Hi, Lo, | |||
7672 | DAG.getConstant(Rotation * Scale, MVT::i8))); | |||
7673 | } | |||
7674 | ||||
7675 | // Default SSE2 implementation | |||
7676 | int LoByteShift = 16 - Rotation * Scale; | |||
7677 | int HiByteShift = Rotation * Scale; | |||
7678 | ||||
7679 | // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ. | |||
7680 | Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Lo); | |||
7681 | Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Hi); | |||
7682 | ||||
7683 | SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo, | |||
7684 | DAG.getConstant(8 * LoByteShift, MVT::i8)); | |||
7685 | SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi, | |||
7686 | DAG.getConstant(8 * HiByteShift, MVT::i8)); | |||
7687 | return DAG.getNode(ISD::BITCAST, DL, VT, | |||
7688 | DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift)); | |||
7689 | } | |||
7690 | ||||
7691 | /// \brief Compute whether each element of a shuffle is zeroable. | |||
7692 | /// | |||
7693 | /// A "zeroable" vector shuffle element is one which can be lowered to zero. | |||
7694 | /// Either it is an undef element in the shuffle mask, the element of the input | |||
7695 | /// referenced is undef, or the element of the input referenced is known to be | |||
7696 | /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle | |||
7697 | /// as many lanes with this technique as possible to simplify the remaining | |||
7698 | /// shuffle. | |||
7699 | static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask, | |||
7700 | SDValue V1, SDValue V2) { | |||
7701 | SmallBitVector Zeroable(Mask.size(), false); | |||
7702 | ||||
7703 | bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode()); | |||
7704 | bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode()); | |||
7705 | ||||
7706 | for (int i = 0, Size = Mask.size(); i < Size; ++i) { | |||
7707 | int M = Mask[i]; | |||
7708 | // Handle the easy cases. | |||
7709 | if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) { | |||
7710 | Zeroable[i] = true; | |||
7711 | continue; | |||
7712 | } | |||
7713 | ||||
7714 | // If this is an index into a build_vector node, dig out the input value and | |||
7715 | // use it. | |||
7716 | SDValue V = M < Size ? V1 : V2; | |||
7717 | if (V.getOpcode() != ISD::BUILD_VECTOR) | |||
7718 | continue; | |||
7719 | ||||
7720 | SDValue Input = V.getOperand(M % Size); | |||
7721 | // The UNDEF opcode check really should be dead code here, but not quite | |||
7722 | // worth asserting on (it isn't invalid, just unexpected). | |||
7723 | if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input)) | |||
7724 | Zeroable[i] = true; | |||
7725 | } | |||
7726 | ||||
7727 | return Zeroable; | |||
7728 | } | |||
7729 | ||||
7730 | /// \brief Try to lower a vector shuffle as a byte shift (shifts in zeros). | |||
7731 | /// | |||
7732 | /// Attempts to match a shuffle mask against the PSRLDQ and PSLLDQ SSE2 | |||
7733 | /// byte-shift instructions. The mask must consist of a shifted sequential | |||
7734 | /// shuffle from one of the input vectors and zeroable elements for the | |||
7735 | /// remaining 'shifted in' elements. | |||
7736 | /// | |||
7737 | /// Note that this only handles 128-bit vector widths currently. | |||
7738 | static SDValue lowerVectorShuffleAsByteShift(SDLoc DL, MVT VT, SDValue V1, | |||
7739 | SDValue V2, ArrayRef<int> Mask, | |||
7740 | SelectionDAG &DAG) { | |||
7741 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7741, __PRETTY_FUNCTION__)); | |||
7742 | ||||
7743 | SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2); | |||
7744 | ||||
7745 | int Size = Mask.size(); | |||
7746 | int Scale = 16 / Size; | |||
7747 | ||||
7748 | for (int Shift = 1; Shift < Size; Shift++) { | |||
7749 | int ByteShift = Shift * Scale; | |||
7750 | ||||
7751 | // PSRLDQ : (little-endian) right byte shift | |||
7752 | // [ 5, 6, 7, zz, zz, zz, zz, zz] | |||
7753 | // [ -1, 5, 6, 7, zz, zz, zz, zz] | |||
7754 | // [ 1, 2, -1, -1, -1, -1, zz, zz] | |||
7755 | bool ZeroableRight = true; | |||
7756 | for (int i = Size - Shift; i < Size; i++) { | |||
7757 | ZeroableRight &= Zeroable[i]; | |||
7758 | } | |||
7759 | ||||
7760 | if (ZeroableRight) { | |||
7761 | bool ValidShiftRight1 = | |||
7762 | isSequentialOrUndefInRange(Mask, 0, Size - Shift, Shift); | |||
7763 | bool ValidShiftRight2 = | |||
7764 | isSequentialOrUndefInRange(Mask, 0, Size - Shift, Size + Shift); | |||
7765 | ||||
7766 | if (ValidShiftRight1 || ValidShiftRight2) { | |||
7767 | // Cast the inputs to v2i64 to match PSRLDQ. | |||
7768 | SDValue &TargetV = ValidShiftRight1 ? V1 : V2; | |||
7769 | SDValue V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, TargetV); | |||
7770 | SDValue Shifted = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, V, | |||
7771 | DAG.getConstant(ByteShift * 8, MVT::i8)); | |||
7772 | return DAG.getNode(ISD::BITCAST, DL, VT, Shifted); | |||
7773 | } | |||
7774 | } | |||
7775 | ||||
7776 | // PSLLDQ : (little-endian) left byte shift | |||
7777 | // [ zz, 0, 1, 2, 3, 4, 5, 6] | |||
7778 | // [ zz, zz, -1, -1, 2, 3, 4, -1] | |||
7779 | // [ zz, zz, zz, zz, zz, zz, -1, 1] | |||
7780 | bool ZeroableLeft = true; | |||
7781 | for (int i = 0; i < Shift; i++) { | |||
7782 | ZeroableLeft &= Zeroable[i]; | |||
7783 | } | |||
7784 | ||||
7785 | if (ZeroableLeft) { | |||
7786 | bool ValidShiftLeft1 = | |||
7787 | isSequentialOrUndefInRange(Mask, Shift, Size - Shift, 0); | |||
7788 | bool ValidShiftLeft2 = | |||
7789 | isSequentialOrUndefInRange(Mask, Shift, Size - Shift, Size); | |||
7790 | ||||
7791 | if (ValidShiftLeft1 || ValidShiftLeft2) { | |||
7792 | // Cast the inputs to v2i64 to match PSLLDQ. | |||
7793 | SDValue &TargetV = ValidShiftLeft1 ? V1 : V2; | |||
7794 | SDValue V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, TargetV); | |||
7795 | SDValue Shifted = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, V, | |||
7796 | DAG.getConstant(ByteShift * 8, MVT::i8)); | |||
7797 | return DAG.getNode(ISD::BITCAST, DL, VT, Shifted); | |||
7798 | } | |||
7799 | } | |||
7800 | } | |||
7801 | ||||
7802 | return SDValue(); | |||
7803 | } | |||
7804 | ||||
7805 | /// \brief Lower a vector shuffle as a zero or any extension. | |||
7806 | /// | |||
7807 | /// Given a specific number of elements, element bit width, and extension | |||
7808 | /// stride, produce either a zero or any extension based on the available | |||
7809 | /// features of the subtarget. | |||
7810 | static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend( | |||
7811 | SDLoc DL, MVT VT, int NumElements, int Scale, bool AnyExt, SDValue InputV, | |||
7812 | const X86Subtarget *Subtarget, SelectionDAG &DAG) { | |||
7813 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7813, __PRETTY_FUNCTION__)); | |||
7814 | int EltBits = VT.getSizeInBits() / NumElements; | |||
7815 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7816, __PRETTY_FUNCTION__)) | |||
7816 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7816, __PRETTY_FUNCTION__)); | |||
7817 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7817, __PRETTY_FUNCTION__)); | |||
7818 | ||||
7819 | // Found a valid zext mask! Try various lowering strategies based on the | |||
7820 | // input type and available ISA extensions. | |||
7821 | if (Subtarget->hasSSE41()) { | |||
7822 | MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements); | |||
7823 | MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale), | |||
7824 | NumElements / Scale); | |||
7825 | InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV); | |||
7826 | return DAG.getNode(ISD::BITCAST, DL, VT, | |||
7827 | DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV)); | |||
7828 | } | |||
7829 | ||||
7830 | // For any extends we can cheat for larger element sizes and use shuffle | |||
7831 | // instructions that can fold with a load and/or copy. | |||
7832 | if (AnyExt && EltBits == 32) { | |||
7833 | int PSHUFDMask[4] = {0, -1, 1, -1}; | |||
7834 | return DAG.getNode( | |||
7835 | ISD::BITCAST, DL, VT, | |||
7836 | DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, | |||
7837 | DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV), | |||
7838 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG))); | |||
7839 | } | |||
7840 | if (AnyExt && EltBits == 16 && Scale > 2) { | |||
7841 | int PSHUFDMask[4] = {0, -1, 0, -1}; | |||
7842 | InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, | |||
7843 | DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV), | |||
7844 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)); | |||
7845 | int PSHUFHWMask[4] = {1, -1, -1, -1}; | |||
7846 | return DAG.getNode( | |||
7847 | ISD::BITCAST, DL, VT, | |||
7848 | DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, | |||
7849 | DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, InputV), | |||
7850 | getV4X86ShuffleImm8ForMask(PSHUFHWMask, DAG))); | |||
7851 | } | |||
7852 | ||||
7853 | // If this would require more than 2 unpack instructions to expand, use | |||
7854 | // pshufb when available. We can only use more than 2 unpack instructions | |||
7855 | // when zero extending i8 elements which also makes it easier to use pshufb. | |||
7856 | if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) { | |||
7857 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7857, __PRETTY_FUNCTION__)); | |||
7858 | SDValue PSHUFBMask[16]; | |||
7859 | for (int i = 0; i < 16; ++i) | |||
7860 | PSHUFBMask[i] = | |||
7861 | DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, MVT::i8); | |||
7862 | InputV = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, InputV); | |||
7863 | return DAG.getNode(ISD::BITCAST, DL, VT, | |||
7864 | DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV, | |||
7865 | DAG.getNode(ISD::BUILD_VECTOR, DL, | |||
7866 | MVT::v16i8, PSHUFBMask))); | |||
7867 | } | |||
7868 | ||||
7869 | // Otherwise emit a sequence of unpacks. | |||
7870 | do { | |||
7871 | MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements); | |||
7872 | SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT) | |||
7873 | : getZeroVector(InputVT, Subtarget, DAG, DL); | |||
7874 | InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV); | |||
7875 | InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext); | |||
7876 | Scale /= 2; | |||
7877 | EltBits *= 2; | |||
7878 | NumElements /= 2; | |||
7879 | } while (Scale > 1); | |||
7880 | return DAG.getNode(ISD::BITCAST, DL, VT, InputV); | |||
7881 | } | |||
7882 | ||||
7883 | /// \brief Try to lower a vector shuffle as a zero extension on any micrarch. | |||
7884 | /// | |||
7885 | /// This routine will try to do everything in its power to cleverly lower | |||
7886 | /// a shuffle which happens to match the pattern of a zero extend. It doesn't | |||
7887 | /// check for the profitability of this lowering, it tries to aggressively | |||
7888 | /// match this pattern. It will use all of the micro-architectural details it | |||
7889 | /// can to emit an efficient lowering. It handles both blends with all-zero | |||
7890 | /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to | |||
7891 | /// masking out later). | |||
7892 | /// | |||
7893 | /// The reason we have dedicated lowering for zext-style shuffles is that they | |||
7894 | /// are both incredibly common and often quite performance sensitive. | |||
7895 | static SDValue lowerVectorShuffleAsZeroOrAnyExtend( | |||
7896 | SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, | |||
7897 | const X86Subtarget *Subtarget, SelectionDAG &DAG) { | |||
7898 | SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2); | |||
7899 | ||||
7900 | int Bits = VT.getSizeInBits(); | |||
7901 | int NumElements = Mask.size(); | |||
7902 | ||||
7903 | // Define a helper function to check a particular ext-scale and lower to it if | |||
7904 | // valid. | |||
7905 | auto Lower = [&](int Scale) -> SDValue { | |||
7906 | SDValue InputV; | |||
7907 | bool AnyExt = true; | |||
7908 | for (int i = 0; i < NumElements; ++i) { | |||
7909 | if (Mask[i] == -1) | |||
7910 | continue; // Valid anywhere but doesn't tell us anything. | |||
7911 | if (i % Scale != 0) { | |||
7912 | // Each of the extended elements need to be zeroable. | |||
7913 | if (!Zeroable[i]) | |||
7914 | return SDValue(); | |||
7915 | ||||
7916 | // We no longer are in the anyext case. | |||
7917 | AnyExt = false; | |||
7918 | continue; | |||
7919 | } | |||
7920 | ||||
7921 | // Each of the base elements needs to be consecutive indices into the | |||
7922 | // same input vector. | |||
7923 | SDValue V = Mask[i] < NumElements ? V1 : V2; | |||
7924 | if (!InputV) | |||
7925 | InputV = V; | |||
7926 | else if (InputV != V) | |||
7927 | return SDValue(); // Flip-flopping inputs. | |||
7928 | ||||
7929 | if (Mask[i] % NumElements != i / Scale) | |||
7930 | return SDValue(); // Non-consecutive strided elements. | |||
7931 | } | |||
7932 | ||||
7933 | // If we fail to find an input, we have a zero-shuffle which should always | |||
7934 | // have already been handled. | |||
7935 | // FIXME: Maybe handle this here in case during blending we end up with one? | |||
7936 | if (!InputV) | |||
7937 | return SDValue(); | |||
7938 | ||||
7939 | return lowerVectorShuffleAsSpecificZeroOrAnyExtend( | |||
7940 | DL, VT, NumElements, Scale, AnyExt, InputV, Subtarget, DAG); | |||
7941 | }; | |||
7942 | ||||
7943 | // The widest scale possible for extending is to a 64-bit integer. | |||
7944 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7945, __PRETTY_FUNCTION__)) | |||
7945 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7945, __PRETTY_FUNCTION__)); | |||
7946 | int NumExtElements = Bits / 64; | |||
7947 | ||||
7948 | // Each iteration, try extending the elements half as much, but into twice as | |||
7949 | // many elements. | |||
7950 | for (; NumExtElements < NumElements; NumExtElements *= 2) { | |||
7951 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7952, __PRETTY_FUNCTION__)) | |||
7952 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 7952, __PRETTY_FUNCTION__)); | |||
7953 | if (SDValue V = Lower(NumElements / NumExtElements)) | |||
7954 | return V; | |||
7955 | } | |||
7956 | ||||
7957 | // No viable ext lowering found. | |||
7958 | return SDValue(); | |||
7959 | } | |||
7960 | ||||
7961 | /// \brief Try to get a scalar value for a specific element of a vector. | |||
7962 | /// | |||
7963 | /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar. | |||
7964 | static SDValue getScalarValueForVectorElement(SDValue V, int Idx, | |||
7965 | SelectionDAG &DAG) { | |||
7966 | MVT VT = V.getSimpleValueType(); | |||
7967 | MVT EltVT = VT.getVectorElementType(); | |||
7968 | while (V.getOpcode() == ISD::BITCAST) | |||
7969 | V = V.getOperand(0); | |||
7970 | // If the bitcasts shift the element size, we can't extract an equivalent | |||
7971 | // element from it. | |||
7972 | MVT NewVT = V.getSimpleValueType(); | |||
7973 | if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits()) | |||
7974 | return SDValue(); | |||
7975 | ||||
7976 | if (V.getOpcode() == ISD::BUILD_VECTOR || | |||
7977 | (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) | |||
7978 | return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, V.getOperand(Idx)); | |||
7979 | ||||
7980 | return SDValue(); | |||
7981 | } | |||
7982 | ||||
7983 | /// \brief Helper to test for a load that can be folded with x86 shuffles. | |||
7984 | /// | |||
7985 | /// This is particularly important because the set of instructions varies | |||
7986 | /// significantly based on whether the operand is a load or not. | |||
7987 | static bool isShuffleFoldableLoad(SDValue V) { | |||
7988 | while (V.getOpcode() == ISD::BITCAST) | |||
7989 | V = V.getOperand(0); | |||
7990 | ||||
7991 | return ISD::isNON_EXTLoad(V.getNode()); | |||
7992 | } | |||
7993 | ||||
7994 | /// \brief Try to lower insertion of a single element into a zero vector. | |||
7995 | /// | |||
7996 | /// This is a common pattern that we have especially efficient patterns to lower | |||
7997 | /// across all subtarget feature sets. | |||
7998 | static SDValue lowerVectorShuffleAsElementInsertion( | |||
7999 | MVT VT, SDLoc DL, SDValue V1, SDValue V2, ArrayRef<int> Mask, | |||
8000 | const X86Subtarget *Subtarget, SelectionDAG &DAG) { | |||
8001 | SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2); | |||
8002 | MVT ExtVT = VT; | |||
8003 | MVT EltVT = VT.getVectorElementType(); | |||
8004 | ||||
8005 | int V2Index = std::find_if(Mask.begin(), Mask.end(), | |||
8006 | [&Mask](int M) { return M >= (int)Mask.size(); }) - | |||
8007 | Mask.begin(); | |||
8008 | bool IsV1Zeroable = true; | |||
8009 | for (int i = 0, Size = Mask.size(); i < Size; ++i) | |||
8010 | if (i != V2Index && !Zeroable[i]) { | |||
8011 | IsV1Zeroable = false; | |||
8012 | break; | |||
8013 | } | |||
8014 | ||||
8015 | // Check for a single input from a SCALAR_TO_VECTOR node. | |||
8016 | // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and | |||
8017 | // all the smarts here sunk into that routine. However, the current | |||
8018 | // lowering of BUILD_VECTOR makes that nearly impossible until the old | |||
8019 | // vector shuffle lowering is dead. | |||
8020 | if (SDValue V2S = getScalarValueForVectorElement( | |||
8021 | V2, Mask[V2Index] - Mask.size(), DAG)) { | |||
8022 | // We need to zext the scalar if it is smaller than an i32. | |||
8023 | V2S = DAG.getNode(ISD::BITCAST, DL, EltVT, V2S); | |||
8024 | if (EltVT == MVT::i8 || EltVT == MVT::i16) { | |||
8025 | // Using zext to expand a narrow element won't work for non-zero | |||
8026 | // insertions. | |||
8027 | if (!IsV1Zeroable) | |||
8028 | return SDValue(); | |||
8029 | ||||
8030 | // Zero-extend directly to i32. | |||
8031 | ExtVT = MVT::v4i32; | |||
8032 | V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S); | |||
8033 | } | |||
8034 | V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S); | |||
8035 | } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 || | |||
8036 | EltVT == MVT::i16) { | |||
8037 | // Either not inserting from the low element of the input or the input | |||
8038 | // element size is too small to use VZEXT_MOVL to clear the high bits. | |||
8039 | return SDValue(); | |||
8040 | } | |||
8041 | ||||
8042 | if (!IsV1Zeroable) { | |||
8043 | // If V1 can't be treated as a zero vector we have fewer options to lower | |||
8044 | // this. We can't support integer vectors or non-zero targets cheaply, and | |||
8045 | // the V1 elements can't be permuted in any way. | |||
8046 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8046, __PRETTY_FUNCTION__)); | |||
8047 | if (!VT.isFloatingPoint() || V2Index != 0) | |||
8048 | return SDValue(); | |||
8049 | SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end()); | |||
8050 | V1Mask[V2Index] = -1; | |||
8051 | if (!isNoopShuffleMask(V1Mask)) | |||
8052 | return SDValue(); | |||
8053 | // This is essentially a special case blend operation, but if we have | |||
8054 | // general purpose blend operations, they are always faster. Bail and let | |||
8055 | // the rest of the lowering handle these as blends. | |||
8056 | if (Subtarget->hasSSE41()) | |||
8057 | return SDValue(); | |||
8058 | ||||
8059 | // Otherwise, use MOVSD or MOVSS. | |||
8060 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8061, __PRETTY_FUNCTION__)) | |||
8061 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8061, __PRETTY_FUNCTION__)); | |||
8062 | return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL, | |||
8063 | ExtVT, V1, V2); | |||
8064 | } | |||
8065 | ||||
8066 | V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2); | |||
8067 | if (ExtVT != VT) | |||
8068 | V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2); | |||
8069 | ||||
8070 | if (V2Index != 0) { | |||
8071 | // If we have 4 or fewer lanes we can cheaply shuffle the element into | |||
8072 | // the desired position. Otherwise it is more efficient to do a vector | |||
8073 | // shift left. We know that we can do a vector shift left because all | |||
8074 | // the inputs are zero. | |||
8075 | if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) { | |||
8076 | SmallVector<int, 4> V2Shuffle(Mask.size(), 1); | |||
8077 | V2Shuffle[V2Index] = 0; | |||
8078 | V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle); | |||
8079 | } else { | |||
8080 | V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V2); | |||
8081 | V2 = DAG.getNode( | |||
8082 | X86ISD::VSHLDQ, DL, MVT::v2i64, V2, | |||
8083 | DAG.getConstant( | |||
8084 | V2Index * EltVT.getSizeInBits(), | |||
8085 | DAG.getTargetLoweringInfo().getScalarShiftAmountTy(MVT::v2i64))); | |||
8086 | V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2); | |||
8087 | } | |||
8088 | } | |||
8089 | return V2; | |||
8090 | } | |||
8091 | ||||
8092 | /// \brief Try to lower broadcast of a single element. | |||
8093 | /// | |||
8094 | /// For convenience, this code also bundles all of the subtarget feature set | |||
8095 | /// filtering. While a little annoying to re-dispatch on type here, there isn't | |||
8096 | /// a convenient way to factor it out. | |||
8097 | static SDValue lowerVectorShuffleAsBroadcast(MVT VT, SDLoc DL, SDValue V, | |||
8098 | ArrayRef<int> Mask, | |||
8099 | const X86Subtarget *Subtarget, | |||
8100 | SelectionDAG &DAG) { | |||
8101 | if (!Subtarget->hasAVX()) | |||
8102 | return SDValue(); | |||
8103 | if (VT.isInteger() && !Subtarget->hasAVX2()) | |||
8104 | return SDValue(); | |||
8105 | ||||
8106 | // Check that the mask is a broadcast. | |||
8107 | int BroadcastIdx = -1; | |||
8108 | for (int M : Mask) | |||
8109 | if (M >= 0 && BroadcastIdx == -1) | |||
8110 | BroadcastIdx = M; | |||
8111 | else if (M >= 0 && M != BroadcastIdx) | |||
8112 | return SDValue(); | |||
8113 | ||||
8114 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8116, __PRETTY_FUNCTION__)) | |||
8115 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8116, __PRETTY_FUNCTION__)) | |||
8116 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8116, __PRETTY_FUNCTION__)); | |||
8117 | ||||
8118 | // Go up the chain of (vector) values to try and find a scalar load that | |||
8119 | // we can combine with the broadcast. | |||
8120 | for (;;) { | |||
8121 | switch (V.getOpcode()) { | |||
8122 | case ISD::CONCAT_VECTORS: { | |||
8123 | int OperandSize = Mask.size() / V.getNumOperands(); | |||
8124 | V = V.getOperand(BroadcastIdx / OperandSize); | |||
8125 | BroadcastIdx %= OperandSize; | |||
8126 | continue; | |||
8127 | } | |||
8128 | ||||
8129 | case ISD::INSERT_SUBVECTOR: { | |||
8130 | SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1); | |||
8131 | auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2)); | |||
8132 | if (!ConstantIdx) | |||
8133 | break; | |||
8134 | ||||
8135 | int BeginIdx = (int)ConstantIdx->getZExtValue(); | |||
8136 | int EndIdx = | |||
8137 | BeginIdx + (int)VInner.getValueType().getVectorNumElements(); | |||
8138 | if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) { | |||
8139 | BroadcastIdx -= BeginIdx; | |||
8140 | V = VInner; | |||
8141 | } else { | |||
8142 | V = VOuter; | |||
8143 | } | |||
8144 | continue; | |||
8145 | } | |||
8146 | } | |||
8147 | break; | |||
8148 | } | |||
8149 | ||||
8150 | // Check if this is a broadcast of a scalar. We special case lowering | |||
8151 | // for scalars so that we can more effectively fold with loads. | |||
8152 | if (V.getOpcode() == ISD::BUILD_VECTOR || | |||
8153 | (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) { | |||
8154 | V = V.getOperand(BroadcastIdx); | |||
8155 | ||||
8156 | // If the scalar isn't a load we can't broadcast from it in AVX1, only with | |||
8157 | // AVX2. | |||
8158 | if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V)) | |||
8159 | return SDValue(); | |||
8160 | } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) { | |||
8161 | // We can't broadcast from a vector register w/o AVX2, and we can only | |||
8162 | // broadcast from the zero-element of a vector register. | |||
8163 | return SDValue(); | |||
8164 | } | |||
8165 | ||||
8166 | return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V); | |||
8167 | } | |||
8168 | ||||
8169 | // Check for whether we can use INSERTPS to perform the shuffle. We only use | |||
8170 | // INSERTPS when the V1 elements are already in the correct locations | |||
8171 | // because otherwise we can just always use two SHUFPS instructions which | |||
8172 | // are much smaller to encode than a SHUFPS and an INSERTPS. We can also | |||
8173 | // perform INSERTPS if a single V1 element is out of place and all V2 | |||
8174 | // elements are zeroable. | |||
8175 | static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2, | |||
8176 | ArrayRef<int> Mask, | |||
8177 | SelectionDAG &DAG) { | |||
8178 | assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!")((Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType() == MVT::v4f32 && \"Bad shuffle type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8178, __PRETTY_FUNCTION__)); | |||
8179 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8179, __PRETTY_FUNCTION__)); | |||
8180 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8180, __PRETTY_FUNCTION__)); | |||
8181 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8181, __PRETTY_FUNCTION__)); | |||
8182 | ||||
8183 | SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2); | |||
8184 | ||||
8185 | unsigned ZMask = 0; | |||
8186 | int V1DstIndex = -1; | |||
8187 | int V2DstIndex = -1; | |||
8188 | bool V1UsedInPlace = false; | |||
8189 | ||||
8190 | for (int i = 0; i < 4; i++) { | |||
8191 | // Synthesize a zero mask from the zeroable elements (includes undefs). | |||
8192 | if (Zeroable[i]) { | |||
8193 | ZMask |= 1 << i; | |||
8194 | continue; | |||
8195 | } | |||
8196 | ||||
8197 | // Flag if we use any V1 inputs in place. | |||
8198 | if (i == Mask[i]) { | |||
8199 | V1UsedInPlace = true; | |||
8200 | continue; | |||
8201 | } | |||
8202 | ||||
8203 | // We can only insert a single non-zeroable element. | |||
8204 | if (V1DstIndex != -1 || V2DstIndex != -1) | |||
8205 | return SDValue(); | |||
8206 | ||||
8207 | if (Mask[i] < 4) { | |||
8208 | // V1 input out of place for insertion. | |||
8209 | V1DstIndex = i; | |||
8210 | } else { | |||
8211 | // V2 input for insertion. | |||
8212 | V2DstIndex = i; | |||
8213 | } | |||
8214 | } | |||
8215 | ||||
8216 | // Don't bother if we have no (non-zeroable) element for insertion. | |||
8217 | if (V1DstIndex == -1 && V2DstIndex == -1) | |||
8218 | return SDValue(); | |||
8219 | ||||
8220 | // Determine element insertion src/dst indices. The src index is from the | |||
8221 | // start of the inserted vector, not the start of the concatenated vector. | |||
8222 | unsigned V2SrcIndex = 0; | |||
8223 | if (V1DstIndex != -1) { | |||
8224 | // If we have a V1 input out of place, we use V1 as the V2 element insertion | |||
8225 | // and don't use the original V2 at all. | |||
8226 | V2SrcIndex = Mask[V1DstIndex]; | |||
8227 | V2DstIndex = V1DstIndex; | |||
8228 | V2 = V1; | |||
8229 | } else { | |||
8230 | V2SrcIndex = Mask[V2DstIndex] - 4; | |||
8231 | } | |||
8232 | ||||
8233 | // If no V1 inputs are used in place, then the result is created only from | |||
8234 | // the zero mask and the V2 insertion - so remove V1 dependency. | |||
8235 | if (!V1UsedInPlace) | |||
8236 | V1 = DAG.getUNDEF(MVT::v4f32); | |||
8237 | ||||
8238 | unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask; | |||
8239 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8239, __PRETTY_FUNCTION__)); | |||
8240 | ||||
8241 | // Insert the V2 element into the desired position. | |||
8242 | SDLoc DL(Op); | |||
8243 | return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2, | |||
8244 | DAG.getConstant(InsertPSMask, MVT::i8)); | |||
8245 | } | |||
8246 | ||||
8247 | /// \brief Handle lowering of 2-lane 64-bit floating point shuffles. | |||
8248 | /// | |||
8249 | /// This is the basis function for the 2-lane 64-bit shuffles as we have full | |||
8250 | /// support for floating point shuffles but not integer shuffles. These | |||
8251 | /// instructions will incur a domain crossing penalty on some chips though so | |||
8252 | /// it is better to avoid lowering through this for integer vectors where | |||
8253 | /// possible. | |||
8254 | static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
8255 | const X86Subtarget *Subtarget, | |||
8256 | SelectionDAG &DAG) { | |||
8257 | SDLoc DL(Op); | |||
8258 | assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!")((Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType() == MVT::v2f64 && \"Bad shuffle type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8258, __PRETTY_FUNCTION__)); | |||
8259 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8259, __PRETTY_FUNCTION__)); | |||
8260 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8260, __PRETTY_FUNCTION__)); | |||
8261 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
8262 | ArrayRef<int> Mask = SVOp->getMask(); | |||
8263 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8263, __PRETTY_FUNCTION__)); | |||
8264 | ||||
8265 | if (isSingleInputShuffleMask(Mask)) { | |||
8266 | // Use low duplicate instructions for masks that match their pattern. | |||
8267 | if (Subtarget->hasSSE3()) | |||
8268 | if (isShuffleEquivalent(Mask, 0, 0)) | |||
8269 | return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1); | |||
8270 | ||||
8271 | // Straight shuffle of a single input vector. Simulate this by using the | |||
8272 | // single input as both of the "inputs" to this instruction.. | |||
8273 | unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1); | |||
8274 | ||||
8275 | if (Subtarget->hasAVX()) { | |||
8276 | // If we have AVX, we can use VPERMILPS which will allow folding a load | |||
8277 | // into the shuffle. | |||
8278 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1, | |||
8279 | DAG.getConstant(SHUFPDMask, MVT::i8)); | |||
8280 | } | |||
8281 | ||||
8282 | return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V1, | |||
8283 | DAG.getConstant(SHUFPDMask, MVT::i8)); | |||
8284 | } | |||
8285 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8285, __PRETTY_FUNCTION__)); | |||
8286 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8286, __PRETTY_FUNCTION__)); | |||
8287 | ||||
8288 | // Use dedicated unpack instructions for masks that match their pattern. | |||
8289 | if (isShuffleEquivalent(Mask, 0, 2)) | |||
8290 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2); | |||
8291 | if (isShuffleEquivalent(Mask, 1, 3)) | |||
8292 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2); | |||
8293 | ||||
8294 | // If we have a single input, insert that into V1 if we can do so cheaply. | |||
8295 | if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) { | |||
8296 | if (SDValue Insertion = lowerVectorShuffleAsElementInsertion( | |||
8297 | MVT::v2f64, DL, V1, V2, Mask, Subtarget, DAG)) | |||
8298 | return Insertion; | |||
8299 | // Try inverting the insertion since for v2 masks it is easy to do and we | |||
8300 | // can't reliably sort the mask one way or the other. | |||
8301 | int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2), | |||
8302 | Mask[1] < 0 ? -1 : (Mask[1] ^ 2)}; | |||
8303 | if (SDValue Insertion = lowerVectorShuffleAsElementInsertion( | |||
8304 | MVT::v2f64, DL, V2, V1, InverseMask, Subtarget, DAG)) | |||
8305 | return Insertion; | |||
8306 | } | |||
8307 | ||||
8308 | // Try to use one of the special instruction patterns to handle two common | |||
8309 | // blend patterns if a zero-blend above didn't work. | |||
8310 | if (isShuffleEquivalent(Mask, 0, 3) || isShuffleEquivalent(Mask, 1, 3)) | |||
8311 | if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG)) | |||
8312 | // We can either use a special instruction to load over the low double or | |||
8313 | // to move just the low double. | |||
8314 | return DAG.getNode( | |||
8315 | isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD, | |||
8316 | DL, MVT::v2f64, V2, | |||
8317 | DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S)); | |||
8318 | ||||
8319 | if (Subtarget->hasSSE41()) | |||
8320 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask, | |||
8321 | Subtarget, DAG)) | |||
8322 | return Blend; | |||
8323 | ||||
8324 | unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1); | |||
8325 | return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V2, | |||
8326 | DAG.getConstant(SHUFPDMask, MVT::i8)); | |||
8327 | } | |||
8328 | ||||
8329 | /// \brief Handle lowering of 2-lane 64-bit integer shuffles. | |||
8330 | /// | |||
8331 | /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by | |||
8332 | /// the integer unit to minimize domain crossing penalties. However, for blends | |||
8333 | /// it falls back to the floating point shuffle operation with appropriate bit | |||
8334 | /// casting. | |||
8335 | static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
8336 | const X86Subtarget *Subtarget, | |||
8337 | SelectionDAG &DAG) { | |||
8338 | SDLoc DL(Op); | |||
8339 | assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!")((Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType() == MVT::v2i64 && \"Bad shuffle type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8339, __PRETTY_FUNCTION__)); | |||
8340 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8340, __PRETTY_FUNCTION__)); | |||
8341 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8341, __PRETTY_FUNCTION__)); | |||
8342 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
8343 | ArrayRef<int> Mask = SVOp->getMask(); | |||
8344 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8344, __PRETTY_FUNCTION__)); | |||
8345 | ||||
8346 | if (isSingleInputShuffleMask(Mask)) { | |||
8347 | // Check for being able to broadcast a single element. | |||
8348 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v2i64, DL, V1, | |||
8349 | Mask, Subtarget, DAG)) | |||
8350 | return Broadcast; | |||
8351 | ||||
8352 | // Straight shuffle of a single input vector. For everything from SSE2 | |||
8353 | // onward this has a single fast instruction with no scary immediates. | |||
8354 | // We have to map the mask as it is actually a v4i32 shuffle instruction. | |||
8355 | V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1); | |||
8356 | int WidenedMask[4] = { | |||
8357 | std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1, | |||
8358 | std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1}; | |||
8359 | return DAG.getNode( | |||
8360 | ISD::BITCAST, DL, MVT::v2i64, | |||
8361 | DAG.getNode(X86ISD::PSHUFD, SDLoc(Op), MVT::v4i32, V1, | |||
8362 | getV4X86ShuffleImm8ForMask(WidenedMask, DAG))); | |||
8363 | } | |||
8364 | ||||
8365 | // Try to use byte shift instructions. | |||
8366 | if (SDValue Shift = lowerVectorShuffleAsByteShift( | |||
8367 | DL, MVT::v2i64, V1, V2, Mask, DAG)) | |||
8368 | return Shift; | |||
8369 | ||||
8370 | // If we have a single input from V2 insert that into V1 if we can do so | |||
8371 | // cheaply. | |||
8372 | if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) { | |||
8373 | if (SDValue Insertion = lowerVectorShuffleAsElementInsertion( | |||
8374 | MVT::v2i64, DL, V1, V2, Mask, Subtarget, DAG)) | |||
8375 | return Insertion; | |||
8376 | // Try inverting the insertion since for v2 masks it is easy to do and we | |||
8377 | // can't reliably sort the mask one way or the other. | |||
8378 | int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2), | |||
8379 | Mask[1] < 0 ? -1 : (Mask[1] ^ 2)}; | |||
8380 | if (SDValue Insertion = lowerVectorShuffleAsElementInsertion( | |||
8381 | MVT::v2i64, DL, V2, V1, InverseMask, Subtarget, DAG)) | |||
8382 | return Insertion; | |||
8383 | } | |||
8384 | ||||
8385 | // Use dedicated unpack instructions for masks that match their pattern. | |||
8386 | if (isShuffleEquivalent(Mask, 0, 2)) | |||
8387 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2); | |||
8388 | if (isShuffleEquivalent(Mask, 1, 3)) | |||
8389 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2); | |||
8390 | ||||
8391 | if (Subtarget->hasSSE41()) | |||
8392 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask, | |||
8393 | Subtarget, DAG)) | |||
8394 | return Blend; | |||
8395 | ||||
8396 | // Try to use byte rotation instructions. | |||
8397 | // Its more profitable for pre-SSSE3 to use shuffles/unpacks. | |||
8398 | if (Subtarget->hasSSSE3()) | |||
8399 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate( | |||
8400 | DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG)) | |||
8401 | return Rotate; | |||
8402 | ||||
8403 | // We implement this with SHUFPD which is pretty lame because it will likely | |||
8404 | // incur 2 cycles of stall for integer vectors on Nehalem and older chips. | |||
8405 | // However, all the alternatives are still more cycles and newer chips don't | |||
8406 | // have this problem. It would be really nice if x86 had better shuffles here. | |||
8407 | V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1); | |||
8408 | V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2); | |||
8409 | return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, | |||
8410 | DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask)); | |||
8411 | } | |||
8412 | ||||
8413 | /// \brief Lower a vector shuffle using the SHUFPS instruction. | |||
8414 | /// | |||
8415 | /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS. | |||
8416 | /// It makes no assumptions about whether this is the *best* lowering, it simply | |||
8417 | /// uses it. | |||
8418 | static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT, | |||
8419 | ArrayRef<int> Mask, SDValue V1, | |||
8420 | SDValue V2, SelectionDAG &DAG) { | |||
8421 | SDValue LowV = V1, HighV = V2; | |||
8422 | int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]}; | |||
8423 | ||||
8424 | int NumV2Elements = | |||
8425 | std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }); | |||
8426 | ||||
8427 | if (NumV2Elements == 1) { | |||
8428 | int V2Index = | |||
8429 | std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) - | |||
8430 | Mask.begin(); | |||
8431 | ||||
8432 | // Compute the index adjacent to V2Index and in the same half by toggling | |||
8433 | // the low bit. | |||
8434 | int V2AdjIndex = V2Index ^ 1; | |||
8435 | ||||
8436 | if (Mask[V2AdjIndex] == -1) { | |||
8437 | // Handles all the cases where we have a single V2 element and an undef. | |||
8438 | // This will only ever happen in the high lanes because we commute the | |||
8439 | // vector otherwise. | |||
8440 | if (V2Index < 2) | |||
8441 | std::swap(LowV, HighV); | |||
8442 | NewMask[V2Index] -= 4; | |||
8443 | } else { | |||
8444 | // Handle the case where the V2 element ends up adjacent to a V1 element. | |||
8445 | // To make this work, blend them together as the first step. | |||
8446 | int V1Index = V2AdjIndex; | |||
8447 | int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0}; | |||
8448 | V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1, | |||
8449 | getV4X86ShuffleImm8ForMask(BlendMask, DAG)); | |||
8450 | ||||
8451 | // Now proceed to reconstruct the final blend as we have the necessary | |||
8452 | // high or low half formed. | |||
8453 | if (V2Index < 2) { | |||
8454 | LowV = V2; | |||
8455 | HighV = V1; | |||
8456 | } else { | |||
8457 | HighV = V2; | |||
8458 | } | |||
8459 | NewMask[V1Index] = 2; // We put the V1 element in V2[2]. | |||
8460 | NewMask[V2Index] = 0; // We shifted the V2 element into V2[0]. | |||
8461 | } | |||
8462 | } else if (NumV2Elements == 2) { | |||
8463 | if (Mask[0] < 4 && Mask[1] < 4) { | |||
8464 | // Handle the easy case where we have V1 in the low lanes and V2 in the | |||
8465 | // high lanes. | |||
8466 | NewMask[2] -= 4; | |||
8467 | NewMask[3] -= 4; | |||
8468 | } else if (Mask[2] < 4 && Mask[3] < 4) { | |||
8469 | // We also handle the reversed case because this utility may get called | |||
8470 | // when we detect a SHUFPS pattern but can't easily commute the shuffle to | |||
8471 | // arrange things in the right direction. | |||
8472 | NewMask[0] -= 4; | |||
8473 | NewMask[1] -= 4; | |||
8474 | HighV = V1; | |||
8475 | LowV = V2; | |||
8476 | } else { | |||
8477 | // We have a mixture of V1 and V2 in both low and high lanes. Rather than | |||
8478 | // trying to place elements directly, just blend them and set up the final | |||
8479 | // shuffle to place them. | |||
8480 | ||||
8481 | // The first two blend mask elements are for V1, the second two are for | |||
8482 | // V2. | |||
8483 | int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1], | |||
8484 | Mask[2] < 4 ? Mask[2] : Mask[3], | |||
8485 | (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4, | |||
8486 | (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4}; | |||
8487 | V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2, | |||
8488 | getV4X86ShuffleImm8ForMask(BlendMask, DAG)); | |||
8489 | ||||
8490 | // Now we do a normal shuffle of V1 by giving V1 as both operands to | |||
8491 | // a blend. | |||
8492 | LowV = HighV = V1; | |||
8493 | NewMask[0] = Mask[0] < 4 ? 0 : 2; | |||
8494 | NewMask[1] = Mask[0] < 4 ? 2 : 0; | |||
8495 | NewMask[2] = Mask[2] < 4 ? 1 : 3; | |||
8496 | NewMask[3] = Mask[2] < 4 ? 3 : 1; | |||
8497 | } | |||
8498 | } | |||
8499 | return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV, | |||
8500 | getV4X86ShuffleImm8ForMask(NewMask, DAG)); | |||
8501 | } | |||
8502 | ||||
8503 | /// \brief Lower 4-lane 32-bit floating point shuffles. | |||
8504 | /// | |||
8505 | /// Uses instructions exclusively from the floating point unit to minimize | |||
8506 | /// domain crossing penalties, as these are sufficient to implement all v4f32 | |||
8507 | /// shuffles. | |||
8508 | static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
8509 | const X86Subtarget *Subtarget, | |||
8510 | SelectionDAG &DAG) { | |||
8511 | SDLoc DL(Op); | |||
8512 | assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!")((Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType() == MVT::v4f32 && \"Bad shuffle type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8512, __PRETTY_FUNCTION__)); | |||
8513 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8513, __PRETTY_FUNCTION__)); | |||
8514 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8514, __PRETTY_FUNCTION__)); | |||
8515 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
8516 | ArrayRef<int> Mask = SVOp->getMask(); | |||
8517 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8517, __PRETTY_FUNCTION__)); | |||
8518 | ||||
8519 | int NumV2Elements = | |||
8520 | std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }); | |||
8521 | ||||
8522 | if (NumV2Elements == 0) { | |||
8523 | // Check for being able to broadcast a single element. | |||
8524 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4f32, DL, V1, | |||
8525 | Mask, Subtarget, DAG)) | |||
8526 | return Broadcast; | |||
8527 | ||||
8528 | // Use even/odd duplicate instructions for masks that match their pattern. | |||
8529 | if (Subtarget->hasSSE3()) { | |||
8530 | if (isShuffleEquivalent(Mask, 0, 0, 2, 2)) | |||
8531 | return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1); | |||
8532 | if (isShuffleEquivalent(Mask, 1, 1, 3, 3)) | |||
8533 | return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1); | |||
8534 | } | |||
8535 | ||||
8536 | if (Subtarget->hasAVX()) { | |||
8537 | // If we have AVX, we can use VPERMILPS which will allow folding a load | |||
8538 | // into the shuffle. | |||
8539 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1, | |||
8540 | getV4X86ShuffleImm8ForMask(Mask, DAG)); | |||
8541 | } | |||
8542 | ||||
8543 | // Otherwise, use a straight shuffle of a single input vector. We pass the | |||
8544 | // input vector to both operands to simulate this with a SHUFPS. | |||
8545 | return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1, | |||
8546 | getV4X86ShuffleImm8ForMask(Mask, DAG)); | |||
8547 | } | |||
8548 | ||||
8549 | // Use dedicated unpack instructions for masks that match their pattern. | |||
8550 | if (isShuffleEquivalent(Mask, 0, 4, 1, 5)) | |||
8551 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2); | |||
8552 | if (isShuffleEquivalent(Mask, 2, 6, 3, 7)) | |||
8553 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2); | |||
8554 | ||||
8555 | // There are special ways we can lower some single-element blends. However, we | |||
8556 | // have custom ways we can lower more complex single-element blends below that | |||
8557 | // we defer to if both this and BLENDPS fail to match, so restrict this to | |||
8558 | // when the V2 input is targeting element 0 of the mask -- that is the fast | |||
8559 | // case here. | |||
8560 | if (NumV2Elements == 1 && Mask[0] >= 4) | |||
8561 | if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v4f32, DL, V1, V2, | |||
8562 | Mask, Subtarget, DAG)) | |||
8563 | return V; | |||
8564 | ||||
8565 | if (Subtarget->hasSSE41()) { | |||
8566 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask, | |||
8567 | Subtarget, DAG)) | |||
8568 | return Blend; | |||
8569 | ||||
8570 | // Use INSERTPS if we can complete the shuffle efficiently. | |||
8571 | if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG)) | |||
8572 | return V; | |||
8573 | } | |||
8574 | ||||
8575 | // Otherwise fall back to a SHUFPS lowering strategy. | |||
8576 | return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG); | |||
8577 | } | |||
8578 | ||||
8579 | /// \brief Lower 4-lane i32 vector shuffles. | |||
8580 | /// | |||
8581 | /// We try to handle these with integer-domain shuffles where we can, but for | |||
8582 | /// blends we use the floating point domain blend instructions. | |||
8583 | static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
8584 | const X86Subtarget *Subtarget, | |||
8585 | SelectionDAG &DAG) { | |||
8586 | SDLoc DL(Op); | |||
8587 | assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!")((Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType() == MVT::v4i32 && \"Bad shuffle type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8587, __PRETTY_FUNCTION__)); | |||
8588 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8588, __PRETTY_FUNCTION__)); | |||
8589 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8589, __PRETTY_FUNCTION__)); | |||
8590 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
8591 | ArrayRef<int> Mask = SVOp->getMask(); | |||
8592 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8592, __PRETTY_FUNCTION__)); | |||
8593 | ||||
8594 | // Whenever we can lower this as a zext, that instruction is strictly faster | |||
8595 | // than any alternative. It also allows us to fold memory operands into the | |||
8596 | // shuffle in many cases. | |||
8597 | if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2, | |||
8598 | Mask, Subtarget, DAG)) | |||
8599 | return ZExt; | |||
8600 | ||||
8601 | int NumV2Elements = | |||
8602 | std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }); | |||
8603 | ||||
8604 | if (NumV2Elements == 0) { | |||
8605 | // Check for being able to broadcast a single element. | |||
8606 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4i32, DL, V1, | |||
8607 | Mask, Subtarget, DAG)) | |||
8608 | return Broadcast; | |||
8609 | ||||
8610 | // Straight shuffle of a single input vector. For everything from SSE2 | |||
8611 | // onward this has a single fast instruction with no scary immediates. | |||
8612 | // We coerce the shuffle pattern to be compatible with UNPCK instructions | |||
8613 | // but we aren't actually going to use the UNPCK instruction because doing | |||
8614 | // so prevents folding a load into this instruction or making a copy. | |||
8615 | const int UnpackLoMask[] = {0, 0, 1, 1}; | |||
8616 | const int UnpackHiMask[] = {2, 2, 3, 3}; | |||
8617 | if (isShuffleEquivalent(Mask, 0, 0, 1, 1)) | |||
8618 | Mask = UnpackLoMask; | |||
8619 | else if (isShuffleEquivalent(Mask, 2, 2, 3, 3)) | |||
8620 | Mask = UnpackHiMask; | |||
8621 | ||||
8622 | return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1, | |||
8623 | getV4X86ShuffleImm8ForMask(Mask, DAG)); | |||
8624 | } | |||
8625 | ||||
8626 | // Try to use byte shift instructions. | |||
8627 | if (SDValue Shift = lowerVectorShuffleAsByteShift( | |||
8628 | DL, MVT::v4i32, V1, V2, Mask, DAG)) | |||
8629 | return Shift; | |||
8630 | ||||
8631 | // There are special ways we can lower some single-element blends. | |||
8632 | if (NumV2Elements == 1) | |||
8633 | if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v4i32, DL, V1, V2, | |||
8634 | Mask, Subtarget, DAG)) | |||
8635 | return V; | |||
8636 | ||||
8637 | // Use dedicated unpack instructions for masks that match their pattern. | |||
8638 | if (isShuffleEquivalent(Mask, 0, 4, 1, 5)) | |||
8639 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2); | |||
8640 | if (isShuffleEquivalent(Mask, 2, 6, 3, 7)) | |||
8641 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2); | |||
8642 | ||||
8643 | if (Subtarget->hasSSE41()) | |||
8644 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask, | |||
8645 | Subtarget, DAG)) | |||
8646 | return Blend; | |||
8647 | ||||
8648 | // Try to use byte rotation instructions. | |||
8649 | // Its more profitable for pre-SSSE3 to use shuffles/unpacks. | |||
8650 | if (Subtarget->hasSSSE3()) | |||
8651 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate( | |||
8652 | DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG)) | |||
8653 | return Rotate; | |||
8654 | ||||
8655 | // We implement this with SHUFPS because it can blend from two vectors. | |||
8656 | // Because we're going to eventually use SHUFPS, we use SHUFPS even to build | |||
8657 | // up the inputs, bypassing domain shift penalties that we would encur if we | |||
8658 | // directly used PSHUFD on Nehalem and older. For newer chips, this isn't | |||
8659 | // relevant. | |||
8660 | return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, | |||
8661 | DAG.getVectorShuffle( | |||
8662 | MVT::v4f32, DL, | |||
8663 | DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1), | |||
8664 | DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask)); | |||
8665 | } | |||
8666 | ||||
8667 | /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2 | |||
8668 | /// shuffle lowering, and the most complex part. | |||
8669 | /// | |||
8670 | /// The lowering strategy is to try to form pairs of input lanes which are | |||
8671 | /// targeted at the same half of the final vector, and then use a dword shuffle | |||
8672 | /// to place them onto the right half, and finally unpack the paired lanes into | |||
8673 | /// their final position. | |||
8674 | /// | |||
8675 | /// The exact breakdown of how to form these dword pairs and align them on the | |||
8676 | /// correct sides is really tricky. See the comments within the function for | |||
8677 | /// more of the details. | |||
8678 | static SDValue lowerV8I16SingleInputVectorShuffle( | |||
8679 | SDLoc DL, SDValue V, MutableArrayRef<int> Mask, | |||
8680 | const X86Subtarget *Subtarget, SelectionDAG &DAG) { | |||
8681 | assert(V.getSimpleValueType() == MVT::v8i16 && "Bad input type!")((V.getSimpleValueType() == MVT::v8i16 && "Bad input type!" ) ? static_cast<void> (0) : __assert_fail ("V.getSimpleValueType() == MVT::v8i16 && \"Bad input type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8681, __PRETTY_FUNCTION__)); | |||
8682 | MutableArrayRef<int> LoMask = Mask.slice(0, 4); | |||
8683 | MutableArrayRef<int> HiMask = Mask.slice(4, 4); | |||
8684 | ||||
8685 | SmallVector<int, 4> LoInputs; | |||
8686 | std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs), | |||
8687 | [](int M) { return M >= 0; }); | |||
8688 | std::sort(LoInputs.begin(), LoInputs.end()); | |||
8689 | LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end()); | |||
8690 | SmallVector<int, 4> HiInputs; | |||
8691 | std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs), | |||
8692 | [](int M) { return M >= 0; }); | |||
8693 | std::sort(HiInputs.begin(), HiInputs.end()); | |||
8694 | HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end()); | |||
8695 | int NumLToL = | |||
8696 | std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin(); | |||
8697 | int NumHToL = LoInputs.size() - NumLToL; | |||
8698 | int NumLToH = | |||
8699 | std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin(); | |||
8700 | int NumHToH = HiInputs.size() - NumLToH; | |||
8701 | MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL); | |||
8702 | MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH); | |||
8703 | MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL); | |||
8704 | MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH); | |||
8705 | ||||
8706 | // Check for being able to broadcast a single element. | |||
8707 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8i16, DL, V, | |||
8708 | Mask, Subtarget, DAG)) | |||
8709 | return Broadcast; | |||
8710 | ||||
8711 | // Try to use byte shift instructions. | |||
8712 | if (SDValue Shift = lowerVectorShuffleAsByteShift( | |||
8713 | DL, MVT::v8i16, V, V, Mask, DAG)) | |||
8714 | return Shift; | |||
8715 | ||||
8716 | // Use dedicated unpack instructions for masks that match their pattern. | |||
8717 | if (isShuffleEquivalent(Mask, 0, 0, 1, 1, 2, 2, 3, 3)) | |||
8718 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V, V); | |||
8719 | if (isShuffleEquivalent(Mask, 4, 4, 5, 5, 6, 6, 7, 7)) | |||
8720 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V, V); | |||
8721 | ||||
8722 | // Try to use byte rotation instructions. | |||
8723 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate( | |||
8724 | DL, MVT::v8i16, V, V, Mask, Subtarget, DAG)) | |||
8725 | return Rotate; | |||
8726 | ||||
8727 | // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all | |||
8728 | // such inputs we can swap two of the dwords across the half mark and end up | |||
8729 | // with <=2 inputs to each half in each half. Once there, we can fall through | |||
8730 | // to the generic code below. For example: | |||
8731 | // | |||
8732 | // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h] | |||
8733 | // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5] | |||
8734 | // | |||
8735 | // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half | |||
8736 | // and an existing 2-into-2 on the other half. In this case we may have to | |||
8737 | // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or | |||
8738 | // 1-into-3 which could cause us to cycle endlessly fixing each side in turn. | |||
8739 | // Fortunately, we don't have to handle anything but a 2-into-2 pattern | |||
8740 | // because any other situation (including a 3-into-1 or 1-into-3 in the other | |||
8741 | // half than the one we target for fixing) will be fixed when we re-enter this | |||
8742 | // path. We will also combine away any sequence of PSHUFD instructions that | |||
8743 | // result into a single instruction. Here is an example of the tricky case: | |||
8744 | // | |||
8745 | // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h] | |||
8746 | // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3] | |||
8747 | // | |||
8748 | // This now has a 1-into-3 in the high half! Instead, we do two shuffles: | |||
8749 | // | |||
8750 | // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h] | |||
8751 | // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6] | |||
8752 | // | |||
8753 | // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h] | |||
8754 | // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6] | |||
8755 | // | |||
8756 | // The result is fine to be handled by the generic logic. | |||
8757 | auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs, | |||
8758 | ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs, | |||
8759 | int AOffset, int BOffset) { | |||
8760 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8761, __PRETTY_FUNCTION__)) | |||
8761 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8761, __PRETTY_FUNCTION__)); | |||
8762 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8763, __PRETTY_FUNCTION__)) | |||
8763 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8763, __PRETTY_FUNCTION__)); | |||
8764 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8765, __PRETTY_FUNCTION__)) | |||
8765 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8765, __PRETTY_FUNCTION__)); | |||
8766 | ||||
8767 | // Compute the index of dword with only one word among the three inputs in | |||
8768 | // a half by taking the sum of the half with three inputs and subtracting | |||
8769 | // the sum of the actual three inputs. The difference is the remaining | |||
8770 | // slot. | |||
8771 | int ADWord, BDWord; | |||
8772 | int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord; | |||
8773 | int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord; | |||
8774 | int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset; | |||
8775 | ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs; | |||
8776 | int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0]; | |||
8777 | int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset); | |||
8778 | int TripleNonInputIdx = | |||
8779 | TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0); | |||
8780 | TripleDWord = TripleNonInputIdx / 2; | |||
8781 | ||||
8782 | // We use xor with one to compute the adjacent DWord to whichever one the | |||
8783 | // OneInput is in. | |||
8784 | OneInputDWord = (OneInput / 2) ^ 1; | |||
8785 | ||||
8786 | // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA | |||
8787 | // and BToA inputs. If there is also such a problem with the BToB and AToB | |||
8788 | // inputs, we don't try to fix it necessarily -- we'll recurse and see it in | |||
8789 | // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it | |||
8790 | // is essential that we don't *create* a 3<-1 as then we might oscillate. | |||
8791 | if (BToBInputs.size() == 2 && AToBInputs.size() == 2) { | |||
8792 | // Compute how many inputs will be flipped by swapping these DWords. We | |||
8793 | // need | |||
8794 | // to balance this to ensure we don't form a 3-1 shuffle in the other | |||
8795 | // half. | |||
8796 | int NumFlippedAToBInputs = | |||
8797 | std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) + | |||
8798 | std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1); | |||
8799 | int NumFlippedBToBInputs = | |||
8800 | std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) + | |||
8801 | std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1); | |||
8802 | if ((NumFlippedAToBInputs == 1 && | |||
8803 | (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) || | |||
8804 | (NumFlippedBToBInputs == 1 && | |||
8805 | (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) { | |||
8806 | // We choose whether to fix the A half or B half based on whether that | |||
8807 | // half has zero flipped inputs. At zero, we may not be able to fix it | |||
8808 | // with that half. We also bias towards fixing the B half because that | |||
8809 | // will more commonly be the high half, and we have to bias one way. | |||
8810 | auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord, | |||
8811 | ArrayRef<int> Inputs) { | |||
8812 | int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot. | |||
8813 | bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(), | |||
8814 | PinnedIdx ^ 1) != Inputs.end(); | |||
8815 | // Determine whether the free index is in the flipped dword or the | |||
8816 | // unflipped dword based on where the pinned index is. We use this bit | |||
8817 | // in an xor to conditionally select the adjacent dword. | |||
8818 | int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord)); | |||
8819 | bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(), | |||
8820 | FixFreeIdx) != Inputs.end(); | |||
8821 | if (IsFixIdxInput == IsFixFreeIdxInput) | |||
8822 | FixFreeIdx += 1; | |||
8823 | IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(), | |||
8824 | FixFreeIdx) != Inputs.end(); | |||
8825 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8826, __PRETTY_FUNCTION__)) | |||
8826 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8826, __PRETTY_FUNCTION__)); | |||
8827 | int PSHUFHalfMask[] = {0, 1, 2, 3}; | |||
8828 | std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]); | |||
8829 | V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL, | |||
8830 | MVT::v8i16, V, | |||
8831 | getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DAG)); | |||
8832 | ||||
8833 | for (int &M : Mask) | |||
8834 | if (M != -1 && M == FixIdx) | |||
8835 | M = FixFreeIdx; | |||
8836 | else if (M != -1 && M == FixFreeIdx) | |||
8837 | M = FixIdx; | |||
8838 | }; | |||
8839 | if (NumFlippedBToBInputs != 0) { | |||
8840 | int BPinnedIdx = | |||
8841 | BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput; | |||
8842 | FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs); | |||
8843 | } else { | |||
8844 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8844, __PRETTY_FUNCTION__)); | |||
8845 | int APinnedIdx = | |||
8846 | AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput; | |||
8847 | FixFlippedInputs(APinnedIdx, ADWord, AToBInputs); | |||
8848 | } | |||
8849 | } | |||
8850 | } | |||
8851 | ||||
8852 | int PSHUFDMask[] = {0, 1, 2, 3}; | |||
8853 | PSHUFDMask[ADWord] = BDWord; | |||
8854 | PSHUFDMask[BDWord] = ADWord; | |||
8855 | V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, | |||
8856 | DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, | |||
8857 | DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V), | |||
8858 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG))); | |||
8859 | ||||
8860 | // Adjust the mask to match the new locations of A and B. | |||
8861 | for (int &M : Mask) | |||
8862 | if (M != -1 && M/2 == ADWord) | |||
8863 | M = 2 * BDWord + M % 2; | |||
8864 | else if (M != -1 && M/2 == BDWord) | |||
8865 | M = 2 * ADWord + M % 2; | |||
8866 | ||||
8867 | // Recurse back into this routine to re-compute state now that this isn't | |||
8868 | // a 3 and 1 problem. | |||
8869 | return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16), | |||
8870 | Mask); | |||
8871 | }; | |||
8872 | if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3)) | |||
8873 | return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4); | |||
8874 | else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3)) | |||
8875 | return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0); | |||
8876 | ||||
8877 | // At this point there are at most two inputs to the low and high halves from | |||
8878 | // each half. That means the inputs can always be grouped into dwords and | |||
8879 | // those dwords can then be moved to the correct half with a dword shuffle. | |||
8880 | // We use at most one low and one high word shuffle to collect these paired | |||
8881 | // inputs into dwords, and finally a dword shuffle to place them. | |||
8882 | int PSHUFLMask[4] = {-1, -1, -1, -1}; | |||
8883 | int PSHUFHMask[4] = {-1, -1, -1, -1}; | |||
8884 | int PSHUFDMask[4] = {-1, -1, -1, -1}; | |||
8885 | ||||
8886 | // First fix the masks for all the inputs that are staying in their | |||
8887 | // original halves. This will then dictate the targets of the cross-half | |||
8888 | // shuffles. | |||
8889 | auto fixInPlaceInputs = | |||
8890 | [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs, | |||
8891 | MutableArrayRef<int> SourceHalfMask, | |||
8892 | MutableArrayRef<int> HalfMask, int HalfOffset) { | |||
8893 | if (InPlaceInputs.empty()) | |||
8894 | return; | |||
8895 | if (InPlaceInputs.size() == 1) { | |||
8896 | SourceHalfMask[InPlaceInputs[0] - HalfOffset] = | |||
8897 | InPlaceInputs[0] - HalfOffset; | |||
8898 | PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2; | |||
8899 | return; | |||
8900 | } | |||
8901 | if (IncomingInputs.empty()) { | |||
8902 | // Just fix all of the in place inputs. | |||
8903 | for (int Input : InPlaceInputs) { | |||
8904 | SourceHalfMask[Input - HalfOffset] = Input - HalfOffset; | |||
8905 | PSHUFDMask[Input / 2] = Input / 2; | |||
8906 | } | |||
8907 | return; | |||
8908 | } | |||
8909 | ||||
8910 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8910, __PRETTY_FUNCTION__)); | |||
8911 | SourceHalfMask[InPlaceInputs[0] - HalfOffset] = | |||
8912 | InPlaceInputs[0] - HalfOffset; | |||
8913 | // Put the second input next to the first so that they are packed into | |||
8914 | // a dword. We find the adjacent index by toggling the low bit. | |||
8915 | int AdjIndex = InPlaceInputs[0] ^ 1; | |||
8916 | SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset; | |||
8917 | std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex); | |||
8918 | PSHUFDMask[AdjIndex / 2] = AdjIndex / 2; | |||
8919 | }; | |||
8920 | fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0); | |||
8921 | fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4); | |||
8922 | ||||
8923 | // Now gather the cross-half inputs and place them into a free dword of | |||
8924 | // their target half. | |||
8925 | // FIXME: This operation could almost certainly be simplified dramatically to | |||
8926 | // look more like the 3-1 fixing operation. | |||
8927 | auto moveInputsToRightHalf = [&PSHUFDMask]( | |||
8928 | MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs, | |||
8929 | MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask, | |||
8930 | MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset, | |||
8931 | int DestOffset) { | |||
8932 | auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) { | |||
8933 | return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word; | |||
8934 | }; | |||
8935 | auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask, | |||
8936 | int Word) { | |||
8937 | int LowWord = Word & ~1; | |||
8938 | int HighWord = Word | 1; | |||
8939 | return isWordClobbered(SourceHalfMask, LowWord) || | |||
8940 | isWordClobbered(SourceHalfMask, HighWord); | |||
8941 | }; | |||
8942 | ||||
8943 | if (IncomingInputs.empty()) | |||
8944 | return; | |||
8945 | ||||
8946 | if (ExistingInputs.empty()) { | |||
8947 | // Map any dwords with inputs from them into the right half. | |||
8948 | for (int Input : IncomingInputs) { | |||
8949 | // If the source half mask maps over the inputs, turn those into | |||
8950 | // swaps and use the swapped lane. | |||
8951 | if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) { | |||
8952 | if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) { | |||
8953 | SourceHalfMask[SourceHalfMask[Input - SourceOffset]] = | |||
8954 | Input - SourceOffset; | |||
8955 | // We have to swap the uses in our half mask in one sweep. | |||
8956 | for (int &M : HalfMask) | |||
8957 | if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset) | |||
8958 | M = Input; | |||
8959 | else if (M == Input) | |||
8960 | M = SourceHalfMask[Input - SourceOffset] + SourceOffset; | |||
8961 | } else { | |||
8962 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8964, __PRETTY_FUNCTION__)) | |||
8963 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8964, __PRETTY_FUNCTION__)) | |||
8964 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8964, __PRETTY_FUNCTION__)); | |||
8965 | } | |||
8966 | // Note that this correctly re-maps both when we do a swap and when | |||
8967 | // we observe the other side of the swap above. We rely on that to | |||
8968 | // avoid swapping the members of the input list directly. | |||
8969 | Input = SourceHalfMask[Input - SourceOffset] + SourceOffset; | |||
8970 | } | |||
8971 | ||||
8972 | // Map the input's dword into the correct half. | |||
8973 | if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1) | |||
8974 | PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2; | |||
8975 | else | |||
8976 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8978, __PRETTY_FUNCTION__)) | |||
8977 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8978, __PRETTY_FUNCTION__)) | |||
8978 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8978, __PRETTY_FUNCTION__)); | |||
8979 | } | |||
8980 | ||||
8981 | // And just directly shift any other-half mask elements to be same-half | |||
8982 | // as we will have mirrored the dword containing the element into the | |||
8983 | // same position within that half. | |||
8984 | for (int &M : HalfMask) | |||
8985 | if (M >= SourceOffset && M < SourceOffset + 4) { | |||
8986 | M = M - SourceOffset + DestOffset; | |||
8987 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 8987, __PRETTY_FUNCTION__)); | |||
8988 | } | |||
8989 | return; | |||
8990 | } | |||
8991 | ||||
8992 | // Ensure we have the input in a viable dword of its current half. This | |||
8993 | // is particularly tricky because the original position may be clobbered | |||
8994 | // by inputs being moved and *staying* in that half. | |||
8995 | if (IncomingInputs.size() == 1) { | |||
8996 | if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) { | |||
8997 | int InputFixed = std::find(std::begin(SourceHalfMask), | |||
8998 | std::end(SourceHalfMask), -1) - | |||
8999 | std::begin(SourceHalfMask) + SourceOffset; | |||
9000 | SourceHalfMask[InputFixed - SourceOffset] = | |||
9001 | IncomingInputs[0] - SourceOffset; | |||
9002 | std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0], | |||
9003 | InputFixed); | |||
9004 | IncomingInputs[0] = InputFixed; | |||
9005 | } | |||
9006 | } else if (IncomingInputs.size() == 2) { | |||
9007 | if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 || | |||
9008 | isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) { | |||
9009 | // We have two non-adjacent or clobbered inputs we need to extract from | |||
9010 | // the source half. To do this, we need to map them into some adjacent | |||
9011 | // dword slot in the source mask. | |||
9012 | int InputsFixed[2] = {IncomingInputs[0] - SourceOffset, | |||
9013 | IncomingInputs[1] - SourceOffset}; | |||
9014 | ||||
9015 | // If there is a free slot in the source half mask adjacent to one of | |||
9016 | // the inputs, place the other input in it. We use (Index XOR 1) to | |||
9017 | // compute an adjacent index. | |||
9018 | if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) && | |||
9019 | SourceHalfMask[InputsFixed[0] ^ 1] == -1) { | |||
9020 | SourceHalfMask[InputsFixed[0]] = InputsFixed[0]; | |||
9021 | SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1]; | |||
9022 | InputsFixed[1] = InputsFixed[0] ^ 1; | |||
9023 | } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) && | |||
9024 | SourceHalfMask[InputsFixed[1] ^ 1] == -1) { | |||
9025 | SourceHalfMask[InputsFixed[1]] = InputsFixed[1]; | |||
9026 | SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0]; | |||
9027 | InputsFixed[0] = InputsFixed[1] ^ 1; | |||
9028 | } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 && | |||
9029 | SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) { | |||
9030 | // The two inputs are in the same DWord but it is clobbered and the | |||
9031 | // adjacent DWord isn't used at all. Move both inputs to the free | |||
9032 | // slot. | |||
9033 | SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0]; | |||
9034 | SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1]; | |||
9035 | InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1); | |||
9036 | InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1; | |||
9037 | } else { | |||
9038 | // The only way we hit this point is if there is no clobbering | |||
9039 | // (because there are no off-half inputs to this half) and there is no | |||
9040 | // free slot adjacent to one of the inputs. In this case, we have to | |||
9041 | // swap an input with a non-input. | |||
9042 | for (int i = 0; i < 4; ++i) | |||
9043 | assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&(((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) && "We can't handle any clobbers here!") ? static_cast<void> (0) : __assert_fail ("(SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) && \"We can't handle any clobbers here!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9044, __PRETTY_FUNCTION__)) | |||
9044 | "We can't handle any clobbers here!")(((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) && "We can't handle any clobbers here!") ? static_cast<void> (0) : __assert_fail ("(SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) && \"We can't handle any clobbers here!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9044, __PRETTY_FUNCTION__)); | |||
9045 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9046, __PRETTY_FUNCTION__)) | |||
9046 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9046, __PRETTY_FUNCTION__)); | |||
9047 | ||||
9048 | SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1]; | |||
9049 | SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1; | |||
9050 | ||||
9051 | // We also have to update the final source mask in this case because | |||
9052 | // it may need to undo the above swap. | |||
9053 | for (int &M : FinalSourceHalfMask) | |||
9054 | if (M == (InputsFixed[0] ^ 1) + SourceOffset) | |||
9055 | M = InputsFixed[1] + SourceOffset; | |||
9056 | else if (M == InputsFixed[1] + SourceOffset) | |||
9057 | M = (InputsFixed[0] ^ 1) + SourceOffset; | |||
9058 | ||||
9059 | InputsFixed[1] = InputsFixed[0] ^ 1; | |||
9060 | } | |||
9061 | ||||
9062 | // Point everything at the fixed inputs. | |||
9063 | for (int &M : HalfMask) | |||
9064 | if (M == IncomingInputs[0]) | |||
9065 | M = InputsFixed[0] + SourceOffset; | |||
9066 | else if (M == IncomingInputs[1]) | |||
9067 | M = InputsFixed[1] + SourceOffset; | |||
9068 | ||||
9069 | IncomingInputs[0] = InputsFixed[0] + SourceOffset; | |||
9070 | IncomingInputs[1] = InputsFixed[1] + SourceOffset; | |||
9071 | } | |||
9072 | } else { | |||
9073 | llvm_unreachable("Unhandled input size!")::llvm::llvm_unreachable_internal("Unhandled input size!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9073); | |||
9074 | } | |||
9075 | ||||
9076 | // Now hoist the DWord down to the right half. | |||
9077 | int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2; | |||
9078 | assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free")((PSHUFDMask[FreeDWord] == -1 && "DWord not free") ? static_cast <void> (0) : __assert_fail ("PSHUFDMask[FreeDWord] == -1 && \"DWord not free\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9078, __PRETTY_FUNCTION__)); | |||
9079 | PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2; | |||
9080 | for (int &M : HalfMask) | |||
9081 | for (int Input : IncomingInputs) | |||
9082 | if (M == Input) | |||
9083 | M = FreeDWord * 2 + Input % 2; | |||
9084 | }; | |||
9085 | moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask, | |||
9086 | /*SourceOffset*/ 4, /*DestOffset*/ 0); | |||
9087 | moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask, | |||
9088 | /*SourceOffset*/ 0, /*DestOffset*/ 4); | |||
9089 | ||||
9090 | // Now enact all the shuffles we've computed to move the inputs into their | |||
9091 | // target half. | |||
9092 | if (!isNoopShuffleMask(PSHUFLMask)) | |||
9093 | V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V, | |||
9094 | getV4X86ShuffleImm8ForMask(PSHUFLMask, DAG)); | |||
9095 | if (!isNoopShuffleMask(PSHUFHMask)) | |||
9096 | V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V, | |||
9097 | getV4X86ShuffleImm8ForMask(PSHUFHMask, DAG)); | |||
9098 | if (!isNoopShuffleMask(PSHUFDMask)) | |||
9099 | V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, | |||
9100 | DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, | |||
9101 | DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V), | |||
9102 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG))); | |||
9103 | ||||
9104 | // At this point, each half should contain all its inputs, and we can then | |||
9105 | // just shuffle them into their final position. | |||
9106 | assert(std::count_if(LoMask.begin(), LoMask.end(),((std::count_if(LoMask.begin(), LoMask.end(), [](int M) { return M >= 4; }) == 0 && "Failed to lift all the high half inputs to the low mask!" ) ? static_cast<void> (0) : __assert_fail ("std::count_if(LoMask.begin(), LoMask.end(), [](int M) { return M >= 4; }) == 0 && \"Failed to lift all the high half inputs to the low mask!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9108, __PRETTY_FUNCTION__)) | |||
9107 | [](int M) { return M >= 4; }) == 0 &&((std::count_if(LoMask.begin(), LoMask.end(), [](int M) { return M >= 4; }) == 0 && "Failed to lift all the high half inputs to the low mask!" ) ? static_cast<void> (0) : __assert_fail ("std::count_if(LoMask.begin(), LoMask.end(), [](int M) { return M >= 4; }) == 0 && \"Failed to lift all the high half inputs to the low mask!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9108, __PRETTY_FUNCTION__)) | |||
9108 | "Failed to lift all the high half inputs to the low mask!")((std::count_if(LoMask.begin(), LoMask.end(), [](int M) { return M >= 4; }) == 0 && "Failed to lift all the high half inputs to the low mask!" ) ? static_cast<void> (0) : __assert_fail ("std::count_if(LoMask.begin(), LoMask.end(), [](int M) { return M >= 4; }) == 0 && \"Failed to lift all the high half inputs to the low mask!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9108, __PRETTY_FUNCTION__)); | |||
9109 | assert(std::count_if(HiMask.begin(), HiMask.end(),((std::count_if(HiMask.begin(), HiMask.end(), [](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 ("std::count_if(HiMask.begin(), HiMask.end(), [](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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9111, __PRETTY_FUNCTION__)) | |||
9110 | [](int M) { return M >= 0 && M < 4; }) == 0 &&((std::count_if(HiMask.begin(), HiMask.end(), [](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 ("std::count_if(HiMask.begin(), HiMask.end(), [](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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9111, __PRETTY_FUNCTION__)) | |||
9111 | "Failed to lift all the low half inputs to the high mask!")((std::count_if(HiMask.begin(), HiMask.end(), [](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 ("std::count_if(HiMask.begin(), HiMask.end(), [](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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9111, __PRETTY_FUNCTION__)); | |||
9112 | ||||
9113 | // Do a half shuffle for the low mask. | |||
9114 | if (!isNoopShuffleMask(LoMask)) | |||
9115 | V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V, | |||
9116 | getV4X86ShuffleImm8ForMask(LoMask, DAG)); | |||
9117 | ||||
9118 | // Do a half shuffle with the high mask after shifting its values down. | |||
9119 | for (int &M : HiMask) | |||
9120 | if (M >= 0) | |||
9121 | M -= 4; | |||
9122 | if (!isNoopShuffleMask(HiMask)) | |||
9123 | V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V, | |||
9124 | getV4X86ShuffleImm8ForMask(HiMask, DAG)); | |||
9125 | ||||
9126 | return V; | |||
9127 | } | |||
9128 | ||||
9129 | /// \brief Detect whether the mask pattern should be lowered through | |||
9130 | /// interleaving. | |||
9131 | /// | |||
9132 | /// This essentially tests whether viewing the mask as an interleaving of two | |||
9133 | /// sub-sequences reduces the cross-input traffic of a blend operation. If so, | |||
9134 | /// lowering it through interleaving is a significantly better strategy. | |||
9135 | static bool shouldLowerAsInterleaving(ArrayRef<int> Mask) { | |||
9136 | int NumEvenInputs[2] = {0, 0}; | |||
9137 | int NumOddInputs[2] = {0, 0}; | |||
9138 | int NumLoInputs[2] = {0, 0}; | |||
9139 | int NumHiInputs[2] = {0, 0}; | |||
9140 | for (int i = 0, Size = Mask.size(); i < Size; ++i) { | |||
9141 | if (Mask[i] < 0) | |||
9142 | continue; | |||
9143 | ||||
9144 | int InputIdx = Mask[i] >= Size; | |||
9145 | ||||
9146 | if (i < Size / 2) | |||
9147 | ++NumLoInputs[InputIdx]; | |||
9148 | else | |||
9149 | ++NumHiInputs[InputIdx]; | |||
9150 | ||||
9151 | if ((i % 2) == 0) | |||
9152 | ++NumEvenInputs[InputIdx]; | |||
9153 | else | |||
9154 | ++NumOddInputs[InputIdx]; | |||
9155 | } | |||
9156 | ||||
9157 | // The minimum number of cross-input results for both the interleaved and | |||
9158 | // split cases. If interleaving results in fewer cross-input results, return | |||
9159 | // true. | |||
9160 | int InterleavedCrosses = std::min(NumEvenInputs[1] + NumOddInputs[0], | |||
9161 | NumEvenInputs[0] + NumOddInputs[1]); | |||
9162 | int SplitCrosses = std::min(NumLoInputs[1] + NumHiInputs[0], | |||
9163 | NumLoInputs[0] + NumHiInputs[1]); | |||
9164 | return InterleavedCrosses < SplitCrosses; | |||
9165 | } | |||
9166 | ||||
9167 | /// \brief Blend two v8i16 vectors using a naive unpack strategy. | |||
9168 | /// | |||
9169 | /// This strategy only works when the inputs from each vector fit into a single | |||
9170 | /// half of that vector, and generally there are not so many inputs as to leave | |||
9171 | /// the in-place shuffles required highly constrained (and thus expensive). It | |||
9172 | /// shifts all the inputs into a single side of both input vectors and then | |||
9173 | /// uses an unpack to interleave these inputs in a single vector. At that | |||
9174 | /// point, we will fall back on the generic single input shuffle lowering. | |||
9175 | static SDValue lowerV8I16BasicBlendVectorShuffle(SDLoc DL, SDValue V1, | |||
9176 | SDValue V2, | |||
9177 | MutableArrayRef<int> Mask, | |||
9178 | const X86Subtarget *Subtarget, | |||
9179 | SelectionDAG &DAG) { | |||
9180 | assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad input type!")((V1.getSimpleValueType() == MVT::v8i16 && "Bad input type!" ) ? static_cast<void> (0) : __assert_fail ("V1.getSimpleValueType() == MVT::v8i16 && \"Bad input type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9180, __PRETTY_FUNCTION__)); | |||
9181 | assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad input type!")((V2.getSimpleValueType() == MVT::v8i16 && "Bad input type!" ) ? static_cast<void> (0) : __assert_fail ("V2.getSimpleValueType() == MVT::v8i16 && \"Bad input type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9181, __PRETTY_FUNCTION__)); | |||
9182 | SmallVector<int, 3> LoV1Inputs, HiV1Inputs, LoV2Inputs, HiV2Inputs; | |||
9183 | for (int i = 0; i < 8; ++i) | |||
9184 | if (Mask[i] >= 0 && Mask[i] < 4) | |||
9185 | LoV1Inputs.push_back(i); | |||
9186 | else if (Mask[i] >= 4 && Mask[i] < 8) | |||
9187 | HiV1Inputs.push_back(i); | |||
9188 | else if (Mask[i] >= 8 && Mask[i] < 12) | |||
9189 | LoV2Inputs.push_back(i); | |||
9190 | else if (Mask[i] >= 12) | |||
9191 | HiV2Inputs.push_back(i); | |||
9192 | ||||
9193 | int NumV1Inputs = LoV1Inputs.size() + HiV1Inputs.size(); | |||
9194 | int NumV2Inputs = LoV2Inputs.size() + HiV2Inputs.size(); | |||
9195 | (void)NumV1Inputs; | |||
9196 | (void)NumV2Inputs; | |||
9197 | assert(NumV1Inputs > 0 && NumV1Inputs <= 3 && "At most 3 inputs supported")((NumV1Inputs > 0 && NumV1Inputs <= 3 && "At most 3 inputs supported") ? static_cast<void> (0) : __assert_fail ("NumV1Inputs > 0 && NumV1Inputs <= 3 && \"At most 3 inputs supported\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9197, __PRETTY_FUNCTION__)); | |||
9198 | assert(NumV2Inputs > 0 && NumV2Inputs <= 3 && "At most 3 inputs supported")((NumV2Inputs > 0 && NumV2Inputs <= 3 && "At most 3 inputs supported") ? static_cast<void> (0) : __assert_fail ("NumV2Inputs > 0 && NumV2Inputs <= 3 && \"At most 3 inputs supported\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9198, __PRETTY_FUNCTION__)); | |||
9199 | assert(NumV1Inputs + NumV2Inputs <= 4 && "At most 4 combined inputs")((NumV1Inputs + NumV2Inputs <= 4 && "At most 4 combined inputs" ) ? static_cast<void> (0) : __assert_fail ("NumV1Inputs + NumV2Inputs <= 4 && \"At most 4 combined inputs\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9199, __PRETTY_FUNCTION__)); | |||
9200 | ||||
9201 | bool MergeFromLo = LoV1Inputs.size() + LoV2Inputs.size() >= | |||
9202 | HiV1Inputs.size() + HiV2Inputs.size(); | |||
9203 | ||||
9204 | auto moveInputsToHalf = [&](SDValue V, ArrayRef<int> LoInputs, | |||
9205 | ArrayRef<int> HiInputs, bool MoveToLo, | |||
9206 | int MaskOffset) { | |||
9207 | ArrayRef<int> GoodInputs = MoveToLo ? LoInputs : HiInputs; | |||
9208 | ArrayRef<int> BadInputs = MoveToLo ? HiInputs : LoInputs; | |||
9209 | if (BadInputs.empty()) | |||
9210 | return V; | |||
9211 | ||||
9212 | int MoveMask[] = {-1, -1, -1, -1, -1, -1, -1, -1}; | |||
9213 | int MoveOffset = MoveToLo ? 0 : 4; | |||
9214 | ||||
9215 | if (GoodInputs.empty()) { | |||
9216 | for (int BadInput : BadInputs) { | |||
9217 | MoveMask[Mask[BadInput] % 4 + MoveOffset] = Mask[BadInput] - MaskOffset; | |||
9218 | Mask[BadInput] = Mask[BadInput] % 4 + MoveOffset + MaskOffset; | |||
9219 | } | |||
9220 | } else { | |||
9221 | if (GoodInputs.size() == 2) { | |||
9222 | // If the low inputs are spread across two dwords, pack them into | |||
9223 | // a single dword. | |||
9224 | MoveMask[MoveOffset] = Mask[GoodInputs[0]] - MaskOffset; | |||
9225 | MoveMask[MoveOffset + 1] = Mask[GoodInputs[1]] - MaskOffset; | |||
9226 | Mask[GoodInputs[0]] = MoveOffset + MaskOffset; | |||
9227 | Mask[GoodInputs[1]] = MoveOffset + 1 + MaskOffset; | |||
9228 | } else { | |||
9229 | // Otherwise pin the good inputs. | |||
9230 | for (int GoodInput : GoodInputs) | |||
9231 | MoveMask[Mask[GoodInput] - MaskOffset] = Mask[GoodInput] - MaskOffset; | |||
9232 | } | |||
9233 | ||||
9234 | if (BadInputs.size() == 2) { | |||
9235 | // If we have two bad inputs then there may be either one or two good | |||
9236 | // inputs fixed in place. Find a fixed input, and then find the *other* | |||
9237 | // two adjacent indices by using modular arithmetic. | |||
9238 | int GoodMaskIdx = | |||
9239 | std::find_if(std::begin(MoveMask) + MoveOffset, std::end(MoveMask), | |||
9240 | [](int M) { return M >= 0; }) - | |||
9241 | std::begin(MoveMask); | |||
9242 | int MoveMaskIdx = | |||
9243 | ((((GoodMaskIdx - MoveOffset) & ~1) + 2) % 4) + MoveOffset; | |||
9244 | assert(MoveMask[MoveMaskIdx] == -1 && "Expected empty slot")((MoveMask[MoveMaskIdx] == -1 && "Expected empty slot" ) ? static_cast<void> (0) : __assert_fail ("MoveMask[MoveMaskIdx] == -1 && \"Expected empty slot\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9244, __PRETTY_FUNCTION__)); | |||
9245 | assert(MoveMask[MoveMaskIdx + 1] == -1 && "Expected empty slot")((MoveMask[MoveMaskIdx + 1] == -1 && "Expected empty slot" ) ? static_cast<void> (0) : __assert_fail ("MoveMask[MoveMaskIdx + 1] == -1 && \"Expected empty slot\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9245, __PRETTY_FUNCTION__)); | |||
9246 | MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset; | |||
9247 | MoveMask[MoveMaskIdx + 1] = Mask[BadInputs[1]] - MaskOffset; | |||
9248 | Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset; | |||
9249 | Mask[BadInputs[1]] = MoveMaskIdx + 1 + MaskOffset; | |||
9250 | } else { | |||
9251 | assert(BadInputs.size() == 1 && "All sizes handled")((BadInputs.size() == 1 && "All sizes handled") ? static_cast <void> (0) : __assert_fail ("BadInputs.size() == 1 && \"All sizes handled\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9251, __PRETTY_FUNCTION__)); | |||
9252 | int MoveMaskIdx = std::find(std::begin(MoveMask) + MoveOffset, | |||
9253 | std::end(MoveMask), -1) - | |||
9254 | std::begin(MoveMask); | |||
9255 | MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset; | |||
9256 | Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset; | |||
9257 | } | |||
9258 | } | |||
9259 | ||||
9260 | return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16), | |||
9261 | MoveMask); | |||
9262 | }; | |||
9263 | V1 = moveInputsToHalf(V1, LoV1Inputs, HiV1Inputs, MergeFromLo, | |||
9264 | /*MaskOffset*/ 0); | |||
9265 | V2 = moveInputsToHalf(V2, LoV2Inputs, HiV2Inputs, MergeFromLo, | |||
9266 | /*MaskOffset*/ 8); | |||
9267 | ||||
9268 | // FIXME: Select an interleaving of the merge of V1 and V2 that minimizes | |||
9269 | // cross-half traffic in the final shuffle. | |||
9270 | ||||
9271 | // Munge the mask to be a single-input mask after the unpack merges the | |||
9272 | // results. | |||
9273 | for (int &M : Mask) | |||
9274 | if (M != -1) | |||
9275 | M = 2 * (M % 4) + (M / 8); | |||
9276 | ||||
9277 | return DAG.getVectorShuffle( | |||
9278 | MVT::v8i16, DL, DAG.getNode(MergeFromLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, | |||
9279 | DL, MVT::v8i16, V1, V2), | |||
9280 | DAG.getUNDEF(MVT::v8i16), Mask); | |||
9281 | } | |||
9282 | ||||
9283 | /// \brief Generic lowering of 8-lane i16 shuffles. | |||
9284 | /// | |||
9285 | /// This handles both single-input shuffles and combined shuffle/blends with | |||
9286 | /// two inputs. The single input shuffles are immediately delegated to | |||
9287 | /// a dedicated lowering routine. | |||
9288 | /// | |||
9289 | /// The blends are lowered in one of three fundamental ways. If there are few | |||
9290 | /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle | |||
9291 | /// of the input is significantly cheaper when lowered as an interleaving of | |||
9292 | /// the two inputs, try to interleave them. Otherwise, blend the low and high | |||
9293 | /// halves of the inputs separately (making them have relatively few inputs) | |||
9294 | /// and then concatenate them. | |||
9295 | static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
9296 | const X86Subtarget *Subtarget, | |||
9297 | SelectionDAG &DAG) { | |||
9298 | SDLoc DL(Op); | |||
9299 | assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!")((Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType() == MVT::v8i16 && \"Bad shuffle type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9299, __PRETTY_FUNCTION__)); | |||
9300 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9300, __PRETTY_FUNCTION__)); | |||
9301 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9301, __PRETTY_FUNCTION__)); | |||
9302 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
9303 | ArrayRef<int> OrigMask = SVOp->getMask(); | |||
9304 | int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3], | |||
9305 | OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]}; | |||
9306 | MutableArrayRef<int> Mask(MaskStorage); | |||
9307 | ||||
9308 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9308, __PRETTY_FUNCTION__)); | |||
9309 | ||||
9310 | // Whenever we can lower this as a zext, that instruction is strictly faster | |||
9311 | // than any alternative. | |||
9312 | if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend( | |||
9313 | DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG)) | |||
9314 | return ZExt; | |||
9315 | ||||
9316 | auto isV1 = [](int M) { return M >= 0 && M < 8; }; | |||
9317 | auto isV2 = [](int M) { return M >= 8; }; | |||
9318 | ||||
9319 | int NumV1Inputs = std::count_if(Mask.begin(), Mask.end(), isV1); | |||
9320 | int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2); | |||
9321 | ||||
9322 | if (NumV2Inputs == 0) | |||
9323 | return lowerV8I16SingleInputVectorShuffle(DL, V1, Mask, Subtarget, DAG); | |||
9324 | ||||
9325 | assert(NumV1Inputs > 0 && "All single-input shuffles should be canonicalized "((NumV1Inputs > 0 && "All single-input shuffles should be canonicalized " "to be V1-input shuffles.") ? static_cast<void> (0) : __assert_fail ("NumV1Inputs > 0 && \"All single-input shuffles should be canonicalized \" \"to be V1-input shuffles.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9326, __PRETTY_FUNCTION__)) | |||
9326 | "to be V1-input shuffles.")((NumV1Inputs > 0 && "All single-input shuffles should be canonicalized " "to be V1-input shuffles.") ? static_cast<void> (0) : __assert_fail ("NumV1Inputs > 0 && \"All single-input shuffles should be canonicalized \" \"to be V1-input shuffles.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9326, __PRETTY_FUNCTION__)); | |||
9327 | ||||
9328 | // Try to use byte shift instructions. | |||
9329 | if (SDValue Shift = lowerVectorShuffleAsByteShift( | |||
9330 | DL, MVT::v8i16, V1, V2, Mask, DAG)) | |||
9331 | return Shift; | |||
9332 | ||||
9333 | // There are special ways we can lower some single-element blends. | |||
9334 | if (NumV2Inputs == 1) | |||
9335 | if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v8i16, DL, V1, V2, | |||
9336 | Mask, Subtarget, DAG)) | |||
9337 | return V; | |||
9338 | ||||
9339 | // Use dedicated unpack instructions for masks that match their pattern. | |||
9340 | if (isShuffleEquivalent(Mask, 0, 8, 1, 9, 2, 10, 3, 11)) | |||
9341 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2); | |||
9342 | if (isShuffleEquivalent(Mask, 4, 12, 5, 13, 6, 14, 7, 15)) | |||
9343 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2); | |||
9344 | ||||
9345 | if (Subtarget->hasSSE41()) | |||
9346 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask, | |||
9347 | Subtarget, DAG)) | |||
9348 | return Blend; | |||
9349 | ||||
9350 | // Try to use byte rotation instructions. | |||
9351 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate( | |||
9352 | DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG)) | |||
9353 | return Rotate; | |||
9354 | ||||
9355 | if (NumV1Inputs + NumV2Inputs <= 4) | |||
9356 | return lowerV8I16BasicBlendVectorShuffle(DL, V1, V2, Mask, Subtarget, DAG); | |||
9357 | ||||
9358 | // Check whether an interleaving lowering is likely to be more efficient. | |||
9359 | // This isn't perfect but it is a strong heuristic that tends to work well on | |||
9360 | // the kinds of shuffles that show up in practice. | |||
9361 | // | |||
9362 | // FIXME: Handle 1x, 2x, and 4x interleaving. | |||
9363 | if (shouldLowerAsInterleaving(Mask)) { | |||
9364 | // FIXME: Figure out whether we should pack these into the low or high | |||
9365 | // halves. | |||
9366 | ||||
9367 | int EMask[8], OMask[8]; | |||
9368 | for (int i = 0; i < 4; ++i) { | |||
9369 | EMask[i] = Mask[2*i]; | |||
9370 | OMask[i] = Mask[2*i + 1]; | |||
9371 | EMask[i + 4] = -1; | |||
9372 | OMask[i + 4] = -1; | |||
9373 | } | |||
9374 | ||||
9375 | SDValue Evens = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, EMask); | |||
9376 | SDValue Odds = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, OMask); | |||
9377 | ||||
9378 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, Evens, Odds); | |||
9379 | } | |||
9380 | ||||
9381 | int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1}; | |||
9382 | int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1}; | |||
9383 | ||||
9384 | for (int i = 0; i < 4; ++i) { | |||
9385 | LoBlendMask[i] = Mask[i]; | |||
9386 | HiBlendMask[i] = Mask[i + 4]; | |||
9387 | } | |||
9388 | ||||
9389 | SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, LoBlendMask); | |||
9390 | SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, HiBlendMask); | |||
9391 | LoV = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, LoV); | |||
9392 | HiV = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, HiV); | |||
9393 | ||||
9394 | return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, | |||
9395 | DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, LoV, HiV)); | |||
9396 | } | |||
9397 | ||||
9398 | /// \brief Check whether a compaction lowering can be done by dropping even | |||
9399 | /// elements and compute how many times even elements must be dropped. | |||
9400 | /// | |||
9401 | /// This handles shuffles which take every Nth element where N is a power of | |||
9402 | /// two. Example shuffle masks: | |||
9403 | /// | |||
9404 | /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14 | |||
9405 | /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30 | |||
9406 | /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12 | |||
9407 | /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28 | |||
9408 | /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8 | |||
9409 | /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24 | |||
9410 | /// | |||
9411 | /// Any of these lanes can of course be undef. | |||
9412 | /// | |||
9413 | /// This routine only supports N <= 3. | |||
9414 | /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here | |||
9415 | /// for larger N. | |||
9416 | /// | |||
9417 | /// \returns N above, or the number of times even elements must be dropped if | |||
9418 | /// there is such a number. Otherwise returns zero. | |||
9419 | static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) { | |||
9420 | // Figure out whether we're looping over two inputs or just one. | |||
9421 | bool IsSingleInput = isSingleInputShuffleMask(Mask); | |||
9422 | ||||
9423 | // The modulus for the shuffle vector entries is based on whether this is | |||
9424 | // a single input or not. | |||
9425 | int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2); | |||
9426 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9427, __PRETTY_FUNCTION__)) | |||
9427 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9427, __PRETTY_FUNCTION__)); | |||
9428 | ||||
9429 | uint64_t ModMask = (uint64_t)ShuffleModulus - 1; | |||
9430 | ||||
9431 | // We track whether the input is viable for all power-of-2 strides 2^1, 2^2, | |||
9432 | // and 2^3 simultaneously. This is because we may have ambiguity with | |||
9433 | // partially undef inputs. | |||
9434 | bool ViableForN[3] = {true, true, true}; | |||
9435 | ||||
9436 | for (int i = 0, e = Mask.size(); i < e; ++i) { | |||
9437 | // Ignore undef lanes, we'll optimistically collapse them to the pattern we | |||
9438 | // want. | |||
9439 | if (Mask[i] == -1) | |||
9440 | continue; | |||
9441 | ||||
9442 | bool IsAnyViable = false; | |||
9443 | for (unsigned j = 0; j != array_lengthof(ViableForN); ++j) | |||
9444 | if (ViableForN[j]) { | |||
9445 | uint64_t N = j + 1; | |||
9446 | ||||
9447 | // The shuffle mask must be equal to (i * 2^N) % M. | |||
9448 | if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask)) | |||
9449 | IsAnyViable = true; | |||
9450 | else | |||
9451 | ViableForN[j] = false; | |||
9452 | } | |||
9453 | // Early exit if we exhaust the possible powers of two. | |||
9454 | if (!IsAnyViable) | |||
9455 | break; | |||
9456 | } | |||
9457 | ||||
9458 | for (unsigned j = 0; j != array_lengthof(ViableForN); ++j) | |||
9459 | if (ViableForN[j]) | |||
9460 | return j + 1; | |||
9461 | ||||
9462 | // Return 0 as there is no viable power of two. | |||
9463 | return 0; | |||
9464 | } | |||
9465 | ||||
9466 | /// \brief Generic lowering of v16i8 shuffles. | |||
9467 | /// | |||
9468 | /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to | |||
9469 | /// detect any complexity reducing interleaving. If that doesn't help, it uses | |||
9470 | /// UNPCK to spread the i8 elements across two i16-element vectors, and uses | |||
9471 | /// the existing lowering for v8i16 blends on each half, finally PACK-ing them | |||
9472 | /// back together. | |||
9473 | static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
9474 | const X86Subtarget *Subtarget, | |||
9475 | SelectionDAG &DAG) { | |||
9476 | SDLoc DL(Op); | |||
9477 | assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!")((Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!" ) ? static_cast<void> (0) : __assert_fail ("Op.getSimpleValueType() == MVT::v16i8 && \"Bad shuffle type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9477, __PRETTY_FUNCTION__)); | |||
9478 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9478, __PRETTY_FUNCTION__)); | |||
9479 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9479, __PRETTY_FUNCTION__)); | |||
9480 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
9481 | ArrayRef<int> OrigMask = SVOp->getMask(); | |||
9482 | assert(OrigMask.size() == 16 && "Unexpected mask size for v16 shuffle!")((OrigMask.size() == 16 && "Unexpected mask size for v16 shuffle!" ) ? static_cast<void> (0) : __assert_fail ("OrigMask.size() == 16 && \"Unexpected mask size for v16 shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9482, __PRETTY_FUNCTION__)); | |||
9483 | ||||
9484 | // Try to use byte shift instructions. | |||
9485 | if (SDValue Shift = lowerVectorShuffleAsByteShift( | |||
9486 | DL, MVT::v16i8, V1, V2, OrigMask, DAG)) | |||
9487 | return Shift; | |||
9488 | ||||
9489 | // Try to use byte rotation instructions. | |||
9490 | if (SDValue Rotate = lowerVectorShuffleAsByteRotate( | |||
9491 | DL, MVT::v16i8, V1, V2, OrigMask, Subtarget, DAG)) | |||
9492 | return Rotate; | |||
9493 | ||||
9494 | // Try to use a zext lowering. | |||
9495 | if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend( | |||
9496 | DL, MVT::v16i8, V1, V2, OrigMask, Subtarget, DAG)) | |||
9497 | return ZExt; | |||
9498 | ||||
9499 | int MaskStorage[16] = { | |||
9500 | OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3], | |||
9501 | OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7], | |||
9502 | OrigMask[8], OrigMask[9], OrigMask[10], OrigMask[11], | |||
9503 | OrigMask[12], OrigMask[13], OrigMask[14], OrigMask[15]}; | |||
9504 | MutableArrayRef<int> Mask(MaskStorage); | |||
9505 | MutableArrayRef<int> LoMask = Mask.slice(0, 8); | |||
9506 | MutableArrayRef<int> HiMask = Mask.slice(8, 8); | |||
9507 | ||||
9508 | int NumV2Elements = | |||
9509 | std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; }); | |||
9510 | ||||
9511 | // For single-input shuffles, there are some nicer lowering tricks we can use. | |||
9512 | if (NumV2Elements == 0) { | |||
9513 | // Check for being able to broadcast a single element. | |||
9514 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v16i8, DL, V1, | |||
9515 | Mask, Subtarget, DAG)) | |||
9516 | return Broadcast; | |||
9517 | ||||
9518 | // Check whether we can widen this to an i16 shuffle by duplicating bytes. | |||
9519 | // Notably, this handles splat and partial-splat shuffles more efficiently. | |||
9520 | // However, it only makes sense if the pre-duplication shuffle simplifies | |||
9521 | // things significantly. Currently, this means we need to be able to | |||
9522 | // express the pre-duplication shuffle as an i16 shuffle. | |||
9523 | // | |||
9524 | // FIXME: We should check for other patterns which can be widened into an | |||
9525 | // i16 shuffle as well. | |||
9526 | auto canWidenViaDuplication = [](ArrayRef<int> Mask) { | |||
9527 | for (int i = 0; i < 16; i += 2) | |||
9528 | if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1]) | |||
9529 | return false; | |||
9530 | ||||
9531 | return true; | |||
9532 | }; | |||
9533 | auto tryToWidenViaDuplication = [&]() -> SDValue { | |||
9534 | if (!canWidenViaDuplication(Mask)) | |||
9535 | return SDValue(); | |||
9536 | SmallVector<int, 4> LoInputs; | |||
9537 | std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs), | |||
9538 | [](int M) { return M >= 0 && M < 8; }); | |||
9539 | std::sort(LoInputs.begin(), LoInputs.end()); | |||
9540 | LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), | |||
9541 | LoInputs.end()); | |||
9542 | SmallVector<int, 4> HiInputs; | |||
9543 | std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs), | |||
9544 | [](int M) { return M >= 8; }); | |||
9545 | std::sort(HiInputs.begin(), HiInputs.end()); | |||
9546 | HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), | |||
9547 | HiInputs.end()); | |||
9548 | ||||
9549 | bool TargetLo = LoInputs.size() >= HiInputs.size(); | |||
9550 | ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs; | |||
9551 | ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs; | |||
9552 | ||||
9553 | int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1}; | |||
9554 | SmallDenseMap<int, int, 8> LaneMap; | |||
9555 | for (int I : InPlaceInputs) { | |||
9556 | PreDupI16Shuffle[I/2] = I/2; | |||
9557 | LaneMap[I] = I; | |||
9558 | } | |||
9559 | int j = TargetLo ? 0 : 4, je = j + 4; | |||
9560 | for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) { | |||
9561 | // Check if j is already a shuffle of this input. This happens when | |||
9562 | // there are two adjacent bytes after we move the low one. | |||
9563 | if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) { | |||
9564 | // If we haven't yet mapped the input, search for a slot into which | |||
9565 | // we can map it. | |||
9566 | while (j < je && PreDupI16Shuffle[j] != -1) | |||
9567 | ++j; | |||
9568 | ||||
9569 | if (j == je) | |||
9570 | // We can't place the inputs into a single half with a simple i16 shuffle, so bail. | |||
9571 | return SDValue(); | |||
9572 | ||||
9573 | // Map this input with the i16 shuffle. | |||
9574 | PreDupI16Shuffle[j] = MovingInputs[i] / 2; | |||
9575 | } | |||
9576 | ||||
9577 | // Update the lane map based on the mapping we ended up with. | |||
9578 | LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2; | |||
9579 | } | |||
9580 | V1 = DAG.getNode( | |||
9581 | ISD::BITCAST, DL, MVT::v16i8, | |||
9582 | DAG.getVectorShuffle(MVT::v8i16, DL, | |||
9583 | DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1), | |||
9584 | DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle)); | |||
9585 | ||||
9586 | // Unpack the bytes to form the i16s that will be shuffled into place. | |||
9587 | V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL, | |||
9588 | MVT::v16i8, V1, V1); | |||
9589 | ||||
9590 | int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1}; | |||
9591 | for (int i = 0; i < 16; ++i) | |||
9592 | if (Mask[i] != -1) { | |||
9593 | int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8); | |||
9594 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9594, __PRETTY_FUNCTION__)); | |||
9595 | if (PostDupI16Shuffle[i / 2] == -1) | |||
9596 | PostDupI16Shuffle[i / 2] = MappedMask; | |||
9597 | else | |||
9598 | assert(PostDupI16Shuffle[i / 2] == MappedMask &&((PostDupI16Shuffle[i / 2] == MappedMask && "Conflicting entrties in the original shuffle!" ) ? static_cast<void> (0) : __assert_fail ("PostDupI16Shuffle[i / 2] == MappedMask && \"Conflicting entrties in the original shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9599, __PRETTY_FUNCTION__)) | |||
9599 | "Conflicting entrties in the original shuffle!")((PostDupI16Shuffle[i / 2] == MappedMask && "Conflicting entrties in the original shuffle!" ) ? static_cast<void> (0) : __assert_fail ("PostDupI16Shuffle[i / 2] == MappedMask && \"Conflicting entrties in the original shuffle!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9599, __PRETTY_FUNCTION__)); | |||
9600 | } | |||
9601 | return DAG.getNode( | |||
9602 | ISD::BITCAST, DL, MVT::v16i8, | |||
9603 | DAG.getVectorShuffle(MVT::v8i16, DL, | |||
9604 | DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1), | |||
9605 | DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle)); | |||
9606 | }; | |||
9607 | if (SDValue V = tryToWidenViaDuplication()) | |||
9608 | return V; | |||
9609 | } | |||
9610 | ||||
9611 | // Check whether an interleaving lowering is likely to be more efficient. | |||
9612 | // This isn't perfect but it is a strong heuristic that tends to work well on | |||
9613 | // the kinds of shuffles that show up in practice. | |||
9614 | // | |||
9615 | // FIXME: We need to handle other interleaving widths (i16, i32, ...). | |||
9616 | if (shouldLowerAsInterleaving(Mask)) { | |||
9617 | int NumLoHalf = std::count_if(Mask.begin(), Mask.end(), [](int M) { | |||
9618 | return (M >= 0 && M < 8) || (M >= 16 && M < 24); | |||
9619 | }); | |||
9620 | int NumHiHalf = std::count_if(Mask.begin(), Mask.end(), [](int M) { | |||
9621 | return (M >= 8 && M < 16) || M >= 24; | |||
9622 | }); | |||
9623 | int EMask[16] = {-1, -1, -1, -1, -1, -1, -1, -1, | |||
9624 | -1, -1, -1, -1, -1, -1, -1, -1}; | |||
9625 | int OMask[16] = {-1, -1, -1, -1, -1, -1, -1, -1, | |||
9626 | -1, -1, -1, -1, -1, -1, -1, -1}; | |||
9627 | bool UnpackLo = NumLoHalf >= NumHiHalf; | |||
9628 | MutableArrayRef<int> TargetEMask(UnpackLo ? EMask : EMask + 8, 8); | |||
9629 | MutableArrayRef<int> TargetOMask(UnpackLo ? OMask : OMask + 8, 8); | |||
9630 | for (int i = 0; i < 8; ++i) { | |||
9631 | TargetEMask[i] = Mask[2 * i]; | |||
9632 | TargetOMask[i] = Mask[2 * i + 1]; | |||
9633 | } | |||
9634 | ||||
9635 | SDValue Evens = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2, EMask); | |||
9636 | SDValue Odds = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2, OMask); | |||
9637 | ||||
9638 | return DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL, | |||
9639 | MVT::v16i8, Evens, Odds); | |||
9640 | } | |||
9641 | ||||
9642 | // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly | |||
9643 | // with PSHUFB. It is important to do this before we attempt to generate any | |||
9644 | // blends but after all of the single-input lowerings. If the single input | |||
9645 | // lowerings can find an instruction sequence that is faster than a PSHUFB, we | |||
9646 | // want to preserve that and we can DAG combine any longer sequences into | |||
9647 | // a PSHUFB in the end. But once we start blending from multiple inputs, | |||
9648 | // the complexity of DAG combining bad patterns back into PSHUFB is too high, | |||
9649 | // and there are *very* few patterns that would actually be faster than the | |||
9650 | // PSHUFB approach because of its ability to zero lanes. | |||
9651 | // | |||
9652 | // FIXME: The only exceptions to the above are blends which are exact | |||
9653 | // interleavings with direct instructions supporting them. We currently don't | |||
9654 | // handle those well here. | |||
9655 | if (Subtarget->hasSSSE3()) { | |||
9656 | SDValue V1Mask[16]; | |||
9657 | SDValue V2Mask[16]; | |||
9658 | bool V1InUse = false; | |||
9659 | bool V2InUse = false; | |||
9660 | SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2); | |||
9661 | ||||
9662 | for (int i = 0; i < 16; ++i) { | |||
9663 | if (Mask[i] == -1) { | |||
9664 | V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8); | |||
9665 | } else { | |||
9666 | const int ZeroMask = 0x80; | |||
9667 | int V1Idx = (Mask[i] < 16 ? Mask[i] : ZeroMask); | |||
9668 | int V2Idx = (Mask[i] < 16 ? ZeroMask : Mask[i] - 16); | |||
9669 | if (Zeroable[i]) | |||
9670 | V1Idx = V2Idx = ZeroMask; | |||
9671 | V1Mask[i] = DAG.getConstant(V1Idx, MVT::i8); | |||
9672 | V2Mask[i] = DAG.getConstant(V2Idx, MVT::i8); | |||
9673 | V1InUse |= (ZeroMask != V1Idx); | |||
9674 | V2InUse |= (ZeroMask != V2Idx); | |||
9675 | } | |||
9676 | } | |||
9677 | ||||
9678 | if (V1InUse) | |||
9679 | V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, V1, | |||
9680 | DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask)); | |||
9681 | if (V2InUse) | |||
9682 | V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, V2, | |||
9683 | DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask)); | |||
9684 | ||||
9685 | // If we need shuffled inputs from both, blend the two. | |||
9686 | if (V1InUse && V2InUse) | |||
9687 | return DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2); | |||
9688 | if (V1InUse) | |||
9689 | return V1; // Single inputs are easy. | |||
9690 | if (V2InUse) | |||
9691 | return V2; // Single inputs are easy. | |||
9692 | // Shuffling to a zeroable vector. | |||
9693 | return getZeroVector(MVT::v16i8, Subtarget, DAG, DL); | |||
9694 | } | |||
9695 | ||||
9696 | // There are special ways we can lower some single-element blends. | |||
9697 | if (NumV2Elements == 1) | |||
9698 | if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v16i8, DL, V1, V2, | |||
9699 | Mask, Subtarget, DAG)) | |||
9700 | return V; | |||
9701 | ||||
9702 | // Check whether a compaction lowering can be done. This handles shuffles | |||
9703 | // which take every Nth element for some even N. See the helper function for | |||
9704 | // details. | |||
9705 | // | |||
9706 | // We special case these as they can be particularly efficiently handled with | |||
9707 | // the PACKUSB instruction on x86 and they show up in common patterns of | |||
9708 | // rearranging bytes to truncate wide elements. | |||
9709 | if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) { | |||
9710 | // NumEvenDrops is the power of two stride of the elements. Another way of | |||
9711 | // thinking about it is that we need to drop the even elements this many | |||
9712 | // times to get the original input. | |||
9713 | bool IsSingleInput = isSingleInputShuffleMask(Mask); | |||
9714 | ||||
9715 | // First we need to zero all the dropped bytes. | |||
9716 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9717, __PRETTY_FUNCTION__)) | |||
9717 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9717, __PRETTY_FUNCTION__)); | |||
9718 | // We use the mask type to pick which bytes are preserved based on how many | |||
9719 | // elements are dropped. | |||
9720 | MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 }; | |||
9721 | SDValue ByteClearMask = | |||
9722 | DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, | |||
9723 | DAG.getConstant(0xFF, MaskVTs[NumEvenDrops - 1])); | |||
9724 | V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask); | |||
9725 | if (!IsSingleInput) | |||
9726 | V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask); | |||
9727 | ||||
9728 | // Now pack things back together. | |||
9729 | V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1); | |||
9730 | V2 = IsSingleInput ? V1 : DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2); | |||
9731 | SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2); | |||
9732 | for (int i = 1; i < NumEvenDrops; ++i) { | |||
9733 | Result = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, Result); | |||
9734 | Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result); | |||
9735 | } | |||
9736 | ||||
9737 | return Result; | |||
9738 | } | |||
9739 | ||||
9740 | int V1LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1}; | |||
9741 | int V1HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1}; | |||
9742 | int V2LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1}; | |||
9743 | int V2HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1}; | |||
9744 | ||||
9745 | auto buildBlendMasks = [](MutableArrayRef<int> HalfMask, | |||
9746 | MutableArrayRef<int> V1HalfBlendMask, | |||
9747 | MutableArrayRef<int> V2HalfBlendMask) { | |||
9748 | for (int i = 0; i < 8; ++i) | |||
9749 | if (HalfMask[i] >= 0 && HalfMask[i] < 16) { | |||
9750 | V1HalfBlendMask[i] = HalfMask[i]; | |||
9751 | HalfMask[i] = i; | |||
9752 | } else if (HalfMask[i] >= 16) { | |||
9753 | V2HalfBlendMask[i] = HalfMask[i] - 16; | |||
9754 | HalfMask[i] = i + 8; | |||
9755 | } | |||
9756 | }; | |||
9757 | buildBlendMasks(LoMask, V1LoBlendMask, V2LoBlendMask); | |||
9758 | buildBlendMasks(HiMask, V1HiBlendMask, V2HiBlendMask); | |||
9759 | ||||
9760 | SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL); | |||
9761 | ||||
9762 | auto buildLoAndHiV8s = [&](SDValue V, MutableArrayRef<int> LoBlendMask, | |||
9763 | MutableArrayRef<int> HiBlendMask) { | |||
9764 | SDValue V1, V2; | |||
9765 | // Check if any of the odd lanes in the v16i8 are used. If not, we can mask | |||
9766 | // them out and avoid using UNPCK{L,H} to extract the elements of V as | |||
9767 | // i16s. | |||
9768 | if (std::none_of(LoBlendMask.begin(), LoBlendMask.end(), | |||
9769 | [](int M) { return M >= 0 && M % 2 == 1; }) && | |||
9770 | std::none_of(HiBlendMask.begin(), HiBlendMask.end(), | |||
9771 | [](int M) { return M >= 0 && M % 2 == 1; })) { | |||
9772 | // Use a mask to drop the high bytes. | |||
9773 | V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V); | |||
9774 | V1 = DAG.getNode(ISD::AND, DL, MVT::v8i16, V1, | |||
9775 | DAG.getConstant(0x00FF, MVT::v8i16)); | |||
9776 | ||||
9777 | // This will be a single vector shuffle instead of a blend so nuke V2. | |||
9778 | V2 = DAG.getUNDEF(MVT::v8i16); | |||
9779 | ||||
9780 | // Squash the masks to point directly into V1. | |||
9781 | for (int &M : LoBlendMask) | |||
9782 | if (M >= 0) | |||
9783 | M /= 2; | |||
9784 | for (int &M : HiBlendMask) | |||
9785 | if (M >= 0) | |||
9786 | M /= 2; | |||
9787 | } else { | |||
9788 | // Otherwise just unpack the low half of V into V1 and the high half into | |||
9789 | // V2 so that we can blend them as i16s. | |||
9790 | V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, | |||
9791 | DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero)); | |||
9792 | V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, | |||
9793 | DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero)); | |||
9794 | } | |||
9795 | ||||
9796 | SDValue BlendedLo = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, LoBlendMask); | |||
9797 | SDValue BlendedHi = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, HiBlendMask); | |||
9798 | return std::make_pair(BlendedLo, BlendedHi); | |||
9799 | }; | |||
9800 | SDValue V1Lo, V1Hi, V2Lo, V2Hi; | |||
9801 | std::tie(V1Lo, V1Hi) = buildLoAndHiV8s(V1, V1LoBlendMask, V1HiBlendMask); | |||
9802 | std::tie(V2Lo, V2Hi) = buildLoAndHiV8s(V2, V2LoBlendMask, V2HiBlendMask); | |||
9803 | ||||
9804 | SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Lo, V2Lo, LoMask); | |||
9805 | SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Hi, V2Hi, HiMask); | |||
9806 | ||||
9807 | return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV); | |||
9808 | } | |||
9809 | ||||
9810 | /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles. | |||
9811 | /// | |||
9812 | /// This routine breaks down the specific type of 128-bit shuffle and | |||
9813 | /// dispatches to the lowering routines accordingly. | |||
9814 | static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
9815 | MVT VT, const X86Subtarget *Subtarget, | |||
9816 | SelectionDAG &DAG) { | |||
9817 | switch (VT.SimpleTy) { | |||
9818 | case MVT::v2i64: | |||
9819 | return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
9820 | case MVT::v2f64: | |||
9821 | return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
9822 | case MVT::v4i32: | |||
9823 | return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
9824 | case MVT::v4f32: | |||
9825 | return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
9826 | case MVT::v8i16: | |||
9827 | return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
9828 | case MVT::v16i8: | |||
9829 | return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
9830 | ||||
9831 | default: | |||
9832 | llvm_unreachable("Unimplemented!")::llvm::llvm_unreachable_internal("Unimplemented!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9832); | |||
9833 | } | |||
9834 | } | |||
9835 | ||||
9836 | /// \brief Helper function to test whether a shuffle mask could be | |||
9837 | /// simplified by widening the elements being shuffled. | |||
9838 | /// | |||
9839 | /// Appends the mask for wider elements in WidenedMask if valid. Otherwise | |||
9840 | /// leaves it in an unspecified state. | |||
9841 | /// | |||
9842 | /// NOTE: This must handle normal vector shuffle masks and *target* vector | |||
9843 | /// shuffle masks. The latter have the special property of a '-2' representing | |||
9844 | /// a zero-ed lane of a vector. | |||
9845 | static bool canWidenShuffleElements(ArrayRef<int> Mask, | |||
9846 | SmallVectorImpl<int> &WidenedMask) { | |||
9847 | for (int i = 0, Size = Mask.size(); i < Size; i += 2) { | |||
9848 | // If both elements are undef, its trivial. | |||
9849 | if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) { | |||
9850 | WidenedMask.push_back(SM_SentinelUndef); | |||
9851 | continue; | |||
9852 | } | |||
9853 | ||||
9854 | // Check for an undef mask and a mask value properly aligned to fit with | |||
9855 | // a pair of values. If we find such a case, use the non-undef mask's value. | |||
9856 | if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) { | |||
9857 | WidenedMask.push_back(Mask[i + 1] / 2); | |||
9858 | continue; | |||
9859 | } | |||
9860 | if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) { | |||
9861 | WidenedMask.push_back(Mask[i] / 2); | |||
9862 | continue; | |||
9863 | } | |||
9864 | ||||
9865 | // When zeroing, we need to spread the zeroing across both lanes to widen. | |||
9866 | if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) { | |||
9867 | if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) && | |||
9868 | (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) { | |||
9869 | WidenedMask.push_back(SM_SentinelZero); | |||
9870 | continue; | |||
9871 | } | |||
9872 | return false; | |||
9873 | } | |||
9874 | ||||
9875 | // Finally check if the two mask values are adjacent and aligned with | |||
9876 | // a pair. | |||
9877 | if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) { | |||
9878 | WidenedMask.push_back(Mask[i] / 2); | |||
9879 | continue; | |||
9880 | } | |||
9881 | ||||
9882 | // Otherwise we can't safely widen the elements used in this shuffle. | |||
9883 | return false; | |||
9884 | } | |||
9885 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9886, __PRETTY_FUNCTION__)) | |||
9886 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9886, __PRETTY_FUNCTION__)); | |||
9887 | ||||
9888 | return true; | |||
9889 | } | |||
9890 | ||||
9891 | /// \brief Generic routine to split ector shuffle into half-sized shuffles. | |||
9892 | /// | |||
9893 | /// This routine just extracts two subvectors, shuffles them independently, and | |||
9894 | /// then concatenates them back together. This should work effectively with all | |||
9895 | /// AVX vector shuffle types. | |||
9896 | static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1, | |||
9897 | SDValue V2, ArrayRef<int> Mask, | |||
9898 | SelectionDAG &DAG) { | |||
9899 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9900, __PRETTY_FUNCTION__)) | |||
9900 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9900, __PRETTY_FUNCTION__)); | |||
9901 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9901, __PRETTY_FUNCTION__)); | |||
9902 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 9902, __PRETTY_FUNCTION__)); | |||
9903 | ||||
9904 | ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2); | |||
9905 | ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2); | |||
9906 | ||||
9907 | int NumElements = VT.getVectorNumElements(); | |||
9908 | int SplitNumElements = NumElements / 2; | |||
9909 | MVT ScalarVT = VT.getScalarType(); | |||
9910 | MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2); | |||
9911 | ||||
9912 | SDValue LoV1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V1, | |||
9913 | DAG.getIntPtrConstant(0)); | |||
9914 | SDValue HiV1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V1, | |||
9915 | DAG.getIntPtrConstant(SplitNumElements)); | |||
9916 | SDValue LoV2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V2, | |||
9917 | DAG.getIntPtrConstant(0)); | |||
9918 | SDValue HiV2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V2, | |||
9919 | DAG.getIntPtrConstant(SplitNumElements)); | |||
9920 | ||||
9921 | // Now create two 4-way blends of these half-width vectors. | |||
9922 | auto HalfBlend = [&](ArrayRef<int> HalfMask) { | |||
9923 | bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false; | |||
9924 | SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask; | |||
9925 | for (int i = 0; i < SplitNumElements; ++i) { | |||
9926 | int M = HalfMask[i]; | |||
9927 | if (M >= NumElements) { | |||
9928 | if (M >= NumElements + SplitNumElements) | |||
9929 | UseHiV2 = true; | |||
9930 | else | |||
9931 | UseLoV2 = true; | |||
9932 | V2BlendMask.push_back(M - NumElements); | |||
9933 | V1BlendMask.push_back(-1); | |||
9934 | BlendMask.push_back(SplitNumElements + i); | |||
9935 | } else if (M >= 0) { | |||
9936 | if (M >= SplitNumElements) | |||
9937 | UseHiV1 = true; | |||
9938 | else | |||
9939 | UseLoV1 = true; | |||
9940 | V2BlendMask.push_back(-1); | |||
9941 | V1BlendMask.push_back(M); | |||
9942 | BlendMask.push_back(i); | |||
9943 | } else { | |||
9944 | V2BlendMask.push_back(-1); | |||
9945 | V1BlendMask.push_back(-1); | |||
9946 | BlendMask.push_back(-1); | |||
9947 | } | |||
9948 | } | |||
9949 | ||||
9950 | // Because the lowering happens after all combining takes place, we need to | |||
9951 | // manually combine these blend masks as much as possible so that we create | |||
9952 | // a minimal number of high-level vector shuffle nodes. | |||
9953 | ||||
9954 | // First try just blending the halves of V1 or V2. | |||
9955 | if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2) | |||
9956 | return DAG.getUNDEF(SplitVT); | |||
9957 | if (!UseLoV2 && !UseHiV2) | |||
9958 | return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask); | |||
9959 | if (!UseLoV1 && !UseHiV1) | |||
9960 | return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask); | |||
9961 | ||||
9962 | SDValue V1Blend, V2Blend; | |||
9963 | if (UseLoV1 && UseHiV1) { | |||
9964 | V1Blend = | |||
9965 | DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask); | |||
9966 | } else { | |||
9967 | // We only use half of V1 so map the usage down into the final blend mask. | |||
9968 | V1Blend = UseLoV1 ? LoV1 : HiV1; | |||
9969 | for (int i = 0; i < SplitNumElements; ++i) | |||
9970 | if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements) | |||
9971 | BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements); | |||
9972 | } | |||
9973 | if (UseLoV2 && UseHiV2) { | |||
9974 | V2Blend = | |||
9975 | DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask); | |||
9976 | } else { | |||
9977 | // We only use half of V2 so map the usage down into the final blend mask. | |||
9978 | V2Blend = UseLoV2 ? LoV2 : HiV2; | |||
9979 | for (int i = 0; i < SplitNumElements; ++i) | |||
9980 | if (BlendMask[i] >= SplitNumElements) | |||
9981 | BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0); | |||
9982 | } | |||
9983 | return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask); | |||
9984 | }; | |||
9985 | SDValue Lo = HalfBlend(LoMask); | |||
9986 | SDValue Hi = HalfBlend(HiMask); | |||
9987 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi); | |||
9988 | } | |||
9989 | ||||
9990 | /// \brief Either split a vector in halves or decompose the shuffles and the | |||
9991 | /// blend. | |||
9992 | /// | |||
9993 | /// This is provided as a good fallback for many lowerings of non-single-input | |||
9994 | /// shuffles with more than one 128-bit lane. In those cases, we want to select | |||
9995 | /// between splitting the shuffle into 128-bit components and stitching those | |||
9996 | /// back together vs. extracting the single-input shuffles and blending those | |||
9997 | /// results. | |||
9998 | static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1, | |||
9999 | SDValue V2, ArrayRef<int> Mask, | |||
10000 | SelectionDAG &DAG) { | |||
10001 | assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "((!isSingleInputShuffleMask(Mask) && "This routine must not be used to " "lower single-input shuffles as it " "could then recurse on itself." ) ? static_cast<void> (0) : __assert_fail ("!isSingleInputShuffleMask(Mask) && \"This routine must not be used to \" \"lower single-input shuffles as it \" \"could then recurse on itself.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10003, __PRETTY_FUNCTION__)) | |||
10002 | "lower single-input shuffles as it "((!isSingleInputShuffleMask(Mask) && "This routine must not be used to " "lower single-input shuffles as it " "could then recurse on itself." ) ? static_cast<void> (0) : __assert_fail ("!isSingleInputShuffleMask(Mask) && \"This routine must not be used to \" \"lower single-input shuffles as it \" \"could then recurse on itself.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10003, __PRETTY_FUNCTION__)) | |||
10003 | "could then recurse on itself.")((!isSingleInputShuffleMask(Mask) && "This routine must not be used to " "lower single-input shuffles as it " "could then recurse on itself." ) ? static_cast<void> (0) : __assert_fail ("!isSingleInputShuffleMask(Mask) && \"This routine must not be used to \" \"lower single-input shuffles as it \" \"could then recurse on itself.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10003, __PRETTY_FUNCTION__)); | |||
10004 | int Size = Mask.size(); | |||
10005 | ||||
10006 | // If this can be modeled as a broadcast of two elements followed by a blend, | |||
10007 | // prefer that lowering. This is especially important because broadcasts can | |||
10008 | // often fold with memory operands. | |||
10009 | auto DoBothBroadcast = [&] { | |||
10010 | int V1BroadcastIdx = -1, V2BroadcastIdx = -1; | |||
10011 | for (int M : Mask) | |||
10012 | if (M >= Size) { | |||
10013 | if (V2BroadcastIdx == -1) | |||
10014 | V2BroadcastIdx = M - Size; | |||
10015 | else if (M - Size != V2BroadcastIdx) | |||
10016 | return false; | |||
10017 | } else if (M >= 0) { | |||
10018 | if (V1BroadcastIdx == -1) | |||
10019 | V1BroadcastIdx = M; | |||
10020 | else if (M != V1BroadcastIdx) | |||
10021 | return false; | |||
10022 | } | |||
10023 | return true; | |||
10024 | }; | |||
10025 | if (DoBothBroadcast()) | |||
10026 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, | |||
10027 | DAG); | |||
10028 | ||||
10029 | // If the inputs all stem from a single 128-bit lane of each input, then we | |||
10030 | // split them rather than blending because the split will decompose to | |||
10031 | // unusually few instructions. | |||
10032 | int LaneCount = VT.getSizeInBits() / 128; | |||
10033 | int LaneSize = Size / LaneCount; | |||
10034 | SmallBitVector LaneInputs[2]; | |||
10035 | LaneInputs[0].resize(LaneCount, false); | |||
10036 | LaneInputs[1].resize(LaneCount, false); | |||
10037 | for (int i = 0; i < Size; ++i) | |||
10038 | if (Mask[i] >= 0) | |||
10039 | LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true; | |||
10040 | if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1) | |||
10041 | return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG); | |||
10042 | ||||
10043 | // Otherwise, just fall back to decomposed shuffles and a blend. This requires | |||
10044 | // that the decomposed single-input shuffles don't end up here. | |||
10045 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG); | |||
10046 | } | |||
10047 | ||||
10048 | /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as | |||
10049 | /// a permutation and blend of those lanes. | |||
10050 | /// | |||
10051 | /// This essentially blends the out-of-lane inputs to each lane into the lane | |||
10052 | /// from a permuted copy of the vector. This lowering strategy results in four | |||
10053 | /// instructions in the worst case for a single-input cross lane shuffle which | |||
10054 | /// is lower than any other fully general cross-lane shuffle strategy I'm aware | |||
10055 | /// of. Special cases for each particular shuffle pattern should be handled | |||
10056 | /// prior to trying this lowering. | |||
10057 | static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT, | |||
10058 | SDValue V1, SDValue V2, | |||
10059 | ArrayRef<int> Mask, | |||
10060 | SelectionDAG &DAG) { | |||
10061 | // FIXME: This should probably be generalized for 512-bit vectors as well. | |||
10062 | assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!")((VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!" ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() == 256 && \"Only for 256-bit vector shuffles!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10062, __PRETTY_FUNCTION__)); | |||
10063 | int LaneSize = Mask.size() / 2; | |||
10064 | ||||
10065 | // If there are only inputs from one 128-bit lane, splitting will in fact be | |||
10066 | // less expensive. The flags track wether the given lane contains an element | |||
10067 | // that crosses to another lane. | |||
10068 | bool LaneCrossing[2] = {false, false}; | |||
10069 | for (int i = 0, Size = Mask.size(); i < Size; ++i) | |||
10070 | if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize) | |||
10071 | LaneCrossing[(Mask[i] % Size) / LaneSize] = true; | |||
10072 | if (!LaneCrossing[0] || !LaneCrossing[1]) | |||
10073 | return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG); | |||
10074 | ||||
10075 | if (isSingleInputShuffleMask(Mask)) { | |||
10076 | SmallVector<int, 32> FlippedBlendMask; | |||
10077 | for (int i = 0, Size = Mask.size(); i < Size; ++i) | |||
10078 | FlippedBlendMask.push_back( | |||
10079 | Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize) | |||
10080 | ? Mask[i] | |||
10081 | : Mask[i] % LaneSize + | |||
10082 | (i / LaneSize) * LaneSize + Size)); | |||
10083 | ||||
10084 | // Flip the vector, and blend the results which should now be in-lane. The | |||
10085 | // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and | |||
10086 | // 5 for the high source. The value 3 selects the high half of source 2 and | |||
10087 | // the value 2 selects the low half of source 2. We only use source 2 to | |||
10088 | // allow folding it into a memory operand. | |||
10089 | unsigned PERMMask = 3 | 2 << 4; | |||
10090 | SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT), | |||
10091 | V1, DAG.getConstant(PERMMask, MVT::i8)); | |||
10092 | return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask); | |||
10093 | } | |||
10094 | ||||
10095 | // This now reduces to two single-input shuffles of V1 and V2 which at worst | |||
10096 | // will be handled by the above logic and a blend of the results, much like | |||
10097 | // other patterns in AVX. | |||
10098 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG); | |||
10099 | } | |||
10100 | ||||
10101 | /// \brief Handle lowering 2-lane 128-bit shuffles. | |||
10102 | static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1, | |||
10103 | SDValue V2, ArrayRef<int> Mask, | |||
10104 | const X86Subtarget *Subtarget, | |||
10105 | SelectionDAG &DAG) { | |||
10106 | // Blends are faster and handle all the non-lane-crossing cases. | |||
10107 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask, | |||
10108 | Subtarget, DAG)) | |||
10109 | return Blend; | |||
10110 | ||||
10111 | MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(), | |||
10112 | VT.getVectorNumElements() / 2); | |||
10113 | // Check for patterns which can be matched with a single insert of a 128-bit | |||
10114 | // subvector. | |||
10115 | if (isShuffleEquivalent(Mask, 0, 1, 0, 1) || | |||
10116 | isShuffleEquivalent(Mask, 0, 1, 4, 5)) { | |||
10117 | SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1, | |||
10118 | DAG.getIntPtrConstant(0)); | |||
10119 | SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, | |||
10120 | Mask[2] < 4 ? V1 : V2, DAG.getIntPtrConstant(0)); | |||
10121 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV); | |||
10122 | } | |||
10123 | if (isShuffleEquivalent(Mask, 0, 1, 6, 7)) { | |||
10124 | SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1, | |||
10125 | DAG.getIntPtrConstant(0)); | |||
10126 | SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V2, | |||
10127 | DAG.getIntPtrConstant(2)); | |||
10128 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV); | |||
10129 | } | |||
10130 | ||||
10131 | // Otherwise form a 128-bit permutation. | |||
10132 | // FIXME: Detect zero-vector inputs and use the VPERM2X128 to zero that half. | |||
10133 | unsigned PermMask = Mask[0] / 2 | (Mask[2] / 2) << 4; | |||
10134 | return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2, | |||
10135 | DAG.getConstant(PermMask, MVT::i8)); | |||
10136 | } | |||
10137 | ||||
10138 | /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then | |||
10139 | /// shuffling each lane. | |||
10140 | /// | |||
10141 | /// This will only succeed when the result of fixing the 128-bit lanes results | |||
10142 | /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in | |||
10143 | /// each 128-bit lanes. This handles many cases where we can quickly blend away | |||
10144 | /// the lane crosses early and then use simpler shuffles within each lane. | |||
10145 | /// | |||
10146 | /// FIXME: It might be worthwhile at some point to support this without | |||
10147 | /// requiring the 128-bit lane-relative shuffles to be repeating, but currently | |||
10148 | /// in x86 only floating point has interesting non-repeating shuffles, and even | |||
10149 | /// those are still *marginally* more expensive. | |||
10150 | static SDValue lowerVectorShuffleByMerging128BitLanes( | |||
10151 | SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, | |||
10152 | const X86Subtarget *Subtarget, SelectionDAG &DAG) { | |||
10153 | assert(!isSingleInputShuffleMask(Mask) &&((!isSingleInputShuffleMask(Mask) && "This is only useful with multiple inputs." ) ? static_cast<void> (0) : __assert_fail ("!isSingleInputShuffleMask(Mask) && \"This is only useful with multiple inputs.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10154, __PRETTY_FUNCTION__)) | |||
10154 | "This is only useful with multiple inputs.")((!isSingleInputShuffleMask(Mask) && "This is only useful with multiple inputs." ) ? static_cast<void> (0) : __assert_fail ("!isSingleInputShuffleMask(Mask) && \"This is only useful with multiple inputs.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10154, __PRETTY_FUNCTION__)); | |||
10155 | ||||
10156 | int Size = Mask.size(); | |||
10157 | int LaneSize = 128 / VT.getScalarSizeInBits(); | |||
10158 | int NumLanes = Size / LaneSize; | |||
10159 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10159, __PRETTY_FUNCTION__)); | |||
10160 | ||||
10161 | // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also | |||
10162 | // check whether the in-128-bit lane shuffles share a repeating pattern. | |||
10163 | SmallVector<int, 4> Lanes; | |||
10164 | Lanes.resize(NumLanes, -1); | |||
10165 | SmallVector<int, 4> InLaneMask; | |||
10166 | InLaneMask.resize(LaneSize, -1); | |||
10167 | for (int i = 0; i < Size; ++i) { | |||
10168 | if (Mask[i] < 0) | |||
10169 | continue; | |||
10170 | ||||
10171 | int j = i / LaneSize; | |||
10172 | ||||
10173 | if (Lanes[j] < 0) { | |||
10174 | // First entry we've seen for this lane. | |||
10175 | Lanes[j] = Mask[i] / LaneSize; | |||
10176 | } else if (Lanes[j] != Mask[i] / LaneSize) { | |||
10177 | // This doesn't match the lane selected previously! | |||
10178 | return SDValue(); | |||
10179 | } | |||
10180 | ||||
10181 | // Check that within each lane we have a consistent shuffle mask. | |||
10182 | int k = i % LaneSize; | |||
10183 | if (InLaneMask[k] < 0) { | |||
10184 | InLaneMask[k] = Mask[i] % LaneSize; | |||
10185 | } else if (InLaneMask[k] != Mask[i] % LaneSize) { | |||
10186 | // This doesn't fit a repeating in-lane mask. | |||
10187 | return SDValue(); | |||
10188 | } | |||
10189 | } | |||
10190 | ||||
10191 | // First shuffle the lanes into place. | |||
10192 | MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64, | |||
10193 | VT.getSizeInBits() / 64); | |||
10194 | SmallVector<int, 8> LaneMask; | |||
10195 | LaneMask.resize(NumLanes * 2, -1); | |||
10196 | for (int i = 0; i < NumLanes; ++i) | |||
10197 | if (Lanes[i] >= 0) { | |||
10198 | LaneMask[2 * i + 0] = 2*Lanes[i] + 0; | |||
10199 | LaneMask[2 * i + 1] = 2*Lanes[i] + 1; | |||
10200 | } | |||
10201 | ||||
10202 | V1 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V1); | |||
10203 | V2 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V2); | |||
10204 | SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask); | |||
10205 | ||||
10206 | // Cast it back to the type we actually want. | |||
10207 | LaneShuffle = DAG.getNode(ISD::BITCAST, DL, VT, LaneShuffle); | |||
10208 | ||||
10209 | // Now do a simple shuffle that isn't lane crossing. | |||
10210 | SmallVector<int, 8> NewMask; | |||
10211 | NewMask.resize(Size, -1); | |||
10212 | for (int i = 0; i < Size; ++i) | |||
10213 | if (Mask[i] >= 0) | |||
10214 | NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize; | |||
10215 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10216, __PRETTY_FUNCTION__)) | |||
10216 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10216, __PRETTY_FUNCTION__)); | |||
10217 | ||||
10218 | return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask); | |||
10219 | } | |||
10220 | ||||
10221 | /// \brief Test whether the specified input (0 or 1) is in-place blended by the | |||
10222 | /// given mask. | |||
10223 | /// | |||
10224 | /// This returns true if the elements from a particular input are already in the | |||
10225 | /// slot required by the given mask and require no permutation. | |||
10226 | static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) { | |||
10227 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10227, __PRETTY_FUNCTION__)); | |||
10228 | int Size = Mask.size(); | |||
10229 | for (int i = 0; i < Size; ++i) | |||
10230 | if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i) | |||
10231 | return false; | |||
10232 | ||||
10233 | return true; | |||
10234 | } | |||
10235 | ||||
10236 | /// \brief Handle lowering of 4-lane 64-bit floating point shuffles. | |||
10237 | /// | |||
10238 | /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2 | |||
10239 | /// isn't available. | |||
10240 | static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
10241 | const X86Subtarget *Subtarget, | |||
10242 | SelectionDAG &DAG) { | |||
10243 | SDLoc DL(Op); | |||
10244 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10244, __PRETTY_FUNCTION__)); | |||
10245 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10245, __PRETTY_FUNCTION__)); | |||
10246 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
10247 | ArrayRef<int> Mask = SVOp->getMask(); | |||
10248 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10248, __PRETTY_FUNCTION__)); | |||
10249 | ||||
10250 | SmallVector<int, 4> WidenedMask; | |||
10251 | if (canWidenShuffleElements(Mask, WidenedMask)) | |||
10252 | return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget, | |||
10253 | DAG); | |||
10254 | ||||
10255 | if (isSingleInputShuffleMask(Mask)) { | |||
10256 | // Check for being able to broadcast a single element. | |||
10257 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4f64, DL, V1, | |||
10258 | Mask, Subtarget, DAG)) | |||
10259 | return Broadcast; | |||
10260 | ||||
10261 | // Use low duplicate instructions for masks that match their pattern. | |||
10262 | if (isShuffleEquivalent(Mask, 0, 0, 2, 2)) | |||
10263 | return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1); | |||
10264 | ||||
10265 | if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) { | |||
10266 | // Non-half-crossing single input shuffles can be lowerid with an | |||
10267 | // interleaved permutation. | |||
10268 | unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) | | |||
10269 | ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3); | |||
10270 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1, | |||
10271 | DAG.getConstant(VPERMILPMask, MVT::i8)); | |||
10272 | } | |||
10273 | ||||
10274 | // With AVX2 we have direct support for this permutation. | |||
10275 | if (Subtarget->hasAVX2()) | |||
10276 | return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1, | |||
10277 | getV4X86ShuffleImm8ForMask(Mask, DAG)); | |||
10278 | ||||
10279 | // Otherwise, fall back. | |||
10280 | return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask, | |||
10281 | DAG); | |||
10282 | } | |||
10283 | ||||
10284 | // X86 has dedicated unpack instructions that can handle specific blend | |||
10285 | // operations: UNPCKH and UNPCKL. | |||
10286 | if (isShuffleEquivalent(Mask, 0, 4, 2, 6)) | |||
10287 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2); | |||
10288 | if (isShuffleEquivalent(Mask, 1, 5, 3, 7)) | |||
10289 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2); | |||
10290 | ||||
10291 | // If we have a single input to the zero element, insert that into V1 if we | |||
10292 | // can do so cheaply. | |||
10293 | int NumV2Elements = | |||
10294 | std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }); | |||
10295 | if (NumV2Elements == 1 && Mask[0] >= 4) | |||
10296 | if (SDValue Insertion = lowerVectorShuffleAsElementInsertion( | |||
10297 | MVT::v4f64, DL, V1, V2, Mask, Subtarget, DAG)) | |||
10298 | return Insertion; | |||
10299 | ||||
10300 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask, | |||
10301 | Subtarget, DAG)) | |||
10302 | return Blend; | |||
10303 | ||||
10304 | // Check if the blend happens to exactly fit that of SHUFPD. | |||
10305 | if ((Mask[0] == -1 || Mask[0] < 2) && | |||
10306 | (Mask[1] == -1 || (Mask[1] >= 4 && Mask[1] < 6)) && | |||
10307 | (Mask[2] == -1 || (Mask[2] >= 2 && Mask[2] < 4)) && | |||
10308 | (Mask[3] == -1 || Mask[3] >= 6)) { | |||
10309 | unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 5) << 1) | | |||
10310 | ((Mask[2] == 3) << 2) | ((Mask[3] == 7) << 3); | |||
10311 | return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V1, V2, | |||
10312 | DAG.getConstant(SHUFPDMask, MVT::i8)); | |||
10313 | } | |||
10314 | if ((Mask[0] == -1 || (Mask[0] >= 4 && Mask[0] < 6)) && | |||
10315 | (Mask[1] == -1 || Mask[1] < 2) && | |||
10316 | (Mask[2] == -1 || Mask[2] >= 6) && | |||
10317 | (Mask[3] == -1 || (Mask[3] >= 2 && Mask[3] < 4))) { | |||
10318 | unsigned SHUFPDMask = (Mask[0] == 5) | ((Mask[1] == 1) << 1) | | |||
10319 | ((Mask[2] == 7) << 2) | ((Mask[3] == 3) << 3); | |||
10320 | return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V2, V1, | |||
10321 | DAG.getConstant(SHUFPDMask, MVT::i8)); | |||
10322 | } | |||
10323 | ||||
10324 | // Try to simplify this by merging 128-bit lanes to enable a lane-based | |||
10325 | // shuffle. However, if we have AVX2 and either inputs are already in place, | |||
10326 | // we will be able to shuffle even across lanes the other input in a single | |||
10327 | // instruction so skip this pattern. | |||
10328 | if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) || | |||
10329 | isShuffleMaskInputInPlace(1, Mask)))) | |||
10330 | if (SDValue Result = lowerVectorShuffleByMerging128BitLanes( | |||
10331 | DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG)) | |||
10332 | return Result; | |||
10333 | ||||
10334 | // If we have AVX2 then we always want to lower with a blend because an v4 we | |||
10335 | // can fully permute the elements. | |||
10336 | if (Subtarget->hasAVX2()) | |||
10337 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2, | |||
10338 | Mask, DAG); | |||
10339 | ||||
10340 | // Otherwise fall back on generic lowering. | |||
10341 | return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG); | |||
10342 | } | |||
10343 | ||||
10344 | /// \brief Handle lowering of 4-lane 64-bit integer shuffles. | |||
10345 | /// | |||
10346 | /// This routine is only called when we have AVX2 and thus a reasonable | |||
10347 | /// instruction set for v4i64 shuffling.. | |||
10348 | static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
10349 | const X86Subtarget *Subtarget, | |||
10350 | SelectionDAG &DAG) { | |||
10351 | SDLoc DL(Op); | |||
10352 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10352, __PRETTY_FUNCTION__)); | |||
10353 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10353, __PRETTY_FUNCTION__)); | |||
10354 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
10355 | ArrayRef<int> Mask = SVOp->getMask(); | |||
10356 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10356, __PRETTY_FUNCTION__)); | |||
10357 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10357, __PRETTY_FUNCTION__)); | |||
10358 | ||||
10359 | SmallVector<int, 4> WidenedMask; | |||
10360 | if (canWidenShuffleElements(Mask, WidenedMask)) | |||
10361 | return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget, | |||
10362 | DAG); | |||
10363 | ||||
10364 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask, | |||
10365 | Subtarget, DAG)) | |||
10366 | return Blend; | |||
10367 | ||||
10368 | // Check for being able to broadcast a single element. | |||
10369 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4i64, DL, V1, | |||
10370 | Mask, Subtarget, DAG)) | |||
10371 | return Broadcast; | |||
10372 | ||||
10373 | // When the shuffle is mirrored between the 128-bit lanes of the unit, we can | |||
10374 | // use lower latency instructions that will operate on both 128-bit lanes. | |||
10375 | SmallVector<int, 2> RepeatedMask; | |||
10376 | if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) { | |||
10377 | if (isSingleInputShuffleMask(Mask)) { | |||
10378 | int PSHUFDMask[] = {-1, -1, -1, -1}; | |||
10379 | for (int i = 0; i < 2; ++i) | |||
10380 | if (RepeatedMask[i] >= 0) { | |||
10381 | PSHUFDMask[2 * i] = 2 * RepeatedMask[i]; | |||
10382 | PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1; | |||
10383 | } | |||
10384 | return DAG.getNode( | |||
10385 | ISD::BITCAST, DL, MVT::v4i64, | |||
10386 | DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, | |||
10387 | DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, V1), | |||
10388 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG))); | |||
10389 | } | |||
10390 | ||||
10391 | // Use dedicated unpack instructions for masks that match their pattern. | |||
10392 | if (isShuffleEquivalent(Mask, 0, 4, 2, 6)) | |||
10393 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2); | |||
10394 | if (isShuffleEquivalent(Mask, 1, 5, 3, 7)) | |||
10395 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2); | |||
10396 | } | |||
10397 | ||||
10398 | // AVX2 provides a direct instruction for permuting a single input across | |||
10399 | // lanes. | |||
10400 | if (isSingleInputShuffleMask(Mask)) | |||
10401 | return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1, | |||
10402 | getV4X86ShuffleImm8ForMask(Mask, DAG)); | |||
10403 | ||||
10404 | // Try to simplify this by merging 128-bit lanes to enable a lane-based | |||
10405 | // shuffle. However, if we have AVX2 and either inputs are already in place, | |||
10406 | // we will be able to shuffle even across lanes the other input in a single | |||
10407 | // instruction so skip this pattern. | |||
10408 | if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) || | |||
10409 | isShuffleMaskInputInPlace(1, Mask)))) | |||
10410 | if (SDValue Result = lowerVectorShuffleByMerging128BitLanes( | |||
10411 | DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG)) | |||
10412 | return Result; | |||
10413 | ||||
10414 | // Otherwise fall back on generic blend lowering. | |||
10415 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2, | |||
10416 | Mask, DAG); | |||
10417 | } | |||
10418 | ||||
10419 | /// \brief Handle lowering of 8-lane 32-bit floating point shuffles. | |||
10420 | /// | |||
10421 | /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2 | |||
10422 | /// isn't available. | |||
10423 | static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
10424 | const X86Subtarget *Subtarget, | |||
10425 | SelectionDAG &DAG) { | |||
10426 | SDLoc DL(Op); | |||
10427 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10427, __PRETTY_FUNCTION__)); | |||
10428 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10428, __PRETTY_FUNCTION__)); | |||
10429 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
10430 | ArrayRef<int> Mask = SVOp->getMask(); | |||
10431 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10431, __PRETTY_FUNCTION__)); | |||
10432 | ||||
10433 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask, | |||
10434 | Subtarget, DAG)) | |||
10435 | return Blend; | |||
10436 | ||||
10437 | // Check for being able to broadcast a single element. | |||
10438 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8f32, DL, V1, | |||
10439 | Mask, Subtarget, DAG)) | |||
10440 | return Broadcast; | |||
10441 | ||||
10442 | // If the shuffle mask is repeated in each 128-bit lane, we have many more | |||
10443 | // options to efficiently lower the shuffle. | |||
10444 | SmallVector<int, 4> RepeatedMask; | |||
10445 | if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) { | |||
10446 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10447, __PRETTY_FUNCTION__)) | |||
10447 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10447, __PRETTY_FUNCTION__)); | |||
10448 | ||||
10449 | // Use even/odd duplicate instructions for masks that match their pattern. | |||
10450 | if (isShuffleEquivalent(Mask, 0, 0, 2, 2, 4, 4, 6, 6)) | |||
10451 | return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1); | |||
10452 | if (isShuffleEquivalent(Mask, 1, 1, 3, 3, 5, 5, 7, 7)) | |||
10453 | return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1); | |||
10454 | ||||
10455 | if (isSingleInputShuffleMask(Mask)) | |||
10456 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1, | |||
10457 | getV4X86ShuffleImm8ForMask(RepeatedMask, DAG)); | |||
10458 | ||||
10459 | // Use dedicated unpack instructions for masks that match their pattern. | |||
10460 | if (isShuffleEquivalent(Mask, 0, 8, 1, 9, 4, 12, 5, 13)) | |||
10461 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2); | |||
10462 | if (isShuffleEquivalent(Mask, 2, 10, 3, 11, 6, 14, 7, 15)) | |||
10463 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2); | |||
10464 | ||||
10465 | // Otherwise, fall back to a SHUFPS sequence. Here it is important that we | |||
10466 | // have already handled any direct blends. We also need to squash the | |||
10467 | // repeated mask into a simulated v4f32 mask. | |||
10468 | for (int i = 0; i < 4; ++i) | |||
10469 | if (RepeatedMask[i] >= 8) | |||
10470 | RepeatedMask[i] -= 4; | |||
10471 | return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG); | |||
10472 | } | |||
10473 | ||||
10474 | // If we have a single input shuffle with different shuffle patterns in the | |||
10475 | // two 128-bit lanes use the variable mask to VPERMILPS. | |||
10476 | if (isSingleInputShuffleMask(Mask)) { | |||
10477 | SDValue VPermMask[8]; | |||
10478 | for (int i = 0; i < 8; ++i) | |||
10479 | VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32) | |||
10480 | : DAG.getConstant(Mask[i], MVT::i32); | |||
10481 | if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask)) | |||
10482 | return DAG.getNode( | |||
10483 | X86ISD::VPERMILPV, DL, MVT::v8f32, V1, | |||
10484 | DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask)); | |||
10485 | ||||
10486 | if (Subtarget->hasAVX2()) | |||
10487 | return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32, | |||
10488 | DAG.getNode(ISD::BITCAST, DL, MVT::v8f32, | |||
10489 | DAG.getNode(ISD::BUILD_VECTOR, DL, | |||
10490 | MVT::v8i32, VPermMask)), | |||
10491 | V1); | |||
10492 | ||||
10493 | // Otherwise, fall back. | |||
10494 | return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask, | |||
10495 | DAG); | |||
10496 | } | |||
10497 | ||||
10498 | // Try to simplify this by merging 128-bit lanes to enable a lane-based | |||
10499 | // shuffle. | |||
10500 | if (SDValue Result = lowerVectorShuffleByMerging128BitLanes( | |||
10501 | DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG)) | |||
10502 | return Result; | |||
10503 | ||||
10504 | // If we have AVX2 then we always want to lower with a blend because at v8 we | |||
10505 | // can fully permute the elements. | |||
10506 | if (Subtarget->hasAVX2()) | |||
10507 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2, | |||
10508 | Mask, DAG); | |||
10509 | ||||
10510 | // Otherwise fall back on generic lowering. | |||
10511 | return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG); | |||
10512 | } | |||
10513 | ||||
10514 | /// \brief Handle lowering of 8-lane 32-bit integer shuffles. | |||
10515 | /// | |||
10516 | /// This routine is only called when we have AVX2 and thus a reasonable | |||
10517 | /// instruction set for v8i32 shuffling.. | |||
10518 | static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
10519 | const X86Subtarget *Subtarget, | |||
10520 | SelectionDAG &DAG) { | |||
10521 | SDLoc DL(Op); | |||
10522 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10522, __PRETTY_FUNCTION__)); | |||
10523 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10523, __PRETTY_FUNCTION__)); | |||
10524 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
10525 | ArrayRef<int> Mask = SVOp->getMask(); | |||
10526 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10526, __PRETTY_FUNCTION__)); | |||
10527 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10527, __PRETTY_FUNCTION__)); | |||
10528 | ||||
10529 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask, | |||
10530 | Subtarget, DAG)) | |||
10531 | return Blend; | |||
10532 | ||||
10533 | // Check for being able to broadcast a single element. | |||
10534 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8i32, DL, V1, | |||
10535 | Mask, Subtarget, DAG)) | |||
10536 | return Broadcast; | |||
10537 | ||||
10538 | // If the shuffle mask is repeated in each 128-bit lane we can use more | |||
10539 | // efficient instructions that mirror the shuffles across the two 128-bit | |||
10540 | // lanes. | |||
10541 | SmallVector<int, 4> RepeatedMask; | |||
10542 | if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) { | |||
10543 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10543, __PRETTY_FUNCTION__)); | |||
10544 | if (isSingleInputShuffleMask(Mask)) | |||
10545 | return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1, | |||
10546 | getV4X86ShuffleImm8ForMask(RepeatedMask, DAG)); | |||
10547 | ||||
10548 | // Use dedicated unpack instructions for masks that match their pattern. | |||
10549 | if (isShuffleEquivalent(Mask, 0, 8, 1, 9, 4, 12, 5, 13)) | |||
10550 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2); | |||
10551 | if (isShuffleEquivalent(Mask, 2, 10, 3, 11, 6, 14, 7, 15)) | |||
10552 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2); | |||
10553 | } | |||
10554 | ||||
10555 | // If the shuffle patterns aren't repeated but it is a single input, directly | |||
10556 | // generate a cross-lane VPERMD instruction. | |||
10557 | if (isSingleInputShuffleMask(Mask)) { | |||
10558 | SDValue VPermMask[8]; | |||
10559 | for (int i = 0; i < 8; ++i) | |||
10560 | VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32) | |||
10561 | : DAG.getConstant(Mask[i], MVT::i32); | |||
10562 | return DAG.getNode( | |||
10563 | X86ISD::VPERMV, DL, MVT::v8i32, | |||
10564 | DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1); | |||
10565 | } | |||
10566 | ||||
10567 | // Try to simplify this by merging 128-bit lanes to enable a lane-based | |||
10568 | // shuffle. | |||
10569 | if (SDValue Result = lowerVectorShuffleByMerging128BitLanes( | |||
10570 | DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG)) | |||
10571 | return Result; | |||
10572 | ||||
10573 | // Otherwise fall back on generic blend lowering. | |||
10574 | return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2, | |||
10575 | Mask, DAG); | |||
10576 | } | |||
10577 | ||||
10578 | /// \brief Handle lowering of 16-lane 16-bit integer shuffles. | |||
10579 | /// | |||
10580 | /// This routine is only called when we have AVX2 and thus a reasonable | |||
10581 | /// instruction set for v16i16 shuffling.. | |||
10582 | static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
10583 | const X86Subtarget *Subtarget, | |||
10584 | SelectionDAG &DAG) { | |||
10585 | SDLoc DL(Op); | |||
10586 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10586, __PRETTY_FUNCTION__)); | |||
10587 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10587, __PRETTY_FUNCTION__)); | |||
10588 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
10589 | ArrayRef<int> Mask = SVOp->getMask(); | |||
10590 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10590, __PRETTY_FUNCTION__)); | |||
10591 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10591, __PRETTY_FUNCTION__)); | |||
10592 | ||||
10593 | // Check for being able to broadcast a single element. | |||
10594 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v16i16, DL, V1, | |||
10595 | Mask, Subtarget, DAG)) | |||
10596 | return Broadcast; | |||
10597 | ||||
10598 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask, | |||
10599 | Subtarget, DAG)) | |||
10600 | return Blend; | |||
10601 | ||||
10602 | // Use dedicated unpack instructions for masks that match their pattern. | |||
10603 | if (isShuffleEquivalent(Mask, | |||
10604 | // First 128-bit lane: | |||
10605 | 0, 16, 1, 17, 2, 18, 3, 19, | |||
10606 | // Second 128-bit lane: | |||
10607 | 8, 24, 9, 25, 10, 26, 11, 27)) | |||
10608 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2); | |||
10609 | if (isShuffleEquivalent(Mask, | |||
10610 | // First 128-bit lane: | |||
10611 | 4, 20, 5, 21, 6, 22, 7, 23, | |||
10612 | // Second 128-bit lane: | |||
10613 | 12, 28, 13, 29, 14, 30, 15, 31)) | |||
10614 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2); | |||
10615 | ||||
10616 | if (isSingleInputShuffleMask(Mask)) { | |||
10617 | // There are no generalized cross-lane shuffle operations available on i16 | |||
10618 | // element types. | |||
10619 | if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask)) | |||
10620 | return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2, | |||
10621 | Mask, DAG); | |||
10622 | ||||
10623 | SDValue PSHUFBMask[32]; | |||
10624 | for (int i = 0; i < 16; ++i) { | |||
10625 | if (Mask[i] == -1) { | |||
10626 | PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8); | |||
10627 | continue; | |||
10628 | } | |||
10629 | ||||
10630 | int M = i < 8 ? Mask[i] : Mask[i] - 8; | |||
10631 | assert(M >= 0 && M < 8 && "Invalid single-input mask!")((M >= 0 && M < 8 && "Invalid single-input mask!" ) ? static_cast<void> (0) : __assert_fail ("M >= 0 && M < 8 && \"Invalid single-input mask!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10631, __PRETTY_FUNCTION__)); | |||
10632 | PSHUFBMask[2 * i] = DAG.getConstant(2 * M, MVT::i8); | |||
10633 | PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, MVT::i8); | |||
10634 | } | |||
10635 | return DAG.getNode( | |||
10636 | ISD::BITCAST, DL, MVT::v16i16, | |||
10637 | DAG.getNode( | |||
10638 | X86ISD::PSHUFB, DL, MVT::v32i8, | |||
10639 | DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1), | |||
10640 | DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask))); | |||
10641 | } | |||
10642 | ||||
10643 | // Try to simplify this by merging 128-bit lanes to enable a lane-based | |||
10644 | // shuffle. | |||
10645 | if (SDValue Result = lowerVectorShuffleByMerging128BitLanes( | |||
10646 | DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG)) | |||
10647 | return Result; | |||
10648 | ||||
10649 | // Otherwise fall back on generic lowering. | |||
10650 | return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG); | |||
10651 | } | |||
10652 | ||||
10653 | /// \brief Handle lowering of 32-lane 8-bit integer shuffles. | |||
10654 | /// | |||
10655 | /// This routine is only called when we have AVX2 and thus a reasonable | |||
10656 | /// instruction set for v32i8 shuffling.. | |||
10657 | static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
10658 | const X86Subtarget *Subtarget, | |||
10659 | SelectionDAG &DAG) { | |||
10660 | SDLoc DL(Op); | |||
10661 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10661, __PRETTY_FUNCTION__)); | |||
10662 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10662, __PRETTY_FUNCTION__)); | |||
10663 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
10664 | ArrayRef<int> Mask = SVOp->getMask(); | |||
10665 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10665, __PRETTY_FUNCTION__)); | |||
10666 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10666, __PRETTY_FUNCTION__)); | |||
10667 | ||||
10668 | // Check for being able to broadcast a single element. | |||
10669 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v32i8, DL, V1, | |||
10670 | Mask, Subtarget, DAG)) | |||
10671 | return Broadcast; | |||
10672 | ||||
10673 | if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask, | |||
10674 | Subtarget, DAG)) | |||
10675 | return Blend; | |||
10676 | ||||
10677 | // Use dedicated unpack instructions for masks that match their pattern. | |||
10678 | // Note that these are repeated 128-bit lane unpacks, not unpacks across all | |||
10679 | // 256-bit lanes. | |||
10680 | if (isShuffleEquivalent( | |||
10681 | Mask, | |||
10682 | // First 128-bit lane: | |||
10683 | 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39, | |||
10684 | // Second 128-bit lane: | |||
10685 | 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55)) | |||
10686 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2); | |||
10687 | if (isShuffleEquivalent( | |||
10688 | Mask, | |||
10689 | // First 128-bit lane: | |||
10690 | 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47, | |||
10691 | // Second 128-bit lane: | |||
10692 | 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63)) | |||
10693 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2); | |||
10694 | ||||
10695 | if (isSingleInputShuffleMask(Mask)) { | |||
10696 | // There are no generalized cross-lane shuffle operations available on i8 | |||
10697 | // element types. | |||
10698 | if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask)) | |||
10699 | return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2, | |||
10700 | Mask, DAG); | |||
10701 | ||||
10702 | SDValue PSHUFBMask[32]; | |||
10703 | for (int i = 0; i < 32; ++i) | |||
10704 | PSHUFBMask[i] = | |||
10705 | Mask[i] < 0 | |||
10706 | ? DAG.getUNDEF(MVT::i8) | |||
10707 | : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, MVT::i8); | |||
10708 | ||||
10709 | return DAG.getNode( | |||
10710 | X86ISD::PSHUFB, DL, MVT::v32i8, V1, | |||
10711 | DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask)); | |||
10712 | } | |||
10713 | ||||
10714 | // Try to simplify this by merging 128-bit lanes to enable a lane-based | |||
10715 | // shuffle. | |||
10716 | if (SDValue Result = lowerVectorShuffleByMerging128BitLanes( | |||
10717 | DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG)) | |||
10718 | return Result; | |||
10719 | ||||
10720 | // Otherwise fall back on generic lowering. | |||
10721 | return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG); | |||
10722 | } | |||
10723 | ||||
10724 | /// \brief High-level routine to lower various 256-bit x86 vector shuffles. | |||
10725 | /// | |||
10726 | /// This routine either breaks down the specific type of a 256-bit x86 vector | |||
10727 | /// shuffle or splits it into two 128-bit shuffles and fuses the results back | |||
10728 | /// together based on the available instructions. | |||
10729 | static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
10730 | MVT VT, const X86Subtarget *Subtarget, | |||
10731 | SelectionDAG &DAG) { | |||
10732 | SDLoc DL(Op); | |||
10733 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
10734 | ArrayRef<int> Mask = SVOp->getMask(); | |||
10735 | ||||
10736 | // There is a really nice hard cut-over between AVX1 and AVX2 that means we can | |||
10737 | // check for those subtargets here and avoid much of the subtarget querying in | |||
10738 | // the per-vector-type lowering routines. With AVX1 we have essentially *zero* | |||
10739 | // ability to manipulate a 256-bit vector with integer types. Since we'll use | |||
10740 | // floating point types there eventually, just immediately cast everything to | |||
10741 | // a float and operate entirely in that domain. | |||
10742 | if (VT.isInteger() && !Subtarget->hasAVX2()) { | |||
10743 | int ElementBits = VT.getScalarSizeInBits(); | |||
10744 | if (ElementBits < 32) | |||
10745 | // No floating point type available, decompose into 128-bit vectors. | |||
10746 | return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG); | |||
10747 | ||||
10748 | MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits), | |||
10749 | VT.getVectorNumElements()); | |||
10750 | V1 = DAG.getNode(ISD::BITCAST, DL, FpVT, V1); | |||
10751 | V2 = DAG.getNode(ISD::BITCAST, DL, FpVT, V2); | |||
10752 | return DAG.getNode(ISD::BITCAST, DL, VT, | |||
10753 | DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask)); | |||
10754 | } | |||
10755 | ||||
10756 | switch (VT.SimpleTy) { | |||
10757 | case MVT::v4f64: | |||
10758 | return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
10759 | case MVT::v4i64: | |||
10760 | return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
10761 | case MVT::v8f32: | |||
10762 | return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
10763 | case MVT::v8i32: | |||
10764 | return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
10765 | case MVT::v16i16: | |||
10766 | return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
10767 | case MVT::v32i8: | |||
10768 | return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
10769 | ||||
10770 | default: | |||
10771 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10771); | |||
10772 | } | |||
10773 | } | |||
10774 | ||||
10775 | /// \brief Handle lowering of 8-lane 64-bit floating point shuffles. | |||
10776 | static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
10777 | const X86Subtarget *Subtarget, | |||
10778 | SelectionDAG &DAG) { | |||
10779 | SDLoc DL(Op); | |||
10780 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10780, __PRETTY_FUNCTION__)); | |||
10781 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10781, __PRETTY_FUNCTION__)); | |||
10782 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
10783 | ArrayRef<int> Mask = SVOp->getMask(); | |||
10784 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10784, __PRETTY_FUNCTION__)); | |||
10785 | ||||
10786 | // X86 has dedicated unpack instructions that can handle specific blend | |||
10787 | // operations: UNPCKH and UNPCKL. | |||
10788 | if (isShuffleEquivalent(Mask, 0, 8, 2, 10, 4, 12, 6, 14)) | |||
10789 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2); | |||
10790 | if (isShuffleEquivalent(Mask, 1, 9, 3, 11, 5, 13, 7, 15)) | |||
10791 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2); | |||
10792 | ||||
10793 | // FIXME: Implement direct support for this type! | |||
10794 | return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG); | |||
10795 | } | |||
10796 | ||||
10797 | /// \brief Handle lowering of 16-lane 32-bit floating point shuffles. | |||
10798 | static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
10799 | const X86Subtarget *Subtarget, | |||
10800 | SelectionDAG &DAG) { | |||
10801 | SDLoc DL(Op); | |||
10802 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10802, __PRETTY_FUNCTION__)); | |||
10803 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10803, __PRETTY_FUNCTION__)); | |||
10804 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
10805 | ArrayRef<int> Mask = SVOp->getMask(); | |||
10806 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10806, __PRETTY_FUNCTION__)); | |||
10807 | ||||
10808 | // Use dedicated unpack instructions for masks that match their pattern. | |||
10809 | if (isShuffleEquivalent(Mask, | |||
10810 | 0, 16, 1, 17, 4, 20, 5, 21, | |||
10811 | 8, 24, 9, 25, 12, 28, 13, 29)) | |||
10812 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2); | |||
10813 | if (isShuffleEquivalent(Mask, | |||
10814 | 2, 18, 3, 19, 6, 22, 7, 23, | |||
10815 | 10, 26, 11, 27, 14, 30, 15, 31)) | |||
10816 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2); | |||
10817 | ||||
10818 | // FIXME: Implement direct support for this type! | |||
10819 | return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG); | |||
10820 | } | |||
10821 | ||||
10822 | /// \brief Handle lowering of 8-lane 64-bit integer shuffles. | |||
10823 | static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
10824 | const X86Subtarget *Subtarget, | |||
10825 | SelectionDAG &DAG) { | |||
10826 | SDLoc DL(Op); | |||
10827 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10827, __PRETTY_FUNCTION__)); | |||
10828 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10828, __PRETTY_FUNCTION__)); | |||
10829 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
10830 | ArrayRef<int> Mask = SVOp->getMask(); | |||
10831 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10831, __PRETTY_FUNCTION__)); | |||
10832 | ||||
10833 | // X86 has dedicated unpack instructions that can handle specific blend | |||
10834 | // operations: UNPCKH and UNPCKL. | |||
10835 | if (isShuffleEquivalent(Mask, 0, 8, 2, 10, 4, 12, 6, 14)) | |||
10836 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2); | |||
10837 | if (isShuffleEquivalent(Mask, 1, 9, 3, 11, 5, 13, 7, 15)) | |||
10838 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2); | |||
10839 | ||||
10840 | // FIXME: Implement direct support for this type! | |||
10841 | return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG); | |||
10842 | } | |||
10843 | ||||
10844 | /// \brief Handle lowering of 16-lane 32-bit integer shuffles. | |||
10845 | static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
10846 | const X86Subtarget *Subtarget, | |||
10847 | SelectionDAG &DAG) { | |||
10848 | SDLoc DL(Op); | |||
10849 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10849, __PRETTY_FUNCTION__)); | |||
10850 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10850, __PRETTY_FUNCTION__)); | |||
10851 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
10852 | ArrayRef<int> Mask = SVOp->getMask(); | |||
10853 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10853, __PRETTY_FUNCTION__)); | |||
10854 | ||||
10855 | // Use dedicated unpack instructions for masks that match their pattern. | |||
10856 | if (isShuffleEquivalent(Mask, | |||
10857 | 0, 16, 1, 17, 4, 20, 5, 21, | |||
10858 | 8, 24, 9, 25, 12, 28, 13, 29)) | |||
10859 | return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2); | |||
10860 | if (isShuffleEquivalent(Mask, | |||
10861 | 2, 18, 3, 19, 6, 22, 7, 23, | |||
10862 | 10, 26, 11, 27, 14, 30, 15, 31)) | |||
10863 | return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2); | |||
10864 | ||||
10865 | // FIXME: Implement direct support for this type! | |||
10866 | return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG); | |||
10867 | } | |||
10868 | ||||
10869 | /// \brief Handle lowering of 32-lane 16-bit integer shuffles. | |||
10870 | static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
10871 | const X86Subtarget *Subtarget, | |||
10872 | SelectionDAG &DAG) { | |||
10873 | SDLoc DL(Op); | |||
10874 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10874, __PRETTY_FUNCTION__)); | |||
10875 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10875, __PRETTY_FUNCTION__)); | |||
10876 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
10877 | ArrayRef<int> Mask = SVOp->getMask(); | |||
10878 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10878, __PRETTY_FUNCTION__)); | |||
10879 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10879, __PRETTY_FUNCTION__)); | |||
10880 | ||||
10881 | // FIXME: Implement direct support for this type! | |||
10882 | return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG); | |||
10883 | } | |||
10884 | ||||
10885 | /// \brief Handle lowering of 64-lane 8-bit integer shuffles. | |||
10886 | static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
10887 | const X86Subtarget *Subtarget, | |||
10888 | SelectionDAG &DAG) { | |||
10889 | SDLoc DL(Op); | |||
10890 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10890, __PRETTY_FUNCTION__)); | |||
10891 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10891, __PRETTY_FUNCTION__)); | |||
10892 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
10893 | ArrayRef<int> Mask = SVOp->getMask(); | |||
10894 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10894, __PRETTY_FUNCTION__)); | |||
10895 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10895, __PRETTY_FUNCTION__)); | |||
10896 | ||||
10897 | // FIXME: Implement direct support for this type! | |||
10898 | return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG); | |||
10899 | } | |||
10900 | ||||
10901 | /// \brief High-level routine to lower various 512-bit x86 vector shuffles. | |||
10902 | /// | |||
10903 | /// This routine either breaks down the specific type of a 512-bit x86 vector | |||
10904 | /// shuffle or splits it into two 256-bit shuffles and fuses the results back | |||
10905 | /// together based on the available instructions. | |||
10906 | static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2, | |||
10907 | MVT VT, const X86Subtarget *Subtarget, | |||
10908 | SelectionDAG &DAG) { | |||
10909 | SDLoc DL(Op); | |||
10910 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
10911 | ArrayRef<int> Mask = SVOp->getMask(); | |||
10912 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10913, __PRETTY_FUNCTION__)) | |||
10913 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10913, __PRETTY_FUNCTION__)); | |||
10914 | ||||
10915 | // Check for being able to broadcast a single element. | |||
10916 | if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(VT.SimpleTy, DL, V1, | |||
10917 | Mask, Subtarget, DAG)) | |||
10918 | return Broadcast; | |||
10919 | ||||
10920 | // Dispatch to each element type for lowering. If we don't have supprot for | |||
10921 | // specific element type shuffles at 512 bits, immediately split them and | |||
10922 | // lower them. Each lowering routine of a given type is allowed to assume that | |||
10923 | // the requisite ISA extensions for that element type are available. | |||
10924 | switch (VT.SimpleTy) { | |||
10925 | case MVT::v8f64: | |||
10926 | return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
10927 | case MVT::v16f32: | |||
10928 | return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
10929 | case MVT::v8i64: | |||
10930 | return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
10931 | case MVT::v16i32: | |||
10932 | return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
10933 | case MVT::v32i16: | |||
10934 | if (Subtarget->hasBWI()) | |||
10935 | return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
10936 | break; | |||
10937 | case MVT::v64i8: | |||
10938 | if (Subtarget->hasBWI()) | |||
10939 | return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG); | |||
10940 | break; | |||
10941 | ||||
10942 | default: | |||
10943 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10943); | |||
10944 | } | |||
10945 | ||||
10946 | // Otherwise fall back on splitting. | |||
10947 | return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG); | |||
10948 | } | |||
10949 | ||||
10950 | /// \brief Top-level lowering for x86 vector shuffles. | |||
10951 | /// | |||
10952 | /// This handles decomposition, canonicalization, and lowering of all x86 | |||
10953 | /// vector shuffles. Most of the specific lowering strategies are encapsulated | |||
10954 | /// above in helper routines. The canonicalization attempts to widen shuffles | |||
10955 | /// to involve fewer lanes of wider elements, consolidate symmetric patterns | |||
10956 | /// s.t. only one of the two inputs needs to be tested, etc. | |||
10957 | static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget, | |||
10958 | SelectionDAG &DAG) { | |||
10959 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
10960 | ArrayRef<int> Mask = SVOp->getMask(); | |||
10961 | SDValue V1 = Op.getOperand(0); | |||
10962 | SDValue V2 = Op.getOperand(1); | |||
10963 | MVT VT = Op.getSimpleValueType(); | |||
10964 | int NumElements = VT.getVectorNumElements(); | |||
10965 | SDLoc dl(Op); | |||
10966 | ||||
10967 | assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles")((VT.getSizeInBits() != 64 && "Can't lower MMX shuffles" ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() != 64 && \"Can't lower MMX shuffles\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 10967, __PRETTY_FUNCTION__)); | |||
10968 | ||||
10969 | bool V1IsUndef = V1.getOpcode() == ISD::UNDEF; | |||
10970 | bool V2IsUndef = V2.getOpcode() == ISD::UNDEF; | |||
10971 | if (V1IsUndef && V2IsUndef) | |||
10972 | return DAG.getUNDEF(VT); | |||
10973 | ||||
10974 | // When we create a shuffle node we put the UNDEF node to second operand, | |||
10975 | // but in some cases the first operand may be transformed to UNDEF. | |||
10976 | // In this case we should just commute the node. | |||
10977 | if (V1IsUndef) | |||
10978 | return DAG.getCommutedVectorShuffle(*SVOp); | |||
10979 | ||||
10980 | // Check for non-undef masks pointing at an undef vector and make the masks | |||
10981 | // undef as well. This makes it easier to match the shuffle based solely on | |||
10982 | // the mask. | |||
10983 | if (V2IsUndef) | |||
10984 | for (int M : Mask) | |||
10985 | if (M >= NumElements) { | |||
10986 | SmallVector<int, 8> NewMask(Mask.begin(), Mask.end()); | |||
10987 | for (int &M : NewMask) | |||
10988 | if (M >= NumElements) | |||
10989 | M = -1; | |||
10990 | return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask); | |||
10991 | } | |||
10992 | ||||
10993 | // Try to collapse shuffles into using a vector type with fewer elements but | |||
10994 | // wider element types. We cap this to not form integers or floating point | |||
10995 | // elements wider than 64 bits, but it might be interesting to form i128 | |||
10996 | // integers to handle flipping the low and high halves of AVX 256-bit vectors. | |||
10997 | SmallVector<int, 16> WidenedMask; | |||
10998 | if (VT.getScalarSizeInBits() < 64 && | |||
10999 | canWidenShuffleElements(Mask, WidenedMask)) { | |||
11000 | MVT NewEltVT = VT.isFloatingPoint() | |||
11001 | ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2) | |||
11002 | : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2); | |||
11003 | MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2); | |||
11004 | // Make sure that the new vector type is legal. For example, v2f64 isn't | |||
11005 | // legal on SSE1. | |||
11006 | if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) { | |||
11007 | V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1); | |||
11008 | V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2); | |||
11009 | return DAG.getNode(ISD::BITCAST, dl, VT, | |||
11010 | DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask)); | |||
11011 | } | |||
11012 | } | |||
11013 | ||||
11014 | int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0; | |||
11015 | for (int M : SVOp->getMask()) | |||
11016 | if (M < 0) | |||
11017 | ++NumUndefElements; | |||
11018 | else if (M < NumElements) | |||
11019 | ++NumV1Elements; | |||
11020 | else | |||
11021 | ++NumV2Elements; | |||
11022 | ||||
11023 | // Commute the shuffle as needed such that more elements come from V1 than | |||
11024 | // V2. This allows us to match the shuffle pattern strictly on how many | |||
11025 | // elements come from V1 without handling the symmetric cases. | |||
11026 | if (NumV2Elements > NumV1Elements) | |||
11027 | return DAG.getCommutedVectorShuffle(*SVOp); | |||
11028 | ||||
11029 | // When the number of V1 and V2 elements are the same, try to minimize the | |||
11030 | // number of uses of V2 in the low half of the vector. When that is tied, | |||
11031 | // ensure that the sum of indices for V1 is equal to or lower than the sum | |||
11032 | // indices for V2. When those are equal, try to ensure that the number of odd | |||
11033 | // indices for V1 is lower than the number of odd indices for V2. | |||
11034 | if (NumV1Elements == NumV2Elements) { | |||
11035 | int LowV1Elements = 0, LowV2Elements = 0; | |||
11036 | for (int M : SVOp->getMask().slice(0, NumElements / 2)) | |||
11037 | if (M >= NumElements) | |||
11038 | ++LowV2Elements; | |||
11039 | else if (M >= 0) | |||
11040 | ++LowV1Elements; | |||
11041 | if (LowV2Elements > LowV1Elements) { | |||
11042 | return DAG.getCommutedVectorShuffle(*SVOp); | |||
11043 | } else if (LowV2Elements == LowV1Elements) { | |||
11044 | int SumV1Indices = 0, SumV2Indices = 0; | |||
11045 | for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i) | |||
11046 | if (SVOp->getMask()[i] >= NumElements) | |||
11047 | SumV2Indices += i; | |||
11048 | else if (SVOp->getMask()[i] >= 0) | |||
11049 | SumV1Indices += i; | |||
11050 | if (SumV2Indices < SumV1Indices) { | |||
11051 | return DAG.getCommutedVectorShuffle(*SVOp); | |||
11052 | } else if (SumV2Indices == SumV1Indices) { | |||
11053 | int NumV1OddIndices = 0, NumV2OddIndices = 0; | |||
11054 | for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i) | |||
11055 | if (SVOp->getMask()[i] >= NumElements) | |||
11056 | NumV2OddIndices += i % 2; | |||
11057 | else if (SVOp->getMask()[i] >= 0) | |||
11058 | NumV1OddIndices += i % 2; | |||
11059 | if (NumV2OddIndices < NumV1OddIndices) | |||
11060 | return DAG.getCommutedVectorShuffle(*SVOp); | |||
11061 | } | |||
11062 | } | |||
11063 | } | |||
11064 | ||||
11065 | // For each vector width, delegate to a specialized lowering routine. | |||
11066 | if (VT.getSizeInBits() == 128) | |||
11067 | return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG); | |||
11068 | ||||
11069 | if (VT.getSizeInBits() == 256) | |||
11070 | return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG); | |||
11071 | ||||
11072 | // Force AVX-512 vectors to be scalarized for now. | |||
11073 | // FIXME: Implement AVX-512 support! | |||
11074 | if (VT.getSizeInBits() == 512) | |||
11075 | return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG); | |||
11076 | ||||
11077 | llvm_unreachable("Unimplemented!")::llvm::llvm_unreachable_internal("Unimplemented!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 11077); | |||
11078 | } | |||
11079 | ||||
11080 | ||||
11081 | //===----------------------------------------------------------------------===// | |||
11082 | // Legacy vector shuffle lowering | |||
11083 | // | |||
11084 | // This code is the legacy code handling vector shuffles until the above | |||
11085 | // replaces its functionality and performance. | |||
11086 | //===----------------------------------------------------------------------===// | |||
11087 | ||||
11088 | static bool isBlendMask(ArrayRef<int> MaskVals, MVT VT, bool hasSSE41, | |||
11089 | bool hasInt256, unsigned *MaskOut = nullptr) { | |||
11090 | MVT EltVT = VT.getVectorElementType(); | |||
11091 | ||||
11092 | // There is no blend with immediate in AVX-512. | |||
11093 | if (VT.is512BitVector()) | |||
11094 | return false; | |||
11095 | ||||
11096 | if (!hasSSE41 || EltVT == MVT::i8) | |||
11097 | return false; | |||
11098 | if (!hasInt256 && VT == MVT::v16i16) | |||
11099 | return false; | |||
11100 | ||||
11101 | unsigned MaskValue = 0; | |||
11102 | unsigned NumElems = VT.getVectorNumElements(); | |||
11103 | // There are 2 lanes if (NumElems > 8), and 1 lane otherwise. | |||
11104 | unsigned NumLanes = (NumElems - 1) / 8 + 1; | |||
11105 | unsigned NumElemsInLane = NumElems / NumLanes; | |||
11106 | ||||
11107 | // Blend for v16i16 should be symetric for the both lanes. | |||
11108 | for (unsigned i = 0; i < NumElemsInLane; ++i) { | |||
11109 | ||||
11110 | int SndLaneEltIdx = (NumLanes == 2) ? MaskVals[i + NumElemsInLane] : -1; | |||
11111 | int EltIdx = MaskVals[i]; | |||
11112 | ||||
11113 | if ((EltIdx < 0 || EltIdx == (int)i) && | |||
11114 | (SndLaneEltIdx < 0 || SndLaneEltIdx == (int)(i + NumElemsInLane))) | |||
11115 | continue; | |||
11116 | ||||
11117 | if (((unsigned)EltIdx == (i + NumElems)) && | |||
11118 | (SndLaneEltIdx < 0 || | |||
11119 | (unsigned)SndLaneEltIdx == i + NumElems + NumElemsInLane)) | |||
11120 | MaskValue |= (1 << i); | |||
11121 | else | |||
11122 | return false; | |||
11123 | } | |||
11124 | ||||
11125 | if (MaskOut) | |||
11126 | *MaskOut = MaskValue; | |||
11127 | return true; | |||
11128 | } | |||
11129 | ||||
11130 | // Try to lower a shuffle node into a simple blend instruction. | |||
11131 | // This function assumes isBlendMask returns true for this | |||
11132 | // SuffleVectorSDNode | |||
11133 | static SDValue LowerVECTOR_SHUFFLEtoBlend(ShuffleVectorSDNode *SVOp, | |||
11134 | unsigned MaskValue, | |||
11135 | const X86Subtarget *Subtarget, | |||
11136 | SelectionDAG &DAG) { | |||
11137 | MVT VT = SVOp->getSimpleValueType(0); | |||
11138 | MVT EltVT = VT.getVectorElementType(); | |||
11139 | assert(isBlendMask(SVOp->getMask(), VT, Subtarget->hasSSE41(),((isBlendMask(SVOp->getMask(), VT, Subtarget->hasSSE41( ), Subtarget->hasInt256() && "Trying to lower a " "VECTOR_SHUFFLE to a Blend but " "with the wrong mask")) ? static_cast<void> (0) : __assert_fail ("isBlendMask(SVOp->getMask(), VT, Subtarget->hasSSE41(), Subtarget->hasInt256() && \"Trying to lower a \" \"VECTOR_SHUFFLE to a Blend but \" \"with the wrong mask\")" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 11142, __PRETTY_FUNCTION__)) | |||
11140 | Subtarget->hasInt256() && "Trying to lower a "((isBlendMask(SVOp->getMask(), VT, Subtarget->hasSSE41( ), Subtarget->hasInt256() && "Trying to lower a " "VECTOR_SHUFFLE to a Blend but " "with the wrong mask")) ? static_cast<void> (0) : __assert_fail ("isBlendMask(SVOp->getMask(), VT, Subtarget->hasSSE41(), Subtarget->hasInt256() && \"Trying to lower a \" \"VECTOR_SHUFFLE to a Blend but \" \"with the wrong mask\")" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 11142, __PRETTY_FUNCTION__)) | |||
11141 | "VECTOR_SHUFFLE to a Blend but "((isBlendMask(SVOp->getMask(), VT, Subtarget->hasSSE41( ), Subtarget->hasInt256() && "Trying to lower a " "VECTOR_SHUFFLE to a Blend but " "with the wrong mask")) ? static_cast<void> (0) : __assert_fail ("isBlendMask(SVOp->getMask(), VT, Subtarget->hasSSE41(), Subtarget->hasInt256() && \"Trying to lower a \" \"VECTOR_SHUFFLE to a Blend but \" \"with the wrong mask\")" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 11142, __PRETTY_FUNCTION__)) | |||
11142 | "with the wrong mask"))((isBlendMask(SVOp->getMask(), VT, Subtarget->hasSSE41( ), Subtarget->hasInt256() && "Trying to lower a " "VECTOR_SHUFFLE to a Blend but " "with the wrong mask")) ? static_cast<void> (0) : __assert_fail ("isBlendMask(SVOp->getMask(), VT, Subtarget->hasSSE41(), Subtarget->hasInt256() && \"Trying to lower a \" \"VECTOR_SHUFFLE to a Blend but \" \"with the wrong mask\")" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 11142, __PRETTY_FUNCTION__)); | |||
11143 | SDValue V1 = SVOp->getOperand(0); | |||
11144 | SDValue V2 = SVOp->getOperand(1); | |||
11145 | SDLoc dl(SVOp); | |||
11146 | unsigned NumElems = VT.getVectorNumElements(); | |||
11147 | ||||
11148 | // Convert i32 vectors to floating point if it is not AVX2. | |||
11149 | // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors. | |||
11150 | MVT BlendVT = VT; | |||
11151 | if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) { | |||
11152 | BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()), | |||
11153 | NumElems); | |||
11154 | V1 = DAG.getNode(ISD::BITCAST, dl, VT, V1); | |||
11155 | V2 = DAG.getNode(ISD::BITCAST, dl, VT, V2); | |||
11156 | } | |||
11157 | ||||
11158 | SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, V1, V2, | |||
11159 | DAG.getConstant(MaskValue, MVT::i32)); | |||
11160 | return DAG.getNode(ISD::BITCAST, dl, VT, Ret); | |||
11161 | } | |||
11162 | ||||
11163 | /// In vector type \p VT, return true if the element at index \p InputIdx | |||
11164 | /// falls on a different 128-bit lane than \p OutputIdx. | |||
11165 | static bool ShuffleCrosses128bitLane(MVT VT, unsigned InputIdx, | |||
11166 | unsigned OutputIdx) { | |||
11167 | unsigned EltSize = VT.getVectorElementType().getSizeInBits(); | |||
11168 | return InputIdx * EltSize / 128 != OutputIdx * EltSize / 128; | |||
11169 | } | |||
11170 | ||||
11171 | /// Generate a PSHUFB if possible. Selects elements from \p V1 according to | |||
11172 | /// \p MaskVals. MaskVals[OutputIdx] = InputIdx specifies that we want to | |||
11173 | /// shuffle the element at InputIdx in V1 to OutputIdx in the result. If \p | |||
11174 | /// MaskVals refers to elements outside of \p V1 or is undef (-1), insert a | |||
11175 | /// zero. | |||
11176 | static SDValue getPSHUFB(ArrayRef<int> MaskVals, SDValue V1, SDLoc &dl, | |||
11177 | SelectionDAG &DAG) { | |||
11178 | MVT VT = V1.getSimpleValueType(); | |||
11179 | assert(VT.is128BitVector() || VT.is256BitVector())((VT.is128BitVector() || VT.is256BitVector()) ? static_cast< void> (0) : __assert_fail ("VT.is128BitVector() || VT.is256BitVector()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 11179, __PRETTY_FUNCTION__)); | |||
11180 | ||||
11181 | MVT EltVT = VT.getVectorElementType(); | |||
11182 | unsigned EltSizeInBytes = EltVT.getSizeInBits() / 8; | |||
11183 | unsigned NumElts = VT.getVectorNumElements(); | |||
11184 | ||||
11185 | SmallVector<SDValue, 32> PshufbMask; | |||
11186 | for (unsigned OutputIdx = 0; OutputIdx < NumElts; ++OutputIdx) { | |||
11187 | int InputIdx = MaskVals[OutputIdx]; | |||
11188 | unsigned InputByteIdx; | |||
11189 | ||||
11190 | if (InputIdx < 0 || NumElts <= (unsigned)InputIdx) | |||
11191 | InputByteIdx = 0x80; | |||
11192 | else { | |||
11193 | // Cross lane is not allowed. | |||
11194 | if (ShuffleCrosses128bitLane(VT, InputIdx, OutputIdx)) | |||
11195 | return SDValue(); | |||
11196 | InputByteIdx = InputIdx * EltSizeInBytes; | |||
11197 | // Index is an byte offset within the 128-bit lane. | |||
11198 | InputByteIdx &= 0xf; | |||
11199 | } | |||
11200 | ||||
11201 | for (unsigned j = 0; j < EltSizeInBytes; ++j) { | |||
11202 | PshufbMask.push_back(DAG.getConstant(InputByteIdx, MVT::i8)); | |||
11203 | if (InputByteIdx != 0x80) | |||
11204 | ++InputByteIdx; | |||
11205 | } | |||
11206 | } | |||
11207 | ||||
11208 | MVT ShufVT = MVT::getVectorVT(MVT::i8, PshufbMask.size()); | |||
11209 | if (ShufVT != VT) | |||
11210 | V1 = DAG.getNode(ISD::BITCAST, dl, ShufVT, V1); | |||
11211 | return DAG.getNode(X86ISD::PSHUFB, dl, ShufVT, V1, | |||
11212 | DAG.getNode(ISD::BUILD_VECTOR, dl, ShufVT, PshufbMask)); | |||
11213 | } | |||
11214 | ||||
11215 | // v8i16 shuffles - Prefer shuffles in the following order: | |||
11216 | // 1. [all] pshuflw, pshufhw, optional move | |||
11217 | // 2. [ssse3] 1 x pshufb | |||
11218 | // 3. [ssse3] 2 x pshufb + 1 x por | |||
11219 | // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw) | |||
11220 | static SDValue | |||
11221 | LowerVECTOR_SHUFFLEv8i16(SDValue Op, const X86Subtarget *Subtarget, | |||
11222 | SelectionDAG &DAG) { | |||
11223 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
11224 | SDValue V1 = SVOp->getOperand(0); | |||
11225 | SDValue V2 = SVOp->getOperand(1); | |||
11226 | SDLoc dl(SVOp); | |||
11227 | SmallVector<int, 8> MaskVals; | |||
11228 | ||||
11229 | // Determine if more than 1 of the words in each of the low and high quadwords | |||
11230 | // of the result come from the same quadword of one of the two inputs. Undef | |||
11231 | // mask values count as coming from any quadword, for better codegen. | |||
11232 | // | |||
11233 | // Lo/HiQuad[i] = j indicates how many words from the ith quad of the input | |||
11234 | // feeds this quad. For i, 0 and 1 refer to V1, 2 and 3 refer to V2. | |||
11235 | unsigned LoQuad[] = { 0, 0, 0, 0 }; | |||
11236 | unsigned HiQuad[] = { 0, 0, 0, 0 }; | |||
11237 | // Indices of quads used. | |||
11238 | std::bitset<4> InputQuads; | |||
11239 | for (unsigned i = 0; i < 8; ++i) { | |||
11240 | unsigned *Quad = i < 4 ? LoQuad : HiQuad; | |||
11241 | int EltIdx = SVOp->getMaskElt(i); | |||
11242 | MaskVals.push_back(EltIdx); | |||
11243 | if (EltIdx < 0) { | |||
11244 | ++Quad[0]; | |||
11245 | ++Quad[1]; | |||
11246 | ++Quad[2]; | |||
11247 | ++Quad[3]; | |||
11248 | continue; | |||
11249 | } | |||
11250 | ++Quad[EltIdx / 4]; | |||
11251 | InputQuads.set(EltIdx / 4); | |||
11252 | } | |||
11253 | ||||
11254 | int BestLoQuad = -1; | |||
11255 | unsigned MaxQuad = 1; | |||
11256 | for (unsigned i = 0; i < 4; ++i) { | |||
11257 | if (LoQuad[i] > MaxQuad) { | |||
11258 | BestLoQuad = i; | |||
11259 | MaxQuad = LoQuad[i]; | |||
11260 | } | |||
11261 | } | |||
11262 | ||||
11263 | int BestHiQuad = -1; | |||
11264 | MaxQuad = 1; | |||
11265 | for (unsigned i = 0; i < 4; ++i) { | |||
11266 | if (HiQuad[i] > MaxQuad) { | |||
11267 | BestHiQuad = i; | |||
11268 | MaxQuad = HiQuad[i]; | |||
11269 | } | |||
11270 | } | |||
11271 | ||||
11272 | // For SSSE3, If all 8 words of the result come from only 1 quadword of each | |||
11273 | // of the two input vectors, shuffle them into one input vector so only a | |||
11274 | // single pshufb instruction is necessary. If there are more than 2 input | |||
11275 | // quads, disable the next transformation since it does not help SSSE3. | |||
11276 | bool V1Used = InputQuads[0] || InputQuads[1]; | |||
11277 | bool V2Used = InputQuads[2] || InputQuads[3]; | |||
11278 | if (Subtarget->hasSSSE3()) { | |||
11279 | if (InputQuads.count() == 2 && V1Used && V2Used) { | |||
11280 | BestLoQuad = InputQuads[0] ? 0 : 1; | |||
11281 | BestHiQuad = InputQuads[2] ? 2 : 3; | |||
11282 | } | |||
11283 | if (InputQuads.count() > 2) { | |||
11284 | BestLoQuad = -1; | |||
11285 | BestHiQuad = -1; | |||
11286 | } | |||
11287 | } | |||
11288 | ||||
11289 | // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update | |||
11290 | // the shuffle mask. If a quad is scored as -1, that means that it contains | |||
11291 | // words from all 4 input quadwords. | |||
11292 | SDValue NewV; | |||
11293 | if (BestLoQuad >= 0 || BestHiQuad >= 0) { | |||
11294 | int MaskV[] = { | |||
11295 | BestLoQuad < 0 ? 0 : BestLoQuad, | |||
11296 | BestHiQuad < 0 ? 1 : BestHiQuad | |||
11297 | }; | |||
11298 | NewV = DAG.getVectorShuffle(MVT::v2i64, dl, | |||
11299 | DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1), | |||
11300 | DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]); | |||
11301 | NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV); | |||
11302 | ||||
11303 | // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the | |||
11304 | // source words for the shuffle, to aid later transformations. | |||
11305 | bool AllWordsInNewV = true; | |||
11306 | bool InOrder[2] = { true, true }; | |||
11307 | for (unsigned i = 0; i != 8; ++i) { | |||
11308 | int idx = MaskVals[i]; | |||
11309 | if (idx != (int)i) | |||
11310 | InOrder[i/4] = false; | |||
11311 | if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad) | |||
11312 | continue; | |||
11313 | AllWordsInNewV = false; | |||
11314 | break; | |||
11315 | } | |||
11316 | ||||
11317 | bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV; | |||
11318 | if (AllWordsInNewV) { | |||
11319 | for (int i = 0; i != 8; ++i) { | |||
11320 | int idx = MaskVals[i]; | |||
11321 | if (idx < 0) | |||
11322 | continue; | |||
11323 | idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4; | |||
11324 | if ((idx != i) && idx < 4) | |||
11325 | pshufhw = false; | |||
11326 | if ((idx != i) && idx > 3) | |||
11327 | pshuflw = false; | |||
11328 | } | |||
11329 | V1 = NewV; | |||
11330 | V2Used = false; | |||
11331 | BestLoQuad = 0; | |||
11332 | BestHiQuad = 1; | |||
11333 | } | |||
11334 | ||||
11335 | // If we've eliminated the use of V2, and the new mask is a pshuflw or | |||
11336 | // pshufhw, that's as cheap as it gets. Return the new shuffle. | |||
11337 | if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) { | |||
11338 | unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW; | |||
11339 | unsigned TargetMask = 0; | |||
11340 | NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, | |||
11341 | DAG.getUNDEF(MVT::v8i16), &MaskVals[0]); | |||
11342 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode()); | |||
11343 | TargetMask = pshufhw ? getShufflePSHUFHWImmediate(SVOp): | |||
11344 | getShufflePSHUFLWImmediate(SVOp); | |||
11345 | V1 = NewV.getOperand(0); | |||
11346 | return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG); | |||
11347 | } | |||
11348 | } | |||
11349 | ||||
11350 | // Promote splats to a larger type which usually leads to more efficient code. | |||
11351 | // FIXME: Is this true if pshufb is available? | |||
11352 | if (SVOp->isSplat()) | |||
11353 | return PromoteSplat(SVOp, DAG); | |||
11354 | ||||
11355 | // If we have SSSE3, and all words of the result are from 1 input vector, | |||
11356 | // case 2 is generated, otherwise case 3 is generated. If no SSSE3 | |||
11357 | // is present, fall back to case 4. | |||
11358 | if (Subtarget->hasSSSE3()) { | |||
11359 | SmallVector<SDValue,16> pshufbMask; | |||
11360 | ||||
11361 | // If we have elements from both input vectors, set the high bit of the | |||
11362 | // shuffle mask element to zero out elements that come from V2 in the V1 | |||
11363 | // mask, and elements that come from V1 in the V2 mask, so that the two | |||
11364 | // results can be OR'd together. | |||
11365 | bool TwoInputs = V1Used && V2Used; | |||
11366 | V1 = getPSHUFB(MaskVals, V1, dl, DAG); | |||
11367 | if (!TwoInputs) | |||
11368 | return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1); | |||
11369 | ||||
11370 | // Calculate the shuffle mask for the second input, shuffle it, and | |||
11371 | // OR it with the first shuffled input. | |||
11372 | CommuteVectorShuffleMask(MaskVals, 8); | |||
11373 | V2 = getPSHUFB(MaskVals, V2, dl, DAG); | |||
11374 | V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2); | |||
11375 | return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1); | |||
11376 | } | |||
11377 | ||||
11378 | // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order, | |||
11379 | // and update MaskVals with new element order. | |||
11380 | std::bitset<8> InOrder; | |||
11381 | if (BestLoQuad >= 0) { | |||
11382 | int MaskV[] = { -1, -1, -1, -1, 4, 5, 6, 7 }; | |||
11383 | for (int i = 0; i != 4; ++i) { | |||
11384 | int idx = MaskVals[i]; | |||
11385 | if (idx < 0) { | |||
11386 | InOrder.set(i); | |||
11387 | } else if ((idx / 4) == BestLoQuad) { | |||
11388 | MaskV[i] = idx & 3; | |||
11389 | InOrder.set(i); | |||
11390 | } | |||
11391 | } | |||
11392 | NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16), | |||
11393 | &MaskV[0]); | |||
11394 | ||||
11395 | if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSE2()) { | |||
11396 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode()); | |||
11397 | NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16, | |||
11398 | NewV.getOperand(0), | |||
11399 | getShufflePSHUFLWImmediate(SVOp), DAG); | |||
11400 | } | |||
11401 | } | |||
11402 | ||||
11403 | // If BestHi >= 0, generate a pshufhw to put the high elements in order, | |||
11404 | // and update MaskVals with the new element order. | |||
11405 | if (BestHiQuad >= 0) { | |||
11406 | int MaskV[] = { 0, 1, 2, 3, -1, -1, -1, -1 }; | |||
11407 | for (unsigned i = 4; i != 8; ++i) { | |||
11408 | int idx = MaskVals[i]; | |||
11409 | if (idx < 0) { | |||
11410 | InOrder.set(i); | |||
11411 | } else if ((idx / 4) == BestHiQuad) { | |||
11412 | MaskV[i] = (idx & 3) + 4; | |||
11413 | InOrder.set(i); | |||
11414 | } | |||
11415 | } | |||
11416 | NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16), | |||
11417 | &MaskV[0]); | |||
11418 | ||||
11419 | if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSE2()) { | |||
11420 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode()); | |||
11421 | NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16, | |||
11422 | NewV.getOperand(0), | |||
11423 | getShufflePSHUFHWImmediate(SVOp), DAG); | |||
11424 | } | |||
11425 | } | |||
11426 | ||||
11427 | // In case BestHi & BestLo were both -1, which means each quadword has a word | |||
11428 | // from each of the four input quadwords, calculate the InOrder bitvector now | |||
11429 | // before falling through to the insert/extract cleanup. | |||
11430 | if (BestLoQuad == -1 && BestHiQuad == -1) { | |||
11431 | NewV = V1; | |||
11432 | for (int i = 0; i != 8; ++i) | |||
11433 | if (MaskVals[i] < 0 || MaskVals[i] == i) | |||
11434 | InOrder.set(i); | |||
11435 | } | |||
11436 | ||||
11437 | // The other elements are put in the right place using pextrw and pinsrw. | |||
11438 | for (unsigned i = 0; i != 8; ++i) { | |||
11439 | if (InOrder[i]) | |||
11440 | continue; | |||
11441 | int EltIdx = MaskVals[i]; | |||
11442 | if (EltIdx < 0) | |||
11443 | continue; | |||
11444 | SDValue ExtOp = (EltIdx < 8) ? | |||
11445 | DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1, | |||
11446 | DAG.getIntPtrConstant(EltIdx)) : | |||
11447 | DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2, | |||
11448 | DAG.getIntPtrConstant(EltIdx - 8)); | |||
11449 | NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp, | |||
11450 | DAG.getIntPtrConstant(i)); | |||
11451 | } | |||
11452 | return NewV; | |||
11453 | } | |||
11454 | ||||
11455 | /// \brief v16i16 shuffles | |||
11456 | /// | |||
11457 | /// FIXME: We only support generation of a single pshufb currently. We can | |||
11458 | /// generalize the other applicable cases from LowerVECTOR_SHUFFLEv8i16 as | |||
11459 | /// well (e.g 2 x pshufb + 1 x por). | |||
11460 | static SDValue | |||
11461 | LowerVECTOR_SHUFFLEv16i16(SDValue Op, SelectionDAG &DAG) { | |||
11462 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
11463 | SDValue V1 = SVOp->getOperand(0); | |||
11464 | SDValue V2 = SVOp->getOperand(1); | |||
11465 | SDLoc dl(SVOp); | |||
11466 | ||||
11467 | if (V2.getOpcode() != ISD::UNDEF) | |||
11468 | return SDValue(); | |||
11469 | ||||
11470 | SmallVector<int, 16> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end()); | |||
11471 | return getPSHUFB(MaskVals, V1, dl, DAG); | |||
11472 | } | |||
11473 | ||||
11474 | // v16i8 shuffles - Prefer shuffles in the following order: | |||
11475 | // 1. [ssse3] 1 x pshufb | |||
11476 | // 2. [ssse3] 2 x pshufb + 1 x por | |||
11477 | // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw | |||
11478 | static SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp, | |||
11479 | const X86Subtarget* Subtarget, | |||
11480 | SelectionDAG &DAG) { | |||
11481 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
11482 | SDValue V1 = SVOp->getOperand(0); | |||
11483 | SDValue V2 = SVOp->getOperand(1); | |||
11484 | SDLoc dl(SVOp); | |||
11485 | ArrayRef<int> MaskVals = SVOp->getMask(); | |||
11486 | ||||
11487 | // Promote splats to a larger type which usually leads to more efficient code. | |||
11488 | // FIXME: Is this true if pshufb is available? | |||
11489 | if (SVOp->isSplat()) | |||
11490 | return PromoteSplat(SVOp, DAG); | |||
11491 | ||||
11492 | // If we have SSSE3, case 1 is generated when all result bytes come from | |||
11493 | // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is | |||
11494 | // present, fall back to case 3. | |||
11495 | ||||
11496 | // If SSSE3, use 1 pshufb instruction per vector with elements in the result. | |||
11497 | if (Subtarget->hasSSSE3()) { | |||
11498 | SmallVector<SDValue,16> pshufbMask; | |||
11499 | ||||
11500 | // If all result elements are from one input vector, then only translate | |||
11501 | // undef mask values to 0x80 (zero out result) in the pshufb mask. | |||
11502 | // | |||
11503 | // Otherwise, we have elements from both input vectors, and must zero out | |||
11504 | // elements that come from V2 in the first mask, and V1 in the second mask | |||
11505 | // so that we can OR them together. | |||
11506 | for (unsigned i = 0; i != 16; ++i) { | |||
11507 | int EltIdx = MaskVals[i]; | |||
11508 | if (EltIdx < 0 || EltIdx >= 16) | |||
11509 | EltIdx = 0x80; | |||
11510 | pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8)); | |||
11511 | } | |||
11512 | V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1, | |||
11513 | DAG.getNode(ISD::BUILD_VECTOR, dl, | |||
11514 | MVT::v16i8, pshufbMask)); | |||
11515 | ||||
11516 | // As PSHUFB will zero elements with negative indices, it's safe to ignore | |||
11517 | // the 2nd operand if it's undefined or zero. | |||
11518 | if (V2.getOpcode() == ISD::UNDEF || | |||
11519 | ISD::isBuildVectorAllZeros(V2.getNode())) | |||
11520 | return V1; | |||
11521 | ||||
11522 | // Calculate the shuffle mask for the second input, shuffle it, and | |||
11523 | // OR it with the first shuffled input. | |||
11524 | pshufbMask.clear(); | |||
11525 | for (unsigned i = 0; i != 16; ++i) { | |||
11526 | int EltIdx = MaskVals[i]; | |||
11527 | EltIdx = (EltIdx < 16) ? 0x80 : EltIdx - 16; | |||
11528 | pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8)); | |||
11529 | } | |||
11530 | V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2, | |||
11531 | DAG.getNode(ISD::BUILD_VECTOR, dl, | |||
11532 | MVT::v16i8, pshufbMask)); | |||
11533 | return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2); | |||
11534 | } | |||
11535 | ||||
11536 | // No SSSE3 - Calculate in place words and then fix all out of place words | |||
11537 | // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from | |||
11538 | // the 16 different words that comprise the two doublequadword input vectors. | |||
11539 | V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1); | |||
11540 | V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2); | |||
11541 | SDValue NewV = V1; | |||
11542 | for (int i = 0; i != 8; ++i) { | |||
11543 | int Elt0 = MaskVals[i*2]; | |||
11544 | int Elt1 = MaskVals[i*2+1]; | |||
11545 | ||||
11546 | // This word of the result is all undef, skip it. | |||
11547 | if (Elt0 < 0 && Elt1 < 0) | |||
11548 | continue; | |||
11549 | ||||
11550 | // This word of the result is already in the correct place, skip it. | |||
11551 | if ((Elt0 == i*2) && (Elt1 == i*2+1)) | |||
11552 | continue; | |||
11553 | ||||
11554 | SDValue Elt0Src = Elt0 < 16 ? V1 : V2; | |||
11555 | SDValue Elt1Src = Elt1 < 16 ? V1 : V2; | |||
11556 | SDValue InsElt; | |||
11557 | ||||
11558 | // If Elt0 and Elt1 are defined, are consecutive, and can be load | |||
11559 | // using a single extract together, load it and store it. | |||
11560 | if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) { | |||
11561 | InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src, | |||
11562 | DAG.getIntPtrConstant(Elt1 / 2)); | |||
11563 | NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt, | |||
11564 | DAG.getIntPtrConstant(i)); | |||
11565 | continue; | |||
11566 | } | |||
11567 | ||||
11568 | // If Elt1 is defined, extract it from the appropriate source. If the | |||
11569 | // source byte is not also odd, shift the extracted word left 8 bits | |||
11570 | // otherwise clear the bottom 8 bits if we need to do an or. | |||
11571 | if (Elt1 >= 0) { | |||
11572 | InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src, | |||
11573 | DAG.getIntPtrConstant(Elt1 / 2)); | |||
11574 | if ((Elt1 & 1) == 0) | |||
11575 | InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt, | |||
11576 | DAG.getConstant(8, | |||
11577 | TLI.getShiftAmountTy(InsElt.getValueType()))); | |||
11578 | else if (Elt0 >= 0) | |||
11579 | InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt, | |||
11580 | DAG.getConstant(0xFF00, MVT::i16)); | |||
11581 | } | |||
11582 | // If Elt0 is defined, extract it from the appropriate source. If the | |||
11583 | // source byte is not also even, shift the extracted word right 8 bits. If | |||
11584 | // Elt1 was also defined, OR the extracted values together before | |||
11585 | // inserting them in the result. | |||
11586 | if (Elt0 >= 0) { | |||
11587 | SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, | |||
11588 | Elt0Src, DAG.getIntPtrConstant(Elt0 / 2)); | |||
11589 | if ((Elt0 & 1) != 0) | |||
11590 | InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0, | |||
11591 | DAG.getConstant(8, | |||
11592 | TLI.getShiftAmountTy(InsElt0.getValueType()))); | |||
11593 | else if (Elt1 >= 0) | |||
11594 | InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0, | |||
11595 | DAG.getConstant(0x00FF, MVT::i16)); | |||
11596 | InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0) | |||
11597 | : InsElt0; | |||
11598 | } | |||
11599 | NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt, | |||
11600 | DAG.getIntPtrConstant(i)); | |||
11601 | } | |||
11602 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV); | |||
11603 | } | |||
11604 | ||||
11605 | // v32i8 shuffles - Translate to VPSHUFB if possible. | |||
11606 | static | |||
11607 | SDValue LowerVECTOR_SHUFFLEv32i8(ShuffleVectorSDNode *SVOp, | |||
11608 | const X86Subtarget *Subtarget, | |||
11609 | SelectionDAG &DAG) { | |||
11610 | MVT VT = SVOp->getSimpleValueType(0); | |||
11611 | SDValue V1 = SVOp->getOperand(0); | |||
11612 | SDValue V2 = SVOp->getOperand(1); | |||
11613 | SDLoc dl(SVOp); | |||
11614 | SmallVector<int, 32> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end()); | |||
11615 | ||||
11616 | bool V2IsUndef = V2.getOpcode() == ISD::UNDEF; | |||
11617 | bool V1IsAllZero = ISD::isBuildVectorAllZeros(V1.getNode()); | |||
11618 | bool V2IsAllZero = ISD::isBuildVectorAllZeros(V2.getNode()); | |||
11619 | ||||
11620 | // VPSHUFB may be generated if | |||
11621 | // (1) one of input vector is undefined or zeroinitializer. | |||
11622 | // The mask value 0x80 puts 0 in the corresponding slot of the vector. | |||
11623 | // And (2) the mask indexes don't cross the 128-bit lane. | |||
11624 | if (VT != MVT::v32i8 || !Subtarget->hasInt256() || | |||
11625 | (!V2IsUndef && !V2IsAllZero && !V1IsAllZero)) | |||
11626 | return SDValue(); | |||
11627 | ||||
11628 | if (V1IsAllZero && !V2IsAllZero) { | |||
11629 | CommuteVectorShuffleMask(MaskVals, 32); | |||
11630 | V1 = V2; | |||
11631 | } | |||
11632 | return getPSHUFB(MaskVals, V1, dl, DAG); | |||
11633 | } | |||
11634 | ||||
11635 | /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide | |||
11636 | /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be | |||
11637 | /// done when every pair / quad of shuffle mask elements point to elements in | |||
11638 | /// the right sequence. e.g. | |||
11639 | /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15> | |||
11640 | static | |||
11641 | SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp, | |||
11642 | SelectionDAG &DAG) { | |||
11643 | MVT VT = SVOp->getSimpleValueType(0); | |||
11644 | SDLoc dl(SVOp); | |||
11645 | unsigned NumElems = VT.getVectorNumElements(); | |||
11646 | MVT NewVT; | |||
11647 | unsigned Scale; | |||
11648 | switch (VT.SimpleTy) { | |||
11649 | default: llvm_unreachable("Unexpected!")::llvm::llvm_unreachable_internal("Unexpected!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 11649); | |||
11650 | case MVT::v2i64: | |||
11651 | case MVT::v2f64: | |||
11652 | return SDValue(SVOp, 0); | |||
11653 | case MVT::v4f32: NewVT = MVT::v2f64; Scale = 2; break; | |||
11654 | case MVT::v4i32: NewVT = MVT::v2i64; Scale = 2; break; | |||
11655 | case MVT::v8i16: NewVT = MVT::v4i32; Scale = 2; break; | |||
11656 | case MVT::v16i8: NewVT = MVT::v4i32; Scale = 4; break; | |||
11657 | case MVT::v16i16: NewVT = MVT::v8i32; Scale = 2; break; | |||
11658 | case MVT::v32i8: NewVT = MVT::v8i32; Scale = 4; break; | |||
11659 | } | |||
11660 | ||||
11661 | SmallVector<int, 8> MaskVec; | |||
11662 | for (unsigned i = 0; i != NumElems; i += Scale) { | |||
11663 | int StartIdx = -1; | |||
11664 | for (unsigned j = 0; j != Scale; ++j) { | |||
11665 | int EltIdx = SVOp->getMaskElt(i+j); | |||
11666 | if (EltIdx < 0) | |||
11667 | continue; | |||
11668 | if (StartIdx < 0) | |||
11669 | StartIdx = (EltIdx / Scale); | |||
11670 | if (EltIdx != (int)(StartIdx*Scale + j)) | |||
11671 | return SDValue(); | |||
11672 | } | |||
11673 | MaskVec.push_back(StartIdx); | |||
11674 | } | |||
11675 | ||||
11676 | SDValue V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(0)); | |||
11677 | SDValue V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(1)); | |||
11678 | return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]); | |||
11679 | } | |||
11680 | ||||
11681 | /// getVZextMovL - Return a zero-extending vector move low node. | |||
11682 | /// | |||
11683 | static SDValue getVZextMovL(MVT VT, MVT OpVT, | |||
11684 | SDValue SrcOp, SelectionDAG &DAG, | |||
11685 | const X86Subtarget *Subtarget, SDLoc dl) { | |||
11686 | if (VT == MVT::v2f64 || VT == MVT::v4f32) { | |||
11687 | LoadSDNode *LD = nullptr; | |||
11688 | if (!isScalarLoadToVector(SrcOp.getNode(), &LD)) | |||
11689 | LD = dyn_cast<LoadSDNode>(SrcOp); | |||
11690 | if (!LD) { | |||
11691 | // movssrr and movsdrr do not clear top bits. Try to use movd, movq | |||
11692 | // instead. | |||
11693 | MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32; | |||
11694 | if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) && | |||
11695 | SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR && | |||
11696 | SrcOp.getOperand(0).getOpcode() == ISD::BITCAST && | |||
11697 | SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) { | |||
11698 | // PR2108 | |||
11699 | OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32; | |||
11700 | return DAG.getNode(ISD::BITCAST, dl, VT, | |||
11701 | DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT, | |||
11702 | DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, | |||
11703 | OpVT, | |||
11704 | SrcOp.getOperand(0) | |||
11705 | .getOperand(0)))); | |||
11706 | } | |||
11707 | } | |||
11708 | } | |||
11709 | ||||
11710 | return DAG.getNode(ISD::BITCAST, dl, VT, | |||
11711 | DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT, | |||
11712 | DAG.getNode(ISD::BITCAST, dl, | |||
11713 | OpVT, SrcOp))); | |||
11714 | } | |||
11715 | ||||
11716 | /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles | |||
11717 | /// which could not be matched by any known target speficic shuffle | |||
11718 | static SDValue | |||
11719 | LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) { | |||
11720 | ||||
11721 | SDValue NewOp = Compact8x32ShuffleNode(SVOp, DAG); | |||
11722 | if (NewOp.getNode()) | |||
11723 | return NewOp; | |||
11724 | ||||
11725 | MVT VT = SVOp->getSimpleValueType(0); | |||
11726 | ||||
11727 | unsigned NumElems = VT.getVectorNumElements(); | |||
11728 | unsigned NumLaneElems = NumElems / 2; | |||
11729 | ||||
11730 | SDLoc dl(SVOp); | |||
11731 | MVT EltVT = VT.getVectorElementType(); | |||
11732 | MVT NVT = MVT::getVectorVT(EltVT, NumLaneElems); | |||
11733 | SDValue Output[2]; | |||
11734 | ||||
11735 | SmallVector<int, 16> Mask; | |||
11736 | for (unsigned l = 0; l < 2; ++l) { | |||
11737 | // Build a shuffle mask for the output, discovering on the fly which | |||
11738 | // input vectors to use as shuffle operands (recorded in InputUsed). | |||
11739 | // If building a suitable shuffle vector proves too hard, then bail | |||
11740 | // out with UseBuildVector set. | |||
11741 | bool UseBuildVector = false; | |||
11742 | int InputUsed[2] = { -1, -1 }; // Not yet discovered. | |||
11743 | unsigned LaneStart = l * NumLaneElems; | |||
11744 | for (unsigned i = 0; i != NumLaneElems; ++i) { | |||
11745 | // The mask element. This indexes into the input. | |||
11746 | int Idx = SVOp->getMaskElt(i+LaneStart); | |||
11747 | if (Idx < 0) { | |||
11748 | // the mask element does not index into any input vector. | |||
11749 | Mask.push_back(-1); | |||
11750 | continue; | |||
11751 | } | |||
11752 | ||||
11753 | // The input vector this mask element indexes into. | |||
11754 | int Input = Idx / NumLaneElems; | |||
11755 | ||||
11756 | // Turn the index into an offset from the start of the input vector. | |||
11757 | Idx -= Input * NumLaneElems; | |||
11758 | ||||
11759 | // Find or create a shuffle vector operand to hold this input. | |||
11760 | unsigned OpNo; | |||
11761 | for (OpNo = 0; OpNo < array_lengthof(InputUsed); ++OpNo) { | |||
11762 | if (InputUsed[OpNo] == Input) | |||
11763 | // This input vector is already an operand. | |||
11764 | break; | |||
11765 | if (InputUsed[OpNo] < 0) { | |||
11766 | // Create a new operand for this input vector. | |||
11767 | InputUsed[OpNo] = Input; | |||
11768 | break; | |||
11769 | } | |||
11770 | } | |||
11771 | ||||
11772 | if (OpNo >= array_lengthof(InputUsed)) { | |||
11773 | // More than two input vectors used! Give up on trying to create a | |||
11774 | // shuffle vector. Insert all elements into a BUILD_VECTOR instead. | |||
11775 | UseBuildVector = true; | |||
11776 | break; | |||
11777 | } | |||
11778 | ||||
11779 | // Add the mask index for the new shuffle vector. | |||
11780 | Mask.push_back(Idx + OpNo * NumLaneElems); | |||
11781 | } | |||
11782 | ||||
11783 | if (UseBuildVector) { | |||
11784 | SmallVector<SDValue, 16> SVOps; | |||
11785 | for (unsigned i = 0; i != NumLaneElems; ++i) { | |||
11786 | // The mask element. This indexes into the input. | |||
11787 | int Idx = SVOp->getMaskElt(i+LaneStart); | |||
11788 | if (Idx < 0) { | |||
11789 | SVOps.push_back(DAG.getUNDEF(EltVT)); | |||
11790 | continue; | |||
11791 | } | |||
11792 | ||||
11793 | // The input vector this mask element indexes into. | |||
11794 | int Input = Idx / NumElems; | |||
11795 | ||||
11796 | // Turn the index into an offset from the start of the input vector. | |||
11797 | Idx -= Input * NumElems; | |||
11798 | ||||
11799 | // Extract the vector element by hand. | |||
11800 | SVOps.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, | |||
11801 | SVOp->getOperand(Input), | |||
11802 | DAG.getIntPtrConstant(Idx))); | |||
11803 | } | |||
11804 | ||||
11805 | // Construct the output using a BUILD_VECTOR. | |||
11806 | Output[l] = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, SVOps); | |||
11807 | } else if (InputUsed[0] < 0) { | |||
11808 | // No input vectors were used! The result is undefined. | |||
11809 | Output[l] = DAG.getUNDEF(NVT); | |||
11810 | } else { | |||
11811 | SDValue Op0 = Extract128BitVector(SVOp->getOperand(InputUsed[0] / 2), | |||
11812 | (InputUsed[0] % 2) * NumLaneElems, | |||
11813 | DAG, dl); | |||
11814 | // If only one input was used, use an undefined vector for the other. | |||
11815 | SDValue Op1 = (InputUsed[1] < 0) ? DAG.getUNDEF(NVT) : | |||
11816 | Extract128BitVector(SVOp->getOperand(InputUsed[1] / 2), | |||
11817 | (InputUsed[1] % 2) * NumLaneElems, DAG, dl); | |||
11818 | // At least one input vector was used. Create a new shuffle vector. | |||
11819 | Output[l] = DAG.getVectorShuffle(NVT, dl, Op0, Op1, &Mask[0]); | |||
11820 | } | |||
11821 | ||||
11822 | Mask.clear(); | |||
11823 | } | |||
11824 | ||||
11825 | // Concatenate the result back | |||
11826 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Output[0], Output[1]); | |||
11827 | } | |||
11828 | ||||
11829 | /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with | |||
11830 | /// 4 elements, and match them with several different shuffle types. | |||
11831 | static SDValue | |||
11832 | LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) { | |||
11833 | SDValue V1 = SVOp->getOperand(0); | |||
11834 | SDValue V2 = SVOp->getOperand(1); | |||
11835 | SDLoc dl(SVOp); | |||
11836 | MVT VT = SVOp->getSimpleValueType(0); | |||
11837 | ||||
11838 | assert(VT.is128BitVector() && "Unsupported vector size")((VT.is128BitVector() && "Unsupported vector size") ? static_cast<void> (0) : __assert_fail ("VT.is128BitVector() && \"Unsupported vector size\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 11838, __PRETTY_FUNCTION__)); | |||
11839 | ||||
11840 | std::pair<int, int> Locs[4]; | |||
11841 | int Mask1[] = { -1, -1, -1, -1 }; | |||
11842 | SmallVector<int, 8> PermMask(SVOp->getMask().begin(), SVOp->getMask().end()); | |||
11843 | ||||
11844 | unsigned NumHi = 0; | |||
11845 | unsigned NumLo = 0; | |||
11846 | for (unsigned i = 0; i != 4; ++i) { | |||
11847 | int Idx = PermMask[i]; | |||
11848 | if (Idx < 0) { | |||
11849 | Locs[i] = std::make_pair(-1, -1); | |||
11850 | } else { | |||
11851 | assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!")((Idx < 8 && "Invalid VECTOR_SHUFFLE index!") ? static_cast <void> (0) : __assert_fail ("Idx < 8 && \"Invalid VECTOR_SHUFFLE index!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 11851, __PRETTY_FUNCTION__)); | |||
11852 | if (Idx < 4) { | |||
11853 | Locs[i] = std::make_pair(0, NumLo); | |||
11854 | Mask1[NumLo] = Idx; | |||
11855 | NumLo++; | |||
11856 | } else { | |||
11857 | Locs[i] = std::make_pair(1, NumHi); | |||
11858 | if (2+NumHi < 4) | |||
11859 | Mask1[2+NumHi] = Idx; | |||
11860 | NumHi++; | |||
11861 | } | |||
11862 | } | |||
11863 | } | |||
11864 | ||||
11865 | if (NumLo <= 2 && NumHi <= 2) { | |||
11866 | // If no more than two elements come from either vector. This can be | |||
11867 | // implemented with two shuffles. First shuffle gather the elements. | |||
11868 | // The second shuffle, which takes the first shuffle as both of its | |||
11869 | // vector operands, put the elements into the right order. | |||
11870 | V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]); | |||
11871 | ||||
11872 | int Mask2[] = { -1, -1, -1, -1 }; | |||
11873 | ||||
11874 | for (unsigned i = 0; i != 4; ++i) | |||
11875 | if (Locs[i].first != -1) { | |||
11876 | unsigned Idx = (i < 2) ? 0 : 4; | |||
11877 | Idx += Locs[i].first * 2 + Locs[i].second; | |||
11878 | Mask2[i] = Idx; | |||
11879 | } | |||
11880 | ||||
11881 | return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]); | |||
11882 | } | |||
11883 | ||||
11884 | if (NumLo == 3 || NumHi == 3) { | |||
11885 | // Otherwise, we must have three elements from one vector, call it X, and | |||
11886 | // one element from the other, call it Y. First, use a shufps to build an | |||
11887 | // intermediate vector with the one element from Y and the element from X | |||
11888 | // that will be in the same half in the final destination (the indexes don't | |||
11889 | // matter). Then, use a shufps to build the final vector, taking the half | |||
11890 | // containing the element from Y from the intermediate, and the other half | |||
11891 | // from X. | |||
11892 | if (NumHi == 3) { | |||
11893 | // Normalize it so the 3 elements come from V1. | |||
11894 | CommuteVectorShuffleMask(PermMask, 4); | |||
11895 | std::swap(V1, V2); | |||
11896 | } | |||
11897 | ||||
11898 | // Find the element from V2. | |||
11899 | unsigned HiIndex; | |||
11900 | for (HiIndex = 0; HiIndex < 3; ++HiIndex) { | |||
11901 | int Val = PermMask[HiIndex]; | |||
11902 | if (Val < 0) | |||
11903 | continue; | |||
11904 | if (Val >= 4) | |||
11905 | break; | |||
11906 | } | |||
11907 | ||||
11908 | Mask1[0] = PermMask[HiIndex]; | |||
11909 | Mask1[1] = -1; | |||
11910 | Mask1[2] = PermMask[HiIndex^1]; | |||
11911 | Mask1[3] = -1; | |||
11912 | V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]); | |||
11913 | ||||
11914 | if (HiIndex >= 2) { | |||
11915 | Mask1[0] = PermMask[0]; | |||
11916 | Mask1[1] = PermMask[1]; | |||
11917 | Mask1[2] = HiIndex & 1 ? 6 : 4; | |||
11918 | Mask1[3] = HiIndex & 1 ? 4 : 6; | |||
11919 | return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]); | |||
11920 | } | |||
11921 | ||||
11922 | Mask1[0] = HiIndex & 1 ? 2 : 0; | |||
11923 | Mask1[1] = HiIndex & 1 ? 0 : 2; | |||
11924 | Mask1[2] = PermMask[2]; | |||
11925 | Mask1[3] = PermMask[3]; | |||
11926 | if (Mask1[2] >= 0) | |||
11927 | Mask1[2] += 4; | |||
11928 | if (Mask1[3] >= 0) | |||
11929 | Mask1[3] += 4; | |||
11930 | return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]); | |||
11931 | } | |||
11932 | ||||
11933 | // Break it into (shuffle shuffle_hi, shuffle_lo). | |||
11934 | int LoMask[] = { -1, -1, -1, -1 }; | |||
11935 | int HiMask[] = { -1, -1, -1, -1 }; | |||
11936 | ||||
11937 | int *MaskPtr = LoMask; | |||
11938 | unsigned MaskIdx = 0; | |||
11939 | unsigned LoIdx = 0; | |||
11940 | unsigned HiIdx = 2; | |||
11941 | for (unsigned i = 0; i != 4; ++i) { | |||
11942 | if (i == 2) { | |||
11943 | MaskPtr = HiMask; | |||
11944 | MaskIdx = 1; | |||
11945 | LoIdx = 0; | |||
11946 | HiIdx = 2; | |||
11947 | } | |||
11948 | int Idx = PermMask[i]; | |||
11949 | if (Idx < 0) { | |||
11950 | Locs[i] = std::make_pair(-1, -1); | |||
11951 | } else if (Idx < 4) { | |||
11952 | Locs[i] = std::make_pair(MaskIdx, LoIdx); | |||
11953 | MaskPtr[LoIdx] = Idx; | |||
11954 | LoIdx++; | |||
11955 | } else { | |||
11956 | Locs[i] = std::make_pair(MaskIdx, HiIdx); | |||
11957 | MaskPtr[HiIdx] = Idx; | |||
11958 | HiIdx++; | |||
11959 | } | |||
11960 | } | |||
11961 | ||||
11962 | SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]); | |||
11963 | SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]); | |||
11964 | int MaskOps[] = { -1, -1, -1, -1 }; | |||
11965 | for (unsigned i = 0; i != 4; ++i) | |||
11966 | if (Locs[i].first != -1) | |||
11967 | MaskOps[i] = Locs[i].first * 4 + Locs[i].second; | |||
11968 | return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]); | |||
11969 | } | |||
11970 | ||||
11971 | static bool MayFoldVectorLoad(SDValue V) { | |||
11972 | while (V.hasOneUse() && V.getOpcode() == ISD::BITCAST) | |||
11973 | V = V.getOperand(0); | |||
11974 | ||||
11975 | if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR) | |||
11976 | V = V.getOperand(0); | |||
11977 | if (V.hasOneUse() && V.getOpcode() == ISD::BUILD_VECTOR && | |||
11978 | V.getNumOperands() == 2 && V.getOperand(1).getOpcode() == ISD::UNDEF) | |||
11979 | // BUILD_VECTOR (load), undef | |||
11980 | V = V.getOperand(0); | |||
11981 | ||||
11982 | return MayFoldLoad(V); | |||
11983 | } | |||
11984 | ||||
11985 | static | |||
11986 | SDValue getMOVDDup(SDValue &Op, SDLoc &dl, SDValue V1, SelectionDAG &DAG) { | |||
11987 | MVT VT = Op.getSimpleValueType(); | |||
11988 | ||||
11989 | // Canonizalize to v2f64. | |||
11990 | V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1); | |||
11991 | return DAG.getNode(ISD::BITCAST, dl, VT, | |||
11992 | getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64, | |||
11993 | V1, DAG)); | |||
11994 | } | |||
11995 | ||||
11996 | static | |||
11997 | SDValue getMOVLowToHigh(SDValue &Op, SDLoc &dl, SelectionDAG &DAG, | |||
11998 | bool HasSSE2) { | |||
11999 | SDValue V1 = Op.getOperand(0); | |||
12000 | SDValue V2 = Op.getOperand(1); | |||
12001 | MVT VT = Op.getSimpleValueType(); | |||
12002 | ||||
12003 | assert(VT != MVT::v2i64 && "unsupported shuffle type")((VT != MVT::v2i64 && "unsupported shuffle type") ? static_cast <void> (0) : __assert_fail ("VT != MVT::v2i64 && \"unsupported shuffle type\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12003, __PRETTY_FUNCTION__)); | |||
12004 | ||||
12005 | if (HasSSE2 && VT == MVT::v2f64) | |||
12006 | return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG); | |||
12007 | ||||
12008 | // v4f32 or v4i32: canonizalized to v4f32 (which is legal for SSE1) | |||
12009 | return DAG.getNode(ISD::BITCAST, dl, VT, | |||
12010 | getTargetShuffleNode(X86ISD::MOVLHPS, dl, MVT::v4f32, | |||
12011 | DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V1), | |||
12012 | DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V2), DAG)); | |||
12013 | } | |||
12014 | ||||
12015 | static | |||
12016 | SDValue getMOVHighToLow(SDValue &Op, SDLoc &dl, SelectionDAG &DAG) { | |||
12017 | SDValue V1 = Op.getOperand(0); | |||
12018 | SDValue V2 = Op.getOperand(1); | |||
12019 | MVT VT = Op.getSimpleValueType(); | |||
12020 | ||||
12021 | assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&(((VT == MVT::v4i32 || VT == MVT::v4f32) && "unsupported shuffle type" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v4i32 || VT == MVT::v4f32) && \"unsupported shuffle type\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12022, __PRETTY_FUNCTION__)) | |||
12022 | "unsupported shuffle type")(((VT == MVT::v4i32 || VT == MVT::v4f32) && "unsupported shuffle type" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v4i32 || VT == MVT::v4f32) && \"unsupported shuffle type\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12022, __PRETTY_FUNCTION__)); | |||
12023 | ||||
12024 | if (V2.getOpcode() == ISD::UNDEF) | |||
12025 | V2 = V1; | |||
12026 | ||||
12027 | // v4i32 or v4f32 | |||
12028 | return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG); | |||
12029 | } | |||
12030 | ||||
12031 | static | |||
12032 | SDValue getMOVLP(SDValue &Op, SDLoc &dl, SelectionDAG &DAG, bool HasSSE2) { | |||
12033 | SDValue V1 = Op.getOperand(0); | |||
12034 | SDValue V2 = Op.getOperand(1); | |||
12035 | MVT VT = Op.getSimpleValueType(); | |||
12036 | unsigned NumElems = VT.getVectorNumElements(); | |||
12037 | ||||
12038 | // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second | |||
12039 | // operand of these instructions is only memory, so check if there's a | |||
12040 | // potencial load folding here, otherwise use SHUFPS or MOVSD to match the | |||
12041 | // same masks. | |||
12042 | bool CanFoldLoad = false; | |||
12043 | ||||
12044 | // Trivial case, when V2 comes from a load. | |||
12045 | if (MayFoldVectorLoad(V2)) | |||
12046 | CanFoldLoad = true; | |||
12047 | ||||
12048 | // When V1 is a load, it can be folded later into a store in isel, example: | |||
12049 | // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1) | |||
12050 | // turns into: | |||
12051 | // (MOVLPSmr addr:$src1, VR128:$src2) | |||
12052 | // So, recognize this potential and also use MOVLPS or MOVLPD | |||
12053 | else if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op)) | |||
12054 | CanFoldLoad = true; | |||
12055 | ||||
12056 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
12057 | if (CanFoldLoad) { | |||
12058 | if (HasSSE2 && NumElems == 2) | |||
12059 | return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG); | |||
12060 | ||||
12061 | if (NumElems == 4) | |||
12062 | // If we don't care about the second element, proceed to use movss. | |||
12063 | if (SVOp->getMaskElt(1) != -1) | |||
12064 | return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG); | |||
12065 | } | |||
12066 | ||||
12067 | // movl and movlp will both match v2i64, but v2i64 is never matched by | |||
12068 | // movl earlier because we make it strict to avoid messing with the movlp load | |||
12069 | // folding logic (see the code above getMOVLP call). Match it here then, | |||
12070 | // this is horrible, but will stay like this until we move all shuffle | |||
12071 | // matching to x86 specific nodes. Note that for the 1st condition all | |||
12072 | // types are matched with movsd. | |||
12073 | if (HasSSE2) { | |||
12074 | // FIXME: isMOVLMask should be checked and matched before getMOVLP, | |||
12075 | // as to remove this logic from here, as much as possible | |||
12076 | if (NumElems == 2 || !isMOVLMask(SVOp->getMask(), VT)) | |||
12077 | return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG); | |||
12078 | return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG); | |||
12079 | } | |||
12080 | ||||
12081 | assert(VT != MVT::v4i32 && "unsupported shuffle type")((VT != MVT::v4i32 && "unsupported shuffle type") ? static_cast <void> (0) : __assert_fail ("VT != MVT::v4i32 && \"unsupported shuffle type\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12081, __PRETTY_FUNCTION__)); | |||
12082 | ||||
12083 | // Invert the operand order and use SHUFPS to match it. | |||
12084 | return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V2, V1, | |||
12085 | getShuffleSHUFImmediate(SVOp), DAG); | |||
12086 | } | |||
12087 | ||||
12088 | static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index, | |||
12089 | SelectionDAG &DAG) { | |||
12090 | SDLoc dl(Load); | |||
12091 | MVT VT = Load->getSimpleValueType(0); | |||
12092 | MVT EVT = VT.getVectorElementType(); | |||
12093 | SDValue Addr = Load->getOperand(1); | |||
12094 | SDValue NewAddr = DAG.getNode( | |||
12095 | ISD::ADD, dl, Addr.getSimpleValueType(), Addr, | |||
12096 | DAG.getConstant(Index * EVT.getStoreSize(), Addr.getSimpleValueType())); | |||
12097 | ||||
12098 | SDValue NewLoad = | |||
12099 | DAG.getLoad(EVT, dl, Load->getChain(), NewAddr, | |||
12100 | DAG.getMachineFunction().getMachineMemOperand( | |||
12101 | Load->getMemOperand(), 0, EVT.getStoreSize())); | |||
12102 | return NewLoad; | |||
12103 | } | |||
12104 | ||||
12105 | // It is only safe to call this function if isINSERTPSMask is true for | |||
12106 | // this shufflevector mask. | |||
12107 | static SDValue getINSERTPS(ShuffleVectorSDNode *SVOp, SDLoc &dl, | |||
12108 | SelectionDAG &DAG) { | |||
12109 | // Generate an insertps instruction when inserting an f32 from memory onto a | |||
12110 | // v4f32 or when copying a member from one v4f32 to another. | |||
12111 | // We also use it for transferring i32 from one register to another, | |||
12112 | // since it simply copies the same bits. | |||
12113 | // If we're transferring an i32 from memory to a specific element in a | |||
12114 | // register, we output a generic DAG that will match the PINSRD | |||
12115 | // instruction. | |||
12116 | MVT VT = SVOp->getSimpleValueType(0); | |||
12117 | MVT EVT = VT.getVectorElementType(); | |||
12118 | SDValue V1 = SVOp->getOperand(0); | |||
12119 | SDValue V2 = SVOp->getOperand(1); | |||
12120 | auto Mask = SVOp->getMask(); | |||
12121 | assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&(((VT == MVT::v4f32 || VT == MVT::v4i32) && "unsupported vector type for insertps/pinsrd" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v4f32 || VT == MVT::v4i32) && \"unsupported vector type for insertps/pinsrd\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12122, __PRETTY_FUNCTION__)) | |||
12122 | "unsupported vector type for insertps/pinsrd")(((VT == MVT::v4f32 || VT == MVT::v4i32) && "unsupported vector type for insertps/pinsrd" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v4f32 || VT == MVT::v4i32) && \"unsupported vector type for insertps/pinsrd\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12122, __PRETTY_FUNCTION__)); | |||
12123 | ||||
12124 | auto FromV1Predicate = [](const int &i) { return i < 4 && i > -1; }; | |||
12125 | auto FromV2Predicate = [](const int &i) { return i >= 4; }; | |||
12126 | int FromV1 = std::count_if(Mask.begin(), Mask.end(), FromV1Predicate); | |||
12127 | ||||
12128 | SDValue From; | |||
12129 | SDValue To; | |||
12130 | unsigned DestIndex; | |||
12131 | if (FromV1 == 1) { | |||
12132 | From = V1; | |||
12133 | To = V2; | |||
12134 | DestIndex = std::find_if(Mask.begin(), Mask.end(), FromV1Predicate) - | |||
12135 | Mask.begin(); | |||
12136 | ||||
12137 | // If we have 1 element from each vector, we have to check if we're | |||
12138 | // changing V1's element's place. If so, we're done. Otherwise, we | |||
12139 | // should assume we're changing V2's element's place and behave | |||
12140 | // accordingly. | |||
12141 | int FromV2 = std::count_if(Mask.begin(), Mask.end(), FromV2Predicate); | |||
12142 | assert(DestIndex <= INT32_MAX && "truncated destination index")((DestIndex <= (2147483647) && "truncated destination index" ) ? static_cast<void> (0) : __assert_fail ("DestIndex <= (2147483647) && \"truncated destination index\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12142, __PRETTY_FUNCTION__)); | |||
12143 | if (FromV1 == FromV2 && | |||
12144 | static_cast<int>(DestIndex) == Mask[DestIndex] % 4) { | |||
12145 | From = V2; | |||
12146 | To = V1; | |||
12147 | DestIndex = | |||
12148 | std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin(); | |||
12149 | } | |||
12150 | } else { | |||
12151 | assert(std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 &&((std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 && "More than one element from V1 and from V2, or no elements from one " "of the vectors. This case should not have returned true from " "isINSERTPSMask") ? static_cast<void> (0) : __assert_fail ("std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 && \"More than one element from V1 and from V2, or no elements from one \" \"of the vectors. This case should not have returned true from \" \"isINSERTPSMask\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12154, __PRETTY_FUNCTION__)) | |||
12152 | "More than one element from V1 and from V2, or no elements from one "((std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 && "More than one element from V1 and from V2, or no elements from one " "of the vectors. This case should not have returned true from " "isINSERTPSMask") ? static_cast<void> (0) : __assert_fail ("std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 && \"More than one element from V1 and from V2, or no elements from one \" \"of the vectors. This case should not have returned true from \" \"isINSERTPSMask\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12154, __PRETTY_FUNCTION__)) | |||
12153 | "of the vectors. This case should not have returned true from "((std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 && "More than one element from V1 and from V2, or no elements from one " "of the vectors. This case should not have returned true from " "isINSERTPSMask") ? static_cast<void> (0) : __assert_fail ("std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 && \"More than one element from V1 and from V2, or no elements from one \" \"of the vectors. This case should not have returned true from \" \"isINSERTPSMask\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12154, __PRETTY_FUNCTION__)) | |||
12154 | "isINSERTPSMask")((std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 && "More than one element from V1 and from V2, or no elements from one " "of the vectors. This case should not have returned true from " "isINSERTPSMask") ? static_cast<void> (0) : __assert_fail ("std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 && \"More than one element from V1 and from V2, or no elements from one \" \"of the vectors. This case should not have returned true from \" \"isINSERTPSMask\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12154, __PRETTY_FUNCTION__)); | |||
12155 | From = V2; | |||
12156 | To = V1; | |||
12157 | DestIndex = | |||
12158 | std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin(); | |||
12159 | } | |||
12160 | ||||
12161 | // Get an index into the source vector in the range [0,4) (the mask is | |||
12162 | // in the range [0,8) because it can address V1 and V2) | |||
12163 | unsigned SrcIndex = Mask[DestIndex] % 4; | |||
12164 | if (MayFoldLoad(From)) { | |||
12165 | // Trivial case, when From comes from a load and is only used by the | |||
12166 | // shuffle. Make it use insertps from the vector that we need from that | |||
12167 | // load. | |||
12168 | SDValue NewLoad = | |||
12169 | NarrowVectorLoadToElement(cast<LoadSDNode>(From), SrcIndex, DAG); | |||
12170 | if (!NewLoad.getNode()) | |||
12171 | return SDValue(); | |||
12172 | ||||
12173 | if (EVT == MVT::f32) { | |||
12174 | // Create this as a scalar to vector to match the instruction pattern. | |||
12175 | SDValue LoadScalarToVector = | |||
12176 | DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, NewLoad); | |||
12177 | SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4); | |||
12178 | return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, LoadScalarToVector, | |||
12179 | InsertpsMask); | |||
12180 | } else { // EVT == MVT::i32 | |||
12181 | // If we're getting an i32 from memory, use an INSERT_VECTOR_ELT | |||
12182 | // instruction, to match the PINSRD instruction, which loads an i32 to a | |||
12183 | // certain vector element. | |||
12184 | return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, To, NewLoad, | |||
12185 | DAG.getConstant(DestIndex, MVT::i32)); | |||
12186 | } | |||
12187 | } | |||
12188 | ||||
12189 | // Vector-element-to-vector | |||
12190 | SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4 | SrcIndex << 6); | |||
12191 | return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, From, InsertpsMask); | |||
12192 | } | |||
12193 | ||||
12194 | // Reduce a vector shuffle to zext. | |||
12195 | static SDValue LowerVectorIntExtend(SDValue Op, const X86Subtarget *Subtarget, | |||
12196 | SelectionDAG &DAG) { | |||
12197 | // PMOVZX is only available from SSE41. | |||
12198 | if (!Subtarget->hasSSE41()) | |||
12199 | return SDValue(); | |||
12200 | ||||
12201 | MVT VT = Op.getSimpleValueType(); | |||
12202 | ||||
12203 | // Only AVX2 support 256-bit vector integer extending. | |||
12204 | if (!Subtarget->hasInt256() && VT.is256BitVector()) | |||
12205 | return SDValue(); | |||
12206 | ||||
12207 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
12208 | SDLoc DL(Op); | |||
12209 | SDValue V1 = Op.getOperand(0); | |||
12210 | SDValue V2 = Op.getOperand(1); | |||
12211 | unsigned NumElems = VT.getVectorNumElements(); | |||
12212 | ||||
12213 | // Extending is an unary operation and the element type of the source vector | |||
12214 | // won't be equal to or larger than i64. | |||
12215 | if (V2.getOpcode() != ISD::UNDEF || !VT.isInteger() || | |||
12216 | VT.getVectorElementType() == MVT::i64) | |||
12217 | return SDValue(); | |||
12218 | ||||
12219 | // Find the expansion ratio, e.g. expanding from i8 to i32 has a ratio of 4. | |||
12220 | unsigned Shift = 1; // Start from 2, i.e. 1 << 1. | |||
12221 | while ((1U << Shift) < NumElems) { | |||
12222 | if (SVOp->getMaskElt(1U << Shift) == 1) | |||
12223 | break; | |||
12224 | Shift += 1; | |||
12225 | // The maximal ratio is 8, i.e. from i8 to i64. | |||
12226 | if (Shift > 3) | |||
12227 | return SDValue(); | |||
12228 | } | |||
12229 | ||||
12230 | // Check the shuffle mask. | |||
12231 | unsigned Mask = (1U << Shift) - 1; | |||
12232 | for (unsigned i = 0; i != NumElems; ++i) { | |||
12233 | int EltIdx = SVOp->getMaskElt(i); | |||
12234 | if ((i & Mask) != 0 && EltIdx != -1) | |||
12235 | return SDValue(); | |||
12236 | if ((i & Mask) == 0 && (unsigned)EltIdx != (i >> Shift)) | |||
12237 | return SDValue(); | |||
12238 | } | |||
12239 | ||||
12240 | unsigned NBits = VT.getVectorElementType().getSizeInBits() << Shift; | |||
12241 | MVT NeVT = MVT::getIntegerVT(NBits); | |||
12242 | MVT NVT = MVT::getVectorVT(NeVT, NumElems >> Shift); | |||
12243 | ||||
12244 | if (!DAG.getTargetLoweringInfo().isTypeLegal(NVT)) | |||
12245 | return SDValue(); | |||
12246 | ||||
12247 | return DAG.getNode(ISD::BITCAST, DL, VT, | |||
12248 | DAG.getNode(X86ISD::VZEXT, DL, NVT, V1)); | |||
12249 | } | |||
12250 | ||||
12251 | static SDValue NormalizeVectorShuffle(SDValue Op, const X86Subtarget *Subtarget, | |||
12252 | SelectionDAG &DAG) { | |||
12253 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
12254 | MVT VT = Op.getSimpleValueType(); | |||
12255 | SDLoc dl(Op); | |||
12256 | SDValue V1 = Op.getOperand(0); | |||
12257 | SDValue V2 = Op.getOperand(1); | |||
12258 | ||||
12259 | if (isZeroShuffle(SVOp)) | |||
12260 | return getZeroVector(VT, Subtarget, DAG, dl); | |||
12261 | ||||
12262 | // Handle splat operations | |||
12263 | if (SVOp->isSplat()) { | |||
12264 | // Use vbroadcast whenever the splat comes from a foldable load | |||
12265 | SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG); | |||
12266 | if (Broadcast.getNode()) | |||
12267 | return Broadcast; | |||
12268 | } | |||
12269 | ||||
12270 | // Check integer expanding shuffles. | |||
12271 | SDValue NewOp = LowerVectorIntExtend(Op, Subtarget, DAG); | |||
12272 | if (NewOp.getNode()) | |||
12273 | return NewOp; | |||
12274 | ||||
12275 | // If the shuffle can be profitably rewritten as a narrower shuffle, then | |||
12276 | // do it! | |||
12277 | if (VT == MVT::v8i16 || VT == MVT::v16i8 || VT == MVT::v16i16 || | |||
12278 | VT == MVT::v32i8) { | |||
12279 | SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG); | |||
12280 | if (NewOp.getNode()) | |||
12281 | return DAG.getNode(ISD::BITCAST, dl, VT, NewOp); | |||
12282 | } else if (VT.is128BitVector() && Subtarget->hasSSE2()) { | |||
12283 | // FIXME: Figure out a cleaner way to do this. | |||
12284 | if (ISD::isBuildVectorAllZeros(V2.getNode())) { | |||
12285 | SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG); | |||
12286 | if (NewOp.getNode()) { | |||
12287 | MVT NewVT = NewOp.getSimpleValueType(); | |||
12288 | if (isCommutedMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(), | |||
12289 | NewVT, true, false)) | |||
12290 | return getVZextMovL(VT, NewVT, NewOp.getOperand(0), DAG, Subtarget, | |||
12291 | dl); | |||
12292 | } | |||
12293 | } else if (ISD::isBuildVectorAllZeros(V1.getNode())) { | |||
12294 | SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG); | |||
12295 | if (NewOp.getNode()) { | |||
12296 | MVT NewVT = NewOp.getSimpleValueType(); | |||
12297 | if (isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(), NewVT)) | |||
12298 | return getVZextMovL(VT, NewVT, NewOp.getOperand(1), DAG, Subtarget, | |||
12299 | dl); | |||
12300 | } | |||
12301 | } | |||
12302 | } | |||
12303 | return SDValue(); | |||
12304 | } | |||
12305 | ||||
12306 | SDValue | |||
12307 | X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const { | |||
12308 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
12309 | SDValue V1 = Op.getOperand(0); | |||
12310 | SDValue V2 = Op.getOperand(1); | |||
12311 | MVT VT = Op.getSimpleValueType(); | |||
12312 | SDLoc dl(Op); | |||
12313 | unsigned NumElems = VT.getVectorNumElements(); | |||
12314 | bool V1IsUndef = V1.getOpcode() == ISD::UNDEF; | |||
12315 | bool V2IsUndef = V2.getOpcode() == ISD::UNDEF; | |||
12316 | bool V1IsSplat = false; | |||
12317 | bool V2IsSplat = false; | |||
12318 | bool HasSSE2 = Subtarget->hasSSE2(); | |||
12319 | bool HasFp256 = Subtarget->hasFp256(); | |||
12320 | bool HasInt256 = Subtarget->hasInt256(); | |||
12321 | MachineFunction &MF = DAG.getMachineFunction(); | |||
12322 | bool OptForSize = MF.getFunction()->getAttributes(). | |||
12323 | hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize); | |||
12324 | ||||
12325 | // Check if we should use the experimental vector shuffle lowering. If so, | |||
12326 | // delegate completely to that code path. | |||
12327 | if (ExperimentalVectorShuffleLowering) | |||
12328 | return lowerVectorShuffle(Op, Subtarget, DAG); | |||
12329 | ||||
12330 | assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles")((VT.getSizeInBits() != 64 && "Can't lower MMX shuffles" ) ? static_cast<void> (0) : __assert_fail ("VT.getSizeInBits() != 64 && \"Can't lower MMX shuffles\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12330, __PRETTY_FUNCTION__)); | |||
12331 | ||||
12332 | if (V1IsUndef && V2IsUndef) | |||
12333 | return DAG.getUNDEF(VT); | |||
12334 | ||||
12335 | // When we create a shuffle node we put the UNDEF node to second operand, | |||
12336 | // but in some cases the first operand may be transformed to UNDEF. | |||
12337 | // In this case we should just commute the node. | |||
12338 | if (V1IsUndef) | |||
12339 | return DAG.getCommutedVectorShuffle(*SVOp); | |||
12340 | ||||
12341 | // Vector shuffle lowering takes 3 steps: | |||
12342 | // | |||
12343 | // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable | |||
12344 | // narrowing and commutation of operands should be handled. | |||
12345 | // 2) Matching of shuffles with known shuffle masks to x86 target specific | |||
12346 | // shuffle nodes. | |||
12347 | // 3) Rewriting of unmatched masks into new generic shuffle operations, | |||
12348 | // so the shuffle can be broken into other shuffles and the legalizer can | |||
12349 | // try the lowering again. | |||
12350 | // | |||
12351 | // The general idea is that no vector_shuffle operation should be left to | |||
12352 | // be matched during isel, all of them must be converted to a target specific | |||
12353 | // node here. | |||
12354 | ||||
12355 | // Normalize the input vectors. Here splats, zeroed vectors, profitable | |||
12356 | // narrowing and commutation of operands should be handled. The actual code | |||
12357 | // doesn't include all of those, work in progress... | |||
12358 | SDValue NewOp = NormalizeVectorShuffle(Op, Subtarget, DAG); | |||
12359 | if (NewOp.getNode()) | |||
12360 | return NewOp; | |||
12361 | ||||
12362 | SmallVector<int, 8> M(SVOp->getMask().begin(), SVOp->getMask().end()); | |||
12363 | ||||
12364 | // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and | |||
12365 | // unpckh_undef). Only use pshufd if speed is more important than size. | |||
12366 | if (OptForSize && isUNPCKL_v_undef_Mask(M, VT, HasInt256)) | |||
12367 | return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG); | |||
12368 | if (OptForSize && isUNPCKH_v_undef_Mask(M, VT, HasInt256)) | |||
12369 | return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG); | |||
12370 | ||||
12371 | if (isMOVDDUPMask(M, VT) && Subtarget->hasSSE3() && | |||
12372 | V2IsUndef && MayFoldVectorLoad(V1)) | |||
12373 | return getMOVDDup(Op, dl, V1, DAG); | |||
12374 | ||||
12375 | if (isMOVHLPS_v_undef_Mask(M, VT)) | |||
12376 | return getMOVHighToLow(Op, dl, DAG); | |||
12377 | ||||
12378 | // Use to match splats | |||
12379 | if (HasSSE2 && isUNPCKHMask(M, VT, HasInt256) && V2IsUndef && | |||
12380 | (VT == MVT::v2f64 || VT == MVT::v2i64)) | |||
12381 | return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG); | |||
12382 | ||||
12383 | if (isPSHUFDMask(M, VT)) { | |||
12384 | // The actual implementation will match the mask in the if above and then | |||
12385 | // during isel it can match several different instructions, not only pshufd | |||
12386 | // as its name says, sad but true, emulate the behavior for now... | |||
12387 | if (isMOVDDUPMask(M, VT) && ((VT == MVT::v4f32 || VT == MVT::v2i64))) | |||
12388 | return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG); | |||
12389 | ||||
12390 | unsigned TargetMask = getShuffleSHUFImmediate(SVOp); | |||
12391 | ||||
12392 | if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32)) | |||
12393 | return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG); | |||
12394 | ||||
12395 | if (HasFp256 && (VT == MVT::v4f32 || VT == MVT::v2f64)) | |||
12396 | return getTargetShuffleNode(X86ISD::VPERMILPI, dl, VT, V1, TargetMask, | |||
12397 | DAG); | |||
12398 | ||||
12399 | return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V1, | |||
12400 | TargetMask, DAG); | |||
12401 | } | |||
12402 | ||||
12403 | if (isPALIGNRMask(M, VT, Subtarget)) | |||
12404 | return getTargetShuffleNode(X86ISD::PALIGNR, dl, VT, V1, V2, | |||
12405 | getShufflePALIGNRImmediate(SVOp), | |||
12406 | DAG); | |||
12407 | ||||
12408 | if (isVALIGNMask(M, VT, Subtarget)) | |||
12409 | return getTargetShuffleNode(X86ISD::VALIGN, dl, VT, V1, V2, | |||
12410 | getShuffleVALIGNImmediate(SVOp), | |||
12411 | DAG); | |||
12412 | ||||
12413 | // Check if this can be converted into a logical shift. | |||
12414 | bool isLeft = false; | |||
12415 | unsigned ShAmt = 0; | |||
12416 | SDValue ShVal; | |||
12417 | bool isShift = HasSSE2 && isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt); | |||
12418 | if (isShift && ShVal.hasOneUse()) { | |||
12419 | // If the shifted value has multiple uses, it may be cheaper to use | |||
12420 | // v_set0 + movlhps or movhlps, etc. | |||
12421 | MVT EltVT = VT.getVectorElementType(); | |||
12422 | ShAmt *= EltVT.getSizeInBits(); | |||
12423 | return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl); | |||
12424 | } | |||
12425 | ||||
12426 | if (isMOVLMask(M, VT)) { | |||
12427 | if (ISD::isBuildVectorAllZeros(V1.getNode())) | |||
12428 | return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl); | |||
12429 | if (!isMOVLPMask(M, VT)) { | |||
12430 | if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64)) | |||
12431 | return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG); | |||
12432 | ||||
12433 | if (VT == MVT::v4i32 || VT == MVT::v4f32) | |||
12434 | return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG); | |||
12435 | } | |||
12436 | } | |||
12437 | ||||
12438 | // FIXME: fold these into legal mask. | |||
12439 | if (isMOVLHPSMask(M, VT) && !isUNPCKLMask(M, VT, HasInt256)) | |||
12440 | return getMOVLowToHigh(Op, dl, DAG, HasSSE2); | |||
12441 | ||||
12442 | if (isMOVHLPSMask(M, VT)) | |||
12443 | return getMOVHighToLow(Op, dl, DAG); | |||
12444 | ||||
12445 | if (V2IsUndef && isMOVSHDUPMask(M, VT, Subtarget)) | |||
12446 | return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG); | |||
12447 | ||||
12448 | if (V2IsUndef && isMOVSLDUPMask(M, VT, Subtarget)) | |||
12449 | return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG); | |||
12450 | ||||
12451 | if (isMOVLPMask(M, VT)) | |||
12452 | return getMOVLP(Op, dl, DAG, HasSSE2); | |||
12453 | ||||
12454 | if (ShouldXformToMOVHLPS(M, VT) || | |||
12455 | ShouldXformToMOVLP(V1.getNode(), V2.getNode(), M, VT)) | |||
12456 | return DAG.getCommutedVectorShuffle(*SVOp); | |||
12457 | ||||
12458 | if (isShift) { | |||
12459 | // No better options. Use a vshldq / vsrldq. | |||
12460 | MVT EltVT = VT.getVectorElementType(); | |||
12461 | ShAmt *= EltVT.getSizeInBits(); | |||
12462 | return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl); | |||
12463 | } | |||
12464 | ||||
12465 | bool Commuted = false; | |||
12466 | // FIXME: This should also accept a bitcast of a splat? Be careful, not | |||
12467 | // 1,1,1,1 -> v8i16 though. | |||
12468 | BitVector UndefElements; | |||
12469 | if (auto *BVOp = dyn_cast<BuildVectorSDNode>(V1.getNode())) | |||
12470 | if (BVOp->getConstantSplatNode(&UndefElements) && UndefElements.none()) | |||
12471 | V1IsSplat = true; | |||
12472 | if (auto *BVOp = dyn_cast<BuildVectorSDNode>(V2.getNode())) | |||
12473 | if (BVOp->getConstantSplatNode(&UndefElements) && UndefElements.none()) | |||
12474 | V2IsSplat = true; | |||
12475 | ||||
12476 | // Canonicalize the splat or undef, if present, to be on the RHS. | |||
12477 | if (!V2IsUndef && V1IsSplat && !V2IsSplat) { | |||
12478 | CommuteVectorShuffleMask(M, NumElems); | |||
12479 | std::swap(V1, V2); | |||
12480 | std::swap(V1IsSplat, V2IsSplat); | |||
12481 | Commuted = true; | |||
12482 | } | |||
12483 | ||||
12484 | if (isCommutedMOVLMask(M, VT, V2IsSplat, V2IsUndef)) { | |||
12485 | // Shuffling low element of v1 into undef, just return v1. | |||
12486 | if (V2IsUndef) | |||
12487 | return V1; | |||
12488 | // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which | |||
12489 | // the instruction selector will not match, so get a canonical MOVL with | |||
12490 | // swapped operands to undo the commute. | |||
12491 | return getMOVL(DAG, dl, VT, V2, V1); | |||
12492 | } | |||
12493 | ||||
12494 | if (isUNPCKLMask(M, VT, HasInt256)) | |||
12495 | return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG); | |||
12496 | ||||
12497 | if (isUNPCKHMask(M, VT, HasInt256)) | |||
12498 | return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG); | |||
12499 | ||||
12500 | if (V2IsSplat) { | |||
12501 | // Normalize mask so all entries that point to V2 points to its first | |||
12502 | // element then try to match unpck{h|l} again. If match, return a | |||
12503 | // new vector_shuffle with the corrected mask.p | |||
12504 | SmallVector<int, 8> NewMask(M.begin(), M.end()); | |||
12505 | NormalizeMask(NewMask, NumElems); | |||
12506 | if (isUNPCKLMask(NewMask, VT, HasInt256, true)) | |||
12507 | return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG); | |||
12508 | if (isUNPCKHMask(NewMask, VT, HasInt256, true)) | |||
12509 | return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG); | |||
12510 | } | |||
12511 | ||||
12512 | if (Commuted) { | |||
12513 | // Commute is back and try unpck* again. | |||
12514 | // FIXME: this seems wrong. | |||
12515 | CommuteVectorShuffleMask(M, NumElems); | |||
12516 | std::swap(V1, V2); | |||
12517 | std::swap(V1IsSplat, V2IsSplat); | |||
12518 | ||||
12519 | if (isUNPCKLMask(M, VT, HasInt256)) | |||
12520 | return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG); | |||
12521 | ||||
12522 | if (isUNPCKHMask(M, VT, HasInt256)) | |||
12523 | return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG); | |||
12524 | } | |||
12525 | ||||
12526 | // Normalize the node to match x86 shuffle ops if needed | |||
12527 | if (!V2IsUndef && (isSHUFPMask(M, VT, /* Commuted */ true))) | |||
12528 | return DAG.getCommutedVectorShuffle(*SVOp); | |||
12529 | ||||
12530 | // The checks below are all present in isShuffleMaskLegal, but they are | |||
12531 | // inlined here right now to enable us to directly emit target specific | |||
12532 | // nodes, and remove one by one until they don't return Op anymore. | |||
12533 | ||||
12534 | if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) && | |||
12535 | SVOp->getSplatIndex() == 0 && V2IsUndef) { | |||
12536 | if (VT == MVT::v2f64 || VT == MVT::v2i64) | |||
12537 | return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG); | |||
12538 | } | |||
12539 | ||||
12540 | if (isPSHUFHWMask(M, VT, HasInt256)) | |||
12541 | return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1, | |||
12542 | getShufflePSHUFHWImmediate(SVOp), | |||
12543 | DAG); | |||
12544 | ||||
12545 | if (isPSHUFLWMask(M, VT, HasInt256)) | |||
12546 | return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1, | |||
12547 | getShufflePSHUFLWImmediate(SVOp), | |||
12548 | DAG); | |||
12549 | ||||
12550 | unsigned MaskValue; | |||
12551 | if (isBlendMask(M, VT, Subtarget->hasSSE41(), Subtarget->hasInt256(), | |||
12552 | &MaskValue)) | |||
12553 | return LowerVECTOR_SHUFFLEtoBlend(SVOp, MaskValue, Subtarget, DAG); | |||
12554 | ||||
12555 | if (isSHUFPMask(M, VT)) | |||
12556 | return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V2, | |||
12557 | getShuffleSHUFImmediate(SVOp), DAG); | |||
12558 | ||||
12559 | if (isUNPCKL_v_undef_Mask(M, VT, HasInt256)) | |||
12560 | return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG); | |||
12561 | if (isUNPCKH_v_undef_Mask(M, VT, HasInt256)) | |||
12562 | return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG); | |||
12563 | ||||
12564 | //===--------------------------------------------------------------------===// | |||
12565 | // Generate target specific nodes for 128 or 256-bit shuffles only | |||
12566 | // supported in the AVX instruction set. | |||
12567 | // | |||
12568 | ||||
12569 | // Handle VMOVDDUPY permutations | |||
12570 | if (V2IsUndef && isMOVDDUPYMask(M, VT, HasFp256)) | |||
12571 | return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG); | |||
12572 | ||||
12573 | // Handle VPERMILPS/D* permutations | |||
12574 | if (isVPERMILPMask(M, VT)) { | |||
12575 | if ((HasInt256 && VT == MVT::v8i32) || VT == MVT::v16i32) | |||
12576 | return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, | |||
12577 | getShuffleSHUFImmediate(SVOp), DAG); | |||
12578 | return getTargetShuffleNode(X86ISD::VPERMILPI, dl, VT, V1, | |||
12579 | getShuffleSHUFImmediate(SVOp), DAG); | |||
12580 | } | |||
12581 | ||||
12582 | unsigned Idx; | |||
12583 | if (VT.is512BitVector() && isINSERT64x4Mask(M, VT, &Idx)) | |||
12584 | return Insert256BitVector(V1, Extract256BitVector(V2, 0, DAG, dl), | |||
12585 | Idx*(NumElems/2), DAG, dl); | |||
12586 | ||||
12587 | // Handle VPERM2F128/VPERM2I128 permutations | |||
12588 | if (isVPERM2X128Mask(M, VT, HasFp256)) | |||
12589 | return getTargetShuffleNode(X86ISD::VPERM2X128, dl, VT, V1, | |||
12590 | V2, getShuffleVPERM2X128Immediate(SVOp), DAG); | |||
12591 | ||||
12592 | if (Subtarget->hasSSE41() && isINSERTPSMask(M, VT)) | |||
12593 | return getINSERTPS(SVOp, dl, DAG); | |||
12594 | ||||
12595 | unsigned Imm8; | |||
12596 | if (V2IsUndef && HasInt256 && isPermImmMask(M, VT, Imm8)) | |||
12597 | return getTargetShuffleNode(X86ISD::VPERMI, dl, VT, V1, Imm8, DAG); | |||
12598 | ||||
12599 | if ((V2IsUndef && HasInt256 && VT.is256BitVector() && NumElems == 8) || | |||
12600 | VT.is512BitVector()) { | |||
12601 | MVT MaskEltVT = MVT::getIntegerVT(VT.getVectorElementType().getSizeInBits()); | |||
12602 | MVT MaskVectorVT = MVT::getVectorVT(MaskEltVT, NumElems); | |||
12603 | SmallVector<SDValue, 16> permclMask; | |||
12604 | for (unsigned i = 0; i != NumElems; ++i) { | |||
12605 | permclMask.push_back(DAG.getConstant((M[i]>=0) ? M[i] : 0, MaskEltVT)); | |||
12606 | } | |||
12607 | ||||
12608 | SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVectorVT, permclMask); | |||
12609 | if (V2IsUndef) | |||
12610 | // Bitcast is for VPERMPS since mask is v8i32 but node takes v8f32 | |||
12611 | return DAG.getNode(X86ISD::VPERMV, dl, VT, | |||
12612 | DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1); | |||
12613 | return DAG.getNode(X86ISD::VPERMV3, dl, VT, V1, | |||
12614 | DAG.getNode(ISD::BITCAST, dl, VT, Mask), V2); | |||
12615 | } | |||
12616 | ||||
12617 | //===--------------------------------------------------------------------===// | |||
12618 | // Since no target specific shuffle was selected for this generic one, | |||
12619 | // lower it into other known shuffles. FIXME: this isn't true yet, but | |||
12620 | // this is the plan. | |||
12621 | // | |||
12622 | ||||
12623 | // Handle v8i16 specifically since SSE can do byte extraction and insertion. | |||
12624 | if (VT == MVT::v8i16) { | |||
12625 | SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, Subtarget, DAG); | |||
12626 | if (NewOp.getNode()) | |||
12627 | return NewOp; | |||
12628 | } | |||
12629 | ||||
12630 | if (VT == MVT::v16i16 && Subtarget->hasInt256()) { | |||
12631 | SDValue NewOp = LowerVECTOR_SHUFFLEv16i16(Op, DAG); | |||
12632 | if (NewOp.getNode()) | |||
12633 | return NewOp; | |||
12634 | } | |||
12635 | ||||
12636 | if (VT == MVT::v16i8) { | |||
12637 | SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, Subtarget, DAG); | |||
12638 | if (NewOp.getNode()) | |||
12639 | return NewOp; | |||
12640 | } | |||
12641 | ||||
12642 | if (VT == MVT::v32i8) { | |||
12643 | SDValue NewOp = LowerVECTOR_SHUFFLEv32i8(SVOp, Subtarget, DAG); | |||
12644 | if (NewOp.getNode()) | |||
12645 | return NewOp; | |||
12646 | } | |||
12647 | ||||
12648 | // Handle all 128-bit wide vectors with 4 elements, and match them with | |||
12649 | // several different shuffle types. | |||
12650 | if (NumElems == 4 && VT.is128BitVector()) | |||
12651 | return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG); | |||
12652 | ||||
12653 | // Handle general 256-bit shuffles | |||
12654 | if (VT.is256BitVector()) | |||
12655 | return LowerVECTOR_SHUFFLE_256(SVOp, DAG); | |||
12656 | ||||
12657 | return SDValue(); | |||
12658 | } | |||
12659 | ||||
12660 | // This function assumes its argument is a BUILD_VECTOR of constants or | |||
12661 | // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is | |||
12662 | // true. | |||
12663 | static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector, | |||
12664 | unsigned &MaskValue) { | |||
12665 | MaskValue = 0; | |||
12666 | unsigned NumElems = BuildVector->getNumOperands(); | |||
12667 | // There are 2 lanes if (NumElems > 8), and 1 lane otherwise. | |||
12668 | unsigned NumLanes = (NumElems - 1) / 8 + 1; | |||
12669 | unsigned NumElemsInLane = NumElems / NumLanes; | |||
12670 | ||||
12671 | // Blend for v16i16 should be symetric for the both lanes. | |||
12672 | for (unsigned i = 0; i < NumElemsInLane; ++i) { | |||
12673 | SDValue EltCond = BuildVector->getOperand(i); | |||
12674 | SDValue SndLaneEltCond = | |||
12675 | (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond; | |||
12676 | ||||
12677 | int Lane1Cond = -1, Lane2Cond = -1; | |||
12678 | if (isa<ConstantSDNode>(EltCond)) | |||
12679 | Lane1Cond = !isZero(EltCond); | |||
12680 | if (isa<ConstantSDNode>(SndLaneEltCond)) | |||
12681 | Lane2Cond = !isZero(SndLaneEltCond); | |||
12682 | ||||
12683 | if (Lane1Cond == Lane2Cond || Lane2Cond < 0) | |||
12684 | // Lane1Cond != 0, means we want the first argument. | |||
12685 | // Lane1Cond == 0, means we want the second argument. | |||
12686 | // The encoding of this argument is 0 for the first argument, 1 | |||
12687 | // for the second. Therefore, invert the condition. | |||
12688 | MaskValue |= !Lane1Cond << i; | |||
12689 | else if (Lane1Cond < 0) | |||
12690 | MaskValue |= !Lane2Cond << i; | |||
12691 | else | |||
12692 | return false; | |||
12693 | } | |||
12694 | return true; | |||
12695 | } | |||
12696 | ||||
12697 | /// \brief Try to lower a VSELECT instruction to an immediate-controlled blend | |||
12698 | /// instruction. | |||
12699 | static SDValue lowerVSELECTtoBLENDI(SDValue Op, const X86Subtarget *Subtarget, | |||
12700 | SelectionDAG &DAG) { | |||
12701 | SDValue Cond = Op.getOperand(0); | |||
12702 | SDValue LHS = Op.getOperand(1); | |||
12703 | SDValue RHS = Op.getOperand(2); | |||
12704 | SDLoc dl(Op); | |||
12705 | MVT VT = Op.getSimpleValueType(); | |||
12706 | MVT EltVT = VT.getVectorElementType(); | |||
12707 | unsigned NumElems = VT.getVectorNumElements(); | |||
12708 | ||||
12709 | // There is no blend with immediate in AVX-512. | |||
12710 | if (VT.is512BitVector()) | |||
12711 | return SDValue(); | |||
12712 | ||||
12713 | if (!Subtarget->hasSSE41() || EltVT == MVT::i8) | |||
12714 | return SDValue(); | |||
12715 | if (!Subtarget->hasInt256() && VT == MVT::v16i16) | |||
12716 | return SDValue(); | |||
12717 | ||||
12718 | if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) | |||
12719 | return SDValue(); | |||
12720 | ||||
12721 | // Check the mask for BLEND and build the value. | |||
12722 | unsigned MaskValue = 0; | |||
12723 | if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue)) | |||
12724 | return SDValue(); | |||
12725 | ||||
12726 | // Convert i32 vectors to floating point if it is not AVX2. | |||
12727 | // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors. | |||
12728 | MVT BlendVT = VT; | |||
12729 | if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) { | |||
12730 | BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()), | |||
12731 | NumElems); | |||
12732 | LHS = DAG.getNode(ISD::BITCAST, dl, VT, LHS); | |||
12733 | RHS = DAG.getNode(ISD::BITCAST, dl, VT, RHS); | |||
12734 | } | |||
12735 | ||||
12736 | SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, LHS, RHS, | |||
12737 | DAG.getConstant(MaskValue, MVT::i32)); | |||
12738 | return DAG.getNode(ISD::BITCAST, dl, VT, Ret); | |||
12739 | } | |||
12740 | ||||
12741 | SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const { | |||
12742 | // A vselect where all conditions and data are constants can be optimized into | |||
12743 | // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR(). | |||
12744 | if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) && | |||
12745 | ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) && | |||
12746 | ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode())) | |||
12747 | return SDValue(); | |||
12748 | ||||
12749 | SDValue BlendOp = lowerVSELECTtoBLENDI(Op, Subtarget, DAG); | |||
12750 | if (BlendOp.getNode()) | |||
12751 | return BlendOp; | |||
12752 | ||||
12753 | // Some types for vselect were previously set to Expand, not Legal or | |||
12754 | // Custom. Return an empty SDValue so we fall-through to Expand, after | |||
12755 | // the Custom lowering phase. | |||
12756 | MVT VT = Op.getSimpleValueType(); | |||
12757 | switch (VT.SimpleTy) { | |||
12758 | default: | |||
12759 | break; | |||
12760 | case MVT::v8i16: | |||
12761 | case MVT::v16i16: | |||
12762 | if (Subtarget->hasBWI() && Subtarget->hasVLX()) | |||
12763 | break; | |||
12764 | return SDValue(); | |||
12765 | } | |||
12766 | ||||
12767 | // We couldn't create a "Blend with immediate" node. | |||
12768 | // This node should still be legal, but we'll have to emit a blendv* | |||
12769 | // instruction. | |||
12770 | return Op; | |||
12771 | } | |||
12772 | ||||
12773 | static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) { | |||
12774 | MVT VT = Op.getSimpleValueType(); | |||
12775 | SDLoc dl(Op); | |||
12776 | ||||
12777 | if (!Op.getOperand(0).getSimpleValueType().is128BitVector()) | |||
12778 | return SDValue(); | |||
12779 | ||||
12780 | if (VT.getSizeInBits() == 8) { | |||
12781 | SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32, | |||
12782 | Op.getOperand(0), Op.getOperand(1)); | |||
12783 | SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract, | |||
12784 | DAG.getValueType(VT)); | |||
12785 | return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert); | |||
12786 | } | |||
12787 | ||||
12788 | if (VT.getSizeInBits() == 16) { | |||
12789 | unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); | |||
12790 | // If Idx is 0, it's cheaper to do a move instead of a pextrw. | |||
12791 | if (Idx == 0) | |||
12792 | return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, | |||
12793 | DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, | |||
12794 | DAG.getNode(ISD::BITCAST, dl, | |||
12795 | MVT::v4i32, | |||
12796 | Op.getOperand(0)), | |||
12797 | Op.getOperand(1))); | |||
12798 | SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32, | |||
12799 | Op.getOperand(0), Op.getOperand(1)); | |||
12800 | SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract, | |||
12801 | DAG.getValueType(VT)); | |||
12802 | return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert); | |||
12803 | } | |||
12804 | ||||
12805 | if (VT == MVT::f32) { | |||
12806 | // EXTRACTPS outputs to a GPR32 register which will require a movd to copy | |||
12807 | // the result back to FR32 register. It's only worth matching if the | |||
12808 | // result has a single use which is a store or a bitcast to i32. And in | |||
12809 | // the case of a store, it's not worth it if the index is a constant 0, | |||
12810 | // because a MOVSSmr can be used instead, which is smaller and faster. | |||
12811 | if (!Op.hasOneUse()) | |||
12812 | return SDValue(); | |||
12813 | SDNode *User = *Op.getNode()->use_begin(); | |||
12814 | if ((User->getOpcode() != ISD::STORE || | |||
12815 | (isa<ConstantSDNode>(Op.getOperand(1)) && | |||
12816 | cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) && | |||
12817 | (User->getOpcode() != ISD::BITCAST || | |||
12818 | User->getValueType(0) != MVT::i32)) | |||
12819 | return SDValue(); | |||
12820 | SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, | |||
12821 | DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, | |||
12822 | Op.getOperand(0)), | |||
12823 | Op.getOperand(1)); | |||
12824 | return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract); | |||
12825 | } | |||
12826 | ||||
12827 | if (VT == MVT::i32 || VT == MVT::i64) { | |||
12828 | // ExtractPS/pextrq works with constant index. | |||
12829 | if (isa<ConstantSDNode>(Op.getOperand(1))) | |||
12830 | return Op; | |||
12831 | } | |||
12832 | return SDValue(); | |||
12833 | } | |||
12834 | ||||
12835 | /// Extract one bit from mask vector, like v16i1 or v8i1. | |||
12836 | /// AVX-512 feature. | |||
12837 | SDValue | |||
12838 | X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const { | |||
12839 | SDValue Vec = Op.getOperand(0); | |||
12840 | SDLoc dl(Vec); | |||
12841 | MVT VecVT = Vec.getSimpleValueType(); | |||
12842 | SDValue Idx = Op.getOperand(1); | |||
12843 | MVT EltVT = Op.getSimpleValueType(); | |||
12844 | ||||
12845 | assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector")(((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector" ) ? static_cast<void> (0) : __assert_fail ("(EltVT == MVT::i1) && \"Unexpected operands in ExtractBitFromMaskVector\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12845, __PRETTY_FUNCTION__)); | |||
12846 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12847, __PRETTY_FUNCTION__)) | |||
12847 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12847, __PRETTY_FUNCTION__)); | |||
12848 | ||||
12849 | // variable index can't be handled in mask registers, | |||
12850 | // extend vector to VR512 | |||
12851 | if (!isa<ConstantSDNode>(Idx)) { | |||
12852 | MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32); | |||
12853 | SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec); | |||
12854 | SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, | |||
12855 | ExtVT.getVectorElementType(), Ext, Idx); | |||
12856 | return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt); | |||
12857 | } | |||
12858 | ||||
12859 | unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); | |||
12860 | const TargetRegisterClass* rc = getRegClassFor(VecVT); | |||
12861 | if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8)) | |||
12862 | rc = getRegClassFor(MVT::v16i1); | |||
12863 | unsigned MaxSift = rc->getSize()*8 - 1; | |||
12864 | Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec, | |||
12865 | DAG.getConstant(MaxSift - IdxVal, MVT::i8)); | |||
12866 | Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec, | |||
12867 | DAG.getConstant(MaxSift, MVT::i8)); | |||
12868 | return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec, | |||
12869 | DAG.getIntPtrConstant(0)); | |||
12870 | } | |||
12871 | ||||
12872 | SDValue | |||
12873 | X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, | |||
12874 | SelectionDAG &DAG) const { | |||
12875 | SDLoc dl(Op); | |||
12876 | SDValue Vec = Op.getOperand(0); | |||
12877 | MVT VecVT = Vec.getSimpleValueType(); | |||
12878 | SDValue Idx = Op.getOperand(1); | |||
12879 | ||||
12880 | if (Op.getSimpleValueType() == MVT::i1) | |||
12881 | return ExtractBitFromMaskVector(Op, DAG); | |||
12882 | ||||
12883 | if (!isa<ConstantSDNode>(Idx)) { | |||
12884 | if (VecVT.is512BitVector() || | |||
12885 | (VecVT.is256BitVector() && Subtarget->hasInt256() && | |||
12886 | VecVT.getVectorElementType().getSizeInBits() == 32)) { | |||
12887 | ||||
12888 | MVT MaskEltVT = | |||
12889 | MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits()); | |||
12890 | MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() / | |||
12891 | MaskEltVT.getSizeInBits()); | |||
12892 | ||||
12893 | Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT); | |||
12894 | SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT, | |||
12895 | getZeroVector(MaskVT, Subtarget, DAG, dl), | |||
12896 | Idx, DAG.getConstant(0, getPointerTy())); | |||
12897 | SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec); | |||
12898 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), | |||
12899 | Perm, DAG.getConstant(0, getPointerTy())); | |||
12900 | } | |||
12901 | return SDValue(); | |||
12902 | } | |||
12903 | ||||
12904 | // If this is a 256-bit vector result, first extract the 128-bit vector and | |||
12905 | // then extract the element from the 128-bit vector. | |||
12906 | if (VecVT.is256BitVector() || VecVT.is512BitVector()) { | |||
12907 | ||||
12908 | unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); | |||
12909 | // Get the 128-bit vector. | |||
12910 | Vec = Extract128BitVector(Vec, IdxVal, DAG, dl); | |||
12911 | MVT EltVT = VecVT.getVectorElementType(); | |||
12912 | ||||
12913 | unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits(); | |||
12914 | ||||
12915 | //if (IdxVal >= NumElems/2) | |||
12916 | // IdxVal -= NumElems/2; | |||
12917 | IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk; | |||
12918 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec, | |||
12919 | DAG.getConstant(IdxVal, MVT::i32)); | |||
12920 | } | |||
12921 | ||||
12922 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 12922, __PRETTY_FUNCTION__)); | |||
12923 | ||||
12924 | if (Subtarget->hasSSE41()) { | |||
12925 | SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG); | |||
12926 | if (Res.getNode()) | |||
12927 | return Res; | |||
12928 | } | |||
12929 | ||||
12930 | MVT VT = Op.getSimpleValueType(); | |||
12931 | // TODO: handle v16i8. | |||
12932 | if (VT.getSizeInBits() == 16) { | |||
12933 | SDValue Vec = Op.getOperand(0); | |||
12934 | unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); | |||
12935 | if (Idx == 0) | |||
12936 | return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, | |||
12937 | DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, | |||
12938 | DAG.getNode(ISD::BITCAST, dl, | |||
12939 | MVT::v4i32, Vec), | |||
12940 | Op.getOperand(1))); | |||
12941 | // Transform it so it match pextrw which produces a 32-bit result. | |||
12942 | MVT EltVT = MVT::i32; | |||
12943 | SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT, | |||
12944 | Op.getOperand(0), Op.getOperand(1)); | |||
12945 | SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract, | |||
12946 | DAG.getValueType(VT)); | |||
12947 | return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert); | |||
12948 | } | |||
12949 | ||||
12950 | if (VT.getSizeInBits() == 32) { | |||
12951 | unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); | |||
12952 | if (Idx == 0) | |||
12953 | return Op; | |||
12954 | ||||
12955 | // SHUFPS the element to the lowest double word, then movss. | |||
12956 | int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 }; | |||
12957 | MVT VVT = Op.getOperand(0).getSimpleValueType(); | |||
12958 | SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0), | |||
12959 | DAG.getUNDEF(VVT), Mask); | |||
12960 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec, | |||
12961 | DAG.getIntPtrConstant(0)); | |||
12962 | } | |||
12963 | ||||
12964 | if (VT.getSizeInBits() == 64) { | |||
12965 | // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b | |||
12966 | // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught | |||
12967 | // to match extract_elt for f64. | |||
12968 | unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); | |||
12969 | if (Idx == 0) | |||
12970 | return Op; | |||
12971 | ||||
12972 | // UNPCKHPD the element to the lowest double word, then movsd. | |||
12973 | // Note if the lower 64 bits of the result of the UNPCKHPD is then stored | |||
12974 | // to a f64mem, the whole operation is folded into a single MOVHPDmr. | |||
12975 | int Mask[2] = { 1, -1 }; | |||
12976 | MVT VVT = Op.getOperand(0).getSimpleValueType(); | |||
12977 | SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0), | |||
12978 | DAG.getUNDEF(VVT), Mask); | |||
12979 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec, | |||
12980 | DAG.getIntPtrConstant(0)); | |||
12981 | } | |||
12982 | ||||
12983 | return SDValue(); | |||
12984 | } | |||
12985 | ||||
12986 | /// Insert one bit to mask vector, like v16i1 or v8i1. | |||
12987 | /// AVX-512 feature. | |||
12988 | SDValue | |||
12989 | X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const { | |||
12990 | SDLoc dl(Op); | |||
12991 | SDValue Vec = Op.getOperand(0); | |||
12992 | SDValue Elt = Op.getOperand(1); | |||
12993 | SDValue Idx = Op.getOperand(2); | |||
12994 | MVT VecVT = Vec.getSimpleValueType(); | |||
12995 | ||||
12996 | if (!isa<ConstantSDNode>(Idx)) { | |||
12997 | // Non constant index. Extend source and destination, | |||
12998 | // insert element and then truncate the result. | |||
12999 | MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32); | |||
13000 | MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32); | |||
13001 | SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT, | |||
13002 | DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec), | |||
13003 | DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx); | |||
13004 | return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp); | |||
13005 | } | |||
13006 | ||||
13007 | unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); | |||
13008 | SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt); | |||
13009 | if (Vec.getOpcode() == ISD::UNDEF) | |||
13010 | return DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec, | |||
13011 | DAG.getConstant(IdxVal, MVT::i8)); | |||
13012 | const TargetRegisterClass* rc = getRegClassFor(VecVT); | |||
13013 | unsigned MaxSift = rc->getSize()*8 - 1; | |||
13014 | EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec, | |||
13015 | DAG.getConstant(MaxSift, MVT::i8)); | |||
13016 | EltInVec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, EltInVec, | |||
13017 | DAG.getConstant(MaxSift - IdxVal, MVT::i8)); | |||
13018 | return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec); | |||
13019 | } | |||
13020 | ||||
13021 | SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, | |||
13022 | SelectionDAG &DAG) const { | |||
13023 | MVT VT = Op.getSimpleValueType(); | |||
13024 | MVT EltVT = VT.getVectorElementType(); | |||
13025 | ||||
13026 | if (EltVT == MVT::i1) | |||
13027 | return InsertBitToMaskVector(Op, DAG); | |||
13028 | ||||
13029 | SDLoc dl(Op); | |||
13030 | SDValue N0 = Op.getOperand(0); | |||
13031 | SDValue N1 = Op.getOperand(1); | |||
13032 | SDValue N2 = Op.getOperand(2); | |||
13033 | if (!isa<ConstantSDNode>(N2)) | |||
13034 | return SDValue(); | |||
13035 | auto *N2C = cast<ConstantSDNode>(N2); | |||
13036 | unsigned IdxVal = N2C->getZExtValue(); | |||
13037 | ||||
13038 | // If the vector is wider than 128 bits, extract the 128-bit subvector, insert | |||
13039 | // into that, and then insert the subvector back into the result. | |||
13040 | if (VT.is256BitVector() || VT.is512BitVector()) { | |||
13041 | // Get the desired 128-bit vector half. | |||
13042 | SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl); | |||
13043 | ||||
13044 | // Insert the element into the desired half. | |||
13045 | unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits(); | |||
13046 | unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128; | |||
13047 | ||||
13048 | V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1, | |||
13049 | DAG.getConstant(IdxIn128, MVT::i32)); | |||
13050 | ||||
13051 | // Insert the changed part back to the 256-bit vector | |||
13052 | return Insert128BitVector(N0, V, IdxVal, DAG, dl); | |||
13053 | } | |||
13054 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 13054, __PRETTY_FUNCTION__)); | |||
13055 | ||||
13056 | if (Subtarget->hasSSE41()) { | |||
13057 | if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) { | |||
13058 | unsigned Opc; | |||
13059 | if (VT == MVT::v8i16) { | |||
13060 | Opc = X86ISD::PINSRW; | |||
13061 | } else { | |||
13062 | assert(VT == MVT::v16i8)((VT == MVT::v16i8) ? static_cast<void> (0) : __assert_fail ("VT == MVT::v16i8", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 13062, __PRETTY_FUNCTION__)); | |||
13063 | Opc = X86ISD::PINSRB; | |||
13064 | } | |||
13065 | ||||
13066 | // Transform it so it match pinsr{b,w} which expects a GR32 as its second | |||
13067 | // argument. | |||
13068 | if (N1.getValueType() != MVT::i32) | |||
13069 | N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1); | |||
13070 | if (N2.getValueType() != MVT::i32) | |||
13071 | N2 = DAG.getIntPtrConstant(IdxVal); | |||
13072 | return DAG.getNode(Opc, dl, VT, N0, N1, N2); | |||
13073 | } | |||
13074 | ||||
13075 | if (EltVT == MVT::f32) { | |||
13076 | // Bits [7:6] of the constant are the source select. This will always be | |||
13077 | // zero here. The DAG Combiner may combine an extract_elt index into | |||
13078 | // these | |||
13079 | // bits. For example (insert (extract, 3), 2) could be matched by | |||
13080 | // putting | |||
13081 | // the '3' into bits [7:6] of X86ISD::INSERTPS. | |||
13082 | // Bits [5:4] of the constant are the destination select. This is the | |||
13083 | // value of the incoming immediate. | |||
13084 | // Bits [3:0] of the constant are the zero mask. The DAG Combiner may | |||
13085 | // combine either bitwise AND or insert of float 0.0 to set these bits. | |||
13086 | N2 = DAG.getIntPtrConstant(IdxVal << 4); | |||
13087 | // Create this as a scalar to vector.. | |||
13088 | N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1); | |||
13089 | return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2); | |||
13090 | } | |||
13091 | ||||
13092 | if (EltVT == MVT::i32 || EltVT == MVT::i64) { | |||
13093 | // PINSR* works with constant index. | |||
13094 | return Op; | |||
13095 | } | |||
13096 | } | |||
13097 | ||||
13098 | if (EltVT == MVT::i8) | |||
13099 | return SDValue(); | |||
13100 | ||||
13101 | if (EltVT.getSizeInBits() == 16) { | |||
13102 | // Transform it so it match pinsrw which expects a 16-bit value in a GR32 | |||
13103 | // as its second argument. | |||
13104 | if (N1.getValueType() != MVT::i32) | |||
13105 | N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1); | |||
13106 | if (N2.getValueType() != MVT::i32) | |||
13107 | N2 = DAG.getIntPtrConstant(IdxVal); | |||
13108 | return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2); | |||
13109 | } | |||
13110 | return SDValue(); | |||
13111 | } | |||
13112 | ||||
13113 | static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) { | |||
13114 | SDLoc dl(Op); | |||
13115 | MVT OpVT = Op.getSimpleValueType(); | |||
13116 | ||||
13117 | // If this is a 256-bit vector result, first insert into a 128-bit | |||
13118 | // vector and then insert into the 256-bit vector. | |||
13119 | if (!OpVT.is128BitVector()) { | |||
13120 | // Insert into a 128-bit vector. | |||
13121 | unsigned SizeFactor = OpVT.getSizeInBits()/128; | |||
13122 | MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(), | |||
13123 | OpVT.getVectorNumElements() / SizeFactor); | |||
13124 | ||||
13125 | Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0)); | |||
13126 | ||||
13127 | // Insert the 128-bit vector. | |||
13128 | return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl); | |||
13129 | } | |||
13130 | ||||
13131 | if (OpVT == MVT::v1i64 && | |||
13132 | Op.getOperand(0).getValueType() == MVT::i64) | |||
13133 | return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0)); | |||
13134 | ||||
13135 | SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0)); | |||
13136 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 13136, __PRETTY_FUNCTION__)); | |||
13137 | return DAG.getNode(ISD::BITCAST, dl, OpVT, | |||
13138 | DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt)); | |||
13139 | } | |||
13140 | ||||
13141 | // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in | |||
13142 | // a simple subregister reference or explicit instructions to grab | |||
13143 | // upper bits of a vector. | |||
13144 | static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget, | |||
13145 | SelectionDAG &DAG) { | |||
13146 | SDLoc dl(Op); | |||
13147 | SDValue In = Op.getOperand(0); | |||
13148 | SDValue Idx = Op.getOperand(1); | |||
13149 | unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); | |||
13150 | MVT ResVT = Op.getSimpleValueType(); | |||
13151 | MVT InVT = In.getSimpleValueType(); | |||
13152 | ||||
13153 | if (Subtarget->hasFp256()) { | |||
13154 | if (ResVT.is128BitVector() && | |||
13155 | (InVT.is256BitVector() || InVT.is512BitVector()) && | |||
13156 | isa<ConstantSDNode>(Idx)) { | |||
13157 | return Extract128BitVector(In, IdxVal, DAG, dl); | |||
13158 | } | |||
13159 | if (ResVT.is256BitVector() && InVT.is512BitVector() && | |||
13160 | isa<ConstantSDNode>(Idx)) { | |||
13161 | return Extract256BitVector(In, IdxVal, DAG, dl); | |||
13162 | } | |||
13163 | } | |||
13164 | return SDValue(); | |||
13165 | } | |||
13166 | ||||
13167 | // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a | |||
13168 | // simple superregister reference or explicit instructions to insert | |||
13169 | // the upper bits of a vector. | |||
13170 | static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget, | |||
13171 | SelectionDAG &DAG) { | |||
13172 | if (!Subtarget->hasAVX()) | |||
13173 | return SDValue(); | |||
13174 | ||||
13175 | SDLoc dl(Op); | |||
13176 | SDValue Vec = Op.getOperand(0); | |||
13177 | SDValue SubVec = Op.getOperand(1); | |||
13178 | SDValue Idx = Op.getOperand(2); | |||
13179 | MVT OpVT = Op.getSimpleValueType(); | |||
13180 | MVT SubVecVT = SubVec.getSimpleValueType(); | |||
13181 | ||||
13182 | if ((OpVT.is256BitVector() || OpVT.is512BitVector()) && | |||
13183 | SubVecVT.is128BitVector() && isa<ConstantSDNode>(Idx)) { | |||
13184 | unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); | |||
13185 | return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl); | |||
13186 | } | |||
13187 | ||||
13188 | if (OpVT.is512BitVector() && | |||
13189 | SubVecVT.is256BitVector() && isa<ConstantSDNode>(Idx)) { | |||
13190 | unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); | |||
13191 | return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl); | |||
13192 | } | |||
13193 | ||||
13194 | return SDValue(); | |||
13195 | } | |||
13196 | ||||
13197 | // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as | |||
13198 | // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is | |||
13199 | // one of the above mentioned nodes. It has to be wrapped because otherwise | |||
13200 | // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only | |||
13201 | // be used to form addressing mode. These wrapped nodes will be selected | |||
13202 | // into MOV32ri. | |||
13203 | SDValue | |||
13204 | X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const { | |||
13205 | ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); | |||
13206 | ||||
13207 | // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the | |||
13208 | // global base reg. | |||
13209 | unsigned char OpFlag = 0; | |||
13210 | unsigned WrapperKind = X86ISD::Wrapper; | |||
13211 | CodeModel::Model M = DAG.getTarget().getCodeModel(); | |||
13212 | ||||
13213 | if (Subtarget->isPICStyleRIPRel() && | |||
13214 | (M == CodeModel::Small || M == CodeModel::Kernel)) | |||
13215 | WrapperKind = X86ISD::WrapperRIP; | |||
13216 | else if (Subtarget->isPICStyleGOT()) | |||
13217 | OpFlag = X86II::MO_GOTOFF; | |||
13218 | else if (Subtarget->isPICStyleStubPIC()) | |||
13219 | OpFlag = X86II::MO_PIC_BASE_OFFSET; | |||
13220 | ||||
13221 | SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(), | |||
13222 | CP->getAlignment(), | |||
13223 | CP->getOffset(), OpFlag); | |||
13224 | SDLoc DL(CP); | |||
13225 | Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result); | |||
13226 | // With PIC, the address is actually $g + Offset. | |||
13227 | if (OpFlag) { | |||
13228 | Result = DAG.getNode(ISD::ADD, DL, getPointerTy(), | |||
13229 | DAG.getNode(X86ISD::GlobalBaseReg, | |||
13230 | SDLoc(), getPointerTy()), | |||
13231 | Result); | |||
13232 | } | |||
13233 | ||||
13234 | return Result; | |||
13235 | } | |||
13236 | ||||
13237 | SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const { | |||
13238 | JumpTableSDNode *JT = cast<JumpTableSDNode>(Op); | |||
13239 | ||||
13240 | // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the | |||
13241 | // global base reg. | |||
13242 | unsigned char OpFlag = 0; | |||
13243 | unsigned WrapperKind = X86ISD::Wrapper; | |||
13244 | CodeModel::Model M = DAG.getTarget().getCodeModel(); | |||
13245 | ||||
13246 | if (Subtarget->isPICStyleRIPRel() && | |||
13247 | (M == CodeModel::Small || M == CodeModel::Kernel)) | |||
13248 | WrapperKind = X86ISD::WrapperRIP; | |||
13249 | else if (Subtarget->isPICStyleGOT()) | |||
13250 | OpFlag = X86II::MO_GOTOFF; | |||
13251 | else if (Subtarget->isPICStyleStubPIC()) | |||
13252 | OpFlag = X86II::MO_PIC_BASE_OFFSET; | |||
13253 | ||||
13254 | SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(), | |||
13255 | OpFlag); | |||
13256 | SDLoc DL(JT); | |||
13257 | Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result); | |||
13258 | ||||
13259 | // With PIC, the address is actually $g + Offset. | |||
13260 | if (OpFlag) | |||
13261 | Result = DAG.getNode(ISD::ADD, DL, getPointerTy(), | |||
13262 | DAG.getNode(X86ISD::GlobalBaseReg, | |||
13263 | SDLoc(), getPointerTy()), | |||
13264 | Result); | |||
13265 | ||||
13266 | return Result; | |||
13267 | } | |||
13268 | ||||
13269 | SDValue | |||
13270 | X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const { | |||
13271 | const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol(); | |||
13272 | ||||
13273 | // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the | |||
13274 | // global base reg. | |||
13275 | unsigned char OpFlag = 0; | |||
13276 | unsigned WrapperKind = X86ISD::Wrapper; | |||
13277 | CodeModel::Model M = DAG.getTarget().getCodeModel(); | |||
13278 | ||||
13279 | if (Subtarget->isPICStyleRIPRel() && | |||
13280 | (M == CodeModel::Small || M == CodeModel::Kernel)) { | |||
13281 | if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF()) | |||
13282 | OpFlag = X86II::MO_GOTPCREL; | |||
13283 | WrapperKind = X86ISD::WrapperRIP; | |||
13284 | } else if (Subtarget->isPICStyleGOT()) { | |||
13285 | OpFlag = X86II::MO_GOT; | |||
13286 | } else if (Subtarget->isPICStyleStubPIC()) { | |||
13287 | OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE; | |||
13288 | } else if (Subtarget->isPICStyleStubNoDynamic()) { | |||
13289 | OpFlag = X86II::MO_DARWIN_NONLAZY; | |||
13290 | } | |||
13291 | ||||
13292 | SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag); | |||
13293 | ||||
13294 | SDLoc DL(Op); | |||
13295 | Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result); | |||
13296 | ||||
13297 | // With PIC, the address is actually $g + Offset. | |||
13298 | if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ && | |||
13299 | !Subtarget->is64Bit()) { | |||
13300 | Result = DAG.getNode(ISD::ADD, DL, getPointerTy(), | |||
13301 | DAG.getNode(X86ISD::GlobalBaseReg, | |||
13302 | SDLoc(), getPointerTy()), | |||
13303 | Result); | |||
13304 | } | |||
13305 | ||||
13306 | // For symbols that require a load from a stub to get the address, emit the | |||
13307 | // load. | |||
13308 | if (isGlobalStubReference(OpFlag)) | |||
13309 | Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result, | |||
13310 | MachinePointerInfo::getGOT(), false, false, false, 0); | |||
13311 | ||||
13312 | return Result; | |||
13313 | } | |||
13314 | ||||
13315 | SDValue | |||
13316 | X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const { | |||
13317 | // Create the TargetBlockAddressAddress node. | |||
13318 | unsigned char OpFlags = | |||
13319 | Subtarget->ClassifyBlockAddressReference(); | |||
13320 | CodeModel::Model M = DAG.getTarget().getCodeModel(); | |||
13321 | const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress(); | |||
13322 | int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset(); | |||
13323 | SDLoc dl(Op); | |||
13324 | SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset, | |||
13325 | OpFlags); | |||
13326 | ||||
13327 | if (Subtarget->isPICStyleRIPRel() && | |||
13328 | (M == CodeModel::Small || M == CodeModel::Kernel)) | |||
13329 | Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result); | |||
13330 | else | |||
13331 | Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result); | |||
13332 | ||||
13333 | // With PIC, the address is actually $g + Offset. | |||
13334 | if (isGlobalRelativeToPICBase(OpFlags)) { | |||
13335 | Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), | |||
13336 | DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()), | |||
13337 | Result); | |||
13338 | } | |||
13339 | ||||
13340 | return Result; | |||
13341 | } | |||
13342 | ||||
13343 | SDValue | |||
13344 | X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl, | |||
13345 | int64_t Offset, SelectionDAG &DAG) const { | |||
13346 | // Create the TargetGlobalAddress node, folding in the constant | |||
13347 | // offset if it is legal. | |||
13348 | unsigned char OpFlags = | |||
13349 | Subtarget->ClassifyGlobalReference(GV, DAG.getTarget()); | |||
13350 | CodeModel::Model M = DAG.getTarget().getCodeModel(); | |||
13351 | SDValue Result; | |||
13352 | if (OpFlags == X86II::MO_NO_FLAG && | |||
13353 | X86::isOffsetSuitableForCodeModel(Offset, M)) { | |||
13354 | // A direct static reference to a global. | |||
13355 | Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset); | |||
13356 | Offset = 0; | |||
13357 | } else { | |||
13358 | Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags); | |||
13359 | } | |||
13360 | ||||
13361 | if (Subtarget->isPICStyleRIPRel() && | |||
13362 | (M == CodeModel::Small || M == CodeModel::Kernel)) | |||
13363 | Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result); | |||
13364 | else | |||
13365 | Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result); | |||
13366 | ||||
13367 | // With PIC, the address is actually $g + Offset. | |||
13368 | if (isGlobalRelativeToPICBase(OpFlags)) { | |||
13369 | Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), | |||
13370 | DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()), | |||
13371 | Result); | |||
13372 | } | |||
13373 | ||||
13374 | // For globals that require a load from a stub to get the address, emit the | |||
13375 | // load. | |||
13376 | if (isGlobalStubReference(OpFlags)) | |||
13377 | Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result, | |||
13378 | MachinePointerInfo::getGOT(), false, false, false, 0); | |||
13379 | ||||
13380 | // If there was a non-zero offset that we didn't fold, create an explicit | |||
13381 | // addition for it. | |||
13382 | if (Offset != 0) | |||
13383 | Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result, | |||
13384 | DAG.getConstant(Offset, getPointerTy())); | |||
13385 | ||||
13386 | return Result; | |||
13387 | } | |||
13388 | ||||
13389 | SDValue | |||
13390 | X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const { | |||
13391 | const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); | |||
13392 | int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset(); | |||
13393 | return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG); | |||
13394 | } | |||
13395 | ||||
13396 | static SDValue | |||
13397 | GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA, | |||
13398 | SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg, | |||
13399 | unsigned char OperandFlags, bool LocalDynamic = false) { | |||
13400 | MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); | |||
13401 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
13402 | SDLoc dl(GA); | |||
13403 | SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, | |||
13404 | GA->getValueType(0), | |||
13405 | GA->getOffset(), | |||
13406 | OperandFlags); | |||
13407 | ||||
13408 | X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR | |||
13409 | : X86ISD::TLSADDR; | |||
13410 | ||||
13411 | if (InFlag) { | |||
13412 | SDValue Ops[] = { Chain, TGA, *InFlag }; | |||
13413 | Chain = DAG.getNode(CallType, dl, NodeTys, Ops); | |||
13414 | } else { | |||
13415 | SDValue Ops[] = { Chain, TGA }; | |||
13416 | Chain = DAG.getNode(CallType, dl, NodeTys, Ops); | |||
13417 | } | |||
13418 | ||||
13419 | // TLSADDR will be codegen'ed as call. Inform MFI that function has calls. | |||
13420 | MFI->setAdjustsStack(true); | |||
13421 | MFI->setHasCalls(true); | |||
13422 | ||||
13423 | SDValue Flag = Chain.getValue(1); | |||
13424 | return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag); | |||
13425 | } | |||
13426 | ||||
13427 | // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit | |||
13428 | static SDValue | |||
13429 | LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG, | |||
13430 | const EVT PtrVT) { | |||
13431 | SDValue InFlag; | |||
13432 | SDLoc dl(GA); // ? function entry point might be better | |||
13433 | SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX, | |||
13434 | DAG.getNode(X86ISD::GlobalBaseReg, | |||
13435 | SDLoc(), PtrVT), InFlag); | |||
13436 | InFlag = Chain.getValue(1); | |||
13437 | ||||
13438 | return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD); | |||
13439 | } | |||
13440 | ||||
13441 | // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit | |||
13442 | static SDValue | |||
13443 | LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG, | |||
13444 | const EVT PtrVT) { | |||
13445 | return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, | |||
13446 | X86::RAX, X86II::MO_TLSGD); | |||
13447 | } | |||
13448 | ||||
13449 | static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA, | |||
13450 | SelectionDAG &DAG, | |||
13451 | const EVT PtrVT, | |||
13452 | bool is64Bit) { | |||
13453 | SDLoc dl(GA); | |||
13454 | ||||
13455 | // Get the start address of the TLS block for this module. | |||
13456 | X86MachineFunctionInfo* MFI = DAG.getMachineFunction() | |||
13457 | .getInfo<X86MachineFunctionInfo>(); | |||
13458 | MFI->incNumLocalDynamicTLSAccesses(); | |||
13459 | ||||
13460 | SDValue Base; | |||
13461 | if (is64Bit) { | |||
13462 | Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX, | |||
13463 | X86II::MO_TLSLD, /*LocalDynamic=*/true); | |||
13464 | } else { | |||
13465 | SDValue InFlag; | |||
13466 | SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX, | |||
13467 | DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag); | |||
13468 | InFlag = Chain.getValue(1); | |||
13469 | Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, | |||
13470 | X86II::MO_TLSLDM, /*LocalDynamic=*/true); | |||
13471 | } | |||
13472 | ||||
13473 | // Note: the CleanupLocalDynamicTLSPass will remove redundant computations | |||
13474 | // of Base. | |||
13475 | ||||
13476 | // Build x@dtpoff. | |||
13477 | unsigned char OperandFlags = X86II::MO_DTPOFF; | |||
13478 | unsigned WrapperKind = X86ISD::Wrapper; | |||
13479 | SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, | |||
13480 | GA->getValueType(0), | |||
13481 | GA->getOffset(), OperandFlags); | |||
13482 | SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA); | |||
13483 | ||||
13484 | // Add x@dtpoff with the base. | |||
13485 | return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base); | |||
13486 | } | |||
13487 | ||||
13488 | // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model. | |||
13489 | static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG, | |||
13490 | const EVT PtrVT, TLSModel::Model model, | |||
13491 | bool is64Bit, bool isPIC) { | |||
13492 | SDLoc dl(GA); | |||
13493 | ||||
13494 | // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit). | |||
13495 | Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(), | |||
13496 | is64Bit ? 257 : 256)); | |||
13497 | ||||
13498 | SDValue ThreadPointer = | |||
13499 | DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0), | |||
13500 | MachinePointerInfo(Ptr), false, false, false, 0); | |||
13501 | ||||
13502 | unsigned char OperandFlags = 0; | |||
13503 | // Most TLS accesses are not RIP relative, even on x86-64. One exception is | |||
13504 | // initialexec. | |||
13505 | unsigned WrapperKind = X86ISD::Wrapper; | |||
13506 | if (model == TLSModel::LocalExec) { | |||
13507 | OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF; | |||
13508 | } else if (model == TLSModel::InitialExec) { | |||
13509 | if (is64Bit) { | |||
13510 | OperandFlags = X86II::MO_GOTTPOFF; | |||
13511 | WrapperKind = X86ISD::WrapperRIP; | |||
13512 | } else { | |||
13513 | OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF; | |||
13514 | } | |||
13515 | } else { | |||
13516 | llvm_unreachable("Unexpected model")::llvm::llvm_unreachable_internal("Unexpected model", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 13516); | |||
13517 | } | |||
13518 | ||||
13519 | // emit "addl x@ntpoff,%eax" (local exec) | |||
13520 | // or "addl x@indntpoff,%eax" (initial exec) | |||
13521 | // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic) | |||
13522 | SDValue TGA = | |||
13523 | DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0), | |||
13524 | GA->getOffset(), OperandFlags); | |||
13525 | SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA); | |||
13526 | ||||
13527 | if (model == TLSModel::InitialExec) { | |||
13528 | if (isPIC && !is64Bit) { | |||
13529 | Offset = DAG.getNode(ISD::ADD, dl, PtrVT, | |||
13530 | DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), | |||
13531 | Offset); | |||
13532 | } | |||
13533 | ||||
13534 | Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset, | |||
13535 | MachinePointerInfo::getGOT(), false, false, false, 0); | |||
13536 | } | |||
13537 | ||||
13538 | // The address of the thread local variable is the add of the thread | |||
13539 | // pointer with the offset of the variable. | |||
13540 | return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset); | |||
13541 | } | |||
13542 | ||||
13543 | SDValue | |||
13544 | X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const { | |||
13545 | ||||
13546 | GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); | |||
13547 | const GlobalValue *GV = GA->getGlobal(); | |||
13548 | ||||
13549 | if (Subtarget->isTargetELF()) { | |||
13550 | TLSModel::Model model = DAG.getTarget().getTLSModel(GV); | |||
13551 | ||||
13552 | switch (model) { | |||
13553 | case TLSModel::GeneralDynamic: | |||
13554 | if (Subtarget->is64Bit()) | |||
13555 | return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy()); | |||
13556 | return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy()); | |||
13557 | case TLSModel::LocalDynamic: | |||
13558 | return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(), | |||
13559 | Subtarget->is64Bit()); | |||
13560 | case TLSModel::InitialExec: | |||
13561 | case TLSModel::LocalExec: | |||
13562 | return LowerToTLSExecModel( | |||
13563 | GA, DAG, getPointerTy(), model, Subtarget->is64Bit(), | |||
13564 | DAG.getTarget().getRelocationModel() == Reloc::PIC_); | |||
13565 | } | |||
13566 | llvm_unreachable("Unknown TLS model.")::llvm::llvm_unreachable_internal("Unknown TLS model.", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 13566); | |||
13567 | } | |||
13568 | ||||
13569 | if (Subtarget->isTargetDarwin()) { | |||
13570 | // Darwin only has one model of TLS. Lower to that. | |||
13571 | unsigned char OpFlag = 0; | |||
13572 | unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ? | |||
13573 | X86ISD::WrapperRIP : X86ISD::Wrapper; | |||
13574 | ||||
13575 | // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the | |||
13576 | // global base reg. | |||
13577 | bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) && | |||
13578 | !Subtarget->is64Bit(); | |||
13579 | if (PIC32) | |||
13580 | OpFlag = X86II::MO_TLVP_PIC_BASE; | |||
13581 | else | |||
13582 | OpFlag = X86II::MO_TLVP; | |||
13583 | SDLoc DL(Op); | |||
13584 | SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL, | |||
13585 | GA->getValueType(0), | |||
13586 | GA->getOffset(), OpFlag); | |||
13587 | SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result); | |||
13588 | ||||
13589 | // With PIC32, the address is actually $g + Offset. | |||
13590 | if (PIC32) | |||
13591 | Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(), | |||
13592 | DAG.getNode(X86ISD::GlobalBaseReg, | |||
13593 | SDLoc(), getPointerTy()), | |||
13594 | Offset); | |||
13595 | ||||
13596 | // Lowering the machine isd will make sure everything is in the right | |||
13597 | // location. | |||
13598 | SDValue Chain = DAG.getEntryNode(); | |||
13599 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
13600 | SDValue Args[] = { Chain, Offset }; | |||
13601 | Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args); | |||
13602 | ||||
13603 | // TLSCALL will be codegen'ed as call. Inform MFI that function has calls. | |||
13604 | MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); | |||
13605 | MFI->setAdjustsStack(true); | |||
13606 | ||||
13607 | // And our return value (tls address) is in the standard call return value | |||
13608 | // location. | |||
13609 | unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX; | |||
13610 | return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(), | |||
13611 | Chain.getValue(1)); | |||
13612 | } | |||
13613 | ||||
13614 | if (Subtarget->isTargetKnownWindowsMSVC() || | |||
13615 | Subtarget->isTargetWindowsGNU()) { | |||
13616 | // Just use the implicit TLS architecture | |||
13617 | // Need to generate someting similar to: | |||
13618 | // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage | |||
13619 | // ; from TEB | |||
13620 | // mov ecx, dword [rel _tls_index]: Load index (from C runtime) | |||
13621 | // mov rcx, qword [rdx+rcx*8] | |||
13622 | // mov eax, .tls$:tlsvar | |||
13623 | // [rax+rcx] contains the address | |||
13624 | // Windows 64bit: gs:0x58 | |||
13625 | // Windows 32bit: fs:__tls_array | |||
13626 | ||||
13627 | SDLoc dl(GA); | |||
13628 | SDValue Chain = DAG.getEntryNode(); | |||
13629 | ||||
13630 | // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or | |||
13631 | // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly | |||
13632 | // use its literal value of 0x2C. | |||
13633 | Value *Ptr = Constant::getNullValue(Subtarget->is64Bit() | |||
13634 | ? Type::getInt8PtrTy(*DAG.getContext(), | |||
13635 | 256) | |||
13636 | : Type::getInt32PtrTy(*DAG.getContext(), | |||
13637 | 257)); | |||
13638 | ||||
13639 | SDValue TlsArray = | |||
13640 | Subtarget->is64Bit() | |||
13641 | ? DAG.getIntPtrConstant(0x58) | |||
13642 | : (Subtarget->isTargetWindowsGNU() | |||
13643 | ? DAG.getIntPtrConstant(0x2C) | |||
13644 | : DAG.getExternalSymbol("_tls_array", getPointerTy())); | |||
13645 | ||||
13646 | SDValue ThreadPointer = | |||
13647 | DAG.getLoad(getPointerTy(), dl, Chain, TlsArray, | |||
13648 | MachinePointerInfo(Ptr), false, false, false, 0); | |||
13649 | ||||
13650 | // Load the _tls_index variable | |||
13651 | SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy()); | |||
13652 | if (Subtarget->is64Bit()) | |||
13653 | IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain, | |||
13654 | IDX, MachinePointerInfo(), MVT::i32, | |||
13655 | false, false, false, 0); | |||
13656 | else | |||
13657 | IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(), | |||
13658 | false, false, false, 0); | |||
13659 | ||||
13660 | SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()), | |||
13661 | getPointerTy()); | |||
13662 | IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale); | |||
13663 | ||||
13664 | SDValue res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX); | |||
13665 | res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(), | |||
13666 | false, false, false, 0); | |||
13667 | ||||
13668 | // Get the offset of start of .tls section | |||
13669 | SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, | |||
13670 | GA->getValueType(0), | |||
13671 | GA->getOffset(), X86II::MO_SECREL); | |||
13672 | SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA); | |||
13673 | ||||
13674 | // The address of the thread local variable is the add of the thread | |||
13675 | // pointer with the offset of the variable. | |||
13676 | return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset); | |||
13677 | } | |||
13678 | ||||
13679 | llvm_unreachable("TLS not implemented for this target.")::llvm::llvm_unreachable_internal("TLS not implemented for this target." , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 13679); | |||
13680 | } | |||
13681 | ||||
13682 | /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values | |||
13683 | /// and take a 2 x i32 value to shift plus a shift amount. | |||
13684 | static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) { | |||
13685 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 13685, __PRETTY_FUNCTION__)); | |||
13686 | MVT VT = Op.getSimpleValueType(); | |||
13687 | unsigned VTBits = VT.getSizeInBits(); | |||
13688 | SDLoc dl(Op); | |||
13689 | bool isSRA = Op.getOpcode() == ISD::SRA_PARTS; | |||
13690 | SDValue ShOpLo = Op.getOperand(0); | |||
13691 | SDValue ShOpHi = Op.getOperand(1); | |||
13692 | SDValue ShAmt = Op.getOperand(2); | |||
13693 | // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the | |||
13694 | // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away | |||
13695 | // during isel. | |||
13696 | SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt, | |||
13697 | DAG.getConstant(VTBits - 1, MVT::i8)); | |||
13698 | SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi, | |||
13699 | DAG.getConstant(VTBits - 1, MVT::i8)) | |||
13700 | : DAG.getConstant(0, VT); | |||
13701 | ||||
13702 | SDValue Tmp2, Tmp3; | |||
13703 | if (Op.getOpcode() == ISD::SHL_PARTS) { | |||
13704 | Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt); | |||
13705 | Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt); | |||
13706 | } else { | |||
13707 | Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt); | |||
13708 | Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt); | |||
13709 | } | |||
13710 | ||||
13711 | // If the shift amount is larger or equal than the width of a part we can't | |||
13712 | // rely on the results of shld/shrd. Insert a test and select the appropriate | |||
13713 | // values for large shift amounts. | |||
13714 | SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt, | |||
13715 | DAG.getConstant(VTBits, MVT::i8)); | |||
13716 | SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32, | |||
13717 | AndNode, DAG.getConstant(0, MVT::i8)); | |||
13718 | ||||
13719 | SDValue Hi, Lo; | |||
13720 | SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8); | |||
13721 | SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond }; | |||
13722 | SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond }; | |||
13723 | ||||
13724 | if (Op.getOpcode() == ISD::SHL_PARTS) { | |||
13725 | Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0); | |||
13726 | Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1); | |||
13727 | } else { | |||
13728 | Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0); | |||
13729 | Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1); | |||
13730 | } | |||
13731 | ||||
13732 | SDValue Ops[2] = { Lo, Hi }; | |||
13733 | return DAG.getMergeValues(Ops, dl); | |||
13734 | } | |||
13735 | ||||
13736 | SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, | |||
13737 | SelectionDAG &DAG) const { | |||
13738 | MVT SrcVT = Op.getOperand(0).getSimpleValueType(); | |||
13739 | SDLoc dl(Op); | |||
13740 | ||||
13741 | if (SrcVT.isVector()) { | |||
13742 | if (SrcVT.getVectorElementType() == MVT::i1) { | |||
13743 | MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements()); | |||
13744 | return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), | |||
13745 | DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, | |||
13746 | Op.getOperand(0))); | |||
13747 | } | |||
13748 | return SDValue(); | |||
13749 | } | |||
13750 | ||||
13751 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 13752, __PRETTY_FUNCTION__)) | |||
13752 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 13752, __PRETTY_FUNCTION__)); | |||
13753 | ||||
13754 | // These are really Legal; return the operand so the caller accepts it as | |||
13755 | // Legal. | |||
13756 | if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType())) | |||
13757 | return Op; | |||
13758 | if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) && | |||
13759 | Subtarget->is64Bit()) { | |||
13760 | return Op; | |||
13761 | } | |||
13762 | ||||
13763 | unsigned Size = SrcVT.getSizeInBits()/8; | |||
13764 | MachineFunction &MF = DAG.getMachineFunction(); | |||
13765 | int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false); | |||
13766 | SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy()); | |||
13767 | SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0), | |||
13768 | StackSlot, | |||
13769 | MachinePointerInfo::getFixedStack(SSFI), | |||
13770 | false, false, 0); | |||
13771 | return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG); | |||
13772 | } | |||
13773 | ||||
13774 | SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, | |||
13775 | SDValue StackSlot, | |||
13776 | SelectionDAG &DAG) const { | |||
13777 | // Build the FILD | |||
13778 | SDLoc DL(Op); | |||
13779 | SDVTList Tys; | |||
13780 | bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType()); | |||
13781 | if (useSSE) | |||
13782 | Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue); | |||
13783 | else | |||
13784 | Tys = DAG.getVTList(Op.getValueType(), MVT::Other); | |||
13785 | ||||
13786 | unsigned ByteSize = SrcVT.getSizeInBits()/8; | |||
13787 | ||||
13788 | FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot); | |||
13789 | MachineMemOperand *MMO; | |||
13790 | if (FI) { | |||
13791 | int SSFI = FI->getIndex(); | |||
13792 | MMO = | |||
13793 | DAG.getMachineFunction() | |||
13794 | .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI), | |||
13795 | MachineMemOperand::MOLoad, ByteSize, ByteSize); | |||
13796 | } else { | |||
13797 | MMO = cast<LoadSDNode>(StackSlot)->getMemOperand(); | |||
13798 | StackSlot = StackSlot.getOperand(1); | |||
13799 | } | |||
13800 | SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) }; | |||
13801 | SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG : | |||
13802 | X86ISD::FILD, DL, | |||
13803 | Tys, Ops, SrcVT, MMO); | |||
13804 | ||||
13805 | if (useSSE) { | |||
13806 | Chain = Result.getValue(1); | |||
13807 | SDValue InFlag = Result.getValue(2); | |||
13808 | ||||
13809 | // FIXME: Currently the FST is flagged to the FILD_FLAG. This | |||
13810 | // shouldn't be necessary except that RFP cannot be live across | |||
13811 | // multiple blocks. When stackifier is fixed, they can be uncoupled. | |||
13812 | MachineFunction &MF = DAG.getMachineFunction(); | |||
13813 | unsigned SSFISize = Op.getValueType().getSizeInBits()/8; | |||
13814 | int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false); | |||
13815 | SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy()); | |||
13816 | Tys = DAG.getVTList(MVT::Other); | |||
13817 | SDValue Ops[] = { | |||
13818 | Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag | |||
13819 | }; | |||
13820 | MachineMemOperand *MMO = | |||
13821 | DAG.getMachineFunction() | |||
13822 | .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI), | |||
13823 | MachineMemOperand::MOStore, SSFISize, SSFISize); | |||
13824 | ||||
13825 | Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys, | |||
13826 | Ops, Op.getValueType(), MMO); | |||
13827 | Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot, | |||
13828 | MachinePointerInfo::getFixedStack(SSFI), | |||
13829 | false, false, false, 0); | |||
13830 | } | |||
13831 | ||||
13832 | return Result; | |||
13833 | } | |||
13834 | ||||
13835 | // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion. | |||
13836 | SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, | |||
13837 | SelectionDAG &DAG) const { | |||
13838 | // This algorithm is not obvious. Here it is what we're trying to output: | |||
13839 | /* | |||
13840 | movq %rax, %xmm0 | |||
13841 | punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U } | |||
13842 | subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 } | |||
13843 | #ifdef __SSE3__ | |||
13844 | haddpd %xmm0, %xmm0 | |||
13845 | #else | |||
13846 | pshufd $0x4e, %xmm0, %xmm1 | |||
13847 | addpd %xmm1, %xmm0 | |||
13848 | #endif | |||
13849 | */ | |||
13850 | ||||
13851 | SDLoc dl(Op); | |||
13852 | LLVMContext *Context = DAG.getContext(); | |||
13853 | ||||
13854 | // Build some magic constants. | |||
13855 | static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 }; | |||
13856 | Constant *C0 = ConstantDataVector::get(*Context, CV0); | |||
13857 | SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16); | |||
13858 | ||||
13859 | SmallVector<Constant*,2> CV1; | |||
13860 | CV1.push_back( | |||
13861 | ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble, | |||
13862 | APInt(64, 0x4330000000000000ULL)))); | |||
13863 | CV1.push_back( | |||
13864 | ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble, | |||
13865 | APInt(64, 0x4530000000000000ULL)))); | |||
13866 | Constant *C1 = ConstantVector::get(CV1); | |||
13867 | SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16); | |||
13868 | ||||
13869 | // Load the 64-bit value into an XMM register. | |||
13870 | SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, | |||
13871 | Op.getOperand(0)); | |||
13872 | SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0, | |||
13873 | MachinePointerInfo::getConstantPool(), | |||
13874 | false, false, false, 16); | |||
13875 | SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, | |||
13876 | DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1), | |||
13877 | CLod0); | |||
13878 | ||||
13879 | SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1, | |||
13880 | MachinePointerInfo::getConstantPool(), | |||
13881 | false, false, false, 16); | |||
13882 | SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1); | |||
13883 | SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1); | |||
13884 | SDValue Result; | |||
13885 | ||||
13886 | if (Subtarget->hasSSE3()) { | |||
13887 | // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'. | |||
13888 | Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub); | |||
13889 | } else { | |||
13890 | SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub); | |||
13891 | SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32, | |||
13892 | S2F, 0x4E, DAG); | |||
13893 | Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64, | |||
13894 | DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle), | |||
13895 | Sub); | |||
13896 | } | |||
13897 | ||||
13898 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result, | |||
13899 | DAG.getIntPtrConstant(0)); | |||
13900 | } | |||
13901 | ||||
13902 | // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion. | |||
13903 | SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, | |||
13904 | SelectionDAG &DAG) const { | |||
13905 | SDLoc dl(Op); | |||
13906 | // FP constant to bias correct the final result. | |||
13907 | SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), | |||
13908 | MVT::f64); | |||
13909 | ||||
13910 | // Load the 32-bit value into an XMM register. | |||
13911 | SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, | |||
13912 | Op.getOperand(0)); | |||
13913 | ||||
13914 | // Zero out the upper parts of the register. | |||
13915 | Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG); | |||
13916 | ||||
13917 | Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, | |||
13918 | DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load), | |||
13919 | DAG.getIntPtrConstant(0)); | |||
13920 | ||||
13921 | // Or the load with the bias. | |||
13922 | SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, | |||
13923 | DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, | |||
13924 | DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, | |||
13925 | MVT::v2f64, Load)), | |||
13926 | DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, | |||
13927 | DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, | |||
13928 | MVT::v2f64, Bias))); | |||
13929 | Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, | |||
13930 | DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or), | |||
13931 | DAG.getIntPtrConstant(0)); | |||
13932 | ||||
13933 | // Subtract the bias. | |||
13934 | SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias); | |||
13935 | ||||
13936 | // Handle final rounding. | |||
13937 | EVT DestVT = Op.getValueType(); | |||
13938 | ||||
13939 | if (DestVT.bitsLT(MVT::f64)) | |||
13940 | return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub, | |||
13941 | DAG.getIntPtrConstant(0)); | |||
13942 | if (DestVT.bitsGT(MVT::f64)) | |||
13943 | return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub); | |||
13944 | ||||
13945 | // Handle final rounding. | |||
13946 | return Sub; | |||
13947 | } | |||
13948 | ||||
13949 | static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG, | |||
13950 | const X86Subtarget &Subtarget) { | |||
13951 | // The algorithm is the following: | |||
13952 | // #ifdef __SSE4_1__ | |||
13953 | // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa); | |||
13954 | // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16), | |||
13955 | // (uint4) 0x53000000, 0xaa); | |||
13956 | // #else | |||
13957 | // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000; | |||
13958 | // uint4 hi = (v >> 16) | (uint4) 0x53000000; | |||
13959 | // #endif | |||
13960 | // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f); | |||
13961 | // return (float4) lo + fhi; | |||
13962 | ||||
13963 | SDLoc DL(Op); | |||
13964 | SDValue V = Op->getOperand(0); | |||
13965 | EVT VecIntVT = V.getValueType(); | |||
13966 | bool Is128 = VecIntVT == MVT::v4i32; | |||
13967 | EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32; | |||
13968 | // If we convert to something else than the supported type, e.g., to v4f64, | |||
13969 | // abort early. | |||
13970 | if (VecFloatVT != Op->getValueType(0)) | |||
13971 | return SDValue(); | |||
13972 | ||||
13973 | unsigned NumElts = VecIntVT.getVectorNumElements(); | |||
13974 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 13975, __PRETTY_FUNCTION__)) | |||
13975 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 13975, __PRETTY_FUNCTION__)); | |||
13976 | assert(NumElts <= 8 && "The size of the constant array must be fixed")((NumElts <= 8 && "The size of the constant array must be fixed" ) ? static_cast<void> (0) : __assert_fail ("NumElts <= 8 && \"The size of the constant array must be fixed\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 13976, __PRETTY_FUNCTION__)); | |||
13977 | ||||
13978 | // In the #idef/#else code, we have in common: | |||
13979 | // - The vector of constants: | |||
13980 | // -- 0x4b000000 | |||
13981 | // -- 0x53000000 | |||
13982 | // - A shift: | |||
13983 | // -- v >> 16 | |||
13984 | ||||
13985 | // Create the splat vector for 0x4b000000. | |||
13986 | SDValue CstLow = DAG.getConstant(0x4b000000, MVT::i32); | |||
13987 | SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow, | |||
13988 | CstLow, CstLow, CstLow, CstLow}; | |||
13989 | SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, | |||
13990 | makeArrayRef(&CstLowArray[0], NumElts)); | |||
13991 | // Create the splat vector for 0x53000000. | |||
13992 | SDValue CstHigh = DAG.getConstant(0x53000000, MVT::i32); | |||
13993 | SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh, | |||
13994 | CstHigh, CstHigh, CstHigh, CstHigh}; | |||
13995 | SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, | |||
13996 | makeArrayRef(&CstHighArray[0], NumElts)); | |||
13997 | ||||
13998 | // Create the right shift. | |||
13999 | SDValue CstShift = DAG.getConstant(16, MVT::i32); | |||
14000 | SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift, | |||
14001 | CstShift, CstShift, CstShift, CstShift}; | |||
14002 | SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, | |||
14003 | makeArrayRef(&CstShiftArray[0], NumElts)); | |||
14004 | SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift); | |||
14005 | ||||
14006 | SDValue Low, High; | |||
14007 | if (Subtarget.hasSSE41()) { | |||
14008 | EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16; | |||
14009 | // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa); | |||
14010 | SDValue VecCstLowBitcast = | |||
14011 | DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstLow); | |||
14012 | SDValue VecBitcast = DAG.getNode(ISD::BITCAST, DL, VecI16VT, V); | |||
14013 | // Low will be bitcasted right away, so do not bother bitcasting back to its | |||
14014 | // original type. | |||
14015 | Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast, | |||
14016 | VecCstLowBitcast, DAG.getConstant(0xaa, MVT::i32)); | |||
14017 | // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16), | |||
14018 | // (uint4) 0x53000000, 0xaa); | |||
14019 | SDValue VecCstHighBitcast = | |||
14020 | DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstHigh); | |||
14021 | SDValue VecShiftBitcast = | |||
14022 | DAG.getNode(ISD::BITCAST, DL, VecI16VT, HighShift); | |||
14023 | // High will be bitcasted right away, so do not bother bitcasting back to | |||
14024 | // its original type. | |||
14025 | High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast, | |||
14026 | VecCstHighBitcast, DAG.getConstant(0xaa, MVT::i32)); | |||
14027 | } else { | |||
14028 | SDValue CstMask = DAG.getConstant(0xffff, MVT::i32); | |||
14029 | SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask, | |||
14030 | CstMask, CstMask, CstMask); | |||
14031 | // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000; | |||
14032 | SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask); | |||
14033 | Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow); | |||
14034 | ||||
14035 | // uint4 hi = (v >> 16) | (uint4) 0x53000000; | |||
14036 | High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh); | |||
14037 | } | |||
14038 | ||||
14039 | // Create the vector constant for -(0x1.0p39f + 0x1.0p23f). | |||
14040 | SDValue CstFAdd = DAG.getConstantFP( | |||
14041 | APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), MVT::f32); | |||
14042 | SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd, | |||
14043 | CstFAdd, CstFAdd, CstFAdd, CstFAdd}; | |||
14044 | SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT, | |||
14045 | makeArrayRef(&CstFAddArray[0], NumElts)); | |||
14046 | ||||
14047 | // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f); | |||
14048 | SDValue HighBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, High); | |||
14049 | SDValue FHigh = | |||
14050 | DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd); | |||
14051 | // return (float4) lo + fhi; | |||
14052 | SDValue LowBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, Low); | |||
14053 | return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh); | |||
14054 | } | |||
14055 | ||||
14056 | SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op, | |||
14057 | SelectionDAG &DAG) const { | |||
14058 | SDValue N0 = Op.getOperand(0); | |||
14059 | MVT SVT = N0.getSimpleValueType(); | |||
14060 | SDLoc dl(Op); | |||
14061 | ||||
14062 | switch (SVT.SimpleTy) { | |||
14063 | default: | |||
14064 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14064); | |||
14065 | case MVT::v4i8: | |||
14066 | case MVT::v4i16: | |||
14067 | case MVT::v8i8: | |||
14068 | case MVT::v8i16: { | |||
14069 | MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements()); | |||
14070 | return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), | |||
14071 | DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0)); | |||
14072 | } | |||
14073 | case MVT::v4i32: | |||
14074 | case MVT::v8i32: | |||
14075 | return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget); | |||
14076 | } | |||
14077 | llvm_unreachable(nullptr)::llvm::llvm_unreachable_internal(nullptr, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14077); | |||
14078 | } | |||
14079 | ||||
14080 | SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, | |||
14081 | SelectionDAG &DAG) const { | |||
14082 | SDValue N0 = Op.getOperand(0); | |||
14083 | SDLoc dl(Op); | |||
14084 | ||||
14085 | if (Op.getValueType().isVector()) | |||
14086 | return lowerUINT_TO_FP_vec(Op, DAG); | |||
14087 | ||||
14088 | // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't | |||
14089 | // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform | |||
14090 | // the optimization here. | |||
14091 | if (DAG.SignBitIsZero(N0)) | |||
14092 | return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0); | |||
14093 | ||||
14094 | MVT SrcVT = N0.getSimpleValueType(); | |||
14095 | MVT DstVT = Op.getSimpleValueType(); | |||
14096 | if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64) | |||
14097 | return LowerUINT_TO_FP_i64(Op, DAG); | |||
14098 | if (SrcVT == MVT::i32 && X86ScalarSSEf64) | |||
14099 | return LowerUINT_TO_FP_i32(Op, DAG); | |||
14100 | if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32) | |||
14101 | return SDValue(); | |||
14102 | ||||
14103 | // Make a 64-bit buffer, and use it to build an FILD. | |||
14104 | SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64); | |||
14105 | if (SrcVT == MVT::i32) { | |||
14106 | SDValue WordOff = DAG.getConstant(4, getPointerTy()); | |||
14107 | SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, | |||
14108 | getPointerTy(), StackSlot, WordOff); | |||
14109 | SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0), | |||
14110 | StackSlot, MachinePointerInfo(), | |||
14111 | false, false, 0); | |||
14112 | SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32), | |||
14113 | OffsetSlot, MachinePointerInfo(), | |||
14114 | false, false, 0); | |||
14115 | SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG); | |||
14116 | return Fild; | |||
14117 | } | |||
14118 | ||||
14119 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14119, __PRETTY_FUNCTION__)); | |||
14120 | SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0), | |||
14121 | StackSlot, MachinePointerInfo(), | |||
14122 | false, false, 0); | |||
14123 | // For i64 source, we need to add the appropriate power of 2 if the input | |||
14124 | // was negative. This is the same as the optimization in | |||
14125 | // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here, | |||
14126 | // we must be careful to do the computation in x87 extended precision, not | |||
14127 | // in SSE. (The generic code can't know it's OK to do this, or how to.) | |||
14128 | int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex(); | |||
14129 | MachineMemOperand *MMO = | |||
14130 | DAG.getMachineFunction() | |||
14131 | .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI), | |||
14132 | MachineMemOperand::MOLoad, 8, 8); | |||
14133 | ||||
14134 | SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other); | |||
14135 | SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) }; | |||
14136 | SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, | |||
14137 | MVT::i64, MMO); | |||
14138 | ||||
14139 | APInt FF(32, 0x5F800000ULL); | |||
14140 | ||||
14141 | // Check whether the sign bit is set. | |||
14142 | SDValue SignSet = DAG.getSetCC(dl, | |||
14143 | getSetCCResultType(*DAG.getContext(), MVT::i64), | |||
14144 | Op.getOperand(0), DAG.getConstant(0, MVT::i64), | |||
14145 | ISD::SETLT); | |||
14146 | ||||
14147 | // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits. | |||
14148 | SDValue FudgePtr = DAG.getConstantPool( | |||
14149 | ConstantInt::get(*DAG.getContext(), FF.zext(64)), | |||
14150 | getPointerTy()); | |||
14151 | ||||
14152 | // Get a pointer to FF if the sign bit was set, or to 0 otherwise. | |||
14153 | SDValue Zero = DAG.getIntPtrConstant(0); | |||
14154 | SDValue Four = DAG.getIntPtrConstant(4); | |||
14155 | SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet, | |||
14156 | Zero, Four); | |||
14157 | FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset); | |||
14158 | ||||
14159 | // Load the value out, extending it from f32 to f80. | |||
14160 | // FIXME: Avoid the extend by constructing the right constant pool? | |||
14161 | SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), | |||
14162 | FudgePtr, MachinePointerInfo::getConstantPool(), | |||
14163 | MVT::f32, false, false, false, 4); | |||
14164 | // Extend everything to 80 bits to force it to be done on x87. | |||
14165 | SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge); | |||
14166 | return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0)); | |||
14167 | } | |||
14168 | ||||
14169 | std::pair<SDValue,SDValue> | |||
14170 | X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, | |||
14171 | bool IsSigned, bool IsReplace) const { | |||
14172 | SDLoc DL(Op); | |||
14173 | ||||
14174 | EVT DstTy = Op.getValueType(); | |||
14175 | ||||
14176 | if (!IsSigned && !isIntegerTypeFTOL(DstTy)) { | |||
14177 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14177, __PRETTY_FUNCTION__)); | |||
14178 | DstTy = MVT::i64; | |||
14179 | } | |||
14180 | ||||
14181 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14183, __PRETTY_FUNCTION__)) | |||
14182 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14183, __PRETTY_FUNCTION__)) | |||
14183 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14183, __PRETTY_FUNCTION__)); | |||
14184 | ||||
14185 | // These are really Legal. | |||
14186 | if (DstTy == MVT::i32 && | |||
14187 | isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) | |||
14188 | return std::make_pair(SDValue(), SDValue()); | |||
14189 | if (Subtarget->is64Bit() && | |||
14190 | DstTy == MVT::i64 && | |||
14191 | isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) | |||
14192 | return std::make_pair(SDValue(), SDValue()); | |||
14193 | ||||
14194 | // We lower FP->int64 either into FISTP64 followed by a load from a temporary | |||
14195 | // stack slot, or into the FTOL runtime function. | |||
14196 | MachineFunction &MF = DAG.getMachineFunction(); | |||
14197 | unsigned MemSize = DstTy.getSizeInBits()/8; | |||
14198 | int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false); | |||
14199 | SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy()); | |||
14200 | ||||
14201 | unsigned Opc; | |||
14202 | if (!IsSigned && isIntegerTypeFTOL(DstTy)) | |||
14203 | Opc = X86ISD::WIN_FTOL; | |||
14204 | else | |||
14205 | switch (DstTy.getSimpleVT().SimpleTy) { | |||
14206 | default: llvm_unreachable("Invalid FP_TO_SINT to lower!")::llvm::llvm_unreachable_internal("Invalid FP_TO_SINT to lower!" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14206); | |||
14207 | case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break; | |||
14208 | case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break; | |||
14209 | case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break; | |||
14210 | } | |||
14211 | ||||
14212 | SDValue Chain = DAG.getEntryNode(); | |||
14213 | SDValue Value = Op.getOperand(0); | |||
14214 | EVT TheVT = Op.getOperand(0).getValueType(); | |||
14215 | // FIXME This causes a redundant load/store if the SSE-class value is already | |||
14216 | // in memory, such as if it is on the callstack. | |||
14217 | if (isScalarFPTypeInSSEReg(TheVT)) { | |||
14218 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14218, __PRETTY_FUNCTION__)); | |||
14219 | Chain = DAG.getStore(Chain, DL, Value, StackSlot, | |||
14220 | MachinePointerInfo::getFixedStack(SSFI), | |||
14221 | false, false, 0); | |||
14222 | SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other); | |||
14223 | SDValue Ops[] = { | |||
14224 | Chain, StackSlot, DAG.getValueType(TheVT) | |||
14225 | }; | |||
14226 | ||||
14227 | MachineMemOperand *MMO = | |||
14228 | MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI), | |||
14229 | MachineMemOperand::MOLoad, MemSize, MemSize); | |||
14230 | Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO); | |||
14231 | Chain = Value.getValue(1); | |||
14232 | SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false); | |||
14233 | StackSlot = DAG.getFrameIndex(SSFI, getPointerTy()); | |||
14234 | } | |||
14235 | ||||
14236 | MachineMemOperand *MMO = | |||
14237 | MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI), | |||
14238 | MachineMemOperand::MOStore, MemSize, MemSize); | |||
14239 | ||||
14240 | if (Opc != X86ISD::WIN_FTOL) { | |||
14241 | // Build the FP_TO_INT*_IN_MEM | |||
14242 | SDValue Ops[] = { Chain, Value, StackSlot }; | |||
14243 | SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other), | |||
14244 | Ops, DstTy, MMO); | |||
14245 | return std::make_pair(FIST, StackSlot); | |||
14246 | } else { | |||
14247 | SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL, | |||
14248 | DAG.getVTList(MVT::Other, MVT::Glue), | |||
14249 | Chain, Value); | |||
14250 | SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX, | |||
14251 | MVT::i32, ftol.getValue(1)); | |||
14252 | SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX, | |||
14253 | MVT::i32, eax.getValue(2)); | |||
14254 | SDValue Ops[] = { eax, edx }; | |||
14255 | SDValue pair = IsReplace | |||
14256 | ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops) | |||
14257 | : DAG.getMergeValues(Ops, DL); | |||
14258 | return std::make_pair(pair, SDValue()); | |||
14259 | } | |||
14260 | } | |||
14261 | ||||
14262 | static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG, | |||
14263 | const X86Subtarget *Subtarget) { | |||
14264 | MVT VT = Op->getSimpleValueType(0); | |||
14265 | SDValue In = Op->getOperand(0); | |||
14266 | MVT InVT = In.getSimpleValueType(); | |||
14267 | SDLoc dl(Op); | |||
14268 | ||||
14269 | // Optimize vectors in AVX mode: | |||
14270 | // | |||
14271 | // v8i16 -> v8i32 | |||
14272 | // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32. | |||
14273 | // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32. | |||
14274 | // Concat upper and lower parts. | |||
14275 | // | |||
14276 | // v4i32 -> v4i64 | |||
14277 | // Use vpunpckldq for 4 lower elements v4i32 -> v2i64. | |||
14278 | // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64. | |||
14279 | // Concat upper and lower parts. | |||
14280 | // | |||
14281 | ||||
14282 | if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) && | |||
14283 | ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) && | |||
14284 | ((VT != MVT::v4i64) || (InVT != MVT::v4i32))) | |||
14285 | return SDValue(); | |||
14286 | ||||
14287 | if (Subtarget->hasInt256()) | |||
14288 | return DAG.getNode(X86ISD::VZEXT, dl, VT, In); | |||
14289 | ||||
14290 | SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl); | |||
14291 | SDValue Undef = DAG.getUNDEF(InVT); | |||
14292 | bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND; | |||
14293 | SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef); | |||
14294 | SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef); | |||
14295 | ||||
14296 | MVT HVT = MVT::getVectorVT(VT.getVectorElementType(), | |||
14297 | VT.getVectorNumElements()/2); | |||
14298 | ||||
14299 | OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo); | |||
14300 | OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi); | |||
14301 | ||||
14302 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi); | |||
14303 | } | |||
14304 | ||||
14305 | static SDValue LowerZERO_EXTEND_AVX512(SDValue Op, | |||
14306 | SelectionDAG &DAG) { | |||
14307 | MVT VT = Op->getSimpleValueType(0); | |||
14308 | SDValue In = Op->getOperand(0); | |||
14309 | MVT InVT = In.getSimpleValueType(); | |||
14310 | SDLoc DL(Op); | |||
14311 | unsigned int NumElts = VT.getVectorNumElements(); | |||
14312 | if (NumElts != 8 && NumElts != 16) | |||
14313 | return SDValue(); | |||
14314 | ||||
14315 | if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) | |||
14316 | return DAG.getNode(X86ISD::VZEXT, DL, VT, In); | |||
14317 | ||||
14318 | EVT ExtVT = (NumElts == 8)? MVT::v8i64 : MVT::v16i32; | |||
14319 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
14320 | // Now we have only mask extension | |||
14321 | assert(InVT.getVectorElementType() == MVT::i1)((InVT.getVectorElementType() == MVT::i1) ? static_cast<void > (0) : __assert_fail ("InVT.getVectorElementType() == MVT::i1" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14321, __PRETTY_FUNCTION__)); | |||
14322 | SDValue Cst = DAG.getTargetConstant(1, ExtVT.getScalarType()); | |||
14323 | const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue(); | |||
14324 | SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy()); | |||
14325 | unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment(); | |||
14326 | SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP, | |||
14327 | MachinePointerInfo::getConstantPool(), | |||
14328 | false, false, false, Alignment); | |||
14329 | ||||
14330 | SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, DL, ExtVT, In, Ld); | |||
14331 | if (VT.is512BitVector()) | |||
14332 | return Brcst; | |||
14333 | return DAG.getNode(X86ISD::VTRUNC, DL, VT, Brcst); | |||
14334 | } | |||
14335 | ||||
14336 | static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget, | |||
14337 | SelectionDAG &DAG) { | |||
14338 | if (Subtarget->hasFp256()) { | |||
14339 | SDValue Res = LowerAVXExtend(Op, DAG, Subtarget); | |||
14340 | if (Res.getNode()) | |||
14341 | return Res; | |||
14342 | } | |||
14343 | ||||
14344 | return SDValue(); | |||
14345 | } | |||
14346 | ||||
14347 | static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget, | |||
14348 | SelectionDAG &DAG) { | |||
14349 | SDLoc DL(Op); | |||
14350 | MVT VT = Op.getSimpleValueType(); | |||
14351 | SDValue In = Op.getOperand(0); | |||
14352 | MVT SVT = In.getSimpleValueType(); | |||
14353 | ||||
14354 | if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1) | |||
14355 | return LowerZERO_EXTEND_AVX512(Op, DAG); | |||
14356 | ||||
14357 | if (Subtarget->hasFp256()) { | |||
14358 | SDValue Res = LowerAVXExtend(Op, DAG, Subtarget); | |||
14359 | if (Res.getNode()) | |||
14360 | return Res; | |||
14361 | } | |||
14362 | ||||
14363 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14364, __PRETTY_FUNCTION__)) | |||
14364 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14364, __PRETTY_FUNCTION__)); | |||
14365 | return SDValue(); | |||
14366 | } | |||
14367 | ||||
14368 | SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const { | |||
14369 | SDLoc DL(Op); | |||
14370 | MVT VT = Op.getSimpleValueType(); | |||
14371 | SDValue In = Op.getOperand(0); | |||
14372 | MVT InVT = In.getSimpleValueType(); | |||
14373 | ||||
14374 | if (VT == MVT::i1) { | |||
14375 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14376, __PRETTY_FUNCTION__)) | |||
14376 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14376, __PRETTY_FUNCTION__)); | |||
14377 | if (InVT.getSizeInBits() >= 32) | |||
14378 | return SDValue(); | |||
14379 | In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In); | |||
14380 | return DAG.getNode(ISD::TRUNCATE, DL, VT, In); | |||
14381 | } | |||
14382 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14383, __PRETTY_FUNCTION__)) | |||
14383 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14383, __PRETTY_FUNCTION__)); | |||
14384 | ||||
14385 | if (InVT.is512BitVector() || VT.getVectorElementType() == MVT::i1) { | |||
14386 | if (VT.getVectorElementType().getSizeInBits() >=8) | |||
14387 | return DAG.getNode(X86ISD::VTRUNC, DL, VT, In); | |||
14388 | ||||
14389 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14389, __PRETTY_FUNCTION__)); | |||
14390 | unsigned NumElts = InVT.getVectorNumElements(); | |||
14391 | assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type")(((NumElts == 8 || NumElts == 16) && "Unexpected vector type" ) ? static_cast<void> (0) : __assert_fail ("(NumElts == 8 || NumElts == 16) && \"Unexpected vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14391, __PRETTY_FUNCTION__)); | |||
14392 | if (InVT.getSizeInBits() < 512) { | |||
14393 | MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64; | |||
14394 | In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In); | |||
14395 | InVT = ExtVT; | |||
14396 | } | |||
14397 | ||||
14398 | SDValue Cst = DAG.getTargetConstant(1, InVT.getVectorElementType()); | |||
14399 | const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue(); | |||
14400 | SDValue CP = DAG.getConstantPool(C, getPointerTy()); | |||
14401 | unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment(); | |||
14402 | SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP, | |||
14403 | MachinePointerInfo::getConstantPool(), | |||
14404 | false, false, false, Alignment); | |||
14405 | SDValue OneV = DAG.getNode(X86ISD::VBROADCAST, DL, InVT, Ld); | |||
14406 | SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In); | |||
14407 | return DAG.getNode(X86ISD::TESTM, DL, VT, And, And); | |||
14408 | } | |||
14409 | ||||
14410 | if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) { | |||
14411 | // On AVX2, v4i64 -> v4i32 becomes VPERMD. | |||
14412 | if (Subtarget->hasInt256()) { | |||
14413 | static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1}; | |||
14414 | In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In); | |||
14415 | In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32), | |||
14416 | ShufMask); | |||
14417 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In, | |||
14418 | DAG.getIntPtrConstant(0)); | |||
14419 | } | |||
14420 | ||||
14421 | SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In, | |||
14422 | DAG.getIntPtrConstant(0)); | |||
14423 | SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In, | |||
14424 | DAG.getIntPtrConstant(2)); | |||
14425 | OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo); | |||
14426 | OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi); | |||
14427 | static const int ShufMask[] = {0, 2, 4, 6}; | |||
14428 | return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask); | |||
14429 | } | |||
14430 | ||||
14431 | if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) { | |||
14432 | // On AVX2, v8i32 -> v8i16 becomed PSHUFB. | |||
14433 | if (Subtarget->hasInt256()) { | |||
14434 | In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In); | |||
14435 | ||||
14436 | SmallVector<SDValue,32> pshufbMask; | |||
14437 | for (unsigned i = 0; i < 2; ++i) { | |||
14438 | pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8)); | |||
14439 | pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8)); | |||
14440 | pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8)); | |||
14441 | pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8)); | |||
14442 | pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8)); | |||
14443 | pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8)); | |||
14444 | pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8)); | |||
14445 | pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8)); | |||
14446 | for (unsigned j = 0; j < 8; ++j) | |||
14447 | pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8)); | |||
14448 | } | |||
14449 | SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask); | |||
14450 | In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV); | |||
14451 | In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In); | |||
14452 | ||||
14453 | static const int ShufMask[] = {0, 2, -1, -1}; | |||
14454 | In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64), | |||
14455 | &ShufMask[0]); | |||
14456 | In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In, | |||
14457 | DAG.getIntPtrConstant(0)); | |||
14458 | return DAG.getNode(ISD::BITCAST, DL, VT, In); | |||
14459 | } | |||
14460 | ||||
14461 | SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In, | |||
14462 | DAG.getIntPtrConstant(0)); | |||
14463 | ||||
14464 | SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In, | |||
14465 | DAG.getIntPtrConstant(4)); | |||
14466 | ||||
14467 | OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo); | |||
14468 | OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi); | |||
14469 | ||||
14470 | // The PSHUFB mask: | |||
14471 | static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13, | |||
14472 | -1, -1, -1, -1, -1, -1, -1, -1}; | |||
14473 | ||||
14474 | SDValue Undef = DAG.getUNDEF(MVT::v16i8); | |||
14475 | OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1); | |||
14476 | OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1); | |||
14477 | ||||
14478 | OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo); | |||
14479 | OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi); | |||
14480 | ||||
14481 | // The MOVLHPS Mask: | |||
14482 | static const int ShufMask2[] = {0, 1, 4, 5}; | |||
14483 | SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2); | |||
14484 | return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res); | |||
14485 | } | |||
14486 | ||||
14487 | // Handle truncation of V256 to V128 using shuffles. | |||
14488 | if (!VT.is128BitVector() || !InVT.is256BitVector()) | |||
14489 | return SDValue(); | |||
14490 | ||||
14491 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14491, __PRETTY_FUNCTION__)); | |||
14492 | ||||
14493 | unsigned NumElems = VT.getVectorNumElements(); | |||
14494 | MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2); | |||
14495 | ||||
14496 | SmallVector<int, 16> MaskVec(NumElems * 2, -1); | |||
14497 | // Prepare truncation shuffle mask | |||
14498 | for (unsigned i = 0; i != NumElems; ++i) | |||
14499 | MaskVec[i] = i * 2; | |||
14500 | SDValue V = DAG.getVectorShuffle(NVT, DL, | |||
14501 | DAG.getNode(ISD::BITCAST, DL, NVT, In), | |||
14502 | DAG.getUNDEF(NVT), &MaskVec[0]); | |||
14503 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V, | |||
14504 | DAG.getIntPtrConstant(0)); | |||
14505 | } | |||
14506 | ||||
14507 | SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, | |||
14508 | SelectionDAG &DAG) const { | |||
14509 | assert(!Op.getSimpleValueType().isVector())((!Op.getSimpleValueType().isVector()) ? static_cast<void> (0) : __assert_fail ("!Op.getSimpleValueType().isVector()", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14509, __PRETTY_FUNCTION__)); | |||
14510 | ||||
14511 | std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, | |||
14512 | /*IsSigned=*/ true, /*IsReplace=*/ false); | |||
14513 | SDValue FIST = Vals.first, StackSlot = Vals.second; | |||
14514 | // If FP_TO_INTHelper failed, the node is actually supposed to be Legal. | |||
14515 | if (!FIST.getNode()) return Op; | |||
14516 | ||||
14517 | if (StackSlot.getNode()) | |||
14518 | // Load the result. | |||
14519 | return DAG.getLoad(Op.getValueType(), SDLoc(Op), | |||
14520 | FIST, StackSlot, MachinePointerInfo(), | |||
14521 | false, false, false, 0); | |||
14522 | ||||
14523 | // The node is the result. | |||
14524 | return FIST; | |||
14525 | } | |||
14526 | ||||
14527 | SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, | |||
14528 | SelectionDAG &DAG) const { | |||
14529 | std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, | |||
14530 | /*IsSigned=*/ false, /*IsReplace=*/ false); | |||
14531 | SDValue FIST = Vals.first, StackSlot = Vals.second; | |||
14532 | assert(FIST.getNode() && "Unexpected failure")((FIST.getNode() && "Unexpected failure") ? static_cast <void> (0) : __assert_fail ("FIST.getNode() && \"Unexpected failure\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14532, __PRETTY_FUNCTION__)); | |||
14533 | ||||
14534 | if (StackSlot.getNode()) | |||
14535 | // Load the result. | |||
14536 | return DAG.getLoad(Op.getValueType(), SDLoc(Op), | |||
14537 | FIST, StackSlot, MachinePointerInfo(), | |||
14538 | false, false, false, 0); | |||
14539 | ||||
14540 | // The node is the result. | |||
14541 | return FIST; | |||
14542 | } | |||
14543 | ||||
14544 | static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) { | |||
14545 | SDLoc DL(Op); | |||
14546 | MVT VT = Op.getSimpleValueType(); | |||
14547 | SDValue In = Op.getOperand(0); | |||
14548 | MVT SVT = In.getSimpleValueType(); | |||
14549 | ||||
14550 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14550, __PRETTY_FUNCTION__)); | |||
14551 | ||||
14552 | return DAG.getNode(X86ISD::VFPEXT, DL, VT, | |||
14553 | DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32, | |||
14554 | In, DAG.getUNDEF(SVT))); | |||
14555 | } | |||
14556 | ||||
14557 | /// The only differences between FABS and FNEG are the mask and the logic op. | |||
14558 | /// FNEG also has a folding opportunity for FNEG(FABS(x)). | |||
14559 | static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) { | |||
14560 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14561, __PRETTY_FUNCTION__)) | |||
14561 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14561, __PRETTY_FUNCTION__)); | |||
14562 | ||||
14563 | bool IsFABS = (Op.getOpcode() == ISD::FABS); | |||
14564 | ||||
14565 | // If this is a FABS and it has an FNEG user, bail out to fold the combination | |||
14566 | // into an FNABS. We'll lower the FABS after that if it is still in use. | |||
14567 | if (IsFABS) | |||
14568 | for (SDNode *User : Op->uses()) | |||
14569 | if (User->getOpcode() == ISD::FNEG) | |||
14570 | return Op; | |||
14571 | ||||
14572 | SDValue Op0 = Op.getOperand(0); | |||
14573 | bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS); | |||
14574 | ||||
14575 | SDLoc dl(Op); | |||
14576 | MVT VT = Op.getSimpleValueType(); | |||
14577 | // Assume scalar op for initialization; update for vector if needed. | |||
14578 | // Note that there are no scalar bitwise logical SSE/AVX instructions, so we | |||
14579 | // generate a 16-byte vector constant and logic op even for the scalar case. | |||
14580 | // Using a 16-byte mask allows folding the load of the mask with | |||
14581 | // the logic op, so it can save (~4 bytes) on code size. | |||
14582 | MVT EltVT = VT; | |||
14583 | unsigned NumElts = VT == MVT::f64 ? 2 : 4; | |||
14584 | // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to | |||
14585 | // decide if we should generate a 16-byte constant mask when we only need 4 or | |||
14586 | // 8 bytes for the scalar case. | |||
14587 | if (VT.isVector()) { | |||
14588 | EltVT = VT.getVectorElementType(); | |||
14589 | NumElts = VT.getVectorNumElements(); | |||
14590 | } | |||
14591 | ||||
14592 | unsigned EltBits = EltVT.getSizeInBits(); | |||
14593 | LLVMContext *Context = DAG.getContext(); | |||
14594 | // For FABS, mask is 0x7f...; for FNEG, mask is 0x80... | |||
14595 | APInt MaskElt = | |||
14596 | IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits); | |||
14597 | Constant *C = ConstantInt::get(*Context, MaskElt); | |||
14598 | C = ConstantVector::getSplat(NumElts, C); | |||
14599 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
14600 | SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy()); | |||
14601 | unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment(); | |||
14602 | SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx, | |||
14603 | MachinePointerInfo::getConstantPool(), | |||
14604 | false, false, false, Alignment); | |||
14605 | ||||
14606 | if (VT.isVector()) { | |||
14607 | // For a vector, cast operands to a vector type, perform the logic op, | |||
14608 | // and cast the result back to the original value type. | |||
14609 | MVT VecVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64); | |||
14610 | SDValue MaskCasted = DAG.getNode(ISD::BITCAST, dl, VecVT, Mask); | |||
14611 | SDValue Operand = IsFNABS ? | |||
14612 | DAG.getNode(ISD::BITCAST, dl, VecVT, Op0.getOperand(0)) : | |||
14613 | DAG.getNode(ISD::BITCAST, dl, VecVT, Op0); | |||
14614 | unsigned BitOp = IsFABS ? ISD::AND : IsFNABS ? ISD::OR : ISD::XOR; | |||
14615 | return DAG.getNode(ISD::BITCAST, dl, VT, | |||
14616 | DAG.getNode(BitOp, dl, VecVT, Operand, MaskCasted)); | |||
14617 | } | |||
14618 | ||||
14619 | // If not vector, then scalar. | |||
14620 | unsigned BitOp = IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR; | |||
14621 | SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0; | |||
14622 | return DAG.getNode(BitOp, dl, VT, Operand, Mask); | |||
14623 | } | |||
14624 | ||||
14625 | static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) { | |||
14626 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
14627 | LLVMContext *Context = DAG.getContext(); | |||
14628 | SDValue Op0 = Op.getOperand(0); | |||
14629 | SDValue Op1 = Op.getOperand(1); | |||
14630 | SDLoc dl(Op); | |||
14631 | MVT VT = Op.getSimpleValueType(); | |||
14632 | MVT SrcVT = Op1.getSimpleValueType(); | |||
14633 | ||||
14634 | // If second operand is smaller, extend it first. | |||
14635 | if (SrcVT.bitsLT(VT)) { | |||
14636 | Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1); | |||
14637 | SrcVT = VT; | |||
14638 | } | |||
14639 | // And if it is bigger, shrink it first. | |||
14640 | if (SrcVT.bitsGT(VT)) { | |||
14641 | Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1)); | |||
14642 | SrcVT = VT; | |||
14643 | } | |||
14644 | ||||
14645 | // At this point the operands and the result should have the same | |||
14646 | // type, and that won't be f80 since that is not custom lowered. | |||
14647 | ||||
14648 | const fltSemantics &Sem = | |||
14649 | VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle; | |||
14650 | const unsigned SizeInBits = VT.getSizeInBits(); | |||
14651 | ||||
14652 | SmallVector<Constant *, 4> CV( | |||
14653 | VT == MVT::f64 ? 2 : 4, | |||
14654 | ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0)))); | |||
14655 | ||||
14656 | // First, clear all bits but the sign bit from the second operand (sign). | |||
14657 | CV[0] = ConstantFP::get(*Context, | |||
14658 | APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1))); | |||
14659 | Constant *C = ConstantVector::get(CV); | |||
14660 | SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16); | |||
14661 | SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx, | |||
14662 | MachinePointerInfo::getConstantPool(), | |||
14663 | false, false, false, 16); | |||
14664 | SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1); | |||
14665 | ||||
14666 | // Next, clear the sign bit from the first operand (magnitude). | |||
14667 | // If it's a constant, we can clear it here. | |||
14668 | if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) { | |||
14669 | APFloat APF = Op0CN->getValueAPF(); | |||
14670 | // If the magnitude is a positive zero, the sign bit alone is enough. | |||
14671 | if (APF.isPosZero()) | |||
14672 | return SignBit; | |||
14673 | APF.clearSign(); | |||
14674 | CV[0] = ConstantFP::get(*Context, APF); | |||
14675 | } else { | |||
14676 | CV[0] = ConstantFP::get( | |||
14677 | *Context, | |||
14678 | APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1))); | |||
14679 | } | |||
14680 | C = ConstantVector::get(CV); | |||
14681 | CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16); | |||
14682 | SDValue Val = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx, | |||
14683 | MachinePointerInfo::getConstantPool(), | |||
14684 | false, false, false, 16); | |||
14685 | // If the magnitude operand wasn't a constant, we need to AND out the sign. | |||
14686 | if (!isa<ConstantFPSDNode>(Op0)) | |||
14687 | Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Val); | |||
14688 | ||||
14689 | // OR the magnitude value with the sign bit. | |||
14690 | return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit); | |||
14691 | } | |||
14692 | ||||
14693 | static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) { | |||
14694 | SDValue N0 = Op.getOperand(0); | |||
14695 | SDLoc dl(Op); | |||
14696 | MVT VT = Op.getSimpleValueType(); | |||
14697 | ||||
14698 | // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1). | |||
14699 | SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0, | |||
14700 | DAG.getConstant(1, VT)); | |||
14701 | return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT)); | |||
14702 | } | |||
14703 | ||||
14704 | // Check whether an OR'd tree is PTEST-able. | |||
14705 | static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget, | |||
14706 | SelectionDAG &DAG) { | |||
14707 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14707, __PRETTY_FUNCTION__)); | |||
14708 | ||||
14709 | if (!Subtarget->hasSSE41()) | |||
14710 | return SDValue(); | |||
14711 | ||||
14712 | if (!Op->hasOneUse()) | |||
14713 | return SDValue(); | |||
14714 | ||||
14715 | SDNode *N = Op.getNode(); | |||
14716 | SDLoc DL(N); | |||
14717 | ||||
14718 | SmallVector<SDValue, 8> Opnds; | |||
14719 | DenseMap<SDValue, unsigned> VecInMap; | |||
14720 | SmallVector<SDValue, 8> VecIns; | |||
14721 | EVT VT = MVT::Other; | |||
14722 | ||||
14723 | // Recognize a special case where a vector is casted into wide integer to | |||
14724 | // test all 0s. | |||
14725 | Opnds.push_back(N->getOperand(0)); | |||
14726 | Opnds.push_back(N->getOperand(1)); | |||
14727 | ||||
14728 | for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) { | |||
14729 | SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot; | |||
14730 | // BFS traverse all OR'd operands. | |||
14731 | if (I->getOpcode() == ISD::OR) { | |||
14732 | Opnds.push_back(I->getOperand(0)); | |||
14733 | Opnds.push_back(I->getOperand(1)); | |||
14734 | // Re-evaluate the number of nodes to be traversed. | |||
14735 | e += 2; // 2 more nodes (LHS and RHS) are pushed. | |||
14736 | continue; | |||
14737 | } | |||
14738 | ||||
14739 | // Quit if a non-EXTRACT_VECTOR_ELT | |||
14740 | if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT) | |||
14741 | return SDValue(); | |||
14742 | ||||
14743 | // Quit if without a constant index. | |||
14744 | SDValue Idx = I->getOperand(1); | |||
14745 | if (!isa<ConstantSDNode>(Idx)) | |||
14746 | return SDValue(); | |||
14747 | ||||
14748 | SDValue ExtractedFromVec = I->getOperand(0); | |||
14749 | DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec); | |||
14750 | if (M == VecInMap.end()) { | |||
14751 | VT = ExtractedFromVec.getValueType(); | |||
14752 | // Quit if not 128/256-bit vector. | |||
14753 | if (!VT.is128BitVector() && !VT.is256BitVector()) | |||
14754 | return SDValue(); | |||
14755 | // Quit if not the same type. | |||
14756 | if (VecInMap.begin() != VecInMap.end() && | |||
14757 | VT != VecInMap.begin()->first.getValueType()) | |||
14758 | return SDValue(); | |||
14759 | M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first; | |||
14760 | VecIns.push_back(ExtractedFromVec); | |||
14761 | } | |||
14762 | M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue(); | |||
14763 | } | |||
14764 | ||||
14765 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14766, __PRETTY_FUNCTION__)) | |||
14766 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14766, __PRETTY_FUNCTION__)); | |||
14767 | ||||
14768 | unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U; | |||
14769 | ||||
14770 | for (DenseMap<SDValue, unsigned>::const_iterator | |||
14771 | I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) { | |||
14772 | // Quit if not all elements are used. | |||
14773 | if (I->second != FullMask) | |||
14774 | return SDValue(); | |||
14775 | } | |||
14776 | ||||
14777 | EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64; | |||
14778 | ||||
14779 | // Cast all vectors into TestVT for PTEST. | |||
14780 | for (unsigned i = 0, e = VecIns.size(); i < e; ++i) | |||
14781 | VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]); | |||
14782 | ||||
14783 | // If more than one full vectors are evaluated, OR them first before PTEST. | |||
14784 | for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) { | |||
14785 | // Each iteration will OR 2 nodes and append the result until there is only | |||
14786 | // 1 node left, i.e. the final OR'd value of all vectors. | |||
14787 | SDValue LHS = VecIns[Slot]; | |||
14788 | SDValue RHS = VecIns[Slot + 1]; | |||
14789 | VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS)); | |||
14790 | } | |||
14791 | ||||
14792 | return DAG.getNode(X86ISD::PTEST, DL, MVT::i32, | |||
14793 | VecIns.back(), VecIns.back()); | |||
14794 | } | |||
14795 | ||||
14796 | /// \brief return true if \c Op has a use that doesn't just read flags. | |||
14797 | static bool hasNonFlagsUse(SDValue Op) { | |||
14798 | for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE; | |||
14799 | ++UI) { | |||
14800 | SDNode *User = *UI; | |||
14801 | unsigned UOpNo = UI.getOperandNo(); | |||
14802 | if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) { | |||
14803 | // Look pass truncate. | |||
14804 | UOpNo = User->use_begin().getOperandNo(); | |||
14805 | User = *User->use_begin(); | |||
14806 | } | |||
14807 | ||||
14808 | if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC && | |||
14809 | !(User->getOpcode() == ISD::SELECT && UOpNo == 0)) | |||
14810 | return true; | |||
14811 | } | |||
14812 | return false; | |||
14813 | } | |||
14814 | ||||
14815 | /// Emit nodes that will be selected as "test Op0,Op0", or something | |||
14816 | /// equivalent. | |||
14817 | SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl, | |||
14818 | SelectionDAG &DAG) const { | |||
14819 | if (Op.getValueType() == MVT::i1) | |||
14820 | // KORTEST instruction should be selected | |||
14821 | return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op, | |||
14822 | DAG.getConstant(0, Op.getValueType())); | |||
14823 | ||||
14824 | // CF and OF aren't always set the way we want. Determine which | |||
14825 | // of these we need. | |||
14826 | bool NeedCF = false; | |||
14827 | bool NeedOF = false; | |||
14828 | switch (X86CC) { | |||
14829 | default: break; | |||
14830 | case X86::COND_A: case X86::COND_AE: | |||
14831 | case X86::COND_B: case X86::COND_BE: | |||
14832 | NeedCF = true; | |||
14833 | break; | |||
14834 | case X86::COND_G: case X86::COND_GE: | |||
14835 | case X86::COND_L: case X86::COND_LE: | |||
14836 | case X86::COND_O: case X86::COND_NO: { | |||
14837 | // Check if we really need to set the | |||
14838 | // Overflow flag. If NoSignedWrap is present | |||
14839 | // that is not actually needed. | |||
14840 | switch (Op->getOpcode()) { | |||
14841 | case ISD::ADD: | |||
14842 | case ISD::SUB: | |||
14843 | case ISD::MUL: | |||
14844 | case ISD::SHL: { | |||
14845 | const BinaryWithFlagsSDNode *BinNode = | |||
14846 | cast<BinaryWithFlagsSDNode>(Op.getNode()); | |||
14847 | if (BinNode->hasNoSignedWrap()) | |||
14848 | break; | |||
14849 | } | |||
14850 | default: | |||
14851 | NeedOF = true; | |||
14852 | break; | |||
14853 | } | |||
14854 | break; | |||
14855 | } | |||
14856 | } | |||
14857 | // See if we can use the EFLAGS value from the operand instead of | |||
14858 | // doing a separate TEST. TEST always sets OF and CF to 0, so unless | |||
14859 | // we prove that the arithmetic won't overflow, we can't use OF or CF. | |||
14860 | if (Op.getResNo() != 0 || NeedOF || NeedCF) { | |||
14861 | // Emit a CMP with 0, which is the TEST pattern. | |||
14862 | //if (Op.getValueType() == MVT::i1) | |||
14863 | // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op, | |||
14864 | // DAG.getConstant(0, MVT::i1)); | |||
14865 | return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op, | |||
14866 | DAG.getConstant(0, Op.getValueType())); | |||
14867 | } | |||
14868 | unsigned Opcode = 0; | |||
14869 | unsigned NumOperands = 0; | |||
14870 | ||||
14871 | // Truncate operations may prevent the merge of the SETCC instruction | |||
14872 | // and the arithmetic instruction before it. Attempt to truncate the operands | |||
14873 | // of the arithmetic instruction and use a reduced bit-width instruction. | |||
14874 | bool NeedTruncation = false; | |||
14875 | SDValue ArithOp = Op; | |||
14876 | if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) { | |||
14877 | SDValue Arith = Op->getOperand(0); | |||
14878 | // Both the trunc and the arithmetic op need to have one user each. | |||
14879 | if (Arith->hasOneUse()) | |||
14880 | switch (Arith.getOpcode()) { | |||
14881 | default: break; | |||
14882 | case ISD::ADD: | |||
14883 | case ISD::SUB: | |||
14884 | case ISD::AND: | |||
14885 | case ISD::OR: | |||
14886 | case ISD::XOR: { | |||
14887 | NeedTruncation = true; | |||
14888 | ArithOp = Arith; | |||
14889 | } | |||
14890 | } | |||
14891 | } | |||
14892 | ||||
14893 | // NOTICE: In the code below we use ArithOp to hold the arithmetic operation | |||
14894 | // which may be the result of a CAST. We use the variable 'Op', which is the | |||
14895 | // non-casted variable when we check for possible users. | |||
14896 | switch (ArithOp.getOpcode()) { | |||
14897 | case ISD::ADD: | |||
14898 | // Due to an isel shortcoming, be conservative if this add is likely to be | |||
14899 | // selected as part of a load-modify-store instruction. When the root node | |||
14900 | // in a match is a store, isel doesn't know how to remap non-chain non-flag | |||
14901 | // uses of other nodes in the match, such as the ADD in this case. This | |||
14902 | // leads to the ADD being left around and reselected, with the result being | |||
14903 | // two adds in the output. Alas, even if none our users are stores, that | |||
14904 | // doesn't prove we're O.K. Ergo, if we have any parents that aren't | |||
14905 | // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require | |||
14906 | // climbing the DAG back to the root, and it doesn't seem to be worth the | |||
14907 | // effort. | |||
14908 | for (SDNode::use_iterator UI = Op.getNode()->use_begin(), | |||
14909 | UE = Op.getNode()->use_end(); UI != UE; ++UI) | |||
14910 | if (UI->getOpcode() != ISD::CopyToReg && | |||
14911 | UI->getOpcode() != ISD::SETCC && | |||
14912 | UI->getOpcode() != ISD::STORE) | |||
14913 | goto default_case; | |||
14914 | ||||
14915 | if (ConstantSDNode *C = | |||
14916 | dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) { | |||
14917 | // An add of one will be selected as an INC. | |||
14918 | if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) { | |||
14919 | Opcode = X86ISD::INC; | |||
14920 | NumOperands = 1; | |||
14921 | break; | |||
14922 | } | |||
14923 | ||||
14924 | // An add of negative one (subtract of one) will be selected as a DEC. | |||
14925 | if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) { | |||
14926 | Opcode = X86ISD::DEC; | |||
14927 | NumOperands = 1; | |||
14928 | break; | |||
14929 | } | |||
14930 | } | |||
14931 | ||||
14932 | // Otherwise use a regular EFLAGS-setting add. | |||
14933 | Opcode = X86ISD::ADD; | |||
14934 | NumOperands = 2; | |||
14935 | break; | |||
14936 | case ISD::SHL: | |||
14937 | case ISD::SRL: | |||
14938 | // If we have a constant logical shift that's only used in a comparison | |||
14939 | // against zero turn it into an equivalent AND. This allows turning it into | |||
14940 | // a TEST instruction later. | |||
14941 | if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() && | |||
14942 | isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) { | |||
14943 | EVT VT = Op.getValueType(); | |||
14944 | unsigned BitWidth = VT.getSizeInBits(); | |||
14945 | unsigned ShAmt = Op->getConstantOperandVal(1); | |||
14946 | if (ShAmt >= BitWidth) // Avoid undefined shifts. | |||
14947 | break; | |||
14948 | APInt Mask = ArithOp.getOpcode() == ISD::SRL | |||
14949 | ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt) | |||
14950 | : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt); | |||
14951 | if (!Mask.isSignedIntN(32)) // Avoid large immediates. | |||
14952 | break; | |||
14953 | SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0), | |||
14954 | DAG.getConstant(Mask, VT)); | |||
14955 | DAG.ReplaceAllUsesWith(Op, New); | |||
14956 | Op = New; | |||
14957 | } | |||
14958 | break; | |||
14959 | ||||
14960 | case ISD::AND: | |||
14961 | // If the primary and result isn't used, don't bother using X86ISD::AND, | |||
14962 | // because a TEST instruction will be better. | |||
14963 | if (!hasNonFlagsUse(Op)) | |||
14964 | break; | |||
14965 | // FALL THROUGH | |||
14966 | case ISD::SUB: | |||
14967 | case ISD::OR: | |||
14968 | case ISD::XOR: | |||
14969 | // Due to the ISEL shortcoming noted above, be conservative if this op is | |||
14970 | // likely to be selected as part of a load-modify-store instruction. | |||
14971 | for (SDNode::use_iterator UI = Op.getNode()->use_begin(), | |||
14972 | UE = Op.getNode()->use_end(); UI != UE; ++UI) | |||
14973 | if (UI->getOpcode() == ISD::STORE) | |||
14974 | goto default_case; | |||
14975 | ||||
14976 | // Otherwise use a regular EFLAGS-setting instruction. | |||
14977 | switch (ArithOp.getOpcode()) { | |||
14978 | default: llvm_unreachable("unexpected operator!")::llvm::llvm_unreachable_internal("unexpected operator!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 14978); | |||
14979 | case ISD::SUB: Opcode = X86ISD::SUB; break; | |||
14980 | case ISD::XOR: Opcode = X86ISD::XOR; break; | |||
14981 | case ISD::AND: Opcode = X86ISD::AND; break; | |||
14982 | case ISD::OR: { | |||
14983 | if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) { | |||
14984 | SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG); | |||
14985 | if (EFLAGS.getNode()) | |||
14986 | return EFLAGS; | |||
14987 | } | |||
14988 | Opcode = X86ISD::OR; | |||
14989 | break; | |||
14990 | } | |||
14991 | } | |||
14992 | ||||
14993 | NumOperands = 2; | |||
14994 | break; | |||
14995 | case X86ISD::ADD: | |||
14996 | case X86ISD::SUB: | |||
14997 | case X86ISD::INC: | |||
14998 | case X86ISD::DEC: | |||
14999 | case X86ISD::OR: | |||
15000 | case X86ISD::XOR: | |||
15001 | case X86ISD::AND: | |||
15002 | return SDValue(Op.getNode(), 1); | |||
15003 | default: | |||
15004 | default_case: | |||
15005 | break; | |||
15006 | } | |||
15007 | ||||
15008 | // If we found that truncation is beneficial, perform the truncation and | |||
15009 | // update 'Op'. | |||
15010 | if (NeedTruncation) { | |||
15011 | EVT VT = Op.getValueType(); | |||
15012 | SDValue WideVal = Op->getOperand(0); | |||
15013 | EVT WideVT = WideVal.getValueType(); | |||
15014 | unsigned ConvertedOp = 0; | |||
15015 | // Use a target machine opcode to prevent further DAGCombine | |||
15016 | // optimizations that may separate the arithmetic operations | |||
15017 | // from the setcc node. | |||
15018 | switch (WideVal.getOpcode()) { | |||
15019 | default: break; | |||
15020 | case ISD::ADD: ConvertedOp = X86ISD::ADD; break; | |||
15021 | case ISD::SUB: ConvertedOp = X86ISD::SUB; break; | |||
15022 | case ISD::AND: ConvertedOp = X86ISD::AND; break; | |||
15023 | case ISD::OR: ConvertedOp = X86ISD::OR; break; | |||
15024 | case ISD::XOR: ConvertedOp = X86ISD::XOR; break; | |||
15025 | } | |||
15026 | ||||
15027 | if (ConvertedOp) { | |||
15028 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
15029 | if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) { | |||
15030 | SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0)); | |||
15031 | SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1)); | |||
15032 | Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1); | |||
15033 | } | |||
15034 | } | |||
15035 | } | |||
15036 | ||||
15037 | if (Opcode == 0) | |||
15038 | // Emit a CMP with 0, which is the TEST pattern. | |||
15039 | return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op, | |||
15040 | DAG.getConstant(0, Op.getValueType())); | |||
15041 | ||||
15042 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32); | |||
15043 | SmallVector<SDValue, 4> Ops; | |||
15044 | for (unsigned i = 0; i != NumOperands; ++i) | |||
15045 | Ops.push_back(Op.getOperand(i)); | |||
15046 | ||||
15047 | SDValue New = DAG.getNode(Opcode, dl, VTs, Ops); | |||
15048 | DAG.ReplaceAllUsesWith(Op, New); | |||
15049 | return SDValue(New.getNode(), 1); | |||
15050 | } | |||
15051 | ||||
15052 | /// Emit nodes that will be selected as "cmp Op0,Op1", or something | |||
15053 | /// equivalent. | |||
15054 | SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, | |||
15055 | SDLoc dl, SelectionDAG &DAG) const { | |||
15056 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) { | |||
15057 | if (C->getAPIntValue() == 0) | |||
15058 | return EmitTest(Op0, X86CC, dl, DAG); | |||
15059 | ||||
15060 | if (Op0.getValueType() == MVT::i1) | |||
15061 | llvm_unreachable("Unexpected comparison operation for MVT::i1 operands")::llvm::llvm_unreachable_internal("Unexpected comparison operation for MVT::i1 operands" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15061); | |||
15062 | } | |||
15063 | ||||
15064 | if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 || | |||
15065 | Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) { | |||
15066 | // Do the comparison at i32 if it's smaller, besides the Atom case. | |||
15067 | // This avoids subregister aliasing issues. Keep the smaller reference | |||
15068 | // if we're optimizing for size, however, as that'll allow better folding | |||
15069 | // of memory operations. | |||
15070 | if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 && | |||
15071 | !DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute( | |||
15072 | AttributeSet::FunctionIndex, Attribute::MinSize) && | |||
15073 | !Subtarget->isAtom()) { | |||
15074 | unsigned ExtendOp = | |||
15075 | isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND; | |||
15076 | Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0); | |||
15077 | Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1); | |||
15078 | } | |||
15079 | // Use SUB instead of CMP to enable CSE between SUB and CMP. | |||
15080 | SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32); | |||
15081 | SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs, | |||
15082 | Op0, Op1); | |||
15083 | return SDValue(Sub.getNode(), 1); | |||
15084 | } | |||
15085 | return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1); | |||
15086 | } | |||
15087 | ||||
15088 | /// Convert a comparison if required by the subtarget. | |||
15089 | SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp, | |||
15090 | SelectionDAG &DAG) const { | |||
15091 | // If the subtarget does not support the FUCOMI instruction, floating-point | |||
15092 | // comparisons have to be converted. | |||
15093 | if (Subtarget->hasCMov() || | |||
15094 | Cmp.getOpcode() != X86ISD::CMP || | |||
15095 | !Cmp.getOperand(0).getValueType().isFloatingPoint() || | |||
15096 | !Cmp.getOperand(1).getValueType().isFloatingPoint()) | |||
15097 | return Cmp; | |||
15098 | ||||
15099 | // The instruction selector will select an FUCOM instruction instead of | |||
15100 | // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence | |||
15101 | // build an SDNode sequence that transfers the result from FPSW into EFLAGS: | |||
15102 | // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8)))) | |||
15103 | SDLoc dl(Cmp); | |||
15104 | SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp); | |||
15105 | SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW); | |||
15106 | SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW, | |||
15107 | DAG.getConstant(8, MVT::i8)); | |||
15108 | SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl); | |||
15109 | return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl); | |||
15110 | } | |||
15111 | ||||
15112 | /// The minimum architected relative accuracy is 2^-12. We need one | |||
15113 | /// Newton-Raphson step to have a good float result (24 bits of precision). | |||
15114 | SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op, | |||
15115 | DAGCombinerInfo &DCI, | |||
15116 | unsigned &RefinementSteps, | |||
15117 | bool &UseOneConstNR) const { | |||
15118 | // FIXME: We should use instruction latency models to calculate the cost of | |||
15119 | // each potential sequence, but this is very hard to do reliably because | |||
15120 | // at least Intel's Core* chips have variable timing based on the number of | |||
15121 | // significant digits in the divisor and/or sqrt operand. | |||
15122 | if (!Subtarget->useSqrtEst()) | |||
15123 | return SDValue(); | |||
15124 | ||||
15125 | EVT VT = Op.getValueType(); | |||
15126 | ||||
15127 | // SSE1 has rsqrtss and rsqrtps. | |||
15128 | // TODO: Add support for AVX512 (v16f32). | |||
15129 | // It is likely not profitable to do this for f64 because a double-precision | |||
15130 | // rsqrt estimate with refinement on x86 prior to FMA requires at least 16 | |||
15131 | // instructions: convert to single, rsqrtss, convert back to double, refine | |||
15132 | // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA | |||
15133 | // along with FMA, this could be a throughput win. | |||
15134 | if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) || | |||
15135 | (Subtarget->hasAVX() && VT == MVT::v8f32)) { | |||
15136 | RefinementSteps = 1; | |||
15137 | UseOneConstNR = false; | |||
15138 | return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op); | |||
15139 | } | |||
15140 | return SDValue(); | |||
15141 | } | |||
15142 | ||||
15143 | /// The minimum architected relative accuracy is 2^-12. We need one | |||
15144 | /// Newton-Raphson step to have a good float result (24 bits of precision). | |||
15145 | SDValue X86TargetLowering::getRecipEstimate(SDValue Op, | |||
15146 | DAGCombinerInfo &DCI, | |||
15147 | unsigned &RefinementSteps) const { | |||
15148 | // FIXME: We should use instruction latency models to calculate the cost of | |||
15149 | // each potential sequence, but this is very hard to do reliably because | |||
15150 | // at least Intel's Core* chips have variable timing based on the number of | |||
15151 | // significant digits in the divisor. | |||
15152 | if (!Subtarget->useReciprocalEst()) | |||
15153 | return SDValue(); | |||
15154 | ||||
15155 | EVT VT = Op.getValueType(); | |||
15156 | ||||
15157 | // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps. | |||
15158 | // TODO: Add support for AVX512 (v16f32). | |||
15159 | // It is likely not profitable to do this for f64 because a double-precision | |||
15160 | // reciprocal estimate with refinement on x86 prior to FMA requires | |||
15161 | // 15 instructions: convert to single, rcpss, convert back to double, refine | |||
15162 | // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA | |||
15163 | // along with FMA, this could be a throughput win. | |||
15164 | if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) || | |||
15165 | (Subtarget->hasAVX() && VT == MVT::v8f32)) { | |||
15166 | RefinementSteps = ReciprocalEstimateRefinementSteps; | |||
15167 | return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op); | |||
15168 | } | |||
15169 | return SDValue(); | |||
15170 | } | |||
15171 | ||||
15172 | static bool isAllOnes(SDValue V) { | |||
15173 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(V); | |||
15174 | return C && C->isAllOnesValue(); | |||
15175 | } | |||
15176 | ||||
15177 | /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node | |||
15178 | /// if it's possible. | |||
15179 | SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC, | |||
15180 | SDLoc dl, SelectionDAG &DAG) const { | |||
15181 | SDValue Op0 = And.getOperand(0); | |||
15182 | SDValue Op1 = And.getOperand(1); | |||
15183 | if (Op0.getOpcode() == ISD::TRUNCATE) | |||
15184 | Op0 = Op0.getOperand(0); | |||
15185 | if (Op1.getOpcode() == ISD::TRUNCATE) | |||
15186 | Op1 = Op1.getOperand(0); | |||
15187 | ||||
15188 | SDValue LHS, RHS; | |||
15189 | if (Op1.getOpcode() == ISD::SHL) | |||
15190 | std::swap(Op0, Op1); | |||
15191 | if (Op0.getOpcode() == ISD::SHL) { | |||
15192 | if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0))) | |||
15193 | if (And00C->getZExtValue() == 1) { | |||
15194 | // If we looked past a truncate, check that it's only truncating away | |||
15195 | // known zeros. | |||
15196 | unsigned BitWidth = Op0.getValueSizeInBits(); | |||
15197 | unsigned AndBitWidth = And.getValueSizeInBits(); | |||
15198 | if (BitWidth > AndBitWidth) { | |||
15199 | APInt Zeros, Ones; | |||
15200 | DAG.computeKnownBits(Op0, Zeros, Ones); | |||
15201 | if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth) | |||
15202 | return SDValue(); | |||
15203 | } | |||
15204 | LHS = Op1; | |||
15205 | RHS = Op0.getOperand(1); | |||
15206 | } | |||
15207 | } else if (Op1.getOpcode() == ISD::Constant) { | |||
15208 | ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1); | |||
15209 | uint64_t AndRHSVal = AndRHS->getZExtValue(); | |||
15210 | SDValue AndLHS = Op0; | |||
15211 | ||||
15212 | if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) { | |||
15213 | LHS = AndLHS.getOperand(0); | |||
15214 | RHS = AndLHS.getOperand(1); | |||
15215 | } | |||
15216 | ||||
15217 | // Use BT if the immediate can't be encoded in a TEST instruction. | |||
15218 | if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) { | |||
15219 | LHS = AndLHS; | |||
15220 | RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), LHS.getValueType()); | |||
15221 | } | |||
15222 | } | |||
15223 | ||||
15224 | if (LHS.getNode()) { | |||
15225 | // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT | |||
15226 | // instruction. Since the shift amount is in-range-or-undefined, we know | |||
15227 | // that doing a bittest on the i32 value is ok. We extend to i32 because | |||
15228 | // the encoding for the i16 version is larger than the i32 version. | |||
15229 | // Also promote i16 to i32 for performance / code size reason. | |||
15230 | if (LHS.getValueType() == MVT::i8 || | |||
15231 | LHS.getValueType() == MVT::i16) | |||
15232 | LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS); | |||
15233 | ||||
15234 | // If the operand types disagree, extend the shift amount to match. Since | |||
15235 | // BT ignores high bits (like shifts) we can use anyextend. | |||
15236 | if (LHS.getValueType() != RHS.getValueType()) | |||
15237 | RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS); | |||
15238 | ||||
15239 | SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS); | |||
15240 | X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B; | |||
15241 | return DAG.getNode(X86ISD::SETCC, dl, MVT::i8, | |||
15242 | DAG.getConstant(Cond, MVT::i8), BT); | |||
15243 | } | |||
15244 | ||||
15245 | return SDValue(); | |||
15246 | } | |||
15247 | ||||
15248 | /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point | |||
15249 | /// mask CMPs. | |||
15250 | static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0, | |||
15251 | SDValue &Op1) { | |||
15252 | unsigned SSECC; | |||
15253 | bool Swap = false; | |||
15254 | ||||
15255 | // SSE Condition code mapping: | |||
15256 | // 0 - EQ | |||
15257 | // 1 - LT | |||
15258 | // 2 - LE | |||
15259 | // 3 - UNORD | |||
15260 | // 4 - NEQ | |||
15261 | // 5 - NLT | |||
15262 | // 6 - NLE | |||
15263 | // 7 - ORD | |||
15264 | switch (SetCCOpcode) { | |||
15265 | default: llvm_unreachable("Unexpected SETCC condition")::llvm::llvm_unreachable_internal("Unexpected SETCC condition" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15265); | |||
15266 | case ISD::SETOEQ: | |||
15267 | case ISD::SETEQ: SSECC = 0; break; | |||
15268 | case ISD::SETOGT: | |||
15269 | case ISD::SETGT: Swap = true; // Fallthrough | |||
15270 | case ISD::SETLT: | |||
15271 | case ISD::SETOLT: SSECC = 1; break; | |||
15272 | case ISD::SETOGE: | |||
15273 | case ISD::SETGE: Swap = true; // Fallthrough | |||
15274 | case ISD::SETLE: | |||
15275 | case ISD::SETOLE: SSECC = 2; break; | |||
15276 | case ISD::SETUO: SSECC = 3; break; | |||
15277 | case ISD::SETUNE: | |||
15278 | case ISD::SETNE: SSECC = 4; break; | |||
15279 | case ISD::SETULE: Swap = true; // Fallthrough | |||
15280 | case ISD::SETUGE: SSECC = 5; break; | |||
15281 | case ISD::SETULT: Swap = true; // Fallthrough | |||
15282 | case ISD::SETUGT: SSECC = 6; break; | |||
15283 | case ISD::SETO: SSECC = 7; break; | |||
15284 | case ISD::SETUEQ: | |||
15285 | case ISD::SETONE: SSECC = 8; break; | |||
15286 | } | |||
15287 | if (Swap) | |||
15288 | std::swap(Op0, Op1); | |||
15289 | ||||
15290 | return SSECC; | |||
15291 | } | |||
15292 | ||||
15293 | // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128 | |||
15294 | // ones, and then concatenate the result back. | |||
15295 | static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) { | |||
15296 | MVT VT = Op.getSimpleValueType(); | |||
15297 | ||||
15298 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15299, __PRETTY_FUNCTION__)) | |||
15299 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15299, __PRETTY_FUNCTION__)); | |||
15300 | ||||
15301 | unsigned NumElems = VT.getVectorNumElements(); | |||
15302 | SDLoc dl(Op); | |||
15303 | SDValue CC = Op.getOperand(2); | |||
15304 | ||||
15305 | // Extract the LHS vectors | |||
15306 | SDValue LHS = Op.getOperand(0); | |||
15307 | SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl); | |||
15308 | SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl); | |||
15309 | ||||
15310 | // Extract the RHS vectors | |||
15311 | SDValue RHS = Op.getOperand(1); | |||
15312 | SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl); | |||
15313 | SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl); | |||
15314 | ||||
15315 | // Issue the operation on the smaller types and concatenate the result back | |||
15316 | MVT EltVT = VT.getVectorElementType(); | |||
15317 | MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2); | |||
15318 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, | |||
15319 | DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC), | |||
15320 | DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC)); | |||
15321 | } | |||
15322 | ||||
15323 | static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG, | |||
15324 | const X86Subtarget *Subtarget) { | |||
15325 | SDValue Op0 = Op.getOperand(0); | |||
15326 | SDValue Op1 = Op.getOperand(1); | |||
15327 | SDValue CC = Op.getOperand(2); | |||
15328 | MVT VT = Op.getSimpleValueType(); | |||
15329 | SDLoc dl(Op); | |||
15330 | ||||
15331 | assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&((Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 && Op.getValueType().getScalarType() == MVT::i1 && "Cannot set masked compare for this operation") ? static_cast <void> (0) : __assert_fail ("Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 && Op.getValueType().getScalarType() == MVT::i1 && \"Cannot set masked compare for this operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15333, __PRETTY_FUNCTION__)) | |||
15332 | Op.getValueType().getScalarType() == MVT::i1 &&((Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 && Op.getValueType().getScalarType() == MVT::i1 && "Cannot set masked compare for this operation") ? static_cast <void> (0) : __assert_fail ("Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 && Op.getValueType().getScalarType() == MVT::i1 && \"Cannot set masked compare for this operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15333, __PRETTY_FUNCTION__)) | |||
15333 | "Cannot set masked compare for this operation")((Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 && Op.getValueType().getScalarType() == MVT::i1 && "Cannot set masked compare for this operation") ? static_cast <void> (0) : __assert_fail ("Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 && Op.getValueType().getScalarType() == MVT::i1 && \"Cannot set masked compare for this operation\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15333, __PRETTY_FUNCTION__)); | |||
15334 | ||||
15335 | ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get(); | |||
15336 | unsigned Opc = 0; | |||
15337 | bool Unsigned = false; | |||
15338 | bool Swap = false; | |||
15339 | unsigned SSECC; | |||
15340 | switch (SetCCOpcode) { | |||
15341 | default: llvm_unreachable("Unexpected SETCC condition")::llvm::llvm_unreachable_internal("Unexpected SETCC condition" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15341); | |||
15342 | case ISD::SETNE: SSECC = 4; break; | |||
15343 | case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break; | |||
15344 | case ISD::SETUGT: SSECC = 6; Unsigned = true; break; | |||
15345 | case ISD::SETLT: Swap = true; //fall-through | |||
15346 | case ISD::SETGT: Opc = X86ISD::PCMPGTM; break; | |||
15347 | case ISD::SETULT: SSECC = 1; Unsigned = true; break; | |||
15348 | case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT | |||
15349 | case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap | |||
15350 | case ISD::SETULE: Unsigned = true; //fall-through | |||
15351 | case ISD::SETLE: SSECC = 2; break; | |||
15352 | } | |||
15353 | ||||
15354 | if (Swap) | |||
15355 | std::swap(Op0, Op1); | |||
15356 | if (Opc) | |||
15357 | return DAG.getNode(Opc, dl, VT, Op0, Op1); | |||
15358 | Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM; | |||
15359 | return DAG.getNode(Opc, dl, VT, Op0, Op1, | |||
15360 | DAG.getConstant(SSECC, MVT::i8)); | |||
15361 | } | |||
15362 | ||||
15363 | /// \brief Try to turn a VSETULT into a VSETULE by modifying its second | |||
15364 | /// operand \p Op1. If non-trivial (for example because it's not constant) | |||
15365 | /// return an empty value. | |||
15366 | static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG) | |||
15367 | { | |||
15368 | BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode()); | |||
15369 | if (!BV) | |||
15370 | return SDValue(); | |||
15371 | ||||
15372 | MVT VT = Op1.getSimpleValueType(); | |||
15373 | MVT EVT = VT.getVectorElementType(); | |||
15374 | unsigned n = VT.getVectorNumElements(); | |||
15375 | SmallVector<SDValue, 8> ULTOp1; | |||
15376 | ||||
15377 | for (unsigned i = 0; i < n; ++i) { | |||
15378 | ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i)); | |||
15379 | if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT) | |||
15380 | return SDValue(); | |||
15381 | ||||
15382 | // Avoid underflow. | |||
15383 | APInt Val = Elt->getAPIntValue(); | |||
15384 | if (Val == 0) | |||
15385 | return SDValue(); | |||
15386 | ||||
15387 | ULTOp1.push_back(DAG.getConstant(Val - 1, EVT)); | |||
15388 | } | |||
15389 | ||||
15390 | return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1); | |||
15391 | } | |||
15392 | ||||
15393 | static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget, | |||
15394 | SelectionDAG &DAG) { | |||
15395 | SDValue Op0 = Op.getOperand(0); | |||
15396 | SDValue Op1 = Op.getOperand(1); | |||
15397 | SDValue CC = Op.getOperand(2); | |||
15398 | MVT VT = Op.getSimpleValueType(); | |||
15399 | ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get(); | |||
15400 | bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint(); | |||
15401 | SDLoc dl(Op); | |||
15402 | ||||
15403 | if (isFP) { | |||
15404 | #ifndef NDEBUG | |||
15405 | MVT EltVT = Op0.getSimpleValueType().getVectorElementType(); | |||
15406 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15406, __PRETTY_FUNCTION__)); | |||
15407 | #endif | |||
15408 | ||||
15409 | unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1); | |||
15410 | unsigned Opc = X86ISD::CMPP; | |||
15411 | if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) { | |||
15412 | assert(VT.getVectorNumElements() <= 16)((VT.getVectorNumElements() <= 16) ? static_cast<void> (0) : __assert_fail ("VT.getVectorNumElements() <= 16", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15412, __PRETTY_FUNCTION__)); | |||
15413 | Opc = X86ISD::CMPM; | |||
15414 | } | |||
15415 | // In the two special cases we can't handle, emit two comparisons. | |||
15416 | if (SSECC == 8) { | |||
15417 | unsigned CC0, CC1; | |||
15418 | unsigned CombineOpc; | |||
15419 | if (SetCCOpcode == ISD::SETUEQ) { | |||
15420 | CC0 = 3; CC1 = 0; CombineOpc = ISD::OR; | |||
15421 | } else { | |||
15422 | assert(SetCCOpcode == ISD::SETONE)((SetCCOpcode == ISD::SETONE) ? static_cast<void> (0) : __assert_fail ("SetCCOpcode == ISD::SETONE", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15422, __PRETTY_FUNCTION__)); | |||
15423 | CC0 = 7; CC1 = 4; CombineOpc = ISD::AND; | |||
15424 | } | |||
15425 | ||||
15426 | SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1, | |||
15427 | DAG.getConstant(CC0, MVT::i8)); | |||
15428 | SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1, | |||
15429 | DAG.getConstant(CC1, MVT::i8)); | |||
15430 | return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1); | |||
15431 | } | |||
15432 | // Handle all other FP comparisons here. | |||
15433 | return DAG.getNode(Opc, dl, VT, Op0, Op1, | |||
15434 | DAG.getConstant(SSECC, MVT::i8)); | |||
15435 | } | |||
15436 | ||||
15437 | // Break 256-bit integer vector compare into smaller ones. | |||
15438 | if (VT.is256BitVector() && !Subtarget->hasInt256()) | |||
15439 | return Lower256IntVSETCC(Op, DAG); | |||
15440 | ||||
15441 | bool MaskResult = (VT.getVectorElementType() == MVT::i1); | |||
15442 | EVT OpVT = Op1.getValueType(); | |||
15443 | if (Subtarget->hasAVX512()) { | |||
15444 | if (Op1.getValueType().is512BitVector() || | |||
15445 | (Subtarget->hasBWI() && Subtarget->hasVLX()) || | |||
15446 | (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32)) | |||
15447 | return LowerIntVSETCC_AVX512(Op, DAG, Subtarget); | |||
15448 | ||||
15449 | // In AVX-512 architecture setcc returns mask with i1 elements, | |||
15450 | // But there is no compare instruction for i8 and i16 elements in KNL. | |||
15451 | // We are not talking about 512-bit operands in this case, these | |||
15452 | // types are illegal. | |||
15453 | if (MaskResult && | |||
15454 | (OpVT.getVectorElementType().getSizeInBits() < 32 && | |||
15455 | OpVT.getVectorElementType().getSizeInBits() >= 8)) | |||
15456 | return DAG.getNode(ISD::TRUNCATE, dl, VT, | |||
15457 | DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC)); | |||
15458 | } | |||
15459 | ||||
15460 | // We are handling one of the integer comparisons here. Since SSE only has | |||
15461 | // GT and EQ comparisons for integer, swapping operands and multiple | |||
15462 | // operations may be required for some comparisons. | |||
15463 | unsigned Opc; | |||
15464 | bool Swap = false, Invert = false, FlipSigns = false, MinMax = false; | |||
15465 | bool Subus = false; | |||
15466 | ||||
15467 | switch (SetCCOpcode) { | |||
15468 | default: llvm_unreachable("Unexpected SETCC condition")::llvm::llvm_unreachable_internal("Unexpected SETCC condition" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15468); | |||
15469 | case ISD::SETNE: Invert = true; | |||
15470 | case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break; | |||
15471 | case ISD::SETLT: Swap = true; | |||
15472 | case ISD::SETGT: Opc = X86ISD::PCMPGT; break; | |||
15473 | case ISD::SETGE: Swap = true; | |||
15474 | case ISD::SETLE: Opc = X86ISD::PCMPGT; | |||
15475 | Invert = true; break; | |||
15476 | case ISD::SETULT: Swap = true; | |||
15477 | case ISD::SETUGT: Opc = X86ISD::PCMPGT; | |||
15478 | FlipSigns = true; break; | |||
15479 | case ISD::SETUGE: Swap = true; | |||
15480 | case ISD::SETULE: Opc = X86ISD::PCMPGT; | |||
15481 | FlipSigns = true; Invert = true; break; | |||
15482 | } | |||
15483 | ||||
15484 | // Special case: Use min/max operations for SETULE/SETUGE | |||
15485 | MVT VET = VT.getVectorElementType(); | |||
15486 | bool hasMinMax = | |||
15487 | (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32)) | |||
15488 | || (Subtarget->hasSSE2() && (VET == MVT::i8)); | |||
15489 | ||||
15490 | if (hasMinMax) { | |||
15491 | switch (SetCCOpcode) { | |||
15492 | default: break; | |||
15493 | case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break; | |||
15494 | case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break; | |||
15495 | } | |||
15496 | ||||
15497 | if (MinMax) { Swap = false; Invert = false; FlipSigns = false; } | |||
15498 | } | |||
15499 | ||||
15500 | bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16); | |||
15501 | if (!MinMax && hasSubus) { | |||
15502 | // As another special case, use PSUBUS[BW] when it's profitable. E.g. for | |||
15503 | // Op0 u<= Op1: | |||
15504 | // t = psubus Op0, Op1 | |||
15505 | // pcmpeq t, <0..0> | |||
15506 | switch (SetCCOpcode) { | |||
15507 | default: break; | |||
15508 | case ISD::SETULT: { | |||
15509 | // If the comparison is against a constant we can turn this into a | |||
15510 | // setule. With psubus, setule does not require a swap. This is | |||
15511 | // beneficial because the constant in the register is no longer | |||
15512 | // destructed as the destination so it can be hoisted out of a loop. | |||
15513 | // Only do this pre-AVX since vpcmp* is no longer destructive. | |||
15514 | if (Subtarget->hasAVX()) | |||
15515 | break; | |||
15516 | SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG); | |||
15517 | if (ULEOp1.getNode()) { | |||
15518 | Op1 = ULEOp1; | |||
15519 | Subus = true; Invert = false; Swap = false; | |||
15520 | } | |||
15521 | break; | |||
15522 | } | |||
15523 | // Psubus is better than flip-sign because it requires no inversion. | |||
15524 | case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break; | |||
15525 | case ISD::SETULE: Subus = true; Invert = false; Swap = false; break; | |||
15526 | } | |||
15527 | ||||
15528 | if (Subus) { | |||
15529 | Opc = X86ISD::SUBUS; | |||
15530 | FlipSigns = false; | |||
15531 | } | |||
15532 | } | |||
15533 | ||||
15534 | if (Swap) | |||
15535 | std::swap(Op0, Op1); | |||
15536 | ||||
15537 | // Check that the operation in question is available (most are plain SSE2, | |||
15538 | // but PCMPGTQ and PCMPEQQ have different requirements). | |||
15539 | if (VT == MVT::v2i64) { | |||
15540 | if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) { | |||
15541 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15541, __PRETTY_FUNCTION__)); | |||
15542 | ||||
15543 | // First cast everything to the right type. | |||
15544 | Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0); | |||
15545 | Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1); | |||
15546 | ||||
15547 | // Since SSE has no unsigned integer comparisons, we need to flip the sign | |||
15548 | // bits of the inputs before performing those operations. The lower | |||
15549 | // compare is always unsigned. | |||
15550 | SDValue SB; | |||
15551 | if (FlipSigns) { | |||
15552 | SB = DAG.getConstant(0x80000000U, MVT::v4i32); | |||
15553 | } else { | |||
15554 | SDValue Sign = DAG.getConstant(0x80000000U, MVT::i32); | |||
15555 | SDValue Zero = DAG.getConstant(0x00000000U, MVT::i32); | |||
15556 | SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, | |||
15557 | Sign, Zero, Sign, Zero); | |||
15558 | } | |||
15559 | Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB); | |||
15560 | Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB); | |||
15561 | ||||
15562 | // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2)) | |||
15563 | SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1); | |||
15564 | SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1); | |||
15565 | ||||
15566 | // Create masks for only the low parts/high parts of the 64 bit integers. | |||
15567 | static const int MaskHi[] = { 1, 1, 3, 3 }; | |||
15568 | static const int MaskLo[] = { 0, 0, 2, 2 }; | |||
15569 | SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi); | |||
15570 | SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo); | |||
15571 | SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi); | |||
15572 | ||||
15573 | SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo); | |||
15574 | Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi); | |||
15575 | ||||
15576 | if (Invert) | |||
15577 | Result = DAG.getNOT(dl, Result, MVT::v4i32); | |||
15578 | ||||
15579 | return DAG.getNode(ISD::BITCAST, dl, VT, Result); | |||
15580 | } | |||
15581 | ||||
15582 | if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) { | |||
15583 | // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with | |||
15584 | // pcmpeqd + pshufd + pand. | |||
15585 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15585, __PRETTY_FUNCTION__)); | |||
15586 | ||||
15587 | // First cast everything to the right type. | |||
15588 | Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0); | |||
15589 | Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1); | |||
15590 | ||||
15591 | // Do the compare. | |||
15592 | SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1); | |||
15593 | ||||
15594 | // Make sure the lower and upper halves are both all-ones. | |||
15595 | static const int Mask[] = { 1, 0, 3, 2 }; | |||
15596 | SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask); | |||
15597 | Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf); | |||
15598 | ||||
15599 | if (Invert) | |||
15600 | Result = DAG.getNOT(dl, Result, MVT::v4i32); | |||
15601 | ||||
15602 | return DAG.getNode(ISD::BITCAST, dl, VT, Result); | |||
15603 | } | |||
15604 | } | |||
15605 | ||||
15606 | // Since SSE has no unsigned integer comparisons, we need to flip the sign | |||
15607 | // bits of the inputs before performing those operations. | |||
15608 | if (FlipSigns) { | |||
15609 | EVT EltVT = VT.getVectorElementType(); | |||
15610 | SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), VT); | |||
15611 | Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB); | |||
15612 | Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB); | |||
15613 | } | |||
15614 | ||||
15615 | SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1); | |||
15616 | ||||
15617 | // If the logical-not of the result is required, perform that now. | |||
15618 | if (Invert) | |||
15619 | Result = DAG.getNOT(dl, Result, VT); | |||
15620 | ||||
15621 | if (MinMax) | |||
15622 | Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result); | |||
15623 | ||||
15624 | if (Subus) | |||
15625 | Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result, | |||
15626 | getZeroVector(VT, Subtarget, DAG, dl)); | |||
15627 | ||||
15628 | return Result; | |||
15629 | } | |||
15630 | ||||
15631 | SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const { | |||
15632 | ||||
15633 | MVT VT = Op.getSimpleValueType(); | |||
15634 | ||||
15635 | if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG); | |||
15636 | ||||
15637 | assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))((((!Subtarget->hasAVX512() && VT == MVT::i8) || ( VT == MVT::i1)) && "SetCC type must be 8-bit or 1-bit integer" ) ? static_cast<void> (0) : __assert_fail ("((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1)) && \"SetCC type must be 8-bit or 1-bit integer\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15638, __PRETTY_FUNCTION__)) | |||
15638 | && "SetCC type must be 8-bit or 1-bit integer")((((!Subtarget->hasAVX512() && VT == MVT::i8) || ( VT == MVT::i1)) && "SetCC type must be 8-bit or 1-bit integer" ) ? static_cast<void> (0) : __assert_fail ("((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1)) && \"SetCC type must be 8-bit or 1-bit integer\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15638, __PRETTY_FUNCTION__)); | |||
15639 | SDValue Op0 = Op.getOperand(0); | |||
15640 | SDValue Op1 = Op.getOperand(1); | |||
15641 | SDLoc dl(Op); | |||
15642 | ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); | |||
15643 | ||||
15644 | // Optimize to BT if possible. | |||
15645 | // Lower (X & (1 << N)) == 0 to BT(X, N). | |||
15646 | // Lower ((X >>u N) & 1) != 0 to BT(X, N). | |||
15647 | // Lower ((X >>s N) & 1) != 0 to BT(X, N). | |||
15648 | if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() && | |||
15649 | Op1.getOpcode() == ISD::Constant && | |||
15650 | cast<ConstantSDNode>(Op1)->isNullValue() && | |||
15651 | (CC == ISD::SETEQ || CC == ISD::SETNE)) { | |||
15652 | SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG); | |||
15653 | if (NewSetCC.getNode()) { | |||
15654 | if (VT == MVT::i1) | |||
15655 | return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC); | |||
15656 | return NewSetCC; | |||
15657 | } | |||
15658 | } | |||
15659 | ||||
15660 | // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of | |||
15661 | // these. | |||
15662 | if (Op1.getOpcode() == ISD::Constant && | |||
15663 | (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 || | |||
15664 | cast<ConstantSDNode>(Op1)->isNullValue()) && | |||
15665 | (CC == ISD::SETEQ || CC == ISD::SETNE)) { | |||
15666 | ||||
15667 | // If the input is a setcc, then reuse the input setcc or use a new one with | |||
15668 | // the inverted condition. | |||
15669 | if (Op0.getOpcode() == X86ISD::SETCC) { | |||
15670 | X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0); | |||
15671 | bool Invert = (CC == ISD::SETNE) ^ | |||
15672 | cast<ConstantSDNode>(Op1)->isNullValue(); | |||
15673 | if (!Invert) | |||
15674 | return Op0; | |||
15675 | ||||
15676 | CCode = X86::GetOppositeBranchCondition(CCode); | |||
15677 | SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, | |||
15678 | DAG.getConstant(CCode, MVT::i8), | |||
15679 | Op0.getOperand(1)); | |||
15680 | if (VT == MVT::i1) | |||
15681 | return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC); | |||
15682 | return SetCC; | |||
15683 | } | |||
15684 | } | |||
15685 | if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) && | |||
15686 | (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) && | |||
15687 | (CC == ISD::SETEQ || CC == ISD::SETNE)) { | |||
15688 | ||||
15689 | ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true); | |||
15690 | return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, MVT::i1), NewCC); | |||
15691 | } | |||
15692 | ||||
15693 | bool isFP = Op1.getSimpleValueType().isFloatingPoint(); | |||
15694 | unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG); | |||
15695 | if (X86CC == X86::COND_INVALID) | |||
15696 | return SDValue(); | |||
15697 | ||||
15698 | SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG); | |||
15699 | EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG); | |||
15700 | SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, | |||
15701 | DAG.getConstant(X86CC, MVT::i8), EFLAGS); | |||
15702 | if (VT == MVT::i1) | |||
15703 | return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC); | |||
15704 | return SetCC; | |||
15705 | } | |||
15706 | ||||
15707 | // isX86LogicalCmp - Return true if opcode is a X86 logical comparison. | |||
15708 | static bool isX86LogicalCmp(SDValue Op) { | |||
15709 | unsigned Opc = Op.getNode()->getOpcode(); | |||
15710 | if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI || | |||
15711 | Opc == X86ISD::SAHF) | |||
15712 | return true; | |||
15713 | if (Op.getResNo() == 1 && | |||
15714 | (Opc == X86ISD::ADD || | |||
15715 | Opc == X86ISD::SUB || | |||
15716 | Opc == X86ISD::ADC || | |||
15717 | Opc == X86ISD::SBB || | |||
15718 | Opc == X86ISD::SMUL || | |||
15719 | Opc == X86ISD::UMUL || | |||
15720 | Opc == X86ISD::INC || | |||
15721 | Opc == X86ISD::DEC || | |||
15722 | Opc == X86ISD::OR || | |||
15723 | Opc == X86ISD::XOR || | |||
15724 | Opc == X86ISD::AND)) | |||
15725 | return true; | |||
15726 | ||||
15727 | if (Op.getResNo() == 2 && Opc == X86ISD::UMUL) | |||
15728 | return true; | |||
15729 | ||||
15730 | return false; | |||
15731 | } | |||
15732 | ||||
15733 | static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) { | |||
15734 | if (V.getOpcode() != ISD::TRUNCATE) | |||
15735 | return false; | |||
15736 | ||||
15737 | SDValue VOp0 = V.getOperand(0); | |||
15738 | unsigned InBits = VOp0.getValueSizeInBits(); | |||
15739 | unsigned Bits = V.getValueSizeInBits(); | |||
15740 | return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits)); | |||
15741 | } | |||
15742 | ||||
15743 | SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const { | |||
15744 | bool addTest = true; | |||
15745 | SDValue Cond = Op.getOperand(0); | |||
15746 | SDValue Op1 = Op.getOperand(1); | |||
15747 | SDValue Op2 = Op.getOperand(2); | |||
15748 | SDLoc DL(Op); | |||
15749 | EVT VT = Op1.getValueType(); | |||
15750 | SDValue CC; | |||
15751 | ||||
15752 | // Lower fp selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops | |||
15753 | // are available. Otherwise fp cmovs get lowered into a less efficient branch | |||
15754 | // sequence later on. | |||
15755 | if (Cond.getOpcode() == ISD::SETCC && | |||
15756 | ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) || | |||
15757 | (Subtarget->hasSSE1() && VT == MVT::f32)) && | |||
15758 | VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) { | |||
15759 | SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1); | |||
15760 | int SSECC = translateX86FSETCC( | |||
15761 | cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1); | |||
15762 | ||||
15763 | if (SSECC != 8) { | |||
15764 | if (Subtarget->hasAVX512()) { | |||
15765 | SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1, | |||
15766 | DAG.getConstant(SSECC, MVT::i8)); | |||
15767 | return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2); | |||
15768 | } | |||
15769 | SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1, | |||
15770 | DAG.getConstant(SSECC, MVT::i8)); | |||
15771 | SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2); | |||
15772 | SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1); | |||
15773 | return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And); | |||
15774 | } | |||
15775 | } | |||
15776 | ||||
15777 | if (Cond.getOpcode() == ISD::SETCC) { | |||
15778 | SDValue NewCond = LowerSETCC(Cond, DAG); | |||
15779 | if (NewCond.getNode()) | |||
15780 | Cond = NewCond; | |||
15781 | } | |||
15782 | ||||
15783 | // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y | |||
15784 | // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y | |||
15785 | // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y | |||
15786 | // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y | |||
15787 | if (Cond.getOpcode() == X86ISD::SETCC && | |||
15788 | Cond.getOperand(1).getOpcode() == X86ISD::CMP && | |||
15789 | isZero(Cond.getOperand(1).getOperand(1))) { | |||
15790 | SDValue Cmp = Cond.getOperand(1); | |||
15791 | ||||
15792 | unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue(); | |||
15793 | ||||
15794 | if ((isAllOnes(Op1) || isAllOnes(Op2)) && | |||
15795 | (CondCode == X86::COND_E || CondCode == X86::COND_NE)) { | |||
15796 | SDValue Y = isAllOnes(Op2) ? Op1 : Op2; | |||
15797 | ||||
15798 | SDValue CmpOp0 = Cmp.getOperand(0); | |||
15799 | // Apply further optimizations for special cases | |||
15800 | // (select (x != 0), -1, 0) -> neg & sbb | |||
15801 | // (select (x == 0), 0, -1) -> neg & sbb | |||
15802 | if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y)) | |||
15803 | if (YC->isNullValue() && | |||
15804 | (isAllOnes(Op1) == (CondCode == X86::COND_NE))) { | |||
15805 | SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32); | |||
15806 | SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs, | |||
15807 | DAG.getConstant(0, CmpOp0.getValueType()), | |||
15808 | CmpOp0); | |||
15809 | SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(), | |||
15810 | DAG.getConstant(X86::COND_B, MVT::i8), | |||
15811 | SDValue(Neg.getNode(), 1)); | |||
15812 | return Res; | |||
15813 | } | |||
15814 | ||||
15815 | Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, | |||
15816 | CmpOp0, DAG.getConstant(1, CmpOp0.getValueType())); | |||
15817 | Cmp = ConvertCmpIfNecessary(Cmp, DAG); | |||
15818 | ||||
15819 | SDValue Res = // Res = 0 or -1. | |||
15820 | DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(), | |||
15821 | DAG.getConstant(X86::COND_B, MVT::i8), Cmp); | |||
15822 | ||||
15823 | if (isAllOnes(Op1) != (CondCode == X86::COND_E)) | |||
15824 | Res = DAG.getNOT(DL, Res, Res.getValueType()); | |||
15825 | ||||
15826 | ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2); | |||
15827 | if (!N2C || !N2C->isNullValue()) | |||
15828 | Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y); | |||
15829 | return Res; | |||
15830 | } | |||
15831 | } | |||
15832 | ||||
15833 | // Look past (and (setcc_carry (cmp ...)), 1). | |||
15834 | if (Cond.getOpcode() == ISD::AND && | |||
15835 | Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) { | |||
15836 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1)); | |||
15837 | if (C && C->getAPIntValue() == 1) | |||
15838 | Cond = Cond.getOperand(0); | |||
15839 | } | |||
15840 | ||||
15841 | // If condition flag is set by a X86ISD::CMP, then use it as the condition | |||
15842 | // setting operand in place of the X86ISD::SETCC. | |||
15843 | unsigned CondOpcode = Cond.getOpcode(); | |||
15844 | if (CondOpcode == X86ISD::SETCC || | |||
15845 | CondOpcode == X86ISD::SETCC_CARRY) { | |||
15846 | CC = Cond.getOperand(0); | |||
15847 | ||||
15848 | SDValue Cmp = Cond.getOperand(1); | |||
15849 | unsigned Opc = Cmp.getOpcode(); | |||
15850 | MVT VT = Op.getSimpleValueType(); | |||
15851 | ||||
15852 | bool IllegalFPCMov = false; | |||
15853 | if (VT.isFloatingPoint() && !VT.isVector() && | |||
15854 | !isScalarFPTypeInSSEReg(VT)) // FPStack? | |||
15855 | IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue()); | |||
15856 | ||||
15857 | if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) || | |||
15858 | Opc == X86ISD::BT) { // FIXME | |||
15859 | Cond = Cmp; | |||
15860 | addTest = false; | |||
15861 | } | |||
15862 | } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO || | |||
15863 | CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO || | |||
15864 | ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) && | |||
15865 | Cond.getOperand(0).getValueType() != MVT::i8)) { | |||
15866 | SDValue LHS = Cond.getOperand(0); | |||
15867 | SDValue RHS = Cond.getOperand(1); | |||
15868 | unsigned X86Opcode; | |||
15869 | unsigned X86Cond; | |||
15870 | SDVTList VTs; | |||
15871 | switch (CondOpcode) { | |||
15872 | case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break; | |||
15873 | case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break; | |||
15874 | case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break; | |||
15875 | case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break; | |||
15876 | case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break; | |||
15877 | case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break; | |||
15878 | default: llvm_unreachable("unexpected overflowing operator")::llvm::llvm_unreachable_internal("unexpected overflowing operator" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15878); | |||
15879 | } | |||
15880 | if (CondOpcode == ISD::UMULO) | |||
15881 | VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(), | |||
15882 | MVT::i32); | |||
15883 | else | |||
15884 | VTs = DAG.getVTList(LHS.getValueType(), MVT::i32); | |||
15885 | ||||
15886 | SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS); | |||
15887 | ||||
15888 | if (CondOpcode == ISD::UMULO) | |||
15889 | Cond = X86Op.getValue(2); | |||
15890 | else | |||
15891 | Cond = X86Op.getValue(1); | |||
15892 | ||||
15893 | CC = DAG.getConstant(X86Cond, MVT::i8); | |||
15894 | addTest = false; | |||
15895 | } | |||
15896 | ||||
15897 | if (addTest) { | |||
15898 | // Look pass the truncate if the high bits are known zero. | |||
15899 | if (isTruncWithZeroHighBitsInput(Cond, DAG)) | |||
15900 | Cond = Cond.getOperand(0); | |||
15901 | ||||
15902 | // We know the result of AND is compared against zero. Try to match | |||
15903 | // it to BT. | |||
15904 | if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) { | |||
15905 | SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG); | |||
15906 | if (NewSetCC.getNode()) { | |||
15907 | CC = NewSetCC.getOperand(0); | |||
15908 | Cond = NewSetCC.getOperand(1); | |||
15909 | addTest = false; | |||
15910 | } | |||
15911 | } | |||
15912 | } | |||
15913 | ||||
15914 | if (addTest) { | |||
15915 | CC = DAG.getConstant(X86::COND_NE, MVT::i8); | |||
15916 | Cond = EmitTest(Cond, X86::COND_NE, DL, DAG); | |||
15917 | } | |||
15918 | ||||
15919 | // a < b ? -1 : 0 -> RES = ~setcc_carry | |||
15920 | // a < b ? 0 : -1 -> RES = setcc_carry | |||
15921 | // a >= b ? -1 : 0 -> RES = setcc_carry | |||
15922 | // a >= b ? 0 : -1 -> RES = ~setcc_carry | |||
15923 | if (Cond.getOpcode() == X86ISD::SUB) { | |||
15924 | Cond = ConvertCmpIfNecessary(Cond, DAG); | |||
15925 | unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue(); | |||
15926 | ||||
15927 | if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) && | |||
15928 | (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) { | |||
15929 | SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(), | |||
15930 | DAG.getConstant(X86::COND_B, MVT::i8), Cond); | |||
15931 | if (isAllOnes(Op1) != (CondCode == X86::COND_B)) | |||
15932 | return DAG.getNOT(DL, Res, Res.getValueType()); | |||
15933 | return Res; | |||
15934 | } | |||
15935 | } | |||
15936 | ||||
15937 | // X86 doesn't have an i8 cmov. If both operands are the result of a truncate | |||
15938 | // widen the cmov and push the truncate through. This avoids introducing a new | |||
15939 | // branch during isel and doesn't add any extensions. | |||
15940 | if (Op.getValueType() == MVT::i8 && | |||
15941 | Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) { | |||
15942 | SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0); | |||
15943 | if (T1.getValueType() == T2.getValueType() && | |||
15944 | // Blacklist CopyFromReg to avoid partial register stalls. | |||
15945 | T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){ | |||
15946 | SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue); | |||
15947 | SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond); | |||
15948 | return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov); | |||
15949 | } | |||
15950 | } | |||
15951 | ||||
15952 | // X86ISD::CMOV means set the result (which is operand 1) to the RHS if | |||
15953 | // condition is true. | |||
15954 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue); | |||
15955 | SDValue Ops[] = { Op2, Op1, CC, Cond }; | |||
15956 | return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops); | |||
15957 | } | |||
15958 | ||||
15959 | static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, const X86Subtarget *Subtarget, | |||
15960 | SelectionDAG &DAG) { | |||
15961 | MVT VT = Op->getSimpleValueType(0); | |||
15962 | SDValue In = Op->getOperand(0); | |||
15963 | MVT InVT = In.getSimpleValueType(); | |||
15964 | MVT VTElt = VT.getVectorElementType(); | |||
15965 | MVT InVTElt = InVT.getVectorElementType(); | |||
15966 | SDLoc dl(Op); | |||
15967 | ||||
15968 | // SKX processor | |||
15969 | if ((InVTElt == MVT::i1) && | |||
15970 | (((Subtarget->hasBWI() && Subtarget->hasVLX() && | |||
15971 | VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) || | |||
15972 | ||||
15973 | ((Subtarget->hasBWI() && VT.is512BitVector() && | |||
15974 | VTElt.getSizeInBits() <= 16)) || | |||
15975 | ||||
15976 | ((Subtarget->hasDQI() && Subtarget->hasVLX() && | |||
15977 | VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) || | |||
15978 | ||||
15979 | ((Subtarget->hasDQI() && VT.is512BitVector() && | |||
15980 | VTElt.getSizeInBits() >= 32)))) | |||
15981 | return DAG.getNode(X86ISD::VSEXT, dl, VT, In); | |||
15982 | ||||
15983 | unsigned int NumElts = VT.getVectorNumElements(); | |||
15984 | ||||
15985 | if (NumElts != 8 && NumElts != 16) | |||
15986 | return SDValue(); | |||
15987 | ||||
15988 | if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) { | |||
15989 | if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT) | |||
15990 | return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0)); | |||
15991 | return DAG.getNode(X86ISD::VSEXT, dl, VT, In); | |||
15992 | } | |||
15993 | ||||
15994 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
15995 | assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type")((InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type" ) ? static_cast<void> (0) : __assert_fail ("InVT.getVectorElementType() == MVT::i1 && \"Unexpected vector type\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 15995, __PRETTY_FUNCTION__)); | |||
15996 | ||||
15997 | MVT ExtVT = (NumElts == 8) ? MVT::v8i64 : MVT::v16i32; | |||
15998 | Constant *C = ConstantInt::get(*DAG.getContext(), | |||
15999 | APInt::getAllOnesValue(ExtVT.getScalarType().getSizeInBits())); | |||
16000 | ||||
16001 | SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy()); | |||
16002 | unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment(); | |||
16003 | SDValue Ld = DAG.getLoad(ExtVT.getScalarType(), dl, DAG.getEntryNode(), CP, | |||
16004 | MachinePointerInfo::getConstantPool(), | |||
16005 | false, false, false, Alignment); | |||
16006 | SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, dl, ExtVT, In, Ld); | |||
16007 | if (VT.is512BitVector()) | |||
16008 | return Brcst; | |||
16009 | return DAG.getNode(X86ISD::VTRUNC, dl, VT, Brcst); | |||
16010 | } | |||
16011 | ||||
16012 | static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget, | |||
16013 | SelectionDAG &DAG) { | |||
16014 | MVT VT = Op->getSimpleValueType(0); | |||
16015 | SDValue In = Op->getOperand(0); | |||
16016 | MVT InVT = In.getSimpleValueType(); | |||
16017 | SDLoc dl(Op); | |||
16018 | ||||
16019 | if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1) | |||
16020 | return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG); | |||
16021 | ||||
16022 | if ((VT != MVT::v4i64 || InVT != MVT::v4i32) && | |||
16023 | (VT != MVT::v8i32 || InVT != MVT::v8i16) && | |||
16024 | (VT != MVT::v16i16 || InVT != MVT::v16i8)) | |||
16025 | return SDValue(); | |||
16026 | ||||
16027 | if (Subtarget->hasInt256()) | |||
16028 | return DAG.getNode(X86ISD::VSEXT, dl, VT, In); | |||
16029 | ||||
16030 | // Optimize vectors in AVX mode | |||
16031 | // Sign extend v8i16 to v8i32 and | |||
16032 | // v4i32 to v4i64 | |||
16033 | // | |||
16034 | // Divide input vector into two parts | |||
16035 | // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1} | |||
16036 | // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32 | |||
16037 | // concat the vectors to original VT | |||
16038 | ||||
16039 | unsigned NumElems = InVT.getVectorNumElements(); | |||
16040 | SDValue Undef = DAG.getUNDEF(InVT); | |||
16041 | ||||
16042 | SmallVector<int,8> ShufMask1(NumElems, -1); | |||
16043 | for (unsigned i = 0; i != NumElems/2; ++i) | |||
16044 | ShufMask1[i] = i; | |||
16045 | ||||
16046 | SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]); | |||
16047 | ||||
16048 | SmallVector<int,8> ShufMask2(NumElems, -1); | |||
16049 | for (unsigned i = 0; i != NumElems/2; ++i) | |||
16050 | ShufMask2[i] = i + NumElems/2; | |||
16051 | ||||
16052 | SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]); | |||
16053 | ||||
16054 | MVT HalfVT = MVT::getVectorVT(VT.getScalarType(), | |||
16055 | VT.getVectorNumElements()/2); | |||
16056 | ||||
16057 | OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo); | |||
16058 | OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi); | |||
16059 | ||||
16060 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi); | |||
16061 | } | |||
16062 | ||||
16063 | // Lower vector extended loads using a shuffle. If SSSE3 is not available we | |||
16064 | // may emit an illegal shuffle but the expansion is still better than scalar | |||
16065 | // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise | |||
16066 | // we'll emit a shuffle and a arithmetic shift. | |||
16067 | // TODO: It is possible to support ZExt by zeroing the undef values during | |||
16068 | // the shuffle phase or after the shuffle. | |||
16069 | static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget, | |||
16070 | SelectionDAG &DAG) { | |||
16071 | MVT RegVT = Op.getSimpleValueType(); | |||
16072 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16072, __PRETTY_FUNCTION__)); | |||
16073 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16074, __PRETTY_FUNCTION__)) | |||
16074 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16074, __PRETTY_FUNCTION__)); | |||
16075 | ||||
16076 | // Nothing useful we can do without SSE2 shuffles. | |||
16077 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16077, __PRETTY_FUNCTION__)); | |||
16078 | ||||
16079 | LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode()); | |||
16080 | SDLoc dl(Ld); | |||
16081 | EVT MemVT = Ld->getMemoryVT(); | |||
16082 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
16083 | unsigned RegSz = RegVT.getSizeInBits(); | |||
16084 | ||||
16085 | ISD::LoadExtType Ext = Ld->getExtensionType(); | |||
16086 | ||||
16087 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16088, __PRETTY_FUNCTION__)) | |||
16088 | && "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16088, __PRETTY_FUNCTION__)); | |||
16089 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16089, __PRETTY_FUNCTION__)); | |||
16090 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16090, __PRETTY_FUNCTION__)); | |||
16091 | ||||
16092 | unsigned NumElems = RegVT.getVectorNumElements(); | |||
16093 | unsigned MemSz = MemVT.getSizeInBits(); | |||
16094 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16094, __PRETTY_FUNCTION__)); | |||
16095 | ||||
16096 | if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) { | |||
16097 | // The only way in which we have a legal 256-bit vector result but not the | |||
16098 | // integer 256-bit operations needed to directly lower a sextload is if we | |||
16099 | // have AVX1 but not AVX2. In that case, we can always emit a sextload to | |||
16100 | // a 128-bit vector and a normal sign_extend to 256-bits that should get | |||
16101 | // correctly legalized. We do this late to allow the canonical form of | |||
16102 | // sextload to persist throughout the rest of the DAG combiner -- it wants | |||
16103 | // to fold together any extensions it can, and so will fuse a sign_extend | |||
16104 | // of an sextload into a sextload targeting a wider value. | |||
16105 | SDValue Load; | |||
16106 | if (MemSz == 128) { | |||
16107 | // Just switch this to a normal load. | |||
16108 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16110, __PRETTY_FUNCTION__)) | |||
16109 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16110, __PRETTY_FUNCTION__)) | |||
16110 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16110, __PRETTY_FUNCTION__)); | |||
16111 | Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(), | |||
16112 | Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(), | |||
16113 | Ld->isInvariant(), Ld->getAlignment()); | |||
16114 | } else { | |||
16115 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16116, __PRETTY_FUNCTION__)) | |||
16116 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16116, __PRETTY_FUNCTION__)); | |||
16117 | // Do an sext load to a 128-bit vector type. We want to use the same | |||
16118 | // number of elements, but elements half as wide. This will end up being | |||
16119 | // recursively lowered by this routine, but will succeed as we definitely | |||
16120 | // have all the necessary features if we're using AVX1. | |||
16121 | EVT HalfEltVT = | |||
16122 | EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2); | |||
16123 | EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems); | |||
16124 | Load = | |||
16125 | DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(), | |||
16126 | Ld->getPointerInfo(), MemVT, Ld->isVolatile(), | |||
16127 | Ld->isNonTemporal(), Ld->isInvariant(), | |||
16128 | Ld->getAlignment()); | |||
16129 | } | |||
16130 | ||||
16131 | // Replace chain users with the new chain. | |||
16132 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16132, __PRETTY_FUNCTION__)); | |||
16133 | DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1)); | |||
16134 | ||||
16135 | // Finally, do a normal sign-extend to the desired register. | |||
16136 | return DAG.getSExtOrTrunc(Load, dl, RegVT); | |||
16137 | } | |||
16138 | ||||
16139 | // All sizes must be a power of two. | |||
16140 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16141, __PRETTY_FUNCTION__)) | |||
16141 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16141, __PRETTY_FUNCTION__)); | |||
16142 | ||||
16143 | // Attempt to load the original value using scalar loads. | |||
16144 | // Find the largest scalar type that divides the total loaded size. | |||
16145 | MVT SclrLoadTy = MVT::i8; | |||
16146 | for (MVT Tp : MVT::integer_valuetypes()) { | |||
16147 | if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) { | |||
16148 | SclrLoadTy = Tp; | |||
16149 | } | |||
16150 | } | |||
16151 | ||||
16152 | // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64. | |||
16153 | if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 && | |||
16154 | (64 <= MemSz)) | |||
16155 | SclrLoadTy = MVT::f64; | |||
16156 | ||||
16157 | // Calculate the number of scalar loads that we need to perform | |||
16158 | // in order to load our vector from memory. | |||
16159 | unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits(); | |||
16160 | ||||
16161 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16162, __PRETTY_FUNCTION__)) | |||
16162 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16162, __PRETTY_FUNCTION__)); | |||
16163 | ||||
16164 | unsigned loadRegZize = RegSz; | |||
16165 | if (Ext == ISD::SEXTLOAD && RegSz == 256) | |||
16166 | loadRegZize /= 2; | |||
16167 | ||||
16168 | // Represent our vector as a sequence of elements which are the | |||
16169 | // largest scalar that we can load. | |||
16170 | EVT LoadUnitVecVT = EVT::getVectorVT( | |||
16171 | *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits()); | |||
16172 | ||||
16173 | // Represent the data using the same element type that is stored in | |||
16174 | // memory. In practice, we ''widen'' MemVT. | |||
16175 | EVT WideVecVT = | |||
16176 | EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(), | |||
16177 | loadRegZize / MemVT.getScalarType().getSizeInBits()); | |||
16178 | ||||
16179 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16180, __PRETTY_FUNCTION__)) | |||
16180 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16180, __PRETTY_FUNCTION__)); | |||
16181 | ||||
16182 | // We can't shuffle using an illegal type. | |||
16183 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16184, __PRETTY_FUNCTION__)) | |||
16184 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16184, __PRETTY_FUNCTION__)); | |||
16185 | ||||
16186 | SmallVector<SDValue, 8> Chains; | |||
16187 | SDValue Ptr = Ld->getBasePtr(); | |||
16188 | SDValue Increment = | |||
16189 | DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, TLI.getPointerTy()); | |||
16190 | SDValue Res = DAG.getUNDEF(LoadUnitVecVT); | |||
16191 | ||||
16192 | for (unsigned i = 0; i < NumLoads; ++i) { | |||
16193 | // Perform a single load. | |||
16194 | SDValue ScalarLoad = | |||
16195 | DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(), | |||
16196 | Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(), | |||
16197 | Ld->getAlignment()); | |||
16198 | Chains.push_back(ScalarLoad.getValue(1)); | |||
16199 | // Create the first element type using SCALAR_TO_VECTOR in order to avoid | |||
16200 | // another round of DAGCombining. | |||
16201 | if (i == 0) | |||
16202 | Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad); | |||
16203 | else | |||
16204 | Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res, | |||
16205 | ScalarLoad, DAG.getIntPtrConstant(i)); | |||
16206 | ||||
16207 | Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment); | |||
16208 | } | |||
16209 | ||||
16210 | SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains); | |||
16211 | ||||
16212 | // Bitcast the loaded value to a vector of the original element type, in | |||
16213 | // the size of the target vector type. | |||
16214 | SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res); | |||
16215 | unsigned SizeRatio = RegSz / MemSz; | |||
16216 | ||||
16217 | if (Ext == ISD::SEXTLOAD) { | |||
16218 | // If we have SSE4.1, we can directly emit a VSEXT node. | |||
16219 | if (Subtarget->hasSSE41()) { | |||
16220 | SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec); | |||
16221 | DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF); | |||
16222 | return Sext; | |||
16223 | } | |||
16224 | ||||
16225 | // Otherwise we'll shuffle the small elements in the high bits of the | |||
16226 | // larger type and perform an arithmetic shift. If the shift is not legal | |||
16227 | // it's better to scalarize. | |||
16228 | assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&((TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) && "We can't implement a sext load without an arithmetic right shift!" ) ? static_cast<void> (0) : __assert_fail ("TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) && \"We can't implement a sext load without an arithmetic right shift!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16229, __PRETTY_FUNCTION__)) | |||
16229 | "We can't implement a sext load without an arithmetic right shift!")((TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) && "We can't implement a sext load without an arithmetic right shift!" ) ? static_cast<void> (0) : __assert_fail ("TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) && \"We can't implement a sext load without an arithmetic right shift!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16229, __PRETTY_FUNCTION__)); | |||
16230 | ||||
16231 | // Redistribute the loaded elements into the different locations. | |||
16232 | SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1); | |||
16233 | for (unsigned i = 0; i != NumElems; ++i) | |||
16234 | ShuffleVec[i * SizeRatio + SizeRatio - 1] = i; | |||
16235 | ||||
16236 | SDValue Shuff = DAG.getVectorShuffle( | |||
16237 | WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]); | |||
16238 | ||||
16239 | Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff); | |||
16240 | ||||
16241 | // Build the arithmetic shift. | |||
16242 | unsigned Amt = RegVT.getVectorElementType().getSizeInBits() - | |||
16243 | MemVT.getVectorElementType().getSizeInBits(); | |||
16244 | Shuff = | |||
16245 | DAG.getNode(ISD::SRA, dl, RegVT, Shuff, DAG.getConstant(Amt, RegVT)); | |||
16246 | ||||
16247 | DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF); | |||
16248 | return Shuff; | |||
16249 | } | |||
16250 | ||||
16251 | // Redistribute the loaded elements into the different locations. | |||
16252 | SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1); | |||
16253 | for (unsigned i = 0; i != NumElems; ++i) | |||
16254 | ShuffleVec[i * SizeRatio] = i; | |||
16255 | ||||
16256 | SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec, | |||
16257 | DAG.getUNDEF(WideVecVT), &ShuffleVec[0]); | |||
16258 | ||||
16259 | // Bitcast to the requested type. | |||
16260 | Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff); | |||
16261 | DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF); | |||
16262 | return Shuff; | |||
16263 | } | |||
16264 | ||||
16265 | // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or | |||
16266 | // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart | |||
16267 | // from the AND / OR. | |||
16268 | static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) { | |||
16269 | Opc = Op.getOpcode(); | |||
16270 | if (Opc != ISD::OR && Opc != ISD::AND) | |||
16271 | return false; | |||
16272 | return (Op.getOperand(0).getOpcode() == X86ISD::SETCC && | |||
16273 | Op.getOperand(0).hasOneUse() && | |||
16274 | Op.getOperand(1).getOpcode() == X86ISD::SETCC && | |||
16275 | Op.getOperand(1).hasOneUse()); | |||
16276 | } | |||
16277 | ||||
16278 | // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and | |||
16279 | // 1 and that the SETCC node has a single use. | |||
16280 | static bool isXor1OfSetCC(SDValue Op) { | |||
16281 | if (Op.getOpcode() != ISD::XOR) | |||
16282 | return false; | |||
16283 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); | |||
16284 | if (N1C && N1C->getAPIntValue() == 1) { | |||
16285 | return Op.getOperand(0).getOpcode() == X86ISD::SETCC && | |||
16286 | Op.getOperand(0).hasOneUse(); | |||
16287 | } | |||
16288 | return false; | |||
16289 | } | |||
16290 | ||||
16291 | SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const { | |||
16292 | bool addTest = true; | |||
16293 | SDValue Chain = Op.getOperand(0); | |||
16294 | SDValue Cond = Op.getOperand(1); | |||
16295 | SDValue Dest = Op.getOperand(2); | |||
16296 | SDLoc dl(Op); | |||
16297 | SDValue CC; | |||
16298 | bool Inverted = false; | |||
16299 | ||||
16300 | if (Cond.getOpcode() == ISD::SETCC) { | |||
16301 | // Check for setcc([su]{add,sub,mul}o == 0). | |||
16302 | if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ && | |||
16303 | isa<ConstantSDNode>(Cond.getOperand(1)) && | |||
16304 | cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() && | |||
16305 | Cond.getOperand(0).getResNo() == 1 && | |||
16306 | (Cond.getOperand(0).getOpcode() == ISD::SADDO || | |||
16307 | Cond.getOperand(0).getOpcode() == ISD::UADDO || | |||
16308 | Cond.getOperand(0).getOpcode() == ISD::SSUBO || | |||
16309 | Cond.getOperand(0).getOpcode() == ISD::USUBO || | |||
16310 | Cond.getOperand(0).getOpcode() == ISD::SMULO || | |||
16311 | Cond.getOperand(0).getOpcode() == ISD::UMULO)) { | |||
16312 | Inverted = true; | |||
16313 | Cond = Cond.getOperand(0); | |||
16314 | } else { | |||
16315 | SDValue NewCond = LowerSETCC(Cond, DAG); | |||
16316 | if (NewCond.getNode()) | |||
16317 | Cond = NewCond; | |||
16318 | } | |||
16319 | } | |||
16320 | #if 0 | |||
16321 | // FIXME: LowerXALUO doesn't handle these!! | |||
16322 | else if (Cond.getOpcode() == X86ISD::ADD || | |||
16323 | Cond.getOpcode() == X86ISD::SUB || | |||
16324 | Cond.getOpcode() == X86ISD::SMUL || | |||
16325 | Cond.getOpcode() == X86ISD::UMUL) | |||
16326 | Cond = LowerXALUO(Cond, DAG); | |||
16327 | #endif | |||
16328 | ||||
16329 | // Look pass (and (setcc_carry (cmp ...)), 1). | |||
16330 | if (Cond.getOpcode() == ISD::AND && | |||
16331 | Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) { | |||
16332 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1)); | |||
16333 | if (C && C->getAPIntValue() == 1) | |||
16334 | Cond = Cond.getOperand(0); | |||
16335 | } | |||
16336 | ||||
16337 | // If condition flag is set by a X86ISD::CMP, then use it as the condition | |||
16338 | // setting operand in place of the X86ISD::SETCC. | |||
16339 | unsigned CondOpcode = Cond.getOpcode(); | |||
16340 | if (CondOpcode == X86ISD::SETCC || | |||
16341 | CondOpcode == X86ISD::SETCC_CARRY) { | |||
16342 | CC = Cond.getOperand(0); | |||
16343 | ||||
16344 | SDValue Cmp = Cond.getOperand(1); | |||
16345 | unsigned Opc = Cmp.getOpcode(); | |||
16346 | // FIXME: WHY THE SPECIAL CASING OF LogicalCmp?? | |||
16347 | if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) { | |||
16348 | Cond = Cmp; | |||
16349 | addTest = false; | |||
16350 | } else { | |||
16351 | switch (cast<ConstantSDNode>(CC)->getZExtValue()) { | |||
16352 | default: break; | |||
16353 | case X86::COND_O: | |||
16354 | case X86::COND_B: | |||
16355 | // These can only come from an arithmetic instruction with overflow, | |||
16356 | // e.g. SADDO, UADDO. | |||
16357 | Cond = Cond.getNode()->getOperand(1); | |||
16358 | addTest = false; | |||
16359 | break; | |||
16360 | } | |||
16361 | } | |||
16362 | } | |||
16363 | CondOpcode = Cond.getOpcode(); | |||
16364 | if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO || | |||
16365 | CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO || | |||
16366 | ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) && | |||
16367 | Cond.getOperand(0).getValueType() != MVT::i8)) { | |||
16368 | SDValue LHS = Cond.getOperand(0); | |||
16369 | SDValue RHS = Cond.getOperand(1); | |||
16370 | unsigned X86Opcode; | |||
16371 | unsigned X86Cond; | |||
16372 | SDVTList VTs; | |||
16373 | // Keep this in sync with LowerXALUO, otherwise we might create redundant | |||
16374 | // instructions that can't be removed afterwards (i.e. X86ISD::ADD and | |||
16375 | // X86ISD::INC). | |||
16376 | switch (CondOpcode) { | |||
16377 | case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break; | |||
16378 | case ISD::SADDO: | |||
16379 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS)) | |||
16380 | if (C->isOne()) { | |||
16381 | X86Opcode = X86ISD::INC; X86Cond = X86::COND_O; | |||
16382 | break; | |||
16383 | } | |||
16384 | X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break; | |||
16385 | case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break; | |||
16386 | case ISD::SSUBO: | |||
16387 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS)) | |||
16388 | if (C->isOne()) { | |||
16389 | X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O; | |||
16390 | break; | |||
16391 | } | |||
16392 | X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break; | |||
16393 | case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break; | |||
16394 | case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break; | |||
16395 | default: llvm_unreachable("unexpected overflowing operator")::llvm::llvm_unreachable_internal("unexpected overflowing operator" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16395); | |||
16396 | } | |||
16397 | if (Inverted) | |||
16398 | X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond); | |||
16399 | if (CondOpcode == ISD::UMULO) | |||
16400 | VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(), | |||
16401 | MVT::i32); | |||
16402 | else | |||
16403 | VTs = DAG.getVTList(LHS.getValueType(), MVT::i32); | |||
16404 | ||||
16405 | SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS); | |||
16406 | ||||
16407 | if (CondOpcode == ISD::UMULO) | |||
16408 | Cond = X86Op.getValue(2); | |||
16409 | else | |||
16410 | Cond = X86Op.getValue(1); | |||
16411 | ||||
16412 | CC = DAG.getConstant(X86Cond, MVT::i8); | |||
16413 | addTest = false; | |||
16414 | } else { | |||
16415 | unsigned CondOpc; | |||
16416 | if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) { | |||
16417 | SDValue Cmp = Cond.getOperand(0).getOperand(1); | |||
16418 | if (CondOpc == ISD::OR) { | |||
16419 | // Also, recognize the pattern generated by an FCMP_UNE. We can emit | |||
16420 | // two branches instead of an explicit OR instruction with a | |||
16421 | // separate test. | |||
16422 | if (Cmp == Cond.getOperand(1).getOperand(1) && | |||
16423 | isX86LogicalCmp(Cmp)) { | |||
16424 | CC = Cond.getOperand(0).getOperand(0); | |||
16425 | Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(), | |||
16426 | Chain, Dest, CC, Cmp); | |||
16427 | CC = Cond.getOperand(1).getOperand(0); | |||
16428 | Cond = Cmp; | |||
16429 | addTest = false; | |||
16430 | } | |||
16431 | } else { // ISD::AND | |||
16432 | // Also, recognize the pattern generated by an FCMP_OEQ. We can emit | |||
16433 | // two branches instead of an explicit AND instruction with a | |||
16434 | // separate test. However, we only do this if this block doesn't | |||
16435 | // have a fall-through edge, because this requires an explicit | |||
16436 | // jmp when the condition is false. | |||
16437 | if (Cmp == Cond.getOperand(1).getOperand(1) && | |||
16438 | isX86LogicalCmp(Cmp) && | |||
16439 | Op.getNode()->hasOneUse()) { | |||
16440 | X86::CondCode CCode = | |||
16441 | (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0); | |||
16442 | CCode = X86::GetOppositeBranchCondition(CCode); | |||
16443 | CC = DAG.getConstant(CCode, MVT::i8); | |||
16444 | SDNode *User = *Op.getNode()->use_begin(); | |||
16445 | // Look for an unconditional branch following this conditional branch. | |||
16446 | // We need this because we need to reverse the successors in order | |||
16447 | // to implement FCMP_OEQ. | |||
16448 | if (User->getOpcode() == ISD::BR) { | |||
16449 | SDValue FalseBB = User->getOperand(1); | |||
16450 | SDNode *NewBR = | |||
16451 | DAG.UpdateNodeOperands(User, User->getOperand(0), Dest); | |||
16452 | assert(NewBR == User)((NewBR == User) ? static_cast<void> (0) : __assert_fail ("NewBR == User", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16452, __PRETTY_FUNCTION__)); | |||
16453 | (void)NewBR; | |||
16454 | Dest = FalseBB; | |||
16455 | ||||
16456 | Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(), | |||
16457 | Chain, Dest, CC, Cmp); | |||
16458 | X86::CondCode CCode = | |||
16459 | (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0); | |||
16460 | CCode = X86::GetOppositeBranchCondition(CCode); | |||
16461 | CC = DAG.getConstant(CCode, MVT::i8); | |||
16462 | Cond = Cmp; | |||
16463 | addTest = false; | |||
16464 | } | |||
16465 | } | |||
16466 | } | |||
16467 | } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) { | |||
16468 | // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition. | |||
16469 | // It should be transformed during dag combiner except when the condition | |||
16470 | // is set by a arithmetics with overflow node. | |||
16471 | X86::CondCode CCode = | |||
16472 | (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0); | |||
16473 | CCode = X86::GetOppositeBranchCondition(CCode); | |||
16474 | CC = DAG.getConstant(CCode, MVT::i8); | |||
16475 | Cond = Cond.getOperand(0).getOperand(1); | |||
16476 | addTest = false; | |||
16477 | } else if (Cond.getOpcode() == ISD::SETCC && | |||
16478 | cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) { | |||
16479 | // For FCMP_OEQ, we can emit | |||
16480 | // two branches instead of an explicit AND instruction with a | |||
16481 | // separate test. However, we only do this if this block doesn't | |||
16482 | // have a fall-through edge, because this requires an explicit | |||
16483 | // jmp when the condition is false. | |||
16484 | if (Op.getNode()->hasOneUse()) { | |||
16485 | SDNode *User = *Op.getNode()->use_begin(); | |||
16486 | // Look for an unconditional branch following this conditional branch. | |||
16487 | // We need this because we need to reverse the successors in order | |||
16488 | // to implement FCMP_OEQ. | |||
16489 | if (User->getOpcode() == ISD::BR) { | |||
16490 | SDValue FalseBB = User->getOperand(1); | |||
16491 | SDNode *NewBR = | |||
16492 | DAG.UpdateNodeOperands(User, User->getOperand(0), Dest); | |||
16493 | assert(NewBR == User)((NewBR == User) ? static_cast<void> (0) : __assert_fail ("NewBR == User", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16493, __PRETTY_FUNCTION__)); | |||
16494 | (void)NewBR; | |||
16495 | Dest = FalseBB; | |||
16496 | ||||
16497 | SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32, | |||
16498 | Cond.getOperand(0), Cond.getOperand(1)); | |||
16499 | Cmp = ConvertCmpIfNecessary(Cmp, DAG); | |||
16500 | CC = DAG.getConstant(X86::COND_NE, MVT::i8); | |||
16501 | Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(), | |||
16502 | Chain, Dest, CC, Cmp); | |||
16503 | CC = DAG.getConstant(X86::COND_P, MVT::i8); | |||
16504 | Cond = Cmp; | |||
16505 | addTest = false; | |||
16506 | } | |||
16507 | } | |||
16508 | } else if (Cond.getOpcode() == ISD::SETCC && | |||
16509 | cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) { | |||
16510 | // For FCMP_UNE, we can emit | |||
16511 | // two branches instead of an explicit AND instruction with a | |||
16512 | // separate test. However, we only do this if this block doesn't | |||
16513 | // have a fall-through edge, because this requires an explicit | |||
16514 | // jmp when the condition is false. | |||
16515 | if (Op.getNode()->hasOneUse()) { | |||
16516 | SDNode *User = *Op.getNode()->use_begin(); | |||
16517 | // Look for an unconditional branch following this conditional branch. | |||
16518 | // We need this because we need to reverse the successors in order | |||
16519 | // to implement FCMP_UNE. | |||
16520 | if (User->getOpcode() == ISD::BR) { | |||
16521 | SDValue FalseBB = User->getOperand(1); | |||
16522 | SDNode *NewBR = | |||
16523 | DAG.UpdateNodeOperands(User, User->getOperand(0), Dest); | |||
16524 | assert(NewBR == User)((NewBR == User) ? static_cast<void> (0) : __assert_fail ("NewBR == User", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16524, __PRETTY_FUNCTION__)); | |||
16525 | (void)NewBR; | |||
16526 | ||||
16527 | SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32, | |||
16528 | Cond.getOperand(0), Cond.getOperand(1)); | |||
16529 | Cmp = ConvertCmpIfNecessary(Cmp, DAG); | |||
16530 | CC = DAG.getConstant(X86::COND_NE, MVT::i8); | |||
16531 | Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(), | |||
16532 | Chain, Dest, CC, Cmp); | |||
16533 | CC = DAG.getConstant(X86::COND_NP, MVT::i8); | |||
16534 | Cond = Cmp; | |||
16535 | addTest = false; | |||
16536 | Dest = FalseBB; | |||
16537 | } | |||
16538 | } | |||
16539 | } | |||
16540 | } | |||
16541 | ||||
16542 | if (addTest) { | |||
16543 | // Look pass the truncate if the high bits are known zero. | |||
16544 | if (isTruncWithZeroHighBitsInput(Cond, DAG)) | |||
16545 | Cond = Cond.getOperand(0); | |||
16546 | ||||
16547 | // We know the result of AND is compared against zero. Try to match | |||
16548 | // it to BT. | |||
16549 | if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) { | |||
16550 | SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG); | |||
16551 | if (NewSetCC.getNode()) { | |||
16552 | CC = NewSetCC.getOperand(0); | |||
16553 | Cond = NewSetCC.getOperand(1); | |||
16554 | addTest = false; | |||
16555 | } | |||
16556 | } | |||
16557 | } | |||
16558 | ||||
16559 | if (addTest) { | |||
16560 | X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE; | |||
16561 | CC = DAG.getConstant(X86Cond, MVT::i8); | |||
16562 | Cond = EmitTest(Cond, X86Cond, dl, DAG); | |||
16563 | } | |||
16564 | Cond = ConvertCmpIfNecessary(Cond, DAG); | |||
16565 | return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(), | |||
16566 | Chain, Dest, CC, Cond); | |||
16567 | } | |||
16568 | ||||
16569 | // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets. | |||
16570 | // Calls to _alloca are needed to probe the stack when allocating more than 4k | |||
16571 | // bytes in one go. Touching the stack at 4K increments is necessary to ensure | |||
16572 | // that the guard pages used by the OS virtual memory manager are allocated in | |||
16573 | // correct sequence. | |||
16574 | SDValue | |||
16575 | X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, | |||
16576 | SelectionDAG &DAG) const { | |||
16577 | MachineFunction &MF = DAG.getMachineFunction(); | |||
16578 | bool SplitStack = MF.shouldSplitStack(); | |||
16579 | bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) || | |||
16580 | SplitStack; | |||
16581 | SDLoc dl(Op); | |||
16582 | ||||
16583 | if (!Lower) { | |||
16584 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
16585 | SDNode* Node = Op.getNode(); | |||
16586 | ||||
16587 | unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore(); | |||
16588 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16589, __PRETTY_FUNCTION__)) | |||
16589 | " 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16589, __PRETTY_FUNCTION__)); | |||
16590 | EVT VT = Node->getValueType(0); | |||
16591 | SDValue Tmp1 = SDValue(Node, 0); | |||
16592 | SDValue Tmp2 = SDValue(Node, 1); | |||
16593 | SDValue Tmp3 = Node->getOperand(2); | |||
16594 | SDValue Chain = Tmp1.getOperand(0); | |||
16595 | ||||
16596 | // Chain the dynamic stack allocation so that it doesn't modify the stack | |||
16597 | // pointer when other instructions are using the stack. | |||
16598 | Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true), | |||
16599 | SDLoc(Node)); | |||
16600 | ||||
16601 | SDValue Size = Tmp2.getOperand(1); | |||
16602 | SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT); | |||
16603 | Chain = SP.getValue(1); | |||
16604 | unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue(); | |||
16605 | const TargetFrameLowering &TFI = *DAG.getSubtarget().getFrameLowering(); | |||
16606 | unsigned StackAlign = TFI.getStackAlignment(); | |||
16607 | Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value | |||
16608 | if (Align > StackAlign) | |||
16609 | Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1, | |||
16610 | DAG.getConstant(-(uint64_t)Align, VT)); | |||
16611 | Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain | |||
16612 | ||||
16613 | Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true), | |||
16614 | DAG.getIntPtrConstant(0, true), SDValue(), | |||
16615 | SDLoc(Node)); | |||
16616 | ||||
16617 | SDValue Ops[2] = { Tmp1, Tmp2 }; | |||
16618 | return DAG.getMergeValues(Ops, dl); | |||
16619 | } | |||
16620 | ||||
16621 | // Get the inputs. | |||
16622 | SDValue Chain = Op.getOperand(0); | |||
16623 | SDValue Size = Op.getOperand(1); | |||
16624 | unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue(); | |||
16625 | EVT VT = Op.getNode()->getValueType(0); | |||
16626 | ||||
16627 | bool Is64Bit = Subtarget->is64Bit(); | |||
16628 | EVT SPTy = getPointerTy(); | |||
16629 | ||||
16630 | if (SplitStack) { | |||
16631 | MachineRegisterInfo &MRI = MF.getRegInfo(); | |||
16632 | ||||
16633 | if (Is64Bit) { | |||
16634 | // The 64 bit implementation of segmented stacks needs to clobber both r10 | |||
16635 | // r11. This makes it impossible to use it along with nested parameters. | |||
16636 | const Function *F = MF.getFunction(); | |||
16637 | ||||
16638 | for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end(); | |||
16639 | I != E; ++I) | |||
16640 | if (I->hasNestAttr()) | |||
16641 | report_fatal_error("Cannot use segmented stacks with functions that " | |||
16642 | "have nested arguments."); | |||
16643 | } | |||
16644 | ||||
16645 | const TargetRegisterClass *AddrRegClass = | |||
16646 | getRegClassFor(getPointerTy()); | |||
16647 | unsigned Vreg = MRI.createVirtualRegister(AddrRegClass); | |||
16648 | Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size); | |||
16649 | SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain, | |||
16650 | DAG.getRegister(Vreg, SPTy)); | |||
16651 | SDValue Ops1[2] = { Value, Chain }; | |||
16652 | return DAG.getMergeValues(Ops1, dl); | |||
16653 | } else { | |||
16654 | SDValue Flag; | |||
16655 | const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX); | |||
16656 | ||||
16657 | Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag); | |||
16658 | Flag = Chain.getValue(1); | |||
16659 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
16660 | ||||
16661 | Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag); | |||
16662 | ||||
16663 | const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>( | |||
16664 | DAG.getSubtarget().getRegisterInfo()); | |||
16665 | unsigned SPReg = RegInfo->getStackRegister(); | |||
16666 | SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy); | |||
16667 | Chain = SP.getValue(1); | |||
16668 | ||||
16669 | if (Align) { | |||
16670 | SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0), | |||
16671 | DAG.getConstant(-(uint64_t)Align, VT)); | |||
16672 | Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP); | |||
16673 | } | |||
16674 | ||||
16675 | SDValue Ops1[2] = { SP, Chain }; | |||
16676 | return DAG.getMergeValues(Ops1, dl); | |||
16677 | } | |||
16678 | } | |||
16679 | ||||
16680 | SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const { | |||
16681 | MachineFunction &MF = DAG.getMachineFunction(); | |||
16682 | X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); | |||
16683 | ||||
16684 | const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); | |||
16685 | SDLoc DL(Op); | |||
16686 | ||||
16687 | if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) { | |||
16688 | // vastart just stores the address of the VarArgsFrameIndex slot into the | |||
16689 | // memory location argument. | |||
16690 | SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), | |||
16691 | getPointerTy()); | |||
16692 | return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1), | |||
16693 | MachinePointerInfo(SV), false, false, 0); | |||
16694 | } | |||
16695 | ||||
16696 | // __va_list_tag: | |||
16697 | // gp_offset (0 - 6 * 8) | |||
16698 | // fp_offset (48 - 48 + 8 * 16) | |||
16699 | // overflow_arg_area (point to parameters coming in memory). | |||
16700 | // reg_save_area | |||
16701 | SmallVector<SDValue, 8> MemOps; | |||
16702 | SDValue FIN = Op.getOperand(1); | |||
16703 | // Store gp_offset | |||
16704 | SDValue Store = DAG.getStore(Op.getOperand(0), DL, | |||
16705 | DAG.getConstant(FuncInfo->getVarArgsGPOffset(), | |||
16706 | MVT::i32), | |||
16707 | FIN, MachinePointerInfo(SV), false, false, 0); | |||
16708 | MemOps.push_back(Store); | |||
16709 | ||||
16710 | // Store fp_offset | |||
16711 | FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), | |||
16712 | FIN, DAG.getIntPtrConstant(4)); | |||
16713 | Store = DAG.getStore(Op.getOperand(0), DL, | |||
16714 | DAG.getConstant(FuncInfo->getVarArgsFPOffset(), | |||
16715 | MVT::i32), | |||
16716 | FIN, MachinePointerInfo(SV, 4), false, false, 0); | |||
16717 | MemOps.push_back(Store); | |||
16718 | ||||
16719 | // Store ptr to overflow_arg_area | |||
16720 | FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), | |||
16721 | FIN, DAG.getIntPtrConstant(4)); | |||
16722 | SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), | |||
16723 | getPointerTy()); | |||
16724 | Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN, | |||
16725 | MachinePointerInfo(SV, 8), | |||
16726 | false, false, 0); | |||
16727 | MemOps.push_back(Store); | |||
16728 | ||||
16729 | // Store ptr to reg_save_area. | |||
16730 | FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), | |||
16731 | FIN, DAG.getIntPtrConstant(8)); | |||
16732 | SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), | |||
16733 | getPointerTy()); | |||
16734 | Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, | |||
16735 | MachinePointerInfo(SV, 16), false, false, 0); | |||
16736 | MemOps.push_back(Store); | |||
16737 | return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps); | |||
16738 | } | |||
16739 | ||||
16740 | SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const { | |||
16741 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16742, __PRETTY_FUNCTION__)) | |||
16742 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16742, __PRETTY_FUNCTION__)); | |||
16743 | assert((Subtarget->isTargetLinux() ||(((Subtarget->isTargetLinux() || Subtarget->isTargetDarwin ()) && "Unhandled target in LowerVAARG") ? static_cast <void> (0) : __assert_fail ("(Subtarget->isTargetLinux() || Subtarget->isTargetDarwin()) && \"Unhandled target in LowerVAARG\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16745, __PRETTY_FUNCTION__)) | |||
16744 | Subtarget->isTargetDarwin()) &&(((Subtarget->isTargetLinux() || Subtarget->isTargetDarwin ()) && "Unhandled target in LowerVAARG") ? static_cast <void> (0) : __assert_fail ("(Subtarget->isTargetLinux() || Subtarget->isTargetDarwin()) && \"Unhandled target in LowerVAARG\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16745, __PRETTY_FUNCTION__)) | |||
16745 | "Unhandled target in LowerVAARG")(((Subtarget->isTargetLinux() || Subtarget->isTargetDarwin ()) && "Unhandled target in LowerVAARG") ? static_cast <void> (0) : __assert_fail ("(Subtarget->isTargetLinux() || Subtarget->isTargetDarwin()) && \"Unhandled target in LowerVAARG\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16745, __PRETTY_FUNCTION__)); | |||
16746 | assert(Op.getNode()->getNumOperands() == 4)((Op.getNode()->getNumOperands() == 4) ? static_cast<void > (0) : __assert_fail ("Op.getNode()->getNumOperands() == 4" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16746, __PRETTY_FUNCTION__)); | |||
16747 | SDValue Chain = Op.getOperand(0); | |||
16748 | SDValue SrcPtr = Op.getOperand(1); | |||
16749 | const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); | |||
16750 | unsigned Align = Op.getConstantOperandVal(3); | |||
16751 | SDLoc dl(Op); | |||
16752 | ||||
16753 | EVT ArgVT = Op.getNode()->getValueType(0); | |||
16754 | Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext()); | |||
16755 | uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy); | |||
16756 | uint8_t ArgMode; | |||
16757 | ||||
16758 | // Decide which area this value should be read from. | |||
16759 | // TODO: Implement the AMD64 ABI in its entirety. This simple | |||
16760 | // selection mechanism works only for the basic types. | |||
16761 | if (ArgVT == MVT::f80) { | |||
16762 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16762); | |||
16763 | } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) { | |||
16764 | ArgMode = 2; // Argument passed in XMM register. Use fp_offset. | |||
16765 | } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) { | |||
16766 | ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset. | |||
16767 | } else { | |||
16768 | llvm_unreachable("Unhandled argument type in LowerVAARG")::llvm::llvm_unreachable_internal("Unhandled argument type in LowerVAARG" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16768); | |||
16769 | } | |||
16770 | ||||
16771 | if (ArgMode == 2) { | |||
16772 | // Sanity Check: Make sure using fp_offset makes sense. | |||
16773 | assert(!DAG.getTarget().Options.UseSoftFloat &&((!DAG.getTarget().Options.UseSoftFloat && !(DAG.getMachineFunction () .getFunction()->getAttributes() .hasAttribute(AttributeSet ::FunctionIndex, Attribute::NoImplicitFloat)) && Subtarget ->hasSSE1()) ? static_cast<void> (0) : __assert_fail ("!DAG.getTarget().Options.UseSoftFloat && !(DAG.getMachineFunction() .getFunction()->getAttributes() .hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat)) && Subtarget->hasSSE1()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16778, __PRETTY_FUNCTION__)) | |||
16774 | !(DAG.getMachineFunction()((!DAG.getTarget().Options.UseSoftFloat && !(DAG.getMachineFunction () .getFunction()->getAttributes() .hasAttribute(AttributeSet ::FunctionIndex, Attribute::NoImplicitFloat)) && Subtarget ->hasSSE1()) ? static_cast<void> (0) : __assert_fail ("!DAG.getTarget().Options.UseSoftFloat && !(DAG.getMachineFunction() .getFunction()->getAttributes() .hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat)) && Subtarget->hasSSE1()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16778, __PRETTY_FUNCTION__)) | |||
16775 | .getFunction()->getAttributes()((!DAG.getTarget().Options.UseSoftFloat && !(DAG.getMachineFunction () .getFunction()->getAttributes() .hasAttribute(AttributeSet ::FunctionIndex, Attribute::NoImplicitFloat)) && Subtarget ->hasSSE1()) ? static_cast<void> (0) : __assert_fail ("!DAG.getTarget().Options.UseSoftFloat && !(DAG.getMachineFunction() .getFunction()->getAttributes() .hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat)) && Subtarget->hasSSE1()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16778, __PRETTY_FUNCTION__)) | |||
16776 | .hasAttribute(AttributeSet::FunctionIndex,((!DAG.getTarget().Options.UseSoftFloat && !(DAG.getMachineFunction () .getFunction()->getAttributes() .hasAttribute(AttributeSet ::FunctionIndex, Attribute::NoImplicitFloat)) && Subtarget ->hasSSE1()) ? static_cast<void> (0) : __assert_fail ("!DAG.getTarget().Options.UseSoftFloat && !(DAG.getMachineFunction() .getFunction()->getAttributes() .hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat)) && Subtarget->hasSSE1()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16778, __PRETTY_FUNCTION__)) | |||
16777 | Attribute::NoImplicitFloat)) &&((!DAG.getTarget().Options.UseSoftFloat && !(DAG.getMachineFunction () .getFunction()->getAttributes() .hasAttribute(AttributeSet ::FunctionIndex, Attribute::NoImplicitFloat)) && Subtarget ->hasSSE1()) ? static_cast<void> (0) : __assert_fail ("!DAG.getTarget().Options.UseSoftFloat && !(DAG.getMachineFunction() .getFunction()->getAttributes() .hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat)) && Subtarget->hasSSE1()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16778, __PRETTY_FUNCTION__)) | |||
16778 | Subtarget->hasSSE1())((!DAG.getTarget().Options.UseSoftFloat && !(DAG.getMachineFunction () .getFunction()->getAttributes() .hasAttribute(AttributeSet ::FunctionIndex, Attribute::NoImplicitFloat)) && Subtarget ->hasSSE1()) ? static_cast<void> (0) : __assert_fail ("!DAG.getTarget().Options.UseSoftFloat && !(DAG.getMachineFunction() .getFunction()->getAttributes() .hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat)) && Subtarget->hasSSE1()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16778, __PRETTY_FUNCTION__)); | |||
16779 | } | |||
16780 | ||||
16781 | // Insert VAARG_64 node into the DAG | |||
16782 | // VAARG_64 returns two values: Variable Argument Address, Chain | |||
16783 | SmallVector<SDValue, 11> InstOps; | |||
16784 | InstOps.push_back(Chain); | |||
16785 | InstOps.push_back(SrcPtr); | |||
16786 | InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32)); | |||
16787 | InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8)); | |||
16788 | InstOps.push_back(DAG.getConstant(Align, MVT::i32)); | |||
16789 | SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other); | |||
16790 | SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl, | |||
16791 | VTs, InstOps, MVT::i64, | |||
16792 | MachinePointerInfo(SV), | |||
16793 | /*Align=*/0, | |||
16794 | /*Volatile=*/false, | |||
16795 | /*ReadMem=*/true, | |||
16796 | /*WriteMem=*/true); | |||
16797 | Chain = VAARG.getValue(1); | |||
16798 | ||||
16799 | // Load the next argument and return it | |||
16800 | return DAG.getLoad(ArgVT, dl, | |||
16801 | Chain, | |||
16802 | VAARG, | |||
16803 | MachinePointerInfo(), | |||
16804 | false, false, false, 0); | |||
16805 | } | |||
16806 | ||||
16807 | static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget, | |||
16808 | SelectionDAG &DAG) { | |||
16809 | // X86-64 va_list is a struct { i32, i32, i8*, i8* }. | |||
16810 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16810, __PRETTY_FUNCTION__)); | |||
16811 | SDValue Chain = Op.getOperand(0); | |||
16812 | SDValue DstPtr = Op.getOperand(1); | |||
16813 | SDValue SrcPtr = Op.getOperand(2); | |||
16814 | const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue(); | |||
16815 | const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue(); | |||
16816 | SDLoc DL(Op); | |||
16817 | ||||
16818 | return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr, | |||
16819 | DAG.getIntPtrConstant(24), 8, /*isVolatile*/false, | |||
16820 | false, | |||
16821 | MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV)); | |||
16822 | } | |||
16823 | ||||
16824 | // getTargetVShiftByConstNode - Handle vector element shifts where the shift | |||
16825 | // amount is a constant. Takes immediate version of shift as input. | |||
16826 | static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT, | |||
16827 | SDValue SrcOp, uint64_t ShiftAmt, | |||
16828 | SelectionDAG &DAG) { | |||
16829 | MVT ElementType = VT.getVectorElementType(); | |||
16830 | ||||
16831 | // Fold this packed shift into its first operand if ShiftAmt is 0. | |||
16832 | if (ShiftAmt == 0) | |||
16833 | return SrcOp; | |||
16834 | ||||
16835 | // Check for ShiftAmt >= element width | |||
16836 | if (ShiftAmt >= ElementType.getSizeInBits()) { | |||
16837 | if (Opc == X86ISD::VSRAI) | |||
16838 | ShiftAmt = ElementType.getSizeInBits() - 1; | |||
16839 | else | |||
16840 | return DAG.getConstant(0, VT); | |||
16841 | } | |||
16842 | ||||
16843 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16844, __PRETTY_FUNCTION__)) | |||
16844 | && "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16844, __PRETTY_FUNCTION__)); | |||
16845 | ||||
16846 | // Fold this packed vector shift into a build vector if SrcOp is a | |||
16847 | // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT. | |||
16848 | if (VT == SrcOp.getSimpleValueType() && | |||
16849 | ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) { | |||
16850 | SmallVector<SDValue, 8> Elts; | |||
16851 | unsigned NumElts = SrcOp->getNumOperands(); | |||
16852 | ConstantSDNode *ND; | |||
16853 | ||||
16854 | switch(Opc) { | |||
16855 | default: llvm_unreachable(nullptr)::llvm::llvm_unreachable_internal(nullptr, "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16855); | |||
16856 | case X86ISD::VSHLI: | |||
16857 | for (unsigned i=0; i!=NumElts; ++i) { | |||
16858 | SDValue CurrentOp = SrcOp->getOperand(i); | |||
16859 | if (CurrentOp->getOpcode() == ISD::UNDEF) { | |||
16860 | Elts.push_back(CurrentOp); | |||
16861 | continue; | |||
16862 | } | |||
16863 | ND = cast<ConstantSDNode>(CurrentOp); | |||
16864 | const APInt &C = ND->getAPIntValue(); | |||
16865 | Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), ElementType)); | |||
16866 | } | |||
16867 | break; | |||
16868 | case X86ISD::VSRLI: | |||
16869 | for (unsigned i=0; i!=NumElts; ++i) { | |||
16870 | SDValue CurrentOp = SrcOp->getOperand(i); | |||
16871 | if (CurrentOp->getOpcode() == ISD::UNDEF) { | |||
16872 | Elts.push_back(CurrentOp); | |||
16873 | continue; | |||
16874 | } | |||
16875 | ND = cast<ConstantSDNode>(CurrentOp); | |||
16876 | const APInt &C = ND->getAPIntValue(); | |||
16877 | Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), ElementType)); | |||
16878 | } | |||
16879 | break; | |||
16880 | case X86ISD::VSRAI: | |||
16881 | for (unsigned i=0; i!=NumElts; ++i) { | |||
16882 | SDValue CurrentOp = SrcOp->getOperand(i); | |||
16883 | if (CurrentOp->getOpcode() == ISD::UNDEF) { | |||
16884 | Elts.push_back(CurrentOp); | |||
16885 | continue; | |||
16886 | } | |||
16887 | ND = cast<ConstantSDNode>(CurrentOp); | |||
16888 | const APInt &C = ND->getAPIntValue(); | |||
16889 | Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), ElementType)); | |||
16890 | } | |||
16891 | break; | |||
16892 | } | |||
16893 | ||||
16894 | return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts); | |||
16895 | } | |||
16896 | ||||
16897 | return DAG.getNode(Opc, dl, VT, SrcOp, DAG.getConstant(ShiftAmt, MVT::i8)); | |||
16898 | } | |||
16899 | ||||
16900 | // getTargetVShiftNode - Handle vector element shifts where the shift amount | |||
16901 | // may or may not be a constant. Takes immediate version of shift as input. | |||
16902 | static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT, | |||
16903 | SDValue SrcOp, SDValue ShAmt, | |||
16904 | SelectionDAG &DAG) { | |||
16905 | MVT SVT = ShAmt.getSimpleValueType(); | |||
16906 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16906, __PRETTY_FUNCTION__)); | |||
16907 | ||||
16908 | // Catch shift-by-constant. | |||
16909 | if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt)) | |||
16910 | return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp, | |||
16911 | CShAmt->getZExtValue(), DAG); | |||
16912 | ||||
16913 | // Change opcode to non-immediate version | |||
16914 | switch (Opc) { | |||
16915 | default: llvm_unreachable("Unknown target vector shift node")::llvm::llvm_unreachable_internal("Unknown target vector shift node" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16915); | |||
16916 | case X86ISD::VSHLI: Opc = X86ISD::VSHL; break; | |||
16917 | case X86ISD::VSRLI: Opc = X86ISD::VSRL; break; | |||
16918 | case X86ISD::VSRAI: Opc = X86ISD::VSRA; break; | |||
16919 | } | |||
16920 | ||||
16921 | const X86Subtarget &Subtarget = | |||
16922 | DAG.getTarget().getSubtarget<X86Subtarget>(); | |||
16923 | if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND && | |||
16924 | ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) { | |||
16925 | // Let the shuffle legalizer expand this shift amount node. | |||
16926 | SDValue Op0 = ShAmt.getOperand(0); | |||
16927 | Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0); | |||
16928 | ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG); | |||
16929 | } else { | |||
16930 | // Need to build a vector containing shift amount. | |||
16931 | // SSE/AVX packed shifts only use the lower 64-bit of the shift count. | |||
16932 | SmallVector<SDValue, 4> ShOps; | |||
16933 | ShOps.push_back(ShAmt); | |||
16934 | if (SVT == MVT::i32) { | |||
16935 | ShOps.push_back(DAG.getConstant(0, SVT)); | |||
16936 | ShOps.push_back(DAG.getUNDEF(SVT)); | |||
16937 | } | |||
16938 | ShOps.push_back(DAG.getUNDEF(SVT)); | |||
16939 | ||||
16940 | MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64; | |||
16941 | ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps); | |||
16942 | } | |||
16943 | ||||
16944 | // The return type has to be a 128-bit type with the same element | |||
16945 | // type as the input type. | |||
16946 | MVT EltVT = VT.getVectorElementType(); | |||
16947 | EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits()); | |||
16948 | ||||
16949 | ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt); | |||
16950 | return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt); | |||
16951 | } | |||
16952 | ||||
16953 | /// \brief Return (and \p Op, \p Mask) for compare instructions or | |||
16954 | /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the | |||
16955 | /// necessary casting for \p Mask when lowering masking intrinsics. | |||
16956 | static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask, | |||
16957 | SDValue PreservedSrc, | |||
16958 | const X86Subtarget *Subtarget, | |||
16959 | SelectionDAG &DAG) { | |||
16960 | EVT VT = Op.getValueType(); | |||
16961 | EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), | |||
16962 | MVT::i1, VT.getVectorNumElements()); | |||
16963 | EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, | |||
16964 | Mask.getValueType().getSizeInBits()); | |||
16965 | SDLoc dl(Op); | |||
16966 | ||||
16967 | assert(MaskVT.isSimple() && "invalid mask type")((MaskVT.isSimple() && "invalid mask type") ? static_cast <void> (0) : __assert_fail ("MaskVT.isSimple() && \"invalid mask type\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 16967, __PRETTY_FUNCTION__)); | |||
16968 | ||||
16969 | if (isAllOnes(Mask)) | |||
16970 | return Op; | |||
16971 | ||||
16972 | // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements | |||
16973 | // are extracted by EXTRACT_SUBVECTOR. | |||
16974 | SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT, | |||
16975 | DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask), | |||
16976 | DAG.getIntPtrConstant(0)); | |||
16977 | ||||
16978 | switch (Op.getOpcode()) { | |||
16979 | default: break; | |||
16980 | case X86ISD::PCMPEQM: | |||
16981 | case X86ISD::PCMPGTM: | |||
16982 | case X86ISD::CMPM: | |||
16983 | case X86ISD::CMPMU: | |||
16984 | return DAG.getNode(ISD::AND, dl, VT, Op, VMask); | |||
16985 | } | |||
16986 | if (PreservedSrc.getOpcode() == ISD::UNDEF) | |||
16987 | PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl); | |||
16988 | return DAG.getNode(ISD::VSELECT, dl, VT, VMask, Op, PreservedSrc); | |||
16989 | } | |||
16990 | ||||
16991 | /// \brief Creates an SDNode for a predicated scalar operation. | |||
16992 | /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc). | |||
16993 | /// The mask is comming as MVT::i8 and it should be truncated | |||
16994 | /// to MVT::i1 while lowering masking intrinsics. | |||
16995 | /// The main difference between ScalarMaskingNode and VectorMaskingNode is using | |||
16996 | /// "X86select" instead of "vselect". We just can't create the "vselect" node for | |||
16997 | /// a scalar instruction. | |||
16998 | static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask, | |||
16999 | SDValue PreservedSrc, | |||
17000 | const X86Subtarget *Subtarget, | |||
17001 | SelectionDAG &DAG) { | |||
17002 | if (isAllOnes(Mask)) | |||
17003 | return Op; | |||
17004 | ||||
17005 | EVT VT = Op.getValueType(); | |||
17006 | SDLoc dl(Op); | |||
17007 | // The mask should be of type MVT::i1 | |||
17008 | SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask); | |||
17009 | ||||
17010 | if (PreservedSrc.getOpcode() == ISD::UNDEF) | |||
17011 | PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl); | |||
17012 | return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc); | |||
17013 | } | |||
17014 | ||||
17015 | static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget, | |||
17016 | SelectionDAG &DAG) { | |||
17017 | SDLoc dl(Op); | |||
17018 | unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); | |||
17019 | EVT VT = Op.getValueType(); | |||
17020 | const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo); | |||
17021 | if (IntrData) { | |||
17022 | switch(IntrData->Type) { | |||
17023 | case INTR_TYPE_1OP: | |||
17024 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1)); | |||
17025 | case INTR_TYPE_2OP: | |||
17026 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1), | |||
17027 | Op.getOperand(2)); | |||
17028 | case INTR_TYPE_3OP: | |||
17029 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1), | |||
17030 | Op.getOperand(2), Op.getOperand(3)); | |||
17031 | case INTR_TYPE_1OP_MASK_RM: { | |||
17032 | SDValue Src = Op.getOperand(1); | |||
17033 | SDValue Src0 = Op.getOperand(2); | |||
17034 | SDValue Mask = Op.getOperand(3); | |||
17035 | SDValue RoundingMode = Op.getOperand(4); | |||
17036 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src, | |||
17037 | RoundingMode), | |||
17038 | Mask, Src0, Subtarget, DAG); | |||
17039 | } | |||
17040 | case INTR_TYPE_SCALAR_MASK_RM: { | |||
17041 | SDValue Src1 = Op.getOperand(1); | |||
17042 | SDValue Src2 = Op.getOperand(2); | |||
17043 | SDValue Src0 = Op.getOperand(3); | |||
17044 | SDValue Mask = Op.getOperand(4); | |||
17045 | SDValue RoundingMode = Op.getOperand(5); | |||
17046 | return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2, | |||
17047 | RoundingMode), | |||
17048 | Mask, Src0, Subtarget, DAG); | |||
17049 | } | |||
17050 | case INTR_TYPE_2OP_MASK: { | |||
17051 | SDValue Mask = Op.getOperand(4); | |||
17052 | SDValue PassThru = Op.getOperand(3); | |||
17053 | unsigned IntrWithRoundingModeOpcode = IntrData->Opc1; | |||
17054 | if (IntrWithRoundingModeOpcode != 0) { | |||
17055 | unsigned Round = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue(); | |||
17056 | if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) { | |||
17057 | return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode, | |||
17058 | dl, Op.getValueType(), | |||
17059 | Op.getOperand(1), Op.getOperand(2), | |||
17060 | Op.getOperand(3), Op.getOperand(5)), | |||
17061 | Mask, PassThru, Subtarget, DAG); | |||
17062 | } | |||
17063 | } | |||
17064 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, | |||
17065 | Op.getOperand(1), | |||
17066 | Op.getOperand(2)), | |||
17067 | Mask, PassThru, Subtarget, DAG); | |||
17068 | } | |||
17069 | case FMA_OP_MASK: { | |||
17070 | SDValue Src1 = Op.getOperand(1); | |||
17071 | SDValue Src2 = Op.getOperand(2); | |||
17072 | SDValue Src3 = Op.getOperand(3); | |||
17073 | SDValue Mask = Op.getOperand(4); | |||
17074 | unsigned IntrWithRoundingModeOpcode = IntrData->Opc1; | |||
17075 | if (IntrWithRoundingModeOpcode != 0) { | |||
17076 | SDValue Rnd = Op.getOperand(5); | |||
17077 | if (cast<ConstantSDNode>(Rnd)->getZExtValue() != | |||
17078 | X86::STATIC_ROUNDING::CUR_DIRECTION) | |||
17079 | return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode, | |||
17080 | dl, Op.getValueType(), | |||
17081 | Src1, Src2, Src3, Rnd), | |||
17082 | Mask, Src1, Subtarget, DAG); | |||
17083 | } | |||
17084 | return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, | |||
17085 | dl, Op.getValueType(), | |||
17086 | Src1, Src2, Src3), | |||
17087 | Mask, Src1, Subtarget, DAG); | |||
17088 | } | |||
17089 | case CMP_MASK: | |||
17090 | case CMP_MASK_CC: { | |||
17091 | // Comparison intrinsics with masks. | |||
17092 | // Example of transformation: | |||
17093 | // (i8 (int_x86_avx512_mask_pcmpeq_q_128 | |||
17094 | // (v2i64 %a), (v2i64 %b), (i8 %mask))) -> | |||
17095 | // (i8 (bitcast | |||
17096 | // (v8i1 (insert_subvector undef, | |||
17097 | // (v2i1 (and (PCMPEQM %a, %b), | |||
17098 | // (extract_subvector | |||
17099 | // (v8i1 (bitcast %mask)), 0))), 0)))) | |||
17100 | EVT VT = Op.getOperand(1).getValueType(); | |||
17101 | EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, | |||
17102 | VT.getVectorNumElements()); | |||
17103 | SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3); | |||
17104 | EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, | |||
17105 | Mask.getValueType().getSizeInBits()); | |||
17106 | SDValue Cmp; | |||
17107 | if (IntrData->Type == CMP_MASK_CC) { | |||
17108 | Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1), | |||
17109 | Op.getOperand(2), Op.getOperand(3)); | |||
17110 | } else { | |||
17111 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17111, __PRETTY_FUNCTION__)); | |||
17112 | Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1), | |||
17113 | Op.getOperand(2)); | |||
17114 | } | |||
17115 | SDValue CmpMask = getVectorMaskingNode(Cmp, Mask, | |||
17116 | DAG.getTargetConstant(0, MaskVT), | |||
17117 | Subtarget, DAG); | |||
17118 | SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT, | |||
17119 | DAG.getUNDEF(BitcastVT), CmpMask, | |||
17120 | DAG.getIntPtrConstant(0)); | |||
17121 | return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res); | |||
17122 | } | |||
17123 | case COMI: { // Comparison intrinsics | |||
17124 | ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1; | |||
17125 | SDValue LHS = Op.getOperand(1); | |||
17126 | SDValue RHS = Op.getOperand(2); | |||
17127 | unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG); | |||
17128 | assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!")((X86CC != X86::COND_INVALID && "Unexpected illegal condition!" ) ? static_cast<void> (0) : __assert_fail ("X86CC != X86::COND_INVALID && \"Unexpected illegal condition!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17128, __PRETTY_FUNCTION__)); | |||
17129 | SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS); | |||
17130 | SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, | |||
17131 | DAG.getConstant(X86CC, MVT::i8), Cond); | |||
17132 | return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC); | |||
17133 | } | |||
17134 | case VSHIFT: | |||
17135 | return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(), | |||
17136 | Op.getOperand(1), Op.getOperand(2), DAG); | |||
17137 | case VSHIFT_MASK: | |||
17138 | return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl, | |||
17139 | Op.getSimpleValueType(), | |||
17140 | Op.getOperand(1), | |||
17141 | Op.getOperand(2), DAG), | |||
17142 | Op.getOperand(4), Op.getOperand(3), Subtarget, | |||
17143 | DAG); | |||
17144 | case COMPRESS_EXPAND_IN_REG: { | |||
17145 | SDValue Mask = Op.getOperand(3); | |||
17146 | SDValue DataToCompress = Op.getOperand(1); | |||
17147 | SDValue PassThru = Op.getOperand(2); | |||
17148 | if (isAllOnes(Mask)) // return data as is | |||
17149 | return Op.getOperand(1); | |||
17150 | EVT VT = Op.getValueType(); | |||
17151 | EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, | |||
17152 | VT.getVectorNumElements()); | |||
17153 | EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, | |||
17154 | Mask.getValueType().getSizeInBits()); | |||
17155 | SDLoc dl(Op); | |||
17156 | SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT, | |||
17157 | DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask), | |||
17158 | DAG.getIntPtrConstant(0)); | |||
17159 | ||||
17160 | return DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToCompress, | |||
17161 | PassThru); | |||
17162 | } | |||
17163 | case BLEND: { | |||
17164 | SDValue Mask = Op.getOperand(3); | |||
17165 | EVT VT = Op.getValueType(); | |||
17166 | EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, | |||
17167 | VT.getVectorNumElements()); | |||
17168 | EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, | |||
17169 | Mask.getValueType().getSizeInBits()); | |||
17170 | SDLoc dl(Op); | |||
17171 | SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT, | |||
17172 | DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask), | |||
17173 | DAG.getIntPtrConstant(0)); | |||
17174 | return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1), | |||
17175 | Op.getOperand(2)); | |||
17176 | } | |||
17177 | default: | |||
17178 | break; | |||
17179 | } | |||
17180 | } | |||
17181 | ||||
17182 | switch (IntNo) { | |||
17183 | default: return SDValue(); // Don't custom lower most intrinsics. | |||
17184 | ||||
17185 | case Intrinsic::x86_avx512_mask_valign_q_512: | |||
17186 | case Intrinsic::x86_avx512_mask_valign_d_512: | |||
17187 | // Vector source operands are swapped. | |||
17188 | return getVectorMaskingNode(DAG.getNode(X86ISD::VALIGN, dl, | |||
17189 | Op.getValueType(), Op.getOperand(2), | |||
17190 | Op.getOperand(1), | |||
17191 | Op.getOperand(3)), | |||
17192 | Op.getOperand(5), Op.getOperand(4), | |||
17193 | Subtarget, DAG); | |||
17194 | ||||
17195 | // ptest and testp intrinsics. The intrinsic these come from are designed to | |||
17196 | // return an integer value, not just an instruction so lower it to the ptest | |||
17197 | // or testp pattern and a setcc for the result. | |||
17198 | case Intrinsic::x86_sse41_ptestz: | |||
17199 | case Intrinsic::x86_sse41_ptestc: | |||
17200 | case Intrinsic::x86_sse41_ptestnzc: | |||
17201 | case Intrinsic::x86_avx_ptestz_256: | |||
17202 | case Intrinsic::x86_avx_ptestc_256: | |||
17203 | case Intrinsic::x86_avx_ptestnzc_256: | |||
17204 | case Intrinsic::x86_avx_vtestz_ps: | |||
17205 | case Intrinsic::x86_avx_vtestc_ps: | |||
17206 | case Intrinsic::x86_avx_vtestnzc_ps: | |||
17207 | case Intrinsic::x86_avx_vtestz_pd: | |||
17208 | case Intrinsic::x86_avx_vtestc_pd: | |||
17209 | case Intrinsic::x86_avx_vtestnzc_pd: | |||
17210 | case Intrinsic::x86_avx_vtestz_ps_256: | |||
17211 | case Intrinsic::x86_avx_vtestc_ps_256: | |||
17212 | case Intrinsic::x86_avx_vtestnzc_ps_256: | |||
17213 | case Intrinsic::x86_avx_vtestz_pd_256: | |||
17214 | case Intrinsic::x86_avx_vtestc_pd_256: | |||
17215 | case Intrinsic::x86_avx_vtestnzc_pd_256: { | |||
17216 | bool IsTestPacked = false; | |||
17217 | unsigned X86CC; | |||
17218 | switch (IntNo) { | |||
17219 | default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.")::llvm::llvm_unreachable_internal("Bad fallthrough in Intrinsic lowering." , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17219); | |||
17220 | case Intrinsic::x86_avx_vtestz_ps: | |||
17221 | case Intrinsic::x86_avx_vtestz_pd: | |||
17222 | case Intrinsic::x86_avx_vtestz_ps_256: | |||
17223 | case Intrinsic::x86_avx_vtestz_pd_256: | |||
17224 | IsTestPacked = true; // Fallthrough | |||
17225 | case Intrinsic::x86_sse41_ptestz: | |||
17226 | case Intrinsic::x86_avx_ptestz_256: | |||
17227 | // ZF = 1 | |||
17228 | X86CC = X86::COND_E; | |||
17229 | break; | |||
17230 | case Intrinsic::x86_avx_vtestc_ps: | |||
17231 | case Intrinsic::x86_avx_vtestc_pd: | |||
17232 | case Intrinsic::x86_avx_vtestc_ps_256: | |||
17233 | case Intrinsic::x86_avx_vtestc_pd_256: | |||
17234 | IsTestPacked = true; // Fallthrough | |||
17235 | case Intrinsic::x86_sse41_ptestc: | |||
17236 | case Intrinsic::x86_avx_ptestc_256: | |||
17237 | // CF = 1 | |||
17238 | X86CC = X86::COND_B; | |||
17239 | break; | |||
17240 | case Intrinsic::x86_avx_vtestnzc_ps: | |||
17241 | case Intrinsic::x86_avx_vtestnzc_pd: | |||
17242 | case Intrinsic::x86_avx_vtestnzc_ps_256: | |||
17243 | case Intrinsic::x86_avx_vtestnzc_pd_256: | |||
17244 | IsTestPacked = true; // Fallthrough | |||
17245 | case Intrinsic::x86_sse41_ptestnzc: | |||
17246 | case Intrinsic::x86_avx_ptestnzc_256: | |||
17247 | // ZF and CF = 0 | |||
17248 | X86CC = X86::COND_A; | |||
17249 | break; | |||
17250 | } | |||
17251 | ||||
17252 | SDValue LHS = Op.getOperand(1); | |||
17253 | SDValue RHS = Op.getOperand(2); | |||
17254 | unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST; | |||
17255 | SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS); | |||
17256 | SDValue CC = DAG.getConstant(X86CC, MVT::i8); | |||
17257 | SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test); | |||
17258 | return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC); | |||
17259 | } | |||
17260 | case Intrinsic::x86_avx512_kortestz_w: | |||
17261 | case Intrinsic::x86_avx512_kortestc_w: { | |||
17262 | unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B; | |||
17263 | SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1)); | |||
17264 | SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2)); | |||
17265 | SDValue CC = DAG.getConstant(X86CC, MVT::i8); | |||
17266 | SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS); | |||
17267 | SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test); | |||
17268 | return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC); | |||
17269 | } | |||
17270 | ||||
17271 | case Intrinsic::x86_sse42_pcmpistria128: | |||
17272 | case Intrinsic::x86_sse42_pcmpestria128: | |||
17273 | case Intrinsic::x86_sse42_pcmpistric128: | |||
17274 | case Intrinsic::x86_sse42_pcmpestric128: | |||
17275 | case Intrinsic::x86_sse42_pcmpistrio128: | |||
17276 | case Intrinsic::x86_sse42_pcmpestrio128: | |||
17277 | case Intrinsic::x86_sse42_pcmpistris128: | |||
17278 | case Intrinsic::x86_sse42_pcmpestris128: | |||
17279 | case Intrinsic::x86_sse42_pcmpistriz128: | |||
17280 | case Intrinsic::x86_sse42_pcmpestriz128: { | |||
17281 | unsigned Opcode; | |||
17282 | unsigned X86CC; | |||
17283 | switch (IntNo) { | |||
17284 | default: llvm_unreachable("Impossible intrinsic")::llvm::llvm_unreachable_internal("Impossible intrinsic", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17284); // Can't reach here. | |||
17285 | case Intrinsic::x86_sse42_pcmpistria128: | |||
17286 | Opcode = X86ISD::PCMPISTRI; | |||
17287 | X86CC = X86::COND_A; | |||
17288 | break; | |||
17289 | case Intrinsic::x86_sse42_pcmpestria128: | |||
17290 | Opcode = X86ISD::PCMPESTRI; | |||
17291 | X86CC = X86::COND_A; | |||
17292 | break; | |||
17293 | case Intrinsic::x86_sse42_pcmpistric128: | |||
17294 | Opcode = X86ISD::PCMPISTRI; | |||
17295 | X86CC = X86::COND_B; | |||
17296 | break; | |||
17297 | case Intrinsic::x86_sse42_pcmpestric128: | |||
17298 | Opcode = X86ISD::PCMPESTRI; | |||
17299 | X86CC = X86::COND_B; | |||
17300 | break; | |||
17301 | case Intrinsic::x86_sse42_pcmpistrio128: | |||
17302 | Opcode = X86ISD::PCMPISTRI; | |||
17303 | X86CC = X86::COND_O; | |||
17304 | break; | |||
17305 | case Intrinsic::x86_sse42_pcmpestrio128: | |||
17306 | Opcode = X86ISD::PCMPESTRI; | |||
17307 | X86CC = X86::COND_O; | |||
17308 | break; | |||
17309 | case Intrinsic::x86_sse42_pcmpistris128: | |||
17310 | Opcode = X86ISD::PCMPISTRI; | |||
17311 | X86CC = X86::COND_S; | |||
17312 | break; | |||
17313 | case Intrinsic::x86_sse42_pcmpestris128: | |||
17314 | Opcode = X86ISD::PCMPESTRI; | |||
17315 | X86CC = X86::COND_S; | |||
17316 | break; | |||
17317 | case Intrinsic::x86_sse42_pcmpistriz128: | |||
17318 | Opcode = X86ISD::PCMPISTRI; | |||
17319 | X86CC = X86::COND_E; | |||
17320 | break; | |||
17321 | case Intrinsic::x86_sse42_pcmpestriz128: | |||
17322 | Opcode = X86ISD::PCMPESTRI; | |||
17323 | X86CC = X86::COND_E; | |||
17324 | break; | |||
17325 | } | |||
17326 | SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end()); | |||
17327 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32); | |||
17328 | SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps); | |||
17329 | SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, | |||
17330 | DAG.getConstant(X86CC, MVT::i8), | |||
17331 | SDValue(PCMP.getNode(), 1)); | |||
17332 | return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC); | |||
17333 | } | |||
17334 | ||||
17335 | case Intrinsic::x86_sse42_pcmpistri128: | |||
17336 | case Intrinsic::x86_sse42_pcmpestri128: { | |||
17337 | unsigned Opcode; | |||
17338 | if (IntNo == Intrinsic::x86_sse42_pcmpistri128) | |||
17339 | Opcode = X86ISD::PCMPISTRI; | |||
17340 | else | |||
17341 | Opcode = X86ISD::PCMPESTRI; | |||
17342 | ||||
17343 | SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end()); | |||
17344 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32); | |||
17345 | return DAG.getNode(Opcode, dl, VTs, NewOps); | |||
17346 | } | |||
17347 | } | |||
17348 | } | |||
17349 | ||||
17350 | static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG, | |||
17351 | SDValue Src, SDValue Mask, SDValue Base, | |||
17352 | SDValue Index, SDValue ScaleOp, SDValue Chain, | |||
17353 | const X86Subtarget * Subtarget) { | |||
17354 | SDLoc dl(Op); | |||
17355 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp); | |||
17356 | assert(C && "Invalid scale type")((C && "Invalid scale type") ? static_cast<void> (0) : __assert_fail ("C && \"Invalid scale type\"", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17356, __PRETTY_FUNCTION__)); | |||
17357 | SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8); | |||
17358 | EVT MaskVT = MVT::getVectorVT(MVT::i1, | |||
17359 | Index.getSimpleValueType().getVectorNumElements()); | |||
17360 | SDValue MaskInReg; | |||
17361 | ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask); | |||
17362 | if (MaskC) | |||
17363 | MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT); | |||
17364 | else | |||
17365 | MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask); | |||
17366 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other); | |||
17367 | SDValue Disp = DAG.getTargetConstant(0, MVT::i32); | |||
17368 | SDValue Segment = DAG.getRegister(0, MVT::i32); | |||
17369 | if (Src.getOpcode() == ISD::UNDEF) | |||
17370 | Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl); | |||
17371 | SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain}; | |||
17372 | SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops); | |||
17373 | SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) }; | |||
17374 | return DAG.getMergeValues(RetOps, dl); | |||
17375 | } | |||
17376 | ||||
17377 | static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG, | |||
17378 | SDValue Src, SDValue Mask, SDValue Base, | |||
17379 | SDValue Index, SDValue ScaleOp, SDValue Chain) { | |||
17380 | SDLoc dl(Op); | |||
17381 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp); | |||
17382 | assert(C && "Invalid scale type")((C && "Invalid scale type") ? static_cast<void> (0) : __assert_fail ("C && \"Invalid scale type\"", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17382, __PRETTY_FUNCTION__)); | |||
17383 | SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8); | |||
17384 | SDValue Disp = DAG.getTargetConstant(0, MVT::i32); | |||
17385 | SDValue Segment = DAG.getRegister(0, MVT::i32); | |||
17386 | EVT MaskVT = MVT::getVectorVT(MVT::i1, | |||
17387 | Index.getSimpleValueType().getVectorNumElements()); | |||
17388 | SDValue MaskInReg; | |||
17389 | ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask); | |||
17390 | if (MaskC) | |||
17391 | MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT); | |||
17392 | else | |||
17393 | MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask); | |||
17394 | SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other); | |||
17395 | SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain}; | |||
17396 | SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops); | |||
17397 | return SDValue(Res, 1); | |||
17398 | } | |||
17399 | ||||
17400 | static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG, | |||
17401 | SDValue Mask, SDValue Base, SDValue Index, | |||
17402 | SDValue ScaleOp, SDValue Chain) { | |||
17403 | SDLoc dl(Op); | |||
17404 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp); | |||
17405 | assert(C && "Invalid scale type")((C && "Invalid scale type") ? static_cast<void> (0) : __assert_fail ("C && \"Invalid scale type\"", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17405, __PRETTY_FUNCTION__)); | |||
17406 | SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8); | |||
17407 | SDValue Disp = DAG.getTargetConstant(0, MVT::i32); | |||
17408 | SDValue Segment = DAG.getRegister(0, MVT::i32); | |||
17409 | EVT MaskVT = | |||
17410 | MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements()); | |||
17411 | SDValue MaskInReg; | |||
17412 | ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask); | |||
17413 | if (MaskC) | |||
17414 | MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT); | |||
17415 | else | |||
17416 | MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask); | |||
17417 | //SDVTList VTs = DAG.getVTList(MVT::Other); | |||
17418 | SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain}; | |||
17419 | SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops); | |||
17420 | return SDValue(Res, 0); | |||
17421 | } | |||
17422 | ||||
17423 | // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that | |||
17424 | // read performance monitor counters (x86_rdpmc). | |||
17425 | static void getReadPerformanceCounter(SDNode *N, SDLoc DL, | |||
17426 | SelectionDAG &DAG, const X86Subtarget *Subtarget, | |||
17427 | SmallVectorImpl<SDValue> &Results) { | |||
17428 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17428, __PRETTY_FUNCTION__)); | |||
17429 | SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
17430 | SDValue LO, HI; | |||
17431 | ||||
17432 | // The ECX register is used to select the index of the performance counter | |||
17433 | // to read. | |||
17434 | SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX, | |||
17435 | N->getOperand(2)); | |||
17436 | SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain); | |||
17437 | ||||
17438 | // Reads the content of a 64-bit performance counter and returns it in the | |||
17439 | // registers EDX:EAX. | |||
17440 | if (Subtarget->is64Bit()) { | |||
17441 | LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1)); | |||
17442 | HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64, | |||
17443 | LO.getValue(2)); | |||
17444 | } else { | |||
17445 | LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1)); | |||
17446 | HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32, | |||
17447 | LO.getValue(2)); | |||
17448 | } | |||
17449 | Chain = HI.getValue(1); | |||
17450 | ||||
17451 | if (Subtarget->is64Bit()) { | |||
17452 | // The EAX register is loaded with the low-order 32 bits. The EDX register | |||
17453 | // is loaded with the supported high-order bits of the counter. | |||
17454 | SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI, | |||
17455 | DAG.getConstant(32, MVT::i8)); | |||
17456 | Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp)); | |||
17457 | Results.push_back(Chain); | |||
17458 | return; | |||
17459 | } | |||
17460 | ||||
17461 | // Use a buildpair to merge the two 32-bit values into a 64-bit one. | |||
17462 | SDValue Ops[] = { LO, HI }; | |||
17463 | SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops); | |||
17464 | Results.push_back(Pair); | |||
17465 | Results.push_back(Chain); | |||
17466 | } | |||
17467 | ||||
17468 | // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that | |||
17469 | // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is | |||
17470 | // also used to custom lower READCYCLECOUNTER nodes. | |||
17471 | static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode, | |||
17472 | SelectionDAG &DAG, const X86Subtarget *Subtarget, | |||
17473 | SmallVectorImpl<SDValue> &Results) { | |||
17474 | SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
17475 | SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0)); | |||
17476 | SDValue LO, HI; | |||
17477 | ||||
17478 | // The processor's time-stamp counter (a 64-bit MSR) is stored into the | |||
17479 | // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR | |||
17480 | // and the EAX register is loaded with the low-order 32 bits. | |||
17481 | if (Subtarget->is64Bit()) { | |||
17482 | LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1)); | |||
17483 | HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64, | |||
17484 | LO.getValue(2)); | |||
17485 | } else { | |||
17486 | LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1)); | |||
17487 | HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32, | |||
17488 | LO.getValue(2)); | |||
17489 | } | |||
17490 | SDValue Chain = HI.getValue(1); | |||
17491 | ||||
17492 | if (Opcode == X86ISD::RDTSCP_DAG) { | |||
17493 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17493, __PRETTY_FUNCTION__)); | |||
17494 | ||||
17495 | // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into | |||
17496 | // the ECX register. Add 'ecx' explicitly to the chain. | |||
17497 | SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32, | |||
17498 | HI.getValue(2)); | |||
17499 | // Explicitly store the content of ECX at the location passed in input | |||
17500 | // to the 'rdtscp' intrinsic. | |||
17501 | Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2), | |||
17502 | MachinePointerInfo(), false, false, 0); | |||
17503 | } | |||
17504 | ||||
17505 | if (Subtarget->is64Bit()) { | |||
17506 | // The EDX register is loaded with the high-order 32 bits of the MSR, and | |||
17507 | // the EAX register is loaded with the low-order 32 bits. | |||
17508 | SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI, | |||
17509 | DAG.getConstant(32, MVT::i8)); | |||
17510 | Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp)); | |||
17511 | Results.push_back(Chain); | |||
17512 | return; | |||
17513 | } | |||
17514 | ||||
17515 | // Use a buildpair to merge the two 32-bit values into a 64-bit one. | |||
17516 | SDValue Ops[] = { LO, HI }; | |||
17517 | SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops); | |||
17518 | Results.push_back(Pair); | |||
17519 | Results.push_back(Chain); | |||
17520 | } | |||
17521 | ||||
17522 | static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget, | |||
17523 | SelectionDAG &DAG) { | |||
17524 | SmallVector<SDValue, 2> Results; | |||
17525 | SDLoc DL(Op); | |||
17526 | getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget, | |||
17527 | Results); | |||
17528 | return DAG.getMergeValues(Results, DL); | |||
17529 | } | |||
17530 | ||||
17531 | ||||
17532 | static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget, | |||
17533 | SelectionDAG &DAG) { | |||
17534 | unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); | |||
17535 | ||||
17536 | const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo); | |||
17537 | if (!IntrData) | |||
17538 | return SDValue(); | |||
17539 | ||||
17540 | SDLoc dl(Op); | |||
17541 | switch(IntrData->Type) { | |||
17542 | default: | |||
17543 | llvm_unreachable("Unknown Intrinsic Type")::llvm::llvm_unreachable_internal("Unknown Intrinsic Type", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17543); | |||
17544 | break; | |||
17545 | case RDSEED: | |||
17546 | case RDRAND: { | |||
17547 | // Emit the node with the right value type. | |||
17548 | SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other); | |||
17549 | SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0)); | |||
17550 | ||||
17551 | // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1. | |||
17552 | // Otherwise return the value from Rand, which is always 0, casted to i32. | |||
17553 | SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)), | |||
17554 | DAG.getConstant(1, Op->getValueType(1)), | |||
17555 | DAG.getConstant(X86::COND_B, MVT::i32), | |||
17556 | SDValue(Result.getNode(), 1) }; | |||
17557 | SDValue isValid = DAG.getNode(X86ISD::CMOV, dl, | |||
17558 | DAG.getVTList(Op->getValueType(1), MVT::Glue), | |||
17559 | Ops); | |||
17560 | ||||
17561 | // Return { result, isValid, chain }. | |||
17562 | return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid, | |||
17563 | SDValue(Result.getNode(), 2)); | |||
17564 | } | |||
17565 | case GATHER: { | |||
17566 | //gather(v1, mask, index, base, scale); | |||
17567 | SDValue Chain = Op.getOperand(0); | |||
17568 | SDValue Src = Op.getOperand(2); | |||
17569 | SDValue Base = Op.getOperand(3); | |||
17570 | SDValue Index = Op.getOperand(4); | |||
17571 | SDValue Mask = Op.getOperand(5); | |||
17572 | SDValue Scale = Op.getOperand(6); | |||
17573 | return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain, | |||
17574 | Subtarget); | |||
17575 | } | |||
17576 | case SCATTER: { | |||
17577 | //scatter(base, mask, index, v1, scale); | |||
17578 | SDValue Chain = Op.getOperand(0); | |||
17579 | SDValue Base = Op.getOperand(2); | |||
17580 | SDValue Mask = Op.getOperand(3); | |||
17581 | SDValue Index = Op.getOperand(4); | |||
17582 | SDValue Src = Op.getOperand(5); | |||
17583 | SDValue Scale = Op.getOperand(6); | |||
17584 | return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain); | |||
17585 | } | |||
17586 | case PREFETCH: { | |||
17587 | SDValue Hint = Op.getOperand(6); | |||
17588 | unsigned HintVal; | |||
17589 | if (dyn_cast<ConstantSDNode> (Hint) == nullptr || | |||
17590 | (HintVal = dyn_cast<ConstantSDNode> (Hint)->getZExtValue()) > 1) | |||
17591 | llvm_unreachable("Wrong prefetch hint in intrinsic: should be 0 or 1")::llvm::llvm_unreachable_internal("Wrong prefetch hint in intrinsic: should be 0 or 1" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17591); | |||
17592 | unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0); | |||
17593 | SDValue Chain = Op.getOperand(0); | |||
17594 | SDValue Mask = Op.getOperand(2); | |||
17595 | SDValue Index = Op.getOperand(3); | |||
17596 | SDValue Base = Op.getOperand(4); | |||
17597 | SDValue Scale = Op.getOperand(5); | |||
17598 | return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain); | |||
17599 | } | |||
17600 | // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP). | |||
17601 | case RDTSC: { | |||
17602 | SmallVector<SDValue, 2> Results; | |||
17603 | getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget, Results); | |||
17604 | return DAG.getMergeValues(Results, dl); | |||
17605 | } | |||
17606 | // Read Performance Monitoring Counters. | |||
17607 | case RDPMC: { | |||
17608 | SmallVector<SDValue, 2> Results; | |||
17609 | getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results); | |||
17610 | return DAG.getMergeValues(Results, dl); | |||
17611 | } | |||
17612 | // XTEST intrinsics. | |||
17613 | case XTEST: { | |||
17614 | SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other); | |||
17615 | SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0)); | |||
17616 | SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, | |||
17617 | DAG.getConstant(X86::COND_NE, MVT::i8), | |||
17618 | InTrans); | |||
17619 | SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC); | |||
17620 | return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), | |||
17621 | Ret, SDValue(InTrans.getNode(), 1)); | |||
17622 | } | |||
17623 | // ADC/ADCX/SBB | |||
17624 | case ADX: { | |||
17625 | SmallVector<SDValue, 2> Results; | |||
17626 | SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other); | |||
17627 | SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other); | |||
17628 | SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2), | |||
17629 | DAG.getConstant(-1, MVT::i8)); | |||
17630 | SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3), | |||
17631 | Op.getOperand(4), GenCF.getValue(1)); | |||
17632 | SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0), | |||
17633 | Op.getOperand(5), MachinePointerInfo(), | |||
17634 | false, false, 0); | |||
17635 | SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, | |||
17636 | DAG.getConstant(X86::COND_B, MVT::i8), | |||
17637 | Res.getValue(1)); | |||
17638 | Results.push_back(SetCC); | |||
17639 | Results.push_back(Store); | |||
17640 | return DAG.getMergeValues(Results, dl); | |||
17641 | } | |||
17642 | case COMPRESS_TO_MEM: { | |||
17643 | SDLoc dl(Op); | |||
17644 | SDValue Mask = Op.getOperand(4); | |||
17645 | SDValue DataToCompress = Op.getOperand(3); | |||
17646 | SDValue Addr = Op.getOperand(2); | |||
17647 | SDValue Chain = Op.getOperand(0); | |||
17648 | ||||
17649 | if (isAllOnes(Mask)) // return just a store | |||
17650 | return DAG.getStore(Chain, dl, DataToCompress, Addr, | |||
17651 | MachinePointerInfo(), false, false, 0); | |||
17652 | ||||
17653 | EVT VT = DataToCompress.getValueType(); | |||
17654 | EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, | |||
17655 | VT.getVectorNumElements()); | |||
17656 | EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, | |||
17657 | Mask.getValueType().getSizeInBits()); | |||
17658 | SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT, | |||
17659 | DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask), | |||
17660 | DAG.getIntPtrConstant(0)); | |||
17661 | ||||
17662 | SDValue Compressed = DAG.getNode(IntrData->Opc0, dl, VT, VMask, | |||
17663 | DataToCompress, DAG.getUNDEF(VT)); | |||
17664 | return DAG.getStore(Chain, dl, Compressed, Addr, | |||
17665 | MachinePointerInfo(), false, false, 0); | |||
17666 | } | |||
17667 | case EXPAND_FROM_MEM: { | |||
17668 | SDLoc dl(Op); | |||
17669 | SDValue Mask = Op.getOperand(4); | |||
17670 | SDValue PathThru = Op.getOperand(3); | |||
17671 | SDValue Addr = Op.getOperand(2); | |||
17672 | SDValue Chain = Op.getOperand(0); | |||
17673 | EVT VT = Op.getValueType(); | |||
17674 | ||||
17675 | if (isAllOnes(Mask)) // return just a load | |||
17676 | return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false, | |||
17677 | false, 0); | |||
17678 | EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, | |||
17679 | VT.getVectorNumElements()); | |||
17680 | EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, | |||
17681 | Mask.getValueType().getSizeInBits()); | |||
17682 | SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT, | |||
17683 | DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask), | |||
17684 | DAG.getIntPtrConstant(0)); | |||
17685 | ||||
17686 | SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), | |||
17687 | false, false, false, 0); | |||
17688 | ||||
17689 | SmallVector<SDValue, 2> Results; | |||
17690 | Results.push_back(DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToExpand, | |||
17691 | PathThru)); | |||
17692 | Results.push_back(Chain); | |||
17693 | return DAG.getMergeValues(Results, dl); | |||
17694 | } | |||
17695 | } | |||
17696 | } | |||
17697 | ||||
17698 | SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, | |||
17699 | SelectionDAG &DAG) const { | |||
17700 | MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); | |||
17701 | MFI->setReturnAddressIsTaken(true); | |||
17702 | ||||
17703 | if (verifyReturnAddressArgumentIsConstant(Op, DAG)) | |||
17704 | return SDValue(); | |||
17705 | ||||
17706 | unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); | |||
17707 | SDLoc dl(Op); | |||
17708 | EVT PtrVT = getPointerTy(); | |||
17709 | ||||
17710 | if (Depth > 0) { | |||
17711 | SDValue FrameAddr = LowerFRAMEADDR(Op, DAG); | |||
17712 | const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>( | |||
17713 | DAG.getSubtarget().getRegisterInfo()); | |||
17714 | SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), PtrVT); | |||
17715 | return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), | |||
17716 | DAG.getNode(ISD::ADD, dl, PtrVT, | |||
17717 | FrameAddr, Offset), | |||
17718 | MachinePointerInfo(), false, false, false, 0); | |||
17719 | } | |||
17720 | ||||
17721 | // Just load the return address. | |||
17722 | SDValue RetAddrFI = getReturnAddressFrameIndex(DAG); | |||
17723 | return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), | |||
17724 | RetAddrFI, MachinePointerInfo(), false, false, false, 0); | |||
17725 | } | |||
17726 | ||||
17727 | SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { | |||
17728 | MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); | |||
17729 | MFI->setFrameAddressIsTaken(true); | |||
17730 | ||||
17731 | EVT VT = Op.getValueType(); | |||
17732 | SDLoc dl(Op); // FIXME probably not meaningful | |||
17733 | unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); | |||
17734 | const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>( | |||
17735 | DAG.getSubtarget().getRegisterInfo()); | |||
17736 | unsigned FrameReg = RegInfo->getPtrSizedFrameRegister( | |||
17737 | DAG.getMachineFunction()); | |||
17738 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17740, __PRETTY_FUNCTION__)) | |||
17739 | (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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17740, __PRETTY_FUNCTION__)) | |||
17740 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17740, __PRETTY_FUNCTION__)); | |||
17741 | SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT); | |||
17742 | while (Depth--) | |||
17743 | FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, | |||
17744 | MachinePointerInfo(), | |||
17745 | false, false, false, 0); | |||
17746 | return FrameAddr; | |||
17747 | } | |||
17748 | ||||
17749 | // FIXME? Maybe this could be a TableGen attribute on some registers and | |||
17750 | // this table could be generated automatically from RegInfo. | |||
17751 | unsigned X86TargetLowering::getRegisterByName(const char* RegName, | |||
17752 | EVT VT) const { | |||
17753 | unsigned Reg = StringSwitch<unsigned>(RegName) | |||
17754 | .Case("esp", X86::ESP) | |||
17755 | .Case("rsp", X86::RSP) | |||
17756 | .Default(0); | |||
17757 | if (Reg) | |||
17758 | return Reg; | |||
17759 | report_fatal_error("Invalid register name global variable"); | |||
17760 | } | |||
17761 | ||||
17762 | SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op, | |||
17763 | SelectionDAG &DAG) const { | |||
17764 | const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>( | |||
17765 | DAG.getSubtarget().getRegisterInfo()); | |||
17766 | return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize()); | |||
17767 | } | |||
17768 | ||||
17769 | SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const { | |||
17770 | SDValue Chain = Op.getOperand(0); | |||
17771 | SDValue Offset = Op.getOperand(1); | |||
17772 | SDValue Handler = Op.getOperand(2); | |||
17773 | SDLoc dl (Op); | |||
17774 | ||||
17775 | EVT PtrVT = getPointerTy(); | |||
17776 | const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>( | |||
17777 | DAG.getSubtarget().getRegisterInfo()); | |||
17778 | unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction()); | |||
17779 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17781, __PRETTY_FUNCTION__)) | |||
17780 | (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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17781, __PRETTY_FUNCTION__)) | |||
17781 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17781, __PRETTY_FUNCTION__)); | |||
17782 | SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT); | |||
17783 | unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX; | |||
17784 | ||||
17785 | SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame, | |||
17786 | DAG.getIntPtrConstant(RegInfo->getSlotSize())); | |||
17787 | StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset); | |||
17788 | Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(), | |||
17789 | false, false, 0); | |||
17790 | Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr); | |||
17791 | ||||
17792 | return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain, | |||
17793 | DAG.getRegister(StoreAddrReg, PtrVT)); | |||
17794 | } | |||
17795 | ||||
17796 | SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op, | |||
17797 | SelectionDAG &DAG) const { | |||
17798 | SDLoc DL(Op); | |||
17799 | return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL, | |||
17800 | DAG.getVTList(MVT::i32, MVT::Other), | |||
17801 | Op.getOperand(0), Op.getOperand(1)); | |||
17802 | } | |||
17803 | ||||
17804 | SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op, | |||
17805 | SelectionDAG &DAG) const { | |||
17806 | SDLoc DL(Op); | |||
17807 | return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other, | |||
17808 | Op.getOperand(0), Op.getOperand(1)); | |||
17809 | } | |||
17810 | ||||
17811 | static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) { | |||
17812 | return Op.getOperand(0); | |||
17813 | } | |||
17814 | ||||
17815 | SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op, | |||
17816 | SelectionDAG &DAG) const { | |||
17817 | SDValue Root = Op.getOperand(0); | |||
17818 | SDValue Trmp = Op.getOperand(1); // trampoline | |||
17819 | SDValue FPtr = Op.getOperand(2); // nested function | |||
17820 | SDValue Nest = Op.getOperand(3); // 'nest' parameter value | |||
17821 | SDLoc dl (Op); | |||
17822 | ||||
17823 | const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue(); | |||
17824 | const TargetRegisterInfo *TRI = DAG.getSubtarget().getRegisterInfo(); | |||
17825 | ||||
17826 | if (Subtarget->is64Bit()) { | |||
17827 | SDValue OutChains[6]; | |||
17828 | ||||
17829 | // Large code-model. | |||
17830 | const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode. | |||
17831 | const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode. | |||
17832 | ||||
17833 | const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7; | |||
17834 | const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7; | |||
17835 | ||||
17836 | const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix | |||
17837 | ||||
17838 | // Load the pointer to the nested function into R11. | |||
17839 | unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11 | |||
17840 | SDValue Addr = Trmp; | |||
17841 | OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16), | |||
17842 | Addr, MachinePointerInfo(TrmpAddr), | |||
17843 | false, false, 0); | |||
17844 | ||||
17845 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp, | |||
17846 | DAG.getConstant(2, MVT::i64)); | |||
17847 | OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, | |||
17848 | MachinePointerInfo(TrmpAddr, 2), | |||
17849 | false, false, 2); | |||
17850 | ||||
17851 | // Load the 'nest' parameter value into R10. | |||
17852 | // R10 is specified in X86CallingConv.td | |||
17853 | OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10 | |||
17854 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp, | |||
17855 | DAG.getConstant(10, MVT::i64)); | |||
17856 | OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16), | |||
17857 | Addr, MachinePointerInfo(TrmpAddr, 10), | |||
17858 | false, false, 0); | |||
17859 | ||||
17860 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp, | |||
17861 | DAG.getConstant(12, MVT::i64)); | |||
17862 | OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, | |||
17863 | MachinePointerInfo(TrmpAddr, 12), | |||
17864 | false, false, 2); | |||
17865 | ||||
17866 | // Jump to the nested function. | |||
17867 | OpCode = (JMP64r << 8) | REX_WB; // jmpq *... | |||
17868 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp, | |||
17869 | DAG.getConstant(20, MVT::i64)); | |||
17870 | OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16), | |||
17871 | Addr, MachinePointerInfo(TrmpAddr, 20), | |||
17872 | false, false, 0); | |||
17873 | ||||
17874 | unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11 | |||
17875 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp, | |||
17876 | DAG.getConstant(22, MVT::i64)); | |||
17877 | OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr, | |||
17878 | MachinePointerInfo(TrmpAddr, 22), | |||
17879 | false, false, 0); | |||
17880 | ||||
17881 | return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains); | |||
17882 | } else { | |||
17883 | const Function *Func = | |||
17884 | cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue()); | |||
17885 | CallingConv::ID CC = Func->getCallingConv(); | |||
17886 | unsigned NestReg; | |||
17887 | ||||
17888 | switch (CC) { | |||
17889 | default: | |||
17890 | llvm_unreachable("Unsupported calling convention")::llvm::llvm_unreachable_internal("Unsupported calling convention" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 17890); | |||
17891 | case CallingConv::C: | |||
17892 | case CallingConv::X86_StdCall: { | |||
17893 | // Pass 'nest' parameter in ECX. | |||
17894 | // Must be kept in sync with X86CallingConv.td | |||
17895 | NestReg = X86::ECX; | |||
17896 | ||||
17897 | // Check that ECX wasn't needed by an 'inreg' parameter. | |||
17898 | FunctionType *FTy = Func->getFunctionType(); | |||
17899 | const AttributeSet &Attrs = Func->getAttributes(); | |||
17900 | ||||
17901 | if (!Attrs.isEmpty() && !Func->isVarArg()) { | |||
17902 | unsigned InRegCount = 0; | |||
17903 | unsigned Idx = 1; | |||
17904 | ||||
17905 | for (FunctionType::param_iterator I = FTy->param_begin(), | |||
17906 | E = FTy->param_end(); I != E; ++I, ++Idx) | |||
17907 | if (Attrs.hasAttribute(Idx, Attribute::InReg)) | |||
17908 | // FIXME: should only count parameters that are lowered to integers. | |||
17909 | InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32; | |||
17910 | ||||
17911 | if (InRegCount > 2) { | |||
17912 | report_fatal_error("Nest register in use - reduce number of inreg" | |||
17913 | " parameters!"); | |||
17914 | } | |||
17915 | } | |||
17916 | break; | |||
17917 | } | |||
17918 | case CallingConv::X86_FastCall: | |||
17919 | case CallingConv::X86_ThisCall: | |||
17920 | case CallingConv::Fast: | |||
17921 | // Pass 'nest' parameter in EAX. | |||
17922 | // Must be kept in sync with X86CallingConv.td | |||
17923 | NestReg = X86::EAX; | |||
17924 | break; | |||
17925 | } | |||
17926 | ||||
17927 | SDValue OutChains[4]; | |||
17928 | SDValue Addr, Disp; | |||
17929 | ||||
17930 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, | |||
17931 | DAG.getConstant(10, MVT::i32)); | |||
17932 | Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr); | |||
17933 | ||||
17934 | // This is storing the opcode for MOV32ri. | |||
17935 | const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte. | |||
17936 | const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7; | |||
17937 | OutChains[0] = DAG.getStore(Root, dl, | |||
17938 | DAG.getConstant(MOV32ri|N86Reg, MVT::i8), | |||
17939 | Trmp, MachinePointerInfo(TrmpAddr), | |||
17940 | false, false, 0); | |||
17941 | ||||
17942 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, | |||
17943 | DAG.getConstant(1, MVT::i32)); | |||
17944 | OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, | |||
17945 | MachinePointerInfo(TrmpAddr, 1), | |||
17946 | false, false, 1); | |||
17947 | ||||
17948 | const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode. | |||
17949 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, | |||
17950 | DAG.getConstant(5, MVT::i32)); | |||
17951 | OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr, | |||
17952 | MachinePointerInfo(TrmpAddr, 5), | |||
17953 | false, false, 1); | |||
17954 | ||||
17955 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, | |||
17956 | DAG.getConstant(6, MVT::i32)); | |||
17957 | OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, | |||
17958 | MachinePointerInfo(TrmpAddr, 6), | |||
17959 | false, false, 1); | |||
17960 | ||||
17961 | return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains); | |||
17962 | } | |||
17963 | } | |||
17964 | ||||
17965 | SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, | |||
17966 | SelectionDAG &DAG) const { | |||
17967 | /* | |||
17968 | The rounding mode is in bits 11:10 of FPSR, and has the following | |||
17969 | settings: | |||
17970 | 00 Round to nearest | |||
17971 | 01 Round to -inf | |||
17972 | 10 Round to +inf | |||
17973 | 11 Round to 0 | |||
17974 | ||||
17975 | FLT_ROUNDS, on the other hand, expects the following: | |||
17976 | -1 Undefined | |||
17977 | 0 Round to 0 | |||
17978 | 1 Round to nearest | |||
17979 | 2 Round to +inf | |||
17980 | 3 Round to -inf | |||
17981 | ||||
17982 | To perform the conversion, we do: | |||
17983 | (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3) | |||
17984 | */ | |||
17985 | ||||
17986 | MachineFunction &MF = DAG.getMachineFunction(); | |||
17987 | const TargetMachine &TM = MF.getTarget(); | |||
17988 | const TargetFrameLowering &TFI = *TM.getSubtargetImpl()->getFrameLowering(); | |||
17989 | unsigned StackAlignment = TFI.getStackAlignment(); | |||
17990 | MVT VT = Op.getSimpleValueType(); | |||
17991 | SDLoc DL(Op); | |||
17992 | ||||
17993 | // Save FP Control Word to stack slot | |||
17994 | int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false); | |||
17995 | SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy()); | |||
17996 | ||||
17997 | MachineMemOperand *MMO = | |||
17998 | MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI), | |||
17999 | MachineMemOperand::MOStore, 2, 2); | |||
18000 | ||||
18001 | SDValue Ops[] = { DAG.getEntryNode(), StackSlot }; | |||
18002 | SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL, | |||
18003 | DAG.getVTList(MVT::Other), | |||
18004 | Ops, MVT::i16, MMO); | |||
18005 | ||||
18006 | // Load FP Control Word from stack slot | |||
18007 | SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot, | |||
18008 | MachinePointerInfo(), false, false, false, 0); | |||
18009 | ||||
18010 | // Transform as necessary | |||
18011 | SDValue CWD1 = | |||
18012 | DAG.getNode(ISD::SRL, DL, MVT::i16, | |||
18013 | DAG.getNode(ISD::AND, DL, MVT::i16, | |||
18014 | CWD, DAG.getConstant(0x800, MVT::i16)), | |||
18015 | DAG.getConstant(11, MVT::i8)); | |||
18016 | SDValue CWD2 = | |||
18017 | DAG.getNode(ISD::SRL, DL, MVT::i16, | |||
18018 | DAG.getNode(ISD::AND, DL, MVT::i16, | |||
18019 | CWD, DAG.getConstant(0x400, MVT::i16)), | |||
18020 | DAG.getConstant(9, MVT::i8)); | |||
18021 | ||||
18022 | SDValue RetVal = | |||
18023 | DAG.getNode(ISD::AND, DL, MVT::i16, | |||
18024 | DAG.getNode(ISD::ADD, DL, MVT::i16, | |||
18025 | DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2), | |||
18026 | DAG.getConstant(1, MVT::i16)), | |||
18027 | DAG.getConstant(3, MVT::i16)); | |||
18028 | ||||
18029 | return DAG.getNode((VT.getSizeInBits() < 16 ? | |||
18030 | ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal); | |||
18031 | } | |||
18032 | ||||
18033 | static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) { | |||
18034 | MVT VT = Op.getSimpleValueType(); | |||
18035 | EVT OpVT = VT; | |||
18036 | unsigned NumBits = VT.getSizeInBits(); | |||
18037 | SDLoc dl(Op); | |||
18038 | ||||
18039 | Op = Op.getOperand(0); | |||
18040 | if (VT == MVT::i8) { | |||
18041 | // Zero extend to i32 since there is not an i8 bsr. | |||
18042 | OpVT = MVT::i32; | |||
18043 | Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op); | |||
18044 | } | |||
18045 | ||||
18046 | // Issue a bsr (scan bits in reverse) which also sets EFLAGS. | |||
18047 | SDVTList VTs = DAG.getVTList(OpVT, MVT::i32); | |||
18048 | Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op); | |||
18049 | ||||
18050 | // If src is zero (i.e. bsr sets ZF), returns NumBits. | |||
18051 | SDValue Ops[] = { | |||
18052 | Op, | |||
18053 | DAG.getConstant(NumBits+NumBits-1, OpVT), | |||
18054 | DAG.getConstant(X86::COND_E, MVT::i8), | |||
18055 | Op.getValue(1) | |||
18056 | }; | |||
18057 | Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops); | |||
18058 | ||||
18059 | // Finally xor with NumBits-1. | |||
18060 | Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT)); | |||
18061 | ||||
18062 | if (VT == MVT::i8) | |||
18063 | Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op); | |||
18064 | return Op; | |||
18065 | } | |||
18066 | ||||
18067 | static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) { | |||
18068 | MVT VT = Op.getSimpleValueType(); | |||
18069 | EVT OpVT = VT; | |||
18070 | unsigned NumBits = VT.getSizeInBits(); | |||
18071 | SDLoc dl(Op); | |||
18072 | ||||
18073 | Op = Op.getOperand(0); | |||
18074 | if (VT == MVT::i8) { | |||
18075 | // Zero extend to i32 since there is not an i8 bsr. | |||
18076 | OpVT = MVT::i32; | |||
18077 | Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op); | |||
18078 | } | |||
18079 | ||||
18080 | // Issue a bsr (scan bits in reverse). | |||
18081 | SDVTList VTs = DAG.getVTList(OpVT, MVT::i32); | |||
18082 | Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op); | |||
18083 | ||||
18084 | // And xor with NumBits-1. | |||
18085 | Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT)); | |||
18086 | ||||
18087 | if (VT == MVT::i8) | |||
18088 | Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op); | |||
18089 | return Op; | |||
18090 | } | |||
18091 | ||||
18092 | static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) { | |||
18093 | MVT VT = Op.getSimpleValueType(); | |||
18094 | unsigned NumBits = VT.getSizeInBits(); | |||
18095 | SDLoc dl(Op); | |||
18096 | Op = Op.getOperand(0); | |||
18097 | ||||
18098 | // Issue a bsf (scan bits forward) which also sets EFLAGS. | |||
18099 | SDVTList VTs = DAG.getVTList(VT, MVT::i32); | |||
18100 | Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op); | |||
18101 | ||||
18102 | // If src is zero (i.e. bsf sets ZF), returns NumBits. | |||
18103 | SDValue Ops[] = { | |||
18104 | Op, | |||
18105 | DAG.getConstant(NumBits, VT), | |||
18106 | DAG.getConstant(X86::COND_E, MVT::i8), | |||
18107 | Op.getValue(1) | |||
18108 | }; | |||
18109 | return DAG.getNode(X86ISD::CMOV, dl, VT, Ops); | |||
18110 | } | |||
18111 | ||||
18112 | // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit | |||
18113 | // ones, and then concatenate the result back. | |||
18114 | static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) { | |||
18115 | MVT VT = Op.getSimpleValueType(); | |||
18116 | ||||
18117 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18118, __PRETTY_FUNCTION__)) | |||
18118 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18118, __PRETTY_FUNCTION__)); | |||
18119 | ||||
18120 | unsigned NumElems = VT.getVectorNumElements(); | |||
18121 | SDLoc dl(Op); | |||
18122 | ||||
18123 | // Extract the LHS vectors | |||
18124 | SDValue LHS = Op.getOperand(0); | |||
18125 | SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl); | |||
18126 | SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl); | |||
18127 | ||||
18128 | // Extract the RHS vectors | |||
18129 | SDValue RHS = Op.getOperand(1); | |||
18130 | SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl); | |||
18131 | SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl); | |||
18132 | ||||
18133 | MVT EltVT = VT.getVectorElementType(); | |||
18134 | MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2); | |||
18135 | ||||
18136 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, | |||
18137 | DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1), | |||
18138 | DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2)); | |||
18139 | } | |||
18140 | ||||
18141 | static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) { | |||
18142 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18144, __PRETTY_FUNCTION__)) | |||
18143 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18144, __PRETTY_FUNCTION__)) | |||
18144 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18144, __PRETTY_FUNCTION__)); | |||
18145 | return Lower256IntArith(Op, DAG); | |||
18146 | } | |||
18147 | ||||
18148 | static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) { | |||
18149 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18151, __PRETTY_FUNCTION__)) | |||
18150 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18151, __PRETTY_FUNCTION__)) | |||
18151 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18151, __PRETTY_FUNCTION__)); | |||
18152 | return Lower256IntArith(Op, DAG); | |||
18153 | } | |||
18154 | ||||
18155 | static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget, | |||
18156 | SelectionDAG &DAG) { | |||
18157 | SDLoc dl(Op); | |||
18158 | MVT VT = Op.getSimpleValueType(); | |||
18159 | ||||
18160 | // Decompose 256-bit ops into smaller 128-bit ops. | |||
18161 | if (VT.is256BitVector() && !Subtarget->hasInt256()) | |||
18162 | return Lower256IntArith(Op, DAG); | |||
18163 | ||||
18164 | SDValue A = Op.getOperand(0); | |||
18165 | SDValue B = Op.getOperand(1); | |||
18166 | ||||
18167 | // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle. | |||
18168 | if (VT == MVT::v4i32) { | |||
18169 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18170, __PRETTY_FUNCTION__)) | |||
18170 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18170, __PRETTY_FUNCTION__)); | |||
18171 | ||||
18172 | // Extract the odd parts. | |||
18173 | static const int UnpackMask[] = { 1, -1, 3, -1 }; | |||
18174 | SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask); | |||
18175 | SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask); | |||
18176 | ||||
18177 | // Multiply the even parts. | |||
18178 | SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B); | |||
18179 | // Now multiply odd parts. | |||
18180 | SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds); | |||
18181 | ||||
18182 | Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens); | |||
18183 | Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds); | |||
18184 | ||||
18185 | // Merge the two vectors back together with a shuffle. This expands into 2 | |||
18186 | // shuffles. | |||
18187 | static const int ShufMask[] = { 0, 4, 2, 6 }; | |||
18188 | return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask); | |||
18189 | } | |||
18190 | ||||
18191 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18192, __PRETTY_FUNCTION__)) | |||
18192 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18192, __PRETTY_FUNCTION__)); | |||
18193 | ||||
18194 | // Ahi = psrlqi(a, 32); | |||
18195 | // Bhi = psrlqi(b, 32); | |||
18196 | // | |||
18197 | // AloBlo = pmuludq(a, b); | |||
18198 | // AloBhi = pmuludq(a, Bhi); | |||
18199 | // AhiBlo = pmuludq(Ahi, b); | |||
18200 | ||||
18201 | // AloBhi = psllqi(AloBhi, 32); | |||
18202 | // AhiBlo = psllqi(AhiBlo, 32); | |||
18203 | // return AloBlo + AloBhi + AhiBlo; | |||
18204 | ||||
18205 | SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG); | |||
18206 | SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG); | |||
18207 | ||||
18208 | // Bit cast to 32-bit vectors for MULUDQ | |||
18209 | EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 : | |||
18210 | (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32; | |||
18211 | A = DAG.getNode(ISD::BITCAST, dl, MulVT, A); | |||
18212 | B = DAG.getNode(ISD::BITCAST, dl, MulVT, B); | |||
18213 | Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi); | |||
18214 | Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi); | |||
18215 | ||||
18216 | SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B); | |||
18217 | SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi); | |||
18218 | SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B); | |||
18219 | ||||
18220 | AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG); | |||
18221 | AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG); | |||
18222 | ||||
18223 | SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi); | |||
18224 | return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo); | |||
18225 | } | |||
18226 | ||||
18227 | SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const { | |||
18228 | assert(Subtarget->isTargetWin64() && "Unexpected target")((Subtarget->isTargetWin64() && "Unexpected target" ) ? static_cast<void> (0) : __assert_fail ("Subtarget->isTargetWin64() && \"Unexpected target\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18228, __PRETTY_FUNCTION__)); | |||
18229 | EVT VT = Op.getValueType(); | |||
18230 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18231, __PRETTY_FUNCTION__)) | |||
18231 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18231, __PRETTY_FUNCTION__)); | |||
18232 | ||||
18233 | RTLIB::Libcall LC; | |||
18234 | bool isSigned; | |||
18235 | switch (Op->getOpcode()) { | |||
18236 | default: llvm_unreachable("Unexpected request for libcall!")::llvm::llvm_unreachable_internal("Unexpected request for libcall!" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18236); | |||
18237 | case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break; | |||
18238 | case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break; | |||
18239 | case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break; | |||
18240 | case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break; | |||
18241 | case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break; | |||
18242 | case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break; | |||
18243 | } | |||
18244 | ||||
18245 | SDLoc dl(Op); | |||
18246 | SDValue InChain = DAG.getEntryNode(); | |||
18247 | ||||
18248 | TargetLowering::ArgListTy Args; | |||
18249 | TargetLowering::ArgListEntry Entry; | |||
18250 | for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) { | |||
18251 | EVT ArgVT = Op->getOperand(i).getValueType(); | |||
18252 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18253, __PRETTY_FUNCTION__)) | |||
18253 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18253, __PRETTY_FUNCTION__)); | |||
18254 | SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16); | |||
18255 | Entry.Node = StackPtr; | |||
18256 | InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(), | |||
18257 | false, false, 16); | |||
18258 | Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext()); | |||
18259 | Entry.Ty = PointerType::get(ArgTy,0); | |||
18260 | Entry.isSExt = false; | |||
18261 | Entry.isZExt = false; | |||
18262 | Args.push_back(Entry); | |||
18263 | } | |||
18264 | ||||
18265 | SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC), | |||
18266 | getPointerTy()); | |||
18267 | ||||
18268 | TargetLowering::CallLoweringInfo CLI(DAG); | |||
18269 | CLI.setDebugLoc(dl).setChain(InChain) | |||
18270 | .setCallee(getLibcallCallingConv(LC), | |||
18271 | static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()), | |||
18272 | Callee, std::move(Args), 0) | |||
18273 | .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned); | |||
18274 | ||||
18275 | std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI); | |||
18276 | return DAG.getNode(ISD::BITCAST, dl, VT, CallInfo.first); | |||
18277 | } | |||
18278 | ||||
18279 | static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget, | |||
18280 | SelectionDAG &DAG) { | |||
18281 | SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1); | |||
18282 | EVT VT = Op0.getValueType(); | |||
18283 | SDLoc dl(Op); | |||
18284 | ||||
18285 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18286, __PRETTY_FUNCTION__)) | |||
18286 | (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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18286, __PRETTY_FUNCTION__)); | |||
18287 | ||||
18288 | // PMULxD operations multiply each even value (starting at 0) of LHS with | |||
18289 | // the related value of RHS and produce a widen result. | |||
18290 | // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h> | |||
18291 | // => <2 x i64> <ae|cg> | |||
18292 | // | |||
18293 | // In other word, to have all the results, we need to perform two PMULxD: | |||
18294 | // 1. one with the even values. | |||
18295 | // 2. one with the odd values. | |||
18296 | // To achieve #2, with need to place the odd values at an even position. | |||
18297 | // | |||
18298 | // Place the odd value at an even position (basically, shift all values 1 | |||
18299 | // step to the left): | |||
18300 | const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1}; | |||
18301 | // <a|b|c|d> => <b|undef|d|undef> | |||
18302 | SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask); | |||
18303 | // <e|f|g|h> => <f|undef|h|undef> | |||
18304 | SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask); | |||
18305 | ||||
18306 | // Emit two multiplies, one for the lower 2 ints and one for the higher 2 | |||
18307 | // ints. | |||
18308 | MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64; | |||
18309 | bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI; | |||
18310 | unsigned Opcode = | |||
18311 | (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ; | |||
18312 | // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h> | |||
18313 | // => <2 x i64> <ae|cg> | |||
18314 | SDValue Mul1 = DAG.getNode(ISD::BITCAST, dl, VT, | |||
18315 | DAG.getNode(Opcode, dl, MulVT, Op0, Op1)); | |||
18316 | // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef> | |||
18317 | // => <2 x i64> <bf|dh> | |||
18318 | SDValue Mul2 = DAG.getNode(ISD::BITCAST, dl, VT, | |||
18319 | DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1)); | |||
18320 | ||||
18321 | // Shuffle it back into the right order. | |||
18322 | SDValue Highs, Lows; | |||
18323 | if (VT == MVT::v8i32) { | |||
18324 | const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15}; | |||
18325 | Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask); | |||
18326 | const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14}; | |||
18327 | Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask); | |||
18328 | } else { | |||
18329 | const int HighMask[] = {1, 5, 3, 7}; | |||
18330 | Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask); | |||
18331 | const int LowMask[] = {0, 4, 2, 6}; | |||
18332 | Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask); | |||
18333 | } | |||
18334 | ||||
18335 | // If we have a signed multiply but no PMULDQ fix up the high parts of a | |||
18336 | // unsigned multiply. | |||
18337 | if (IsSigned && !Subtarget->hasSSE41()) { | |||
18338 | SDValue ShAmt = | |||
18339 | DAG.getConstant(31, DAG.getTargetLoweringInfo().getShiftAmountTy(VT)); | |||
18340 | SDValue T1 = DAG.getNode(ISD::AND, dl, VT, | |||
18341 | DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1); | |||
18342 | SDValue T2 = DAG.getNode(ISD::AND, dl, VT, | |||
18343 | DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0); | |||
18344 | ||||
18345 | SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2); | |||
18346 | Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup); | |||
18347 | } | |||
18348 | ||||
18349 | // The first result of MUL_LOHI is actually the low value, followed by the | |||
18350 | // high value. | |||
18351 | SDValue Ops[] = {Lows, Highs}; | |||
18352 | return DAG.getMergeValues(Ops, dl); | |||
18353 | } | |||
18354 | ||||
18355 | static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG, | |||
18356 | const X86Subtarget *Subtarget) { | |||
18357 | MVT VT = Op.getSimpleValueType(); | |||
18358 | SDLoc dl(Op); | |||
18359 | SDValue R = Op.getOperand(0); | |||
18360 | SDValue Amt = Op.getOperand(1); | |||
18361 | ||||
18362 | // Optimize shl/srl/sra with constant shift amount. | |||
18363 | if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) { | |||
18364 | if (auto *ShiftConst = BVAmt->getConstantSplatNode()) { | |||
18365 | uint64_t ShiftAmt = ShiftConst->getZExtValue(); | |||
18366 | ||||
18367 | if (VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 || | |||
18368 | (Subtarget->hasInt256() && | |||
18369 | (VT == MVT::v4i64 || VT == MVT::v8i32 || VT == MVT::v16i16)) || | |||
18370 | (Subtarget->hasAVX512() && | |||
18371 | (VT == MVT::v8i64 || VT == MVT::v16i32))) { | |||
18372 | if (Op.getOpcode() == ISD::SHL) | |||
18373 | return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt, | |||
18374 | DAG); | |||
18375 | if (Op.getOpcode() == ISD::SRL) | |||
18376 | return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, | |||
18377 | DAG); | |||
18378 | if (Op.getOpcode() == ISD::SRA && VT != MVT::v2i64 && VT != MVT::v4i64) | |||
18379 | return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt, | |||
18380 | DAG); | |||
18381 | } | |||
18382 | ||||
18383 | if (VT == MVT::v16i8) { | |||
18384 | if (Op.getOpcode() == ISD::SHL) { | |||
18385 | // Make a large shift. | |||
18386 | SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, | |||
18387 | MVT::v8i16, R, ShiftAmt, | |||
18388 | DAG); | |||
18389 | SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL); | |||
18390 | // Zero out the rightmost bits. | |||
18391 | SmallVector<SDValue, 16> V(16, | |||
18392 | DAG.getConstant(uint8_t(-1U << ShiftAmt), | |||
18393 | MVT::i8)); | |||
18394 | return DAG.getNode(ISD::AND, dl, VT, SHL, | |||
18395 | DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V)); | |||
18396 | } | |||
18397 | if (Op.getOpcode() == ISD::SRL) { | |||
18398 | // Make a large shift. | |||
18399 | SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, | |||
18400 | MVT::v8i16, R, ShiftAmt, | |||
18401 | DAG); | |||
18402 | SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL); | |||
18403 | // Zero out the leftmost bits. | |||
18404 | SmallVector<SDValue, 16> V(16, | |||
18405 | DAG.getConstant(uint8_t(-1U) >> ShiftAmt, | |||
18406 | MVT::i8)); | |||
18407 | return DAG.getNode(ISD::AND, dl, VT, SRL, | |||
18408 | DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V)); | |||
18409 | } | |||
18410 | if (Op.getOpcode() == ISD::SRA) { | |||
18411 | if (ShiftAmt == 7) { | |||
18412 | // R s>> 7 === R s< 0 | |||
18413 | SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl); | |||
18414 | return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R); | |||
18415 | } | |||
18416 | ||||
18417 | // R s>> a === ((R u>> a) ^ m) - m | |||
18418 | SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt); | |||
18419 | SmallVector<SDValue, 16> V(16, DAG.getConstant(128 >> ShiftAmt, | |||
18420 | MVT::i8)); | |||
18421 | SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V); | |||
18422 | Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask); | |||
18423 | Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask); | |||
18424 | return Res; | |||
18425 | } | |||
18426 | llvm_unreachable("Unknown shift opcode.")::llvm::llvm_unreachable_internal("Unknown shift opcode.", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18426); | |||
18427 | } | |||
18428 | ||||
18429 | if (Subtarget->hasInt256() && VT == MVT::v32i8) { | |||
18430 | if (Op.getOpcode() == ISD::SHL) { | |||
18431 | // Make a large shift. | |||
18432 | SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, | |||
18433 | MVT::v16i16, R, ShiftAmt, | |||
18434 | DAG); | |||
18435 | SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL); | |||
18436 | // Zero out the rightmost bits. | |||
18437 | SmallVector<SDValue, 32> V(32, | |||
18438 | DAG.getConstant(uint8_t(-1U << ShiftAmt), | |||
18439 | MVT::i8)); | |||
18440 | return DAG.getNode(ISD::AND, dl, VT, SHL, | |||
18441 | DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V)); | |||
18442 | } | |||
18443 | if (Op.getOpcode() == ISD::SRL) { | |||
18444 | // Make a large shift. | |||
18445 | SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, | |||
18446 | MVT::v16i16, R, ShiftAmt, | |||
18447 | DAG); | |||
18448 | SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL); | |||
18449 | // Zero out the leftmost bits. | |||
18450 | SmallVector<SDValue, 32> V(32, | |||
18451 | DAG.getConstant(uint8_t(-1U) >> ShiftAmt, | |||
18452 | MVT::i8)); | |||
18453 | return DAG.getNode(ISD::AND, dl, VT, SRL, | |||
18454 | DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V)); | |||
18455 | } | |||
18456 | if (Op.getOpcode() == ISD::SRA) { | |||
18457 | if (ShiftAmt == 7) { | |||
18458 | // R s>> 7 === R s< 0 | |||
18459 | SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl); | |||
18460 | return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R); | |||
18461 | } | |||
18462 | ||||
18463 | // R s>> a === ((R u>> a) ^ m) - m | |||
18464 | SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt); | |||
18465 | SmallVector<SDValue, 32> V(32, DAG.getConstant(128 >> ShiftAmt, | |||
18466 | MVT::i8)); | |||
18467 | SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V); | |||
18468 | Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask); | |||
18469 | Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask); | |||
18470 | return Res; | |||
18471 | } | |||
18472 | llvm_unreachable("Unknown shift opcode.")::llvm::llvm_unreachable_internal("Unknown shift opcode.", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18472); | |||
18473 | } | |||
18474 | } | |||
18475 | } | |||
18476 | ||||
18477 | // Special case in 32-bit mode, where i64 is expanded into high and low parts. | |||
18478 | if (!Subtarget->is64Bit() && | |||
18479 | (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) && | |||
18480 | Amt.getOpcode() == ISD::BITCAST && | |||
18481 | Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) { | |||
18482 | Amt = Amt.getOperand(0); | |||
18483 | unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() / | |||
18484 | VT.getVectorNumElements(); | |||
18485 | unsigned RatioInLog2 = Log2_32_Ceil(Ratio); | |||
18486 | uint64_t ShiftAmt = 0; | |||
18487 | for (unsigned i = 0; i != Ratio; ++i) { | |||
18488 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i)); | |||
18489 | if (!C) | |||
18490 | return SDValue(); | |||
18491 | // 6 == Log2(64) | |||
18492 | ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2))); | |||
18493 | } | |||
18494 | // Check remaining shift amounts. | |||
18495 | for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) { | |||
18496 | uint64_t ShAmt = 0; | |||
18497 | for (unsigned j = 0; j != Ratio; ++j) { | |||
18498 | ConstantSDNode *C = | |||
18499 | dyn_cast<ConstantSDNode>(Amt.getOperand(i + j)); | |||
18500 | if (!C) | |||
18501 | return SDValue(); | |||
18502 | // 6 == Log2(64) | |||
18503 | ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2))); | |||
18504 | } | |||
18505 | if (ShAmt != ShiftAmt) | |||
18506 | return SDValue(); | |||
18507 | } | |||
18508 | switch (Op.getOpcode()) { | |||
18509 | default: | |||
18510 | llvm_unreachable("Unknown shift opcode!")::llvm::llvm_unreachable_internal("Unknown shift opcode!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18510); | |||
18511 | case ISD::SHL: | |||
18512 | return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt, | |||
18513 | DAG); | |||
18514 | case ISD::SRL: | |||
18515 | return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, | |||
18516 | DAG); | |||
18517 | case ISD::SRA: | |||
18518 | return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt, | |||
18519 | DAG); | |||
18520 | } | |||
18521 | } | |||
18522 | ||||
18523 | return SDValue(); | |||
18524 | } | |||
18525 | ||||
18526 | static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG, | |||
18527 | const X86Subtarget* Subtarget) { | |||
18528 | MVT VT = Op.getSimpleValueType(); | |||
18529 | SDLoc dl(Op); | |||
18530 | SDValue R = Op.getOperand(0); | |||
18531 | SDValue Amt = Op.getOperand(1); | |||
18532 | ||||
18533 | if ((VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) || | |||
18534 | VT == MVT::v4i32 || VT == MVT::v8i16 || | |||
18535 | (Subtarget->hasInt256() && | |||
18536 | ((VT == MVT::v4i64 && Op.getOpcode() != ISD::SRA) || | |||
18537 | VT == MVT::v8i32 || VT == MVT::v16i16)) || | |||
18538 | (Subtarget->hasAVX512() && (VT == MVT::v8i64 || VT == MVT::v16i32))) { | |||
18539 | SDValue BaseShAmt; | |||
18540 | EVT EltVT = VT.getVectorElementType(); | |||
18541 | ||||
18542 | if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) { | |||
18543 | // Check if this build_vector node is doing a splat. | |||
18544 | // If so, then set BaseShAmt equal to the splat value. | |||
18545 | BaseShAmt = BV->getSplatValue(); | |||
18546 | if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF) | |||
18547 | BaseShAmt = SDValue(); | |||
18548 | } else { | |||
18549 | if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR) | |||
18550 | Amt = Amt.getOperand(0); | |||
18551 | ||||
18552 | ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt); | |||
18553 | if (SVN && SVN->isSplat()) { | |||
18554 | unsigned SplatIdx = (unsigned)SVN->getSplatIndex(); | |||
18555 | SDValue InVec = Amt.getOperand(0); | |||
18556 | if (InVec.getOpcode() == ISD::BUILD_VECTOR) { | |||
18557 | assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&(((SplatIdx < InVec.getValueType().getVectorNumElements()) && "Unexpected shuffle index found!") ? static_cast< void> (0) : __assert_fail ("(SplatIdx < InVec.getValueType().getVectorNumElements()) && \"Unexpected shuffle index found!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18558, __PRETTY_FUNCTION__)) | |||
18558 | "Unexpected shuffle index found!")(((SplatIdx < InVec.getValueType().getVectorNumElements()) && "Unexpected shuffle index found!") ? static_cast< void> (0) : __assert_fail ("(SplatIdx < InVec.getValueType().getVectorNumElements()) && \"Unexpected shuffle index found!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18558, __PRETTY_FUNCTION__)); | |||
18559 | BaseShAmt = InVec.getOperand(SplatIdx); | |||
18560 | } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) { | |||
18561 | if (ConstantSDNode *C = | |||
18562 | dyn_cast<ConstantSDNode>(InVec.getOperand(2))) { | |||
18563 | if (C->getZExtValue() == SplatIdx) | |||
18564 | BaseShAmt = InVec.getOperand(1); | |||
18565 | } | |||
18566 | } | |||
18567 | ||||
18568 | if (!BaseShAmt) | |||
18569 | // Avoid introducing an extract element from a shuffle. | |||
18570 | BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec, | |||
18571 | DAG.getIntPtrConstant(SplatIdx)); | |||
18572 | } | |||
18573 | } | |||
18574 | ||||
18575 | if (BaseShAmt.getNode()) { | |||
18576 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18576, __PRETTY_FUNCTION__)); | |||
18577 | if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32)) | |||
18578 | BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt); | |||
18579 | else if (EltVT.bitsLT(MVT::i32)) | |||
18580 | BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt); | |||
18581 | ||||
18582 | switch (Op.getOpcode()) { | |||
18583 | default: | |||
18584 | llvm_unreachable("Unknown shift opcode!")::llvm::llvm_unreachable_internal("Unknown shift opcode!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18584); | |||
18585 | case ISD::SHL: | |||
18586 | switch (VT.SimpleTy) { | |||
18587 | default: return SDValue(); | |||
18588 | case MVT::v2i64: | |||
18589 | case MVT::v4i32: | |||
18590 | case MVT::v8i16: | |||
18591 | case MVT::v4i64: | |||
18592 | case MVT::v8i32: | |||
18593 | case MVT::v16i16: | |||
18594 | case MVT::v16i32: | |||
18595 | case MVT::v8i64: | |||
18596 | return getTargetVShiftNode(X86ISD::VSHLI, dl, VT, R, BaseShAmt, DAG); | |||
18597 | } | |||
18598 | case ISD::SRA: | |||
18599 | switch (VT.SimpleTy) { | |||
18600 | default: return SDValue(); | |||
18601 | case MVT::v4i32: | |||
18602 | case MVT::v8i16: | |||
18603 | case MVT::v8i32: | |||
18604 | case MVT::v16i16: | |||
18605 | case MVT::v16i32: | |||
18606 | case MVT::v8i64: | |||
18607 | return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, R, BaseShAmt, DAG); | |||
18608 | } | |||
18609 | case ISD::SRL: | |||
18610 | switch (VT.SimpleTy) { | |||
18611 | default: return SDValue(); | |||
18612 | case MVT::v2i64: | |||
18613 | case MVT::v4i32: | |||
18614 | case MVT::v8i16: | |||
18615 | case MVT::v4i64: | |||
18616 | case MVT::v8i32: | |||
18617 | case MVT::v16i16: | |||
18618 | case MVT::v16i32: | |||
18619 | case MVT::v8i64: | |||
18620 | return getTargetVShiftNode(X86ISD::VSRLI, dl, VT, R, BaseShAmt, DAG); | |||
18621 | } | |||
18622 | } | |||
18623 | } | |||
18624 | } | |||
18625 | ||||
18626 | // Special case in 32-bit mode, where i64 is expanded into high and low parts. | |||
18627 | if (!Subtarget->is64Bit() && | |||
18628 | (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64) || | |||
18629 | (Subtarget->hasAVX512() && VT == MVT::v8i64)) && | |||
18630 | Amt.getOpcode() == ISD::BITCAST && | |||
18631 | Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) { | |||
18632 | Amt = Amt.getOperand(0); | |||
18633 | unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() / | |||
18634 | VT.getVectorNumElements(); | |||
18635 | std::vector<SDValue> Vals(Ratio); | |||
18636 | for (unsigned i = 0; i != Ratio; ++i) | |||
18637 | Vals[i] = Amt.getOperand(i); | |||
18638 | for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) { | |||
18639 | for (unsigned j = 0; j != Ratio; ++j) | |||
18640 | if (Vals[j] != Amt.getOperand(i + j)) | |||
18641 | return SDValue(); | |||
18642 | } | |||
18643 | switch (Op.getOpcode()) { | |||
18644 | default: | |||
18645 | llvm_unreachable("Unknown shift opcode!")::llvm::llvm_unreachable_internal("Unknown shift opcode!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18645); | |||
18646 | case ISD::SHL: | |||
18647 | return DAG.getNode(X86ISD::VSHL, dl, VT, R, Op.getOperand(1)); | |||
18648 | case ISD::SRL: | |||
18649 | return DAG.getNode(X86ISD::VSRL, dl, VT, R, Op.getOperand(1)); | |||
18650 | case ISD::SRA: | |||
18651 | return DAG.getNode(X86ISD::VSRA, dl, VT, R, Op.getOperand(1)); | |||
18652 | } | |||
18653 | } | |||
18654 | ||||
18655 | return SDValue(); | |||
18656 | } | |||
18657 | ||||
18658 | static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget, | |||
18659 | SelectionDAG &DAG) { | |||
18660 | MVT VT = Op.getSimpleValueType(); | |||
18661 | SDLoc dl(Op); | |||
18662 | SDValue R = Op.getOperand(0); | |||
18663 | SDValue Amt = Op.getOperand(1); | |||
18664 | SDValue V; | |||
18665 | ||||
18666 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18666, __PRETTY_FUNCTION__)); | |||
18667 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18667, __PRETTY_FUNCTION__)); | |||
18668 | ||||
18669 | V = LowerScalarImmediateShift(Op, DAG, Subtarget); | |||
18670 | if (V.getNode()) | |||
18671 | return V; | |||
18672 | ||||
18673 | V = LowerScalarVariableShift(Op, DAG, Subtarget); | |||
18674 | if (V.getNode()) | |||
18675 | return V; | |||
18676 | ||||
18677 | if (Subtarget->hasAVX512() && (VT == MVT::v16i32 || VT == MVT::v8i64)) | |||
18678 | return Op; | |||
18679 | // AVX2 has VPSLLV/VPSRAV/VPSRLV. | |||
18680 | if (Subtarget->hasInt256()) { | |||
18681 | if (Op.getOpcode() == ISD::SRL && | |||
18682 | (VT == MVT::v2i64 || VT == MVT::v4i32 || | |||
18683 | VT == MVT::v4i64 || VT == MVT::v8i32)) | |||
18684 | return Op; | |||
18685 | if (Op.getOpcode() == ISD::SHL && | |||
18686 | (VT == MVT::v2i64 || VT == MVT::v4i32 || | |||
18687 | VT == MVT::v4i64 || VT == MVT::v8i32)) | |||
18688 | return Op; | |||
18689 | if (Op.getOpcode() == ISD::SRA && (VT == MVT::v4i32 || VT == MVT::v8i32)) | |||
18690 | return Op; | |||
18691 | } | |||
18692 | ||||
18693 | // If possible, lower this packed shift into a vector multiply instead of | |||
18694 | // expanding it into a sequence of scalar shifts. | |||
18695 | // Do this only if the vector shift count is a constant build_vector. | |||
18696 | if (Op.getOpcode() == ISD::SHL && | |||
18697 | (VT == MVT::v8i16 || VT == MVT::v4i32 || | |||
18698 | (Subtarget->hasInt256() && VT == MVT::v16i16)) && | |||
18699 | ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) { | |||
18700 | SmallVector<SDValue, 8> Elts; | |||
18701 | EVT SVT = VT.getScalarType(); | |||
18702 | unsigned SVTBits = SVT.getSizeInBits(); | |||
18703 | const APInt &One = APInt(SVTBits, 1); | |||
18704 | unsigned NumElems = VT.getVectorNumElements(); | |||
18705 | ||||
18706 | for (unsigned i=0; i !=NumElems; ++i) { | |||
18707 | SDValue Op = Amt->getOperand(i); | |||
18708 | if (Op->getOpcode() == ISD::UNDEF) { | |||
18709 | Elts.push_back(Op); | |||
18710 | continue; | |||
18711 | } | |||
18712 | ||||
18713 | ConstantSDNode *ND = cast<ConstantSDNode>(Op); | |||
18714 | const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue()); | |||
18715 | uint64_t ShAmt = C.getZExtValue(); | |||
18716 | if (ShAmt >= SVTBits) { | |||
18717 | Elts.push_back(DAG.getUNDEF(SVT)); | |||
18718 | continue; | |||
18719 | } | |||
18720 | Elts.push_back(DAG.getConstant(One.shl(ShAmt), SVT)); | |||
18721 | } | |||
18722 | SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts); | |||
18723 | return DAG.getNode(ISD::MUL, dl, VT, R, BV); | |||
18724 | } | |||
18725 | ||||
18726 | // Lower SHL with variable shift amount. | |||
18727 | if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) { | |||
18728 | Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, VT)); | |||
18729 | ||||
18730 | Op = DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getConstant(0x3f800000U, VT)); | |||
18731 | Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op); | |||
18732 | Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op); | |||
18733 | return DAG.getNode(ISD::MUL, dl, VT, Op, R); | |||
18734 | } | |||
18735 | ||||
18736 | // If possible, lower this shift as a sequence of two shifts by | |||
18737 | // constant plus a MOVSS/MOVSD instead of scalarizing it. | |||
18738 | // Example: | |||
18739 | // (v4i32 (srl A, (build_vector < X, Y, Y, Y>))) | |||
18740 | // | |||
18741 | // Could be rewritten as: | |||
18742 | // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>))) | |||
18743 | // | |||
18744 | // The advantage is that the two shifts from the example would be | |||
18745 | // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing | |||
18746 | // the vector shift into four scalar shifts plus four pairs of vector | |||
18747 | // insert/extract. | |||
18748 | if ((VT == MVT::v8i16 || VT == MVT::v4i32) && | |||
18749 | ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) { | |||
18750 | unsigned TargetOpcode = X86ISD::MOVSS; | |||
18751 | bool CanBeSimplified; | |||
18752 | // The splat value for the first packed shift (the 'X' from the example). | |||
18753 | SDValue Amt1 = Amt->getOperand(0); | |||
18754 | // The splat value for the second packed shift (the 'Y' from the example). | |||
18755 | SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) : | |||
18756 | Amt->getOperand(2); | |||
18757 | ||||
18758 | // See if it is possible to replace this node with a sequence of | |||
18759 | // two shifts followed by a MOVSS/MOVSD | |||
18760 | if (VT == MVT::v4i32) { | |||
18761 | // Check if it is legal to use a MOVSS. | |||
18762 | CanBeSimplified = Amt2 == Amt->getOperand(2) && | |||
18763 | Amt2 == Amt->getOperand(3); | |||
18764 | if (!CanBeSimplified) { | |||
18765 | // Otherwise, check if we can still simplify this node using a MOVSD. | |||
18766 | CanBeSimplified = Amt1 == Amt->getOperand(1) && | |||
18767 | Amt->getOperand(2) == Amt->getOperand(3); | |||
18768 | TargetOpcode = X86ISD::MOVSD; | |||
18769 | Amt2 = Amt->getOperand(2); | |||
18770 | } | |||
18771 | } else { | |||
18772 | // Do similar checks for the case where the machine value type | |||
18773 | // is MVT::v8i16. | |||
18774 | CanBeSimplified = Amt1 == Amt->getOperand(1); | |||
18775 | for (unsigned i=3; i != 8 && CanBeSimplified; ++i) | |||
18776 | CanBeSimplified = Amt2 == Amt->getOperand(i); | |||
18777 | ||||
18778 | if (!CanBeSimplified) { | |||
18779 | TargetOpcode = X86ISD::MOVSD; | |||
18780 | CanBeSimplified = true; | |||
18781 | Amt2 = Amt->getOperand(4); | |||
18782 | for (unsigned i=0; i != 4 && CanBeSimplified; ++i) | |||
18783 | CanBeSimplified = Amt1 == Amt->getOperand(i); | |||
18784 | for (unsigned j=4; j != 8 && CanBeSimplified; ++j) | |||
18785 | CanBeSimplified = Amt2 == Amt->getOperand(j); | |||
18786 | } | |||
18787 | } | |||
18788 | ||||
18789 | if (CanBeSimplified && isa<ConstantSDNode>(Amt1) && | |||
18790 | isa<ConstantSDNode>(Amt2)) { | |||
18791 | // Replace this node with two shifts followed by a MOVSS/MOVSD. | |||
18792 | EVT CastVT = MVT::v4i32; | |||
18793 | SDValue Splat1 = | |||
18794 | DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), VT); | |||
18795 | SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1); | |||
18796 | SDValue Splat2 = | |||
18797 | DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), VT); | |||
18798 | SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2); | |||
18799 | if (TargetOpcode == X86ISD::MOVSD) | |||
18800 | CastVT = MVT::v2i64; | |||
18801 | SDValue BitCast1 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift1); | |||
18802 | SDValue BitCast2 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift2); | |||
18803 | SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2, | |||
18804 | BitCast1, DAG); | |||
18805 | return DAG.getNode(ISD::BITCAST, dl, VT, Result); | |||
18806 | } | |||
18807 | } | |||
18808 | ||||
18809 | if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) { | |||
18810 | assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.")((Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq." ) ? static_cast<void> (0) : __assert_fail ("Subtarget->hasSSE2() && \"Need SSE2 for pslli/pcmpeq.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18810, __PRETTY_FUNCTION__)); | |||
18811 | ||||
18812 | // a = a << 5; | |||
18813 | Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, VT)); | |||
18814 | Op = DAG.getNode(ISD::BITCAST, dl, VT, Op); | |||
18815 | ||||
18816 | // Turn 'a' into a mask suitable for VSELECT | |||
18817 | SDValue VSelM = DAG.getConstant(0x80, VT); | |||
18818 | SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op); | |||
18819 | OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM); | |||
18820 | ||||
18821 | SDValue CM1 = DAG.getConstant(0x0f, VT); | |||
18822 | SDValue CM2 = DAG.getConstant(0x3f, VT); | |||
18823 | ||||
18824 | // r = VSELECT(r, psllw(r & (char16)15, 4), a); | |||
18825 | SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1); | |||
18826 | M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 4, DAG); | |||
18827 | M = DAG.getNode(ISD::BITCAST, dl, VT, M); | |||
18828 | R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R); | |||
18829 | ||||
18830 | // a += a | |||
18831 | Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op); | |||
18832 | OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op); | |||
18833 | OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM); | |||
18834 | ||||
18835 | // r = VSELECT(r, psllw(r & (char16)63, 2), a); | |||
18836 | M = DAG.getNode(ISD::AND, dl, VT, R, CM2); | |||
18837 | M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 2, DAG); | |||
18838 | M = DAG.getNode(ISD::BITCAST, dl, VT, M); | |||
18839 | R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R); | |||
18840 | ||||
18841 | // a += a | |||
18842 | Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op); | |||
18843 | OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op); | |||
18844 | OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM); | |||
18845 | ||||
18846 | // return VSELECT(r, r+r, a); | |||
18847 | R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, | |||
18848 | DAG.getNode(ISD::ADD, dl, VT, R, R), R); | |||
18849 | return R; | |||
18850 | } | |||
18851 | ||||
18852 | // It's worth extending once and using the v8i32 shifts for 16-bit types, but | |||
18853 | // the extra overheads to get from v16i8 to v8i32 make the existing SSE | |||
18854 | // solution better. | |||
18855 | if (Subtarget->hasInt256() && VT == MVT::v8i16) { | |||
18856 | MVT NewVT = VT == MVT::v8i16 ? MVT::v8i32 : MVT::v16i16; | |||
18857 | unsigned ExtOpc = | |||
18858 | Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; | |||
18859 | R = DAG.getNode(ExtOpc, dl, NewVT, R); | |||
18860 | Amt = DAG.getNode(ISD::ANY_EXTEND, dl, NewVT, Amt); | |||
18861 | return DAG.getNode(ISD::TRUNCATE, dl, VT, | |||
18862 | DAG.getNode(Op.getOpcode(), dl, NewVT, R, Amt)); | |||
18863 | } | |||
18864 | ||||
18865 | // Decompose 256-bit shifts into smaller 128-bit shifts. | |||
18866 | if (VT.is256BitVector()) { | |||
18867 | unsigned NumElems = VT.getVectorNumElements(); | |||
18868 | MVT EltVT = VT.getVectorElementType(); | |||
18869 | EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2); | |||
18870 | ||||
18871 | // Extract the two vectors | |||
18872 | SDValue V1 = Extract128BitVector(R, 0, DAG, dl); | |||
18873 | SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl); | |||
18874 | ||||
18875 | // Recreate the shift amount vectors | |||
18876 | SDValue Amt1, Amt2; | |||
18877 | if (Amt.getOpcode() == ISD::BUILD_VECTOR) { | |||
18878 | // Constant shift amount | |||
18879 | SmallVector<SDValue, 4> Amt1Csts; | |||
18880 | SmallVector<SDValue, 4> Amt2Csts; | |||
18881 | for (unsigned i = 0; i != NumElems/2; ++i) | |||
18882 | Amt1Csts.push_back(Amt->getOperand(i)); | |||
18883 | for (unsigned i = NumElems/2; i != NumElems; ++i) | |||
18884 | Amt2Csts.push_back(Amt->getOperand(i)); | |||
18885 | ||||
18886 | Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts); | |||
18887 | Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts); | |||
18888 | } else { | |||
18889 | // Variable shift amount | |||
18890 | Amt1 = Extract128BitVector(Amt, 0, DAG, dl); | |||
18891 | Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl); | |||
18892 | } | |||
18893 | ||||
18894 | // Issue new vector shifts for the smaller types | |||
18895 | V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1); | |||
18896 | V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2); | |||
18897 | ||||
18898 | // Concatenate the result back | |||
18899 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2); | |||
18900 | } | |||
18901 | ||||
18902 | return SDValue(); | |||
18903 | } | |||
18904 | ||||
18905 | static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) { | |||
18906 | // Lower the "add/sub/mul with overflow" instruction into a regular ins plus | |||
18907 | // a "setcc" instruction that checks the overflow flag. The "brcond" lowering | |||
18908 | // looks for this combo and may remove the "setcc" instruction if the "setcc" | |||
18909 | // has only one use. | |||
18910 | SDNode *N = Op.getNode(); | |||
18911 | SDValue LHS = N->getOperand(0); | |||
18912 | SDValue RHS = N->getOperand(1); | |||
18913 | unsigned BaseOp = 0; | |||
18914 | unsigned Cond = 0; | |||
18915 | SDLoc DL(Op); | |||
18916 | switch (Op.getOpcode()) { | |||
18917 | default: llvm_unreachable("Unknown ovf instruction!")::llvm::llvm_unreachable_internal("Unknown ovf instruction!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 18917); | |||
18918 | case ISD::SADDO: | |||
18919 | // A subtract of one will be selected as a INC. Note that INC doesn't | |||
18920 | // set CF, so we can't do this for UADDO. | |||
18921 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS)) | |||
18922 | if (C->isOne()) { | |||
18923 | BaseOp = X86ISD::INC; | |||
18924 | Cond = X86::COND_O; | |||
18925 | break; | |||
18926 | } | |||
18927 | BaseOp = X86ISD::ADD; | |||
18928 | Cond = X86::COND_O; | |||
18929 | break; | |||
18930 | case ISD::UADDO: | |||
18931 | BaseOp = X86ISD::ADD; | |||
18932 | Cond = X86::COND_B; | |||
18933 | break; | |||
18934 | case ISD::SSUBO: | |||
18935 | // A subtract of one will be selected as a DEC. Note that DEC doesn't | |||
18936 | // set CF, so we can't do this for USUBO. | |||
18937 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS)) | |||
18938 | if (C->isOne()) { | |||
18939 | BaseOp = X86ISD::DEC; | |||
18940 | Cond = X86::COND_O; | |||
18941 | break; | |||
18942 | } | |||
18943 | BaseOp = X86ISD::SUB; | |||
18944 | Cond = X86::COND_O; | |||
18945 | break; | |||
18946 | case ISD::USUBO: | |||
18947 | BaseOp = X86ISD::SUB; | |||
18948 | Cond = X86::COND_B; | |||
18949 | break; | |||
18950 | case ISD::SMULO: | |||
18951 | BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL; | |||
18952 | Cond = X86::COND_O; | |||
18953 | break; | |||
18954 | case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs | |||
18955 | if (N->getValueType(0) == MVT::i8) { | |||
18956 | BaseOp = X86ISD::UMUL8; | |||
18957 | Cond = X86::COND_O; | |||
18958 | break; | |||
18959 | } | |||
18960 | SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0), | |||
18961 | MVT::i32); | |||
18962 | SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS); | |||
18963 | ||||
18964 | SDValue SetCC = | |||
18965 | DAG.getNode(X86ISD::SETCC, DL, MVT::i8, | |||
18966 | DAG.getConstant(X86::COND_O, MVT::i32), | |||
18967 | SDValue(Sum.getNode(), 2)); | |||
18968 | ||||
18969 | return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC); | |||
18970 | } | |||
18971 | } | |||
18972 | ||||
18973 | // Also sets EFLAGS. | |||
18974 | SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32); | |||
18975 | SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS); | |||
18976 | ||||
18977 | SDValue SetCC = | |||
18978 | DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1), | |||
18979 | DAG.getConstant(Cond, MVT::i32), | |||
18980 | SDValue(Sum.getNode(), 1)); | |||
18981 | ||||
18982 | return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC); | |||
18983 | } | |||
18984 | ||||
18985 | // Sign extension of the low part of vector elements. This may be used either | |||
18986 | // when sign extend instructions are not available or if the vector element | |||
18987 | // sizes already match the sign-extended size. If the vector elements are in | |||
18988 | // their pre-extended size and sign extend instructions are available, that will | |||
18989 | // be handled by LowerSIGN_EXTEND. | |||
18990 | SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, | |||
18991 | SelectionDAG &DAG) const { | |||
18992 | SDLoc dl(Op); | |||
18993 | EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); | |||
18994 | MVT VT = Op.getSimpleValueType(); | |||
18995 | ||||
18996 | if (!Subtarget->hasSSE2() || !VT.isVector()) | |||
18997 | return SDValue(); | |||
18998 | ||||
18999 | unsigned BitsDiff = VT.getScalarType().getSizeInBits() - | |||
19000 | ExtraVT.getScalarType().getSizeInBits(); | |||
19001 | ||||
19002 | switch (VT.SimpleTy) { | |||
19003 | default: return SDValue(); | |||
19004 | case MVT::v8i32: | |||
19005 | case MVT::v16i16: | |||
19006 | if (!Subtarget->hasFp256()) | |||
19007 | return SDValue(); | |||
19008 | if (!Subtarget->hasInt256()) { | |||
19009 | // needs to be split | |||
19010 | unsigned NumElems = VT.getVectorNumElements(); | |||
19011 | ||||
19012 | // Extract the LHS vectors | |||
19013 | SDValue LHS = Op.getOperand(0); | |||
19014 | SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl); | |||
19015 | SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl); | |||
19016 | ||||
19017 | MVT EltVT = VT.getVectorElementType(); | |||
19018 | EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2); | |||
19019 | ||||
19020 | EVT ExtraEltVT = ExtraVT.getVectorElementType(); | |||
19021 | unsigned ExtraNumElems = ExtraVT.getVectorNumElements(); | |||
19022 | ExtraVT = EVT::getVectorVT(*DAG.getContext(), ExtraEltVT, | |||
19023 | ExtraNumElems/2); | |||
19024 | SDValue Extra = DAG.getValueType(ExtraVT); | |||
19025 | ||||
19026 | LHS1 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, Extra); | |||
19027 | LHS2 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, Extra); | |||
19028 | ||||
19029 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, LHS1, LHS2); | |||
19030 | } | |||
19031 | // fall through | |||
19032 | case MVT::v4i32: | |||
19033 | case MVT::v8i16: { | |||
19034 | SDValue Op0 = Op.getOperand(0); | |||
19035 | ||||
19036 | // This is a sign extension of some low part of vector elements without | |||
19037 | // changing the size of the vector elements themselves: | |||
19038 | // Shift-Left + Shift-Right-Algebraic. | |||
19039 | SDValue Shl = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Op0, | |||
19040 | BitsDiff, DAG); | |||
19041 | return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, Shl, BitsDiff, | |||
19042 | DAG); | |||
19043 | } | |||
19044 | } | |||
19045 | } | |||
19046 | ||||
19047 | /// Returns true if the operand type is exactly twice the native width, and | |||
19048 | /// the corresponding cmpxchg8b or cmpxchg16b instruction is available. | |||
19049 | /// Used to know whether to use cmpxchg8/16b when expanding atomic operations | |||
19050 | /// (otherwise we leave them alone to become __sync_fetch_and_... calls). | |||
19051 | bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const { | |||
19052 | const X86Subtarget &Subtarget = | |||
19053 | getTargetMachine().getSubtarget<X86Subtarget>(); | |||
19054 | unsigned OpWidth = MemType->getPrimitiveSizeInBits(); | |||
19055 | ||||
19056 | if (OpWidth == 64) | |||
19057 | return !Subtarget.is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b | |||
19058 | else if (OpWidth == 128) | |||
19059 | return Subtarget.hasCmpxchg16b(); | |||
19060 | else | |||
19061 | return false; | |||
19062 | } | |||
19063 | ||||
19064 | bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const { | |||
19065 | return needsCmpXchgNb(SI->getValueOperand()->getType()); | |||
19066 | } | |||
19067 | ||||
19068 | // Note: this turns large loads into lock cmpxchg8b/16b. | |||
19069 | // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b. | |||
19070 | bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const { | |||
19071 | auto PTy = cast<PointerType>(LI->getPointerOperand()->getType()); | |||
19072 | return needsCmpXchgNb(PTy->getElementType()); | |||
19073 | } | |||
19074 | ||||
19075 | bool X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const { | |||
19076 | const X86Subtarget &Subtarget = | |||
19077 | getTargetMachine().getSubtarget<X86Subtarget>(); | |||
19078 | unsigned NativeWidth = Subtarget.is64Bit() ? 64 : 32; | |||
19079 | const Type *MemType = AI->getType(); | |||
19080 | ||||
19081 | // If the operand is too big, we must see if cmpxchg8/16b is available | |||
19082 | // and default to library calls otherwise. | |||
19083 | if (MemType->getPrimitiveSizeInBits() > NativeWidth) | |||
19084 | return needsCmpXchgNb(MemType); | |||
19085 | ||||
19086 | AtomicRMWInst::BinOp Op = AI->getOperation(); | |||
19087 | switch (Op) { | |||
19088 | default: | |||
19089 | llvm_unreachable("Unknown atomic operation")::llvm::llvm_unreachable_internal("Unknown atomic operation", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19089); | |||
19090 | case AtomicRMWInst::Xchg: | |||
19091 | case AtomicRMWInst::Add: | |||
19092 | case AtomicRMWInst::Sub: | |||
19093 | // It's better to use xadd, xsub or xchg for these in all cases. | |||
19094 | return false; | |||
19095 | case AtomicRMWInst::Or: | |||
19096 | case AtomicRMWInst::And: | |||
19097 | case AtomicRMWInst::Xor: | |||
19098 | // If the atomicrmw's result isn't actually used, we can just add a "lock" | |||
19099 | // prefix to a normal instruction for these operations. | |||
19100 | return !AI->use_empty(); | |||
19101 | case AtomicRMWInst::Nand: | |||
19102 | case AtomicRMWInst::Max: | |||
19103 | case AtomicRMWInst::Min: | |||
19104 | case AtomicRMWInst::UMax: | |||
19105 | case AtomicRMWInst::UMin: | |||
19106 | // These always require a non-trivial set of data operations on x86. We must | |||
19107 | // use a cmpxchg loop. | |||
19108 | return true; | |||
19109 | } | |||
19110 | } | |||
19111 | ||||
19112 | static bool hasMFENCE(const X86Subtarget& Subtarget) { | |||
19113 | // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for | |||
19114 | // no-sse2). There isn't any reason to disable it if the target processor | |||
19115 | // supports it. | |||
19116 | return Subtarget.hasSSE2() || Subtarget.is64Bit(); | |||
19117 | } | |||
19118 | ||||
19119 | LoadInst * | |||
19120 | X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const { | |||
19121 | const X86Subtarget &Subtarget = | |||
19122 | getTargetMachine().getSubtarget<X86Subtarget>(); | |||
19123 | unsigned NativeWidth = Subtarget.is64Bit() ? 64 : 32; | |||
19124 | const Type *MemType = AI->getType(); | |||
19125 | // Accesses larger than the native width are turned into cmpxchg/libcalls, so | |||
19126 | // there is no benefit in turning such RMWs into loads, and it is actually | |||
19127 | // harmful as it introduces a mfence. | |||
19128 | if (MemType->getPrimitiveSizeInBits() > NativeWidth) | |||
19129 | return nullptr; | |||
19130 | ||||
19131 | auto Builder = IRBuilder<>(AI); | |||
19132 | Module *M = Builder.GetInsertBlock()->getParent()->getParent(); | |||
19133 | auto SynchScope = AI->getSynchScope(); | |||
19134 | // We must restrict the ordering to avoid generating loads with Release or | |||
19135 | // ReleaseAcquire orderings. | |||
19136 | auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering()); | |||
19137 | auto Ptr = AI->getPointerOperand(); | |||
19138 | ||||
19139 | // Before the load we need a fence. Here is an example lifted from | |||
19140 | // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence | |||
19141 | // is required: | |||
19142 | // Thread 0: | |||
19143 | // x.store(1, relaxed); | |||
19144 | // r1 = y.fetch_add(0, release); | |||
19145 | // Thread 1: | |||
19146 | // y.fetch_add(42, acquire); | |||
19147 | // r2 = x.load(relaxed); | |||
19148 | // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is | |||
19149 | // lowered to just a load without a fence. A mfence flushes the store buffer, | |||
19150 | // making the optimization clearly correct. | |||
19151 | // FIXME: it is required if isAtLeastRelease(Order) but it is not clear | |||
19152 | // otherwise, we might be able to be more agressive on relaxed idempotent | |||
19153 | // rmw. In practice, they do not look useful, so we don't try to be | |||
19154 | // especially clever. | |||
19155 | if (SynchScope == SingleThread) { | |||
19156 | // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at | |||
19157 | // the IR level, so we must wrap it in an intrinsic. | |||
19158 | return nullptr; | |||
19159 | } else if (hasMFENCE(Subtarget)) { | |||
19160 | Function *MFence = llvm::Intrinsic::getDeclaration(M, | |||
19161 | Intrinsic::x86_sse2_mfence); | |||
19162 | Builder.CreateCall(MFence); | |||
19163 | } else { | |||
19164 | // FIXME: it might make sense to use a locked operation here but on a | |||
19165 | // different cache-line to prevent cache-line bouncing. In practice it | |||
19166 | // is probably a small win, and x86 processors without mfence are rare | |||
19167 | // enough that we do not bother. | |||
19168 | return nullptr; | |||
19169 | } | |||
19170 | ||||
19171 | // Finally we can emit the atomic load. | |||
19172 | LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr, | |||
19173 | AI->getType()->getPrimitiveSizeInBits()); | |||
19174 | Loaded->setAtomic(Order, SynchScope); | |||
19175 | AI->replaceAllUsesWith(Loaded); | |||
19176 | AI->eraseFromParent(); | |||
19177 | return Loaded; | |||
19178 | } | |||
19179 | ||||
19180 | static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget, | |||
19181 | SelectionDAG &DAG) { | |||
19182 | SDLoc dl(Op); | |||
19183 | AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>( | |||
19184 | cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()); | |||
19185 | SynchronizationScope FenceScope = static_cast<SynchronizationScope>( | |||
19186 | cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue()); | |||
19187 | ||||
19188 | // The only fence that needs an instruction is a sequentially-consistent | |||
19189 | // cross-thread fence. | |||
19190 | if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) { | |||
19191 | if (hasMFENCE(*Subtarget)) | |||
19192 | return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0)); | |||
19193 | ||||
19194 | SDValue Chain = Op.getOperand(0); | |||
19195 | SDValue Zero = DAG.getConstant(0, MVT::i32); | |||
19196 | SDValue Ops[] = { | |||
19197 | DAG.getRegister(X86::ESP, MVT::i32), // Base | |||
19198 | DAG.getTargetConstant(1, MVT::i8), // Scale | |||
19199 | DAG.getRegister(0, MVT::i32), // Index | |||
19200 | DAG.getTargetConstant(0, MVT::i32), // Disp | |||
19201 | DAG.getRegister(0, MVT::i32), // Segment. | |||
19202 | Zero, | |||
19203 | Chain | |||
19204 | }; | |||
19205 | SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops); | |||
19206 | return SDValue(Res, 0); | |||
19207 | } | |||
19208 | ||||
19209 | // MEMBARRIER is a compiler barrier; it codegens to a no-op. | |||
19210 | return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0)); | |||
19211 | } | |||
19212 | ||||
19213 | static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget, | |||
19214 | SelectionDAG &DAG) { | |||
19215 | MVT T = Op.getSimpleValueType(); | |||
19216 | SDLoc DL(Op); | |||
19217 | unsigned Reg = 0; | |||
19218 | unsigned size = 0; | |||
19219 | switch(T.SimpleTy) { | |||
19220 | default: llvm_unreachable("Invalid value type!")::llvm::llvm_unreachable_internal("Invalid value type!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19220); | |||
19221 | case MVT::i8: Reg = X86::AL; size = 1; break; | |||
19222 | case MVT::i16: Reg = X86::AX; size = 2; break; | |||
19223 | case MVT::i32: Reg = X86::EAX; size = 4; break; | |||
19224 | case MVT::i64: | |||
19225 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19225, __PRETTY_FUNCTION__)); | |||
19226 | Reg = X86::RAX; size = 8; | |||
19227 | break; | |||
19228 | } | |||
19229 | SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg, | |||
19230 | Op.getOperand(2), SDValue()); | |||
19231 | SDValue Ops[] = { cpIn.getValue(0), | |||
19232 | Op.getOperand(1), | |||
19233 | Op.getOperand(3), | |||
19234 | DAG.getTargetConstant(size, MVT::i8), | |||
19235 | cpIn.getValue(1) }; | |||
19236 | SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
19237 | MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand(); | |||
19238 | SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys, | |||
19239 | Ops, T, MMO); | |||
19240 | ||||
19241 | SDValue cpOut = | |||
19242 | DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1)); | |||
19243 | SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS, | |||
19244 | MVT::i32, cpOut.getValue(2)); | |||
19245 | SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1), | |||
19246 | DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS); | |||
19247 | ||||
19248 | DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut); | |||
19249 | DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success); | |||
19250 | DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1)); | |||
19251 | return SDValue(); | |||
19252 | } | |||
19253 | ||||
19254 | static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget, | |||
19255 | SelectionDAG &DAG) { | |||
19256 | MVT SrcVT = Op.getOperand(0).getSimpleValueType(); | |||
19257 | MVT DstVT = Op.getSimpleValueType(); | |||
19258 | ||||
19259 | if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) { | |||
19260 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19260, __PRETTY_FUNCTION__)); | |||
19261 | if (DstVT != MVT::f64) | |||
19262 | // This conversion needs to be expanded. | |||
19263 | return SDValue(); | |||
19264 | ||||
19265 | SDValue InVec = Op->getOperand(0); | |||
19266 | SDLoc dl(Op); | |||
19267 | unsigned NumElts = SrcVT.getVectorNumElements(); | |||
19268 | EVT SVT = SrcVT.getVectorElementType(); | |||
19269 | ||||
19270 | // Widen the vector in input in the case of MVT::v2i32. | |||
19271 | // Example: from MVT::v2i32 to MVT::v4i32. | |||
19272 | SmallVector<SDValue, 16> Elts; | |||
19273 | for (unsigned i = 0, e = NumElts; i != e; ++i) | |||
19274 | Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec, | |||
19275 | DAG.getIntPtrConstant(i))); | |||
19276 | ||||
19277 | // Explicitly mark the extra elements as Undef. | |||
19278 | SDValue Undef = DAG.getUNDEF(SVT); | |||
19279 | for (unsigned i = NumElts, e = NumElts * 2; i != e; ++i) | |||
19280 | Elts.push_back(Undef); | |||
19281 | ||||
19282 | EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2); | |||
19283 | SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts); | |||
19284 | SDValue ToV2F64 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, BV); | |||
19285 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64, | |||
19286 | DAG.getIntPtrConstant(0)); | |||
19287 | } | |||
19288 | ||||
19289 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19290, __PRETTY_FUNCTION__)) | |||
19290 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19290, __PRETTY_FUNCTION__)); | |||
19291 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19293, __PRETTY_FUNCTION__)) | |||
19292 | (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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19293, __PRETTY_FUNCTION__)) | |||
19293 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19293, __PRETTY_FUNCTION__)); | |||
19294 | // i64 <=> MMX conversions are Legal. | |||
19295 | if (SrcVT==MVT::i64 && DstVT.isVector()) | |||
19296 | return Op; | |||
19297 | if (DstVT==MVT::i64 && SrcVT.isVector()) | |||
19298 | return Op; | |||
19299 | // MMX <=> MMX conversions are Legal. | |||
19300 | if (SrcVT.isVector() && DstVT.isVector()) | |||
19301 | return Op; | |||
19302 | // All other conversions need to be expanded. | |||
19303 | return SDValue(); | |||
19304 | } | |||
19305 | ||||
19306 | static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget, | |||
19307 | SelectionDAG &DAG) { | |||
19308 | SDNode *Node = Op.getNode(); | |||
19309 | SDLoc dl(Node); | |||
19310 | ||||
19311 | Op = Op.getOperand(0); | |||
19312 | EVT VT = Op.getValueType(); | |||
19313 | assert((VT.is128BitVector() || VT.is256BitVector()) &&(((VT.is128BitVector() || VT.is256BitVector()) && "CTPOP lowering only implemented for 128/256-bit wide vector types" ) ? static_cast<void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector()) && \"CTPOP lowering only implemented for 128/256-bit wide vector types\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19314, __PRETTY_FUNCTION__)) | |||
19314 | "CTPOP lowering only implemented for 128/256-bit wide vector types")(((VT.is128BitVector() || VT.is256BitVector()) && "CTPOP lowering only implemented for 128/256-bit wide vector types" ) ? static_cast<void> (0) : __assert_fail ("(VT.is128BitVector() || VT.is256BitVector()) && \"CTPOP lowering only implemented for 128/256-bit wide vector types\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19314, __PRETTY_FUNCTION__)); | |||
19315 | ||||
19316 | unsigned NumElts = VT.getVectorNumElements(); | |||
19317 | EVT EltVT = VT.getVectorElementType(); | |||
19318 | unsigned Len = EltVT.getSizeInBits(); | |||
19319 | ||||
19320 | // This is the vectorized version of the "best" algorithm from | |||
19321 | // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel | |||
19322 | // with a minor tweak to use a series of adds + shifts instead of vector | |||
19323 | // multiplications. Implemented for the v2i64, v4i64, v4i32, v8i32 types: | |||
19324 | // | |||
19325 | // v2i64, v4i64, v4i32 => Only profitable w/ popcnt disabled | |||
19326 | // v8i32 => Always profitable | |||
19327 | // | |||
19328 | // FIXME: There a couple of possible improvements: | |||
19329 | // | |||
19330 | // 1) Support for i8 and i16 vectors (needs measurements if popcnt enabled). | |||
19331 | // 2) Use strategies from http://wm.ite.pl/articles/sse-popcount.html | |||
19332 | // | |||
19333 | assert(EltVT.isInteger() && (Len == 32 || Len == 64) && Len % 8 == 0 &&((EltVT.isInteger() && (Len == 32 || Len == 64) && Len % 8 == 0 && "CTPOP not implemented for this vector element type." ) ? static_cast<void> (0) : __assert_fail ("EltVT.isInteger() && (Len == 32 || Len == 64) && Len % 8 == 0 && \"CTPOP not implemented for this vector element type.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19334, __PRETTY_FUNCTION__)) | |||
19334 | "CTPOP not implemented for this vector element type.")((EltVT.isInteger() && (Len == 32 || Len == 64) && Len % 8 == 0 && "CTPOP not implemented for this vector element type." ) ? static_cast<void> (0) : __assert_fail ("EltVT.isInteger() && (Len == 32 || Len == 64) && Len % 8 == 0 && \"CTPOP not implemented for this vector element type.\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19334, __PRETTY_FUNCTION__)); | |||
19335 | ||||
19336 | // X86 canonicalize ANDs to vXi64, generate the appropriate bitcasts to avoid | |||
19337 | // extra legalization. | |||
19338 | bool NeedsBitcast = EltVT == MVT::i32; | |||
19339 | MVT BitcastVT = VT.is256BitVector() ? MVT::v4i64 : MVT::v2i64; | |||
19340 | ||||
19341 | SDValue Cst55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), EltVT); | |||
19342 | SDValue Cst33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), EltVT); | |||
19343 | SDValue Cst0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), EltVT); | |||
19344 | ||||
19345 | // v = v - ((v >> 1) & 0x55555555...) | |||
19346 | SmallVector<SDValue, 8> Ones(NumElts, DAG.getConstant(1, EltVT)); | |||
19347 | SDValue OnesV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ones); | |||
19348 | SDValue Srl = DAG.getNode(ISD::SRL, dl, VT, Op, OnesV); | |||
19349 | if (NeedsBitcast) | |||
19350 | Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl); | |||
19351 | ||||
19352 | SmallVector<SDValue, 8> Mask55(NumElts, Cst55); | |||
19353 | SDValue M55 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask55); | |||
19354 | if (NeedsBitcast) | |||
19355 | M55 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M55); | |||
19356 | ||||
19357 | SDValue And = DAG.getNode(ISD::AND, dl, Srl.getValueType(), Srl, M55); | |||
19358 | if (VT != And.getValueType()) | |||
19359 | And = DAG.getNode(ISD::BITCAST, dl, VT, And); | |||
19360 | SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, Op, And); | |||
19361 | ||||
19362 | // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...) | |||
19363 | SmallVector<SDValue, 8> Mask33(NumElts, Cst33); | |||
19364 | SDValue M33 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask33); | |||
19365 | SmallVector<SDValue, 8> Twos(NumElts, DAG.getConstant(2, EltVT)); | |||
19366 | SDValue TwosV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Twos); | |||
19367 | ||||
19368 | Srl = DAG.getNode(ISD::SRL, dl, VT, Sub, TwosV); | |||
19369 | if (NeedsBitcast) { | |||
19370 | Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl); | |||
19371 | M33 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M33); | |||
19372 | Sub = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Sub); | |||
19373 | } | |||
19374 | ||||
19375 | SDValue AndRHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Srl, M33); | |||
19376 | SDValue AndLHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Sub, M33); | |||
19377 | if (VT != AndRHS.getValueType()) { | |||
19378 | AndRHS = DAG.getNode(ISD::BITCAST, dl, VT, AndRHS); | |||
19379 | AndLHS = DAG.getNode(ISD::BITCAST, dl, VT, AndLHS); | |||
19380 | } | |||
19381 | SDValue Add = DAG.getNode(ISD::ADD, dl, VT, AndLHS, AndRHS); | |||
19382 | ||||
19383 | // v = (v + (v >> 4)) & 0x0F0F0F0F... | |||
19384 | SmallVector<SDValue, 8> Fours(NumElts, DAG.getConstant(4, EltVT)); | |||
19385 | SDValue FoursV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Fours); | |||
19386 | Srl = DAG.getNode(ISD::SRL, dl, VT, Add, FoursV); | |||
19387 | Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl); | |||
19388 | ||||
19389 | SmallVector<SDValue, 8> Mask0F(NumElts, Cst0F); | |||
19390 | SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask0F); | |||
19391 | if (NeedsBitcast) { | |||
19392 | Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add); | |||
19393 | M0F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M0F); | |||
19394 | } | |||
19395 | And = DAG.getNode(ISD::AND, dl, M0F.getValueType(), Add, M0F); | |||
19396 | if (VT != And.getValueType()) | |||
19397 | And = DAG.getNode(ISD::BITCAST, dl, VT, And); | |||
19398 | ||||
19399 | // The algorithm mentioned above uses: | |||
19400 | // v = (v * 0x01010101...) >> (Len - 8) | |||
19401 | // | |||
19402 | // Change it to use vector adds + vector shifts which yield faster results on | |||
19403 | // Haswell than using vector integer multiplication. | |||
19404 | // | |||
19405 | // For i32 elements: | |||
19406 | // v = v + (v >> 8) | |||
19407 | // v = v + (v >> 16) | |||
19408 | // | |||
19409 | // For i64 elements: | |||
19410 | // v = v + (v >> 8) | |||
19411 | // v = v + (v >> 16) | |||
19412 | // v = v + (v >> 32) | |||
19413 | // | |||
19414 | Add = And; | |||
19415 | SmallVector<SDValue, 8> Csts; | |||
19416 | for (unsigned i = 8; i <= Len/2; i *= 2) { | |||
19417 | Csts.assign(NumElts, DAG.getConstant(i, EltVT)); | |||
19418 | SDValue CstsV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Csts); | |||
19419 | Srl = DAG.getNode(ISD::SRL, dl, VT, Add, CstsV); | |||
19420 | Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl); | |||
19421 | Csts.clear(); | |||
19422 | } | |||
19423 | ||||
19424 | // The result is on the least significant 6-bits on i32 and 7-bits on i64. | |||
19425 | SDValue Cst3F = DAG.getConstant(APInt(Len, Len == 32 ? 0x3F : 0x7F), EltVT); | |||
19426 | SmallVector<SDValue, 8> Cst3FV(NumElts, Cst3F); | |||
19427 | SDValue M3F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Cst3FV); | |||
19428 | if (NeedsBitcast) { | |||
19429 | Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add); | |||
19430 | M3F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M3F); | |||
19431 | } | |||
19432 | And = DAG.getNode(ISD::AND, dl, M3F.getValueType(), Add, M3F); | |||
19433 | if (VT != And.getValueType()) | |||
19434 | And = DAG.getNode(ISD::BITCAST, dl, VT, And); | |||
19435 | ||||
19436 | return And; | |||
19437 | } | |||
19438 | ||||
19439 | static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) { | |||
19440 | SDNode *Node = Op.getNode(); | |||
19441 | SDLoc dl(Node); | |||
19442 | EVT T = Node->getValueType(0); | |||
19443 | SDValue negOp = DAG.getNode(ISD::SUB, dl, T, | |||
19444 | DAG.getConstant(0, T), Node->getOperand(2)); | |||
19445 | return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl, | |||
19446 | cast<AtomicSDNode>(Node)->getMemoryVT(), | |||
19447 | Node->getOperand(0), | |||
19448 | Node->getOperand(1), negOp, | |||
19449 | cast<AtomicSDNode>(Node)->getMemOperand(), | |||
19450 | cast<AtomicSDNode>(Node)->getOrdering(), | |||
19451 | cast<AtomicSDNode>(Node)->getSynchScope()); | |||
19452 | } | |||
19453 | ||||
19454 | static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) { | |||
19455 | SDNode *Node = Op.getNode(); | |||
19456 | SDLoc dl(Node); | |||
19457 | EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT(); | |||
19458 | ||||
19459 | // Convert seq_cst store -> xchg | |||
19460 | // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b) | |||
19461 | // FIXME: On 32-bit, store -> fist or movq would be more efficient | |||
19462 | // (The only way to get a 16-byte store is cmpxchg16b) | |||
19463 | // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment. | |||
19464 | if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent || | |||
19465 | !DAG.getTargetLoweringInfo().isTypeLegal(VT)) { | |||
19466 | SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl, | |||
19467 | cast<AtomicSDNode>(Node)->getMemoryVT(), | |||
19468 | Node->getOperand(0), | |||
19469 | Node->getOperand(1), Node->getOperand(2), | |||
19470 | cast<AtomicSDNode>(Node)->getMemOperand(), | |||
19471 | cast<AtomicSDNode>(Node)->getOrdering(), | |||
19472 | cast<AtomicSDNode>(Node)->getSynchScope()); | |||
19473 | return Swap.getValue(1); | |||
19474 | } | |||
19475 | // Other atomic stores have a simple pattern. | |||
19476 | return Op; | |||
19477 | } | |||
19478 | ||||
19479 | static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) { | |||
19480 | EVT VT = Op.getNode()->getSimpleValueType(0); | |||
19481 | ||||
19482 | // Let legalize expand this if it isn't a legal type yet. | |||
19483 | if (!DAG.getTargetLoweringInfo().isTypeLegal(VT)) | |||
19484 | return SDValue(); | |||
19485 | ||||
19486 | SDVTList VTs = DAG.getVTList(VT, MVT::i32); | |||
19487 | ||||
19488 | unsigned Opc; | |||
19489 | bool ExtraOp = false; | |||
19490 | switch (Op.getOpcode()) { | |||
19491 | default: llvm_unreachable("Invalid code")::llvm::llvm_unreachable_internal("Invalid code", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19491); | |||
19492 | case ISD::ADDC: Opc = X86ISD::ADD; break; | |||
19493 | case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break; | |||
19494 | case ISD::SUBC: Opc = X86ISD::SUB; break; | |||
19495 | case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break; | |||
19496 | } | |||
19497 | ||||
19498 | if (!ExtraOp) | |||
19499 | return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), | |||
19500 | Op.getOperand(1)); | |||
19501 | return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), | |||
19502 | Op.getOperand(1), Op.getOperand(2)); | |||
19503 | } | |||
19504 | ||||
19505 | static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget, | |||
19506 | SelectionDAG &DAG) { | |||
19507 | assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit())((Subtarget->isTargetDarwin() && Subtarget->is64Bit ()) ? static_cast<void> (0) : __assert_fail ("Subtarget->isTargetDarwin() && Subtarget->is64Bit()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19507, __PRETTY_FUNCTION__)); | |||
19508 | ||||
19509 | // For MacOSX, we want to call an alternative entry point: __sincos_stret, | |||
19510 | // which returns the values as { float, float } (in XMM0) or | |||
19511 | // { double, double } (which is returned in XMM0, XMM1). | |||
19512 | SDLoc dl(Op); | |||
19513 | SDValue Arg = Op.getOperand(0); | |||
19514 | EVT ArgVT = Arg.getValueType(); | |||
19515 | Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext()); | |||
19516 | ||||
19517 | TargetLowering::ArgListTy Args; | |||
19518 | TargetLowering::ArgListEntry Entry; | |||
19519 | ||||
19520 | Entry.Node = Arg; | |||
19521 | Entry.Ty = ArgTy; | |||
19522 | Entry.isSExt = false; | |||
19523 | Entry.isZExt = false; | |||
19524 | Args.push_back(Entry); | |||
19525 | ||||
19526 | bool isF64 = ArgVT == MVT::f64; | |||
19527 | // Only optimize x86_64 for now. i386 is a bit messy. For f32, | |||
19528 | // the small struct {f32, f32} is returned in (eax, edx). For f64, | |||
19529 | // the results are returned via SRet in memory. | |||
19530 | const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret"; | |||
19531 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
19532 | SDValue Callee = DAG.getExternalSymbol(LibcallName, TLI.getPointerTy()); | |||
19533 | ||||
19534 | Type *RetTy = isF64 | |||
19535 | ? (Type*)StructType::get(ArgTy, ArgTy, nullptr) | |||
19536 | : (Type*)VectorType::get(ArgTy, 4); | |||
19537 | ||||
19538 | TargetLowering::CallLoweringInfo CLI(DAG); | |||
19539 | CLI.setDebugLoc(dl).setChain(DAG.getEntryNode()) | |||
19540 | .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0); | |||
19541 | ||||
19542 | std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI); | |||
19543 | ||||
19544 | if (isF64) | |||
19545 | // Returned in xmm0 and xmm1. | |||
19546 | return CallResult.first; | |||
19547 | ||||
19548 | // Returned in bits 0:31 and 32:64 xmm0. | |||
19549 | SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT, | |||
19550 | CallResult.first, DAG.getIntPtrConstant(0)); | |||
19551 | SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT, | |||
19552 | CallResult.first, DAG.getIntPtrConstant(1)); | |||
19553 | SDVTList Tys = DAG.getVTList(ArgVT, ArgVT); | |||
19554 | return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal); | |||
19555 | } | |||
19556 | ||||
19557 | /// LowerOperation - Provide custom lowering hooks for some operations. | |||
19558 | /// | |||
19559 | SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { | |||
19560 | switch (Op.getOpcode()) { | |||
19561 | default: llvm_unreachable("Should not custom lower this!")::llvm::llvm_unreachable_internal("Should not custom lower this!" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19561); | |||
19562 | case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG); | |||
19563 | case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG); | |||
19564 | case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: | |||
19565 | return LowerCMP_SWAP(Op, Subtarget, DAG); | |||
19566 | case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG); | |||
19567 | case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG); | |||
19568 | case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG); | |||
19569 | case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG); | |||
19570 | case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG); | |||
19571 | case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG); | |||
19572 | case ISD::VSELECT: return LowerVSELECT(Op, DAG); | |||
19573 | case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG); | |||
19574 | case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG); | |||
19575 | case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG); | |||
19576 | case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG); | |||
19577 | case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG); | |||
19578 | case ISD::ConstantPool: return LowerConstantPool(Op, DAG); | |||
19579 | case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG); | |||
19580 | case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); | |||
19581 | case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG); | |||
19582 | case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); | |||
19583 | case ISD::SHL_PARTS: | |||
19584 | case ISD::SRA_PARTS: | |||
19585 | case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG); | |||
19586 | case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG); | |||
19587 | case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG); | |||
19588 | case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG); | |||
19589 | case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG); | |||
19590 | case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG); | |||
19591 | case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG); | |||
19592 | case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG); | |||
19593 | case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG); | |||
19594 | case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG); | |||
19595 | case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG); | |||
19596 | case ISD::FABS: | |||
19597 | case ISD::FNEG: return LowerFABSorFNEG(Op, DAG); | |||
19598 | case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG); | |||
19599 | case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG); | |||
19600 | case ISD::SETCC: return LowerSETCC(Op, DAG); | |||
19601 | case ISD::SELECT: return LowerSELECT(Op, DAG); | |||
19602 | case ISD::BRCOND: return LowerBRCOND(Op, DAG); | |||
19603 | case ISD::JumpTable: return LowerJumpTable(Op, DAG); | |||
19604 | case ISD::VASTART: return LowerVASTART(Op, DAG); | |||
19605 | case ISD::VAARG: return LowerVAARG(Op, DAG); | |||
19606 | case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG); | |||
19607 | case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG); | |||
19608 | case ISD::INTRINSIC_VOID: | |||
19609 | case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG); | |||
19610 | case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG); | |||
19611 | case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); | |||
19612 | case ISD::FRAME_TO_ARGS_OFFSET: | |||
19613 | return LowerFRAME_TO_ARGS_OFFSET(Op, DAG); | |||
19614 | case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG); | |||
19615 | case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG); | |||
19616 | case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG); | |||
19617 | case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG); | |||
19618 | case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG); | |||
19619 | case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG); | |||
19620 | case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG); | |||
19621 | case ISD::CTLZ: return LowerCTLZ(Op, DAG); | |||
19622 | case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG); | |||
19623 | case ISD::CTTZ: return LowerCTTZ(Op, DAG); | |||
19624 | case ISD::MUL: return LowerMUL(Op, Subtarget, DAG); | |||
19625 | case ISD::UMUL_LOHI: | |||
19626 | case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG); | |||
19627 | case ISD::SRA: | |||
19628 | case ISD::SRL: | |||
19629 | case ISD::SHL: return LowerShift(Op, Subtarget, DAG); | |||
19630 | case ISD::SADDO: | |||
19631 | case ISD::UADDO: | |||
19632 | case ISD::SSUBO: | |||
19633 | case ISD::USUBO: | |||
19634 | case ISD::SMULO: | |||
19635 | case ISD::UMULO: return LowerXALUO(Op, DAG); | |||
19636 | case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG); | |||
19637 | case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG); | |||
19638 | case ISD::ADDC: | |||
19639 | case ISD::ADDE: | |||
19640 | case ISD::SUBC: | |||
19641 | case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG); | |||
19642 | case ISD::ADD: return LowerADD(Op, DAG); | |||
19643 | case ISD::SUB: return LowerSUB(Op, DAG); | |||
19644 | case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG); | |||
19645 | } | |||
19646 | } | |||
19647 | ||||
19648 | /// ReplaceNodeResults - Replace a node with an illegal result type | |||
19649 | /// with a new node built out of custom code. | |||
19650 | void X86TargetLowering::ReplaceNodeResults(SDNode *N, | |||
19651 | SmallVectorImpl<SDValue>&Results, | |||
19652 | SelectionDAG &DAG) const { | |||
19653 | SDLoc dl(N); | |||
19654 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
19655 | switch (N->getOpcode()) { | |||
19656 | default: | |||
19657 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19657); | |||
19658 | // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32. | |||
19659 | case X86ISD::FMINC: | |||
19660 | case X86ISD::FMIN: | |||
19661 | case X86ISD::FMAXC: | |||
19662 | case X86ISD::FMAX: { | |||
19663 | EVT VT = N->getValueType(0); | |||
19664 | if (VT != MVT::v2f32) | |||
19665 | llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.")::llvm::llvm_unreachable_internal("Unexpected type (!= v2f32) on FMIN/FMAX." , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19665); | |||
19666 | SDValue UNDEF = DAG.getUNDEF(VT); | |||
19667 | SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32, | |||
19668 | N->getOperand(0), UNDEF); | |||
19669 | SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32, | |||
19670 | N->getOperand(1), UNDEF); | |||
19671 | Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS)); | |||
19672 | return; | |||
19673 | } | |||
19674 | case ISD::SIGN_EXTEND_INREG: | |||
19675 | case ISD::ADDC: | |||
19676 | case ISD::ADDE: | |||
19677 | case ISD::SUBC: | |||
19678 | case ISD::SUBE: | |||
19679 | // We don't want to expand or promote these. | |||
19680 | return; | |||
19681 | case ISD::SDIV: | |||
19682 | case ISD::UDIV: | |||
19683 | case ISD::SREM: | |||
19684 | case ISD::UREM: | |||
19685 | case ISD::SDIVREM: | |||
19686 | case ISD::UDIVREM: { | |||
19687 | SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG); | |||
19688 | Results.push_back(V); | |||
19689 | return; | |||
19690 | } | |||
19691 | case ISD::FP_TO_SINT: | |||
19692 | case ISD::FP_TO_UINT: { | |||
19693 | bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT; | |||
19694 | ||||
19695 | if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType())) | |||
19696 | return; | |||
19697 | ||||
19698 | std::pair<SDValue,SDValue> Vals = | |||
19699 | FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true); | |||
19700 | SDValue FIST = Vals.first, StackSlot = Vals.second; | |||
19701 | if (FIST.getNode()) { | |||
19702 | EVT VT = N->getValueType(0); | |||
19703 | // Return a load from the stack slot. | |||
19704 | if (StackSlot.getNode()) | |||
19705 | Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, | |||
19706 | MachinePointerInfo(), | |||
19707 | false, false, false, 0)); | |||
19708 | else | |||
19709 | Results.push_back(FIST); | |||
19710 | } | |||
19711 | return; | |||
19712 | } | |||
19713 | case ISD::UINT_TO_FP: { | |||
19714 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19714, __PRETTY_FUNCTION__)); | |||
19715 | if (N->getOperand(0).getValueType() != MVT::v2i32 || | |||
19716 | N->getValueType(0) != MVT::v2f32) | |||
19717 | return; | |||
19718 | SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64, | |||
19719 | N->getOperand(0)); | |||
19720 | SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), | |||
19721 | MVT::f64); | |||
19722 | SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias); | |||
19723 | SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn, | |||
19724 | DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias)); | |||
19725 | Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or); | |||
19726 | SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias); | |||
19727 | Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub)); | |||
19728 | return; | |||
19729 | } | |||
19730 | case ISD::FP_ROUND: { | |||
19731 | if (!TLI.isTypeLegal(N->getOperand(0).getValueType())) | |||
19732 | return; | |||
19733 | SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0)); | |||
19734 | Results.push_back(V); | |||
19735 | return; | |||
19736 | } | |||
19737 | case ISD::INTRINSIC_W_CHAIN: { | |||
19738 | unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); | |||
19739 | switch (IntNo) { | |||
19740 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19741) | |||
19741 | "legalize this intrinsic operation!")::llvm::llvm_unreachable_internal("Do not know how to custom type " "legalize this intrinsic operation!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19741); | |||
19742 | case Intrinsic::x86_rdtsc: | |||
19743 | return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget, | |||
19744 | Results); | |||
19745 | case Intrinsic::x86_rdtscp: | |||
19746 | return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget, | |||
19747 | Results); | |||
19748 | case Intrinsic::x86_rdpmc: | |||
19749 | return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results); | |||
19750 | } | |||
19751 | } | |||
19752 | case ISD::READCYCLECOUNTER: { | |||
19753 | return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget, | |||
19754 | Results); | |||
19755 | } | |||
19756 | case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: { | |||
19757 | EVT T = N->getValueType(0); | |||
19758 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19758, __PRETTY_FUNCTION__)); | |||
19759 | bool Regs64bit = T == MVT::i128; | |||
19760 | EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32; | |||
19761 | SDValue cpInL, cpInH; | |||
19762 | cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2), | |||
19763 | DAG.getConstant(0, HalfT)); | |||
19764 | cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2), | |||
19765 | DAG.getConstant(1, HalfT)); | |||
19766 | cpInL = DAG.getCopyToReg(N->getOperand(0), dl, | |||
19767 | Regs64bit ? X86::RAX : X86::EAX, | |||
19768 | cpInL, SDValue()); | |||
19769 | cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, | |||
19770 | Regs64bit ? X86::RDX : X86::EDX, | |||
19771 | cpInH, cpInL.getValue(1)); | |||
19772 | SDValue swapInL, swapInH; | |||
19773 | swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3), | |||
19774 | DAG.getConstant(0, HalfT)); | |||
19775 | swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3), | |||
19776 | DAG.getConstant(1, HalfT)); | |||
19777 | swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, | |||
19778 | Regs64bit ? X86::RBX : X86::EBX, | |||
19779 | swapInL, cpInH.getValue(1)); | |||
19780 | swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, | |||
19781 | Regs64bit ? X86::RCX : X86::ECX, | |||
19782 | swapInH, swapInL.getValue(1)); | |||
19783 | SDValue Ops[] = { swapInH.getValue(0), | |||
19784 | N->getOperand(1), | |||
19785 | swapInH.getValue(1) }; | |||
19786 | SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
19787 | MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand(); | |||
19788 | unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG : | |||
19789 | X86ISD::LCMPXCHG8_DAG; | |||
19790 | SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO); | |||
19791 | SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, | |||
19792 | Regs64bit ? X86::RAX : X86::EAX, | |||
19793 | HalfT, Result.getValue(1)); | |||
19794 | SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, | |||
19795 | Regs64bit ? X86::RDX : X86::EDX, | |||
19796 | HalfT, cpOutL.getValue(2)); | |||
19797 | SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)}; | |||
19798 | ||||
19799 | SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS, | |||
19800 | MVT::i32, cpOutH.getValue(2)); | |||
19801 | SDValue Success = | |||
19802 | DAG.getNode(X86ISD::SETCC, dl, MVT::i8, | |||
19803 | DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS); | |||
19804 | Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1)); | |||
19805 | ||||
19806 | Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF)); | |||
19807 | Results.push_back(Success); | |||
19808 | Results.push_back(EFLAGS.getValue(1)); | |||
19809 | return; | |||
19810 | } | |||
19811 | case ISD::ATOMIC_SWAP: | |||
19812 | case ISD::ATOMIC_LOAD_ADD: | |||
19813 | case ISD::ATOMIC_LOAD_SUB: | |||
19814 | case ISD::ATOMIC_LOAD_AND: | |||
19815 | case ISD::ATOMIC_LOAD_OR: | |||
19816 | case ISD::ATOMIC_LOAD_XOR: | |||
19817 | case ISD::ATOMIC_LOAD_NAND: | |||
19818 | case ISD::ATOMIC_LOAD_MIN: | |||
19819 | case ISD::ATOMIC_LOAD_MAX: | |||
19820 | case ISD::ATOMIC_LOAD_UMIN: | |||
19821 | case ISD::ATOMIC_LOAD_UMAX: | |||
19822 | case ISD::ATOMIC_LOAD: { | |||
19823 | // Delegate to generic TypeLegalization. Situations we can really handle | |||
19824 | // should have already been dealt with by AtomicExpandPass.cpp. | |||
19825 | break; | |||
19826 | } | |||
19827 | case ISD::BITCAST: { | |||
19828 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 19828, __PRETTY_FUNCTION__)); | |||
19829 | EVT DstVT = N->getValueType(0); | |||
19830 | EVT SrcVT = N->getOperand(0)->getValueType(0); | |||
19831 | ||||
19832 | if (SrcVT != MVT::f64 || | |||
19833 | (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8)) | |||
19834 | return; | |||
19835 | ||||
19836 | unsigned NumElts = DstVT.getVectorNumElements(); | |||
19837 | EVT SVT = DstVT.getVectorElementType(); | |||
19838 | EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2); | |||
19839 | SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, | |||
19840 | MVT::v2f64, N->getOperand(0)); | |||
19841 | SDValue ToVecInt = DAG.getNode(ISD::BITCAST, dl, WiderVT, Expanded); | |||
19842 | ||||
19843 | if (ExperimentalVectorWideningLegalization) { | |||
19844 | // If we are legalizing vectors by widening, we already have the desired | |||
19845 | // legal vector type, just return it. | |||
19846 | Results.push_back(ToVecInt); | |||
19847 | return; | |||
19848 | } | |||
19849 | ||||
19850 | SmallVector<SDValue, 8> Elts; | |||
19851 | for (unsigned i = 0, e = NumElts; i != e; ++i) | |||
19852 | Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, | |||
19853 | ToVecInt, DAG.getIntPtrConstant(i))); | |||
19854 | ||||
19855 | Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts)); | |||
19856 | } | |||
19857 | } | |||
19858 | } | |||
19859 | ||||
19860 | const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const { | |||
19861 | switch (Opcode) { | |||
19862 | default: return nullptr; | |||
19863 | case X86ISD::BSF: return "X86ISD::BSF"; | |||
19864 | case X86ISD::BSR: return "X86ISD::BSR"; | |||
19865 | case X86ISD::SHLD: return "X86ISD::SHLD"; | |||
19866 | case X86ISD::SHRD: return "X86ISD::SHRD"; | |||
19867 | case X86ISD::FAND: return "X86ISD::FAND"; | |||
19868 | case X86ISD::FANDN: return "X86ISD::FANDN"; | |||
19869 | case X86ISD::FOR: return "X86ISD::FOR"; | |||
19870 | case X86ISD::FXOR: return "X86ISD::FXOR"; | |||
19871 | case X86ISD::FSRL: return "X86ISD::FSRL"; | |||
19872 | case X86ISD::FILD: return "X86ISD::FILD"; | |||
19873 | case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG"; | |||
19874 | case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM"; | |||
19875 | case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM"; | |||
19876 | case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM"; | |||
19877 | case X86ISD::FLD: return "X86ISD::FLD"; | |||
19878 | case X86ISD::FST: return "X86ISD::FST"; | |||
19879 | case X86ISD::CALL: return "X86ISD::CALL"; | |||
19880 | case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG"; | |||
19881 | case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG"; | |||
19882 | case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG"; | |||
19883 | case X86ISD::BT: return "X86ISD::BT"; | |||
19884 | case X86ISD::CMP: return "X86ISD::CMP"; | |||
19885 | case X86ISD::COMI: return "X86ISD::COMI"; | |||
19886 | case X86ISD::UCOMI: return "X86ISD::UCOMI"; | |||
19887 | case X86ISD::CMPM: return "X86ISD::CMPM"; | |||
19888 | case X86ISD::CMPMU: return "X86ISD::CMPMU"; | |||
19889 | case X86ISD::SETCC: return "X86ISD::SETCC"; | |||
19890 | case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY"; | |||
19891 | case X86ISD::FSETCC: return "X86ISD::FSETCC"; | |||
19892 | case X86ISD::CMOV: return "X86ISD::CMOV"; | |||
19893 | case X86ISD::BRCOND: return "X86ISD::BRCOND"; | |||
19894 | case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG"; | |||
19895 | case X86ISD::REP_STOS: return "X86ISD::REP_STOS"; | |||
19896 | case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS"; | |||
19897 | case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg"; | |||
19898 | case X86ISD::Wrapper: return "X86ISD::Wrapper"; | |||
19899 | case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP"; | |||
19900 | case X86ISD::PEXTRB: return "X86ISD::PEXTRB"; | |||
19901 | case X86ISD::PEXTRW: return "X86ISD::PEXTRW"; | |||
19902 | case X86ISD::INSERTPS: return "X86ISD::INSERTPS"; | |||
19903 | case X86ISD::PINSRB: return "X86ISD::PINSRB"; | |||
19904 | case X86ISD::PINSRW: return "X86ISD::PINSRW"; | |||
19905 | case X86ISD::PSHUFB: return "X86ISD::PSHUFB"; | |||
19906 | case X86ISD::ANDNP: return "X86ISD::ANDNP"; | |||
19907 | case X86ISD::PSIGN: return "X86ISD::PSIGN"; | |||
19908 | case X86ISD::BLENDI: return "X86ISD::BLENDI"; | |||
19909 | case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND"; | |||
19910 | case X86ISD::SUBUS: return "X86ISD::SUBUS"; | |||
19911 | case X86ISD::HADD: return "X86ISD::HADD"; | |||
19912 | case X86ISD::HSUB: return "X86ISD::HSUB"; | |||
19913 | case X86ISD::FHADD: return "X86ISD::FHADD"; | |||
19914 | case X86ISD::FHSUB: return "X86ISD::FHSUB"; | |||
19915 | case X86ISD::UMAX: return "X86ISD::UMAX"; | |||
19916 | case X86ISD::UMIN: return "X86ISD::UMIN"; | |||
19917 | case X86ISD::SMAX: return "X86ISD::SMAX"; | |||
19918 | case X86ISD::SMIN: return "X86ISD::SMIN"; | |||
19919 | case X86ISD::FMAX: return "X86ISD::FMAX"; | |||
19920 | case X86ISD::FMIN: return "X86ISD::FMIN"; | |||
19921 | case X86ISD::FMAXC: return "X86ISD::FMAXC"; | |||
19922 | case X86ISD::FMINC: return "X86ISD::FMINC"; | |||
19923 | case X86ISD::FRSQRT: return "X86ISD::FRSQRT"; | |||
19924 | case X86ISD::FRCP: return "X86ISD::FRCP"; | |||
19925 | case X86ISD::TLSADDR: return "X86ISD::TLSADDR"; | |||
19926 | case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR"; | |||
19927 | case X86ISD::TLSCALL: return "X86ISD::TLSCALL"; | |||
19928 | case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP"; | |||
19929 | case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP"; | |||
19930 | case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN"; | |||
19931 | case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN"; | |||
19932 | case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m"; | |||
19933 | case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r"; | |||
19934 | case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG"; | |||
19935 | case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG"; | |||
19936 | case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG"; | |||
19937 | case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL"; | |||
19938 | case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD"; | |||
19939 | case X86ISD::VZEXT: return "X86ISD::VZEXT"; | |||
19940 | case X86ISD::VSEXT: return "X86ISD::VSEXT"; | |||
19941 | case X86ISD::VTRUNC: return "X86ISD::VTRUNC"; | |||
19942 | case X86ISD::VTRUNCM: return "X86ISD::VTRUNCM"; | |||
19943 | case X86ISD::VINSERT: return "X86ISD::VINSERT"; | |||
19944 | case X86ISD::VFPEXT: return "X86ISD::VFPEXT"; | |||
19945 | case X86ISD::VFPROUND: return "X86ISD::VFPROUND"; | |||
19946 | case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ"; | |||
19947 | case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ"; | |||
19948 | case X86ISD::VSHL: return "X86ISD::VSHL"; | |||
19949 | case X86ISD::VSRL: return "X86ISD::VSRL"; | |||
19950 | case X86ISD::VSRA: return "X86ISD::VSRA"; | |||
19951 | case X86ISD::VSHLI: return "X86ISD::VSHLI"; | |||
19952 | case X86ISD::VSRLI: return "X86ISD::VSRLI"; | |||
19953 | case X86ISD::VSRAI: return "X86ISD::VSRAI"; | |||
19954 | case X86ISD::CMPP: return "X86ISD::CMPP"; | |||
19955 | case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ"; | |||
19956 | case X86ISD::PCMPGT: return "X86ISD::PCMPGT"; | |||
19957 | case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM"; | |||
19958 | case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM"; | |||
19959 | case X86ISD::ADD: return "X86ISD::ADD"; | |||
19960 | case X86ISD::SUB: return "X86ISD::SUB"; | |||
19961 | case X86ISD::ADC: return "X86ISD::ADC"; | |||
19962 | case X86ISD::SBB: return "X86ISD::SBB"; | |||
19963 | case X86ISD::SMUL: return "X86ISD::SMUL"; | |||
19964 | case X86ISD::UMUL: return "X86ISD::UMUL"; | |||
19965 | case X86ISD::SMUL8: return "X86ISD::SMUL8"; | |||
19966 | case X86ISD::UMUL8: return "X86ISD::UMUL8"; | |||
19967 | case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG"; | |||
19968 | case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG"; | |||
19969 | case X86ISD::INC: return "X86ISD::INC"; | |||
19970 | case X86ISD::DEC: return "X86ISD::DEC"; | |||
19971 | case X86ISD::OR: return "X86ISD::OR"; | |||
19972 | case X86ISD::XOR: return "X86ISD::XOR"; | |||
19973 | case X86ISD::AND: return "X86ISD::AND"; | |||
19974 | case X86ISD::BEXTR: return "X86ISD::BEXTR"; | |||
19975 | case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM"; | |||
19976 | case X86ISD::PTEST: return "X86ISD::PTEST"; | |||
19977 | case X86ISD::TESTP: return "X86ISD::TESTP"; | |||
19978 | case X86ISD::TESTM: return "X86ISD::TESTM"; | |||
19979 | case X86ISD::TESTNM: return "X86ISD::TESTNM"; | |||
19980 | case X86ISD::KORTEST: return "X86ISD::KORTEST"; | |||
19981 | case X86ISD::PACKSS: return "X86ISD::PACKSS"; | |||
19982 | case X86ISD::PACKUS: return "X86ISD::PACKUS"; | |||
19983 | case X86ISD::PALIGNR: return "X86ISD::PALIGNR"; | |||
19984 | case X86ISD::VALIGN: return "X86ISD::VALIGN"; | |||
19985 | case X86ISD::PSHUFD: return "X86ISD::PSHUFD"; | |||
19986 | case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW"; | |||
19987 | case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW"; | |||
19988 | case X86ISD::SHUFP: return "X86ISD::SHUFP"; | |||
19989 | case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS"; | |||
19990 | case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD"; | |||
19991 | case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS"; | |||
19992 | case X86ISD::MOVLPS: return "X86ISD::MOVLPS"; | |||
19993 | case X86ISD::MOVLPD: return "X86ISD::MOVLPD"; | |||
19994 | case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP"; | |||
19995 | case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP"; | |||
19996 | case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP"; | |||
19997 | case X86ISD::MOVSD: return "X86ISD::MOVSD"; | |||
19998 | case X86ISD::MOVSS: return "X86ISD::MOVSS"; | |||
19999 | case X86ISD::UNPCKL: return "X86ISD::UNPCKL"; | |||
20000 | case X86ISD::UNPCKH: return "X86ISD::UNPCKH"; | |||
20001 | case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST"; | |||
20002 | case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM"; | |||
20003 | case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT"; | |||
20004 | case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI"; | |||
20005 | case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128"; | |||
20006 | case X86ISD::VPERMV: return "X86ISD::VPERMV"; | |||
20007 | case X86ISD::VPERMV3: return "X86ISD::VPERMV3"; | |||
20008 | case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3"; | |||
20009 | case X86ISD::VPERMI: return "X86ISD::VPERMI"; | |||
20010 | case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ"; | |||
20011 | case X86ISD::PMULDQ: return "X86ISD::PMULDQ"; | |||
20012 | case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS"; | |||
20013 | case X86ISD::VAARG_64: return "X86ISD::VAARG_64"; | |||
20014 | case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA"; | |||
20015 | case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER"; | |||
20016 | case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA"; | |||
20017 | case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL"; | |||
20018 | case X86ISD::SAHF: return "X86ISD::SAHF"; | |||
20019 | case X86ISD::RDRAND: return "X86ISD::RDRAND"; | |||
20020 | case X86ISD::RDSEED: return "X86ISD::RDSEED"; | |||
20021 | case X86ISD::FMADD: return "X86ISD::FMADD"; | |||
20022 | case X86ISD::FMSUB: return "X86ISD::FMSUB"; | |||
20023 | case X86ISD::FNMADD: return "X86ISD::FNMADD"; | |||
20024 | case X86ISD::FNMSUB: return "X86ISD::FNMSUB"; | |||
20025 | case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB"; | |||
20026 | case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD"; | |||
20027 | case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI"; | |||
20028 | case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI"; | |||
20029 | case X86ISD::XTEST: return "X86ISD::XTEST"; | |||
20030 | case X86ISD::COMPRESS: return "X86ISD::COMPRESS"; | |||
20031 | case X86ISD::EXPAND: return "X86ISD::EXPAND"; | |||
20032 | case X86ISD::SELECT: return "X86ISD::SELECT"; | |||
20033 | case X86ISD::ADDSUB: return "X86ISD::ADDSUB"; | |||
20034 | case X86ISD::RCP28: return "X86ISD::RCP28"; | |||
20035 | case X86ISD::RSQRT28: return "X86ISD::RSQRT28"; | |||
20036 | } | |||
20037 | } | |||
20038 | ||||
20039 | // isLegalAddressingMode - Return true if the addressing mode represented | |||
20040 | // by AM is legal for this target, for a load/store of the specified type. | |||
20041 | bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM, | |||
20042 | Type *Ty) const { | |||
20043 | // X86 supports extremely general addressing modes. | |||
20044 | CodeModel::Model M = getTargetMachine().getCodeModel(); | |||
20045 | Reloc::Model R = getTargetMachine().getRelocationModel(); | |||
20046 | ||||
20047 | // X86 allows a sign-extended 32-bit immediate field as a displacement. | |||
20048 | if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr)) | |||
20049 | return false; | |||
20050 | ||||
20051 | if (AM.BaseGV) { | |||
20052 | unsigned GVFlags = | |||
20053 | Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine()); | |||
20054 | ||||
20055 | // If a reference to this global requires an extra load, we can't fold it. | |||
20056 | if (isGlobalStubReference(GVFlags)) | |||
20057 | return false; | |||
20058 | ||||
20059 | // If BaseGV requires a register for the PIC base, we cannot also have a | |||
20060 | // BaseReg specified. | |||
20061 | if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags)) | |||
20062 | return false; | |||
20063 | ||||
20064 | // If lower 4G is not available, then we must use rip-relative addressing. | |||
20065 | if ((M != CodeModel::Small || R != Reloc::Static) && | |||
20066 | Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1)) | |||
20067 | return false; | |||
20068 | } | |||
20069 | ||||
20070 | switch (AM.Scale) { | |||
20071 | case 0: | |||
20072 | case 1: | |||
20073 | case 2: | |||
20074 | case 4: | |||
20075 | case 8: | |||
20076 | // These scales always work. | |||
20077 | break; | |||
20078 | case 3: | |||
20079 | case 5: | |||
20080 | case 9: | |||
20081 | // These scales are formed with basereg+scalereg. Only accept if there is | |||
20082 | // no basereg yet. | |||
20083 | if (AM.HasBaseReg) | |||
20084 | return false; | |||
20085 | break; | |||
20086 | default: // Other stuff never works. | |||
20087 | return false; | |||
20088 | } | |||
20089 | ||||
20090 | return true; | |||
20091 | } | |||
20092 | ||||
20093 | bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const { | |||
20094 | unsigned Bits = Ty->getScalarSizeInBits(); | |||
20095 | ||||
20096 | // 8-bit shifts are always expensive, but versions with a scalar amount aren't | |||
20097 | // particularly cheaper than those without. | |||
20098 | if (Bits == 8) | |||
20099 | return false; | |||
20100 | ||||
20101 | // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make | |||
20102 | // variable shifts just as cheap as scalar ones. | |||
20103 | if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64)) | |||
20104 | return false; | |||
20105 | ||||
20106 | // Otherwise, it's significantly cheaper to shift by a scalar amount than by a | |||
20107 | // fully general vector. | |||
20108 | return true; | |||
20109 | } | |||
20110 | ||||
20111 | bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const { | |||
20112 | if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy()) | |||
20113 | return false; | |||
20114 | unsigned NumBits1 = Ty1->getPrimitiveSizeInBits(); | |||
20115 | unsigned NumBits2 = Ty2->getPrimitiveSizeInBits(); | |||
20116 | return NumBits1 > NumBits2; | |||
20117 | } | |||
20118 | ||||
20119 | bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const { | |||
20120 | if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy()) | |||
20121 | return false; | |||
20122 | ||||
20123 | if (!isTypeLegal(EVT::getEVT(Ty1))) | |||
20124 | return false; | |||
20125 | ||||
20126 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 20126, __PRETTY_FUNCTION__)); | |||
20127 | ||||
20128 | // Assuming the caller doesn't have a zeroext or signext return parameter, | |||
20129 | // truncation all the way down to i1 is valid. | |||
20130 | return true; | |||
20131 | } | |||
20132 | ||||
20133 | bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const { | |||
20134 | return isInt<32>(Imm); | |||
20135 | } | |||
20136 | ||||
20137 | bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const { | |||
20138 | // Can also use sub to handle negated immediates. | |||
20139 | return isInt<32>(Imm); | |||
20140 | } | |||
20141 | ||||
20142 | bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const { | |||
20143 | if (!VT1.isInteger() || !VT2.isInteger()) | |||
20144 | return false; | |||
20145 | unsigned NumBits1 = VT1.getSizeInBits(); | |||
20146 | unsigned NumBits2 = VT2.getSizeInBits(); | |||
20147 | return NumBits1 > NumBits2; | |||
20148 | } | |||
20149 | ||||
20150 | bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const { | |||
20151 | // x86-64 implicitly zero-extends 32-bit results in 64-bit registers. | |||
20152 | return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit(); | |||
20153 | } | |||
20154 | ||||
20155 | bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const { | |||
20156 | // x86-64 implicitly zero-extends 32-bit results in 64-bit registers. | |||
20157 | return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit(); | |||
20158 | } | |||
20159 | ||||
20160 | bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const { | |||
20161 | EVT VT1 = Val.getValueType(); | |||
20162 | if (isZExtFree(VT1, VT2)) | |||
20163 | return true; | |||
20164 | ||||
20165 | if (Val.getOpcode() != ISD::LOAD) | |||
20166 | return false; | |||
20167 | ||||
20168 | if (!VT1.isSimple() || !VT1.isInteger() || | |||
20169 | !VT2.isSimple() || !VT2.isInteger()) | |||
20170 | return false; | |||
20171 | ||||
20172 | switch (VT1.getSimpleVT().SimpleTy) { | |||
20173 | default: break; | |||
20174 | case MVT::i8: | |||
20175 | case MVT::i16: | |||
20176 | case MVT::i32: | |||
20177 | // X86 has 8, 16, and 32-bit zero-extending loads. | |||
20178 | return true; | |||
20179 | } | |||
20180 | ||||
20181 | return false; | |||
20182 | } | |||
20183 | ||||
20184 | bool | |||
20185 | X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const { | |||
20186 | if (!(Subtarget->hasFMA() || Subtarget->hasFMA4())) | |||
20187 | return false; | |||
20188 | ||||
20189 | VT = VT.getScalarType(); | |||
20190 | ||||
20191 | if (!VT.isSimple()) | |||
20192 | return false; | |||
20193 | ||||
20194 | switch (VT.getSimpleVT().SimpleTy) { | |||
20195 | case MVT::f32: | |||
20196 | case MVT::f64: | |||
20197 | return true; | |||
20198 | default: | |||
20199 | break; | |||
20200 | } | |||
20201 | ||||
20202 | return false; | |||
20203 | } | |||
20204 | ||||
20205 | bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const { | |||
20206 | // i16 instructions are longer (0x66 prefix) and potentially slower. | |||
20207 | return !(VT1 == MVT::i32 && VT2 == MVT::i16); | |||
20208 | } | |||
20209 | ||||
20210 | /// isShuffleMaskLegal - Targets can use this to indicate that they only | |||
20211 | /// support *some* VECTOR_SHUFFLE operations, those with specific masks. | |||
20212 | /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values | |||
20213 | /// are assumed to be legal. | |||
20214 | bool | |||
20215 | X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M, | |||
20216 | EVT VT) const { | |||
20217 | if (!VT.isSimple()) | |||
20218 | return false; | |||
20219 | ||||
20220 | MVT SVT = VT.getSimpleVT(); | |||
20221 | ||||
20222 | // Very little shuffling can be done for 64-bit vectors right now. | |||
20223 | if (VT.getSizeInBits() == 64) | |||
20224 | return false; | |||
20225 | ||||
20226 | // This is an experimental legality test that is tailored to match the | |||
20227 | // legality test of the experimental lowering more closely. They are gated | |||
20228 | // separately to ease testing of performance differences. | |||
20229 | if (ExperimentalVectorShuffleLegality) | |||
20230 | // We only care that the types being shuffled are legal. The lowering can | |||
20231 | // handle any possible shuffle mask that results. | |||
20232 | return isTypeLegal(SVT); | |||
20233 | ||||
20234 | // If this is a single-input shuffle with no 128 bit lane crossings we can | |||
20235 | // lower it into pshufb. | |||
20236 | if ((SVT.is128BitVector() && Subtarget->hasSSSE3()) || | |||
20237 | (SVT.is256BitVector() && Subtarget->hasInt256())) { | |||
20238 | bool isLegal = true; | |||
20239 | for (unsigned I = 0, E = M.size(); I != E; ++I) { | |||
20240 | if (M[I] >= (int)SVT.getVectorNumElements() || | |||
20241 | ShuffleCrosses128bitLane(SVT, I, M[I])) { | |||
20242 | isLegal = false; | |||
20243 | break; | |||
20244 | } | |||
20245 | } | |||
20246 | if (isLegal) | |||
20247 | return true; | |||
20248 | } | |||
20249 | ||||
20250 | // FIXME: blends, shifts. | |||
20251 | return (SVT.getVectorNumElements() == 2 || | |||
20252 | ShuffleVectorSDNode::isSplatMask(&M[0], VT) || | |||
20253 | isMOVLMask(M, SVT) || | |||
20254 | isCommutedMOVLMask(M, SVT) || | |||
20255 | isMOVHLPSMask(M, SVT) || | |||
20256 | isSHUFPMask(M, SVT) || | |||
20257 | isSHUFPMask(M, SVT, /* Commuted */ true) || | |||
20258 | isPSHUFDMask(M, SVT) || | |||
20259 | isPSHUFDMask(M, SVT, /* SecondOperand */ true) || | |||
20260 | isPSHUFHWMask(M, SVT, Subtarget->hasInt256()) || | |||
20261 | isPSHUFLWMask(M, SVT, Subtarget->hasInt256()) || | |||
20262 | isPALIGNRMask(M, SVT, Subtarget) || | |||
20263 | isUNPCKLMask(M, SVT, Subtarget->hasInt256()) || | |||
20264 | isUNPCKHMask(M, SVT, Subtarget->hasInt256()) || | |||
20265 | isUNPCKL_v_undef_Mask(M, SVT, Subtarget->hasInt256()) || | |||
20266 | isUNPCKH_v_undef_Mask(M, SVT, Subtarget->hasInt256()) || | |||
20267 | isBlendMask(M, SVT, Subtarget->hasSSE41(), Subtarget->hasInt256()) || | |||
20268 | (Subtarget->hasSSE41() && isINSERTPSMask(M, SVT))); | |||
20269 | } | |||
20270 | ||||
20271 | bool | |||
20272 | X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask, | |||
20273 | EVT VT) const { | |||
20274 | if (!VT.isSimple()) | |||
20275 | return false; | |||
20276 | ||||
20277 | MVT SVT = VT.getSimpleVT(); | |||
20278 | ||||
20279 | // This is an experimental legality test that is tailored to match the | |||
20280 | // legality test of the experimental lowering more closely. They are gated | |||
20281 | // separately to ease testing of performance differences. | |||
20282 | if (ExperimentalVectorShuffleLegality) | |||
20283 | // The new vector shuffle lowering is very good at managing zero-inputs. | |||
20284 | return isShuffleMaskLegal(Mask, VT); | |||
20285 | ||||
20286 | unsigned NumElts = SVT.getVectorNumElements(); | |||
20287 | // FIXME: This collection of masks seems suspect. | |||
20288 | if (NumElts == 2) | |||
20289 | return true; | |||
20290 | if (NumElts == 4 && SVT.is128BitVector()) { | |||
20291 | return (isMOVLMask(Mask, SVT) || | |||
20292 | isCommutedMOVLMask(Mask, SVT, true) || | |||
20293 | isSHUFPMask(Mask, SVT) || | |||
20294 | isSHUFPMask(Mask, SVT, /* Commuted */ true) || | |||
20295 | isBlendMask(Mask, SVT, Subtarget->hasSSE41(), | |||
20296 | Subtarget->hasInt256())); | |||
20297 | } | |||
20298 | return false; | |||
20299 | } | |||
20300 | ||||
20301 | //===----------------------------------------------------------------------===// | |||
20302 | // X86 Scheduler Hooks | |||
20303 | //===----------------------------------------------------------------------===// | |||
20304 | ||||
20305 | /// Utility function to emit xbegin specifying the start of an RTM region. | |||
20306 | static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB, | |||
20307 | const TargetInstrInfo *TII) { | |||
20308 | DebugLoc DL = MI->getDebugLoc(); | |||
20309 | ||||
20310 | const BasicBlock *BB = MBB->getBasicBlock(); | |||
20311 | MachineFunction::iterator I = MBB; | |||
20312 | ++I; | |||
20313 | ||||
20314 | // For the v = xbegin(), we generate | |||
20315 | // | |||
20316 | // thisMBB: | |||
20317 | // xbegin sinkMBB | |||
20318 | // | |||
20319 | // mainMBB: | |||
20320 | // eax = -1 | |||
20321 | // | |||
20322 | // sinkMBB: | |||
20323 | // v = eax | |||
20324 | ||||
20325 | MachineBasicBlock *thisMBB = MBB; | |||
20326 | MachineFunction *MF = MBB->getParent(); | |||
20327 | MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB); | |||
20328 | MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB); | |||
20329 | MF->insert(I, mainMBB); | |||
20330 | MF->insert(I, sinkMBB); | |||
20331 | ||||
20332 | // Transfer the remainder of BB and its successor edges to sinkMBB. | |||
20333 | sinkMBB->splice(sinkMBB->begin(), MBB, | |||
20334 | std::next(MachineBasicBlock::iterator(MI)), MBB->end()); | |||
20335 | sinkMBB->transferSuccessorsAndUpdatePHIs(MBB); | |||
20336 | ||||
20337 | // thisMBB: | |||
20338 | // xbegin sinkMBB | |||
20339 | // # fallthrough to mainMBB | |||
20340 | // # abortion to sinkMBB | |||
20341 | BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB); | |||
20342 | thisMBB->addSuccessor(mainMBB); | |||
20343 | thisMBB->addSuccessor(sinkMBB); | |||
20344 | ||||
20345 | // mainMBB: | |||
20346 | // EAX = -1 | |||
20347 | BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1); | |||
20348 | mainMBB->addSuccessor(sinkMBB); | |||
20349 | ||||
20350 | // sinkMBB: | |||
20351 | // EAX is live into the sinkMBB | |||
20352 | sinkMBB->addLiveIn(X86::EAX); | |||
20353 | BuildMI(*sinkMBB, sinkMBB->begin(), DL, | |||
20354 | TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg()) | |||
20355 | .addReg(X86::EAX); | |||
20356 | ||||
20357 | MI->eraseFromParent(); | |||
20358 | return sinkMBB; | |||
20359 | } | |||
20360 | ||||
20361 | // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8 | |||
20362 | // or XMM0_V32I8 in AVX all of this code can be replaced with that | |||
20363 | // in the .td file. | |||
20364 | static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB, | |||
20365 | const TargetInstrInfo *TII) { | |||
20366 | unsigned Opc; | |||
20367 | switch (MI->getOpcode()) { | |||
20368 | default: llvm_unreachable("illegal opcode!")::llvm::llvm_unreachable_internal("illegal opcode!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 20368); | |||
20369 | case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break; | |||
20370 | case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break; | |||
20371 | case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break; | |||
20372 | case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break; | |||
20373 | case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break; | |||
20374 | case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break; | |||
20375 | case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break; | |||
20376 | case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break; | |||
20377 | } | |||
20378 | ||||
20379 | DebugLoc dl = MI->getDebugLoc(); | |||
20380 | MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc)); | |||
20381 | ||||
20382 | unsigned NumArgs = MI->getNumOperands(); | |||
20383 | for (unsigned i = 1; i < NumArgs; ++i) { | |||
20384 | MachineOperand &Op = MI->getOperand(i); | |||
20385 | if (!(Op.isReg() && Op.isImplicit())) | |||
20386 | MIB.addOperand(Op); | |||
20387 | } | |||
20388 | if (MI->hasOneMemOperand()) | |||
20389 | MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); | |||
20390 | ||||
20391 | BuildMI(*BB, MI, dl, | |||
20392 | TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg()) | |||
20393 | .addReg(X86::XMM0); | |||
20394 | ||||
20395 | MI->eraseFromParent(); | |||
20396 | return BB; | |||
20397 | } | |||
20398 | ||||
20399 | // FIXME: Custom handling because TableGen doesn't support multiple implicit | |||
20400 | // defs in an instruction pattern | |||
20401 | static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB, | |||
20402 | const TargetInstrInfo *TII) { | |||
20403 | unsigned Opc; | |||
20404 | switch (MI->getOpcode()) { | |||
20405 | default: llvm_unreachable("illegal opcode!")::llvm::llvm_unreachable_internal("illegal opcode!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 20405); | |||
20406 | case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break; | |||
20407 | case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break; | |||
20408 | case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break; | |||
20409 | case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break; | |||
20410 | case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break; | |||
20411 | case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break; | |||
20412 | case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break; | |||
20413 | case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break; | |||
20414 | } | |||
20415 | ||||
20416 | DebugLoc dl = MI->getDebugLoc(); | |||
20417 | MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc)); | |||
20418 | ||||
20419 | unsigned NumArgs = MI->getNumOperands(); // remove the results | |||
20420 | for (unsigned i = 1; i < NumArgs; ++i) { | |||
20421 | MachineOperand &Op = MI->getOperand(i); | |||
20422 | if (!(Op.isReg() && Op.isImplicit())) | |||
20423 | MIB.addOperand(Op); | |||
20424 | } | |||
20425 | if (MI->hasOneMemOperand()) | |||
20426 | MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); | |||
20427 | ||||
20428 | BuildMI(*BB, MI, dl, | |||
20429 | TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg()) | |||
20430 | .addReg(X86::ECX); | |||
20431 | ||||
20432 | MI->eraseFromParent(); | |||
20433 | return BB; | |||
20434 | } | |||
20435 | ||||
20436 | static MachineBasicBlock * EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB, | |||
20437 | const TargetInstrInfo *TII, | |||
20438 | const X86Subtarget* Subtarget) { | |||
20439 | DebugLoc dl = MI->getDebugLoc(); | |||
20440 | ||||
20441 | // Address into RAX/EAX, other two args into ECX, EDX. | |||
20442 | unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r; | |||
20443 | unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX; | |||
20444 | MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg); | |||
20445 | for (int i = 0; i < X86::AddrNumOperands; ++i) | |||
20446 | MIB.addOperand(MI->getOperand(i)); | |||
20447 | ||||
20448 | unsigned ValOps = X86::AddrNumOperands; | |||
20449 | BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX) | |||
20450 | .addReg(MI->getOperand(ValOps).getReg()); | |||
20451 | BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX) | |||
20452 | .addReg(MI->getOperand(ValOps+1).getReg()); | |||
20453 | ||||
20454 | // The instruction doesn't actually take any operands though. | |||
20455 | BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr)); | |||
20456 | ||||
20457 | MI->eraseFromParent(); // The pseudo is gone now. | |||
20458 | return BB; | |||
20459 | } | |||
20460 | ||||
20461 | MachineBasicBlock * | |||
20462 | X86TargetLowering::EmitVAARG64WithCustomInserter( | |||
20463 | MachineInstr *MI, | |||
20464 | MachineBasicBlock *MBB) const { | |||
20465 | // Emit va_arg instruction on X86-64. | |||
20466 | ||||
20467 | // Operands to this pseudo-instruction: | |||
20468 | // 0 ) Output : destination address (reg) | |||
20469 | // 1-5) Input : va_list address (addr, i64mem) | |||
20470 | // 6 ) ArgSize : Size (in bytes) of vararg type | |||
20471 | // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset | |||
20472 | // 8 ) Align : Alignment of type | |||
20473 | // 9 ) EFLAGS (implicit-def) | |||
20474 | ||||
20475 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 20475, __PRETTY_FUNCTION__)); | |||
20476 | assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands")((X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands" ) ? static_cast<void> (0) : __assert_fail ("X86::AddrNumOperands == 5 && \"VAARG_64 assumes 5 address operands\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 20476, __PRETTY_FUNCTION__)); | |||
20477 | ||||
20478 | unsigned DestReg = MI->getOperand(0).getReg(); | |||
20479 | MachineOperand &Base = MI->getOperand(1); | |||
20480 | MachineOperand &Scale = MI->getOperand(2); | |||
20481 | MachineOperand &Index = MI->getOperand(3); | |||
20482 | MachineOperand &Disp = MI->getOperand(4); | |||
20483 | MachineOperand &Segment = MI->getOperand(5); | |||
20484 | unsigned ArgSize = MI->getOperand(6).getImm(); | |||
20485 | unsigned ArgMode = MI->getOperand(7).getImm(); | |||
20486 | unsigned Align = MI->getOperand(8).getImm(); | |||
20487 | ||||
20488 | // Memory Reference | |||
20489 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 20489, __PRETTY_FUNCTION__)); | |||
20490 | MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin(); | |||
20491 | MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end(); | |||
20492 | ||||
20493 | // Machine Information | |||
20494 | const TargetInstrInfo *TII = MBB->getParent()->getSubtarget().getInstrInfo(); | |||
20495 | MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); | |||
20496 | const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64); | |||
20497 | const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32); | |||
20498 | DebugLoc DL = MI->getDebugLoc(); | |||
20499 | ||||
20500 | // struct va_list { | |||
20501 | // i32 gp_offset | |||
20502 | // i32 fp_offset | |||
20503 | // i64 overflow_area (address) | |||
20504 | // i64 reg_save_area (address) | |||
20505 | // } | |||
20506 | // sizeof(va_list) = 24 | |||
20507 | // alignment(va_list) = 8 | |||
20508 | ||||
20509 | unsigned TotalNumIntRegs = 6; | |||
20510 | unsigned TotalNumXMMRegs = 8; | |||
20511 | bool UseGPOffset = (ArgMode == 1); | |||
20512 | bool UseFPOffset = (ArgMode == 2); | |||
20513 | unsigned MaxOffset = TotalNumIntRegs * 8 + | |||
20514 | (UseFPOffset ? TotalNumXMMRegs * 16 : 0); | |||
20515 | ||||
20516 | /* Align ArgSize to a multiple of 8 */ | |||
20517 | unsigned ArgSizeA8 = (ArgSize + 7) & ~7; | |||
20518 | bool NeedsAlign = (Align > 8); | |||
20519 | ||||
20520 | MachineBasicBlock *thisMBB = MBB; | |||
20521 | MachineBasicBlock *overflowMBB; | |||
20522 | MachineBasicBlock *offsetMBB; | |||
20523 | MachineBasicBlock *endMBB; | |||
20524 | ||||
20525 | unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB | |||
20526 | unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB | |||
20527 | unsigned OffsetReg = 0; | |||
20528 | ||||
20529 | if (!UseGPOffset && !UseFPOffset) { | |||
20530 | // If we only pull from the overflow region, we don't create a branch. | |||
20531 | // We don't need to alter control flow. | |||
20532 | OffsetDestReg = 0; // unused | |||
20533 | OverflowDestReg = DestReg; | |||
20534 | ||||
20535 | offsetMBB = nullptr; | |||
20536 | overflowMBB = thisMBB; | |||
20537 | endMBB = thisMBB; | |||
20538 | } else { | |||
20539 | // First emit code to check if gp_offset (or fp_offset) is below the bound. | |||
20540 | // If so, pull the argument from reg_save_area. (branch to offsetMBB) | |||
20541 | // If not, pull from overflow_area. (branch to overflowMBB) | |||
20542 | // | |||
20543 | // thisMBB | |||
20544 | // | . | |||
20545 | // | . | |||
20546 | // offsetMBB overflowMBB | |||
20547 | // | . | |||
20548 | // | . | |||
20549 | // endMBB | |||
20550 | ||||
20551 | // Registers for the PHI in endMBB | |||
20552 | OffsetDestReg = MRI.createVirtualRegister(AddrRegClass); | |||
20553 | OverflowDestReg = MRI.createVirtualRegister(AddrRegClass); | |||
20554 | ||||
20555 | const BasicBlock *LLVM_BB = MBB->getBasicBlock(); | |||
20556 | MachineFunction *MF = MBB->getParent(); | |||
20557 | overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB); | |||
20558 | offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB); | |||
20559 | endMBB = MF->CreateMachineBasicBlock(LLVM_BB); | |||
20560 | ||||
20561 | MachineFunction::iterator MBBIter = MBB; | |||
20562 | ++MBBIter; | |||
20563 | ||||
20564 | // Insert the new basic blocks | |||
20565 | MF->insert(MBBIter, offsetMBB); | |||
20566 | MF->insert(MBBIter, overflowMBB); | |||
20567 | MF->insert(MBBIter, endMBB); | |||
20568 | ||||
20569 | // Transfer the remainder of MBB and its successor edges to endMBB. | |||
20570 | endMBB->splice(endMBB->begin(), thisMBB, | |||
20571 | std::next(MachineBasicBlock::iterator(MI)), thisMBB->end()); | |||
20572 | endMBB->transferSuccessorsAndUpdatePHIs(thisMBB); | |||
20573 | ||||
20574 | // Make offsetMBB and overflowMBB successors of thisMBB | |||
20575 | thisMBB->addSuccessor(offsetMBB); | |||
20576 | thisMBB->addSuccessor(overflowMBB); | |||
20577 | ||||
20578 | // endMBB is a successor of both offsetMBB and overflowMBB | |||
20579 | offsetMBB->addSuccessor(endMBB); | |||
20580 | overflowMBB->addSuccessor(endMBB); | |||
20581 | ||||
20582 | // Load the offset value into a register | |||
20583 | OffsetReg = MRI.createVirtualRegister(OffsetRegClass); | |||
20584 | BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg) | |||
20585 | .addOperand(Base) | |||
20586 | .addOperand(Scale) | |||
20587 | .addOperand(Index) | |||
20588 | .addDisp(Disp, UseFPOffset ? 4 : 0) | |||
20589 | .addOperand(Segment) | |||
20590 | .setMemRefs(MMOBegin, MMOEnd); | |||
20591 | ||||
20592 | // Check if there is enough room left to pull this argument. | |||
20593 | BuildMI(thisMBB, DL, TII->get(X86::CMP32ri)) | |||
20594 | .addReg(OffsetReg) | |||
20595 | .addImm(MaxOffset + 8 - ArgSizeA8); | |||
20596 | ||||
20597 | // Branch to "overflowMBB" if offset >= max | |||
20598 | // Fall through to "offsetMBB" otherwise | |||
20599 | BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE))) | |||
20600 | .addMBB(overflowMBB); | |||
20601 | } | |||
20602 | ||||
20603 | // In offsetMBB, emit code to use the reg_save_area. | |||
20604 | if (offsetMBB) { | |||
20605 | assert(OffsetReg != 0)((OffsetReg != 0) ? static_cast<void> (0) : __assert_fail ("OffsetReg != 0", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 20605, __PRETTY_FUNCTION__)); | |||
20606 | ||||
20607 | // Read the reg_save_area address. | |||
20608 | unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass); | |||
20609 | BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg) | |||
20610 | .addOperand(Base) | |||
20611 | .addOperand(Scale) | |||
20612 | .addOperand(Index) | |||
20613 | .addDisp(Disp, 16) | |||
20614 | .addOperand(Segment) | |||
20615 | .setMemRefs(MMOBegin, MMOEnd); | |||
20616 | ||||
20617 | // Zero-extend the offset | |||
20618 | unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass); | |||
20619 | BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64) | |||
20620 | .addImm(0) | |||
20621 | .addReg(OffsetReg) | |||
20622 | .addImm(X86::sub_32bit); | |||
20623 | ||||
20624 | // Add the offset to the reg_save_area to get the final address. | |||
20625 | BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg) | |||
20626 | .addReg(OffsetReg64) | |||
20627 | .addReg(RegSaveReg); | |||
20628 | ||||
20629 | // Compute the offset for the next argument | |||
20630 | unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass); | |||
20631 | BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg) | |||
20632 | .addReg(OffsetReg) | |||
20633 | .addImm(UseFPOffset ? 16 : 8); | |||
20634 | ||||
20635 | // Store it back into the va_list. | |||
20636 | BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr)) | |||
20637 | .addOperand(Base) | |||
20638 | .addOperand(Scale) | |||
20639 | .addOperand(Index) | |||
20640 | .addDisp(Disp, UseFPOffset ? 4 : 0) | |||
20641 | .addOperand(Segment) | |||
20642 | .addReg(NextOffsetReg) | |||
20643 | .setMemRefs(MMOBegin, MMOEnd); | |||
20644 | ||||
20645 | // Jump to endMBB | |||
20646 | BuildMI(offsetMBB, DL, TII->get(X86::JMP_1)) | |||
20647 | .addMBB(endMBB); | |||
20648 | } | |||
20649 | ||||
20650 | // | |||
20651 | // Emit code to use overflow area | |||
20652 | // | |||
20653 | ||||
20654 | // Load the overflow_area address into a register. | |||
20655 | unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass); | |||
20656 | BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg) | |||
20657 | .addOperand(Base) | |||
20658 | .addOperand(Scale) | |||
20659 | .addOperand(Index) | |||
20660 | .addDisp(Disp, 8) | |||
20661 | .addOperand(Segment) | |||
20662 | .setMemRefs(MMOBegin, MMOEnd); | |||
20663 | ||||
20664 | // If we need to align it, do so. Otherwise, just copy the address | |||
20665 | // to OverflowDestReg. | |||
20666 | if (NeedsAlign) { | |||
20667 | // Align the overflow address | |||
20668 | assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2")(((Align & (Align-1)) == 0 && "Alignment must be a power of 2" ) ? static_cast<void> (0) : __assert_fail ("(Align & (Align-1)) == 0 && \"Alignment must be a power of 2\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 20668, __PRETTY_FUNCTION__)); | |||
20669 | unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass); | |||
20670 | ||||
20671 | // aligned_addr = (addr + (align-1)) & ~(align-1) | |||
20672 | BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg) | |||
20673 | .addReg(OverflowAddrReg) | |||
20674 | .addImm(Align-1); | |||
20675 | ||||
20676 | BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg) | |||
20677 | .addReg(TmpReg) | |||
20678 | .addImm(~(uint64_t)(Align-1)); | |||
20679 | } else { | |||
20680 | BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg) | |||
20681 | .addReg(OverflowAddrReg); | |||
20682 | } | |||
20683 | ||||
20684 | // Compute the next overflow address after this argument. | |||
20685 | // (the overflow address should be kept 8-byte aligned) | |||
20686 | unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass); | |||
20687 | BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg) | |||
20688 | .addReg(OverflowDestReg) | |||
20689 | .addImm(ArgSizeA8); | |||
20690 | ||||
20691 | // Store the new overflow address. | |||
20692 | BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr)) | |||
20693 | .addOperand(Base) | |||
20694 | .addOperand(Scale) | |||
20695 | .addOperand(Index) | |||
20696 | .addDisp(Disp, 8) | |||
20697 | .addOperand(Segment) | |||
20698 | .addReg(NextAddrReg) | |||
20699 | .setMemRefs(MMOBegin, MMOEnd); | |||
20700 | ||||
20701 | // If we branched, emit the PHI to the front of endMBB. | |||
20702 | if (offsetMBB) { | |||
20703 | BuildMI(*endMBB, endMBB->begin(), DL, | |||
20704 | TII->get(X86::PHI), DestReg) | |||
20705 | .addReg(OffsetDestReg).addMBB(offsetMBB) | |||
20706 | .addReg(OverflowDestReg).addMBB(overflowMBB); | |||
20707 | } | |||
20708 | ||||
20709 | // Erase the pseudo instruction | |||
20710 | MI->eraseFromParent(); | |||
20711 | ||||
20712 | return endMBB; | |||
20713 | } | |||
20714 | ||||
20715 | MachineBasicBlock * | |||
20716 | X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter( | |||
20717 | MachineInstr *MI, | |||
20718 | MachineBasicBlock *MBB) const { | |||
20719 | // Emit code to save XMM registers to the stack. The ABI says that the | |||
20720 | // number of registers to save is given in %al, so it's theoretically | |||
20721 | // possible to do an indirect jump trick to avoid saving all of them, | |||
20722 | // however this code takes a simpler approach and just executes all | |||
20723 | // of the stores if %al is non-zero. It's less code, and it's probably | |||
20724 | // easier on the hardware branch predictor, and stores aren't all that | |||
20725 | // expensive anyway. | |||
20726 | ||||
20727 | // Create the new basic blocks. One block contains all the XMM stores, | |||
20728 | // and one block is the final destination regardless of whether any | |||
20729 | // stores were performed. | |||
20730 | const BasicBlock *LLVM_BB = MBB->getBasicBlock(); | |||
20731 | MachineFunction *F = MBB->getParent(); | |||
20732 | MachineFunction::iterator MBBIter = MBB; | |||
20733 | ++MBBIter; | |||
20734 | MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB); | |||
20735 | MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB); | |||
20736 | F->insert(MBBIter, XMMSaveMBB); | |||
20737 | F->insert(MBBIter, EndMBB); | |||
20738 | ||||
20739 | // Transfer the remainder of MBB and its successor edges to EndMBB. | |||
20740 | EndMBB->splice(EndMBB->begin(), MBB, | |||
20741 | std::next(MachineBasicBlock::iterator(MI)), MBB->end()); | |||
20742 | EndMBB->transferSuccessorsAndUpdatePHIs(MBB); | |||
20743 | ||||
20744 | // The original block will now fall through to the XMM save block. | |||
20745 | MBB->addSuccessor(XMMSaveMBB); | |||
20746 | // The XMMSaveMBB will fall through to the end block. | |||
20747 | XMMSaveMBB->addSuccessor(EndMBB); | |||
20748 | ||||
20749 | // Now add the instructions. | |||
20750 | const TargetInstrInfo *TII = MBB->getParent()->getSubtarget().getInstrInfo(); | |||
20751 | DebugLoc DL = MI->getDebugLoc(); | |||
20752 | ||||
20753 | unsigned CountReg = MI->getOperand(0).getReg(); | |||
20754 | int64_t RegSaveFrameIndex = MI->getOperand(1).getImm(); | |||
20755 | int64_t VarArgsFPOffset = MI->getOperand(2).getImm(); | |||
20756 | ||||
20757 | if (!Subtarget->isTargetWin64()) { | |||
20758 | // If %al is 0, branch around the XMM save block. | |||
20759 | BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg); | |||
20760 | BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB); | |||
20761 | MBB->addSuccessor(EndMBB); | |||
20762 | } | |||
20763 | ||||
20764 | // Make sure the last operand is EFLAGS, which gets clobbered by the branch | |||
20765 | // that was just emitted, but clearly shouldn't be "saved". | |||
20766 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 20769, __PRETTY_FUNCTION__)) | |||
20767 | !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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 20769, __PRETTY_FUNCTION__)) | |||
20768 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 20769, __PRETTY_FUNCTION__)) | |||
20769 | && "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 20769, __PRETTY_FUNCTION__)); | |||
20770 | unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr; | |||
20771 | // In the XMM save block, save all the XMM argument registers. | |||
20772 | for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) { | |||
20773 | int64_t Offset = (i - 3) * 16 + VarArgsFPOffset; | |||
20774 | MachineMemOperand *MMO = | |||
20775 | F->getMachineMemOperand( | |||
20776 | MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset), | |||
20777 | MachineMemOperand::MOStore, | |||
20778 | /*Size=*/16, /*Align=*/16); | |||
20779 | BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc)) | |||
20780 | .addFrameIndex(RegSaveFrameIndex) | |||
20781 | .addImm(/*Scale=*/1) | |||
20782 | .addReg(/*IndexReg=*/0) | |||
20783 | .addImm(/*Disp=*/Offset) | |||
20784 | .addReg(/*Segment=*/0) | |||
20785 | .addReg(MI->getOperand(i).getReg()) | |||
20786 | .addMemOperand(MMO); | |||
20787 | } | |||
20788 | ||||
20789 | MI->eraseFromParent(); // The pseudo instruction is gone now. | |||
20790 | ||||
20791 | return EndMBB; | |||
20792 | } | |||
20793 | ||||
20794 | // The EFLAGS operand of SelectItr might be missing a kill marker | |||
20795 | // because there were multiple uses of EFLAGS, and ISel didn't know | |||
20796 | // which to mark. Figure out whether SelectItr should have had a | |||
20797 | // kill marker, and set it if it should. Returns the correct kill | |||
20798 | // marker value. | |||
20799 | static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr, | |||
20800 | MachineBasicBlock* BB, | |||
20801 | const TargetRegisterInfo* TRI) { | |||
20802 | // Scan forward through BB for a use/def of EFLAGS. | |||
20803 | MachineBasicBlock::iterator miI(std::next(SelectItr)); | |||
20804 | for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) { | |||
20805 | const MachineInstr& mi = *miI; | |||
20806 | if (mi.readsRegister(X86::EFLAGS)) | |||
20807 | return false; | |||
20808 | if (mi.definesRegister(X86::EFLAGS)) | |||
20809 | break; // Should have kill-flag - update below. | |||
20810 | } | |||
20811 | ||||
20812 | // If we hit the end of the block, check whether EFLAGS is live into a | |||
20813 | // successor. | |||
20814 | if (miI == BB->end()) { | |||
20815 | for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(), | |||
20816 | sEnd = BB->succ_end(); | |||
20817 | sItr != sEnd; ++sItr) { | |||
20818 | MachineBasicBlock* succ = *sItr; | |||
20819 | if (succ->isLiveIn(X86::EFLAGS)) | |||
20820 | return false; | |||
20821 | } | |||
20822 | } | |||
20823 | ||||
20824 | // We found a def, or hit the end of the basic block and EFLAGS wasn't live | |||
20825 | // out. SelectMI should have a kill flag on EFLAGS. | |||
20826 | SelectItr->addRegisterKilled(X86::EFLAGS, TRI); | |||
20827 | return true; | |||
20828 | } | |||
20829 | ||||
20830 | MachineBasicBlock * | |||
20831 | X86TargetLowering::EmitLoweredSelect(MachineInstr *MI, | |||
20832 | MachineBasicBlock *BB) const { | |||
20833 | const TargetInstrInfo *TII = BB->getParent()->getSubtarget().getInstrInfo(); | |||
20834 | DebugLoc DL = MI->getDebugLoc(); | |||
20835 | ||||
20836 | // To "insert" a SELECT_CC instruction, we actually have to insert the | |||
20837 | // diamond control-flow pattern. The incoming instruction knows the | |||
20838 | // destination vreg to set, the condition code register to branch on, the | |||
20839 | // true/false values to select between, and a branch opcode to use. | |||
20840 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); | |||
20841 | MachineFunction::iterator It = BB; | |||
20842 | ++It; | |||
20843 | ||||
20844 | // thisMBB: | |||
20845 | // ... | |||
20846 | // TrueVal = ... | |||
20847 | // cmpTY ccX, r1, r2 | |||
20848 | // bCC copy1MBB | |||
20849 | // fallthrough --> copy0MBB | |||
20850 | MachineBasicBlock *thisMBB = BB; | |||
20851 | MachineFunction *F = BB->getParent(); | |||
20852 | MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); | |||
20853 | MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); | |||
20854 | F->insert(It, copy0MBB); | |||
20855 | F->insert(It, sinkMBB); | |||
20856 | ||||
20857 | // If the EFLAGS register isn't dead in the terminator, then claim that it's | |||
20858 | // live into the sink and copy blocks. | |||
20859 | const TargetRegisterInfo *TRI = | |||
20860 | BB->getParent()->getSubtarget().getRegisterInfo(); | |||
20861 | if (!MI->killsRegister(X86::EFLAGS) && | |||
20862 | !checkAndUpdateEFLAGSKill(MI, BB, TRI)) { | |||
20863 | copy0MBB->addLiveIn(X86::EFLAGS); | |||
20864 | sinkMBB->addLiveIn(X86::EFLAGS); | |||
20865 | } | |||
20866 | ||||
20867 | // Transfer the remainder of BB and its successor edges to sinkMBB. | |||
20868 | sinkMBB->splice(sinkMBB->begin(), BB, | |||
20869 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); | |||
20870 | sinkMBB->transferSuccessorsAndUpdatePHIs(BB); | |||
20871 | ||||
20872 | // Add the true and fallthrough blocks as its successors. | |||
20873 | BB->addSuccessor(copy0MBB); | |||
20874 | BB->addSuccessor(sinkMBB); | |||
20875 | ||||
20876 | // Create the conditional branch instruction. | |||
20877 | unsigned Opc = | |||
20878 | X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm()); | |||
20879 | BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB); | |||
20880 | ||||
20881 | // copy0MBB: | |||
20882 | // %FalseValue = ... | |||
20883 | // # fallthrough to sinkMBB | |||
20884 | copy0MBB->addSuccessor(sinkMBB); | |||
20885 | ||||
20886 | // sinkMBB: | |||
20887 | // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] | |||
20888 | // ... | |||
20889 | BuildMI(*sinkMBB, sinkMBB->begin(), DL, | |||
20890 | TII->get(X86::PHI), MI->getOperand(0).getReg()) | |||
20891 | .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB) | |||
20892 | .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB); | |||
20893 | ||||
20894 | MI->eraseFromParent(); // The pseudo instruction is gone now. | |||
20895 | return sinkMBB; | |||
20896 | } | |||
20897 | ||||
20898 | MachineBasicBlock * | |||
20899 | X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI, | |||
20900 | MachineBasicBlock *BB) const { | |||
20901 | MachineFunction *MF = BB->getParent(); | |||
20902 | const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo(); | |||
20903 | DebugLoc DL = MI->getDebugLoc(); | |||
20904 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); | |||
20905 | ||||
20906 | assert(MF->shouldSplitStack())((MF->shouldSplitStack()) ? static_cast<void> (0) : __assert_fail ("MF->shouldSplitStack()", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 20906, __PRETTY_FUNCTION__)); | |||
20907 | ||||
20908 | const bool Is64Bit = Subtarget->is64Bit(); | |||
20909 | const bool IsLP64 = Subtarget->isTarget64BitLP64(); | |||
20910 | ||||
20911 | const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS; | |||
20912 | const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30; | |||
20913 | ||||
20914 | // BB: | |||
20915 | // ... [Till the alloca] | |||
20916 | // If stacklet is not large enough, jump to mallocMBB | |||
20917 | // | |||
20918 | // bumpMBB: | |||
20919 | // Allocate by subtracting from RSP | |||
20920 | // Jump to continueMBB | |||
20921 | // | |||
20922 | // mallocMBB: | |||
20923 | // Allocate by call to runtime | |||
20924 | // | |||
20925 | // continueMBB: | |||
20926 | // ... | |||
20927 | // [rest of original BB] | |||
20928 | // | |||
20929 | ||||
20930 | MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB); | |||
20931 | MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB); | |||
20932 | MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB); | |||
20933 | ||||
20934 | MachineRegisterInfo &MRI = MF->getRegInfo(); | |||
20935 | const TargetRegisterClass *AddrRegClass = | |||
20936 | getRegClassFor(getPointerTy()); | |||
20937 | ||||
20938 | unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass), | |||
20939 | bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass), | |||
20940 | tmpSPVReg = MRI.createVirtualRegister(AddrRegClass), | |||
20941 | SPLimitVReg = MRI.createVirtualRegister(AddrRegClass), | |||
20942 | sizeVReg = MI->getOperand(1).getReg(), | |||
20943 | physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP; | |||
20944 | ||||
20945 | MachineFunction::iterator MBBIter = BB; | |||
20946 | ++MBBIter; | |||
20947 | ||||
20948 | MF->insert(MBBIter, bumpMBB); | |||
20949 | MF->insert(MBBIter, mallocMBB); | |||
20950 | MF->insert(MBBIter, continueMBB); | |||
20951 | ||||
20952 | continueMBB->splice(continueMBB->begin(), BB, | |||
20953 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); | |||
20954 | continueMBB->transferSuccessorsAndUpdatePHIs(BB); | |||
20955 | ||||
20956 | // Add code to the main basic block to check if the stack limit has been hit, | |||
20957 | // and if so, jump to mallocMBB otherwise to bumpMBB. | |||
20958 | BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg); | |||
20959 | BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg) | |||
20960 | .addReg(tmpSPVReg).addReg(sizeVReg); | |||
20961 | BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr)) | |||
20962 | .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg) | |||
20963 | .addReg(SPLimitVReg); | |||
20964 | BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB); | |||
20965 | ||||
20966 | // bumpMBB simply decreases the stack pointer, since we know the current | |||
20967 | // stacklet has enough space. | |||
20968 | BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg) | |||
20969 | .addReg(SPLimitVReg); | |||
20970 | BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg) | |||
20971 | .addReg(SPLimitVReg); | |||
20972 | BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB); | |||
20973 | ||||
20974 | // Calls into a routine in libgcc to allocate more space from the heap. | |||
20975 | const uint32_t *RegMask = MF->getTarget() | |||
20976 | .getSubtargetImpl() | |||
20977 | ->getRegisterInfo() | |||
20978 | ->getCallPreservedMask(CallingConv::C); | |||
20979 | if (IsLP64) { | |||
20980 | BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI) | |||
20981 | .addReg(sizeVReg); | |||
20982 | BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32)) | |||
20983 | .addExternalSymbol("__morestack_allocate_stack_space") | |||
20984 | .addRegMask(RegMask) | |||
20985 | .addReg(X86::RDI, RegState::Implicit) | |||
20986 | .addReg(X86::RAX, RegState::ImplicitDefine); | |||
20987 | } else if (Is64Bit) { | |||
20988 | BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI) | |||
20989 | .addReg(sizeVReg); | |||
20990 | BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32)) | |||
20991 | .addExternalSymbol("__morestack_allocate_stack_space") | |||
20992 | .addRegMask(RegMask) | |||
20993 | .addReg(X86::EDI, RegState::Implicit) | |||
20994 | .addReg(X86::EAX, RegState::ImplicitDefine); | |||
20995 | } else { | |||
20996 | BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg) | |||
20997 | .addImm(12); | |||
20998 | BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg); | |||
20999 | BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32)) | |||
21000 | .addExternalSymbol("__morestack_allocate_stack_space") | |||
21001 | .addRegMask(RegMask) | |||
21002 | .addReg(X86::EAX, RegState::ImplicitDefine); | |||
21003 | } | |||
21004 | ||||
21005 | if (!Is64Bit) | |||
21006 | BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg) | |||
21007 | .addImm(16); | |||
21008 | ||||
21009 | BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg) | |||
21010 | .addReg(IsLP64 ? X86::RAX : X86::EAX); | |||
21011 | BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB); | |||
21012 | ||||
21013 | // Set up the CFG correctly. | |||
21014 | BB->addSuccessor(bumpMBB); | |||
21015 | BB->addSuccessor(mallocMBB); | |||
21016 | mallocMBB->addSuccessor(continueMBB); | |||
21017 | bumpMBB->addSuccessor(continueMBB); | |||
21018 | ||||
21019 | // Take care of the PHI nodes. | |||
21020 | BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI), | |||
21021 | MI->getOperand(0).getReg()) | |||
21022 | .addReg(mallocPtrVReg).addMBB(mallocMBB) | |||
21023 | .addReg(bumpSPPtrVReg).addMBB(bumpMBB); | |||
21024 | ||||
21025 | // Delete the original pseudo instruction. | |||
21026 | MI->eraseFromParent(); | |||
21027 | ||||
21028 | // And we're done. | |||
21029 | return continueMBB; | |||
21030 | } | |||
21031 | ||||
21032 | MachineBasicBlock * | |||
21033 | X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI, | |||
21034 | MachineBasicBlock *BB) const { | |||
21035 | DebugLoc DL = MI->getDebugLoc(); | |||
21036 | ||||
21037 | assert(!Subtarget->isTargetMachO())((!Subtarget->isTargetMachO()) ? static_cast<void> ( 0) : __assert_fail ("!Subtarget->isTargetMachO()", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21037, __PRETTY_FUNCTION__)); | |||
21038 | ||||
21039 | X86FrameLowering::emitStackProbeCall(*BB->getParent(), *BB, MI, DL); | |||
21040 | ||||
21041 | MI->eraseFromParent(); // The pseudo instruction is gone now. | |||
21042 | return BB; | |||
21043 | } | |||
21044 | ||||
21045 | MachineBasicBlock * | |||
21046 | X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI, | |||
21047 | MachineBasicBlock *BB) const { | |||
21048 | // This is pretty easy. We're taking the value that we received from | |||
21049 | // our load from the relocation, sticking it in either RDI (x86-64) | |||
21050 | // or EAX and doing an indirect call. The return value will then | |||
21051 | // be in the normal return register. | |||
21052 | MachineFunction *F = BB->getParent(); | |||
21053 | const X86InstrInfo *TII = | |||
21054 | static_cast<const X86InstrInfo *>(F->getSubtarget().getInstrInfo()); | |||
21055 | DebugLoc DL = MI->getDebugLoc(); | |||
21056 | ||||
21057 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21057, __PRETTY_FUNCTION__)); | |||
21058 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21058, __PRETTY_FUNCTION__)); | |||
21059 | ||||
21060 | // Get a register mask for the lowered call. | |||
21061 | // FIXME: The 32-bit calls have non-standard calling conventions. Use a | |||
21062 | // proper register mask. | |||
21063 | const uint32_t *RegMask = F->getTarget() | |||
21064 | .getSubtargetImpl() | |||
21065 | ->getRegisterInfo() | |||
21066 | ->getCallPreservedMask(CallingConv::C); | |||
21067 | if (Subtarget->is64Bit()) { | |||
21068 | MachineInstrBuilder MIB = BuildMI(*BB, MI, DL, | |||
21069 | TII->get(X86::MOV64rm), X86::RDI) | |||
21070 | .addReg(X86::RIP) | |||
21071 | .addImm(0).addReg(0) | |||
21072 | .addGlobalAddress(MI->getOperand(3).getGlobal(), 0, | |||
21073 | MI->getOperand(3).getTargetFlags()) | |||
21074 | .addReg(0); | |||
21075 | MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m)); | |||
21076 | addDirectMem(MIB, X86::RDI); | |||
21077 | MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask); | |||
21078 | } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) { | |||
21079 | MachineInstrBuilder MIB = BuildMI(*BB, MI, DL, | |||
21080 | TII->get(X86::MOV32rm), X86::EAX) | |||
21081 | .addReg(0) | |||
21082 | .addImm(0).addReg(0) | |||
21083 | .addGlobalAddress(MI->getOperand(3).getGlobal(), 0, | |||
21084 | MI->getOperand(3).getTargetFlags()) | |||
21085 | .addReg(0); | |||
21086 | MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m)); | |||
21087 | addDirectMem(MIB, X86::EAX); | |||
21088 | MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask); | |||
21089 | } else { | |||
21090 | MachineInstrBuilder MIB = BuildMI(*BB, MI, DL, | |||
21091 | TII->get(X86::MOV32rm), X86::EAX) | |||
21092 | .addReg(TII->getGlobalBaseReg(F)) | |||
21093 | .addImm(0).addReg(0) | |||
21094 | .addGlobalAddress(MI->getOperand(3).getGlobal(), 0, | |||
21095 | MI->getOperand(3).getTargetFlags()) | |||
21096 | .addReg(0); | |||
21097 | MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m)); | |||
21098 | addDirectMem(MIB, X86::EAX); | |||
21099 | MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask); | |||
21100 | } | |||
21101 | ||||
21102 | MI->eraseFromParent(); // The pseudo instruction is gone now. | |||
21103 | return BB; | |||
21104 | } | |||
21105 | ||||
21106 | MachineBasicBlock * | |||
21107 | X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI, | |||
21108 | MachineBasicBlock *MBB) const { | |||
21109 | DebugLoc DL = MI->getDebugLoc(); | |||
21110 | MachineFunction *MF = MBB->getParent(); | |||
21111 | const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo(); | |||
21112 | MachineRegisterInfo &MRI = MF->getRegInfo(); | |||
21113 | ||||
21114 | const BasicBlock *BB = MBB->getBasicBlock(); | |||
21115 | MachineFunction::iterator I = MBB; | |||
21116 | ++I; | |||
21117 | ||||
21118 | // Memory Reference | |||
21119 | MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin(); | |||
21120 | MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end(); | |||
21121 | ||||
21122 | unsigned DstReg; | |||
21123 | unsigned MemOpndSlot = 0; | |||
21124 | ||||
21125 | unsigned CurOp = 0; | |||
21126 | ||||
21127 | DstReg = MI->getOperand(CurOp++).getReg(); | |||
21128 | const TargetRegisterClass *RC = MRI.getRegClass(DstReg); | |||
21129 | assert(RC->hasType(MVT::i32) && "Invalid destination!")((RC->hasType(MVT::i32) && "Invalid destination!") ? static_cast<void> (0) : __assert_fail ("RC->hasType(MVT::i32) && \"Invalid destination!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21129, __PRETTY_FUNCTION__)); | |||
21130 | unsigned mainDstReg = MRI.createVirtualRegister(RC); | |||
21131 | unsigned restoreDstReg = MRI.createVirtualRegister(RC); | |||
21132 | ||||
21133 | MemOpndSlot = CurOp; | |||
21134 | ||||
21135 | MVT PVT = getPointerTy(); | |||
21136 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21137, __PRETTY_FUNCTION__)) | |||
21137 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21137, __PRETTY_FUNCTION__)); | |||
21138 | ||||
21139 | // For v = setjmp(buf), we generate | |||
21140 | // | |||
21141 | // thisMBB: | |||
21142 | // buf[LabelOffset] = restoreMBB | |||
21143 | // SjLjSetup restoreMBB | |||
21144 | // | |||
21145 | // mainMBB: | |||
21146 | // v_main = 0 | |||
21147 | // | |||
21148 | // sinkMBB: | |||
21149 | // v = phi(main, restore) | |||
21150 | // | |||
21151 | // restoreMBB: | |||
21152 | // if base pointer being used, load it from frame | |||
21153 | // v_restore = 1 | |||
21154 | ||||
21155 | MachineBasicBlock *thisMBB = MBB; | |||
21156 | MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB); | |||
21157 | MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB); | |||
21158 | MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB); | |||
21159 | MF->insert(I, mainMBB); | |||
21160 | MF->insert(I, sinkMBB); | |||
21161 | MF->push_back(restoreMBB); | |||
21162 | ||||
21163 | MachineInstrBuilder MIB; | |||
21164 | ||||
21165 | // Transfer the remainder of BB and its successor edges to sinkMBB. | |||
21166 | sinkMBB->splice(sinkMBB->begin(), MBB, | |||
21167 | std::next(MachineBasicBlock::iterator(MI)), MBB->end()); | |||
21168 | sinkMBB->transferSuccessorsAndUpdatePHIs(MBB); | |||
21169 | ||||
21170 | // thisMBB: | |||
21171 | unsigned PtrStoreOpc = 0; | |||
21172 | unsigned LabelReg = 0; | |||
21173 | const int64_t LabelOffset = 1 * PVT.getStoreSize(); | |||
21174 | Reloc::Model RM = MF->getTarget().getRelocationModel(); | |||
21175 | bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) && | |||
21176 | (RM == Reloc::Static || RM == Reloc::DynamicNoPIC); | |||
21177 | ||||
21178 | // Prepare IP either in reg or imm. | |||
21179 | if (!UseImmLabel) { | |||
21180 | PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr; | |||
21181 | const TargetRegisterClass *PtrRC = getRegClassFor(PVT); | |||
21182 | LabelReg = MRI.createVirtualRegister(PtrRC); | |||
21183 | if (Subtarget->is64Bit()) { | |||
21184 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg) | |||
21185 | .addReg(X86::RIP) | |||
21186 | .addImm(0) | |||
21187 | .addReg(0) | |||
21188 | .addMBB(restoreMBB) | |||
21189 | .addReg(0); | |||
21190 | } else { | |||
21191 | const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII); | |||
21192 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg) | |||
21193 | .addReg(XII->getGlobalBaseReg(MF)) | |||
21194 | .addImm(0) | |||
21195 | .addReg(0) | |||
21196 | .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference()) | |||
21197 | .addReg(0); | |||
21198 | } | |||
21199 | } else | |||
21200 | PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi; | |||
21201 | // Store IP | |||
21202 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc)); | |||
21203 | for (unsigned i = 0; i < X86::AddrNumOperands; ++i) { | |||
21204 | if (i == X86::AddrDisp) | |||
21205 | MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset); | |||
21206 | else | |||
21207 | MIB.addOperand(MI->getOperand(MemOpndSlot + i)); | |||
21208 | } | |||
21209 | if (!UseImmLabel) | |||
21210 | MIB.addReg(LabelReg); | |||
21211 | else | |||
21212 | MIB.addMBB(restoreMBB); | |||
21213 | MIB.setMemRefs(MMOBegin, MMOEnd); | |||
21214 | // Setup | |||
21215 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup)) | |||
21216 | .addMBB(restoreMBB); | |||
21217 | ||||
21218 | const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>( | |||
21219 | MF->getSubtarget().getRegisterInfo()); | |||
21220 | MIB.addRegMask(RegInfo->getNoPreservedMask()); | |||
21221 | thisMBB->addSuccessor(mainMBB); | |||
21222 | thisMBB->addSuccessor(restoreMBB); | |||
21223 | ||||
21224 | // mainMBB: | |||
21225 | // EAX = 0 | |||
21226 | BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg); | |||
21227 | mainMBB->addSuccessor(sinkMBB); | |||
21228 | ||||
21229 | // sinkMBB: | |||
21230 | BuildMI(*sinkMBB, sinkMBB->begin(), DL, | |||
21231 | TII->get(X86::PHI), DstReg) | |||
21232 | .addReg(mainDstReg).addMBB(mainMBB) | |||
21233 | .addReg(restoreDstReg).addMBB(restoreMBB); | |||
21234 | ||||
21235 | // restoreMBB: | |||
21236 | if (RegInfo->hasBasePointer(*MF)) { | |||
21237 | const X86Subtarget &STI = MF->getTarget().getSubtarget<X86Subtarget>(); | |||
21238 | const bool Uses64BitFramePtr = STI.isTarget64BitLP64() || STI.isTargetNaCl64(); | |||
21239 | X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>(); | |||
21240 | X86FI->setRestoreBasePointer(MF); | |||
21241 | unsigned FramePtr = RegInfo->getFrameRegister(*MF); | |||
21242 | unsigned BasePtr = RegInfo->getBaseRegister(); | |||
21243 | unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm; | |||
21244 | addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr), | |||
21245 | FramePtr, true, X86FI->getRestoreBasePointerOffset()) | |||
21246 | .setMIFlag(MachineInstr::FrameSetup); | |||
21247 | } | |||
21248 | BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1); | |||
21249 | BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB); | |||
21250 | restoreMBB->addSuccessor(sinkMBB); | |||
21251 | ||||
21252 | MI->eraseFromParent(); | |||
21253 | return sinkMBB; | |||
21254 | } | |||
21255 | ||||
21256 | MachineBasicBlock * | |||
21257 | X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI, | |||
21258 | MachineBasicBlock *MBB) const { | |||
21259 | DebugLoc DL = MI->getDebugLoc(); | |||
21260 | MachineFunction *MF = MBB->getParent(); | |||
21261 | const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo(); | |||
21262 | MachineRegisterInfo &MRI = MF->getRegInfo(); | |||
21263 | ||||
21264 | // Memory Reference | |||
21265 | MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin(); | |||
21266 | MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end(); | |||
21267 | ||||
21268 | MVT PVT = getPointerTy(); | |||
21269 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21270, __PRETTY_FUNCTION__)) | |||
21270 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21270, __PRETTY_FUNCTION__)); | |||
21271 | ||||
21272 | const TargetRegisterClass *RC = | |||
21273 | (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass; | |||
21274 | unsigned Tmp = MRI.createVirtualRegister(RC); | |||
21275 | // Since FP is only updated here but NOT referenced, it's treated as GPR. | |||
21276 | const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>( | |||
21277 | MF->getSubtarget().getRegisterInfo()); | |||
21278 | unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP; | |||
21279 | unsigned SP = RegInfo->getStackRegister(); | |||
21280 | ||||
21281 | MachineInstrBuilder MIB; | |||
21282 | ||||
21283 | const int64_t LabelOffset = 1 * PVT.getStoreSize(); | |||
21284 | const int64_t SPOffset = 2 * PVT.getStoreSize(); | |||
21285 | ||||
21286 | unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm; | |||
21287 | unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r; | |||
21288 | ||||
21289 | // Reload FP | |||
21290 | MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP); | |||
21291 | for (unsigned i = 0; i < X86::AddrNumOperands; ++i) | |||
21292 | MIB.addOperand(MI->getOperand(i)); | |||
21293 | MIB.setMemRefs(MMOBegin, MMOEnd); | |||
21294 | // Reload IP | |||
21295 | MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp); | |||
21296 | for (unsigned i = 0; i < X86::AddrNumOperands; ++i) { | |||
21297 | if (i == X86::AddrDisp) | |||
21298 | MIB.addDisp(MI->getOperand(i), LabelOffset); | |||
21299 | else | |||
21300 | MIB.addOperand(MI->getOperand(i)); | |||
21301 | } | |||
21302 | MIB.setMemRefs(MMOBegin, MMOEnd); | |||
21303 | // Reload SP | |||
21304 | MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP); | |||
21305 | for (unsigned i = 0; i < X86::AddrNumOperands; ++i) { | |||
21306 | if (i == X86::AddrDisp) | |||
21307 | MIB.addDisp(MI->getOperand(i), SPOffset); | |||
21308 | else | |||
21309 | MIB.addOperand(MI->getOperand(i)); | |||
21310 | } | |||
21311 | MIB.setMemRefs(MMOBegin, MMOEnd); | |||
21312 | // Jump | |||
21313 | BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp); | |||
21314 | ||||
21315 | MI->eraseFromParent(); | |||
21316 | return MBB; | |||
21317 | } | |||
21318 | ||||
21319 | // Replace 213-type (isel default) FMA3 instructions with 231-type for | |||
21320 | // accumulator loops. Writing back to the accumulator allows the coalescer | |||
21321 | // to remove extra copies in the loop. | |||
21322 | MachineBasicBlock * | |||
21323 | X86TargetLowering::emitFMA3Instr(MachineInstr *MI, | |||
21324 | MachineBasicBlock *MBB) const { | |||
21325 | MachineOperand &AddendOp = MI->getOperand(3); | |||
21326 | ||||
21327 | // Bail out early if the addend isn't a register - we can't switch these. | |||
21328 | if (!AddendOp.isReg()) | |||
21329 | return MBB; | |||
21330 | ||||
21331 | MachineFunction &MF = *MBB->getParent(); | |||
21332 | MachineRegisterInfo &MRI = MF.getRegInfo(); | |||
21333 | ||||
21334 | // Check whether the addend is defined by a PHI: | |||
21335 | assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?")((MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?" ) ? static_cast<void> (0) : __assert_fail ("MRI.hasOneDef(AddendOp.getReg()) && \"Multiple defs in SSA?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21335, __PRETTY_FUNCTION__)); | |||
21336 | MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg()); | |||
21337 | if (!AddendDef.isPHI()) | |||
21338 | return MBB; | |||
21339 | ||||
21340 | // Look for the following pattern: | |||
21341 | // loop: | |||
21342 | // %addend = phi [%entry, 0], [%loop, %result] | |||
21343 | // ... | |||
21344 | // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend | |||
21345 | ||||
21346 | // Replace with: | |||
21347 | // loop: | |||
21348 | // %addend = phi [%entry, 0], [%loop, %result] | |||
21349 | // ... | |||
21350 | // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2 | |||
21351 | ||||
21352 | for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) { | |||
21353 | assert(AddendDef.getOperand(i).isReg())((AddendDef.getOperand(i).isReg()) ? static_cast<void> ( 0) : __assert_fail ("AddendDef.getOperand(i).isReg()", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21353, __PRETTY_FUNCTION__)); | |||
21354 | MachineOperand PHISrcOp = AddendDef.getOperand(i); | |||
21355 | MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg()); | |||
21356 | if (&PHISrcInst == MI) { | |||
21357 | // Found a matching instruction. | |||
21358 | unsigned NewFMAOpc = 0; | |||
21359 | switch (MI->getOpcode()) { | |||
21360 | case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break; | |||
21361 | case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break; | |||
21362 | case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break; | |||
21363 | case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break; | |||
21364 | case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break; | |||
21365 | case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break; | |||
21366 | case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break; | |||
21367 | case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break; | |||
21368 | case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break; | |||
21369 | case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break; | |||
21370 | case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break; | |||
21371 | case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break; | |||
21372 | case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break; | |||
21373 | case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break; | |||
21374 | case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break; | |||
21375 | case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break; | |||
21376 | case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break; | |||
21377 | case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break; | |||
21378 | case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break; | |||
21379 | case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break; | |||
21380 | ||||
21381 | case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break; | |||
21382 | case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break; | |||
21383 | case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break; | |||
21384 | case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break; | |||
21385 | case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break; | |||
21386 | case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break; | |||
21387 | case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break; | |||
21388 | case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break; | |||
21389 | case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break; | |||
21390 | case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break; | |||
21391 | case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break; | |||
21392 | case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break; | |||
21393 | default: llvm_unreachable("Unrecognized FMA variant.")::llvm::llvm_unreachable_internal("Unrecognized FMA variant." , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21393); | |||
21394 | } | |||
21395 | ||||
21396 | const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo(); | |||
21397 | MachineInstrBuilder MIB = | |||
21398 | BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc)) | |||
21399 | .addOperand(MI->getOperand(0)) | |||
21400 | .addOperand(MI->getOperand(3)) | |||
21401 | .addOperand(MI->getOperand(2)) | |||
21402 | .addOperand(MI->getOperand(1)); | |||
21403 | MBB->insert(MachineBasicBlock::iterator(MI), MIB); | |||
21404 | MI->eraseFromParent(); | |||
21405 | } | |||
21406 | } | |||
21407 | ||||
21408 | return MBB; | |||
21409 | } | |||
21410 | ||||
21411 | MachineBasicBlock * | |||
21412 | X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, | |||
21413 | MachineBasicBlock *BB) const { | |||
21414 | switch (MI->getOpcode()) { | |||
21415 | default: llvm_unreachable("Unexpected instr type to insert")::llvm::llvm_unreachable_internal("Unexpected instr type to insert" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21415); | |||
21416 | case X86::TAILJMPd64: | |||
21417 | case X86::TAILJMPr64: | |||
21418 | case X86::TAILJMPm64: | |||
21419 | case X86::TAILJMPd64_REX: | |||
21420 | case X86::TAILJMPr64_REX: | |||
21421 | case X86::TAILJMPm64_REX: | |||
21422 | llvm_unreachable("TAILJMP64 would not be touched here.")::llvm::llvm_unreachable_internal("TAILJMP64 would not be touched here." , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21422); | |||
21423 | case X86::TCRETURNdi64: | |||
21424 | case X86::TCRETURNri64: | |||
21425 | case X86::TCRETURNmi64: | |||
21426 | return BB; | |||
21427 | case X86::WIN_ALLOCA: | |||
21428 | return EmitLoweredWinAlloca(MI, BB); | |||
21429 | case X86::SEG_ALLOCA_32: | |||
21430 | case X86::SEG_ALLOCA_64: | |||
21431 | return EmitLoweredSegAlloca(MI, BB); | |||
21432 | case X86::TLSCall_32: | |||
21433 | case X86::TLSCall_64: | |||
21434 | return EmitLoweredTLSCall(MI, BB); | |||
21435 | case X86::CMOV_GR8: | |||
21436 | case X86::CMOV_FR32: | |||
21437 | case X86::CMOV_FR64: | |||
21438 | case X86::CMOV_V4F32: | |||
21439 | case X86::CMOV_V2F64: | |||
21440 | case X86::CMOV_V2I64: | |||
21441 | case X86::CMOV_V8F32: | |||
21442 | case X86::CMOV_V4F64: | |||
21443 | case X86::CMOV_V4I64: | |||
21444 | case X86::CMOV_V16F32: | |||
21445 | case X86::CMOV_V8F64: | |||
21446 | case X86::CMOV_V8I64: | |||
21447 | case X86::CMOV_GR16: | |||
21448 | case X86::CMOV_GR32: | |||
21449 | case X86::CMOV_RFP32: | |||
21450 | case X86::CMOV_RFP64: | |||
21451 | case X86::CMOV_RFP80: | |||
21452 | return EmitLoweredSelect(MI, BB); | |||
21453 | ||||
21454 | case X86::FP32_TO_INT16_IN_MEM: | |||
21455 | case X86::FP32_TO_INT32_IN_MEM: | |||
21456 | case X86::FP32_TO_INT64_IN_MEM: | |||
21457 | case X86::FP64_TO_INT16_IN_MEM: | |||
21458 | case X86::FP64_TO_INT32_IN_MEM: | |||
21459 | case X86::FP64_TO_INT64_IN_MEM: | |||
21460 | case X86::FP80_TO_INT16_IN_MEM: | |||
21461 | case X86::FP80_TO_INT32_IN_MEM: | |||
21462 | case X86::FP80_TO_INT64_IN_MEM: { | |||
21463 | MachineFunction *F = BB->getParent(); | |||
21464 | const TargetInstrInfo *TII = F->getSubtarget().getInstrInfo(); | |||
21465 | DebugLoc DL = MI->getDebugLoc(); | |||
21466 | ||||
21467 | // Change the floating point control register to use "round towards zero" | |||
21468 | // mode when truncating to an integer value. | |||
21469 | int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false); | |||
21470 | addFrameReference(BuildMI(*BB, MI, DL, | |||
21471 | TII->get(X86::FNSTCW16m)), CWFrameIdx); | |||
21472 | ||||
21473 | // Load the old value of the high byte of the control word... | |||
21474 | unsigned OldCW = | |||
21475 | F->getRegInfo().createVirtualRegister(&X86::GR16RegClass); | |||
21476 | addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW), | |||
21477 | CWFrameIdx); | |||
21478 | ||||
21479 | // Set the high part to be round to zero... | |||
21480 | addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx) | |||
21481 | .addImm(0xC7F); | |||
21482 | ||||
21483 | // Reload the modified control word now... | |||
21484 | addFrameReference(BuildMI(*BB, MI, DL, | |||
21485 | TII->get(X86::FLDCW16m)), CWFrameIdx); | |||
21486 | ||||
21487 | // Restore the memory image of control word to original value | |||
21488 | addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx) | |||
21489 | .addReg(OldCW); | |||
21490 | ||||
21491 | // Get the X86 opcode to use. | |||
21492 | unsigned Opc; | |||
21493 | switch (MI->getOpcode()) { | |||
21494 | default: llvm_unreachable("illegal opcode!")::llvm::llvm_unreachable_internal("illegal opcode!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21494); | |||
21495 | case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break; | |||
21496 | case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break; | |||
21497 | case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break; | |||
21498 | case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break; | |||
21499 | case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break; | |||
21500 | case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break; | |||
21501 | case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break; | |||
21502 | case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break; | |||
21503 | case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break; | |||
21504 | } | |||
21505 | ||||
21506 | X86AddressMode AM; | |||
21507 | MachineOperand &Op = MI->getOperand(0); | |||
21508 | if (Op.isReg()) { | |||
21509 | AM.BaseType = X86AddressMode::RegBase; | |||
21510 | AM.Base.Reg = Op.getReg(); | |||
21511 | } else { | |||
21512 | AM.BaseType = X86AddressMode::FrameIndexBase; | |||
21513 | AM.Base.FrameIndex = Op.getIndex(); | |||
21514 | } | |||
21515 | Op = MI->getOperand(1); | |||
21516 | if (Op.isImm()) | |||
21517 | AM.Scale = Op.getImm(); | |||
21518 | Op = MI->getOperand(2); | |||
21519 | if (Op.isImm()) | |||
21520 | AM.IndexReg = Op.getImm(); | |||
21521 | Op = MI->getOperand(3); | |||
21522 | if (Op.isGlobal()) { | |||
21523 | AM.GV = Op.getGlobal(); | |||
21524 | } else { | |||
21525 | AM.Disp = Op.getImm(); | |||
21526 | } | |||
21527 | addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM) | |||
21528 | .addReg(MI->getOperand(X86::AddrNumOperands).getReg()); | |||
21529 | ||||
21530 | // Reload the original control word now. | |||
21531 | addFrameReference(BuildMI(*BB, MI, DL, | |||
21532 | TII->get(X86::FLDCW16m)), CWFrameIdx); | |||
21533 | ||||
21534 | MI->eraseFromParent(); // The pseudo instruction is gone now. | |||
21535 | return BB; | |||
21536 | } | |||
21537 | // String/text processing lowering. | |||
21538 | case X86::PCMPISTRM128REG: | |||
21539 | case X86::VPCMPISTRM128REG: | |||
21540 | case X86::PCMPISTRM128MEM: | |||
21541 | case X86::VPCMPISTRM128MEM: | |||
21542 | case X86::PCMPESTRM128REG: | |||
21543 | case X86::VPCMPESTRM128REG: | |||
21544 | case X86::PCMPESTRM128MEM: | |||
21545 | case X86::VPCMPESTRM128MEM: | |||
21546 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21547, __PRETTY_FUNCTION__)) | |||
21547 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21547, __PRETTY_FUNCTION__)); | |||
21548 | return EmitPCMPSTRM(MI, BB, BB->getParent()->getSubtarget().getInstrInfo()); | |||
21549 | ||||
21550 | // String/text processing lowering. | |||
21551 | case X86::PCMPISTRIREG: | |||
21552 | case X86::VPCMPISTRIREG: | |||
21553 | case X86::PCMPISTRIMEM: | |||
21554 | case X86::VPCMPISTRIMEM: | |||
21555 | case X86::PCMPESTRIREG: | |||
21556 | case X86::VPCMPESTRIREG: | |||
21557 | case X86::PCMPESTRIMEM: | |||
21558 | case X86::VPCMPESTRIMEM: | |||
21559 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21560, __PRETTY_FUNCTION__)) | |||
21560 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21560, __PRETTY_FUNCTION__)); | |||
21561 | return EmitPCMPSTRI(MI, BB, BB->getParent()->getSubtarget().getInstrInfo()); | |||
21562 | ||||
21563 | // Thread synchronization. | |||
21564 | case X86::MONITOR: | |||
21565 | return EmitMonitor(MI, BB, BB->getParent()->getSubtarget().getInstrInfo(), | |||
21566 | Subtarget); | |||
21567 | ||||
21568 | // xbegin | |||
21569 | case X86::XBEGIN: | |||
21570 | return EmitXBegin(MI, BB, BB->getParent()->getSubtarget().getInstrInfo()); | |||
21571 | ||||
21572 | case X86::VASTART_SAVE_XMM_REGS: | |||
21573 | return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB); | |||
21574 | ||||
21575 | case X86::VAARG_64: | |||
21576 | return EmitVAARG64WithCustomInserter(MI, BB); | |||
21577 | ||||
21578 | case X86::EH_SjLj_SetJmp32: | |||
21579 | case X86::EH_SjLj_SetJmp64: | |||
21580 | return emitEHSjLjSetJmp(MI, BB); | |||
21581 | ||||
21582 | case X86::EH_SjLj_LongJmp32: | |||
21583 | case X86::EH_SjLj_LongJmp64: | |||
21584 | return emitEHSjLjLongJmp(MI, BB); | |||
21585 | ||||
21586 | case TargetOpcode::STATEPOINT: | |||
21587 | // As an implementation detail, STATEPOINT shares the STACKMAP format at | |||
21588 | // this point in the process. We diverge later. | |||
21589 | return emitPatchPoint(MI, BB); | |||
21590 | ||||
21591 | case TargetOpcode::STACKMAP: | |||
21592 | case TargetOpcode::PATCHPOINT: | |||
21593 | return emitPatchPoint(MI, BB); | |||
21594 | ||||
21595 | case X86::VFMADDPDr213r: | |||
21596 | case X86::VFMADDPSr213r: | |||
21597 | case X86::VFMADDSDr213r: | |||
21598 | case X86::VFMADDSSr213r: | |||
21599 | case X86::VFMSUBPDr213r: | |||
21600 | case X86::VFMSUBPSr213r: | |||
21601 | case X86::VFMSUBSDr213r: | |||
21602 | case X86::VFMSUBSSr213r: | |||
21603 | case X86::VFNMADDPDr213r: | |||
21604 | case X86::VFNMADDPSr213r: | |||
21605 | case X86::VFNMADDSDr213r: | |||
21606 | case X86::VFNMADDSSr213r: | |||
21607 | case X86::VFNMSUBPDr213r: | |||
21608 | case X86::VFNMSUBPSr213r: | |||
21609 | case X86::VFNMSUBSDr213r: | |||
21610 | case X86::VFNMSUBSSr213r: | |||
21611 | case X86::VFMADDSUBPDr213r: | |||
21612 | case X86::VFMADDSUBPSr213r: | |||
21613 | case X86::VFMSUBADDPDr213r: | |||
21614 | case X86::VFMSUBADDPSr213r: | |||
21615 | case X86::VFMADDPDr213rY: | |||
21616 | case X86::VFMADDPSr213rY: | |||
21617 | case X86::VFMSUBPDr213rY: | |||
21618 | case X86::VFMSUBPSr213rY: | |||
21619 | case X86::VFNMADDPDr213rY: | |||
21620 | case X86::VFNMADDPSr213rY: | |||
21621 | case X86::VFNMSUBPDr213rY: | |||
21622 | case X86::VFNMSUBPSr213rY: | |||
21623 | case X86::VFMADDSUBPDr213rY: | |||
21624 | case X86::VFMADDSUBPSr213rY: | |||
21625 | case X86::VFMSUBADDPDr213rY: | |||
21626 | case X86::VFMSUBADDPSr213rY: | |||
21627 | return emitFMA3Instr(MI, BB); | |||
21628 | } | |||
21629 | } | |||
21630 | ||||
21631 | //===----------------------------------------------------------------------===// | |||
21632 | // X86 Optimization Hooks | |||
21633 | //===----------------------------------------------------------------------===// | |||
21634 | ||||
21635 | void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op, | |||
21636 | APInt &KnownZero, | |||
21637 | APInt &KnownOne, | |||
21638 | const SelectionDAG &DAG, | |||
21639 | unsigned Depth) const { | |||
21640 | unsigned BitWidth = KnownZero.getBitWidth(); | |||
21641 | unsigned Opc = Op.getOpcode(); | |||
21642 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21647, __PRETTY_FUNCTION__)) | |||
21643 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21647, __PRETTY_FUNCTION__)) | |||
21644 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21647, __PRETTY_FUNCTION__)) | |||
21645 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21647, __PRETTY_FUNCTION__)) | |||
21646 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21647, __PRETTY_FUNCTION__)) | |||
21647 | " 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21647, __PRETTY_FUNCTION__)); | |||
21648 | ||||
21649 | KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything. | |||
21650 | switch (Opc) { | |||
21651 | default: break; | |||
21652 | case X86ISD::ADD: | |||
21653 | case X86ISD::SUB: | |||
21654 | case X86ISD::ADC: | |||
21655 | case X86ISD::SBB: | |||
21656 | case X86ISD::SMUL: | |||
21657 | case X86ISD::UMUL: | |||
21658 | case X86ISD::INC: | |||
21659 | case X86ISD::DEC: | |||
21660 | case X86ISD::OR: | |||
21661 | case X86ISD::XOR: | |||
21662 | case X86ISD::AND: | |||
21663 | // These nodes' second result is a boolean. | |||
21664 | if (Op.getResNo() == 0) | |||
21665 | break; | |||
21666 | // Fallthrough | |||
21667 | case X86ISD::SETCC: | |||
21668 | KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1); | |||
21669 | break; | |||
21670 | case ISD::INTRINSIC_WO_CHAIN: { | |||
21671 | unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); | |||
21672 | unsigned NumLoBits = 0; | |||
21673 | switch (IntId) { | |||
21674 | default: break; | |||
21675 | case Intrinsic::x86_sse_movmsk_ps: | |||
21676 | case Intrinsic::x86_avx_movmsk_ps_256: | |||
21677 | case Intrinsic::x86_sse2_movmsk_pd: | |||
21678 | case Intrinsic::x86_avx_movmsk_pd_256: | |||
21679 | case Intrinsic::x86_mmx_pmovmskb: | |||
21680 | case Intrinsic::x86_sse2_pmovmskb_128: | |||
21681 | case Intrinsic::x86_avx2_pmovmskb: { | |||
21682 | // High bits of movmskp{s|d}, pmovmskb are known zero. | |||
21683 | switch (IntId) { | |||
21684 | default: llvm_unreachable("Impossible intrinsic")::llvm::llvm_unreachable_internal("Impossible intrinsic", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21684); // Can't reach here. | |||
21685 | case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break; | |||
21686 | case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break; | |||
21687 | case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break; | |||
21688 | case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break; | |||
21689 | case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break; | |||
21690 | case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break; | |||
21691 | case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break; | |||
21692 | } | |||
21693 | KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits); | |||
21694 | break; | |||
21695 | } | |||
21696 | } | |||
21697 | break; | |||
21698 | } | |||
21699 | } | |||
21700 | } | |||
21701 | ||||
21702 | unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode( | |||
21703 | SDValue Op, | |||
21704 | const SelectionDAG &, | |||
21705 | unsigned Depth) const { | |||
21706 | // SETCC_CARRY sets the dest to ~0 for true or 0 for false. | |||
21707 | if (Op.getOpcode() == X86ISD::SETCC_CARRY) | |||
21708 | return Op.getValueType().getScalarType().getSizeInBits(); | |||
21709 | ||||
21710 | // Fallback case. | |||
21711 | return 1; | |||
21712 | } | |||
21713 | ||||
21714 | /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the | |||
21715 | /// node is a GlobalAddress + offset. | |||
21716 | bool X86TargetLowering::isGAPlusOffset(SDNode *N, | |||
21717 | const GlobalValue* &GA, | |||
21718 | int64_t &Offset) const { | |||
21719 | if (N->getOpcode() == X86ISD::Wrapper) { | |||
21720 | if (isa<GlobalAddressSDNode>(N->getOperand(0))) { | |||
21721 | GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal(); | |||
21722 | Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset(); | |||
21723 | return true; | |||
21724 | } | |||
21725 | } | |||
21726 | return TargetLowering::isGAPlusOffset(N, GA, Offset); | |||
21727 | } | |||
21728 | ||||
21729 | /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the | |||
21730 | /// same as extracting the high 128-bit part of 256-bit vector and then | |||
21731 | /// inserting the result into the low part of a new 256-bit vector | |||
21732 | static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) { | |||
21733 | EVT VT = SVOp->getValueType(0); | |||
21734 | unsigned NumElems = VT.getVectorNumElements(); | |||
21735 | ||||
21736 | // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u> | |||
21737 | for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j) | |||
21738 | if (!isUndefOrEqual(SVOp->getMaskElt(i), j) || | |||
21739 | SVOp->getMaskElt(j) >= 0) | |||
21740 | return false; | |||
21741 | ||||
21742 | return true; | |||
21743 | } | |||
21744 | ||||
21745 | /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the | |||
21746 | /// same as extracting the low 128-bit part of 256-bit vector and then | |||
21747 | /// inserting the result into the high part of a new 256-bit vector | |||
21748 | static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) { | |||
21749 | EVT VT = SVOp->getValueType(0); | |||
21750 | unsigned NumElems = VT.getVectorNumElements(); | |||
21751 | ||||
21752 | // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1> | |||
21753 | for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j) | |||
21754 | if (!isUndefOrEqual(SVOp->getMaskElt(i), j) || | |||
21755 | SVOp->getMaskElt(j) >= 0) | |||
21756 | return false; | |||
21757 | ||||
21758 | return true; | |||
21759 | } | |||
21760 | ||||
21761 | /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors. | |||
21762 | static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG, | |||
21763 | TargetLowering::DAGCombinerInfo &DCI, | |||
21764 | const X86Subtarget* Subtarget) { | |||
21765 | SDLoc dl(N); | |||
21766 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N); | |||
21767 | SDValue V1 = SVOp->getOperand(0); | |||
21768 | SDValue V2 = SVOp->getOperand(1); | |||
21769 | EVT VT = SVOp->getValueType(0); | |||
21770 | unsigned NumElems = VT.getVectorNumElements(); | |||
21771 | ||||
21772 | if (V1.getOpcode() == ISD::CONCAT_VECTORS && | |||
21773 | V2.getOpcode() == ISD::CONCAT_VECTORS) { | |||
21774 | // | |||
21775 | // 0,0,0,... | |||
21776 | // | | |||
21777 | // V UNDEF BUILD_VECTOR UNDEF | |||
21778 | // \ / \ / | |||
21779 | // CONCAT_VECTOR CONCAT_VECTOR | |||
21780 | // \ / | |||
21781 | // \ / | |||
21782 | // RESULT: V + zero extended | |||
21783 | // | |||
21784 | if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR || | |||
21785 | V2.getOperand(1).getOpcode() != ISD::UNDEF || | |||
21786 | V1.getOperand(1).getOpcode() != ISD::UNDEF) | |||
21787 | return SDValue(); | |||
21788 | ||||
21789 | if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode())) | |||
21790 | return SDValue(); | |||
21791 | ||||
21792 | // To match the shuffle mask, the first half of the mask should | |||
21793 | // be exactly the first vector, and all the rest a splat with the | |||
21794 | // first element of the second one. | |||
21795 | for (unsigned i = 0; i != NumElems/2; ++i) | |||
21796 | if (!isUndefOrEqual(SVOp->getMaskElt(i), i) || | |||
21797 | !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems)) | |||
21798 | return SDValue(); | |||
21799 | ||||
21800 | // If V1 is coming from a vector load then just fold to a VZEXT_LOAD. | |||
21801 | if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) { | |||
21802 | if (Ld->hasNUsesOfValue(1, 0)) { | |||
21803 | SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other); | |||
21804 | SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() }; | |||
21805 | SDValue ResNode = | |||
21806 | DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, | |||
21807 | Ld->getMemoryVT(), | |||
21808 | Ld->getPointerInfo(), | |||
21809 | Ld->getAlignment(), | |||
21810 | false/*isVolatile*/, true/*ReadMem*/, | |||
21811 | false/*WriteMem*/); | |||
21812 | ||||
21813 | // Make sure the newly-created LOAD is in the same position as Ld in | |||
21814 | // terms of dependency. We create a TokenFactor for Ld and ResNode, | |||
21815 | // and update uses of Ld's output chain to use the TokenFactor. | |||
21816 | if (Ld->hasAnyUseOfValue(1)) { | |||
21817 | SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, | |||
21818 | SDValue(Ld, 1), SDValue(ResNode.getNode(), 1)); | |||
21819 | DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain); | |||
21820 | DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1), | |||
21821 | SDValue(ResNode.getNode(), 1)); | |||
21822 | } | |||
21823 | ||||
21824 | return DAG.getNode(ISD::BITCAST, dl, VT, ResNode); | |||
21825 | } | |||
21826 | } | |||
21827 | ||||
21828 | // Emit a zeroed vector and insert the desired subvector on its | |||
21829 | // first half. | |||
21830 | SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl); | |||
21831 | SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl); | |||
21832 | return DCI.CombineTo(N, InsV); | |||
21833 | } | |||
21834 | ||||
21835 | //===--------------------------------------------------------------------===// | |||
21836 | // Combine some shuffles into subvector extracts and inserts: | |||
21837 | // | |||
21838 | ||||
21839 | // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u> | |||
21840 | if (isShuffleHigh128VectorInsertLow(SVOp)) { | |||
21841 | SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl); | |||
21842 | SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl); | |||
21843 | return DCI.CombineTo(N, InsV); | |||
21844 | } | |||
21845 | ||||
21846 | // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1> | |||
21847 | if (isShuffleLow128VectorInsertHigh(SVOp)) { | |||
21848 | SDValue V = Extract128BitVector(V1, 0, DAG, dl); | |||
21849 | SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl); | |||
21850 | return DCI.CombineTo(N, InsV); | |||
21851 | } | |||
21852 | ||||
21853 | return SDValue(); | |||
21854 | } | |||
21855 | ||||
21856 | /// \brief Combine an arbitrary chain of shuffles into a single instruction if | |||
21857 | /// possible. | |||
21858 | /// | |||
21859 | /// This is the leaf of the recursive combinine below. When we have found some | |||
21860 | /// chain of single-use x86 shuffle instructions and accumulated the combined | |||
21861 | /// shuffle mask represented by them, this will try to pattern match that mask | |||
21862 | /// into either a single instruction if there is a special purpose instruction | |||
21863 | /// for this operation, or into a PSHUFB instruction which is a fully general | |||
21864 | /// instruction but should only be used to replace chains over a certain depth. | |||
21865 | static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask, | |||
21866 | int Depth, bool HasPSHUFB, SelectionDAG &DAG, | |||
21867 | TargetLowering::DAGCombinerInfo &DCI, | |||
21868 | const X86Subtarget *Subtarget) { | |||
21869 | assert(!Mask.empty() && "Cannot combine an empty shuffle mask!")((!Mask.empty() && "Cannot combine an empty shuffle mask!" ) ? static_cast<void> (0) : __assert_fail ("!Mask.empty() && \"Cannot combine an empty shuffle mask!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21869, __PRETTY_FUNCTION__)); | |||
21870 | ||||
21871 | // Find the operand that enters the chain. Note that multiple uses are OK | |||
21872 | // here, we're not going to remove the operand we find. | |||
21873 | SDValue Input = Op.getOperand(0); | |||
21874 | while (Input.getOpcode() == ISD::BITCAST) | |||
21875 | Input = Input.getOperand(0); | |||
21876 | ||||
21877 | MVT VT = Input.getSimpleValueType(); | |||
21878 | MVT RootVT = Root.getSimpleValueType(); | |||
21879 | SDLoc DL(Root); | |||
21880 | ||||
21881 | // Just remove no-op shuffle masks. | |||
21882 | if (Mask.size() == 1) { | |||
21883 | DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Input), | |||
21884 | /*AddTo*/ true); | |||
21885 | return true; | |||
21886 | } | |||
21887 | ||||
21888 | // Use the float domain if the operand type is a floating point type. | |||
21889 | bool FloatDomain = VT.isFloatingPoint(); | |||
21890 | ||||
21891 | // For floating point shuffles, we don't have free copies in the shuffle | |||
21892 | // instructions or the ability to load as part of the instruction, so | |||
21893 | // canonicalize their shuffles to UNPCK or MOV variants. | |||
21894 | // | |||
21895 | // Note that even with AVX we prefer the PSHUFD form of shuffle for integer | |||
21896 | // vectors because it can have a load folded into it that UNPCK cannot. This | |||
21897 | // doesn't preclude something switching to the shorter encoding post-RA. | |||
21898 | if (FloatDomain) { | |||
21899 | if (Mask.equals(0, 0) || Mask.equals(1, 1)) { | |||
21900 | bool Lo = Mask.equals(0, 0); | |||
21901 | unsigned Shuffle; | |||
21902 | MVT ShuffleVT; | |||
21903 | // Check if we have SSE3 which will let us use MOVDDUP. That instruction | |||
21904 | // is no slower than UNPCKLPD but has the option to fold the input operand | |||
21905 | // into even an unaligned memory load. | |||
21906 | if (Lo && Subtarget->hasSSE3()) { | |||
21907 | Shuffle = X86ISD::MOVDDUP; | |||
21908 | ShuffleVT = MVT::v2f64; | |||
21909 | } else { | |||
21910 | // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller | |||
21911 | // than the UNPCK variants. | |||
21912 | Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS; | |||
21913 | ShuffleVT = MVT::v4f32; | |||
21914 | } | |||
21915 | if (Depth == 1 && Root->getOpcode() == Shuffle) | |||
21916 | return false; // Nothing to do! | |||
21917 | Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input); | |||
21918 | DCI.AddToWorklist(Op.getNode()); | |||
21919 | if (Shuffle == X86ISD::MOVDDUP) | |||
21920 | Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op); | |||
21921 | else | |||
21922 | Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op); | |||
21923 | DCI.AddToWorklist(Op.getNode()); | |||
21924 | DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op), | |||
21925 | /*AddTo*/ true); | |||
21926 | return true; | |||
21927 | } | |||
21928 | if (Subtarget->hasSSE3() && | |||
21929 | (Mask.equals(0, 0, 2, 2) || Mask.equals(1, 1, 3, 3))) { | |||
21930 | bool Lo = Mask.equals(0, 0, 2, 2); | |||
21931 | unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP; | |||
21932 | MVT ShuffleVT = MVT::v4f32; | |||
21933 | if (Depth == 1 && Root->getOpcode() == Shuffle) | |||
21934 | return false; // Nothing to do! | |||
21935 | Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input); | |||
21936 | DCI.AddToWorklist(Op.getNode()); | |||
21937 | Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op); | |||
21938 | DCI.AddToWorklist(Op.getNode()); | |||
21939 | DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op), | |||
21940 | /*AddTo*/ true); | |||
21941 | return true; | |||
21942 | } | |||
21943 | if (Mask.equals(0, 0, 1, 1) || Mask.equals(2, 2, 3, 3)) { | |||
21944 | bool Lo = Mask.equals(0, 0, 1, 1); | |||
21945 | unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH; | |||
21946 | MVT ShuffleVT = MVT::v4f32; | |||
21947 | if (Depth == 1 && Root->getOpcode() == Shuffle) | |||
21948 | return false; // Nothing to do! | |||
21949 | Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input); | |||
21950 | DCI.AddToWorklist(Op.getNode()); | |||
21951 | Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op); | |||
21952 | DCI.AddToWorklist(Op.getNode()); | |||
21953 | DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op), | |||
21954 | /*AddTo*/ true); | |||
21955 | return true; | |||
21956 | } | |||
21957 | } | |||
21958 | ||||
21959 | // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK | |||
21960 | // variants as none of these have single-instruction variants that are | |||
21961 | // superior to the UNPCK formulation. | |||
21962 | if (!FloatDomain && | |||
21963 | (Mask.equals(0, 0, 1, 1, 2, 2, 3, 3) || | |||
21964 | Mask.equals(4, 4, 5, 5, 6, 6, 7, 7) || | |||
21965 | Mask.equals(0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7) || | |||
21966 | Mask.equals(8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, | |||
21967 | 15))) { | |||
21968 | bool Lo = Mask[0] == 0; | |||
21969 | unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH; | |||
21970 | if (Depth == 1 && Root->getOpcode() == Shuffle) | |||
21971 | return false; // Nothing to do! | |||
21972 | MVT ShuffleVT; | |||
21973 | switch (Mask.size()) { | |||
21974 | case 8: | |||
21975 | ShuffleVT = MVT::v8i16; | |||
21976 | break; | |||
21977 | case 16: | |||
21978 | ShuffleVT = MVT::v16i8; | |||
21979 | break; | |||
21980 | default: | |||
21981 | llvm_unreachable("Impossible mask size!")::llvm::llvm_unreachable_internal("Impossible mask size!", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 21981); | |||
21982 | }; | |||
21983 | Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input); | |||
21984 | DCI.AddToWorklist(Op.getNode()); | |||
21985 | Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op); | |||
21986 | DCI.AddToWorklist(Op.getNode()); | |||
21987 | DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op), | |||
21988 | /*AddTo*/ true); | |||
21989 | return true; | |||
21990 | } | |||
21991 | ||||
21992 | // Don't try to re-form single instruction chains under any circumstances now | |||
21993 | // that we've done encoding canonicalization for them. | |||
21994 | if (Depth < 2) | |||
21995 | return false; | |||
21996 | ||||
21997 | // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we | |||
21998 | // can replace them with a single PSHUFB instruction profitably. Intel's | |||
21999 | // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but | |||
22000 | // in practice PSHUFB tends to be *very* fast so we're more aggressive. | |||
22001 | if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) { | |||
22002 | SmallVector<SDValue, 16> PSHUFBMask; | |||
22003 | assert(Mask.size() <= 16 && "Can't shuffle elements smaller than bytes!")((Mask.size() <= 16 && "Can't shuffle elements smaller than bytes!" ) ? static_cast<void> (0) : __assert_fail ("Mask.size() <= 16 && \"Can't shuffle elements smaller than bytes!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22003, __PRETTY_FUNCTION__)); | |||
22004 | int Ratio = 16 / Mask.size(); | |||
22005 | for (unsigned i = 0; i < 16; ++i) { | |||
22006 | if (Mask[i / Ratio] == SM_SentinelUndef) { | |||
22007 | PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8)); | |||
22008 | continue; | |||
22009 | } | |||
22010 | int M = Mask[i / Ratio] != SM_SentinelZero | |||
22011 | ? Ratio * Mask[i / Ratio] + i % Ratio | |||
22012 | : 255; | |||
22013 | PSHUFBMask.push_back(DAG.getConstant(M, MVT::i8)); | |||
22014 | } | |||
22015 | Op = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Input); | |||
22016 | DCI.AddToWorklist(Op.getNode()); | |||
22017 | SDValue PSHUFBMaskOp = | |||
22018 | DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, PSHUFBMask); | |||
22019 | DCI.AddToWorklist(PSHUFBMaskOp.getNode()); | |||
22020 | Op = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, Op, PSHUFBMaskOp); | |||
22021 | DCI.AddToWorklist(Op.getNode()); | |||
22022 | DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op), | |||
22023 | /*AddTo*/ true); | |||
22024 | return true; | |||
22025 | } | |||
22026 | ||||
22027 | // Failed to find any combines. | |||
22028 | return false; | |||
22029 | } | |||
22030 | ||||
22031 | /// \brief Fully generic combining of x86 shuffle instructions. | |||
22032 | /// | |||
22033 | /// This should be the last combine run over the x86 shuffle instructions. Once | |||
22034 | /// they have been fully optimized, this will recursively consider all chains | |||
22035 | /// of single-use shuffle instructions, build a generic model of the cumulative | |||
22036 | /// shuffle operation, and check for simpler instructions which implement this | |||
22037 | /// operation. We use this primarily for two purposes: | |||
22038 | /// | |||
22039 | /// 1) Collapse generic shuffles to specialized single instructions when | |||
22040 | /// equivalent. In most cases, this is just an encoding size win, but | |||
22041 | /// sometimes we will collapse multiple generic shuffles into a single | |||
22042 | /// special-purpose shuffle. | |||
22043 | /// 2) Look for sequences of shuffle instructions with 3 or more total | |||
22044 | /// instructions, and replace them with the slightly more expensive SSSE3 | |||
22045 | /// PSHUFB instruction if available. We do this as the last combining step | |||
22046 | /// to ensure we avoid using PSHUFB if we can implement the shuffle with | |||
22047 | /// a suitable short sequence of other instructions. The PHUFB will either | |||
22048 | /// use a register or have to read from memory and so is slightly (but only | |||
22049 | /// slightly) more expensive than the other shuffle instructions. | |||
22050 | /// | |||
22051 | /// Because this is inherently a quadratic operation (for each shuffle in | |||
22052 | /// a chain, we recurse up the chain), the depth is limited to 8 instructions. | |||
22053 | /// This should never be an issue in practice as the shuffle lowering doesn't | |||
22054 | /// produce sequences of more than 8 instructions. | |||
22055 | /// | |||
22056 | /// FIXME: We will currently miss some cases where the redundant shuffling | |||
22057 | /// would simplify under the threshold for PSHUFB formation because of | |||
22058 | /// combine-ordering. To fix this, we should do the redundant instruction | |||
22059 | /// combining in this recursive walk. | |||
22060 | static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root, | |||
22061 | ArrayRef<int> RootMask, | |||
22062 | int Depth, bool HasPSHUFB, | |||
22063 | SelectionDAG &DAG, | |||
22064 | TargetLowering::DAGCombinerInfo &DCI, | |||
22065 | const X86Subtarget *Subtarget) { | |||
22066 | // Bound the depth of our recursive combine because this is ultimately | |||
22067 | // quadratic in nature. | |||
22068 | if (Depth > 8) | |||
22069 | return false; | |||
22070 | ||||
22071 | // Directly rip through bitcasts to find the underlying operand. | |||
22072 | while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse()) | |||
22073 | Op = Op.getOperand(0); | |||
22074 | ||||
22075 | MVT VT = Op.getSimpleValueType(); | |||
22076 | if (!VT.isVector()) | |||
22077 | return false; // Bail if we hit a non-vector. | |||
22078 | // FIXME: This routine should be taught about 256-bit shuffles, or a 256-bit | |||
22079 | // version should be added. | |||
22080 | if (VT.getSizeInBits() != 128) | |||
22081 | return false; | |||
22082 | ||||
22083 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22084, __PRETTY_FUNCTION__)) | |||
22084 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22084, __PRETTY_FUNCTION__)); | |||
22085 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22086, __PRETTY_FUNCTION__)) | |||
22086 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22086, __PRETTY_FUNCTION__)); | |||
22087 | ||||
22088 | if (!isTargetShuffle(Op.getOpcode())) | |||
22089 | return false; | |||
22090 | SmallVector<int, 16> OpMask; | |||
22091 | bool IsUnary; | |||
22092 | bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary); | |||
22093 | // We only can combine unary shuffles which we can decode the mask for. | |||
22094 | if (!HaveMask || !IsUnary) | |||
22095 | return false; | |||
22096 | ||||
22097 | assert(VT.getVectorNumElements() == OpMask.size() &&((VT.getVectorNumElements() == OpMask.size() && "Different mask size from vector size!" ) ? static_cast<void> (0) : __assert_fail ("VT.getVectorNumElements() == OpMask.size() && \"Different mask size from vector size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22098, __PRETTY_FUNCTION__)) | |||
22098 | "Different mask size from vector size!")((VT.getVectorNumElements() == OpMask.size() && "Different mask size from vector size!" ) ? static_cast<void> (0) : __assert_fail ("VT.getVectorNumElements() == OpMask.size() && \"Different mask size from vector size!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22098, __PRETTY_FUNCTION__)); | |||
22099 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22104, __PRETTY_FUNCTION__)) | |||
22100 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22104, __PRETTY_FUNCTION__)) | |||
22101 | (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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22104, __PRETTY_FUNCTION__)) | |||
22102 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22104, __PRETTY_FUNCTION__)) | |||
22103 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22104, __PRETTY_FUNCTION__)) | |||
22104 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22104, __PRETTY_FUNCTION__)); | |||
22105 | int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size()); | |||
22106 | int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size()); | |||
22107 | assert(((RootRatio == 1 && OpRatio == 1) ||((((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) || (RootRatio == 1) != (OpRatio == 1)) && \"Must not have a ratio for both incoming and op masks!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22109, __PRETTY_FUNCTION__)) | |||
22108 | (RootRatio == 1) != (OpRatio == 1)) &&((((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) || (RootRatio == 1) != (OpRatio == 1)) && \"Must not have a ratio for both incoming and op masks!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22109, __PRETTY_FUNCTION__)) | |||
22109 | "Must not have a ratio for both incoming and op masks!")((((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) || (RootRatio == 1) != (OpRatio == 1)) && \"Must not have a ratio for both incoming and op masks!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22109, __PRETTY_FUNCTION__)); | |||
22110 | ||||
22111 | SmallVector<int, 16> Mask; | |||
22112 | Mask.reserve(std::max(OpMask.size(), RootMask.size())); | |||
22113 | ||||
22114 | // Merge this shuffle operation's mask into our accumulated mask. Note that | |||
22115 | // this shuffle's mask will be the first applied to the input, followed by the | |||
22116 | // root mask to get us all the way to the root value arrangement. The reason | |||
22117 | // for this order is that we are recursing up the operation chain. | |||
22118 | for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) { | |||
22119 | int RootIdx = i / RootRatio; | |||
22120 | if (RootMask[RootIdx] < 0) { | |||
22121 | // This is a zero or undef lane, we're done. | |||
22122 | Mask.push_back(RootMask[RootIdx]); | |||
22123 | continue; | |||
22124 | } | |||
22125 | ||||
22126 | int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio; | |||
22127 | int OpIdx = RootMaskedIdx / OpRatio; | |||
22128 | if (OpMask[OpIdx] < 0) { | |||
22129 | // The incoming lanes are zero or undef, it doesn't matter which ones we | |||
22130 | // are using. | |||
22131 | Mask.push_back(OpMask[OpIdx]); | |||
22132 | continue; | |||
22133 | } | |||
22134 | ||||
22135 | // Ok, we have non-zero lanes, map them through. | |||
22136 | Mask.push_back(OpMask[OpIdx] * OpRatio + | |||
22137 | RootMaskedIdx % OpRatio); | |||
22138 | } | |||
22139 | ||||
22140 | // See if we can recurse into the operand to combine more things. | |||
22141 | switch (Op.getOpcode()) { | |||
22142 | case X86ISD::PSHUFB: | |||
22143 | HasPSHUFB = true; | |||
22144 | case X86ISD::PSHUFD: | |||
22145 | case X86ISD::PSHUFHW: | |||
22146 | case X86ISD::PSHUFLW: | |||
22147 | if (Op.getOperand(0).hasOneUse() && | |||
22148 | combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1, | |||
22149 | HasPSHUFB, DAG, DCI, Subtarget)) | |||
22150 | return true; | |||
22151 | break; | |||
22152 | ||||
22153 | case X86ISD::UNPCKL: | |||
22154 | case X86ISD::UNPCKH: | |||
22155 | assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!")((Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!" ) ? static_cast<void> (0) : __assert_fail ("Op.getOperand(0) == Op.getOperand(1) && \"We only combine unary shuffles!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22155, __PRETTY_FUNCTION__)); | |||
22156 | // We can't check for single use, we have to check that this shuffle is the only user. | |||
22157 | if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) && | |||
22158 | combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1, | |||
22159 | HasPSHUFB, DAG, DCI, Subtarget)) | |||
22160 | return true; | |||
22161 | break; | |||
22162 | } | |||
22163 | ||||
22164 | // Minor canonicalization of the accumulated shuffle mask to make it easier | |||
22165 | // to match below. All this does is detect masks with squential pairs of | |||
22166 | // elements, and shrink them to the half-width mask. It does this in a loop | |||
22167 | // so it will reduce the size of the mask to the minimal width mask which | |||
22168 | // performs an equivalent shuffle. | |||
22169 | SmallVector<int, 16> WidenedMask; | |||
22170 | while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) { | |||
22171 | Mask = std::move(WidenedMask); | |||
22172 | WidenedMask.clear(); | |||
22173 | } | |||
22174 | ||||
22175 | return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI, | |||
22176 | Subtarget); | |||
22177 | } | |||
22178 | ||||
22179 | /// \brief Get the PSHUF-style mask from PSHUF node. | |||
22180 | /// | |||
22181 | /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4 | |||
22182 | /// PSHUF-style masks that can be reused with such instructions. | |||
22183 | static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) { | |||
22184 | SmallVector<int, 4> Mask; | |||
22185 | bool IsUnary; | |||
22186 | bool HaveMask = getTargetShuffleMask(N.getNode(), N.getSimpleValueType(), Mask, IsUnary); | |||
22187 | (void)HaveMask; | |||
22188 | assert(HaveMask)((HaveMask) ? static_cast<void> (0) : __assert_fail ("HaveMask" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22188, __PRETTY_FUNCTION__)); | |||
22189 | ||||
22190 | switch (N.getOpcode()) { | |||
22191 | case X86ISD::PSHUFD: | |||
22192 | return Mask; | |||
22193 | case X86ISD::PSHUFLW: | |||
22194 | Mask.resize(4); | |||
22195 | return Mask; | |||
22196 | case X86ISD::PSHUFHW: | |||
22197 | Mask.erase(Mask.begin(), Mask.begin() + 4); | |||
22198 | for (int &M : Mask) | |||
22199 | M -= 4; | |||
22200 | return Mask; | |||
22201 | default: | |||
22202 | llvm_unreachable("No valid shuffle instruction found!")::llvm::llvm_unreachable_internal("No valid shuffle instruction found!" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22202); | |||
22203 | } | |||
22204 | } | |||
22205 | ||||
22206 | /// \brief Search for a combinable shuffle across a chain ending in pshufd. | |||
22207 | /// | |||
22208 | /// We walk up the chain and look for a combinable shuffle, skipping over | |||
22209 | /// shuffles that we could hoist this shuffle's transformation past without | |||
22210 | /// altering anything. | |||
22211 | static SDValue | |||
22212 | combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask, | |||
22213 | SelectionDAG &DAG, | |||
22214 | TargetLowering::DAGCombinerInfo &DCI) { | |||
22215 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22216, __PRETTY_FUNCTION__)) | |||
22216 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22216, __PRETTY_FUNCTION__)); | |||
22217 | SDLoc DL(N); | |||
22218 | ||||
22219 | // Walk up a single-use chain looking for a combinable shuffle. Keep a stack | |||
22220 | // of the shuffles in the chain so that we can form a fresh chain to replace | |||
22221 | // this one. | |||
22222 | SmallVector<SDValue, 8> Chain; | |||
22223 | SDValue V = N.getOperand(0); | |||
22224 | for (; V.hasOneUse(); V = V.getOperand(0)) { | |||
22225 | switch (V.getOpcode()) { | |||
22226 | default: | |||
22227 | return SDValue(); // Nothing combined! | |||
22228 | ||||
22229 | case ISD::BITCAST: | |||
22230 | // Skip bitcasts as we always know the type for the target specific | |||
22231 | // instructions. | |||
22232 | continue; | |||
22233 | ||||
22234 | case X86ISD::PSHUFD: | |||
22235 | // Found another dword shuffle. | |||
22236 | break; | |||
22237 | ||||
22238 | case X86ISD::PSHUFLW: | |||
22239 | // Check that the low words (being shuffled) are the identity in the | |||
22240 | // dword shuffle, and the high words are self-contained. | |||
22241 | if (Mask[0] != 0 || Mask[1] != 1 || | |||
22242 | !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4)) | |||
22243 | return SDValue(); | |||
22244 | ||||
22245 | Chain.push_back(V); | |||
22246 | continue; | |||
22247 | ||||
22248 | case X86ISD::PSHUFHW: | |||
22249 | // Check that the high words (being shuffled) are the identity in the | |||
22250 | // dword shuffle, and the low words are self-contained. | |||
22251 | if (Mask[2] != 2 || Mask[3] != 3 || | |||
22252 | !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2)) | |||
22253 | return SDValue(); | |||
22254 | ||||
22255 | Chain.push_back(V); | |||
22256 | continue; | |||
22257 | ||||
22258 | case X86ISD::UNPCKL: | |||
22259 | case X86ISD::UNPCKH: | |||
22260 | // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword | |||
22261 | // shuffle into a preceding word shuffle. | |||
22262 | if (V.getValueType() != MVT::v16i8 && V.getValueType() != MVT::v8i16) | |||
22263 | return SDValue(); | |||
22264 | ||||
22265 | // Search for a half-shuffle which we can combine with. | |||
22266 | unsigned CombineOp = | |||
22267 | V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW; | |||
22268 | if (V.getOperand(0) != V.getOperand(1) || | |||
22269 | !V->isOnlyUserOf(V.getOperand(0).getNode())) | |||
22270 | return SDValue(); | |||
22271 | Chain.push_back(V); | |||
22272 | V = V.getOperand(0); | |||
22273 | do { | |||
22274 | switch (V.getOpcode()) { | |||
22275 | default: | |||
22276 | return SDValue(); // Nothing to combine. | |||
22277 | ||||
22278 | case X86ISD::PSHUFLW: | |||
22279 | case X86ISD::PSHUFHW: | |||
22280 | if (V.getOpcode() == CombineOp) | |||
22281 | break; | |||
22282 | ||||
22283 | Chain.push_back(V); | |||
22284 | ||||
22285 | // Fallthrough! | |||
22286 | case ISD::BITCAST: | |||
22287 | V = V.getOperand(0); | |||
22288 | continue; | |||
22289 | } | |||
22290 | break; | |||
22291 | } while (V.hasOneUse()); | |||
22292 | break; | |||
22293 | } | |||
22294 | // Break out of the loop if we break out of the switch. | |||
22295 | break; | |||
22296 | } | |||
22297 | ||||
22298 | if (!V.hasOneUse()) | |||
22299 | // We fell out of the loop without finding a viable combining instruction. | |||
22300 | return SDValue(); | |||
22301 | ||||
22302 | // Merge this node's mask and our incoming mask. | |||
22303 | SmallVector<int, 4> VMask = getPSHUFShuffleMask(V); | |||
22304 | for (int &M : Mask) | |||
22305 | M = VMask[M]; | |||
22306 | V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0), | |||
22307 | getV4X86ShuffleImm8ForMask(Mask, DAG)); | |||
22308 | ||||
22309 | // Rebuild the chain around this new shuffle. | |||
22310 | while (!Chain.empty()) { | |||
22311 | SDValue W = Chain.pop_back_val(); | |||
22312 | ||||
22313 | if (V.getValueType() != W.getOperand(0).getValueType()) | |||
22314 | V = DAG.getNode(ISD::BITCAST, DL, W.getOperand(0).getValueType(), V); | |||
22315 | ||||
22316 | switch (W.getOpcode()) { | |||
22317 | default: | |||
22318 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22318); | |||
22319 | ||||
22320 | case X86ISD::UNPCKL: | |||
22321 | case X86ISD::UNPCKH: | |||
22322 | V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V); | |||
22323 | break; | |||
22324 | ||||
22325 | case X86ISD::PSHUFD: | |||
22326 | case X86ISD::PSHUFLW: | |||
22327 | case X86ISD::PSHUFHW: | |||
22328 | V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1)); | |||
22329 | break; | |||
22330 | } | |||
22331 | } | |||
22332 | if (V.getValueType() != N.getValueType()) | |||
22333 | V = DAG.getNode(ISD::BITCAST, DL, N.getValueType(), V); | |||
22334 | ||||
22335 | // Return the new chain to replace N. | |||
22336 | return V; | |||
22337 | } | |||
22338 | ||||
22339 | /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw. | |||
22340 | /// | |||
22341 | /// We walk up the chain, skipping shuffles of the other half and looking | |||
22342 | /// through shuffles which switch halves trying to find a shuffle of the same | |||
22343 | /// pair of dwords. | |||
22344 | static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask, | |||
22345 | SelectionDAG &DAG, | |||
22346 | TargetLowering::DAGCombinerInfo &DCI) { | |||
22347 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22349, __PRETTY_FUNCTION__)) | |||
22348 | (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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22349, __PRETTY_FUNCTION__)) | |||
22349 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22349, __PRETTY_FUNCTION__)); | |||
22350 | SDLoc DL(N); | |||
22351 | unsigned CombineOpcode = N.getOpcode(); | |||
22352 | ||||
22353 | // Walk up a single-use chain looking for a combinable shuffle. | |||
22354 | SDValue V = N.getOperand(0); | |||
22355 | for (; V.hasOneUse(); V = V.getOperand(0)) { | |||
22356 | switch (V.getOpcode()) { | |||
22357 | default: | |||
22358 | return false; // Nothing combined! | |||
22359 | ||||
22360 | case ISD::BITCAST: | |||
22361 | // Skip bitcasts as we always know the type for the target specific | |||
22362 | // instructions. | |||
22363 | continue; | |||
22364 | ||||
22365 | case X86ISD::PSHUFLW: | |||
22366 | case X86ISD::PSHUFHW: | |||
22367 | if (V.getOpcode() == CombineOpcode) | |||
22368 | break; | |||
22369 | ||||
22370 | // Other-half shuffles are no-ops. | |||
22371 | continue; | |||
22372 | } | |||
22373 | // Break out of the loop if we break out of the switch. | |||
22374 | break; | |||
22375 | } | |||
22376 | ||||
22377 | if (!V.hasOneUse()) | |||
22378 | // We fell out of the loop without finding a viable combining instruction. | |||
22379 | return false; | |||
22380 | ||||
22381 | // Combine away the bottom node as its shuffle will be accumulated into | |||
22382 | // a preceding shuffle. | |||
22383 | DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true); | |||
22384 | ||||
22385 | // Record the old value. | |||
22386 | SDValue Old = V; | |||
22387 | ||||
22388 | // Merge this node's mask and our incoming mask (adjusted to account for all | |||
22389 | // the pshufd instructions encountered). | |||
22390 | SmallVector<int, 4> VMask = getPSHUFShuffleMask(V); | |||
22391 | for (int &M : Mask) | |||
22392 | M = VMask[M]; | |||
22393 | V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0), | |||
22394 | getV4X86ShuffleImm8ForMask(Mask, DAG)); | |||
22395 | ||||
22396 | // Check that the shuffles didn't cancel each other out. If not, we need to | |||
22397 | // combine to the new one. | |||
22398 | if (Old != V) | |||
22399 | // Replace the combinable shuffle with the combined one, updating all users | |||
22400 | // so that we re-evaluate the chain here. | |||
22401 | DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true); | |||
22402 | ||||
22403 | return true; | |||
22404 | } | |||
22405 | ||||
22406 | /// \brief Try to combine x86 target specific shuffles. | |||
22407 | static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG, | |||
22408 | TargetLowering::DAGCombinerInfo &DCI, | |||
22409 | const X86Subtarget *Subtarget) { | |||
22410 | SDLoc DL(N); | |||
22411 | MVT VT = N.getSimpleValueType(); | |||
22412 | SmallVector<int, 4> Mask; | |||
22413 | ||||
22414 | switch (N.getOpcode()) { | |||
22415 | case X86ISD::PSHUFD: | |||
22416 | case X86ISD::PSHUFLW: | |||
22417 | case X86ISD::PSHUFHW: | |||
22418 | Mask = getPSHUFShuffleMask(N); | |||
22419 | assert(Mask.size() == 4)((Mask.size() == 4) ? static_cast<void> (0) : __assert_fail ("Mask.size() == 4", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22419, __PRETTY_FUNCTION__)); | |||
22420 | break; | |||
22421 | default: | |||
22422 | return SDValue(); | |||
22423 | } | |||
22424 | ||||
22425 | // Nuke no-op shuffles that show up after combining. | |||
22426 | if (isNoopShuffleMask(Mask)) | |||
22427 | return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true); | |||
22428 | ||||
22429 | // Look for simplifications involving one or two shuffle instructions. | |||
22430 | SDValue V = N.getOperand(0); | |||
22431 | switch (N.getOpcode()) { | |||
22432 | default: | |||
22433 | break; | |||
22434 | case X86ISD::PSHUFLW: | |||
22435 | case X86ISD::PSHUFHW: | |||
22436 | assert(VT == MVT::v8i16)((VT == MVT::v8i16) ? static_cast<void> (0) : __assert_fail ("VT == MVT::v8i16", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22436, __PRETTY_FUNCTION__)); | |||
22437 | (void)VT; | |||
22438 | ||||
22439 | if (combineRedundantHalfShuffle(N, Mask, DAG, DCI)) | |||
22440 | return SDValue(); // We combined away this shuffle, so we're done. | |||
22441 | ||||
22442 | // See if this reduces to a PSHUFD which is no more expensive and can | |||
22443 | // combine with more operations. Note that it has to at least flip the | |||
22444 | // dwords as otherwise it would have been removed as a no-op. | |||
22445 | if (Mask[0] == 2 && Mask[1] == 3 && Mask[2] == 0 && Mask[3] == 1) { | |||
22446 | int DMask[] = {0, 1, 2, 3}; | |||
22447 | int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2; | |||
22448 | DMask[DOffset + 0] = DOffset + 1; | |||
22449 | DMask[DOffset + 1] = DOffset + 0; | |||
22450 | V = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V); | |||
22451 | DCI.AddToWorklist(V.getNode()); | |||
22452 | V = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V, | |||
22453 | getV4X86ShuffleImm8ForMask(DMask, DAG)); | |||
22454 | DCI.AddToWorklist(V.getNode()); | |||
22455 | return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V); | |||
22456 | } | |||
22457 | ||||
22458 | // Look for shuffle patterns which can be implemented as a single unpack. | |||
22459 | // FIXME: This doesn't handle the location of the PSHUFD generically, and | |||
22460 | // only works when we have a PSHUFD followed by two half-shuffles. | |||
22461 | if (Mask[0] == Mask[1] && Mask[2] == Mask[3] && | |||
22462 | (V.getOpcode() == X86ISD::PSHUFLW || | |||
22463 | V.getOpcode() == X86ISD::PSHUFHW) && | |||
22464 | V.getOpcode() != N.getOpcode() && | |||
22465 | V.hasOneUse()) { | |||
22466 | SDValue D = V.getOperand(0); | |||
22467 | while (D.getOpcode() == ISD::BITCAST && D.hasOneUse()) | |||
22468 | D = D.getOperand(0); | |||
22469 | if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) { | |||
22470 | SmallVector<int, 4> VMask = getPSHUFShuffleMask(V); | |||
22471 | SmallVector<int, 4> DMask = getPSHUFShuffleMask(D); | |||
22472 | int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4; | |||
22473 | int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4; | |||
22474 | int WordMask[8]; | |||
22475 | for (int i = 0; i < 4; ++i) { | |||
22476 | WordMask[i + NOffset] = Mask[i] + NOffset; | |||
22477 | WordMask[i + VOffset] = VMask[i] + VOffset; | |||
22478 | } | |||
22479 | // Map the word mask through the DWord mask. | |||
22480 | int MappedMask[8]; | |||
22481 | for (int i = 0; i < 8; ++i) | |||
22482 | MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2; | |||
22483 | const int UnpackLoMask[] = {0, 0, 1, 1, 2, 2, 3, 3}; | |||
22484 | const int UnpackHiMask[] = {4, 4, 5, 5, 6, 6, 7, 7}; | |||
22485 | if (std::equal(std::begin(MappedMask), std::end(MappedMask), | |||
22486 | std::begin(UnpackLoMask)) || | |||
22487 | std::equal(std::begin(MappedMask), std::end(MappedMask), | |||
22488 | std::begin(UnpackHiMask))) { | |||
22489 | // We can replace all three shuffles with an unpack. | |||
22490 | V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, D.getOperand(0)); | |||
22491 | DCI.AddToWorklist(V.getNode()); | |||
22492 | return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL | |||
22493 | : X86ISD::UNPCKH, | |||
22494 | DL, MVT::v8i16, V, V); | |||
22495 | } | |||
22496 | } | |||
22497 | } | |||
22498 | ||||
22499 | break; | |||
22500 | ||||
22501 | case X86ISD::PSHUFD: | |||
22502 | if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI)) | |||
22503 | return NewN; | |||
22504 | ||||
22505 | break; | |||
22506 | } | |||
22507 | ||||
22508 | return SDValue(); | |||
22509 | } | |||
22510 | ||||
22511 | /// \brief Try to combine a shuffle into a target-specific add-sub node. | |||
22512 | /// | |||
22513 | /// We combine this directly on the abstract vector shuffle nodes so it is | |||
22514 | /// easier to generically match. We also insert dummy vector shuffle nodes for | |||
22515 | /// the operands which explicitly discard the lanes which are unused by this | |||
22516 | /// operation to try to flow through the rest of the combiner the fact that | |||
22517 | /// they're unused. | |||
22518 | static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) { | |||
22519 | SDLoc DL(N); | |||
22520 | EVT VT = N->getValueType(0); | |||
22521 | ||||
22522 | // We only handle target-independent shuffles. | |||
22523 | // FIXME: It would be easy and harmless to use the target shuffle mask | |||
22524 | // extraction tool to support more. | |||
22525 | if (N->getOpcode() != ISD::VECTOR_SHUFFLE) | |||
22526 | return SDValue(); | |||
22527 | ||||
22528 | auto *SVN = cast<ShuffleVectorSDNode>(N); | |||
22529 | ArrayRef<int> Mask = SVN->getMask(); | |||
22530 | SDValue V1 = N->getOperand(0); | |||
22531 | SDValue V2 = N->getOperand(1); | |||
22532 | ||||
22533 | // We require the first shuffle operand to be the SUB node, and the second to | |||
22534 | // be the ADD node. | |||
22535 | // FIXME: We should support the commuted patterns. | |||
22536 | if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD) | |||
22537 | return SDValue(); | |||
22538 | ||||
22539 | // If there are other uses of these operations we can't fold them. | |||
22540 | if (!V1->hasOneUse() || !V2->hasOneUse()) | |||
22541 | return SDValue(); | |||
22542 | ||||
22543 | // Ensure that both operations have the same operands. Note that we can | |||
22544 | // commute the FADD operands. | |||
22545 | SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1); | |||
22546 | if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) && | |||
22547 | (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS)) | |||
22548 | return SDValue(); | |||
22549 | ||||
22550 | // We're looking for blends between FADD and FSUB nodes. We insist on these | |||
22551 | // nodes being lined up in a specific expected pattern. | |||
22552 | if (!(isShuffleEquivalent(Mask, 0, 3) || | |||
22553 | isShuffleEquivalent(Mask, 0, 5, 2, 7) || | |||
22554 | isShuffleEquivalent(Mask, 0, 9, 2, 11, 4, 13, 6, 15))) | |||
22555 | return SDValue(); | |||
22556 | ||||
22557 | // Only specific types are legal at this point, assert so we notice if and | |||
22558 | // when these change. | |||
22559 | assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||(((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 || VT == MVT::v4f64) && "Unknown vector type encountered!" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 || VT == MVT::v4f64) && \"Unknown vector type encountered!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22561, __PRETTY_FUNCTION__)) | |||
22560 | VT == MVT::v4f64) &&(((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 || VT == MVT::v4f64) && "Unknown vector type encountered!" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 || VT == MVT::v4f64) && \"Unknown vector type encountered!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22561, __PRETTY_FUNCTION__)) | |||
22561 | "Unknown vector type encountered!")(((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 || VT == MVT::v4f64) && "Unknown vector type encountered!" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 || VT == MVT::v4f64) && \"Unknown vector type encountered!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 22561, __PRETTY_FUNCTION__)); | |||
22562 | ||||
22563 | return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS); | |||
22564 | } | |||
22565 | ||||
22566 | /// PerformShuffleCombine - Performs several different shuffle combines. | |||
22567 | static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG, | |||
22568 | TargetLowering::DAGCombinerInfo &DCI, | |||
22569 | const X86Subtarget *Subtarget) { | |||
22570 | SDLoc dl(N); | |||
22571 | SDValue N0 = N->getOperand(0); | |||
22572 | SDValue N1 = N->getOperand(1); | |||
22573 | EVT VT = N->getValueType(0); | |||
22574 | ||||
22575 | // Don't create instructions with illegal types after legalize types has run. | |||
22576 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
22577 | if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType())) | |||
| ||||
22578 | return SDValue(); | |||
22579 | ||||
22580 | // If we have legalized the vector types, look for blends of FADD and FSUB | |||
22581 | // nodes that we can fuse into an ADDSUB node. | |||
22582 | if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3()) | |||
22583 | if (SDValue AddSub = combineShuffleToAddSub(N, DAG)) | |||
22584 | return AddSub; | |||
22585 | ||||
22586 | // Combine 256-bit vector shuffles. This is only profitable when in AVX mode | |||
22587 | if (Subtarget->hasFp256() && VT.is256BitVector() && | |||
22588 | N->getOpcode() == ISD::VECTOR_SHUFFLE) | |||
22589 | return PerformShuffleCombine256(N, DAG, DCI, Subtarget); | |||
22590 | ||||
22591 | // During Type Legalization, when promoting illegal vector types, | |||
22592 | // the backend might introduce new shuffle dag nodes and bitcasts. | |||
22593 | // | |||
22594 | // This code performs the following transformation: | |||
22595 | // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) -> | |||
22596 | // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>) | |||
22597 | // | |||
22598 | // We do this only if both the bitcast and the BINOP dag nodes have | |||
22599 | // one use. Also, perform this transformation only if the new binary | |||
22600 | // operation is legal. This is to avoid introducing dag nodes that | |||
22601 | // potentially need to be further expanded (or custom lowered) into a | |||
22602 | // less optimal sequence of dag nodes. | |||
22603 | if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() && | |||
22604 | N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() && | |||
22605 | N0.getOpcode() == ISD::BITCAST) { | |||
22606 | SDValue BC0 = N0.getOperand(0); | |||
22607 | EVT SVT = BC0.getValueType(); | |||
22608 | unsigned Opcode = BC0.getOpcode(); | |||
22609 | unsigned NumElts = VT.getVectorNumElements(); | |||
22610 | ||||
22611 | if (BC0.hasOneUse() && SVT.isVector() && | |||
22612 | SVT.getVectorNumElements() * 2 == NumElts && | |||
22613 | TLI.isOperationLegal(Opcode, VT)) { | |||
22614 | bool CanFold = false; | |||
22615 | switch (Opcode) { | |||
22616 | default : break; | |||
22617 | case ISD::ADD : | |||
22618 | case ISD::FADD : | |||
22619 | case ISD::SUB : | |||
22620 | case ISD::FSUB : | |||
22621 | case ISD::MUL : | |||
22622 | case ISD::FMUL : | |||
22623 | CanFold = true; | |||
22624 | } | |||
22625 | ||||
22626 | unsigned SVTNumElts = SVT.getVectorNumElements(); | |||
22627 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N); | |||
22628 | for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i) | |||
22629 | CanFold = SVOp->getMaskElt(i) == (int)(i * 2); | |||
22630 | for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i) | |||
22631 | CanFold = SVOp->getMaskElt(i) < 0; | |||
22632 | ||||
22633 | if (CanFold) { | |||
22634 | SDValue BC00 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(0)); | |||
22635 | SDValue BC01 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(1)); | |||
22636 | SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01); | |||
22637 | return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]); | |||
22638 | } | |||
22639 | } | |||
22640 | } | |||
22641 | ||||
22642 | // Only handle 128 wide vector from here on. | |||
22643 | if (!VT.is128BitVector()) | |||
22644 | return SDValue(); | |||
22645 | ||||
22646 | // Combine a vector_shuffle that is equal to build_vector load1, load2, load3, | |||
22647 | // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are | |||
22648 | // consecutive, non-overlapping, and in the right order. | |||
22649 | SmallVector<SDValue, 16> Elts; | |||
22650 | for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) | |||
22651 | Elts.push_back(getShuffleScalarElt(N, i, DAG, 0)); | |||
22652 | ||||
22653 | SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true); | |||
22654 | if (LD.getNode()) | |||
22655 | return LD; | |||
22656 | ||||
22657 | if (isTargetShuffle(N->getOpcode())) { | |||
22658 | SDValue Shuffle = | |||
22659 | PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget); | |||
22660 | if (Shuffle.getNode()) | |||
22661 | return Shuffle; | |||
22662 | ||||
22663 | // Try recursively combining arbitrary sequences of x86 shuffle | |||
22664 | // instructions into higher-order shuffles. We do this after combining | |||
22665 | // specific PSHUF instruction sequences into their minimal form so that we | |||
22666 | // can evaluate how many specialized shuffle instructions are involved in | |||
22667 | // a particular chain. | |||
22668 | SmallVector<int, 1> NonceMask; // Just a placeholder. | |||
22669 | NonceMask.push_back(0); | |||
22670 | if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask, | |||
22671 | /*Depth*/ 1, /*HasPSHUFB*/ false, DAG, | |||
22672 | DCI, Subtarget)) | |||
22673 | return SDValue(); // This routine will use CombineTo to replace N. | |||
22674 | } | |||
22675 | ||||
22676 | return SDValue(); | |||
22677 | } | |||
22678 | ||||
22679 | /// PerformTruncateCombine - Converts truncate operation to | |||
22680 | /// a sequence of vector shuffle operations. | |||
22681 | /// It is possible when we truncate 256-bit vector to 128-bit vector | |||
22682 | static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG, | |||
22683 | TargetLowering::DAGCombinerInfo &DCI, | |||
22684 | const X86Subtarget *Subtarget) { | |||
22685 | return SDValue(); | |||
22686 | } | |||
22687 | ||||
22688 | /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target | |||
22689 | /// specific shuffle of a load can be folded into a single element load. | |||
22690 | /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but | |||
22691 | /// shuffles have been custom lowered so we need to handle those here. | |||
22692 | static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG, | |||
22693 | TargetLowering::DAGCombinerInfo &DCI) { | |||
22694 | if (DCI.isBeforeLegalizeOps()) | |||
22695 | return SDValue(); | |||
22696 | ||||
22697 | SDValue InVec = N->getOperand(0); | |||
22698 | SDValue EltNo = N->getOperand(1); | |||
22699 | ||||
22700 | if (!isa<ConstantSDNode>(EltNo)) | |||
22701 | return SDValue(); | |||
22702 | ||||
22703 | EVT OriginalVT = InVec.getValueType(); | |||
22704 | ||||
22705 | if (InVec.getOpcode() == ISD::BITCAST) { | |||
22706 | // Don't duplicate a load with other uses. | |||
22707 | if (!InVec.hasOneUse()) | |||
22708 | return SDValue(); | |||
22709 | EVT BCVT = InVec.getOperand(0).getValueType(); | |||
22710 | if (BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements()) | |||
22711 | return SDValue(); | |||
22712 | InVec = InVec.getOperand(0); | |||
22713 | } | |||
22714 | ||||
22715 | EVT CurrentVT = InVec.getValueType(); | |||
22716 | ||||
22717 | if (!isTargetShuffle(InVec.getOpcode())) | |||
22718 | return SDValue(); | |||
22719 | ||||
22720 | // Don't duplicate a load with other uses. | |||
22721 | if (!InVec.hasOneUse()) | |||
22722 | return SDValue(); | |||
22723 | ||||
22724 | SmallVector<int, 16> ShuffleMask; | |||
22725 | bool UnaryShuffle; | |||
22726 | if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(), | |||
22727 | ShuffleMask, UnaryShuffle)) | |||
22728 | return SDValue(); | |||
22729 | ||||
22730 | // Select the input vector, guarding against out of range extract vector. | |||
22731 | unsigned NumElems = CurrentVT.getVectorNumElements(); | |||
22732 | int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue(); | |||
22733 | int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt]; | |||
22734 | SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0) | |||
22735 | : InVec.getOperand(1); | |||
22736 | ||||
22737 | // If inputs to shuffle are the same for both ops, then allow 2 uses | |||
22738 | unsigned AllowedUses = InVec.getNumOperands() > 1 && | |||
22739 | InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1; | |||
22740 | ||||
22741 | if (LdNode.getOpcode() == ISD::BITCAST) { | |||
22742 | // Don't duplicate a load with other uses. | |||
22743 | if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0)) | |||
22744 | return SDValue(); | |||
22745 | ||||
22746 | AllowedUses = 1; // only allow 1 load use if we have a bitcast | |||
22747 | LdNode = LdNode.getOperand(0); | |||
22748 | } | |||
22749 | ||||
22750 | if (!ISD::isNormalLoad(LdNode.getNode())) | |||
22751 | return SDValue(); | |||
22752 | ||||
22753 | LoadSDNode *LN0 = cast<LoadSDNode>(LdNode); | |||
22754 | ||||
22755 | if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile()) | |||
22756 | return SDValue(); | |||
22757 | ||||
22758 | EVT EltVT = N->getValueType(0); | |||
22759 | // If there's a bitcast before the shuffle, check if the load type and | |||
22760 | // alignment is valid. | |||
22761 | unsigned Align = LN0->getAlignment(); | |||
22762 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
22763 | unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment( | |||
22764 | EltVT.getTypeForEVT(*DAG.getContext())); | |||
22765 | ||||
22766 | if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT)) | |||
22767 | return SDValue(); | |||
22768 | ||||
22769 | // All checks match so transform back to vector_shuffle so that DAG combiner | |||
22770 | // can finish the job | |||
22771 | SDLoc dl(N); | |||
22772 | ||||
22773 | // Create shuffle node taking into account the case that its a unary shuffle | |||
22774 | SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT) | |||
22775 | : InVec.getOperand(1); | |||
22776 | Shuffle = DAG.getVectorShuffle(CurrentVT, dl, | |||
22777 | InVec.getOperand(0), Shuffle, | |||
22778 | &ShuffleMask[0]); | |||
22779 | Shuffle = DAG.getNode(ISD::BITCAST, dl, OriginalVT, Shuffle); | |||
22780 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle, | |||
22781 | EltNo); | |||
22782 | } | |||
22783 | ||||
22784 | /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index | |||
22785 | /// generation and convert it from being a bunch of shuffles and extracts | |||
22786 | /// into a somewhat faster sequence. For i686, the best sequence is apparently | |||
22787 | /// storing the value and loading scalars back, while for x64 we should | |||
22788 | /// use 64-bit extracts and shifts. | |||
22789 | static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG, | |||
22790 | TargetLowering::DAGCombinerInfo &DCI) { | |||
22791 | SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI); | |||
22792 | if (NewOp.getNode()) | |||
22793 | return NewOp; | |||
22794 | ||||
22795 | SDValue InputVector = N->getOperand(0); | |||
22796 | ||||
22797 | // Detect whether we are trying to convert from mmx to i32 and the bitcast | |||
22798 | // from mmx to v2i32 has a single usage. | |||
22799 | if (InputVector.getNode()->getOpcode() == llvm::ISD::BITCAST && | |||
22800 | InputVector.getNode()->getOperand(0).getValueType() == MVT::x86mmx && | |||
22801 | InputVector.hasOneUse() && N->getValueType(0) == MVT::i32) | |||
22802 | return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector), | |||
22803 | N->getValueType(0), | |||
22804 | InputVector.getNode()->getOperand(0)); | |||
22805 | ||||
22806 | // Only operate on vectors of 4 elements, where the alternative shuffling | |||
22807 | // gets to be more expensive. | |||
22808 | if (InputVector.getValueType() != MVT::v4i32) | |||
22809 | return SDValue(); | |||
22810 | ||||
22811 | // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a | |||
22812 | // single use which is a sign-extend or zero-extend, and all elements are | |||
22813 | // used. | |||
22814 | SmallVector<SDNode *, 4> Uses; | |||
22815 | unsigned ExtractedElements = 0; | |||
22816 | for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(), | |||
22817 | UE = InputVector.getNode()->use_end(); UI != UE; ++UI) { | |||
22818 | if (UI.getUse().getResNo() != InputVector.getResNo()) | |||
22819 | return SDValue(); | |||
22820 | ||||
22821 | SDNode *Extract = *UI; | |||
22822 | if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT) | |||
22823 | return SDValue(); | |||
22824 | ||||
22825 | if (Extract->getValueType(0) != MVT::i32) | |||
22826 | return SDValue(); | |||
22827 | if (!Extract->hasOneUse()) | |||
22828 | return SDValue(); | |||
22829 | if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND && | |||
22830 | Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND) | |||
22831 | return SDValue(); | |||
22832 | if (!isa<ConstantSDNode>(Extract->getOperand(1))) | |||
22833 | return SDValue(); | |||
22834 | ||||
22835 | // Record which element was extracted. | |||
22836 | ExtractedElements |= | |||
22837 | 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue(); | |||
22838 | ||||
22839 | Uses.push_back(Extract); | |||
22840 | } | |||
22841 | ||||
22842 | // If not all the elements were used, this may not be worthwhile. | |||
22843 | if (ExtractedElements != 15) | |||
22844 | return SDValue(); | |||
22845 | ||||
22846 | // Ok, we've now decided to do the transformation. | |||
22847 | // If 64-bit shifts are legal, use the extract-shift sequence, | |||
22848 | // otherwise bounce the vector off the cache. | |||
22849 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
22850 | SDValue Vals[4]; | |||
22851 | SDLoc dl(InputVector); | |||
22852 | ||||
22853 | if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) { | |||
22854 | SDValue Cst = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, InputVector); | |||
22855 | EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(); | |||
22856 | SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst, | |||
22857 | DAG.getConstant(0, VecIdxTy)); | |||
22858 | SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst, | |||
22859 | DAG.getConstant(1, VecIdxTy)); | |||
22860 | ||||
22861 | SDValue ShAmt = DAG.getConstant(32, | |||
22862 | DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64)); | |||
22863 | Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf); | |||
22864 | Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, | |||
22865 | DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt)); | |||
22866 | Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf); | |||
22867 | Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, | |||
22868 | DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt)); | |||
22869 | } else { | |||
22870 | // Store the value to a temporary stack slot. | |||
22871 | SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType()); | |||
22872 | SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, | |||
22873 | MachinePointerInfo(), false, false, 0); | |||
22874 | ||||
22875 | EVT ElementType = InputVector.getValueType().getVectorElementType(); | |||
22876 | unsigned EltSize = ElementType.getSizeInBits() / 8; | |||
22877 | ||||
22878 | // Replace each use (extract) with a load of the appropriate element. | |||
22879 | for (unsigned i = 0; i < 4; ++i) { | |||
22880 | uint64_t Offset = EltSize * i; | |||
22881 | SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy()); | |||
22882 | ||||
22883 | SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), | |||
22884 | StackPtr, OffsetVal); | |||
22885 | ||||
22886 | // Load the scalar. | |||
22887 | Vals[i] = DAG.getLoad(ElementType, dl, Ch, | |||
22888 | ScalarAddr, MachinePointerInfo(), | |||
22889 | false, false, false, 0); | |||
22890 | ||||
22891 | } | |||
22892 | } | |||
22893 | ||||
22894 | // Replace the extracts | |||
22895 | for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(), | |||
22896 | UE = Uses.end(); UI != UE; ++UI) { | |||
22897 | SDNode *Extract = *UI; | |||
22898 | ||||
22899 | SDValue Idx = Extract->getOperand(1); | |||
22900 | uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); | |||
22901 | DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]); | |||
22902 | } | |||
22903 | ||||
22904 | // The replacement was made in place; don't return anything. | |||
22905 | return SDValue(); | |||
22906 | } | |||
22907 | ||||
22908 | /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match. | |||
22909 | static std::pair<unsigned, bool> | |||
22910 | matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS, | |||
22911 | SelectionDAG &DAG, const X86Subtarget *Subtarget) { | |||
22912 | if (!VT.isVector()) | |||
22913 | return std::make_pair(0, false); | |||
22914 | ||||
22915 | bool NeedSplit = false; | |||
22916 | switch (VT.getSimpleVT().SimpleTy) { | |||
22917 | default: return std::make_pair(0, false); | |||
22918 | case MVT::v4i64: | |||
22919 | case MVT::v2i64: | |||
22920 | if (!Subtarget->hasVLX()) | |||
22921 | return std::make_pair(0, false); | |||
22922 | break; | |||
22923 | case MVT::v64i8: | |||
22924 | case MVT::v32i16: | |||
22925 | if (!Subtarget->hasBWI()) | |||
22926 | return std::make_pair(0, false); | |||
22927 | break; | |||
22928 | case MVT::v16i32: | |||
22929 | case MVT::v8i64: | |||
22930 | if (!Subtarget->hasAVX512()) | |||
22931 | return std::make_pair(0, false); | |||
22932 | break; | |||
22933 | case MVT::v32i8: | |||
22934 | case MVT::v16i16: | |||
22935 | case MVT::v8i32: | |||
22936 | if (!Subtarget->hasAVX2()) | |||
22937 | NeedSplit = true; | |||
22938 | if (!Subtarget->hasAVX()) | |||
22939 | return std::make_pair(0, false); | |||
22940 | break; | |||
22941 | case MVT::v16i8: | |||
22942 | case MVT::v8i16: | |||
22943 | case MVT::v4i32: | |||
22944 | if (!Subtarget->hasSSE2()) | |||
22945 | return std::make_pair(0, false); | |||
22946 | } | |||
22947 | ||||
22948 | // SSE2 has only a small subset of the operations. | |||
22949 | bool hasUnsigned = Subtarget->hasSSE41() || | |||
22950 | (Subtarget->hasSSE2() && VT == MVT::v16i8); | |||
22951 | bool hasSigned = Subtarget->hasSSE41() || | |||
22952 | (Subtarget->hasSSE2() && VT == MVT::v8i16); | |||
22953 | ||||
22954 | ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get(); | |||
22955 | ||||
22956 | unsigned Opc = 0; | |||
22957 | // Check for x CC y ? x : y. | |||
22958 | if (DAG.isEqualTo(LHS, Cond.getOperand(0)) && | |||
22959 | DAG.isEqualTo(RHS, Cond.getOperand(1))) { | |||
22960 | switch (CC) { | |||
22961 | default: break; | |||
22962 | case ISD::SETULT: | |||
22963 | case ISD::SETULE: | |||
22964 | Opc = hasUnsigned ? X86ISD::UMIN : 0; break; | |||
22965 | case ISD::SETUGT: | |||
22966 | case ISD::SETUGE: | |||
22967 | Opc = hasUnsigned ? X86ISD::UMAX : 0; break; | |||
22968 | case ISD::SETLT: | |||
22969 | case ISD::SETLE: | |||
22970 | Opc = hasSigned ? X86ISD::SMIN : 0; break; | |||
22971 | case ISD::SETGT: | |||
22972 | case ISD::SETGE: | |||
22973 | Opc = hasSigned ? X86ISD::SMAX : 0; break; | |||
22974 | } | |||
22975 | // Check for x CC y ? y : x -- a min/max with reversed arms. | |||
22976 | } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) && | |||
22977 | DAG.isEqualTo(RHS, Cond.getOperand(0))) { | |||
22978 | switch (CC) { | |||
22979 | default: break; | |||
22980 | case ISD::SETULT: | |||
22981 | case ISD::SETULE: | |||
22982 | Opc = hasUnsigned ? X86ISD::UMAX : 0; break; | |||
22983 | case ISD::SETUGT: | |||
22984 | case ISD::SETUGE: | |||
22985 | Opc = hasUnsigned ? X86ISD::UMIN : 0; break; | |||
22986 | case ISD::SETLT: | |||
22987 | case ISD::SETLE: | |||
22988 | Opc = hasSigned ? X86ISD::SMAX : 0; break; | |||
22989 | case ISD::SETGT: | |||
22990 | case ISD::SETGE: | |||
22991 | Opc = hasSigned ? X86ISD::SMIN : 0; break; | |||
22992 | } | |||
22993 | } | |||
22994 | ||||
22995 | return std::make_pair(Opc, NeedSplit); | |||
22996 | } | |||
22997 | ||||
22998 | static SDValue | |||
22999 | transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG, | |||
23000 | const X86Subtarget *Subtarget) { | |||
23001 | SDLoc dl(N); | |||
23002 | SDValue Cond = N->getOperand(0); | |||
23003 | SDValue LHS = N->getOperand(1); | |||
23004 | SDValue RHS = N->getOperand(2); | |||
23005 | ||||
23006 | if (Cond.getOpcode() == ISD::SIGN_EXTEND) { | |||
23007 | SDValue CondSrc = Cond->getOperand(0); | |||
23008 | if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG) | |||
23009 | Cond = CondSrc->getOperand(0); | |||
23010 | } | |||
23011 | ||||
23012 | if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) | |||
23013 | return SDValue(); | |||
23014 | ||||
23015 | // A vselect where all conditions and data are constants can be optimized into | |||
23016 | // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR(). | |||
23017 | if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) && | |||
23018 | ISD::isBuildVectorOfConstantSDNodes(RHS.getNode())) | |||
23019 | return SDValue(); | |||
23020 | ||||
23021 | unsigned MaskValue = 0; | |||
23022 | if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue)) | |||
23023 | return SDValue(); | |||
23024 | ||||
23025 | MVT VT = N->getSimpleValueType(0); | |||
23026 | unsigned NumElems = VT.getVectorNumElements(); | |||
23027 | SmallVector<int, 8> ShuffleMask(NumElems, -1); | |||
23028 | for (unsigned i = 0; i < NumElems; ++i) { | |||
23029 | // Be sure we emit undef where we can. | |||
23030 | if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF) | |||
23031 | ShuffleMask[i] = -1; | |||
23032 | else | |||
23033 | ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1); | |||
23034 | } | |||
23035 | ||||
23036 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
23037 | if (!TLI.isShuffleMaskLegal(ShuffleMask, VT)) | |||
23038 | return SDValue(); | |||
23039 | return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]); | |||
23040 | } | |||
23041 | ||||
23042 | /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT | |||
23043 | /// nodes. | |||
23044 | static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG, | |||
23045 | TargetLowering::DAGCombinerInfo &DCI, | |||
23046 | const X86Subtarget *Subtarget) { | |||
23047 | SDLoc DL(N); | |||
23048 | SDValue Cond = N->getOperand(0); | |||
23049 | // Get the LHS/RHS of the select. | |||
23050 | SDValue LHS = N->getOperand(1); | |||
23051 | SDValue RHS = N->getOperand(2); | |||
23052 | EVT VT = LHS.getValueType(); | |||
23053 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
23054 | ||||
23055 | // If we have SSE[12] support, try to form min/max nodes. SSE min/max | |||
23056 | // instructions match the semantics of the common C idiom x<y?x:y but not | |||
23057 | // x<=y?x:y, because of how they handle negative zero (which can be | |||
23058 | // ignored in unsafe-math mode). | |||
23059 | // We also try to create v2f32 min/max nodes, which we later widen to v4f32. | |||
23060 | if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() && | |||
23061 | VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) && | |||
23062 | (Subtarget->hasSSE2() || | |||
23063 | (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) { | |||
23064 | ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get(); | |||
23065 | ||||
23066 | unsigned Opcode = 0; | |||
23067 | // Check for x CC y ? x : y. | |||
23068 | if (DAG.isEqualTo(LHS, Cond.getOperand(0)) && | |||
23069 | DAG.isEqualTo(RHS, Cond.getOperand(1))) { | |||
23070 | switch (CC) { | |||
23071 | default: break; | |||
23072 | case ISD::SETULT: | |||
23073 | // Converting this to a min would handle NaNs incorrectly, and swapping | |||
23074 | // the operands would cause it to handle comparisons between positive | |||
23075 | // and negative zero incorrectly. | |||
23076 | if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) { | |||
23077 | if (!DAG.getTarget().Options.UnsafeFPMath && | |||
23078 | !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) | |||
23079 | break; | |||
23080 | std::swap(LHS, RHS); | |||
23081 | } | |||
23082 | Opcode = X86ISD::FMIN; | |||
23083 | break; | |||
23084 | case ISD::SETOLE: | |||
23085 | // Converting this to a min would handle comparisons between positive | |||
23086 | // and negative zero incorrectly. | |||
23087 | if (!DAG.getTarget().Options.UnsafeFPMath && | |||
23088 | !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) | |||
23089 | break; | |||
23090 | Opcode = X86ISD::FMIN; | |||
23091 | break; | |||
23092 | case ISD::SETULE: | |||
23093 | // Converting this to a min would handle both negative zeros and NaNs | |||
23094 | // incorrectly, but we can swap the operands to fix both. | |||
23095 | std::swap(LHS, RHS); | |||
23096 | case ISD::SETOLT: | |||
23097 | case ISD::SETLT: | |||
23098 | case ISD::SETLE: | |||
23099 | Opcode = X86ISD::FMIN; | |||
23100 | break; | |||
23101 | ||||
23102 | case ISD::SETOGE: | |||
23103 | // Converting this to a max would handle comparisons between positive | |||
23104 | // and negative zero incorrectly. | |||
23105 | if (!DAG.getTarget().Options.UnsafeFPMath && | |||
23106 | !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) | |||
23107 | break; | |||
23108 | Opcode = X86ISD::FMAX; | |||
23109 | break; | |||
23110 | case ISD::SETUGT: | |||
23111 | // Converting this to a max would handle NaNs incorrectly, and swapping | |||
23112 | // the operands would cause it to handle comparisons between positive | |||
23113 | // and negative zero incorrectly. | |||
23114 | if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) { | |||
23115 | if (!DAG.getTarget().Options.UnsafeFPMath && | |||
23116 | !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) | |||
23117 | break; | |||
23118 | std::swap(LHS, RHS); | |||
23119 | } | |||
23120 | Opcode = X86ISD::FMAX; | |||
23121 | break; | |||
23122 | case ISD::SETUGE: | |||
23123 | // Converting this to a max would handle both negative zeros and NaNs | |||
23124 | // incorrectly, but we can swap the operands to fix both. | |||
23125 | std::swap(LHS, RHS); | |||
23126 | case ISD::SETOGT: | |||
23127 | case ISD::SETGT: | |||
23128 | case ISD::SETGE: | |||
23129 | Opcode = X86ISD::FMAX; | |||
23130 | break; | |||
23131 | } | |||
23132 | // Check for x CC y ? y : x -- a min/max with reversed arms. | |||
23133 | } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) && | |||
23134 | DAG.isEqualTo(RHS, Cond.getOperand(0))) { | |||
23135 | switch (CC) { | |||
23136 | default: break; | |||
23137 | case ISD::SETOGE: | |||
23138 | // Converting this to a min would handle comparisons between positive | |||
23139 | // and negative zero incorrectly, and swapping the operands would | |||
23140 | // cause it to handle NaNs incorrectly. | |||
23141 | if (!DAG.getTarget().Options.UnsafeFPMath && | |||
23142 | !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) { | |||
23143 | if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) | |||
23144 | break; | |||
23145 | std::swap(LHS, RHS); | |||
23146 | } | |||
23147 | Opcode = X86ISD::FMIN; | |||
23148 | break; | |||
23149 | case ISD::SETUGT: | |||
23150 | // Converting this to a min would handle NaNs incorrectly. | |||
23151 | if (!DAG.getTarget().Options.UnsafeFPMath && | |||
23152 | (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) | |||
23153 | break; | |||
23154 | Opcode = X86ISD::FMIN; | |||
23155 | break; | |||
23156 | case ISD::SETUGE: | |||
23157 | // Converting this to a min would handle both negative zeros and NaNs | |||
23158 | // incorrectly, but we can swap the operands to fix both. | |||
23159 | std::swap(LHS, RHS); | |||
23160 | case ISD::SETOGT: | |||
23161 | case ISD::SETGT: | |||
23162 | case ISD::SETGE: | |||
23163 | Opcode = X86ISD::FMIN; | |||
23164 | break; | |||
23165 | ||||
23166 | case ISD::SETULT: | |||
23167 | // Converting this to a max would handle NaNs incorrectly. | |||
23168 | if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) | |||
23169 | break; | |||
23170 | Opcode = X86ISD::FMAX; | |||
23171 | break; | |||
23172 | case ISD::SETOLE: | |||
23173 | // Converting this to a max would handle comparisons between positive | |||
23174 | // and negative zero incorrectly, and swapping the operands would | |||
23175 | // cause it to handle NaNs incorrectly. | |||
23176 | if (!DAG.getTarget().Options.UnsafeFPMath && | |||
23177 | !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) { | |||
23178 | if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) | |||
23179 | break; | |||
23180 | std::swap(LHS, RHS); | |||
23181 | } | |||
23182 | Opcode = X86ISD::FMAX; | |||
23183 | break; | |||
23184 | case ISD::SETULE: | |||
23185 | // Converting this to a max would handle both negative zeros and NaNs | |||
23186 | // incorrectly, but we can swap the operands to fix both. | |||
23187 | std::swap(LHS, RHS); | |||
23188 | case ISD::SETOLT: | |||
23189 | case ISD::SETLT: | |||
23190 | case ISD::SETLE: | |||
23191 | Opcode = X86ISD::FMAX; | |||
23192 | break; | |||
23193 | } | |||
23194 | } | |||
23195 | ||||
23196 | if (Opcode) | |||
23197 | return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS); | |||
23198 | } | |||
23199 | ||||
23200 | EVT CondVT = Cond.getValueType(); | |||
23201 | if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() && | |||
23202 | CondVT.getVectorElementType() == MVT::i1) { | |||
23203 | // v16i8 (select v16i1, v16i8, v16i8) does not have a proper | |||
23204 | // lowering on KNL. In this case we convert it to | |||
23205 | // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction. | |||
23206 | // The same situation for all 128 and 256-bit vectors of i8 and i16. | |||
23207 | // Since SKX these selects have a proper lowering. | |||
23208 | EVT OpVT = LHS.getValueType(); | |||
23209 | if ((OpVT.is128BitVector() || OpVT.is256BitVector()) && | |||
23210 | (OpVT.getVectorElementType() == MVT::i8 || | |||
23211 | OpVT.getVectorElementType() == MVT::i16) && | |||
23212 | !(Subtarget->hasBWI() && Subtarget->hasVLX())) { | |||
23213 | Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond); | |||
23214 | DCI.AddToWorklist(Cond.getNode()); | |||
23215 | return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS); | |||
23216 | } | |||
23217 | } | |||
23218 | // If this is a select between two integer constants, try to do some | |||
23219 | // optimizations. | |||
23220 | if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) { | |||
23221 | if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS)) | |||
23222 | // Don't do this for crazy integer types. | |||
23223 | if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) { | |||
23224 | // If this is efficiently invertible, canonicalize the LHSC/RHSC values | |||
23225 | // so that TrueC (the true value) is larger than FalseC. | |||
23226 | bool NeedsCondInvert = false; | |||
23227 | ||||
23228 | if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) && | |||
23229 | // Efficiently invertible. | |||
23230 | (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible. | |||
23231 | (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible. | |||
23232 | isa<ConstantSDNode>(Cond.getOperand(1))))) { | |||
23233 | NeedsCondInvert = true; | |||
23234 | std::swap(TrueC, FalseC); | |||
23235 | } | |||
23236 | ||||
23237 | // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0. | |||
23238 | if (FalseC->getAPIntValue() == 0 && | |||
23239 | TrueC->getAPIntValue().isPowerOf2()) { | |||
23240 | if (NeedsCondInvert) // Invert the condition if needed. | |||
23241 | Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond, | |||
23242 | DAG.getConstant(1, Cond.getValueType())); | |||
23243 | ||||
23244 | // Zero extend the condition if needed. | |||
23245 | Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond); | |||
23246 | ||||
23247 | unsigned ShAmt = TrueC->getAPIntValue().logBase2(); | |||
23248 | return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond, | |||
23249 | DAG.getConstant(ShAmt, MVT::i8)); | |||
23250 | } | |||
23251 | ||||
23252 | // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. | |||
23253 | if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) { | |||
23254 | if (NeedsCondInvert) // Invert the condition if needed. | |||
23255 | Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond, | |||
23256 | DAG.getConstant(1, Cond.getValueType())); | |||
23257 | ||||
23258 | // Zero extend the condition if needed. | |||
23259 | Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, | |||
23260 | FalseC->getValueType(0), Cond); | |||
23261 | return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond, | |||
23262 | SDValue(FalseC, 0)); | |||
23263 | } | |||
23264 | ||||
23265 | // Optimize cases that will turn into an LEA instruction. This requires | |||
23266 | // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9). | |||
23267 | if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) { | |||
23268 | uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue(); | |||
23269 | if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff; | |||
23270 | ||||
23271 | bool isFastMultiplier = false; | |||
23272 | if (Diff < 10) { | |||
23273 | switch ((unsigned char)Diff) { | |||
23274 | default: break; | |||
23275 | case 1: // result = add base, cond | |||
23276 | case 2: // result = lea base( , cond*2) | |||
23277 | case 3: // result = lea base(cond, cond*2) | |||
23278 | case 4: // result = lea base( , cond*4) | |||
23279 | case 5: // result = lea base(cond, cond*4) | |||
23280 | case 8: // result = lea base( , cond*8) | |||
23281 | case 9: // result = lea base(cond, cond*8) | |||
23282 | isFastMultiplier = true; | |||
23283 | break; | |||
23284 | } | |||
23285 | } | |||
23286 | ||||
23287 | if (isFastMultiplier) { | |||
23288 | APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue(); | |||
23289 | if (NeedsCondInvert) // Invert the condition if needed. | |||
23290 | Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond, | |||
23291 | DAG.getConstant(1, Cond.getValueType())); | |||
23292 | ||||
23293 | // Zero extend the condition if needed. | |||
23294 | Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0), | |||
23295 | Cond); | |||
23296 | // Scale the condition by the difference. | |||
23297 | if (Diff != 1) | |||
23298 | Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond, | |||
23299 | DAG.getConstant(Diff, Cond.getValueType())); | |||
23300 | ||||
23301 | // Add the base if non-zero. | |||
23302 | if (FalseC->getAPIntValue() != 0) | |||
23303 | Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond, | |||
23304 | SDValue(FalseC, 0)); | |||
23305 | return Cond; | |||
23306 | } | |||
23307 | } | |||
23308 | } | |||
23309 | } | |||
23310 | ||||
23311 | // Canonicalize max and min: | |||
23312 | // (x > y) ? x : y -> (x >= y) ? x : y | |||
23313 | // (x < y) ? x : y -> (x <= y) ? x : y | |||
23314 | // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates | |||
23315 | // the need for an extra compare | |||
23316 | // against zero. e.g. | |||
23317 | // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0 | |||
23318 | // subl %esi, %edi | |||
23319 | // testl %edi, %edi | |||
23320 | // movl $0, %eax | |||
23321 | // cmovgl %edi, %eax | |||
23322 | // => | |||
23323 | // xorl %eax, %eax | |||
23324 | // subl %esi, $edi | |||
23325 | // cmovsl %eax, %edi | |||
23326 | if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC && | |||
23327 | DAG.isEqualTo(LHS, Cond.getOperand(0)) && | |||
23328 | DAG.isEqualTo(RHS, Cond.getOperand(1))) { | |||
23329 | ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get(); | |||
23330 | switch (CC) { | |||
23331 | default: break; | |||
23332 | case ISD::SETLT: | |||
23333 | case ISD::SETGT: { | |||
23334 | ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE; | |||
23335 | Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(), | |||
23336 | Cond.getOperand(0), Cond.getOperand(1), NewCC); | |||
23337 | return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS); | |||
23338 | } | |||
23339 | } | |||
23340 | } | |||
23341 | ||||
23342 | // Early exit check | |||
23343 | if (!TLI.isTypeLegal(VT)) | |||
23344 | return SDValue(); | |||
23345 | ||||
23346 | // Match VSELECTs into subs with unsigned saturation. | |||
23347 | if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC && | |||
23348 | // psubus is available in SSE2 and AVX2 for i8 and i16 vectors. | |||
23349 | ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) || | |||
23350 | (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) { | |||
23351 | ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get(); | |||
23352 | ||||
23353 | // Check if one of the arms of the VSELECT is a zero vector. If it's on the | |||
23354 | // left side invert the predicate to simplify logic below. | |||
23355 | SDValue Other; | |||
23356 | if (ISD::isBuildVectorAllZeros(LHS.getNode())) { | |||
23357 | Other = RHS; | |||
23358 | CC = ISD::getSetCCInverse(CC, true); | |||
23359 | } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) { | |||
23360 | Other = LHS; | |||
23361 | } | |||
23362 | ||||
23363 | if (Other.getNode() && Other->getNumOperands() == 2 && | |||
23364 | DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) { | |||
23365 | SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1); | |||
23366 | SDValue CondRHS = Cond->getOperand(1); | |||
23367 | ||||
23368 | // Look for a general sub with unsigned saturation first. | |||
23369 | // x >= y ? x-y : 0 --> subus x, y | |||
23370 | // x > y ? x-y : 0 --> subus x, y | |||
23371 | if ((CC == ISD::SETUGE || CC == ISD::SETUGT) && | |||
23372 | Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS)) | |||
23373 | return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS); | |||
23374 | ||||
23375 | if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS)) | |||
23376 | if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) { | |||
23377 | if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS)) | |||
23378 | if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode()) | |||
23379 | // If the RHS is a constant we have to reverse the const | |||
23380 | // canonicalization. | |||
23381 | // x > C-1 ? x+-C : 0 --> subus x, C | |||
23382 | if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD && | |||
23383 | CondRHSConst->getAPIntValue() == | |||
23384 | (-OpRHSConst->getAPIntValue() - 1)) | |||
23385 | return DAG.getNode( | |||
23386 | X86ISD::SUBUS, DL, VT, OpLHS, | |||
23387 | DAG.getConstant(-OpRHSConst->getAPIntValue(), VT)); | |||
23388 | ||||
23389 | // Another special case: If C was a sign bit, the sub has been | |||
23390 | // canonicalized into a xor. | |||
23391 | // FIXME: Would it be better to use computeKnownBits to determine | |||
23392 | // whether it's safe to decanonicalize the xor? | |||
23393 | // x s< 0 ? x^C : 0 --> subus x, C | |||
23394 | if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR && | |||
23395 | ISD::isBuildVectorAllZeros(CondRHS.getNode()) && | |||
23396 | OpRHSConst->getAPIntValue().isSignBit()) | |||
23397 | // Note that we have to rebuild the RHS constant here to ensure we | |||
23398 | // don't rely on particular values of undef lanes. | |||
23399 | return DAG.getNode( | |||
23400 | X86ISD::SUBUS, DL, VT, OpLHS, | |||
23401 | DAG.getConstant(OpRHSConst->getAPIntValue(), VT)); | |||
23402 | } | |||
23403 | } | |||
23404 | } | |||
23405 | ||||
23406 | // Try to match a min/max vector operation. | |||
23407 | if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) { | |||
23408 | std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget); | |||
23409 | unsigned Opc = ret.first; | |||
23410 | bool NeedSplit = ret.second; | |||
23411 | ||||
23412 | if (Opc && NeedSplit) { | |||
23413 | unsigned NumElems = VT.getVectorNumElements(); | |||
23414 | // Extract the LHS vectors | |||
23415 | SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL); | |||
23416 | SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL); | |||
23417 | ||||
23418 | // Extract the RHS vectors | |||
23419 | SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL); | |||
23420 | SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL); | |||
23421 | ||||
23422 | // Create min/max for each subvector | |||
23423 | LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1); | |||
23424 | RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2); | |||
23425 | ||||
23426 | // Merge the result | |||
23427 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS); | |||
23428 | } else if (Opc) | |||
23429 | return DAG.getNode(Opc, DL, VT, LHS, RHS); | |||
23430 | } | |||
23431 | ||||
23432 | // Simplify vector selection if condition value type matches vselect | |||
23433 | // operand type | |||
23434 | if (N->getOpcode() == ISD::VSELECT && CondVT == VT) { | |||
23435 | assert(Cond.getValueType().isVector() &&((Cond.getValueType().isVector() && "vector select expects a vector selector!" ) ? static_cast<void> (0) : __assert_fail ("Cond.getValueType().isVector() && \"vector select expects a vector selector!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 23436, __PRETTY_FUNCTION__)) | |||
23436 | "vector select expects a vector selector!")((Cond.getValueType().isVector() && "vector select expects a vector selector!" ) ? static_cast<void> (0) : __assert_fail ("Cond.getValueType().isVector() && \"vector select expects a vector selector!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 23436, __PRETTY_FUNCTION__)); | |||
23437 | ||||
23438 | bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode()); | |||
23439 | bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode()); | |||
23440 | ||||
23441 | // Try invert the condition if true value is not all 1s and false value | |||
23442 | // is not all 0s. | |||
23443 | if (!TValIsAllOnes && !FValIsAllZeros && | |||
23444 | // Check if the selector will be produced by CMPP*/PCMP* | |||
23445 | Cond.getOpcode() == ISD::SETCC && | |||
23446 | // Check if SETCC has already been promoted | |||
23447 | TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT) { | |||
23448 | bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode()); | |||
23449 | bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode()); | |||
23450 | ||||
23451 | if (TValIsAllZeros || FValIsAllOnes) { | |||
23452 | SDValue CC = Cond.getOperand(2); | |||
23453 | ISD::CondCode NewCC = | |||
23454 | ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(), | |||
23455 | Cond.getOperand(0).getValueType().isInteger()); | |||
23456 | Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC); | |||
23457 | std::swap(LHS, RHS); | |||
23458 | TValIsAllOnes = FValIsAllOnes; | |||
23459 | FValIsAllZeros = TValIsAllZeros; | |||
23460 | } | |||
23461 | } | |||
23462 | ||||
23463 | if (TValIsAllOnes || FValIsAllZeros) { | |||
23464 | SDValue Ret; | |||
23465 | ||||
23466 | if (TValIsAllOnes && FValIsAllZeros) | |||
23467 | Ret = Cond; | |||
23468 | else if (TValIsAllOnes) | |||
23469 | Ret = DAG.getNode(ISD::OR, DL, CondVT, Cond, | |||
23470 | DAG.getNode(ISD::BITCAST, DL, CondVT, RHS)); | |||
23471 | else if (FValIsAllZeros) | |||
23472 | Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond, | |||
23473 | DAG.getNode(ISD::BITCAST, DL, CondVT, LHS)); | |||
23474 | ||||
23475 | return DAG.getNode(ISD::BITCAST, DL, VT, Ret); | |||
23476 | } | |||
23477 | } | |||
23478 | ||||
23479 | // If we know that this node is legal then we know that it is going to be | |||
23480 | // matched by one of the SSE/AVX BLEND instructions. These instructions only | |||
23481 | // depend on the highest bit in each word. Try to use SimplifyDemandedBits | |||
23482 | // to simplify previous instructions. | |||
23483 | if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() && | |||
23484 | !DCI.isBeforeLegalize() && | |||
23485 | // We explicitly check against v8i16 and v16i16 because, although | |||
23486 | // they're marked as Custom, they might only be legal when Cond is a | |||
23487 | // build_vector of constants. This will be taken care in a later | |||
23488 | // condition. | |||
23489 | (TLI.isOperationLegalOrCustom(ISD::VSELECT, VT) && VT != MVT::v16i16 && | |||
23490 | VT != MVT::v8i16) && | |||
23491 | // Don't optimize vector of constants. Those are handled by | |||
23492 | // the generic code and all the bits must be properly set for | |||
23493 | // the generic optimizer. | |||
23494 | !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) { | |||
23495 | unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits(); | |||
23496 | ||||
23497 | // Don't optimize vector selects that map to mask-registers. | |||
23498 | if (BitWidth == 1) | |||
23499 | return SDValue(); | |||
23500 | ||||
23501 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 23501, __PRETTY_FUNCTION__)); | |||
23502 | APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1); | |||
23503 | ||||
23504 | APInt KnownZero, KnownOne; | |||
23505 | TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(), | |||
23506 | DCI.isBeforeLegalizeOps()); | |||
23507 | if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) || | |||
23508 | TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne, | |||
23509 | TLO)) { | |||
23510 | // If we changed the computation somewhere in the DAG, this change | |||
23511 | // will affect all users of Cond. | |||
23512 | // Make sure it is fine and update all the nodes so that we do not | |||
23513 | // use the generic VSELECT anymore. Otherwise, we may perform | |||
23514 | // wrong optimizations as we messed up with the actual expectation | |||
23515 | // for the vector boolean values. | |||
23516 | if (Cond != TLO.Old) { | |||
23517 | // Check all uses of that condition operand to check whether it will be | |||
23518 | // consumed by non-BLEND instructions, which may depend on all bits are | |||
23519 | // set properly. | |||
23520 | for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end(); | |||
23521 | I != E; ++I) | |||
23522 | if (I->getOpcode() != ISD::VSELECT) | |||
23523 | // TODO: Add other opcodes eventually lowered into BLEND. | |||
23524 | return SDValue(); | |||
23525 | ||||
23526 | // Update all the users of the condition, before committing the change, | |||
23527 | // so that the VSELECT optimizations that expect the correct vector | |||
23528 | // boolean value will not be triggered. | |||
23529 | for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end(); | |||
23530 | I != E; ++I) | |||
23531 | DAG.ReplaceAllUsesOfValueWith( | |||
23532 | SDValue(*I, 0), | |||
23533 | DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0), | |||
23534 | Cond, I->getOperand(1), I->getOperand(2))); | |||
23535 | DCI.CommitTargetLoweringOpt(TLO); | |||
23536 | return SDValue(); | |||
23537 | } | |||
23538 | // At this point, only Cond is changed. Change the condition | |||
23539 | // just for N to keep the opportunity to optimize all other | |||
23540 | // users their own way. | |||
23541 | DAG.ReplaceAllUsesOfValueWith( | |||
23542 | SDValue(N, 0), | |||
23543 | DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0), | |||
23544 | TLO.New, N->getOperand(1), N->getOperand(2))); | |||
23545 | return SDValue(); | |||
23546 | } | |||
23547 | } | |||
23548 | ||||
23549 | // We should generate an X86ISD::BLENDI from a vselect if its argument | |||
23550 | // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of | |||
23551 | // constants. This specific pattern gets generated when we split a | |||
23552 | // selector for a 512 bit vector in a machine without AVX512 (but with | |||
23553 | // 256-bit vectors), during legalization: | |||
23554 | // | |||
23555 | // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS) | |||
23556 | // | |||
23557 | // Iff we find this pattern and the build_vectors are built from | |||
23558 | // constants, we translate the vselect into a shuffle_vector that we | |||
23559 | // know will be matched by LowerVECTOR_SHUFFLEtoBlend. | |||
23560 | if ((N->getOpcode() == ISD::VSELECT || | |||
23561 | N->getOpcode() == X86ISD::SHRUNKBLEND) && | |||
23562 | !DCI.isBeforeLegalize()) { | |||
23563 | SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget); | |||
23564 | if (Shuffle.getNode()) | |||
23565 | return Shuffle; | |||
23566 | } | |||
23567 | ||||
23568 | return SDValue(); | |||
23569 | } | |||
23570 | ||||
23571 | // Check whether a boolean test is testing a boolean value generated by | |||
23572 | // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition | |||
23573 | // code. | |||
23574 | // | |||
23575 | // Simplify the following patterns: | |||
23576 | // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or | |||
23577 | // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ) | |||
23578 | // to (Op EFLAGS Cond) | |||
23579 | // | |||
23580 | // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or | |||
23581 | // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ) | |||
23582 | // to (Op EFLAGS !Cond) | |||
23583 | // | |||
23584 | // where Op could be BRCOND or CMOV. | |||
23585 | // | |||
23586 | static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) { | |||
23587 | // Quit if not CMP and SUB with its value result used. | |||
23588 | if (Cmp.getOpcode() != X86ISD::CMP && | |||
23589 | (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0))) | |||
23590 | return SDValue(); | |||
23591 | ||||
23592 | // Quit if not used as a boolean value. | |||
23593 | if (CC != X86::COND_E && CC != X86::COND_NE) | |||
23594 | return SDValue(); | |||
23595 | ||||
23596 | // Check CMP operands. One of them should be 0 or 1 and the other should be | |||
23597 | // an SetCC or extended from it. | |||
23598 | SDValue Op1 = Cmp.getOperand(0); | |||
23599 | SDValue Op2 = Cmp.getOperand(1); | |||
23600 | ||||
23601 | SDValue SetCC; | |||
23602 | const ConstantSDNode* C = nullptr; | |||
23603 | bool needOppositeCond = (CC == X86::COND_E); | |||
23604 | bool checkAgainstTrue = false; // Is it a comparison against 1? | |||
23605 | ||||
23606 | if ((C = dyn_cast<ConstantSDNode>(Op1))) | |||
23607 | SetCC = Op2; | |||
23608 | else if ((C = dyn_cast<ConstantSDNode>(Op2))) | |||
23609 | SetCC = Op1; | |||
23610 | else // Quit if all operands are not constants. | |||
23611 | return SDValue(); | |||
23612 | ||||
23613 | if (C->getZExtValue() == 1) { | |||
23614 | needOppositeCond = !needOppositeCond; | |||
23615 | checkAgainstTrue = true; | |||
23616 | } else if (C->getZExtValue() != 0) | |||
23617 | // Quit if the constant is neither 0 or 1. | |||
23618 | return SDValue(); | |||
23619 | ||||
23620 | bool truncatedToBoolWithAnd = false; | |||
23621 | // Skip (zext $x), (trunc $x), or (and $x, 1) node. | |||
23622 | while (SetCC.getOpcode() == ISD::ZERO_EXTEND || | |||
23623 | SetCC.getOpcode() == ISD::TRUNCATE || | |||
23624 | SetCC.getOpcode() == ISD::AND) { | |||
23625 | if (SetCC.getOpcode() == ISD::AND) { | |||
23626 | int OpIdx = -1; | |||
23627 | ConstantSDNode *CS; | |||
23628 | if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) && | |||
23629 | CS->getZExtValue() == 1) | |||
23630 | OpIdx = 1; | |||
23631 | if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) && | |||
23632 | CS->getZExtValue() == 1) | |||
23633 | OpIdx = 0; | |||
23634 | if (OpIdx == -1) | |||
23635 | break; | |||
23636 | SetCC = SetCC.getOperand(OpIdx); | |||
23637 | truncatedToBoolWithAnd = true; | |||
23638 | } else | |||
23639 | SetCC = SetCC.getOperand(0); | |||
23640 | } | |||
23641 | ||||
23642 | switch (SetCC.getOpcode()) { | |||
23643 | case X86ISD::SETCC_CARRY: | |||
23644 | // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to | |||
23645 | // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1, | |||
23646 | // i.e. it's a comparison against true but the result of SETCC_CARRY is not | |||
23647 | // truncated to i1 using 'and'. | |||
23648 | if (checkAgainstTrue && !truncatedToBoolWithAnd) | |||
23649 | break; | |||
23650 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 23651, __PRETTY_FUNCTION__)) | |||
23651 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 23651, __PRETTY_FUNCTION__)); | |||
23652 | // FALL THROUGH | |||
23653 | case X86ISD::SETCC: | |||
23654 | // Set the condition code or opposite one if necessary. | |||
23655 | CC = X86::CondCode(SetCC.getConstantOperandVal(0)); | |||
23656 | if (needOppositeCond) | |||
23657 | CC = X86::GetOppositeBranchCondition(CC); | |||
23658 | return SetCC.getOperand(1); | |||
23659 | case X86ISD::CMOV: { | |||
23660 | // Check whether false/true value has canonical one, i.e. 0 or 1. | |||
23661 | ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0)); | |||
23662 | ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1)); | |||
23663 | // Quit if true value is not a constant. | |||
23664 | if (!TVal) | |||
23665 | return SDValue(); | |||
23666 | // Quit if false value is not a constant. | |||
23667 | if (!FVal) { | |||
23668 | SDValue Op = SetCC.getOperand(0); | |||
23669 | // Skip 'zext' or 'trunc' node. | |||
23670 | if (Op.getOpcode() == ISD::ZERO_EXTEND || | |||
23671 | Op.getOpcode() == ISD::TRUNCATE) | |||
23672 | Op = Op.getOperand(0); | |||
23673 | // A special case for rdrand/rdseed, where 0 is set if false cond is | |||
23674 | // found. | |||
23675 | if ((Op.getOpcode() != X86ISD::RDRAND && | |||
23676 | Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0) | |||
23677 | return SDValue(); | |||
23678 | } | |||
23679 | // Quit if false value is not the constant 0 or 1. | |||
23680 | bool FValIsFalse = true; | |||
23681 | if (FVal && FVal->getZExtValue() != 0) { | |||
23682 | if (FVal->getZExtValue() != 1) | |||
23683 | return SDValue(); | |||
23684 | // If FVal is 1, opposite cond is needed. | |||
23685 | needOppositeCond = !needOppositeCond; | |||
23686 | FValIsFalse = false; | |||
23687 | } | |||
23688 | // Quit if TVal is not the constant opposite of FVal. | |||
23689 | if (FValIsFalse && TVal->getZExtValue() != 1) | |||
23690 | return SDValue(); | |||
23691 | if (!FValIsFalse && TVal->getZExtValue() != 0) | |||
23692 | return SDValue(); | |||
23693 | CC = X86::CondCode(SetCC.getConstantOperandVal(2)); | |||
23694 | if (needOppositeCond) | |||
23695 | CC = X86::GetOppositeBranchCondition(CC); | |||
23696 | return SetCC.getOperand(3); | |||
23697 | } | |||
23698 | } | |||
23699 | ||||
23700 | return SDValue(); | |||
23701 | } | |||
23702 | ||||
23703 | /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL] | |||
23704 | static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG, | |||
23705 | TargetLowering::DAGCombinerInfo &DCI, | |||
23706 | const X86Subtarget *Subtarget) { | |||
23707 | SDLoc DL(N); | |||
23708 | ||||
23709 | // If the flag operand isn't dead, don't touch this CMOV. | |||
23710 | if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty()) | |||
23711 | return SDValue(); | |||
23712 | ||||
23713 | SDValue FalseOp = N->getOperand(0); | |||
23714 | SDValue TrueOp = N->getOperand(1); | |||
23715 | X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2); | |||
23716 | SDValue Cond = N->getOperand(3); | |||
23717 | ||||
23718 | if (CC == X86::COND_E || CC == X86::COND_NE) { | |||
23719 | switch (Cond.getOpcode()) { | |||
23720 | default: break; | |||
23721 | case X86ISD::BSR: | |||
23722 | case X86ISD::BSF: | |||
23723 | // If operand of BSR / BSF are proven never zero, then ZF cannot be set. | |||
23724 | if (DAG.isKnownNeverZero(Cond.getOperand(0))) | |||
23725 | return (CC == X86::COND_E) ? FalseOp : TrueOp; | |||
23726 | } | |||
23727 | } | |||
23728 | ||||
23729 | SDValue Flags; | |||
23730 | ||||
23731 | Flags = checkBoolTestSetCCCombine(Cond, CC); | |||
23732 | if (Flags.getNode() && | |||
23733 | // Extra check as FCMOV only supports a subset of X86 cond. | |||
23734 | (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) { | |||
23735 | SDValue Ops[] = { FalseOp, TrueOp, | |||
23736 | DAG.getConstant(CC, MVT::i8), Flags }; | |||
23737 | return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops); | |||
23738 | } | |||
23739 | ||||
23740 | // If this is a select between two integer constants, try to do some | |||
23741 | // optimizations. Note that the operands are ordered the opposite of SELECT | |||
23742 | // operands. | |||
23743 | if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) { | |||
23744 | if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) { | |||
23745 | // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is | |||
23746 | // larger than FalseC (the false value). | |||
23747 | if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) { | |||
23748 | CC = X86::GetOppositeBranchCondition(CC); | |||
23749 | std::swap(TrueC, FalseC); | |||
23750 | std::swap(TrueOp, FalseOp); | |||
23751 | } | |||
23752 | ||||
23753 | // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0. | |||
23754 | // This is efficient for any integer data type (including i8/i16) and | |||
23755 | // shift amount. | |||
23756 | if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) { | |||
23757 | Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8, | |||
23758 | DAG.getConstant(CC, MVT::i8), Cond); | |||
23759 | ||||
23760 | // Zero extend the condition if needed. | |||
23761 | Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond); | |||
23762 | ||||
23763 | unsigned ShAmt = TrueC->getAPIntValue().logBase2(); | |||
23764 | Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond, | |||
23765 | DAG.getConstant(ShAmt, MVT::i8)); | |||
23766 | if (N->getNumValues() == 2) // Dead flag value? | |||
23767 | return DCI.CombineTo(N, Cond, SDValue()); | |||
23768 | return Cond; | |||
23769 | } | |||
23770 | ||||
23771 | // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient | |||
23772 | // for any integer data type, including i8/i16. | |||
23773 | if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) { | |||
23774 | Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8, | |||
23775 | DAG.getConstant(CC, MVT::i8), Cond); | |||
23776 | ||||
23777 | // Zero extend the condition if needed. | |||
23778 | Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, | |||
23779 | FalseC->getValueType(0), Cond); | |||
23780 | Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond, | |||
23781 | SDValue(FalseC, 0)); | |||
23782 | ||||
23783 | if (N->getNumValues() == 2) // Dead flag value? | |||
23784 | return DCI.CombineTo(N, Cond, SDValue()); | |||
23785 | return Cond; | |||
23786 | } | |||
23787 | ||||
23788 | // Optimize cases that will turn into an LEA instruction. This requires | |||
23789 | // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9). | |||
23790 | if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) { | |||
23791 | uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue(); | |||
23792 | if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff; | |||
23793 | ||||
23794 | bool isFastMultiplier = false; | |||
23795 | if (Diff < 10) { | |||
23796 | switch ((unsigned char)Diff) { | |||
23797 | default: break; | |||
23798 | case 1: // result = add base, cond | |||
23799 | case 2: // result = lea base( , cond*2) | |||
23800 | case 3: // result = lea base(cond, cond*2) | |||
23801 | case 4: // result = lea base( , cond*4) | |||
23802 | case 5: // result = lea base(cond, cond*4) | |||
23803 | case 8: // result = lea base( , cond*8) | |||
23804 | case 9: // result = lea base(cond, cond*8) | |||
23805 | isFastMultiplier = true; | |||
23806 | break; | |||
23807 | } | |||
23808 | } | |||
23809 | ||||
23810 | if (isFastMultiplier) { | |||
23811 | APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue(); | |||
23812 | Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8, | |||
23813 | DAG.getConstant(CC, MVT::i8), Cond); | |||
23814 | // Zero extend the condition if needed. | |||
23815 | Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0), | |||
23816 | Cond); | |||
23817 | // Scale the condition by the difference. | |||
23818 | if (Diff != 1) | |||
23819 | Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond, | |||
23820 | DAG.getConstant(Diff, Cond.getValueType())); | |||
23821 | ||||
23822 | // Add the base if non-zero. | |||
23823 | if (FalseC->getAPIntValue() != 0) | |||
23824 | Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond, | |||
23825 | SDValue(FalseC, 0)); | |||
23826 | if (N->getNumValues() == 2) // Dead flag value? | |||
23827 | return DCI.CombineTo(N, Cond, SDValue()); | |||
23828 | return Cond; | |||
23829 | } | |||
23830 | } | |||
23831 | } | |||
23832 | } | |||
23833 | ||||
23834 | // Handle these cases: | |||
23835 | // (select (x != c), e, c) -> select (x != c), e, x), | |||
23836 | // (select (x == c), c, e) -> select (x == c), x, e) | |||
23837 | // where the c is an integer constant, and the "select" is the combination | |||
23838 | // of CMOV and CMP. | |||
23839 | // | |||
23840 | // The rationale for this change is that the conditional-move from a constant | |||
23841 | // needs two instructions, however, conditional-move from a register needs | |||
23842 | // only one instruction. | |||
23843 | // | |||
23844 | // CAVEAT: By replacing a constant with a symbolic value, it may obscure | |||
23845 | // some instruction-combining opportunities. This opt needs to be | |||
23846 | // postponed as late as possible. | |||
23847 | // | |||
23848 | if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) { | |||
23849 | // the DCI.xxxx conditions are provided to postpone the optimization as | |||
23850 | // late as possible. | |||
23851 | ||||
23852 | ConstantSDNode *CmpAgainst = nullptr; | |||
23853 | if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) && | |||
23854 | (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) && | |||
23855 | !isa<ConstantSDNode>(Cond.getOperand(0))) { | |||
23856 | ||||
23857 | if (CC == X86::COND_NE && | |||
23858 | CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) { | |||
23859 | CC = X86::GetOppositeBranchCondition(CC); | |||
23860 | std::swap(TrueOp, FalseOp); | |||
23861 | } | |||
23862 | ||||
23863 | if (CC == X86::COND_E && | |||
23864 | CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) { | |||
23865 | SDValue Ops[] = { FalseOp, Cond.getOperand(0), | |||
23866 | DAG.getConstant(CC, MVT::i8), Cond }; | |||
23867 | return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops); | |||
23868 | } | |||
23869 | } | |||
23870 | } | |||
23871 | ||||
23872 | return SDValue(); | |||
23873 | } | |||
23874 | ||||
23875 | static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG, | |||
23876 | const X86Subtarget *Subtarget) { | |||
23877 | unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); | |||
23878 | switch (IntNo) { | |||
23879 | default: return SDValue(); | |||
23880 | // SSE/AVX/AVX2 blend intrinsics. | |||
23881 | case Intrinsic::x86_avx2_pblendvb: | |||
23882 | case Intrinsic::x86_avx2_pblendw: | |||
23883 | case Intrinsic::x86_avx2_pblendd_128: | |||
23884 | case Intrinsic::x86_avx2_pblendd_256: | |||
23885 | // Don't try to simplify this intrinsic if we don't have AVX2. | |||
23886 | if (!Subtarget->hasAVX2()) | |||
23887 | return SDValue(); | |||
23888 | // FALL-THROUGH | |||
23889 | case Intrinsic::x86_avx_blend_pd_256: | |||
23890 | case Intrinsic::x86_avx_blend_ps_256: | |||
23891 | case Intrinsic::x86_avx_blendv_pd_256: | |||
23892 | case Intrinsic::x86_avx_blendv_ps_256: | |||
23893 | // Don't try to simplify this intrinsic if we don't have AVX. | |||
23894 | if (!Subtarget->hasAVX()) | |||
23895 | return SDValue(); | |||
23896 | // FALL-THROUGH | |||
23897 | case Intrinsic::x86_sse41_pblendw: | |||
23898 | case Intrinsic::x86_sse41_blendpd: | |||
23899 | case Intrinsic::x86_sse41_blendps: | |||
23900 | case Intrinsic::x86_sse41_blendvps: | |||
23901 | case Intrinsic::x86_sse41_blendvpd: | |||
23902 | case Intrinsic::x86_sse41_pblendvb: { | |||
23903 | SDValue Op0 = N->getOperand(1); | |||
23904 | SDValue Op1 = N->getOperand(2); | |||
23905 | SDValue Mask = N->getOperand(3); | |||
23906 | ||||
23907 | // Don't try to simplify this intrinsic if we don't have SSE4.1. | |||
23908 | if (!Subtarget->hasSSE41()) | |||
23909 | return SDValue(); | |||
23910 | ||||
23911 | // fold (blend A, A, Mask) -> A | |||
23912 | if (Op0 == Op1) | |||
23913 | return Op0; | |||
23914 | // fold (blend A, B, allZeros) -> A | |||
23915 | if (ISD::isBuildVectorAllZeros(Mask.getNode())) | |||
23916 | return Op0; | |||
23917 | // fold (blend A, B, allOnes) -> B | |||
23918 | if (ISD::isBuildVectorAllOnes(Mask.getNode())) | |||
23919 | return Op1; | |||
23920 | ||||
23921 | // Simplify the case where the mask is a constant i32 value. | |||
23922 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) { | |||
23923 | if (C->isNullValue()) | |||
23924 | return Op0; | |||
23925 | if (C->isAllOnesValue()) | |||
23926 | return Op1; | |||
23927 | } | |||
23928 | ||||
23929 | return SDValue(); | |||
23930 | } | |||
23931 | ||||
23932 | // Packed SSE2/AVX2 arithmetic shift immediate intrinsics. | |||
23933 | case Intrinsic::x86_sse2_psrai_w: | |||
23934 | case Intrinsic::x86_sse2_psrai_d: | |||
23935 | case Intrinsic::x86_avx2_psrai_w: | |||
23936 | case Intrinsic::x86_avx2_psrai_d: | |||
23937 | case Intrinsic::x86_sse2_psra_w: | |||
23938 | case Intrinsic::x86_sse2_psra_d: | |||
23939 | case Intrinsic::x86_avx2_psra_w: | |||
23940 | case Intrinsic::x86_avx2_psra_d: { | |||
23941 | SDValue Op0 = N->getOperand(1); | |||
23942 | SDValue Op1 = N->getOperand(2); | |||
23943 | EVT VT = Op0.getValueType(); | |||
23944 | assert(VT.isVector() && "Expected a vector type!")((VT.isVector() && "Expected a vector type!") ? static_cast <void> (0) : __assert_fail ("VT.isVector() && \"Expected a vector type!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 23944, __PRETTY_FUNCTION__)); | |||
23945 | ||||
23946 | if (isa<BuildVectorSDNode>(Op1)) | |||
23947 | Op1 = Op1.getOperand(0); | |||
23948 | ||||
23949 | if (!isa<ConstantSDNode>(Op1)) | |||
23950 | return SDValue(); | |||
23951 | ||||
23952 | EVT SVT = VT.getVectorElementType(); | |||
23953 | unsigned SVTBits = SVT.getSizeInBits(); | |||
23954 | ||||
23955 | ConstantSDNode *CND = cast<ConstantSDNode>(Op1); | |||
23956 | const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue()); | |||
23957 | uint64_t ShAmt = C.getZExtValue(); | |||
23958 | ||||
23959 | // Don't try to convert this shift into a ISD::SRA if the shift | |||
23960 | // count is bigger than or equal to the element size. | |||
23961 | if (ShAmt >= SVTBits) | |||
23962 | return SDValue(); | |||
23963 | ||||
23964 | // Trivial case: if the shift count is zero, then fold this | |||
23965 | // into the first operand. | |||
23966 | if (ShAmt == 0) | |||
23967 | return Op0; | |||
23968 | ||||
23969 | // Replace this packed shift intrinsic with a target independent | |||
23970 | // shift dag node. | |||
23971 | SDValue Splat = DAG.getConstant(C, VT); | |||
23972 | return DAG.getNode(ISD::SRA, SDLoc(N), VT, Op0, Splat); | |||
23973 | } | |||
23974 | } | |||
23975 | } | |||
23976 | ||||
23977 | /// PerformMulCombine - Optimize a single multiply with constant into two | |||
23978 | /// in order to implement it with two cheaper instructions, e.g. | |||
23979 | /// LEA + SHL, LEA + LEA. | |||
23980 | static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG, | |||
23981 | TargetLowering::DAGCombinerInfo &DCI) { | |||
23982 | if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer()) | |||
23983 | return SDValue(); | |||
23984 | ||||
23985 | EVT VT = N->getValueType(0); | |||
23986 | if (VT != MVT::i64 && VT != MVT::i32) | |||
23987 | return SDValue(); | |||
23988 | ||||
23989 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)); | |||
23990 | if (!C) | |||
23991 | return SDValue(); | |||
23992 | uint64_t MulAmt = C->getZExtValue(); | |||
23993 | if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9) | |||
23994 | return SDValue(); | |||
23995 | ||||
23996 | uint64_t MulAmt1 = 0; | |||
23997 | uint64_t MulAmt2 = 0; | |||
23998 | if ((MulAmt % 9) == 0) { | |||
23999 | MulAmt1 = 9; | |||
24000 | MulAmt2 = MulAmt / 9; | |||
24001 | } else if ((MulAmt % 5) == 0) { | |||
24002 | MulAmt1 = 5; | |||
24003 | MulAmt2 = MulAmt / 5; | |||
24004 | } else if ((MulAmt % 3) == 0) { | |||
24005 | MulAmt1 = 3; | |||
24006 | MulAmt2 = MulAmt / 3; | |||
24007 | } | |||
24008 | if (MulAmt2 && | |||
24009 | (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){ | |||
24010 | SDLoc DL(N); | |||
24011 | ||||
24012 | if (isPowerOf2_64(MulAmt2) && | |||
24013 | !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD)) | |||
24014 | // If second multiplifer is pow2, issue it first. We want the multiply by | |||
24015 | // 3, 5, or 9 to be folded into the addressing mode unless the lone use | |||
24016 | // is an add. | |||
24017 | std::swap(MulAmt1, MulAmt2); | |||
24018 | ||||
24019 | SDValue NewMul; | |||
24020 | if (isPowerOf2_64(MulAmt1)) | |||
24021 | NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0), | |||
24022 | DAG.getConstant(Log2_64(MulAmt1), MVT::i8)); | |||
24023 | else | |||
24024 | NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0), | |||
24025 | DAG.getConstant(MulAmt1, VT)); | |||
24026 | ||||
24027 | if (isPowerOf2_64(MulAmt2)) | |||
24028 | NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul, | |||
24029 | DAG.getConstant(Log2_64(MulAmt2), MVT::i8)); | |||
24030 | else | |||
24031 | NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul, | |||
24032 | DAG.getConstant(MulAmt2, VT)); | |||
24033 | ||||
24034 | // Do not add new nodes to DAG combiner worklist. | |||
24035 | DCI.CombineTo(N, NewMul, false); | |||
24036 | } | |||
24037 | return SDValue(); | |||
24038 | } | |||
24039 | ||||
24040 | static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) { | |||
24041 | SDValue N0 = N->getOperand(0); | |||
24042 | SDValue N1 = N->getOperand(1); | |||
24043 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); | |||
24044 | EVT VT = N0.getValueType(); | |||
24045 | ||||
24046 | // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2)) | |||
24047 | // since the result of setcc_c is all zero's or all ones. | |||
24048 | if (VT.isInteger() && !VT.isVector() && | |||
24049 | N1C && N0.getOpcode() == ISD::AND && | |||
24050 | N0.getOperand(1).getOpcode() == ISD::Constant) { | |||
24051 | SDValue N00 = N0.getOperand(0); | |||
24052 | if (N00.getOpcode() == X86ISD::SETCC_CARRY || | |||
24053 | ((N00.getOpcode() == ISD::ANY_EXTEND || | |||
24054 | N00.getOpcode() == ISD::ZERO_EXTEND) && | |||
24055 | N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) { | |||
24056 | APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); | |||
24057 | APInt ShAmt = N1C->getAPIntValue(); | |||
24058 | Mask = Mask.shl(ShAmt); | |||
24059 | if (Mask != 0) | |||
24060 | return DAG.getNode(ISD::AND, SDLoc(N), VT, | |||
24061 | N00, DAG.getConstant(Mask, VT)); | |||
24062 | } | |||
24063 | } | |||
24064 | ||||
24065 | // Hardware support for vector shifts is sparse which makes us scalarize the | |||
24066 | // vector operations in many cases. Also, on sandybridge ADD is faster than | |||
24067 | // shl. | |||
24068 | // (shl V, 1) -> add V,V | |||
24069 | if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1)) | |||
24070 | if (auto *N1SplatC = N1BV->getConstantSplatNode()) { | |||
24071 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24071, __PRETTY_FUNCTION__)); | |||
24072 | // We shift all of the values by one. In many cases we do not have | |||
24073 | // hardware support for this operation. This is better expressed as an ADD | |||
24074 | // of two values. | |||
24075 | if (N1SplatC->getZExtValue() == 1) | |||
24076 | return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0); | |||
24077 | } | |||
24078 | ||||
24079 | return SDValue(); | |||
24080 | } | |||
24081 | ||||
24082 | /// \brief Returns a vector of 0s if the node in input is a vector logical | |||
24083 | /// shift by a constant amount which is known to be bigger than or equal | |||
24084 | /// to the vector element size in bits. | |||
24085 | static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG, | |||
24086 | const X86Subtarget *Subtarget) { | |||
24087 | EVT VT = N->getValueType(0); | |||
24088 | ||||
24089 | if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 && | |||
24090 | (!Subtarget->hasInt256() || | |||
24091 | (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16))) | |||
24092 | return SDValue(); | |||
24093 | ||||
24094 | SDValue Amt = N->getOperand(1); | |||
24095 | SDLoc DL(N); | |||
24096 | if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt)) | |||
24097 | if (auto *AmtSplat = AmtBV->getConstantSplatNode()) { | |||
24098 | APInt ShiftAmt = AmtSplat->getAPIntValue(); | |||
24099 | unsigned MaxAmount = VT.getVectorElementType().getSizeInBits(); | |||
24100 | ||||
24101 | // SSE2/AVX2 logical shifts always return a vector of 0s | |||
24102 | // if the shift amount is bigger than or equal to | |||
24103 | // the element size. The constant shift amount will be | |||
24104 | // encoded as a 8-bit immediate. | |||
24105 | if (ShiftAmt.trunc(8).uge(MaxAmount)) | |||
24106 | return getZeroVector(VT, Subtarget, DAG, DL); | |||
24107 | } | |||
24108 | ||||
24109 | return SDValue(); | |||
24110 | } | |||
24111 | ||||
24112 | /// PerformShiftCombine - Combine shifts. | |||
24113 | static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG, | |||
24114 | TargetLowering::DAGCombinerInfo &DCI, | |||
24115 | const X86Subtarget *Subtarget) { | |||
24116 | if (N->getOpcode() == ISD::SHL) { | |||
24117 | SDValue V = PerformSHLCombine(N, DAG); | |||
24118 | if (V.getNode()) return V; | |||
24119 | } | |||
24120 | ||||
24121 | if (N->getOpcode() != ISD::SRA) { | |||
24122 | // Try to fold this logical shift into a zero vector. | |||
24123 | SDValue V = performShiftToAllZeros(N, DAG, Subtarget); | |||
24124 | if (V.getNode()) return V; | |||
24125 | } | |||
24126 | ||||
24127 | return SDValue(); | |||
24128 | } | |||
24129 | ||||
24130 | // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..)) | |||
24131 | // where both setccs reference the same FP CMP, and rewrite for CMPEQSS | |||
24132 | // and friends. Likewise for OR -> CMPNEQSS. | |||
24133 | static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG, | |||
24134 | TargetLowering::DAGCombinerInfo &DCI, | |||
24135 | const X86Subtarget *Subtarget) { | |||
24136 | unsigned opcode; | |||
24137 | ||||
24138 | // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but | |||
24139 | // we're requiring SSE2 for both. | |||
24140 | if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) { | |||
24141 | SDValue N0 = N->getOperand(0); | |||
24142 | SDValue N1 = N->getOperand(1); | |||
24143 | SDValue CMP0 = N0->getOperand(1); | |||
24144 | SDValue CMP1 = N1->getOperand(1); | |||
24145 | SDLoc DL(N); | |||
24146 | ||||
24147 | // The SETCCs should both refer to the same CMP. | |||
24148 | if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1) | |||
24149 | return SDValue(); | |||
24150 | ||||
24151 | SDValue CMP00 = CMP0->getOperand(0); | |||
24152 | SDValue CMP01 = CMP0->getOperand(1); | |||
24153 | EVT VT = CMP00.getValueType(); | |||
24154 | ||||
24155 | if (VT == MVT::f32 || VT == MVT::f64) { | |||
24156 | bool ExpectingFlags = false; | |||
24157 | // Check for any users that want flags: | |||
24158 | for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); | |||
24159 | !ExpectingFlags && UI != UE; ++UI) | |||
24160 | switch (UI->getOpcode()) { | |||
24161 | default: | |||
24162 | case ISD::BR_CC: | |||
24163 | case ISD::BRCOND: | |||
24164 | case ISD::SELECT: | |||
24165 | ExpectingFlags = true; | |||
24166 | break; | |||
24167 | case ISD::CopyToReg: | |||
24168 | case ISD::SIGN_EXTEND: | |||
24169 | case ISD::ZERO_EXTEND: | |||
24170 | case ISD::ANY_EXTEND: | |||
24171 | break; | |||
24172 | } | |||
24173 | ||||
24174 | if (!ExpectingFlags) { | |||
24175 | enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0); | |||
24176 | enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0); | |||
24177 | ||||
24178 | if (cc1 == X86::COND_E || cc1 == X86::COND_NE) { | |||
24179 | X86::CondCode tmp = cc0; | |||
24180 | cc0 = cc1; | |||
24181 | cc1 = tmp; | |||
24182 | } | |||
24183 | ||||
24184 | if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) || | |||
24185 | (cc0 == X86::COND_NE && cc1 == X86::COND_P)) { | |||
24186 | // FIXME: need symbolic constants for these magic numbers. | |||
24187 | // See X86ATTInstPrinter.cpp:printSSECC(). | |||
24188 | unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4; | |||
24189 | if (Subtarget->hasAVX512()) { | |||
24190 | SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00, | |||
24191 | CMP01, DAG.getConstant(x86cc, MVT::i8)); | |||
24192 | if (N->getValueType(0) != MVT::i1) | |||
24193 | return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0), | |||
24194 | FSetCC); | |||
24195 | return FSetCC; | |||
24196 | } | |||
24197 | SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL, | |||
24198 | CMP00.getValueType(), CMP00, CMP01, | |||
24199 | DAG.getConstant(x86cc, MVT::i8)); | |||
24200 | ||||
24201 | bool is64BitFP = (CMP00.getValueType() == MVT::f64); | |||
24202 | MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32; | |||
24203 | ||||
24204 | if (is64BitFP && !Subtarget->is64Bit()) { | |||
24205 | // On a 32-bit target, we cannot bitcast the 64-bit float to a | |||
24206 | // 64-bit integer, since that's not a legal type. Since | |||
24207 | // OnesOrZeroesF is all ones of all zeroes, we don't need all the | |||
24208 | // bits, but can do this little dance to extract the lowest 32 bits | |||
24209 | // and work with those going forward. | |||
24210 | SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, | |||
24211 | OnesOrZeroesF); | |||
24212 | SDValue Vector32 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, | |||
24213 | Vector64); | |||
24214 | OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, | |||
24215 | Vector32, DAG.getIntPtrConstant(0)); | |||
24216 | IntVT = MVT::i32; | |||
24217 | } | |||
24218 | ||||
24219 | SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT, OnesOrZeroesF); | |||
24220 | SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI, | |||
24221 | DAG.getConstant(1, IntVT)); | |||
24222 | SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed); | |||
24223 | return OneBitOfTruth; | |||
24224 | } | |||
24225 | } | |||
24226 | } | |||
24227 | } | |||
24228 | return SDValue(); | |||
24229 | } | |||
24230 | ||||
24231 | /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector | |||
24232 | /// so it can be folded inside ANDNP. | |||
24233 | static bool CanFoldXORWithAllOnes(const SDNode *N) { | |||
24234 | EVT VT = N->getValueType(0); | |||
24235 | ||||
24236 | // Match direct AllOnes for 128 and 256-bit vectors | |||
24237 | if (ISD::isBuildVectorAllOnes(N)) | |||
24238 | return true; | |||
24239 | ||||
24240 | // Look through a bit convert. | |||
24241 | if (N->getOpcode() == ISD::BITCAST) | |||
24242 | N = N->getOperand(0).getNode(); | |||
24243 | ||||
24244 | // Sometimes the operand may come from a insert_subvector building a 256-bit | |||
24245 | // allones vector | |||
24246 | if (VT.is256BitVector() && | |||
24247 | N->getOpcode() == ISD::INSERT_SUBVECTOR) { | |||
24248 | SDValue V1 = N->getOperand(0); | |||
24249 | SDValue V2 = N->getOperand(1); | |||
24250 | ||||
24251 | if (V1.getOpcode() == ISD::INSERT_SUBVECTOR && | |||
24252 | V1.getOperand(0).getOpcode() == ISD::UNDEF && | |||
24253 | ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) && | |||
24254 | ISD::isBuildVectorAllOnes(V2.getNode())) | |||
24255 | return true; | |||
24256 | } | |||
24257 | ||||
24258 | return false; | |||
24259 | } | |||
24260 | ||||
24261 | // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized | |||
24262 | // register. In most cases we actually compare or select YMM-sized registers | |||
24263 | // and mixing the two types creates horrible code. This method optimizes | |||
24264 | // some of the transition sequences. | |||
24265 | static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG, | |||
24266 | TargetLowering::DAGCombinerInfo &DCI, | |||
24267 | const X86Subtarget *Subtarget) { | |||
24268 | EVT VT = N->getValueType(0); | |||
24269 | if (!VT.is256BitVector()) | |||
24270 | return SDValue(); | |||
24271 | ||||
24272 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24274, __PRETTY_FUNCTION__)) | |||
24273 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24274, __PRETTY_FUNCTION__)) | |||
24274 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24274, __PRETTY_FUNCTION__)); | |||
24275 | ||||
24276 | SDValue Narrow = N->getOperand(0); | |||
24277 | EVT NarrowVT = Narrow->getValueType(0); | |||
24278 | if (!NarrowVT.is128BitVector()) | |||
24279 | return SDValue(); | |||
24280 | ||||
24281 | if (Narrow->getOpcode() != ISD::XOR && | |||
24282 | Narrow->getOpcode() != ISD::AND && | |||
24283 | Narrow->getOpcode() != ISD::OR) | |||
24284 | return SDValue(); | |||
24285 | ||||
24286 | SDValue N0 = Narrow->getOperand(0); | |||
24287 | SDValue N1 = Narrow->getOperand(1); | |||
24288 | SDLoc DL(Narrow); | |||
24289 | ||||
24290 | // The Left side has to be a trunc. | |||
24291 | if (N0.getOpcode() != ISD::TRUNCATE) | |||
24292 | return SDValue(); | |||
24293 | ||||
24294 | // The type of the truncated inputs. | |||
24295 | EVT WideVT = N0->getOperand(0)->getValueType(0); | |||
24296 | if (WideVT != VT) | |||
24297 | return SDValue(); | |||
24298 | ||||
24299 | // The right side has to be a 'trunc' or a constant vector. | |||
24300 | bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE; | |||
24301 | ConstantSDNode *RHSConstSplat = nullptr; | |||
24302 | if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1)) | |||
24303 | RHSConstSplat = RHSBV->getConstantSplatNode(); | |||
24304 | if (!RHSTrunc && !RHSConstSplat) | |||
24305 | return SDValue(); | |||
24306 | ||||
24307 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
24308 | ||||
24309 | if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT)) | |||
24310 | return SDValue(); | |||
24311 | ||||
24312 | // Set N0 and N1 to hold the inputs to the new wide operation. | |||
24313 | N0 = N0->getOperand(0); | |||
24314 | if (RHSConstSplat) { | |||
24315 | N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(), | |||
24316 | SDValue(RHSConstSplat, 0)); | |||
24317 | SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1); | |||
24318 | N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C); | |||
24319 | } else if (RHSTrunc) { | |||
24320 | N1 = N1->getOperand(0); | |||
24321 | } | |||
24322 | ||||
24323 | // Generate the wide operation. | |||
24324 | SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1); | |||
24325 | unsigned Opcode = N->getOpcode(); | |||
24326 | switch (Opcode) { | |||
24327 | case ISD::ANY_EXTEND: | |||
24328 | return Op; | |||
24329 | case ISD::ZERO_EXTEND: { | |||
24330 | unsigned InBits = NarrowVT.getScalarType().getSizeInBits(); | |||
24331 | APInt Mask = APInt::getAllOnesValue(InBits); | |||
24332 | Mask = Mask.zext(VT.getScalarType().getSizeInBits()); | |||
24333 | return DAG.getNode(ISD::AND, DL, VT, | |||
24334 | Op, DAG.getConstant(Mask, VT)); | |||
24335 | } | |||
24336 | case ISD::SIGN_EXTEND: | |||
24337 | return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, | |||
24338 | Op, DAG.getValueType(NarrowVT)); | |||
24339 | default: | |||
24340 | llvm_unreachable("Unexpected opcode")::llvm::llvm_unreachable_internal("Unexpected opcode", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24340); | |||
24341 | } | |||
24342 | } | |||
24343 | ||||
24344 | static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG, | |||
24345 | TargetLowering::DAGCombinerInfo &DCI, | |||
24346 | const X86Subtarget *Subtarget) { | |||
24347 | EVT VT = N->getValueType(0); | |||
24348 | if (DCI.isBeforeLegalizeOps()) | |||
24349 | return SDValue(); | |||
24350 | ||||
24351 | SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget); | |||
24352 | if (R.getNode()) | |||
24353 | return R; | |||
24354 | ||||
24355 | // Create BEXTR instructions | |||
24356 | // BEXTR is ((X >> imm) & (2**size-1)) | |||
24357 | if (VT == MVT::i32 || VT == MVT::i64) { | |||
24358 | SDValue N0 = N->getOperand(0); | |||
24359 | SDValue N1 = N->getOperand(1); | |||
24360 | SDLoc DL(N); | |||
24361 | ||||
24362 | // Check for BEXTR. | |||
24363 | if ((Subtarget->hasBMI() || Subtarget->hasTBM()) && | |||
24364 | (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) { | |||
24365 | ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1); | |||
24366 | ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1)); | |||
24367 | if (MaskNode && ShiftNode) { | |||
24368 | uint64_t Mask = MaskNode->getZExtValue(); | |||
24369 | uint64_t Shift = ShiftNode->getZExtValue(); | |||
24370 | if (isMask_64(Mask)) { | |||
24371 | uint64_t MaskSize = CountPopulation_64(Mask); | |||
24372 | if (Shift + MaskSize <= VT.getSizeInBits()) | |||
24373 | return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0), | |||
24374 | DAG.getConstant(Shift | (MaskSize << 8), VT)); | |||
24375 | } | |||
24376 | } | |||
24377 | } // BEXTR | |||
24378 | ||||
24379 | return SDValue(); | |||
24380 | } | |||
24381 | ||||
24382 | // Want to form ANDNP nodes: | |||
24383 | // 1) In the hopes of then easily combining them with OR and AND nodes | |||
24384 | // to form PBLEND/PSIGN. | |||
24385 | // 2) To match ANDN packed intrinsics | |||
24386 | if (VT != MVT::v2i64 && VT != MVT::v4i64) | |||
24387 | return SDValue(); | |||
24388 | ||||
24389 | SDValue N0 = N->getOperand(0); | |||
24390 | SDValue N1 = N->getOperand(1); | |||
24391 | SDLoc DL(N); | |||
24392 | ||||
24393 | // Check LHS for vnot | |||
24394 | if (N0.getOpcode() == ISD::XOR && | |||
24395 | //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode())) | |||
24396 | CanFoldXORWithAllOnes(N0.getOperand(1).getNode())) | |||
24397 | return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1); | |||
24398 | ||||
24399 | // Check RHS for vnot | |||
24400 | if (N1.getOpcode() == ISD::XOR && | |||
24401 | //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode())) | |||
24402 | CanFoldXORWithAllOnes(N1.getOperand(1).getNode())) | |||
24403 | return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0); | |||
24404 | ||||
24405 | return SDValue(); | |||
24406 | } | |||
24407 | ||||
24408 | static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG, | |||
24409 | TargetLowering::DAGCombinerInfo &DCI, | |||
24410 | const X86Subtarget *Subtarget) { | |||
24411 | if (DCI.isBeforeLegalizeOps()) | |||
24412 | return SDValue(); | |||
24413 | ||||
24414 | SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget); | |||
24415 | if (R.getNode()) | |||
24416 | return R; | |||
24417 | ||||
24418 | SDValue N0 = N->getOperand(0); | |||
24419 | SDValue N1 = N->getOperand(1); | |||
24420 | EVT VT = N->getValueType(0); | |||
24421 | ||||
24422 | // look for psign/blend | |||
24423 | if (VT == MVT::v2i64 || VT == MVT::v4i64) { | |||
24424 | if (!Subtarget->hasSSSE3() || | |||
24425 | (VT == MVT::v4i64 && !Subtarget->hasInt256())) | |||
24426 | return SDValue(); | |||
24427 | ||||
24428 | // Canonicalize pandn to RHS | |||
24429 | if (N0.getOpcode() == X86ISD::ANDNP) | |||
24430 | std::swap(N0, N1); | |||
24431 | // or (and (m, y), (pandn m, x)) | |||
24432 | if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) { | |||
24433 | SDValue Mask = N1.getOperand(0); | |||
24434 | SDValue X = N1.getOperand(1); | |||
24435 | SDValue Y; | |||
24436 | if (N0.getOperand(0) == Mask) | |||
24437 | Y = N0.getOperand(1); | |||
24438 | if (N0.getOperand(1) == Mask) | |||
24439 | Y = N0.getOperand(0); | |||
24440 | ||||
24441 | // Check to see if the mask appeared in both the AND and ANDNP and | |||
24442 | if (!Y.getNode()) | |||
24443 | return SDValue(); | |||
24444 | ||||
24445 | // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them. | |||
24446 | // Look through mask bitcast. | |||
24447 | if (Mask.getOpcode() == ISD::BITCAST) | |||
24448 | Mask = Mask.getOperand(0); | |||
24449 | if (X.getOpcode() == ISD::BITCAST) | |||
24450 | X = X.getOperand(0); | |||
24451 | if (Y.getOpcode() == ISD::BITCAST) | |||
24452 | Y = Y.getOperand(0); | |||
24453 | ||||
24454 | EVT MaskVT = Mask.getValueType(); | |||
24455 | ||||
24456 | // Validate that the Mask operand is a vector sra node. | |||
24457 | // FIXME: what to do for bytes, since there is a psignb/pblendvb, but | |||
24458 | // there is no psrai.b | |||
24459 | unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits(); | |||
24460 | unsigned SraAmt = ~0; | |||
24461 | if (Mask.getOpcode() == ISD::SRA) { | |||
24462 | if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1))) | |||
24463 | if (auto *AmtConst = AmtBV->getConstantSplatNode()) | |||
24464 | SraAmt = AmtConst->getZExtValue(); | |||
24465 | } else if (Mask.getOpcode() == X86ISD::VSRAI) { | |||
24466 | SDValue SraC = Mask.getOperand(1); | |||
24467 | SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue(); | |||
24468 | } | |||
24469 | if ((SraAmt + 1) != EltBits) | |||
24470 | return SDValue(); | |||
24471 | ||||
24472 | SDLoc DL(N); | |||
24473 | ||||
24474 | // Now we know we at least have a plendvb with the mask val. See if | |||
24475 | // we can form a psignb/w/d. | |||
24476 | // psign = x.type == y.type == mask.type && y = sub(0, x); | |||
24477 | if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X && | |||
24478 | ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) && | |||
24479 | X.getValueType() == MaskVT && Y.getValueType() == MaskVT) { | |||
24480 | assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&(((EltBits == 8 || EltBits == 16 || EltBits == 32) && "Unsupported VT for PSIGN") ? static_cast<void> (0) : __assert_fail ("(EltBits == 8 || EltBits == 16 || EltBits == 32) && \"Unsupported VT for PSIGN\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24481, __PRETTY_FUNCTION__)) | |||
24481 | "Unsupported VT for PSIGN")(((EltBits == 8 || EltBits == 16 || EltBits == 32) && "Unsupported VT for PSIGN") ? static_cast<void> (0) : __assert_fail ("(EltBits == 8 || EltBits == 16 || EltBits == 32) && \"Unsupported VT for PSIGN\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24481, __PRETTY_FUNCTION__)); | |||
24482 | Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0)); | |||
24483 | return DAG.getNode(ISD::BITCAST, DL, VT, Mask); | |||
24484 | } | |||
24485 | // PBLENDVB only available on SSE 4.1 | |||
24486 | if (!Subtarget->hasSSE41()) | |||
24487 | return SDValue(); | |||
24488 | ||||
24489 | EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8; | |||
24490 | ||||
24491 | X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X); | |||
24492 | Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y); | |||
24493 | Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask); | |||
24494 | Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X); | |||
24495 | return DAG.getNode(ISD::BITCAST, DL, VT, Mask); | |||
24496 | } | |||
24497 | } | |||
24498 | ||||
24499 | if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64) | |||
24500 | return SDValue(); | |||
24501 | ||||
24502 | // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c) | |||
24503 | MachineFunction &MF = DAG.getMachineFunction(); | |||
24504 | bool OptForSize = MF.getFunction()->getAttributes(). | |||
24505 | hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize); | |||
24506 | ||||
24507 | // SHLD/SHRD instructions have lower register pressure, but on some | |||
24508 | // platforms they have higher latency than the equivalent | |||
24509 | // series of shifts/or that would otherwise be generated. | |||
24510 | // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions | |||
24511 | // have higher latencies and we are not optimizing for size. | |||
24512 | if (!OptForSize && Subtarget->isSHLDSlow()) | |||
24513 | return SDValue(); | |||
24514 | ||||
24515 | if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL) | |||
24516 | std::swap(N0, N1); | |||
24517 | if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL) | |||
24518 | return SDValue(); | |||
24519 | if (!N0.hasOneUse() || !N1.hasOneUse()) | |||
24520 | return SDValue(); | |||
24521 | ||||
24522 | SDValue ShAmt0 = N0.getOperand(1); | |||
24523 | if (ShAmt0.getValueType() != MVT::i8) | |||
24524 | return SDValue(); | |||
24525 | SDValue ShAmt1 = N1.getOperand(1); | |||
24526 | if (ShAmt1.getValueType() != MVT::i8) | |||
24527 | return SDValue(); | |||
24528 | if (ShAmt0.getOpcode() == ISD::TRUNCATE) | |||
24529 | ShAmt0 = ShAmt0.getOperand(0); | |||
24530 | if (ShAmt1.getOpcode() == ISD::TRUNCATE) | |||
24531 | ShAmt1 = ShAmt1.getOperand(0); | |||
24532 | ||||
24533 | SDLoc DL(N); | |||
24534 | unsigned Opc = X86ISD::SHLD; | |||
24535 | SDValue Op0 = N0.getOperand(0); | |||
24536 | SDValue Op1 = N1.getOperand(0); | |||
24537 | if (ShAmt0.getOpcode() == ISD::SUB) { | |||
24538 | Opc = X86ISD::SHRD; | |||
24539 | std::swap(Op0, Op1); | |||
24540 | std::swap(ShAmt0, ShAmt1); | |||
24541 | } | |||
24542 | ||||
24543 | unsigned Bits = VT.getSizeInBits(); | |||
24544 | if (ShAmt1.getOpcode() == ISD::SUB) { | |||
24545 | SDValue Sum = ShAmt1.getOperand(0); | |||
24546 | if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) { | |||
24547 | SDValue ShAmt1Op1 = ShAmt1.getOperand(1); | |||
24548 | if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE) | |||
24549 | ShAmt1Op1 = ShAmt1Op1.getOperand(0); | |||
24550 | if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0) | |||
24551 | return DAG.getNode(Opc, DL, VT, | |||
24552 | Op0, Op1, | |||
24553 | DAG.getNode(ISD::TRUNCATE, DL, | |||
24554 | MVT::i8, ShAmt0)); | |||
24555 | } | |||
24556 | } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) { | |||
24557 | ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0); | |||
24558 | if (ShAmt0C && | |||
24559 | ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits) | |||
24560 | return DAG.getNode(Opc, DL, VT, | |||
24561 | N0.getOperand(0), N1.getOperand(0), | |||
24562 | DAG.getNode(ISD::TRUNCATE, DL, | |||
24563 | MVT::i8, ShAmt0)); | |||
24564 | } | |||
24565 | ||||
24566 | return SDValue(); | |||
24567 | } | |||
24568 | ||||
24569 | // Generate NEG and CMOV for integer abs. | |||
24570 | static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) { | |||
24571 | EVT VT = N->getValueType(0); | |||
24572 | ||||
24573 | // Since X86 does not have CMOV for 8-bit integer, we don't convert | |||
24574 | // 8-bit integer abs to NEG and CMOV. | |||
24575 | if (VT.isInteger() && VT.getSizeInBits() == 8) | |||
24576 | return SDValue(); | |||
24577 | ||||
24578 | SDValue N0 = N->getOperand(0); | |||
24579 | SDValue N1 = N->getOperand(1); | |||
24580 | SDLoc DL(N); | |||
24581 | ||||
24582 | // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1) | |||
24583 | // and change it to SUB and CMOV. | |||
24584 | if (VT.isInteger() && N->getOpcode() == ISD::XOR && | |||
24585 | N0.getOpcode() == ISD::ADD && | |||
24586 | N0.getOperand(1) == N1 && | |||
24587 | N1.getOpcode() == ISD::SRA && | |||
24588 | N1.getOperand(0) == N0.getOperand(0)) | |||
24589 | if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1))) | |||
24590 | if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) { | |||
24591 | // Generate SUB & CMOV. | |||
24592 | SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32), | |||
24593 | DAG.getConstant(0, VT), N0.getOperand(0)); | |||
24594 | ||||
24595 | SDValue Ops[] = { N0.getOperand(0), Neg, | |||
24596 | DAG.getConstant(X86::COND_GE, MVT::i8), | |||
24597 | SDValue(Neg.getNode(), 1) }; | |||
24598 | return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops); | |||
24599 | } | |||
24600 | return SDValue(); | |||
24601 | } | |||
24602 | ||||
24603 | // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes | |||
24604 | static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG, | |||
24605 | TargetLowering::DAGCombinerInfo &DCI, | |||
24606 | const X86Subtarget *Subtarget) { | |||
24607 | if (DCI.isBeforeLegalizeOps()) | |||
24608 | return SDValue(); | |||
24609 | ||||
24610 | if (Subtarget->hasCMov()) { | |||
24611 | SDValue RV = performIntegerAbsCombine(N, DAG); | |||
24612 | if (RV.getNode()) | |||
24613 | return RV; | |||
24614 | } | |||
24615 | ||||
24616 | return SDValue(); | |||
24617 | } | |||
24618 | ||||
24619 | /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes. | |||
24620 | static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG, | |||
24621 | TargetLowering::DAGCombinerInfo &DCI, | |||
24622 | const X86Subtarget *Subtarget) { | |||
24623 | LoadSDNode *Ld = cast<LoadSDNode>(N); | |||
24624 | EVT RegVT = Ld->getValueType(0); | |||
24625 | EVT MemVT = Ld->getMemoryVT(); | |||
24626 | SDLoc dl(Ld); | |||
24627 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
24628 | ||||
24629 | // For chips with slow 32-byte unaligned loads, break the 32-byte operation | |||
24630 | // into two 16-byte operations. | |||
24631 | ISD::LoadExtType Ext = Ld->getExtensionType(); | |||
24632 | unsigned Alignment = Ld->getAlignment(); | |||
24633 | bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8; | |||
24634 | if (RegVT.is256BitVector() && Subtarget->isUnalignedMem32Slow() && | |||
24635 | !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) { | |||
24636 | unsigned NumElems = RegVT.getVectorNumElements(); | |||
24637 | if (NumElems < 2) | |||
24638 | return SDValue(); | |||
24639 | ||||
24640 | SDValue Ptr = Ld->getBasePtr(); | |||
24641 | SDValue Increment = DAG.getConstant(16, TLI.getPointerTy()); | |||
24642 | ||||
24643 | EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(), | |||
24644 | NumElems/2); | |||
24645 | SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr, | |||
24646 | Ld->getPointerInfo(), Ld->isVolatile(), | |||
24647 | Ld->isNonTemporal(), Ld->isInvariant(), | |||
24648 | Alignment); | |||
24649 | Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment); | |||
24650 | SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr, | |||
24651 | Ld->getPointerInfo(), Ld->isVolatile(), | |||
24652 | Ld->isNonTemporal(), Ld->isInvariant(), | |||
24653 | std::min(16U, Alignment)); | |||
24654 | SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, | |||
24655 | Load1.getValue(1), | |||
24656 | Load2.getValue(1)); | |||
24657 | ||||
24658 | SDValue NewVec = DAG.getUNDEF(RegVT); | |||
24659 | NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl); | |||
24660 | NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl); | |||
24661 | return DCI.CombineTo(N, NewVec, TF, true); | |||
24662 | } | |||
24663 | ||||
24664 | return SDValue(); | |||
24665 | } | |||
24666 | ||||
24667 | /// PerformMLOADCombine - Resolve extending loads | |||
24668 | static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG, | |||
24669 | TargetLowering::DAGCombinerInfo &DCI, | |||
24670 | const X86Subtarget *Subtarget) { | |||
24671 | MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N); | |||
24672 | if (Mld->getExtensionType() != ISD::SEXTLOAD) | |||
24673 | return SDValue(); | |||
24674 | ||||
24675 | EVT VT = Mld->getValueType(0); | |||
24676 | unsigned NumElems = VT.getVectorNumElements(); | |||
24677 | EVT LdVT = Mld->getMemoryVT(); | |||
24678 | SDLoc dl(Mld); | |||
24679 | ||||
24680 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24680, __PRETTY_FUNCTION__)); | |||
24681 | unsigned ToSz = VT.getVectorElementType().getSizeInBits(); | |||
24682 | unsigned FromSz = LdVT.getVectorElementType().getSizeInBits(); | |||
24683 | // From, To sizes and ElemCount must be pow of two | |||
24684 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24685, __PRETTY_FUNCTION__)) | |||
24685 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24685, __PRETTY_FUNCTION__)); | |||
24686 | ||||
24687 | unsigned SizeRatio = ToSz / FromSz; | |||
24688 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24688, __PRETTY_FUNCTION__)); | |||
24689 | ||||
24690 | // Create a type on which we perform the shuffle | |||
24691 | EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), | |||
24692 | LdVT.getScalarType(), NumElems*SizeRatio); | |||
24693 | assert(WideVecVT.getSizeInBits() == VT.getSizeInBits())((WideVecVT.getSizeInBits() == VT.getSizeInBits()) ? static_cast <void> (0) : __assert_fail ("WideVecVT.getSizeInBits() == VT.getSizeInBits()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24693, __PRETTY_FUNCTION__)); | |||
24694 | ||||
24695 | // Convert Src0 value | |||
24696 | SDValue WideSrc0 = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mld->getSrc0()); | |||
24697 | if (Mld->getSrc0().getOpcode() != ISD::UNDEF) { | |||
24698 | SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1); | |||
24699 | for (unsigned i = 0; i != NumElems; ++i) | |||
24700 | ShuffleVec[i] = i * SizeRatio; | |||
24701 | ||||
24702 | // Can't shuffle using an illegal type. | |||
24703 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24704, __PRETTY_FUNCTION__)) | |||
24704 | && "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24704, __PRETTY_FUNCTION__)); | |||
24705 | WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0, | |||
24706 | DAG.getUNDEF(WideVecVT), &ShuffleVec[0]); | |||
24707 | } | |||
24708 | // Prepare the new mask | |||
24709 | SDValue NewMask; | |||
24710 | SDValue Mask = Mld->getMask(); | |||
24711 | if (Mask.getValueType() == VT) { | |||
24712 | // Mask and original value have the same type | |||
24713 | NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask); | |||
24714 | SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1); | |||
24715 | for (unsigned i = 0; i != NumElems; ++i) | |||
24716 | ShuffleVec[i] = i * SizeRatio; | |||
24717 | for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i) | |||
24718 | ShuffleVec[i] = NumElems*SizeRatio; | |||
24719 | NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask, | |||
24720 | DAG.getConstant(0, WideVecVT), | |||
24721 | &ShuffleVec[0]); | |||
24722 | } | |||
24723 | else { | |||
24724 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24724, __PRETTY_FUNCTION__)); | |||
24725 | unsigned WidenNumElts = NumElems*SizeRatio; | |||
24726 | unsigned MaskNumElts = VT.getVectorNumElements(); | |||
24727 | EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, | |||
24728 | WidenNumElts); | |||
24729 | ||||
24730 | unsigned NumConcat = WidenNumElts / MaskNumElts; | |||
24731 | SmallVector<SDValue, 16> Ops(NumConcat); | |||
24732 | SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType()); | |||
24733 | Ops[0] = Mask; | |||
24734 | for (unsigned i = 1; i != NumConcat; ++i) | |||
24735 | Ops[i] = ZeroVal; | |||
24736 | ||||
24737 | NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops); | |||
24738 | } | |||
24739 | ||||
24740 | SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(), | |||
24741 | Mld->getBasePtr(), NewMask, WideSrc0, | |||
24742 | Mld->getMemoryVT(), Mld->getMemOperand(), | |||
24743 | ISD::NON_EXTLOAD); | |||
24744 | SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd); | |||
24745 | return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true); | |||
24746 | ||||
24747 | } | |||
24748 | /// PerformMSTORECombine - Resolve truncating stores | |||
24749 | static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG, | |||
24750 | const X86Subtarget *Subtarget) { | |||
24751 | MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N); | |||
24752 | if (!Mst->isTruncatingStore()) | |||
24753 | return SDValue(); | |||
24754 | ||||
24755 | EVT VT = Mst->getValue().getValueType(); | |||
24756 | unsigned NumElems = VT.getVectorNumElements(); | |||
24757 | EVT StVT = Mst->getMemoryVT(); | |||
24758 | SDLoc dl(Mst); | |||
24759 | ||||
24760 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24760, __PRETTY_FUNCTION__)); | |||
24761 | unsigned FromSz = VT.getVectorElementType().getSizeInBits(); | |||
24762 | unsigned ToSz = StVT.getVectorElementType().getSizeInBits(); | |||
24763 | ||||
24764 | // From, To sizes and ElemCount must be pow of two | |||
24765 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24766, __PRETTY_FUNCTION__)) | |||
24766 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24766, __PRETTY_FUNCTION__)); | |||
24767 | // We are going to use the original vector elt for storing. | |||
24768 | // Accumulated smaller vector elements must be a multiple of the store size. | |||
24769 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24770, __PRETTY_FUNCTION__)) | |||
24770 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24770, __PRETTY_FUNCTION__)); | |||
24771 | ||||
24772 | unsigned SizeRatio = FromSz / ToSz; | |||
24773 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24773, __PRETTY_FUNCTION__)); | |||
24774 | ||||
24775 | // Create a type on which we perform the shuffle | |||
24776 | EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), | |||
24777 | StVT.getScalarType(), NumElems*SizeRatio); | |||
24778 | ||||
24779 | assert(WideVecVT.getSizeInBits() == VT.getSizeInBits())((WideVecVT.getSizeInBits() == VT.getSizeInBits()) ? static_cast <void> (0) : __assert_fail ("WideVecVT.getSizeInBits() == VT.getSizeInBits()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24779, __PRETTY_FUNCTION__)); | |||
24780 | ||||
24781 | SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mst->getValue()); | |||
24782 | SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1); | |||
24783 | for (unsigned i = 0; i != NumElems; ++i) | |||
24784 | ShuffleVec[i] = i * SizeRatio; | |||
24785 | ||||
24786 | // Can't shuffle using an illegal type. | |||
24787 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24788, __PRETTY_FUNCTION__)) | |||
24788 | && "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24788, __PRETTY_FUNCTION__)); | |||
24789 | ||||
24790 | SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec, | |||
24791 | DAG.getUNDEF(WideVecVT), | |||
24792 | &ShuffleVec[0]); | |||
24793 | ||||
24794 | SDValue NewMask; | |||
24795 | SDValue Mask = Mst->getMask(); | |||
24796 | if (Mask.getValueType() == VT) { | |||
24797 | // Mask and original value have the same type | |||
24798 | NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask); | |||
24799 | for (unsigned i = 0; i != NumElems; ++i) | |||
24800 | ShuffleVec[i] = i * SizeRatio; | |||
24801 | for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i) | |||
24802 | ShuffleVec[i] = NumElems*SizeRatio; | |||
24803 | NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask, | |||
24804 | DAG.getConstant(0, WideVecVT), | |||
24805 | &ShuffleVec[0]); | |||
24806 | } | |||
24807 | else { | |||
24808 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24808, __PRETTY_FUNCTION__)); | |||
24809 | unsigned WidenNumElts = NumElems*SizeRatio; | |||
24810 | unsigned MaskNumElts = VT.getVectorNumElements(); | |||
24811 | EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, | |||
24812 | WidenNumElts); | |||
24813 | ||||
24814 | unsigned NumConcat = WidenNumElts / MaskNumElts; | |||
24815 | SmallVector<SDValue, 16> Ops(NumConcat); | |||
24816 | SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType()); | |||
24817 | Ops[0] = Mask; | |||
24818 | for (unsigned i = 1; i != NumConcat; ++i) | |||
24819 | Ops[i] = ZeroVal; | |||
24820 | ||||
24821 | NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops); | |||
24822 | } | |||
24823 | ||||
24824 | return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(), | |||
24825 | NewMask, StVT, Mst->getMemOperand(), false); | |||
24826 | } | |||
24827 | /// PerformSTORECombine - Do target-specific dag combines on STORE nodes. | |||
24828 | static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG, | |||
24829 | const X86Subtarget *Subtarget) { | |||
24830 | StoreSDNode *St = cast<StoreSDNode>(N); | |||
24831 | EVT VT = St->getValue().getValueType(); | |||
24832 | EVT StVT = St->getMemoryVT(); | |||
24833 | SDLoc dl(St); | |||
24834 | SDValue StoredVal = St->getOperand(1); | |||
24835 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
24836 | ||||
24837 | // If we are saving a concatenation of two XMM registers and 32-byte stores | |||
24838 | // are slow, such as on Sandy Bridge, perform two 16-byte stores. | |||
24839 | unsigned Alignment = St->getAlignment(); | |||
24840 | bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8; | |||
24841 | if (VT.is256BitVector() && Subtarget->isUnalignedMem32Slow() && | |||
24842 | StVT == VT && !IsAligned) { | |||
24843 | unsigned NumElems = VT.getVectorNumElements(); | |||
24844 | if (NumElems < 2) | |||
24845 | return SDValue(); | |||
24846 | ||||
24847 | SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl); | |||
24848 | SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl); | |||
24849 | ||||
24850 | SDValue Stride = DAG.getConstant(16, TLI.getPointerTy()); | |||
24851 | SDValue Ptr0 = St->getBasePtr(); | |||
24852 | SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride); | |||
24853 | ||||
24854 | SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0, | |||
24855 | St->getPointerInfo(), St->isVolatile(), | |||
24856 | St->isNonTemporal(), Alignment); | |||
24857 | SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1, | |||
24858 | St->getPointerInfo(), St->isVolatile(), | |||
24859 | St->isNonTemporal(), | |||
24860 | std::min(16U, Alignment)); | |||
24861 | return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1); | |||
24862 | } | |||
24863 | ||||
24864 | // Optimize trunc store (of multiple scalars) to shuffle and store. | |||
24865 | // First, pack all of the elements in one place. Next, store to memory | |||
24866 | // in fewer chunks. | |||
24867 | if (St->isTruncatingStore() && VT.isVector()) { | |||
24868 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
24869 | unsigned NumElems = VT.getVectorNumElements(); | |||
24870 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24870, __PRETTY_FUNCTION__)); | |||
24871 | unsigned FromSz = VT.getVectorElementType().getSizeInBits(); | |||
24872 | unsigned ToSz = StVT.getVectorElementType().getSizeInBits(); | |||
24873 | ||||
24874 | // From, To sizes and ElemCount must be pow of two | |||
24875 | if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue(); | |||
24876 | // We are going to use the original vector elt for storing. | |||
24877 | // Accumulated smaller vector elements must be a multiple of the store size. | |||
24878 | if (0 != (NumElems * FromSz) % ToSz) return SDValue(); | |||
24879 | ||||
24880 | unsigned SizeRatio = FromSz / ToSz; | |||
24881 | ||||
24882 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24882, __PRETTY_FUNCTION__)); | |||
24883 | ||||
24884 | // Create a type on which we perform the shuffle | |||
24885 | EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), | |||
24886 | StVT.getScalarType(), NumElems*SizeRatio); | |||
24887 | ||||
24888 | assert(WideVecVT.getSizeInBits() == VT.getSizeInBits())((WideVecVT.getSizeInBits() == VT.getSizeInBits()) ? static_cast <void> (0) : __assert_fail ("WideVecVT.getSizeInBits() == VT.getSizeInBits()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24888, __PRETTY_FUNCTION__)); | |||
24889 | ||||
24890 | SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue()); | |||
24891 | SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1); | |||
24892 | for (unsigned i = 0; i != NumElems; ++i) | |||
24893 | ShuffleVec[i] = i * SizeRatio; | |||
24894 | ||||
24895 | // Can't shuffle using an illegal type. | |||
24896 | if (!TLI.isTypeLegal(WideVecVT)) | |||
24897 | return SDValue(); | |||
24898 | ||||
24899 | SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec, | |||
24900 | DAG.getUNDEF(WideVecVT), | |||
24901 | &ShuffleVec[0]); | |||
24902 | // At this point all of the data is stored at the bottom of the | |||
24903 | // register. We now need to save it to mem. | |||
24904 | ||||
24905 | // Find the largest store unit | |||
24906 | MVT StoreType = MVT::i8; | |||
24907 | for (MVT Tp : MVT::integer_valuetypes()) { | |||
24908 | if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz) | |||
24909 | StoreType = Tp; | |||
24910 | } | |||
24911 | ||||
24912 | // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64. | |||
24913 | if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 && | |||
24914 | (64 <= NumElems * ToSz)) | |||
24915 | StoreType = MVT::f64; | |||
24916 | ||||
24917 | // Bitcast the original vector into a vector of store-size units | |||
24918 | EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(), | |||
24919 | StoreType, VT.getSizeInBits()/StoreType.getSizeInBits()); | |||
24920 | assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits())((StoreVecVT.getSizeInBits() == VT.getSizeInBits()) ? static_cast <void> (0) : __assert_fail ("StoreVecVT.getSizeInBits() == VT.getSizeInBits()" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 24920, __PRETTY_FUNCTION__)); | |||
24921 | SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff); | |||
24922 | SmallVector<SDValue, 8> Chains; | |||
24923 | SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8, | |||
24924 | TLI.getPointerTy()); | |||
24925 | SDValue Ptr = St->getBasePtr(); | |||
24926 | ||||
24927 | // Perform one or more big stores into memory. | |||
24928 | for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) { | |||
24929 | SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, | |||
24930 | StoreType, ShuffWide, | |||
24931 | DAG.getIntPtrConstant(i)); | |||
24932 | SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr, | |||
24933 | St->getPointerInfo(), St->isVolatile(), | |||
24934 | St->isNonTemporal(), St->getAlignment()); | |||
24935 | Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment); | |||
24936 | Chains.push_back(Ch); | |||
24937 | } | |||
24938 | ||||
24939 | return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains); | |||
24940 | } | |||
24941 | ||||
24942 | // Turn load->store of MMX types into GPR load/stores. This avoids clobbering | |||
24943 | // the FP state in cases where an emms may be missing. | |||
24944 | // A preferable solution to the general problem is to figure out the right | |||
24945 | // places to insert EMMS. This qualifies as a quick hack. | |||
24946 | ||||
24947 | // Similarly, turn load->store of i64 into double load/stores in 32-bit mode. | |||
24948 | if (VT.getSizeInBits() != 64) | |||
24949 | return SDValue(); | |||
24950 | ||||
24951 | const Function *F = DAG.getMachineFunction().getFunction(); | |||
24952 | bool NoImplicitFloatOps = F->getAttributes(). | |||
24953 | hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat); | |||
24954 | bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps | |||
24955 | && Subtarget->hasSSE2(); | |||
24956 | if ((VT.isVector() || | |||
24957 | (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) && | |||
24958 | isa<LoadSDNode>(St->getValue()) && | |||
24959 | !cast<LoadSDNode>(St->getValue())->isVolatile() && | |||
24960 | St->getChain().hasOneUse() && !St->isVolatile()) { | |||
24961 | SDNode* LdVal = St->getValue().getNode(); | |||
24962 | LoadSDNode *Ld = nullptr; | |||
24963 | int TokenFactorIndex = -1; | |||
24964 | SmallVector<SDValue, 8> Ops; | |||
24965 | SDNode* ChainVal = St->getChain().getNode(); | |||
24966 | // Must be a store of a load. We currently handle two cases: the load | |||
24967 | // is a direct child, and it's under an intervening TokenFactor. It is | |||
24968 | // possible to dig deeper under nested TokenFactors. | |||
24969 | if (ChainVal == LdVal) | |||
24970 | Ld = cast<LoadSDNode>(St->getChain()); | |||
24971 | else if (St->getValue().hasOneUse() && | |||
24972 | ChainVal->getOpcode() == ISD::TokenFactor) { | |||
24973 | for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) { | |||
24974 | if (ChainVal->getOperand(i).getNode() == LdVal) { | |||
24975 | TokenFactorIndex = i; | |||
24976 | Ld = cast<LoadSDNode>(St->getValue()); | |||
24977 | } else | |||
24978 | Ops.push_back(ChainVal->getOperand(i)); | |||
24979 | } | |||
24980 | } | |||
24981 | ||||
24982 | if (!Ld || !ISD::isNormalLoad(Ld)) | |||
24983 | return SDValue(); | |||
24984 | ||||
24985 | // If this is not the MMX case, i.e. we are just turning i64 load/store | |||
24986 | // into f64 load/store, avoid the transformation if there are multiple | |||
24987 | // uses of the loaded value. | |||
24988 | if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0)) | |||
24989 | return SDValue(); | |||
24990 | ||||
24991 | SDLoc LdDL(Ld); | |||
24992 | SDLoc StDL(N); | |||
24993 | // If we are a 64-bit capable x86, lower to a single movq load/store pair. | |||
24994 | // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store | |||
24995 | // pair instead. | |||
24996 | if (Subtarget->is64Bit() || F64IsLegal) { | |||
24997 | EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64; | |||
24998 | SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(), | |||
24999 | Ld->getPointerInfo(), Ld->isVolatile(), | |||
25000 | Ld->isNonTemporal(), Ld->isInvariant(), | |||
25001 | Ld->getAlignment()); | |||
25002 | SDValue NewChain = NewLd.getValue(1); | |||
25003 | if (TokenFactorIndex != -1) { | |||
25004 | Ops.push_back(NewChain); | |||
25005 | NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops); | |||
25006 | } | |||
25007 | return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(), | |||
25008 | St->getPointerInfo(), | |||
25009 | St->isVolatile(), St->isNonTemporal(), | |||
25010 | St->getAlignment()); | |||
25011 | } | |||
25012 | ||||
25013 | // Otherwise, lower to two pairs of 32-bit loads / stores. | |||
25014 | SDValue LoAddr = Ld->getBasePtr(); | |||
25015 | SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr, | |||
25016 | DAG.getConstant(4, MVT::i32)); | |||
25017 | ||||
25018 | SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr, | |||
25019 | Ld->getPointerInfo(), | |||
25020 | Ld->isVolatile(), Ld->isNonTemporal(), | |||
25021 | Ld->isInvariant(), Ld->getAlignment()); | |||
25022 | SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr, | |||
25023 | Ld->getPointerInfo().getWithOffset(4), | |||
25024 | Ld->isVolatile(), Ld->isNonTemporal(), | |||
25025 | Ld->isInvariant(), | |||
25026 | MinAlign(Ld->getAlignment(), 4)); | |||
25027 | ||||
25028 | SDValue NewChain = LoLd.getValue(1); | |||
25029 | if (TokenFactorIndex != -1) { | |||
25030 | Ops.push_back(LoLd); | |||
25031 | Ops.push_back(HiLd); | |||
25032 | NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops); | |||
25033 | } | |||
25034 | ||||
25035 | LoAddr = St->getBasePtr(); | |||
25036 | HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr, | |||
25037 | DAG.getConstant(4, MVT::i32)); | |||
25038 | ||||
25039 | SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr, | |||
25040 | St->getPointerInfo(), | |||
25041 | St->isVolatile(), St->isNonTemporal(), | |||
25042 | St->getAlignment()); | |||
25043 | SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr, | |||
25044 | St->getPointerInfo().getWithOffset(4), | |||
25045 | St->isVolatile(), | |||
25046 | St->isNonTemporal(), | |||
25047 | MinAlign(St->getAlignment(), 4)); | |||
25048 | return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt); | |||
25049 | } | |||
25050 | return SDValue(); | |||
25051 | } | |||
25052 | ||||
25053 | /// Return 'true' if this vector operation is "horizontal" | |||
25054 | /// and return the operands for the horizontal operation in LHS and RHS. A | |||
25055 | /// horizontal operation performs the binary operation on successive elements | |||
25056 | /// of its first operand, then on successive elements of its second operand, | |||
25057 | /// returning the resulting values in a vector. For example, if | |||
25058 | /// A = < float a0, float a1, float a2, float a3 > | |||
25059 | /// and | |||
25060 | /// B = < float b0, float b1, float b2, float b3 > | |||
25061 | /// then the result of doing a horizontal operation on A and B is | |||
25062 | /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >. | |||
25063 | /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form | |||
25064 | /// A horizontal-op B, for some already available A and B, and if so then LHS is | |||
25065 | /// set to A, RHS to B, and the routine returns 'true'. | |||
25066 | /// Note that the binary operation should have the property that if one of the | |||
25067 | /// operands is UNDEF then the result is UNDEF. | |||
25068 | static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) { | |||
25069 | // Look for the following pattern: if | |||
25070 | // A = < float a0, float a1, float a2, float a3 > | |||
25071 | // B = < float b0, float b1, float b2, float b3 > | |||
25072 | // and | |||
25073 | // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6> | |||
25074 | // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7> | |||
25075 | // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 > | |||
25076 | // which is A horizontal-op B. | |||
25077 | ||||
25078 | // At least one of the operands should be a vector shuffle. | |||
25079 | if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE && | |||
25080 | RHS.getOpcode() != ISD::VECTOR_SHUFFLE) | |||
25081 | return false; | |||
25082 | ||||
25083 | MVT VT = LHS.getSimpleValueType(); | |||
25084 | ||||
25085 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 25086, __PRETTY_FUNCTION__)) | |||
25086 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 25086, __PRETTY_FUNCTION__)); | |||
25087 | ||||
25088 | // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to | |||
25089 | // operate independently on 128-bit lanes. | |||
25090 | unsigned NumElts = VT.getVectorNumElements(); | |||
25091 | unsigned NumLanes = VT.getSizeInBits()/128; | |||
25092 | unsigned NumLaneElts = NumElts / NumLanes; | |||
25093 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 25094, __PRETTY_FUNCTION__)) | |||
25094 | "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 25094, __PRETTY_FUNCTION__)); | |||
25095 | unsigned HalfLaneElts = NumLaneElts/2; | |||
25096 | ||||
25097 | // View LHS in the form | |||
25098 | // LHS = VECTOR_SHUFFLE A, B, LMask | |||
25099 | // If LHS is not a shuffle then pretend it is the shuffle | |||
25100 | // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1> | |||
25101 | // NOTE: in what follows a default initialized SDValue represents an UNDEF of | |||
25102 | // type VT. | |||
25103 | SDValue A, B; | |||
25104 | SmallVector<int, 16> LMask(NumElts); | |||
25105 | if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) { | |||
25106 | if (LHS.getOperand(0).getOpcode() != ISD::UNDEF) | |||
25107 | A = LHS.getOperand(0); | |||
25108 | if (LHS.getOperand(1).getOpcode() != ISD::UNDEF) | |||
25109 | B = LHS.getOperand(1); | |||
25110 | ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask(); | |||
25111 | std::copy(Mask.begin(), Mask.end(), LMask.begin()); | |||
25112 | } else { | |||
25113 | if (LHS.getOpcode() != ISD::UNDEF) | |||
25114 | A = LHS; | |||
25115 | for (unsigned i = 0; i != NumElts; ++i) | |||
25116 | LMask[i] = i; | |||
25117 | } | |||
25118 | ||||
25119 | // Likewise, view RHS in the form | |||
25120 | // RHS = VECTOR_SHUFFLE C, D, RMask | |||
25121 | SDValue C, D; | |||
25122 | SmallVector<int, 16> RMask(NumElts); | |||
25123 | if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) { | |||
25124 | if (RHS.getOperand(0).getOpcode() != ISD::UNDEF) | |||
25125 | C = RHS.getOperand(0); | |||
25126 | if (RHS.getOperand(1).getOpcode() != ISD::UNDEF) | |||
25127 | D = RHS.getOperand(1); | |||
25128 | ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask(); | |||
25129 | std::copy(Mask.begin(), Mask.end(), RMask.begin()); | |||
25130 | } else { | |||
25131 | if (RHS.getOpcode() != ISD::UNDEF) | |||
25132 | C = RHS; | |||
25133 | for (unsigned i = 0; i != NumElts; ++i) | |||
25134 | RMask[i] = i; | |||
25135 | } | |||
25136 | ||||
25137 | // Check that the shuffles are both shuffling the same vectors. | |||
25138 | if (!(A == C && B == D) && !(A == D && B == C)) | |||
25139 | return false; | |||
25140 | ||||
25141 | // If everything is UNDEF then bail out: it would be better to fold to UNDEF. | |||
25142 | if (!A.getNode() && !B.getNode()) | |||
25143 | return false; | |||
25144 | ||||
25145 | // If A and B occur in reverse order in RHS, then "swap" them (which means | |||
25146 | // rewriting the mask). | |||
25147 | if (A != C) | |||
25148 | CommuteVectorShuffleMask(RMask, NumElts); | |||
25149 | ||||
25150 | // At this point LHS and RHS are equivalent to | |||
25151 | // LHS = VECTOR_SHUFFLE A, B, LMask | |||
25152 | // RHS = VECTOR_SHUFFLE A, B, RMask | |||
25153 | // Check that the masks correspond to performing a horizontal operation. | |||
25154 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { | |||
25155 | for (unsigned i = 0; i != NumLaneElts; ++i) { | |||
25156 | int LIdx = LMask[i+l], RIdx = RMask[i+l]; | |||
25157 | ||||
25158 | // Ignore any UNDEF components. | |||
25159 | if (LIdx < 0 || RIdx < 0 || | |||
25160 | (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) || | |||
25161 | (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts))) | |||
25162 | continue; | |||
25163 | ||||
25164 | // Check that successive elements are being operated on. If not, this is | |||
25165 | // not a horizontal operation. | |||
25166 | unsigned Src = (i/HalfLaneElts); // each lane is split between srcs | |||
25167 | int Index = 2*(i%HalfLaneElts) + NumElts*Src + l; | |||
25168 | if (!(LIdx == Index && RIdx == Index + 1) && | |||
25169 | !(IsCommutative && LIdx == Index + 1 && RIdx == Index)) | |||
25170 | return false; | |||
25171 | } | |||
25172 | } | |||
25173 | ||||
25174 | LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it. | |||
25175 | RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it. | |||
25176 | return true; | |||
25177 | } | |||
25178 | ||||
25179 | /// Do target-specific dag combines on floating point adds. | |||
25180 | static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG, | |||
25181 | const X86Subtarget *Subtarget) { | |||
25182 | EVT VT = N->getValueType(0); | |||
25183 | SDValue LHS = N->getOperand(0); | |||
25184 | SDValue RHS = N->getOperand(1); | |||
25185 | ||||
25186 | // Try to synthesize horizontal adds from adds of shuffles. | |||
25187 | if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) || | |||
25188 | (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) && | |||
25189 | isHorizontalBinOp(LHS, RHS, true)) | |||
25190 | return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS); | |||
25191 | return SDValue(); | |||
25192 | } | |||
25193 | ||||
25194 | /// Do target-specific dag combines on floating point subs. | |||
25195 | static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG, | |||
25196 | const X86Subtarget *Subtarget) { | |||
25197 | EVT VT = N->getValueType(0); | |||
25198 | SDValue LHS = N->getOperand(0); | |||
25199 | SDValue RHS = N->getOperand(1); | |||
25200 | ||||
25201 | // Try to synthesize horizontal subs from subs of shuffles. | |||
25202 | if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) || | |||
25203 | (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) && | |||
25204 | isHorizontalBinOp(LHS, RHS, false)) | |||
25205 | return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS); | |||
25206 | return SDValue(); | |||
25207 | } | |||
25208 | ||||
25209 | /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes. | |||
25210 | static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) { | |||
25211 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 25211, __PRETTY_FUNCTION__)); | |||
25212 | // F[X]OR(0.0, x) -> x | |||
25213 | // F[X]OR(x, 0.0) -> x | |||
25214 | if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0))) | |||
25215 | if (C->getValueAPF().isPosZero()) | |||
25216 | return N->getOperand(1); | |||
25217 | if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1))) | |||
25218 | if (C->getValueAPF().isPosZero()) | |||
25219 | return N->getOperand(0); | |||
25220 | return SDValue(); | |||
25221 | } | |||
25222 | ||||
25223 | /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes. | |||
25224 | static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) { | |||
25225 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 25225, __PRETTY_FUNCTION__)); | |||
25226 | ||||
25227 | // Only perform optimizations if UnsafeMath is used. | |||
25228 | if (!DAG.getTarget().Options.UnsafeFPMath) | |||
25229 | return SDValue(); | |||
25230 | ||||
25231 | // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes | |||
25232 | // into FMINC and FMAXC, which are Commutative operations. | |||
25233 | unsigned NewOp = 0; | |||
25234 | switch (N->getOpcode()) { | |||
25235 | default: llvm_unreachable("unknown opcode")::llvm::llvm_unreachable_internal("unknown opcode", "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 25235); | |||
25236 | case X86ISD::FMIN: NewOp = X86ISD::FMINC; break; | |||
25237 | case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break; | |||
25238 | } | |||
25239 | ||||
25240 | return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0), | |||
25241 | N->getOperand(0), N->getOperand(1)); | |||
25242 | } | |||
25243 | ||||
25244 | /// Do target-specific dag combines on X86ISD::FAND nodes. | |||
25245 | static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) { | |||
25246 | // FAND(0.0, x) -> 0.0 | |||
25247 | // FAND(x, 0.0) -> 0.0 | |||
25248 | if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0))) | |||
25249 | if (C->getValueAPF().isPosZero()) | |||
25250 | return N->getOperand(0); | |||
25251 | if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1))) | |||
25252 | if (C->getValueAPF().isPosZero()) | |||
25253 | return N->getOperand(1); | |||
25254 | return SDValue(); | |||
25255 | } | |||
25256 | ||||
25257 | /// Do target-specific dag combines on X86ISD::FANDN nodes | |||
25258 | static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) { | |||
25259 | // FANDN(x, 0.0) -> 0.0 | |||
25260 | // FANDN(0.0, x) -> x | |||
25261 | if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0))) | |||
25262 | if (C->getValueAPF().isPosZero()) | |||
25263 | return N->getOperand(1); | |||
25264 | if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1))) | |||
25265 | if (C->getValueAPF().isPosZero()) | |||
25266 | return N->getOperand(1); | |||
25267 | return SDValue(); | |||
25268 | } | |||
25269 | ||||
25270 | static SDValue PerformBTCombine(SDNode *N, | |||
25271 | SelectionDAG &DAG, | |||
25272 | TargetLowering::DAGCombinerInfo &DCI) { | |||
25273 | // BT ignores high bits in the bit index operand. | |||
25274 | SDValue Op1 = N->getOperand(1); | |||
25275 | if (Op1.hasOneUse()) { | |||
25276 | unsigned BitWidth = Op1.getValueSizeInBits(); | |||
25277 | APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth)); | |||
25278 | APInt KnownZero, KnownOne; | |||
25279 | TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(), | |||
25280 | !DCI.isBeforeLegalizeOps()); | |||
25281 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
25282 | if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) || | |||
25283 | TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO)) | |||
25284 | DCI.CommitTargetLoweringOpt(TLO); | |||
25285 | } | |||
25286 | return SDValue(); | |||
25287 | } | |||
25288 | ||||
25289 | static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) { | |||
25290 | SDValue Op = N->getOperand(0); | |||
25291 | if (Op.getOpcode() == ISD::BITCAST) | |||
25292 | Op = Op.getOperand(0); | |||
25293 | EVT VT = N->getValueType(0), OpVT = Op.getValueType(); | |||
25294 | if (Op.getOpcode() == X86ISD::VZEXT_LOAD && | |||
25295 | VT.getVectorElementType().getSizeInBits() == | |||
25296 | OpVT.getVectorElementType().getSizeInBits()) { | |||
25297 | return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op); | |||
25298 | } | |||
25299 | return SDValue(); | |||
25300 | } | |||
25301 | ||||
25302 | static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG, | |||
25303 | const X86Subtarget *Subtarget) { | |||
25304 | EVT VT = N->getValueType(0); | |||
25305 | if (!VT.isVector()) | |||
25306 | return SDValue(); | |||
25307 | ||||
25308 | SDValue N0 = N->getOperand(0); | |||
25309 | SDValue N1 = N->getOperand(1); | |||
25310 | EVT ExtraVT = cast<VTSDNode>(N1)->getVT(); | |||
25311 | SDLoc dl(N); | |||
25312 | ||||
25313 | // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the | |||
25314 | // both SSE and AVX2 since there is no sign-extended shift right | |||
25315 | // operation on a vector with 64-bit elements. | |||
25316 | //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) -> | |||
25317 | // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT))) | |||
25318 | if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND || | |||
25319 | N0.getOpcode() == ISD::SIGN_EXTEND)) { | |||
25320 | SDValue N00 = N0.getOperand(0); | |||
25321 | ||||
25322 | // EXTLOAD has a better solution on AVX2, | |||
25323 | // it may be replaced with X86ISD::VSEXT node. | |||
25324 | if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256()) | |||
25325 | if (!ISD::isNormalLoad(N00.getNode())) | |||
25326 | return SDValue(); | |||
25327 | ||||
25328 | if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) { | |||
25329 | SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32, | |||
25330 | N00, N1); | |||
25331 | return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp); | |||
25332 | } | |||
25333 | } | |||
25334 | return SDValue(); | |||
25335 | } | |||
25336 | ||||
25337 | static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG, | |||
25338 | TargetLowering::DAGCombinerInfo &DCI, | |||
25339 | const X86Subtarget *Subtarget) { | |||
25340 | SDValue N0 = N->getOperand(0); | |||
25341 | EVT VT = N->getValueType(0); | |||
25342 | ||||
25343 | // (i8,i32 sext (sdivrem (i8 x, i8 y)) -> | |||
25344 | // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y) | |||
25345 | // This exposes the sext to the sdivrem lowering, so that it directly extends | |||
25346 | // from AH (which we otherwise need to do contortions to access). | |||
25347 | if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 && | |||
25348 | N0.getValueType() == MVT::i8 && VT == MVT::i32) { | |||
25349 | SDLoc dl(N); | |||
25350 | SDVTList NodeTys = DAG.getVTList(MVT::i8, VT); | |||
25351 | SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, dl, NodeTys, | |||
25352 | N0.getOperand(0), N0.getOperand(1)); | |||
25353 | DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0)); | |||
25354 | return R.getValue(1); | |||
25355 | } | |||
25356 | ||||
25357 | if (!DCI.isBeforeLegalizeOps()) | |||
25358 | return SDValue(); | |||
25359 | ||||
25360 | if (!Subtarget->hasFp256()) | |||
25361 | return SDValue(); | |||
25362 | ||||
25363 | if (VT.isVector() && VT.getSizeInBits() == 256) { | |||
25364 | SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget); | |||
25365 | if (R.getNode()) | |||
25366 | return R; | |||
25367 | } | |||
25368 | ||||
25369 | return SDValue(); | |||
25370 | } | |||
25371 | ||||
25372 | static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG, | |||
25373 | const X86Subtarget* Subtarget) { | |||
25374 | SDLoc dl(N); | |||
25375 | EVT VT = N->getValueType(0); | |||
25376 | ||||
25377 | // Let legalize expand this if it isn't a legal type yet. | |||
25378 | if (!DAG.getTargetLoweringInfo().isTypeLegal(VT)) | |||
25379 | return SDValue(); | |||
25380 | ||||
25381 | EVT ScalarVT = VT.getScalarType(); | |||
25382 | if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) || | |||
25383 | (!Subtarget->hasFMA() && !Subtarget->hasFMA4())) | |||
25384 | return SDValue(); | |||
25385 | ||||
25386 | SDValue A = N->getOperand(0); | |||
25387 | SDValue B = N->getOperand(1); | |||
25388 | SDValue C = N->getOperand(2); | |||
25389 | ||||
25390 | bool NegA = (A.getOpcode() == ISD::FNEG); | |||
25391 | bool NegB = (B.getOpcode() == ISD::FNEG); | |||
25392 | bool NegC = (C.getOpcode() == ISD::FNEG); | |||
25393 | ||||
25394 | // Negative multiplication when NegA xor NegB | |||
25395 | bool NegMul = (NegA != NegB); | |||
25396 | if (NegA) | |||
25397 | A = A.getOperand(0); | |||
25398 | if (NegB) | |||
25399 | B = B.getOperand(0); | |||
25400 | if (NegC) | |||
25401 | C = C.getOperand(0); | |||
25402 | ||||
25403 | unsigned Opcode; | |||
25404 | if (!NegMul) | |||
25405 | Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB; | |||
25406 | else | |||
25407 | Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB; | |||
25408 | ||||
25409 | return DAG.getNode(Opcode, dl, VT, A, B, C); | |||
25410 | } | |||
25411 | ||||
25412 | static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG, | |||
25413 | TargetLowering::DAGCombinerInfo &DCI, | |||
25414 | const X86Subtarget *Subtarget) { | |||
25415 | // (i32 zext (and (i8 x86isd::setcc_carry), 1)) -> | |||
25416 | // (and (i32 x86isd::setcc_carry), 1) | |||
25417 | // This eliminates the zext. This transformation is necessary because | |||
25418 | // ISD::SETCC is always legalized to i8. | |||
25419 | SDLoc dl(N); | |||
25420 | SDValue N0 = N->getOperand(0); | |||
25421 | EVT VT = N->getValueType(0); | |||
25422 | ||||
25423 | if (N0.getOpcode() == ISD::AND && | |||
25424 | N0.hasOneUse() && | |||
25425 | N0.getOperand(0).hasOneUse()) { | |||
25426 | SDValue N00 = N0.getOperand(0); | |||
25427 | if (N00.getOpcode() == X86ISD::SETCC_CARRY) { | |||
25428 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); | |||
25429 | if (!C || C->getZExtValue() != 1) | |||
25430 | return SDValue(); | |||
25431 | return DAG.getNode(ISD::AND, dl, VT, | |||
25432 | DAG.getNode(X86ISD::SETCC_CARRY, dl, VT, | |||
25433 | N00.getOperand(0), N00.getOperand(1)), | |||
25434 | DAG.getConstant(1, VT)); | |||
25435 | } | |||
25436 | } | |||
25437 | ||||
25438 | if (N0.getOpcode() == ISD::TRUNCATE && | |||
25439 | N0.hasOneUse() && | |||
25440 | N0.getOperand(0).hasOneUse()) { | |||
25441 | SDValue N00 = N0.getOperand(0); | |||
25442 | if (N00.getOpcode() == X86ISD::SETCC_CARRY) { | |||
25443 | return DAG.getNode(ISD::AND, dl, VT, | |||
25444 | DAG.getNode(X86ISD::SETCC_CARRY, dl, VT, | |||
25445 | N00.getOperand(0), N00.getOperand(1)), | |||
25446 | DAG.getConstant(1, VT)); | |||
25447 | } | |||
25448 | } | |||
25449 | if (VT.is256BitVector()) { | |||
25450 | SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget); | |||
25451 | if (R.getNode()) | |||
25452 | return R; | |||
25453 | } | |||
25454 | ||||
25455 | // (i8,i32 zext (udivrem (i8 x, i8 y)) -> | |||
25456 | // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y) | |||
25457 | // This exposes the zext to the udivrem lowering, so that it directly extends | |||
25458 | // from AH (which we otherwise need to do contortions to access). | |||
25459 | if (N0.getOpcode() == ISD::UDIVREM && | |||
25460 | N0.getResNo() == 1 && N0.getValueType() == MVT::i8 && | |||
25461 | (VT == MVT::i32 || VT == MVT::i64)) { | |||
25462 | SDVTList NodeTys = DAG.getVTList(MVT::i8, VT); | |||
25463 | SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys, | |||
25464 | N0.getOperand(0), N0.getOperand(1)); | |||
25465 | DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0)); | |||
25466 | return R.getValue(1); | |||
25467 | } | |||
25468 | ||||
25469 | return SDValue(); | |||
25470 | } | |||
25471 | ||||
25472 | // Optimize x == -y --> x+y == 0 | |||
25473 | // x != -y --> x+y != 0 | |||
25474 | static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG, | |||
25475 | const X86Subtarget* Subtarget) { | |||
25476 | ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get(); | |||
25477 | SDValue LHS = N->getOperand(0); | |||
25478 | SDValue RHS = N->getOperand(1); | |||
25479 | EVT VT = N->getValueType(0); | |||
25480 | SDLoc DL(N); | |||
25481 | ||||
25482 | if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB) | |||
25483 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0))) | |||
25484 | if (C->getAPIntValue() == 0 && LHS.hasOneUse()) { | |||
25485 | SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N), | |||
25486 | LHS.getValueType(), RHS, LHS.getOperand(1)); | |||
25487 | return DAG.getSetCC(SDLoc(N), N->getValueType(0), | |||
25488 | addV, DAG.getConstant(0, addV.getValueType()), CC); | |||
25489 | } | |||
25490 | if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB) | |||
25491 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0))) | |||
25492 | if (C->getAPIntValue() == 0 && RHS.hasOneUse()) { | |||
25493 | SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N), | |||
25494 | RHS.getValueType(), LHS, RHS.getOperand(1)); | |||
25495 | return DAG.getSetCC(SDLoc(N), N->getValueType(0), | |||
25496 | addV, DAG.getConstant(0, addV.getValueType()), CC); | |||
25497 | } | |||
25498 | ||||
25499 | if (VT.getScalarType() == MVT::i1) { | |||
25500 | bool IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) && | |||
25501 | (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1); | |||
25502 | bool IsVZero0 = ISD::isBuildVectorAllZeros(LHS.getNode()); | |||
25503 | if (!IsSEXT0 && !IsVZero0) | |||
25504 | return SDValue(); | |||
25505 | bool IsSEXT1 = (RHS.getOpcode() == ISD::SIGN_EXTEND) && | |||
25506 | (RHS.getOperand(0).getValueType().getScalarType() == MVT::i1); | |||
25507 | bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode()); | |||
25508 | ||||
25509 | if (!IsSEXT1 && !IsVZero1) | |||
25510 | return SDValue(); | |||
25511 | ||||
25512 | if (IsSEXT0 && IsVZero1) { | |||
25513 | assert(VT == LHS.getOperand(0).getValueType() && "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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 25513, __PRETTY_FUNCTION__)); | |||
25514 | if (CC == ISD::SETEQ) | |||
25515 | return DAG.getNOT(DL, LHS.getOperand(0), VT); | |||
25516 | return LHS.getOperand(0); | |||
25517 | } | |||
25518 | if (IsSEXT1 && IsVZero0) { | |||
25519 | assert(VT == RHS.getOperand(0).getValueType() && "Uexpected operand type")((VT == RHS.getOperand(0).getValueType() && "Uexpected operand type" ) ? static_cast<void> (0) : __assert_fail ("VT == RHS.getOperand(0).getValueType() && \"Uexpected operand type\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 25519, __PRETTY_FUNCTION__)); | |||
25520 | if (CC == ISD::SETEQ) | |||
25521 | return DAG.getNOT(DL, RHS.getOperand(0), VT); | |||
25522 | return RHS.getOperand(0); | |||
25523 | } | |||
25524 | } | |||
25525 | ||||
25526 | return SDValue(); | |||
25527 | } | |||
25528 | ||||
25529 | static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG, | |||
25530 | const X86Subtarget *Subtarget) { | |||
25531 | SDLoc dl(N); | |||
25532 | MVT VT = N->getOperand(1)->getSimpleValueType(0); | |||
25533 | assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&(((VT == MVT::v4f32 || VT == MVT::v4i32) && "X86insertps is only defined for v4x32" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v4f32 || VT == MVT::v4i32) && \"X86insertps is only defined for v4x32\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 25534, __PRETTY_FUNCTION__)) | |||
25534 | "X86insertps is only defined for v4x32")(((VT == MVT::v4f32 || VT == MVT::v4i32) && "X86insertps is only defined for v4x32" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v4f32 || VT == MVT::v4i32) && \"X86insertps is only defined for v4x32\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 25534, __PRETTY_FUNCTION__)); | |||
25535 | ||||
25536 | SDValue Ld = N->getOperand(1); | |||
25537 | if (MayFoldLoad(Ld)) { | |||
25538 | // Extract the countS bits from the immediate so we can get the proper | |||
25539 | // address when narrowing the vector load to a specific element. | |||
25540 | // When the second source op is a memory address, interps doesn't use | |||
25541 | // countS and just gets an f32 from that address. | |||
25542 | unsigned DestIndex = | |||
25543 | cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6; | |||
25544 | Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG); | |||
25545 | } else | |||
25546 | return SDValue(); | |||
25547 | ||||
25548 | // Create this as a scalar to vector to match the instruction pattern. | |||
25549 | SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld); | |||
25550 | // countS bits are ignored when loading from memory on insertps, which | |||
25551 | // means we don't need to explicitly set them to 0. | |||
25552 | return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0), | |||
25553 | LoadScalarToVector, N->getOperand(2)); | |||
25554 | } | |||
25555 | ||||
25556 | // Helper function of PerformSETCCCombine. It is to materialize "setb reg" | |||
25557 | // as "sbb reg,reg", since it can be extended without zext and produces | |||
25558 | // an all-ones bit which is more useful than 0/1 in some cases. | |||
25559 | static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG, | |||
25560 | MVT VT) { | |||
25561 | if (VT == MVT::i8) | |||
25562 | return DAG.getNode(ISD::AND, DL, VT, | |||
25563 | DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8, | |||
25564 | DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS), | |||
25565 | DAG.getConstant(1, VT)); | |||
25566 | assert (VT == MVT::i1 && "Unexpected type for SECCC node")((VT == MVT::i1 && "Unexpected type for SECCC node") ? static_cast<void> (0) : __assert_fail ("VT == MVT::i1 && \"Unexpected type for SECCC node\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 25566, __PRETTY_FUNCTION__)); | |||
25567 | return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, | |||
25568 | DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8, | |||
25569 | DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS)); | |||
25570 | } | |||
25571 | ||||
25572 | // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT | |||
25573 | static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG, | |||
25574 | TargetLowering::DAGCombinerInfo &DCI, | |||
25575 | const X86Subtarget *Subtarget) { | |||
25576 | SDLoc DL(N); | |||
25577 | X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0)); | |||
25578 | SDValue EFLAGS = N->getOperand(1); | |||
25579 | ||||
25580 | if (CC == X86::COND_A) { | |||
25581 | // Try to convert COND_A into COND_B in an attempt to facilitate | |||
25582 | // materializing "setb reg". | |||
25583 | // | |||
25584 | // Do not flip "e > c", where "c" is a constant, because Cmp instruction | |||
25585 | // cannot take an immediate as its first operand. | |||
25586 | // | |||
25587 | if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() && | |||
25588 | EFLAGS.getValueType().isInteger() && | |||
25589 | !isa<ConstantSDNode>(EFLAGS.getOperand(1))) { | |||
25590 | SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS), | |||
25591 | EFLAGS.getNode()->getVTList(), | |||
25592 | EFLAGS.getOperand(1), EFLAGS.getOperand(0)); | |||
25593 | SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo()); | |||
25594 | return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0)); | |||
25595 | } | |||
25596 | } | |||
25597 | ||||
25598 | // Materialize "setb reg" as "sbb reg,reg", since it can be extended without | |||
25599 | // a zext and produces an all-ones bit which is more useful than 0/1 in some | |||
25600 | // cases. | |||
25601 | if (CC == X86::COND_B) | |||
25602 | return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0)); | |||
25603 | ||||
25604 | SDValue Flags; | |||
25605 | ||||
25606 | Flags = checkBoolTestSetCCCombine(EFLAGS, CC); | |||
25607 | if (Flags.getNode()) { | |||
25608 | SDValue Cond = DAG.getConstant(CC, MVT::i8); | |||
25609 | return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags); | |||
25610 | } | |||
25611 | ||||
25612 | return SDValue(); | |||
25613 | } | |||
25614 | ||||
25615 | // Optimize branch condition evaluation. | |||
25616 | // | |||
25617 | static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG, | |||
25618 | TargetLowering::DAGCombinerInfo &DCI, | |||
25619 | const X86Subtarget *Subtarget) { | |||
25620 | SDLoc DL(N); | |||
25621 | SDValue Chain = N->getOperand(0); | |||
25622 | SDValue Dest = N->getOperand(1); | |||
25623 | SDValue EFLAGS = N->getOperand(3); | |||
25624 | X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2)); | |||
25625 | ||||
25626 | SDValue Flags; | |||
25627 | ||||
25628 | Flags = checkBoolTestSetCCCombine(EFLAGS, CC); | |||
25629 | if (Flags.getNode()) { | |||
25630 | SDValue Cond = DAG.getConstant(CC, MVT::i8); | |||
25631 | return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond, | |||
25632 | Flags); | |||
25633 | } | |||
25634 | ||||
25635 | return SDValue(); | |||
25636 | } | |||
25637 | ||||
25638 | static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N, | |||
25639 | SelectionDAG &DAG) { | |||
25640 | // Take advantage of vector comparisons producing 0 or -1 in each lane to | |||
25641 | // optimize away operation when it's from a constant. | |||
25642 | // | |||
25643 | // The general transformation is: | |||
25644 | // UNARYOP(AND(VECTOR_CMP(x,y), constant)) --> | |||
25645 | // AND(VECTOR_CMP(x,y), constant2) | |||
25646 | // constant2 = UNARYOP(constant) | |||
25647 | ||||
25648 | // Early exit if this isn't a vector operation, the operand of the | |||
25649 | // unary operation isn't a bitwise AND, or if the sizes of the operations | |||
25650 | // aren't the same. | |||
25651 | EVT VT = N->getValueType(0); | |||
25652 | if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND || | |||
25653 | N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC || | |||
25654 | VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits()) | |||
25655 | return SDValue(); | |||
25656 | ||||
25657 | // Now check that the other operand of the AND is a constant. We could | |||
25658 | // make the transformation for non-constant splats as well, but it's unclear | |||
25659 | // that would be a benefit as it would not eliminate any operations, just | |||
25660 | // perform one more step in scalar code before moving to the vector unit. | |||
25661 | if (BuildVectorSDNode *BV = | |||
25662 | dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) { | |||
25663 | // Bail out if the vector isn't a constant. | |||
25664 | if (!BV->isConstant()) | |||
25665 | return SDValue(); | |||
25666 | ||||
25667 | // Everything checks out. Build up the new and improved node. | |||
25668 | SDLoc DL(N); | |||
25669 | EVT IntVT = BV->getValueType(0); | |||
25670 | // Create a new constant of the appropriate type for the transformed | |||
25671 | // DAG. | |||
25672 | SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0)); | |||
25673 | // The AND node needs bitcasts to/from an integer vector type around it. | |||
25674 | SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst); | |||
25675 | SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT, | |||
25676 | N->getOperand(0)->getOperand(0), MaskConst); | |||
25677 | SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd); | |||
25678 | return Res; | |||
25679 | } | |||
25680 | ||||
25681 | return SDValue(); | |||
25682 | } | |||
25683 | ||||
25684 | static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG, | |||
25685 | const X86TargetLowering *XTLI) { | |||
25686 | // First try to optimize away the conversion entirely when it's | |||
25687 | // conditionally from a constant. Vectors only. | |||
25688 | SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG); | |||
25689 | if (Res != SDValue()) | |||
25690 | return Res; | |||
25691 | ||||
25692 | // Now move on to more general possibilities. | |||
25693 | SDValue Op0 = N->getOperand(0); | |||
25694 | EVT InVT = Op0->getValueType(0); | |||
25695 | ||||
25696 | // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32)) | |||
25697 | if (InVT == MVT::v8i8 || InVT == MVT::v4i8) { | |||
25698 | SDLoc dl(N); | |||
25699 | MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32; | |||
25700 | SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0); | |||
25701 | return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P); | |||
25702 | } | |||
25703 | ||||
25704 | // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have | |||
25705 | // a 32-bit target where SSE doesn't support i64->FP operations. | |||
25706 | if (Op0.getOpcode() == ISD::LOAD) { | |||
25707 | LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode()); | |||
25708 | EVT VT = Ld->getValueType(0); | |||
25709 | if (!Ld->isVolatile() && !N->getValueType(0).isVector() && | |||
25710 | ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() && | |||
25711 | !XTLI->getSubtarget()->is64Bit() && | |||
25712 | VT == MVT::i64) { | |||
25713 | SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0), | |||
25714 | Ld->getChain(), Op0, DAG); | |||
25715 | DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1)); | |||
25716 | return FILDChain; | |||
25717 | } | |||
25718 | } | |||
25719 | return SDValue(); | |||
25720 | } | |||
25721 | ||||
25722 | // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS | |||
25723 | static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG, | |||
25724 | X86TargetLowering::DAGCombinerInfo &DCI) { | |||
25725 | // If the LHS and RHS of the ADC node are zero, then it can't overflow and | |||
25726 | // the result is either zero or one (depending on the input carry bit). | |||
25727 | // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1. | |||
25728 | if (X86::isZeroNode(N->getOperand(0)) && | |||
25729 | X86::isZeroNode(N->getOperand(1)) && | |||
25730 | // We don't have a good way to replace an EFLAGS use, so only do this when | |||
25731 | // dead right now. | |||
25732 | SDValue(N, 1).use_empty()) { | |||
25733 | SDLoc DL(N); | |||
25734 | EVT VT = N->getValueType(0); | |||
25735 | SDValue CarryOut = DAG.getConstant(0, N->getValueType(1)); | |||
25736 | SDValue Res1 = DAG.getNode(ISD::AND, DL, VT, | |||
25737 | DAG.getNode(X86ISD::SETCC_CARRY, DL, VT, | |||
25738 | DAG.getConstant(X86::COND_B,MVT::i8), | |||
25739 | N->getOperand(2)), | |||
25740 | DAG.getConstant(1, VT)); | |||
25741 | return DCI.CombineTo(N, Res1, CarryOut); | |||
25742 | } | |||
25743 | ||||
25744 | return SDValue(); | |||
25745 | } | |||
25746 | ||||
25747 | // fold (add Y, (sete X, 0)) -> adc 0, Y | |||
25748 | // (add Y, (setne X, 0)) -> sbb -1, Y | |||
25749 | // (sub (sete X, 0), Y) -> sbb 0, Y | |||
25750 | // (sub (setne X, 0), Y) -> adc -1, Y | |||
25751 | static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) { | |||
25752 | SDLoc DL(N); | |||
25753 | ||||
25754 | // Look through ZExts. | |||
25755 | SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0); | |||
25756 | if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse()) | |||
25757 | return SDValue(); | |||
25758 | ||||
25759 | SDValue SetCC = Ext.getOperand(0); | |||
25760 | if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse()) | |||
25761 | return SDValue(); | |||
25762 | ||||
25763 | X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0); | |||
25764 | if (CC != X86::COND_E && CC != X86::COND_NE) | |||
25765 | return SDValue(); | |||
25766 | ||||
25767 | SDValue Cmp = SetCC.getOperand(1); | |||
25768 | if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() || | |||
25769 | !X86::isZeroNode(Cmp.getOperand(1)) || | |||
25770 | !Cmp.getOperand(0).getValueType().isInteger()) | |||
25771 | return SDValue(); | |||
25772 | ||||
25773 | SDValue CmpOp0 = Cmp.getOperand(0); | |||
25774 | SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0, | |||
25775 | DAG.getConstant(1, CmpOp0.getValueType())); | |||
25776 | ||||
25777 | SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1); | |||
25778 | if (CC == X86::COND_NE) | |||
25779 | return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB, | |||
25780 | DL, OtherVal.getValueType(), OtherVal, | |||
25781 | DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp); | |||
25782 | return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC, | |||
25783 | DL, OtherVal.getValueType(), OtherVal, | |||
25784 | DAG.getConstant(0, OtherVal.getValueType()), NewCmp); | |||
25785 | } | |||
25786 | ||||
25787 | /// PerformADDCombine - Do target-specific dag combines on integer adds. | |||
25788 | static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG, | |||
25789 | const X86Subtarget *Subtarget) { | |||
25790 | EVT VT = N->getValueType(0); | |||
25791 | SDValue Op0 = N->getOperand(0); | |||
25792 | SDValue Op1 = N->getOperand(1); | |||
25793 | ||||
25794 | // Try to synthesize horizontal adds from adds of shuffles. | |||
25795 | if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) || | |||
25796 | (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) && | |||
25797 | isHorizontalBinOp(Op0, Op1, true)) | |||
25798 | return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1); | |||
25799 | ||||
25800 | return OptimizeConditionalInDecrement(N, DAG); | |||
25801 | } | |||
25802 | ||||
25803 | static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG, | |||
25804 | const X86Subtarget *Subtarget) { | |||
25805 | SDValue Op0 = N->getOperand(0); | |||
25806 | SDValue Op1 = N->getOperand(1); | |||
25807 | ||||
25808 | // X86 can't encode an immediate LHS of a sub. See if we can push the | |||
25809 | // negation into a preceding instruction. | |||
25810 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) { | |||
25811 | // If the RHS of the sub is a XOR with one use and a constant, invert the | |||
25812 | // immediate. Then add one to the LHS of the sub so we can turn | |||
25813 | // X-Y -> X+~Y+1, saving one register. | |||
25814 | if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR && | |||
25815 | isa<ConstantSDNode>(Op1.getOperand(1))) { | |||
25816 | APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue(); | |||
25817 | EVT VT = Op0.getValueType(); | |||
25818 | SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT, | |||
25819 | Op1.getOperand(0), | |||
25820 | DAG.getConstant(~XorC, VT)); | |||
25821 | return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor, | |||
25822 | DAG.getConstant(C->getAPIntValue()+1, VT)); | |||
25823 | } | |||
25824 | } | |||
25825 | ||||
25826 | // Try to synthesize horizontal adds from adds of shuffles. | |||
25827 | EVT VT = N->getValueType(0); | |||
25828 | if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) || | |||
25829 | (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) && | |||
25830 | isHorizontalBinOp(Op0, Op1, true)) | |||
25831 | return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1); | |||
25832 | ||||
25833 | return OptimizeConditionalInDecrement(N, DAG); | |||
25834 | } | |||
25835 | ||||
25836 | /// performVZEXTCombine - Performs build vector combines | |||
25837 | static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG, | |||
25838 | TargetLowering::DAGCombinerInfo &DCI, | |||
25839 | const X86Subtarget *Subtarget) { | |||
25840 | SDLoc DL(N); | |||
25841 | MVT VT = N->getSimpleValueType(0); | |||
25842 | SDValue Op = N->getOperand(0); | |||
25843 | MVT OpVT = Op.getSimpleValueType(); | |||
25844 | MVT OpEltVT = OpVT.getVectorElementType(); | |||
25845 | unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements(); | |||
25846 | ||||
25847 | // (vzext (bitcast (vzext (x)) -> (vzext x) | |||
25848 | SDValue V = Op; | |||
25849 | while (V.getOpcode() == ISD::BITCAST) | |||
25850 | V = V.getOperand(0); | |||
25851 | ||||
25852 | if (V != Op && V.getOpcode() == X86ISD::VZEXT) { | |||
25853 | MVT InnerVT = V.getSimpleValueType(); | |||
25854 | MVT InnerEltVT = InnerVT.getVectorElementType(); | |||
25855 | ||||
25856 | // If the element sizes match exactly, we can just do one larger vzext. This | |||
25857 | // is always an exact type match as vzext operates on integer types. | |||
25858 | if (OpEltVT == InnerEltVT) { | |||
25859 | 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-3.7~svn227765/lib/Target/X86/X86ISelLowering.cpp" , 25859, __PRETTY_FUNCTION__)); | |||
25860 | return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0)); | |||
25861 | } | |||
25862 | ||||
25863 | // The only other way we can combine them is if only a single element of the | |||
25864 | // inner vzext is used in the input to the outer vzext. | |||
25865 | if (InnerEltVT.getSizeInBits() < InputBits) | |||
25866 | return SDValue(); | |||
25867 | ||||
25868 | // In this case, the inner vzext is completely dead because we're going to | |||
25869 | // only look at bits inside of the low element. Just do the outer vzext on | |||
25870 | // a bitcast of the input to the inner. | |||
25871 | return DAG.getNode(X86ISD::VZEXT, DL, VT, | |||
25872 | DAG.getNode(ISD::BITCAST, DL, OpVT, V)); | |||
25873 | } | |||
25874 | ||||
25875 | // Check if we can bypass extracting and re-inserting an element of an input | |||
25876 | // vector. Essentialy: | |||
25877 | // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x) | |||
25878 | if (V.getOpcode() == ISD::SCALAR_TO_VECTOR && | |||
25879 | V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT && | |||
25880 | V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) { | |||
25881 | SDValue ExtractedV = V.getOperand(0); | |||
25882 | SDValue OrigV = ExtractedV.getOperand(0); | |||
25883 | if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1))) | |||
25884 | if (ExtractIdx->getZExtValue() == 0) { | |||
25885 | MVT OrigVT = OrigV.getSimpleValueType(); | |||
25886 | // Extract a subvector if necessary... | |||
25887 | if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) { | |||
25888 | int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits(); | |||
25889 | OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(), | |||
25890 | OrigVT.getVectorNumElements() / Ratio); | |||
25891 | OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV, | |||
25892 | DAG.getIntPtrConstant(0)); | |||
25893 | } | |||
25894 | Op = DAG.getNode(ISD::BITCAST, DL, OpVT, OrigV); | |||
25895 | return DAG.getNode(X86ISD::VZEXT, DL, VT, Op); | |||
25896 | } | |||
25897 | } | |||
25898 | ||||
25899 | return SDValue(); | |||
25900 | } | |||
25901 | ||||
25902 | SDValue X86TargetLowering::PerformDAGCombine(SDNode *N, | |||
25903 | DAGCombinerInfo &DCI) const { | |||
25904 | SelectionDAG &DAG = DCI.DAG; | |||
25905 | switch (N->getOpcode()) { | |||
25906 | default: break; | |||
25907 | case ISD::EXTRACT_VECTOR_ELT: | |||
25908 | return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI); | |||
25909 | case ISD::VSELECT: | |||
25910 | case ISD::SELECT: | |||
25911 | case X86ISD::SHRUNKBLEND: | |||
25912 | return PerformSELECTCombine(N, DAG, DCI, Subtarget); | |||
25913 | case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget); | |||
25914 | case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget); | |||
25915 | case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget); | |||
25916 | case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI); | |||
25917 | case ISD::MUL: return PerformMulCombine(N, DAG, DCI); | |||
25918 | case ISD::SHL: | |||
25919 | case ISD::SRA: | |||
25920 | case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget); | |||
25921 | case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget); | |||
25922 | case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget); | |||
25923 | case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget); | |||
25924 | case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget); | |||
25925 | case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget); | |||
25926 | case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget); | |||
25927 | case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget); | |||
25928 | case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this); | |||
25929 | case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget); | |||
25930 | case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget); | |||
25931 | case X86ISD::FXOR: | |||
25932 | case X86ISD::FOR: return PerformFORCombine(N, DAG); | |||
25933 | case X86ISD::FMIN: | |||
25934 | case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG); | |||
25935 | case X86ISD::FAND: return PerformFANDCombine(N, DAG); | |||
25936 | case X86ISD::FANDN: return PerformFANDNCombine(N, DAG); | |||
25937 | case X86ISD::BT: return PerformBTCombine(N, DAG, DCI); | |||
25938 | case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG); | |||
25939 | case ISD::ANY_EXTEND: | |||
25940 | case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget); | |||
25941 | case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget); | |||
25942 | case ISD::SIGN_EXTEND_INREG: | |||
25943 | return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget); | |||
25944 | case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget); | |||
25945 | case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget); | |||
25946 | case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget); | |||
25947 | case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget); | |||
25948 | case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget); | |||
25949 | case X86ISD::SHUFP: // Handle all target specific shuffles | |||
25950 | case X86ISD::PALIGNR: | |||
25951 | case X86ISD::UNPCKH: | |||
25952 | case X86ISD::UNPCKL: | |||
25953 | case X86ISD::MOVHLPS: | |||
25954 | case X86ISD::MOVLHPS: | |||
25955 | case X86ISD::PSHUFB: | |||
25956 | case X86ISD::PSHUFD: | |||
25957 | case X86ISD::PSHUFHW: | |||
25958 | case X86ISD::PSHUFLW: | |||
25959 | case X86ISD::MOVSS: | |||
25960 | case X86ISD::MOVSD: | |||
25961 | case X86ISD::VPERMILPI: | |||
25962 | case X86ISD::VPERM2X128: | |||
25963 | case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget); | |||
25964 | case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget); | |||
25965 | case ISD::INTRINSIC_WO_CHAIN: | |||
25966 | return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget); | |||
25967 | case X86ISD::INSERTPS: { | |||
25968 | if (getTargetMachine().getOptLevel() > CodeGenOpt::None) | |||
25969 | return PerformINSERTPSCombine(N, DAG, Subtarget); | |||
25970 | break; | |||
25971 | } | |||
25972 | case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DAG, Subtarget); | |||
25973 | } | |||
25974 | ||||
25975 | return SDValue(); | |||
25976 | } | |||
25977 | ||||
25978 | /// isTypeDesirableForOp - Return true if the target has native support for | |||
25979 | /// the specified value type and it is 'desirable' to use the type for the | |||
25980 | /// given node type. e.g. On x86 i16 is legal, but undesirable since i16 | |||
25981 | /// instruction encodings are longer and some i16 instructions are slow. | |||
25982 | bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const { | |||
25983 | if (!isTypeLegal(VT)) | |||
25984 | return false; | |||
25985 | if (VT != MVT::i16) | |||
25986 | return true; | |||
25987 | ||||
25988 | switch (Opc) { | |||
25989 | default: | |||
25990 | return true; | |||
25991 | case ISD::LOAD: | |||
25992 | case ISD::SIGN_EXTEND: | |||
25993 | case ISD::ZERO_EXTEND: | |||
25994 | case ISD::ANY_EXTEND: | |||
25995 | case ISD::SHL: | |||
25996 | case ISD::SRL: | |||
25997 | case ISD::SUB: | |||
25998 | case ISD::ADD: | |||
25999 | case ISD::MUL: | |||
26000 | case ISD::AND: | |||
26001 | case ISD::OR: | |||
26002 | case ISD::XOR: | |||
26003 | return false; | |||
26004 | } | |||
26005 | } | |||
26006 | ||||
26007 | /// IsDesirableToPromoteOp - This method query the target whether it is | |||
26008 | /// beneficial for dag combiner to promote the specified node. If true, it | |||
26009 | /// should return the desired promotion type by reference. | |||
26010 | bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const { | |||
26011 | EVT VT = Op.getValueType(); | |||
26012 | if (VT != MVT::i16) | |||
26013 | return false; | |||
26014 | ||||
26015 | bool Promote = false; | |||
26016 | bool Commute = false; | |||
26017 | switch (Op.getOpcode()) { | |||
26018 | default: break; | |||
26019 | case ISD::LOAD: { | |||
26020 | LoadSDNode *LD = cast<LoadSDNode>(Op); | |||
26021 | // If the non-extending load has a single use and it's not live out, then it | |||
26022 | // might be folded. | |||
26023 | if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&& | |||
26024 | Op.hasOneUse()*/) { | |||
26025 | for (SDNode::use_iterator UI = Op.getNode()->use_begin(), | |||
26026 | UE = Op.getNode()->use_end(); UI != UE; ++UI) { | |||
26027 | // The only case where we'd want to promote LOAD (rather then it being | |||
26028 | // promoted as an operand is when it's only use is liveout. | |||
26029 | if (UI->getOpcode() != ISD::CopyToReg) | |||
26030 | return false; | |||
26031 | } | |||
26032 | } | |||
26033 | Promote = true; | |||
26034 | break; | |||
26035 | } | |||
26036 | case ISD::SIGN_EXTEND: | |||
26037 | case ISD::ZERO_EXTEND: | |||
26038 | case ISD::ANY_EXTEND: | |||
26039 | Promote = true; | |||
26040 | break; | |||
26041 | case ISD::SHL: | |||
26042 | case ISD::SRL: { | |||
26043 | SDValue N0 = Op.getOperand(0); | |||
26044 | // Look out for (store (shl (load), x)). | |||
26045 | if (MayFoldLoad(N0) && MayFoldIntoStore(Op)) | |||
26046 | return false; | |||
26047 | Promote = true; | |||
26048 | break; | |||
26049 | } | |||
26050 | case ISD::ADD: | |||
26051 | case ISD::MUL: | |||
26052 | case ISD::AND: | |||
26053 | case ISD::OR: | |||
26054 | case ISD::XOR: | |||
26055 | Commute = true; | |||
26056 | // fallthrough | |||
26057 | case ISD::SUB: { | |||
26058 | SDValue N0 = Op.getOperand(0); | |||
26059 | SDValue N1 = Op.getOperand(1); | |||
26060 | if (!Commute && MayFoldLoad(N1)) | |||
26061 | return false; | |||
26062 | // Avoid disabling potential load folding opportunities. | |||
26063 | if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op))) | |||
26064 | return false; | |||
26065 | if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op))) | |||
26066 | return false; | |||
26067 | Promote = true; | |||
26068 | } | |||
26069 | } | |||
26070 | ||||
26071 | PVT = MVT::i32; | |||
26072 | return Promote; | |||
26073 | } | |||
26074 | ||||
26075 | //===----------------------------------------------------------------------===// | |||
26076 | // X86 Inline Assembly Support | |||
26077 | //===----------------------------------------------------------------------===// | |||
26078 | ||||
26079 | namespace { | |||
26080 | // Helper to match a string separated by whitespace. | |||
26081 | bool matchAsmImpl(StringRef s, ArrayRef<const StringRef *> args) { | |||
26082 | s = s.substr(s.find_first_not_of(" \t")); // Skip leading whitespace. | |||
26083 | ||||
26084 | for (unsigned i = 0, e = args.size(); i != e; ++i) { | |||
26085 | StringRef piece(*args[i]); | |||
26086 | if (!s.startswith(piece)) // Check if the piece matches. | |||
26087 | return false; | |||
26088 | ||||
26089 | s = s.substr(piece.size()); | |||
26090 | StringRef::size_type pos = s.find_first_not_of(" \t"); | |||
26091 | if (pos == 0) // We matched a prefix. | |||
26092 | return false; | |||
26093 | ||||
26094 | s = s.substr(pos); | |||
26095 | } | |||
26096 | ||||
26097 | return s.empty(); | |||
26098 | } | |||
26099 | const VariadicFunction1<bool, StringRef, StringRef, matchAsmImpl> matchAsm={}; | |||
26100 | } | |||
26101 | ||||
26102 | static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) { | |||
26103 | ||||
26104 | if (AsmPieces.size() == 3 || AsmPieces.size() == 4) { | |||
26105 | if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") && | |||
26106 | std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") && | |||
26107 | std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) { | |||
26108 | ||||
26109 | if (AsmPieces.size() == 3) | |||
26110 | return true; | |||
26111 | else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}")) | |||
26112 | return true; | |||
26113 | } | |||
26114 | } | |||
26115 | return false; | |||
26116 | } | |||
26117 | ||||
26118 | bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const { | |||
26119 | InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue()); | |||
26120 | ||||
26121 | std::string AsmStr = IA->getAsmString(); | |||
26122 | ||||
26123 | IntegerType *Ty = dyn_cast<IntegerType>(CI->getType()); | |||
26124 | if (!Ty || Ty->getBitWidth() % 16 != 0) | |||
26125 | return false; | |||
26126 | ||||
26127 | // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a" | |||
26128 | SmallVector<StringRef, 4> AsmPieces; | |||
26129 | SplitString(AsmStr, AsmPieces, ";\n"); | |||
26130 | ||||
26131 | switch (AsmPieces.size()) { | |||
26132 | default: return false; | |||
26133 | case 1: | |||
26134 | // FIXME: this should verify that we are targeting a 486 or better. If not, | |||
26135 | // we will turn this bswap into something that will be lowered to logical | |||
26136 | // ops instead of emitting the bswap asm. For now, we don't support 486 or | |||
26137 | // lower so don't worry about this. | |||
26138 | // bswap $0 | |||
26139 | if (matchAsm(AsmPieces[0], "bswap", "$0") || | |||
26140 | matchAsm(AsmPieces[0], "bswapl", "$0") || | |||
26141 | matchAsm(AsmPieces[0], "bswapq", "$0") || | |||
26142 | matchAsm(AsmPieces[0], "bswap", "${0:q}") || | |||
26143 | matchAsm(AsmPieces[0], "bswapl", "${0:q}") || | |||
26144 | matchAsm(AsmPieces[0], "bswapq", "${0:q}")) { | |||
26145 | // No need to check constraints, nothing other than the equivalent of | |||
26146 | // "=r,0" would be valid here. | |||
26147 | return IntrinsicLowering::LowerToByteSwap(CI); | |||
26148 | } | |||
26149 | ||||
26150 | // rorw $$8, ${0:w} --> llvm.bswap.i16 | |||
26151 | if (CI->getType()->isIntegerTy(16) && | |||
26152 | IA->getConstraintString().compare(0, 5, "=r,0,") == 0 && | |||
26153 | (matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") || | |||
26154 | matchAsm(AsmPieces[0], "rolw", "$$8,", "${0:w}"))) { | |||
26155 | AsmPieces.clear(); | |||
26156 | const std::string &ConstraintsStr = IA->getConstraintString(); | |||
26157 | SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ","); | |||
26158 | array_pod_sort(AsmPieces.begin(), AsmPieces.end()); | |||
26159 | if (clobbersFlagRegisters(AsmPieces)) | |||
26160 | return IntrinsicLowering::LowerToByteSwap(CI); | |||
26161 | } | |||
26162 | break; | |||
26163 | case 3: | |||
26164 | if (CI->getType()->isIntegerTy(32) && | |||
26165 | IA->getConstraintString().compare(0, 5, "=r,0,") == 0 && | |||
26166 | matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") && | |||
26167 | matchAsm(AsmPieces[1], "rorl", "$$16,", "$0") && | |||
26168 | matchAsm(AsmPieces[2], "rorw", "$$8,", "${0:w}")) { | |||
26169 | AsmPieces.clear(); | |||
26170 | const std::string &ConstraintsStr = IA->getConstraintString(); | |||
26171 | SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ","); | |||
26172 | array_pod_sort(AsmPieces.begin(), AsmPieces.end()); | |||
26173 | if (clobbersFlagRegisters(AsmPieces)) | |||
26174 | return IntrinsicLowering::LowerToByteSwap(CI); | |||
26175 | } | |||
26176 | ||||
26177 | if (CI->getType()->isIntegerTy(64)) { | |||
26178 | InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints(); | |||
26179 | if (Constraints.size() >= 2 && | |||
26180 | Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" && | |||
26181 | Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") { | |||
26182 | // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64 | |||
26183 | if (matchAsm(AsmPieces[0], "bswap", "%eax") && | |||
26184 | matchAsm(AsmPieces[1], "bswap", "%edx") && | |||
26185 | matchAsm(AsmPieces[2], "xchgl", "%eax,", "%edx")) | |||
26186 | return IntrinsicLowering::LowerToByteSwap(CI); | |||
26187 | } | |||
26188 | } | |||
26189 | break; | |||
26190 | } | |||
26191 | return false; | |||
26192 | } | |||
26193 | ||||
26194 | /// getConstraintType - Given a constraint letter, return the type of | |||
26195 | /// constraint it is for this target. | |||
26196 | X86TargetLowering::ConstraintType | |||
26197 | X86TargetLowering::getConstraintType(const std::string &Constraint) const { | |||
26198 | if (Constraint.size() == 1) { | |||
26199 | switch (Constraint[0]) { | |||
26200 | case 'R': | |||
26201 | case 'q': | |||
26202 | case 'Q': | |||
26203 | case 'f': | |||
26204 | case 't': | |||
26205 | case 'u': | |||
26206 | case 'y': | |||
26207 | case 'x': | |||
26208 | case 'Y': | |||
26209 | case 'l': | |||
26210 | return C_RegisterClass; | |||
26211 | case 'a': | |||
26212 | case 'b': | |||
26213 | case 'c': | |||
26214 | case 'd': | |||
26215 | case 'S': | |||
26216 | case 'D': | |||
26217 | case 'A': | |||
26218 | return C_Register; | |||
26219 | case 'I': | |||
26220 | case 'J': | |||
26221 | case 'K': | |||
26222 | case 'L': | |||
26223 | case 'M': | |||
26224 | case 'N': | |||
26225 | case 'G': | |||
26226 | case 'C': | |||
26227 | case 'e': | |||
26228 | case 'Z': | |||
26229 | return C_Other; | |||
26230 | default: | |||
26231 | break; | |||
26232 | } | |||
26233 | } | |||
26234 | return TargetLowering::getConstraintType(Constraint); | |||
26235 | } | |||
26236 | ||||
26237 | /// Examine constraint type and operand type and determine a weight value. | |||
26238 | /// This object must already have been set up with the operand type | |||
26239 | /// and the current alternative constraint selected. | |||
26240 | TargetLowering::ConstraintWeight | |||
26241 | X86TargetLowering::getSingleConstraintMatchWeight( | |||
26242 | AsmOperandInfo &info, const char *constraint) const { | |||
26243 | ConstraintWeight weight = CW_Invalid; | |||
26244 | Value *CallOperandVal = info.CallOperandVal; | |||
26245 | // If we don't have a value, we can't do a match, | |||
26246 | // but allow it at the lowest weight. | |||
26247 | if (!CallOperandVal) | |||
26248 | return CW_Default; | |||
26249 | Type *type = CallOperandVal->getType(); | |||
26250 | // Look at the constraint type. | |||
26251 | switch (*constraint) { | |||
26252 | default: | |||
26253 | weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); | |||
26254 | case 'R': | |||
26255 | case 'q': | |||
26256 | case 'Q': | |||
26257 | case 'a': | |||
26258 | case 'b': | |||
26259 | case 'c': | |||
26260 | case 'd': | |||
26261 | case 'S': | |||
26262 | case 'D': | |||
26263 | case 'A': | |||
26264 | if (CallOperandVal->getType()->isIntegerTy()) | |||
26265 | weight = CW_SpecificReg; | |||
26266 | break; | |||
26267 | case 'f': | |||
26268 | case 't': | |||
26269 | case 'u': | |||
26270 | if (type->isFloatingPointTy()) | |||
26271 | weight = CW_SpecificReg; | |||
26272 | break; | |||
26273 | case 'y': | |||
26274 | if (type->isX86_MMXTy() && Subtarget->hasMMX()) | |||
26275 | weight = CW_SpecificReg; | |||
26276 | break; | |||
26277 | case 'x': | |||
26278 | case 'Y': | |||
26279 | if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) || | |||
26280 | ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256())) | |||
26281 | weight = CW_Register; | |||
26282 | break; | |||
26283 | case 'I': | |||
26284 | if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) { | |||
26285 | if (C->getZExtValue() <= 31) | |||
26286 | weight = CW_Constant; | |||
26287 | } | |||
26288 | break; | |||
26289 | case 'J': | |||
26290 | if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) { | |||
26291 | if (C->getZExtValue() <= 63) | |||
26292 | weight = CW_Constant; | |||
26293 | } | |||
26294 | break; | |||
26295 | case 'K': | |||
26296 | if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) { | |||
26297 | if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f)) | |||
26298 | weight = CW_Constant; | |||
26299 | } | |||
26300 | break; | |||
26301 | case 'L': | |||
26302 | if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) { | |||
26303 | if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff)) | |||
26304 | weight = CW_Constant; | |||
26305 | } | |||
26306 | break; | |||
26307 | case 'M': | |||
26308 | if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) { | |||
26309 | if (C->getZExtValue() <= 3) | |||
26310 | weight = CW_Constant; | |||
26311 | } | |||
26312 | break; | |||
26313 | case 'N': | |||
26314 | if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) { | |||
26315 | if (C->getZExtValue() <= 0xff) | |||
26316 | weight = CW_Constant; | |||
26317 | } | |||
26318 | break; | |||
26319 | case 'G': | |||
26320 | case 'C': | |||
26321 | if (dyn_cast<ConstantFP>(CallOperandVal)) { | |||
26322 | weight = CW_Constant; | |||
26323 | } | |||
26324 | break; | |||
26325 | case 'e': | |||
26326 | if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) { | |||
26327 | if ((C->getSExtValue() >= -0x80000000LL) && | |||
26328 | (C->getSExtValue() <= 0x7fffffffLL)) | |||
26329 | weight = CW_Constant; | |||
26330 | } | |||
26331 | break; | |||
26332 | case 'Z': | |||
26333 | if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) { | |||
26334 | if (C->getZExtValue() <= 0xffffffff) | |||
26335 | weight = CW_Constant; | |||
26336 | } | |||
26337 | break; | |||
26338 | } | |||
26339 | return weight; | |||
26340 | } | |||
26341 | ||||
26342 | /// LowerXConstraint - try to replace an X constraint, which matches anything, | |||
26343 | /// with another that has more specific requirements based on the type of the | |||
26344 | /// corresponding operand. | |||
26345 | const char *X86TargetLowering:: | |||
26346 | LowerXConstraint(EVT ConstraintVT) const { | |||
26347 | // FP X constraints get lowered to SSE1/2 registers if available, otherwise | |||
26348 | // 'f' like normal targets. | |||
26349 | if (ConstraintVT.isFloatingPoint()) { | |||
26350 | if (Subtarget->hasSSE2()) | |||
26351 | return "Y"; | |||
26352 | if (Subtarget->hasSSE1()) | |||
26353 | return "x"; | |||
26354 | } | |||
26355 | ||||
26356 | return TargetLowering::LowerXConstraint(ConstraintVT); | |||
26357 | } | |||
26358 | ||||
26359 | /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops | |||
26360 | /// vector. If it is invalid, don't add anything to Ops. | |||
26361 | void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op, | |||
26362 | std::string &Constraint, | |||
26363 | std::vector<SDValue>&Ops, | |||
26364 | SelectionDAG &DAG) const { | |||
26365 | SDValue Result; | |||
26366 | ||||
26367 | // Only support length 1 constraints for now. | |||
26368 | if (Constraint.length() > 1) return; | |||
26369 | ||||
26370 | char ConstraintLetter = Constraint[0]; | |||
26371 | switch (ConstraintLetter) { | |||
26372 | default: break; | |||
26373 | case 'I': | |||
26374 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { | |||
26375 | if (C->getZExtValue() <= 31) { | |||
26376 | Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType()); | |||
26377 | break; | |||
26378 | } | |||
26379 | } | |||
26380 | return; | |||
26381 | case 'J': | |||
26382 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { | |||
26383 | if (C->getZExtValue() <= 63) { | |||
26384 | Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType()); | |||
26385 | break; | |||
26386 | } | |||
26387 | } | |||
26388 | return; | |||
26389 | case 'K': | |||
26390 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { | |||
26391 | if (isInt<8>(C->getSExtValue())) { | |||
26392 | Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType()); | |||
26393 | break; | |||
26394 | } | |||
26395 | } | |||
26396 | return; | |||
26397 | case 'L': | |||
26398 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { | |||
26399 | if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff || | |||
26400 | (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) { | |||
26401 | Result = DAG.getTargetConstant(C->getSExtValue(), Op.getValueType()); | |||
26402 | break; | |||
26403 | } | |||
26404 | } | |||
26405 | return; | |||
26406 | case 'M': | |||
26407 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { | |||
26408 | if (C->getZExtValue() <= 3) { | |||
26409 | Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType()); | |||
26410 | break; | |||
26411 | } | |||
26412 | } | |||
26413 | return; | |||
26414 | case 'N': | |||
26415 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { | |||
26416 | if (C->getZExtValue() <= 255) { | |||
26417 | Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType()); | |||
26418 | break; | |||
26419 | } | |||
26420 | } | |||
26421 | return; | |||
26422 | case 'O': | |||
26423 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { | |||
26424 | if (C->getZExtValue() <= 127) { | |||
26425 | Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType()); | |||
26426 | break; | |||
26427 | } | |||
26428 | } | |||
26429 | return; | |||
26430 | case 'e': { | |||
26431 | // 32-bit signed value | |||
26432 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { | |||
26433 | if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()), | |||
26434 | C->getSExtValue())) { | |||
26435 | // Widen to 64 bits here to get it sign extended. | |||
26436 | Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64); | |||
26437 | break; | |||
26438 | } | |||
26439 | // FIXME gcc accepts some relocatable values here too, but only in certain | |||
26440 | // memory models; it's complicated. | |||
26441 | } | |||
26442 | return; | |||
26443 | } | |||
26444 | case 'Z': { | |||
26445 | // 32-bit unsigned value | |||
26446 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { | |||
26447 | if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()), | |||
26448 | C->getZExtValue())) { | |||
26449 | Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType()); | |||
26450 | break; | |||
26451 | } | |||
26452 | } | |||
26453 | // FIXME gcc accepts some relocatable values here too, but only in certain | |||
26454 | // memory models; it's complicated. | |||
26455 | return; | |||
26456 | } | |||
26457 | case 'i': { | |||
26458 | // Literal immediates are always ok. | |||
26459 | if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) { | |||
26460 | // Widen to 64 bits here to get it sign extended. | |||
26461 | Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64); | |||
26462 | break; | |||
26463 | } | |||
26464 | ||||
26465 | // In any sort of PIC mode addresses need to be computed at runtime by | |||
26466 | // adding in a register or some sort of table lookup. These can't | |||
26467 | // be used as immediates. | |||
26468 | if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC()) | |||
26469 | return; | |||
26470 | ||||
26471 | // If we are in non-pic codegen mode, we allow the address of a global (with | |||
26472 | // an optional displacement) to be used with 'i'. | |||
26473 | GlobalAddressSDNode *GA = nullptr; | |||
26474 | int64_t Offset = 0; | |||
26475 | ||||
26476 | // Match either (GA), (GA+C), (GA+C1+C2), etc. | |||
26477 | while (1) { | |||
26478 | if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) { | |||
26479 | Offset += GA->getOffset(); | |||
26480 | break; | |||
26481 | } else if (Op.getOpcode() == ISD::ADD) { | |||
26482 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { | |||
26483 | Offset += C->getZExtValue(); | |||
26484 | Op = Op.getOperand(0); | |||
26485 | continue; | |||
26486 | } | |||
26487 | } else if (Op.getOpcode() == ISD::SUB) { | |||
26488 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { | |||
26489 | Offset += -C->getZExtValue(); | |||
26490 | Op = Op.getOperand(0); | |||
26491 | continue; | |||
26492 | } | |||
26493 | } | |||
26494 | ||||
26495 | // Otherwise, this isn't something we can handle, reject it. | |||
26496 | return; | |||
26497 | } | |||
26498 | ||||
26499 | const GlobalValue *GV = GA->getGlobal(); | |||
26500 | // If we require an extra load to get this address, as in PIC mode, we | |||
26501 | // can't accept it. | |||
26502 | if (isGlobalStubReference( | |||
26503 | Subtarget->ClassifyGlobalReference(GV, DAG.getTarget()))) | |||
26504 | return; | |||
26505 | ||||
26506 | Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op), | |||
26507 | GA->getValueType(0), Offset); | |||
26508 | break; | |||
26509 | } | |||
26510 | } | |||
26511 | ||||
26512 | if (Result.getNode()) { | |||
26513 | Ops.push_back(Result); | |||
26514 | return; | |||
26515 | } | |||
26516 | return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); | |||
26517 | } | |||
26518 | ||||
26519 | std::pair<unsigned, const TargetRegisterClass*> | |||
26520 | X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint, | |||
26521 | MVT VT) const { | |||
26522 | // First, see if this is a constraint that directly corresponds to an LLVM | |||
26523 | // register class. | |||
26524 | if (Constraint.size() == 1) { | |||
26525 | // GCC Constraint Letters | |||
26526 | switch (Constraint[0]) { | |||
26527 | default: break; | |||
26528 | // TODO: Slight differences here in allocation order and leaving | |||
26529 | // RIP in the class. Do they matter any more here than they do | |||
26530 | // in the normal allocation? | |||
26531 | case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode. | |||
26532 | if (Subtarget->is64Bit()) { | |||
26533 | if (VT == MVT::i32 || VT == MVT::f32) | |||
26534 | return std::make_pair(0U, &X86::GR32RegClass); | |||
26535 | if (VT == MVT::i16) | |||
26536 | return std::make_pair(0U, &X86::GR16RegClass); | |||
26537 | if (VT == MVT::i8 || VT == MVT::i1) | |||
26538 | return std::make_pair(0U, &X86::GR8RegClass); | |||
26539 | if (VT == MVT::i64 || VT == MVT::f64) | |||
26540 | return std::make_pair(0U, &X86::GR64RegClass); | |||
26541 | break; | |||
26542 | } | |||
26543 | // 32-bit fallthrough | |||
26544 | case 'Q': // Q_REGS | |||
26545 | if (VT == MVT::i32 || VT == MVT::f32) | |||
26546 | return std::make_pair(0U, &X86::GR32_ABCDRegClass); | |||
26547 | if (VT == MVT::i16) | |||
26548 | return std::make_pair(0U, &X86::GR16_ABCDRegClass); | |||
26549 | if (VT == MVT::i8 || VT == MVT::i1) | |||
26550 | return std::make_pair(0U, &X86::GR8_ABCD_LRegClass); | |||
26551 | if (VT == MVT::i64) | |||
26552 | return std::make_pair(0U, &X86::GR64_ABCDRegClass); | |||
26553 | break; | |||
26554 | case 'r': // GENERAL_REGS | |||
26555 | case 'l': // INDEX_REGS | |||
26556 | if (VT == MVT::i8 || VT == MVT::i1) | |||
26557 | return std::make_pair(0U, &X86::GR8RegClass); | |||
26558 | if (VT == MVT::i16) | |||
26559 | return std::make_pair(0U, &X86::GR16RegClass); | |||
26560 | if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit()) | |||
26561 | return std::make_pair(0U, &X86::GR32RegClass); | |||
26562 | return std::make_pair(0U, &X86::GR64RegClass); | |||
26563 | case 'R': // LEGACY_REGS | |||
26564 | if (VT == MVT::i8 || VT == MVT::i1) | |||
26565 | return std::make_pair(0U, &X86::GR8_NOREXRegClass); | |||
26566 | if (VT == MVT::i16) | |||
26567 | return std::make_pair(0U, &X86::GR16_NOREXRegClass); | |||
26568 | if (VT == MVT::i32 || !Subtarget->is64Bit()) | |||
26569 | return std::make_pair(0U, &X86::GR32_NOREXRegClass); | |||
26570 | return std::make_pair(0U, &X86::GR64_NOREXRegClass); | |||
26571 | case 'f': // FP Stack registers. | |||
26572 | // If SSE is enabled for this VT, use f80 to ensure the isel moves the | |||
26573 | // value to the correct fpstack register class. | |||
26574 | if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT)) | |||
26575 | return std::make_pair(0U, &X86::RFP32RegClass); | |||
26576 | if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT)) | |||
26577 | return std::make_pair(0U, &X86::RFP64RegClass); | |||
26578 | return std::make_pair(0U, &X86::RFP80RegClass); | |||
26579 | case 'y': // MMX_REGS if MMX allowed. | |||
26580 | if (!Subtarget->hasMMX()) break; | |||
26581 | return std::make_pair(0U, &X86::VR64RegClass); | |||
26582 | case 'Y': // SSE_REGS if SSE2 allowed | |||
26583 | if (!Subtarget->hasSSE2()) break; | |||
26584 | // FALL THROUGH. | |||
26585 | case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed | |||
26586 | if (!Subtarget->hasSSE1()) break; | |||
26587 | ||||
26588 | switch (VT.SimpleTy) { | |||
26589 | default: break; | |||
26590 | // Scalar SSE types. | |||
26591 | case MVT::f32: | |||
26592 | case MVT::i32: | |||
26593 | return std::make_pair(0U, &X86::FR32RegClass); | |||
26594 | case MVT::f64: | |||
26595 | case MVT::i64: | |||
26596 | return std::make_pair(0U, &X86::FR64RegClass); | |||
26597 | // Vector types. | |||
26598 | case MVT::v16i8: | |||
26599 | case MVT::v8i16: | |||
26600 | case MVT::v4i32: | |||
26601 | case MVT::v2i64: | |||
26602 | case MVT::v4f32: | |||
26603 | case MVT::v2f64: | |||
26604 | return std::make_pair(0U, &X86::VR128RegClass); | |||
26605 | // AVX types. | |||
26606 | case MVT::v32i8: | |||
26607 | case MVT::v16i16: | |||
26608 | case MVT::v8i32: | |||
26609 | case MVT::v4i64: | |||
26610 | case MVT::v8f32: | |||
26611 | case MVT::v4f64: | |||
26612 | return std::make_pair(0U, &X86::VR256RegClass); | |||
26613 | case MVT::v8f64: | |||
26614 | case MVT::v16f32: | |||
26615 | case MVT::v16i32: | |||
26616 | case MVT::v8i64: | |||
26617 | return std::make_pair(0U, &X86::VR512RegClass); | |||
26618 | } | |||
26619 | break; | |||
26620 | } | |||
26621 | } | |||
26622 | ||||
26623 | // Use the default implementation in TargetLowering to convert the register | |||
26624 | // constraint into a member of a register class. | |||
26625 | std::pair<unsigned, const TargetRegisterClass*> Res; | |||
26626 | Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); | |||
26627 | ||||
26628 | // Not found as a standard register? | |||
26629 | if (!Res.second) { | |||
26630 | // Map st(0) -> st(7) -> ST0 | |||
26631 | if (Constraint.size() == 7 && Constraint[0] == '{' && | |||
26632 | tolower(Constraint[1]) == 's' && | |||
26633 | tolower(Constraint[2]) == 't' && | |||
26634 | Constraint[3] == '(' && | |||
26635 | (Constraint[4] >= '0' && Constraint[4] <= '7') && | |||
26636 | Constraint[5] == ')' && | |||
26637 | Constraint[6] == '}') { | |||
26638 | ||||
26639 | Res.first = X86::FP0+Constraint[4]-'0'; | |||
26640 | Res.second = &X86::RFP80RegClass; | |||
26641 | return Res; | |||
26642 | } | |||
26643 | ||||
26644 | // GCC allows "st(0)" to be called just plain "st". | |||
26645 | if (StringRef("{st}").equals_lower(Constraint)) { | |||
26646 | Res.first = X86::FP0; | |||
26647 | Res.second = &X86::RFP80RegClass; | |||
26648 | return Res; | |||
26649 | } | |||
26650 | ||||
26651 | // flags -> EFLAGS | |||
26652 | if (StringRef("{flags}").equals_lower(Constraint)) { | |||
26653 | Res.first = X86::EFLAGS; | |||
26654 | Res.second = &X86::CCRRegClass; | |||
26655 | return Res; | |||
26656 | } | |||
26657 | ||||
26658 | // 'A' means EAX + EDX. | |||
26659 | if (Constraint == "A") { | |||
26660 | Res.first = X86::EAX; | |||
26661 | Res.second = &X86::GR32_ADRegClass; | |||
26662 | return Res; | |||
26663 | } | |||
26664 | return Res; | |||
26665 | } | |||
26666 | ||||
26667 | // Otherwise, check to see if this is a register class of the wrong value | |||
26668 | // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to | |||
26669 | // turn into {ax},{dx}. | |||
26670 | if (Res.second->hasType(VT)) | |||
26671 | return Res; // Correct type already, nothing to do. | |||
26672 | ||||
26673 | // All of the single-register GCC register classes map their values onto | |||
26674 | // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we | |||
26675 | // really want an 8-bit or 32-bit register, map to the appropriate register | |||
26676 | // class and return the appropriate register. | |||
26677 | if (Res.second == &X86::GR16RegClass) { | |||
26678 | if (VT == MVT::i8 || VT == MVT::i1) { | |||
26679 | unsigned DestReg = 0; | |||
26680 | switch (Res.first) { | |||
26681 | default: break; | |||
26682 | case X86::AX: DestReg = X86::AL; break; | |||
26683 | case X86::DX: DestReg = X86::DL; break; | |||
26684 | case X86::CX: DestReg = X86::CL; break; | |||
26685 | case X86::BX: DestReg = X86::BL; break; | |||
26686 | } | |||
26687 | if (DestReg) { | |||
26688 | Res.first = DestReg; | |||
26689 | Res.second = &X86::GR8RegClass; | |||
26690 | } | |||
26691 | } else if (VT == MVT::i32 || VT == MVT::f32) { | |||
26692 | unsigned DestReg = 0; | |||
26693 | switch (Res.first) { | |||
26694 | default: break; | |||
26695 | case X86::AX: DestReg = X86::EAX; break; | |||
26696 | case X86::DX: DestReg = X86::EDX; break; | |||
26697 | case X86::CX: DestReg = X86::ECX; break; | |||
26698 | case X86::BX: DestReg = X86::EBX; break; | |||
26699 | case X86::SI: DestReg = X86::ESI; break; | |||
26700 | case X86::DI: DestReg = X86::EDI; break; | |||
26701 | case X86::BP: DestReg = X86::EBP; break; | |||
26702 | case X86::SP: DestReg = X86::ESP; break; | |||
26703 | } | |||
26704 | if (DestReg) { | |||
26705 | Res.first = DestReg; | |||
26706 | Res.second = &X86::GR32RegClass; | |||
26707 | } | |||
26708 | } else if (VT == MVT::i64 || VT == MVT::f64) { | |||
26709 | unsigned DestReg = 0; | |||
26710 | switch (Res.first) { | |||
26711 | default: break; | |||
26712 | case X86::AX: DestReg = X86::RAX; break; | |||
26713 | case X86::DX: DestReg = X86::RDX; break; | |||
26714 | case X86::CX: DestReg = X86::RCX; break; | |||
26715 | case X86::BX: DestReg = X86::RBX; break; | |||
26716 | case X86::SI: DestReg = X86::RSI; break; | |||
26717 | case X86::DI: DestReg = X86::RDI; break; | |||
26718 | case X86::BP: DestReg = X86::RBP; break; | |||
26719 | case X86::SP: DestReg = X86::RSP; break; | |||
26720 | } | |||
26721 | if (DestReg) { | |||
26722 | Res.first = DestReg; | |||
26723 | Res.second = &X86::GR64RegClass; | |||
26724 | } | |||
26725 | } | |||
26726 | } else if (Res.second == &X86::FR32RegClass || | |||
26727 | Res.second == &X86::FR64RegClass || | |||
26728 | Res.second == &X86::VR128RegClass || | |||
26729 | Res.second == &X86::VR256RegClass || | |||
26730 | Res.second == &X86::FR32XRegClass || | |||
26731 | Res.second == &X86::FR64XRegClass || | |||
26732 | Res.second == &X86::VR128XRegClass || | |||
26733 | Res.second == &X86::VR256XRegClass || | |||
26734 | Res.second == &X86::VR512RegClass) { | |||
26735 | // Handle references to XMM physical registers that got mapped into the | |||
26736 | // wrong class. This can happen with constraints like {xmm0} where the | |||
26737 | // target independent register mapper will just pick the first match it can | |||
26738 | // find, ignoring the required type. | |||
26739 | ||||
26740 | if (VT == MVT::f32 || VT == MVT::i32) | |||
26741 | Res.second = &X86::FR32RegClass; | |||
26742 | else if (VT == MVT::f64 || VT == MVT::i64) | |||
26743 | Res.second = &X86::FR64RegClass; | |||
26744 | else if (X86::VR128RegClass.hasType(VT)) | |||
26745 | Res.second = &X86::VR128RegClass; | |||
26746 | else if (X86::VR256RegClass.hasType(VT)) | |||
26747 | Res.second = &X86::VR256RegClass; | |||
26748 | else if (X86::VR512RegClass.hasType(VT)) | |||
26749 | Res.second = &X86::VR512RegClass; | |||
26750 | } | |||
26751 | ||||
26752 | return Res; | |||
26753 | } | |||
26754 | ||||
26755 | int X86TargetLowering::getScalingFactorCost(const AddrMode &AM, | |||
26756 | Type *Ty) const { | |||
26757 | // Scaling factors are not free at all. | |||
26758 | // An indexed folded instruction, i.e., inst (reg1, reg2, scale), | |||
26759 | // will take 2 allocations in the out of order engine instead of 1 | |||
26760 | // for plain addressing mode, i.e. inst (reg1). | |||
26761 | // E.g., | |||
26762 | // vaddps (%rsi,%drx), %ymm0, %ymm1 | |||
26763 | // Requires two allocations (one for the load, one for the computation) | |||
26764 | // whereas: | |||
26765 | // vaddps (%rsi), %ymm0, %ymm1 | |||
26766 | // Requires just 1 allocation, i.e., freeing allocations for other operations | |||
26767 | // and having less micro operations to execute. | |||
26768 | // | |||
26769 | // For some X86 architectures, this is even worse because for instance for | |||
26770 | // stores, the complex addressing mode forces the instruction to use the | |||
26771 | // "load" ports instead of the dedicated "store" port. | |||
26772 | // E.g., on Haswell: | |||
26773 | // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3. | |||
26774 | // vmovaps %ymm1, (%r8) can use port 2, 3, or 7. | |||
26775 | if (isLegalAddressingMode(AM, Ty)) | |||
26776 | // Scale represents reg2 * scale, thus account for 1 | |||
26777 | // as soon as we use a second register. | |||
26778 | return AM.Scale != 0; | |||
26779 | return -1; | |||
26780 | } | |||
26781 | ||||
26782 | bool X86TargetLowering::isTargetFTOL() const { | |||
26783 | return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit(); | |||
26784 | } |