File: | llvm/lib/Target/PowerPC/PPCISelLowering.cpp |
Warning: | line 9266, column 36 Although the value stored to 'SplatBits' is used in the enclosing expression, the value is never actually read from 'SplatBits' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===-- PPCISelLowering.cpp - PPC DAG Lowering Implementation -------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file implements the PPCISelLowering class. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "PPCISelLowering.h" |
14 | #include "MCTargetDesc/PPCPredicates.h" |
15 | #include "PPC.h" |
16 | #include "PPCCCState.h" |
17 | #include "PPCCallingConv.h" |
18 | #include "PPCFrameLowering.h" |
19 | #include "PPCInstrInfo.h" |
20 | #include "PPCMachineFunctionInfo.h" |
21 | #include "PPCPerfectShuffle.h" |
22 | #include "PPCRegisterInfo.h" |
23 | #include "PPCSubtarget.h" |
24 | #include "PPCTargetMachine.h" |
25 | #include "llvm/ADT/APFloat.h" |
26 | #include "llvm/ADT/APInt.h" |
27 | #include "llvm/ADT/ArrayRef.h" |
28 | #include "llvm/ADT/DenseMap.h" |
29 | #include "llvm/ADT/None.h" |
30 | #include "llvm/ADT/STLExtras.h" |
31 | #include "llvm/ADT/SmallPtrSet.h" |
32 | #include "llvm/ADT/SmallSet.h" |
33 | #include "llvm/ADT/SmallVector.h" |
34 | #include "llvm/ADT/Statistic.h" |
35 | #include "llvm/ADT/StringRef.h" |
36 | #include "llvm/ADT/StringSwitch.h" |
37 | #include "llvm/CodeGen/CallingConvLower.h" |
38 | #include "llvm/CodeGen/ISDOpcodes.h" |
39 | #include "llvm/CodeGen/MachineBasicBlock.h" |
40 | #include "llvm/CodeGen/MachineFrameInfo.h" |
41 | #include "llvm/CodeGen/MachineFunction.h" |
42 | #include "llvm/CodeGen/MachineInstr.h" |
43 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
44 | #include "llvm/CodeGen/MachineJumpTableInfo.h" |
45 | #include "llvm/CodeGen/MachineLoopInfo.h" |
46 | #include "llvm/CodeGen/MachineMemOperand.h" |
47 | #include "llvm/CodeGen/MachineModuleInfo.h" |
48 | #include "llvm/CodeGen/MachineOperand.h" |
49 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
50 | #include "llvm/CodeGen/RuntimeLibcalls.h" |
51 | #include "llvm/CodeGen/SelectionDAG.h" |
52 | #include "llvm/CodeGen/SelectionDAGNodes.h" |
53 | #include "llvm/CodeGen/TargetInstrInfo.h" |
54 | #include "llvm/CodeGen/TargetLowering.h" |
55 | #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h" |
56 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
57 | #include "llvm/CodeGen/ValueTypes.h" |
58 | #include "llvm/IR/CallingConv.h" |
59 | #include "llvm/IR/Constant.h" |
60 | #include "llvm/IR/Constants.h" |
61 | #include "llvm/IR/DataLayout.h" |
62 | #include "llvm/IR/DebugLoc.h" |
63 | #include "llvm/IR/DerivedTypes.h" |
64 | #include "llvm/IR/Function.h" |
65 | #include "llvm/IR/GlobalValue.h" |
66 | #include "llvm/IR/IRBuilder.h" |
67 | #include "llvm/IR/Instructions.h" |
68 | #include "llvm/IR/Intrinsics.h" |
69 | #include "llvm/IR/IntrinsicsPowerPC.h" |
70 | #include "llvm/IR/Module.h" |
71 | #include "llvm/IR/Type.h" |
72 | #include "llvm/IR/Use.h" |
73 | #include "llvm/IR/Value.h" |
74 | #include "llvm/MC/MCContext.h" |
75 | #include "llvm/MC/MCExpr.h" |
76 | #include "llvm/MC/MCRegisterInfo.h" |
77 | #include "llvm/MC/MCSectionXCOFF.h" |
78 | #include "llvm/MC/MCSymbolXCOFF.h" |
79 | #include "llvm/Support/AtomicOrdering.h" |
80 | #include "llvm/Support/BranchProbability.h" |
81 | #include "llvm/Support/Casting.h" |
82 | #include "llvm/Support/CodeGen.h" |
83 | #include "llvm/Support/CommandLine.h" |
84 | #include "llvm/Support/Compiler.h" |
85 | #include "llvm/Support/Debug.h" |
86 | #include "llvm/Support/ErrorHandling.h" |
87 | #include "llvm/Support/Format.h" |
88 | #include "llvm/Support/KnownBits.h" |
89 | #include "llvm/Support/MachineValueType.h" |
90 | #include "llvm/Support/MathExtras.h" |
91 | #include "llvm/Support/raw_ostream.h" |
92 | #include "llvm/Target/TargetMachine.h" |
93 | #include "llvm/Target/TargetOptions.h" |
94 | #include <algorithm> |
95 | #include <cassert> |
96 | #include <cstdint> |
97 | #include <iterator> |
98 | #include <list> |
99 | #include <utility> |
100 | #include <vector> |
101 | |
102 | using namespace llvm; |
103 | |
104 | #define DEBUG_TYPE"ppc-lowering" "ppc-lowering" |
105 | |
106 | static cl::opt<bool> DisablePPCPreinc("disable-ppc-preinc", |
107 | cl::desc("disable preincrement load/store generation on PPC"), cl::Hidden); |
108 | |
109 | static cl::opt<bool> DisableILPPref("disable-ppc-ilp-pref", |
110 | cl::desc("disable setting the node scheduling preference to ILP on PPC"), cl::Hidden); |
111 | |
112 | static cl::opt<bool> DisablePPCUnaligned("disable-ppc-unaligned", |
113 | cl::desc("disable unaligned load/store generation on PPC"), cl::Hidden); |
114 | |
115 | static cl::opt<bool> DisableSCO("disable-ppc-sco", |
116 | cl::desc("disable sibling call optimization on ppc"), cl::Hidden); |
117 | |
118 | static cl::opt<bool> DisableInnermostLoopAlign32("disable-ppc-innermost-loop-align32", |
119 | cl::desc("don't always align innermost loop to 32 bytes on ppc"), cl::Hidden); |
120 | |
121 | static cl::opt<bool> UseAbsoluteJumpTables("ppc-use-absolute-jumptables", |
122 | cl::desc("use absolute jump tables on ppc"), cl::Hidden); |
123 | |
124 | STATISTIC(NumTailCalls, "Number of tail calls")static llvm::Statistic NumTailCalls = {"ppc-lowering", "NumTailCalls" , "Number of tail calls"}; |
125 | STATISTIC(NumSiblingCalls, "Number of sibling calls")static llvm::Statistic NumSiblingCalls = {"ppc-lowering", "NumSiblingCalls" , "Number of sibling calls"}; |
126 | STATISTIC(ShufflesHandledWithVPERM, "Number of shuffles lowered to a VPERM")static llvm::Statistic ShufflesHandledWithVPERM = {"ppc-lowering" , "ShufflesHandledWithVPERM", "Number of shuffles lowered to a VPERM" }; |
127 | STATISTIC(NumDynamicAllocaProbed, "Number of dynamic stack allocation probed")static llvm::Statistic NumDynamicAllocaProbed = {"ppc-lowering" , "NumDynamicAllocaProbed", "Number of dynamic stack allocation probed" }; |
128 | |
129 | static bool isNByteElemShuffleMask(ShuffleVectorSDNode *, unsigned, int); |
130 | |
131 | static SDValue widenVec(SelectionDAG &DAG, SDValue Vec, const SDLoc &dl); |
132 | |
133 | // FIXME: Remove this once the bug has been fixed! |
134 | extern cl::opt<bool> ANDIGlueBug; |
135 | |
136 | PPCTargetLowering::PPCTargetLowering(const PPCTargetMachine &TM, |
137 | const PPCSubtarget &STI) |
138 | : TargetLowering(TM), Subtarget(STI) { |
139 | // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all |
140 | // arguments are at least 4/8 bytes aligned. |
141 | bool isPPC64 = Subtarget.isPPC64(); |
142 | setMinStackArgumentAlignment(isPPC64 ? Align(8) : Align(4)); |
143 | |
144 | // Set up the register classes. |
145 | addRegisterClass(MVT::i32, &PPC::GPRCRegClass); |
146 | if (!useSoftFloat()) { |
147 | if (hasSPE()) { |
148 | addRegisterClass(MVT::f32, &PPC::GPRCRegClass); |
149 | addRegisterClass(MVT::f64, &PPC::SPERCRegClass); |
150 | } else { |
151 | addRegisterClass(MVT::f32, &PPC::F4RCRegClass); |
152 | addRegisterClass(MVT::f64, &PPC::F8RCRegClass); |
153 | } |
154 | } |
155 | |
156 | // Match BITREVERSE to customized fast code sequence in the td file. |
157 | setOperationAction(ISD::BITREVERSE, MVT::i32, Legal); |
158 | setOperationAction(ISD::BITREVERSE, MVT::i64, Legal); |
159 | |
160 | // Sub-word ATOMIC_CMP_SWAP need to ensure that the input is zero-extended. |
161 | setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom); |
162 | |
163 | // PowerPC has an i16 but no i8 (or i1) SEXTLOAD. |
164 | for (MVT VT : MVT::integer_valuetypes()) { |
165 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote); |
166 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Expand); |
167 | } |
168 | |
169 | if (Subtarget.isISA3_0()) { |
170 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Legal); |
171 | setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Legal); |
172 | setTruncStoreAction(MVT::f64, MVT::f16, Legal); |
173 | setTruncStoreAction(MVT::f32, MVT::f16, Legal); |
174 | } else { |
175 | // No extending loads from f16 or HW conversions back and forth. |
176 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand); |
177 | setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand); |
178 | setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand); |
179 | setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand); |
180 | setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand); |
181 | setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand); |
182 | setTruncStoreAction(MVT::f64, MVT::f16, Expand); |
183 | setTruncStoreAction(MVT::f32, MVT::f16, Expand); |
184 | } |
185 | |
186 | setTruncStoreAction(MVT::f64, MVT::f32, Expand); |
187 | |
188 | // PowerPC has pre-inc load and store's. |
189 | setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal); |
190 | setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal); |
191 | setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal); |
192 | setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal); |
193 | setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal); |
194 | setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal); |
195 | setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal); |
196 | setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal); |
197 | setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal); |
198 | setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal); |
199 | if (!Subtarget.hasSPE()) { |
200 | setIndexedLoadAction(ISD::PRE_INC, MVT::f32, Legal); |
201 | setIndexedLoadAction(ISD::PRE_INC, MVT::f64, Legal); |
202 | setIndexedStoreAction(ISD::PRE_INC, MVT::f32, Legal); |
203 | setIndexedStoreAction(ISD::PRE_INC, MVT::f64, Legal); |
204 | } |
205 | |
206 | // PowerPC uses ADDC/ADDE/SUBC/SUBE to propagate carry. |
207 | const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 }; |
208 | for (MVT VT : ScalarIntVTs) { |
209 | setOperationAction(ISD::ADDC, VT, Legal); |
210 | setOperationAction(ISD::ADDE, VT, Legal); |
211 | setOperationAction(ISD::SUBC, VT, Legal); |
212 | setOperationAction(ISD::SUBE, VT, Legal); |
213 | } |
214 | |
215 | if (Subtarget.useCRBits()) { |
216 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); |
217 | |
218 | if (isPPC64 || Subtarget.hasFPCVT()) { |
219 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i1, Promote); |
220 | AddPromotedToType(ISD::STRICT_SINT_TO_FP, MVT::i1, |
221 | isPPC64 ? MVT::i64 : MVT::i32); |
222 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i1, Promote); |
223 | AddPromotedToType(ISD::STRICT_UINT_TO_FP, MVT::i1, |
224 | isPPC64 ? MVT::i64 : MVT::i32); |
225 | |
226 | setOperationAction(ISD::SINT_TO_FP, MVT::i1, Promote); |
227 | AddPromotedToType (ISD::SINT_TO_FP, MVT::i1, |
228 | isPPC64 ? MVT::i64 : MVT::i32); |
229 | setOperationAction(ISD::UINT_TO_FP, MVT::i1, Promote); |
230 | AddPromotedToType(ISD::UINT_TO_FP, MVT::i1, |
231 | isPPC64 ? MVT::i64 : MVT::i32); |
232 | } else { |
233 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i1, Custom); |
234 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i1, Custom); |
235 | setOperationAction(ISD::SINT_TO_FP, MVT::i1, Custom); |
236 | setOperationAction(ISD::UINT_TO_FP, MVT::i1, Custom); |
237 | } |
238 | |
239 | // PowerPC does not support direct load/store of condition registers. |
240 | setOperationAction(ISD::LOAD, MVT::i1, Custom); |
241 | setOperationAction(ISD::STORE, MVT::i1, Custom); |
242 | |
243 | // FIXME: Remove this once the ANDI glue bug is fixed: |
244 | if (ANDIGlueBug) |
245 | setOperationAction(ISD::TRUNCATE, MVT::i1, Custom); |
246 | |
247 | for (MVT VT : MVT::integer_valuetypes()) { |
248 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote); |
249 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote); |
250 | setTruncStoreAction(VT, MVT::i1, Expand); |
251 | } |
252 | |
253 | addRegisterClass(MVT::i1, &PPC::CRBITRCRegClass); |
254 | } |
255 | |
256 | // Expand ppcf128 to i32 by hand for the benefit of llvm-gcc bootstrap on |
257 | // PPC (the libcall is not available). |
258 | setOperationAction(ISD::FP_TO_SINT, MVT::ppcf128, Custom); |
259 | setOperationAction(ISD::FP_TO_UINT, MVT::ppcf128, Custom); |
260 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::ppcf128, Custom); |
261 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::ppcf128, Custom); |
262 | |
263 | // We do not currently implement these libm ops for PowerPC. |
264 | setOperationAction(ISD::FFLOOR, MVT::ppcf128, Expand); |
265 | setOperationAction(ISD::FCEIL, MVT::ppcf128, Expand); |
266 | setOperationAction(ISD::FTRUNC, MVT::ppcf128, Expand); |
267 | setOperationAction(ISD::FRINT, MVT::ppcf128, Expand); |
268 | setOperationAction(ISD::FNEARBYINT, MVT::ppcf128, Expand); |
269 | setOperationAction(ISD::FREM, MVT::ppcf128, Expand); |
270 | |
271 | // PowerPC has no SREM/UREM instructions unless we are on P9 |
272 | // On P9 we may use a hardware instruction to compute the remainder. |
273 | // When the result of both the remainder and the division is required it is |
274 | // more efficient to compute the remainder from the result of the division |
275 | // rather than use the remainder instruction. The instructions are legalized |
276 | // directly because the DivRemPairsPass performs the transformation at the IR |
277 | // level. |
278 | if (Subtarget.isISA3_0()) { |
279 | setOperationAction(ISD::SREM, MVT::i32, Legal); |
280 | setOperationAction(ISD::UREM, MVT::i32, Legal); |
281 | setOperationAction(ISD::SREM, MVT::i64, Legal); |
282 | setOperationAction(ISD::UREM, MVT::i64, Legal); |
283 | } else { |
284 | setOperationAction(ISD::SREM, MVT::i32, Expand); |
285 | setOperationAction(ISD::UREM, MVT::i32, Expand); |
286 | setOperationAction(ISD::SREM, MVT::i64, Expand); |
287 | setOperationAction(ISD::UREM, MVT::i64, Expand); |
288 | } |
289 | |
290 | // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM. |
291 | setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand); |
292 | setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand); |
293 | setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand); |
294 | setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand); |
295 | setOperationAction(ISD::UDIVREM, MVT::i32, Expand); |
296 | setOperationAction(ISD::SDIVREM, MVT::i32, Expand); |
297 | setOperationAction(ISD::UDIVREM, MVT::i64, Expand); |
298 | setOperationAction(ISD::SDIVREM, MVT::i64, Expand); |
299 | |
300 | // Handle constrained floating-point operations of scalar. |
301 | // TODO: Handle SPE specific operation. |
302 | setOperationAction(ISD::STRICT_FADD, MVT::f32, Legal); |
303 | setOperationAction(ISD::STRICT_FSUB, MVT::f32, Legal); |
304 | setOperationAction(ISD::STRICT_FMUL, MVT::f32, Legal); |
305 | setOperationAction(ISD::STRICT_FDIV, MVT::f32, Legal); |
306 | setOperationAction(ISD::STRICT_FMA, MVT::f32, Legal); |
307 | setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Legal); |
308 | |
309 | setOperationAction(ISD::STRICT_FADD, MVT::f64, Legal); |
310 | setOperationAction(ISD::STRICT_FSUB, MVT::f64, Legal); |
311 | setOperationAction(ISD::STRICT_FMUL, MVT::f64, Legal); |
312 | setOperationAction(ISD::STRICT_FDIV, MVT::f64, Legal); |
313 | setOperationAction(ISD::STRICT_FMA, MVT::f64, Legal); |
314 | if (Subtarget.hasVSX()) { |
315 | setOperationAction(ISD::STRICT_FRINT, MVT::f32, Legal); |
316 | setOperationAction(ISD::STRICT_FRINT, MVT::f64, Legal); |
317 | } |
318 | |
319 | if (Subtarget.hasFSQRT()) { |
320 | setOperationAction(ISD::STRICT_FSQRT, MVT::f32, Legal); |
321 | setOperationAction(ISD::STRICT_FSQRT, MVT::f64, Legal); |
322 | } |
323 | |
324 | if (Subtarget.hasFPRND()) { |
325 | setOperationAction(ISD::STRICT_FFLOOR, MVT::f32, Legal); |
326 | setOperationAction(ISD::STRICT_FCEIL, MVT::f32, Legal); |
327 | setOperationAction(ISD::STRICT_FTRUNC, MVT::f32, Legal); |
328 | setOperationAction(ISD::STRICT_FROUND, MVT::f32, Legal); |
329 | |
330 | setOperationAction(ISD::STRICT_FFLOOR, MVT::f64, Legal); |
331 | setOperationAction(ISD::STRICT_FCEIL, MVT::f64, Legal); |
332 | setOperationAction(ISD::STRICT_FTRUNC, MVT::f64, Legal); |
333 | setOperationAction(ISD::STRICT_FROUND, MVT::f64, Legal); |
334 | } |
335 | |
336 | // We don't support sin/cos/sqrt/fmod/pow |
337 | setOperationAction(ISD::FSIN , MVT::f64, Expand); |
338 | setOperationAction(ISD::FCOS , MVT::f64, Expand); |
339 | setOperationAction(ISD::FSINCOS, MVT::f64, Expand); |
340 | setOperationAction(ISD::FREM , MVT::f64, Expand); |
341 | setOperationAction(ISD::FPOW , MVT::f64, Expand); |
342 | setOperationAction(ISD::FSIN , MVT::f32, Expand); |
343 | setOperationAction(ISD::FCOS , MVT::f32, Expand); |
344 | setOperationAction(ISD::FSINCOS, MVT::f32, Expand); |
345 | setOperationAction(ISD::FREM , MVT::f32, Expand); |
346 | setOperationAction(ISD::FPOW , MVT::f32, Expand); |
347 | if (Subtarget.hasSPE()) { |
348 | setOperationAction(ISD::FMA , MVT::f64, Expand); |
349 | setOperationAction(ISD::FMA , MVT::f32, Expand); |
350 | } else { |
351 | setOperationAction(ISD::FMA , MVT::f64, Legal); |
352 | setOperationAction(ISD::FMA , MVT::f32, Legal); |
353 | } |
354 | |
355 | if (Subtarget.hasSPE()) |
356 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand); |
357 | |
358 | setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom); |
359 | |
360 | // If we're enabling GP optimizations, use hardware square root |
361 | if (!Subtarget.hasFSQRT() && |
362 | !(TM.Options.UnsafeFPMath && Subtarget.hasFRSQRTE() && |
363 | Subtarget.hasFRE())) |
364 | setOperationAction(ISD::FSQRT, MVT::f64, Expand); |
365 | |
366 | if (!Subtarget.hasFSQRT() && |
367 | !(TM.Options.UnsafeFPMath && Subtarget.hasFRSQRTES() && |
368 | Subtarget.hasFRES())) |
369 | setOperationAction(ISD::FSQRT, MVT::f32, Expand); |
370 | |
371 | if (Subtarget.hasFCPSGN()) { |
372 | setOperationAction(ISD::FCOPYSIGN, MVT::f64, Legal); |
373 | setOperationAction(ISD::FCOPYSIGN, MVT::f32, Legal); |
374 | } else { |
375 | setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand); |
376 | setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand); |
377 | } |
378 | |
379 | if (Subtarget.hasFPRND()) { |
380 | setOperationAction(ISD::FFLOOR, MVT::f64, Legal); |
381 | setOperationAction(ISD::FCEIL, MVT::f64, Legal); |
382 | setOperationAction(ISD::FTRUNC, MVT::f64, Legal); |
383 | setOperationAction(ISD::FROUND, MVT::f64, Legal); |
384 | |
385 | setOperationAction(ISD::FFLOOR, MVT::f32, Legal); |
386 | setOperationAction(ISD::FCEIL, MVT::f32, Legal); |
387 | setOperationAction(ISD::FTRUNC, MVT::f32, Legal); |
388 | setOperationAction(ISD::FROUND, MVT::f32, Legal); |
389 | } |
390 | |
391 | // PowerPC does not have BSWAP, but we can use vector BSWAP instruction xxbrd |
392 | // to speed up scalar BSWAP64. |
393 | // CTPOP or CTTZ were introduced in P8/P9 respectively |
394 | setOperationAction(ISD::BSWAP, MVT::i32 , Expand); |
395 | if (Subtarget.hasP9Vector()) |
396 | setOperationAction(ISD::BSWAP, MVT::i64 , Custom); |
397 | else |
398 | setOperationAction(ISD::BSWAP, MVT::i64 , Expand); |
399 | if (Subtarget.isISA3_0()) { |
400 | setOperationAction(ISD::CTTZ , MVT::i32 , Legal); |
401 | setOperationAction(ISD::CTTZ , MVT::i64 , Legal); |
402 | } else { |
403 | setOperationAction(ISD::CTTZ , MVT::i32 , Expand); |
404 | setOperationAction(ISD::CTTZ , MVT::i64 , Expand); |
405 | } |
406 | |
407 | if (Subtarget.hasPOPCNTD() == PPCSubtarget::POPCNTD_Fast) { |
408 | setOperationAction(ISD::CTPOP, MVT::i32 , Legal); |
409 | setOperationAction(ISD::CTPOP, MVT::i64 , Legal); |
410 | } else { |
411 | setOperationAction(ISD::CTPOP, MVT::i32 , Expand); |
412 | setOperationAction(ISD::CTPOP, MVT::i64 , Expand); |
413 | } |
414 | |
415 | // PowerPC does not have ROTR |
416 | setOperationAction(ISD::ROTR, MVT::i32 , Expand); |
417 | setOperationAction(ISD::ROTR, MVT::i64 , Expand); |
418 | |
419 | if (!Subtarget.useCRBits()) { |
420 | // PowerPC does not have Select |
421 | setOperationAction(ISD::SELECT, MVT::i32, Expand); |
422 | setOperationAction(ISD::SELECT, MVT::i64, Expand); |
423 | setOperationAction(ISD::SELECT, MVT::f32, Expand); |
424 | setOperationAction(ISD::SELECT, MVT::f64, Expand); |
425 | } |
426 | |
427 | // PowerPC wants to turn select_cc of FP into fsel when possible. |
428 | setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); |
429 | setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); |
430 | |
431 | // PowerPC wants to optimize integer setcc a bit |
432 | if (!Subtarget.useCRBits()) |
433 | setOperationAction(ISD::SETCC, MVT::i32, Custom); |
434 | |
435 | if (Subtarget.hasFPU()) { |
436 | setOperationAction(ISD::STRICT_FSETCC, MVT::f32, Legal); |
437 | setOperationAction(ISD::STRICT_FSETCC, MVT::f64, Legal); |
438 | setOperationAction(ISD::STRICT_FSETCC, MVT::f128, Legal); |
439 | |
440 | setOperationAction(ISD::STRICT_FSETCCS, MVT::f32, Legal); |
441 | setOperationAction(ISD::STRICT_FSETCCS, MVT::f64, Legal); |
442 | setOperationAction(ISD::STRICT_FSETCCS, MVT::f128, Legal); |
443 | } |
444 | |
445 | // PowerPC does not have BRCOND which requires SetCC |
446 | if (!Subtarget.useCRBits()) |
447 | setOperationAction(ISD::BRCOND, MVT::Other, Expand); |
448 | |
449 | setOperationAction(ISD::BR_JT, MVT::Other, Expand); |
450 | |
451 | if (Subtarget.hasSPE()) { |
452 | // SPE has built-in conversions |
453 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Legal); |
454 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Legal); |
455 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i32, Legal); |
456 | setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal); |
457 | setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal); |
458 | setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal); |
459 | } else { |
460 | // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores. |
461 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Custom); |
462 | setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); |
463 | |
464 | // PowerPC does not have [U|S]INT_TO_FP |
465 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Expand); |
466 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i32, Expand); |
467 | setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand); |
468 | setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand); |
469 | } |
470 | |
471 | if (Subtarget.hasDirectMove() && isPPC64) { |
472 | setOperationAction(ISD::BITCAST, MVT::f32, Legal); |
473 | setOperationAction(ISD::BITCAST, MVT::i32, Legal); |
474 | setOperationAction(ISD::BITCAST, MVT::i64, Legal); |
475 | setOperationAction(ISD::BITCAST, MVT::f64, Legal); |
476 | if (TM.Options.UnsafeFPMath) { |
477 | setOperationAction(ISD::LRINT, MVT::f64, Legal); |
478 | setOperationAction(ISD::LRINT, MVT::f32, Legal); |
479 | setOperationAction(ISD::LLRINT, MVT::f64, Legal); |
480 | setOperationAction(ISD::LLRINT, MVT::f32, Legal); |
481 | setOperationAction(ISD::LROUND, MVT::f64, Legal); |
482 | setOperationAction(ISD::LROUND, MVT::f32, Legal); |
483 | setOperationAction(ISD::LLROUND, MVT::f64, Legal); |
484 | setOperationAction(ISD::LLROUND, MVT::f32, Legal); |
485 | } |
486 | } else { |
487 | setOperationAction(ISD::BITCAST, MVT::f32, Expand); |
488 | setOperationAction(ISD::BITCAST, MVT::i32, Expand); |
489 | setOperationAction(ISD::BITCAST, MVT::i64, Expand); |
490 | setOperationAction(ISD::BITCAST, MVT::f64, Expand); |
491 | } |
492 | |
493 | // We cannot sextinreg(i1). Expand to shifts. |
494 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); |
495 | |
496 | // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support |
497 | // SjLj exception handling but a light-weight setjmp/longjmp replacement to |
498 | // support continuation, user-level threading, and etc.. As a result, no |
499 | // other SjLj exception interfaces are implemented and please don't build |
500 | // your own exception handling based on them. |
501 | // LLVM/Clang supports zero-cost DWARF exception handling. |
502 | setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom); |
503 | setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom); |
504 | |
505 | // We want to legalize GlobalAddress and ConstantPool nodes into the |
506 | // appropriate instructions to materialize the address. |
507 | setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); |
508 | setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); |
509 | setOperationAction(ISD::BlockAddress, MVT::i32, Custom); |
510 | setOperationAction(ISD::ConstantPool, MVT::i32, Custom); |
511 | setOperationAction(ISD::JumpTable, MVT::i32, Custom); |
512 | setOperationAction(ISD::GlobalAddress, MVT::i64, Custom); |
513 | setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom); |
514 | setOperationAction(ISD::BlockAddress, MVT::i64, Custom); |
515 | setOperationAction(ISD::ConstantPool, MVT::i64, Custom); |
516 | setOperationAction(ISD::JumpTable, MVT::i64, Custom); |
517 | |
518 | // TRAP is legal. |
519 | setOperationAction(ISD::TRAP, MVT::Other, Legal); |
520 | |
521 | // TRAMPOLINE is custom lowered. |
522 | setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom); |
523 | setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom); |
524 | |
525 | // VASTART needs to be custom lowered to use the VarArgsFrameIndex |
526 | setOperationAction(ISD::VASTART , MVT::Other, Custom); |
527 | |
528 | if (Subtarget.is64BitELFABI()) { |
529 | // VAARG always uses double-word chunks, so promote anything smaller. |
530 | setOperationAction(ISD::VAARG, MVT::i1, Promote); |
531 | AddPromotedToType(ISD::VAARG, MVT::i1, MVT::i64); |
532 | setOperationAction(ISD::VAARG, MVT::i8, Promote); |
533 | AddPromotedToType(ISD::VAARG, MVT::i8, MVT::i64); |
534 | setOperationAction(ISD::VAARG, MVT::i16, Promote); |
535 | AddPromotedToType(ISD::VAARG, MVT::i16, MVT::i64); |
536 | setOperationAction(ISD::VAARG, MVT::i32, Promote); |
537 | AddPromotedToType(ISD::VAARG, MVT::i32, MVT::i64); |
538 | setOperationAction(ISD::VAARG, MVT::Other, Expand); |
539 | } else if (Subtarget.is32BitELFABI()) { |
540 | // VAARG is custom lowered with the 32-bit SVR4 ABI. |
541 | setOperationAction(ISD::VAARG, MVT::Other, Custom); |
542 | setOperationAction(ISD::VAARG, MVT::i64, Custom); |
543 | } else |
544 | setOperationAction(ISD::VAARG, MVT::Other, Expand); |
545 | |
546 | // VACOPY is custom lowered with the 32-bit SVR4 ABI. |
547 | if (Subtarget.is32BitELFABI()) |
548 | setOperationAction(ISD::VACOPY , MVT::Other, Custom); |
549 | else |
550 | setOperationAction(ISD::VACOPY , MVT::Other, Expand); |
551 | |
552 | // Use the default implementation. |
553 | setOperationAction(ISD::VAEND , MVT::Other, Expand); |
554 | setOperationAction(ISD::STACKSAVE , MVT::Other, Expand); |
555 | setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom); |
556 | setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom); |
557 | setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom); |
558 | setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, MVT::i32, Custom); |
559 | setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, MVT::i64, Custom); |
560 | setOperationAction(ISD::EH_DWARF_CFA, MVT::i32, Custom); |
561 | setOperationAction(ISD::EH_DWARF_CFA, MVT::i64, Custom); |
562 | |
563 | // We want to custom lower some of our intrinsics. |
564 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); |
565 | |
566 | // To handle counter-based loop conditions. |
567 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i1, Custom); |
568 | |
569 | setOperationAction(ISD::INTRINSIC_VOID, MVT::i8, Custom); |
570 | setOperationAction(ISD::INTRINSIC_VOID, MVT::i16, Custom); |
571 | setOperationAction(ISD::INTRINSIC_VOID, MVT::i32, Custom); |
572 | setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom); |
573 | |
574 | // Comparisons that require checking two conditions. |
575 | if (Subtarget.hasSPE()) { |
576 | setCondCodeAction(ISD::SETO, MVT::f32, Expand); |
577 | setCondCodeAction(ISD::SETO, MVT::f64, Expand); |
578 | setCondCodeAction(ISD::SETUO, MVT::f32, Expand); |
579 | setCondCodeAction(ISD::SETUO, MVT::f64, Expand); |
580 | } |
581 | setCondCodeAction(ISD::SETULT, MVT::f32, Expand); |
582 | setCondCodeAction(ISD::SETULT, MVT::f64, Expand); |
583 | setCondCodeAction(ISD::SETUGT, MVT::f32, Expand); |
584 | setCondCodeAction(ISD::SETUGT, MVT::f64, Expand); |
585 | setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand); |
586 | setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand); |
587 | setCondCodeAction(ISD::SETOGE, MVT::f32, Expand); |
588 | setCondCodeAction(ISD::SETOGE, MVT::f64, Expand); |
589 | setCondCodeAction(ISD::SETOLE, MVT::f32, Expand); |
590 | setCondCodeAction(ISD::SETOLE, MVT::f64, Expand); |
591 | setCondCodeAction(ISD::SETONE, MVT::f32, Expand); |
592 | setCondCodeAction(ISD::SETONE, MVT::f64, Expand); |
593 | |
594 | if (Subtarget.has64BitSupport()) { |
595 | // They also have instructions for converting between i64 and fp. |
596 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i64, Custom); |
597 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i64, Expand); |
598 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i64, Custom); |
599 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i64, Expand); |
600 | setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom); |
601 | setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand); |
602 | setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom); |
603 | setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand); |
604 | // This is just the low 32 bits of a (signed) fp->i64 conversion. |
605 | // We cannot do this with Promote because i64 is not a legal type. |
606 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Custom); |
607 | setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom); |
608 | |
609 | if (Subtarget.hasLFIWAX() || Subtarget.isPPC64()) { |
610 | setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom); |
611 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Custom); |
612 | } |
613 | } else { |
614 | // PowerPC does not have FP_TO_UINT on 32-bit implementations. |
615 | if (Subtarget.hasSPE()) { |
616 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Legal); |
617 | setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal); |
618 | } else { |
619 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Expand); |
620 | setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand); |
621 | } |
622 | } |
623 | |
624 | // With the instructions enabled under FPCVT, we can do everything. |
625 | if (Subtarget.hasFPCVT()) { |
626 | if (Subtarget.has64BitSupport()) { |
627 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i64, Custom); |
628 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i64, Custom); |
629 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i64, Custom); |
630 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i64, Custom); |
631 | setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom); |
632 | setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom); |
633 | setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom); |
634 | setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom); |
635 | } |
636 | |
637 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Custom); |
638 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Custom); |
639 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Custom); |
640 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i32, Custom); |
641 | setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); |
642 | setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom); |
643 | setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom); |
644 | setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom); |
645 | } |
646 | |
647 | if (Subtarget.use64BitRegs()) { |
648 | // 64-bit PowerPC implementations can support i64 types directly |
649 | addRegisterClass(MVT::i64, &PPC::G8RCRegClass); |
650 | // BUILD_PAIR can't be handled natively, and should be expanded to shl/or |
651 | setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand); |
652 | // 64-bit PowerPC wants to expand i128 shifts itself. |
653 | setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom); |
654 | setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom); |
655 | setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom); |
656 | } else { |
657 | // 32-bit PowerPC wants to expand i64 shifts itself. |
658 | setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom); |
659 | setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom); |
660 | setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom); |
661 | } |
662 | |
663 | // PowerPC has better expansions for funnel shifts than the generic |
664 | // TargetLowering::expandFunnelShift. |
665 | if (Subtarget.has64BitSupport()) { |
666 | setOperationAction(ISD::FSHL, MVT::i64, Custom); |
667 | setOperationAction(ISD::FSHR, MVT::i64, Custom); |
668 | } |
669 | setOperationAction(ISD::FSHL, MVT::i32, Custom); |
670 | setOperationAction(ISD::FSHR, MVT::i32, Custom); |
671 | |
672 | if (Subtarget.hasVSX()) { |
673 | setOperationAction(ISD::FMAXNUM_IEEE, MVT::f64, Legal); |
674 | setOperationAction(ISD::FMAXNUM_IEEE, MVT::f32, Legal); |
675 | setOperationAction(ISD::FMINNUM_IEEE, MVT::f64, Legal); |
676 | setOperationAction(ISD::FMINNUM_IEEE, MVT::f32, Legal); |
677 | } |
678 | |
679 | if (Subtarget.hasAltivec()) { |
680 | for (MVT VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) { |
681 | setOperationAction(ISD::SADDSAT, VT, Legal); |
682 | setOperationAction(ISD::SSUBSAT, VT, Legal); |
683 | setOperationAction(ISD::UADDSAT, VT, Legal); |
684 | setOperationAction(ISD::USUBSAT, VT, Legal); |
685 | } |
686 | // First set operation action for all vector types to expand. Then we |
687 | // will selectively turn on ones that can be effectively codegen'd. |
688 | for (MVT VT : MVT::fixedlen_vector_valuetypes()) { |
689 | // add/sub are legal for all supported vector VT's. |
690 | setOperationAction(ISD::ADD, VT, Legal); |
691 | setOperationAction(ISD::SUB, VT, Legal); |
692 | |
693 | // For v2i64, these are only valid with P8Vector. This is corrected after |
694 | // the loop. |
695 | if (VT.getSizeInBits() <= 128 && VT.getScalarSizeInBits() <= 64) { |
696 | setOperationAction(ISD::SMAX, VT, Legal); |
697 | setOperationAction(ISD::SMIN, VT, Legal); |
698 | setOperationAction(ISD::UMAX, VT, Legal); |
699 | setOperationAction(ISD::UMIN, VT, Legal); |
700 | } |
701 | else { |
702 | setOperationAction(ISD::SMAX, VT, Expand); |
703 | setOperationAction(ISD::SMIN, VT, Expand); |
704 | setOperationAction(ISD::UMAX, VT, Expand); |
705 | setOperationAction(ISD::UMIN, VT, Expand); |
706 | } |
707 | |
708 | if (Subtarget.hasVSX()) { |
709 | setOperationAction(ISD::FMAXNUM, VT, Legal); |
710 | setOperationAction(ISD::FMINNUM, VT, Legal); |
711 | } |
712 | |
713 | // Vector instructions introduced in P8 |
714 | if (Subtarget.hasP8Altivec() && (VT.SimpleTy != MVT::v1i128)) { |
715 | setOperationAction(ISD::CTPOP, VT, Legal); |
716 | setOperationAction(ISD::CTLZ, VT, Legal); |
717 | } |
718 | else { |
719 | setOperationAction(ISD::CTPOP, VT, Expand); |
720 | setOperationAction(ISD::CTLZ, VT, Expand); |
721 | } |
722 | |
723 | // Vector instructions introduced in P9 |
724 | if (Subtarget.hasP9Altivec() && (VT.SimpleTy != MVT::v1i128)) |
725 | setOperationAction(ISD::CTTZ, VT, Legal); |
726 | else |
727 | setOperationAction(ISD::CTTZ, VT, Expand); |
728 | |
729 | // We promote all shuffles to v16i8. |
730 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote); |
731 | AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8); |
732 | |
733 | // We promote all non-typed operations to v4i32. |
734 | setOperationAction(ISD::AND , VT, Promote); |
735 | AddPromotedToType (ISD::AND , VT, MVT::v4i32); |
736 | setOperationAction(ISD::OR , VT, Promote); |
737 | AddPromotedToType (ISD::OR , VT, MVT::v4i32); |
738 | setOperationAction(ISD::XOR , VT, Promote); |
739 | AddPromotedToType (ISD::XOR , VT, MVT::v4i32); |
740 | setOperationAction(ISD::LOAD , VT, Promote); |
741 | AddPromotedToType (ISD::LOAD , VT, MVT::v4i32); |
742 | setOperationAction(ISD::SELECT, VT, Promote); |
743 | AddPromotedToType (ISD::SELECT, VT, MVT::v4i32); |
744 | setOperationAction(ISD::VSELECT, VT, Legal); |
745 | setOperationAction(ISD::SELECT_CC, VT, Promote); |
746 | AddPromotedToType (ISD::SELECT_CC, VT, MVT::v4i32); |
747 | setOperationAction(ISD::STORE, VT, Promote); |
748 | AddPromotedToType (ISD::STORE, VT, MVT::v4i32); |
749 | |
750 | // No other operations are legal. |
751 | setOperationAction(ISD::MUL , VT, Expand); |
752 | setOperationAction(ISD::SDIV, VT, Expand); |
753 | setOperationAction(ISD::SREM, VT, Expand); |
754 | setOperationAction(ISD::UDIV, VT, Expand); |
755 | setOperationAction(ISD::UREM, VT, Expand); |
756 | setOperationAction(ISD::FDIV, VT, Expand); |
757 | setOperationAction(ISD::FREM, VT, Expand); |
758 | setOperationAction(ISD::FNEG, VT, Expand); |
759 | setOperationAction(ISD::FSQRT, VT, Expand); |
760 | setOperationAction(ISD::FLOG, VT, Expand); |
761 | setOperationAction(ISD::FLOG10, VT, Expand); |
762 | setOperationAction(ISD::FLOG2, VT, Expand); |
763 | setOperationAction(ISD::FEXP, VT, Expand); |
764 | setOperationAction(ISD::FEXP2, VT, Expand); |
765 | setOperationAction(ISD::FSIN, VT, Expand); |
766 | setOperationAction(ISD::FCOS, VT, Expand); |
767 | setOperationAction(ISD::FABS, VT, Expand); |
768 | setOperationAction(ISD::FFLOOR, VT, Expand); |
769 | setOperationAction(ISD::FCEIL, VT, Expand); |
770 | setOperationAction(ISD::FTRUNC, VT, Expand); |
771 | setOperationAction(ISD::FRINT, VT, Expand); |
772 | setOperationAction(ISD::FNEARBYINT, VT, Expand); |
773 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand); |
774 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand); |
775 | setOperationAction(ISD::BUILD_VECTOR, VT, Expand); |
776 | setOperationAction(ISD::MULHU, VT, Expand); |
777 | setOperationAction(ISD::MULHS, VT, Expand); |
778 | setOperationAction(ISD::UMUL_LOHI, VT, Expand); |
779 | setOperationAction(ISD::SMUL_LOHI, VT, Expand); |
780 | setOperationAction(ISD::UDIVREM, VT, Expand); |
781 | setOperationAction(ISD::SDIVREM, VT, Expand); |
782 | setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand); |
783 | setOperationAction(ISD::FPOW, VT, Expand); |
784 | setOperationAction(ISD::BSWAP, VT, Expand); |
785 | setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand); |
786 | setOperationAction(ISD::ROTL, VT, Expand); |
787 | setOperationAction(ISD::ROTR, VT, Expand); |
788 | |
789 | for (MVT InnerVT : MVT::fixedlen_vector_valuetypes()) { |
790 | setTruncStoreAction(VT, InnerVT, Expand); |
791 | setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand); |
792 | setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand); |
793 | setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand); |
794 | } |
795 | } |
796 | setOperationAction(ISD::SELECT_CC, MVT::v4i32, Expand); |
797 | if (!Subtarget.hasP8Vector()) { |
798 | setOperationAction(ISD::SMAX, MVT::v2i64, Expand); |
799 | setOperationAction(ISD::SMIN, MVT::v2i64, Expand); |
800 | setOperationAction(ISD::UMAX, MVT::v2i64, Expand); |
801 | setOperationAction(ISD::UMIN, MVT::v2i64, Expand); |
802 | } |
803 | |
804 | for (auto VT : {MVT::v2i64, MVT::v4i32, MVT::v8i16, MVT::v16i8}) |
805 | setOperationAction(ISD::ABS, VT, Custom); |
806 | |
807 | // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle |
808 | // with merges, splats, etc. |
809 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom); |
810 | |
811 | // Vector truncates to sub-word integer that fit in an Altivec/VSX register |
812 | // are cheap, so handle them before they get expanded to scalar. |
813 | setOperationAction(ISD::TRUNCATE, MVT::v8i8, Custom); |
814 | setOperationAction(ISD::TRUNCATE, MVT::v4i8, Custom); |
815 | setOperationAction(ISD::TRUNCATE, MVT::v2i8, Custom); |
816 | setOperationAction(ISD::TRUNCATE, MVT::v4i16, Custom); |
817 | setOperationAction(ISD::TRUNCATE, MVT::v2i16, Custom); |
818 | |
819 | setOperationAction(ISD::AND , MVT::v4i32, Legal); |
820 | setOperationAction(ISD::OR , MVT::v4i32, Legal); |
821 | setOperationAction(ISD::XOR , MVT::v4i32, Legal); |
822 | setOperationAction(ISD::LOAD , MVT::v4i32, Legal); |
823 | setOperationAction(ISD::SELECT, MVT::v4i32, |
824 | Subtarget.useCRBits() ? Legal : Expand); |
825 | setOperationAction(ISD::STORE , MVT::v4i32, Legal); |
826 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v4i32, Legal); |
827 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v4i32, Legal); |
828 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v4i32, Legal); |
829 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4i32, Legal); |
830 | setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal); |
831 | setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal); |
832 | setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal); |
833 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal); |
834 | setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal); |
835 | setOperationAction(ISD::FCEIL, MVT::v4f32, Legal); |
836 | setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal); |
837 | setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal); |
838 | |
839 | // Without hasP8Altivec set, v2i64 SMAX isn't available. |
840 | // But ABS custom lowering requires SMAX support. |
841 | if (!Subtarget.hasP8Altivec()) |
842 | setOperationAction(ISD::ABS, MVT::v2i64, Expand); |
843 | |
844 | // Custom lowering ROTL v1i128 to VECTOR_SHUFFLE v16i8. |
845 | setOperationAction(ISD::ROTL, MVT::v1i128, Custom); |
846 | // With hasAltivec set, we can lower ISD::ROTL to vrl(b|h|w). |
847 | if (Subtarget.hasAltivec()) |
848 | for (auto VT : {MVT::v4i32, MVT::v8i16, MVT::v16i8}) |
849 | setOperationAction(ISD::ROTL, VT, Legal); |
850 | // With hasP8Altivec set, we can lower ISD::ROTL to vrld. |
851 | if (Subtarget.hasP8Altivec()) |
852 | setOperationAction(ISD::ROTL, MVT::v2i64, Legal); |
853 | |
854 | addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass); |
855 | addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass); |
856 | addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass); |
857 | addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass); |
858 | |
859 | setOperationAction(ISD::MUL, MVT::v4f32, Legal); |
860 | setOperationAction(ISD::FMA, MVT::v4f32, Legal); |
861 | |
862 | if (Subtarget.hasVSX()) { |
863 | setOperationAction(ISD::FDIV, MVT::v4f32, Legal); |
864 | setOperationAction(ISD::FSQRT, MVT::v4f32, Legal); |
865 | } |
866 | |
867 | if (Subtarget.hasP8Altivec()) |
868 | setOperationAction(ISD::MUL, MVT::v4i32, Legal); |
869 | else |
870 | setOperationAction(ISD::MUL, MVT::v4i32, Custom); |
871 | |
872 | if (Subtarget.isISA3_1()) { |
873 | setOperationAction(ISD::MUL, MVT::v2i64, Legal); |
874 | setOperationAction(ISD::MULHS, MVT::v2i64, Legal); |
875 | setOperationAction(ISD::MULHU, MVT::v2i64, Legal); |
876 | setOperationAction(ISD::MULHS, MVT::v4i32, Legal); |
877 | setOperationAction(ISD::MULHU, MVT::v4i32, Legal); |
878 | setOperationAction(ISD::UDIV, MVT::v2i64, Legal); |
879 | setOperationAction(ISD::SDIV, MVT::v2i64, Legal); |
880 | setOperationAction(ISD::UDIV, MVT::v4i32, Legal); |
881 | setOperationAction(ISD::SDIV, MVT::v4i32, Legal); |
882 | setOperationAction(ISD::UREM, MVT::v2i64, Legal); |
883 | setOperationAction(ISD::SREM, MVT::v2i64, Legal); |
884 | setOperationAction(ISD::UREM, MVT::v4i32, Legal); |
885 | setOperationAction(ISD::SREM, MVT::v4i32, Legal); |
886 | setOperationAction(ISD::UREM, MVT::v1i128, Legal); |
887 | setOperationAction(ISD::SREM, MVT::v1i128, Legal); |
888 | setOperationAction(ISD::UDIV, MVT::v1i128, Legal); |
889 | setOperationAction(ISD::SDIV, MVT::v1i128, Legal); |
890 | setOperationAction(ISD::ROTL, MVT::v1i128, Legal); |
891 | } |
892 | |
893 | setOperationAction(ISD::MUL, MVT::v8i16, Legal); |
894 | setOperationAction(ISD::MUL, MVT::v16i8, Custom); |
895 | |
896 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom); |
897 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom); |
898 | |
899 | setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom); |
900 | setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom); |
901 | setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom); |
902 | setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom); |
903 | |
904 | // Altivec does not contain unordered floating-point compare instructions |
905 | setCondCodeAction(ISD::SETUO, MVT::v4f32, Expand); |
906 | setCondCodeAction(ISD::SETUEQ, MVT::v4f32, Expand); |
907 | setCondCodeAction(ISD::SETO, MVT::v4f32, Expand); |
908 | setCondCodeAction(ISD::SETONE, MVT::v4f32, Expand); |
909 | |
910 | if (Subtarget.hasVSX()) { |
911 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f64, Legal); |
912 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Legal); |
913 | if (Subtarget.hasP8Vector()) { |
914 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Legal); |
915 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Legal); |
916 | } |
917 | if (Subtarget.hasDirectMove() && isPPC64) { |
918 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Legal); |
919 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Legal); |
920 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Legal); |
921 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i64, Legal); |
922 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Legal); |
923 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Legal); |
924 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Legal); |
925 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal); |
926 | } |
927 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Legal); |
928 | |
929 | // The nearbyint variants are not allowed to raise the inexact exception |
930 | // so we can only code-gen them with unsafe math. |
931 | if (TM.Options.UnsafeFPMath) { |
932 | setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal); |
933 | setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal); |
934 | } |
935 | |
936 | setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal); |
937 | setOperationAction(ISD::FCEIL, MVT::v2f64, Legal); |
938 | setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal); |
939 | setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal); |
940 | setOperationAction(ISD::FRINT, MVT::v2f64, Legal); |
941 | setOperationAction(ISD::FROUND, MVT::v2f64, Legal); |
942 | setOperationAction(ISD::FROUND, MVT::f64, Legal); |
943 | setOperationAction(ISD::FRINT, MVT::f64, Legal); |
944 | |
945 | setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal); |
946 | setOperationAction(ISD::FRINT, MVT::v4f32, Legal); |
947 | setOperationAction(ISD::FROUND, MVT::v4f32, Legal); |
948 | setOperationAction(ISD::FROUND, MVT::f32, Legal); |
949 | setOperationAction(ISD::FRINT, MVT::f32, Legal); |
950 | |
951 | setOperationAction(ISD::MUL, MVT::v2f64, Legal); |
952 | setOperationAction(ISD::FMA, MVT::v2f64, Legal); |
953 | |
954 | setOperationAction(ISD::FDIV, MVT::v2f64, Legal); |
955 | setOperationAction(ISD::FSQRT, MVT::v2f64, Legal); |
956 | |
957 | // Share the Altivec comparison restrictions. |
958 | setCondCodeAction(ISD::SETUO, MVT::v2f64, Expand); |
959 | setCondCodeAction(ISD::SETUEQ, MVT::v2f64, Expand); |
960 | setCondCodeAction(ISD::SETO, MVT::v2f64, Expand); |
961 | setCondCodeAction(ISD::SETONE, MVT::v2f64, Expand); |
962 | |
963 | setOperationAction(ISD::LOAD, MVT::v2f64, Legal); |
964 | setOperationAction(ISD::STORE, MVT::v2f64, Legal); |
965 | |
966 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Legal); |
967 | |
968 | if (Subtarget.hasP8Vector()) |
969 | addRegisterClass(MVT::f32, &PPC::VSSRCRegClass); |
970 | |
971 | addRegisterClass(MVT::f64, &PPC::VSFRCRegClass); |
972 | |
973 | addRegisterClass(MVT::v4i32, &PPC::VSRCRegClass); |
974 | addRegisterClass(MVT::v4f32, &PPC::VSRCRegClass); |
975 | addRegisterClass(MVT::v2f64, &PPC::VSRCRegClass); |
976 | |
977 | if (Subtarget.hasP8Altivec()) { |
978 | setOperationAction(ISD::SHL, MVT::v2i64, Legal); |
979 | setOperationAction(ISD::SRA, MVT::v2i64, Legal); |
980 | setOperationAction(ISD::SRL, MVT::v2i64, Legal); |
981 | |
982 | // 128 bit shifts can be accomplished via 3 instructions for SHL and |
983 | // SRL, but not for SRA because of the instructions available: |
984 | // VS{RL} and VS{RL}O. However due to direct move costs, it's not worth |
985 | // doing |
986 | setOperationAction(ISD::SHL, MVT::v1i128, Expand); |
987 | setOperationAction(ISD::SRL, MVT::v1i128, Expand); |
988 | setOperationAction(ISD::SRA, MVT::v1i128, Expand); |
989 | |
990 | setOperationAction(ISD::SETCC, MVT::v2i64, Legal); |
991 | } |
992 | else { |
993 | setOperationAction(ISD::SHL, MVT::v2i64, Expand); |
994 | setOperationAction(ISD::SRA, MVT::v2i64, Expand); |
995 | setOperationAction(ISD::SRL, MVT::v2i64, Expand); |
996 | |
997 | setOperationAction(ISD::SETCC, MVT::v2i64, Custom); |
998 | |
999 | // VSX v2i64 only supports non-arithmetic operations. |
1000 | setOperationAction(ISD::ADD, MVT::v2i64, Expand); |
1001 | setOperationAction(ISD::SUB, MVT::v2i64, Expand); |
1002 | } |
1003 | |
1004 | if (Subtarget.isISA3_1()) |
1005 | setOperationAction(ISD::SETCC, MVT::v1i128, Legal); |
1006 | else |
1007 | setOperationAction(ISD::SETCC, MVT::v1i128, Expand); |
1008 | |
1009 | setOperationAction(ISD::LOAD, MVT::v2i64, Promote); |
1010 | AddPromotedToType (ISD::LOAD, MVT::v2i64, MVT::v2f64); |
1011 | setOperationAction(ISD::STORE, MVT::v2i64, Promote); |
1012 | AddPromotedToType (ISD::STORE, MVT::v2i64, MVT::v2f64); |
1013 | |
1014 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Legal); |
1015 | |
1016 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v2i64, Legal); |
1017 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v2i64, Legal); |
1018 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v2i64, Legal); |
1019 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v2i64, Legal); |
1020 | setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal); |
1021 | setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal); |
1022 | setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal); |
1023 | setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal); |
1024 | |
1025 | // Custom handling for partial vectors of integers converted to |
1026 | // floating point. We already have optimal handling for v2i32 through |
1027 | // the DAG combine, so those aren't necessary. |
1028 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v2i8, Custom); |
1029 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4i8, Custom); |
1030 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v2i16, Custom); |
1031 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4i16, Custom); |
1032 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v2i8, Custom); |
1033 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v4i8, Custom); |
1034 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v2i16, Custom); |
1035 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v4i16, Custom); |
1036 | setOperationAction(ISD::UINT_TO_FP, MVT::v2i8, Custom); |
1037 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom); |
1038 | setOperationAction(ISD::UINT_TO_FP, MVT::v2i16, Custom); |
1039 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom); |
1040 | setOperationAction(ISD::SINT_TO_FP, MVT::v2i8, Custom); |
1041 | setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Custom); |
1042 | setOperationAction(ISD::SINT_TO_FP, MVT::v2i16, Custom); |
1043 | setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom); |
1044 | |
1045 | setOperationAction(ISD::FNEG, MVT::v4f32, Legal); |
1046 | setOperationAction(ISD::FNEG, MVT::v2f64, Legal); |
1047 | setOperationAction(ISD::FABS, MVT::v4f32, Legal); |
1048 | setOperationAction(ISD::FABS, MVT::v2f64, Legal); |
1049 | setOperationAction(ISD::FCOPYSIGN, MVT::v4f32, Legal); |
1050 | setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Legal); |
1051 | |
1052 | if (Subtarget.hasDirectMove()) |
1053 | setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom); |
1054 | setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom); |
1055 | |
1056 | // Handle constrained floating-point operations of vector. |
1057 | // The predictor is `hasVSX` because altivec instruction has |
1058 | // no exception but VSX vector instruction has. |
1059 | setOperationAction(ISD::STRICT_FADD, MVT::v4f32, Legal); |
1060 | setOperationAction(ISD::STRICT_FSUB, MVT::v4f32, Legal); |
1061 | setOperationAction(ISD::STRICT_FMUL, MVT::v4f32, Legal); |
1062 | setOperationAction(ISD::STRICT_FDIV, MVT::v4f32, Legal); |
1063 | setOperationAction(ISD::STRICT_FMA, MVT::v4f32, Legal); |
1064 | setOperationAction(ISD::STRICT_FSQRT, MVT::v4f32, Legal); |
1065 | setOperationAction(ISD::STRICT_FMAXNUM, MVT::v4f32, Legal); |
1066 | setOperationAction(ISD::STRICT_FMINNUM, MVT::v4f32, Legal); |
1067 | setOperationAction(ISD::STRICT_FRINT, MVT::v4f32, Legal); |
1068 | setOperationAction(ISD::STRICT_FFLOOR, MVT::v4f32, Legal); |
1069 | setOperationAction(ISD::STRICT_FCEIL, MVT::v4f32, Legal); |
1070 | setOperationAction(ISD::STRICT_FTRUNC, MVT::v4f32, Legal); |
1071 | setOperationAction(ISD::STRICT_FROUND, MVT::v4f32, Legal); |
1072 | |
1073 | setOperationAction(ISD::STRICT_FADD, MVT::v2f64, Legal); |
1074 | setOperationAction(ISD::STRICT_FSUB, MVT::v2f64, Legal); |
1075 | setOperationAction(ISD::STRICT_FMUL, MVT::v2f64, Legal); |
1076 | setOperationAction(ISD::STRICT_FDIV, MVT::v2f64, Legal); |
1077 | setOperationAction(ISD::STRICT_FMA, MVT::v2f64, Legal); |
1078 | setOperationAction(ISD::STRICT_FSQRT, MVT::v2f64, Legal); |
1079 | setOperationAction(ISD::STRICT_FMAXNUM, MVT::v2f64, Legal); |
1080 | setOperationAction(ISD::STRICT_FMINNUM, MVT::v2f64, Legal); |
1081 | setOperationAction(ISD::STRICT_FRINT, MVT::v2f64, Legal); |
1082 | setOperationAction(ISD::STRICT_FFLOOR, MVT::v2f64, Legal); |
1083 | setOperationAction(ISD::STRICT_FCEIL, MVT::v2f64, Legal); |
1084 | setOperationAction(ISD::STRICT_FTRUNC, MVT::v2f64, Legal); |
1085 | setOperationAction(ISD::STRICT_FROUND, MVT::v2f64, Legal); |
1086 | |
1087 | addRegisterClass(MVT::v2i64, &PPC::VSRCRegClass); |
1088 | } |
1089 | |
1090 | if (Subtarget.hasP8Altivec()) { |
1091 | addRegisterClass(MVT::v2i64, &PPC::VRRCRegClass); |
1092 | addRegisterClass(MVT::v1i128, &PPC::VRRCRegClass); |
1093 | } |
1094 | |
1095 | if (Subtarget.hasP9Vector()) { |
1096 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom); |
1097 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom); |
1098 | |
1099 | // 128 bit shifts can be accomplished via 3 instructions for SHL and |
1100 | // SRL, but not for SRA because of the instructions available: |
1101 | // VS{RL} and VS{RL}O. |
1102 | setOperationAction(ISD::SHL, MVT::v1i128, Legal); |
1103 | setOperationAction(ISD::SRL, MVT::v1i128, Legal); |
1104 | setOperationAction(ISD::SRA, MVT::v1i128, Expand); |
1105 | |
1106 | addRegisterClass(MVT::f128, &PPC::VRRCRegClass); |
1107 | setOperationAction(ISD::FADD, MVT::f128, Legal); |
1108 | setOperationAction(ISD::FSUB, MVT::f128, Legal); |
1109 | setOperationAction(ISD::FDIV, MVT::f128, Legal); |
1110 | setOperationAction(ISD::FMUL, MVT::f128, Legal); |
1111 | setOperationAction(ISD::FP_EXTEND, MVT::f128, Legal); |
1112 | // No extending loads to f128 on PPC. |
1113 | for (MVT FPT : MVT::fp_valuetypes()) |
1114 | setLoadExtAction(ISD::EXTLOAD, MVT::f128, FPT, Expand); |
1115 | setOperationAction(ISD::FMA, MVT::f128, Legal); |
1116 | setCondCodeAction(ISD::SETULT, MVT::f128, Expand); |
1117 | setCondCodeAction(ISD::SETUGT, MVT::f128, Expand); |
1118 | setCondCodeAction(ISD::SETUEQ, MVT::f128, Expand); |
1119 | setCondCodeAction(ISD::SETOGE, MVT::f128, Expand); |
1120 | setCondCodeAction(ISD::SETOLE, MVT::f128, Expand); |
1121 | setCondCodeAction(ISD::SETONE, MVT::f128, Expand); |
1122 | |
1123 | setOperationAction(ISD::FTRUNC, MVT::f128, Legal); |
1124 | setOperationAction(ISD::FRINT, MVT::f128, Legal); |
1125 | setOperationAction(ISD::FFLOOR, MVT::f128, Legal); |
1126 | setOperationAction(ISD::FCEIL, MVT::f128, Legal); |
1127 | setOperationAction(ISD::FNEARBYINT, MVT::f128, Legal); |
1128 | setOperationAction(ISD::FROUND, MVT::f128, Legal); |
1129 | |
1130 | setOperationAction(ISD::SELECT, MVT::f128, Expand); |
1131 | setOperationAction(ISD::FP_ROUND, MVT::f64, Legal); |
1132 | setOperationAction(ISD::FP_ROUND, MVT::f32, Legal); |
1133 | setTruncStoreAction(MVT::f128, MVT::f64, Expand); |
1134 | setTruncStoreAction(MVT::f128, MVT::f32, Expand); |
1135 | setOperationAction(ISD::BITCAST, MVT::i128, Custom); |
1136 | // No implementation for these ops for PowerPC. |
1137 | setOperationAction(ISD::FSIN, MVT::f128, Expand); |
1138 | setOperationAction(ISD::FCOS, MVT::f128, Expand); |
1139 | setOperationAction(ISD::FPOW, MVT::f128, Expand); |
1140 | setOperationAction(ISD::FPOWI, MVT::f128, Expand); |
1141 | setOperationAction(ISD::FREM, MVT::f128, Expand); |
1142 | |
1143 | // Handle constrained floating-point operations of fp128 |
1144 | setOperationAction(ISD::STRICT_FADD, MVT::f128, Legal); |
1145 | setOperationAction(ISD::STRICT_FSUB, MVT::f128, Legal); |
1146 | setOperationAction(ISD::STRICT_FMUL, MVT::f128, Legal); |
1147 | setOperationAction(ISD::STRICT_FDIV, MVT::f128, Legal); |
1148 | setOperationAction(ISD::STRICT_FMA, MVT::f128, Legal); |
1149 | setOperationAction(ISD::STRICT_FSQRT, MVT::f128, Legal); |
1150 | setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f128, Legal); |
1151 | setOperationAction(ISD::STRICT_FP_ROUND, MVT::f64, Legal); |
1152 | setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Legal); |
1153 | setOperationAction(ISD::STRICT_FRINT, MVT::f128, Legal); |
1154 | setOperationAction(ISD::STRICT_FNEARBYINT, MVT::f128, Legal); |
1155 | setOperationAction(ISD::STRICT_FFLOOR, MVT::f128, Legal); |
1156 | setOperationAction(ISD::STRICT_FCEIL, MVT::f128, Legal); |
1157 | setOperationAction(ISD::STRICT_FTRUNC, MVT::f128, Legal); |
1158 | setOperationAction(ISD::STRICT_FROUND, MVT::f128, Legal); |
1159 | setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom); |
1160 | setOperationAction(ISD::BSWAP, MVT::v8i16, Legal); |
1161 | setOperationAction(ISD::BSWAP, MVT::v4i32, Legal); |
1162 | setOperationAction(ISD::BSWAP, MVT::v2i64, Legal); |
1163 | setOperationAction(ISD::BSWAP, MVT::v1i128, Legal); |
1164 | } |
1165 | |
1166 | if (Subtarget.hasP9Altivec()) { |
1167 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom); |
1168 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom); |
1169 | |
1170 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Legal); |
1171 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Legal); |
1172 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i32, Legal); |
1173 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Legal); |
1174 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Legal); |
1175 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i32, Legal); |
1176 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i64, Legal); |
1177 | } |
1178 | } |
1179 | |
1180 | if (Subtarget.pairedVectorMemops()) { |
1181 | addRegisterClass(MVT::v256i1, &PPC::VSRpRCRegClass); |
1182 | setOperationAction(ISD::LOAD, MVT::v256i1, Custom); |
1183 | setOperationAction(ISD::STORE, MVT::v256i1, Custom); |
1184 | } |
1185 | if (Subtarget.hasMMA()) { |
1186 | addRegisterClass(MVT::v512i1, &PPC::UACCRCRegClass); |
1187 | setOperationAction(ISD::LOAD, MVT::v512i1, Custom); |
1188 | setOperationAction(ISD::STORE, MVT::v512i1, Custom); |
1189 | setOperationAction(ISD::BUILD_VECTOR, MVT::v512i1, Custom); |
1190 | } |
1191 | |
1192 | if (Subtarget.has64BitSupport()) |
1193 | setOperationAction(ISD::PREFETCH, MVT::Other, Legal); |
1194 | |
1195 | if (Subtarget.isISA3_1()) |
1196 | setOperationAction(ISD::SRA, MVT::v1i128, Legal); |
1197 | |
1198 | setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, isPPC64 ? Legal : Custom); |
1199 | |
1200 | if (!isPPC64) { |
1201 | setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand); |
1202 | setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand); |
1203 | } |
1204 | |
1205 | setBooleanContents(ZeroOrOneBooleanContent); |
1206 | |
1207 | if (Subtarget.hasAltivec()) { |
1208 | // Altivec instructions set fields to all zeros or all ones. |
1209 | setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); |
1210 | } |
1211 | |
1212 | if (!isPPC64) { |
1213 | // These libcalls are not available in 32-bit. |
1214 | setLibcallName(RTLIB::SHL_I128, nullptr); |
1215 | setLibcallName(RTLIB::SRL_I128, nullptr); |
1216 | setLibcallName(RTLIB::SRA_I128, nullptr); |
1217 | } |
1218 | |
1219 | if (!isPPC64) |
1220 | setMaxAtomicSizeInBitsSupported(32); |
1221 | |
1222 | setStackPointerRegisterToSaveRestore(isPPC64 ? PPC::X1 : PPC::R1); |
1223 | |
1224 | // We have target-specific dag combine patterns for the following nodes: |
1225 | setTargetDAGCombine(ISD::ADD); |
1226 | setTargetDAGCombine(ISD::SHL); |
1227 | setTargetDAGCombine(ISD::SRA); |
1228 | setTargetDAGCombine(ISD::SRL); |
1229 | setTargetDAGCombine(ISD::MUL); |
1230 | setTargetDAGCombine(ISD::FMA); |
1231 | setTargetDAGCombine(ISD::SINT_TO_FP); |
1232 | setTargetDAGCombine(ISD::BUILD_VECTOR); |
1233 | if (Subtarget.hasFPCVT()) |
1234 | setTargetDAGCombine(ISD::UINT_TO_FP); |
1235 | setTargetDAGCombine(ISD::LOAD); |
1236 | setTargetDAGCombine(ISD::STORE); |
1237 | setTargetDAGCombine(ISD::BR_CC); |
1238 | if (Subtarget.useCRBits()) |
1239 | setTargetDAGCombine(ISD::BRCOND); |
1240 | setTargetDAGCombine(ISD::BSWAP); |
1241 | setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN); |
1242 | setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN); |
1243 | setTargetDAGCombine(ISD::INTRINSIC_VOID); |
1244 | |
1245 | setTargetDAGCombine(ISD::SIGN_EXTEND); |
1246 | setTargetDAGCombine(ISD::ZERO_EXTEND); |
1247 | setTargetDAGCombine(ISD::ANY_EXTEND); |
1248 | |
1249 | setTargetDAGCombine(ISD::TRUNCATE); |
1250 | setTargetDAGCombine(ISD::VECTOR_SHUFFLE); |
1251 | |
1252 | |
1253 | if (Subtarget.useCRBits()) { |
1254 | setTargetDAGCombine(ISD::TRUNCATE); |
1255 | setTargetDAGCombine(ISD::SETCC); |
1256 | setTargetDAGCombine(ISD::SELECT_CC); |
1257 | } |
1258 | |
1259 | if (Subtarget.hasP9Altivec()) { |
1260 | setTargetDAGCombine(ISD::ABS); |
1261 | setTargetDAGCombine(ISD::VSELECT); |
1262 | } |
1263 | |
1264 | setLibcallName(RTLIB::LOG_F128, "logf128"); |
1265 | setLibcallName(RTLIB::LOG2_F128, "log2f128"); |
1266 | setLibcallName(RTLIB::LOG10_F128, "log10f128"); |
1267 | setLibcallName(RTLIB::EXP_F128, "expf128"); |
1268 | setLibcallName(RTLIB::EXP2_F128, "exp2f128"); |
1269 | setLibcallName(RTLIB::SIN_F128, "sinf128"); |
1270 | setLibcallName(RTLIB::COS_F128, "cosf128"); |
1271 | setLibcallName(RTLIB::POW_F128, "powf128"); |
1272 | setLibcallName(RTLIB::FMIN_F128, "fminf128"); |
1273 | setLibcallName(RTLIB::FMAX_F128, "fmaxf128"); |
1274 | setLibcallName(RTLIB::POWI_F128, "__powikf2"); |
1275 | setLibcallName(RTLIB::REM_F128, "fmodf128"); |
1276 | |
1277 | // With 32 condition bits, we don't need to sink (and duplicate) compares |
1278 | // aggressively in CodeGenPrep. |
1279 | if (Subtarget.useCRBits()) { |
1280 | setHasMultipleConditionRegisters(); |
1281 | setJumpIsExpensive(); |
1282 | } |
1283 | |
1284 | setMinFunctionAlignment(Align(4)); |
1285 | |
1286 | switch (Subtarget.getCPUDirective()) { |
1287 | default: break; |
1288 | case PPC::DIR_970: |
1289 | case PPC::DIR_A2: |
1290 | case PPC::DIR_E500: |
1291 | case PPC::DIR_E500mc: |
1292 | case PPC::DIR_E5500: |
1293 | case PPC::DIR_PWR4: |
1294 | case PPC::DIR_PWR5: |
1295 | case PPC::DIR_PWR5X: |
1296 | case PPC::DIR_PWR6: |
1297 | case PPC::DIR_PWR6X: |
1298 | case PPC::DIR_PWR7: |
1299 | case PPC::DIR_PWR8: |
1300 | case PPC::DIR_PWR9: |
1301 | case PPC::DIR_PWR10: |
1302 | case PPC::DIR_PWR_FUTURE: |
1303 | setPrefLoopAlignment(Align(16)); |
1304 | setPrefFunctionAlignment(Align(16)); |
1305 | break; |
1306 | } |
1307 | |
1308 | if (Subtarget.enableMachineScheduler()) |
1309 | setSchedulingPreference(Sched::Source); |
1310 | else |
1311 | setSchedulingPreference(Sched::Hybrid); |
1312 | |
1313 | computeRegisterProperties(STI.getRegisterInfo()); |
1314 | |
1315 | // The Freescale cores do better with aggressive inlining of memcpy and |
1316 | // friends. GCC uses same threshold of 128 bytes (= 32 word stores). |
1317 | if (Subtarget.getCPUDirective() == PPC::DIR_E500mc || |
1318 | Subtarget.getCPUDirective() == PPC::DIR_E5500) { |
1319 | MaxStoresPerMemset = 32; |
1320 | MaxStoresPerMemsetOptSize = 16; |
1321 | MaxStoresPerMemcpy = 32; |
1322 | MaxStoresPerMemcpyOptSize = 8; |
1323 | MaxStoresPerMemmove = 32; |
1324 | MaxStoresPerMemmoveOptSize = 8; |
1325 | } else if (Subtarget.getCPUDirective() == PPC::DIR_A2) { |
1326 | // The A2 also benefits from (very) aggressive inlining of memcpy and |
1327 | // friends. The overhead of a the function call, even when warm, can be |
1328 | // over one hundred cycles. |
1329 | MaxStoresPerMemset = 128; |
1330 | MaxStoresPerMemcpy = 128; |
1331 | MaxStoresPerMemmove = 128; |
1332 | MaxLoadsPerMemcmp = 128; |
1333 | } else { |
1334 | MaxLoadsPerMemcmp = 8; |
1335 | MaxLoadsPerMemcmpOptSize = 4; |
1336 | } |
1337 | |
1338 | IsStrictFPEnabled = true; |
1339 | |
1340 | // Let the subtarget (CPU) decide if a predictable select is more expensive |
1341 | // than the corresponding branch. This information is used in CGP to decide |
1342 | // when to convert selects into branches. |
1343 | PredictableSelectIsExpensive = Subtarget.isPredictableSelectIsExpensive(); |
1344 | } |
1345 | |
1346 | /// getMaxByValAlign - Helper for getByValTypeAlignment to determine |
1347 | /// the desired ByVal argument alignment. |
1348 | static void getMaxByValAlign(Type *Ty, Align &MaxAlign, Align MaxMaxAlign) { |
1349 | if (MaxAlign == MaxMaxAlign) |
1350 | return; |
1351 | if (VectorType *VTy = dyn_cast<VectorType>(Ty)) { |
1352 | if (MaxMaxAlign >= 32 && |
1353 | VTy->getPrimitiveSizeInBits().getFixedSize() >= 256) |
1354 | MaxAlign = Align(32); |
1355 | else if (VTy->getPrimitiveSizeInBits().getFixedSize() >= 128 && |
1356 | MaxAlign < 16) |
1357 | MaxAlign = Align(16); |
1358 | } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { |
1359 | Align EltAlign; |
1360 | getMaxByValAlign(ATy->getElementType(), EltAlign, MaxMaxAlign); |
1361 | if (EltAlign > MaxAlign) |
1362 | MaxAlign = EltAlign; |
1363 | } else if (StructType *STy = dyn_cast<StructType>(Ty)) { |
1364 | for (auto *EltTy : STy->elements()) { |
1365 | Align EltAlign; |
1366 | getMaxByValAlign(EltTy, EltAlign, MaxMaxAlign); |
1367 | if (EltAlign > MaxAlign) |
1368 | MaxAlign = EltAlign; |
1369 | if (MaxAlign == MaxMaxAlign) |
1370 | break; |
1371 | } |
1372 | } |
1373 | } |
1374 | |
1375 | /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate |
1376 | /// function arguments in the caller parameter area. |
1377 | unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty, |
1378 | const DataLayout &DL) const { |
1379 | // 16byte and wider vectors are passed on 16byte boundary. |
1380 | // The rest is 8 on PPC64 and 4 on PPC32 boundary. |
1381 | Align Alignment = Subtarget.isPPC64() ? Align(8) : Align(4); |
1382 | if (Subtarget.hasAltivec()) |
1383 | getMaxByValAlign(Ty, Alignment, Align(16)); |
1384 | return Alignment.value(); |
1385 | } |
1386 | |
1387 | bool PPCTargetLowering::useSoftFloat() const { |
1388 | return Subtarget.useSoftFloat(); |
1389 | } |
1390 | |
1391 | bool PPCTargetLowering::hasSPE() const { |
1392 | return Subtarget.hasSPE(); |
1393 | } |
1394 | |
1395 | bool PPCTargetLowering::preferIncOfAddToSubOfNot(EVT VT) const { |
1396 | return VT.isScalarInteger(); |
1397 | } |
1398 | |
1399 | const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const { |
1400 | switch ((PPCISD::NodeType)Opcode) { |
1401 | case PPCISD::FIRST_NUMBER: break; |
1402 | case PPCISD::FSEL: return "PPCISD::FSEL"; |
1403 | case PPCISD::XSMAXCDP: return "PPCISD::XSMAXCDP"; |
1404 | case PPCISD::XSMINCDP: return "PPCISD::XSMINCDP"; |
1405 | case PPCISD::FCFID: return "PPCISD::FCFID"; |
1406 | case PPCISD::FCFIDU: return "PPCISD::FCFIDU"; |
1407 | case PPCISD::FCFIDS: return "PPCISD::FCFIDS"; |
1408 | case PPCISD::FCFIDUS: return "PPCISD::FCFIDUS"; |
1409 | case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ"; |
1410 | case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ"; |
1411 | case PPCISD::FCTIDUZ: return "PPCISD::FCTIDUZ"; |
1412 | case PPCISD::FCTIWUZ: return "PPCISD::FCTIWUZ"; |
1413 | case PPCISD::FP_TO_UINT_IN_VSR: |
1414 | return "PPCISD::FP_TO_UINT_IN_VSR,"; |
1415 | case PPCISD::FP_TO_SINT_IN_VSR: |
1416 | return "PPCISD::FP_TO_SINT_IN_VSR"; |
1417 | case PPCISD::FRE: return "PPCISD::FRE"; |
1418 | case PPCISD::FRSQRTE: return "PPCISD::FRSQRTE"; |
1419 | case PPCISD::STFIWX: return "PPCISD::STFIWX"; |
1420 | case PPCISD::VPERM: return "PPCISD::VPERM"; |
1421 | case PPCISD::XXSPLT: return "PPCISD::XXSPLT"; |
1422 | case PPCISD::XXSPLTI_SP_TO_DP: |
1423 | return "PPCISD::XXSPLTI_SP_TO_DP"; |
1424 | case PPCISD::XXSPLTI32DX: |
1425 | return "PPCISD::XXSPLTI32DX"; |
1426 | case PPCISD::VECINSERT: return "PPCISD::VECINSERT"; |
1427 | case PPCISD::XXPERMDI: return "PPCISD::XXPERMDI"; |
1428 | case PPCISD::VECSHL: return "PPCISD::VECSHL"; |
1429 | case PPCISD::CMPB: return "PPCISD::CMPB"; |
1430 | case PPCISD::Hi: return "PPCISD::Hi"; |
1431 | case PPCISD::Lo: return "PPCISD::Lo"; |
1432 | case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY"; |
1433 | case PPCISD::ATOMIC_CMP_SWAP_8: return "PPCISD::ATOMIC_CMP_SWAP_8"; |
1434 | case PPCISD::ATOMIC_CMP_SWAP_16: return "PPCISD::ATOMIC_CMP_SWAP_16"; |
1435 | case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC"; |
1436 | case PPCISD::DYNAREAOFFSET: return "PPCISD::DYNAREAOFFSET"; |
1437 | case PPCISD::PROBED_ALLOCA: return "PPCISD::PROBED_ALLOCA"; |
1438 | case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg"; |
1439 | case PPCISD::SRL: return "PPCISD::SRL"; |
1440 | case PPCISD::SRA: return "PPCISD::SRA"; |
1441 | case PPCISD::SHL: return "PPCISD::SHL"; |
1442 | case PPCISD::SRA_ADDZE: return "PPCISD::SRA_ADDZE"; |
1443 | case PPCISD::CALL: return "PPCISD::CALL"; |
1444 | case PPCISD::CALL_NOP: return "PPCISD::CALL_NOP"; |
1445 | case PPCISD::CALL_NOTOC: return "PPCISD::CALL_NOTOC"; |
1446 | case PPCISD::MTCTR: return "PPCISD::MTCTR"; |
1447 | case PPCISD::BCTRL: return "PPCISD::BCTRL"; |
1448 | case PPCISD::BCTRL_LOAD_TOC: return "PPCISD::BCTRL_LOAD_TOC"; |
1449 | case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG"; |
1450 | case PPCISD::READ_TIME_BASE: return "PPCISD::READ_TIME_BASE"; |
1451 | case PPCISD::EH_SJLJ_SETJMP: return "PPCISD::EH_SJLJ_SETJMP"; |
1452 | case PPCISD::EH_SJLJ_LONGJMP: return "PPCISD::EH_SJLJ_LONGJMP"; |
1453 | case PPCISD::MFOCRF: return "PPCISD::MFOCRF"; |
1454 | case PPCISD::MFVSR: return "PPCISD::MFVSR"; |
1455 | case PPCISD::MTVSRA: return "PPCISD::MTVSRA"; |
1456 | case PPCISD::MTVSRZ: return "PPCISD::MTVSRZ"; |
1457 | case PPCISD::SINT_VEC_TO_FP: return "PPCISD::SINT_VEC_TO_FP"; |
1458 | case PPCISD::UINT_VEC_TO_FP: return "PPCISD::UINT_VEC_TO_FP"; |
1459 | case PPCISD::SCALAR_TO_VECTOR_PERMUTED: |
1460 | return "PPCISD::SCALAR_TO_VECTOR_PERMUTED"; |
1461 | case PPCISD::ANDI_rec_1_EQ_BIT: |
1462 | return "PPCISD::ANDI_rec_1_EQ_BIT"; |
1463 | case PPCISD::ANDI_rec_1_GT_BIT: |
1464 | return "PPCISD::ANDI_rec_1_GT_BIT"; |
1465 | case PPCISD::VCMP: return "PPCISD::VCMP"; |
1466 | case PPCISD::VCMPo: return "PPCISD::VCMPo"; |
1467 | case PPCISD::LBRX: return "PPCISD::LBRX"; |
1468 | case PPCISD::STBRX: return "PPCISD::STBRX"; |
1469 | case PPCISD::LFIWAX: return "PPCISD::LFIWAX"; |
1470 | case PPCISD::LFIWZX: return "PPCISD::LFIWZX"; |
1471 | case PPCISD::LXSIZX: return "PPCISD::LXSIZX"; |
1472 | case PPCISD::STXSIX: return "PPCISD::STXSIX"; |
1473 | case PPCISD::VEXTS: return "PPCISD::VEXTS"; |
1474 | case PPCISD::LXVD2X: return "PPCISD::LXVD2X"; |
1475 | case PPCISD::STXVD2X: return "PPCISD::STXVD2X"; |
1476 | case PPCISD::LOAD_VEC_BE: return "PPCISD::LOAD_VEC_BE"; |
1477 | case PPCISD::STORE_VEC_BE: return "PPCISD::STORE_VEC_BE"; |
1478 | case PPCISD::ST_VSR_SCAL_INT: |
1479 | return "PPCISD::ST_VSR_SCAL_INT"; |
1480 | case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH"; |
1481 | case PPCISD::BDNZ: return "PPCISD::BDNZ"; |
1482 | case PPCISD::BDZ: return "PPCISD::BDZ"; |
1483 | case PPCISD::MFFS: return "PPCISD::MFFS"; |
1484 | case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ"; |
1485 | case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN"; |
1486 | case PPCISD::CR6SET: return "PPCISD::CR6SET"; |
1487 | case PPCISD::CR6UNSET: return "PPCISD::CR6UNSET"; |
1488 | case PPCISD::PPC32_GOT: return "PPCISD::PPC32_GOT"; |
1489 | case PPCISD::PPC32_PICGOT: return "PPCISD::PPC32_PICGOT"; |
1490 | case PPCISD::ADDIS_GOT_TPREL_HA: return "PPCISD::ADDIS_GOT_TPREL_HA"; |
1491 | case PPCISD::LD_GOT_TPREL_L: return "PPCISD::LD_GOT_TPREL_L"; |
1492 | case PPCISD::ADD_TLS: return "PPCISD::ADD_TLS"; |
1493 | case PPCISD::ADDIS_TLSGD_HA: return "PPCISD::ADDIS_TLSGD_HA"; |
1494 | case PPCISD::ADDI_TLSGD_L: return "PPCISD::ADDI_TLSGD_L"; |
1495 | case PPCISD::GET_TLS_ADDR: return "PPCISD::GET_TLS_ADDR"; |
1496 | case PPCISD::ADDI_TLSGD_L_ADDR: return "PPCISD::ADDI_TLSGD_L_ADDR"; |
1497 | case PPCISD::ADDIS_TLSLD_HA: return "PPCISD::ADDIS_TLSLD_HA"; |
1498 | case PPCISD::ADDI_TLSLD_L: return "PPCISD::ADDI_TLSLD_L"; |
1499 | case PPCISD::GET_TLSLD_ADDR: return "PPCISD::GET_TLSLD_ADDR"; |
1500 | case PPCISD::ADDI_TLSLD_L_ADDR: return "PPCISD::ADDI_TLSLD_L_ADDR"; |
1501 | case PPCISD::ADDIS_DTPREL_HA: return "PPCISD::ADDIS_DTPREL_HA"; |
1502 | case PPCISD::ADDI_DTPREL_L: return "PPCISD::ADDI_DTPREL_L"; |
1503 | case PPCISD::PADDI_DTPREL: |
1504 | return "PPCISD::PADDI_DTPREL"; |
1505 | case PPCISD::VADD_SPLAT: return "PPCISD::VADD_SPLAT"; |
1506 | case PPCISD::SC: return "PPCISD::SC"; |
1507 | case PPCISD::CLRBHRB: return "PPCISD::CLRBHRB"; |
1508 | case PPCISD::MFBHRBE: return "PPCISD::MFBHRBE"; |
1509 | case PPCISD::RFEBB: return "PPCISD::RFEBB"; |
1510 | case PPCISD::XXSWAPD: return "PPCISD::XXSWAPD"; |
1511 | case PPCISD::SWAP_NO_CHAIN: return "PPCISD::SWAP_NO_CHAIN"; |
1512 | case PPCISD::VABSD: return "PPCISD::VABSD"; |
1513 | case PPCISD::BUILD_FP128: return "PPCISD::BUILD_FP128"; |
1514 | case PPCISD::BUILD_SPE64: return "PPCISD::BUILD_SPE64"; |
1515 | case PPCISD::EXTRACT_SPE: return "PPCISD::EXTRACT_SPE"; |
1516 | case PPCISD::EXTSWSLI: return "PPCISD::EXTSWSLI"; |
1517 | case PPCISD::LD_VSX_LH: return "PPCISD::LD_VSX_LH"; |
1518 | case PPCISD::FP_EXTEND_HALF: return "PPCISD::FP_EXTEND_HALF"; |
1519 | case PPCISD::MAT_PCREL_ADDR: return "PPCISD::MAT_PCREL_ADDR"; |
1520 | case PPCISD::TLS_DYNAMIC_MAT_PCREL_ADDR: |
1521 | return "PPCISD::TLS_DYNAMIC_MAT_PCREL_ADDR"; |
1522 | case PPCISD::TLS_LOCAL_EXEC_MAT_ADDR: |
1523 | return "PPCISD::TLS_LOCAL_EXEC_MAT_ADDR"; |
1524 | case PPCISD::ACC_BUILD: return "PPCISD::ACC_BUILD"; |
1525 | case PPCISD::PAIR_BUILD: return "PPCISD::PAIR_BUILD"; |
1526 | case PPCISD::EXTRACT_VSX_REG: return "PPCISD::EXTRACT_VSX_REG"; |
1527 | case PPCISD::XXMFACC: return "PPCISD::XXMFACC"; |
1528 | case PPCISD::LD_SPLAT: return "PPCISD::LD_SPLAT"; |
1529 | case PPCISD::FNMSUB: return "PPCISD::FNMSUB"; |
1530 | case PPCISD::STRICT_FADDRTZ: |
1531 | return "PPCISD::STRICT_FADDRTZ"; |
1532 | case PPCISD::STRICT_FCTIDZ: |
1533 | return "PPCISD::STRICT_FCTIDZ"; |
1534 | case PPCISD::STRICT_FCTIWZ: |
1535 | return "PPCISD::STRICT_FCTIWZ"; |
1536 | case PPCISD::STRICT_FCTIDUZ: |
1537 | return "PPCISD::STRICT_FCTIDUZ"; |
1538 | case PPCISD::STRICT_FCTIWUZ: |
1539 | return "PPCISD::STRICT_FCTIWUZ"; |
1540 | case PPCISD::STRICT_FCFID: |
1541 | return "PPCISD::STRICT_FCFID"; |
1542 | case PPCISD::STRICT_FCFIDU: |
1543 | return "PPCISD::STRICT_FCFIDU"; |
1544 | case PPCISD::STRICT_FCFIDS: |
1545 | return "PPCISD::STRICT_FCFIDS"; |
1546 | case PPCISD::STRICT_FCFIDUS: |
1547 | return "PPCISD::STRICT_FCFIDUS"; |
1548 | case PPCISD::LXVRZX: return "PPCISD::LXVRZX"; |
1549 | } |
1550 | return nullptr; |
1551 | } |
1552 | |
1553 | EVT PPCTargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &C, |
1554 | EVT VT) const { |
1555 | if (!VT.isVector()) |
1556 | return Subtarget.useCRBits() ? MVT::i1 : MVT::i32; |
1557 | |
1558 | return VT.changeVectorElementTypeToInteger(); |
1559 | } |
1560 | |
1561 | bool PPCTargetLowering::enableAggressiveFMAFusion(EVT VT) const { |
1562 | assert(VT.isFloatingPoint() && "Non-floating-point FMA?")((VT.isFloatingPoint() && "Non-floating-point FMA?") ? static_cast<void> (0) : __assert_fail ("VT.isFloatingPoint() && \"Non-floating-point FMA?\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1562, __PRETTY_FUNCTION__)); |
1563 | return true; |
1564 | } |
1565 | |
1566 | //===----------------------------------------------------------------------===// |
1567 | // Node matching predicates, for use by the tblgen matching code. |
1568 | //===----------------------------------------------------------------------===// |
1569 | |
1570 | /// isFloatingPointZero - Return true if this is 0.0 or -0.0. |
1571 | static bool isFloatingPointZero(SDValue Op) { |
1572 | if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op)) |
1573 | return CFP->getValueAPF().isZero(); |
1574 | else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) { |
1575 | // Maybe this has already been legalized into the constant pool? |
1576 | if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1))) |
1577 | if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal())) |
1578 | return CFP->getValueAPF().isZero(); |
1579 | } |
1580 | return false; |
1581 | } |
1582 | |
1583 | /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return |
1584 | /// true if Op is undef or if it matches the specified value. |
1585 | static bool isConstantOrUndef(int Op, int Val) { |
1586 | return Op < 0 || Op == Val; |
1587 | } |
1588 | |
1589 | /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a |
1590 | /// VPKUHUM instruction. |
1591 | /// The ShuffleKind distinguishes between big-endian operations with |
1592 | /// two different inputs (0), either-endian operations with two identical |
1593 | /// inputs (1), and little-endian operations with two different inputs (2). |
1594 | /// For the latter, the input operands are swapped (see PPCInstrAltivec.td). |
1595 | bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind, |
1596 | SelectionDAG &DAG) { |
1597 | bool IsLE = DAG.getDataLayout().isLittleEndian(); |
1598 | if (ShuffleKind == 0) { |
1599 | if (IsLE) |
1600 | return false; |
1601 | for (unsigned i = 0; i != 16; ++i) |
1602 | if (!isConstantOrUndef(N->getMaskElt(i), i*2+1)) |
1603 | return false; |
1604 | } else if (ShuffleKind == 2) { |
1605 | if (!IsLE) |
1606 | return false; |
1607 | for (unsigned i = 0; i != 16; ++i) |
1608 | if (!isConstantOrUndef(N->getMaskElt(i), i*2)) |
1609 | return false; |
1610 | } else if (ShuffleKind == 1) { |
1611 | unsigned j = IsLE ? 0 : 1; |
1612 | for (unsigned i = 0; i != 8; ++i) |
1613 | if (!isConstantOrUndef(N->getMaskElt(i), i*2+j) || |
1614 | !isConstantOrUndef(N->getMaskElt(i+8), i*2+j)) |
1615 | return false; |
1616 | } |
1617 | return true; |
1618 | } |
1619 | |
1620 | /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a |
1621 | /// VPKUWUM instruction. |
1622 | /// The ShuffleKind distinguishes between big-endian operations with |
1623 | /// two different inputs (0), either-endian operations with two identical |
1624 | /// inputs (1), and little-endian operations with two different inputs (2). |
1625 | /// For the latter, the input operands are swapped (see PPCInstrAltivec.td). |
1626 | bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind, |
1627 | SelectionDAG &DAG) { |
1628 | bool IsLE = DAG.getDataLayout().isLittleEndian(); |
1629 | if (ShuffleKind == 0) { |
1630 | if (IsLE) |
1631 | return false; |
1632 | for (unsigned i = 0; i != 16; i += 2) |
1633 | if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) || |
1634 | !isConstantOrUndef(N->getMaskElt(i+1), i*2+3)) |
1635 | return false; |
1636 | } else if (ShuffleKind == 2) { |
1637 | if (!IsLE) |
1638 | return false; |
1639 | for (unsigned i = 0; i != 16; i += 2) |
1640 | if (!isConstantOrUndef(N->getMaskElt(i ), i*2) || |
1641 | !isConstantOrUndef(N->getMaskElt(i+1), i*2+1)) |
1642 | return false; |
1643 | } else if (ShuffleKind == 1) { |
1644 | unsigned j = IsLE ? 0 : 2; |
1645 | for (unsigned i = 0; i != 8; i += 2) |
1646 | if (!isConstantOrUndef(N->getMaskElt(i ), i*2+j) || |
1647 | !isConstantOrUndef(N->getMaskElt(i+1), i*2+j+1) || |
1648 | !isConstantOrUndef(N->getMaskElt(i+8), i*2+j) || |
1649 | !isConstantOrUndef(N->getMaskElt(i+9), i*2+j+1)) |
1650 | return false; |
1651 | } |
1652 | return true; |
1653 | } |
1654 | |
1655 | /// isVPKUDUMShuffleMask - Return true if this is the shuffle mask for a |
1656 | /// VPKUDUM instruction, AND the VPKUDUM instruction exists for the |
1657 | /// current subtarget. |
1658 | /// |
1659 | /// The ShuffleKind distinguishes between big-endian operations with |
1660 | /// two different inputs (0), either-endian operations with two identical |
1661 | /// inputs (1), and little-endian operations with two different inputs (2). |
1662 | /// For the latter, the input operands are swapped (see PPCInstrAltivec.td). |
1663 | bool PPC::isVPKUDUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind, |
1664 | SelectionDAG &DAG) { |
1665 | const PPCSubtarget& Subtarget = |
1666 | static_cast<const PPCSubtarget&>(DAG.getSubtarget()); |
1667 | if (!Subtarget.hasP8Vector()) |
1668 | return false; |
1669 | |
1670 | bool IsLE = DAG.getDataLayout().isLittleEndian(); |
1671 | if (ShuffleKind == 0) { |
1672 | if (IsLE) |
1673 | return false; |
1674 | for (unsigned i = 0; i != 16; i += 4) |
1675 | if (!isConstantOrUndef(N->getMaskElt(i ), i*2+4) || |
1676 | !isConstantOrUndef(N->getMaskElt(i+1), i*2+5) || |
1677 | !isConstantOrUndef(N->getMaskElt(i+2), i*2+6) || |
1678 | !isConstantOrUndef(N->getMaskElt(i+3), i*2+7)) |
1679 | return false; |
1680 | } else if (ShuffleKind == 2) { |
1681 | if (!IsLE) |
1682 | return false; |
1683 | for (unsigned i = 0; i != 16; i += 4) |
1684 | if (!isConstantOrUndef(N->getMaskElt(i ), i*2) || |
1685 | !isConstantOrUndef(N->getMaskElt(i+1), i*2+1) || |
1686 | !isConstantOrUndef(N->getMaskElt(i+2), i*2+2) || |
1687 | !isConstantOrUndef(N->getMaskElt(i+3), i*2+3)) |
1688 | return false; |
1689 | } else if (ShuffleKind == 1) { |
1690 | unsigned j = IsLE ? 0 : 4; |
1691 | for (unsigned i = 0; i != 8; i += 4) |
1692 | if (!isConstantOrUndef(N->getMaskElt(i ), i*2+j) || |
1693 | !isConstantOrUndef(N->getMaskElt(i+1), i*2+j+1) || |
1694 | !isConstantOrUndef(N->getMaskElt(i+2), i*2+j+2) || |
1695 | !isConstantOrUndef(N->getMaskElt(i+3), i*2+j+3) || |
1696 | !isConstantOrUndef(N->getMaskElt(i+8), i*2+j) || |
1697 | !isConstantOrUndef(N->getMaskElt(i+9), i*2+j+1) || |
1698 | !isConstantOrUndef(N->getMaskElt(i+10), i*2+j+2) || |
1699 | !isConstantOrUndef(N->getMaskElt(i+11), i*2+j+3)) |
1700 | return false; |
1701 | } |
1702 | return true; |
1703 | } |
1704 | |
1705 | /// isVMerge - Common function, used to match vmrg* shuffles. |
1706 | /// |
1707 | static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize, |
1708 | unsigned LHSStart, unsigned RHSStart) { |
1709 | if (N->getValueType(0) != MVT::v16i8) |
1710 | return false; |
1711 | assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&(((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) && "Unsupported merge size!") ? static_cast<void> (0) : __assert_fail ("(UnitSize == 1 || UnitSize == 2 || UnitSize == 4) && \"Unsupported merge size!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1712, __PRETTY_FUNCTION__)) |
1712 | "Unsupported merge size!")(((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) && "Unsupported merge size!") ? static_cast<void> (0) : __assert_fail ("(UnitSize == 1 || UnitSize == 2 || UnitSize == 4) && \"Unsupported merge size!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1712, __PRETTY_FUNCTION__)); |
1713 | |
1714 | for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units |
1715 | for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit |
1716 | if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j), |
1717 | LHSStart+j+i*UnitSize) || |
1718 | !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j), |
1719 | RHSStart+j+i*UnitSize)) |
1720 | return false; |
1721 | } |
1722 | return true; |
1723 | } |
1724 | |
1725 | /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for |
1726 | /// a VMRGL* instruction with the specified unit size (1,2 or 4 bytes). |
1727 | /// The ShuffleKind distinguishes between big-endian merges with two |
1728 | /// different inputs (0), either-endian merges with two identical inputs (1), |
1729 | /// and little-endian merges with two different inputs (2). For the latter, |
1730 | /// the input operands are swapped (see PPCInstrAltivec.td). |
1731 | bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize, |
1732 | unsigned ShuffleKind, SelectionDAG &DAG) { |
1733 | if (DAG.getDataLayout().isLittleEndian()) { |
1734 | if (ShuffleKind == 1) // unary |
1735 | return isVMerge(N, UnitSize, 0, 0); |
1736 | else if (ShuffleKind == 2) // swapped |
1737 | return isVMerge(N, UnitSize, 0, 16); |
1738 | else |
1739 | return false; |
1740 | } else { |
1741 | if (ShuffleKind == 1) // unary |
1742 | return isVMerge(N, UnitSize, 8, 8); |
1743 | else if (ShuffleKind == 0) // normal |
1744 | return isVMerge(N, UnitSize, 8, 24); |
1745 | else |
1746 | return false; |
1747 | } |
1748 | } |
1749 | |
1750 | /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for |
1751 | /// a VMRGH* instruction with the specified unit size (1,2 or 4 bytes). |
1752 | /// The ShuffleKind distinguishes between big-endian merges with two |
1753 | /// different inputs (0), either-endian merges with two identical inputs (1), |
1754 | /// and little-endian merges with two different inputs (2). For the latter, |
1755 | /// the input operands are swapped (see PPCInstrAltivec.td). |
1756 | bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize, |
1757 | unsigned ShuffleKind, SelectionDAG &DAG) { |
1758 | if (DAG.getDataLayout().isLittleEndian()) { |
1759 | if (ShuffleKind == 1) // unary |
1760 | return isVMerge(N, UnitSize, 8, 8); |
1761 | else if (ShuffleKind == 2) // swapped |
1762 | return isVMerge(N, UnitSize, 8, 24); |
1763 | else |
1764 | return false; |
1765 | } else { |
1766 | if (ShuffleKind == 1) // unary |
1767 | return isVMerge(N, UnitSize, 0, 0); |
1768 | else if (ShuffleKind == 0) // normal |
1769 | return isVMerge(N, UnitSize, 0, 16); |
1770 | else |
1771 | return false; |
1772 | } |
1773 | } |
1774 | |
1775 | /** |
1776 | * Common function used to match vmrgew and vmrgow shuffles |
1777 | * |
1778 | * The indexOffset determines whether to look for even or odd words in |
1779 | * the shuffle mask. This is based on the of the endianness of the target |
1780 | * machine. |
1781 | * - Little Endian: |
1782 | * - Use offset of 0 to check for odd elements |
1783 | * - Use offset of 4 to check for even elements |
1784 | * - Big Endian: |
1785 | * - Use offset of 0 to check for even elements |
1786 | * - Use offset of 4 to check for odd elements |
1787 | * A detailed description of the vector element ordering for little endian and |
1788 | * big endian can be found at |
1789 | * http://www.ibm.com/developerworks/library/l-ibm-xl-c-cpp-compiler/index.html |
1790 | * Targeting your applications - what little endian and big endian IBM XL C/C++ |
1791 | * compiler differences mean to you |
1792 | * |
1793 | * The mask to the shuffle vector instruction specifies the indices of the |
1794 | * elements from the two input vectors to place in the result. The elements are |
1795 | * numbered in array-access order, starting with the first vector. These vectors |
1796 | * are always of type v16i8, thus each vector will contain 16 elements of size |
1797 | * 8. More info on the shuffle vector can be found in the |
1798 | * http://llvm.org/docs/LangRef.html#shufflevector-instruction |
1799 | * Language Reference. |
1800 | * |
1801 | * The RHSStartValue indicates whether the same input vectors are used (unary) |
1802 | * or two different input vectors are used, based on the following: |
1803 | * - If the instruction uses the same vector for both inputs, the range of the |
1804 | * indices will be 0 to 15. In this case, the RHSStart value passed should |
1805 | * be 0. |
1806 | * - If the instruction has two different vectors then the range of the |
1807 | * indices will be 0 to 31. In this case, the RHSStart value passed should |
1808 | * be 16 (indices 0-15 specify elements in the first vector while indices 16 |
1809 | * to 31 specify elements in the second vector). |
1810 | * |
1811 | * \param[in] N The shuffle vector SD Node to analyze |
1812 | * \param[in] IndexOffset Specifies whether to look for even or odd elements |
1813 | * \param[in] RHSStartValue Specifies the starting index for the righthand input |
1814 | * vector to the shuffle_vector instruction |
1815 | * \return true iff this shuffle vector represents an even or odd word merge |
1816 | */ |
1817 | static bool isVMerge(ShuffleVectorSDNode *N, unsigned IndexOffset, |
1818 | unsigned RHSStartValue) { |
1819 | if (N->getValueType(0) != MVT::v16i8) |
1820 | return false; |
1821 | |
1822 | for (unsigned i = 0; i < 2; ++i) |
1823 | for (unsigned j = 0; j < 4; ++j) |
1824 | if (!isConstantOrUndef(N->getMaskElt(i*4+j), |
1825 | i*RHSStartValue+j+IndexOffset) || |
1826 | !isConstantOrUndef(N->getMaskElt(i*4+j+8), |
1827 | i*RHSStartValue+j+IndexOffset+8)) |
1828 | return false; |
1829 | return true; |
1830 | } |
1831 | |
1832 | /** |
1833 | * Determine if the specified shuffle mask is suitable for the vmrgew or |
1834 | * vmrgow instructions. |
1835 | * |
1836 | * \param[in] N The shuffle vector SD Node to analyze |
1837 | * \param[in] CheckEven Check for an even merge (true) or an odd merge (false) |
1838 | * \param[in] ShuffleKind Identify the type of merge: |
1839 | * - 0 = big-endian merge with two different inputs; |
1840 | * - 1 = either-endian merge with two identical inputs; |
1841 | * - 2 = little-endian merge with two different inputs (inputs are swapped for |
1842 | * little-endian merges). |
1843 | * \param[in] DAG The current SelectionDAG |
1844 | * \return true iff this shuffle mask |
1845 | */ |
1846 | bool PPC::isVMRGEOShuffleMask(ShuffleVectorSDNode *N, bool CheckEven, |
1847 | unsigned ShuffleKind, SelectionDAG &DAG) { |
1848 | if (DAG.getDataLayout().isLittleEndian()) { |
1849 | unsigned indexOffset = CheckEven ? 4 : 0; |
1850 | if (ShuffleKind == 1) // Unary |
1851 | return isVMerge(N, indexOffset, 0); |
1852 | else if (ShuffleKind == 2) // swapped |
1853 | return isVMerge(N, indexOffset, 16); |
1854 | else |
1855 | return false; |
1856 | } |
1857 | else { |
1858 | unsigned indexOffset = CheckEven ? 0 : 4; |
1859 | if (ShuffleKind == 1) // Unary |
1860 | return isVMerge(N, indexOffset, 0); |
1861 | else if (ShuffleKind == 0) // Normal |
1862 | return isVMerge(N, indexOffset, 16); |
1863 | else |
1864 | return false; |
1865 | } |
1866 | return false; |
1867 | } |
1868 | |
1869 | /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift |
1870 | /// amount, otherwise return -1. |
1871 | /// The ShuffleKind distinguishes between big-endian operations with two |
1872 | /// different inputs (0), either-endian operations with two identical inputs |
1873 | /// (1), and little-endian operations with two different inputs (2). For the |
1874 | /// latter, the input operands are swapped (see PPCInstrAltivec.td). |
1875 | int PPC::isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind, |
1876 | SelectionDAG &DAG) { |
1877 | if (N->getValueType(0) != MVT::v16i8) |
1878 | return -1; |
1879 | |
1880 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N); |
1881 | |
1882 | // Find the first non-undef value in the shuffle mask. |
1883 | unsigned i; |
1884 | for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i) |
1885 | /*search*/; |
1886 | |
1887 | if (i == 16) return -1; // all undef. |
1888 | |
1889 | // Otherwise, check to see if the rest of the elements are consecutively |
1890 | // numbered from this value. |
1891 | unsigned ShiftAmt = SVOp->getMaskElt(i); |
1892 | if (ShiftAmt < i) return -1; |
1893 | |
1894 | ShiftAmt -= i; |
1895 | bool isLE = DAG.getDataLayout().isLittleEndian(); |
1896 | |
1897 | if ((ShuffleKind == 0 && !isLE) || (ShuffleKind == 2 && isLE)) { |
1898 | // Check the rest of the elements to see if they are consecutive. |
1899 | for (++i; i != 16; ++i) |
1900 | if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i)) |
1901 | return -1; |
1902 | } else if (ShuffleKind == 1) { |
1903 | // Check the rest of the elements to see if they are consecutive. |
1904 | for (++i; i != 16; ++i) |
1905 | if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15)) |
1906 | return -1; |
1907 | } else |
1908 | return -1; |
1909 | |
1910 | if (isLE) |
1911 | ShiftAmt = 16 - ShiftAmt; |
1912 | |
1913 | return ShiftAmt; |
1914 | } |
1915 | |
1916 | /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand |
1917 | /// specifies a splat of a single element that is suitable for input to |
1918 | /// one of the splat operations (VSPLTB/VSPLTH/VSPLTW/XXSPLTW/LXVDSX/etc.). |
1919 | bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) { |
1920 | assert(N->getValueType(0) == MVT::v16i8 && isPowerOf2_32(EltSize) &&((N->getValueType(0) == MVT::v16i8 && isPowerOf2_32 (EltSize) && EltSize <= 8 && "Can only handle 1,2,4,8 byte element sizes" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::v16i8 && isPowerOf2_32(EltSize) && EltSize <= 8 && \"Can only handle 1,2,4,8 byte element sizes\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1921, __PRETTY_FUNCTION__)) |
1921 | EltSize <= 8 && "Can only handle 1,2,4,8 byte element sizes")((N->getValueType(0) == MVT::v16i8 && isPowerOf2_32 (EltSize) && EltSize <= 8 && "Can only handle 1,2,4,8 byte element sizes" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::v16i8 && isPowerOf2_32(EltSize) && EltSize <= 8 && \"Can only handle 1,2,4,8 byte element sizes\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1921, __PRETTY_FUNCTION__)); |
1922 | |
1923 | // The consecutive indices need to specify an element, not part of two |
1924 | // different elements. So abandon ship early if this isn't the case. |
1925 | if (N->getMaskElt(0) % EltSize != 0) |
1926 | return false; |
1927 | |
1928 | // This is a splat operation if each element of the permute is the same, and |
1929 | // if the value doesn't reference the second vector. |
1930 | unsigned ElementBase = N->getMaskElt(0); |
1931 | |
1932 | // FIXME: Handle UNDEF elements too! |
1933 | if (ElementBase >= 16) |
1934 | return false; |
1935 | |
1936 | // Check that the indices are consecutive, in the case of a multi-byte element |
1937 | // splatted with a v16i8 mask. |
1938 | for (unsigned i = 1; i != EltSize; ++i) |
1939 | if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase)) |
1940 | return false; |
1941 | |
1942 | for (unsigned i = EltSize, e = 16; i != e; i += EltSize) { |
1943 | if (N->getMaskElt(i) < 0) continue; |
1944 | for (unsigned j = 0; j != EltSize; ++j) |
1945 | if (N->getMaskElt(i+j) != N->getMaskElt(j)) |
1946 | return false; |
1947 | } |
1948 | return true; |
1949 | } |
1950 | |
1951 | /// Check that the mask is shuffling N byte elements. Within each N byte |
1952 | /// element of the mask, the indices could be either in increasing or |
1953 | /// decreasing order as long as they are consecutive. |
1954 | /// \param[in] N the shuffle vector SD Node to analyze |
1955 | /// \param[in] Width the element width in bytes, could be 2/4/8/16 (HalfWord/ |
1956 | /// Word/DoubleWord/QuadWord). |
1957 | /// \param[in] StepLen the delta indices number among the N byte element, if |
1958 | /// the mask is in increasing/decreasing order then it is 1/-1. |
1959 | /// \return true iff the mask is shuffling N byte elements. |
1960 | static bool isNByteElemShuffleMask(ShuffleVectorSDNode *N, unsigned Width, |
1961 | int StepLen) { |
1962 | assert((Width == 2 || Width == 4 || Width == 8 || Width == 16) &&(((Width == 2 || Width == 4 || Width == 8 || Width == 16) && "Unexpected element width.") ? static_cast<void> (0) : __assert_fail ("(Width == 2 || Width == 4 || Width == 8 || Width == 16) && \"Unexpected element width.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1963, __PRETTY_FUNCTION__)) |
1963 | "Unexpected element width.")(((Width == 2 || Width == 4 || Width == 8 || Width == 16) && "Unexpected element width.") ? static_cast<void> (0) : __assert_fail ("(Width == 2 || Width == 4 || Width == 8 || Width == 16) && \"Unexpected element width.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1963, __PRETTY_FUNCTION__)); |
1964 | assert((StepLen == 1 || StepLen == -1) && "Unexpected element width.")(((StepLen == 1 || StepLen == -1) && "Unexpected element width." ) ? static_cast<void> (0) : __assert_fail ("(StepLen == 1 || StepLen == -1) && \"Unexpected element width.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1964, __PRETTY_FUNCTION__)); |
1965 | |
1966 | unsigned NumOfElem = 16 / Width; |
1967 | unsigned MaskVal[16]; // Width is never greater than 16 |
1968 | for (unsigned i = 0; i < NumOfElem; ++i) { |
1969 | MaskVal[0] = N->getMaskElt(i * Width); |
1970 | if ((StepLen == 1) && (MaskVal[0] % Width)) { |
1971 | return false; |
1972 | } else if ((StepLen == -1) && ((MaskVal[0] + 1) % Width)) { |
1973 | return false; |
1974 | } |
1975 | |
1976 | for (unsigned int j = 1; j < Width; ++j) { |
1977 | MaskVal[j] = N->getMaskElt(i * Width + j); |
1978 | if (MaskVal[j] != MaskVal[j-1] + StepLen) { |
1979 | return false; |
1980 | } |
1981 | } |
1982 | } |
1983 | |
1984 | return true; |
1985 | } |
1986 | |
1987 | bool PPC::isXXINSERTWMask(ShuffleVectorSDNode *N, unsigned &ShiftElts, |
1988 | unsigned &InsertAtByte, bool &Swap, bool IsLE) { |
1989 | if (!isNByteElemShuffleMask(N, 4, 1)) |
1990 | return false; |
1991 | |
1992 | // Now we look at mask elements 0,4,8,12 |
1993 | unsigned M0 = N->getMaskElt(0) / 4; |
1994 | unsigned M1 = N->getMaskElt(4) / 4; |
1995 | unsigned M2 = N->getMaskElt(8) / 4; |
1996 | unsigned M3 = N->getMaskElt(12) / 4; |
1997 | unsigned LittleEndianShifts[] = { 2, 1, 0, 3 }; |
1998 | unsigned BigEndianShifts[] = { 3, 0, 1, 2 }; |
1999 | |
2000 | // Below, let H and L be arbitrary elements of the shuffle mask |
2001 | // where H is in the range [4,7] and L is in the range [0,3]. |
2002 | // H, 1, 2, 3 or L, 5, 6, 7 |
2003 | if ((M0 > 3 && M1 == 1 && M2 == 2 && M3 == 3) || |
2004 | (M0 < 4 && M1 == 5 && M2 == 6 && M3 == 7)) { |
2005 | ShiftElts = IsLE ? LittleEndianShifts[M0 & 0x3] : BigEndianShifts[M0 & 0x3]; |
2006 | InsertAtByte = IsLE ? 12 : 0; |
2007 | Swap = M0 < 4; |
2008 | return true; |
2009 | } |
2010 | // 0, H, 2, 3 or 4, L, 6, 7 |
2011 | if ((M1 > 3 && M0 == 0 && M2 == 2 && M3 == 3) || |
2012 | (M1 < 4 && M0 == 4 && M2 == 6 && M3 == 7)) { |
2013 | ShiftElts = IsLE ? LittleEndianShifts[M1 & 0x3] : BigEndianShifts[M1 & 0x3]; |
2014 | InsertAtByte = IsLE ? 8 : 4; |
2015 | Swap = M1 < 4; |
2016 | return true; |
2017 | } |
2018 | // 0, 1, H, 3 or 4, 5, L, 7 |
2019 | if ((M2 > 3 && M0 == 0 && M1 == 1 && M3 == 3) || |
2020 | (M2 < 4 && M0 == 4 && M1 == 5 && M3 == 7)) { |
2021 | ShiftElts = IsLE ? LittleEndianShifts[M2 & 0x3] : BigEndianShifts[M2 & 0x3]; |
2022 | InsertAtByte = IsLE ? 4 : 8; |
2023 | Swap = M2 < 4; |
2024 | return true; |
2025 | } |
2026 | // 0, 1, 2, H or 4, 5, 6, L |
2027 | if ((M3 > 3 && M0 == 0 && M1 == 1 && M2 == 2) || |
2028 | (M3 < 4 && M0 == 4 && M1 == 5 && M2 == 6)) { |
2029 | ShiftElts = IsLE ? LittleEndianShifts[M3 & 0x3] : BigEndianShifts[M3 & 0x3]; |
2030 | InsertAtByte = IsLE ? 0 : 12; |
2031 | Swap = M3 < 4; |
2032 | return true; |
2033 | } |
2034 | |
2035 | // If both vector operands for the shuffle are the same vector, the mask will |
2036 | // contain only elements from the first one and the second one will be undef. |
2037 | if (N->getOperand(1).isUndef()) { |
2038 | ShiftElts = 0; |
2039 | Swap = true; |
2040 | unsigned XXINSERTWSrcElem = IsLE ? 2 : 1; |
2041 | if (M0 == XXINSERTWSrcElem && M1 == 1 && M2 == 2 && M3 == 3) { |
2042 | InsertAtByte = IsLE ? 12 : 0; |
2043 | return true; |
2044 | } |
2045 | if (M0 == 0 && M1 == XXINSERTWSrcElem && M2 == 2 && M3 == 3) { |
2046 | InsertAtByte = IsLE ? 8 : 4; |
2047 | return true; |
2048 | } |
2049 | if (M0 == 0 && M1 == 1 && M2 == XXINSERTWSrcElem && M3 == 3) { |
2050 | InsertAtByte = IsLE ? 4 : 8; |
2051 | return true; |
2052 | } |
2053 | if (M0 == 0 && M1 == 1 && M2 == 2 && M3 == XXINSERTWSrcElem) { |
2054 | InsertAtByte = IsLE ? 0 : 12; |
2055 | return true; |
2056 | } |
2057 | } |
2058 | |
2059 | return false; |
2060 | } |
2061 | |
2062 | bool PPC::isXXSLDWIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts, |
2063 | bool &Swap, bool IsLE) { |
2064 | assert(N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8")((N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::v16i8 && \"Shuffle vector expects v16i8\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2064, __PRETTY_FUNCTION__)); |
2065 | // Ensure each byte index of the word is consecutive. |
2066 | if (!isNByteElemShuffleMask(N, 4, 1)) |
2067 | return false; |
2068 | |
2069 | // Now we look at mask elements 0,4,8,12, which are the beginning of words. |
2070 | unsigned M0 = N->getMaskElt(0) / 4; |
2071 | unsigned M1 = N->getMaskElt(4) / 4; |
2072 | unsigned M2 = N->getMaskElt(8) / 4; |
2073 | unsigned M3 = N->getMaskElt(12) / 4; |
2074 | |
2075 | // If both vector operands for the shuffle are the same vector, the mask will |
2076 | // contain only elements from the first one and the second one will be undef. |
2077 | if (N->getOperand(1).isUndef()) { |
2078 | assert(M0 < 4 && "Indexing into an undef vector?")((M0 < 4 && "Indexing into an undef vector?") ? static_cast <void> (0) : __assert_fail ("M0 < 4 && \"Indexing into an undef vector?\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2078, __PRETTY_FUNCTION__)); |
2079 | if (M1 != (M0 + 1) % 4 || M2 != (M1 + 1) % 4 || M3 != (M2 + 1) % 4) |
2080 | return false; |
2081 | |
2082 | ShiftElts = IsLE ? (4 - M0) % 4 : M0; |
2083 | Swap = false; |
2084 | return true; |
2085 | } |
2086 | |
2087 | // Ensure each word index of the ShuffleVector Mask is consecutive. |
2088 | if (M1 != (M0 + 1) % 8 || M2 != (M1 + 1) % 8 || M3 != (M2 + 1) % 8) |
2089 | return false; |
2090 | |
2091 | if (IsLE) { |
2092 | if (M0 == 0 || M0 == 7 || M0 == 6 || M0 == 5) { |
2093 | // Input vectors don't need to be swapped if the leading element |
2094 | // of the result is one of the 3 left elements of the second vector |
2095 | // (or if there is no shift to be done at all). |
2096 | Swap = false; |
2097 | ShiftElts = (8 - M0) % 8; |
2098 | } else if (M0 == 4 || M0 == 3 || M0 == 2 || M0 == 1) { |
2099 | // Input vectors need to be swapped if the leading element |
2100 | // of the result is one of the 3 left elements of the first vector |
2101 | // (or if we're shifting by 4 - thereby simply swapping the vectors). |
2102 | Swap = true; |
2103 | ShiftElts = (4 - M0) % 4; |
2104 | } |
2105 | |
2106 | return true; |
2107 | } else { // BE |
2108 | if (M0 == 0 || M0 == 1 || M0 == 2 || M0 == 3) { |
2109 | // Input vectors don't need to be swapped if the leading element |
2110 | // of the result is one of the 4 elements of the first vector. |
2111 | Swap = false; |
2112 | ShiftElts = M0; |
2113 | } else if (M0 == 4 || M0 == 5 || M0 == 6 || M0 == 7) { |
2114 | // Input vectors need to be swapped if the leading element |
2115 | // of the result is one of the 4 elements of the right vector. |
2116 | Swap = true; |
2117 | ShiftElts = M0 - 4; |
2118 | } |
2119 | |
2120 | return true; |
2121 | } |
2122 | } |
2123 | |
2124 | bool static isXXBRShuffleMaskHelper(ShuffleVectorSDNode *N, int Width) { |
2125 | assert(N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8")((N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::v16i8 && \"Shuffle vector expects v16i8\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2125, __PRETTY_FUNCTION__)); |
2126 | |
2127 | if (!isNByteElemShuffleMask(N, Width, -1)) |
2128 | return false; |
2129 | |
2130 | for (int i = 0; i < 16; i += Width) |
2131 | if (N->getMaskElt(i) != i + Width - 1) |
2132 | return false; |
2133 | |
2134 | return true; |
2135 | } |
2136 | |
2137 | bool PPC::isXXBRHShuffleMask(ShuffleVectorSDNode *N) { |
2138 | return isXXBRShuffleMaskHelper(N, 2); |
2139 | } |
2140 | |
2141 | bool PPC::isXXBRWShuffleMask(ShuffleVectorSDNode *N) { |
2142 | return isXXBRShuffleMaskHelper(N, 4); |
2143 | } |
2144 | |
2145 | bool PPC::isXXBRDShuffleMask(ShuffleVectorSDNode *N) { |
2146 | return isXXBRShuffleMaskHelper(N, 8); |
2147 | } |
2148 | |
2149 | bool PPC::isXXBRQShuffleMask(ShuffleVectorSDNode *N) { |
2150 | return isXXBRShuffleMaskHelper(N, 16); |
2151 | } |
2152 | |
2153 | /// Can node \p N be lowered to an XXPERMDI instruction? If so, set \p Swap |
2154 | /// if the inputs to the instruction should be swapped and set \p DM to the |
2155 | /// value for the immediate. |
2156 | /// Specifically, set \p Swap to true only if \p N can be lowered to XXPERMDI |
2157 | /// AND element 0 of the result comes from the first input (LE) or second input |
2158 | /// (BE). Set \p DM to the calculated result (0-3) only if \p N can be lowered. |
2159 | /// \return true iff the given mask of shuffle node \p N is a XXPERMDI shuffle |
2160 | /// mask. |
2161 | bool PPC::isXXPERMDIShuffleMask(ShuffleVectorSDNode *N, unsigned &DM, |
2162 | bool &Swap, bool IsLE) { |
2163 | assert(N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8")((N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::v16i8 && \"Shuffle vector expects v16i8\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2163, __PRETTY_FUNCTION__)); |
2164 | |
2165 | // Ensure each byte index of the double word is consecutive. |
2166 | if (!isNByteElemShuffleMask(N, 8, 1)) |
2167 | return false; |
2168 | |
2169 | unsigned M0 = N->getMaskElt(0) / 8; |
2170 | unsigned M1 = N->getMaskElt(8) / 8; |
2171 | assert(((M0 | M1) < 4) && "A mask element out of bounds?")((((M0 | M1) < 4) && "A mask element out of bounds?" ) ? static_cast<void> (0) : __assert_fail ("((M0 | M1) < 4) && \"A mask element out of bounds?\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2171, __PRETTY_FUNCTION__)); |
2172 | |
2173 | // If both vector operands for the shuffle are the same vector, the mask will |
2174 | // contain only elements from the first one and the second one will be undef. |
2175 | if (N->getOperand(1).isUndef()) { |
2176 | if ((M0 | M1) < 2) { |
2177 | DM = IsLE ? (((~M1) & 1) << 1) + ((~M0) & 1) : (M0 << 1) + (M1 & 1); |
2178 | Swap = false; |
2179 | return true; |
2180 | } else |
2181 | return false; |
2182 | } |
2183 | |
2184 | if (IsLE) { |
2185 | if (M0 > 1 && M1 < 2) { |
2186 | Swap = false; |
2187 | } else if (M0 < 2 && M1 > 1) { |
2188 | M0 = (M0 + 2) % 4; |
2189 | M1 = (M1 + 2) % 4; |
2190 | Swap = true; |
2191 | } else |
2192 | return false; |
2193 | |
2194 | // Note: if control flow comes here that means Swap is already set above |
2195 | DM = (((~M1) & 1) << 1) + ((~M0) & 1); |
2196 | return true; |
2197 | } else { // BE |
2198 | if (M0 < 2 && M1 > 1) { |
2199 | Swap = false; |
2200 | } else if (M0 > 1 && M1 < 2) { |
2201 | M0 = (M0 + 2) % 4; |
2202 | M1 = (M1 + 2) % 4; |
2203 | Swap = true; |
2204 | } else |
2205 | return false; |
2206 | |
2207 | // Note: if control flow comes here that means Swap is already set above |
2208 | DM = (M0 << 1) + (M1 & 1); |
2209 | return true; |
2210 | } |
2211 | } |
2212 | |
2213 | |
2214 | /// getSplatIdxForPPCMnemonics - Return the splat index as a value that is |
2215 | /// appropriate for PPC mnemonics (which have a big endian bias - namely |
2216 | /// elements are counted from the left of the vector register). |
2217 | unsigned PPC::getSplatIdxForPPCMnemonics(SDNode *N, unsigned EltSize, |
2218 | SelectionDAG &DAG) { |
2219 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N); |
2220 | assert(isSplatShuffleMask(SVOp, EltSize))((isSplatShuffleMask(SVOp, EltSize)) ? static_cast<void> (0) : __assert_fail ("isSplatShuffleMask(SVOp, EltSize)", "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2220, __PRETTY_FUNCTION__)); |
2221 | if (DAG.getDataLayout().isLittleEndian()) |
2222 | return (16 / EltSize) - 1 - (SVOp->getMaskElt(0) / EltSize); |
2223 | else |
2224 | return SVOp->getMaskElt(0) / EltSize; |
2225 | } |
2226 | |
2227 | /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed |
2228 | /// by using a vspltis[bhw] instruction of the specified element size, return |
2229 | /// the constant being splatted. The ByteSize field indicates the number of |
2230 | /// bytes of each element [124] -> [bhw]. |
2231 | SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) { |
2232 | SDValue OpVal(nullptr, 0); |
2233 | |
2234 | // If ByteSize of the splat is bigger than the element size of the |
2235 | // build_vector, then we have a case where we are checking for a splat where |
2236 | // multiple elements of the buildvector are folded together into a single |
2237 | // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8). |
2238 | unsigned EltSize = 16/N->getNumOperands(); |
2239 | if (EltSize < ByteSize) { |
2240 | unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval. |
2241 | SDValue UniquedVals[4]; |
2242 | assert(Multiple > 1 && Multiple <= 4 && "How can this happen?")((Multiple > 1 && Multiple <= 4 && "How can this happen?" ) ? static_cast<void> (0) : __assert_fail ("Multiple > 1 && Multiple <= 4 && \"How can this happen?\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2242, __PRETTY_FUNCTION__)); |
2243 | |
2244 | // See if all of the elements in the buildvector agree across. |
2245 | for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { |
2246 | if (N->getOperand(i).isUndef()) continue; |
2247 | // If the element isn't a constant, bail fully out. |
2248 | if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue(); |
2249 | |
2250 | if (!UniquedVals[i&(Multiple-1)].getNode()) |
2251 | UniquedVals[i&(Multiple-1)] = N->getOperand(i); |
2252 | else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i)) |
2253 | return SDValue(); // no match. |
2254 | } |
2255 | |
2256 | // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains |
2257 | // either constant or undef values that are identical for each chunk. See |
2258 | // if these chunks can form into a larger vspltis*. |
2259 | |
2260 | // Check to see if all of the leading entries are either 0 or -1. If |
2261 | // neither, then this won't fit into the immediate field. |
2262 | bool LeadingZero = true; |
2263 | bool LeadingOnes = true; |
2264 | for (unsigned i = 0; i != Multiple-1; ++i) { |
2265 | if (!UniquedVals[i].getNode()) continue; // Must have been undefs. |
2266 | |
2267 | LeadingZero &= isNullConstant(UniquedVals[i]); |
2268 | LeadingOnes &= isAllOnesConstant(UniquedVals[i]); |
2269 | } |
2270 | // Finally, check the least significant entry. |
2271 | if (LeadingZero) { |
2272 | if (!UniquedVals[Multiple-1].getNode()) |
2273 | return DAG.getTargetConstant(0, SDLoc(N), MVT::i32); // 0,0,0,undef |
2274 | int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue(); |
2275 | if (Val < 16) // 0,0,0,4 -> vspltisw(4) |
2276 | return DAG.getTargetConstant(Val, SDLoc(N), MVT::i32); |
2277 | } |
2278 | if (LeadingOnes) { |
2279 | if (!UniquedVals[Multiple-1].getNode()) |
2280 | return DAG.getTargetConstant(~0U, SDLoc(N), MVT::i32); // -1,-1,-1,undef |
2281 | int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue(); |
2282 | if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2) |
2283 | return DAG.getTargetConstant(Val, SDLoc(N), MVT::i32); |
2284 | } |
2285 | |
2286 | return SDValue(); |
2287 | } |
2288 | |
2289 | // Check to see if this buildvec has a single non-undef value in its elements. |
2290 | for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { |
2291 | if (N->getOperand(i).isUndef()) continue; |
2292 | if (!OpVal.getNode()) |
2293 | OpVal = N->getOperand(i); |
2294 | else if (OpVal != N->getOperand(i)) |
2295 | return SDValue(); |
2296 | } |
2297 | |
2298 | if (!OpVal.getNode()) return SDValue(); // All UNDEF: use implicit def. |
2299 | |
2300 | unsigned ValSizeInBytes = EltSize; |
2301 | uint64_t Value = 0; |
2302 | if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) { |
2303 | Value = CN->getZExtValue(); |
2304 | } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) { |
2305 | assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!")((CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!" ) ? static_cast<void> (0) : __assert_fail ("CN->getValueType(0) == MVT::f32 && \"Only one legal FP vector type!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2305, __PRETTY_FUNCTION__)); |
2306 | Value = FloatToBits(CN->getValueAPF().convertToFloat()); |
2307 | } |
2308 | |
2309 | // If the splat value is larger than the element value, then we can never do |
2310 | // this splat. The only case that we could fit the replicated bits into our |
2311 | // immediate field for would be zero, and we prefer to use vxor for it. |
2312 | if (ValSizeInBytes < ByteSize) return SDValue(); |
2313 | |
2314 | // If the element value is larger than the splat value, check if it consists |
2315 | // of a repeated bit pattern of size ByteSize. |
2316 | if (!APInt(ValSizeInBytes * 8, Value).isSplat(ByteSize * 8)) |
2317 | return SDValue(); |
2318 | |
2319 | // Properly sign extend the value. |
2320 | int MaskVal = SignExtend32(Value, ByteSize * 8); |
2321 | |
2322 | // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros. |
2323 | if (MaskVal == 0) return SDValue(); |
2324 | |
2325 | // Finally, if this value fits in a 5 bit sext field, return it |
2326 | if (SignExtend32<5>(MaskVal) == MaskVal) |
2327 | return DAG.getTargetConstant(MaskVal, SDLoc(N), MVT::i32); |
2328 | return SDValue(); |
2329 | } |
2330 | |
2331 | /// isQVALIGNIShuffleMask - If this is a qvaligni shuffle mask, return the shift |
2332 | /// amount, otherwise return -1. |
2333 | int PPC::isQVALIGNIShuffleMask(SDNode *N) { |
2334 | EVT VT = N->getValueType(0); |
2335 | if (VT != MVT::v4f64 && VT != MVT::v4f32 && VT != MVT::v4i1) |
2336 | return -1; |
2337 | |
2338 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N); |
2339 | |
2340 | // Find the first non-undef value in the shuffle mask. |
2341 | unsigned i; |
2342 | for (i = 0; i != 4 && SVOp->getMaskElt(i) < 0; ++i) |
2343 | /*search*/; |
2344 | |
2345 | if (i == 4) return -1; // all undef. |
2346 | |
2347 | // Otherwise, check to see if the rest of the elements are consecutively |
2348 | // numbered from this value. |
2349 | unsigned ShiftAmt = SVOp->getMaskElt(i); |
2350 | if (ShiftAmt < i) return -1; |
2351 | ShiftAmt -= i; |
2352 | |
2353 | // Check the rest of the elements to see if they are consecutive. |
2354 | for (++i; i != 4; ++i) |
2355 | if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i)) |
2356 | return -1; |
2357 | |
2358 | return ShiftAmt; |
2359 | } |
2360 | |
2361 | //===----------------------------------------------------------------------===// |
2362 | // Addressing Mode Selection |
2363 | //===----------------------------------------------------------------------===// |
2364 | |
2365 | /// isIntS16Immediate - This method tests to see if the node is either a 32-bit |
2366 | /// or 64-bit immediate, and if the value can be accurately represented as a |
2367 | /// sign extension from a 16-bit value. If so, this returns true and the |
2368 | /// immediate. |
2369 | bool llvm::isIntS16Immediate(SDNode *N, int16_t &Imm) { |
2370 | if (!isa<ConstantSDNode>(N)) |
2371 | return false; |
2372 | |
2373 | Imm = (int16_t)cast<ConstantSDNode>(N)->getZExtValue(); |
2374 | if (N->getValueType(0) == MVT::i32) |
2375 | return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue(); |
2376 | else |
2377 | return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue(); |
2378 | } |
2379 | bool llvm::isIntS16Immediate(SDValue Op, int16_t &Imm) { |
2380 | return isIntS16Immediate(Op.getNode(), Imm); |
2381 | } |
2382 | |
2383 | |
2384 | /// SelectAddressEVXRegReg - Given the specified address, check to see if it can |
2385 | /// be represented as an indexed [r+r] operation. |
2386 | bool PPCTargetLowering::SelectAddressEVXRegReg(SDValue N, SDValue &Base, |
2387 | SDValue &Index, |
2388 | SelectionDAG &DAG) const { |
2389 | for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end(); |
2390 | UI != E; ++UI) { |
2391 | if (MemSDNode *Memop = dyn_cast<MemSDNode>(*UI)) { |
2392 | if (Memop->getMemoryVT() == MVT::f64) { |
2393 | Base = N.getOperand(0); |
2394 | Index = N.getOperand(1); |
2395 | return true; |
2396 | } |
2397 | } |
2398 | } |
2399 | return false; |
2400 | } |
2401 | |
2402 | /// SelectAddressRegReg - Given the specified addressed, check to see if it |
2403 | /// can be represented as an indexed [r+r] operation. Returns false if it |
2404 | /// can be more efficiently represented as [r+imm]. If \p EncodingAlignment is |
2405 | /// non-zero and N can be represented by a base register plus a signed 16-bit |
2406 | /// displacement, make a more precise judgement by checking (displacement % \p |
2407 | /// EncodingAlignment). |
2408 | bool PPCTargetLowering::SelectAddressRegReg( |
2409 | SDValue N, SDValue &Base, SDValue &Index, SelectionDAG &DAG, |
2410 | MaybeAlign EncodingAlignment) const { |
2411 | // If we have a PC Relative target flag don't select as [reg+reg]. It will be |
2412 | // a [pc+imm]. |
2413 | if (SelectAddressPCRel(N, Base)) |
2414 | return false; |
2415 | |
2416 | int16_t Imm = 0; |
2417 | if (N.getOpcode() == ISD::ADD) { |
2418 | // Is there any SPE load/store (f64), which can't handle 16bit offset? |
2419 | // SPE load/store can only handle 8-bit offsets. |
2420 | if (hasSPE() && SelectAddressEVXRegReg(N, Base, Index, DAG)) |
2421 | return true; |
2422 | if (isIntS16Immediate(N.getOperand(1), Imm) && |
2423 | (!EncodingAlignment || isAligned(*EncodingAlignment, Imm))) |
2424 | return false; // r+i |
2425 | if (N.getOperand(1).getOpcode() == PPCISD::Lo) |
2426 | return false; // r+i |
2427 | |
2428 | Base = N.getOperand(0); |
2429 | Index = N.getOperand(1); |
2430 | return true; |
2431 | } else if (N.getOpcode() == ISD::OR) { |
2432 | if (isIntS16Immediate(N.getOperand(1), Imm) && |
2433 | (!EncodingAlignment || isAligned(*EncodingAlignment, Imm))) |
2434 | return false; // r+i can fold it if we can. |
2435 | |
2436 | // If this is an or of disjoint bitfields, we can codegen this as an add |
2437 | // (for better address arithmetic) if the LHS and RHS of the OR are provably |
2438 | // disjoint. |
2439 | KnownBits LHSKnown = DAG.computeKnownBits(N.getOperand(0)); |
2440 | |
2441 | if (LHSKnown.Zero.getBoolValue()) { |
2442 | KnownBits RHSKnown = DAG.computeKnownBits(N.getOperand(1)); |
2443 | // If all of the bits are known zero on the LHS or RHS, the add won't |
2444 | // carry. |
2445 | if (~(LHSKnown.Zero | RHSKnown.Zero) == 0) { |
2446 | Base = N.getOperand(0); |
2447 | Index = N.getOperand(1); |
2448 | return true; |
2449 | } |
2450 | } |
2451 | } |
2452 | |
2453 | return false; |
2454 | } |
2455 | |
2456 | // If we happen to be doing an i64 load or store into a stack slot that has |
2457 | // less than a 4-byte alignment, then the frame-index elimination may need to |
2458 | // use an indexed load or store instruction (because the offset may not be a |
2459 | // multiple of 4). The extra register needed to hold the offset comes from the |
2460 | // register scavenger, and it is possible that the scavenger will need to use |
2461 | // an emergency spill slot. As a result, we need to make sure that a spill slot |
2462 | // is allocated when doing an i64 load/store into a less-than-4-byte-aligned |
2463 | // stack slot. |
2464 | static void fixupFuncForFI(SelectionDAG &DAG, int FrameIdx, EVT VT) { |
2465 | // FIXME: This does not handle the LWA case. |
2466 | if (VT != MVT::i64) |
2467 | return; |
2468 | |
2469 | // NOTE: We'll exclude negative FIs here, which come from argument |
2470 | // lowering, because there are no known test cases triggering this problem |
2471 | // using packed structures (or similar). We can remove this exclusion if |
2472 | // we find such a test case. The reason why this is so test-case driven is |
2473 | // because this entire 'fixup' is only to prevent crashes (from the |
2474 | // register scavenger) on not-really-valid inputs. For example, if we have: |
2475 | // %a = alloca i1 |
2476 | // %b = bitcast i1* %a to i64* |
2477 | // store i64* a, i64 b |
2478 | // then the store should really be marked as 'align 1', but is not. If it |
2479 | // were marked as 'align 1' then the indexed form would have been |
2480 | // instruction-selected initially, and the problem this 'fixup' is preventing |
2481 | // won't happen regardless. |
2482 | if (FrameIdx < 0) |
2483 | return; |
2484 | |
2485 | MachineFunction &MF = DAG.getMachineFunction(); |
2486 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
2487 | |
2488 | if (MFI.getObjectAlign(FrameIdx) >= Align(4)) |
2489 | return; |
2490 | |
2491 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
2492 | FuncInfo->setHasNonRISpills(); |
2493 | } |
2494 | |
2495 | /// Returns true if the address N can be represented by a base register plus |
2496 | /// a signed 16-bit displacement [r+imm], and if it is not better |
2497 | /// represented as reg+reg. If \p EncodingAlignment is non-zero, only accept |
2498 | /// displacements that are multiples of that value. |
2499 | bool PPCTargetLowering::SelectAddressRegImm( |
2500 | SDValue N, SDValue &Disp, SDValue &Base, SelectionDAG &DAG, |
2501 | MaybeAlign EncodingAlignment) const { |
2502 | // FIXME dl should come from parent load or store, not from address |
2503 | SDLoc dl(N); |
2504 | |
2505 | // If we have a PC Relative target flag don't select as [reg+imm]. It will be |
2506 | // a [pc+imm]. |
2507 | if (SelectAddressPCRel(N, Base)) |
2508 | return false; |
2509 | |
2510 | // If this can be more profitably realized as r+r, fail. |
2511 | if (SelectAddressRegReg(N, Disp, Base, DAG, EncodingAlignment)) |
2512 | return false; |
2513 | |
2514 | if (N.getOpcode() == ISD::ADD) { |
2515 | int16_t imm = 0; |
2516 | if (isIntS16Immediate(N.getOperand(1), imm) && |
2517 | (!EncodingAlignment || isAligned(*EncodingAlignment, imm))) { |
2518 | Disp = DAG.getTargetConstant(imm, dl, N.getValueType()); |
2519 | if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) { |
2520 | Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType()); |
2521 | fixupFuncForFI(DAG, FI->getIndex(), N.getValueType()); |
2522 | } else { |
2523 | Base = N.getOperand(0); |
2524 | } |
2525 | return true; // [r+i] |
2526 | } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) { |
2527 | // Match LOAD (ADD (X, Lo(G))). |
2528 | assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()((!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))-> getZExtValue() && "Cannot handle constant offsets yet!" ) ? static_cast<void> (0) : __assert_fail ("!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue() && \"Cannot handle constant offsets yet!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2529, __PRETTY_FUNCTION__)) |
2529 | && "Cannot handle constant offsets yet!")((!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))-> getZExtValue() && "Cannot handle constant offsets yet!" ) ? static_cast<void> (0) : __assert_fail ("!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue() && \"Cannot handle constant offsets yet!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2529, __PRETTY_FUNCTION__)); |
2530 | Disp = N.getOperand(1).getOperand(0); // The global address. |
2531 | assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||((Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode () == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD:: TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable ) ? static_cast<void> (0) : __assert_fail ("Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode() == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD::TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2534, __PRETTY_FUNCTION__)) |
2532 | Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||((Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode () == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD:: TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable ) ? static_cast<void> (0) : __assert_fail ("Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode() == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD::TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2534, __PRETTY_FUNCTION__)) |
2533 | Disp.getOpcode() == ISD::TargetConstantPool ||((Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode () == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD:: TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable ) ? static_cast<void> (0) : __assert_fail ("Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode() == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD::TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2534, __PRETTY_FUNCTION__)) |
2534 | Disp.getOpcode() == ISD::TargetJumpTable)((Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode () == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD:: TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable ) ? static_cast<void> (0) : __assert_fail ("Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode() == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD::TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2534, __PRETTY_FUNCTION__)); |
2535 | Base = N.getOperand(0); |
2536 | return true; // [&g+r] |
2537 | } |
2538 | } else if (N.getOpcode() == ISD::OR) { |
2539 | int16_t imm = 0; |
2540 | if (isIntS16Immediate(N.getOperand(1), imm) && |
2541 | (!EncodingAlignment || isAligned(*EncodingAlignment, imm))) { |
2542 | // If this is an or of disjoint bitfields, we can codegen this as an add |
2543 | // (for better address arithmetic) if the LHS and RHS of the OR are |
2544 | // provably disjoint. |
2545 | KnownBits LHSKnown = DAG.computeKnownBits(N.getOperand(0)); |
2546 | |
2547 | if ((LHSKnown.Zero.getZExtValue()|~(uint64_t)imm) == ~0ULL) { |
2548 | // If all of the bits are known zero on the LHS or RHS, the add won't |
2549 | // carry. |
2550 | if (FrameIndexSDNode *FI = |
2551 | dyn_cast<FrameIndexSDNode>(N.getOperand(0))) { |
2552 | Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType()); |
2553 | fixupFuncForFI(DAG, FI->getIndex(), N.getValueType()); |
2554 | } else { |
2555 | Base = N.getOperand(0); |
2556 | } |
2557 | Disp = DAG.getTargetConstant(imm, dl, N.getValueType()); |
2558 | return true; |
2559 | } |
2560 | } |
2561 | } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) { |
2562 | // Loading from a constant address. |
2563 | |
2564 | // If this address fits entirely in a 16-bit sext immediate field, codegen |
2565 | // this as "d, 0" |
2566 | int16_t Imm; |
2567 | if (isIntS16Immediate(CN, Imm) && |
2568 | (!EncodingAlignment || isAligned(*EncodingAlignment, Imm))) { |
2569 | Disp = DAG.getTargetConstant(Imm, dl, CN->getValueType(0)); |
2570 | Base = DAG.getRegister(Subtarget.isPPC64() ? PPC::ZERO8 : PPC::ZERO, |
2571 | CN->getValueType(0)); |
2572 | return true; |
2573 | } |
2574 | |
2575 | // Handle 32-bit sext immediates with LIS + addr mode. |
2576 | if ((CN->getValueType(0) == MVT::i32 || |
2577 | (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) && |
2578 | (!EncodingAlignment || |
2579 | isAligned(*EncodingAlignment, CN->getZExtValue()))) { |
2580 | int Addr = (int)CN->getZExtValue(); |
2581 | |
2582 | // Otherwise, break this down into an LIS + disp. |
2583 | Disp = DAG.getTargetConstant((short)Addr, dl, MVT::i32); |
2584 | |
2585 | Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, dl, |
2586 | MVT::i32); |
2587 | unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8; |
2588 | Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0); |
2589 | return true; |
2590 | } |
2591 | } |
2592 | |
2593 | Disp = DAG.getTargetConstant(0, dl, getPointerTy(DAG.getDataLayout())); |
2594 | if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N)) { |
2595 | Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType()); |
2596 | fixupFuncForFI(DAG, FI->getIndex(), N.getValueType()); |
2597 | } else |
2598 | Base = N; |
2599 | return true; // [r+0] |
2600 | } |
2601 | |
2602 | /// SelectAddressRegRegOnly - Given the specified addressed, force it to be |
2603 | /// represented as an indexed [r+r] operation. |
2604 | bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base, |
2605 | SDValue &Index, |
2606 | SelectionDAG &DAG) const { |
2607 | // Check to see if we can easily represent this as an [r+r] address. This |
2608 | // will fail if it thinks that the address is more profitably represented as |
2609 | // reg+imm, e.g. where imm = 0. |
2610 | if (SelectAddressRegReg(N, Base, Index, DAG)) |
2611 | return true; |
2612 | |
2613 | // If the address is the result of an add, we will utilize the fact that the |
2614 | // address calculation includes an implicit add. However, we can reduce |
2615 | // register pressure if we do not materialize a constant just for use as the |
2616 | // index register. We only get rid of the add if it is not an add of a |
2617 | // value and a 16-bit signed constant and both have a single use. |
2618 | int16_t imm = 0; |
2619 | if (N.getOpcode() == ISD::ADD && |
2620 | (!isIntS16Immediate(N.getOperand(1), imm) || |
2621 | !N.getOperand(1).hasOneUse() || !N.getOperand(0).hasOneUse())) { |
2622 | Base = N.getOperand(0); |
2623 | Index = N.getOperand(1); |
2624 | return true; |
2625 | } |
2626 | |
2627 | // Otherwise, do it the hard way, using R0 as the base register. |
2628 | Base = DAG.getRegister(Subtarget.isPPC64() ? PPC::ZERO8 : PPC::ZERO, |
2629 | N.getValueType()); |
2630 | Index = N; |
2631 | return true; |
2632 | } |
2633 | |
2634 | template <typename Ty> static bool isValidPCRelNode(SDValue N) { |
2635 | Ty *PCRelCand = dyn_cast<Ty>(N); |
2636 | return PCRelCand && (PCRelCand->getTargetFlags() & PPCII::MO_PCREL_FLAG); |
2637 | } |
2638 | |
2639 | /// Returns true if this address is a PC Relative address. |
2640 | /// PC Relative addresses are marked with the flag PPCII::MO_PCREL_FLAG |
2641 | /// or if the node opcode is PPCISD::MAT_PCREL_ADDR. |
2642 | bool PPCTargetLowering::SelectAddressPCRel(SDValue N, SDValue &Base) const { |
2643 | // This is a materialize PC Relative node. Always select this as PC Relative. |
2644 | Base = N; |
2645 | if (N.getOpcode() == PPCISD::MAT_PCREL_ADDR) |
2646 | return true; |
2647 | if (isValidPCRelNode<ConstantPoolSDNode>(N) || |
2648 | isValidPCRelNode<GlobalAddressSDNode>(N) || |
2649 | isValidPCRelNode<JumpTableSDNode>(N) || |
2650 | isValidPCRelNode<BlockAddressSDNode>(N)) |
2651 | return true; |
2652 | return false; |
2653 | } |
2654 | |
2655 | /// Returns true if we should use a direct load into vector instruction |
2656 | /// (such as lxsd or lfd), instead of a load into gpr + direct move sequence. |
2657 | static bool usePartialVectorLoads(SDNode *N, const PPCSubtarget& ST) { |
2658 | |
2659 | // If there are any other uses other than scalar to vector, then we should |
2660 | // keep it as a scalar load -> direct move pattern to prevent multiple |
2661 | // loads. |
2662 | LoadSDNode *LD = dyn_cast<LoadSDNode>(N); |
2663 | if (!LD) |
2664 | return false; |
2665 | |
2666 | EVT MemVT = LD->getMemoryVT(); |
2667 | if (!MemVT.isSimple()) |
2668 | return false; |
2669 | switch(MemVT.getSimpleVT().SimpleTy) { |
2670 | case MVT::i64: |
2671 | break; |
2672 | case MVT::i32: |
2673 | if (!ST.hasP8Vector()) |
2674 | return false; |
2675 | break; |
2676 | case MVT::i16: |
2677 | case MVT::i8: |
2678 | if (!ST.hasP9Vector()) |
2679 | return false; |
2680 | break; |
2681 | default: |
2682 | return false; |
2683 | } |
2684 | |
2685 | SDValue LoadedVal(N, 0); |
2686 | if (!LoadedVal.hasOneUse()) |
2687 | return false; |
2688 | |
2689 | for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end(); |
2690 | UI != UE; ++UI) |
2691 | if (UI.getUse().get().getResNo() == 0 && |
2692 | UI->getOpcode() != ISD::SCALAR_TO_VECTOR && |
2693 | UI->getOpcode() != PPCISD::SCALAR_TO_VECTOR_PERMUTED) |
2694 | return false; |
2695 | |
2696 | return true; |
2697 | } |
2698 | |
2699 | /// getPreIndexedAddressParts - returns true by value, base pointer and |
2700 | /// offset pointer and addressing mode by reference if the node's address |
2701 | /// can be legally represented as pre-indexed load / store address. |
2702 | bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base, |
2703 | SDValue &Offset, |
2704 | ISD::MemIndexedMode &AM, |
2705 | SelectionDAG &DAG) const { |
2706 | if (DisablePPCPreinc) return false; |
2707 | |
2708 | bool isLoad = true; |
2709 | SDValue Ptr; |
2710 | EVT VT; |
2711 | unsigned Alignment; |
2712 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { |
2713 | Ptr = LD->getBasePtr(); |
2714 | VT = LD->getMemoryVT(); |
2715 | Alignment = LD->getAlignment(); |
2716 | } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { |
2717 | Ptr = ST->getBasePtr(); |
2718 | VT = ST->getMemoryVT(); |
2719 | Alignment = ST->getAlignment(); |
2720 | isLoad = false; |
2721 | } else |
2722 | return false; |
2723 | |
2724 | // Do not generate pre-inc forms for specific loads that feed scalar_to_vector |
2725 | // instructions because we can fold these into a more efficient instruction |
2726 | // instead, (such as LXSD). |
2727 | if (isLoad && usePartialVectorLoads(N, Subtarget)) { |
2728 | return false; |
2729 | } |
2730 | |
2731 | // PowerPC doesn't have preinc load/store instructions for vectors |
2732 | if (VT.isVector()) |
2733 | return false; |
2734 | |
2735 | if (SelectAddressRegReg(Ptr, Base, Offset, DAG)) { |
2736 | // Common code will reject creating a pre-inc form if the base pointer |
2737 | // is a frame index, or if N is a store and the base pointer is either |
2738 | // the same as or a predecessor of the value being stored. Check for |
2739 | // those situations here, and try with swapped Base/Offset instead. |
2740 | bool Swap = false; |
2741 | |
2742 | if (isa<FrameIndexSDNode>(Base) || isa<RegisterSDNode>(Base)) |
2743 | Swap = true; |
2744 | else if (!isLoad) { |
2745 | SDValue Val = cast<StoreSDNode>(N)->getValue(); |
2746 | if (Val == Base || Base.getNode()->isPredecessorOf(Val.getNode())) |
2747 | Swap = true; |
2748 | } |
2749 | |
2750 | if (Swap) |
2751 | std::swap(Base, Offset); |
2752 | |
2753 | AM = ISD::PRE_INC; |
2754 | return true; |
2755 | } |
2756 | |
2757 | // LDU/STU can only handle immediates that are a multiple of 4. |
2758 | if (VT != MVT::i64) { |
2759 | if (!SelectAddressRegImm(Ptr, Offset, Base, DAG, None)) |
2760 | return false; |
2761 | } else { |
2762 | // LDU/STU need an address with at least 4-byte alignment. |
2763 | if (Alignment < 4) |
2764 | return false; |
2765 | |
2766 | if (!SelectAddressRegImm(Ptr, Offset, Base, DAG, Align(4))) |
2767 | return false; |
2768 | } |
2769 | |
2770 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { |
2771 | // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of |
2772 | // sext i32 to i64 when addr mode is r+i. |
2773 | if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 && |
2774 | LD->getExtensionType() == ISD::SEXTLOAD && |
2775 | isa<ConstantSDNode>(Offset)) |
2776 | return false; |
2777 | } |
2778 | |
2779 | AM = ISD::PRE_INC; |
2780 | return true; |
2781 | } |
2782 | |
2783 | //===----------------------------------------------------------------------===// |
2784 | // LowerOperation implementation |
2785 | //===----------------------------------------------------------------------===// |
2786 | |
2787 | /// Return true if we should reference labels using a PICBase, set the HiOpFlags |
2788 | /// and LoOpFlags to the target MO flags. |
2789 | static void getLabelAccessInfo(bool IsPIC, const PPCSubtarget &Subtarget, |
2790 | unsigned &HiOpFlags, unsigned &LoOpFlags, |
2791 | const GlobalValue *GV = nullptr) { |
2792 | HiOpFlags = PPCII::MO_HA; |
2793 | LoOpFlags = PPCII::MO_LO; |
2794 | |
2795 | // Don't use the pic base if not in PIC relocation model. |
2796 | if (IsPIC) { |
2797 | HiOpFlags |= PPCII::MO_PIC_FLAG; |
2798 | LoOpFlags |= PPCII::MO_PIC_FLAG; |
2799 | } |
2800 | } |
2801 | |
2802 | static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC, |
2803 | SelectionDAG &DAG) { |
2804 | SDLoc DL(HiPart); |
2805 | EVT PtrVT = HiPart.getValueType(); |
2806 | SDValue Zero = DAG.getConstant(0, DL, PtrVT); |
2807 | |
2808 | SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero); |
2809 | SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero); |
2810 | |
2811 | // With PIC, the first instruction is actually "GR+hi(&G)". |
2812 | if (isPIC) |
2813 | Hi = DAG.getNode(ISD::ADD, DL, PtrVT, |
2814 | DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi); |
2815 | |
2816 | // Generate non-pic code that has direct accesses to the constant pool. |
2817 | // The address of the global is just (hi(&g)+lo(&g)). |
2818 | return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo); |
2819 | } |
2820 | |
2821 | static void setUsesTOCBasePtr(MachineFunction &MF) { |
2822 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
2823 | FuncInfo->setUsesTOCBasePtr(); |
2824 | } |
2825 | |
2826 | static void setUsesTOCBasePtr(SelectionDAG &DAG) { |
2827 | setUsesTOCBasePtr(DAG.getMachineFunction()); |
2828 | } |
2829 | |
2830 | SDValue PPCTargetLowering::getTOCEntry(SelectionDAG &DAG, const SDLoc &dl, |
2831 | SDValue GA) const { |
2832 | const bool Is64Bit = Subtarget.isPPC64(); |
2833 | EVT VT = Is64Bit ? MVT::i64 : MVT::i32; |
2834 | SDValue Reg = Is64Bit ? DAG.getRegister(PPC::X2, VT) |
2835 | : Subtarget.isAIXABI() |
2836 | ? DAG.getRegister(PPC::R2, VT) |
2837 | : DAG.getNode(PPCISD::GlobalBaseReg, dl, VT); |
2838 | SDValue Ops[] = { GA, Reg }; |
2839 | return DAG.getMemIntrinsicNode( |
2840 | PPCISD::TOC_ENTRY, dl, DAG.getVTList(VT, MVT::Other), Ops, VT, |
2841 | MachinePointerInfo::getGOT(DAG.getMachineFunction()), None, |
2842 | MachineMemOperand::MOLoad); |
2843 | } |
2844 | |
2845 | SDValue PPCTargetLowering::LowerConstantPool(SDValue Op, |
2846 | SelectionDAG &DAG) const { |
2847 | EVT PtrVT = Op.getValueType(); |
2848 | ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); |
2849 | const Constant *C = CP->getConstVal(); |
2850 | |
2851 | // 64-bit SVR4 ABI and AIX ABI code are always position-independent. |
2852 | // The actual address of the GlobalValue is stored in the TOC. |
2853 | if (Subtarget.is64BitELFABI() || Subtarget.isAIXABI()) { |
2854 | if (Subtarget.isUsingPCRelativeCalls()) { |
2855 | SDLoc DL(CP); |
2856 | EVT Ty = getPointerTy(DAG.getDataLayout()); |
2857 | SDValue ConstPool = DAG.getTargetConstantPool( |
2858 | C, Ty, CP->getAlign(), CP->getOffset(), PPCII::MO_PCREL_FLAG); |
2859 | return DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, Ty, ConstPool); |
2860 | } |
2861 | setUsesTOCBasePtr(DAG); |
2862 | SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlign(), 0); |
2863 | return getTOCEntry(DAG, SDLoc(CP), GA); |
2864 | } |
2865 | |
2866 | unsigned MOHiFlag, MOLoFlag; |
2867 | bool IsPIC = isPositionIndependent(); |
2868 | getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag); |
2869 | |
2870 | if (IsPIC && Subtarget.isSVR4ABI()) { |
2871 | SDValue GA = |
2872 | DAG.getTargetConstantPool(C, PtrVT, CP->getAlign(), PPCII::MO_PIC_FLAG); |
2873 | return getTOCEntry(DAG, SDLoc(CP), GA); |
2874 | } |
2875 | |
2876 | SDValue CPIHi = |
2877 | DAG.getTargetConstantPool(C, PtrVT, CP->getAlign(), 0, MOHiFlag); |
2878 | SDValue CPILo = |
2879 | DAG.getTargetConstantPool(C, PtrVT, CP->getAlign(), 0, MOLoFlag); |
2880 | return LowerLabelRef(CPIHi, CPILo, IsPIC, DAG); |
2881 | } |
2882 | |
2883 | // For 64-bit PowerPC, prefer the more compact relative encodings. |
2884 | // This trades 32 bits per jump table entry for one or two instructions |
2885 | // on the jump site. |
2886 | unsigned PPCTargetLowering::getJumpTableEncoding() const { |
2887 | if (isJumpTableRelative()) |
2888 | return MachineJumpTableInfo::EK_LabelDifference32; |
2889 | |
2890 | return TargetLowering::getJumpTableEncoding(); |
2891 | } |
2892 | |
2893 | bool PPCTargetLowering::isJumpTableRelative() const { |
2894 | if (UseAbsoluteJumpTables) |
2895 | return false; |
2896 | if (Subtarget.isPPC64() || Subtarget.isAIXABI()) |
2897 | return true; |
2898 | return TargetLowering::isJumpTableRelative(); |
2899 | } |
2900 | |
2901 | SDValue PPCTargetLowering::getPICJumpTableRelocBase(SDValue Table, |
2902 | SelectionDAG &DAG) const { |
2903 | if (!Subtarget.isPPC64() || Subtarget.isAIXABI()) |
2904 | return TargetLowering::getPICJumpTableRelocBase(Table, DAG); |
2905 | |
2906 | switch (getTargetMachine().getCodeModel()) { |
2907 | case CodeModel::Small: |
2908 | case CodeModel::Medium: |
2909 | return TargetLowering::getPICJumpTableRelocBase(Table, DAG); |
2910 | default: |
2911 | return DAG.getNode(PPCISD::GlobalBaseReg, SDLoc(), |
2912 | getPointerTy(DAG.getDataLayout())); |
2913 | } |
2914 | } |
2915 | |
2916 | const MCExpr * |
2917 | PPCTargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF, |
2918 | unsigned JTI, |
2919 | MCContext &Ctx) const { |
2920 | if (!Subtarget.isPPC64() || Subtarget.isAIXABI()) |
2921 | return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx); |
2922 | |
2923 | switch (getTargetMachine().getCodeModel()) { |
2924 | case CodeModel::Small: |
2925 | case CodeModel::Medium: |
2926 | return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx); |
2927 | default: |
2928 | return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx); |
2929 | } |
2930 | } |
2931 | |
2932 | SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const { |
2933 | EVT PtrVT = Op.getValueType(); |
2934 | JumpTableSDNode *JT = cast<JumpTableSDNode>(Op); |
2935 | |
2936 | // isUsingPCRelativeCalls() returns true when PCRelative is enabled |
2937 | if (Subtarget.isUsingPCRelativeCalls()) { |
2938 | SDLoc DL(JT); |
2939 | EVT Ty = getPointerTy(DAG.getDataLayout()); |
2940 | SDValue GA = |
2941 | DAG.getTargetJumpTable(JT->getIndex(), Ty, PPCII::MO_PCREL_FLAG); |
2942 | SDValue MatAddr = DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, Ty, GA); |
2943 | return MatAddr; |
2944 | } |
2945 | |
2946 | // 64-bit SVR4 ABI and AIX ABI code are always position-independent. |
2947 | // The actual address of the GlobalValue is stored in the TOC. |
2948 | if (Subtarget.is64BitELFABI() || Subtarget.isAIXABI()) { |
2949 | setUsesTOCBasePtr(DAG); |
2950 | SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT); |
2951 | return getTOCEntry(DAG, SDLoc(JT), GA); |
2952 | } |
2953 | |
2954 | unsigned MOHiFlag, MOLoFlag; |
2955 | bool IsPIC = isPositionIndependent(); |
2956 | getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag); |
2957 | |
2958 | if (IsPIC && Subtarget.isSVR4ABI()) { |
2959 | SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, |
2960 | PPCII::MO_PIC_FLAG); |
2961 | return getTOCEntry(DAG, SDLoc(GA), GA); |
2962 | } |
2963 | |
2964 | SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag); |
2965 | SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag); |
2966 | return LowerLabelRef(JTIHi, JTILo, IsPIC, DAG); |
2967 | } |
2968 | |
2969 | SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op, |
2970 | SelectionDAG &DAG) const { |
2971 | EVT PtrVT = Op.getValueType(); |
2972 | BlockAddressSDNode *BASDN = cast<BlockAddressSDNode>(Op); |
2973 | const BlockAddress *BA = BASDN->getBlockAddress(); |
2974 | |
2975 | // isUsingPCRelativeCalls() returns true when PCRelative is enabled |
2976 | if (Subtarget.isUsingPCRelativeCalls()) { |
2977 | SDLoc DL(BASDN); |
2978 | EVT Ty = getPointerTy(DAG.getDataLayout()); |
2979 | SDValue GA = DAG.getTargetBlockAddress(BA, Ty, BASDN->getOffset(), |
2980 | PPCII::MO_PCREL_FLAG); |
2981 | SDValue MatAddr = DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, Ty, GA); |
2982 | return MatAddr; |
2983 | } |
2984 | |
2985 | // 64-bit SVR4 ABI and AIX ABI code are always position-independent. |
2986 | // The actual BlockAddress is stored in the TOC. |
2987 | if (Subtarget.is64BitELFABI() || Subtarget.isAIXABI()) { |
2988 | setUsesTOCBasePtr(DAG); |
2989 | SDValue GA = DAG.getTargetBlockAddress(BA, PtrVT, BASDN->getOffset()); |
2990 | return getTOCEntry(DAG, SDLoc(BASDN), GA); |
2991 | } |
2992 | |
2993 | // 32-bit position-independent ELF stores the BlockAddress in the .got. |
2994 | if (Subtarget.is32BitELFABI() && isPositionIndependent()) |
2995 | return getTOCEntry( |
2996 | DAG, SDLoc(BASDN), |
2997 | DAG.getTargetBlockAddress(BA, PtrVT, BASDN->getOffset())); |
2998 | |
2999 | unsigned MOHiFlag, MOLoFlag; |
3000 | bool IsPIC = isPositionIndependent(); |
3001 | getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag); |
3002 | SDValue TgtBAHi = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOHiFlag); |
3003 | SDValue TgtBALo = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOLoFlag); |
3004 | return LowerLabelRef(TgtBAHi, TgtBALo, IsPIC, DAG); |
3005 | } |
3006 | |
3007 | SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op, |
3008 | SelectionDAG &DAG) const { |
3009 | // FIXME: TLS addresses currently use medium model code sequences, |
3010 | // which is the most useful form. Eventually support for small and |
3011 | // large models could be added if users need it, at the cost of |
3012 | // additional complexity. |
3013 | GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); |
3014 | if (DAG.getTarget().useEmulatedTLS()) |
3015 | return LowerToTLSEmulatedModel(GA, DAG); |
3016 | |
3017 | SDLoc dl(GA); |
3018 | const GlobalValue *GV = GA->getGlobal(); |
3019 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
3020 | bool is64bit = Subtarget.isPPC64(); |
3021 | const Module *M = DAG.getMachineFunction().getFunction().getParent(); |
3022 | PICLevel::Level picLevel = M->getPICLevel(); |
3023 | |
3024 | const TargetMachine &TM = getTargetMachine(); |
3025 | TLSModel::Model Model = TM.getTLSModel(GV); |
3026 | |
3027 | if (Model == TLSModel::LocalExec) { |
3028 | if (Subtarget.isUsingPCRelativeCalls()) { |
3029 | SDValue TLSReg = DAG.getRegister(PPC::X13, MVT::i64); |
3030 | SDValue TGA = DAG.getTargetGlobalAddress( |
3031 | GV, dl, PtrVT, 0, (PPCII::MO_PCREL_FLAG | PPCII::MO_TPREL_FLAG)); |
3032 | SDValue MatAddr = |
3033 | DAG.getNode(PPCISD::TLS_LOCAL_EXEC_MAT_ADDR, dl, PtrVT, TGA); |
3034 | return DAG.getNode(PPCISD::ADD_TLS, dl, PtrVT, TLSReg, MatAddr); |
3035 | } |
3036 | |
3037 | SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, |
3038 | PPCII::MO_TPREL_HA); |
3039 | SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, |
3040 | PPCII::MO_TPREL_LO); |
3041 | SDValue TLSReg = is64bit ? DAG.getRegister(PPC::X13, MVT::i64) |
3042 | : DAG.getRegister(PPC::R2, MVT::i32); |
3043 | |
3044 | SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, TGAHi, TLSReg); |
3045 | return DAG.getNode(PPCISD::Lo, dl, PtrVT, TGALo, Hi); |
3046 | } |
3047 | |
3048 | if (Model == TLSModel::InitialExec) { |
3049 | bool IsPCRel = Subtarget.isUsingPCRelativeCalls(); |
3050 | SDValue TGA = DAG.getTargetGlobalAddress( |
3051 | GV, dl, PtrVT, 0, IsPCRel ? PPCII::MO_GOT_TPREL_PCREL_FLAG : 0); |
3052 | SDValue TGATLS = DAG.getTargetGlobalAddress( |
3053 | GV, dl, PtrVT, 0, |
3054 | IsPCRel ? (PPCII::MO_TLS | PPCII::MO_PCREL_FLAG) : PPCII::MO_TLS); |
3055 | SDValue TPOffset; |
3056 | if (IsPCRel) { |
3057 | SDValue MatPCRel = DAG.getNode(PPCISD::MAT_PCREL_ADDR, dl, PtrVT, TGA); |
3058 | TPOffset = DAG.getLoad(MVT::i64, dl, DAG.getEntryNode(), MatPCRel, |
3059 | MachinePointerInfo()); |
3060 | } else { |
3061 | SDValue GOTPtr; |
3062 | if (is64bit) { |
3063 | setUsesTOCBasePtr(DAG); |
3064 | SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64); |
3065 | GOTPtr = |
3066 | DAG.getNode(PPCISD::ADDIS_GOT_TPREL_HA, dl, PtrVT, GOTReg, TGA); |
3067 | } else { |
3068 | if (!TM.isPositionIndependent()) |
3069 | GOTPtr = DAG.getNode(PPCISD::PPC32_GOT, dl, PtrVT); |
3070 | else if (picLevel == PICLevel::SmallPIC) |
3071 | GOTPtr = DAG.getNode(PPCISD::GlobalBaseReg, dl, PtrVT); |
3072 | else |
3073 | GOTPtr = DAG.getNode(PPCISD::PPC32_PICGOT, dl, PtrVT); |
3074 | } |
3075 | TPOffset = DAG.getNode(PPCISD::LD_GOT_TPREL_L, dl, PtrVT, TGA, GOTPtr); |
3076 | } |
3077 | return DAG.getNode(PPCISD::ADD_TLS, dl, PtrVT, TPOffset, TGATLS); |
3078 | } |
3079 | |
3080 | if (Model == TLSModel::GeneralDynamic) { |
3081 | if (Subtarget.isUsingPCRelativeCalls()) { |
3082 | SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, |
3083 | PPCII::MO_GOT_TLSGD_PCREL_FLAG); |
3084 | return DAG.getNode(PPCISD::TLS_DYNAMIC_MAT_PCREL_ADDR, dl, PtrVT, TGA); |
3085 | } |
3086 | |
3087 | SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0); |
3088 | SDValue GOTPtr; |
3089 | if (is64bit) { |
3090 | setUsesTOCBasePtr(DAG); |
3091 | SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64); |
3092 | GOTPtr = DAG.getNode(PPCISD::ADDIS_TLSGD_HA, dl, PtrVT, |
3093 | GOTReg, TGA); |
3094 | } else { |
3095 | if (picLevel == PICLevel::SmallPIC) |
3096 | GOTPtr = DAG.getNode(PPCISD::GlobalBaseReg, dl, PtrVT); |
3097 | else |
3098 | GOTPtr = DAG.getNode(PPCISD::PPC32_PICGOT, dl, PtrVT); |
3099 | } |
3100 | return DAG.getNode(PPCISD::ADDI_TLSGD_L_ADDR, dl, PtrVT, |
3101 | GOTPtr, TGA, TGA); |
3102 | } |
3103 | |
3104 | if (Model == TLSModel::LocalDynamic) { |
3105 | if (Subtarget.isUsingPCRelativeCalls()) { |
3106 | SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, |
3107 | PPCII::MO_GOT_TLSLD_PCREL_FLAG); |
3108 | SDValue MatPCRel = |
3109 | DAG.getNode(PPCISD::TLS_DYNAMIC_MAT_PCREL_ADDR, dl, PtrVT, TGA); |
3110 | return DAG.getNode(PPCISD::PADDI_DTPREL, dl, PtrVT, MatPCRel, TGA); |
3111 | } |
3112 | |
3113 | SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0); |
3114 | SDValue GOTPtr; |
3115 | if (is64bit) { |
3116 | setUsesTOCBasePtr(DAG); |
3117 | SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64); |
3118 | GOTPtr = DAG.getNode(PPCISD::ADDIS_TLSLD_HA, dl, PtrVT, |
3119 | GOTReg, TGA); |
3120 | } else { |
3121 | if (picLevel == PICLevel::SmallPIC) |
3122 | GOTPtr = DAG.getNode(PPCISD::GlobalBaseReg, dl, PtrVT); |
3123 | else |
3124 | GOTPtr = DAG.getNode(PPCISD::PPC32_PICGOT, dl, PtrVT); |
3125 | } |
3126 | SDValue TLSAddr = DAG.getNode(PPCISD::ADDI_TLSLD_L_ADDR, dl, |
3127 | PtrVT, GOTPtr, TGA, TGA); |
3128 | SDValue DtvOffsetHi = DAG.getNode(PPCISD::ADDIS_DTPREL_HA, dl, |
3129 | PtrVT, TLSAddr, TGA); |
3130 | return DAG.getNode(PPCISD::ADDI_DTPREL_L, dl, PtrVT, DtvOffsetHi, TGA); |
3131 | } |
3132 | |
3133 | llvm_unreachable("Unknown TLS model!")::llvm::llvm_unreachable_internal("Unknown TLS model!", "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3133); |
3134 | } |
3135 | |
3136 | SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op, |
3137 | SelectionDAG &DAG) const { |
3138 | EVT PtrVT = Op.getValueType(); |
3139 | GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op); |
3140 | SDLoc DL(GSDN); |
3141 | const GlobalValue *GV = GSDN->getGlobal(); |
3142 | |
3143 | // 64-bit SVR4 ABI & AIX ABI code is always position-independent. |
3144 | // The actual address of the GlobalValue is stored in the TOC. |
3145 | if (Subtarget.is64BitELFABI() || Subtarget.isAIXABI()) { |
3146 | if (Subtarget.isUsingPCRelativeCalls()) { |
3147 | EVT Ty = getPointerTy(DAG.getDataLayout()); |
3148 | if (isAccessedAsGotIndirect(Op)) { |
3149 | SDValue GA = DAG.getTargetGlobalAddress(GV, DL, Ty, GSDN->getOffset(), |
3150 | PPCII::MO_PCREL_FLAG | |
3151 | PPCII::MO_GOT_FLAG); |
3152 | SDValue MatPCRel = DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, Ty, GA); |
3153 | SDValue Load = DAG.getLoad(MVT::i64, DL, DAG.getEntryNode(), MatPCRel, |
3154 | MachinePointerInfo()); |
3155 | return Load; |
3156 | } else { |
3157 | SDValue GA = DAG.getTargetGlobalAddress(GV, DL, Ty, GSDN->getOffset(), |
3158 | PPCII::MO_PCREL_FLAG); |
3159 | return DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, Ty, GA); |
3160 | } |
3161 | } |
3162 | setUsesTOCBasePtr(DAG); |
3163 | SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset()); |
3164 | return getTOCEntry(DAG, DL, GA); |
3165 | } |
3166 | |
3167 | unsigned MOHiFlag, MOLoFlag; |
3168 | bool IsPIC = isPositionIndependent(); |
3169 | getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag, GV); |
3170 | |
3171 | if (IsPIC && Subtarget.isSVR4ABI()) { |
3172 | SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, |
3173 | GSDN->getOffset(), |
3174 | PPCII::MO_PIC_FLAG); |
3175 | return getTOCEntry(DAG, DL, GA); |
3176 | } |
3177 | |
3178 | SDValue GAHi = |
3179 | DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag); |
3180 | SDValue GALo = |
3181 | DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag); |
3182 | |
3183 | return LowerLabelRef(GAHi, GALo, IsPIC, DAG); |
3184 | } |
3185 | |
3186 | SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const { |
3187 | ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); |
3188 | SDLoc dl(Op); |
3189 | |
3190 | if (Op.getValueType() == MVT::v2i64) { |
3191 | // When the operands themselves are v2i64 values, we need to do something |
3192 | // special because VSX has no underlying comparison operations for these. |
3193 | if (Op.getOperand(0).getValueType() == MVT::v2i64) { |
3194 | // Equality can be handled by casting to the legal type for Altivec |
3195 | // comparisons, everything else needs to be expanded. |
3196 | if (CC == ISD::SETEQ || CC == ISD::SETNE) { |
3197 | return DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, |
3198 | DAG.getSetCC(dl, MVT::v4i32, |
3199 | DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op.getOperand(0)), |
3200 | DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op.getOperand(1)), |
3201 | CC)); |
3202 | } |
3203 | |
3204 | return SDValue(); |
3205 | } |
3206 | |
3207 | // We handle most of these in the usual way. |
3208 | return Op; |
3209 | } |
3210 | |
3211 | // If we're comparing for equality to zero, expose the fact that this is |
3212 | // implemented as a ctlz/srl pair on ppc, so that the dag combiner can |
3213 | // fold the new nodes. |
3214 | if (SDValue V = lowerCmpEqZeroToCtlzSrl(Op, DAG)) |
3215 | return V; |
3216 | |
3217 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { |
3218 | // Leave comparisons against 0 and -1 alone for now, since they're usually |
3219 | // optimized. FIXME: revisit this when we can custom lower all setcc |
3220 | // optimizations. |
3221 | if (C->isAllOnesValue() || C->isNullValue()) |
3222 | return SDValue(); |
3223 | } |
3224 | |
3225 | // If we have an integer seteq/setne, turn it into a compare against zero |
3226 | // by xor'ing the rhs with the lhs, which is faster than setting a |
3227 | // condition register, reading it back out, and masking the correct bit. The |
3228 | // normal approach here uses sub to do this instead of xor. Using xor exposes |
3229 | // the result to other bit-twiddling opportunities. |
3230 | EVT LHSVT = Op.getOperand(0).getValueType(); |
3231 | if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) { |
3232 | EVT VT = Op.getValueType(); |
3233 | SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0), |
3234 | Op.getOperand(1)); |
3235 | return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, dl, LHSVT), CC); |
3236 | } |
3237 | return SDValue(); |
3238 | } |
3239 | |
3240 | SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const { |
3241 | SDNode *Node = Op.getNode(); |
3242 | EVT VT = Node->getValueType(0); |
3243 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
3244 | SDValue InChain = Node->getOperand(0); |
3245 | SDValue VAListPtr = Node->getOperand(1); |
3246 | const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue(); |
3247 | SDLoc dl(Node); |
3248 | |
3249 | assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only")((!Subtarget.isPPC64() && "LowerVAARG is PPC32 only") ? static_cast<void> (0) : __assert_fail ("!Subtarget.isPPC64() && \"LowerVAARG is PPC32 only\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3249, __PRETTY_FUNCTION__)); |
3250 | |
3251 | // gpr_index |
3252 | SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain, |
3253 | VAListPtr, MachinePointerInfo(SV), MVT::i8); |
3254 | InChain = GprIndex.getValue(1); |
3255 | |
3256 | if (VT == MVT::i64) { |
3257 | // Check if GprIndex is even |
3258 | SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex, |
3259 | DAG.getConstant(1, dl, MVT::i32)); |
3260 | SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd, |
3261 | DAG.getConstant(0, dl, MVT::i32), ISD::SETNE); |
3262 | SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex, |
3263 | DAG.getConstant(1, dl, MVT::i32)); |
3264 | // Align GprIndex to be even if it isn't |
3265 | GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne, |
3266 | GprIndex); |
3267 | } |
3268 | |
3269 | // fpr index is 1 byte after gpr |
3270 | SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr, |
3271 | DAG.getConstant(1, dl, MVT::i32)); |
3272 | |
3273 | // fpr |
3274 | SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain, |
3275 | FprPtr, MachinePointerInfo(SV), MVT::i8); |
3276 | InChain = FprIndex.getValue(1); |
3277 | |
3278 | SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr, |
3279 | DAG.getConstant(8, dl, MVT::i32)); |
3280 | |
3281 | SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr, |
3282 | DAG.getConstant(4, dl, MVT::i32)); |
3283 | |
3284 | // areas |
3285 | SDValue OverflowArea = |
3286 | DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr, MachinePointerInfo()); |
3287 | InChain = OverflowArea.getValue(1); |
3288 | |
3289 | SDValue RegSaveArea = |
3290 | DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr, MachinePointerInfo()); |
3291 | InChain = RegSaveArea.getValue(1); |
3292 | |
3293 | // select overflow_area if index > 8 |
3294 | SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex, |
3295 | DAG.getConstant(8, dl, MVT::i32), ISD::SETLT); |
3296 | |
3297 | // adjustment constant gpr_index * 4/8 |
3298 | SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32, |
3299 | VT.isInteger() ? GprIndex : FprIndex, |
3300 | DAG.getConstant(VT.isInteger() ? 4 : 8, dl, |
3301 | MVT::i32)); |
3302 | |
3303 | // OurReg = RegSaveArea + RegConstant |
3304 | SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea, |
3305 | RegConstant); |
3306 | |
3307 | // Floating types are 32 bytes into RegSaveArea |
3308 | if (VT.isFloatingPoint()) |
3309 | OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg, |
3310 | DAG.getConstant(32, dl, MVT::i32)); |
3311 | |
3312 | // increase {f,g}pr_index by 1 (or 2 if VT is i64) |
3313 | SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32, |
3314 | VT.isInteger() ? GprIndex : FprIndex, |
3315 | DAG.getConstant(VT == MVT::i64 ? 2 : 1, dl, |
3316 | MVT::i32)); |
3317 | |
3318 | InChain = DAG.getTruncStore(InChain, dl, IndexPlus1, |
3319 | VT.isInteger() ? VAListPtr : FprPtr, |
3320 | MachinePointerInfo(SV), MVT::i8); |
3321 | |
3322 | // determine if we should load from reg_save_area or overflow_area |
3323 | SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea); |
3324 | |
3325 | // increase overflow_area by 4/8 if gpr/fpr > 8 |
3326 | SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea, |
3327 | DAG.getConstant(VT.isInteger() ? 4 : 8, |
3328 | dl, MVT::i32)); |
3329 | |
3330 | OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea, |
3331 | OverflowAreaPlusN); |
3332 | |
3333 | InChain = DAG.getTruncStore(InChain, dl, OverflowArea, OverflowAreaPtr, |
3334 | MachinePointerInfo(), MVT::i32); |
3335 | |
3336 | return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo()); |
3337 | } |
3338 | |
3339 | SDValue PPCTargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const { |
3340 | assert(!Subtarget.isPPC64() && "LowerVACOPY is PPC32 only")((!Subtarget.isPPC64() && "LowerVACOPY is PPC32 only" ) ? static_cast<void> (0) : __assert_fail ("!Subtarget.isPPC64() && \"LowerVACOPY is PPC32 only\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3340, __PRETTY_FUNCTION__)); |
3341 | |
3342 | // We have to copy the entire va_list struct: |
3343 | // 2*sizeof(char) + 2 Byte alignment + 2*sizeof(char*) = 12 Byte |
3344 | return DAG.getMemcpy(Op.getOperand(0), Op, Op.getOperand(1), Op.getOperand(2), |
3345 | DAG.getConstant(12, SDLoc(Op), MVT::i32), Align(8), |
3346 | false, true, false, MachinePointerInfo(), |
3347 | MachinePointerInfo()); |
3348 | } |
3349 | |
3350 | SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op, |
3351 | SelectionDAG &DAG) const { |
3352 | if (Subtarget.isAIXABI()) |
3353 | report_fatal_error("ADJUST_TRAMPOLINE operation is not supported on AIX."); |
3354 | |
3355 | return Op.getOperand(0); |
3356 | } |
3357 | |
3358 | SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op, |
3359 | SelectionDAG &DAG) const { |
3360 | if (Subtarget.isAIXABI()) |
3361 | report_fatal_error("INIT_TRAMPOLINE operation is not supported on AIX."); |
3362 | |
3363 | SDValue Chain = Op.getOperand(0); |
3364 | SDValue Trmp = Op.getOperand(1); // trampoline |
3365 | SDValue FPtr = Op.getOperand(2); // nested function |
3366 | SDValue Nest = Op.getOperand(3); // 'nest' parameter value |
3367 | SDLoc dl(Op); |
3368 | |
3369 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
3370 | bool isPPC64 = (PtrVT == MVT::i64); |
3371 | Type *IntPtrTy = DAG.getDataLayout().getIntPtrType(*DAG.getContext()); |
3372 | |
3373 | TargetLowering::ArgListTy Args; |
3374 | TargetLowering::ArgListEntry Entry; |
3375 | |
3376 | Entry.Ty = IntPtrTy; |
3377 | Entry.Node = Trmp; Args.push_back(Entry); |
3378 | |
3379 | // TrampSize == (isPPC64 ? 48 : 40); |
3380 | Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40, dl, |
3381 | isPPC64 ? MVT::i64 : MVT::i32); |
3382 | Args.push_back(Entry); |
3383 | |
3384 | Entry.Node = FPtr; Args.push_back(Entry); |
3385 | Entry.Node = Nest; Args.push_back(Entry); |
3386 | |
3387 | // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg) |
3388 | TargetLowering::CallLoweringInfo CLI(DAG); |
3389 | CLI.setDebugLoc(dl).setChain(Chain).setLibCallee( |
3390 | CallingConv::C, Type::getVoidTy(*DAG.getContext()), |
3391 | DAG.getExternalSymbol("__trampoline_setup", PtrVT), std::move(Args)); |
3392 | |
3393 | std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); |
3394 | return CallResult.second; |
3395 | } |
3396 | |
3397 | SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const { |
3398 | MachineFunction &MF = DAG.getMachineFunction(); |
3399 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
3400 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
3401 | |
3402 | SDLoc dl(Op); |
3403 | |
3404 | if (Subtarget.isPPC64() || Subtarget.isAIXABI()) { |
3405 | // vastart just stores the address of the VarArgsFrameIndex slot into the |
3406 | // memory location argument. |
3407 | SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); |
3408 | const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); |
3409 | return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), |
3410 | MachinePointerInfo(SV)); |
3411 | } |
3412 | |
3413 | // For the 32-bit SVR4 ABI we follow the layout of the va_list struct. |
3414 | // We suppose the given va_list is already allocated. |
3415 | // |
3416 | // typedef struct { |
3417 | // char gpr; /* index into the array of 8 GPRs |
3418 | // * stored in the register save area |
3419 | // * gpr=0 corresponds to r3, |
3420 | // * gpr=1 to r4, etc. |
3421 | // */ |
3422 | // char fpr; /* index into the array of 8 FPRs |
3423 | // * stored in the register save area |
3424 | // * fpr=0 corresponds to f1, |
3425 | // * fpr=1 to f2, etc. |
3426 | // */ |
3427 | // char *overflow_arg_area; |
3428 | // /* location on stack that holds |
3429 | // * the next overflow argument |
3430 | // */ |
3431 | // char *reg_save_area; |
3432 | // /* where r3:r10 and f1:f8 (if saved) |
3433 | // * are stored |
3434 | // */ |
3435 | // } va_list[1]; |
3436 | |
3437 | SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), dl, MVT::i32); |
3438 | SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), dl, MVT::i32); |
3439 | SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(), |
3440 | PtrVT); |
3441 | SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), |
3442 | PtrVT); |
3443 | |
3444 | uint64_t FrameOffset = PtrVT.getSizeInBits()/8; |
3445 | SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, dl, PtrVT); |
3446 | |
3447 | uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1; |
3448 | SDValue ConstStackOffset = DAG.getConstant(StackOffset, dl, PtrVT); |
3449 | |
3450 | uint64_t FPROffset = 1; |
3451 | SDValue ConstFPROffset = DAG.getConstant(FPROffset, dl, PtrVT); |
3452 | |
3453 | const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); |
3454 | |
3455 | // Store first byte : number of int regs |
3456 | SDValue firstStore = |
3457 | DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR, Op.getOperand(1), |
3458 | MachinePointerInfo(SV), MVT::i8); |
3459 | uint64_t nextOffset = FPROffset; |
3460 | SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1), |
3461 | ConstFPROffset); |
3462 | |
3463 | // Store second byte : number of float regs |
3464 | SDValue secondStore = |
3465 | DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr, |
3466 | MachinePointerInfo(SV, nextOffset), MVT::i8); |
3467 | nextOffset += StackOffset; |
3468 | nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset); |
3469 | |
3470 | // Store second word : arguments given on stack |
3471 | SDValue thirdStore = DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr, |
3472 | MachinePointerInfo(SV, nextOffset)); |
3473 | nextOffset += FrameOffset; |
3474 | nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset); |
3475 | |
3476 | // Store third word : arguments given in registers |
3477 | return DAG.getStore(thirdStore, dl, FR, nextPtr, |
3478 | MachinePointerInfo(SV, nextOffset)); |
3479 | } |
3480 | |
3481 | /// FPR - The set of FP registers that should be allocated for arguments |
3482 | /// on Darwin and AIX. |
3483 | static const MCPhysReg FPR[] = {PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, |
3484 | PPC::F6, PPC::F7, PPC::F8, PPC::F9, PPC::F10, |
3485 | PPC::F11, PPC::F12, PPC::F13}; |
3486 | |
3487 | /// CalculateStackSlotSize - Calculates the size reserved for this argument on |
3488 | /// the stack. |
3489 | static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags, |
3490 | unsigned PtrByteSize) { |
3491 | unsigned ArgSize = ArgVT.getStoreSize(); |
3492 | if (Flags.isByVal()) |
3493 | ArgSize = Flags.getByValSize(); |
3494 | |
3495 | // Round up to multiples of the pointer size, except for array members, |
3496 | // which are always packed. |
3497 | if (!Flags.isInConsecutiveRegs()) |
3498 | ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
3499 | |
3500 | return ArgSize; |
3501 | } |
3502 | |
3503 | /// CalculateStackSlotAlignment - Calculates the alignment of this argument |
3504 | /// on the stack. |
3505 | static Align CalculateStackSlotAlignment(EVT ArgVT, EVT OrigVT, |
3506 | ISD::ArgFlagsTy Flags, |
3507 | unsigned PtrByteSize) { |
3508 | Align Alignment(PtrByteSize); |
3509 | |
3510 | // Altivec parameters are padded to a 16 byte boundary. |
3511 | if (ArgVT == MVT::v4f32 || ArgVT == MVT::v4i32 || |
3512 | ArgVT == MVT::v8i16 || ArgVT == MVT::v16i8 || |
3513 | ArgVT == MVT::v2f64 || ArgVT == MVT::v2i64 || |
3514 | ArgVT == MVT::v1i128 || ArgVT == MVT::f128) |
3515 | Alignment = Align(16); |
3516 | |
3517 | // ByVal parameters are aligned as requested. |
3518 | if (Flags.isByVal()) { |
3519 | auto BVAlign = Flags.getNonZeroByValAlign(); |
3520 | if (BVAlign > PtrByteSize) { |
3521 | if (BVAlign.value() % PtrByteSize != 0) |
3522 | llvm_unreachable(::llvm::llvm_unreachable_internal("ByVal alignment is not a multiple of the pointer size" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3523) |
3523 | "ByVal alignment is not a multiple of the pointer size")::llvm::llvm_unreachable_internal("ByVal alignment is not a multiple of the pointer size" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3523); |
3524 | |
3525 | Alignment = BVAlign; |
3526 | } |
3527 | } |
3528 | |
3529 | // Array members are always packed to their original alignment. |
3530 | if (Flags.isInConsecutiveRegs()) { |
3531 | // If the array member was split into multiple registers, the first |
3532 | // needs to be aligned to the size of the full type. (Except for |
3533 | // ppcf128, which is only aligned as its f64 components.) |
3534 | if (Flags.isSplit() && OrigVT != MVT::ppcf128) |
3535 | Alignment = Align(OrigVT.getStoreSize()); |
3536 | else |
3537 | Alignment = Align(ArgVT.getStoreSize()); |
3538 | } |
3539 | |
3540 | return Alignment; |
3541 | } |
3542 | |
3543 | /// CalculateStackSlotUsed - Return whether this argument will use its |
3544 | /// stack slot (instead of being passed in registers). ArgOffset, |
3545 | /// AvailableFPRs, and AvailableVRs must hold the current argument |
3546 | /// position, and will be updated to account for this argument. |
3547 | static bool CalculateStackSlotUsed(EVT ArgVT, EVT OrigVT, ISD::ArgFlagsTy Flags, |
3548 | unsigned PtrByteSize, unsigned LinkageSize, |
3549 | unsigned ParamAreaSize, unsigned &ArgOffset, |
3550 | unsigned &AvailableFPRs, |
3551 | unsigned &AvailableVRs) { |
3552 | bool UseMemory = false; |
3553 | |
3554 | // Respect alignment of argument on the stack. |
3555 | Align Alignment = |
3556 | CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize); |
3557 | ArgOffset = alignTo(ArgOffset, Alignment); |
3558 | // If there's no space left in the argument save area, we must |
3559 | // use memory (this check also catches zero-sized arguments). |
3560 | if (ArgOffset >= LinkageSize + ParamAreaSize) |
3561 | UseMemory = true; |
3562 | |
3563 | // Allocate argument on the stack. |
3564 | ArgOffset += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize); |
3565 | if (Flags.isInConsecutiveRegsLast()) |
3566 | ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
3567 | // If we overran the argument save area, we must use memory |
3568 | // (this check catches arguments passed partially in memory) |
3569 | if (ArgOffset > LinkageSize + ParamAreaSize) |
3570 | UseMemory = true; |
3571 | |
3572 | // However, if the argument is actually passed in an FPR or a VR, |
3573 | // we don't use memory after all. |
3574 | if (!Flags.isByVal()) { |
3575 | if (ArgVT == MVT::f32 || ArgVT == MVT::f64) |
3576 | if (AvailableFPRs > 0) { |
3577 | --AvailableFPRs; |
3578 | return false; |
3579 | } |
3580 | if (ArgVT == MVT::v4f32 || ArgVT == MVT::v4i32 || |
3581 | ArgVT == MVT::v8i16 || ArgVT == MVT::v16i8 || |
3582 | ArgVT == MVT::v2f64 || ArgVT == MVT::v2i64 || |
3583 | ArgVT == MVT::v1i128 || ArgVT == MVT::f128) |
3584 | if (AvailableVRs > 0) { |
3585 | --AvailableVRs; |
3586 | return false; |
3587 | } |
3588 | } |
3589 | |
3590 | return UseMemory; |
3591 | } |
3592 | |
3593 | /// EnsureStackAlignment - Round stack frame size up from NumBytes to |
3594 | /// ensure minimum alignment required for target. |
3595 | static unsigned EnsureStackAlignment(const PPCFrameLowering *Lowering, |
3596 | unsigned NumBytes) { |
3597 | return alignTo(NumBytes, Lowering->getStackAlign()); |
3598 | } |
3599 | |
3600 | SDValue PPCTargetLowering::LowerFormalArguments( |
3601 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
3602 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
3603 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
3604 | if (Subtarget.isAIXABI()) |
3605 | return LowerFormalArguments_AIX(Chain, CallConv, isVarArg, Ins, dl, DAG, |
3606 | InVals); |
3607 | if (Subtarget.is64BitELFABI()) |
3608 | return LowerFormalArguments_64SVR4(Chain, CallConv, isVarArg, Ins, dl, DAG, |
3609 | InVals); |
3610 | if (Subtarget.is32BitELFABI()) |
3611 | return LowerFormalArguments_32SVR4(Chain, CallConv, isVarArg, Ins, dl, DAG, |
3612 | InVals); |
3613 | |
3614 | return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins, dl, DAG, |
3615 | InVals); |
3616 | } |
3617 | |
3618 | SDValue PPCTargetLowering::LowerFormalArguments_32SVR4( |
3619 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
3620 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
3621 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
3622 | |
3623 | // 32-bit SVR4 ABI Stack Frame Layout: |
3624 | // +-----------------------------------+ |
3625 | // +--> | Back chain | |
3626 | // | +-----------------------------------+ |
3627 | // | | Floating-point register save area | |
3628 | // | +-----------------------------------+ |
3629 | // | | General register save area | |
3630 | // | +-----------------------------------+ |
3631 | // | | CR save word | |
3632 | // | +-----------------------------------+ |
3633 | // | | VRSAVE save word | |
3634 | // | +-----------------------------------+ |
3635 | // | | Alignment padding | |
3636 | // | +-----------------------------------+ |
3637 | // | | Vector register save area | |
3638 | // | +-----------------------------------+ |
3639 | // | | Local variable space | |
3640 | // | +-----------------------------------+ |
3641 | // | | Parameter list area | |
3642 | // | +-----------------------------------+ |
3643 | // | | LR save word | |
3644 | // | +-----------------------------------+ |
3645 | // SP--> +--- | Back chain | |
3646 | // +-----------------------------------+ |
3647 | // |
3648 | // Specifications: |
3649 | // System V Application Binary Interface PowerPC Processor Supplement |
3650 | // AltiVec Technology Programming Interface Manual |
3651 | |
3652 | MachineFunction &MF = DAG.getMachineFunction(); |
3653 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
3654 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
3655 | |
3656 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
3657 | // Potential tail calls could cause overwriting of argument stack slots. |
3658 | bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt && |
3659 | (CallConv == CallingConv::Fast)); |
3660 | const Align PtrAlign(4); |
3661 | |
3662 | // Assign locations to all of the incoming arguments. |
3663 | SmallVector<CCValAssign, 16> ArgLocs; |
3664 | PPCCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs, |
3665 | *DAG.getContext()); |
3666 | |
3667 | // Reserve space for the linkage area on the stack. |
3668 | unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
3669 | CCInfo.AllocateStack(LinkageSize, PtrAlign); |
3670 | if (useSoftFloat()) |
3671 | CCInfo.PreAnalyzeFormalArguments(Ins); |
3672 | |
3673 | CCInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4); |
3674 | CCInfo.clearWasPPCF128(); |
3675 | |
3676 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { |
3677 | CCValAssign &VA = ArgLocs[i]; |
3678 | |
3679 | // Arguments stored in registers. |
3680 | if (VA.isRegLoc()) { |
3681 | const TargetRegisterClass *RC; |
3682 | EVT ValVT = VA.getValVT(); |
3683 | |
3684 | switch (ValVT.getSimpleVT().SimpleTy) { |
3685 | default: |
3686 | llvm_unreachable("ValVT not supported by formal arguments Lowering")::llvm::llvm_unreachable_internal("ValVT not supported by formal arguments Lowering" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3686); |
3687 | case MVT::i1: |
3688 | case MVT::i32: |
3689 | RC = &PPC::GPRCRegClass; |
3690 | break; |
3691 | case MVT::f32: |
3692 | if (Subtarget.hasP8Vector()) |
3693 | RC = &PPC::VSSRCRegClass; |
3694 | else if (Subtarget.hasSPE()) |
3695 | RC = &PPC::GPRCRegClass; |
3696 | else |
3697 | RC = &PPC::F4RCRegClass; |
3698 | break; |
3699 | case MVT::f64: |
3700 | if (Subtarget.hasVSX()) |
3701 | RC = &PPC::VSFRCRegClass; |
3702 | else if (Subtarget.hasSPE()) |
3703 | // SPE passes doubles in GPR pairs. |
3704 | RC = &PPC::GPRCRegClass; |
3705 | else |
3706 | RC = &PPC::F8RCRegClass; |
3707 | break; |
3708 | case MVT::v16i8: |
3709 | case MVT::v8i16: |
3710 | case MVT::v4i32: |
3711 | RC = &PPC::VRRCRegClass; |
3712 | break; |
3713 | case MVT::v4f32: |
3714 | RC = &PPC::VRRCRegClass; |
3715 | break; |
3716 | case MVT::v2f64: |
3717 | case MVT::v2i64: |
3718 | RC = &PPC::VRRCRegClass; |
3719 | break; |
3720 | } |
3721 | |
3722 | SDValue ArgValue; |
3723 | // Transform the arguments stored in physical registers into |
3724 | // virtual ones. |
3725 | if (VA.getLocVT() == MVT::f64 && Subtarget.hasSPE()) { |
3726 | assert(i + 1 < e && "No second half of double precision argument")((i + 1 < e && "No second half of double precision argument" ) ? static_cast<void> (0) : __assert_fail ("i + 1 < e && \"No second half of double precision argument\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3726, __PRETTY_FUNCTION__)); |
3727 | unsigned RegLo = MF.addLiveIn(VA.getLocReg(), RC); |
3728 | unsigned RegHi = MF.addLiveIn(ArgLocs[++i].getLocReg(), RC); |
3729 | SDValue ArgValueLo = DAG.getCopyFromReg(Chain, dl, RegLo, MVT::i32); |
3730 | SDValue ArgValueHi = DAG.getCopyFromReg(Chain, dl, RegHi, MVT::i32); |
3731 | if (!Subtarget.isLittleEndian()) |
3732 | std::swap (ArgValueLo, ArgValueHi); |
3733 | ArgValue = DAG.getNode(PPCISD::BUILD_SPE64, dl, MVT::f64, ArgValueLo, |
3734 | ArgValueHi); |
3735 | } else { |
3736 | unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); |
3737 | ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, |
3738 | ValVT == MVT::i1 ? MVT::i32 : ValVT); |
3739 | if (ValVT == MVT::i1) |
3740 | ArgValue = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, ArgValue); |
3741 | } |
3742 | |
3743 | InVals.push_back(ArgValue); |
3744 | } else { |
3745 | // Argument stored in memory. |
3746 | assert(VA.isMemLoc())((VA.isMemLoc()) ? static_cast<void> (0) : __assert_fail ("VA.isMemLoc()", "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3746, __PRETTY_FUNCTION__)); |
3747 | |
3748 | // Get the extended size of the argument type in stack |
3749 | unsigned ArgSize = VA.getLocVT().getStoreSize(); |
3750 | // Get the actual size of the argument type |
3751 | unsigned ObjSize = VA.getValVT().getStoreSize(); |
3752 | unsigned ArgOffset = VA.getLocMemOffset(); |
3753 | // Stack objects in PPC32 are right justified. |
3754 | ArgOffset += ArgSize - ObjSize; |
3755 | int FI = MFI.CreateFixedObject(ArgSize, ArgOffset, isImmutable); |
3756 | |
3757 | // Create load nodes to retrieve arguments from the stack. |
3758 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
3759 | InVals.push_back( |
3760 | DAG.getLoad(VA.getValVT(), dl, Chain, FIN, MachinePointerInfo())); |
3761 | } |
3762 | } |
3763 | |
3764 | // Assign locations to all of the incoming aggregate by value arguments. |
3765 | // Aggregates passed by value are stored in the local variable space of the |
3766 | // caller's stack frame, right above the parameter list area. |
3767 | SmallVector<CCValAssign, 16> ByValArgLocs; |
3768 | CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(), |
3769 | ByValArgLocs, *DAG.getContext()); |
3770 | |
3771 | // Reserve stack space for the allocations in CCInfo. |
3772 | CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrAlign); |
3773 | |
3774 | CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4_ByVal); |
3775 | |
3776 | // Area that is at least reserved in the caller of this function. |
3777 | unsigned MinReservedArea = CCByValInfo.getNextStackOffset(); |
3778 | MinReservedArea = std::max(MinReservedArea, LinkageSize); |
3779 | |
3780 | // Set the size that is at least reserved in caller of this function. Tail |
3781 | // call optimized function's reserved stack space needs to be aligned so that |
3782 | // taking the difference between two stack areas will result in an aligned |
3783 | // stack. |
3784 | MinReservedArea = |
3785 | EnsureStackAlignment(Subtarget.getFrameLowering(), MinReservedArea); |
3786 | FuncInfo->setMinReservedArea(MinReservedArea); |
3787 | |
3788 | SmallVector<SDValue, 8> MemOps; |
3789 | |
3790 | // If the function takes variable number of arguments, make a frame index for |
3791 | // the start of the first vararg value... for expansion of llvm.va_start. |
3792 | if (isVarArg) { |
3793 | static const MCPhysReg GPArgRegs[] = { |
3794 | PPC::R3, PPC::R4, PPC::R5, PPC::R6, |
3795 | PPC::R7, PPC::R8, PPC::R9, PPC::R10, |
3796 | }; |
3797 | const unsigned NumGPArgRegs = array_lengthof(GPArgRegs); |
3798 | |
3799 | static const MCPhysReg FPArgRegs[] = { |
3800 | PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7, |
3801 | PPC::F8 |
3802 | }; |
3803 | unsigned NumFPArgRegs = array_lengthof(FPArgRegs); |
3804 | |
3805 | if (useSoftFloat() || hasSPE()) |
3806 | NumFPArgRegs = 0; |
3807 | |
3808 | FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs)); |
3809 | FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs)); |
3810 | |
3811 | // Make room for NumGPArgRegs and NumFPArgRegs. |
3812 | int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 + |
3813 | NumFPArgRegs * MVT(MVT::f64).getSizeInBits()/8; |
3814 | |
3815 | FuncInfo->setVarArgsStackOffset( |
3816 | MFI.CreateFixedObject(PtrVT.getSizeInBits()/8, |
3817 | CCInfo.getNextStackOffset(), true)); |
3818 | |
3819 | FuncInfo->setVarArgsFrameIndex( |
3820 | MFI.CreateStackObject(Depth, Align(8), false)); |
3821 | SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); |
3822 | |
3823 | // The fixed integer arguments of a variadic function are stored to the |
3824 | // VarArgsFrameIndex on the stack so that they may be loaded by |
3825 | // dereferencing the result of va_next. |
3826 | for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) { |
3827 | // Get an existing live-in vreg, or add a new one. |
3828 | unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]); |
3829 | if (!VReg) |
3830 | VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass); |
3831 | |
3832 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
3833 | SDValue Store = |
3834 | DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo()); |
3835 | MemOps.push_back(Store); |
3836 | // Increment the address by four for the next argument to store |
3837 | SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, dl, PtrVT); |
3838 | FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff); |
3839 | } |
3840 | |
3841 | // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6 |
3842 | // is set. |
3843 | // The double arguments are stored to the VarArgsFrameIndex |
3844 | // on the stack. |
3845 | for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) { |
3846 | // Get an existing live-in vreg, or add a new one. |
3847 | unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]); |
3848 | if (!VReg) |
3849 | VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass); |
3850 | |
3851 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64); |
3852 | SDValue Store = |
3853 | DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo()); |
3854 | MemOps.push_back(Store); |
3855 | // Increment the address by eight for the next argument to store |
3856 | SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8, dl, |
3857 | PtrVT); |
3858 | FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff); |
3859 | } |
3860 | } |
3861 | |
3862 | if (!MemOps.empty()) |
3863 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps); |
3864 | |
3865 | return Chain; |
3866 | } |
3867 | |
3868 | // PPC64 passes i8, i16, and i32 values in i64 registers. Promote |
3869 | // value to MVT::i64 and then truncate to the correct register size. |
3870 | SDValue PPCTargetLowering::extendArgForPPC64(ISD::ArgFlagsTy Flags, |
3871 | EVT ObjectVT, SelectionDAG &DAG, |
3872 | SDValue ArgVal, |
3873 | const SDLoc &dl) const { |
3874 | if (Flags.isSExt()) |
3875 | ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal, |
3876 | DAG.getValueType(ObjectVT)); |
3877 | else if (Flags.isZExt()) |
3878 | ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal, |
3879 | DAG.getValueType(ObjectVT)); |
3880 | |
3881 | return DAG.getNode(ISD::TRUNCATE, dl, ObjectVT, ArgVal); |
3882 | } |
3883 | |
3884 | SDValue PPCTargetLowering::LowerFormalArguments_64SVR4( |
3885 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
3886 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
3887 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
3888 | // TODO: add description of PPC stack frame format, or at least some docs. |
3889 | // |
3890 | bool isELFv2ABI = Subtarget.isELFv2ABI(); |
3891 | bool isLittleEndian = Subtarget.isLittleEndian(); |
3892 | MachineFunction &MF = DAG.getMachineFunction(); |
3893 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
3894 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
3895 | |
3896 | assert(!(CallConv == CallingConv::Fast && isVarArg) &&((!(CallConv == CallingConv::Fast && isVarArg) && "fastcc not supported on varargs functions") ? static_cast< void> (0) : __assert_fail ("!(CallConv == CallingConv::Fast && isVarArg) && \"fastcc not supported on varargs functions\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3897, __PRETTY_FUNCTION__)) |
3897 | "fastcc not supported on varargs functions")((!(CallConv == CallingConv::Fast && isVarArg) && "fastcc not supported on varargs functions") ? static_cast< void> (0) : __assert_fail ("!(CallConv == CallingConv::Fast && isVarArg) && \"fastcc not supported on varargs functions\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3897, __PRETTY_FUNCTION__)); |
3898 | |
3899 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
3900 | // Potential tail calls could cause overwriting of argument stack slots. |
3901 | bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt && |
3902 | (CallConv == CallingConv::Fast)); |
3903 | unsigned PtrByteSize = 8; |
3904 | unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
3905 | |
3906 | static const MCPhysReg GPR[] = { |
3907 | PPC::X3, PPC::X4, PPC::X5, PPC::X6, |
3908 | PPC::X7, PPC::X8, PPC::X9, PPC::X10, |
3909 | }; |
3910 | static const MCPhysReg VR[] = { |
3911 | PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8, |
3912 | PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13 |
3913 | }; |
3914 | |
3915 | const unsigned Num_GPR_Regs = array_lengthof(GPR); |
3916 | const unsigned Num_FPR_Regs = useSoftFloat() ? 0 : 13; |
3917 | const unsigned Num_VR_Regs = array_lengthof(VR); |
3918 | |
3919 | // Do a first pass over the arguments to determine whether the ABI |
3920 | // guarantees that our caller has allocated the parameter save area |
3921 | // on its stack frame. In the ELFv1 ABI, this is always the case; |
3922 | // in the ELFv2 ABI, it is true if this is a vararg function or if |
3923 | // any parameter is located in a stack slot. |
3924 | |
3925 | bool HasParameterArea = !isELFv2ABI || isVarArg; |
3926 | unsigned ParamAreaSize = Num_GPR_Regs * PtrByteSize; |
3927 | unsigned NumBytes = LinkageSize; |
3928 | unsigned AvailableFPRs = Num_FPR_Regs; |
3929 | unsigned AvailableVRs = Num_VR_Regs; |
3930 | for (unsigned i = 0, e = Ins.size(); i != e; ++i) { |
3931 | if (Ins[i].Flags.isNest()) |
3932 | continue; |
3933 | |
3934 | if (CalculateStackSlotUsed(Ins[i].VT, Ins[i].ArgVT, Ins[i].Flags, |
3935 | PtrByteSize, LinkageSize, ParamAreaSize, |
3936 | NumBytes, AvailableFPRs, AvailableVRs)) |
3937 | HasParameterArea = true; |
3938 | } |
3939 | |
3940 | // Add DAG nodes to load the arguments or copy them out of registers. On |
3941 | // entry to a function on PPC, the arguments start after the linkage area, |
3942 | // although the first ones are often in registers. |
3943 | |
3944 | unsigned ArgOffset = LinkageSize; |
3945 | unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0; |
3946 | SmallVector<SDValue, 8> MemOps; |
3947 | Function::const_arg_iterator FuncArg = MF.getFunction().arg_begin(); |
3948 | unsigned CurArgIdx = 0; |
3949 | for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) { |
3950 | SDValue ArgVal; |
3951 | bool needsLoad = false; |
3952 | EVT ObjectVT = Ins[ArgNo].VT; |
3953 | EVT OrigVT = Ins[ArgNo].ArgVT; |
3954 | unsigned ObjSize = ObjectVT.getStoreSize(); |
3955 | unsigned ArgSize = ObjSize; |
3956 | ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags; |
3957 | if (Ins[ArgNo].isOrigArg()) { |
3958 | std::advance(FuncArg, Ins[ArgNo].getOrigArgIndex() - CurArgIdx); |
3959 | CurArgIdx = Ins[ArgNo].getOrigArgIndex(); |
3960 | } |
3961 | // We re-align the argument offset for each argument, except when using the |
3962 | // fast calling convention, when we need to make sure we do that only when |
3963 | // we'll actually use a stack slot. |
3964 | unsigned CurArgOffset; |
3965 | Align Alignment; |
3966 | auto ComputeArgOffset = [&]() { |
3967 | /* Respect alignment of argument on the stack. */ |
3968 | Alignment = |
3969 | CalculateStackSlotAlignment(ObjectVT, OrigVT, Flags, PtrByteSize); |
3970 | ArgOffset = alignTo(ArgOffset, Alignment); |
3971 | CurArgOffset = ArgOffset; |
3972 | }; |
3973 | |
3974 | if (CallConv != CallingConv::Fast) { |
3975 | ComputeArgOffset(); |
3976 | |
3977 | /* Compute GPR index associated with argument offset. */ |
3978 | GPR_idx = (ArgOffset - LinkageSize) / PtrByteSize; |
3979 | GPR_idx = std::min(GPR_idx, Num_GPR_Regs); |
3980 | } |
3981 | |
3982 | // FIXME the codegen can be much improved in some cases. |
3983 | // We do not have to keep everything in memory. |
3984 | if (Flags.isByVal()) { |
3985 | assert(Ins[ArgNo].isOrigArg() && "Byval arguments cannot be implicit")((Ins[ArgNo].isOrigArg() && "Byval arguments cannot be implicit" ) ? static_cast<void> (0) : __assert_fail ("Ins[ArgNo].isOrigArg() && \"Byval arguments cannot be implicit\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3985, __PRETTY_FUNCTION__)); |
3986 | |
3987 | if (CallConv == CallingConv::Fast) |
3988 | ComputeArgOffset(); |
3989 | |
3990 | // ObjSize is the true size, ArgSize rounded up to multiple of registers. |
3991 | ObjSize = Flags.getByValSize(); |
3992 | ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
3993 | // Empty aggregate parameters do not take up registers. Examples: |
3994 | // struct { } a; |
3995 | // union { } b; |
3996 | // int c[0]; |
3997 | // etc. However, we have to provide a place-holder in InVals, so |
3998 | // pretend we have an 8-byte item at the current address for that |
3999 | // purpose. |
4000 | if (!ObjSize) { |
4001 | int FI = MFI.CreateFixedObject(PtrByteSize, ArgOffset, true); |
4002 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
4003 | InVals.push_back(FIN); |
4004 | continue; |
4005 | } |
4006 | |
4007 | // Create a stack object covering all stack doublewords occupied |
4008 | // by the argument. If the argument is (fully or partially) on |
4009 | // the stack, or if the argument is fully in registers but the |
4010 | // caller has allocated the parameter save anyway, we can refer |
4011 | // directly to the caller's stack frame. Otherwise, create a |
4012 | // local copy in our own frame. |
4013 | int FI; |
4014 | if (HasParameterArea || |
4015 | ArgSize + ArgOffset > LinkageSize + Num_GPR_Regs * PtrByteSize) |
4016 | FI = MFI.CreateFixedObject(ArgSize, ArgOffset, false, true); |
4017 | else |
4018 | FI = MFI.CreateStackObject(ArgSize, Alignment, false); |
4019 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
4020 | |
4021 | // Handle aggregates smaller than 8 bytes. |
4022 | if (ObjSize < PtrByteSize) { |
4023 | // The value of the object is its address, which differs from the |
4024 | // address of the enclosing doubleword on big-endian systems. |
4025 | SDValue Arg = FIN; |
4026 | if (!isLittleEndian) { |
4027 | SDValue ArgOff = DAG.getConstant(PtrByteSize - ObjSize, dl, PtrVT); |
4028 | Arg = DAG.getNode(ISD::ADD, dl, ArgOff.getValueType(), Arg, ArgOff); |
4029 | } |
4030 | InVals.push_back(Arg); |
4031 | |
4032 | if (GPR_idx != Num_GPR_Regs) { |
4033 | unsigned VReg = MF.addLiveIn(GPR[GPR_idx++], &PPC::G8RCRegClass); |
4034 | FuncInfo->addLiveInAttr(VReg, Flags); |
4035 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
4036 | SDValue Store; |
4037 | |
4038 | if (ObjSize==1 || ObjSize==2 || ObjSize==4) { |
4039 | EVT ObjType = (ObjSize == 1 ? MVT::i8 : |
4040 | (ObjSize == 2 ? MVT::i16 : MVT::i32)); |
4041 | Store = DAG.getTruncStore(Val.getValue(1), dl, Val, Arg, |
4042 | MachinePointerInfo(&*FuncArg), ObjType); |
4043 | } else { |
4044 | // For sizes that don't fit a truncating store (3, 5, 6, 7), |
4045 | // store the whole register as-is to the parameter save area |
4046 | // slot. |
4047 | Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, |
4048 | MachinePointerInfo(&*FuncArg)); |
4049 | } |
4050 | |
4051 | MemOps.push_back(Store); |
4052 | } |
4053 | // Whether we copied from a register or not, advance the offset |
4054 | // into the parameter save area by a full doubleword. |
4055 | ArgOffset += PtrByteSize; |
4056 | continue; |
4057 | } |
4058 | |
4059 | // The value of the object is its address, which is the address of |
4060 | // its first stack doubleword. |
4061 | InVals.push_back(FIN); |
4062 | |
4063 | // Store whatever pieces of the object are in registers to memory. |
4064 | for (unsigned j = 0; j < ArgSize; j += PtrByteSize) { |
4065 | if (GPR_idx == Num_GPR_Regs) |
4066 | break; |
4067 | |
4068 | unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass); |
4069 | FuncInfo->addLiveInAttr(VReg, Flags); |
4070 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
4071 | SDValue Addr = FIN; |
4072 | if (j) { |
4073 | SDValue Off = DAG.getConstant(j, dl, PtrVT); |
4074 | Addr = DAG.getNode(ISD::ADD, dl, Off.getValueType(), Addr, Off); |
4075 | } |
4076 | SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, Addr, |
4077 | MachinePointerInfo(&*FuncArg, j)); |
4078 | MemOps.push_back(Store); |
4079 | ++GPR_idx; |
4080 | } |
4081 | ArgOffset += ArgSize; |
4082 | continue; |
4083 | } |
4084 | |
4085 | switch (ObjectVT.getSimpleVT().SimpleTy) { |
4086 | default: llvm_unreachable("Unhandled argument type!")::llvm::llvm_unreachable_internal("Unhandled argument type!", "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 4086); |
4087 | case MVT::i1: |
4088 | case MVT::i32: |
4089 | case MVT::i64: |
4090 | if (Flags.isNest()) { |
4091 | // The 'nest' parameter, if any, is passed in R11. |
4092 | unsigned VReg = MF.addLiveIn(PPC::X11, &PPC::G8RCRegClass); |
4093 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64); |
4094 | |
4095 | if (ObjectVT == MVT::i32 || ObjectVT == MVT::i1) |
4096 | ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl); |
4097 | |
4098 | break; |
4099 | } |
4100 | |
4101 | // These can be scalar arguments or elements of an integer array type |
4102 | // passed directly. Clang may use those instead of "byval" aggregate |
4103 | // types to avoid forcing arguments to memory unnecessarily. |
4104 | if (GPR_idx != Num_GPR_Regs) { |
4105 | unsigned VReg = MF.addLiveIn(GPR[GPR_idx++], &PPC::G8RCRegClass); |
4106 | FuncInfo->addLiveInAttr(VReg, Flags); |
4107 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64); |
4108 | |
4109 | if (ObjectVT == MVT::i32 || ObjectVT == MVT::i1) |
4110 | // PPC64 passes i8, i16, and i32 values in i64 registers. Promote |
4111 | // value to MVT::i64 and then truncate to the correct register size. |
4112 | ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl); |
4113 | } else { |
4114 | if (CallConv == CallingConv::Fast) |
4115 | ComputeArgOffset(); |
4116 | |
4117 | needsLoad = true; |
4118 | ArgSize = PtrByteSize; |
4119 | } |
4120 | if (CallConv != CallingConv::Fast || needsLoad) |
4121 | ArgOffset += 8; |
4122 | break; |
4123 | |
4124 | case MVT::f32: |
4125 | case MVT::f64: |
4126 | // These can be scalar arguments or elements of a float array type |
4127 | // passed directly. The latter are used to implement ELFv2 homogenous |
4128 | // float aggregates. |
4129 | if (FPR_idx != Num_FPR_Regs) { |
4130 | unsigned VReg; |
4131 | |
4132 | if (ObjectVT == MVT::f32) |
4133 | VReg = MF.addLiveIn(FPR[FPR_idx], |
4134 | Subtarget.hasP8Vector() |
4135 | ? &PPC::VSSRCRegClass |
4136 | : &PPC::F4RCRegClass); |
4137 | else |
4138 | VReg = MF.addLiveIn(FPR[FPR_idx], Subtarget.hasVSX() |
4139 | ? &PPC::VSFRCRegClass |
4140 | : &PPC::F8RCRegClass); |
4141 | |
4142 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT); |
4143 | ++FPR_idx; |
4144 | } else if (GPR_idx != Num_GPR_Regs && CallConv != CallingConv::Fast) { |
4145 | // FIXME: We may want to re-enable this for CallingConv::Fast on the P8 |
4146 | // once we support fp <-> gpr moves. |
4147 | |
4148 | // This can only ever happen in the presence of f32 array types, |
4149 | // since otherwise we never run out of FPRs before running out |
4150 | // of GPRs. |
4151 | unsigned VReg = MF.addLiveIn(GPR[GPR_idx++], &PPC::G8RCRegClass); |
4152 | FuncInfo->addLiveInAttr(VReg, Flags); |
4153 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64); |
4154 | |
4155 | if (ObjectVT == MVT::f32) { |
4156 | if ((ArgOffset % PtrByteSize) == (isLittleEndian ? 4 : 0)) |
4157 | ArgVal = DAG.getNode(ISD::SRL, dl, MVT::i64, ArgVal, |
4158 | DAG.getConstant(32, dl, MVT::i32)); |
4159 | ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal); |
4160 | } |
4161 | |
4162 | ArgVal = DAG.getNode(ISD::BITCAST, dl, ObjectVT, ArgVal); |
4163 | } else { |
4164 | if (CallConv == CallingConv::Fast) |
4165 | ComputeArgOffset(); |
4166 | |
4167 | needsLoad = true; |
4168 | } |
4169 | |
4170 | // When passing an array of floats, the array occupies consecutive |
4171 | // space in the argument area; only round up to the next doubleword |
4172 | // at the end of the array. Otherwise, each float takes 8 bytes. |
4173 | if (CallConv != CallingConv::Fast || needsLoad) { |
4174 | ArgSize = Flags.isInConsecutiveRegs() ? ObjSize : PtrByteSize; |
4175 | ArgOffset += ArgSize; |
4176 | if (Flags.isInConsecutiveRegsLast()) |
4177 | ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
4178 | } |
4179 | break; |
4180 | case MVT::v4f32: |
4181 | case MVT::v4i32: |
4182 | case MVT::v8i16: |
4183 | case MVT::v16i8: |
4184 | case MVT::v2f64: |
4185 | case MVT::v2i64: |
4186 | case MVT::v1i128: |
4187 | case MVT::f128: |
4188 | // These can be scalar arguments or elements of a vector array type |
4189 | // passed directly. The latter are used to implement ELFv2 homogenous |
4190 | // vector aggregates. |
4191 | if (VR_idx != Num_VR_Regs) { |
4192 | unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass); |
4193 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT); |
4194 | ++VR_idx; |
4195 | } else { |
4196 | if (CallConv == CallingConv::Fast) |
4197 | ComputeArgOffset(); |
4198 | needsLoad = true; |
4199 | } |
4200 | if (CallConv != CallingConv::Fast || needsLoad) |
4201 | ArgOffset += 16; |
4202 | break; |
4203 | } |
4204 | |
4205 | // We need to load the argument to a virtual register if we determined |
4206 | // above that we ran out of physical registers of the appropriate type. |
4207 | if (needsLoad) { |
4208 | if (ObjSize < ArgSize && !isLittleEndian) |
4209 | CurArgOffset += ArgSize - ObjSize; |
4210 | int FI = MFI.CreateFixedObject(ObjSize, CurArgOffset, isImmutable); |
4211 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
4212 | ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo()); |
4213 | } |
4214 | |
4215 | InVals.push_back(ArgVal); |
4216 | } |
4217 | |
4218 | // Area that is at least reserved in the caller of this function. |
4219 | unsigned MinReservedArea; |
4220 | if (HasParameterArea) |
4221 | MinReservedArea = std::max(ArgOffset, LinkageSize + 8 * PtrByteSize); |
4222 | else |
4223 | MinReservedArea = LinkageSize; |
4224 | |
4225 | // Set the size that is at least reserved in caller of this function. Tail |
4226 | // call optimized functions' reserved stack space needs to be aligned so that |
4227 | // taking the difference between two stack areas will result in an aligned |
4228 | // stack. |
4229 | MinReservedArea = |
4230 | EnsureStackAlignment(Subtarget.getFrameLowering(), MinReservedArea); |
4231 | FuncInfo->setMinReservedArea(MinReservedArea); |
4232 | |
4233 | // If the function takes variable number of arguments, make a frame index for |
4234 | // the start of the first vararg value... for expansion of llvm.va_start. |
4235 | // On ELFv2ABI spec, it writes: |
4236 | // C programs that are intended to be *portable* across different compilers |
4237 | // and architectures must use the header file <stdarg.h> to deal with variable |
4238 | // argument lists. |
4239 | if (isVarArg && MFI.hasVAStart()) { |
4240 | int Depth = ArgOffset; |
4241 | |
4242 | FuncInfo->setVarArgsFrameIndex( |
4243 | MFI.CreateFixedObject(PtrByteSize, Depth, true)); |
4244 | SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); |
4245 | |
4246 | // If this function is vararg, store any remaining integer argument regs |
4247 | // to their spots on the stack so that they may be loaded by dereferencing |
4248 | // the result of va_next. |
4249 | for (GPR_idx = (ArgOffset - LinkageSize) / PtrByteSize; |
4250 | GPR_idx < Num_GPR_Regs; ++GPR_idx) { |
4251 | unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass); |
4252 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
4253 | SDValue Store = |
4254 | DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo()); |
4255 | MemOps.push_back(Store); |
4256 | // Increment the address by four for the next argument to store |
4257 | SDValue PtrOff = DAG.getConstant(PtrByteSize, dl, PtrVT); |
4258 | FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff); |
4259 | } |
4260 | } |
4261 | |
4262 | if (!MemOps.empty()) |
4263 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps); |
4264 | |
4265 | return Chain; |
4266 | } |
4267 | |
4268 | SDValue PPCTargetLowering::LowerFormalArguments_Darwin( |
4269 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
4270 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
4271 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
4272 | // TODO: add description of PPC stack frame format, or at least some docs. |
4273 | // |
4274 | MachineFunction &MF = DAG.getMachineFunction(); |
4275 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
4276 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
4277 | |
4278 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
4279 | bool isPPC64 = PtrVT == MVT::i64; |
4280 | // Potential tail calls could cause overwriting of argument stack slots. |
4281 | bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt && |
4282 | (CallConv == CallingConv::Fast)); |
4283 | unsigned PtrByteSize = isPPC64 ? 8 : 4; |
4284 | unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
4285 | unsigned ArgOffset = LinkageSize; |
4286 | // Area that is at least reserved in caller of this function. |
4287 | unsigned MinReservedArea = ArgOffset; |
4288 | |
4289 | static const MCPhysReg GPR_32[] = { // 32-bit registers. |
4290 | PPC::R3, PPC::R4, PPC::R5, PPC::R6, |
4291 | PPC::R7, PPC::R8, PPC::R9, PPC::R10, |
4292 | }; |
4293 | static const MCPhysReg GPR_64[] = { // 64-bit registers. |
4294 | PPC::X3, PPC::X4, PPC::X5, PPC::X6, |
4295 | PPC::X7, PPC::X8, PPC::X9, PPC::X10, |
4296 | }; |
4297 | static const MCPhysReg VR[] = { |
4298 | PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8, |
4299 | PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13 |
4300 | }; |
4301 | |
4302 | const unsigned Num_GPR_Regs = array_lengthof(GPR_32); |
4303 | const unsigned Num_FPR_Regs = useSoftFloat() ? 0 : 13; |
4304 | const unsigned Num_VR_Regs = array_lengthof( VR); |
4305 | |
4306 | unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0; |
4307 | |
4308 | const MCPhysReg *GPR = isPPC64 ? GPR_64 : GPR_32; |
4309 | |
4310 | // In 32-bit non-varargs functions, the stack space for vectors is after the |
4311 | // stack space for non-vectors. We do not use this space unless we have |
4312 | // too many vectors to fit in registers, something that only occurs in |
4313 | // constructed examples:), but we have to walk the arglist to figure |
4314 | // that out...for the pathological case, compute VecArgOffset as the |
4315 | // start of the vector parameter area. Computing VecArgOffset is the |
4316 | // entire point of the following loop. |
4317 | unsigned VecArgOffset = ArgOffset; |
4318 | if (!isVarArg && !isPPC64) { |
4319 | for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; |
4320 | ++ArgNo) { |
4321 | EVT ObjectVT = Ins[ArgNo].VT; |
4322 | ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags; |
4323 | |
4324 | if (Flags.isByVal()) { |
4325 | // ObjSize is the true size, ArgSize rounded up to multiple of regs. |
4326 | unsigned ObjSize = Flags.getByValSize(); |
4327 | unsigned ArgSize = |
4328 | ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
4329 | VecArgOffset += ArgSize; |
4330 | continue; |
4331 | } |
4332 | |
4333 | switch(ObjectVT.getSimpleVT().SimpleTy) { |
4334 | default: llvm_unreachable("Unhandled argument type!")::llvm::llvm_unreachable_internal("Unhandled argument type!", "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 4334); |
4335 | case MVT::i1: |
4336 | case MVT::i32: |
4337 | case MVT::f32: |
4338 | VecArgOffset += 4; |
4339 | break; |
4340 | case MVT::i64: // PPC64 |
4341 | case MVT::f64: |
4342 | // FIXME: We are guaranteed to be !isPPC64 at this point. |
4343 | // Does MVT::i64 apply? |
4344 | VecArgOffset += 8; |
4345 | break; |
4346 | case MVT::v4f32: |
4347 | case MVT::v4i32: |
4348 | case MVT::v8i16: |
4349 | case MVT::v16i8: |
4350 | // Nothing to do, we're only looking at Nonvector args here. |
4351 | break; |
4352 | } |
4353 | } |
4354 | } |
4355 | // We've found where the vector parameter area in memory is. Skip the |
4356 | // first 12 parameters; these don't use that memory. |
4357 | VecArgOffset = ((VecArgOffset+15)/16)*16; |
4358 | VecArgOffset += 12*16; |
4359 | |
4360 | // Add DAG nodes to load the arguments or copy them out of registers. On |
4361 | // entry to a function on PPC, the arguments start after the linkage area, |
4362 | // although the first ones are often in registers. |
4363 | |
4364 | SmallVector<SDValue, 8> MemOps; |
4365 | unsigned nAltivecParamsAtEnd = 0; |
4366 | Function::const_arg_iterator FuncArg = MF.getFunction().arg_begin(); |
4367 | unsigned CurArgIdx = 0; |
4368 | for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) { |
4369 | SDValue ArgVal; |
4370 | bool needsLoad = false; |
4371 | EVT ObjectVT = Ins[ArgNo].VT; |
4372 | unsigned ObjSize = ObjectVT.getSizeInBits()/8; |
4373 | unsigned ArgSize = ObjSize; |
4374 | ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags; |
4375 | if (Ins[ArgNo].isOrigArg()) { |
4376 | std::advance(FuncArg, Ins[ArgNo].getOrigArgIndex() - CurArgIdx); |
4377 | CurArgIdx = Ins[ArgNo].getOrigArgIndex(); |
4378 | } |
4379 | unsigned CurArgOffset = ArgOffset; |
4380 | |
4381 | // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary. |
4382 | if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 || |
4383 | ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) { |
4384 | if (isVarArg || isPPC64) { |
4385 | MinReservedArea = ((MinReservedArea+15)/16)*16; |
4386 | MinReservedArea += CalculateStackSlotSize(ObjectVT, |
4387 | Flags, |
4388 | PtrByteSize); |
4389 | } else nAltivecParamsAtEnd++; |
4390 | } else |
4391 | // Calculate min reserved area. |
4392 | MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT, |
4393 | Flags, |
4394 | PtrByteSize); |
4395 | |
4396 | // FIXME the codegen can be much improved in some cases. |
4397 | // We do not have to keep everything in memory. |
4398 | if (Flags.isByVal()) { |
4399 | assert(Ins[ArgNo].isOrigArg() && "Byval arguments cannot be implicit")((Ins[ArgNo].isOrigArg() && "Byval arguments cannot be implicit" ) ? static_cast<void> (0) : __assert_fail ("Ins[ArgNo].isOrigArg() && \"Byval arguments cannot be implicit\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 4399, __PRETTY_FUNCTION__)); |
4400 | |
4401 | // ObjSize is the true size, ArgSize rounded up to multiple of registers. |
4402 | ObjSize = Flags.getByValSize(); |
4403 | ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
4404 | // Objects of size 1 and 2 are right justified, everything else is |
4405 | // left justified. This means the memory address is adjusted forwards. |
4406 | if (ObjSize==1 || ObjSize==2) { |
4407 | CurArgOffset = CurArgOffset + (4 - ObjSize); |
4408 | } |
4409 | // The value of the object is its address. |
4410 | int FI = MFI.CreateFixedObject(ObjSize, CurArgOffset, false, true); |
4411 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
4412 | InVals.push_back(FIN); |
4413 | if (ObjSize==1 || ObjSize==2) { |
4414 | if (GPR_idx != Num_GPR_Regs) { |
4415 | unsigned VReg; |
4416 | if (isPPC64) |
4417 | VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass); |
4418 | else |
4419 | VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass); |
4420 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
4421 | EVT ObjType = ObjSize == 1 ? MVT::i8 : MVT::i16; |
4422 | SDValue Store = |
4423 | DAG.getTruncStore(Val.getValue(1), dl, Val, FIN, |
4424 | MachinePointerInfo(&*FuncArg), ObjType); |
4425 | MemOps.push_back(Store); |
4426 | ++GPR_idx; |
4427 | } |
4428 | |
4429 | ArgOffset += PtrByteSize; |
4430 | |
4431 | continue; |
4432 | } |
4433 | for (unsigned j = 0; j < ArgSize; j += PtrByteSize) { |
4434 | // Store whatever pieces of the object are in registers |
4435 | // to memory. ArgOffset will be the address of the beginning |
4436 | // of the object. |
4437 | if (GPR_idx != Num_GPR_Regs) { |
4438 | unsigned VReg; |
4439 | if (isPPC64) |
4440 | VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass); |
4441 | else |
4442 | VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass); |
4443 | int FI = MFI.CreateFixedObject(PtrByteSize, ArgOffset, true); |
4444 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
4445 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
4446 | SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, |
4447 | MachinePointerInfo(&*FuncArg, j)); |
4448 | MemOps.push_back(Store); |
4449 | ++GPR_idx; |
4450 | ArgOffset += PtrByteSize; |
4451 | } else { |
4452 | ArgOffset += ArgSize - (ArgOffset-CurArgOffset); |
4453 | break; |
4454 | } |
4455 | } |
4456 | continue; |
4457 | } |
4458 | |
4459 | switch (ObjectVT.getSimpleVT().SimpleTy) { |
4460 | default: llvm_unreachable("Unhandled argument type!")::llvm::llvm_unreachable_internal("Unhandled argument type!", "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 4460); |
4461 | case MVT::i1: |
4462 | case MVT::i32: |
4463 | if (!isPPC64) { |
4464 | if (GPR_idx != Num_GPR_Regs) { |
4465 | unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass); |
4466 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32); |
4467 | |
4468 | if (ObjectVT == MVT::i1) |
4469 | ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, ArgVal); |
4470 | |
4471 | ++GPR_idx; |
4472 | } else { |
4473 | needsLoad = true; |
4474 | ArgSize = PtrByteSize; |
4475 | } |
4476 | // All int arguments reserve stack space in the Darwin ABI. |
4477 | ArgOffset += PtrByteSize; |
4478 | break; |
4479 | } |
4480 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
4481 | case MVT::i64: // PPC64 |
4482 | if (GPR_idx != Num_GPR_Regs) { |
4483 | unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass); |
4484 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64); |
4485 | |
4486 | if (ObjectVT == MVT::i32 || ObjectVT == MVT::i1) |
4487 | // PPC64 passes i8, i16, and i32 values in i64 registers. Promote |
4488 | // value to MVT::i64 and then truncate to the correct register size. |
4489 | ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl); |
4490 | |
4491 | ++GPR_idx; |
4492 | } else { |
4493 | needsLoad = true; |
4494 | ArgSize = PtrByteSize; |
4495 | } |
4496 | // All int arguments reserve stack space in the Darwin ABI. |
4497 | ArgOffset += 8; |
4498 | break; |
4499 | |
4500 | case MVT::f32: |
4501 | case MVT::f64: |
4502 | // Every 4 bytes of argument space consumes one of the GPRs available for |
4503 | // argument passing. |
4504 | if (GPR_idx != Num_GPR_Regs) { |
4505 | ++GPR_idx; |
4506 | if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64) |
4507 | ++GPR_idx; |
4508 | } |
4509 | if (FPR_idx != Num_FPR_Regs) { |
4510 | unsigned VReg; |
4511 | |
4512 | if (ObjectVT == MVT::f32) |
4513 | VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass); |
4514 | else |
4515 | VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass); |
4516 | |
4517 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT); |
4518 | ++FPR_idx; |
4519 | } else { |
4520 | needsLoad = true; |
4521 | } |
4522 | |
4523 | // All FP arguments reserve stack space in the Darwin ABI. |
4524 | ArgOffset += isPPC64 ? 8 : ObjSize; |
4525 | break; |
4526 | case MVT::v4f32: |
4527 | case MVT::v4i32: |
4528 | case MVT::v8i16: |
4529 | case MVT::v16i8: |
4530 | // Note that vector arguments in registers don't reserve stack space, |
4531 | // except in varargs functions. |
4532 | if (VR_idx != Num_VR_Regs) { |
4533 | unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass); |
4534 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT); |
4535 | if (isVarArg) { |
4536 | while ((ArgOffset % 16) != 0) { |
4537 | ArgOffset += PtrByteSize; |
4538 | if (GPR_idx != Num_GPR_Regs) |
4539 | GPR_idx++; |
4540 | } |
4541 | ArgOffset += 16; |
4542 | GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64? |
4543 | } |
4544 | ++VR_idx; |
4545 | } else { |
4546 | if (!isVarArg && !isPPC64) { |
4547 | // Vectors go after all the nonvectors. |
4548 | CurArgOffset = VecArgOffset; |
4549 | VecArgOffset += 16; |
4550 | } else { |
4551 | // Vectors are aligned. |
4552 | ArgOffset = ((ArgOffset+15)/16)*16; |
4553 | CurArgOffset = ArgOffset; |
4554 | ArgOffset += 16; |
4555 | } |
4556 | needsLoad = true; |
4557 | } |
4558 | break; |
4559 | } |
4560 | |
4561 | // We need to load the argument to a virtual register if we determined above |
4562 | // that we ran out of physical registers of the appropriate type. |
4563 | if (needsLoad) { |
4564 | int FI = MFI.CreateFixedObject(ObjSize, |
4565 | CurArgOffset + (ArgSize - ObjSize), |
4566 | isImmutable); |
4567 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
4568 | ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo()); |
4569 | } |
4570 | |
4571 | InVals.push_back(ArgVal); |
4572 | } |
4573 | |
4574 | // Allow for Altivec parameters at the end, if needed. |
4575 | if (nAltivecParamsAtEnd) { |
4576 | MinReservedArea = ((MinReservedArea+15)/16)*16; |
4577 | MinReservedArea += 16*nAltivecParamsAtEnd; |
4578 | } |
4579 | |
4580 | // Area that is at least reserved in the caller of this function. |
4581 | MinReservedArea = std::max(MinReservedArea, LinkageSize + 8 * PtrByteSize); |
4582 | |
4583 | // Set the size that is at least reserved in caller of this function. Tail |
4584 | // call optimized functions' reserved stack space needs to be aligned so that |
4585 | // taking the difference between two stack areas will result in an aligned |
4586 | // stack. |
4587 | MinReservedArea = |
4588 | EnsureStackAlignment(Subtarget.getFrameLowering(), MinReservedArea); |
4589 | FuncInfo->setMinReservedArea(MinReservedArea); |
4590 | |
4591 | // If the function takes variable number of arguments, make a frame index for |
4592 | // the start of the first vararg value... for expansion of llvm.va_start. |
4593 | if (isVarArg) { |
4594 | int Depth = ArgOffset; |
4595 | |
4596 | FuncInfo->setVarArgsFrameIndex( |
4597 | MFI.CreateFixedObject(PtrVT.getSizeInBits()/8, |
4598 | Depth, true)); |
4599 | SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); |
4600 | |
4601 | // If this function is vararg, store any remaining integer argument regs |
4602 | // to their spots on the stack so that they may be loaded by dereferencing |
4603 | // the result of va_next. |
4604 | for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) { |
4605 | unsigned VReg; |
4606 | |
4607 | if (isPPC64) |
4608 | VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass); |
4609 | else |
4610 | VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass); |
4611 | |
4612 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
4613 | SDValue Store = |
4614 | DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo()); |
4615 | MemOps.push_back(Store); |
4616 | // Increment the address by four for the next argument to store |
4617 | SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, dl, PtrVT); |
4618 | FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff); |
4619 | } |
4620 | } |
4621 | |
4622 | if (!MemOps.empty()) |
4623 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps); |
4624 | |
4625 | return Chain; |
4626 | } |
4627 | |
4628 | /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be |
4629 | /// adjusted to accommodate the arguments for the tailcall. |
4630 | static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall, |
4631 | unsigned ParamSize) { |
4632 | |
4633 | if (!isTailCall) return 0; |
4634 | |
4635 | PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>(); |
4636 | unsigned CallerMinReservedArea = FI->getMinReservedArea(); |
4637 | int SPDiff = (int)CallerMinReservedArea - (int)ParamSize; |
4638 | // Remember only if the new adjustment is bigger. |
4639 | if (SPDiff < FI->getTailCallSPDelta()) |
4640 | FI->setTailCallSPDelta(SPDiff); |
4641 | |
4642 | return SPDiff; |
4643 | } |
4644 | |
4645 | static bool isFunctionGlobalAddress(SDValue Callee); |
4646 | |
4647 | static bool callsShareTOCBase(const Function *Caller, SDValue Callee, |
4648 | const TargetMachine &TM) { |
4649 | // It does not make sense to call callsShareTOCBase() with a caller that |
4650 | // is PC Relative since PC Relative callers do not have a TOC. |
4651 | #ifndef NDEBUG |
4652 | const PPCSubtarget *STICaller = &TM.getSubtarget<PPCSubtarget>(*Caller); |
4653 | assert(!STICaller->isUsingPCRelativeCalls() &&((!STICaller->isUsingPCRelativeCalls() && "PC Relative callers do not have a TOC and cannot share a TOC Base" ) ? static_cast<void> (0) : __assert_fail ("!STICaller->isUsingPCRelativeCalls() && \"PC Relative callers do not have a TOC and cannot share a TOC Base\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 4654, __PRETTY_FUNCTION__)) |
4654 | "PC Relative callers do not have a TOC and cannot share a TOC Base")((!STICaller->isUsingPCRelativeCalls() && "PC Relative callers do not have a TOC and cannot share a TOC Base" ) ? static_cast<void> (0) : __assert_fail ("!STICaller->isUsingPCRelativeCalls() && \"PC Relative callers do not have a TOC and cannot share a TOC Base\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 4654, __PRETTY_FUNCTION__)); |
4655 | #endif |
4656 | |
4657 | // Callee is either a GlobalAddress or an ExternalSymbol. ExternalSymbols |
4658 | // don't have enough information to determine if the caller and callee share |
4659 | // the same TOC base, so we have to pessimistically assume they don't for |
4660 | // correctness. |
4661 | GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee); |
4662 | if (!G) |
4663 | return false; |
4664 | |
4665 | const GlobalValue *GV = G->getGlobal(); |
4666 | |
4667 | // If the callee is preemptable, then the static linker will use a plt-stub |
4668 | // which saves the toc to the stack, and needs a nop after the call |
4669 | // instruction to convert to a toc-restore. |
4670 | if (!TM.shouldAssumeDSOLocal(*Caller->getParent(), GV)) |
4671 | return false; |
4672 | |
4673 | // Functions with PC Relative enabled may clobber the TOC in the same DSO. |
4674 | // We may need a TOC restore in the situation where the caller requires a |
4675 | // valid TOC but the callee is PC Relative and does not. |
4676 | const Function *F = dyn_cast<Function>(GV); |
4677 | const GlobalAlias *Alias = dyn_cast<GlobalAlias>(GV); |
4678 | |
4679 | // If we have an Alias we can try to get the function from there. |
4680 | if (Alias) { |
4681 | const GlobalObject *GlobalObj = Alias->getBaseObject(); |
4682 | F = dyn_cast<Function>(GlobalObj); |
4683 | } |
4684 | |
4685 | // If we still have no valid function pointer we do not have enough |
4686 | // information to determine if the callee uses PC Relative calls so we must |
4687 | // assume that it does. |
4688 | if (!F) |
4689 | return false; |
4690 | |
4691 | // If the callee uses PC Relative we cannot guarantee that the callee won't |
4692 | // clobber the TOC of the caller and so we must assume that the two |
4693 | // functions do not share a TOC base. |
4694 | const PPCSubtarget *STICallee = &TM.getSubtarget<PPCSubtarget>(*F); |
4695 | if (STICallee->isUsingPCRelativeCalls()) |
4696 | return false; |
4697 | |
4698 | // The medium and large code models are expected to provide a sufficiently |
4699 | // large TOC to provide all data addressing needs of a module with a |
4700 | // single TOC. |
4701 | if (CodeModel::Medium == TM.getCodeModel() || |
4702 | CodeModel::Large == TM.getCodeModel()) |
4703 | return true; |
4704 | |
4705 | // Otherwise we need to ensure callee and caller are in the same section, |
4706 | // since the linker may allocate multiple TOCs, and we don't know which |
4707 | // sections will belong to the same TOC base. |
4708 | if (!GV->isStrongDefinitionForLinker()) |
4709 | return false; |
4710 | |
4711 | // Any explicitly-specified sections and section prefixes must also match. |
4712 | // Also, if we're using -ffunction-sections, then each function is always in |
4713 | // a different section (the same is true for COMDAT functions). |
4714 | if (TM.getFunctionSections() || GV->hasComdat() || Caller->hasComdat() || |
4715 | GV->getSection() != Caller->getSection()) |
4716 | return false; |
4717 | if (const auto *F = dyn_cast<Function>(GV)) { |
4718 | if (F->getSectionPrefix() != Caller->getSectionPrefix()) |
4719 | return false; |
4720 | } |
4721 | |
4722 | return true; |
4723 | } |
4724 | |
4725 | static bool |
4726 | needStackSlotPassParameters(const PPCSubtarget &Subtarget, |
4727 | const SmallVectorImpl<ISD::OutputArg> &Outs) { |
4728 | assert(Subtarget.is64BitELFABI())((Subtarget.is64BitELFABI()) ? static_cast<void> (0) : __assert_fail ("Subtarget.is64BitELFABI()", "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 4728, __PRETTY_FUNCTION__)); |
4729 | |
4730 | const unsigned PtrByteSize = 8; |
4731 | const unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
4732 | |
4733 | static const MCPhysReg GPR[] = { |
4734 | PPC::X3, PPC::X4, PPC::X5, PPC::X6, |
4735 | PPC::X7, PPC::X8, PPC::X9, PPC::X10, |
4736 | }; |
4737 | static const MCPhysReg VR[] = { |
4738 | PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8, |
4739 | PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13 |
4740 | }; |
4741 | |
4742 | const unsigned NumGPRs = array_lengthof(GPR); |
4743 | const unsigned NumFPRs = 13; |
4744 | const unsigned NumVRs = array_lengthof(VR); |
4745 | const unsigned ParamAreaSize = NumGPRs * PtrByteSize; |
4746 | |
4747 | unsigned NumBytes = LinkageSize; |
4748 | unsigned AvailableFPRs = NumFPRs; |
4749 | unsigned AvailableVRs = NumVRs; |
4750 | |
4751 | for (const ISD::OutputArg& Param : Outs) { |
4752 | if (Param.Flags.isNest()) continue; |
4753 | |
4754 | if (CalculateStackSlotUsed(Param.VT, Param.ArgVT, Param.Flags, PtrByteSize, |
4755 | LinkageSize, ParamAreaSize, NumBytes, |
4756 | AvailableFPRs, AvailableVRs)) |
4757 | return true; |
4758 | } |
4759 | return false; |
4760 | } |
4761 | |
4762 | static bool hasSameArgumentList(const Function *CallerFn, const CallBase &CB) { |
4763 | if (CB.arg_size() != CallerFn->arg_size()) |
4764 | return false; |
4765 | |
4766 | auto CalleeArgIter = CB.arg_begin(); |
4767 | auto CalleeArgEnd = CB.arg_end(); |
4768 | Function::const_arg_iterator CallerArgIter = CallerFn->arg_begin(); |
4769 | |
4770 | for (; CalleeArgIter != CalleeArgEnd; ++CalleeArgIter, ++CallerArgIter) { |
4771 | const Value* CalleeArg = *CalleeArgIter; |
4772 | const Value* CallerArg = &(*CallerArgIter); |
4773 | if (CalleeArg == CallerArg) |
4774 | continue; |
4775 | |
4776 | // e.g. @caller([4 x i64] %a, [4 x i64] %b) { |
4777 | // tail call @callee([4 x i64] undef, [4 x i64] %b) |
4778 | // } |
4779 | // 1st argument of callee is undef and has the same type as caller. |
4780 | if (CalleeArg->getType() == CallerArg->getType() && |
4781 | isa<UndefValue>(CalleeArg)) |
4782 | continue; |
4783 | |
4784 | return false; |
4785 | } |
4786 | |
4787 | return true; |
4788 | } |
4789 | |
4790 | // Returns true if TCO is possible between the callers and callees |
4791 | // calling conventions. |
4792 | static bool |
4793 | areCallingConvEligibleForTCO_64SVR4(CallingConv::ID CallerCC, |
4794 | CallingConv::ID CalleeCC) { |
4795 | // Tail calls are possible with fastcc and ccc. |
4796 | auto isTailCallableCC = [] (CallingConv::ID CC){ |
4797 | return CC == CallingConv::C || CC == CallingConv::Fast; |
4798 | }; |
4799 | if (!isTailCallableCC(CallerCC) || !isTailCallableCC(CalleeCC)) |
4800 | return false; |
4801 | |
4802 | // We can safely tail call both fastcc and ccc callees from a c calling |
4803 | // convention caller. If the caller is fastcc, we may have less stack space |
4804 | // than a non-fastcc caller with the same signature so disable tail-calls in |
4805 | // that case. |
4806 | return CallerCC == CallingConv::C || CallerCC == CalleeCC; |
4807 | } |
4808 | |
4809 | bool PPCTargetLowering::IsEligibleForTailCallOptimization_64SVR4( |
4810 | SDValue Callee, CallingConv::ID CalleeCC, const CallBase *CB, bool isVarArg, |
4811 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
4812 | const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const { |
4813 | bool TailCallOpt = getTargetMachine().Options.GuaranteedTailCallOpt; |
4814 | |
4815 | if (DisableSCO && !TailCallOpt) return false; |
4816 | |
4817 | // Variadic argument functions are not supported. |
4818 | if (isVarArg) return false; |
4819 | |
4820 | auto &Caller = DAG.getMachineFunction().getFunction(); |
4821 | // Check that the calling conventions are compatible for tco. |
4822 | if (!areCallingConvEligibleForTCO_64SVR4(Caller.getCallingConv(), CalleeCC)) |
4823 | return false; |
4824 | |
4825 | // Caller contains any byval parameter is not supported. |
4826 | if (any_of(Ins, [](const ISD::InputArg &IA) { return IA.Flags.isByVal(); })) |
4827 | return false; |
4828 | |
4829 | // Callee contains any byval parameter is not supported, too. |
4830 | // Note: This is a quick work around, because in some cases, e.g. |
4831 | // caller's stack size > callee's stack size, we are still able to apply |
4832 | // sibling call optimization. For example, gcc is able to do SCO for caller1 |
4833 | // in the following example, but not for caller2. |
4834 | // struct test { |
4835 | // long int a; |
4836 | // char ary[56]; |
4837 | // } gTest; |
4838 | // __attribute__((noinline)) int callee(struct test v, struct test *b) { |
4839 | // b->a = v.a; |
4840 | // return 0; |
4841 | // } |
4842 | // void caller1(struct test a, struct test c, struct test *b) { |
4843 | // callee(gTest, b); } |
4844 | // void caller2(struct test *b) { callee(gTest, b); } |
4845 | if (any_of(Outs, [](const ISD::OutputArg& OA) { return OA.Flags.isByVal(); })) |
4846 | return false; |
4847 | |
4848 | // If callee and caller use different calling conventions, we cannot pass |
4849 | // parameters on stack since offsets for the parameter area may be different. |
4850 | if (Caller.getCallingConv() != CalleeCC && |
4851 | needStackSlotPassParameters(Subtarget, Outs)) |
4852 | return false; |
4853 | |
4854 | // All variants of 64-bit ELF ABIs without PC-Relative addressing require that |
4855 | // the caller and callee share the same TOC for TCO/SCO. If the caller and |
4856 | // callee potentially have different TOC bases then we cannot tail call since |
4857 | // we need to restore the TOC pointer after the call. |
4858 | // ref: https://bugzilla.mozilla.org/show_bug.cgi?id=973977 |
4859 | // We cannot guarantee this for indirect calls or calls to external functions. |
4860 | // When PC-Relative addressing is used, the concept of the TOC is no longer |
4861 | // applicable so this check is not required. |
4862 | // Check first for indirect calls. |
4863 | if (!Subtarget.isUsingPCRelativeCalls() && |
4864 | !isFunctionGlobalAddress(Callee) && !isa<ExternalSymbolSDNode>(Callee)) |
4865 | return false; |
4866 | |
4867 | // Check if we share the TOC base. |
4868 | if (!Subtarget.isUsingPCRelativeCalls() && |
4869 | !callsShareTOCBase(&Caller, Callee, getTargetMachine())) |
4870 | return false; |
4871 | |
4872 | // TCO allows altering callee ABI, so we don't have to check further. |
4873 | if (CalleeCC == CallingConv::Fast && TailCallOpt) |
4874 | return true; |
4875 | |
4876 | if (DisableSCO) return false; |
4877 | |
4878 | // If callee use the same argument list that caller is using, then we can |
4879 | // apply SCO on this case. If it is not, then we need to check if callee needs |
4880 | // stack for passing arguments. |
4881 | // PC Relative tail calls may not have a CallBase. |
4882 | // If there is no CallBase we cannot verify if we have the same argument |
4883 | // list so assume that we don't have the same argument list. |
4884 | if (CB && !hasSameArgumentList(&Caller, *CB) && |
4885 | needStackSlotPassParameters(Subtarget, Outs)) |
4886 | return false; |
4887 | else if (!CB && needStackSlotPassParameters(Subtarget, Outs)) |
4888 | return false; |
4889 | |
4890 | return true; |
4891 | } |
4892 | |
4893 | /// IsEligibleForTailCallOptimization - Check whether the call is eligible |
4894 | /// for tail call optimization. Targets which want to do tail call |
4895 | /// optimization should implement this function. |
4896 | bool |
4897 | PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee, |
4898 | CallingConv::ID CalleeCC, |
4899 | bool isVarArg, |
4900 | const SmallVectorImpl<ISD::InputArg> &Ins, |
4901 | SelectionDAG& DAG) const { |
4902 | if (!getTargetMachine().Options.GuaranteedTailCallOpt) |
4903 | return false; |
4904 | |
4905 | // Variable argument functions are not supported. |
4906 | if (isVarArg) |
4907 | return false; |
4908 | |
4909 | MachineFunction &MF = DAG.getMachineFunction(); |
4910 | CallingConv::ID CallerCC = MF.getFunction().getCallingConv(); |
4911 | if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) { |
4912 | // Functions containing by val parameters are not supported. |
4913 | for (unsigned i = 0; i != Ins.size(); i++) { |
4914 | ISD::ArgFlagsTy Flags = Ins[i].Flags; |
4915 | if (Flags.isByVal()) return false; |
4916 | } |
4917 | |
4918 | // Non-PIC/GOT tail calls are supported. |
4919 | if (getTargetMachine().getRelocationModel() != Reloc::PIC_) |
4920 | return true; |
4921 | |
4922 | // At the moment we can only do local tail calls (in same module, hidden |
4923 | // or protected) if we are generating PIC. |
4924 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) |
4925 | return G->getGlobal()->hasHiddenVisibility() |
4926 | || G->getGlobal()->hasProtectedVisibility(); |
4927 | } |
4928 | |
4929 | return false; |
4930 | } |
4931 | |
4932 | /// isCallCompatibleAddress - Return the immediate to use if the specified |
4933 | /// 32-bit value is representable in the immediate field of a BxA instruction. |
4934 | static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) { |
4935 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); |
4936 | if (!C) return nullptr; |
4937 | |
4938 | int Addr = C->getZExtValue(); |
4939 | if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero. |
4940 | SignExtend32<26>(Addr) != Addr) |
4941 | return nullptr; // Top 6 bits have to be sext of immediate. |
4942 | |
4943 | return DAG |
4944 | .getConstant( |
4945 | (int)C->getZExtValue() >> 2, SDLoc(Op), |
4946 | DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout())) |
4947 | .getNode(); |
4948 | } |
4949 | |
4950 | namespace { |
4951 | |
4952 | struct TailCallArgumentInfo { |
4953 | SDValue Arg; |
4954 | SDValue FrameIdxOp; |
4955 | int FrameIdx = 0; |
4956 | |
4957 | TailCallArgumentInfo() = default; |
4958 | }; |
4959 | |
4960 | } // end anonymous namespace |
4961 | |
4962 | /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot. |
4963 | static void StoreTailCallArgumentsToStackSlot( |
4964 | SelectionDAG &DAG, SDValue Chain, |
4965 | const SmallVectorImpl<TailCallArgumentInfo> &TailCallArgs, |
4966 | SmallVectorImpl<SDValue> &MemOpChains, const SDLoc &dl) { |
4967 | for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) { |
4968 | SDValue Arg = TailCallArgs[i].Arg; |
4969 | SDValue FIN = TailCallArgs[i].FrameIdxOp; |
4970 | int FI = TailCallArgs[i].FrameIdx; |
4971 | // Store relative to framepointer. |
4972 | MemOpChains.push_back(DAG.getStore( |
4973 | Chain, dl, Arg, FIN, |
4974 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI))); |
4975 | } |
4976 | } |
4977 | |
4978 | /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to |
4979 | /// the appropriate stack slot for the tail call optimized function call. |
4980 | static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG, SDValue Chain, |
4981 | SDValue OldRetAddr, SDValue OldFP, |
4982 | int SPDiff, const SDLoc &dl) { |
4983 | if (SPDiff) { |
4984 | // Calculate the new stack slot for the return address. |
4985 | MachineFunction &MF = DAG.getMachineFunction(); |
4986 | const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); |
4987 | const PPCFrameLowering *FL = Subtarget.getFrameLowering(); |
4988 | bool isPPC64 = Subtarget.isPPC64(); |
4989 | int SlotSize = isPPC64 ? 8 : 4; |
4990 | int NewRetAddrLoc = SPDiff + FL->getReturnSaveOffset(); |
4991 | int NewRetAddr = MF.getFrameInfo().CreateFixedObject(SlotSize, |
4992 | NewRetAddrLoc, true); |
4993 | EVT VT = isPPC64 ? MVT::i64 : MVT::i32; |
4994 | SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT); |
4995 | Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx, |
4996 | MachinePointerInfo::getFixedStack(MF, NewRetAddr)); |
4997 | } |
4998 | return Chain; |
4999 | } |
5000 | |
5001 | /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate |
5002 | /// the position of the argument. |
5003 | static void |
5004 | CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64, |
5005 | SDValue Arg, int SPDiff, unsigned ArgOffset, |
5006 | SmallVectorImpl<TailCallArgumentInfo>& TailCallArguments) { |
5007 | int Offset = ArgOffset + SPDiff; |
5008 | uint32_t OpSize = (Arg.getValueSizeInBits() + 7) / 8; |
5009 | int FI = MF.getFrameInfo().CreateFixedObject(OpSize, Offset, true); |
5010 | EVT VT = isPPC64 ? MVT::i64 : MVT::i32; |
5011 | SDValue FIN = DAG.getFrameIndex(FI, VT); |
5012 | TailCallArgumentInfo Info; |
5013 | Info.Arg = Arg; |
5014 | Info.FrameIdxOp = FIN; |
5015 | Info.FrameIdx = FI; |
5016 | TailCallArguments.push_back(Info); |
5017 | } |
5018 | |
5019 | /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address |
5020 | /// stack slot. Returns the chain as result and the loaded frame pointers in |
5021 | /// LROpOut/FPOpout. Used when tail calling. |
5022 | SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr( |
5023 | SelectionDAG &DAG, int SPDiff, SDValue Chain, SDValue &LROpOut, |
5024 | SDValue &FPOpOut, const SDLoc &dl) const { |
5025 | if (SPDiff) { |
5026 | // Load the LR and FP stack slot for later adjusting. |
5027 | EVT VT = Subtarget.isPPC64() ? MVT::i64 : MVT::i32; |
5028 | LROpOut = getReturnAddrFrameIndex(DAG); |
5029 | LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo()); |
5030 | Chain = SDValue(LROpOut.getNode(), 1); |
5031 | } |
5032 | return Chain; |
5033 | } |
5034 | |
5035 | /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified |
5036 | /// by "Src" to address "Dst" of size "Size". Alignment information is |
5037 | /// specified by the specific parameter attribute. The copy will be passed as |
5038 | /// a byval function parameter. |
5039 | /// Sometimes what we are copying is the end of a larger object, the part that |
5040 | /// does not fit in registers. |
5041 | static SDValue CreateCopyOfByValArgument(SDValue Src, SDValue Dst, |
5042 | SDValue Chain, ISD::ArgFlagsTy Flags, |
5043 | SelectionDAG &DAG, const SDLoc &dl) { |
5044 | SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32); |
5045 | return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, |
5046 | Flags.getNonZeroByValAlign(), false, false, false, |
5047 | MachinePointerInfo(), MachinePointerInfo()); |
5048 | } |
5049 | |
5050 | /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of |
5051 | /// tail calls. |
5052 | static void LowerMemOpCallTo( |
5053 | SelectionDAG &DAG, MachineFunction &MF, SDValue Chain, SDValue Arg, |
5054 | SDValue PtrOff, int SPDiff, unsigned ArgOffset, bool isPPC64, |
5055 | bool isTailCall, bool isVector, SmallVectorImpl<SDValue> &MemOpChains, |
5056 | SmallVectorImpl<TailCallArgumentInfo> &TailCallArguments, const SDLoc &dl) { |
5057 | EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout()); |
5058 | if (!isTailCall) { |
5059 | if (isVector) { |
5060 | SDValue StackPtr; |
5061 | if (isPPC64) |
5062 | StackPtr = DAG.getRegister(PPC::X1, MVT::i64); |
5063 | else |
5064 | StackPtr = DAG.getRegister(PPC::R1, MVT::i32); |
5065 | PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, |
5066 | DAG.getConstant(ArgOffset, dl, PtrVT)); |
5067 | } |
5068 | MemOpChains.push_back( |
5069 | DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo())); |
5070 | // Calculate and remember argument location. |
5071 | } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset, |
5072 | TailCallArguments); |
5073 | } |
5074 | |
5075 | static void |
5076 | PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain, |
5077 | const SDLoc &dl, int SPDiff, unsigned NumBytes, SDValue LROp, |
5078 | SDValue FPOp, |
5079 | SmallVectorImpl<TailCallArgumentInfo> &TailCallArguments) { |
5080 | // Emit a sequence of copyto/copyfrom virtual registers for arguments that |
5081 | // might overwrite each other in case of tail call optimization. |
5082 | SmallVector<SDValue, 8> MemOpChains2; |
5083 | // Do not flag preceding copytoreg stuff together with the following stuff. |
5084 | InFlag = SDValue(); |
5085 | StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments, |
5086 | MemOpChains2, dl); |
5087 | if (!MemOpChains2.empty()) |
5088 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2); |
5089 | |
5090 | // Store the return address to the appropriate stack slot. |
5091 | Chain = EmitTailCallStoreFPAndRetAddr(DAG, Chain, LROp, FPOp, SPDiff, dl); |
5092 | |
5093 | // Emit callseq_end just before tailcall node. |
5094 | Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true), |
5095 | DAG.getIntPtrConstant(0, dl, true), InFlag, dl); |
5096 | InFlag = Chain.getValue(1); |
5097 | } |
5098 | |
5099 | // Is this global address that of a function that can be called by name? (as |
5100 | // opposed to something that must hold a descriptor for an indirect call). |
5101 | static bool isFunctionGlobalAddress(SDValue Callee) { |
5102 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { |
5103 | if (Callee.getOpcode() == ISD::GlobalTLSAddress || |
5104 | Callee.getOpcode() == ISD::TargetGlobalTLSAddress) |
5105 | return false; |
5106 | |
5107 | return G->getGlobal()->getValueType()->isFunctionTy(); |
5108 | } |
5109 | |
5110 | return false; |
5111 | } |
5112 | |
5113 | SDValue PPCTargetLowering::LowerCallResult( |
5114 | SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg, |
5115 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
5116 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
5117 | SmallVector<CCValAssign, 16> RVLocs; |
5118 | CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, |
5119 | *DAG.getContext()); |
5120 | |
5121 | CCRetInfo.AnalyzeCallResult( |
5122 | Ins, (Subtarget.isSVR4ABI() && CallConv == CallingConv::Cold) |
5123 | ? RetCC_PPC_Cold |
5124 | : RetCC_PPC); |
5125 | |
5126 | // Copy all of the result registers out of their specified physreg. |
5127 | for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) { |
5128 | CCValAssign &VA = RVLocs[i]; |
5129 | 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!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5129, __PRETTY_FUNCTION__)); |
5130 | |
5131 | SDValue Val; |
5132 | |
5133 | if (Subtarget.hasSPE() && VA.getLocVT() == MVT::f64) { |
5134 | SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, |
5135 | InFlag); |
5136 | Chain = Lo.getValue(1); |
5137 | InFlag = Lo.getValue(2); |
5138 | VA = RVLocs[++i]; // skip ahead to next loc |
5139 | SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, |
5140 | InFlag); |
5141 | Chain = Hi.getValue(1); |
5142 | InFlag = Hi.getValue(2); |
5143 | if (!Subtarget.isLittleEndian()) |
5144 | std::swap (Lo, Hi); |
5145 | Val = DAG.getNode(PPCISD::BUILD_SPE64, dl, MVT::f64, Lo, Hi); |
5146 | } else { |
5147 | Val = DAG.getCopyFromReg(Chain, dl, |
5148 | VA.getLocReg(), VA.getLocVT(), InFlag); |
5149 | Chain = Val.getValue(1); |
5150 | InFlag = Val.getValue(2); |
5151 | } |
5152 | |
5153 | switch (VA.getLocInfo()) { |
5154 | default: llvm_unreachable("Unknown loc info!")::llvm::llvm_unreachable_internal("Unknown loc info!", "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5154); |
5155 | case CCValAssign::Full: break; |
5156 | case CCValAssign::AExt: |
5157 | Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val); |
5158 | break; |
5159 | case CCValAssign::ZExt: |
5160 | Val = DAG.getNode(ISD::AssertZext, dl, VA.getLocVT(), Val, |
5161 | DAG.getValueType(VA.getValVT())); |
5162 | Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val); |
5163 | break; |
5164 | case CCValAssign::SExt: |
5165 | Val = DAG.getNode(ISD::AssertSext, dl, VA.getLocVT(), Val, |
5166 | DAG.getValueType(VA.getValVT())); |
5167 | Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val); |
5168 | break; |
5169 | } |
5170 | |
5171 | InVals.push_back(Val); |
5172 | } |
5173 | |
5174 | return Chain; |
5175 | } |
5176 | |
5177 | static bool isIndirectCall(const SDValue &Callee, SelectionDAG &DAG, |
5178 | const PPCSubtarget &Subtarget, bool isPatchPoint) { |
5179 | // PatchPoint calls are not indirect. |
5180 | if (isPatchPoint) |
5181 | return false; |
5182 | |
5183 | if (isFunctionGlobalAddress(Callee) || dyn_cast<ExternalSymbolSDNode>(Callee)) |
5184 | return false; |
5185 | |
5186 | // Darwin, and 32-bit ELF can use a BLA. The descriptor based ABIs can not |
5187 | // becuase the immediate function pointer points to a descriptor instead of |
5188 | // a function entry point. The ELFv2 ABI cannot use a BLA because the function |
5189 | // pointer immediate points to the global entry point, while the BLA would |
5190 | // need to jump to the local entry point (see rL211174). |
5191 | if (!Subtarget.usesFunctionDescriptors() && !Subtarget.isELFv2ABI() && |
5192 | isBLACompatibleAddress(Callee, DAG)) |
5193 | return false; |
5194 | |
5195 | return true; |
5196 | } |
5197 | |
5198 | // AIX and 64-bit ELF ABIs w/o PCRel require a TOC save/restore around calls. |
5199 | static inline bool isTOCSaveRestoreRequired(const PPCSubtarget &Subtarget) { |
5200 | return Subtarget.isAIXABI() || |
5201 | (Subtarget.is64BitELFABI() && !Subtarget.isUsingPCRelativeCalls()); |
5202 | } |
5203 | |
5204 | static unsigned getCallOpcode(PPCTargetLowering::CallFlags CFlags, |
5205 | const Function &Caller, |
5206 | const SDValue &Callee, |
5207 | const PPCSubtarget &Subtarget, |
5208 | const TargetMachine &TM) { |
5209 | if (CFlags.IsTailCall) |
5210 | return PPCISD::TC_RETURN; |
5211 | |
5212 | // This is a call through a function pointer. |
5213 | if (CFlags.IsIndirect) { |
5214 | // AIX and the 64-bit ELF ABIs need to maintain the TOC pointer accross |
5215 | // indirect calls. The save of the caller's TOC pointer to the stack will be |
5216 | // inserted into the DAG as part of call lowering. The restore of the TOC |
5217 | // pointer is modeled by using a pseudo instruction for the call opcode that |
5218 | // represents the 2 instruction sequence of an indirect branch and link, |
5219 | // immediately followed by a load of the TOC pointer from the the stack save |
5220 | // slot into gpr2. For 64-bit ELFv2 ABI with PCRel, do not restore the TOC |
5221 | // as it is not saved or used. |
5222 | return isTOCSaveRestoreRequired(Subtarget) ? PPCISD::BCTRL_LOAD_TOC |
5223 | : PPCISD::BCTRL; |
5224 | } |
5225 | |
5226 | if (Subtarget.isUsingPCRelativeCalls()) { |
5227 | assert(Subtarget.is64BitELFABI() && "PC Relative is only on ELF ABI.")((Subtarget.is64BitELFABI() && "PC Relative is only on ELF ABI." ) ? static_cast<void> (0) : __assert_fail ("Subtarget.is64BitELFABI() && \"PC Relative is only on ELF ABI.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5227, __PRETTY_FUNCTION__)); |
5228 | return PPCISD::CALL_NOTOC; |
5229 | } |
5230 | |
5231 | // The ABIs that maintain a TOC pointer accross calls need to have a nop |
5232 | // immediately following the call instruction if the caller and callee may |
5233 | // have different TOC bases. At link time if the linker determines the calls |
5234 | // may not share a TOC base, the call is redirected to a trampoline inserted |
5235 | // by the linker. The trampoline will (among other things) save the callers |
5236 | // TOC pointer at an ABI designated offset in the linkage area and the linker |
5237 | // will rewrite the nop to be a load of the TOC pointer from the linkage area |
5238 | // into gpr2. |
5239 | if (Subtarget.isAIXABI() || Subtarget.is64BitELFABI()) |
5240 | return callsShareTOCBase(&Caller, Callee, TM) ? PPCISD::CALL |
5241 | : PPCISD::CALL_NOP; |
5242 | |
5243 | return PPCISD::CALL; |
5244 | } |
5245 | |
5246 | static SDValue transformCallee(const SDValue &Callee, SelectionDAG &DAG, |
5247 | const SDLoc &dl, const PPCSubtarget &Subtarget) { |
5248 | if (!Subtarget.usesFunctionDescriptors() && !Subtarget.isELFv2ABI()) |
5249 | if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) |
5250 | return SDValue(Dest, 0); |
5251 | |
5252 | // Returns true if the callee is local, and false otherwise. |
5253 | auto isLocalCallee = [&]() { |
5254 | const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee); |
5255 | const Module *Mod = DAG.getMachineFunction().getFunction().getParent(); |
5256 | const GlobalValue *GV = G ? G->getGlobal() : nullptr; |
5257 | |
5258 | return DAG.getTarget().shouldAssumeDSOLocal(*Mod, GV) && |
5259 | !dyn_cast_or_null<GlobalIFunc>(GV); |
5260 | }; |
5261 | |
5262 | // The PLT is only used in 32-bit ELF PIC mode. Attempting to use the PLT in |
5263 | // a static relocation model causes some versions of GNU LD (2.17.50, at |
5264 | // least) to force BSS-PLT, instead of secure-PLT, even if all objects are |
5265 | // built with secure-PLT. |
5266 | bool UsePlt = |
5267 | Subtarget.is32BitELFABI() && !isLocalCallee() && |
5268 | Subtarget.getTargetMachine().getRelocationModel() == Reloc::PIC_; |
5269 | |
5270 | const auto getAIXFuncEntryPointSymbolSDNode = [&](const GlobalValue *GV) { |
5271 | const TargetMachine &TM = Subtarget.getTargetMachine(); |
5272 | const TargetLoweringObjectFile *TLOF = TM.getObjFileLowering(); |
5273 | MCSymbolXCOFF *S = |
5274 | cast<MCSymbolXCOFF>(TLOF->getFunctionEntryPointSymbol(GV, TM)); |
5275 | |
5276 | MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout()); |
5277 | return DAG.getMCSymbol(S, PtrVT); |
5278 | }; |
5279 | |
5280 | if (isFunctionGlobalAddress(Callee)) { |
5281 | const GlobalValue *GV = cast<GlobalAddressSDNode>(Callee)->getGlobal(); |
5282 | |
5283 | if (Subtarget.isAIXABI()) { |
5284 | assert(!isa<GlobalIFunc>(GV) && "IFunc is not supported on AIX.")((!isa<GlobalIFunc>(GV) && "IFunc is not supported on AIX." ) ? static_cast<void> (0) : __assert_fail ("!isa<GlobalIFunc>(GV) && \"IFunc is not supported on AIX.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5284, __PRETTY_FUNCTION__)); |
5285 | return getAIXFuncEntryPointSymbolSDNode(GV); |
5286 | } |
5287 | return DAG.getTargetGlobalAddress(GV, dl, Callee.getValueType(), 0, |
5288 | UsePlt ? PPCII::MO_PLT : 0); |
5289 | } |
5290 | |
5291 | if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { |
5292 | const char *SymName = S->getSymbol(); |
5293 | if (Subtarget.isAIXABI()) { |
5294 | // If there exists a user-declared function whose name is the same as the |
5295 | // ExternalSymbol's, then we pick up the user-declared version. |
5296 | const Module *Mod = DAG.getMachineFunction().getFunction().getParent(); |
5297 | if (const Function *F = |
5298 | dyn_cast_or_null<Function>(Mod->getNamedValue(SymName))) |
5299 | return getAIXFuncEntryPointSymbolSDNode(F); |
5300 | |
5301 | // On AIX, direct function calls reference the symbol for the function's |
5302 | // entry point, which is named by prepending a "." before the function's |
5303 | // C-linkage name. A Qualname is returned here because an external |
5304 | // function entry point is a csect with XTY_ER property. |
5305 | const auto getExternalFunctionEntryPointSymbol = [&](StringRef SymName) { |
5306 | auto &Context = DAG.getMachineFunction().getMMI().getContext(); |
5307 | MCSectionXCOFF *Sec = Context.getXCOFFSection( |
5308 | (Twine(".") + Twine(SymName)).str(), XCOFF::XMC_PR, XCOFF::XTY_ER, |
5309 | SectionKind::getMetadata()); |
5310 | return Sec->getQualNameSymbol(); |
5311 | }; |
5312 | |
5313 | SymName = getExternalFunctionEntryPointSymbol(SymName)->getName().data(); |
5314 | } |
5315 | return DAG.getTargetExternalSymbol(SymName, Callee.getValueType(), |
5316 | UsePlt ? PPCII::MO_PLT : 0); |
5317 | } |
5318 | |
5319 | // No transformation needed. |
5320 | assert(Callee.getNode() && "What no callee?")((Callee.getNode() && "What no callee?") ? static_cast <void> (0) : __assert_fail ("Callee.getNode() && \"What no callee?\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5320, __PRETTY_FUNCTION__)); |
5321 | return Callee; |
5322 | } |
5323 | |
5324 | static SDValue getOutputChainFromCallSeq(SDValue CallSeqStart) { |
5325 | assert(CallSeqStart.getOpcode() == ISD::CALLSEQ_START &&((CallSeqStart.getOpcode() == ISD::CALLSEQ_START && "Expected a CALLSEQ_STARTSDNode." ) ? static_cast<void> (0) : __assert_fail ("CallSeqStart.getOpcode() == ISD::CALLSEQ_START && \"Expected a CALLSEQ_STARTSDNode.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5326, __PRETTY_FUNCTION__)) |
5326 | "Expected a CALLSEQ_STARTSDNode.")((CallSeqStart.getOpcode() == ISD::CALLSEQ_START && "Expected a CALLSEQ_STARTSDNode." ) ? static_cast<void> (0) : __assert_fail ("CallSeqStart.getOpcode() == ISD::CALLSEQ_START && \"Expected a CALLSEQ_STARTSDNode.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5326, __PRETTY_FUNCTION__)); |
5327 | |
5328 | // The last operand is the chain, except when the node has glue. If the node |
5329 | // has glue, then the last operand is the glue, and the chain is the second |
5330 | // last operand. |
5331 | SDValue LastValue = CallSeqStart.getValue(CallSeqStart->getNumValues() - 1); |
5332 | if (LastValue.getValueType() != MVT::Glue) |
5333 | return LastValue; |
5334 | |
5335 | return CallSeqStart.getValue(CallSeqStart->getNumValues() - 2); |
5336 | } |
5337 | |
5338 | // Creates the node that moves a functions address into the count register |
5339 | // to prepare for an indirect call instruction. |
5340 | static void prepareIndirectCall(SelectionDAG &DAG, SDValue &Callee, |
5341 | SDValue &Glue, SDValue &Chain, |
5342 | const SDLoc &dl) { |
5343 | SDValue MTCTROps[] = {Chain, Callee, Glue}; |
5344 | EVT ReturnTypes[] = {MVT::Other, MVT::Glue}; |
5345 | Chain = DAG.getNode(PPCISD::MTCTR, dl, makeArrayRef(ReturnTypes, 2), |
5346 | makeArrayRef(MTCTROps, Glue.getNode() ? 3 : 2)); |
5347 | // The glue is the second value produced. |
5348 | Glue = Chain.getValue(1); |
5349 | } |
5350 | |
5351 | static void prepareDescriptorIndirectCall(SelectionDAG &DAG, SDValue &Callee, |
5352 | SDValue &Glue, SDValue &Chain, |
5353 | SDValue CallSeqStart, |
5354 | const CallBase *CB, const SDLoc &dl, |
5355 | bool hasNest, |
5356 | const PPCSubtarget &Subtarget) { |
5357 | // Function pointers in the 64-bit SVR4 ABI do not point to the function |
5358 | // entry point, but to the function descriptor (the function entry point |
5359 | // address is part of the function descriptor though). |
5360 | // The function descriptor is a three doubleword structure with the |
5361 | // following fields: function entry point, TOC base address and |
5362 | // environment pointer. |
5363 | // Thus for a call through a function pointer, the following actions need |
5364 | // to be performed: |
5365 | // 1. Save the TOC of the caller in the TOC save area of its stack |
5366 | // frame (this is done in LowerCall_Darwin() or LowerCall_64SVR4()). |
5367 | // 2. Load the address of the function entry point from the function |
5368 | // descriptor. |
5369 | // 3. Load the TOC of the callee from the function descriptor into r2. |
5370 | // 4. Load the environment pointer from the function descriptor into |
5371 | // r11. |
5372 | // 5. Branch to the function entry point address. |
5373 | // 6. On return of the callee, the TOC of the caller needs to be |
5374 | // restored (this is done in FinishCall()). |
5375 | // |
5376 | // The loads are scheduled at the beginning of the call sequence, and the |
5377 | // register copies are flagged together to ensure that no other |
5378 | // operations can be scheduled in between. E.g. without flagging the |
5379 | // copies together, a TOC access in the caller could be scheduled between |
5380 | // the assignment of the callee TOC and the branch to the callee, which leads |
5381 | // to incorrect code. |
5382 | |
5383 | // Start by loading the function address from the descriptor. |
5384 | SDValue LDChain = getOutputChainFromCallSeq(CallSeqStart); |
5385 | auto MMOFlags = Subtarget.hasInvariantFunctionDescriptors() |
5386 | ? (MachineMemOperand::MODereferenceable | |
5387 | MachineMemOperand::MOInvariant) |
5388 | : MachineMemOperand::MONone; |
5389 | |
5390 | MachinePointerInfo MPI(CB ? CB->getCalledOperand() : nullptr); |
5391 | |
5392 | // Registers used in building the DAG. |
5393 | const MCRegister EnvPtrReg = Subtarget.getEnvironmentPointerRegister(); |
5394 | const MCRegister TOCReg = Subtarget.getTOCPointerRegister(); |
5395 | |
5396 | // Offsets of descriptor members. |
5397 | const unsigned TOCAnchorOffset = Subtarget.descriptorTOCAnchorOffset(); |
5398 | const unsigned EnvPtrOffset = Subtarget.descriptorEnvironmentPointerOffset(); |
5399 | |
5400 | const MVT RegVT = Subtarget.isPPC64() ? MVT::i64 : MVT::i32; |
5401 | const unsigned Alignment = Subtarget.isPPC64() ? 8 : 4; |
5402 | |
5403 | // One load for the functions entry point address. |
5404 | SDValue LoadFuncPtr = DAG.getLoad(RegVT, dl, LDChain, Callee, MPI, |
5405 | Alignment, MMOFlags); |
5406 | |
5407 | // One for loading the TOC anchor for the module that contains the called |
5408 | // function. |
5409 | SDValue TOCOff = DAG.getIntPtrConstant(TOCAnchorOffset, dl); |
5410 | SDValue AddTOC = DAG.getNode(ISD::ADD, dl, RegVT, Callee, TOCOff); |
5411 | SDValue TOCPtr = |
5412 | DAG.getLoad(RegVT, dl, LDChain, AddTOC, |
5413 | MPI.getWithOffset(TOCAnchorOffset), Alignment, MMOFlags); |
5414 | |
5415 | // One for loading the environment pointer. |
5416 | SDValue PtrOff = DAG.getIntPtrConstant(EnvPtrOffset, dl); |
5417 | SDValue AddPtr = DAG.getNode(ISD::ADD, dl, RegVT, Callee, PtrOff); |
5418 | SDValue LoadEnvPtr = |
5419 | DAG.getLoad(RegVT, dl, LDChain, AddPtr, |
5420 | MPI.getWithOffset(EnvPtrOffset), Alignment, MMOFlags); |
5421 | |
5422 | |
5423 | // Then copy the newly loaded TOC anchor to the TOC pointer. |
5424 | SDValue TOCVal = DAG.getCopyToReg(Chain, dl, TOCReg, TOCPtr, Glue); |
5425 | Chain = TOCVal.getValue(0); |
5426 | Glue = TOCVal.getValue(1); |
5427 | |
5428 | // If the function call has an explicit 'nest' parameter, it takes the |
5429 | // place of the environment pointer. |
5430 | assert((!hasNest || !Subtarget.isAIXABI()) &&(((!hasNest || !Subtarget.isAIXABI()) && "Nest parameter is not supported on AIX." ) ? static_cast<void> (0) : __assert_fail ("(!hasNest || !Subtarget.isAIXABI()) && \"Nest parameter is not supported on AIX.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5431, __PRETTY_FUNCTION__)) |
5431 | "Nest parameter is not supported on AIX.")(((!hasNest || !Subtarget.isAIXABI()) && "Nest parameter is not supported on AIX." ) ? static_cast<void> (0) : __assert_fail ("(!hasNest || !Subtarget.isAIXABI()) && \"Nest parameter is not supported on AIX.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5431, __PRETTY_FUNCTION__)); |
5432 | if (!hasNest) { |
5433 | SDValue EnvVal = DAG.getCopyToReg(Chain, dl, EnvPtrReg, LoadEnvPtr, Glue); |
5434 | Chain = EnvVal.getValue(0); |
5435 | Glue = EnvVal.getValue(1); |
5436 | } |
5437 | |
5438 | // The rest of the indirect call sequence is the same as the non-descriptor |
5439 | // DAG. |
5440 | prepareIndirectCall(DAG, LoadFuncPtr, Glue, Chain, dl); |
5441 | } |
5442 | |
5443 | static void |
5444 | buildCallOperands(SmallVectorImpl<SDValue> &Ops, |
5445 | PPCTargetLowering::CallFlags CFlags, const SDLoc &dl, |
5446 | SelectionDAG &DAG, |
5447 | SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass, |
5448 | SDValue Glue, SDValue Chain, SDValue &Callee, int SPDiff, |
5449 | const PPCSubtarget &Subtarget) { |
5450 | const bool IsPPC64 = Subtarget.isPPC64(); |
5451 | // MVT for a general purpose register. |
5452 | const MVT RegVT = IsPPC64 ? MVT::i64 : MVT::i32; |
5453 | |
5454 | // First operand is always the chain. |
5455 | Ops.push_back(Chain); |
5456 | |
5457 | // If it's a direct call pass the callee as the second operand. |
5458 | if (!CFlags.IsIndirect) |
5459 | Ops.push_back(Callee); |
5460 | else { |
5461 | assert(!CFlags.IsPatchPoint && "Patch point calls are not indirect.")((!CFlags.IsPatchPoint && "Patch point calls are not indirect." ) ? static_cast<void> (0) : __assert_fail ("!CFlags.IsPatchPoint && \"Patch point calls are not indirect.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5461, __PRETTY_FUNCTION__)); |
5462 | |
5463 | // For the TOC based ABIs, we have saved the TOC pointer to the linkage area |
5464 | // on the stack (this would have been done in `LowerCall_64SVR4` or |
5465 | // `LowerCall_AIX`). The call instruction is a pseudo instruction that |
5466 | // represents both the indirect branch and a load that restores the TOC |
5467 | // pointer from the linkage area. The operand for the TOC restore is an add |
5468 | // of the TOC save offset to the stack pointer. This must be the second |
5469 | // operand: after the chain input but before any other variadic arguments. |
5470 | // For 64-bit ELFv2 ABI with PCRel, do not restore the TOC as it is not |
5471 | // saved or used. |
5472 | if (isTOCSaveRestoreRequired(Subtarget)) { |
5473 | const MCRegister StackPtrReg = Subtarget.getStackPointerRegister(); |
5474 | |
5475 | SDValue StackPtr = DAG.getRegister(StackPtrReg, RegVT); |
5476 | unsigned TOCSaveOffset = Subtarget.getFrameLowering()->getTOCSaveOffset(); |
5477 | SDValue TOCOff = DAG.getIntPtrConstant(TOCSaveOffset, dl); |
5478 | SDValue AddTOC = DAG.getNode(ISD::ADD, dl, RegVT, StackPtr, TOCOff); |
5479 | Ops.push_back(AddTOC); |
5480 | } |
5481 | |
5482 | // Add the register used for the environment pointer. |
5483 | if (Subtarget.usesFunctionDescriptors() && !CFlags.HasNest) |
5484 | Ops.push_back(DAG.getRegister(Subtarget.getEnvironmentPointerRegister(), |
5485 | RegVT)); |
5486 | |
5487 | |
5488 | // Add CTR register as callee so a bctr can be emitted later. |
5489 | if (CFlags.IsTailCall) |
5490 | Ops.push_back(DAG.getRegister(IsPPC64 ? PPC::CTR8 : PPC::CTR, RegVT)); |
5491 | } |
5492 | |
5493 | // If this is a tail call add stack pointer delta. |
5494 | if (CFlags.IsTailCall) |
5495 | Ops.push_back(DAG.getConstant(SPDiff, dl, MVT::i32)); |
5496 | |
5497 | // Add argument registers to the end of the list so that they are known live |
5498 | // into the call. |
5499 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) |
5500 | Ops.push_back(DAG.getRegister(RegsToPass[i].first, |
5501 | RegsToPass[i].second.getValueType())); |
5502 | |
5503 | // We cannot add R2/X2 as an operand here for PATCHPOINT, because there is |
5504 | // no way to mark dependencies as implicit here. |
5505 | // We will add the R2/X2 dependency in EmitInstrWithCustomInserter. |
5506 | if ((Subtarget.is64BitELFABI() || Subtarget.isAIXABI()) && |
5507 | !CFlags.IsPatchPoint && !Subtarget.isUsingPCRelativeCalls()) |
5508 | Ops.push_back(DAG.getRegister(Subtarget.getTOCPointerRegister(), RegVT)); |
5509 | |
5510 | // Add implicit use of CR bit 6 for 32-bit SVR4 vararg calls |
5511 | if (CFlags.IsVarArg && Subtarget.is32BitELFABI()) |
5512 | Ops.push_back(DAG.getRegister(PPC::CR1EQ, MVT::i32)); |
5513 | |
5514 | // Add a register mask operand representing the call-preserved registers. |
5515 | const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); |
5516 | const uint32_t *Mask = |
5517 | TRI->getCallPreservedMask(DAG.getMachineFunction(), CFlags.CallConv); |
5518 | 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\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5518, __PRETTY_FUNCTION__)); |
5519 | Ops.push_back(DAG.getRegisterMask(Mask)); |
5520 | |
5521 | // If the glue is valid, it is the last operand. |
5522 | if (Glue.getNode()) |
5523 | Ops.push_back(Glue); |
5524 | } |
5525 | |
5526 | SDValue PPCTargetLowering::FinishCall( |
5527 | CallFlags CFlags, const SDLoc &dl, SelectionDAG &DAG, |
5528 | SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass, SDValue Glue, |
5529 | SDValue Chain, SDValue CallSeqStart, SDValue &Callee, int SPDiff, |
5530 | unsigned NumBytes, const SmallVectorImpl<ISD::InputArg> &Ins, |
5531 | SmallVectorImpl<SDValue> &InVals, const CallBase *CB) const { |
5532 | |
5533 | if ((Subtarget.is64BitELFABI() && !Subtarget.isUsingPCRelativeCalls()) || |
5534 | Subtarget.isAIXABI()) |
5535 | setUsesTOCBasePtr(DAG); |
5536 | |
5537 | unsigned CallOpc = |
5538 | getCallOpcode(CFlags, DAG.getMachineFunction().getFunction(), Callee, |
5539 | Subtarget, DAG.getTarget()); |
5540 | |
5541 | if (!CFlags.IsIndirect) |
5542 | Callee = transformCallee(Callee, DAG, dl, Subtarget); |
5543 | else if (Subtarget.usesFunctionDescriptors()) |
5544 | prepareDescriptorIndirectCall(DAG, Callee, Glue, Chain, CallSeqStart, CB, |
5545 | dl, CFlags.HasNest, Subtarget); |
5546 | else |
5547 | prepareIndirectCall(DAG, Callee, Glue, Chain, dl); |
5548 | |
5549 | // Build the operand list for the call instruction. |
5550 | SmallVector<SDValue, 8> Ops; |
5551 | buildCallOperands(Ops, CFlags, dl, DAG, RegsToPass, Glue, Chain, Callee, |
5552 | SPDiff, Subtarget); |
5553 | |
5554 | // Emit tail call. |
5555 | if (CFlags.IsTailCall) { |
5556 | // Indirect tail call when using PC Relative calls do not have the same |
5557 | // constraints. |
5558 | assert(((Callee.getOpcode() == ISD::Register &&((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5566, __PRETTY_FUNCTION__)) |
5559 | cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5566, __PRETTY_FUNCTION__)) |
5560 | Callee.getOpcode() == ISD::TargetExternalSymbol ||((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5566, __PRETTY_FUNCTION__)) |
5561 | Callee.getOpcode() == ISD::TargetGlobalAddress ||((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5566, __PRETTY_FUNCTION__)) |
5562 | isa<ConstantSDNode>(Callee) ||((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5566, __PRETTY_FUNCTION__)) |
5563 | (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) &&((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5566, __PRETTY_FUNCTION__)) |
5564 | "Expecting a global address, external symbol, absolute value, "((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5566, __PRETTY_FUNCTION__)) |
5565 | "register or an indirect tail call when PC Relative calls are "((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5566, __PRETTY_FUNCTION__)) |
5566 | "used.")((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5566, __PRETTY_FUNCTION__)); |
5567 | // PC Relative calls also use TC_RETURN as the way to mark tail calls. |
5568 | assert(CallOpc == PPCISD::TC_RETURN &&((CallOpc == PPCISD::TC_RETURN && "Unexpected call opcode for a tail call." ) ? static_cast<void> (0) : __assert_fail ("CallOpc == PPCISD::TC_RETURN && \"Unexpected call opcode for a tail call.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5569, __PRETTY_FUNCTION__)) |
5569 | "Unexpected call opcode for a tail call.")((CallOpc == PPCISD::TC_RETURN && "Unexpected call opcode for a tail call." ) ? static_cast<void> (0) : __assert_fail ("CallOpc == PPCISD::TC_RETURN && \"Unexpected call opcode for a tail call.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5569, __PRETTY_FUNCTION__)); |
5570 | DAG.getMachineFunction().getFrameInfo().setHasTailCall(); |
5571 | return DAG.getNode(CallOpc, dl, MVT::Other, Ops); |
5572 | } |
5573 | |
5574 | std::array<EVT, 2> ReturnTypes = {{MVT::Other, MVT::Glue}}; |
5575 | Chain = DAG.getNode(CallOpc, dl, ReturnTypes, Ops); |
5576 | DAG.addNoMergeSiteInfo(Chain.getNode(), CFlags.NoMerge); |
5577 | Glue = Chain.getValue(1); |
5578 | |
5579 | // When performing tail call optimization the callee pops its arguments off |
5580 | // the stack. Account for this here so these bytes can be pushed back on in |
5581 | // PPCFrameLowering::eliminateCallFramePseudoInstr. |
5582 | int BytesCalleePops = (CFlags.CallConv == CallingConv::Fast && |
5583 | getTargetMachine().Options.GuaranteedTailCallOpt) |
5584 | ? NumBytes |
5585 | : 0; |
5586 | |
5587 | Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true), |
5588 | DAG.getIntPtrConstant(BytesCalleePops, dl, true), |
5589 | Glue, dl); |
5590 | Glue = Chain.getValue(1); |
5591 | |
5592 | return LowerCallResult(Chain, Glue, CFlags.CallConv, CFlags.IsVarArg, Ins, dl, |
5593 | DAG, InVals); |
5594 | } |
5595 | |
5596 | SDValue |
5597 | PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, |
5598 | SmallVectorImpl<SDValue> &InVals) const { |
5599 | SelectionDAG &DAG = CLI.DAG; |
5600 | SDLoc &dl = CLI.DL; |
5601 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; |
5602 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; |
5603 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; |
5604 | SDValue Chain = CLI.Chain; |
5605 | SDValue Callee = CLI.Callee; |
5606 | bool &isTailCall = CLI.IsTailCall; |
5607 | CallingConv::ID CallConv = CLI.CallConv; |
5608 | bool isVarArg = CLI.IsVarArg; |
5609 | bool isPatchPoint = CLI.IsPatchPoint; |
5610 | const CallBase *CB = CLI.CB; |
5611 | |
5612 | if (isTailCall) { |
5613 | if (Subtarget.useLongCalls() && !(CB && CB->isMustTailCall())) |
5614 | isTailCall = false; |
5615 | else if (Subtarget.isSVR4ABI() && Subtarget.isPPC64()) |
5616 | isTailCall = IsEligibleForTailCallOptimization_64SVR4( |
5617 | Callee, CallConv, CB, isVarArg, Outs, Ins, DAG); |
5618 | else |
5619 | isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg, |
5620 | Ins, DAG); |
5621 | if (isTailCall) { |
5622 | ++NumTailCalls; |
5623 | if (!getTargetMachine().Options.GuaranteedTailCallOpt) |
5624 | ++NumSiblingCalls; |
5625 | |
5626 | // PC Relative calls no longer guarantee that the callee is a Global |
5627 | // Address Node. The callee could be an indirect tail call in which |
5628 | // case the SDValue for the callee could be a load (to load the address |
5629 | // of a function pointer) or it may be a register copy (to move the |
5630 | // address of the callee from a function parameter into a virtual |
5631 | // register). It may also be an ExternalSymbolSDNode (ex memcopy). |
5632 | assert((Subtarget.isUsingPCRelativeCalls() ||(((Subtarget.isUsingPCRelativeCalls() || isa<GlobalAddressSDNode >(Callee)) && "Callee should be an llvm::Function object." ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.isUsingPCRelativeCalls() || isa<GlobalAddressSDNode>(Callee)) && \"Callee should be an llvm::Function object.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5634, __PRETTY_FUNCTION__)) |
5633 | isa<GlobalAddressSDNode>(Callee)) &&(((Subtarget.isUsingPCRelativeCalls() || isa<GlobalAddressSDNode >(Callee)) && "Callee should be an llvm::Function object." ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.isUsingPCRelativeCalls() || isa<GlobalAddressSDNode>(Callee)) && \"Callee should be an llvm::Function object.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5634, __PRETTY_FUNCTION__)) |
5634 | "Callee should be an llvm::Function object.")(((Subtarget.isUsingPCRelativeCalls() || isa<GlobalAddressSDNode >(Callee)) && "Callee should be an llvm::Function object." ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.isUsingPCRelativeCalls() || isa<GlobalAddressSDNode>(Callee)) && \"Callee should be an llvm::Function object.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5634, __PRETTY_FUNCTION__)); |
5635 | |
5636 | LLVM_DEBUG(dbgs() << "TCO caller: " << DAG.getMachineFunction().getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-lowering")) { dbgs() << "TCO caller: " << DAG .getMachineFunction().getName() << "\nTCO callee: "; } } while (false) |
5637 | << "\nTCO callee: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-lowering")) { dbgs() << "TCO caller: " << DAG .getMachineFunction().getName() << "\nTCO callee: "; } } while (false); |
5638 | LLVM_DEBUG(Callee.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-lowering")) { Callee.dump(); } } while (false); |
5639 | } |
5640 | } |
5641 | |
5642 | if (!isTailCall && CB && CB->isMustTailCall()) |
5643 | report_fatal_error("failed to perform tail call elimination on a call " |
5644 | "site marked musttail"); |
5645 | |
5646 | // When long calls (i.e. indirect calls) are always used, calls are always |
5647 | // made via function pointer. If we have a function name, first translate it |
5648 | // into a pointer. |
5649 | if (Subtarget.useLongCalls() && isa<GlobalAddressSDNode>(Callee) && |
5650 | !isTailCall) |
5651 | Callee = LowerGlobalAddress(Callee, DAG); |
5652 | |
5653 | CallFlags CFlags( |
5654 | CallConv, isTailCall, isVarArg, isPatchPoint, |
5655 | isIndirectCall(Callee, DAG, Subtarget, isPatchPoint), |
5656 | // hasNest |
5657 | Subtarget.is64BitELFABI() && |
5658 | any_of(Outs, [](ISD::OutputArg Arg) { return Arg.Flags.isNest(); }), |
5659 | CLI.NoMerge); |
5660 | |
5661 | if (Subtarget.isSVR4ABI() && Subtarget.isPPC64()) |
5662 | return LowerCall_64SVR4(Chain, Callee, CFlags, Outs, OutVals, Ins, dl, DAG, |
5663 | InVals, CB); |
5664 | |
5665 | if (Subtarget.isSVR4ABI()) |
5666 | return LowerCall_32SVR4(Chain, Callee, CFlags, Outs, OutVals, Ins, dl, DAG, |
5667 | InVals, CB); |
5668 | |
5669 | if (Subtarget.isAIXABI()) |
5670 | return LowerCall_AIX(Chain, Callee, CFlags, Outs, OutVals, Ins, dl, DAG, |
5671 | InVals, CB); |
5672 | |
5673 | return LowerCall_Darwin(Chain, Callee, CFlags, Outs, OutVals, Ins, dl, DAG, |
5674 | InVals, CB); |
5675 | } |
5676 | |
5677 | SDValue PPCTargetLowering::LowerCall_32SVR4( |
5678 | SDValue Chain, SDValue Callee, CallFlags CFlags, |
5679 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
5680 | const SmallVectorImpl<SDValue> &OutVals, |
5681 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
5682 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, |
5683 | const CallBase *CB) const { |
5684 | // See PPCTargetLowering::LowerFormalArguments_32SVR4() for a description |
5685 | // of the 32-bit SVR4 ABI stack frame layout. |
5686 | |
5687 | const CallingConv::ID CallConv = CFlags.CallConv; |
5688 | const bool IsVarArg = CFlags.IsVarArg; |
5689 | const bool IsTailCall = CFlags.IsTailCall; |
5690 | |
5691 | assert((CallConv == CallingConv::C ||(((CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && "Unknown calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && \"Unknown calling convention!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5693, __PRETTY_FUNCTION__)) |
5692 | CallConv == CallingConv::Cold ||(((CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && "Unknown calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && \"Unknown calling convention!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5693, __PRETTY_FUNCTION__)) |
5693 | CallConv == CallingConv::Fast) && "Unknown calling convention!")(((CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && "Unknown calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && \"Unknown calling convention!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5693, __PRETTY_FUNCTION__)); |
5694 | |
5695 | const Align PtrAlign(4); |
5696 | |
5697 | MachineFunction &MF = DAG.getMachineFunction(); |
5698 | |
5699 | // Mark this function as potentially containing a function that contains a |
5700 | // tail call. As a consequence the frame pointer will be used for dynamicalloc |
5701 | // and restoring the callers stack pointer in this functions epilog. This is |
5702 | // done because by tail calling the called function might overwrite the value |
5703 | // in this function's (MF) stack pointer stack slot 0(SP). |
5704 | if (getTargetMachine().Options.GuaranteedTailCallOpt && |
5705 | CallConv == CallingConv::Fast) |
5706 | MF.getInfo<PPCFunctionInfo>()->setHasFastCall(); |
5707 | |
5708 | // Count how many bytes are to be pushed on the stack, including the linkage |
5709 | // area, parameter list area and the part of the local variable space which |
5710 | // contains copies of aggregates which are passed by value. |
5711 | |
5712 | // Assign locations to all of the outgoing arguments. |
5713 | SmallVector<CCValAssign, 16> ArgLocs; |
5714 | PPCCCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext()); |
5715 | |
5716 | // Reserve space for the linkage area on the stack. |
5717 | CCInfo.AllocateStack(Subtarget.getFrameLowering()->getLinkageSize(), |
5718 | PtrAlign); |
5719 | if (useSoftFloat()) |
5720 | CCInfo.PreAnalyzeCallOperands(Outs); |
5721 | |
5722 | if (IsVarArg) { |
5723 | // Handle fixed and variable vector arguments differently. |
5724 | // Fixed vector arguments go into registers as long as registers are |
5725 | // available. Variable vector arguments always go into memory. |
5726 | unsigned NumArgs = Outs.size(); |
5727 | |
5728 | for (unsigned i = 0; i != NumArgs; ++i) { |
5729 | MVT ArgVT = Outs[i].VT; |
5730 | ISD::ArgFlagsTy ArgFlags = Outs[i].Flags; |
5731 | bool Result; |
5732 | |
5733 | if (Outs[i].IsFixed) { |
5734 | Result = CC_PPC32_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, |
5735 | CCInfo); |
5736 | } else { |
5737 | Result = CC_PPC32_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full, |
5738 | ArgFlags, CCInfo); |
5739 | } |
5740 | |
5741 | if (Result) { |
5742 | #ifndef NDEBUG |
5743 | errs() << "Call operand #" << i << " has unhandled type " |
5744 | << EVT(ArgVT).getEVTString() << "\n"; |
5745 | #endif |
5746 | llvm_unreachable(nullptr)::llvm::llvm_unreachable_internal(nullptr, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5746); |
5747 | } |
5748 | } |
5749 | } else { |
5750 | // All arguments are treated the same. |
5751 | CCInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4); |
5752 | } |
5753 | CCInfo.clearWasPPCF128(); |
5754 | |
5755 | // Assign locations to all of the outgoing aggregate by value arguments. |
5756 | SmallVector<CCValAssign, 16> ByValArgLocs; |
5757 | CCState CCByValInfo(CallConv, IsVarArg, MF, ByValArgLocs, *DAG.getContext()); |
5758 | |
5759 | // Reserve stack space for the allocations in CCInfo. |
5760 | CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrAlign); |
5761 | |
5762 | CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4_ByVal); |
5763 | |
5764 | // Size of the linkage area, parameter list area and the part of the local |
5765 | // space variable where copies of aggregates which are passed by value are |
5766 | // stored. |
5767 | unsigned NumBytes = CCByValInfo.getNextStackOffset(); |
5768 | |
5769 | // Calculate by how many bytes the stack has to be adjusted in case of tail |
5770 | // call optimization. |
5771 | int SPDiff = CalculateTailCallSPDiff(DAG, IsTailCall, NumBytes); |
5772 | |
5773 | // Adjust the stack pointer for the new arguments... |
5774 | // These operations are automatically eliminated by the prolog/epilog pass |
5775 | Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl); |
5776 | SDValue CallSeqStart = Chain; |
5777 | |
5778 | // Load the return address and frame pointer so it can be moved somewhere else |
5779 | // later. |
5780 | SDValue LROp, FPOp; |
5781 | Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, dl); |
5782 | |
5783 | // Set up a copy of the stack pointer for use loading and storing any |
5784 | // arguments that may not fit in the registers available for argument |
5785 | // passing. |
5786 | SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32); |
5787 | |
5788 | SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; |
5789 | SmallVector<TailCallArgumentInfo, 8> TailCallArguments; |
5790 | SmallVector<SDValue, 8> MemOpChains; |
5791 | |
5792 | bool seenFloatArg = false; |
5793 | // Walk the register/memloc assignments, inserting copies/loads. |
5794 | // i - Tracks the index into the list of registers allocated for the call |
5795 | // RealArgIdx - Tracks the index into the list of actual function arguments |
5796 | // j - Tracks the index into the list of byval arguments |
5797 | for (unsigned i = 0, RealArgIdx = 0, j = 0, e = ArgLocs.size(); |
5798 | i != e; |
5799 | ++i, ++RealArgIdx) { |
5800 | CCValAssign &VA = ArgLocs[i]; |
5801 | SDValue Arg = OutVals[RealArgIdx]; |
5802 | ISD::ArgFlagsTy Flags = Outs[RealArgIdx].Flags; |
5803 | |
5804 | if (Flags.isByVal()) { |
5805 | // Argument is an aggregate which is passed by value, thus we need to |
5806 | // create a copy of it in the local variable space of the current stack |
5807 | // frame (which is the stack frame of the caller) and pass the address of |
5808 | // this copy to the callee. |
5809 | assert((j < ByValArgLocs.size()) && "Index out of bounds!")(((j < ByValArgLocs.size()) && "Index out of bounds!" ) ? static_cast<void> (0) : __assert_fail ("(j < ByValArgLocs.size()) && \"Index out of bounds!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5809, __PRETTY_FUNCTION__)); |
5810 | CCValAssign &ByValVA = ByValArgLocs[j++]; |
5811 | assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!")(((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!" ) ? static_cast<void> (0) : __assert_fail ("(VA.getValNo() == ByValVA.getValNo()) && \"ValNo mismatch!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5811, __PRETTY_FUNCTION__)); |
5812 | |
5813 | // Memory reserved in the local variable space of the callers stack frame. |
5814 | unsigned LocMemOffset = ByValVA.getLocMemOffset(); |
5815 | |
5816 | SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl); |
5817 | PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(MF.getDataLayout()), |
5818 | StackPtr, PtrOff); |
5819 | |
5820 | // Create a copy of the argument in the local area of the current |
5821 | // stack frame. |
5822 | SDValue MemcpyCall = |
5823 | CreateCopyOfByValArgument(Arg, PtrOff, |
5824 | CallSeqStart.getNode()->getOperand(0), |
5825 | Flags, DAG, dl); |
5826 | |
5827 | // This must go outside the CALLSEQ_START..END. |
5828 | SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall, NumBytes, 0, |
5829 | SDLoc(MemcpyCall)); |
5830 | DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), |
5831 | NewCallSeqStart.getNode()); |
5832 | Chain = CallSeqStart = NewCallSeqStart; |
5833 | |
5834 | // Pass the address of the aggregate copy on the stack either in a |
5835 | // physical register or in the parameter list area of the current stack |
5836 | // frame to the callee. |
5837 | Arg = PtrOff; |
5838 | } |
5839 | |
5840 | // When useCRBits() is true, there can be i1 arguments. |
5841 | // It is because getRegisterType(MVT::i1) => MVT::i1, |
5842 | // and for other integer types getRegisterType() => MVT::i32. |
5843 | // Extend i1 and ensure callee will get i32. |
5844 | if (Arg.getValueType() == MVT::i1) |
5845 | Arg = DAG.getNode(Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND, |
5846 | dl, MVT::i32, Arg); |
5847 | |
5848 | if (VA.isRegLoc()) { |
5849 | seenFloatArg |= VA.getLocVT().isFloatingPoint(); |
5850 | // Put argument in a physical register. |
5851 | if (Subtarget.hasSPE() && Arg.getValueType() == MVT::f64) { |
5852 | bool IsLE = Subtarget.isLittleEndian(); |
5853 | SDValue SVal = DAG.getNode(PPCISD::EXTRACT_SPE, dl, MVT::i32, Arg, |
5854 | DAG.getIntPtrConstant(IsLE ? 0 : 1, dl)); |
5855 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), SVal.getValue(0))); |
5856 | SVal = DAG.getNode(PPCISD::EXTRACT_SPE, dl, MVT::i32, Arg, |
5857 | DAG.getIntPtrConstant(IsLE ? 1 : 0, dl)); |
5858 | RegsToPass.push_back(std::make_pair(ArgLocs[++i].getLocReg(), |
5859 | SVal.getValue(0))); |
5860 | } else |
5861 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); |
5862 | } else { |
5863 | // Put argument in the parameter list area of the current stack frame. |
5864 | assert(VA.isMemLoc())((VA.isMemLoc()) ? static_cast<void> (0) : __assert_fail ("VA.isMemLoc()", "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5864, __PRETTY_FUNCTION__)); |
5865 | unsigned LocMemOffset = VA.getLocMemOffset(); |
5866 | |
5867 | if (!IsTailCall) { |
5868 | SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl); |
5869 | PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(MF.getDataLayout()), |
5870 | StackPtr, PtrOff); |
5871 | |
5872 | MemOpChains.push_back( |
5873 | DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo())); |
5874 | } else { |
5875 | // Calculate and remember argument location. |
5876 | CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset, |
5877 | TailCallArguments); |
5878 | } |
5879 | } |
5880 | } |
5881 | |
5882 | if (!MemOpChains.empty()) |
5883 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains); |
5884 | |
5885 | // Build a sequence of copy-to-reg nodes chained together with token chain |
5886 | // and flag operands which copy the outgoing args into the appropriate regs. |
5887 | SDValue InFlag; |
5888 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { |
5889 | Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, |
5890 | RegsToPass[i].second, InFlag); |
5891 | InFlag = Chain.getValue(1); |
5892 | } |
5893 | |
5894 | // Set CR bit 6 to true if this is a vararg call with floating args passed in |
5895 | // registers. |
5896 | if (IsVarArg) { |
5897 | SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue); |
5898 | SDValue Ops[] = { Chain, InFlag }; |
5899 | |
5900 | Chain = DAG.getNode(seenFloatArg ? PPCISD::CR6SET : PPCISD::CR6UNSET, |
5901 | dl, VTs, makeArrayRef(Ops, InFlag.getNode() ? 2 : 1)); |
5902 | |
5903 | InFlag = Chain.getValue(1); |
5904 | } |
5905 | |
5906 | if (IsTailCall) |
5907 | PrepareTailCall(DAG, InFlag, Chain, dl, SPDiff, NumBytes, LROp, FPOp, |
5908 | TailCallArguments); |
5909 | |
5910 | return FinishCall(CFlags, dl, DAG, RegsToPass, InFlag, Chain, CallSeqStart, |
5911 | Callee, SPDiff, NumBytes, Ins, InVals, CB); |
5912 | } |
5913 | |
5914 | // Copy an argument into memory, being careful to do this outside the |
5915 | // call sequence for the call to which the argument belongs. |
5916 | SDValue PPCTargetLowering::createMemcpyOutsideCallSeq( |
5917 | SDValue Arg, SDValue PtrOff, SDValue CallSeqStart, ISD::ArgFlagsTy Flags, |
5918 | SelectionDAG &DAG, const SDLoc &dl) const { |
5919 | SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff, |
5920 | CallSeqStart.getNode()->getOperand(0), |
5921 | Flags, DAG, dl); |
5922 | // The MEMCPY must go outside the CALLSEQ_START..END. |
5923 | int64_t FrameSize = CallSeqStart.getConstantOperandVal(1); |
5924 | SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall, FrameSize, 0, |
5925 | SDLoc(MemcpyCall)); |
5926 | DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), |
5927 | NewCallSeqStart.getNode()); |
5928 | return NewCallSeqStart; |
5929 | } |
5930 | |
5931 | SDValue PPCTargetLowering::LowerCall_64SVR4( |
5932 | SDValue Chain, SDValue Callee, CallFlags CFlags, |
5933 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
5934 | const SmallVectorImpl<SDValue> &OutVals, |
5935 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
5936 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, |
5937 | const CallBase *CB) const { |
5938 | bool isELFv2ABI = Subtarget.isELFv2ABI(); |
5939 | bool isLittleEndian = Subtarget.isLittleEndian(); |
5940 | unsigned NumOps = Outs.size(); |
5941 | bool IsSibCall = false; |
5942 | bool IsFastCall = CFlags.CallConv == CallingConv::Fast; |
5943 | |
5944 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
5945 | unsigned PtrByteSize = 8; |
5946 | |
5947 | MachineFunction &MF = DAG.getMachineFunction(); |
5948 | |
5949 | if (CFlags.IsTailCall && !getTargetMachine().Options.GuaranteedTailCallOpt) |
5950 | IsSibCall = true; |
5951 | |
5952 | // Mark this function as potentially containing a function that contains a |
5953 | // tail call. As a consequence the frame pointer will be used for dynamicalloc |
5954 | // and restoring the callers stack pointer in this functions epilog. This is |
5955 | // done because by tail calling the called function might overwrite the value |
5956 | // in this function's (MF) stack pointer stack slot 0(SP). |
5957 | if (getTargetMachine().Options.GuaranteedTailCallOpt && IsFastCall) |
5958 | MF.getInfo<PPCFunctionInfo>()->setHasFastCall(); |
5959 | |
5960 | assert(!(IsFastCall && CFlags.IsVarArg) &&((!(IsFastCall && CFlags.IsVarArg) && "fastcc not supported on varargs functions" ) ? static_cast<void> (0) : __assert_fail ("!(IsFastCall && CFlags.IsVarArg) && \"fastcc not supported on varargs functions\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5961, __PRETTY_FUNCTION__)) |
5961 | "fastcc not supported on varargs functions")((!(IsFastCall && CFlags.IsVarArg) && "fastcc not supported on varargs functions" ) ? static_cast<void> (0) : __assert_fail ("!(IsFastCall && CFlags.IsVarArg) && \"fastcc not supported on varargs functions\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5961, __PRETTY_FUNCTION__)); |
5962 | |
5963 | // Count how many bytes are to be pushed on the stack, including the linkage |
5964 | // area, and parameter passing area. On ELFv1, the linkage area is 48 bytes |
5965 | // reserved space for [SP][CR][LR][2 x unused][TOC]; on ELFv2, the linkage |
5966 | // area is 32 bytes reserved space for [SP][CR][LR][TOC]. |
5967 | unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
5968 | unsigned NumBytes = LinkageSize; |
5969 | unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0; |
5970 | |
5971 | static const MCPhysReg GPR[] = { |
5972 | PPC::X3, PPC::X4, PPC::X5, PPC::X6, |
5973 | PPC::X7, PPC::X8, PPC::X9, PPC::X10, |
5974 | }; |
5975 | static const MCPhysReg VR[] = { |
5976 | PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8, |
5977 | PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13 |
5978 | }; |
5979 | |
5980 | const unsigned NumGPRs = array_lengthof(GPR); |
5981 | const unsigned NumFPRs = useSoftFloat() ? 0 : 13; |
5982 | const unsigned NumVRs = array_lengthof(VR); |
5983 | |
5984 | // On ELFv2, we can avoid allocating the parameter area if all the arguments |
5985 | // can be passed to the callee in registers. |
5986 | // For the fast calling convention, there is another check below. |
5987 | // Note: We should keep consistent with LowerFormalArguments_64SVR4() |
5988 | bool HasParameterArea = !isELFv2ABI || CFlags.IsVarArg || IsFastCall; |
5989 | if (!HasParameterArea) { |
5990 | unsigned ParamAreaSize = NumGPRs * PtrByteSize; |
5991 | unsigned AvailableFPRs = NumFPRs; |
5992 | unsigned AvailableVRs = NumVRs; |
5993 | unsigned NumBytesTmp = NumBytes; |
5994 | for (unsigned i = 0; i != NumOps; ++i) { |
5995 | if (Outs[i].Flags.isNest()) continue; |
5996 | if (CalculateStackSlotUsed(Outs[i].VT, Outs[i].ArgVT, Outs[i].Flags, |
5997 | PtrByteSize, LinkageSize, ParamAreaSize, |
5998 | NumBytesTmp, AvailableFPRs, AvailableVRs)) |
5999 | HasParameterArea = true; |
6000 | } |
6001 | } |
6002 | |
6003 | // When using the fast calling convention, we don't provide backing for |
6004 | // arguments that will be in registers. |
6005 | unsigned NumGPRsUsed = 0, NumFPRsUsed = 0, NumVRsUsed = 0; |
6006 | |
6007 | // Avoid allocating parameter area for fastcc functions if all the arguments |
6008 | // can be passed in the registers. |
6009 | if (IsFastCall) |
6010 | HasParameterArea = false; |
6011 | |
6012 | // Add up all the space actually used. |
6013 | for (unsigned i = 0; i != NumOps; ++i) { |
6014 | ISD::ArgFlagsTy Flags = Outs[i].Flags; |
6015 | EVT ArgVT = Outs[i].VT; |
6016 | EVT OrigVT = Outs[i].ArgVT; |
6017 | |
6018 | if (Flags.isNest()) |
6019 | continue; |
6020 | |
6021 | if (IsFastCall) { |
6022 | if (Flags.isByVal()) { |
6023 | NumGPRsUsed += (Flags.getByValSize()+7)/8; |
6024 | if (NumGPRsUsed > NumGPRs) |
6025 | HasParameterArea = true; |
6026 | } else { |
6027 | switch (ArgVT.getSimpleVT().SimpleTy) { |
6028 | default: llvm_unreachable("Unexpected ValueType for argument!")::llvm::llvm_unreachable_internal("Unexpected ValueType for argument!" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6028); |
6029 | case MVT::i1: |
6030 | case MVT::i32: |
6031 | case MVT::i64: |
6032 | if (++NumGPRsUsed <= NumGPRs) |
6033 | continue; |
6034 | break; |
6035 | case MVT::v4i32: |
6036 | case MVT::v8i16: |
6037 | case MVT::v16i8: |
6038 | case MVT::v2f64: |
6039 | case MVT::v2i64: |
6040 | case MVT::v1i128: |
6041 | case MVT::f128: |
6042 | if (++NumVRsUsed <= NumVRs) |
6043 | continue; |
6044 | break; |
6045 | case MVT::v4f32: |
6046 | if (++NumVRsUsed <= NumVRs) |
6047 | continue; |
6048 | break; |
6049 | case MVT::f32: |
6050 | case MVT::f64: |
6051 | if (++NumFPRsUsed <= NumFPRs) |
6052 | continue; |
6053 | break; |
6054 | } |
6055 | HasParameterArea = true; |
6056 | } |
6057 | } |
6058 | |
6059 | /* Respect alignment of argument on the stack. */ |
6060 | auto Alignement = |
6061 | CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize); |
6062 | NumBytes = alignTo(NumBytes, Alignement); |
6063 | |
6064 | NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize); |
6065 | if (Flags.isInConsecutiveRegsLast()) |
6066 | NumBytes = ((NumBytes + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
6067 | } |
6068 | |
6069 | unsigned NumBytesActuallyUsed = NumBytes; |
6070 | |
6071 | // In the old ELFv1 ABI, |
6072 | // the prolog code of the callee may store up to 8 GPR argument registers to |
6073 | // the stack, allowing va_start to index over them in memory if its varargs. |
6074 | // Because we cannot tell if this is needed on the caller side, we have to |
6075 | // conservatively assume that it is needed. As such, make sure we have at |
6076 | // least enough stack space for the caller to store the 8 GPRs. |
6077 | // In the ELFv2 ABI, we allocate the parameter area iff a callee |
6078 | // really requires memory operands, e.g. a vararg function. |
6079 | if (HasParameterArea) |
6080 | NumBytes = std::max(NumBytes, LinkageSize + 8 * PtrByteSize); |
6081 | else |
6082 | NumBytes = LinkageSize; |
6083 | |
6084 | // Tail call needs the stack to be aligned. |
6085 | if (getTargetMachine().Options.GuaranteedTailCallOpt && IsFastCall) |
6086 | NumBytes = EnsureStackAlignment(Subtarget.getFrameLowering(), NumBytes); |
6087 | |
6088 | int SPDiff = 0; |
6089 | |
6090 | // Calculate by how many bytes the stack has to be adjusted in case of tail |
6091 | // call optimization. |
6092 | if (!IsSibCall) |
6093 | SPDiff = CalculateTailCallSPDiff(DAG, CFlags.IsTailCall, NumBytes); |
6094 | |
6095 | // To protect arguments on the stack from being clobbered in a tail call, |
6096 | // force all the loads to happen before doing any other lowering. |
6097 | if (CFlags.IsTailCall) |
6098 | Chain = DAG.getStackArgumentTokenFactor(Chain); |
6099 | |
6100 | // Adjust the stack pointer for the new arguments... |
6101 | // These operations are automatically eliminated by the prolog/epilog pass |
6102 | if (!IsSibCall) |
6103 | Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl); |
6104 | SDValue CallSeqStart = Chain; |
6105 | |
6106 | // Load the return address and frame pointer so it can be move somewhere else |
6107 | // later. |
6108 | SDValue LROp, FPOp; |
6109 | Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, dl); |
6110 | |
6111 | // Set up a copy of the stack pointer for use loading and storing any |
6112 | // arguments that may not fit in the registers available for argument |
6113 | // passing. |
6114 | SDValue StackPtr = DAG.getRegister(PPC::X1, MVT::i64); |
6115 | |
6116 | // Figure out which arguments are going to go in registers, and which in |
6117 | // memory. Also, if this is a vararg function, floating point operations |
6118 | // must be stored to our stack, and loaded into integer regs as well, if |
6119 | // any integer regs are available for argument passing. |
6120 | unsigned ArgOffset = LinkageSize; |
6121 | |
6122 | SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; |
6123 | SmallVector<TailCallArgumentInfo, 8> TailCallArguments; |
6124 | |
6125 | SmallVector<SDValue, 8> MemOpChains; |
6126 | for (unsigned i = 0; i != NumOps; ++i) { |
6127 | SDValue Arg = OutVals[i]; |
6128 | ISD::ArgFlagsTy Flags = Outs[i].Flags; |
6129 | EVT ArgVT = Outs[i].VT; |
6130 | EVT OrigVT = Outs[i].ArgVT; |
6131 | |
6132 | // PtrOff will be used to store the current argument to the stack if a |
6133 | // register cannot be found for it. |
6134 | SDValue PtrOff; |
6135 | |
6136 | // We re-align the argument offset for each argument, except when using the |
6137 | // fast calling convention, when we need to make sure we do that only when |
6138 | // we'll actually use a stack slot. |
6139 | auto ComputePtrOff = [&]() { |
6140 | /* Respect alignment of argument on the stack. */ |
6141 | auto Alignment = |
6142 | CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize); |
6143 | ArgOffset = alignTo(ArgOffset, Alignment); |
6144 | |
6145 | PtrOff = DAG.getConstant(ArgOffset, dl, StackPtr.getValueType()); |
6146 | |
6147 | PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff); |
6148 | }; |
6149 | |
6150 | if (!IsFastCall) { |
6151 | ComputePtrOff(); |
6152 | |
6153 | /* Compute GPR index associated with argument offset. */ |
6154 | GPR_idx = (ArgOffset - LinkageSize) / PtrByteSize; |
6155 | GPR_idx = std::min(GPR_idx, NumGPRs); |
6156 | } |
6157 | |
6158 | // Promote integers to 64-bit values. |
6159 | if (Arg.getValueType() == MVT::i32 || Arg.getValueType() == MVT::i1) { |
6160 | // FIXME: Should this use ANY_EXTEND if neither sext nor zext? |
6161 | unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; |
6162 | Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg); |
6163 | } |
6164 | |
6165 | // FIXME memcpy is used way more than necessary. Correctness first. |
6166 | // Note: "by value" is code for passing a structure by value, not |
6167 | // basic types. |
6168 | if (Flags.isByVal()) { |
6169 | // Note: Size includes alignment padding, so |
6170 | // struct x { short a; char b; } |
6171 | // will have Size = 4. With #pragma pack(1), it will have Size = 3. |
6172 | // These are the proper values we need for right-justifying the |
6173 | // aggregate in a parameter register. |
6174 | unsigned Size = Flags.getByValSize(); |
6175 | |
6176 | // An empty aggregate parameter takes up no storage and no |
6177 | // registers. |
6178 | if (Size == 0) |
6179 | continue; |
6180 | |
6181 | if (IsFastCall) |
6182 | ComputePtrOff(); |
6183 | |
6184 | // All aggregates smaller than 8 bytes must be passed right-justified. |
6185 | if (Size==1 || Size==2 || Size==4) { |
6186 | EVT VT = (Size==1) ? MVT::i8 : ((Size==2) ? MVT::i16 : MVT::i32); |
6187 | if (GPR_idx != NumGPRs) { |
6188 | SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg, |
6189 | MachinePointerInfo(), VT); |
6190 | MemOpChains.push_back(Load.getValue(1)); |
6191 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6192 | |
6193 | ArgOffset += PtrByteSize; |
6194 | continue; |
6195 | } |
6196 | } |
6197 | |
6198 | if (GPR_idx == NumGPRs && Size < 8) { |
6199 | SDValue AddPtr = PtrOff; |
6200 | if (!isLittleEndian) { |
6201 | SDValue Const = DAG.getConstant(PtrByteSize - Size, dl, |
6202 | PtrOff.getValueType()); |
6203 | AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const); |
6204 | } |
6205 | Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr, |
6206 | CallSeqStart, |
6207 | Flags, DAG, dl); |
6208 | ArgOffset += PtrByteSize; |
6209 | continue; |
6210 | } |
6211 | // Copy entire object into memory. There are cases where gcc-generated |
6212 | // code assumes it is there, even if it could be put entirely into |
6213 | // registers. (This is not what the doc says.) |
6214 | |
6215 | // FIXME: The above statement is likely due to a misunderstanding of the |
6216 | // documents. All arguments must be copied into the parameter area BY |
6217 | // THE CALLEE in the event that the callee takes the address of any |
6218 | // formal argument. That has not yet been implemented. However, it is |
6219 | // reasonable to use the stack area as a staging area for the register |
6220 | // load. |
6221 | |
6222 | // Skip this for small aggregates, as we will use the same slot for a |
6223 | // right-justified copy, below. |
6224 | if (Size >= 8) |
6225 | Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff, |
6226 | CallSeqStart, |
6227 | Flags, DAG, dl); |
6228 | |
6229 | // When a register is available, pass a small aggregate right-justified. |
6230 | if (Size < 8 && GPR_idx != NumGPRs) { |
6231 | // The easiest way to get this right-justified in a register |
6232 | // is to copy the structure into the rightmost portion of a |
6233 | // local variable slot, then load the whole slot into the |
6234 | // register. |
6235 | // FIXME: The memcpy seems to produce pretty awful code for |
6236 | // small aggregates, particularly for packed ones. |
6237 | // FIXME: It would be preferable to use the slot in the |
6238 | // parameter save area instead of a new local variable. |
6239 | SDValue AddPtr = PtrOff; |
6240 | if (!isLittleEndian) { |
6241 | SDValue Const = DAG.getConstant(8 - Size, dl, PtrOff.getValueType()); |
6242 | AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const); |
6243 | } |
6244 | Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr, |
6245 | CallSeqStart, |
6246 | Flags, DAG, dl); |
6247 | |
6248 | // Load the slot into the register. |
6249 | SDValue Load = |
6250 | DAG.getLoad(PtrVT, dl, Chain, PtrOff, MachinePointerInfo()); |
6251 | MemOpChains.push_back(Load.getValue(1)); |
6252 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6253 | |
6254 | // Done with this argument. |
6255 | ArgOffset += PtrByteSize; |
6256 | continue; |
6257 | } |
6258 | |
6259 | // For aggregates larger than PtrByteSize, copy the pieces of the |
6260 | // object that fit into registers from the parameter save area. |
6261 | for (unsigned j=0; j<Size; j+=PtrByteSize) { |
6262 | SDValue Const = DAG.getConstant(j, dl, PtrOff.getValueType()); |
6263 | SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const); |
6264 | if (GPR_idx != NumGPRs) { |
6265 | SDValue Load = |
6266 | DAG.getLoad(PtrVT, dl, Chain, AddArg, MachinePointerInfo()); |
6267 | MemOpChains.push_back(Load.getValue(1)); |
6268 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6269 | ArgOffset += PtrByteSize; |
6270 | } else { |
6271 | ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize; |
6272 | break; |
6273 | } |
6274 | } |
6275 | continue; |
6276 | } |
6277 | |
6278 | switch (Arg.getSimpleValueType().SimpleTy) { |
6279 | default: llvm_unreachable("Unexpected ValueType for argument!")::llvm::llvm_unreachable_internal("Unexpected ValueType for argument!" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6279); |
6280 | case MVT::i1: |
6281 | case MVT::i32: |
6282 | case MVT::i64: |
6283 | if (Flags.isNest()) { |
6284 | // The 'nest' parameter, if any, is passed in R11. |
6285 | RegsToPass.push_back(std::make_pair(PPC::X11, Arg)); |
6286 | break; |
6287 | } |
6288 | |
6289 | // These can be scalar arguments or elements of an integer array type |
6290 | // passed directly. Clang may use those instead of "byval" aggregate |
6291 | // types to avoid forcing arguments to memory unnecessarily. |
6292 | if (GPR_idx != NumGPRs) { |
6293 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg)); |
6294 | } else { |
6295 | if (IsFastCall) |
6296 | ComputePtrOff(); |
6297 | |
6298 | assert(HasParameterArea &&((HasParameterArea && "Parameter area must exist to pass an argument in memory." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist to pass an argument in memory.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6299, __PRETTY_FUNCTION__)) |
6299 | "Parameter area must exist to pass an argument in memory.")((HasParameterArea && "Parameter area must exist to pass an argument in memory." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist to pass an argument in memory.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6299, __PRETTY_FUNCTION__)); |
6300 | LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset, |
6301 | true, CFlags.IsTailCall, false, MemOpChains, |
6302 | TailCallArguments, dl); |
6303 | if (IsFastCall) |
6304 | ArgOffset += PtrByteSize; |
6305 | } |
6306 | if (!IsFastCall) |
6307 | ArgOffset += PtrByteSize; |
6308 | break; |
6309 | case MVT::f32: |
6310 | case MVT::f64: { |
6311 | // These can be scalar arguments or elements of a float array type |
6312 | // passed directly. The latter are used to implement ELFv2 homogenous |
6313 | // float aggregates. |
6314 | |
6315 | // Named arguments go into FPRs first, and once they overflow, the |
6316 | // remaining arguments go into GPRs and then the parameter save area. |
6317 | // Unnamed arguments for vararg functions always go to GPRs and |
6318 | // then the parameter save area. For now, put all arguments to vararg |
6319 | // routines always in both locations (FPR *and* GPR or stack slot). |
6320 | bool NeedGPROrStack = CFlags.IsVarArg || FPR_idx == NumFPRs; |
6321 | bool NeededLoad = false; |
6322 | |
6323 | // First load the argument into the next available FPR. |
6324 | if (FPR_idx != NumFPRs) |
6325 | RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg)); |
6326 | |
6327 | // Next, load the argument into GPR or stack slot if needed. |
6328 | if (!NeedGPROrStack) |
6329 | ; |
6330 | else if (GPR_idx != NumGPRs && !IsFastCall) { |
6331 | // FIXME: We may want to re-enable this for CallingConv::Fast on the P8 |
6332 | // once we support fp <-> gpr moves. |
6333 | |
6334 | // In the non-vararg case, this can only ever happen in the |
6335 | // presence of f32 array types, since otherwise we never run |
6336 | // out of FPRs before running out of GPRs. |
6337 | SDValue ArgVal; |
6338 | |
6339 | // Double values are always passed in a single GPR. |
6340 | if (Arg.getValueType() != MVT::f32) { |
6341 | ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg); |
6342 | |
6343 | // Non-array float values are extended and passed in a GPR. |
6344 | } else if (!Flags.isInConsecutiveRegs()) { |
6345 | ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg); |
6346 | ArgVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i64, ArgVal); |
6347 | |
6348 | // If we have an array of floats, we collect every odd element |
6349 | // together with its predecessor into one GPR. |
6350 | } else if (ArgOffset % PtrByteSize != 0) { |
6351 | SDValue Lo, Hi; |
6352 | Lo = DAG.getNode(ISD::BITCAST, dl, MVT::i32, OutVals[i - 1]); |
6353 | Hi = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg); |
6354 | if (!isLittleEndian) |
6355 | std::swap(Lo, Hi); |
6356 | ArgVal = DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi); |
6357 | |
6358 | // The final element, if even, goes into the first half of a GPR. |
6359 | } else if (Flags.isInConsecutiveRegsLast()) { |
6360 | ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg); |
6361 | ArgVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i64, ArgVal); |
6362 | if (!isLittleEndian) |
6363 | ArgVal = DAG.getNode(ISD::SHL, dl, MVT::i64, ArgVal, |
6364 | DAG.getConstant(32, dl, MVT::i32)); |
6365 | |
6366 | // Non-final even elements are skipped; they will be handled |
6367 | // together the with subsequent argument on the next go-around. |
6368 | } else |
6369 | ArgVal = SDValue(); |
6370 | |
6371 | if (ArgVal.getNode()) |
6372 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], ArgVal)); |
6373 | } else { |
6374 | if (IsFastCall) |
6375 | ComputePtrOff(); |
6376 | |
6377 | // Single-precision floating-point values are mapped to the |
6378 | // second (rightmost) word of the stack doubleword. |
6379 | if (Arg.getValueType() == MVT::f32 && |
6380 | !isLittleEndian && !Flags.isInConsecutiveRegs()) { |
6381 | SDValue ConstFour = DAG.getConstant(4, dl, PtrOff.getValueType()); |
6382 | PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour); |
6383 | } |
6384 | |
6385 | assert(HasParameterArea &&((HasParameterArea && "Parameter area must exist to pass an argument in memory." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist to pass an argument in memory.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6386, __PRETTY_FUNCTION__)) |
6386 | "Parameter area must exist to pass an argument in memory.")((HasParameterArea && "Parameter area must exist to pass an argument in memory." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist to pass an argument in memory.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6386, __PRETTY_FUNCTION__)); |
6387 | LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset, |
6388 | true, CFlags.IsTailCall, false, MemOpChains, |
6389 | TailCallArguments, dl); |
6390 | |
6391 | NeededLoad = true; |
6392 | } |
6393 | // When passing an array of floats, the array occupies consecutive |
6394 | // space in the argument area; only round up to the next doubleword |
6395 | // at the end of the array. Otherwise, each float takes 8 bytes. |
6396 | if (!IsFastCall || NeededLoad) { |
6397 | ArgOffset += (Arg.getValueType() == MVT::f32 && |
6398 | Flags.isInConsecutiveRegs()) ? 4 : 8; |
6399 | if (Flags.isInConsecutiveRegsLast()) |
6400 | ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
6401 | } |
6402 | break; |
6403 | } |
6404 | case MVT::v4f32: |
6405 | case MVT::v4i32: |
6406 | case MVT::v8i16: |
6407 | case MVT::v16i8: |
6408 | case MVT::v2f64: |
6409 | case MVT::v2i64: |
6410 | case MVT::v1i128: |
6411 | case MVT::f128: |
6412 | // These can be scalar arguments or elements of a vector array type |
6413 | // passed directly. The latter are used to implement ELFv2 homogenous |
6414 | // vector aggregates. |
6415 | |
6416 | // For a varargs call, named arguments go into VRs or on the stack as |
6417 | // usual; unnamed arguments always go to the stack or the corresponding |
6418 | // GPRs when within range. For now, we always put the value in both |
6419 | // locations (or even all three). |
6420 | if (CFlags.IsVarArg) { |
6421 | assert(HasParameterArea &&((HasParameterArea && "Parameter area must exist if we have a varargs call." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist if we have a varargs call.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6422, __PRETTY_FUNCTION__)) |
6422 | "Parameter area must exist if we have a varargs call.")((HasParameterArea && "Parameter area must exist if we have a varargs call." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist if we have a varargs call.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6422, __PRETTY_FUNCTION__)); |
6423 | // We could elide this store in the case where the object fits |
6424 | // entirely in R registers. Maybe later. |
6425 | SDValue Store = |
6426 | DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo()); |
6427 | MemOpChains.push_back(Store); |
6428 | if (VR_idx != NumVRs) { |
6429 | SDValue Load = |
6430 | DAG.getLoad(MVT::v4f32, dl, Store, PtrOff, MachinePointerInfo()); |
6431 | MemOpChains.push_back(Load.getValue(1)); |
6432 | RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load)); |
6433 | } |
6434 | ArgOffset += 16; |
6435 | for (unsigned i=0; i<16; i+=PtrByteSize) { |
6436 | if (GPR_idx == NumGPRs) |
6437 | break; |
6438 | SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, |
6439 | DAG.getConstant(i, dl, PtrVT)); |
6440 | SDValue Load = |
6441 | DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo()); |
6442 | MemOpChains.push_back(Load.getValue(1)); |
6443 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6444 | } |
6445 | break; |
6446 | } |
6447 | |
6448 | // Non-varargs Altivec params go into VRs or on the stack. |
6449 | if (VR_idx != NumVRs) { |
6450 | RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg)); |
6451 | } else { |
6452 | if (IsFastCall) |
6453 | ComputePtrOff(); |
6454 | |
6455 | assert(HasParameterArea &&((HasParameterArea && "Parameter area must exist to pass an argument in memory." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist to pass an argument in memory.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6456, __PRETTY_FUNCTION__)) |
6456 | "Parameter area must exist to pass an argument in memory.")((HasParameterArea && "Parameter area must exist to pass an argument in memory." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist to pass an argument in memory.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6456, __PRETTY_FUNCTION__)); |
6457 | LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset, |
6458 | true, CFlags.IsTailCall, true, MemOpChains, |
6459 | TailCallArguments, dl); |
6460 | if (IsFastCall) |
6461 | ArgOffset += 16; |
6462 | } |
6463 | |
6464 | if (!IsFastCall) |
6465 | ArgOffset += 16; |
6466 | break; |
6467 | } |
6468 | } |
6469 | |
6470 | assert((!HasParameterArea || NumBytesActuallyUsed == ArgOffset) &&(((!HasParameterArea || NumBytesActuallyUsed == ArgOffset) && "mismatch in size of parameter area") ? static_cast<void> (0) : __assert_fail ("(!HasParameterArea || NumBytesActuallyUsed == ArgOffset) && \"mismatch in size of parameter area\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6471, __PRETTY_FUNCTION__)) |
6471 | "mismatch in size of parameter area")(((!HasParameterArea || NumBytesActuallyUsed == ArgOffset) && "mismatch in size of parameter area") ? static_cast<void> (0) : __assert_fail ("(!HasParameterArea || NumBytesActuallyUsed == ArgOffset) && \"mismatch in size of parameter area\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6471, __PRETTY_FUNCTION__)); |
6472 | (void)NumBytesActuallyUsed; |
6473 | |
6474 | if (!MemOpChains.empty()) |
6475 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains); |
6476 | |
6477 | // Check if this is an indirect call (MTCTR/BCTRL). |
6478 | // See prepareDescriptorIndirectCall and buildCallOperands for more |
6479 | // information about calls through function pointers in the 64-bit SVR4 ABI. |
6480 | if (CFlags.IsIndirect) { |
6481 | // For 64-bit ELFv2 ABI with PCRel, do not save the TOC of the |
6482 | // caller in the TOC save area. |
6483 | if (isTOCSaveRestoreRequired(Subtarget)) { |
6484 | assert(!CFlags.IsTailCall && "Indirect tails calls not supported")((!CFlags.IsTailCall && "Indirect tails calls not supported" ) ? static_cast<void> (0) : __assert_fail ("!CFlags.IsTailCall && \"Indirect tails calls not supported\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6484, __PRETTY_FUNCTION__)); |
6485 | // Load r2 into a virtual register and store it to the TOC save area. |
6486 | setUsesTOCBasePtr(DAG); |
6487 | SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64); |
6488 | // TOC save area offset. |
6489 | unsigned TOCSaveOffset = Subtarget.getFrameLowering()->getTOCSaveOffset(); |
6490 | SDValue PtrOff = DAG.getIntPtrConstant(TOCSaveOffset, dl); |
6491 | SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff); |
6492 | Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, |
6493 | MachinePointerInfo::getStack( |
6494 | DAG.getMachineFunction(), TOCSaveOffset)); |
6495 | } |
6496 | // In the ELFv2 ABI, R12 must contain the address of an indirect callee. |
6497 | // This does not mean the MTCTR instruction must use R12; it's easier |
6498 | // to model this as an extra parameter, so do that. |
6499 | if (isELFv2ABI && !CFlags.IsPatchPoint) |
6500 | RegsToPass.push_back(std::make_pair((unsigned)PPC::X12, Callee)); |
6501 | } |
6502 | |
6503 | // Build a sequence of copy-to-reg nodes chained together with token chain |
6504 | // and flag operands which copy the outgoing args into the appropriate regs. |
6505 | SDValue InFlag; |
6506 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { |
6507 | Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, |
6508 | RegsToPass[i].second, InFlag); |
6509 | InFlag = Chain.getValue(1); |
6510 | } |
6511 | |
6512 | if (CFlags.IsTailCall && !IsSibCall) |
6513 | PrepareTailCall(DAG, InFlag, Chain, dl, SPDiff, NumBytes, LROp, FPOp, |
6514 | TailCallArguments); |
6515 | |
6516 | return FinishCall(CFlags, dl, DAG, RegsToPass, InFlag, Chain, CallSeqStart, |
6517 | Callee, SPDiff, NumBytes, Ins, InVals, CB); |
6518 | } |
6519 | |
6520 | SDValue PPCTargetLowering::LowerCall_Darwin( |
6521 | SDValue Chain, SDValue Callee, CallFlags CFlags, |
6522 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
6523 | const SmallVectorImpl<SDValue> &OutVals, |
6524 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
6525 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, |
6526 | const CallBase *CB) const { |
6527 | unsigned NumOps = Outs.size(); |
6528 | |
6529 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
6530 | bool isPPC64 = PtrVT == MVT::i64; |
6531 | unsigned PtrByteSize = isPPC64 ? 8 : 4; |
6532 | |
6533 | MachineFunction &MF = DAG.getMachineFunction(); |
6534 | |
6535 | // Mark this function as potentially containing a function that contains a |
6536 | // tail call. As a consequence the frame pointer will be used for dynamicalloc |
6537 | // and restoring the callers stack pointer in this functions epilog. This is |
6538 | // done because by tail calling the called function might overwrite the value |
6539 | // in this function's (MF) stack pointer stack slot 0(SP). |
6540 | if (getTargetMachine().Options.GuaranteedTailCallOpt && |
6541 | CFlags.CallConv == CallingConv::Fast) |
6542 | MF.getInfo<PPCFunctionInfo>()->setHasFastCall(); |
6543 | |
6544 | // Count how many bytes are to be pushed on the stack, including the linkage |
6545 | // area, and parameter passing area. We start with 24/48 bytes, which is |
6546 | // prereserved space for [SP][CR][LR][3 x unused]. |
6547 | unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
6548 | unsigned NumBytes = LinkageSize; |
6549 | |
6550 | // Add up all the space actually used. |
6551 | // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually |
6552 | // they all go in registers, but we must reserve stack space for them for |
6553 | // possible use by the caller. In varargs or 64-bit calls, parameters are |
6554 | // assigned stack space in order, with padding so Altivec parameters are |
6555 | // 16-byte aligned. |
6556 | unsigned nAltivecParamsAtEnd = 0; |
6557 | for (unsigned i = 0; i != NumOps; ++i) { |
6558 | ISD::ArgFlagsTy Flags = Outs[i].Flags; |
6559 | EVT ArgVT = Outs[i].VT; |
6560 | // Varargs Altivec parameters are padded to a 16 byte boundary. |
6561 | if (ArgVT == MVT::v4f32 || ArgVT == MVT::v4i32 || |
6562 | ArgVT == MVT::v8i16 || ArgVT == MVT::v16i8 || |
6563 | ArgVT == MVT::v2f64 || ArgVT == MVT::v2i64) { |
6564 | if (!CFlags.IsVarArg && !isPPC64) { |
6565 | // Non-varargs Altivec parameters go after all the non-Altivec |
6566 | // parameters; handle those later so we know how much padding we need. |
6567 | nAltivecParamsAtEnd++; |
6568 | continue; |
6569 | } |
6570 | // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary. |
6571 | NumBytes = ((NumBytes+15)/16)*16; |
6572 | } |
6573 | NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize); |
6574 | } |
6575 | |
6576 | // Allow for Altivec parameters at the end, if needed. |
6577 | if (nAltivecParamsAtEnd) { |
6578 | NumBytes = ((NumBytes+15)/16)*16; |
6579 | NumBytes += 16*nAltivecParamsAtEnd; |
6580 | } |
6581 | |
6582 | // The prolog code of the callee may store up to 8 GPR argument registers to |
6583 | // the stack, allowing va_start to index over them in memory if its varargs. |
6584 | // Because we cannot tell if this is needed on the caller side, we have to |
6585 | // conservatively assume that it is needed. As such, make sure we have at |
6586 | // least enough stack space for the caller to store the 8 GPRs. |
6587 | NumBytes = std::max(NumBytes, LinkageSize + 8 * PtrByteSize); |
6588 | |
6589 | // Tail call needs the stack to be aligned. |
6590 | if (getTargetMachine().Options.GuaranteedTailCallOpt && |
6591 | CFlags.CallConv == CallingConv::Fast) |
6592 | NumBytes = EnsureStackAlignment(Subtarget.getFrameLowering(), NumBytes); |
6593 | |
6594 | // Calculate by how many bytes the stack has to be adjusted in case of tail |
6595 | // call optimization. |
6596 | int SPDiff = CalculateTailCallSPDiff(DAG, CFlags.IsTailCall, NumBytes); |
6597 | |
6598 | // To protect arguments on the stack from being clobbered in a tail call, |
6599 | // force all the loads to happen before doing any other lowering. |
6600 | if (CFlags.IsTailCall) |
6601 | Chain = DAG.getStackArgumentTokenFactor(Chain); |
6602 | |
6603 | // Adjust the stack pointer for the new arguments... |
6604 | // These operations are automatically eliminated by the prolog/epilog pass |
6605 | Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl); |
6606 | SDValue CallSeqStart = Chain; |
6607 | |
6608 | // Load the return address and frame pointer so it can be move somewhere else |
6609 | // later. |
6610 | SDValue LROp, FPOp; |
6611 | Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, dl); |
6612 | |
6613 | // Set up a copy of the stack pointer for use loading and storing any |
6614 | // arguments that may not fit in the registers available for argument |
6615 | // passing. |
6616 | SDValue StackPtr; |
6617 | if (isPPC64) |
6618 | StackPtr = DAG.getRegister(PPC::X1, MVT::i64); |
6619 | else |
6620 | StackPtr = DAG.getRegister(PPC::R1, MVT::i32); |
6621 | |
6622 | // Figure out which arguments are going to go in registers, and which in |
6623 | // memory. Also, if this is a vararg function, floating point operations |
6624 | // must be stored to our stack, and loaded into integer regs as well, if |
6625 | // any integer regs are available for argument passing. |
6626 | unsigned ArgOffset = LinkageSize; |
6627 | unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0; |
6628 | |
6629 | static const MCPhysReg GPR_32[] = { // 32-bit registers. |
6630 | PPC::R3, PPC::R4, PPC::R5, PPC::R6, |
6631 | PPC::R7, PPC::R8, PPC::R9, PPC::R10, |
6632 | }; |
6633 | static const MCPhysReg GPR_64[] = { // 64-bit registers. |
6634 | PPC::X3, PPC::X4, PPC::X5, PPC::X6, |
6635 | PPC::X7, PPC::X8, PPC::X9, PPC::X10, |
6636 | }; |
6637 | static const MCPhysReg VR[] = { |
6638 | PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8, |
6639 | PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13 |
6640 | }; |
6641 | const unsigned NumGPRs = array_lengthof(GPR_32); |
6642 | const unsigned NumFPRs = 13; |
6643 | const unsigned NumVRs = array_lengthof(VR); |
6644 | |
6645 | const MCPhysReg *GPR = isPPC64 ? GPR_64 : GPR_32; |
6646 | |
6647 | SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; |
6648 | SmallVector<TailCallArgumentInfo, 8> TailCallArguments; |
6649 | |
6650 | SmallVector<SDValue, 8> MemOpChains; |
6651 | for (unsigned i = 0; i != NumOps; ++i) { |
6652 | SDValue Arg = OutVals[i]; |
6653 | ISD::ArgFlagsTy Flags = Outs[i].Flags; |
6654 | |
6655 | // PtrOff will be used to store the current argument to the stack if a |
6656 | // register cannot be found for it. |
6657 | SDValue PtrOff; |
6658 | |
6659 | PtrOff = DAG.getConstant(ArgOffset, dl, StackPtr.getValueType()); |
6660 | |
6661 | PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff); |
6662 | |
6663 | // On PPC64, promote integers to 64-bit values. |
6664 | if (isPPC64 && Arg.getValueType() == MVT::i32) { |
6665 | // FIXME: Should this use ANY_EXTEND if neither sext nor zext? |
6666 | unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; |
6667 | Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg); |
6668 | } |
6669 | |
6670 | // FIXME memcpy is used way more than necessary. Correctness first. |
6671 | // Note: "by value" is code for passing a structure by value, not |
6672 | // basic types. |
6673 | if (Flags.isByVal()) { |
6674 | unsigned Size = Flags.getByValSize(); |
6675 | // Very small objects are passed right-justified. Everything else is |
6676 | // passed left-justified. |
6677 | if (Size==1 || Size==2) { |
6678 | EVT VT = (Size==1) ? MVT::i8 : MVT::i16; |
6679 | if (GPR_idx != NumGPRs) { |
6680 | SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg, |
6681 | MachinePointerInfo(), VT); |
6682 | MemOpChains.push_back(Load.getValue(1)); |
6683 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6684 | |
6685 | ArgOffset += PtrByteSize; |
6686 | } else { |
6687 | SDValue Const = DAG.getConstant(PtrByteSize - Size, dl, |
6688 | PtrOff.getValueType()); |
6689 | SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const); |
6690 | Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr, |
6691 | CallSeqStart, |
6692 | Flags, DAG, dl); |
6693 | ArgOffset += PtrByteSize; |
6694 | } |
6695 | continue; |
6696 | } |
6697 | // Copy entire object into memory. There are cases where gcc-generated |
6698 | // code assumes it is there, even if it could be put entirely into |
6699 | // registers. (This is not what the doc says.) |
6700 | Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff, |
6701 | CallSeqStart, |
6702 | Flags, DAG, dl); |
6703 | |
6704 | // For small aggregates (Darwin only) and aggregates >= PtrByteSize, |
6705 | // copy the pieces of the object that fit into registers from the |
6706 | // parameter save area. |
6707 | for (unsigned j=0; j<Size; j+=PtrByteSize) { |
6708 | SDValue Const = DAG.getConstant(j, dl, PtrOff.getValueType()); |
6709 | SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const); |
6710 | if (GPR_idx != NumGPRs) { |
6711 | SDValue Load = |
6712 | DAG.getLoad(PtrVT, dl, Chain, AddArg, MachinePointerInfo()); |
6713 | MemOpChains.push_back(Load.getValue(1)); |
6714 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6715 | ArgOffset += PtrByteSize; |
6716 | } else { |
6717 | ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize; |
6718 | break; |
6719 | } |
6720 | } |
6721 | continue; |
6722 | } |
6723 | |
6724 | switch (Arg.getSimpleValueType().SimpleTy) { |
6725 | default: llvm_unreachable("Unexpected ValueType for argument!")::llvm::llvm_unreachable_internal("Unexpected ValueType for argument!" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6725); |
6726 | case MVT::i1: |
6727 | case MVT::i32: |
6728 | case MVT::i64: |
6729 | if (GPR_idx != NumGPRs) { |
6730 | if (Arg.getValueType() == MVT::i1) |
6731 | Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, PtrVT, Arg); |
6732 | |
6733 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg)); |
6734 | } else { |
6735 | LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset, |
6736 | isPPC64, CFlags.IsTailCall, false, MemOpChains, |
6737 | TailCallArguments, dl); |
6738 | } |
6739 | ArgOffset += PtrByteSize; |
6740 | break; |
6741 | case MVT::f32: |
6742 | case MVT::f64: |
6743 | if (FPR_idx != NumFPRs) { |
6744 | RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg)); |
6745 | |
6746 | if (CFlags.IsVarArg) { |
6747 | SDValue Store = |
6748 | DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo()); |
6749 | MemOpChains.push_back(Store); |
6750 | |
6751 | // Float varargs are always shadowed in available integer registers |
6752 | if (GPR_idx != NumGPRs) { |
6753 | SDValue Load = |
6754 | DAG.getLoad(PtrVT, dl, Store, PtrOff, MachinePointerInfo()); |
6755 | MemOpChains.push_back(Load.getValue(1)); |
6756 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6757 | } |
6758 | if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){ |
6759 | SDValue ConstFour = DAG.getConstant(4, dl, PtrOff.getValueType()); |
6760 | PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour); |
6761 | SDValue Load = |
6762 | DAG.getLoad(PtrVT, dl, Store, PtrOff, MachinePointerInfo()); |
6763 | MemOpChains.push_back(Load.getValue(1)); |
6764 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6765 | } |
6766 | } else { |
6767 | // If we have any FPRs remaining, we may also have GPRs remaining. |
6768 | // Args passed in FPRs consume either 1 (f32) or 2 (f64) available |
6769 | // GPRs. |
6770 | if (GPR_idx != NumGPRs) |
6771 | ++GPR_idx; |
6772 | if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && |
6773 | !isPPC64) // PPC64 has 64-bit GPR's obviously :) |
6774 | ++GPR_idx; |
6775 | } |
6776 | } else |
6777 | LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset, |
6778 | isPPC64, CFlags.IsTailCall, false, MemOpChains, |
6779 | TailCallArguments, dl); |
6780 | if (isPPC64) |
6781 | ArgOffset += 8; |
6782 | else |
6783 | ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8; |
6784 | break; |
6785 | case MVT::v4f32: |
6786 | case MVT::v4i32: |
6787 | case MVT::v8i16: |
6788 | case MVT::v16i8: |
6789 | if (CFlags.IsVarArg) { |
6790 | // These go aligned on the stack, or in the corresponding R registers |
6791 | // when within range. The Darwin PPC ABI doc claims they also go in |
6792 | // V registers; in fact gcc does this only for arguments that are |
6793 | // prototyped, not for those that match the ... We do it for all |
6794 | // arguments, seems to work. |
6795 | while (ArgOffset % 16 !=0) { |
6796 | ArgOffset += PtrByteSize; |
6797 | if (GPR_idx != NumGPRs) |
6798 | GPR_idx++; |
6799 | } |
6800 | // We could elide this store in the case where the object fits |
6801 | // entirely in R registers. Maybe later. |
6802 | PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, |
6803 | DAG.getConstant(ArgOffset, dl, PtrVT)); |
6804 | SDValue Store = |
6805 | DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo()); |
6806 | MemOpChains.push_back(Store); |
6807 | if (VR_idx != NumVRs) { |
6808 | SDValue Load = |
6809 | DAG.getLoad(MVT::v4f32, dl, Store, PtrOff, MachinePointerInfo()); |
6810 | MemOpChains.push_back(Load.getValue(1)); |
6811 | RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load)); |
6812 | } |
6813 | ArgOffset += 16; |
6814 | for (unsigned i=0; i<16; i+=PtrByteSize) { |
6815 | if (GPR_idx == NumGPRs) |
6816 | break; |
6817 | SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, |
6818 | DAG.getConstant(i, dl, PtrVT)); |
6819 | SDValue Load = |
6820 | DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo()); |
6821 | MemOpChains.push_back(Load.getValue(1)); |
6822 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6823 | } |
6824 | break; |
6825 | } |
6826 | |
6827 | // Non-varargs Altivec params generally go in registers, but have |
6828 | // stack space allocated at the end. |
6829 | if (VR_idx != NumVRs) { |
6830 | // Doesn't have GPR space allocated. |
6831 | RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg)); |
6832 | } else if (nAltivecParamsAtEnd==0) { |
6833 | // We are emitting Altivec params in order. |
6834 | LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset, |
6835 | isPPC64, CFlags.IsTailCall, true, MemOpChains, |
6836 | TailCallArguments, dl); |
6837 | ArgOffset += 16; |
6838 | } |
6839 | break; |
6840 | } |
6841 | } |
6842 | // If all Altivec parameters fit in registers, as they usually do, |
6843 | // they get stack space following the non-Altivec parameters. We |
6844 | // don't track this here because nobody below needs it. |
6845 | // If there are more Altivec parameters than fit in registers emit |
6846 | // the stores here. |
6847 | if (!CFlags.IsVarArg && nAltivecParamsAtEnd > NumVRs) { |
6848 | unsigned j = 0; |
6849 | // Offset is aligned; skip 1st 12 params which go in V registers. |
6850 | ArgOffset = ((ArgOffset+15)/16)*16; |
6851 | ArgOffset += 12*16; |
6852 | for (unsigned i = 0; i != NumOps; ++i) { |
6853 | SDValue Arg = OutVals[i]; |
6854 | EVT ArgType = Outs[i].VT; |
6855 | if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 || |
6856 | ArgType==MVT::v8i16 || ArgType==MVT::v16i8) { |
6857 | if (++j > NumVRs) { |
6858 | SDValue PtrOff; |
6859 | // We are emitting Altivec params in order. |
6860 | LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset, |
6861 | isPPC64, CFlags.IsTailCall, true, MemOpChains, |
6862 | TailCallArguments, dl); |
6863 | ArgOffset += 16; |
6864 | } |
6865 | } |
6866 | } |
6867 | } |
6868 | |
6869 | if (!MemOpChains.empty()) |
6870 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains); |
6871 | |
6872 | // On Darwin, R12 must contain the address of an indirect callee. This does |
6873 | // not mean the MTCTR instruction must use R12; it's easier to model this as |
6874 | // an extra parameter, so do that. |
6875 | if (CFlags.IsIndirect) { |
6876 | assert(!CFlags.IsTailCall && "Indirect tail-calls not supported.")((!CFlags.IsTailCall && "Indirect tail-calls not supported." ) ? static_cast<void> (0) : __assert_fail ("!CFlags.IsTailCall && \"Indirect tail-calls not supported.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6876, __PRETTY_FUNCTION__)); |
6877 | RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 : |
6878 | PPC::R12), Callee)); |
6879 | } |
6880 | |
6881 | // Build a sequence of copy-to-reg nodes chained together with token chain |
6882 | // and flag operands which copy the outgoing args into the appropriate regs. |
6883 | SDValue InFlag; |
6884 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { |
6885 | Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, |
6886 | RegsToPass[i].second, InFlag); |
6887 | InFlag = Chain.getValue(1); |
6888 | } |
6889 | |
6890 | if (CFlags.IsTailCall) |
6891 | PrepareTailCall(DAG, InFlag, Chain, dl, SPDiff, NumBytes, LROp, FPOp, |
6892 | TailCallArguments); |
6893 | |
6894 | return FinishCall(CFlags, dl, DAG, RegsToPass, InFlag, Chain, CallSeqStart, |
6895 | Callee, SPDiff, NumBytes, Ins, InVals, CB); |
6896 | } |
6897 | |
6898 | static bool CC_AIX(unsigned ValNo, MVT ValVT, MVT LocVT, |
6899 | CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, |
6900 | CCState &State) { |
6901 | |
6902 | const PPCSubtarget &Subtarget = static_cast<const PPCSubtarget &>( |
6903 | State.getMachineFunction().getSubtarget()); |
6904 | const bool IsPPC64 = Subtarget.isPPC64(); |
6905 | const Align PtrAlign = IsPPC64 ? Align(8) : Align(4); |
6906 | const MVT RegVT = IsPPC64 ? MVT::i64 : MVT::i32; |
6907 | |
6908 | assert((!ValVT.isInteger() ||(((!ValVT.isInteger() || (ValVT.getFixedSizeInBits() <= RegVT .getFixedSizeInBits())) && "Integer argument exceeds register size: should have been legalized" ) ? static_cast<void> (0) : __assert_fail ("(!ValVT.isInteger() || (ValVT.getFixedSizeInBits() <= RegVT.getFixedSizeInBits())) && \"Integer argument exceeds register size: should have been legalized\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6910, __PRETTY_FUNCTION__)) |
6909 | (ValVT.getFixedSizeInBits() <= RegVT.getFixedSizeInBits())) &&(((!ValVT.isInteger() || (ValVT.getFixedSizeInBits() <= RegVT .getFixedSizeInBits())) && "Integer argument exceeds register size: should have been legalized" ) ? static_cast<void> (0) : __assert_fail ("(!ValVT.isInteger() || (ValVT.getFixedSizeInBits() <= RegVT.getFixedSizeInBits())) && \"Integer argument exceeds register size: should have been legalized\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6910, __PRETTY_FUNCTION__)) |
6910 | "Integer argument exceeds register size: should have been legalized")(((!ValVT.isInteger() || (ValVT.getFixedSizeInBits() <= RegVT .getFixedSizeInBits())) && "Integer argument exceeds register size: should have been legalized" ) ? static_cast<void> (0) : __assert_fail ("(!ValVT.isInteger() || (ValVT.getFixedSizeInBits() <= RegVT.getFixedSizeInBits())) && \"Integer argument exceeds register size: should have been legalized\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6910, __PRETTY_FUNCTION__)); |
6911 | |
6912 | if (ValVT == MVT::f128) |
6913 | report_fatal_error("f128 is unimplemented on AIX."); |
6914 | |
6915 | if (ArgFlags.isNest()) |
6916 | report_fatal_error("Nest arguments are unimplemented."); |
6917 | |
6918 | if (ValVT.isVector() || LocVT.isVector()) |
6919 | report_fatal_error("Vector arguments are unimplemented on AIX."); |
6920 | |
6921 | static const MCPhysReg GPR_32[] = {// 32-bit registers. |
6922 | PPC::R3, PPC::R4, PPC::R5, PPC::R6, |
6923 | PPC::R7, PPC::R8, PPC::R9, PPC::R10}; |
6924 | static const MCPhysReg GPR_64[] = {// 64-bit registers. |
6925 | PPC::X3, PPC::X4, PPC::X5, PPC::X6, |
6926 | PPC::X7, PPC::X8, PPC::X9, PPC::X10}; |
6927 | |
6928 | if (ArgFlags.isByVal()) { |
6929 | if (ArgFlags.getNonZeroByValAlign() > PtrAlign) |
6930 | report_fatal_error("Pass-by-value arguments with alignment greater than " |
6931 | "register width are not supported."); |
6932 | |
6933 | const unsigned ByValSize = ArgFlags.getByValSize(); |
6934 | |
6935 | // An empty aggregate parameter takes up no storage and no registers, |
6936 | // but needs a MemLoc for a stack slot for the formal arguments side. |
6937 | if (ByValSize == 0) { |
6938 | State.addLoc(CCValAssign::getMem(ValNo, MVT::INVALID_SIMPLE_VALUE_TYPE, |
6939 | State.getNextStackOffset(), RegVT, |
6940 | LocInfo)); |
6941 | return false; |
6942 | } |
6943 | |
6944 | const unsigned StackSize = alignTo(ByValSize, PtrAlign); |
6945 | unsigned Offset = State.AllocateStack(StackSize, PtrAlign); |
6946 | for (const unsigned E = Offset + StackSize; Offset < E; |
6947 | Offset += PtrAlign.value()) { |
6948 | if (unsigned Reg = State.AllocateReg(IsPPC64 ? GPR_64 : GPR_32)) |
6949 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, RegVT, LocInfo)); |
6950 | else { |
6951 | State.addLoc(CCValAssign::getMem(ValNo, MVT::INVALID_SIMPLE_VALUE_TYPE, |
6952 | Offset, MVT::INVALID_SIMPLE_VALUE_TYPE, |
6953 | LocInfo)); |
6954 | break; |
6955 | } |
6956 | } |
6957 | return false; |
6958 | } |
6959 | |
6960 | // Arguments always reserve parameter save area. |
6961 | switch (ValVT.SimpleTy) { |
6962 | default: |
6963 | report_fatal_error("Unhandled value type for argument."); |
6964 | case MVT::i64: |
6965 | // i64 arguments should have been split to i32 for PPC32. |
6966 | assert(IsPPC64 && "PPC32 should have split i64 values.")((IsPPC64 && "PPC32 should have split i64 values.") ? static_cast<void> (0) : __assert_fail ("IsPPC64 && \"PPC32 should have split i64 values.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6966, __PRETTY_FUNCTION__)); |
6967 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
6968 | case MVT::i1: |
6969 | case MVT::i32: { |
6970 | const unsigned Offset = State.AllocateStack(PtrAlign.value(), PtrAlign); |
6971 | // AIX integer arguments are always passed in register width. |
6972 | if (ValVT.getFixedSizeInBits() < RegVT.getFixedSizeInBits()) |
6973 | LocInfo = ArgFlags.isSExt() ? CCValAssign::LocInfo::SExt |
6974 | : CCValAssign::LocInfo::ZExt; |
6975 | if (unsigned Reg = State.AllocateReg(IsPPC64 ? GPR_64 : GPR_32)) |
6976 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, RegVT, LocInfo)); |
6977 | else |
6978 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, RegVT, LocInfo)); |
6979 | |
6980 | return false; |
6981 | } |
6982 | case MVT::f32: |
6983 | case MVT::f64: { |
6984 | // Parameter save area (PSA) is reserved even if the float passes in fpr. |
6985 | const unsigned StoreSize = LocVT.getStoreSize(); |
6986 | // Floats are always 4-byte aligned in the PSA on AIX. |
6987 | // This includes f64 in 64-bit mode for ABI compatibility. |
6988 | const unsigned Offset = |
6989 | State.AllocateStack(IsPPC64 ? 8 : StoreSize, Align(4)); |
6990 | unsigned FReg = State.AllocateReg(FPR); |
6991 | if (FReg) |
6992 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, FReg, LocVT, LocInfo)); |
6993 | |
6994 | // Reserve and initialize GPRs or initialize the PSA as required. |
6995 | for (unsigned I = 0; I < StoreSize; I += PtrAlign.value()) { |
6996 | if (unsigned Reg = State.AllocateReg(IsPPC64 ? GPR_64 : GPR_32)) { |
6997 | assert(FReg && "An FPR should be available when a GPR is reserved.")((FReg && "An FPR should be available when a GPR is reserved." ) ? static_cast<void> (0) : __assert_fail ("FReg && \"An FPR should be available when a GPR is reserved.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6997, __PRETTY_FUNCTION__)); |
6998 | if (State.isVarArg()) { |
6999 | // Successfully reserved GPRs are only initialized for vararg calls. |
7000 | // Custom handling is required for: |
7001 | // f64 in PPC32 needs to be split into 2 GPRs. |
7002 | // f32 in PPC64 needs to occupy only lower 32 bits of 64-bit GPR. |
7003 | State.addLoc( |
7004 | CCValAssign::getCustomReg(ValNo, ValVT, Reg, RegVT, LocInfo)); |
7005 | } |
7006 | } else { |
7007 | // If there are insufficient GPRs, the PSA needs to be initialized. |
7008 | // Initialization occurs even if an FPR was initialized for |
7009 | // compatibility with the AIX XL compiler. The full memory for the |
7010 | // argument will be initialized even if a prior word is saved in GPR. |
7011 | // A custom memLoc is used when the argument also passes in FPR so |
7012 | // that the callee handling can skip over it easily. |
7013 | State.addLoc( |
7014 | FReg ? CCValAssign::getCustomMem(ValNo, ValVT, Offset, LocVT, |
7015 | LocInfo) |
7016 | : CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
7017 | break; |
7018 | } |
7019 | } |
7020 | |
7021 | return false; |
7022 | } |
7023 | } |
7024 | return true; |
7025 | } |
7026 | |
7027 | static const TargetRegisterClass *getRegClassForSVT(MVT::SimpleValueType SVT, |
7028 | bool IsPPC64) { |
7029 | assert((IsPPC64 || SVT != MVT::i64) &&(((IsPPC64 || SVT != MVT::i64) && "i64 should have been split for 32-bit codegen." ) ? static_cast<void> (0) : __assert_fail ("(IsPPC64 || SVT != MVT::i64) && \"i64 should have been split for 32-bit codegen.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7030, __PRETTY_FUNCTION__)) |
7030 | "i64 should have been split for 32-bit codegen.")(((IsPPC64 || SVT != MVT::i64) && "i64 should have been split for 32-bit codegen." ) ? static_cast<void> (0) : __assert_fail ("(IsPPC64 || SVT != MVT::i64) && \"i64 should have been split for 32-bit codegen.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7030, __PRETTY_FUNCTION__)); |
7031 | |
7032 | switch (SVT) { |
7033 | default: |
7034 | report_fatal_error("Unexpected value type for formal argument"); |
7035 | case MVT::i1: |
7036 | case MVT::i32: |
7037 | case MVT::i64: |
7038 | return IsPPC64 ? &PPC::G8RCRegClass : &PPC::GPRCRegClass; |
7039 | case MVT::f32: |
7040 | return &PPC::F4RCRegClass; |
7041 | case MVT::f64: |
7042 | return &PPC::F8RCRegClass; |
7043 | } |
7044 | } |
7045 | |
7046 | static SDValue truncateScalarIntegerArg(ISD::ArgFlagsTy Flags, EVT ValVT, |
7047 | SelectionDAG &DAG, SDValue ArgValue, |
7048 | MVT LocVT, const SDLoc &dl) { |
7049 | assert(ValVT.isScalarInteger() && LocVT.isScalarInteger())((ValVT.isScalarInteger() && LocVT.isScalarInteger()) ? static_cast<void> (0) : __assert_fail ("ValVT.isScalarInteger() && LocVT.isScalarInteger()" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7049, __PRETTY_FUNCTION__)); |
7050 | assert(ValVT.getFixedSizeInBits() < LocVT.getFixedSizeInBits())((ValVT.getFixedSizeInBits() < LocVT.getFixedSizeInBits()) ? static_cast<void> (0) : __assert_fail ("ValVT.getFixedSizeInBits() < LocVT.getFixedSizeInBits()" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7050, __PRETTY_FUNCTION__)); |
7051 | |
7052 | if (Flags.isSExt()) |
7053 | ArgValue = DAG.getNode(ISD::AssertSext, dl, LocVT, ArgValue, |
7054 | DAG.getValueType(ValVT)); |
7055 | else if (Flags.isZExt()) |
7056 | ArgValue = DAG.getNode(ISD::AssertZext, dl, LocVT, ArgValue, |
7057 | DAG.getValueType(ValVT)); |
7058 | |
7059 | return DAG.getNode(ISD::TRUNCATE, dl, ValVT, ArgValue); |
7060 | } |
7061 | |
7062 | static unsigned mapArgRegToOffsetAIX(unsigned Reg, const PPCFrameLowering *FL) { |
7063 | const unsigned LASize = FL->getLinkageSize(); |
7064 | |
7065 | if (PPC::GPRCRegClass.contains(Reg)) { |
7066 | assert(Reg >= PPC::R3 && Reg <= PPC::R10 &&((Reg >= PPC::R3 && Reg <= PPC::R10 && "Reg must be a valid argument register!" ) ? static_cast<void> (0) : __assert_fail ("Reg >= PPC::R3 && Reg <= PPC::R10 && \"Reg must be a valid argument register!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7067, __PRETTY_FUNCTION__)) |
7067 | "Reg must be a valid argument register!")((Reg >= PPC::R3 && Reg <= PPC::R10 && "Reg must be a valid argument register!" ) ? static_cast<void> (0) : __assert_fail ("Reg >= PPC::R3 && Reg <= PPC::R10 && \"Reg must be a valid argument register!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7067, __PRETTY_FUNCTION__)); |
7068 | return LASize + 4 * (Reg - PPC::R3); |
7069 | } |
7070 | |
7071 | if (PPC::G8RCRegClass.contains(Reg)) { |
7072 | assert(Reg >= PPC::X3 && Reg <= PPC::X10 &&((Reg >= PPC::X3 && Reg <= PPC::X10 && "Reg must be a valid argument register!" ) ? static_cast<void> (0) : __assert_fail ("Reg >= PPC::X3 && Reg <= PPC::X10 && \"Reg must be a valid argument register!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7073, __PRETTY_FUNCTION__)) |
7073 | "Reg must be a valid argument register!")((Reg >= PPC::X3 && Reg <= PPC::X10 && "Reg must be a valid argument register!" ) ? static_cast<void> (0) : __assert_fail ("Reg >= PPC::X3 && Reg <= PPC::X10 && \"Reg must be a valid argument register!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7073, __PRETTY_FUNCTION__)); |
7074 | return LASize + 8 * (Reg - PPC::X3); |
7075 | } |
7076 | |
7077 | llvm_unreachable("Only general purpose registers expected.")::llvm::llvm_unreachable_internal("Only general purpose registers expected." , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7077); |
7078 | } |
7079 | |
7080 | // AIX ABI Stack Frame Layout: |
7081 | // |
7082 | // Low Memory +--------------------------------------------+ |
7083 | // SP +---> | Back chain | ---+ |
7084 | // | +--------------------------------------------+ | |
7085 | // | | Saved Condition Register | | |
7086 | // | +--------------------------------------------+ | |
7087 | // | | Saved Linkage Register | | |
7088 | // | +--------------------------------------------+ | Linkage Area |
7089 | // | | Reserved for compilers | | |
7090 | // | +--------------------------------------------+ | |
7091 | // | | Reserved for binders | | |
7092 | // | +--------------------------------------------+ | |
7093 | // | | Saved TOC pointer | ---+ |
7094 | // | +--------------------------------------------+ |
7095 | // | | Parameter save area | |
7096 | // | +--------------------------------------------+ |
7097 | // | | Alloca space | |
7098 | // | +--------------------------------------------+ |
7099 | // | | Local variable space | |
7100 | // | +--------------------------------------------+ |
7101 | // | | Float/int conversion temporary | |
7102 | // | +--------------------------------------------+ |
7103 | // | | Save area for AltiVec registers | |
7104 | // | +--------------------------------------------+ |
7105 | // | | AltiVec alignment padding | |
7106 | // | +--------------------------------------------+ |
7107 | // | | Save area for VRSAVE register | |
7108 | // | +--------------------------------------------+ |
7109 | // | | Save area for General Purpose registers | |
7110 | // | +--------------------------------------------+ |
7111 | // | | Save area for Floating Point registers | |
7112 | // | +--------------------------------------------+ |
7113 | // +---- | Back chain | |
7114 | // High Memory +--------------------------------------------+ |
7115 | // |
7116 | // Specifications: |
7117 | // AIX 7.2 Assembler Language Reference |
7118 | // Subroutine linkage convention |
7119 | |
7120 | SDValue PPCTargetLowering::LowerFormalArguments_AIX( |
7121 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
7122 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
7123 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
7124 | |
7125 | assert((CallConv == CallingConv::C || CallConv == CallingConv::Cold ||(((CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && "Unexpected calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && \"Unexpected calling convention!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7127, __PRETTY_FUNCTION__)) |
7126 | CallConv == CallingConv::Fast) &&(((CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && "Unexpected calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && \"Unexpected calling convention!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7127, __PRETTY_FUNCTION__)) |
7127 | "Unexpected calling convention!")(((CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && "Unexpected calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && \"Unexpected calling convention!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7127, __PRETTY_FUNCTION__)); |
7128 | |
7129 | if (getTargetMachine().Options.GuaranteedTailCallOpt) |
7130 | report_fatal_error("Tail call support is unimplemented on AIX."); |
7131 | |
7132 | if (useSoftFloat()) |
7133 | report_fatal_error("Soft float support is unimplemented on AIX."); |
7134 | |
7135 | const PPCSubtarget &Subtarget = |
7136 | static_cast<const PPCSubtarget &>(DAG.getSubtarget()); |
7137 | |
7138 | const bool IsPPC64 = Subtarget.isPPC64(); |
7139 | const unsigned PtrByteSize = IsPPC64 ? 8 : 4; |
7140 | |
7141 | // Assign locations to all of the incoming arguments. |
7142 | SmallVector<CCValAssign, 16> ArgLocs; |
7143 | MachineFunction &MF = DAG.getMachineFunction(); |
7144 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
7145 | CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext()); |
7146 | |
7147 | const EVT PtrVT = getPointerTy(MF.getDataLayout()); |
7148 | // Reserve space for the linkage area on the stack. |
7149 | const unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
7150 | CCInfo.AllocateStack(LinkageSize, Align(PtrByteSize)); |
7151 | CCInfo.AnalyzeFormalArguments(Ins, CC_AIX); |
7152 | |
7153 | SmallVector<SDValue, 8> MemOps; |
7154 | |
7155 | for (size_t I = 0, End = ArgLocs.size(); I != End; /* No increment here */) { |
7156 | CCValAssign &VA = ArgLocs[I++]; |
7157 | MVT LocVT = VA.getLocVT(); |
7158 | ISD::ArgFlagsTy Flags = Ins[VA.getValNo()].Flags; |
7159 | |
7160 | // For compatibility with the AIX XL compiler, the float args in the |
7161 | // parameter save area are initialized even if the argument is available |
7162 | // in register. The caller is required to initialize both the register |
7163 | // and memory, however, the callee can choose to expect it in either. |
7164 | // The memloc is dismissed here because the argument is retrieved from |
7165 | // the register. |
7166 | if (VA.isMemLoc() && VA.needsCustom()) |
7167 | continue; |
7168 | |
7169 | if (Flags.isByVal() && VA.isMemLoc()) { |
7170 | const unsigned Size = |
7171 | alignTo(Flags.getByValSize() ? Flags.getByValSize() : PtrByteSize, |
7172 | PtrByteSize); |
7173 | const int FI = MF.getFrameInfo().CreateFixedObject( |
7174 | Size, VA.getLocMemOffset(), /* IsImmutable */ false, |
7175 | /* IsAliased */ true); |
7176 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
7177 | InVals.push_back(FIN); |
7178 | |
7179 | continue; |
7180 | } |
7181 | |
7182 | if (Flags.isByVal()) { |
7183 | assert(VA.isRegLoc() && "MemLocs should already be handled.")((VA.isRegLoc() && "MemLocs should already be handled." ) ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && \"MemLocs should already be handled.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7183, __PRETTY_FUNCTION__)); |
7184 | |
7185 | const MCPhysReg ArgReg = VA.getLocReg(); |
7186 | const PPCFrameLowering *FL = Subtarget.getFrameLowering(); |
7187 | |
7188 | if (Flags.getNonZeroByValAlign() > PtrByteSize) |
7189 | report_fatal_error("Over aligned byvals not supported yet."); |
7190 | |
7191 | const unsigned StackSize = alignTo(Flags.getByValSize(), PtrByteSize); |
7192 | const int FI = MF.getFrameInfo().CreateFixedObject( |
7193 | StackSize, mapArgRegToOffsetAIX(ArgReg, FL), /* IsImmutable */ false, |
7194 | /* IsAliased */ true); |
7195 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
7196 | InVals.push_back(FIN); |
7197 | |
7198 | // Add live ins for all the RegLocs for the same ByVal. |
7199 | const TargetRegisterClass *RegClass = |
7200 | IsPPC64 ? &PPC::G8RCRegClass : &PPC::GPRCRegClass; |
7201 | |
7202 | auto HandleRegLoc = [&, RegClass, LocVT](const MCPhysReg PhysReg, |
7203 | unsigned Offset) { |
7204 | const unsigned VReg = MF.addLiveIn(PhysReg, RegClass); |
7205 | // Since the callers side has left justified the aggregate in the |
7206 | // register, we can simply store the entire register into the stack |
7207 | // slot. |
7208 | SDValue CopyFrom = DAG.getCopyFromReg(Chain, dl, VReg, LocVT); |
7209 | // The store to the fixedstack object is needed becuase accessing a |
7210 | // field of the ByVal will use a gep and load. Ideally we will optimize |
7211 | // to extracting the value from the register directly, and elide the |
7212 | // stores when the arguments address is not taken, but that will need to |
7213 | // be future work. |
7214 | SDValue Store = DAG.getStore( |
7215 | CopyFrom.getValue(1), dl, CopyFrom, |
7216 | DAG.getObjectPtrOffset(dl, FIN, TypeSize::Fixed(Offset)), |
7217 | MachinePointerInfo::getFixedStack(MF, FI, Offset)); |
7218 | |
7219 | MemOps.push_back(Store); |
7220 | }; |
7221 | |
7222 | unsigned Offset = 0; |
7223 | HandleRegLoc(VA.getLocReg(), Offset); |
7224 | Offset += PtrByteSize; |
7225 | for (; Offset != StackSize && ArgLocs[I].isRegLoc(); |
7226 | Offset += PtrByteSize) { |
7227 | assert(ArgLocs[I].getValNo() == VA.getValNo() &&((ArgLocs[I].getValNo() == VA.getValNo() && "RegLocs should be for ByVal argument." ) ? static_cast<void> (0) : __assert_fail ("ArgLocs[I].getValNo() == VA.getValNo() && \"RegLocs should be for ByVal argument.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7228, __PRETTY_FUNCTION__)) |
7228 | "RegLocs should be for ByVal argument.")((ArgLocs[I].getValNo() == VA.getValNo() && "RegLocs should be for ByVal argument." ) ? static_cast<void> (0) : __assert_fail ("ArgLocs[I].getValNo() == VA.getValNo() && \"RegLocs should be for ByVal argument.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7228, __PRETTY_FUNCTION__)); |
7229 | |
7230 | const CCValAssign RL = ArgLocs[I++]; |
7231 | HandleRegLoc(RL.getLocReg(), Offset); |
7232 | } |
7233 | |
7234 | if (Offset != StackSize) { |
7235 | assert(ArgLocs[I].getValNo() == VA.getValNo() &&((ArgLocs[I].getValNo() == VA.getValNo() && "Expected MemLoc for remaining bytes." ) ? static_cast<void> (0) : __assert_fail ("ArgLocs[I].getValNo() == VA.getValNo() && \"Expected MemLoc for remaining bytes.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7236, __PRETTY_FUNCTION__)) |
7236 | "Expected MemLoc for remaining bytes.")((ArgLocs[I].getValNo() == VA.getValNo() && "Expected MemLoc for remaining bytes." ) ? static_cast<void> (0) : __assert_fail ("ArgLocs[I].getValNo() == VA.getValNo() && \"Expected MemLoc for remaining bytes.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7236, __PRETTY_FUNCTION__)); |
7237 | assert(ArgLocs[I].isMemLoc() && "Expected MemLoc for remaining bytes.")((ArgLocs[I].isMemLoc() && "Expected MemLoc for remaining bytes." ) ? static_cast<void> (0) : __assert_fail ("ArgLocs[I].isMemLoc() && \"Expected MemLoc for remaining bytes.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7237, __PRETTY_FUNCTION__)); |
7238 | // Consume the MemLoc.The InVal has already been emitted, so nothing |
7239 | // more needs to be done. |
7240 | ++I; |
7241 | } |
7242 | |
7243 | continue; |
7244 | } |
7245 | |
7246 | EVT ValVT = VA.getValVT(); |
7247 | if (VA.isRegLoc() && !VA.needsCustom()) { |
7248 | MVT::SimpleValueType SVT = ValVT.getSimpleVT().SimpleTy; |
7249 | unsigned VReg = |
7250 | MF.addLiveIn(VA.getLocReg(), getRegClassForSVT(SVT, IsPPC64)); |
7251 | SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, VReg, LocVT); |
7252 | if (ValVT.isScalarInteger() && |
7253 | (ValVT.getFixedSizeInBits() < LocVT.getFixedSizeInBits())) { |
7254 | ArgValue = |
7255 | truncateScalarIntegerArg(Flags, ValVT, DAG, ArgValue, LocVT, dl); |
7256 | } |
7257 | InVals.push_back(ArgValue); |
7258 | continue; |
7259 | } |
7260 | if (VA.isMemLoc()) { |
7261 | const unsigned LocSize = LocVT.getStoreSize(); |
7262 | const unsigned ValSize = ValVT.getStoreSize(); |
7263 | assert((ValSize <= LocSize) &&(((ValSize <= LocSize) && "Object size is larger than size of MemLoc" ) ? static_cast<void> (0) : __assert_fail ("(ValSize <= LocSize) && \"Object size is larger than size of MemLoc\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7264, __PRETTY_FUNCTION__)) |
7264 | "Object size is larger than size of MemLoc")(((ValSize <= LocSize) && "Object size is larger than size of MemLoc" ) ? static_cast<void> (0) : __assert_fail ("(ValSize <= LocSize) && \"Object size is larger than size of MemLoc\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7264, __PRETTY_FUNCTION__)); |
7265 | int CurArgOffset = VA.getLocMemOffset(); |
7266 | // Objects are right-justified because AIX is big-endian. |
7267 | if (LocSize > ValSize) |
7268 | CurArgOffset += LocSize - ValSize; |
7269 | // Potential tail calls could cause overwriting of argument stack slots. |
7270 | const bool IsImmutable = |
7271 | !(getTargetMachine().Options.GuaranteedTailCallOpt && |
7272 | (CallConv == CallingConv::Fast)); |
7273 | int FI = MFI.CreateFixedObject(ValSize, CurArgOffset, IsImmutable); |
7274 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
7275 | SDValue ArgValue = |
7276 | DAG.getLoad(ValVT, dl, Chain, FIN, MachinePointerInfo()); |
7277 | InVals.push_back(ArgValue); |
7278 | continue; |
7279 | } |
7280 | } |
7281 | |
7282 | // On AIX a minimum of 8 words is saved to the parameter save area. |
7283 | const unsigned MinParameterSaveArea = 8 * PtrByteSize; |
7284 | // Area that is at least reserved in the caller of this function. |
7285 | unsigned CallerReservedArea = |
7286 | std::max(CCInfo.getNextStackOffset(), LinkageSize + MinParameterSaveArea); |
7287 | |
7288 | // Set the size that is at least reserved in caller of this function. Tail |
7289 | // call optimized function's reserved stack space needs to be aligned so |
7290 | // that taking the difference between two stack areas will result in an |
7291 | // aligned stack. |
7292 | CallerReservedArea = |
7293 | EnsureStackAlignment(Subtarget.getFrameLowering(), CallerReservedArea); |
7294 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
7295 | FuncInfo->setMinReservedArea(CallerReservedArea); |
7296 | |
7297 | if (isVarArg) { |
7298 | FuncInfo->setVarArgsFrameIndex( |
7299 | MFI.CreateFixedObject(PtrByteSize, CCInfo.getNextStackOffset(), true)); |
7300 | SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); |
7301 | |
7302 | static const MCPhysReg GPR_32[] = {PPC::R3, PPC::R4, PPC::R5, PPC::R6, |
7303 | PPC::R7, PPC::R8, PPC::R9, PPC::R10}; |
7304 | |
7305 | static const MCPhysReg GPR_64[] = {PPC::X3, PPC::X4, PPC::X5, PPC::X6, |
7306 | PPC::X7, PPC::X8, PPC::X9, PPC::X10}; |
7307 | const unsigned NumGPArgRegs = array_lengthof(IsPPC64 ? GPR_64 : GPR_32); |
7308 | |
7309 | // The fixed integer arguments of a variadic function are stored to the |
7310 | // VarArgsFrameIndex on the stack so that they may be loaded by |
7311 | // dereferencing the result of va_next. |
7312 | for (unsigned GPRIndex = |
7313 | (CCInfo.getNextStackOffset() - LinkageSize) / PtrByteSize; |
7314 | GPRIndex < NumGPArgRegs; ++GPRIndex) { |
7315 | |
7316 | const unsigned VReg = |
7317 | IsPPC64 ? MF.addLiveIn(GPR_64[GPRIndex], &PPC::G8RCRegClass) |
7318 | : MF.addLiveIn(GPR_32[GPRIndex], &PPC::GPRCRegClass); |
7319 | |
7320 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
7321 | SDValue Store = |
7322 | DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo()); |
7323 | MemOps.push_back(Store); |
7324 | // Increment the address for the next argument to store. |
7325 | SDValue PtrOff = DAG.getConstant(PtrByteSize, dl, PtrVT); |
7326 | FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff); |
7327 | } |
7328 | } |
7329 | |
7330 | if (!MemOps.empty()) |
7331 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps); |
7332 | |
7333 | return Chain; |
7334 | } |
7335 | |
7336 | SDValue PPCTargetLowering::LowerCall_AIX( |
7337 | SDValue Chain, SDValue Callee, CallFlags CFlags, |
7338 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
7339 | const SmallVectorImpl<SDValue> &OutVals, |
7340 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
7341 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, |
7342 | const CallBase *CB) const { |
7343 | // See PPCTargetLowering::LowerFormalArguments_AIX() for a description of the |
7344 | // AIX ABI stack frame layout. |
7345 | |
7346 | assert((CFlags.CallConv == CallingConv::C ||(((CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv ::Cold || CFlags.CallConv == CallingConv::Fast) && "Unexpected calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv::Cold || CFlags.CallConv == CallingConv::Fast) && \"Unexpected calling convention!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7349, __PRETTY_FUNCTION__)) |
7347 | CFlags.CallConv == CallingConv::Cold ||(((CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv ::Cold || CFlags.CallConv == CallingConv::Fast) && "Unexpected calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv::Cold || CFlags.CallConv == CallingConv::Fast) && \"Unexpected calling convention!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7349, __PRETTY_FUNCTION__)) |
7348 | CFlags.CallConv == CallingConv::Fast) &&(((CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv ::Cold || CFlags.CallConv == CallingConv::Fast) && "Unexpected calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv::Cold || CFlags.CallConv == CallingConv::Fast) && \"Unexpected calling convention!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7349, __PRETTY_FUNCTION__)) |
7349 | "Unexpected calling convention!")(((CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv ::Cold || CFlags.CallConv == CallingConv::Fast) && "Unexpected calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv::Cold || CFlags.CallConv == CallingConv::Fast) && \"Unexpected calling convention!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7349, __PRETTY_FUNCTION__)); |
7350 | |
7351 | if (CFlags.IsPatchPoint) |
7352 | report_fatal_error("This call type is unimplemented on AIX."); |
7353 | |
7354 | const PPCSubtarget& Subtarget = |
7355 | static_cast<const PPCSubtarget&>(DAG.getSubtarget()); |
7356 | if (Subtarget.hasAltivec()) |
7357 | report_fatal_error("Altivec support is unimplemented on AIX."); |
7358 | |
7359 | MachineFunction &MF = DAG.getMachineFunction(); |
7360 | SmallVector<CCValAssign, 16> ArgLocs; |
7361 | CCState CCInfo(CFlags.CallConv, CFlags.IsVarArg, MF, ArgLocs, |
7362 | *DAG.getContext()); |
7363 | |
7364 | // Reserve space for the linkage save area (LSA) on the stack. |
7365 | // In both PPC32 and PPC64 there are 6 reserved slots in the LSA: |
7366 | // [SP][CR][LR][2 x reserved][TOC]. |
7367 | // The LSA is 24 bytes (6x4) in PPC32 and 48 bytes (6x8) in PPC64. |
7368 | const unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
7369 | const bool IsPPC64 = Subtarget.isPPC64(); |
7370 | const EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
7371 | const unsigned PtrByteSize = IsPPC64 ? 8 : 4; |
7372 | CCInfo.AllocateStack(LinkageSize, Align(PtrByteSize)); |
7373 | CCInfo.AnalyzeCallOperands(Outs, CC_AIX); |
7374 | |
7375 | // The prolog code of the callee may store up to 8 GPR argument registers to |
7376 | // the stack, allowing va_start to index over them in memory if the callee |
7377 | // is variadic. |
7378 | // Because we cannot tell if this is needed on the caller side, we have to |
7379 | // conservatively assume that it is needed. As such, make sure we have at |
7380 | // least enough stack space for the caller to store the 8 GPRs. |
7381 | const unsigned MinParameterSaveAreaSize = 8 * PtrByteSize; |
7382 | const unsigned NumBytes = std::max(LinkageSize + MinParameterSaveAreaSize, |
7383 | CCInfo.getNextStackOffset()); |
7384 | |
7385 | // Adjust the stack pointer for the new arguments... |
7386 | // These operations are automatically eliminated by the prolog/epilog pass. |
7387 | Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl); |
7388 | SDValue CallSeqStart = Chain; |
7389 | |
7390 | SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; |
7391 | SmallVector<SDValue, 8> MemOpChains; |
7392 | |
7393 | // Set up a copy of the stack pointer for loading and storing any |
7394 | // arguments that may not fit in the registers available for argument |
7395 | // passing. |
7396 | const SDValue StackPtr = IsPPC64 ? DAG.getRegister(PPC::X1, MVT::i64) |
7397 | : DAG.getRegister(PPC::R1, MVT::i32); |
7398 | |
7399 | for (unsigned I = 0, E = ArgLocs.size(); I != E;) { |
7400 | const unsigned ValNo = ArgLocs[I].getValNo(); |
7401 | SDValue Arg = OutVals[ValNo]; |
7402 | ISD::ArgFlagsTy Flags = Outs[ValNo].Flags; |
7403 | |
7404 | if (Flags.isByVal()) { |
7405 | const unsigned ByValSize = Flags.getByValSize(); |
7406 | |
7407 | // Nothing to do for zero-sized ByVals on the caller side. |
7408 | if (!ByValSize) { |
7409 | ++I; |
7410 | continue; |
7411 | } |
7412 | |
7413 | auto GetLoad = [&](EVT VT, unsigned LoadOffset) { |
7414 | return DAG.getExtLoad( |
7415 | ISD::ZEXTLOAD, dl, PtrVT, Chain, |
7416 | (LoadOffset != 0) |
7417 | ? DAG.getObjectPtrOffset(dl, Arg, TypeSize::Fixed(LoadOffset)) |
7418 | : Arg, |
7419 | MachinePointerInfo(), VT); |
7420 | }; |
7421 | |
7422 | unsigned LoadOffset = 0; |
7423 | |
7424 | // Initialize registers, which are fully occupied by the by-val argument. |
7425 | while (LoadOffset + PtrByteSize <= ByValSize && ArgLocs[I].isRegLoc()) { |
7426 | SDValue Load = GetLoad(PtrVT, LoadOffset); |
7427 | MemOpChains.push_back(Load.getValue(1)); |
7428 | LoadOffset += PtrByteSize; |
7429 | const CCValAssign &ByValVA = ArgLocs[I++]; |
7430 | assert(ByValVA.getValNo() == ValNo &&((ByValVA.getValNo() == ValNo && "Unexpected location for pass-by-value argument." ) ? static_cast<void> (0) : __assert_fail ("ByValVA.getValNo() == ValNo && \"Unexpected location for pass-by-value argument.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7431, __PRETTY_FUNCTION__)) |
7431 | "Unexpected location for pass-by-value argument.")((ByValVA.getValNo() == ValNo && "Unexpected location for pass-by-value argument." ) ? static_cast<void> (0) : __assert_fail ("ByValVA.getValNo() == ValNo && \"Unexpected location for pass-by-value argument.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7431, __PRETTY_FUNCTION__)); |
7432 | RegsToPass.push_back(std::make_pair(ByValVA.getLocReg(), Load)); |
7433 | } |
7434 | |
7435 | if (LoadOffset == ByValSize) |
7436 | continue; |
7437 | |
7438 | // There must be one more loc to handle the remainder. |
7439 | assert(ArgLocs[I].getValNo() == ValNo &&((ArgLocs[I].getValNo() == ValNo && "Expected additional location for by-value argument." ) ? static_cast<void> (0) : __assert_fail ("ArgLocs[I].getValNo() == ValNo && \"Expected additional location for by-value argument.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7440, __PRETTY_FUNCTION__)) |
7440 | "Expected additional location for by-value argument.")((ArgLocs[I].getValNo() == ValNo && "Expected additional location for by-value argument." ) ? static_cast<void> (0) : __assert_fail ("ArgLocs[I].getValNo() == ValNo && \"Expected additional location for by-value argument.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7440, __PRETTY_FUNCTION__)); |
7441 | |
7442 | if (ArgLocs[I].isMemLoc()) { |
7443 | assert(LoadOffset < ByValSize && "Unexpected memloc for by-val arg.")((LoadOffset < ByValSize && "Unexpected memloc for by-val arg." ) ? static_cast<void> (0) : __assert_fail ("LoadOffset < ByValSize && \"Unexpected memloc for by-val arg.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7443, __PRETTY_FUNCTION__)); |
7444 | const CCValAssign &ByValVA = ArgLocs[I++]; |
7445 | ISD::ArgFlagsTy MemcpyFlags = Flags; |
7446 | // Only memcpy the bytes that don't pass in register. |
7447 | MemcpyFlags.setByValSize(ByValSize - LoadOffset); |
7448 | Chain = CallSeqStart = createMemcpyOutsideCallSeq( |
7449 | (LoadOffset != 0) |
7450 | ? DAG.getObjectPtrOffset(dl, Arg, TypeSize::Fixed(LoadOffset)) |
7451 | : Arg, |
7452 | DAG.getObjectPtrOffset(dl, StackPtr, |
7453 | TypeSize::Fixed(ByValVA.getLocMemOffset())), |
7454 | CallSeqStart, MemcpyFlags, DAG, dl); |
7455 | continue; |
7456 | } |
7457 | |
7458 | // Initialize the final register residue. |
7459 | // Any residue that occupies the final by-val arg register must be |
7460 | // left-justified on AIX. Loads must be a power-of-2 size and cannot be |
7461 | // larger than the ByValSize. For example: a 7 byte by-val arg requires 4, |
7462 | // 2 and 1 byte loads. |
7463 | const unsigned ResidueBytes = ByValSize % PtrByteSize; |
7464 | assert(ResidueBytes != 0 && LoadOffset + PtrByteSize > ByValSize &&((ResidueBytes != 0 && LoadOffset + PtrByteSize > ByValSize && "Unexpected register residue for by-value argument." ) ? static_cast<void> (0) : __assert_fail ("ResidueBytes != 0 && LoadOffset + PtrByteSize > ByValSize && \"Unexpected register residue for by-value argument.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7465, __PRETTY_FUNCTION__)) |
7465 | "Unexpected register residue for by-value argument.")((ResidueBytes != 0 && LoadOffset + PtrByteSize > ByValSize && "Unexpected register residue for by-value argument." ) ? static_cast<void> (0) : __assert_fail ("ResidueBytes != 0 && LoadOffset + PtrByteSize > ByValSize && \"Unexpected register residue for by-value argument.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7465, __PRETTY_FUNCTION__)); |
7466 | SDValue ResidueVal; |
7467 | for (unsigned Bytes = 0; Bytes != ResidueBytes;) { |
7468 | const unsigned N = PowerOf2Floor(ResidueBytes - Bytes); |
7469 | const MVT VT = |
7470 | N == 1 ? MVT::i8 |
7471 | : ((N == 2) ? MVT::i16 : (N == 4 ? MVT::i32 : MVT::i64)); |
7472 | SDValue Load = GetLoad(VT, LoadOffset); |
7473 | MemOpChains.push_back(Load.getValue(1)); |
7474 | LoadOffset += N; |
7475 | Bytes += N; |
7476 | |
7477 | // By-val arguments are passed left-justfied in register. |
7478 | // Every load here needs to be shifted, otherwise a full register load |
7479 | // should have been used. |
7480 | assert(PtrVT.getSimpleVT().getSizeInBits() > (Bytes * 8) &&((PtrVT.getSimpleVT().getSizeInBits() > (Bytes * 8) && "Unexpected load emitted during handling of pass-by-value " "argument." ) ? static_cast<void> (0) : __assert_fail ("PtrVT.getSimpleVT().getSizeInBits() > (Bytes * 8) && \"Unexpected load emitted during handling of pass-by-value \" \"argument.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7482, __PRETTY_FUNCTION__)) |
7481 | "Unexpected load emitted during handling of pass-by-value "((PtrVT.getSimpleVT().getSizeInBits() > (Bytes * 8) && "Unexpected load emitted during handling of pass-by-value " "argument." ) ? static_cast<void> (0) : __assert_fail ("PtrVT.getSimpleVT().getSizeInBits() > (Bytes * 8) && \"Unexpected load emitted during handling of pass-by-value \" \"argument.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7482, __PRETTY_FUNCTION__)) |
7482 | "argument.")((PtrVT.getSimpleVT().getSizeInBits() > (Bytes * 8) && "Unexpected load emitted during handling of pass-by-value " "argument." ) ? static_cast<void> (0) : __assert_fail ("PtrVT.getSimpleVT().getSizeInBits() > (Bytes * 8) && \"Unexpected load emitted during handling of pass-by-value \" \"argument.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7482, __PRETTY_FUNCTION__)); |
7483 | unsigned NumSHLBits = PtrVT.getSimpleVT().getSizeInBits() - (Bytes * 8); |
7484 | EVT ShiftAmountTy = |
7485 | getShiftAmountTy(Load->getValueType(0), DAG.getDataLayout()); |
7486 | SDValue SHLAmt = DAG.getConstant(NumSHLBits, dl, ShiftAmountTy); |
7487 | SDValue ShiftedLoad = |
7488 | DAG.getNode(ISD::SHL, dl, Load.getValueType(), Load, SHLAmt); |
7489 | ResidueVal = ResidueVal ? DAG.getNode(ISD::OR, dl, PtrVT, ResidueVal, |
7490 | ShiftedLoad) |
7491 | : ShiftedLoad; |
7492 | } |
7493 | |
7494 | const CCValAssign &ByValVA = ArgLocs[I++]; |
7495 | RegsToPass.push_back(std::make_pair(ByValVA.getLocReg(), ResidueVal)); |
7496 | continue; |
7497 | } |
7498 | |
7499 | CCValAssign &VA = ArgLocs[I++]; |
7500 | const MVT LocVT = VA.getLocVT(); |
7501 | const MVT ValVT = VA.getValVT(); |
7502 | |
7503 | switch (VA.getLocInfo()) { |
7504 | default: |
7505 | report_fatal_error("Unexpected argument extension type."); |
7506 | case CCValAssign::Full: |
7507 | break; |
7508 | case CCValAssign::ZExt: |
7509 | Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg); |
7510 | break; |
7511 | case CCValAssign::SExt: |
7512 | Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg); |
7513 | break; |
7514 | } |
7515 | |
7516 | if (VA.isRegLoc() && !VA.needsCustom()) { |
7517 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); |
7518 | continue; |
7519 | } |
7520 | |
7521 | if (VA.isMemLoc()) { |
7522 | SDValue PtrOff = |
7523 | DAG.getConstant(VA.getLocMemOffset(), dl, StackPtr.getValueType()); |
7524 | PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff); |
7525 | MemOpChains.push_back( |
7526 | DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo())); |
7527 | |
7528 | continue; |
7529 | } |
7530 | |
7531 | // Custom handling is used for GPR initializations for vararg float |
7532 | // arguments. |
7533 | assert(VA.isRegLoc() && VA.needsCustom() && CFlags.IsVarArg &&((VA.isRegLoc() && VA.needsCustom() && CFlags .IsVarArg && ValVT.isFloatingPoint() && LocVT .isInteger() && "Unexpected register handling for calling convention." ) ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && VA.needsCustom() && CFlags.IsVarArg && ValVT.isFloatingPoint() && LocVT.isInteger() && \"Unexpected register handling for calling convention.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7535, __PRETTY_FUNCTION__)) |
7534 | ValVT.isFloatingPoint() && LocVT.isInteger() &&((VA.isRegLoc() && VA.needsCustom() && CFlags .IsVarArg && ValVT.isFloatingPoint() && LocVT .isInteger() && "Unexpected register handling for calling convention." ) ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && VA.needsCustom() && CFlags.IsVarArg && ValVT.isFloatingPoint() && LocVT.isInteger() && \"Unexpected register handling for calling convention.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7535, __PRETTY_FUNCTION__)) |
7535 | "Unexpected register handling for calling convention.")((VA.isRegLoc() && VA.needsCustom() && CFlags .IsVarArg && ValVT.isFloatingPoint() && LocVT .isInteger() && "Unexpected register handling for calling convention." ) ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && VA.needsCustom() && CFlags.IsVarArg && ValVT.isFloatingPoint() && LocVT.isInteger() && \"Unexpected register handling for calling convention.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7535, __PRETTY_FUNCTION__)); |
7536 | |
7537 | SDValue ArgAsInt = |
7538 | DAG.getBitcast(MVT::getIntegerVT(ValVT.getSizeInBits()), Arg); |
7539 | |
7540 | if (Arg.getValueType().getStoreSize() == LocVT.getStoreSize()) |
7541 | // f32 in 32-bit GPR |
7542 | // f64 in 64-bit GPR |
7543 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgAsInt)); |
7544 | else if (Arg.getValueType().getFixedSizeInBits() < |
7545 | LocVT.getFixedSizeInBits()) |
7546 | // f32 in 64-bit GPR. |
7547 | RegsToPass.push_back(std::make_pair( |
7548 | VA.getLocReg(), DAG.getZExtOrTrunc(ArgAsInt, dl, LocVT))); |
7549 | else { |
7550 | // f64 in two 32-bit GPRs |
7551 | // The 2 GPRs are marked custom and expected to be adjacent in ArgLocs. |
7552 | assert(Arg.getValueType() == MVT::f64 && CFlags.IsVarArg && !IsPPC64 &&((Arg.getValueType() == MVT::f64 && CFlags.IsVarArg && !IsPPC64 && "Unexpected custom register for argument!" ) ? static_cast<void> (0) : __assert_fail ("Arg.getValueType() == MVT::f64 && CFlags.IsVarArg && !IsPPC64 && \"Unexpected custom register for argument!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7553, __PRETTY_FUNCTION__)) |
7553 | "Unexpected custom register for argument!")((Arg.getValueType() == MVT::f64 && CFlags.IsVarArg && !IsPPC64 && "Unexpected custom register for argument!" ) ? static_cast<void> (0) : __assert_fail ("Arg.getValueType() == MVT::f64 && CFlags.IsVarArg && !IsPPC64 && \"Unexpected custom register for argument!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7553, __PRETTY_FUNCTION__)); |
7554 | CCValAssign &GPR1 = VA; |
7555 | SDValue MSWAsI64 = DAG.getNode(ISD::SRL, dl, MVT::i64, ArgAsInt, |
7556 | DAG.getConstant(32, dl, MVT::i8)); |
7557 | RegsToPass.push_back(std::make_pair( |
7558 | GPR1.getLocReg(), DAG.getZExtOrTrunc(MSWAsI64, dl, MVT::i32))); |
7559 | |
7560 | if (I != E) { |
7561 | // If only 1 GPR was available, there will only be one custom GPR and |
7562 | // the argument will also pass in memory. |
7563 | CCValAssign &PeekArg = ArgLocs[I]; |
7564 | if (PeekArg.isRegLoc() && PeekArg.getValNo() == PeekArg.getValNo()) { |
7565 | assert(PeekArg.needsCustom() && "A second custom GPR is expected.")((PeekArg.needsCustom() && "A second custom GPR is expected." ) ? static_cast<void> (0) : __assert_fail ("PeekArg.needsCustom() && \"A second custom GPR is expected.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7565, __PRETTY_FUNCTION__)); |
7566 | CCValAssign &GPR2 = ArgLocs[I++]; |
7567 | RegsToPass.push_back(std::make_pair( |
7568 | GPR2.getLocReg(), DAG.getZExtOrTrunc(ArgAsInt, dl, MVT::i32))); |
7569 | } |
7570 | } |
7571 | } |
7572 | } |
7573 | |
7574 | if (!MemOpChains.empty()) |
7575 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains); |
7576 | |
7577 | // For indirect calls, we need to save the TOC base to the stack for |
7578 | // restoration after the call. |
7579 | if (CFlags.IsIndirect) { |
7580 | assert(!CFlags.IsTailCall && "Indirect tail-calls not supported.")((!CFlags.IsTailCall && "Indirect tail-calls not supported." ) ? static_cast<void> (0) : __assert_fail ("!CFlags.IsTailCall && \"Indirect tail-calls not supported.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7580, __PRETTY_FUNCTION__)); |
7581 | const MCRegister TOCBaseReg = Subtarget.getTOCPointerRegister(); |
7582 | const MCRegister StackPtrReg = Subtarget.getStackPointerRegister(); |
7583 | const MVT PtrVT = Subtarget.isPPC64() ? MVT::i64 : MVT::i32; |
7584 | const unsigned TOCSaveOffset = |
7585 | Subtarget.getFrameLowering()->getTOCSaveOffset(); |
7586 | |
7587 | setUsesTOCBasePtr(DAG); |
7588 | SDValue Val = DAG.getCopyFromReg(Chain, dl, TOCBaseReg, PtrVT); |
7589 | SDValue PtrOff = DAG.getIntPtrConstant(TOCSaveOffset, dl); |
7590 | SDValue StackPtr = DAG.getRegister(StackPtrReg, PtrVT); |
7591 | SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff); |
7592 | Chain = DAG.getStore( |
7593 | Val.getValue(1), dl, Val, AddPtr, |
7594 | MachinePointerInfo::getStack(DAG.getMachineFunction(), TOCSaveOffset)); |
7595 | } |
7596 | |
7597 | // Build a sequence of copy-to-reg nodes chained together with token chain |
7598 | // and flag operands which copy the outgoing args into the appropriate regs. |
7599 | SDValue InFlag; |
7600 | for (auto Reg : RegsToPass) { |
7601 | Chain = DAG.getCopyToReg(Chain, dl, Reg.first, Reg.second, InFlag); |
7602 | InFlag = Chain.getValue(1); |
7603 | } |
7604 | |
7605 | const int SPDiff = 0; |
7606 | return FinishCall(CFlags, dl, DAG, RegsToPass, InFlag, Chain, CallSeqStart, |
7607 | Callee, SPDiff, NumBytes, Ins, InVals, CB); |
7608 | } |
7609 | |
7610 | bool |
7611 | PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv, |
7612 | MachineFunction &MF, bool isVarArg, |
7613 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
7614 | LLVMContext &Context) const { |
7615 | SmallVector<CCValAssign, 16> RVLocs; |
7616 | CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context); |
7617 | return CCInfo.CheckReturn( |
7618 | Outs, (Subtarget.isSVR4ABI() && CallConv == CallingConv::Cold) |
7619 | ? RetCC_PPC_Cold |
7620 | : RetCC_PPC); |
7621 | } |
7622 | |
7623 | SDValue |
7624 | PPCTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, |
7625 | bool isVarArg, |
7626 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
7627 | const SmallVectorImpl<SDValue> &OutVals, |
7628 | const SDLoc &dl, SelectionDAG &DAG) const { |
7629 | SmallVector<CCValAssign, 16> RVLocs; |
7630 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, |
7631 | *DAG.getContext()); |
7632 | CCInfo.AnalyzeReturn(Outs, |
7633 | (Subtarget.isSVR4ABI() && CallConv == CallingConv::Cold) |
7634 | ? RetCC_PPC_Cold |
7635 | : RetCC_PPC); |
7636 | |
7637 | SDValue Flag; |
7638 | SmallVector<SDValue, 4> RetOps(1, Chain); |
7639 | |
7640 | // Copy the result values into the output registers. |
7641 | for (unsigned i = 0, RealResIdx = 0; i != RVLocs.size(); ++i, ++RealResIdx) { |
7642 | CCValAssign &VA = RVLocs[i]; |
7643 | 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!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7643, __PRETTY_FUNCTION__)); |
7644 | |
7645 | SDValue Arg = OutVals[RealResIdx]; |
7646 | |
7647 | if (Subtarget.isAIXABI() && |
7648 | (VA.getLocVT().isVector() || VA.getValVT().isVector())) |
7649 | report_fatal_error("Returning vector types not yet supported on AIX."); |
7650 | |
7651 | switch (VA.getLocInfo()) { |
7652 | default: llvm_unreachable("Unknown loc info!")::llvm::llvm_unreachable_internal("Unknown loc info!", "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7652); |
7653 | case CCValAssign::Full: break; |
7654 | case CCValAssign::AExt: |
7655 | Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg); |
7656 | break; |
7657 | case CCValAssign::ZExt: |
7658 | Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg); |
7659 | break; |
7660 | case CCValAssign::SExt: |
7661 | Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg); |
7662 | break; |
7663 | } |
7664 | if (Subtarget.hasSPE() && VA.getLocVT() == MVT::f64) { |
7665 | bool isLittleEndian = Subtarget.isLittleEndian(); |
7666 | // Legalize ret f64 -> ret 2 x i32. |
7667 | SDValue SVal = |
7668 | DAG.getNode(PPCISD::EXTRACT_SPE, dl, MVT::i32, Arg, |
7669 | DAG.getIntPtrConstant(isLittleEndian ? 0 : 1, dl)); |
7670 | Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), SVal, Flag); |
7671 | RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); |
7672 | SVal = DAG.getNode(PPCISD::EXTRACT_SPE, dl, MVT::i32, Arg, |
7673 | DAG.getIntPtrConstant(isLittleEndian ? 1 : 0, dl)); |
7674 | Flag = Chain.getValue(1); |
7675 | VA = RVLocs[++i]; // skip ahead to next loc |
7676 | Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), SVal, Flag); |
7677 | } else |
7678 | Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag); |
7679 | Flag = Chain.getValue(1); |
7680 | RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); |
7681 | } |
7682 | |
7683 | RetOps[0] = Chain; // Update chain. |
7684 | |
7685 | // Add the flag if we have it. |
7686 | if (Flag.getNode()) |
7687 | RetOps.push_back(Flag); |
7688 | |
7689 | return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, RetOps); |
7690 | } |
7691 | |
7692 | SDValue |
7693 | PPCTargetLowering::LowerGET_DYNAMIC_AREA_OFFSET(SDValue Op, |
7694 | SelectionDAG &DAG) const { |
7695 | SDLoc dl(Op); |
7696 | |
7697 | // Get the correct type for integers. |
7698 | EVT IntVT = Op.getValueType(); |
7699 | |
7700 | // Get the inputs. |
7701 | SDValue Chain = Op.getOperand(0); |
7702 | SDValue FPSIdx = getFramePointerFrameIndex(DAG); |
7703 | // Build a DYNAREAOFFSET node. |
7704 | SDValue Ops[2] = {Chain, FPSIdx}; |
7705 | SDVTList VTs = DAG.getVTList(IntVT); |
7706 | return DAG.getNode(PPCISD::DYNAREAOFFSET, dl, VTs, Ops); |
7707 | } |
7708 | |
7709 | SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, |
7710 | SelectionDAG &DAG) const { |
7711 | // When we pop the dynamic allocation we need to restore the SP link. |
7712 | SDLoc dl(Op); |
7713 | |
7714 | // Get the correct type for pointers. |
7715 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
7716 | |
7717 | // Construct the stack pointer operand. |
7718 | bool isPPC64 = Subtarget.isPPC64(); |
7719 | unsigned SP = isPPC64 ? PPC::X1 : PPC::R1; |
7720 | SDValue StackPtr = DAG.getRegister(SP, PtrVT); |
7721 | |
7722 | // Get the operands for the STACKRESTORE. |
7723 | SDValue Chain = Op.getOperand(0); |
7724 | SDValue SaveSP = Op.getOperand(1); |
7725 | |
7726 | // Load the old link SP. |
7727 | SDValue LoadLinkSP = |
7728 | DAG.getLoad(PtrVT, dl, Chain, StackPtr, MachinePointerInfo()); |
7729 | |
7730 | // Restore the stack pointer. |
7731 | Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP); |
7732 | |
7733 | // Store the old link SP. |
7734 | return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo()); |
7735 | } |
7736 | |
7737 | SDValue PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG &DAG) const { |
7738 | MachineFunction &MF = DAG.getMachineFunction(); |
7739 | bool isPPC64 = Subtarget.isPPC64(); |
7740 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
7741 | |
7742 | // Get current frame pointer save index. The users of this index will be |
7743 | // primarily DYNALLOC instructions. |
7744 | PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>(); |
7745 | int RASI = FI->getReturnAddrSaveIndex(); |
7746 | |
7747 | // If the frame pointer save index hasn't been defined yet. |
7748 | if (!RASI) { |
7749 | // Find out what the fix offset of the frame pointer save area. |
7750 | int LROffset = Subtarget.getFrameLowering()->getReturnSaveOffset(); |
7751 | // Allocate the frame index for frame pointer save area. |
7752 | RASI = MF.getFrameInfo().CreateFixedObject(isPPC64? 8 : 4, LROffset, false); |
7753 | // Save the result. |
7754 | FI->setReturnAddrSaveIndex(RASI); |
7755 | } |
7756 | return DAG.getFrameIndex(RASI, PtrVT); |
7757 | } |
7758 | |
7759 | SDValue |
7760 | PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const { |
7761 | MachineFunction &MF = DAG.getMachineFunction(); |
7762 | bool isPPC64 = Subtarget.isPPC64(); |
7763 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
7764 | |
7765 | // Get current frame pointer save index. The users of this index will be |
7766 | // primarily DYNALLOC instructions. |
7767 | PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>(); |
7768 | int FPSI = FI->getFramePointerSaveIndex(); |
7769 | |
7770 | // If the frame pointer save index hasn't been defined yet. |
7771 | if (!FPSI) { |
7772 | // Find out what the fix offset of the frame pointer save area. |
7773 | int FPOffset = Subtarget.getFrameLowering()->getFramePointerSaveOffset(); |
7774 | // Allocate the frame index for frame pointer save area. |
7775 | FPSI = MF.getFrameInfo().CreateFixedObject(isPPC64? 8 : 4, FPOffset, true); |
7776 | // Save the result. |
7777 | FI->setFramePointerSaveIndex(FPSI); |
7778 | } |
7779 | return DAG.getFrameIndex(FPSI, PtrVT); |
7780 | } |
7781 | |
7782 | SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, |
7783 | SelectionDAG &DAG) const { |
7784 | MachineFunction &MF = DAG.getMachineFunction(); |
7785 | // Get the inputs. |
7786 | SDValue Chain = Op.getOperand(0); |
7787 | SDValue Size = Op.getOperand(1); |
7788 | SDLoc dl(Op); |
7789 | |
7790 | // Get the correct type for pointers. |
7791 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
7792 | // Negate the size. |
7793 | SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT, |
7794 | DAG.getConstant(0, dl, PtrVT), Size); |
7795 | // Construct a node for the frame pointer save index. |
7796 | SDValue FPSIdx = getFramePointerFrameIndex(DAG); |
7797 | SDValue Ops[3] = { Chain, NegSize, FPSIdx }; |
7798 | SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other); |
7799 | if (hasInlineStackProbe(MF)) |
7800 | return DAG.getNode(PPCISD::PROBED_ALLOCA, dl, VTs, Ops); |
7801 | return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops); |
7802 | } |
7803 | |
7804 | SDValue PPCTargetLowering::LowerEH_DWARF_CFA(SDValue Op, |
7805 | SelectionDAG &DAG) const { |
7806 | MachineFunction &MF = DAG.getMachineFunction(); |
7807 | |
7808 | bool isPPC64 = Subtarget.isPPC64(); |
7809 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
7810 | |
7811 | int FI = MF.getFrameInfo().CreateFixedObject(isPPC64 ? 8 : 4, 0, false); |
7812 | return DAG.getFrameIndex(FI, PtrVT); |
7813 | } |
7814 | |
7815 | SDValue PPCTargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op, |
7816 | SelectionDAG &DAG) const { |
7817 | SDLoc DL(Op); |
7818 | return DAG.getNode(PPCISD::EH_SJLJ_SETJMP, DL, |
7819 | DAG.getVTList(MVT::i32, MVT::Other), |
7820 | Op.getOperand(0), Op.getOperand(1)); |
7821 | } |
7822 | |
7823 | SDValue PPCTargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op, |
7824 | SelectionDAG &DAG) const { |
7825 | SDLoc DL(Op); |
7826 | return DAG.getNode(PPCISD::EH_SJLJ_LONGJMP, DL, MVT::Other, |
7827 | Op.getOperand(0), Op.getOperand(1)); |
7828 | } |
7829 | |
7830 | SDValue PPCTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const { |
7831 | if (Op.getValueType().isVector()) |
7832 | return LowerVectorLoad(Op, DAG); |
7833 | |
7834 | assert(Op.getValueType() == MVT::i1 &&((Op.getValueType() == MVT::i1 && "Custom lowering only for i1 loads" ) ? static_cast<void> (0) : __assert_fail ("Op.getValueType() == MVT::i1 && \"Custom lowering only for i1 loads\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7835, __PRETTY_FUNCTION__)) |
7835 | "Custom lowering only for i1 loads")((Op.getValueType() == MVT::i1 && "Custom lowering only for i1 loads" ) ? static_cast<void> (0) : __assert_fail ("Op.getValueType() == MVT::i1 && \"Custom lowering only for i1 loads\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7835, __PRETTY_FUNCTION__)); |
7836 | |
7837 | // First, load 8 bits into 32 bits, then truncate to 1 bit. |
7838 | |
7839 | SDLoc dl(Op); |
7840 | LoadSDNode *LD = cast<LoadSDNode>(Op); |
7841 | |
7842 | SDValue Chain = LD->getChain(); |
7843 | SDValue BasePtr = LD->getBasePtr(); |
7844 | MachineMemOperand *MMO = LD->getMemOperand(); |
7845 | |
7846 | SDValue NewLD = |
7847 | DAG.getExtLoad(ISD::EXTLOAD, dl, getPointerTy(DAG.getDataLayout()), Chain, |
7848 | BasePtr, MVT::i8, MMO); |
7849 | SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewLD); |
7850 | |
7851 | SDValue Ops[] = { Result, SDValue(NewLD.getNode(), 1) }; |
7852 | return DAG.getMergeValues(Ops, dl); |
7853 | } |
7854 | |
7855 | SDValue PPCTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const { |
7856 | if (Op.getOperand(1).getValueType().isVector()) |
7857 | return LowerVectorStore(Op, DAG); |
7858 | |
7859 | assert(Op.getOperand(1).getValueType() == MVT::i1 &&((Op.getOperand(1).getValueType() == MVT::i1 && "Custom lowering only for i1 stores" ) ? static_cast<void> (0) : __assert_fail ("Op.getOperand(1).getValueType() == MVT::i1 && \"Custom lowering only for i1 stores\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7860, __PRETTY_FUNCTION__)) |
7860 | "Custom lowering only for i1 stores")((Op.getOperand(1).getValueType() == MVT::i1 && "Custom lowering only for i1 stores" ) ? static_cast<void> (0) : __assert_fail ("Op.getOperand(1).getValueType() == MVT::i1 && \"Custom lowering only for i1 stores\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7860, __PRETTY_FUNCTION__)); |
7861 | |
7862 | // First, zero extend to 32 bits, then use a truncating store to 8 bits. |
7863 | |
7864 | SDLoc dl(Op); |
7865 | StoreSDNode *ST = cast<StoreSDNode>(Op); |
7866 | |
7867 | SDValue Chain = ST->getChain(); |
7868 | SDValue BasePtr = ST->getBasePtr(); |
7869 | SDValue Value = ST->getValue(); |
7870 | MachineMemOperand *MMO = ST->getMemOperand(); |
7871 | |
7872 | Value = DAG.getNode(ISD::ZERO_EXTEND, dl, getPointerTy(DAG.getDataLayout()), |
7873 | Value); |
7874 | return DAG.getTruncStore(Chain, dl, Value, BasePtr, MVT::i8, MMO); |
7875 | } |
7876 | |
7877 | // FIXME: Remove this once the ANDI glue bug is fixed: |
7878 | SDValue PPCTargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const { |
7879 | assert(Op.getValueType() == MVT::i1 &&((Op.getValueType() == MVT::i1 && "Custom lowering only for i1 results" ) ? static_cast<void> (0) : __assert_fail ("Op.getValueType() == MVT::i1 && \"Custom lowering only for i1 results\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7880, __PRETTY_FUNCTION__)) |
7880 | "Custom lowering only for i1 results")((Op.getValueType() == MVT::i1 && "Custom lowering only for i1 results" ) ? static_cast<void> (0) : __assert_fail ("Op.getValueType() == MVT::i1 && \"Custom lowering only for i1 results\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7880, __PRETTY_FUNCTION__)); |
7881 | |
7882 | SDLoc DL(Op); |
7883 | return DAG.getNode(PPCISD::ANDI_rec_1_GT_BIT, DL, MVT::i1, Op.getOperand(0)); |
7884 | } |
7885 | |
7886 | SDValue PPCTargetLowering::LowerTRUNCATEVector(SDValue Op, |
7887 | SelectionDAG &DAG) const { |
7888 | |
7889 | // Implements a vector truncate that fits in a vector register as a shuffle. |
7890 | // We want to legalize vector truncates down to where the source fits in |
7891 | // a vector register (and target is therefore smaller than vector register |
7892 | // size). At that point legalization will try to custom lower the sub-legal |
7893 | // result and get here - where we can contain the truncate as a single target |
7894 | // operation. |
7895 | |
7896 | // For example a trunc <2 x i16> to <2 x i8> could be visualized as follows: |
7897 | // <MSB1|LSB1, MSB2|LSB2> to <LSB1, LSB2> |
7898 | // |
7899 | // We will implement it for big-endian ordering as this (where x denotes |
7900 | // undefined): |
7901 | // < MSB1|LSB1, MSB2|LSB2, uu, uu, uu, uu, uu, uu> to |
7902 | // < LSB1, LSB2, u, u, u, u, u, u, u, u, u, u, u, u, u, u> |
7903 | // |
7904 | // The same operation in little-endian ordering will be: |
7905 | // <uu, uu, uu, uu, uu, uu, LSB2|MSB2, LSB1|MSB1> to |
7906 | // <u, u, u, u, u, u, u, u, u, u, u, u, u, u, LSB2, LSB1> |
7907 | |
7908 | EVT TrgVT = Op.getValueType(); |
7909 | assert(TrgVT.isVector() && "Vector type expected.")((TrgVT.isVector() && "Vector type expected.") ? static_cast <void> (0) : __assert_fail ("TrgVT.isVector() && \"Vector type expected.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7909, __PRETTY_FUNCTION__)); |
7910 | unsigned TrgNumElts = TrgVT.getVectorNumElements(); |
7911 | EVT EltVT = TrgVT.getVectorElementType(); |
7912 | if (!isOperationCustom(Op.getOpcode(), TrgVT) || |
7913 | TrgVT.getSizeInBits() > 128 || !isPowerOf2_32(TrgNumElts) || |
7914 | !isPowerOf2_32(EltVT.getSizeInBits())) |
7915 | return SDValue(); |
7916 | |
7917 | SDValue N1 = Op.getOperand(0); |
7918 | EVT SrcVT = N1.getValueType(); |
7919 | unsigned SrcSize = SrcVT.getSizeInBits(); |
7920 | if (SrcSize > 256 || |
7921 | !isPowerOf2_32(SrcVT.getVectorNumElements()) || |
7922 | !isPowerOf2_32(SrcVT.getVectorElementType().getSizeInBits())) |
7923 | return SDValue(); |
7924 | if (SrcSize == 256 && SrcVT.getVectorNumElements() < 2) |
7925 | return SDValue(); |
7926 | |
7927 | unsigned WideNumElts = 128 / EltVT.getSizeInBits(); |
7928 | EVT WideVT = EVT::getVectorVT(*DAG.getContext(), EltVT, WideNumElts); |
7929 | |
7930 | SDLoc DL(Op); |
7931 | SDValue Op1, Op2; |
7932 | if (SrcSize == 256) { |
7933 | EVT VecIdxTy = getVectorIdxTy(DAG.getDataLayout()); |
7934 | EVT SplitVT = |
7935 | N1.getValueType().getHalfNumVectorElementsVT(*DAG.getContext()); |
7936 | unsigned SplitNumElts = SplitVT.getVectorNumElements(); |
7937 | Op1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, N1, |
7938 | DAG.getConstant(0, DL, VecIdxTy)); |
7939 | Op2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, N1, |
7940 | DAG.getConstant(SplitNumElts, DL, VecIdxTy)); |
7941 | } |
7942 | else { |
7943 | Op1 = SrcSize == 128 ? N1 : widenVec(DAG, N1, DL); |
7944 | Op2 = DAG.getUNDEF(WideVT); |
7945 | } |
7946 | |
7947 | // First list the elements we want to keep. |
7948 | unsigned SizeMult = SrcSize / TrgVT.getSizeInBits(); |
7949 | SmallVector<int, 16> ShuffV; |
7950 | if (Subtarget.isLittleEndian()) |
7951 | for (unsigned i = 0; i < TrgNumElts; ++i) |
7952 | ShuffV.push_back(i * SizeMult); |
7953 | else |
7954 | for (unsigned i = 1; i <= TrgNumElts; ++i) |
7955 | ShuffV.push_back(i * SizeMult - 1); |
7956 | |
7957 | // Populate the remaining elements with undefs. |
7958 | for (unsigned i = TrgNumElts; i < WideNumElts; ++i) |
7959 | // ShuffV.push_back(i + WideNumElts); |
7960 | ShuffV.push_back(WideNumElts + 1); |
7961 | |
7962 | Op1 = DAG.getNode(ISD::BITCAST, DL, WideVT, Op1); |
7963 | Op2 = DAG.getNode(ISD::BITCAST, DL, WideVT, Op2); |
7964 | return DAG.getVectorShuffle(WideVT, DL, Op1, Op2, ShuffV); |
7965 | } |
7966 | |
7967 | /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when |
7968 | /// possible. |
7969 | SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const { |
7970 | // Not FP, or using SPE? Not a fsel. |
7971 | if (!Op.getOperand(0).getValueType().isFloatingPoint() || |
7972 | !Op.getOperand(2).getValueType().isFloatingPoint() || Subtarget.hasSPE()) |
7973 | return Op; |
7974 | |
7975 | ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get(); |
7976 | |
7977 | EVT ResVT = Op.getValueType(); |
7978 | EVT CmpVT = Op.getOperand(0).getValueType(); |
7979 | SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1); |
7980 | SDValue TV = Op.getOperand(2), FV = Op.getOperand(3); |
7981 | SDLoc dl(Op); |
7982 | SDNodeFlags Flags = Op.getNode()->getFlags(); |
7983 | |
7984 | // We have xsmaxcdp/xsmincdp which are OK to emit even in the |
7985 | // presence of infinities. |
7986 | if (Subtarget.hasP9Vector() && LHS == TV && RHS == FV) { |
7987 | switch (CC) { |
7988 | default: |
7989 | break; |
7990 | case ISD::SETOGT: |
7991 | case ISD::SETGT: |
7992 | return DAG.getNode(PPCISD::XSMAXCDP, dl, Op.getValueType(), LHS, RHS); |
7993 | case ISD::SETOLT: |
7994 | case ISD::SETLT: |
7995 | return DAG.getNode(PPCISD::XSMINCDP, dl, Op.getValueType(), LHS, RHS); |
7996 | } |
7997 | } |
7998 | |
7999 | // We might be able to do better than this under some circumstances, but in |
8000 | // general, fsel-based lowering of select is a finite-math-only optimization. |
8001 | // For more information, see section F.3 of the 2.06 ISA specification. |
8002 | // With ISA 3.0 |
8003 | if ((!DAG.getTarget().Options.NoInfsFPMath && !Flags.hasNoInfs()) || |
8004 | (!DAG.getTarget().Options.NoNaNsFPMath && !Flags.hasNoNaNs())) |
8005 | return Op; |
8006 | |
8007 | // If the RHS of the comparison is a 0.0, we don't need to do the |
8008 | // subtraction at all. |
8009 | SDValue Sel1; |
8010 | if (isFloatingPointZero(RHS)) |
8011 | switch (CC) { |
8012 | default: break; // SETUO etc aren't handled by fsel. |
8013 | case ISD::SETNE: |
8014 | std::swap(TV, FV); |
8015 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
8016 | case ISD::SETEQ: |
8017 | if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits |
8018 | LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS); |
8019 | Sel1 = DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV); |
8020 | if (Sel1.getValueType() == MVT::f32) // Comparison is always 64-bits |
8021 | Sel1 = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Sel1); |
8022 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, |
8023 | DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), Sel1, FV); |
8024 | case ISD::SETULT: |
8025 | case ISD::SETLT: |
8026 | std::swap(TV, FV); // fsel is natively setge, swap operands for setlt |
8027 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
8028 | case ISD::SETOGE: |
8029 | case ISD::SETGE: |
8030 | if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits |
8031 | LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS); |
8032 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV); |
8033 | case ISD::SETUGT: |
8034 | case ISD::SETGT: |
8035 | std::swap(TV, FV); // fsel is natively setge, swap operands for setlt |
8036 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
8037 | case ISD::SETOLE: |
8038 | case ISD::SETLE: |
8039 | if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits |
8040 | LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS); |
8041 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, |
8042 | DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV); |
8043 | } |
8044 | |
8045 | SDValue Cmp; |
8046 | switch (CC) { |
8047 | default: break; // SETUO etc aren't handled by fsel. |
8048 | case ISD::SETNE: |
8049 | std::swap(TV, FV); |
8050 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
8051 | case ISD::SETEQ: |
8052 | Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS, Flags); |
8053 | if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits |
8054 | Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp); |
8055 | Sel1 = DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV); |
8056 | if (Sel1.getValueType() == MVT::f32) // Comparison is always 64-bits |
8057 | Sel1 = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Sel1); |
8058 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, |
8059 | DAG.getNode(ISD::FNEG, dl, MVT::f64, Cmp), Sel1, FV); |
8060 | case ISD::SETULT: |
8061 | case ISD::SETLT: |
8062 | Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS, Flags); |
8063 | if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits |
8064 | Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp); |
8065 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV); |
8066 | case ISD::SETOGE: |
8067 | case ISD::SETGE: |
8068 | Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS, Flags); |
8069 | if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits |
8070 | Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp); |
8071 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV); |
8072 | case ISD::SETUGT: |
8073 | case ISD::SETGT: |
8074 | Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS, Flags); |
8075 | if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits |
8076 | Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp); |
8077 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV); |
8078 | case ISD::SETOLE: |
8079 | case ISD::SETLE: |
8080 | Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS, Flags); |
8081 | if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits |
8082 | Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp); |
8083 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV); |
8084 | } |
8085 | return Op; |
8086 | } |
8087 | |
8088 | static unsigned getPPCStrictOpcode(unsigned Opc) { |
8089 | switch (Opc) { |
8090 | default: |
8091 | llvm_unreachable("No strict version of this opcode!")::llvm::llvm_unreachable_internal("No strict version of this opcode!" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8091); |
8092 | case PPCISD::FCTIDZ: |
8093 | return PPCISD::STRICT_FCTIDZ; |
8094 | case PPCISD::FCTIWZ: |
8095 | return PPCISD::STRICT_FCTIWZ; |
8096 | case PPCISD::FCTIDUZ: |
8097 | return PPCISD::STRICT_FCTIDUZ; |
8098 | case PPCISD::FCTIWUZ: |
8099 | return PPCISD::STRICT_FCTIWUZ; |
8100 | case PPCISD::FCFID: |
8101 | return PPCISD::STRICT_FCFID; |
8102 | case PPCISD::FCFIDU: |
8103 | return PPCISD::STRICT_FCFIDU; |
8104 | case PPCISD::FCFIDS: |
8105 | return PPCISD::STRICT_FCFIDS; |
8106 | case PPCISD::FCFIDUS: |
8107 | return PPCISD::STRICT_FCFIDUS; |
8108 | } |
8109 | } |
8110 | |
8111 | static SDValue convertFPToInt(SDValue Op, SelectionDAG &DAG, |
8112 | const PPCSubtarget &Subtarget) { |
8113 | SDLoc dl(Op); |
8114 | bool IsStrict = Op->isStrictFPOpcode(); |
8115 | bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT || |
8116 | Op.getOpcode() == ISD::STRICT_FP_TO_SINT; |
8117 | |
8118 | // TODO: Any other flags to propagate? |
8119 | SDNodeFlags Flags; |
8120 | Flags.setNoFPExcept(Op->getFlags().hasNoFPExcept()); |
8121 | |
8122 | // For strict nodes, source is the second operand. |
8123 | SDValue Src = Op.getOperand(IsStrict ? 1 : 0); |
8124 | SDValue Chain = IsStrict ? Op.getOperand(0) : SDValue(); |
8125 | assert(Src.getValueType().isFloatingPoint())((Src.getValueType().isFloatingPoint()) ? static_cast<void > (0) : __assert_fail ("Src.getValueType().isFloatingPoint()" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8125, __PRETTY_FUNCTION__)); |
8126 | if (Src.getValueType() == MVT::f32) { |
8127 | if (IsStrict) { |
8128 | Src = |
8129 | DAG.getNode(ISD::STRICT_FP_EXTEND, dl, |
8130 | DAG.getVTList(MVT::f64, MVT::Other), {Chain, Src}, Flags); |
8131 | Chain = Src.getValue(1); |
8132 | } else |
8133 | Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src); |
8134 | } |
8135 | SDValue Conv; |
8136 | unsigned Opc = ISD::DELETED_NODE; |
8137 | switch (Op.getSimpleValueType().SimpleTy) { |
8138 | default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!")::llvm::llvm_unreachable_internal("Unhandled FP_TO_INT type in custom expander!" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8138); |
8139 | case MVT::i32: |
8140 | Opc = IsSigned ? PPCISD::FCTIWZ |
8141 | : (Subtarget.hasFPCVT() ? PPCISD::FCTIWUZ : PPCISD::FCTIDZ); |
8142 | break; |
8143 | case MVT::i64: |
8144 | assert((IsSigned || Subtarget.hasFPCVT()) &&(((IsSigned || Subtarget.hasFPCVT()) && "i64 FP_TO_UINT is supported only with FPCVT" ) ? static_cast<void> (0) : __assert_fail ("(IsSigned || Subtarget.hasFPCVT()) && \"i64 FP_TO_UINT is supported only with FPCVT\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8145, __PRETTY_FUNCTION__)) |
8145 | "i64 FP_TO_UINT is supported only with FPCVT")(((IsSigned || Subtarget.hasFPCVT()) && "i64 FP_TO_UINT is supported only with FPCVT" ) ? static_cast<void> (0) : __assert_fail ("(IsSigned || Subtarget.hasFPCVT()) && \"i64 FP_TO_UINT is supported only with FPCVT\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8145, __PRETTY_FUNCTION__)); |
8146 | Opc = IsSigned ? PPCISD::FCTIDZ : PPCISD::FCTIDUZ; |
8147 | } |
8148 | if (IsStrict) { |
8149 | Opc = getPPCStrictOpcode(Opc); |
8150 | Conv = DAG.getNode(Opc, dl, DAG.getVTList(MVT::f64, MVT::Other), |
8151 | {Chain, Src}, Flags); |
8152 | } else { |
8153 | Conv = DAG.getNode(Opc, dl, MVT::f64, Src); |
8154 | } |
8155 | return Conv; |
8156 | } |
8157 | |
8158 | void PPCTargetLowering::LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI, |
8159 | SelectionDAG &DAG, |
8160 | const SDLoc &dl) const { |
8161 | SDValue Tmp = convertFPToInt(Op, DAG, Subtarget); |
8162 | bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT || |
8163 | Op.getOpcode() == ISD::STRICT_FP_TO_SINT; |
8164 | bool IsStrict = Op->isStrictFPOpcode(); |
8165 | |
8166 | // Convert the FP value to an int value through memory. |
8167 | bool i32Stack = Op.getValueType() == MVT::i32 && Subtarget.hasSTFIWX() && |
8168 | (IsSigned || Subtarget.hasFPCVT()); |
8169 | SDValue FIPtr = DAG.CreateStackTemporary(i32Stack ? MVT::i32 : MVT::f64); |
8170 | int FI = cast<FrameIndexSDNode>(FIPtr)->getIndex(); |
8171 | MachinePointerInfo MPI = |
8172 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI); |
8173 | |
8174 | // Emit a store to the stack slot. |
8175 | SDValue Chain = IsStrict ? Tmp.getValue(1) : DAG.getEntryNode(); |
8176 | Align Alignment(DAG.getEVTAlign(Tmp.getValueType())); |
8177 | if (i32Stack) { |
8178 | MachineFunction &MF = DAG.getMachineFunction(); |
8179 | Alignment = Align(4); |
8180 | MachineMemOperand *MMO = |
8181 | MF.getMachineMemOperand(MPI, MachineMemOperand::MOStore, 4, Alignment); |
8182 | SDValue Ops[] = { Chain, Tmp, FIPtr }; |
8183 | Chain = DAG.getMemIntrinsicNode(PPCISD::STFIWX, dl, |
8184 | DAG.getVTList(MVT::Other), Ops, MVT::i32, MMO); |
8185 | } else |
8186 | Chain = DAG.getStore(Chain, dl, Tmp, FIPtr, MPI, Alignment); |
8187 | |
8188 | // Result is a load from the stack slot. If loading 4 bytes, make sure to |
8189 | // add in a bias on big endian. |
8190 | if (Op.getValueType() == MVT::i32 && !i32Stack) { |
8191 | FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, |
8192 | DAG.getConstant(4, dl, FIPtr.getValueType())); |
8193 | MPI = MPI.getWithOffset(Subtarget.isLittleEndian() ? 0 : 4); |
8194 | } |
8195 | |
8196 | RLI.Chain = Chain; |
8197 | RLI.Ptr = FIPtr; |
8198 | RLI.MPI = MPI; |
8199 | RLI.Alignment = Alignment; |
8200 | } |
8201 | |
8202 | /// Custom lowers floating point to integer conversions to use |
8203 | /// the direct move instructions available in ISA 2.07 to avoid the |
8204 | /// need for load/store combinations. |
8205 | SDValue PPCTargetLowering::LowerFP_TO_INTDirectMove(SDValue Op, |
8206 | SelectionDAG &DAG, |
8207 | const SDLoc &dl) const { |
8208 | SDValue Conv = convertFPToInt(Op, DAG, Subtarget); |
8209 | SDValue Mov = DAG.getNode(PPCISD::MFVSR, dl, Op.getValueType(), Conv); |
8210 | if (Op->isStrictFPOpcode()) |
8211 | return DAG.getMergeValues({Mov, Conv.getValue(1)}, dl); |
8212 | else |
8213 | return Mov; |
8214 | } |
8215 | |
8216 | SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG, |
8217 | const SDLoc &dl) const { |
8218 | bool IsStrict = Op->isStrictFPOpcode(); |
8219 | bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT || |
8220 | Op.getOpcode() == ISD::STRICT_FP_TO_SINT; |
8221 | SDValue Src = Op.getOperand(IsStrict ? 1 : 0); |
8222 | EVT SrcVT = Src.getValueType(); |
8223 | EVT DstVT = Op.getValueType(); |
8224 | |
8225 | // FP to INT conversions are legal for f128. |
8226 | if (SrcVT == MVT::f128) |
8227 | return Op; |
8228 | |
8229 | // Expand ppcf128 to i32 by hand for the benefit of llvm-gcc bootstrap on |
8230 | // PPC (the libcall is not available). |
8231 | if (SrcVT == MVT::ppcf128) { |
8232 | if (DstVT == MVT::i32) { |
8233 | // TODO: Conservatively pass only nofpexcept flag here. Need to check and |
8234 | // set other fast-math flags to FP operations in both strict and |
8235 | // non-strict cases. (FP_TO_SINT, FSUB) |
8236 | SDNodeFlags Flags; |
8237 | Flags.setNoFPExcept(Op->getFlags().hasNoFPExcept()); |
8238 | |
8239 | if (IsSigned) { |
8240 | SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::f64, Src, |
8241 | DAG.getIntPtrConstant(0, dl)); |
8242 | SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::f64, Src, |
8243 | DAG.getIntPtrConstant(1, dl)); |
8244 | |
8245 | // Add the two halves of the long double in round-to-zero mode, and use |
8246 | // a smaller FP_TO_SINT. |
8247 | if (IsStrict) { |
8248 | SDValue Res = DAG.getNode(PPCISD::STRICT_FADDRTZ, dl, |
8249 | DAG.getVTList(MVT::f64, MVT::Other), |
8250 | {Op.getOperand(0), Lo, Hi}, Flags); |
8251 | return DAG.getNode(ISD::STRICT_FP_TO_SINT, dl, |
8252 | DAG.getVTList(MVT::i32, MVT::Other), |
8253 | {Res.getValue(1), Res}, Flags); |
8254 | } else { |
8255 | SDValue Res = DAG.getNode(PPCISD::FADDRTZ, dl, MVT::f64, Lo, Hi); |
8256 | return DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Res); |
8257 | } |
8258 | } else { |
8259 | const uint64_t TwoE31[] = {0x41e0000000000000LL, 0}; |
8260 | APFloat APF = APFloat(APFloat::PPCDoubleDouble(), APInt(128, TwoE31)); |
8261 | SDValue Cst = DAG.getConstantFP(APF, dl, SrcVT); |
8262 | SDValue SignMask = DAG.getConstant(0x80000000, dl, DstVT); |
8263 | if (IsStrict) { |
8264 | // Sel = Src < 0x80000000 |
8265 | // FltOfs = select Sel, 0.0, 0x80000000 |
8266 | // IntOfs = select Sel, 0, 0x80000000 |
8267 | // Result = fp_to_sint(Src - FltOfs) ^ IntOfs |
8268 | SDValue Chain = Op.getOperand(0); |
8269 | EVT SetCCVT = |
8270 | getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), SrcVT); |
8271 | EVT DstSetCCVT = |
8272 | getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), DstVT); |
8273 | SDValue Sel = DAG.getSetCC(dl, SetCCVT, Src, Cst, ISD::SETLT, |
8274 | Chain, true); |
8275 | Chain = Sel.getValue(1); |
8276 | |
8277 | SDValue FltOfs = DAG.getSelect( |
8278 | dl, SrcVT, Sel, DAG.getConstantFP(0.0, dl, SrcVT), Cst); |
8279 | Sel = DAG.getBoolExtOrTrunc(Sel, dl, DstSetCCVT, DstVT); |
8280 | |
8281 | SDValue Val = DAG.getNode(ISD::STRICT_FSUB, dl, |
8282 | DAG.getVTList(SrcVT, MVT::Other), |
8283 | {Chain, Src, FltOfs}, Flags); |
8284 | Chain = Val.getValue(1); |
8285 | SDValue SInt = DAG.getNode(ISD::STRICT_FP_TO_SINT, dl, |
8286 | DAG.getVTList(DstVT, MVT::Other), |
8287 | {Chain, Val}, Flags); |
8288 | Chain = SInt.getValue(1); |
8289 | SDValue IntOfs = DAG.getSelect( |
8290 | dl, DstVT, Sel, DAG.getConstant(0, dl, DstVT), SignMask); |
8291 | SDValue Result = DAG.getNode(ISD::XOR, dl, DstVT, SInt, IntOfs); |
8292 | return DAG.getMergeValues({Result, Chain}, dl); |
8293 | } else { |
8294 | // X>=2^31 ? (int)(X-2^31)+0x80000000 : (int)X |
8295 | // FIXME: generated code sucks. |
8296 | SDValue True = DAG.getNode(ISD::FSUB, dl, MVT::ppcf128, Src, Cst); |
8297 | True = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, True); |
8298 | True = DAG.getNode(ISD::ADD, dl, MVT::i32, True, SignMask); |
8299 | SDValue False = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Src); |
8300 | return DAG.getSelectCC(dl, Src, Cst, True, False, ISD::SETGE); |
8301 | } |
8302 | } |
8303 | } |
8304 | |
8305 | return SDValue(); |
8306 | } |
8307 | |
8308 | if (Subtarget.hasDirectMove() && Subtarget.isPPC64()) |
8309 | return LowerFP_TO_INTDirectMove(Op, DAG, dl); |
8310 | |
8311 | ReuseLoadInfo RLI; |
8312 | LowerFP_TO_INTForReuse(Op, RLI, DAG, dl); |
8313 | |
8314 | return DAG.getLoad(Op.getValueType(), dl, RLI.Chain, RLI.Ptr, RLI.MPI, |
8315 | RLI.Alignment, RLI.MMOFlags(), RLI.AAInfo, RLI.Ranges); |
8316 | } |
8317 | |
8318 | // We're trying to insert a regular store, S, and then a load, L. If the |
8319 | // incoming value, O, is a load, we might just be able to have our load use the |
8320 | // address used by O. However, we don't know if anything else will store to |
8321 | // that address before we can load from it. To prevent this situation, we need |
8322 | // to insert our load, L, into the chain as a peer of O. To do this, we give L |
8323 | // the same chain operand as O, we create a token factor from the chain results |
8324 | // of O and L, and we replace all uses of O's chain result with that token |
8325 | // factor (see spliceIntoChain below for this last part). |
8326 | bool PPCTargetLowering::canReuseLoadAddress(SDValue Op, EVT MemVT, |
8327 | ReuseLoadInfo &RLI, |
8328 | SelectionDAG &DAG, |
8329 | ISD::LoadExtType ET) const { |
8330 | // Conservatively skip reusing for constrained FP nodes. |
8331 | if (Op->isStrictFPOpcode()) |
8332 | return false; |
8333 | |
8334 | SDLoc dl(Op); |
8335 | bool ValidFPToUint = Op.getOpcode() == ISD::FP_TO_UINT && |
8336 | (Subtarget.hasFPCVT() || Op.getValueType() == MVT::i32); |
8337 | if (ET == ISD::NON_EXTLOAD && |
8338 | (ValidFPToUint || Op.getOpcode() == ISD::FP_TO_SINT) && |
8339 | isOperationLegalOrCustom(Op.getOpcode(), |
8340 | Op.getOperand(0).getValueType())) { |
8341 | |
8342 | LowerFP_TO_INTForReuse(Op, RLI, DAG, dl); |
8343 | return true; |
8344 | } |
8345 | |
8346 | LoadSDNode *LD = dyn_cast<LoadSDNode>(Op); |
8347 | if (!LD || LD->getExtensionType() != ET || LD->isVolatile() || |
8348 | LD->isNonTemporal()) |
8349 | return false; |
8350 | if (LD->getMemoryVT() != MemVT) |
8351 | return false; |
8352 | |
8353 | RLI.Ptr = LD->getBasePtr(); |
8354 | if (LD->isIndexed() && !LD->getOffset().isUndef()) { |
8355 | assert(LD->getAddressingMode() == ISD::PRE_INC &&((LD->getAddressingMode() == ISD::PRE_INC && "Non-pre-inc AM on PPC?" ) ? static_cast<void> (0) : __assert_fail ("LD->getAddressingMode() == ISD::PRE_INC && \"Non-pre-inc AM on PPC?\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8356, __PRETTY_FUNCTION__)) |
8356 | "Non-pre-inc AM on PPC?")((LD->getAddressingMode() == ISD::PRE_INC && "Non-pre-inc AM on PPC?" ) ? static_cast<void> (0) : __assert_fail ("LD->getAddressingMode() == ISD::PRE_INC && \"Non-pre-inc AM on PPC?\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8356, __PRETTY_FUNCTION__)); |
8357 | RLI.Ptr = DAG.getNode(ISD::ADD, dl, RLI.Ptr.getValueType(), RLI.Ptr, |
8358 | LD->getOffset()); |
8359 | } |
8360 | |
8361 | RLI.Chain = LD->getChain(); |
8362 | RLI.MPI = LD->getPointerInfo(); |
8363 | RLI.IsDereferenceable = LD->isDereferenceable(); |
8364 | RLI.IsInvariant = LD->isInvariant(); |
8365 | RLI.Alignment = LD->getAlign(); |
8366 | RLI.AAInfo = LD->getAAInfo(); |
8367 | RLI.Ranges = LD->getRanges(); |
8368 | |
8369 | RLI.ResChain = SDValue(LD, LD->isIndexed() ? 2 : 1); |
8370 | return true; |
8371 | } |
8372 | |
8373 | // Given the head of the old chain, ResChain, insert a token factor containing |
8374 | // it and NewResChain, and make users of ResChain now be users of that token |
8375 | // factor. |
8376 | // TODO: Remove and use DAG::makeEquivalentMemoryOrdering() instead. |
8377 | void PPCTargetLowering::spliceIntoChain(SDValue ResChain, |
8378 | SDValue NewResChain, |
8379 | SelectionDAG &DAG) const { |
8380 | if (!ResChain) |
8381 | return; |
8382 | |
8383 | SDLoc dl(NewResChain); |
8384 | |
8385 | SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, |
8386 | NewResChain, DAG.getUNDEF(MVT::Other)); |
8387 | assert(TF.getNode() != NewResChain.getNode() &&((TF.getNode() != NewResChain.getNode() && "A new TF really is required here" ) ? static_cast<void> (0) : __assert_fail ("TF.getNode() != NewResChain.getNode() && \"A new TF really is required here\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8388, __PRETTY_FUNCTION__)) |
8388 | "A new TF really is required here")((TF.getNode() != NewResChain.getNode() && "A new TF really is required here" ) ? static_cast<void> (0) : __assert_fail ("TF.getNode() != NewResChain.getNode() && \"A new TF really is required here\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8388, __PRETTY_FUNCTION__)); |
8389 | |
8390 | DAG.ReplaceAllUsesOfValueWith(ResChain, TF); |
8391 | DAG.UpdateNodeOperands(TF.getNode(), ResChain, NewResChain); |
8392 | } |
8393 | |
8394 | /// Analyze profitability of direct move |
8395 | /// prefer float load to int load plus direct move |
8396 | /// when there is no integer use of int load |
8397 | bool PPCTargetLowering::directMoveIsProfitable(const SDValue &Op) const { |
8398 | SDNode *Origin = Op.getOperand(0).getNode(); |
8399 | if (Origin->getOpcode() != ISD::LOAD) |
8400 | return true; |
8401 | |
8402 | // If there is no LXSIBZX/LXSIHZX, like Power8, |
8403 | // prefer direct move if the memory size is 1 or 2 bytes. |
8404 | MachineMemOperand *MMO = cast<LoadSDNode>(Origin)->getMemOperand(); |
8405 | if (!Subtarget.hasP9Vector() && MMO->getSize() <= 2) |
8406 | return true; |
8407 | |
8408 | for (SDNode::use_iterator UI = Origin->use_begin(), |
8409 | UE = Origin->use_end(); |
8410 | UI != UE; ++UI) { |
8411 | |
8412 | // Only look at the users of the loaded value. |
8413 | if (UI.getUse().get().getResNo() != 0) |
8414 | continue; |
8415 | |
8416 | if (UI->getOpcode() != ISD::SINT_TO_FP && |
8417 | UI->getOpcode() != ISD::UINT_TO_FP && |
8418 | UI->getOpcode() != ISD::STRICT_SINT_TO_FP && |
8419 | UI->getOpcode() != ISD::STRICT_UINT_TO_FP) |
8420 | return true; |
8421 | } |
8422 | |
8423 | return false; |
8424 | } |
8425 | |
8426 | static SDValue convertIntToFP(SDValue Op, SDValue Src, SelectionDAG &DAG, |
8427 | const PPCSubtarget &Subtarget, |
8428 | SDValue Chain = SDValue()) { |
8429 | bool IsSigned = Op.getOpcode() == ISD::SINT_TO_FP || |
8430 | Op.getOpcode() == ISD::STRICT_SINT_TO_FP; |
8431 | SDLoc dl(Op); |
8432 | |
8433 | // TODO: Any other flags to propagate? |
8434 | SDNodeFlags Flags; |
8435 | Flags.setNoFPExcept(Op->getFlags().hasNoFPExcept()); |
8436 | |
8437 | // If we have FCFIDS, then use it when converting to single-precision. |
8438 | // Otherwise, convert to double-precision and then round. |
8439 | bool IsSingle = Op.getValueType() == MVT::f32 && Subtarget.hasFPCVT(); |
8440 | unsigned ConvOpc = IsSingle ? (IsSigned ? PPCISD::FCFIDS : PPCISD::FCFIDUS) |
8441 | : (IsSigned ? PPCISD::FCFID : PPCISD::FCFIDU); |
8442 | EVT ConvTy = IsSingle ? MVT::f32 : MVT::f64; |
8443 | if (Op->isStrictFPOpcode()) { |
8444 | if (!Chain) |
8445 | Chain = Op.getOperand(0); |
8446 | return DAG.getNode(getPPCStrictOpcode(ConvOpc), dl, |
8447 | DAG.getVTList(ConvTy, MVT::Other), {Chain, Src}, Flags); |
8448 | } else |
8449 | return DAG.getNode(ConvOpc, dl, ConvTy, Src); |
8450 | } |
8451 | |
8452 | /// Custom lowers integer to floating point conversions to use |
8453 | /// the direct move instructions available in ISA 2.07 to avoid the |
8454 | /// need for load/store combinations. |
8455 | SDValue PPCTargetLowering::LowerINT_TO_FPDirectMove(SDValue Op, |
8456 | SelectionDAG &DAG, |
8457 | const SDLoc &dl) const { |
8458 | assert((Op.getValueType() == MVT::f32 ||(((Op.getValueType() == MVT::f32 || Op.getValueType() == MVT:: f64) && "Invalid floating point type as target of conversion" ) ? static_cast<void> (0) : __assert_fail ("(Op.getValueType() == MVT::f32 || Op.getValueType() == MVT::f64) && \"Invalid floating point type as target of conversion\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8460, __PRETTY_FUNCTION__)) |
8459 | Op.getValueType() == MVT::f64) &&(((Op.getValueType() == MVT::f32 || Op.getValueType() == MVT:: f64) && "Invalid floating point type as target of conversion" ) ? static_cast<void> (0) : __assert_fail ("(Op.getValueType() == MVT::f32 || Op.getValueType() == MVT::f64) && \"Invalid floating point type as target of conversion\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8460, __PRETTY_FUNCTION__)) |
8460 | "Invalid floating point type as target of conversion")(((Op.getValueType() == MVT::f32 || Op.getValueType() == MVT:: f64) && "Invalid floating point type as target of conversion" ) ? static_cast<void> (0) : __assert_fail ("(Op.getValueType() == MVT::f32 || Op.getValueType() == MVT::f64) && \"Invalid floating point type as target of conversion\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8460, __PRETTY_FUNCTION__)); |
8461 | assert(Subtarget.hasFPCVT() &&((Subtarget.hasFPCVT() && "Int to FP conversions with direct moves require FPCVT" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasFPCVT() && \"Int to FP conversions with direct moves require FPCVT\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8462, __PRETTY_FUNCTION__)) |
8462 | "Int to FP conversions with direct moves require FPCVT")((Subtarget.hasFPCVT() && "Int to FP conversions with direct moves require FPCVT" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasFPCVT() && \"Int to FP conversions with direct moves require FPCVT\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8462, __PRETTY_FUNCTION__)); |
8463 | SDValue Src = Op.getOperand(Op->isStrictFPOpcode() ? 1 : 0); |
8464 | bool WordInt = Src.getSimpleValueType().SimpleTy == MVT::i32; |
8465 | bool Signed = Op.getOpcode() == ISD::SINT_TO_FP || |
8466 | Op.getOpcode() == ISD::STRICT_SINT_TO_FP; |
8467 | unsigned MovOpc = (WordInt && !Signed) ? PPCISD::MTVSRZ : PPCISD::MTVSRA; |
8468 | SDValue Mov = DAG.getNode(MovOpc, dl, MVT::f64, Src); |
8469 | return convertIntToFP(Op, Mov, DAG, Subtarget); |
8470 | } |
8471 | |
8472 | static SDValue widenVec(SelectionDAG &DAG, SDValue Vec, const SDLoc &dl) { |
8473 | |
8474 | EVT VecVT = Vec.getValueType(); |
8475 | assert(VecVT.isVector() && "Expected a vector type.")((VecVT.isVector() && "Expected a vector type.") ? static_cast <void> (0) : __assert_fail ("VecVT.isVector() && \"Expected a vector type.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8475, __PRETTY_FUNCTION__)); |
8476 | assert(VecVT.getSizeInBits() < 128 && "Vector is already full width.")((VecVT.getSizeInBits() < 128 && "Vector is already full width." ) ? static_cast<void> (0) : __assert_fail ("VecVT.getSizeInBits() < 128 && \"Vector is already full width.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8476, __PRETTY_FUNCTION__)); |
8477 | |
8478 | EVT EltVT = VecVT.getVectorElementType(); |
8479 | unsigned WideNumElts = 128 / EltVT.getSizeInBits(); |
8480 | EVT WideVT = EVT::getVectorVT(*DAG.getContext(), EltVT, WideNumElts); |
8481 | |
8482 | unsigned NumConcat = WideNumElts / VecVT.getVectorNumElements(); |
8483 | SmallVector<SDValue, 16> Ops(NumConcat); |
8484 | Ops[0] = Vec; |
8485 | SDValue UndefVec = DAG.getUNDEF(VecVT); |
8486 | for (unsigned i = 1; i < NumConcat; ++i) |
8487 | Ops[i] = UndefVec; |
8488 | |
8489 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, WideVT, Ops); |
8490 | } |
8491 | |
8492 | SDValue PPCTargetLowering::LowerINT_TO_FPVector(SDValue Op, SelectionDAG &DAG, |
8493 | const SDLoc &dl) const { |
8494 | bool IsStrict = Op->isStrictFPOpcode(); |
8495 | unsigned Opc = Op.getOpcode(); |
8496 | SDValue Src = Op.getOperand(IsStrict ? 1 : 0); |
8497 | assert((Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP ||(((Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_UINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP) && "Unexpected conversion type") ? static_cast<void> (0) : __assert_fail ("(Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_UINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP) && \"Unexpected conversion type\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8499, __PRETTY_FUNCTION__)) |
8498 | Opc == ISD::STRICT_UINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP) &&(((Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_UINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP) && "Unexpected conversion type") ? static_cast<void> (0) : __assert_fail ("(Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_UINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP) && \"Unexpected conversion type\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8499, __PRETTY_FUNCTION__)) |
8499 | "Unexpected conversion type")(((Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_UINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP) && "Unexpected conversion type") ? static_cast<void> (0) : __assert_fail ("(Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_UINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP) && \"Unexpected conversion type\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8499, __PRETTY_FUNCTION__)); |
8500 | assert((Op.getValueType() == MVT::v2f64 || Op.getValueType() == MVT::v4f32) &&(((Op.getValueType() == MVT::v2f64 || Op.getValueType() == MVT ::v4f32) && "Supports conversions to v2f64/v4f32 only." ) ? static_cast<void> (0) : __assert_fail ("(Op.getValueType() == MVT::v2f64 || Op.getValueType() == MVT::v4f32) && \"Supports conversions to v2f64/v4f32 only.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8501, __PRETTY_FUNCTION__)) |
8501 | "Supports conversions to v2f64/v4f32 only.")(((Op.getValueType() == MVT::v2f64 || Op.getValueType() == MVT ::v4f32) && "Supports conversions to v2f64/v4f32 only." ) ? static_cast<void> (0) : __assert_fail ("(Op.getValueType() == MVT::v2f64 || Op.getValueType() == MVT::v4f32) && \"Supports conversions to v2f64/v4f32 only.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8501, __PRETTY_FUNCTION__)); |
8502 | |
8503 | // TODO: Any other flags to propagate? |
8504 | SDNodeFlags Flags; |
8505 | Flags.setNoFPExcept(Op->getFlags().hasNoFPExcept()); |
8506 | |
8507 | bool SignedConv = Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP; |
8508 | bool FourEltRes = Op.getValueType() == MVT::v4f32; |
8509 | |
8510 | SDValue Wide = widenVec(DAG, Src, dl); |
8511 | EVT WideVT = Wide.getValueType(); |
8512 | unsigned WideNumElts = WideVT.getVectorNumElements(); |
8513 | MVT IntermediateVT = FourEltRes ? MVT::v4i32 : MVT::v2i64; |
8514 | |
8515 | SmallVector<int, 16> ShuffV; |
8516 | for (unsigned i = 0; i < WideNumElts; ++i) |
8517 | ShuffV.push_back(i + WideNumElts); |
8518 | |
8519 | int Stride = FourEltRes ? WideNumElts / 4 : WideNumElts / 2; |
8520 | int SaveElts = FourEltRes ? 4 : 2; |
8521 | if (Subtarget.isLittleEndian()) |
8522 | for (int i = 0; i < SaveElts; i++) |
8523 | ShuffV[i * Stride] = i; |
8524 | else |
8525 | for (int i = 1; i <= SaveElts; i++) |
8526 | ShuffV[i * Stride - 1] = i - 1; |
8527 | |
8528 | SDValue ShuffleSrc2 = |
8529 | SignedConv ? DAG.getUNDEF(WideVT) : DAG.getConstant(0, dl, WideVT); |
8530 | SDValue Arrange = DAG.getVectorShuffle(WideVT, dl, Wide, ShuffleSrc2, ShuffV); |
8531 | |
8532 | SDValue Extend; |
8533 | if (SignedConv) { |
8534 | Arrange = DAG.getBitcast(IntermediateVT, Arrange); |
8535 | EVT ExtVT = Src.getValueType(); |
8536 | if (Subtarget.hasP9Altivec()) |
8537 | ExtVT = EVT::getVectorVT(*DAG.getContext(), WideVT.getVectorElementType(), |
8538 | IntermediateVT.getVectorNumElements()); |
8539 | |
8540 | Extend = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, IntermediateVT, Arrange, |
8541 | DAG.getValueType(ExtVT)); |
8542 | } else |
8543 | Extend = DAG.getNode(ISD::BITCAST, dl, IntermediateVT, Arrange); |
8544 | |
8545 | if (IsStrict) |
8546 | return DAG.getNode(Opc, dl, DAG.getVTList(Op.getValueType(), MVT::Other), |
8547 | {Op.getOperand(0), Extend}, Flags); |
8548 | |
8549 | return DAG.getNode(Opc, dl, Op.getValueType(), Extend); |
8550 | } |
8551 | |
8552 | SDValue PPCTargetLowering::LowerINT_TO_FP(SDValue Op, |
8553 | SelectionDAG &DAG) const { |
8554 | SDLoc dl(Op); |
8555 | bool IsSigned = Op.getOpcode() == ISD::SINT_TO_FP || |
8556 | Op.getOpcode() == ISD::STRICT_SINT_TO_FP; |
8557 | bool IsStrict = Op->isStrictFPOpcode(); |
8558 | SDValue Src = Op.getOperand(IsStrict ? 1 : 0); |
8559 | SDValue Chain = IsStrict ? Op.getOperand(0) : DAG.getEntryNode(); |
8560 | |
8561 | // TODO: Any other flags to propagate? |
8562 | SDNodeFlags Flags; |
8563 | Flags.setNoFPExcept(Op->getFlags().hasNoFPExcept()); |
8564 | |
8565 | EVT InVT = Src.getValueType(); |
8566 | EVT OutVT = Op.getValueType(); |
8567 | if (OutVT.isVector() && OutVT.isFloatingPoint() && |
8568 | isOperationCustom(Op.getOpcode(), InVT)) |
8569 | return LowerINT_TO_FPVector(Op, DAG, dl); |
8570 | |
8571 | // Conversions to f128 are legal. |
8572 | if (Op.getValueType() == MVT::f128) |
8573 | return Op; |
8574 | |
8575 | // Don't handle ppc_fp128 here; let it be lowered to a libcall. |
8576 | if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64) |
8577 | return SDValue(); |
8578 | |
8579 | if (Src.getValueType() == MVT::i1) |
8580 | return DAG.getNode(ISD::SELECT, dl, Op.getValueType(), Src, |
8581 | DAG.getConstantFP(1.0, dl, Op.getValueType()), |
8582 | DAG.getConstantFP(0.0, dl, Op.getValueType())); |
8583 | |
8584 | // If we have direct moves, we can do all the conversion, skip the store/load |
8585 | // however, without FPCVT we can't do most conversions. |
8586 | if (Subtarget.hasDirectMove() && directMoveIsProfitable(Op) && |
8587 | Subtarget.isPPC64() && Subtarget.hasFPCVT()) |
8588 | return LowerINT_TO_FPDirectMove(Op, DAG, dl); |
8589 | |
8590 | assert((IsSigned || Subtarget.hasFPCVT()) &&(((IsSigned || Subtarget.hasFPCVT()) && "UINT_TO_FP is supported only with FPCVT" ) ? static_cast<void> (0) : __assert_fail ("(IsSigned || Subtarget.hasFPCVT()) && \"UINT_TO_FP is supported only with FPCVT\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8591, __PRETTY_FUNCTION__)) |
8591 | "UINT_TO_FP is supported only with FPCVT")(((IsSigned || Subtarget.hasFPCVT()) && "UINT_TO_FP is supported only with FPCVT" ) ? static_cast<void> (0) : __assert_fail ("(IsSigned || Subtarget.hasFPCVT()) && \"UINT_TO_FP is supported only with FPCVT\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8591, __PRETTY_FUNCTION__)); |
8592 | |
8593 | if (Src.getValueType() == MVT::i64) { |
8594 | SDValue SINT = Src; |
8595 | // When converting to single-precision, we actually need to convert |
8596 | // to double-precision first and then round to single-precision. |
8597 | // To avoid double-rounding effects during that operation, we have |
8598 | // to prepare the input operand. Bits that might be truncated when |
8599 | // converting to double-precision are replaced by a bit that won't |
8600 | // be lost at this stage, but is below the single-precision rounding |
8601 | // position. |
8602 | // |
8603 | // However, if -enable-unsafe-fp-math is in effect, accept double |
8604 | // rounding to avoid the extra overhead. |
8605 | if (Op.getValueType() == MVT::f32 && |
8606 | !Subtarget.hasFPCVT() && |
8607 | !DAG.getTarget().Options.UnsafeFPMath) { |
8608 | |
8609 | // Twiddle input to make sure the low 11 bits are zero. (If this |
8610 | // is the case, we are guaranteed the value will fit into the 53 bit |
8611 | // mantissa of an IEEE double-precision value without rounding.) |
8612 | // If any of those low 11 bits were not zero originally, make sure |
8613 | // bit 12 (value 2048) is set instead, so that the final rounding |
8614 | // to single-precision gets the correct result. |
8615 | SDValue Round = DAG.getNode(ISD::AND, dl, MVT::i64, |
8616 | SINT, DAG.getConstant(2047, dl, MVT::i64)); |
8617 | Round = DAG.getNode(ISD::ADD, dl, MVT::i64, |
8618 | Round, DAG.getConstant(2047, dl, MVT::i64)); |
8619 | Round = DAG.getNode(ISD::OR, dl, MVT::i64, Round, SINT); |
8620 | Round = DAG.getNode(ISD::AND, dl, MVT::i64, |
8621 | Round, DAG.getConstant(-2048, dl, MVT::i64)); |
8622 | |
8623 | // However, we cannot use that value unconditionally: if the magnitude |
8624 | // of the input value is small, the bit-twiddling we did above might |
8625 | // end up visibly changing the output. Fortunately, in that case, we |
8626 | // don't need to twiddle bits since the original input will convert |
8627 | // exactly to double-precision floating-point already. Therefore, |
8628 | // construct a conditional to use the original value if the top 11 |
8629 | // bits are all sign-bit copies, and use the rounded value computed |
8630 | // above otherwise. |
8631 | SDValue Cond = DAG.getNode(ISD::SRA, dl, MVT::i64, |
8632 | SINT, DAG.getConstant(53, dl, MVT::i32)); |
8633 | Cond = DAG.getNode(ISD::ADD, dl, MVT::i64, |
8634 | Cond, DAG.getConstant(1, dl, MVT::i64)); |
8635 | Cond = DAG.getSetCC( |
8636 | dl, |
8637 | getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64), |
8638 | Cond, DAG.getConstant(1, dl, MVT::i64), ISD::SETUGT); |
8639 | |
8640 | SINT = DAG.getNode(ISD::SELECT, dl, MVT::i64, Cond, Round, SINT); |
8641 | } |
8642 | |
8643 | ReuseLoadInfo RLI; |
8644 | SDValue Bits; |
8645 | |
8646 | MachineFunction &MF = DAG.getMachineFunction(); |
8647 | if (canReuseLoadAddress(SINT, MVT::i64, RLI, DAG)) { |
8648 | Bits = DAG.getLoad(MVT::f64, dl, RLI.Chain, RLI.Ptr, RLI.MPI, |
8649 | RLI.Alignment, RLI.MMOFlags(), RLI.AAInfo, RLI.Ranges); |
8650 | spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG); |
8651 | } else if (Subtarget.hasLFIWAX() && |
8652 | canReuseLoadAddress(SINT, MVT::i32, RLI, DAG, ISD::SEXTLOAD)) { |
8653 | MachineMemOperand *MMO = |
8654 | MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4, |
8655 | RLI.Alignment, RLI.AAInfo, RLI.Ranges); |
8656 | SDValue Ops[] = { RLI.Chain, RLI.Ptr }; |
8657 | Bits = DAG.getMemIntrinsicNode(PPCISD::LFIWAX, dl, |
8658 | DAG.getVTList(MVT::f64, MVT::Other), |
8659 | Ops, MVT::i32, MMO); |
8660 | spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG); |
8661 | } else if (Subtarget.hasFPCVT() && |
8662 | canReuseLoadAddress(SINT, MVT::i32, RLI, DAG, ISD::ZEXTLOAD)) { |
8663 | MachineMemOperand *MMO = |
8664 | MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4, |
8665 | RLI.Alignment, RLI.AAInfo, RLI.Ranges); |
8666 | SDValue Ops[] = { RLI.Chain, RLI.Ptr }; |
8667 | Bits = DAG.getMemIntrinsicNode(PPCISD::LFIWZX, dl, |
8668 | DAG.getVTList(MVT::f64, MVT::Other), |
8669 | Ops, MVT::i32, MMO); |
8670 | spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG); |
8671 | } else if (((Subtarget.hasLFIWAX() && |
8672 | SINT.getOpcode() == ISD::SIGN_EXTEND) || |
8673 | (Subtarget.hasFPCVT() && |
8674 | SINT.getOpcode() == ISD::ZERO_EXTEND)) && |
8675 | SINT.getOperand(0).getValueType() == MVT::i32) { |
8676 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
8677 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
8678 | |
8679 | int FrameIdx = MFI.CreateStackObject(4, Align(4), false); |
8680 | SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT); |
8681 | |
8682 | SDValue Store = DAG.getStore(Chain, dl, SINT.getOperand(0), FIdx, |
8683 | MachinePointerInfo::getFixedStack( |
8684 | DAG.getMachineFunction(), FrameIdx)); |
8685 | Chain = Store; |
8686 | |
8687 | assert(cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 &&((cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && "Expected an i32 store") ? static_cast<void> (0) : __assert_fail ("cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && \"Expected an i32 store\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8688, __PRETTY_FUNCTION__)) |
8688 | "Expected an i32 store")((cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && "Expected an i32 store") ? static_cast<void> (0) : __assert_fail ("cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && \"Expected an i32 store\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8688, __PRETTY_FUNCTION__)); |
8689 | |
8690 | RLI.Ptr = FIdx; |
8691 | RLI.Chain = Chain; |
8692 | RLI.MPI = |
8693 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx); |
8694 | RLI.Alignment = Align(4); |
8695 | |
8696 | MachineMemOperand *MMO = |
8697 | MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4, |
8698 | RLI.Alignment, RLI.AAInfo, RLI.Ranges); |
8699 | SDValue Ops[] = { RLI.Chain, RLI.Ptr }; |
8700 | Bits = DAG.getMemIntrinsicNode(SINT.getOpcode() == ISD::ZERO_EXTEND ? |
8701 | PPCISD::LFIWZX : PPCISD::LFIWAX, |
8702 | dl, DAG.getVTList(MVT::f64, MVT::Other), |
8703 | Ops, MVT::i32, MMO); |
8704 | Chain = Bits.getValue(1); |
8705 | } else |
8706 | Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, SINT); |
8707 | |
8708 | SDValue FP = convertIntToFP(Op, Bits, DAG, Subtarget, Chain); |
8709 | if (IsStrict) |
8710 | Chain = FP.getValue(1); |
8711 | |
8712 | if (Op.getValueType() == MVT::f32 && !Subtarget.hasFPCVT()) { |
8713 | if (IsStrict) |
8714 | FP = DAG.getNode(ISD::STRICT_FP_ROUND, dl, |
8715 | DAG.getVTList(MVT::f32, MVT::Other), |
8716 | {Chain, FP, DAG.getIntPtrConstant(0, dl)}, Flags); |
8717 | else |
8718 | FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, |
8719 | DAG.getIntPtrConstant(0, dl)); |
8720 | } |
8721 | return FP; |
8722 | } |
8723 | |
8724 | assert(Src.getValueType() == MVT::i32 &&((Src.getValueType() == MVT::i32 && "Unhandled INT_TO_FP type in custom expander!" ) ? static_cast<void> (0) : __assert_fail ("Src.getValueType() == MVT::i32 && \"Unhandled INT_TO_FP type in custom expander!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8725, __PRETTY_FUNCTION__)) |
8725 | "Unhandled INT_TO_FP type in custom expander!")((Src.getValueType() == MVT::i32 && "Unhandled INT_TO_FP type in custom expander!" ) ? static_cast<void> (0) : __assert_fail ("Src.getValueType() == MVT::i32 && \"Unhandled INT_TO_FP type in custom expander!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8725, __PRETTY_FUNCTION__)); |
8726 | // Since we only generate this in 64-bit mode, we can take advantage of |
8727 | // 64-bit registers. In particular, sign extend the input value into the |
8728 | // 64-bit register with extsw, store the WHOLE 64-bit value into the stack |
8729 | // then lfd it and fcfid it. |
8730 | MachineFunction &MF = DAG.getMachineFunction(); |
8731 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
8732 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
8733 | |
8734 | SDValue Ld; |
8735 | if (Subtarget.hasLFIWAX() || Subtarget.hasFPCVT()) { |
8736 | ReuseLoadInfo RLI; |
8737 | bool ReusingLoad; |
8738 | if (!(ReusingLoad = canReuseLoadAddress(Src, MVT::i32, RLI, DAG))) { |
8739 | int FrameIdx = MFI.CreateStackObject(4, Align(4), false); |
8740 | SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT); |
8741 | |
8742 | SDValue Store = DAG.getStore(Chain, dl, Src, FIdx, |
8743 | MachinePointerInfo::getFixedStack( |
8744 | DAG.getMachineFunction(), FrameIdx)); |
8745 | Chain = Store; |
8746 | |
8747 | assert(cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 &&((cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && "Expected an i32 store") ? static_cast<void> (0) : __assert_fail ("cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && \"Expected an i32 store\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8748, __PRETTY_FUNCTION__)) |
8748 | "Expected an i32 store")((cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && "Expected an i32 store") ? static_cast<void> (0) : __assert_fail ("cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && \"Expected an i32 store\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8748, __PRETTY_FUNCTION__)); |
8749 | |
8750 | RLI.Ptr = FIdx; |
8751 | RLI.Chain = Chain; |
8752 | RLI.MPI = |
8753 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx); |
8754 | RLI.Alignment = Align(4); |
8755 | } |
8756 | |
8757 | MachineMemOperand *MMO = |
8758 | MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4, |
8759 | RLI.Alignment, RLI.AAInfo, RLI.Ranges); |
8760 | SDValue Ops[] = { RLI.Chain, RLI.Ptr }; |
8761 | Ld = DAG.getMemIntrinsicNode(IsSigned ? PPCISD::LFIWAX : PPCISD::LFIWZX, dl, |
8762 | DAG.getVTList(MVT::f64, MVT::Other), Ops, |
8763 | MVT::i32, MMO); |
8764 | Chain = Ld.getValue(1); |
8765 | if (ReusingLoad) |
8766 | spliceIntoChain(RLI.ResChain, Ld.getValue(1), DAG); |
8767 | } else { |
8768 | assert(Subtarget.isPPC64() &&((Subtarget.isPPC64() && "i32->FP without LFIWAX supported only on PPC64" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.isPPC64() && \"i32->FP without LFIWAX supported only on PPC64\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8769, __PRETTY_FUNCTION__)) |
8769 | "i32->FP without LFIWAX supported only on PPC64")((Subtarget.isPPC64() && "i32->FP without LFIWAX supported only on PPC64" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.isPPC64() && \"i32->FP without LFIWAX supported only on PPC64\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8769, __PRETTY_FUNCTION__)); |
8770 | |
8771 | int FrameIdx = MFI.CreateStackObject(8, Align(8), false); |
8772 | SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT); |
8773 | |
8774 | SDValue Ext64 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i64, Src); |
8775 | |
8776 | // STD the extended value into the stack slot. |
8777 | SDValue Store = DAG.getStore( |
8778 | Chain, dl, Ext64, FIdx, |
8779 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx)); |
8780 | Chain = Store; |
8781 | |
8782 | // Load the value as a double. |
8783 | Ld = DAG.getLoad( |
8784 | MVT::f64, dl, Chain, FIdx, |
8785 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx)); |
8786 | Chain = Ld.getValue(1); |
8787 | } |
8788 | |
8789 | // FCFID it and return it. |
8790 | SDValue FP = convertIntToFP(Op, Ld, DAG, Subtarget, Chain); |
8791 | if (IsStrict) |
8792 | Chain = FP.getValue(1); |
8793 | if (Op.getValueType() == MVT::f32 && !Subtarget.hasFPCVT()) { |
8794 | if (IsStrict) |
8795 | FP = DAG.getNode(ISD::STRICT_FP_ROUND, dl, |
8796 | DAG.getVTList(MVT::f32, MVT::Other), |
8797 | {Chain, FP, DAG.getIntPtrConstant(0, dl)}, Flags); |
8798 | else |
8799 | FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, |
8800 | DAG.getIntPtrConstant(0, dl)); |
8801 | } |
8802 | return FP; |
8803 | } |
8804 | |
8805 | SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op, |
8806 | SelectionDAG &DAG) const { |
8807 | SDLoc dl(Op); |
8808 | /* |
8809 | The rounding mode is in bits 30:31 of FPSR, and has the following |
8810 | settings: |
8811 | 00 Round to nearest |
8812 | 01 Round to 0 |
8813 | 10 Round to +inf |
8814 | 11 Round to -inf |
8815 | |
8816 | FLT_ROUNDS, on the other hand, expects the following: |
8817 | -1 Undefined |
8818 | 0 Round to 0 |
8819 | 1 Round to nearest |
8820 | 2 Round to +inf |
8821 | 3 Round to -inf |
8822 | |
8823 | To perform the conversion, we do: |
8824 | ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1)) |
8825 | */ |
8826 | |
8827 | MachineFunction &MF = DAG.getMachineFunction(); |
8828 | EVT VT = Op.getValueType(); |
8829 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
8830 | |
8831 | // Save FP Control Word to register |
8832 | SDValue Chain = Op.getOperand(0); |
8833 | SDValue MFFS = DAG.getNode(PPCISD::MFFS, dl, {MVT::f64, MVT::Other}, Chain); |
8834 | Chain = MFFS.getValue(1); |
8835 | |
8836 | // Save FP register to stack slot |
8837 | int SSFI = MF.getFrameInfo().CreateStackObject(8, Align(8), false); |
8838 | SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT); |
8839 | Chain = DAG.getStore(Chain, dl, MFFS, StackSlot, MachinePointerInfo()); |
8840 | |
8841 | // Load FP Control Word from low 32 bits of stack slot. |
8842 | SDValue Four = DAG.getConstant(4, dl, PtrVT); |
8843 | SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four); |
8844 | SDValue CWD = DAG.getLoad(MVT::i32, dl, Chain, Addr, MachinePointerInfo()); |
8845 | Chain = CWD.getValue(1); |
8846 | |
8847 | // Transform as necessary |
8848 | SDValue CWD1 = |
8849 | DAG.getNode(ISD::AND, dl, MVT::i32, |
8850 | CWD, DAG.getConstant(3, dl, MVT::i32)); |
8851 | SDValue CWD2 = |
8852 | DAG.getNode(ISD::SRL, dl, MVT::i32, |
8853 | DAG.getNode(ISD::AND, dl, MVT::i32, |
8854 | DAG.getNode(ISD::XOR, dl, MVT::i32, |
8855 | CWD, DAG.getConstant(3, dl, MVT::i32)), |
8856 | DAG.getConstant(3, dl, MVT::i32)), |
8857 | DAG.getConstant(1, dl, MVT::i32)); |
8858 | |
8859 | SDValue RetVal = |
8860 | DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2); |
8861 | |
8862 | RetVal = |
8863 | DAG.getNode((VT.getSizeInBits() < 16 ? ISD::TRUNCATE : ISD::ZERO_EXTEND), |
8864 | dl, VT, RetVal); |
8865 | |
8866 | return DAG.getMergeValues({RetVal, Chain}, dl); |
8867 | } |
8868 | |
8869 | SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const { |
8870 | EVT VT = Op.getValueType(); |
8871 | unsigned BitWidth = VT.getSizeInBits(); |
8872 | SDLoc dl(Op); |
8873 | assert(Op.getNumOperands() == 3 &&((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SHL!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SHL!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8875, __PRETTY_FUNCTION__)) |
8874 | VT == Op.getOperand(1).getValueType() &&((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SHL!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SHL!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8875, __PRETTY_FUNCTION__)) |
8875 | "Unexpected SHL!")((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SHL!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SHL!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8875, __PRETTY_FUNCTION__)); |
8876 | |
8877 | // Expand into a bunch of logical ops. Note that these ops |
8878 | // depend on the PPC behavior for oversized shift amounts. |
8879 | SDValue Lo = Op.getOperand(0); |
8880 | SDValue Hi = Op.getOperand(1); |
8881 | SDValue Amt = Op.getOperand(2); |
8882 | EVT AmtVT = Amt.getValueType(); |
8883 | |
8884 | SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT, |
8885 | DAG.getConstant(BitWidth, dl, AmtVT), Amt); |
8886 | SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt); |
8887 | SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1); |
8888 | SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3); |
8889 | SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt, |
8890 | DAG.getConstant(-BitWidth, dl, AmtVT)); |
8891 | SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5); |
8892 | SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6); |
8893 | SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt); |
8894 | SDValue OutOps[] = { OutLo, OutHi }; |
8895 | return DAG.getMergeValues(OutOps, dl); |
8896 | } |
8897 | |
8898 | SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const { |
8899 | EVT VT = Op.getValueType(); |
8900 | SDLoc dl(Op); |
8901 | unsigned BitWidth = VT.getSizeInBits(); |
8902 | assert(Op.getNumOperands() == 3 &&((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SRL!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SRL!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8904, __PRETTY_FUNCTION__)) |
8903 | VT == Op.getOperand(1).getValueType() &&((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SRL!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SRL!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8904, __PRETTY_FUNCTION__)) |
8904 | "Unexpected SRL!")((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SRL!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SRL!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8904, __PRETTY_FUNCTION__)); |
8905 | |
8906 | // Expand into a bunch of logical ops. Note that these ops |
8907 | // depend on the PPC behavior for oversized shift amounts. |
8908 | SDValue Lo = Op.getOperand(0); |
8909 | SDValue Hi = Op.getOperand(1); |
8910 | SDValue Amt = Op.getOperand(2); |
8911 | EVT AmtVT = Amt.getValueType(); |
8912 | |
8913 | SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT, |
8914 | DAG.getConstant(BitWidth, dl, AmtVT), Amt); |
8915 | SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt); |
8916 | SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1); |
8917 | SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3); |
8918 | SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt, |
8919 | DAG.getConstant(-BitWidth, dl, AmtVT)); |
8920 | SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5); |
8921 | SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6); |
8922 | SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt); |
8923 | SDValue OutOps[] = { OutLo, OutHi }; |
8924 | return DAG.getMergeValues(OutOps, dl); |
8925 | } |
8926 | |
8927 | SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const { |
8928 | SDLoc dl(Op); |
8929 | EVT VT = Op.getValueType(); |
8930 | unsigned BitWidth = VT.getSizeInBits(); |
8931 | assert(Op.getNumOperands() == 3 &&((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SRA!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SRA!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8933, __PRETTY_FUNCTION__)) |
8932 | VT == Op.getOperand(1).getValueType() &&((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SRA!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SRA!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8933, __PRETTY_FUNCTION__)) |
8933 | "Unexpected SRA!")((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SRA!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SRA!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8933, __PRETTY_FUNCTION__)); |
8934 | |
8935 | // Expand into a bunch of logical ops, followed by a select_cc. |
8936 | SDValue Lo = Op.getOperand(0); |
8937 | SDValue Hi = Op.getOperand(1); |
8938 | SDValue Amt = Op.getOperand(2); |
8939 | EVT AmtVT = Amt.getValueType(); |
8940 | |
8941 | SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT, |
8942 | DAG.getConstant(BitWidth, dl, AmtVT), Amt); |
8943 | SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt); |
8944 | SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1); |
8945 | SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3); |
8946 | SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt, |
8947 | DAG.getConstant(-BitWidth, dl, AmtVT)); |
8948 | SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5); |
8949 | SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt); |
8950 | SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, dl, AmtVT), |
8951 | Tmp4, Tmp6, ISD::SETLE); |
8952 | SDValue OutOps[] = { OutLo, OutHi }; |
8953 | return DAG.getMergeValues(OutOps, dl); |
8954 | } |
8955 | |
8956 | SDValue PPCTargetLowering::LowerFunnelShift(SDValue Op, |
8957 | SelectionDAG &DAG) const { |
8958 | SDLoc dl(Op); |
8959 | EVT VT = Op.getValueType(); |
8960 | unsigned BitWidth = VT.getSizeInBits(); |
8961 | |
8962 | bool IsFSHL = Op.getOpcode() == ISD::FSHL; |
8963 | SDValue X = Op.getOperand(0); |
8964 | SDValue Y = Op.getOperand(1); |
8965 | SDValue Z = Op.getOperand(2); |
8966 | EVT AmtVT = Z.getValueType(); |
8967 | |
8968 | // fshl: (X << (Z % BW)) | (Y >> (BW - (Z % BW))) |
8969 | // fshr: (X << (BW - (Z % BW))) | (Y >> (Z % BW)) |
8970 | // This is simpler than TargetLowering::expandFunnelShift because we can rely |
8971 | // on PowerPC shift by BW being well defined. |
8972 | Z = DAG.getNode(ISD::AND, dl, AmtVT, Z, |
8973 | DAG.getConstant(BitWidth - 1, dl, AmtVT)); |
8974 | SDValue SubZ = |
8975 | DAG.getNode(ISD::SUB, dl, AmtVT, DAG.getConstant(BitWidth, dl, AmtVT), Z); |
8976 | X = DAG.getNode(PPCISD::SHL, dl, VT, X, IsFSHL ? Z : SubZ); |
8977 | Y = DAG.getNode(PPCISD::SRL, dl, VT, Y, IsFSHL ? SubZ : Z); |
8978 | return DAG.getNode(ISD::OR, dl, VT, X, Y); |
8979 | } |
8980 | |
8981 | //===----------------------------------------------------------------------===// |
8982 | // Vector related lowering. |
8983 | // |
8984 | |
8985 | /// getCanonicalConstSplat - Build a canonical splat immediate of Val with an |
8986 | /// element size of SplatSize. Cast the result to VT. |
8987 | static SDValue getCanonicalConstSplat(uint64_t Val, unsigned SplatSize, EVT VT, |
8988 | SelectionDAG &DAG, const SDLoc &dl) { |
8989 | static const MVT VTys[] = { // canonical VT to use for each size. |
8990 | MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32 |
8991 | }; |
8992 | |
8993 | EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1]; |
8994 | |
8995 | // For a splat with all ones, turn it to vspltisb 0xFF to canonicalize. |
8996 | if (Val == ((1LU << (SplatSize * 8)) - 1)) { |
8997 | SplatSize = 1; |
8998 | Val = 0xFF; |
8999 | } |
9000 | |
9001 | EVT CanonicalVT = VTys[SplatSize-1]; |
9002 | |
9003 | // Build a canonical splat for this value. |
9004 | return DAG.getBitcast(ReqVT, DAG.getConstant(Val, dl, CanonicalVT)); |
9005 | } |
9006 | |
9007 | /// BuildIntrinsicOp - Return a unary operator intrinsic node with the |
9008 | /// specified intrinsic ID. |
9009 | static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op, SelectionDAG &DAG, |
9010 | const SDLoc &dl, EVT DestVT = MVT::Other) { |
9011 | if (DestVT == MVT::Other) DestVT = Op.getValueType(); |
9012 | return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT, |
9013 | DAG.getConstant(IID, dl, MVT::i32), Op); |
9014 | } |
9015 | |
9016 | /// BuildIntrinsicOp - Return a binary operator intrinsic node with the |
9017 | /// specified intrinsic ID. |
9018 | static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS, |
9019 | SelectionDAG &DAG, const SDLoc &dl, |
9020 | EVT DestVT = MVT::Other) { |
9021 | if (DestVT == MVT::Other) DestVT = LHS.getValueType(); |
9022 | return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT, |
9023 | DAG.getConstant(IID, dl, MVT::i32), LHS, RHS); |
9024 | } |
9025 | |
9026 | /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the |
9027 | /// specified intrinsic ID. |
9028 | static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1, |
9029 | SDValue Op2, SelectionDAG &DAG, const SDLoc &dl, |
9030 | EVT DestVT = MVT::Other) { |
9031 | if (DestVT == MVT::Other) DestVT = Op0.getValueType(); |
9032 | return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT, |
9033 | DAG.getConstant(IID, dl, MVT::i32), Op0, Op1, Op2); |
9034 | } |
9035 | |
9036 | /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified |
9037 | /// amount. The result has the specified value type. |
9038 | static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt, EVT VT, |
9039 | SelectionDAG &DAG, const SDLoc &dl) { |
9040 | // Force LHS/RHS to be the right type. |
9041 | LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS); |
9042 | RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS); |
9043 | |
9044 | int Ops[16]; |
9045 | for (unsigned i = 0; i != 16; ++i) |
9046 | Ops[i] = i + Amt; |
9047 | SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops); |
9048 | return DAG.getNode(ISD::BITCAST, dl, VT, T); |
9049 | } |
9050 | |
9051 | /// Do we have an efficient pattern in a .td file for this node? |
9052 | /// |
9053 | /// \param V - pointer to the BuildVectorSDNode being matched |
9054 | /// \param HasDirectMove - does this subtarget have VSR <-> GPR direct moves? |
9055 | /// |
9056 | /// There are some patterns where it is beneficial to keep a BUILD_VECTOR |
9057 | /// node as a BUILD_VECTOR node rather than expanding it. The patterns where |
9058 | /// the opposite is true (expansion is beneficial) are: |
9059 | /// - The node builds a vector out of integers that are not 32 or 64-bits |
9060 | /// - The node builds a vector out of constants |
9061 | /// - The node is a "load-and-splat" |
9062 | /// In all other cases, we will choose to keep the BUILD_VECTOR. |
9063 | static bool haveEfficientBuildVectorPattern(BuildVectorSDNode *V, |
9064 | bool HasDirectMove, |
9065 | bool HasP8Vector) { |
9066 | EVT VecVT = V->getValueType(0); |
9067 | bool RightType = VecVT == MVT::v2f64 || |
9068 | (HasP8Vector && VecVT == MVT::v4f32) || |
9069 | (HasDirectMove && (VecVT == MVT::v2i64 || VecVT == MVT::v4i32)); |
9070 | if (!RightType) |
9071 | return false; |
9072 | |
9073 | bool IsSplat = true; |
9074 | bool IsLoad = false; |
9075 | SDValue Op0 = V->getOperand(0); |
9076 | |
9077 | // This function is called in a block that confirms the node is not a constant |
9078 | // splat. So a constant BUILD_VECTOR here means the vector is built out of |
9079 | // different constants. |
9080 | if (V->isConstant()) |
9081 | return false; |
9082 | for (int i = 0, e = V->getNumOperands(); i < e; ++i) { |
9083 | if (V->getOperand(i).isUndef()) |
9084 | return false; |
9085 | // We want to expand nodes that represent load-and-splat even if the |
9086 | // loaded value is a floating point truncation or conversion to int. |
9087 | if (V->getOperand(i).getOpcode() == ISD::LOAD || |
9088 | (V->getOperand(i).getOpcode() == ISD::FP_ROUND && |
9089 | V->getOperand(i).getOperand(0).getOpcode() == ISD::LOAD) || |
9090 | (V->getOperand(i).getOpcode() == ISD::FP_TO_SINT && |
9091 | V->getOperand(i).getOperand(0).getOpcode() == ISD::LOAD) || |
9092 | (V->getOperand(i).getOpcode() == ISD::FP_TO_UINT && |
9093 | V->getOperand(i).getOperand(0).getOpcode() == ISD::LOAD)) |
9094 | IsLoad = true; |
9095 | // If the operands are different or the input is not a load and has more |
9096 | // uses than just this BV node, then it isn't a splat. |
9097 | if (V->getOperand(i) != Op0 || |
9098 | (!IsLoad && !V->isOnlyUserOf(V->getOperand(i).getNode()))) |
9099 | IsSplat = false; |
9100 | } |
9101 | return !(IsSplat && IsLoad); |
9102 | } |
9103 | |
9104 | // Lower BITCAST(f128, (build_pair i64, i64)) to BUILD_FP128. |
9105 | SDValue PPCTargetLowering::LowerBITCAST(SDValue Op, SelectionDAG &DAG) const { |
9106 | |
9107 | SDLoc dl(Op); |
9108 | SDValue Op0 = Op->getOperand(0); |
9109 | |
9110 | if ((Op.getValueType() != MVT::f128) || |
9111 | (Op0.getOpcode() != ISD::BUILD_PAIR) || |
9112 | (Op0.getOperand(0).getValueType() != MVT::i64) || |
9113 | (Op0.getOperand(1).getValueType() != MVT::i64)) |
9114 | return SDValue(); |
9115 | |
9116 | return DAG.getNode(PPCISD::BUILD_FP128, dl, MVT::f128, Op0.getOperand(0), |
9117 | Op0.getOperand(1)); |
9118 | } |
9119 | |
9120 | static const SDValue *getNormalLoadInput(const SDValue &Op, bool &IsPermuted) { |
9121 | const SDValue *InputLoad = &Op; |
9122 | if (InputLoad->getOpcode() == ISD::BITCAST) |
9123 | InputLoad = &InputLoad->getOperand(0); |
9124 | if (InputLoad->getOpcode() == ISD::SCALAR_TO_VECTOR || |
9125 | InputLoad->getOpcode() == PPCISD::SCALAR_TO_VECTOR_PERMUTED) { |
9126 | IsPermuted = InputLoad->getOpcode() == PPCISD::SCALAR_TO_VECTOR_PERMUTED; |
9127 | InputLoad = &InputLoad->getOperand(0); |
9128 | } |
9129 | if (InputLoad->getOpcode() != ISD::LOAD) |
9130 | return nullptr; |
9131 | LoadSDNode *LD = cast<LoadSDNode>(*InputLoad); |
9132 | return ISD::isNormalLoad(LD) ? InputLoad : nullptr; |
9133 | } |
9134 | |
9135 | // Convert the argument APFloat to a single precision APFloat if there is no |
9136 | // loss in information during the conversion to single precision APFloat and the |
9137 | // resulting number is not a denormal number. Return true if successful. |
9138 | bool llvm::convertToNonDenormSingle(APFloat &ArgAPFloat) { |
9139 | APFloat APFloatToConvert = ArgAPFloat; |
9140 | bool LosesInfo = true; |
9141 | APFloatToConvert.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven, |
9142 | &LosesInfo); |
9143 | bool Success = (!LosesInfo && !APFloatToConvert.isDenormal()); |
9144 | if (Success) |
9145 | ArgAPFloat = APFloatToConvert; |
9146 | return Success; |
9147 | } |
9148 | |
9149 | // Bitcast the argument APInt to a double and convert it to a single precision |
9150 | // APFloat, bitcast the APFloat to an APInt and assign it to the original |
9151 | // argument if there is no loss in information during the conversion from |
9152 | // double to single precision APFloat and the resulting number is not a denormal |
9153 | // number. Return true if successful. |
9154 | bool llvm::convertToNonDenormSingle(APInt &ArgAPInt) { |
9155 | double DpValue = ArgAPInt.bitsToDouble(); |
9156 | APFloat APFloatDp(DpValue); |
9157 | bool Success = convertToNonDenormSingle(APFloatDp); |
9158 | if (Success) |
9159 | ArgAPInt = APFloatDp.bitcastToAPInt(); |
9160 | return Success; |
9161 | } |
9162 | |
9163 | // If this is a case we can't handle, return null and let the default |
9164 | // expansion code take care of it. If we CAN select this case, and if it |
9165 | // selects to a single instruction, return Op. Otherwise, if we can codegen |
9166 | // this case more efficiently than a constant pool load, lower it to the |
9167 | // sequence of ops that should be used. |
9168 | SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op, |
9169 | SelectionDAG &DAG) const { |
9170 | SDLoc dl(Op); |
9171 | BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode()); |
9172 | assert(BVN && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR")((BVN && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR" ) ? static_cast<void> (0) : __assert_fail ("BVN && \"Expected a BuildVectorSDNode in LowerBUILD_VECTOR\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9172, __PRETTY_FUNCTION__)); |
9173 | |
9174 | // Check if this is a splat of a constant value. |
9175 | APInt APSplatBits, APSplatUndef; |
9176 | unsigned SplatBitSize; |
9177 | bool HasAnyUndefs; |
9178 | bool BVNIsConstantSplat = |
9179 | BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize, |
9180 | HasAnyUndefs, 0, !Subtarget.isLittleEndian()); |
9181 | |
9182 | // If it is a splat of a double, check if we can shrink it to a 32 bit |
9183 | // non-denormal float which when converted back to double gives us the same |
9184 | // double. This is to exploit the XXSPLTIDP instruction. |
9185 | if (BVNIsConstantSplat && Subtarget.hasPrefixInstrs() && |
9186 | (SplatBitSize == 64) && (Op->getValueType(0) == MVT::v2f64) && |
9187 | convertToNonDenormSingle(APSplatBits)) { |
9188 | SDValue SplatNode = DAG.getNode( |
9189 | PPCISD::XXSPLTI_SP_TO_DP, dl, MVT::v2f64, |
9190 | DAG.getTargetConstant(APSplatBits.getZExtValue(), dl, MVT::i32)); |
9191 | return DAG.getBitcast(Op.getValueType(), SplatNode); |
9192 | } |
9193 | |
9194 | if (!BVNIsConstantSplat || SplatBitSize > 32) { |
9195 | |
9196 | bool IsPermutedLoad = false; |
9197 | const SDValue *InputLoad = |
9198 | getNormalLoadInput(Op.getOperand(0), IsPermutedLoad); |
9199 | // Handle load-and-splat patterns as we have instructions that will do this |
9200 | // in one go. |
9201 | if (InputLoad && DAG.isSplatValue(Op, true)) { |
9202 | LoadSDNode *LD = cast<LoadSDNode>(*InputLoad); |
9203 | |
9204 | // We have handling for 4 and 8 byte elements. |
9205 | unsigned ElementSize = LD->getMemoryVT().getScalarSizeInBits(); |
9206 | |
9207 | // Checking for a single use of this load, we have to check for vector |
9208 | // width (128 bits) / ElementSize uses (since each operand of the |
9209 | // BUILD_VECTOR is a separate use of the value. |
9210 | unsigned NumUsesOfInputLD = 128 / ElementSize; |
9211 | for (SDValue BVInOp : Op->ops()) |
9212 | if (BVInOp.isUndef()) |
9213 | NumUsesOfInputLD--; |
9214 | assert(NumUsesOfInputLD > 0 && "No uses of input LD of a build_vector?")((NumUsesOfInputLD > 0 && "No uses of input LD of a build_vector?" ) ? static_cast<void> (0) : __assert_fail ("NumUsesOfInputLD > 0 && \"No uses of input LD of a build_vector?\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9214, __PRETTY_FUNCTION__)); |
9215 | if (InputLoad->getNode()->hasNUsesOfValue(NumUsesOfInputLD, 0) && |
9216 | ((Subtarget.hasVSX() && ElementSize == 64) || |
9217 | (Subtarget.hasP9Vector() && ElementSize == 32))) { |
9218 | SDValue Ops[] = { |
9219 | LD->getChain(), // Chain |
9220 | LD->getBasePtr(), // Ptr |
9221 | DAG.getValueType(Op.getValueType()) // VT |
9222 | }; |
9223 | SDValue LdSplt = DAG.getMemIntrinsicNode( |
9224 | PPCISD::LD_SPLAT, dl, DAG.getVTList(Op.getValueType(), MVT::Other), |
9225 | Ops, LD->getMemoryVT(), LD->getMemOperand()); |
9226 | // Replace all uses of the output chain of the original load with the |
9227 | // output chain of the new load. |
9228 | DAG.ReplaceAllUsesOfValueWith(InputLoad->getValue(1), |
9229 | LdSplt.getValue(1)); |
9230 | return LdSplt; |
9231 | } |
9232 | } |
9233 | |
9234 | // BUILD_VECTOR nodes that are not constant splats of up to 32-bits can be |
9235 | // lowered to VSX instructions under certain conditions. |
9236 | // Without VSX, there is no pattern more efficient than expanding the node. |
9237 | if (Subtarget.hasVSX() && |
9238 | haveEfficientBuildVectorPattern(BVN, Subtarget.hasDirectMove(), |
9239 | Subtarget.hasP8Vector())) |
9240 | return Op; |
9241 | return SDValue(); |
9242 | } |
9243 | |
9244 | uint64_t SplatBits = APSplatBits.getZExtValue(); |
9245 | uint64_t SplatUndef = APSplatUndef.getZExtValue(); |
9246 | unsigned SplatSize = SplatBitSize / 8; |
9247 | |
9248 | // First, handle single instruction cases. |
9249 | |
9250 | // All zeros? |
9251 | if (SplatBits == 0) { |
9252 | // Canonicalize all zero vectors to be v4i32. |
9253 | if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) { |
9254 | SDValue Z = DAG.getConstant(0, dl, MVT::v4i32); |
9255 | Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z); |
9256 | } |
9257 | return Op; |
9258 | } |
9259 | |
9260 | // We have XXSPLTIW for constant splats four bytes wide. |
9261 | // Given vector length is a multiple of 4, 2-byte splats can be replaced |
9262 | // with 4-byte splats. We replicate the SplatBits in case of 2-byte splat to |
9263 | // make a 4-byte splat element. For example: 2-byte splat of 0xABAB can be |
9264 | // turned into a 4-byte splat of 0xABABABAB. |
9265 | if (Subtarget.hasPrefixInstrs() && SplatSize == 2) |
9266 | return getCanonicalConstSplat((SplatBits |= SplatBits << 16), SplatSize * 2, |
Although the value stored to 'SplatBits' is used in the enclosing expression, the value is never actually read from 'SplatBits' | |
9267 | Op.getValueType(), DAG, dl); |
9268 | |
9269 | if (Subtarget.hasPrefixInstrs() && SplatSize == 4) |
9270 | return getCanonicalConstSplat(SplatBits, SplatSize, Op.getValueType(), DAG, |
9271 | dl); |
9272 | |
9273 | // We have XXSPLTIB for constant splats one byte wide. |
9274 | if (Subtarget.hasP9Vector() && SplatSize == 1) |
9275 | return getCanonicalConstSplat(SplatBits, SplatSize, Op.getValueType(), DAG, |
9276 | dl); |
9277 | |
9278 | // If the sign extended value is in the range [-16,15], use VSPLTI[bhw]. |
9279 | int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >> |
9280 | (32-SplatBitSize)); |
9281 | if (SextVal >= -16 && SextVal <= 15) |
9282 | return getCanonicalConstSplat(SextVal, SplatSize, Op.getValueType(), DAG, |
9283 | dl); |
9284 | |
9285 | // Two instruction sequences. |
9286 | |
9287 | // If this value is in the range [-32,30] and is even, use: |
9288 | // VSPLTI[bhw](val/2) + VSPLTI[bhw](val/2) |
9289 | // If this value is in the range [17,31] and is odd, use: |
9290 | // VSPLTI[bhw](val-16) - VSPLTI[bhw](-16) |
9291 | // If this value is in the range [-31,-17] and is odd, use: |
9292 | // VSPLTI[bhw](val+16) + VSPLTI[bhw](-16) |
9293 | // Note the last two are three-instruction sequences. |
9294 | if (SextVal >= -32 && SextVal <= 31) { |
9295 | // To avoid having these optimizations undone by constant folding, |
9296 | // we convert to a pseudo that will be expanded later into one of |
9297 | // the above forms. |
9298 | SDValue Elt = DAG.getConstant(SextVal, dl, MVT::i32); |
9299 | EVT VT = (SplatSize == 1 ? MVT::v16i8 : |
9300 | (SplatSize == 2 ? MVT::v8i16 : MVT::v4i32)); |
9301 | SDValue EltSize = DAG.getConstant(SplatSize, dl, MVT::i32); |
9302 | SDValue RetVal = DAG.getNode(PPCISD::VADD_SPLAT, dl, VT, Elt, EltSize); |
9303 | if (VT == Op.getValueType()) |
9304 | return RetVal; |
9305 | else |
9306 | return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), RetVal); |
9307 | } |
9308 | |
9309 | // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is |
9310 | // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important |
9311 | // for fneg/fabs. |
9312 | if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) { |
9313 | // Make -1 and vspltisw -1: |
9314 | SDValue OnesV = getCanonicalConstSplat(-1, 4, MVT::v4i32, DAG, dl); |
9315 | |
9316 | // Make the VSLW intrinsic, computing 0x8000_0000. |
9317 | SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV, |
9318 | OnesV, DAG, dl); |
9319 | |
9320 | // xor by OnesV to invert it. |
9321 | Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV); |
9322 | return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res); |
9323 | } |
9324 | |
9325 | // Check to see if this is a wide variety of vsplti*, binop self cases. |
9326 | static const signed char SplatCsts[] = { |
9327 | -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7, |
9328 | -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16 |
9329 | }; |
9330 | |
9331 | for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) { |
9332 | // Indirect through the SplatCsts array so that we favor 'vsplti -1' for |
9333 | // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1' |
9334 | int i = SplatCsts[idx]; |
9335 | |
9336 | // Figure out what shift amount will be used by altivec if shifted by i in |
9337 | // this splat size. |
9338 | unsigned TypeShiftAmt = i & (SplatBitSize-1); |
9339 | |
9340 | // vsplti + shl self. |
9341 | if (SextVal == (int)((unsigned)i << TypeShiftAmt)) { |
9342 | SDValue Res = getCanonicalConstSplat(i, SplatSize, MVT::Other, DAG, dl); |
9343 | static const unsigned IIDs[] = { // Intrinsic to use for each size. |
9344 | Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0, |
9345 | Intrinsic::ppc_altivec_vslw |
9346 | }; |
9347 | Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl); |
9348 | return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res); |
9349 | } |
9350 | |
9351 | // vsplti + srl self. |
9352 | if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) { |
9353 | SDValue Res = getCanonicalConstSplat(i, SplatSize, MVT::Other, DAG, dl); |
9354 | static const unsigned IIDs[] = { // Intrinsic to use for each size. |
9355 | Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0, |
9356 | Intrinsic::ppc_altivec_vsrw |
9357 | }; |
9358 | Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl); |
9359 | return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res); |
9360 | } |
9361 | |
9362 | // vsplti + sra self. |
9363 | if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) { |
9364 | SDValue Res = getCanonicalConstSplat(i, SplatSize, MVT::Other, DAG, dl); |
9365 | static const unsigned IIDs[] = { // Intrinsic to use for each size. |
9366 | Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0, |
9367 | Intrinsic::ppc_altivec_vsraw |
9368 | }; |
9369 | Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl); |
9370 | return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res); |
9371 | } |
9372 | |
9373 | // vsplti + rol self. |
9374 | if (SextVal == (int)(((unsigned)i << TypeShiftAmt) | |
9375 | ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) { |
9376 | SDValue Res = getCanonicalConstSplat(i, SplatSize, MVT::Other, DAG, dl); |
9377 | static const unsigned IIDs[] = { // Intrinsic to use for each size. |
9378 | Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0, |
9379 | Intrinsic::ppc_altivec_vrlw |
9380 | }; |
9381 | Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl); |
9382 | return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res); |
9383 | } |
9384 | |
9385 | // t = vsplti c, result = vsldoi t, t, 1 |
9386 | if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) { |
9387 | SDValue T = getCanonicalConstSplat(i, SplatSize, MVT::v16i8, DAG, dl); |
9388 | unsigned Amt = Subtarget.isLittleEndian() ? 15 : 1; |
9389 | return BuildVSLDOI(T, T, Amt, Op.getValueType(), DAG, dl); |
9390 | } |
9391 | // t = vsplti c, result = vsldoi t, t, 2 |
9392 | if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) { |
9393 | SDValue T = getCanonicalConstSplat(i, SplatSize, MVT::v16i8, DAG, dl); |
9394 | unsigned Amt = Subtarget.isLittleEndian() ? 14 : 2; |
9395 | return BuildVSLDOI(T, T, Amt, Op.getValueType(), DAG, dl); |
9396 | } |
9397 | // t = vsplti c, result = vsldoi t, t, 3 |
9398 | if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) { |
9399 | SDValue T = getCanonicalConstSplat(i, SplatSize, MVT::v16i8, DAG, dl); |
9400 | unsigned Amt = Subtarget.isLittleEndian() ? 13 : 3; |
9401 | return BuildVSLDOI(T, T, Amt, Op.getValueType(), DAG, dl); |
9402 | } |
9403 | } |
9404 | |
9405 | return SDValue(); |
9406 | } |
9407 | |
9408 | /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit |
9409 | /// the specified operations to build the shuffle. |
9410 | static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS, |
9411 | SDValue RHS, SelectionDAG &DAG, |
9412 | const SDLoc &dl) { |
9413 | unsigned OpNum = (PFEntry >> 26) & 0x0F; |
9414 | unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1); |
9415 | unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1); |
9416 | |
9417 | enum { |
9418 | OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3> |
9419 | OP_VMRGHW, |
9420 | OP_VMRGLW, |
9421 | OP_VSPLTISW0, |
9422 | OP_VSPLTISW1, |
9423 | OP_VSPLTISW2, |
9424 | OP_VSPLTISW3, |
9425 | OP_VSLDOI4, |
9426 | OP_VSLDOI8, |
9427 | OP_VSLDOI12 |
9428 | }; |
9429 | |
9430 | if (OpNum == OP_COPY) { |
9431 | if (LHSID == (1*9+2)*9+3) return LHS; |
9432 | assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!")((LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!") ? static_cast<void> (0) : __assert_fail ("LHSID == ((4*9+5)*9+6)*9+7 && \"Illegal OP_COPY!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9432, __PRETTY_FUNCTION__)); |
9433 | return RHS; |
9434 | } |
9435 | |
9436 | SDValue OpLHS, OpRHS; |
9437 | OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl); |
9438 | OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl); |
9439 | |
9440 | int ShufIdxs[16]; |
9441 | switch (OpNum) { |
9442 | default: llvm_unreachable("Unknown i32 permute!")::llvm::llvm_unreachable_internal("Unknown i32 permute!", "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9442); |
9443 | case OP_VMRGHW: |
9444 | ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3; |
9445 | ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19; |
9446 | ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7; |
9447 | ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23; |
9448 | break; |
9449 | case OP_VMRGLW: |
9450 | ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11; |
9451 | ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27; |
9452 | ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15; |
9453 | ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31; |
9454 | break; |
9455 | case OP_VSPLTISW0: |
9456 | for (unsigned i = 0; i != 16; ++i) |
9457 | ShufIdxs[i] = (i&3)+0; |
9458 | break; |
9459 | case OP_VSPLTISW1: |
9460 | for (unsigned i = 0; i != 16; ++i) |
9461 | ShufIdxs[i] = (i&3)+4; |
9462 | break; |
9463 | case OP_VSPLTISW2: |
9464 | for (unsigned i = 0; i != 16; ++i) |
9465 | ShufIdxs[i] = (i&3)+8; |
9466 | break; |
9467 | case OP_VSPLTISW3: |
9468 | for (unsigned i = 0; i != 16; ++i) |
9469 | ShufIdxs[i] = (i&3)+12; |
9470 | break; |
9471 | case OP_VSLDOI4: |
9472 | return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl); |
9473 | case OP_VSLDOI8: |
9474 | return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl); |
9475 | case OP_VSLDOI12: |
9476 | return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl); |
9477 | } |
9478 | EVT VT = OpLHS.getValueType(); |
9479 | OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS); |
9480 | OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS); |
9481 | SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs); |
9482 | return DAG.getNode(ISD::BITCAST, dl, VT, T); |
9483 | } |
9484 | |
9485 | /// lowerToVINSERTB - Return the SDValue if this VECTOR_SHUFFLE can be handled |
9486 | /// by the VINSERTB instruction introduced in ISA 3.0, else just return default |
9487 | /// SDValue. |
9488 | SDValue PPCTargetLowering::lowerToVINSERTB(ShuffleVectorSDNode *N, |
9489 | SelectionDAG &DAG) const { |
9490 | const unsigned BytesInVector = 16; |
9491 | bool IsLE = Subtarget.isLittleEndian(); |
9492 | SDLoc dl(N); |
9493 | SDValue V1 = N->getOperand(0); |
9494 | SDValue V2 = N->getOperand(1); |
9495 | unsigned ShiftElts = 0, InsertAtByte = 0; |
9496 | bool Swap = false; |
9497 | |
9498 | // Shifts required to get the byte we want at element 7. |
9499 | unsigned LittleEndianShifts[] = {8, 7, 6, 5, 4, 3, 2, 1, |
9500 | 0, 15, 14, 13, 12, 11, 10, 9}; |
9501 | unsigned BigEndianShifts[] = {9, 10, 11, 12, 13, 14, 15, 0, |
9502 | 1, 2, 3, 4, 5, 6, 7, 8}; |
9503 | |
9504 | ArrayRef<int> Mask = N->getMask(); |
9505 | int OriginalOrder[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15}; |
9506 | |
9507 | // For each mask element, find out if we're just inserting something |
9508 | // from V2 into V1 or vice versa. |
9509 | // Possible permutations inserting an element from V2 into V1: |
9510 | // X, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 |
9511 | // 0, X, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 |
9512 | // ... |
9513 | // 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, X |
9514 | // Inserting from V1 into V2 will be similar, except mask range will be |
9515 | // [16,31]. |
9516 | |
9517 | bool FoundCandidate = false; |
9518 | // If both vector operands for the shuffle are the same vector, the mask |
9519 | // will contain only elements from the first one and the second one will be |
9520 | // undef. |
9521 | unsigned VINSERTBSrcElem = IsLE ? 8 : 7; |
9522 | // Go through the mask of half-words to find an element that's being moved |
9523 | // from one vector to the other. |
9524 | for (unsigned i = 0; i < BytesInVector; ++i) { |
9525 | unsigned CurrentElement = Mask[i]; |
9526 | // If 2nd operand is undefined, we should only look for element 7 in the |
9527 | // Mask. |
9528 | if (V2.isUndef() && CurrentElement != VINSERTBSrcElem) |
9529 | continue; |
9530 | |
9531 | bool OtherElementsInOrder = true; |
9532 | // Examine the other elements in the Mask to see if they're in original |
9533 | // order. |
9534 | for (unsigned j = 0; j < BytesInVector; ++j) { |
9535 | if (j == i) |
9536 | continue; |
9537 | // If CurrentElement is from V1 [0,15], then we the rest of the Mask to be |
9538 | // from V2 [16,31] and vice versa. Unless the 2nd operand is undefined, |
9539 | // in which we always assume we're always picking from the 1st operand. |
9540 | int MaskOffset = |
9541 | (!V2.isUndef() && CurrentElement < BytesInVector) ? BytesInVector : 0; |
9542 | if (Mask[j] != OriginalOrder[j] + MaskOffset) { |
9543 | OtherElementsInOrder = false; |
9544 | break; |
9545 | } |
9546 | } |
9547 | // If other elements are in original order, we record the number of shifts |
9548 | // we need to get the element we want into element 7. Also record which byte |
9549 | // in the vector we should insert into. |
9550 | if (OtherElementsInOrder) { |
9551 | // If 2nd operand is undefined, we assume no shifts and no swapping. |
9552 | if (V2.isUndef()) { |
9553 | ShiftElts = 0; |
9554 | Swap = false; |
9555 | } else { |
9556 | // Only need the last 4-bits for shifts because operands will be swapped if CurrentElement is >= 2^4. |
9557 | ShiftElts = IsLE ? LittleEndianShifts[CurrentElement & 0xF] |
9558 | : BigEndianShifts[CurrentElement & 0xF]; |
9559 | Swap = CurrentElement < BytesInVector; |
9560 | } |
9561 | InsertAtByte = IsLE ? BytesInVector - (i + 1) : i; |
9562 | FoundCandidate = true; |
9563 | break; |
9564 | } |
9565 | } |
9566 | |
9567 | if (!FoundCandidate) |
9568 | return SDValue(); |
9569 | |
9570 | // Candidate found, construct the proper SDAG sequence with VINSERTB, |
9571 | // optionally with VECSHL if shift is required. |
9572 | if (Swap) |
9573 | std::swap(V1, V2); |
9574 | if (V2.isUndef()) |
9575 | V2 = V1; |
9576 | if (ShiftElts) { |
9577 | SDValue Shl = DAG.getNode(PPCISD::VECSHL, dl, MVT::v16i8, V2, V2, |
9578 | DAG.getConstant(ShiftElts, dl, MVT::i32)); |
9579 | return DAG.getNode(PPCISD::VECINSERT, dl, MVT::v16i8, V1, Shl, |
9580 | DAG.getConstant(InsertAtByte, dl, MVT::i32)); |
9581 | } |
9582 | return DAG.getNode(PPCISD::VECINSERT, dl, MVT::v16i8, V1, V2, |
9583 | DAG.getConstant(InsertAtByte, dl, MVT::i32)); |
9584 | } |
9585 | |
9586 | /// lowerToVINSERTH - Return the SDValue if this VECTOR_SHUFFLE can be handled |
9587 | /// by the VINSERTH instruction introduced in ISA 3.0, else just return default |
9588 | /// SDValue. |
9589 | SDValue PPCTargetLowering::lowerToVINSERTH(ShuffleVectorSDNode *N, |
9590 | SelectionDAG &DAG) const { |
9591 | const unsigned NumHalfWords = 8; |
9592 | const unsigned BytesInVector = NumHalfWords * 2; |
9593 | // Check that the shuffle is on half-words. |
9594 | if (!isNByteElemShuffleMask(N, 2, 1)) |
9595 | return SDValue(); |
9596 | |
9597 | bool IsLE = Subtarget.isLittleEndian(); |
9598 | SDLoc dl(N); |
9599 | SDValue V1 = N->getOperand(0); |
9600 | SDValue V2 = N->getOperand(1); |
9601 | unsigned ShiftElts = 0, InsertAtByte = 0; |
9602 | bool Swap = false; |
9603 | |
9604 | // Shifts required to get the half-word we want at element 3. |
9605 | unsigned LittleEndianShifts[] = {4, 3, 2, 1, 0, 7, 6, 5}; |
9606 | unsigned BigEndianShifts[] = {5, 6, 7, 0, 1, 2, 3, 4}; |
9607 | |
9608 | uint32_t Mask = 0; |
9609 | uint32_t OriginalOrderLow = 0x1234567; |
9610 | uint32_t OriginalOrderHigh = 0x89ABCDEF; |
9611 | // Now we look at mask elements 0,2,4,6,8,10,12,14. Pack the mask into a |
9612 | // 32-bit space, only need 4-bit nibbles per element. |
9613 | for (unsigned i = 0; i < NumHalfWords; ++i) { |
9614 | unsigned MaskShift = (NumHalfWords - 1 - i) * 4; |
9615 | Mask |= ((uint32_t)(N->getMaskElt(i * 2) / 2) << MaskShift); |
9616 | } |
9617 | |
9618 | // For each mask element, find out if we're just inserting something |
9619 | // from V2 into V1 or vice versa. Possible permutations inserting an element |
9620 | // from V2 into V1: |
9621 | // X, 1, 2, 3, 4, 5, 6, 7 |
9622 | // 0, X, 2, 3, 4, 5, 6, 7 |
9623 | // 0, 1, X, 3, 4, 5, 6, 7 |
9624 | // 0, 1, 2, X, 4, 5, 6, 7 |
9625 | // 0, 1, 2, 3, X, 5, 6, 7 |
9626 | // 0, 1, 2, 3, 4, X, 6, 7 |
9627 | // 0, 1, 2, 3, 4, 5, X, 7 |
9628 | // 0, 1, 2, 3, 4, 5, 6, X |
9629 | // Inserting from V1 into V2 will be similar, except mask range will be [8,15]. |
9630 | |
9631 | bool FoundCandidate = false; |
9632 | // Go through the mask of half-words to find an element that's being moved |
9633 | // from one vector to the other. |
9634 | for (unsigned i = 0; i < NumHalfWords; ++i) { |
9635 | unsigned MaskShift = (NumHalfWords - 1 - i) * 4; |
9636 | uint32_t MaskOneElt = (Mask >> MaskShift) & 0xF; |
9637 | uint32_t MaskOtherElts = ~(0xF << MaskShift); |
9638 | uint32_t TargetOrder = 0x0; |
9639 | |
9640 | // If both vector operands for the shuffle are the same vector, the mask |
9641 | // will contain only elements from the first one and the second one will be |
9642 | // undef. |
9643 | if (V2.isUndef()) { |
9644 | ShiftElts = 0; |
9645 | unsigned VINSERTHSrcElem = IsLE ? 4 : 3; |
9646 | TargetOrder = OriginalOrderLow; |
9647 | Swap = false; |
9648 | // Skip if not the correct element or mask of other elements don't equal |
9649 | // to our expected order. |
9650 | if (MaskOneElt == VINSERTHSrcElem && |
9651 | (Mask & MaskOtherElts) == (TargetOrder & MaskOtherElts)) { |
9652 | InsertAtByte = IsLE ? BytesInVector - (i + 1) * 2 : i * 2; |
9653 | FoundCandidate = true; |
9654 | break; |
9655 | } |
9656 | } else { // If both operands are defined. |
9657 | // Target order is [8,15] if the current mask is between [0,7]. |
9658 | TargetOrder = |
9659 | (MaskOneElt < NumHalfWords) ? OriginalOrderHigh : OriginalOrderLow; |
9660 | // Skip if mask of other elements don't equal our expected order. |
9661 | if ((Mask & MaskOtherElts) == (TargetOrder & MaskOtherElts)) { |
9662 | // We only need the last 3 bits for the number of shifts. |
9663 | ShiftElts = IsLE ? LittleEndianShifts[MaskOneElt & 0x7] |
9664 | : BigEndianShifts[MaskOneElt & 0x7]; |
9665 | InsertAtByte = IsLE ? BytesInVector - (i + 1) * 2 : i * 2; |
9666 | Swap = MaskOneElt < NumHalfWords; |
9667 | FoundCandidate = true; |
9668 | break; |
9669 | } |
9670 | } |
9671 | } |
9672 | |
9673 | if (!FoundCandidate) |
9674 | return SDValue(); |
9675 | |
9676 | // Candidate found, construct the proper SDAG sequence with VINSERTH, |
9677 | // optionally with VECSHL if shift is required. |
9678 | if (Swap) |
9679 | std::swap(V1, V2); |
9680 | if (V2.isUndef()) |
9681 | V2 = V1; |
9682 | SDValue Conv1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1); |
9683 | if (ShiftElts) { |
9684 | // Double ShiftElts because we're left shifting on v16i8 type. |
9685 | SDValue Shl = DAG.getNode(PPCISD::VECSHL, dl, MVT::v16i8, V2, V2, |
9686 | DAG.getConstant(2 * ShiftElts, dl, MVT::i32)); |
9687 | SDValue Conv2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, Shl); |
9688 | SDValue Ins = DAG.getNode(PPCISD::VECINSERT, dl, MVT::v8i16, Conv1, Conv2, |
9689 | DAG.getConstant(InsertAtByte, dl, MVT::i32)); |
9690 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Ins); |
9691 | } |
9692 | SDValue Conv2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2); |
9693 | SDValue Ins = DAG.getNode(PPCISD::VECINSERT, dl, MVT::v8i16, Conv1, Conv2, |
9694 | DAG.getConstant(InsertAtByte, dl, MVT::i32)); |
9695 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Ins); |
9696 | } |
9697 | |
9698 | /// lowerToXXSPLTI32DX - Return the SDValue if this VECTOR_SHUFFLE can be |
9699 | /// handled by the XXSPLTI32DX instruction introduced in ISA 3.1, otherwise |
9700 | /// return the default SDValue. |
9701 | SDValue PPCTargetLowering::lowerToXXSPLTI32DX(ShuffleVectorSDNode *SVN, |
9702 | SelectionDAG &DAG) const { |
9703 | // The LHS and RHS may be bitcasts to v16i8 as we canonicalize shuffles |
9704 | // to v16i8. Peek through the bitcasts to get the actual operands. |
9705 | SDValue LHS = peekThroughBitcasts(SVN->getOperand(0)); |
9706 | SDValue RHS = peekThroughBitcasts(SVN->getOperand(1)); |
9707 | |
9708 | auto ShuffleMask = SVN->getMask(); |
9709 | SDValue VecShuffle(SVN, 0); |
9710 | SDLoc DL(SVN); |
9711 | |
9712 | // Check that we have a four byte shuffle. |
9713 | if (!isNByteElemShuffleMask(SVN, 4, 1)) |
9714 | return SDValue(); |
9715 | |
9716 | // Canonicalize the RHS being a BUILD_VECTOR when lowering to xxsplti32dx. |
9717 | if (RHS->getOpcode() != ISD::BUILD_VECTOR) { |
9718 | std::swap(LHS, RHS); |
9719 | VecShuffle = DAG.getCommutedVectorShuffle(*SVN); |
9720 | ShuffleMask = cast<ShuffleVectorSDNode>(VecShuffle)->getMask(); |
9721 | } |
9722 | |
9723 | // Ensure that the RHS is a vector of constants. |
9724 | BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode()); |
9725 | if (!BVN) |
9726 | return SDValue(); |
9727 | |
9728 | // Check if RHS is a splat of 4-bytes (or smaller). |
9729 | APInt APSplatValue, APSplatUndef; |
9730 | unsigned SplatBitSize; |
9731 | bool HasAnyUndefs; |
9732 | if (!BVN->isConstantSplat(APSplatValue, APSplatUndef, SplatBitSize, |
9733 | HasAnyUndefs, 0, !Subtarget.isLittleEndian()) || |
9734 | SplatBitSize > 32) |
9735 | return SDValue(); |
9736 | |
9737 | // Check that the shuffle mask matches the semantics of XXSPLTI32DX. |
9738 | // The instruction splats a constant C into two words of the source vector |
9739 | // producing { C, Unchanged, C, Unchanged } or { Unchanged, C, Unchanged, C }. |
9740 | // Thus we check that the shuffle mask is the equivalent of |
9741 | // <0, [4-7], 2, [4-7]> or <[4-7], 1, [4-7], 3> respectively. |
9742 | // Note: the check above of isNByteElemShuffleMask() ensures that the bytes |
9743 | // within each word are consecutive, so we only need to check the first byte. |
9744 | SDValue Index; |
9745 | bool IsLE = Subtarget.isLittleEndian(); |
9746 | if ((ShuffleMask[0] == 0 && ShuffleMask[8] == 8) && |
9747 | (ShuffleMask[4] % 4 == 0 && ShuffleMask[12] % 4 == 0 && |
9748 | ShuffleMask[4] > 15 && ShuffleMask[12] > 15)) |
9749 | Index = DAG.getTargetConstant(IsLE ? 0 : 1, DL, MVT::i32); |
9750 | else if ((ShuffleMask[4] == 4 && ShuffleMask[12] == 12) && |
9751 | (ShuffleMask[0] % 4 == 0 && ShuffleMask[8] % 4 == 0 && |
9752 | ShuffleMask[0] > 15 && ShuffleMask[8] > 15)) |
9753 | Index = DAG.getTargetConstant(IsLE ? 1 : 0, DL, MVT::i32); |
9754 | else |
9755 | return SDValue(); |
9756 | |
9757 | // If the splat is narrower than 32-bits, we need to get the 32-bit value |
9758 | // for XXSPLTI32DX. |
9759 | unsigned SplatVal = APSplatValue.getZExtValue(); |
9760 | for (; SplatBitSize < 32; SplatBitSize <<= 1) |
9761 | SplatVal |= (SplatVal << SplatBitSize); |
9762 | |
9763 | SDValue SplatNode = DAG.getNode( |
9764 | PPCISD::XXSPLTI32DX, DL, MVT::v2i64, DAG.getBitcast(MVT::v2i64, LHS), |
9765 | Index, DAG.getTargetConstant(SplatVal, DL, MVT::i32)); |
9766 | return DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, SplatNode); |
9767 | } |
9768 | |
9769 | /// LowerROTL - Custom lowering for ROTL(v1i128) to vector_shuffle(v16i8). |
9770 | /// We lower ROTL(v1i128) to vector_shuffle(v16i8) only if shift amount is |
9771 | /// a multiple of 8. Otherwise convert it to a scalar rotation(i128) |
9772 | /// i.e (or (shl x, C1), (srl x, 128-C1)). |
9773 | SDValue PPCTargetLowering::LowerROTL(SDValue Op, SelectionDAG &DAG) const { |
9774 | assert(Op.getOpcode() == ISD::ROTL && "Should only be called for ISD::ROTL")((Op.getOpcode() == ISD::ROTL && "Should only be called for ISD::ROTL" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::ROTL && \"Should only be called for ISD::ROTL\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9774, __PRETTY_FUNCTION__)); |
9775 | assert(Op.getValueType() == MVT::v1i128 &&((Op.getValueType() == MVT::v1i128 && "Only set v1i128 as custom, other type shouldn't reach here!" ) ? static_cast<void> (0) : __assert_fail ("Op.getValueType() == MVT::v1i128 && \"Only set v1i128 as custom, other type shouldn't reach here!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9776, __PRETTY_FUNCTION__)) |
9776 | "Only set v1i128 as custom, other type shouldn't reach here!")((Op.getValueType() == MVT::v1i128 && "Only set v1i128 as custom, other type shouldn't reach here!" ) ? static_cast<void> (0) : __assert_fail ("Op.getValueType() == MVT::v1i128 && \"Only set v1i128 as custom, other type shouldn't reach here!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9776, __PRETTY_FUNCTION__)); |
9777 | SDLoc dl(Op); |
9778 | SDValue N0 = peekThroughBitcasts(Op.getOperand(0)); |
9779 | SDValue N1 = peekThroughBitcasts(Op.getOperand(1)); |
9780 | unsigned SHLAmt = N1.getConstantOperandVal(0); |
9781 | if (SHLAmt % 8 == 0) { |
9782 | SmallVector<int, 16> Mask(16, 0); |
9783 | std::iota(Mask.begin(), Mask.end(), 0); |
9784 | std::rotate(Mask.begin(), Mask.begin() + SHLAmt / 8, Mask.end()); |
9785 | if (SDValue Shuffle = |
9786 | DAG.getVectorShuffle(MVT::v16i8, dl, DAG.getBitcast(MVT::v16i8, N0), |
9787 | DAG.getUNDEF(MVT::v16i8), Mask)) |
9788 | return DAG.getNode(ISD::BITCAST, dl, MVT::v1i128, Shuffle); |
9789 | } |
9790 | SDValue ArgVal = DAG.getBitcast(MVT::i128, N0); |
9791 | SDValue SHLOp = DAG.getNode(ISD::SHL, dl, MVT::i128, ArgVal, |
9792 | DAG.getConstant(SHLAmt, dl, MVT::i32)); |
9793 | SDValue SRLOp = DAG.getNode(ISD::SRL, dl, MVT::i128, ArgVal, |
9794 | DAG.getConstant(128 - SHLAmt, dl, MVT::i32)); |
9795 | SDValue OROp = DAG.getNode(ISD::OR, dl, MVT::i128, SHLOp, SRLOp); |
9796 | return DAG.getNode(ISD::BITCAST, dl, MVT::v1i128, OROp); |
9797 | } |
9798 | |
9799 | /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this |
9800 | /// is a shuffle we can handle in a single instruction, return it. Otherwise, |
9801 | /// return the code it can be lowered into. Worst case, it can always be |
9802 | /// lowered into a vperm. |
9803 | SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, |
9804 | SelectionDAG &DAG) const { |
9805 | SDLoc dl(Op); |
9806 | SDValue V1 = Op.getOperand(0); |
9807 | SDValue V2 = Op.getOperand(1); |
9808 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); |
9809 | |
9810 | // Any nodes that were combined in the target-independent combiner prior |
9811 | // to vector legalization will not be sent to the target combine. Try to |
9812 | // combine it here. |
9813 | if (SDValue NewShuffle = combineVectorShuffle(SVOp, DAG)) { |
9814 | if (!isa<ShuffleVectorSDNode>(NewShuffle)) |
9815 | return NewShuffle; |
9816 | Op = NewShuffle; |
9817 | SVOp = cast<ShuffleVectorSDNode>(Op); |
9818 | V1 = Op.getOperand(0); |
9819 | V2 = Op.getOperand(1); |
9820 | } |
9821 | EVT VT = Op.getValueType(); |
9822 | bool isLittleEndian = Subtarget.isLittleEndian(); |
9823 | |
9824 | unsigned ShiftElts, InsertAtByte; |
9825 | bool Swap = false; |
9826 | |
9827 | // If this is a load-and-splat, we can do that with a single instruction |
9828 | // in some cases. However if the load has multiple uses, we don't want to |
9829 | // combine it because that will just produce multiple loads. |
9830 | bool IsPermutedLoad = false; |
9831 | const SDValue *InputLoad = getNormalLoadInput(V1, IsPermutedLoad); |
9832 | if (InputLoad && Subtarget.hasVSX() && V2.isUndef() && |
9833 | (PPC::isSplatShuffleMask(SVOp, 4) || PPC::isSplatShuffleMask(SVOp, 8)) && |
9834 | InputLoad->hasOneUse()) { |
9835 | bool IsFourByte = PPC::isSplatShuffleMask(SVOp, 4); |
9836 | int SplatIdx = |
9837 | PPC::getSplatIdxForPPCMnemonics(SVOp, IsFourByte ? 4 : 8, DAG); |
9838 | |
9839 | // The splat index for permuted loads will be in the left half of the vector |
9840 | // which is strictly wider than the loaded value by 8 bytes. So we need to |
9841 | // adjust the splat index to point to the correct address in memory. |
9842 | if (IsPermutedLoad) { |
9843 | assert(isLittleEndian && "Unexpected permuted load on big endian target")((isLittleEndian && "Unexpected permuted load on big endian target" ) ? static_cast<void> (0) : __assert_fail ("isLittleEndian && \"Unexpected permuted load on big endian target\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9843, __PRETTY_FUNCTION__)); |
9844 | SplatIdx += IsFourByte ? 2 : 1; |
9845 | assert((SplatIdx < (IsFourByte ? 4 : 2)) &&(((SplatIdx < (IsFourByte ? 4 : 2)) && "Splat of a value outside of the loaded memory" ) ? static_cast<void> (0) : __assert_fail ("(SplatIdx < (IsFourByte ? 4 : 2)) && \"Splat of a value outside of the loaded memory\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9846, __PRETTY_FUNCTION__)) |
9846 | "Splat of a value outside of the loaded memory")(((SplatIdx < (IsFourByte ? 4 : 2)) && "Splat of a value outside of the loaded memory" ) ? static_cast<void> (0) : __assert_fail ("(SplatIdx < (IsFourByte ? 4 : 2)) && \"Splat of a value outside of the loaded memory\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9846, __PRETTY_FUNCTION__)); |
9847 | } |
9848 | |
9849 | LoadSDNode *LD = cast<LoadSDNode>(*InputLoad); |
9850 | // For 4-byte load-and-splat, we need Power9. |
9851 | if ((IsFourByte && Subtarget.hasP9Vector()) || !IsFourByte) { |
9852 | uint64_t Offset = 0; |
9853 | if (IsFourByte) |
9854 | Offset = isLittleEndian ? (3 - SplatIdx) * 4 : SplatIdx * 4; |
9855 | else |
9856 | Offset = isLittleEndian ? (1 - SplatIdx) * 8 : SplatIdx * 8; |
9857 | |
9858 | SDValue BasePtr = LD->getBasePtr(); |
9859 | if (Offset != 0) |
9860 | BasePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()), |
9861 | BasePtr, DAG.getIntPtrConstant(Offset, dl)); |
9862 | SDValue Ops[] = { |
9863 | LD->getChain(), // Chain |
9864 | BasePtr, // BasePtr |
9865 | DAG.getValueType(Op.getValueType()) // VT |
9866 | }; |
9867 | SDVTList VTL = |
9868 | DAG.getVTList(IsFourByte ? MVT::v4i32 : MVT::v2i64, MVT::Other); |
9869 | SDValue LdSplt = |
9870 | DAG.getMemIntrinsicNode(PPCISD::LD_SPLAT, dl, VTL, |
9871 | Ops, LD->getMemoryVT(), LD->getMemOperand()); |
9872 | DAG.ReplaceAllUsesOfValueWith(InputLoad->getValue(1), LdSplt.getValue(1)); |
9873 | if (LdSplt.getValueType() != SVOp->getValueType(0)) |
9874 | LdSplt = DAG.getBitcast(SVOp->getValueType(0), LdSplt); |
9875 | return LdSplt; |
9876 | } |
9877 | } |
9878 | if (Subtarget.hasP9Vector() && |
9879 | PPC::isXXINSERTWMask(SVOp, ShiftElts, InsertAtByte, Swap, |
9880 | isLittleEndian)) { |
9881 | if (Swap) |
9882 | std::swap(V1, V2); |
9883 | SDValue Conv1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1); |
9884 | SDValue Conv2 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V2); |
9885 | if (ShiftElts) { |
9886 | SDValue Shl = DAG.getNode(PPCISD::VECSHL, dl, MVT::v4i32, Conv2, Conv2, |
9887 | DAG.getConstant(ShiftElts, dl, MVT::i32)); |
9888 | SDValue Ins = DAG.getNode(PPCISD::VECINSERT, dl, MVT::v4i32, Conv1, Shl, |
9889 | DAG.getConstant(InsertAtByte, dl, MVT::i32)); |
9890 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Ins); |
9891 | } |
9892 | SDValue Ins = DAG.getNode(PPCISD::VECINSERT, dl, MVT::v4i32, Conv1, Conv2, |
9893 | DAG.getConstant(InsertAtByte, dl, MVT::i32)); |
9894 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Ins); |
9895 | } |
9896 | |
9897 | if (Subtarget.hasPrefixInstrs()) { |
9898 | SDValue SplatInsertNode; |
9899 | if ((SplatInsertNode = lowerToXXSPLTI32DX(SVOp, DAG))) |
9900 | return SplatInsertNode; |
9901 | } |
9902 | |
9903 | if (Subtarget.hasP9Altivec()) { |
9904 | SDValue NewISDNode; |
9905 | if ((NewISDNode = lowerToVINSERTH(SVOp, DAG))) |
9906 | return NewISDNode; |
9907 | |
9908 | if ((NewISDNode = lowerToVINSERTB(SVOp, DAG))) |
9909 | return NewISDNode; |
9910 | } |
9911 | |
9912 | if (Subtarget.hasVSX() && |
9913 | PPC::isXXSLDWIShuffleMask(SVOp, ShiftElts, Swap, isLittleEndian)) { |
9914 | if (Swap) |
9915 | std::swap(V1, V2); |
9916 | SDValue Conv1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1); |
9917 | SDValue Conv2 = |
9918 | DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V2.isUndef() ? V1 : V2); |
9919 | |
9920 | SDValue Shl = DAG.getNode(PPCISD::VECSHL, dl, MVT::v4i32, Conv1, Conv2, |
9921 | DAG.getConstant(ShiftElts, dl, MVT::i32)); |
9922 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Shl); |
9923 | } |
9924 | |
9925 | if (Subtarget.hasVSX() && |
9926 | PPC::isXXPERMDIShuffleMask(SVOp, ShiftElts, Swap, isLittleEndian)) { |
9927 | if (Swap) |
9928 | std::swap(V1, V2); |
9929 | SDValue Conv1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1); |
9930 | SDValue Conv2 = |
9931 | DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2.isUndef() ? V1 : V2); |
9932 | |
9933 | SDValue PermDI = DAG.getNode(PPCISD::XXPERMDI, dl, MVT::v2i64, Conv1, Conv2, |
9934 | DAG.getConstant(ShiftElts, dl, MVT::i32)); |
9935 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, PermDI); |
9936 | } |
9937 | |
9938 | if (Subtarget.hasP9Vector()) { |
9939 | if (PPC::isXXBRHShuffleMask(SVOp)) { |
9940 | SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1); |
9941 | SDValue ReveHWord = DAG.getNode(ISD::BSWAP, dl, MVT::v8i16, Conv); |
9942 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, ReveHWord); |
9943 | } else if (PPC::isXXBRWShuffleMask(SVOp)) { |
9944 | SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1); |
9945 | SDValue ReveWord = DAG.getNode(ISD::BSWAP, dl, MVT::v4i32, Conv); |
9946 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, ReveWord); |
9947 | } else if (PPC::isXXBRDShuffleMask(SVOp)) { |
9948 | SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1); |
9949 | SDValue ReveDWord = DAG.getNode(ISD::BSWAP, dl, MVT::v2i64, Conv); |
9950 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, ReveDWord); |
9951 | } else if (PPC::isXXBRQShuffleMask(SVOp)) { |
9952 | SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v1i128, V1); |
9953 | SDValue ReveQWord = DAG.getNode(ISD::BSWAP, dl, MVT::v1i128, Conv); |
9954 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, ReveQWord); |
9955 | } |
9956 | } |
9957 | |
9958 | if (Subtarget.hasVSX()) { |
9959 | if (V2.isUndef() && PPC::isSplatShuffleMask(SVOp, 4)) { |
9960 | int SplatIdx = PPC::getSplatIdxForPPCMnemonics(SVOp, 4, DAG); |
9961 | |
9962 | SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1); |
9963 | SDValue Splat = DAG.getNode(PPCISD::XXSPLT, dl, MVT::v4i32, Conv, |
9964 | DAG.getConstant(SplatIdx, dl, MVT::i32)); |
9965 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Splat); |
9966 | } |
9967 | |
9968 | // Left shifts of 8 bytes are actually swaps. Convert accordingly. |
9969 | if (V2.isUndef() && PPC::isVSLDOIShuffleMask(SVOp, 1, DAG) == 8) { |
9970 | SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1); |
9971 | SDValue Swap = DAG.getNode(PPCISD::SWAP_NO_CHAIN, dl, MVT::v2f64, Conv); |
9972 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Swap); |
9973 | } |
9974 | } |
9975 | |
9976 | // Cases that are handled by instructions that take permute immediates |
9977 | // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be |
9978 | // selected by the instruction selector. |
9979 | if (V2.isUndef()) { |
9980 | if (PPC::isSplatShuffleMask(SVOp, 1) || |
9981 | PPC::isSplatShuffleMask(SVOp, 2) || |
9982 | PPC::isSplatShuffleMask(SVOp, 4) || |
9983 | PPC::isVPKUWUMShuffleMask(SVOp, 1, DAG) || |
9984 | PPC::isVPKUHUMShuffleMask(SVOp, 1, DAG) || |
9985 | PPC::isVSLDOIShuffleMask(SVOp, 1, DAG) != -1 || |
9986 | PPC::isVMRGLShuffleMask(SVOp, 1, 1, DAG) || |
9987 | PPC::isVMRGLShuffleMask(SVOp, 2, 1, DAG) || |
9988 | PPC::isVMRGLShuffleMask(SVOp, 4, 1, DAG) || |
9989 | PPC::isVMRGHShuffleMask(SVOp, 1, 1, DAG) || |
9990 | PPC::isVMRGHShuffleMask(SVOp, 2, 1, DAG) || |
9991 | PPC::isVMRGHShuffleMask(SVOp, 4, 1, DAG) || |
9992 | (Subtarget.hasP8Altivec() && ( |
9993 | PPC::isVPKUDUMShuffleMask(SVOp, 1, DAG) || |
9994 | PPC::isVMRGEOShuffleMask(SVOp, true, 1, DAG) || |
9995 | PPC::isVMRGEOShuffleMask(SVOp, false, 1, DAG)))) { |
9996 | return Op; |
9997 | } |
9998 | } |
9999 | |
10000 | // Altivec has a variety of "shuffle immediates" that take two vector inputs |
10001 | // and produce a fixed permutation. If any of these match, do not lower to |
10002 | // VPERM. |
10003 | unsigned int ShuffleKind = isLittleEndian ? 2 : 0; |
10004 | if (PPC::isVPKUWUMShuffleMask(SVOp, ShuffleKind, DAG) || |
10005 | PPC::isVPKUHUMShuffleMask(SVOp, ShuffleKind, DAG) || |
10006 | PPC::isVSLDOIShuffleMask(SVOp, ShuffleKind, DAG) != -1 || |
10007 | PPC::isVMRGLShuffleMask(SVOp, 1, ShuffleKind, DAG) || |
10008 | PPC::isVMRGLShuffleMask(SVOp, 2, ShuffleKind, DAG) || |
10009 | PPC::isVMRGLShuffleMask(SVOp, 4, ShuffleKind, DAG) || |
10010 | PPC::isVMRGHShuffleMask(SVOp, 1, ShuffleKind, DAG) || |
10011 | PPC::isVMRGHShuffleMask(SVOp, 2, ShuffleKind, DAG) || |
10012 | PPC::isVMRGHShuffleMask(SVOp, 4, ShuffleKind, DAG) || |
10013 | (Subtarget.hasP8Altivec() && ( |
10014 | PPC::isVPKUDUMShuffleMask(SVOp, ShuffleKind, DAG) || |
10015 | PPC::isVMRGEOShuffleMask(SVOp, true, ShuffleKind, DAG) || |
10016 | PPC::isVMRGEOShuffleMask(SVOp, false, ShuffleKind, DAG)))) |
10017 | return Op; |
10018 | |
10019 | // Check to see if this is a shuffle of 4-byte values. If so, we can use our |
10020 | // perfect shuffle table to emit an optimal matching sequence. |
10021 | ArrayRef<int> PermMask = SVOp->getMask(); |
10022 | |
10023 | unsigned PFIndexes[4]; |
10024 | bool isFourElementShuffle = true; |
10025 | for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number |
10026 | unsigned EltNo = 8; // Start out undef. |
10027 | for (unsigned j = 0; j != 4; ++j) { // Intra-element byte. |
10028 | if (PermMask[i*4+j] < 0) |
10029 | continue; // Undef, ignore it. |
10030 | |
10031 | unsigned ByteSource = PermMask[i*4+j]; |
10032 | if ((ByteSource & 3) != j) { |
10033 | isFourElementShuffle = false; |
10034 | break; |
10035 | } |
10036 | |
10037 | if (EltNo == 8) { |
10038 | EltNo = ByteSource/4; |
10039 | } else if (EltNo != ByteSource/4) { |
10040 | isFourElementShuffle = false; |
10041 | break; |
10042 | } |
10043 | } |
10044 | PFIndexes[i] = EltNo; |
10045 | } |
10046 | |
10047 | // If this shuffle can be expressed as a shuffle of 4-byte elements, use the |
10048 | // perfect shuffle vector to determine if it is cost effective to do this as |
10049 | // discrete instructions, or whether we should use a vperm. |
10050 | // For now, we skip this for little endian until such time as we have a |
10051 | // little-endian perfect shuffle table. |
10052 | if (isFourElementShuffle && !isLittleEndian) { |
10053 | // Compute the index in the perfect shuffle table. |
10054 | unsigned PFTableIndex = |
10055 | PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3]; |
10056 | |
10057 | unsigned PFEntry = PerfectShuffleTable[PFTableIndex]; |
10058 | unsigned Cost = (PFEntry >> 30); |
10059 | |
10060 | // Determining when to avoid vperm is tricky. Many things affect the cost |
10061 | // of vperm, particularly how many times the perm mask needs to be computed. |
10062 | // For example, if the perm mask can be hoisted out of a loop or is already |
10063 | // used (perhaps because there are multiple permutes with the same shuffle |
10064 | // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of |
10065 | // the loop requires an extra register. |
10066 | // |
10067 | // As a compromise, we only emit discrete instructions if the shuffle can be |
10068 | // generated in 3 or fewer operations. When we have loop information |
10069 | // available, if this block is within a loop, we should avoid using vperm |
10070 | // for 3-operation perms and use a constant pool load instead. |
10071 | if (Cost < 3) |
10072 | return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl); |
10073 | } |
10074 | |
10075 | // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant |
10076 | // vector that will get spilled to the constant pool. |
10077 | if (V2.isUndef()) V2 = V1; |
10078 | |
10079 | // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except |
10080 | // that it is in input element units, not in bytes. Convert now. |
10081 | |
10082 | // For little endian, the order of the input vectors is reversed, and |
10083 | // the permutation mask is complemented with respect to 31. This is |
10084 | // necessary to produce proper semantics with the big-endian-biased vperm |
10085 | // instruction. |
10086 | EVT EltVT = V1.getValueType().getVectorElementType(); |
10087 | unsigned BytesPerElement = EltVT.getSizeInBits()/8; |
10088 | |
10089 | SmallVector<SDValue, 16> ResultMask; |
10090 | for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) { |
10091 | unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i]; |
10092 | |
10093 | for (unsigned j = 0; j != BytesPerElement; ++j) |
10094 | if (isLittleEndian) |
10095 | ResultMask.push_back(DAG.getConstant(31 - (SrcElt*BytesPerElement + j), |
10096 | dl, MVT::i32)); |
10097 | else |
10098 | ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement + j, dl, |
10099 | MVT::i32)); |
10100 | } |
10101 | |
10102 | ShufflesHandledWithVPERM++; |
10103 | SDValue VPermMask = DAG.getBuildVector(MVT::v16i8, dl, ResultMask); |
10104 | LLVM_DEBUG(dbgs() << "Emitting a VPERM for the following shuffle:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-lowering")) { dbgs() << "Emitting a VPERM for the following shuffle:\n" ; } } while (false); |
10105 | LLVM_DEBUG(SVOp->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-lowering")) { SVOp->dump(); } } while (false); |
10106 | LLVM_DEBUG(dbgs() << "With the following permute control vector:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-lowering")) { dbgs() << "With the following permute control vector:\n" ; } } while (false); |
10107 | LLVM_DEBUG(VPermMask.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-lowering")) { VPermMask.dump(); } } while (false); |
10108 | |
10109 | if (isLittleEndian) |
10110 | return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), |
10111 | V2, V1, VPermMask); |
10112 | else |
10113 | return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), |
10114 | V1, V2, VPermMask); |
10115 | } |
10116 | |
10117 | /// getVectorCompareInfo - Given an intrinsic, return false if it is not a |
10118 | /// vector comparison. If it is, return true and fill in Opc/isDot with |
10119 | /// information about the intrinsic. |
10120 | static bool getVectorCompareInfo(SDValue Intrin, int &CompareOpc, |
10121 | bool &isDot, const PPCSubtarget &Subtarget) { |
10122 | unsigned IntrinsicID = |
10123 | cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue(); |
10124 | CompareOpc = -1; |
10125 | isDot = false; |
10126 | switch (IntrinsicID) { |
10127 | default: |
10128 | return false; |
10129 | // Comparison predicates. |
10130 | case Intrinsic::ppc_altivec_vcmpbfp_p: |
10131 | CompareOpc = 966; |
10132 | isDot = true; |
10133 | break; |
10134 | case Intrinsic::ppc_altivec_vcmpeqfp_p: |
10135 | CompareOpc = 198; |
10136 | isDot = true; |
10137 | break; |
10138 | case Intrinsic::ppc_altivec_vcmpequb_p: |
10139 | CompareOpc = 6; |
10140 | isDot = true; |
10141 | break; |
10142 | case Intrinsic::ppc_altivec_vcmpequh_p: |
10143 | CompareOpc = 70; |
10144 | isDot = true; |
10145 | break; |
10146 | case Intrinsic::ppc_altivec_vcmpequw_p: |
10147 | CompareOpc = 134; |
10148 | isDot = true; |
10149 | break; |
10150 | case Intrinsic::ppc_altivec_vcmpequd_p: |
10151 | if (Subtarget.hasP8Altivec()) { |
10152 | CompareOpc = 199; |
10153 | isDot = true; |
10154 | } else |
10155 | return false; |
10156 | break; |
10157 | case Intrinsic::ppc_altivec_vcmpneb_p: |
10158 | case Intrinsic::ppc_altivec_vcmpneh_p: |
10159 | case Intrinsic::ppc_altivec_vcmpnew_p: |
10160 | case Intrinsic::ppc_altivec_vcmpnezb_p: |
10161 | case Intrinsic::ppc_altivec_vcmpnezh_p: |
10162 | case Intrinsic::ppc_altivec_vcmpnezw_p: |
10163 | if (Subtarget.hasP9Altivec()) { |
10164 | switch (IntrinsicID) { |
10165 | default: |
10166 | llvm_unreachable("Unknown comparison intrinsic.")::llvm::llvm_unreachable_internal("Unknown comparison intrinsic." , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10166); |
10167 | case Intrinsic::ppc_altivec_vcmpneb_p: |
10168 | CompareOpc = 7; |
10169 | break; |
10170 | case Intrinsic::ppc_altivec_vcmpneh_p: |
10171 | CompareOpc = 71; |
10172 | break; |
10173 | case Intrinsic::ppc_altivec_vcmpnew_p: |
10174 | CompareOpc = 135; |
10175 | break; |
10176 | case Intrinsic::ppc_altivec_vcmpnezb_p: |
10177 | CompareOpc = 263; |
10178 | break; |
10179 | case Intrinsic::ppc_altivec_vcmpnezh_p: |
10180 | CompareOpc = 327; |
10181 | break; |
10182 | case Intrinsic::ppc_altivec_vcmpnezw_p: |
10183 | CompareOpc = 391; |
10184 | break; |
10185 | } |
10186 | isDot = true; |
10187 | } else |
10188 | return false; |
10189 | break; |
10190 | case Intrinsic::ppc_altivec_vcmpgefp_p: |
10191 | CompareOpc = 454; |
10192 | isDot = true; |
10193 | break; |
10194 | case Intrinsic::ppc_altivec_vcmpgtfp_p: |
10195 | CompareOpc = 710; |
10196 | isDot = true; |
10197 | break; |
10198 | case Intrinsic::ppc_altivec_vcmpgtsb_p: |
10199 | CompareOpc = 774; |
10200 | isDot = true; |
10201 | break; |
10202 | case Intrinsic::ppc_altivec_vcmpgtsh_p: |
10203 | CompareOpc = 838; |
10204 | isDot = true; |
10205 | break; |
10206 | case Intrinsic::ppc_altivec_vcmpgtsw_p: |
10207 | CompareOpc = 902; |
10208 | isDot = true; |
10209 | break; |
10210 | case Intrinsic::ppc_altivec_vcmpgtsd_p: |
10211 | if (Subtarget.hasP8Altivec()) { |
10212 | CompareOpc = 967; |
10213 | isDot = true; |
10214 | } else |
10215 | return false; |
10216 | break; |
10217 | case Intrinsic::ppc_altivec_vcmpgtub_p: |
10218 | CompareOpc = 518; |
10219 | isDot = true; |
10220 | break; |
10221 | case Intrinsic::ppc_altivec_vcmpgtuh_p: |
10222 | CompareOpc = 582; |
10223 | isDot = true; |
10224 | break; |
10225 | case Intrinsic::ppc_altivec_vcmpgtuw_p: |
10226 | CompareOpc = 646; |
10227 | isDot = true; |
10228 | break; |
10229 | case Intrinsic::ppc_altivec_vcmpgtud_p: |
10230 | if (Subtarget.hasP8Altivec()) { |
10231 | CompareOpc = 711; |
10232 | isDot = true; |
10233 | } else |
10234 | return false; |
10235 | break; |
10236 | |
10237 | case Intrinsic::ppc_altivec_vcmpequq: |
10238 | case Intrinsic::ppc_altivec_vcmpgtsq: |
10239 | case Intrinsic::ppc_altivec_vcmpgtuq: |
10240 | if (!Subtarget.isISA3_1()) |
10241 | return false; |
10242 | switch (IntrinsicID) { |
10243 | default: |
10244 | llvm_unreachable("Unknown comparison intrinsic.")::llvm::llvm_unreachable_internal("Unknown comparison intrinsic." , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10244); |
10245 | case Intrinsic::ppc_altivec_vcmpequq: |
10246 | CompareOpc = 455; |
10247 | break; |
10248 | case Intrinsic::ppc_altivec_vcmpgtsq: |
10249 | CompareOpc = 903; |
10250 | break; |
10251 | case Intrinsic::ppc_altivec_vcmpgtuq: |
10252 | CompareOpc = 647; |
10253 | break; |
10254 | } |
10255 | break; |
10256 | |
10257 | // VSX predicate comparisons use the same infrastructure |
10258 | case Intrinsic::ppc_vsx_xvcmpeqdp_p: |
10259 | case Intrinsic::ppc_vsx_xvcmpgedp_p: |
10260 | case Intrinsic::ppc_vsx_xvcmpgtdp_p: |
10261 | case Intrinsic::ppc_vsx_xvcmpeqsp_p: |
10262 | case Intrinsic::ppc_vsx_xvcmpgesp_p: |
10263 | case Intrinsic::ppc_vsx_xvcmpgtsp_p: |
10264 | if (Subtarget.hasVSX()) { |
10265 | switch (IntrinsicID) { |
10266 | case Intrinsic::ppc_vsx_xvcmpeqdp_p: |
10267 | CompareOpc = 99; |
10268 | break; |
10269 | case Intrinsic::ppc_vsx_xvcmpgedp_p: |
10270 | CompareOpc = 115; |
10271 | break; |
10272 | case Intrinsic::ppc_vsx_xvcmpgtdp_p: |
10273 | CompareOpc = 107; |
10274 | break; |
10275 | case Intrinsic::ppc_vsx_xvcmpeqsp_p: |
10276 | CompareOpc = 67; |
10277 | break; |
10278 | case Intrinsic::ppc_vsx_xvcmpgesp_p: |
10279 | CompareOpc = 83; |
10280 | break; |
10281 | case Intrinsic::ppc_vsx_xvcmpgtsp_p: |
10282 | CompareOpc = 75; |
10283 | break; |
10284 | } |
10285 | isDot = true; |
10286 | } else |
10287 | return false; |
10288 | break; |
10289 | |
10290 | // Normal Comparisons. |
10291 | case Intrinsic::ppc_altivec_vcmpbfp: |
10292 | CompareOpc = 966; |
10293 | break; |
10294 | case Intrinsic::ppc_altivec_vcmpeqfp: |
10295 | CompareOpc = 198; |
10296 | break; |
10297 | case Intrinsic::ppc_altivec_vcmpequb: |
10298 | CompareOpc = 6; |
10299 | break; |
10300 | case Intrinsic::ppc_altivec_vcmpequh: |
10301 | CompareOpc = 70; |
10302 | break; |
10303 | case Intrinsic::ppc_altivec_vcmpequw: |
10304 | CompareOpc = 134; |
10305 | break; |
10306 | case Intrinsic::ppc_altivec_vcmpequd: |
10307 | if (Subtarget.hasP8Altivec()) |
10308 | CompareOpc = 199; |
10309 | else |
10310 | return false; |
10311 | break; |
10312 | case Intrinsic::ppc_altivec_vcmpneb: |
10313 | case Intrinsic::ppc_altivec_vcmpneh: |
10314 | case Intrinsic::ppc_altivec_vcmpnew: |
10315 | case Intrinsic::ppc_altivec_vcmpnezb: |
10316 | case Intrinsic::ppc_altivec_vcmpnezh: |
10317 | case Intrinsic::ppc_altivec_vcmpnezw: |
10318 | if (Subtarget.hasP9Altivec()) |
10319 | switch (IntrinsicID) { |
10320 | default: |
10321 | llvm_unreachable("Unknown comparison intrinsic.")::llvm::llvm_unreachable_internal("Unknown comparison intrinsic." , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10321); |
10322 | case Intrinsic::ppc_altivec_vcmpneb: |
10323 | CompareOpc = 7; |
10324 | break; |
10325 | case Intrinsic::ppc_altivec_vcmpneh: |
10326 | CompareOpc = 71; |
10327 | break; |
10328 | case Intrinsic::ppc_altivec_vcmpnew: |
10329 | CompareOpc = 135; |
10330 | break; |
10331 | case Intrinsic::ppc_altivec_vcmpnezb: |
10332 | CompareOpc = 263; |
10333 | break; |
10334 | case Intrinsic::ppc_altivec_vcmpnezh: |
10335 | CompareOpc = 327; |
10336 | break; |
10337 | case Intrinsic::ppc_altivec_vcmpnezw: |
10338 | CompareOpc = 391; |
10339 | break; |
10340 | } |
10341 | else |
10342 | return false; |
10343 | break; |
10344 | case Intrinsic::ppc_altivec_vcmpgefp: |
10345 | CompareOpc = 454; |
10346 | break; |
10347 | case Intrinsic::ppc_altivec_vcmpgtfp: |
10348 | CompareOpc = 710; |
10349 | break; |
10350 | case Intrinsic::ppc_altivec_vcmpgtsb: |
10351 | CompareOpc = 774; |
10352 | break; |
10353 | case Intrinsic::ppc_altivec_vcmpgtsh: |
10354 | CompareOpc = 838; |
10355 | break; |
10356 | case Intrinsic::ppc_altivec_vcmpgtsw: |
10357 | CompareOpc = 902; |
10358 | break; |
10359 | case Intrinsic::ppc_altivec_vcmpgtsd: |
10360 | if (Subtarget.hasP8Altivec()) |
10361 | CompareOpc = 967; |
10362 | else |
10363 | return false; |
10364 | break; |
10365 | case Intrinsic::ppc_altivec_vcmpgtub: |
10366 | CompareOpc = 518; |
10367 | break; |
10368 | case Intrinsic::ppc_altivec_vcmpgtuh: |
10369 | CompareOpc = 582; |
10370 | break; |
10371 | case Intrinsic::ppc_altivec_vcmpgtuw: |
10372 | CompareOpc = 646; |
10373 | break; |
10374 | case Intrinsic::ppc_altivec_vcmpgtud: |
10375 | if (Subtarget.hasP8Altivec()) |
10376 | CompareOpc = 711; |
10377 | else |
10378 | return false; |
10379 | break; |
10380 | case Intrinsic::ppc_altivec_vcmpequq_p: |
10381 | case Intrinsic::ppc_altivec_vcmpgtsq_p: |
10382 | case Intrinsic::ppc_altivec_vcmpgtuq_p: |
10383 | if (!Subtarget.isISA3_1()) |
10384 | return false; |
10385 | switch (IntrinsicID) { |
10386 | default: |
10387 | llvm_unreachable("Unknown comparison intrinsic.")::llvm::llvm_unreachable_internal("Unknown comparison intrinsic." , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10387); |
10388 | case Intrinsic::ppc_altivec_vcmpequq_p: |
10389 | CompareOpc = 455; |
10390 | break; |
10391 | case Intrinsic::ppc_altivec_vcmpgtsq_p: |
10392 | CompareOpc = 903; |
10393 | break; |
10394 | case Intrinsic::ppc_altivec_vcmpgtuq_p: |
10395 | CompareOpc = 647; |
10396 | break; |
10397 | } |
10398 | isDot = true; |
10399 | break; |
10400 | } |
10401 | return true; |
10402 | } |
10403 | |
10404 | /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom |
10405 | /// lower, do it, otherwise return null. |
10406 | SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, |
10407 | SelectionDAG &DAG) const { |
10408 | unsigned IntrinsicID = |
10409 | cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); |
10410 | |
10411 | SDLoc dl(Op); |
10412 | |
10413 | switch (IntrinsicID) { |
10414 | case Intrinsic::thread_pointer: |
10415 | // Reads the thread pointer register, used for __builtin_thread_pointer. |
10416 | if (Subtarget.isPPC64()) |
10417 | return DAG.getRegister(PPC::X13, MVT::i64); |
10418 | return DAG.getRegister(PPC::R2, MVT::i32); |
10419 | |
10420 | case Intrinsic::ppc_mma_disassemble_acc: |
10421 | case Intrinsic::ppc_mma_disassemble_pair: { |
10422 | int NumVecs = 2; |
10423 | SDValue WideVec = Op.getOperand(1); |
10424 | if (IntrinsicID == Intrinsic::ppc_mma_disassemble_acc) { |
10425 | NumVecs = 4; |
10426 | WideVec = DAG.getNode(PPCISD::XXMFACC, dl, MVT::v512i1, WideVec); |
10427 | } |
10428 | SmallVector<SDValue, 4> RetOps; |
10429 | for (int VecNo = 0; VecNo < NumVecs; VecNo++) { |
10430 | SDValue Extract = DAG.getNode( |
10431 | PPCISD::EXTRACT_VSX_REG, dl, MVT::v16i8, WideVec, |
10432 | DAG.getConstant(Subtarget.isLittleEndian() ? NumVecs - 1 - VecNo |
10433 | : VecNo, |
10434 | dl, MVT::i64)); |
10435 | RetOps.push_back(Extract); |
10436 | } |
10437 | return DAG.getMergeValues(RetOps, dl); |
10438 | } |
10439 | } |
10440 | |
10441 | // If this is a lowered altivec predicate compare, CompareOpc is set to the |
10442 | // opcode number of the comparison. |
10443 | int CompareOpc; |
10444 | bool isDot; |
10445 | if (!getVectorCompareInfo(Op, CompareOpc, isDot, Subtarget)) |
10446 | return SDValue(); // Don't custom lower most intrinsics. |
10447 | |
10448 | // If this is a non-dot comparison, make the VCMP node and we are done. |
10449 | if (!isDot) { |
10450 | SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(), |
10451 | Op.getOperand(1), Op.getOperand(2), |
10452 | DAG.getConstant(CompareOpc, dl, MVT::i32)); |
10453 | return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp); |
10454 | } |
10455 | |
10456 | // Create the PPCISD altivec 'dot' comparison node. |
10457 | SDValue Ops[] = { |
10458 | Op.getOperand(2), // LHS |
10459 | Op.getOperand(3), // RHS |
10460 | DAG.getConstant(CompareOpc, dl, MVT::i32) |
10461 | }; |
10462 | EVT VTs[] = { Op.getOperand(2).getValueType(), MVT::Glue }; |
10463 | SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops); |
10464 | |
10465 | // Now that we have the comparison, emit a copy from the CR to a GPR. |
10466 | // This is flagged to the above dot comparison. |
10467 | SDValue Flags = DAG.getNode(PPCISD::MFOCRF, dl, MVT::i32, |
10468 | DAG.getRegister(PPC::CR6, MVT::i32), |
10469 | CompNode.getValue(1)); |
10470 | |
10471 | // Unpack the result based on how the target uses it. |
10472 | unsigned BitNo; // Bit # of CR6. |
10473 | bool InvertBit; // Invert result? |
10474 | switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) { |
10475 | default: // Can't happen, don't crash on invalid number though. |
10476 | case 0: // Return the value of the EQ bit of CR6. |
10477 | BitNo = 0; InvertBit = false; |
10478 | break; |
10479 | case 1: // Return the inverted value of the EQ bit of CR6. |
10480 | BitNo = 0; InvertBit = true; |
10481 | break; |
10482 | case 2: // Return the value of the LT bit of CR6. |
10483 | BitNo = 2; InvertBit = false; |
10484 | break; |
10485 | case 3: // Return the inverted value of the LT bit of CR6. |
10486 | BitNo = 2; InvertBit = true; |
10487 | break; |
10488 | } |
10489 | |
10490 | // Shift the bit into the low position. |
10491 | Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags, |
10492 | DAG.getConstant(8 - (3 - BitNo), dl, MVT::i32)); |
10493 | // Isolate the bit. |
10494 | Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags, |
10495 | DAG.getConstant(1, dl, MVT::i32)); |
10496 | |
10497 | // If we are supposed to, toggle the bit. |
10498 | if (InvertBit) |
10499 | Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags, |
10500 | DAG.getConstant(1, dl, MVT::i32)); |
10501 | return Flags; |
10502 | } |
10503 | |
10504 | SDValue PPCTargetLowering::LowerINTRINSIC_VOID(SDValue Op, |
10505 | SelectionDAG &DAG) const { |
10506 | // SelectionDAGBuilder::visitTargetIntrinsic may insert one extra chain to |
10507 | // the beginning of the argument list. |
10508 | int ArgStart = isa<ConstantSDNode>(Op.getOperand(0)) ? 0 : 1; |
10509 | SDLoc DL(Op); |
10510 | switch (cast<ConstantSDNode>(Op.getOperand(ArgStart))->getZExtValue()) { |
10511 | case Intrinsic::ppc_cfence: { |
10512 | assert(ArgStart == 1 && "llvm.ppc.cfence must carry a chain argument.")((ArgStart == 1 && "llvm.ppc.cfence must carry a chain argument." ) ? static_cast<void> (0) : __assert_fail ("ArgStart == 1 && \"llvm.ppc.cfence must carry a chain argument.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10512, __PRETTY_FUNCTION__)); |
10513 | assert(Subtarget.isPPC64() && "Only 64-bit is supported for now.")((Subtarget.isPPC64() && "Only 64-bit is supported for now." ) ? static_cast<void> (0) : __assert_fail ("Subtarget.isPPC64() && \"Only 64-bit is supported for now.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10513, __PRETTY_FUNCTION__)); |
10514 | return SDValue(DAG.getMachineNode(PPC::CFENCE8, DL, MVT::Other, |
10515 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, |
10516 | Op.getOperand(ArgStart + 1)), |
10517 | Op.getOperand(0)), |
10518 | 0); |
10519 | } |
10520 | default: |
10521 | break; |
10522 | } |
10523 | return SDValue(); |
10524 | } |
10525 | |
10526 | // Lower scalar BSWAP64 to xxbrd. |
10527 | SDValue PPCTargetLowering::LowerBSWAP(SDValue Op, SelectionDAG &DAG) const { |
10528 | SDLoc dl(Op); |
10529 | // MTVSRDD |
10530 | Op = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64, Op.getOperand(0), |
10531 | Op.getOperand(0)); |
10532 | // XXBRD |
10533 | Op = DAG.getNode(ISD::BSWAP, dl, MVT::v2i64, Op); |
10534 | // MFVSRD |
10535 | int VectorIndex = 0; |
10536 | if (Subtarget.isLittleEndian()) |
10537 | VectorIndex = 1; |
10538 | Op = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Op, |
10539 | DAG.getTargetConstant(VectorIndex, dl, MVT::i32)); |
10540 | return Op; |
10541 | } |
10542 | |
10543 | // ATOMIC_CMP_SWAP for i8/i16 needs to zero-extend its input since it will be |
10544 | // compared to a value that is atomically loaded (atomic loads zero-extend). |
10545 | SDValue PPCTargetLowering::LowerATOMIC_CMP_SWAP(SDValue Op, |
10546 | SelectionDAG &DAG) const { |
10547 | assert(Op.getOpcode() == ISD::ATOMIC_CMP_SWAP &&((Op.getOpcode() == ISD::ATOMIC_CMP_SWAP && "Expecting an atomic compare-and-swap here." ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::ATOMIC_CMP_SWAP && \"Expecting an atomic compare-and-swap here.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10548, __PRETTY_FUNCTION__)) |
10548 | "Expecting an atomic compare-and-swap here.")((Op.getOpcode() == ISD::ATOMIC_CMP_SWAP && "Expecting an atomic compare-and-swap here." ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::ATOMIC_CMP_SWAP && \"Expecting an atomic compare-and-swap here.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10548, __PRETTY_FUNCTION__)); |
10549 | SDLoc dl(Op); |
10550 | auto *AtomicNode = cast<AtomicSDNode>(Op.getNode()); |
10551 | EVT MemVT = AtomicNode->getMemoryVT(); |
10552 | if (MemVT.getSizeInBits() >= 32) |
10553 | return Op; |
10554 | |
10555 | SDValue CmpOp = Op.getOperand(2); |
10556 | // If this is already correctly zero-extended, leave it alone. |
10557 | auto HighBits = APInt::getHighBitsSet(32, 32 - MemVT.getSizeInBits()); |
10558 | if (DAG.MaskedValueIsZero(CmpOp, HighBits)) |
10559 | return Op; |
10560 | |
10561 | // Clear the high bits of the compare operand. |
10562 | unsigned MaskVal = (1 << MemVT.getSizeInBits()) - 1; |
10563 | SDValue NewCmpOp = |
10564 | DAG.getNode(ISD::AND, dl, MVT::i32, CmpOp, |
10565 | DAG.getConstant(MaskVal, dl, MVT::i32)); |
10566 | |
10567 | // Replace the existing compare operand with the properly zero-extended one. |
10568 | SmallVector<SDValue, 4> Ops; |
10569 | for (int i = 0, e = AtomicNode->getNumOperands(); i < e; i++) |
10570 | Ops.push_back(AtomicNode->getOperand(i)); |
10571 | Ops[2] = NewCmpOp; |
10572 | MachineMemOperand *MMO = AtomicNode->getMemOperand(); |
10573 | SDVTList Tys = DAG.getVTList(MVT::i32, MVT::Other); |
10574 | auto NodeTy = |
10575 | (MemVT == MVT::i8) ? PPCISD::ATOMIC_CMP_SWAP_8 : PPCISD::ATOMIC_CMP_SWAP_16; |
10576 | return DAG.getMemIntrinsicNode(NodeTy, dl, Tys, Ops, MemVT, MMO); |
10577 | } |
10578 | |
10579 | SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, |
10580 | SelectionDAG &DAG) const { |
10581 | SDLoc dl(Op); |
10582 | // Create a stack slot that is 16-byte aligned. |
10583 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
10584 | int FrameIdx = MFI.CreateStackObject(16, Align(16), false); |
10585 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
10586 | SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT); |
10587 | |
10588 | // Store the input value into Value#0 of the stack slot. |
10589 | SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0), FIdx, |
10590 | MachinePointerInfo()); |
10591 | // Load it out. |
10592 | return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo()); |
10593 | } |
10594 | |
10595 | SDValue PPCTargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, |
10596 | SelectionDAG &DAG) const { |
10597 | assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT &&((Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Should only be called for ISD::INSERT_VECTOR_ELT" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::INSERT_VECTOR_ELT && \"Should only be called for ISD::INSERT_VECTOR_ELT\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10598, __PRETTY_FUNCTION__)) |
10598 | "Should only be called for ISD::INSERT_VECTOR_ELT")((Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Should only be called for ISD::INSERT_VECTOR_ELT" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::INSERT_VECTOR_ELT && \"Should only be called for ISD::INSERT_VECTOR_ELT\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10598, __PRETTY_FUNCTION__)); |
10599 | |
10600 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(2)); |
10601 | // We have legal lowering for constant indices but not for variable ones. |
10602 | if (!C) |
10603 | return SDValue(); |
10604 | |
10605 | EVT VT = Op.getValueType(); |
10606 | SDLoc dl(Op); |
10607 | SDValue V1 = Op.getOperand(0); |
10608 | SDValue V2 = Op.getOperand(1); |
10609 | // We can use MTVSRZ + VECINSERT for v8i16 and v16i8 types. |
10610 | if (VT == MVT::v8i16 || VT == MVT::v16i8) { |
10611 | SDValue Mtvsrz = DAG.getNode(PPCISD::MTVSRZ, dl, VT, V2); |
10612 | unsigned BytesInEachElement = VT.getVectorElementType().getSizeInBits() / 8; |
10613 | unsigned InsertAtElement = C->getZExtValue(); |
10614 | unsigned InsertAtByte = InsertAtElement * BytesInEachElement; |
10615 | if (Subtarget.isLittleEndian()) { |
10616 | InsertAtByte = (16 - BytesInEachElement) - InsertAtByte; |
10617 | } |
10618 | return DAG.getNode(PPCISD::VECINSERT, dl, VT, V1, Mtvsrz, |
10619 | DAG.getConstant(InsertAtByte, dl, MVT::i32)); |
10620 | } |
10621 | return Op; |
10622 | } |
10623 | |
10624 | SDValue PPCTargetLowering::LowerVectorLoad(SDValue Op, |
10625 | SelectionDAG &DAG) const { |
10626 | SDLoc dl(Op); |
10627 | LoadSDNode *LN = cast<LoadSDNode>(Op.getNode()); |
10628 | SDValue LoadChain = LN->getChain(); |
10629 | SDValue BasePtr = LN->getBasePtr(); |
10630 | EVT VT = Op.getValueType(); |
10631 | |
10632 | if (VT != MVT::v256i1 && VT != MVT::v512i1) |
10633 | return Op; |
10634 | |
10635 | // Type v256i1 is used for pairs and v512i1 is used for accumulators. |
10636 | // Here we create 2 or 4 v16i8 loads to load the pair or accumulator value in |
10637 | // 2 or 4 vsx registers. |
10638 | assert((VT != MVT::v512i1 || Subtarget.hasMMA()) &&(((VT != MVT::v512i1 || Subtarget.hasMMA()) && "Type unsupported without MMA" ) ? static_cast<void> (0) : __assert_fail ("(VT != MVT::v512i1 || Subtarget.hasMMA()) && \"Type unsupported without MMA\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10639, __PRETTY_FUNCTION__)) |
10639 | "Type unsupported without MMA")(((VT != MVT::v512i1 || Subtarget.hasMMA()) && "Type unsupported without MMA" ) ? static_cast<void> (0) : __assert_fail ("(VT != MVT::v512i1 || Subtarget.hasMMA()) && \"Type unsupported without MMA\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10639, __PRETTY_FUNCTION__)); |
10640 | assert((VT != MVT::v256i1 || Subtarget.pairedVectorMemops()) &&(((VT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && "Type unsupported without paired vector support") ? static_cast <void> (0) : __assert_fail ("(VT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && \"Type unsupported without paired vector support\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10641, __PRETTY_FUNCTION__)) |
10641 | "Type unsupported without paired vector support")(((VT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && "Type unsupported without paired vector support") ? static_cast <void> (0) : __assert_fail ("(VT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && \"Type unsupported without paired vector support\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10641, __PRETTY_FUNCTION__)); |
10642 | Align Alignment = LN->getAlign(); |
10643 | SmallVector<SDValue, 4> Loads; |
10644 | SmallVector<SDValue, 4> LoadChains; |
10645 | unsigned NumVecs = VT.getSizeInBits() / 128; |
10646 | for (unsigned Idx = 0; Idx < NumVecs; ++Idx) { |
10647 | SDValue Load = |
10648 | DAG.getLoad(MVT::v16i8, dl, LoadChain, BasePtr, |
10649 | LN->getPointerInfo().getWithOffset(Idx * 16), |
10650 | commonAlignment(Alignment, Idx * 16), |
10651 | LN->getMemOperand()->getFlags(), LN->getAAInfo()); |
10652 | BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr, |
10653 | DAG.getConstant(16, dl, BasePtr.getValueType())); |
10654 | Loads.push_back(Load); |
10655 | LoadChains.push_back(Load.getValue(1)); |
10656 | } |
10657 | if (Subtarget.isLittleEndian()) { |
10658 | std::reverse(Loads.begin(), Loads.end()); |
10659 | std::reverse(LoadChains.begin(), LoadChains.end()); |
10660 | } |
10661 | SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains); |
10662 | SDValue Value = |
10663 | DAG.getNode(VT == MVT::v512i1 ? PPCISD::ACC_BUILD : PPCISD::PAIR_BUILD, |
10664 | dl, VT, Loads); |
10665 | SDValue RetOps[] = {Value, TF}; |
10666 | return DAG.getMergeValues(RetOps, dl); |
10667 | } |
10668 | |
10669 | SDValue PPCTargetLowering::LowerVectorStore(SDValue Op, |
10670 | SelectionDAG &DAG) const { |
10671 | SDLoc dl(Op); |
10672 | StoreSDNode *SN = cast<StoreSDNode>(Op.getNode()); |
10673 | SDValue StoreChain = SN->getChain(); |
10674 | SDValue BasePtr = SN->getBasePtr(); |
10675 | SDValue Value = SN->getValue(); |
10676 | EVT StoreVT = Value.getValueType(); |
10677 | |
10678 | if (StoreVT != MVT::v256i1 && StoreVT != MVT::v512i1) |
10679 | return Op; |
10680 | |
10681 | // Type v256i1 is used for pairs and v512i1 is used for accumulators. |
10682 | // Here we create 2 or 4 v16i8 stores to store the pair or accumulator |
10683 | // underlying registers individually. |
10684 | assert((StoreVT != MVT::v512i1 || Subtarget.hasMMA()) &&(((StoreVT != MVT::v512i1 || Subtarget.hasMMA()) && "Type unsupported without MMA" ) ? static_cast<void> (0) : __assert_fail ("(StoreVT != MVT::v512i1 || Subtarget.hasMMA()) && \"Type unsupported without MMA\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10685, __PRETTY_FUNCTION__)) |
10685 | "Type unsupported without MMA")(((StoreVT != MVT::v512i1 || Subtarget.hasMMA()) && "Type unsupported without MMA" ) ? static_cast<void> (0) : __assert_fail ("(StoreVT != MVT::v512i1 || Subtarget.hasMMA()) && \"Type unsupported without MMA\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10685, __PRETTY_FUNCTION__)); |
10686 | assert((StoreVT != MVT::v256i1 || Subtarget.pairedVectorMemops()) &&(((StoreVT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && "Type unsupported without paired vector support") ? static_cast <void> (0) : __assert_fail ("(StoreVT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && \"Type unsupported without paired vector support\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10687, __PRETTY_FUNCTION__)) |
10687 | "Type unsupported without paired vector support")(((StoreVT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && "Type unsupported without paired vector support") ? static_cast <void> (0) : __assert_fail ("(StoreVT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && \"Type unsupported without paired vector support\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10687, __PRETTY_FUNCTION__)); |
10688 | Align Alignment = SN->getAlign(); |
10689 | SmallVector<SDValue, 4> Stores; |
10690 | unsigned NumVecs = 2; |
10691 | if (StoreVT == MVT::v512i1) { |
10692 | Value = DAG.getNode(PPCISD::XXMFACC, dl, MVT::v512i1, Value); |
10693 | NumVecs = 4; |
10694 | } |
10695 | for (unsigned Idx = 0; Idx < NumVecs; ++Idx) { |
10696 | unsigned VecNum = Subtarget.isLittleEndian() ? NumVecs - 1 - Idx : Idx; |
10697 | SDValue Elt = DAG.getNode(PPCISD::EXTRACT_VSX_REG, dl, MVT::v16i8, Value, |
10698 | DAG.getConstant(VecNum, dl, MVT::i64)); |
10699 | SDValue Store = |
10700 | DAG.getStore(StoreChain, dl, Elt, BasePtr, |
10701 | SN->getPointerInfo().getWithOffset(Idx * 16), |
10702 | commonAlignment(Alignment, Idx * 16), |
10703 | SN->getMemOperand()->getFlags(), SN->getAAInfo()); |
10704 | BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr, |
10705 | DAG.getConstant(16, dl, BasePtr.getValueType())); |
10706 | Stores.push_back(Store); |
10707 | } |
10708 | SDValue TF = DAG.getTokenFactor(dl, Stores); |
10709 | return TF; |
10710 | } |
10711 | |
10712 | SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const { |
10713 | SDLoc dl(Op); |
10714 | if (Op.getValueType() == MVT::v4i32) { |
10715 | SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1); |
10716 | |
10717 | SDValue Zero = getCanonicalConstSplat(0, 1, MVT::v4i32, DAG, dl); |
10718 | // +16 as shift amt. |
10719 | SDValue Neg16 = getCanonicalConstSplat(-16, 4, MVT::v4i32, DAG, dl); |
10720 | SDValue RHSSwap = // = vrlw RHS, 16 |
10721 | BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl); |
10722 | |
10723 | // Shrinkify inputs to v8i16. |
10724 | LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS); |
10725 | RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS); |
10726 | RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap); |
10727 | |
10728 | // Low parts multiplied together, generating 32-bit results (we ignore the |
10729 | // top parts). |
10730 | SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh, |
10731 | LHS, RHS, DAG, dl, MVT::v4i32); |
10732 | |
10733 | SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm, |
10734 | LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32); |
10735 | // Shift the high parts up 16 bits. |
10736 | HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd, |
10737 | Neg16, DAG, dl); |
10738 | return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd); |
10739 | } else if (Op.getValueType() == MVT::v16i8) { |
10740 | SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1); |
10741 | bool isLittleEndian = Subtarget.isLittleEndian(); |
10742 | |
10743 | // Multiply the even 8-bit parts, producing 16-bit sums. |
10744 | SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub, |
10745 | LHS, RHS, DAG, dl, MVT::v8i16); |
10746 | EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts); |
10747 | |
10748 | // Multiply the odd 8-bit parts, producing 16-bit sums. |
10749 | SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub, |
10750 | LHS, RHS, DAG, dl, MVT::v8i16); |
10751 | OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts); |
10752 | |
10753 | // Merge the results together. Because vmuleub and vmuloub are |
10754 | // instructions with a big-endian bias, we must reverse the |
10755 | // element numbering and reverse the meaning of "odd" and "even" |
10756 | // when generating little endian code. |
10757 | int Ops[16]; |
10758 | for (unsigned i = 0; i != 8; ++i) { |
10759 | if (isLittleEndian) { |
10760 | Ops[i*2 ] = 2*i; |
10761 | Ops[i*2+1] = 2*i+16; |
10762 | } else { |
10763 | Ops[i*2 ] = 2*i+1; |
10764 | Ops[i*2+1] = 2*i+1+16; |
10765 | } |
10766 | } |
10767 | if (isLittleEndian) |
10768 | return DAG.getVectorShuffle(MVT::v16i8, dl, OddParts, EvenParts, Ops); |
10769 | else |
10770 | return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops); |
10771 | } else { |
10772 | llvm_unreachable("Unknown mul to lower!")::llvm::llvm_unreachable_internal("Unknown mul to lower!", "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10772); |
10773 | } |
10774 | } |
10775 | |
10776 | SDValue PPCTargetLowering::LowerABS(SDValue Op, SelectionDAG &DAG) const { |
10777 | |
10778 | assert(Op.getOpcode() == ISD::ABS && "Should only be called for ISD::ABS")((Op.getOpcode() == ISD::ABS && "Should only be called for ISD::ABS" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::ABS && \"Should only be called for ISD::ABS\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10778, __PRETTY_FUNCTION__)); |
10779 | |
10780 | EVT VT = Op.getValueType(); |
10781 | assert(VT.isVector() &&((VT.isVector() && "Only set vector abs as custom, scalar abs shouldn't reach here!" ) ? static_cast<void> (0) : __assert_fail ("VT.isVector() && \"Only set vector abs as custom, scalar abs shouldn't reach here!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10782, __PRETTY_FUNCTION__)) |
10782 | "Only set vector abs as custom, scalar abs shouldn't reach here!")((VT.isVector() && "Only set vector abs as custom, scalar abs shouldn't reach here!" ) ? static_cast<void> (0) : __assert_fail ("VT.isVector() && \"Only set vector abs as custom, scalar abs shouldn't reach here!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10782, __PRETTY_FUNCTION__)); |
10783 | assert((VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||(((VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 || VT == MVT::v16i8) && "Unexpected vector element type!" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 || VT == MVT::v16i8) && \"Unexpected vector element type!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10785, __PRETTY_FUNCTION__)) |
10784 | VT == MVT::v16i8) &&(((VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 || VT == MVT::v16i8) && "Unexpected vector element type!" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 || VT == MVT::v16i8) && \"Unexpected vector element type!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10785, __PRETTY_FUNCTION__)) |
10785 | "Unexpected vector element type!")(((VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 || VT == MVT::v16i8) && "Unexpected vector element type!" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 || VT == MVT::v16i8) && \"Unexpected vector element type!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10785, __PRETTY_FUNCTION__)); |
10786 | assert((VT != MVT::v2i64 || Subtarget.hasP8Altivec()) &&(((VT != MVT::v2i64 || Subtarget.hasP8Altivec()) && "Current subtarget doesn't support smax v2i64!" ) ? static_cast<void> (0) : __assert_fail ("(VT != MVT::v2i64 || Subtarget.hasP8Altivec()) && \"Current subtarget doesn't support smax v2i64!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10787, __PRETTY_FUNCTION__)) |
10787 | "Current subtarget doesn't support smax v2i64!")(((VT != MVT::v2i64 || Subtarget.hasP8Altivec()) && "Current subtarget doesn't support smax v2i64!" ) ? static_cast<void> (0) : __assert_fail ("(VT != MVT::v2i64 || Subtarget.hasP8Altivec()) && \"Current subtarget doesn't support smax v2i64!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10787, __PRETTY_FUNCTION__)); |
10788 | |
10789 | // For vector abs, it can be lowered to: |
10790 | // abs x |
10791 | // ==> |
10792 | // y = -x |
10793 | // smax(x, y) |
10794 | |
10795 | SDLoc dl(Op); |
10796 | SDValue X = Op.getOperand(0); |
10797 | SDValue Zero = DAG.getConstant(0, dl, VT); |
10798 | SDValue Y = DAG.getNode(ISD::SUB, dl, VT, Zero, X); |
10799 | |
10800 | // SMAX patch https://reviews.llvm.org/D47332 |
10801 | // hasn't landed yet, so use intrinsic first here. |
10802 | // TODO: Should use SMAX directly once SMAX patch landed |
10803 | Intrinsic::ID BifID = Intrinsic::ppc_altivec_vmaxsw; |
10804 | if (VT == MVT::v2i64) |
10805 | BifID = Intrinsic::ppc_altivec_vmaxsd; |
10806 | else if (VT == MVT::v8i16) |
10807 | BifID = Intrinsic::ppc_altivec_vmaxsh; |
10808 | else if (VT == MVT::v16i8) |
10809 | BifID = Intrinsic::ppc_altivec_vmaxsb; |
10810 | |
10811 | return BuildIntrinsicOp(BifID, X, Y, DAG, dl, VT); |
10812 | } |
10813 | |
10814 | // Custom lowering for fpext vf32 to v2f64 |
10815 | SDValue PPCTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const { |
10816 | |
10817 | assert(Op.getOpcode() == ISD::FP_EXTEND &&((Op.getOpcode() == ISD::FP_EXTEND && "Should only be called for ISD::FP_EXTEND" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::FP_EXTEND && \"Should only be called for ISD::FP_EXTEND\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10818, __PRETTY_FUNCTION__)) |
10818 | "Should only be called for ISD::FP_EXTEND")((Op.getOpcode() == ISD::FP_EXTEND && "Should only be called for ISD::FP_EXTEND" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::FP_EXTEND && \"Should only be called for ISD::FP_EXTEND\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10818, __PRETTY_FUNCTION__)); |
10819 | |
10820 | // FIXME: handle extends from half precision float vectors on P9. |
10821 | // We only want to custom lower an extend from v2f32 to v2f64. |
10822 | if (Op.getValueType() != MVT::v2f64 || |
10823 | Op.getOperand(0).getValueType() != MVT::v2f32) |
10824 | return SDValue(); |
10825 | |
10826 | SDLoc dl(Op); |
10827 | SDValue Op0 = Op.getOperand(0); |
10828 | |
10829 | switch (Op0.getOpcode()) { |
10830 | default: |
10831 | return SDValue(); |
10832 | case ISD::EXTRACT_SUBVECTOR: { |
10833 | assert(Op0.getNumOperands() == 2 &&((Op0.getNumOperands() == 2 && isa<ConstantSDNode> (Op0->getOperand(1)) && "Node should have 2 operands with second one being a constant!" ) ? static_cast<void> (0) : __assert_fail ("Op0.getNumOperands() == 2 && isa<ConstantSDNode>(Op0->getOperand(1)) && \"Node should have 2 operands with second one being a constant!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10835, __PRETTY_FUNCTION__)) |
10834 | isa<ConstantSDNode>(Op0->getOperand(1)) &&((Op0.getNumOperands() == 2 && isa<ConstantSDNode> (Op0->getOperand(1)) && "Node should have 2 operands with second one being a constant!" ) ? static_cast<void> (0) : __assert_fail ("Op0.getNumOperands() == 2 && isa<ConstantSDNode>(Op0->getOperand(1)) && \"Node should have 2 operands with second one being a constant!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10835, __PRETTY_FUNCTION__)) |
10835 | "Node should have 2 operands with second one being a constant!")((Op0.getNumOperands() == 2 && isa<ConstantSDNode> (Op0->getOperand(1)) && "Node should have 2 operands with second one being a constant!" ) ? static_cast<void> (0) : __assert_fail ("Op0.getNumOperands() == 2 && isa<ConstantSDNode>(Op0->getOperand(1)) && \"Node should have 2 operands with second one being a constant!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10835, __PRETTY_FUNCTION__)); |
10836 | |
10837 | if (Op0.getOperand(0).getValueType() != MVT::v4f32) |
10838 | return SDValue(); |
10839 | |
10840 | // Custom lower is only done for high or low doubleword. |
10841 | int Idx = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue(); |
10842 | if (Idx % 2 != 0) |
10843 | return SDValue(); |
10844 | |
10845 | // Since input is v4f32, at this point Idx is either 0 or 2. |
10846 | // Shift to get the doubleword position we want. |
10847 | int DWord = Idx >> 1; |
10848 | |
10849 | // High and low word positions are different on little endian. |
10850 | if (Subtarget.isLittleEndian()) |
10851 | DWord ^= 0x1; |
10852 | |
10853 | return DAG.getNode(PPCISD::FP_EXTEND_HALF, dl, MVT::v2f64, |
10854 | Op0.getOperand(0), DAG.getConstant(DWord, dl, MVT::i32)); |
10855 | } |
10856 | case ISD::FADD: |
10857 | case ISD::FMUL: |
10858 | case ISD::FSUB: { |
10859 | SDValue NewLoad[2]; |
10860 | for (unsigned i = 0, ie = Op0.getNumOperands(); i != ie; ++i) { |
10861 | // Ensure both input are loads. |
10862 | SDValue LdOp = Op0.getOperand(i); |
10863 | if (LdOp.getOpcode() != ISD::LOAD) |
10864 | return SDValue(); |
10865 | // Generate new load node. |
10866 | LoadSDNode *LD = cast<LoadSDNode>(LdOp); |
10867 | SDValue LoadOps[] = {LD->getChain(), LD->getBasePtr()}; |
10868 | NewLoad[i] = DAG.getMemIntrinsicNode( |
10869 | PPCISD::LD_VSX_LH, dl, DAG.getVTList(MVT::v4f32, MVT::Other), LoadOps, |
10870 | LD->getMemoryVT(), LD->getMemOperand()); |
10871 | } |
10872 | SDValue NewOp = |
10873 | DAG.getNode(Op0.getOpcode(), SDLoc(Op0), MVT::v4f32, NewLoad[0], |
10874 | NewLoad[1], Op0.getNode()->getFlags()); |
10875 | return DAG.getNode(PPCISD::FP_EXTEND_HALF, dl, MVT::v2f64, NewOp, |
10876 | DAG.getConstant(0, dl, MVT::i32)); |
10877 | } |
10878 | case ISD::LOAD: { |
10879 | LoadSDNode *LD = cast<LoadSDNode>(Op0); |
10880 | SDValue LoadOps[] = {LD->getChain(), LD->getBasePtr()}; |
10881 | SDValue NewLd = DAG.getMemIntrinsicNode( |
10882 | PPCISD::LD_VSX_LH, dl, DAG.getVTList(MVT::v4f32, MVT::Other), LoadOps, |
10883 | LD->getMemoryVT(), LD->getMemOperand()); |
10884 | return DAG.getNode(PPCISD::FP_EXTEND_HALF, dl, MVT::v2f64, NewLd, |
10885 | DAG.getConstant(0, dl, MVT::i32)); |
10886 | } |
10887 | } |
10888 | llvm_unreachable("ERROR:Should return for all cases within swtich.")::llvm::llvm_unreachable_internal("ERROR:Should return for all cases within swtich." , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10888); |
10889 | } |
10890 | |
10891 | /// LowerOperation - Provide custom lowering hooks for some operations. |
10892 | /// |
10893 | SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { |
10894 | switch (Op.getOpcode()) { |
10895 | default: llvm_unreachable("Wasn't expecting to be able to lower this!")::llvm::llvm_unreachable_internal("Wasn't expecting to be able to lower this!" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10895); |
10896 | case ISD::ConstantPool: return LowerConstantPool(Op, DAG); |
10897 | case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); |
10898 | case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG); |
10899 | case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); |
10900 | case ISD::JumpTable: return LowerJumpTable(Op, DAG); |
10901 | case ISD::SETCC: return LowerSETCC(Op, DAG); |
10902 | case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG); |
10903 | case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG); |
10904 | |
10905 | // Variable argument lowering. |
10906 | case ISD::VASTART: return LowerVASTART(Op, DAG); |
10907 | case ISD::VAARG: return LowerVAARG(Op, DAG); |
10908 | case ISD::VACOPY: return LowerVACOPY(Op, DAG); |
10909 | |
10910 | case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG); |
10911 | case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG); |
10912 | case ISD::GET_DYNAMIC_AREA_OFFSET: |
10913 | return LowerGET_DYNAMIC_AREA_OFFSET(Op, DAG); |
10914 | |
10915 | // Exception handling lowering. |
10916 | case ISD::EH_DWARF_CFA: return LowerEH_DWARF_CFA(Op, DAG); |
10917 | case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG); |
10918 | case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG); |
10919 | |
10920 | case ISD::LOAD: return LowerLOAD(Op, DAG); |
10921 | case ISD::STORE: return LowerSTORE(Op, DAG); |
10922 | case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG); |
10923 | case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG); |
10924 | case ISD::STRICT_FP_TO_UINT: |
10925 | case ISD::STRICT_FP_TO_SINT: |
10926 | case ISD::FP_TO_UINT: |
10927 | case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG, SDLoc(Op)); |
10928 | case ISD::STRICT_UINT_TO_FP: |
10929 | case ISD::STRICT_SINT_TO_FP: |
10930 | case ISD::UINT_TO_FP: |
10931 | case ISD::SINT_TO_FP: return LowerINT_TO_FP(Op, DAG); |
10932 | case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG); |
10933 | |
10934 | // Lower 64-bit shifts. |
10935 | case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG); |
10936 | case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG); |
10937 | case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG); |
10938 | |
10939 | case ISD::FSHL: return LowerFunnelShift(Op, DAG); |
10940 | case ISD::FSHR: return LowerFunnelShift(Op, DAG); |
10941 | |
10942 | // Vector-related lowering. |
10943 | case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG); |
10944 | case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG); |
10945 | case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG); |
10946 | case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG); |
10947 | case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG); |
10948 | case ISD::MUL: return LowerMUL(Op, DAG); |
10949 | case ISD::ABS: return LowerABS(Op, DAG); |
10950 | case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG); |
10951 | case ISD::ROTL: return LowerROTL(Op, DAG); |
10952 | |
10953 | // For counter-based loop handling. |
10954 | case ISD::INTRINSIC_W_CHAIN: return SDValue(); |
10955 | |
10956 | case ISD::BITCAST: return LowerBITCAST(Op, DAG); |
10957 | |
10958 | // Frame & Return address. |
10959 | case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG); |
10960 | case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); |
10961 | |
10962 | case ISD::INTRINSIC_VOID: |
10963 | return LowerINTRINSIC_VOID(Op, DAG); |
10964 | case ISD::BSWAP: |
10965 | return LowerBSWAP(Op, DAG); |
10966 | case ISD::ATOMIC_CMP_SWAP: |
10967 | return LowerATOMIC_CMP_SWAP(Op, DAG); |
10968 | } |
10969 | } |
10970 | |
10971 | void PPCTargetLowering::LowerOperationWrapper(SDNode *N, |
10972 | SmallVectorImpl<SDValue> &Results, |
10973 | SelectionDAG &DAG) const { |
10974 | SDValue Res = LowerOperation(SDValue(N, 0), DAG); |
10975 | |
10976 | if (!Res.getNode()) |
10977 | return; |
10978 | |
10979 | // Take the return value as-is if original node has only one result. |
10980 | if (N->getNumValues() == 1) { |
10981 | Results.push_back(Res); |
10982 | return; |
10983 | } |
10984 | |
10985 | // New node should have the same number of results. |
10986 | assert((N->getNumValues() == Res->getNumValues()) &&(((N->getNumValues() == Res->getNumValues()) && "Lowering returned the wrong number of results!") ? static_cast <void> (0) : __assert_fail ("(N->getNumValues() == Res->getNumValues()) && \"Lowering returned the wrong number of results!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10987, __PRETTY_FUNCTION__)) |
10987 | "Lowering returned the wrong number of results!")(((N->getNumValues() == Res->getNumValues()) && "Lowering returned the wrong number of results!") ? static_cast <void> (0) : __assert_fail ("(N->getNumValues() == Res->getNumValues()) && \"Lowering returned the wrong number of results!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10987, __PRETTY_FUNCTION__)); |
10988 | |
10989 | for (unsigned i = 0; i < N->getNumValues(); ++i) |
10990 | Results.push_back(Res.getValue(i)); |
10991 | } |
10992 | |
10993 | void PPCTargetLowering::ReplaceNodeResults(SDNode *N, |
10994 | SmallVectorImpl<SDValue>&Results, |
10995 | SelectionDAG &DAG) const { |
10996 | SDLoc dl(N); |
10997 | switch (N->getOpcode()) { |
10998 | default: |
10999 | 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!" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10999); |
11000 | case ISD::READCYCLECOUNTER: { |
11001 | SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other); |
11002 | SDValue RTB = DAG.getNode(PPCISD::READ_TIME_BASE, dl, VTs, N->getOperand(0)); |
11003 | |
11004 | Results.push_back( |
11005 | DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, RTB, RTB.getValue(1))); |
11006 | Results.push_back(RTB.getValue(2)); |
11007 | break; |
11008 | } |
11009 | case ISD::INTRINSIC_W_CHAIN: { |
11010 | if (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != |
11011 | Intrinsic::loop_decrement) |
11012 | break; |
11013 | |
11014 | assert(N->getValueType(0) == MVT::i1 &&((N->getValueType(0) == MVT::i1 && "Unexpected result type for CTR decrement intrinsic" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::i1 && \"Unexpected result type for CTR decrement intrinsic\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11015, __PRETTY_FUNCTION__)) |
11015 | "Unexpected result type for CTR decrement intrinsic")((N->getValueType(0) == MVT::i1 && "Unexpected result type for CTR decrement intrinsic" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::i1 && \"Unexpected result type for CTR decrement intrinsic\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11015, __PRETTY_FUNCTION__)); |
11016 | EVT SVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), |
11017 | N->getValueType(0)); |
11018 | SDVTList VTs = DAG.getVTList(SVT, MVT::Other); |
11019 | SDValue NewInt = DAG.getNode(N->getOpcode(), dl, VTs, N->getOperand(0), |
11020 | N->getOperand(1)); |
11021 | |
11022 | Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewInt)); |
11023 | Results.push_back(NewInt.getValue(1)); |
11024 | break; |
11025 | } |
11026 | case ISD::VAARG: { |
11027 | if (!Subtarget.isSVR4ABI() || Subtarget.isPPC64()) |
11028 | return; |
11029 | |
11030 | EVT VT = N->getValueType(0); |
11031 | |
11032 | if (VT == MVT::i64) { |
11033 | SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG); |
11034 | |
11035 | Results.push_back(NewNode); |
11036 | Results.push_back(NewNode.getValue(1)); |
11037 | } |
11038 | return; |
11039 | } |
11040 | case ISD::STRICT_FP_TO_SINT: |
11041 | case ISD::STRICT_FP_TO_UINT: |
11042 | case ISD::FP_TO_SINT: |
11043 | case ISD::FP_TO_UINT: |
11044 | // LowerFP_TO_INT() can only handle f32 and f64. |
11045 | if (N->getOperand(N->isStrictFPOpcode() ? 1 : 0).getValueType() == |
11046 | MVT::ppcf128) |
11047 | return; |
11048 | Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl)); |
11049 | return; |
11050 | case ISD::TRUNCATE: { |
11051 | if (!N->getValueType(0).isVector()) |
11052 | return; |
11053 | SDValue Lowered = LowerTRUNCATEVector(SDValue(N, 0), DAG); |
11054 | if (Lowered) |
11055 | Results.push_back(Lowered); |
11056 | return; |
11057 | } |
11058 | case ISD::FSHL: |
11059 | case ISD::FSHR: |
11060 | // Don't handle funnel shifts here. |
11061 | return; |
11062 | case ISD::BITCAST: |
11063 | // Don't handle bitcast here. |
11064 | return; |
11065 | case ISD::FP_EXTEND: |
11066 | SDValue Lowered = LowerFP_EXTEND(SDValue(N, 0), DAG); |
11067 | if (Lowered) |
11068 | Results.push_back(Lowered); |
11069 | return; |
11070 | } |
11071 | } |
11072 | |
11073 | //===----------------------------------------------------------------------===// |
11074 | // Other Lowering Code |
11075 | //===----------------------------------------------------------------------===// |
11076 | |
11077 | static Instruction* callIntrinsic(IRBuilder<> &Builder, Intrinsic::ID Id) { |
11078 | Module *M = Builder.GetInsertBlock()->getParent()->getParent(); |
11079 | Function *Func = Intrinsic::getDeclaration(M, Id); |
11080 | return Builder.CreateCall(Func, {}); |
11081 | } |
11082 | |
11083 | // The mappings for emitLeading/TrailingFence is taken from |
11084 | // http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html |
11085 | Instruction *PPCTargetLowering::emitLeadingFence(IRBuilder<> &Builder, |
11086 | Instruction *Inst, |
11087 | AtomicOrdering Ord) const { |
11088 | if (Ord == AtomicOrdering::SequentiallyConsistent) |
11089 | return callIntrinsic(Builder, Intrinsic::ppc_sync); |
11090 | if (isReleaseOrStronger(Ord)) |
11091 | return callIntrinsic(Builder, Intrinsic::ppc_lwsync); |
11092 | return nullptr; |
11093 | } |
11094 | |
11095 | Instruction *PPCTargetLowering::emitTrailingFence(IRBuilder<> &Builder, |
11096 | Instruction *Inst, |
11097 | AtomicOrdering Ord) const { |
11098 | if (Inst->hasAtomicLoad() && isAcquireOrStronger(Ord)) { |
11099 | // See http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html and |
11100 | // http://www.rdrop.com/users/paulmck/scalability/paper/N2745r.2011.03.04a.html |
11101 | // and http://www.cl.cam.ac.uk/~pes20/cppppc/ for justification. |
11102 | if (isa<LoadInst>(Inst) && Subtarget.isPPC64()) |
11103 | return Builder.CreateCall( |
11104 | Intrinsic::getDeclaration( |
11105 | Builder.GetInsertBlock()->getParent()->getParent(), |
11106 | Intrinsic::ppc_cfence, {Inst->getType()}), |
11107 | {Inst}); |
11108 | // FIXME: Can use isync for rmw operation. |
11109 | return callIntrinsic(Builder, Intrinsic::ppc_lwsync); |
11110 | } |
11111 | return nullptr; |
11112 | } |
11113 | |
11114 | MachineBasicBlock * |
11115 | PPCTargetLowering::EmitAtomicBinary(MachineInstr &MI, MachineBasicBlock *BB, |
11116 | unsigned AtomicSize, |
11117 | unsigned BinOpcode, |
11118 | unsigned CmpOpcode, |
11119 | unsigned CmpPred) const { |
11120 | // This also handles ATOMIC_SWAP, indicated by BinOpcode==0. |
11121 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
11122 | |
11123 | auto LoadMnemonic = PPC::LDARX; |
11124 | auto StoreMnemonic = PPC::STDCX; |
11125 | switch (AtomicSize) { |
11126 | default: |
11127 | llvm_unreachable("Unexpected size of atomic entity")::llvm::llvm_unreachable_internal("Unexpected size of atomic entity" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11127); |
11128 | case 1: |
11129 | LoadMnemonic = PPC::LBARX; |
11130 | StoreMnemonic = PPC::STBCX; |
11131 | assert(Subtarget.hasPartwordAtomics() && "Call this only with size >=4")((Subtarget.hasPartwordAtomics() && "Call this only with size >=4" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasPartwordAtomics() && \"Call this only with size >=4\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11131, __PRETTY_FUNCTION__)); |
11132 | break; |
11133 | case 2: |
11134 | LoadMnemonic = PPC::LHARX; |
11135 | StoreMnemonic = PPC::STHCX; |
11136 | assert(Subtarget.hasPartwordAtomics() && "Call this only with size >=4")((Subtarget.hasPartwordAtomics() && "Call this only with size >=4" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasPartwordAtomics() && \"Call this only with size >=4\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11136, __PRETTY_FUNCTION__)); |
11137 | break; |
11138 | case 4: |
11139 | LoadMnemonic = PPC::LWARX; |
11140 | StoreMnemonic = PPC::STWCX; |
11141 | break; |
11142 | case 8: |
11143 | LoadMnemonic = PPC::LDARX; |
11144 | StoreMnemonic = PPC::STDCX; |
11145 | break; |
11146 | } |
11147 | |
11148 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
11149 | MachineFunction *F = BB->getParent(); |
11150 | MachineFunction::iterator It = ++BB->getIterator(); |
11151 | |
11152 | Register dest = MI.getOperand(0).getReg(); |
11153 | Register ptrA = MI.getOperand(1).getReg(); |
11154 | Register ptrB = MI.getOperand(2).getReg(); |
11155 | Register incr = MI.getOperand(3).getReg(); |
11156 | DebugLoc dl = MI.getDebugLoc(); |
11157 | |
11158 | MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB); |
11159 | MachineBasicBlock *loop2MBB = |
11160 | CmpOpcode ? F->CreateMachineBasicBlock(LLVM_BB) : nullptr; |
11161 | MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB); |
11162 | F->insert(It, loopMBB); |
11163 | if (CmpOpcode) |
11164 | F->insert(It, loop2MBB); |
11165 | F->insert(It, exitMBB); |
11166 | exitMBB->splice(exitMBB->begin(), BB, |
11167 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); |
11168 | exitMBB->transferSuccessorsAndUpdatePHIs(BB); |
11169 | |
11170 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
11171 | Register TmpReg = (!BinOpcode) ? incr : |
11172 | RegInfo.createVirtualRegister( AtomicSize == 8 ? &PPC::G8RCRegClass |
11173 | : &PPC::GPRCRegClass); |
11174 | |
11175 | // thisMBB: |
11176 | // ... |
11177 | // fallthrough --> loopMBB |
11178 | BB->addSuccessor(loopMBB); |
11179 | |
11180 | // loopMBB: |
11181 | // l[wd]arx dest, ptr |
11182 | // add r0, dest, incr |
11183 | // st[wd]cx. r0, ptr |
11184 | // bne- loopMBB |
11185 | // fallthrough --> exitMBB |
11186 | |
11187 | // For max/min... |
11188 | // loopMBB: |
11189 | // l[wd]arx dest, ptr |
11190 | // cmpl?[wd] incr, dest |
11191 | // bgt exitMBB |
11192 | // loop2MBB: |
11193 | // st[wd]cx. dest, ptr |
11194 | // bne- loopMBB |
11195 | // fallthrough --> exitMBB |
11196 | |
11197 | BB = loopMBB; |
11198 | BuildMI(BB, dl, TII->get(LoadMnemonic), dest) |
11199 | .addReg(ptrA).addReg(ptrB); |
11200 | if (BinOpcode) |
11201 | BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest); |
11202 | if (CmpOpcode) { |
11203 | // Signed comparisons of byte or halfword values must be sign-extended. |
11204 | if (CmpOpcode == PPC::CMPW && AtomicSize < 4) { |
11205 | Register ExtReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass); |
11206 | BuildMI(BB, dl, TII->get(AtomicSize == 1 ? PPC::EXTSB : PPC::EXTSH), |
11207 | ExtReg).addReg(dest); |
11208 | BuildMI(BB, dl, TII->get(CmpOpcode), PPC::CR0) |
11209 | .addReg(incr).addReg(ExtReg); |
11210 | } else |
11211 | BuildMI(BB, dl, TII->get(CmpOpcode), PPC::CR0) |
11212 | .addReg(incr).addReg(dest); |
11213 | |
11214 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
11215 | .addImm(CmpPred).addReg(PPC::CR0).addMBB(exitMBB); |
11216 | BB->addSuccessor(loop2MBB); |
11217 | BB->addSuccessor(exitMBB); |
11218 | BB = loop2MBB; |
11219 | } |
11220 | BuildMI(BB, dl, TII->get(StoreMnemonic)) |
11221 | .addReg(TmpReg).addReg(ptrA).addReg(ptrB); |
11222 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
11223 | .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB); |
11224 | BB->addSuccessor(loopMBB); |
11225 | BB->addSuccessor(exitMBB); |
11226 | |
11227 | // exitMBB: |
11228 | // ... |
11229 | BB = exitMBB; |
11230 | return BB; |
11231 | } |
11232 | |
11233 | MachineBasicBlock *PPCTargetLowering::EmitPartwordAtomicBinary( |
11234 | MachineInstr &MI, MachineBasicBlock *BB, |
11235 | bool is8bit, // operation |
11236 | unsigned BinOpcode, unsigned CmpOpcode, unsigned CmpPred) const { |
11237 | // If we support part-word atomic mnemonics, just use them |
11238 | if (Subtarget.hasPartwordAtomics()) |
11239 | return EmitAtomicBinary(MI, BB, is8bit ? 1 : 2, BinOpcode, CmpOpcode, |
11240 | CmpPred); |
11241 | |
11242 | // This also handles ATOMIC_SWAP, indicated by BinOpcode==0. |
11243 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
11244 | // In 64 bit mode we have to use 64 bits for addresses, even though the |
11245 | // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address |
11246 | // registers without caring whether they're 32 or 64, but here we're |
11247 | // doing actual arithmetic on the addresses. |
11248 | bool is64bit = Subtarget.isPPC64(); |
11249 | bool isLittleEndian = Subtarget.isLittleEndian(); |
11250 | unsigned ZeroReg = is64bit ? PPC::ZERO8 : PPC::ZERO; |
11251 | |
11252 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
11253 | MachineFunction *F = BB->getParent(); |
11254 | MachineFunction::iterator It = ++BB->getIterator(); |
11255 | |
11256 | Register dest = MI.getOperand(0).getReg(); |
11257 | Register ptrA = MI.getOperand(1).getReg(); |
11258 | Register ptrB = MI.getOperand(2).getReg(); |
11259 | Register incr = MI.getOperand(3).getReg(); |
11260 | DebugLoc dl = MI.getDebugLoc(); |
11261 | |
11262 | MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB); |
11263 | MachineBasicBlock *loop2MBB = |
11264 | CmpOpcode ? F->CreateMachineBasicBlock(LLVM_BB) : nullptr; |
11265 | MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB); |
11266 | F->insert(It, loopMBB); |
11267 | if (CmpOpcode) |
11268 | F->insert(It, loop2MBB); |
11269 | F->insert(It, exitMBB); |
11270 | exitMBB->splice(exitMBB->begin(), BB, |
11271 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); |
11272 | exitMBB->transferSuccessorsAndUpdatePHIs(BB); |
11273 | |
11274 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
11275 | const TargetRegisterClass *RC = |
11276 | is64bit ? &PPC::G8RCRegClass : &PPC::GPRCRegClass; |
11277 | const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; |
11278 | |
11279 | Register PtrReg = RegInfo.createVirtualRegister(RC); |
11280 | Register Shift1Reg = RegInfo.createVirtualRegister(GPRC); |
11281 | Register ShiftReg = |
11282 | isLittleEndian ? Shift1Reg : RegInfo.createVirtualRegister(GPRC); |
11283 | Register Incr2Reg = RegInfo.createVirtualRegister(GPRC); |
11284 | Register MaskReg = RegInfo.createVirtualRegister(GPRC); |
11285 | Register Mask2Reg = RegInfo.createVirtualRegister(GPRC); |
11286 | Register Mask3Reg = RegInfo.createVirtualRegister(GPRC); |
11287 | Register Tmp2Reg = RegInfo.createVirtualRegister(GPRC); |
11288 | Register Tmp3Reg = RegInfo.createVirtualRegister(GPRC); |
11289 | Register Tmp4Reg = RegInfo.createVirtualRegister(GPRC); |
11290 | Register TmpDestReg = RegInfo.createVirtualRegister(GPRC); |
11291 | Register Ptr1Reg; |
11292 | Register TmpReg = |
11293 | (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(GPRC); |
11294 | |
11295 | // thisMBB: |
11296 | // ... |
11297 | // fallthrough --> loopMBB |
11298 | BB->addSuccessor(loopMBB); |
11299 | |
11300 | // The 4-byte load must be aligned, while a char or short may be |
11301 | // anywhere in the word. Hence all this nasty bookkeeping code. |
11302 | // add ptr1, ptrA, ptrB [copy if ptrA==0] |
11303 | // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27] |
11304 | // xori shift, shift1, 24 [16] |
11305 | // rlwinm ptr, ptr1, 0, 0, 29 |
11306 | // slw incr2, incr, shift |
11307 | // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535] |
11308 | // slw mask, mask2, shift |
11309 | // loopMBB: |
11310 | // lwarx tmpDest, ptr |
11311 | // add tmp, tmpDest, incr2 |
11312 | // andc tmp2, tmpDest, mask |
11313 | // and tmp3, tmp, mask |
11314 | // or tmp4, tmp3, tmp2 |
11315 | // stwcx. tmp4, ptr |
11316 | // bne- loopMBB |
11317 | // fallthrough --> exitMBB |
11318 | // srw dest, tmpDest, shift |
11319 | if (ptrA != ZeroReg) { |
11320 | Ptr1Reg = RegInfo.createVirtualRegister(RC); |
11321 | BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg) |
11322 | .addReg(ptrA) |
11323 | .addReg(ptrB); |
11324 | } else { |
11325 | Ptr1Reg = ptrB; |
11326 | } |
11327 | // We need use 32-bit subregister to avoid mismatch register class in 64-bit |
11328 | // mode. |
11329 | BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg) |
11330 | .addReg(Ptr1Reg, 0, is64bit ? PPC::sub_32 : 0) |
11331 | .addImm(3) |
11332 | .addImm(27) |
11333 | .addImm(is8bit ? 28 : 27); |
11334 | if (!isLittleEndian) |
11335 | BuildMI(BB, dl, TII->get(PPC::XORI), ShiftReg) |
11336 | .addReg(Shift1Reg) |
11337 | .addImm(is8bit ? 24 : 16); |
11338 | if (is64bit) |
11339 | BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg) |
11340 | .addReg(Ptr1Reg) |
11341 | .addImm(0) |
11342 | .addImm(61); |
11343 | else |
11344 | BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg) |
11345 | .addReg(Ptr1Reg) |
11346 | .addImm(0) |
11347 | .addImm(0) |
11348 | .addImm(29); |
11349 | BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg).addReg(incr).addReg(ShiftReg); |
11350 | if (is8bit) |
11351 | BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255); |
11352 | else { |
11353 | BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0); |
11354 | BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg) |
11355 | .addReg(Mask3Reg) |
11356 | .addImm(65535); |
11357 | } |
11358 | BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg) |
11359 | .addReg(Mask2Reg) |
11360 | .addReg(ShiftReg); |
11361 | |
11362 | BB = loopMBB; |
11363 | BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg) |
11364 | .addReg(ZeroReg) |
11365 | .addReg(PtrReg); |
11366 | if (BinOpcode) |
11367 | BuildMI(BB, dl, TII->get(BinOpcode), TmpReg) |
11368 | .addReg(Incr2Reg) |
11369 | .addReg(TmpDestReg); |
11370 | BuildMI(BB, dl, TII->get(PPC::ANDC), Tmp2Reg) |
11371 | .addReg(TmpDestReg) |
11372 | .addReg(MaskReg); |
11373 | BuildMI(BB, dl, TII->get(PPC::AND), Tmp3Reg).addReg(TmpReg).addReg(MaskReg); |
11374 | if (CmpOpcode) { |
11375 | // For unsigned comparisons, we can directly compare the shifted values. |
11376 | // For signed comparisons we shift and sign extend. |
11377 | Register SReg = RegInfo.createVirtualRegister(GPRC); |
11378 | BuildMI(BB, dl, TII->get(PPC::AND), SReg) |
11379 | .addReg(TmpDestReg) |
11380 | .addReg(MaskReg); |
11381 | unsigned ValueReg = SReg; |
11382 | unsigned CmpReg = Incr2Reg; |
11383 | if (CmpOpcode == PPC::CMPW) { |
11384 | ValueReg = RegInfo.createVirtualRegister(GPRC); |
11385 | BuildMI(BB, dl, TII->get(PPC::SRW), ValueReg) |
11386 | .addReg(SReg) |
11387 | .addReg(ShiftReg); |
11388 | Register ValueSReg = RegInfo.createVirtualRegister(GPRC); |
11389 | BuildMI(BB, dl, TII->get(is8bit ? PPC::EXTSB : PPC::EXTSH), ValueSReg) |
11390 | .addReg(ValueReg); |
11391 | ValueReg = ValueSReg; |
11392 | CmpReg = incr; |
11393 | } |
11394 | BuildMI(BB, dl, TII->get(CmpOpcode), PPC::CR0) |
11395 | .addReg(CmpReg) |
11396 | .addReg(ValueReg); |
11397 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
11398 | .addImm(CmpPred) |
11399 | .addReg(PPC::CR0) |
11400 | .addMBB(exitMBB); |
11401 | BB->addSuccessor(loop2MBB); |
11402 | BB->addSuccessor(exitMBB); |
11403 | BB = loop2MBB; |
11404 | } |
11405 | BuildMI(BB, dl, TII->get(PPC::OR), Tmp4Reg).addReg(Tmp3Reg).addReg(Tmp2Reg); |
11406 | BuildMI(BB, dl, TII->get(PPC::STWCX)) |
11407 | .addReg(Tmp4Reg) |
11408 | .addReg(ZeroReg) |
11409 | .addReg(PtrReg); |
11410 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
11411 | .addImm(PPC::PRED_NE) |
11412 | .addReg(PPC::CR0) |
11413 | .addMBB(loopMBB); |
11414 | BB->addSuccessor(loopMBB); |
11415 | BB->addSuccessor(exitMBB); |
11416 | |
11417 | // exitMBB: |
11418 | // ... |
11419 | BB = exitMBB; |
11420 | BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest) |
11421 | .addReg(TmpDestReg) |
11422 | .addReg(ShiftReg); |
11423 | return BB; |
11424 | } |
11425 | |
11426 | llvm::MachineBasicBlock * |
11427 | PPCTargetLowering::emitEHSjLjSetJmp(MachineInstr &MI, |
11428 | MachineBasicBlock *MBB) const { |
11429 | DebugLoc DL = MI.getDebugLoc(); |
11430 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
11431 | const PPCRegisterInfo *TRI = Subtarget.getRegisterInfo(); |
11432 | |
11433 | MachineFunction *MF = MBB->getParent(); |
11434 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
11435 | |
11436 | const BasicBlock *BB = MBB->getBasicBlock(); |
11437 | MachineFunction::iterator I = ++MBB->getIterator(); |
11438 | |
11439 | Register DstReg = MI.getOperand(0).getReg(); |
11440 | const TargetRegisterClass *RC = MRI.getRegClass(DstReg); |
11441 | assert(TRI->isTypeLegalForClass(*RC, MVT::i32) && "Invalid destination!")((TRI->isTypeLegalForClass(*RC, MVT::i32) && "Invalid destination!" ) ? static_cast<void> (0) : __assert_fail ("TRI->isTypeLegalForClass(*RC, MVT::i32) && \"Invalid destination!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11441, __PRETTY_FUNCTION__)); |
11442 | Register mainDstReg = MRI.createVirtualRegister(RC); |
11443 | Register restoreDstReg = MRI.createVirtualRegister(RC); |
11444 | |
11445 | MVT PVT = getPointerTy(MF->getDataLayout()); |
11446 | 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!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11447, __PRETTY_FUNCTION__)) |
11447 | "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!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11447, __PRETTY_FUNCTION__)); |
11448 | // For v = setjmp(buf), we generate |
11449 | // |
11450 | // thisMBB: |
11451 | // SjLjSetup mainMBB |
11452 | // bl mainMBB |
11453 | // v_restore = 1 |
11454 | // b sinkMBB |
11455 | // |
11456 | // mainMBB: |
11457 | // buf[LabelOffset] = LR |
11458 | // v_main = 0 |
11459 | // |
11460 | // sinkMBB: |
11461 | // v = phi(main, restore) |
11462 | // |
11463 | |
11464 | MachineBasicBlock *thisMBB = MBB; |
11465 | MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB); |
11466 | MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB); |
11467 | MF->insert(I, mainMBB); |
11468 | MF->insert(I, sinkMBB); |
11469 | |
11470 | MachineInstrBuilder MIB; |
11471 | |
11472 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
11473 | sinkMBB->splice(sinkMBB->begin(), MBB, |
11474 | std::next(MachineBasicBlock::iterator(MI)), MBB->end()); |
11475 | sinkMBB->transferSuccessorsAndUpdatePHIs(MBB); |
11476 | |
11477 | // Note that the structure of the jmp_buf used here is not compatible |
11478 | // with that used by libc, and is not designed to be. Specifically, it |
11479 | // stores only those 'reserved' registers that LLVM does not otherwise |
11480 | // understand how to spill. Also, by convention, by the time this |
11481 | // intrinsic is called, Clang has already stored the frame address in the |
11482 | // first slot of the buffer and stack address in the third. Following the |
11483 | // X86 target code, we'll store the jump address in the second slot. We also |
11484 | // need to save the TOC pointer (R2) to handle jumps between shared |
11485 | // libraries, and that will be stored in the fourth slot. The thread |
11486 | // identifier (R13) is not affected. |
11487 | |
11488 | // thisMBB: |
11489 | const int64_t LabelOffset = 1 * PVT.getStoreSize(); |
11490 | const int64_t TOCOffset = 3 * PVT.getStoreSize(); |
11491 | const int64_t BPOffset = 4 * PVT.getStoreSize(); |
11492 | |
11493 | // Prepare IP either in reg. |
11494 | const TargetRegisterClass *PtrRC = getRegClassFor(PVT); |
11495 | Register LabelReg = MRI.createVirtualRegister(PtrRC); |
11496 | Register BufReg = MI.getOperand(1).getReg(); |
11497 | |
11498 | if (Subtarget.is64BitELFABI()) { |
11499 | setUsesTOCBasePtr(*MBB->getParent()); |
11500 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::STD)) |
11501 | .addReg(PPC::X2) |
11502 | .addImm(TOCOffset) |
11503 | .addReg(BufReg) |
11504 | .cloneMemRefs(MI); |
11505 | } |
11506 | |
11507 | // Naked functions never have a base pointer, and so we use r1. For all |
11508 | // other functions, this decision must be delayed until during PEI. |
11509 | unsigned BaseReg; |
11510 | if (MF->getFunction().hasFnAttribute(Attribute::Naked)) |
11511 | BaseReg = Subtarget.isPPC64() ? PPC::X1 : PPC::R1; |
11512 | else |
11513 | BaseReg = Subtarget.isPPC64() ? PPC::BP8 : PPC::BP; |
11514 | |
11515 | MIB = BuildMI(*thisMBB, MI, DL, |
11516 | TII->get(Subtarget.isPPC64() ? PPC::STD : PPC::STW)) |
11517 | .addReg(BaseReg) |
11518 | .addImm(BPOffset) |
11519 | .addReg(BufReg) |
11520 | .cloneMemRefs(MI); |
11521 | |
11522 | // Setup |
11523 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::BCLalways)).addMBB(mainMBB); |
11524 | MIB.addRegMask(TRI->getNoPreservedMask()); |
11525 | |
11526 | BuildMI(*thisMBB, MI, DL, TII->get(PPC::LI), restoreDstReg).addImm(1); |
11527 | |
11528 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::EH_SjLj_Setup)) |
11529 | .addMBB(mainMBB); |
11530 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::B)).addMBB(sinkMBB); |
11531 | |
11532 | thisMBB->addSuccessor(mainMBB, BranchProbability::getZero()); |
11533 | thisMBB->addSuccessor(sinkMBB, BranchProbability::getOne()); |
11534 | |
11535 | // mainMBB: |
11536 | // mainDstReg = 0 |
11537 | MIB = |
11538 | BuildMI(mainMBB, DL, |
11539 | TII->get(Subtarget.isPPC64() ? PPC::MFLR8 : PPC::MFLR), LabelReg); |
11540 | |
11541 | // Store IP |
11542 | if (Subtarget.isPPC64()) { |
11543 | MIB = BuildMI(mainMBB, DL, TII->get(PPC::STD)) |
11544 | .addReg(LabelReg) |
11545 | .addImm(LabelOffset) |
11546 | .addReg(BufReg); |
11547 | } else { |
11548 | MIB = BuildMI(mainMBB, DL, TII->get(PPC::STW)) |
11549 | .addReg(LabelReg) |
11550 | .addImm(LabelOffset) |
11551 | .addReg(BufReg); |
11552 | } |
11553 | MIB.cloneMemRefs(MI); |
11554 | |
11555 | BuildMI(mainMBB, DL, TII->get(PPC::LI), mainDstReg).addImm(0); |
11556 | mainMBB->addSuccessor(sinkMBB); |
11557 | |
11558 | // sinkMBB: |
11559 | BuildMI(*sinkMBB, sinkMBB->begin(), DL, |
11560 | TII->get(PPC::PHI), DstReg) |
11561 | .addReg(mainDstReg).addMBB(mainMBB) |
11562 | .addReg(restoreDstReg).addMBB(thisMBB); |
11563 | |
11564 | MI.eraseFromParent(); |
11565 | return sinkMBB; |
11566 | } |
11567 | |
11568 | MachineBasicBlock * |
11569 | PPCTargetLowering::emitEHSjLjLongJmp(MachineInstr &MI, |
11570 | MachineBasicBlock *MBB) const { |
11571 | DebugLoc DL = MI.getDebugLoc(); |
11572 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
11573 | |
11574 | MachineFunction *MF = MBB->getParent(); |
11575 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
11576 | |
11577 | MVT PVT = getPointerTy(MF->getDataLayout()); |
11578 | 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!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11579, __PRETTY_FUNCTION__)) |
11579 | "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!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11579, __PRETTY_FUNCTION__)); |
11580 | |
11581 | const TargetRegisterClass *RC = |
11582 | (PVT == MVT::i64) ? &PPC::G8RCRegClass : &PPC::GPRCRegClass; |
11583 | Register Tmp = MRI.createVirtualRegister(RC); |
11584 | // Since FP is only updated here but NOT referenced, it's treated as GPR. |
11585 | unsigned FP = (PVT == MVT::i64) ? PPC::X31 : PPC::R31; |
11586 | unsigned SP = (PVT == MVT::i64) ? PPC::X1 : PPC::R1; |
11587 | unsigned BP = |
11588 | (PVT == MVT::i64) |
11589 | ? PPC::X30 |
11590 | : (Subtarget.isSVR4ABI() && isPositionIndependent() ? PPC::R29 |
11591 | : PPC::R30); |
11592 | |
11593 | MachineInstrBuilder MIB; |
11594 | |
11595 | const int64_t LabelOffset = 1 * PVT.getStoreSize(); |
11596 | const int64_t SPOffset = 2 * PVT.getStoreSize(); |
11597 | const int64_t TOCOffset = 3 * PVT.getStoreSize(); |
11598 | const int64_t BPOffset = 4 * PVT.getStoreSize(); |
11599 | |
11600 | Register BufReg = MI.getOperand(0).getReg(); |
11601 | |
11602 | // Reload FP (the jumped-to function may not have had a |
11603 | // frame pointer, and if so, then its r31 will be restored |
11604 | // as necessary). |
11605 | if (PVT == MVT::i64) { |
11606 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), FP) |
11607 | .addImm(0) |
11608 | .addReg(BufReg); |
11609 | } else { |
11610 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), FP) |
11611 | .addImm(0) |
11612 | .addReg(BufReg); |
11613 | } |
11614 | MIB.cloneMemRefs(MI); |
11615 | |
11616 | // Reload IP |
11617 | if (PVT == MVT::i64) { |
11618 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), Tmp) |
11619 | .addImm(LabelOffset) |
11620 | .addReg(BufReg); |
11621 | } else { |
11622 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), Tmp) |
11623 | .addImm(LabelOffset) |
11624 | .addReg(BufReg); |
11625 | } |
11626 | MIB.cloneMemRefs(MI); |
11627 | |
11628 | // Reload SP |
11629 | if (PVT == MVT::i64) { |
11630 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), SP) |
11631 | .addImm(SPOffset) |
11632 | .addReg(BufReg); |
11633 | } else { |
11634 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), SP) |
11635 | .addImm(SPOffset) |
11636 | .addReg(BufReg); |
11637 | } |
11638 | MIB.cloneMemRefs(MI); |
11639 | |
11640 | // Reload BP |
11641 | if (PVT == MVT::i64) { |
11642 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), BP) |
11643 | .addImm(BPOffset) |
11644 | .addReg(BufReg); |
11645 | } else { |
11646 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), BP) |
11647 | .addImm(BPOffset) |
11648 | .addReg(BufReg); |
11649 | } |
11650 | MIB.cloneMemRefs(MI); |
11651 | |
11652 | // Reload TOC |
11653 | if (PVT == MVT::i64 && Subtarget.isSVR4ABI()) { |
11654 | setUsesTOCBasePtr(*MBB->getParent()); |
11655 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), PPC::X2) |
11656 | .addImm(TOCOffset) |
11657 | .addReg(BufReg) |
11658 | .cloneMemRefs(MI); |
11659 | } |
11660 | |
11661 | // Jump |
11662 | BuildMI(*MBB, MI, DL, |
11663 | TII->get(PVT == MVT::i64 ? PPC::MTCTR8 : PPC::MTCTR)).addReg(Tmp); |
11664 | BuildMI(*MBB, MI, DL, TII->get(PVT == MVT::i64 ? PPC::BCTR8 : PPC::BCTR)); |
11665 | |
11666 | MI.eraseFromParent(); |
11667 | return MBB; |
11668 | } |
11669 | |
11670 | bool PPCTargetLowering::hasInlineStackProbe(MachineFunction &MF) const { |
11671 | // If the function specifically requests inline stack probes, emit them. |
11672 | if (MF.getFunction().hasFnAttribute("probe-stack")) |
11673 | return MF.getFunction().getFnAttribute("probe-stack").getValueAsString() == |
11674 | "inline-asm"; |
11675 | return false; |
11676 | } |
11677 | |
11678 | unsigned PPCTargetLowering::getStackProbeSize(MachineFunction &MF) const { |
11679 | const TargetFrameLowering *TFI = Subtarget.getFrameLowering(); |
11680 | unsigned StackAlign = TFI->getStackAlignment(); |
11681 | assert(StackAlign >= 1 && isPowerOf2_32(StackAlign) &&((StackAlign >= 1 && isPowerOf2_32(StackAlign) && "Unexpected stack alignment") ? static_cast<void> (0) : __assert_fail ("StackAlign >= 1 && isPowerOf2_32(StackAlign) && \"Unexpected stack alignment\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11682, __PRETTY_FUNCTION__)) |
11682 | "Unexpected stack alignment")((StackAlign >= 1 && isPowerOf2_32(StackAlign) && "Unexpected stack alignment") ? static_cast<void> (0) : __assert_fail ("StackAlign >= 1 && isPowerOf2_32(StackAlign) && \"Unexpected stack alignment\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11682, __PRETTY_FUNCTION__)); |
11683 | // The default stack probe size is 4096 if the function has no |
11684 | // stack-probe-size attribute. |
11685 | unsigned StackProbeSize = 4096; |
11686 | const Function &Fn = MF.getFunction(); |
11687 | if (Fn.hasFnAttribute("stack-probe-size")) |
11688 | Fn.getFnAttribute("stack-probe-size") |
11689 | .getValueAsString() |
11690 | .getAsInteger(0, StackProbeSize); |
11691 | // Round down to the stack alignment. |
11692 | StackProbeSize &= ~(StackAlign - 1); |
11693 | return StackProbeSize ? StackProbeSize : StackAlign; |
11694 | } |
11695 | |
11696 | // Lower dynamic stack allocation with probing. `emitProbedAlloca` is splitted |
11697 | // into three phases. In the first phase, it uses pseudo instruction |
11698 | // PREPARE_PROBED_ALLOCA to get the future result of actual FramePointer and |
11699 | // FinalStackPtr. In the second phase, it generates a loop for probing blocks. |
11700 | // At last, it uses pseudo instruction DYNAREAOFFSET to get the future result of |
11701 | // MaxCallFrameSize so that it can calculate correct data area pointer. |
11702 | MachineBasicBlock * |
11703 | PPCTargetLowering::emitProbedAlloca(MachineInstr &MI, |
11704 | MachineBasicBlock *MBB) const { |
11705 | const bool isPPC64 = Subtarget.isPPC64(); |
11706 | MachineFunction *MF = MBB->getParent(); |
11707 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
11708 | DebugLoc DL = MI.getDebugLoc(); |
11709 | const unsigned ProbeSize = getStackProbeSize(*MF); |
11710 | const BasicBlock *ProbedBB = MBB->getBasicBlock(); |
11711 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
11712 | // The CFG of probing stack looks as |
11713 | // +-----+ |
11714 | // | MBB | |
11715 | // +--+--+ |
11716 | // | |
11717 | // +----v----+ |
11718 | // +--->+ TestMBB +---+ |
11719 | // | +----+----+ | |
11720 | // | | | |
11721 | // | +-----v----+ | |
11722 | // +---+ BlockMBB | | |
11723 | // +----------+ | |
11724 | // | |
11725 | // +---------+ | |
11726 | // | TailMBB +<--+ |
11727 | // +---------+ |
11728 | // In MBB, calculate previous frame pointer and final stack pointer. |
11729 | // In TestMBB, test if sp is equal to final stack pointer, if so, jump to |
11730 | // TailMBB. In BlockMBB, update the sp atomically and jump back to TestMBB. |
11731 | // TailMBB is spliced via \p MI. |
11732 | MachineBasicBlock *TestMBB = MF->CreateMachineBasicBlock(ProbedBB); |
11733 | MachineBasicBlock *TailMBB = MF->CreateMachineBasicBlock(ProbedBB); |
11734 | MachineBasicBlock *BlockMBB = MF->CreateMachineBasicBlock(ProbedBB); |
11735 | |
11736 | MachineFunction::iterator MBBIter = ++MBB->getIterator(); |
11737 | MF->insert(MBBIter, TestMBB); |
11738 | MF->insert(MBBIter, BlockMBB); |
11739 | MF->insert(MBBIter, TailMBB); |
11740 | |
11741 | const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; |
11742 | const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; |
11743 | |
11744 | Register DstReg = MI.getOperand(0).getReg(); |
11745 | Register NegSizeReg = MI.getOperand(1).getReg(); |
11746 | Register SPReg = isPPC64 ? PPC::X1 : PPC::R1; |
11747 | Register FinalStackPtr = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11748 | Register FramePointer = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11749 | Register ActualNegSizeReg = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11750 | |
11751 | // Since value of NegSizeReg might be realigned in prologepilog, insert a |
11752 | // PREPARE_PROBED_ALLOCA pseudo instruction to get actual FramePointer and |
11753 | // NegSize. |
11754 | unsigned ProbeOpc; |
11755 | if (!MRI.hasOneNonDBGUse(NegSizeReg)) |
11756 | ProbeOpc = |
11757 | isPPC64 ? PPC::PREPARE_PROBED_ALLOCA_64 : PPC::PREPARE_PROBED_ALLOCA_32; |
11758 | else |
11759 | // By introducing PREPARE_PROBED_ALLOCA_NEGSIZE_OPT, ActualNegSizeReg |
11760 | // and NegSizeReg will be allocated in the same phyreg to avoid |
11761 | // redundant copy when NegSizeReg has only one use which is current MI and |
11762 | // will be replaced by PREPARE_PROBED_ALLOCA then. |
11763 | ProbeOpc = isPPC64 ? PPC::PREPARE_PROBED_ALLOCA_NEGSIZE_SAME_REG_64 |
11764 | : PPC::PREPARE_PROBED_ALLOCA_NEGSIZE_SAME_REG_32; |
11765 | BuildMI(*MBB, {MI}, DL, TII->get(ProbeOpc), FramePointer) |
11766 | .addDef(ActualNegSizeReg) |
11767 | .addReg(NegSizeReg) |
11768 | .add(MI.getOperand(2)) |
11769 | .add(MI.getOperand(3)); |
11770 | |
11771 | // Calculate final stack pointer, which equals to SP + ActualNegSize. |
11772 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::ADD8 : PPC::ADD4), |
11773 | FinalStackPtr) |
11774 | .addReg(SPReg) |
11775 | .addReg(ActualNegSizeReg); |
11776 | |
11777 | // Materialize a scratch register for update. |
11778 | int64_t NegProbeSize = -(int64_t)ProbeSize; |
11779 | assert(isInt<32>(NegProbeSize) && "Unhandled probe size!")((isInt<32>(NegProbeSize) && "Unhandled probe size!" ) ? static_cast<void> (0) : __assert_fail ("isInt<32>(NegProbeSize) && \"Unhandled probe size!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11779, __PRETTY_FUNCTION__)); |
11780 | Register ScratchReg = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11781 | if (!isInt<16>(NegProbeSize)) { |
11782 | Register TempReg = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11783 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::LIS8 : PPC::LIS), TempReg) |
11784 | .addImm(NegProbeSize >> 16); |
11785 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::ORI8 : PPC::ORI), |
11786 | ScratchReg) |
11787 | .addReg(TempReg) |
11788 | .addImm(NegProbeSize & 0xFFFF); |
11789 | } else |
11790 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::LI8 : PPC::LI), ScratchReg) |
11791 | .addImm(NegProbeSize); |
11792 | |
11793 | { |
11794 | // Probing leading residual part. |
11795 | Register Div = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11796 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::DIVD : PPC::DIVW), Div) |
11797 | .addReg(ActualNegSizeReg) |
11798 | .addReg(ScratchReg); |
11799 | Register Mul = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11800 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::MULLD : PPC::MULLW), Mul) |
11801 | .addReg(Div) |
11802 | .addReg(ScratchReg); |
11803 | Register NegMod = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11804 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::SUBF8 : PPC::SUBF), NegMod) |
11805 | .addReg(Mul) |
11806 | .addReg(ActualNegSizeReg); |
11807 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::STDUX : PPC::STWUX), SPReg) |
11808 | .addReg(FramePointer) |
11809 | .addReg(SPReg) |
11810 | .addReg(NegMod); |
11811 | } |
11812 | |
11813 | { |
11814 | // Remaining part should be multiple of ProbeSize. |
11815 | Register CmpResult = MRI.createVirtualRegister(&PPC::CRRCRegClass); |
11816 | BuildMI(TestMBB, DL, TII->get(isPPC64 ? PPC::CMPD : PPC::CMPW), CmpResult) |
11817 | .addReg(SPReg) |
11818 | .addReg(FinalStackPtr); |
11819 | BuildMI(TestMBB, DL, TII->get(PPC::BCC)) |
11820 | .addImm(PPC::PRED_EQ) |
11821 | .addReg(CmpResult) |
11822 | .addMBB(TailMBB); |
11823 | TestMBB->addSuccessor(BlockMBB); |
11824 | TestMBB->addSuccessor(TailMBB); |
11825 | } |
11826 | |
11827 | { |
11828 | // Touch the block. |
11829 | // |P...|P...|P... |
11830 | BuildMI(BlockMBB, DL, TII->get(isPPC64 ? PPC::STDUX : PPC::STWUX), SPReg) |
11831 | .addReg(FramePointer) |
11832 | .addReg(SPReg) |
11833 | .addReg(ScratchReg); |
11834 | BuildMI(BlockMBB, DL, TII->get(PPC::B)).addMBB(TestMBB); |
11835 | BlockMBB->addSuccessor(TestMBB); |
11836 | } |
11837 | |
11838 | // Calculation of MaxCallFrameSize is deferred to prologepilog, use |
11839 | // DYNAREAOFFSET pseudo instruction to get the future result. |
11840 | Register MaxCallFrameSizeReg = |
11841 | MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11842 | BuildMI(TailMBB, DL, |
11843 | TII->get(isPPC64 ? PPC::DYNAREAOFFSET8 : PPC::DYNAREAOFFSET), |
11844 | MaxCallFrameSizeReg) |
11845 | .add(MI.getOperand(2)) |
11846 | .add(MI.getOperand(3)); |
11847 | BuildMI(TailMBB, DL, TII->get(isPPC64 ? PPC::ADD8 : PPC::ADD4), DstReg) |
11848 | .addReg(SPReg) |
11849 | .addReg(MaxCallFrameSizeReg); |
11850 | |
11851 | // Splice instructions after MI to TailMBB. |
11852 | TailMBB->splice(TailMBB->end(), MBB, |
11853 | std::next(MachineBasicBlock::iterator(MI)), MBB->end()); |
11854 | TailMBB->transferSuccessorsAndUpdatePHIs(MBB); |
11855 | MBB->addSuccessor(TestMBB); |
11856 | |
11857 | // Delete the pseudo instruction. |
11858 | MI.eraseFromParent(); |
11859 | |
11860 | ++NumDynamicAllocaProbed; |
11861 | return TailMBB; |
11862 | } |
11863 | |
11864 | MachineBasicBlock * |
11865 | PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI, |
11866 | MachineBasicBlock *BB) const { |
11867 | if (MI.getOpcode() == TargetOpcode::STACKMAP || |
11868 | MI.getOpcode() == TargetOpcode::PATCHPOINT) { |
11869 | if (Subtarget.is64BitELFABI() && |
11870 | MI.getOpcode() == TargetOpcode::PATCHPOINT && |
11871 | !Subtarget.isUsingPCRelativeCalls()) { |
11872 | // Call lowering should have added an r2 operand to indicate a dependence |
11873 | // on the TOC base pointer value. It can't however, because there is no |
11874 | // way to mark the dependence as implicit there, and so the stackmap code |
11875 | // will confuse it with a regular operand. Instead, add the dependence |
11876 | // here. |
11877 | MI.addOperand(MachineOperand::CreateReg(PPC::X2, false, true)); |
11878 | } |
11879 | |
11880 | return emitPatchPoint(MI, BB); |
11881 | } |
11882 | |
11883 | if (MI.getOpcode() == PPC::EH_SjLj_SetJmp32 || |
11884 | MI.getOpcode() == PPC::EH_SjLj_SetJmp64) { |
11885 | return emitEHSjLjSetJmp(MI, BB); |
11886 | } else if (MI.getOpcode() == PPC::EH_SjLj_LongJmp32 || |
11887 | MI.getOpcode() == PPC::EH_SjLj_LongJmp64) { |
11888 | return emitEHSjLjLongJmp(MI, BB); |
11889 | } |
11890 | |
11891 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
11892 | |
11893 | // To "insert" these instructions we actually have to insert their |
11894 | // control-flow patterns. |
11895 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
11896 | MachineFunction::iterator It = ++BB->getIterator(); |
11897 | |
11898 | MachineFunction *F = BB->getParent(); |
11899 | |
11900 | if (MI.getOpcode() == PPC::SELECT_CC_I4 || |
11901 | MI.getOpcode() == PPC::SELECT_CC_I8 || MI.getOpcode() == PPC::SELECT_I4 || |
11902 | MI.getOpcode() == PPC::SELECT_I8) { |
11903 | SmallVector<MachineOperand, 2> Cond; |
11904 | if (MI.getOpcode() == PPC::SELECT_CC_I4 || |
11905 | MI.getOpcode() == PPC::SELECT_CC_I8) |
11906 | Cond.push_back(MI.getOperand(4)); |
11907 | else |
11908 | Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET)); |
11909 | Cond.push_back(MI.getOperand(1)); |
11910 | |
11911 | DebugLoc dl = MI.getDebugLoc(); |
11912 | TII->insertSelect(*BB, MI, dl, MI.getOperand(0).getReg(), Cond, |
11913 | MI.getOperand(2).getReg(), MI.getOperand(3).getReg()); |
11914 | } else if (MI.getOpcode() == PPC::SELECT_CC_F4 || |
11915 | MI.getOpcode() == PPC::SELECT_CC_F8 || |
11916 | MI.getOpcode() == PPC::SELECT_CC_F16 || |
11917 | MI.getOpcode() == PPC::SELECT_CC_VRRC || |
11918 | MI.getOpcode() == PPC::SELECT_CC_VSFRC || |
11919 | MI.getOpcode() == PPC::SELECT_CC_VSSRC || |
11920 | MI.getOpcode() == PPC::SELECT_CC_VSRC || |
11921 | MI.getOpcode() == PPC::SELECT_CC_SPE4 || |
11922 | MI.getOpcode() == PPC::SELECT_CC_SPE || |
11923 | MI.getOpcode() == PPC::SELECT_F4 || |
11924 | MI.getOpcode() == PPC::SELECT_F8 || |
11925 | MI.getOpcode() == PPC::SELECT_F16 || |
11926 | MI.getOpcode() == PPC::SELECT_SPE || |
11927 | MI.getOpcode() == PPC::SELECT_SPE4 || |
11928 | MI.getOpcode() == PPC::SELECT_VRRC || |
11929 | MI.getOpcode() == PPC::SELECT_VSFRC || |
11930 | MI.getOpcode() == PPC::SELECT_VSSRC || |
11931 | MI.getOpcode() == PPC::SELECT_VSRC) { |
11932 | // The incoming instruction knows the destination vreg to set, the |
11933 | // condition code register to branch on, the true/false values to |
11934 | // select between, and a branch opcode to use. |
11935 | |
11936 | // thisMBB: |
11937 | // ... |
11938 | // TrueVal = ... |
11939 | // cmpTY ccX, r1, r2 |
11940 | // bCC copy1MBB |
11941 | // fallthrough --> copy0MBB |
11942 | MachineBasicBlock *thisMBB = BB; |
11943 | MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); |
11944 | MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); |
11945 | DebugLoc dl = MI.getDebugLoc(); |
11946 | F->insert(It, copy0MBB); |
11947 | F->insert(It, sinkMBB); |
11948 | |
11949 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
11950 | sinkMBB->splice(sinkMBB->begin(), BB, |
11951 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); |
11952 | sinkMBB->transferSuccessorsAndUpdatePHIs(BB); |
11953 | |
11954 | // Next, add the true and fallthrough blocks as its successors. |
11955 | BB->addSuccessor(copy0MBB); |
11956 | BB->addSuccessor(sinkMBB); |
11957 | |
11958 | if (MI.getOpcode() == PPC::SELECT_I4 || MI.getOpcode() == PPC::SELECT_I8 || |
11959 | MI.getOpcode() == PPC::SELECT_F4 || MI.getOpcode() == PPC::SELECT_F8 || |
11960 | MI.getOpcode() == PPC::SELECT_F16 || |
11961 | MI.getOpcode() == PPC::SELECT_SPE4 || |
11962 | MI.getOpcode() == PPC::SELECT_SPE || |
11963 | MI.getOpcode() == PPC::SELECT_VRRC || |
11964 | MI.getOpcode() == PPC::SELECT_VSFRC || |
11965 | MI.getOpcode() == PPC::SELECT_VSSRC || |
11966 | MI.getOpcode() == PPC::SELECT_VSRC) { |
11967 | BuildMI(BB, dl, TII->get(PPC::BC)) |
11968 | .addReg(MI.getOperand(1).getReg()) |
11969 | .addMBB(sinkMBB); |
11970 | } else { |
11971 | unsigned SelectPred = MI.getOperand(4).getImm(); |
11972 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
11973 | .addImm(SelectPred) |
11974 | .addReg(MI.getOperand(1).getReg()) |
11975 | .addMBB(sinkMBB); |
11976 | } |
11977 | |
11978 | // copy0MBB: |
11979 | // %FalseValue = ... |
11980 | // # fallthrough to sinkMBB |
11981 | BB = copy0MBB; |
11982 | |
11983 | // Update machine-CFG edges |
11984 | BB->addSuccessor(sinkMBB); |
11985 | |
11986 | // sinkMBB: |
11987 | // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] |
11988 | // ... |
11989 | BB = sinkMBB; |
11990 | BuildMI(*BB, BB->begin(), dl, TII->get(PPC::PHI), MI.getOperand(0).getReg()) |
11991 | .addReg(MI.getOperand(3).getReg()) |
11992 | .addMBB(copy0MBB) |
11993 | .addReg(MI.getOperand(2).getReg()) |
11994 | .addMBB(thisMBB); |
11995 | } else if (MI.getOpcode() == PPC::ReadTB) { |
11996 | // To read the 64-bit time-base register on a 32-bit target, we read the |
11997 | // two halves. Should the counter have wrapped while it was being read, we |
11998 | // need to try again. |
11999 | // ... |
12000 | // readLoop: |
12001 | // mfspr Rx,TBU # load from TBU |
12002 | // mfspr Ry,TB # load from TB |
12003 | // mfspr Rz,TBU # load from TBU |
12004 | // cmpw crX,Rx,Rz # check if 'old'='new' |
12005 | // bne readLoop # branch if they're not equal |
12006 | // ... |
12007 | |
12008 | MachineBasicBlock *readMBB = F->CreateMachineBasicBlock(LLVM_BB); |
12009 | MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); |
12010 | DebugLoc dl = MI.getDebugLoc(); |
12011 | F->insert(It, readMBB); |
12012 | F->insert(It, sinkMBB); |
12013 | |
12014 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
12015 | sinkMBB->splice(sinkMBB->begin(), BB, |
12016 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); |
12017 | sinkMBB->transferSuccessorsAndUpdatePHIs(BB); |
12018 | |
12019 | BB->addSuccessor(readMBB); |
12020 | BB = readMBB; |
12021 | |
12022 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
12023 | Register ReadAgainReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass); |
12024 | Register LoReg = MI.getOperand(0).getReg(); |
12025 | Register HiReg = MI.getOperand(1).getReg(); |
12026 | |
12027 | BuildMI(BB, dl, TII->get(PPC::MFSPR), HiReg).addImm(269); |
12028 | BuildMI(BB, dl, TII->get(PPC::MFSPR), LoReg).addImm(268); |
12029 | BuildMI(BB, dl, TII->get(PPC::MFSPR), ReadAgainReg).addImm(269); |
12030 | |
12031 | Register CmpReg = RegInfo.createVirtualRegister(&PPC::CRRCRegClass); |
12032 | |
12033 | BuildMI(BB, dl, TII->get(PPC::CMPW), CmpReg) |
12034 | .addReg(HiReg) |
12035 | .addReg(ReadAgainReg); |
12036 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
12037 | .addImm(PPC::PRED_NE) |
12038 | .addReg(CmpReg) |
12039 | .addMBB(readMBB); |
12040 | |
12041 | BB->addSuccessor(readMBB); |
12042 | BB->addSuccessor(sinkMBB); |
12043 | } else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I8) |
12044 | BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4); |
12045 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I16) |
12046 | BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4); |
12047 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I32) |
12048 | BB = EmitAtomicBinary(MI, BB, 4, PPC::ADD4); |
12049 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I64) |
12050 | BB = EmitAtomicBinary(MI, BB, 8, PPC::ADD8); |
12051 | |
12052 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I8) |
12053 | BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND); |
12054 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I16) |
12055 | BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND); |
12056 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I32) |
12057 | BB = EmitAtomicBinary(MI, BB, 4, PPC::AND); |
12058 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I64) |
12059 | BB = EmitAtomicBinary(MI, BB, 8, PPC::AND8); |
12060 | |
12061 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I8) |
12062 | BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR); |
12063 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I16) |
12064 | BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR); |
12065 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I32) |
12066 | BB = EmitAtomicBinary(MI, BB, 4, PPC::OR); |
12067 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I64) |
12068 | BB = EmitAtomicBinary(MI, BB, 8, PPC::OR8); |
12069 | |
12070 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I8) |
12071 | BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR); |
12072 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I16) |
12073 | BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR); |
12074 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I32) |
12075 | BB = EmitAtomicBinary(MI, BB, 4, PPC::XOR); |
12076 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I64) |
12077 | BB = EmitAtomicBinary(MI, BB, 8, PPC::XOR8); |
12078 | |
12079 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I8) |
12080 | BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::NAND); |
12081 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I16) |
12082 | BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::NAND); |
12083 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I32) |
12084 | BB = EmitAtomicBinary(MI, BB, 4, PPC::NAND); |
12085 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I64) |
12086 | BB = EmitAtomicBinary(MI, BB, 8, PPC::NAND8); |
12087 | |
12088 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I8) |
12089 | BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF); |
12090 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I16) |
12091 | BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF); |
12092 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I32) |
12093 | BB = EmitAtomicBinary(MI, BB, 4, PPC::SUBF); |
12094 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I64) |
12095 | BB = EmitAtomicBinary(MI, BB, 8, PPC::SUBF8); |
12096 | |
12097 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I8) |
12098 | BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPW, PPC::PRED_GE); |
12099 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I16) |
12100 | BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPW, PPC::PRED_GE); |
12101 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I32) |
12102 | BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPW, PPC::PRED_GE); |
12103 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I64) |
12104 | BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPD, PPC::PRED_GE); |
12105 | |
12106 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I8) |
12107 | BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPW, PPC::PRED_LE); |
12108 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I16) |
12109 | BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPW, PPC::PRED_LE); |
12110 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I32) |
12111 | BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPW, PPC::PRED_LE); |
12112 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I64) |
12113 | BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPD, PPC::PRED_LE); |
12114 | |
12115 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I8) |
12116 | BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPLW, PPC::PRED_GE); |
12117 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I16) |
12118 | BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPLW, PPC::PRED_GE); |
12119 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I32) |
12120 | BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPLW, PPC::PRED_GE); |
12121 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I64) |
12122 | BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPLD, PPC::PRED_GE); |
12123 | |
12124 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I8) |
12125 | BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPLW, PPC::PRED_LE); |
12126 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I16) |
12127 | BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPLW, PPC::PRED_LE); |
12128 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I32) |
12129 | BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPLW, PPC::PRED_LE); |
12130 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I64) |
12131 | BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPLD, PPC::PRED_LE); |
12132 | |
12133 | else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I8) |
12134 | BB = EmitPartwordAtomicBinary(MI, BB, true, 0); |
12135 | else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I16) |
12136 | BB = EmitPartwordAtomicBinary(MI, BB, false, 0); |
12137 | else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I32) |
12138 | BB = EmitAtomicBinary(MI, BB, 4, 0); |
12139 | else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I64) |
12140 | BB = EmitAtomicBinary(MI, BB, 8, 0); |
12141 | else if (MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 || |
12142 | MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I64 || |
12143 | (Subtarget.hasPartwordAtomics() && |
12144 | MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I8) || |
12145 | (Subtarget.hasPartwordAtomics() && |
12146 | MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I16)) { |
12147 | bool is64bit = MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I64; |
12148 | |
12149 | auto LoadMnemonic = PPC::LDARX; |
12150 | auto StoreMnemonic = PPC::STDCX; |
12151 | switch (MI.getOpcode()) { |
12152 | default: |
12153 | llvm_unreachable("Compare and swap of unknown size")::llvm::llvm_unreachable_internal("Compare and swap of unknown size" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12153); |
12154 | case PPC::ATOMIC_CMP_SWAP_I8: |
12155 | LoadMnemonic = PPC::LBARX; |
12156 | StoreMnemonic = PPC::STBCX; |
12157 | assert(Subtarget.hasPartwordAtomics() && "No support partword atomics.")((Subtarget.hasPartwordAtomics() && "No support partword atomics." ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasPartwordAtomics() && \"No support partword atomics.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12157, __PRETTY_FUNCTION__)); |
12158 | break; |
12159 | case PPC::ATOMIC_CMP_SWAP_I16: |
12160 | LoadMnemonic = PPC::LHARX; |
12161 | StoreMnemonic = PPC::STHCX; |
12162 | assert(Subtarget.hasPartwordAtomics() && "No support partword atomics.")((Subtarget.hasPartwordAtomics() && "No support partword atomics." ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasPartwordAtomics() && \"No support partword atomics.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12162, __PRETTY_FUNCTION__)); |
12163 | break; |
12164 | case PPC::ATOMIC_CMP_SWAP_I32: |
12165 | LoadMnemonic = PPC::LWARX; |
12166 | StoreMnemonic = PPC::STWCX; |
12167 | break; |
12168 | case PPC::ATOMIC_CMP_SWAP_I64: |
12169 | LoadMnemonic = PPC::LDARX; |
12170 | StoreMnemonic = PPC::STDCX; |
12171 | break; |
12172 | } |
12173 | Register dest = MI.getOperand(0).getReg(); |
12174 | Register ptrA = MI.getOperand(1).getReg(); |
12175 | Register ptrB = MI.getOperand(2).getReg(); |
12176 | Register oldval = MI.getOperand(3).getReg(); |
12177 | Register newval = MI.getOperand(4).getReg(); |
12178 | DebugLoc dl = MI.getDebugLoc(); |
12179 | |
12180 | MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB); |
12181 | MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB); |
12182 | MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB); |
12183 | MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB); |
12184 | F->insert(It, loop1MBB); |
12185 | F->insert(It, loop2MBB); |
12186 | F->insert(It, midMBB); |
12187 | F->insert(It, exitMBB); |
12188 | exitMBB->splice(exitMBB->begin(), BB, |
12189 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); |
12190 | exitMBB->transferSuccessorsAndUpdatePHIs(BB); |
12191 | |
12192 | // thisMBB: |
12193 | // ... |
12194 | // fallthrough --> loopMBB |
12195 | BB->addSuccessor(loop1MBB); |
12196 | |
12197 | // loop1MBB: |
12198 | // l[bhwd]arx dest, ptr |
12199 | // cmp[wd] dest, oldval |
12200 | // bne- midMBB |
12201 | // loop2MBB: |
12202 | // st[bhwd]cx. newval, ptr |
12203 | // bne- loopMBB |
12204 | // b exitBB |
12205 | // midMBB: |
12206 | // st[bhwd]cx. dest, ptr |
12207 | // exitBB: |
12208 | BB = loop1MBB; |
12209 | BuildMI(BB, dl, TII->get(LoadMnemonic), dest).addReg(ptrA).addReg(ptrB); |
12210 | BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0) |
12211 | .addReg(oldval) |
12212 | .addReg(dest); |
12213 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
12214 | .addImm(PPC::PRED_NE) |
12215 | .addReg(PPC::CR0) |
12216 | .addMBB(midMBB); |
12217 | BB->addSuccessor(loop2MBB); |
12218 | BB->addSuccessor(midMBB); |
12219 | |
12220 | BB = loop2MBB; |
12221 | BuildMI(BB, dl, TII->get(StoreMnemonic)) |
12222 | .addReg(newval) |
12223 | .addReg(ptrA) |
12224 | .addReg(ptrB); |
12225 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
12226 | .addImm(PPC::PRED_NE) |
12227 | .addReg(PPC::CR0) |
12228 | .addMBB(loop1MBB); |
12229 | BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB); |
12230 | BB->addSuccessor(loop1MBB); |
12231 | BB->addSuccessor(exitMBB); |
12232 | |
12233 | BB = midMBB; |
12234 | BuildMI(BB, dl, TII->get(StoreMnemonic)) |
12235 | .addReg(dest) |
12236 | .addReg(ptrA) |
12237 | .addReg(ptrB); |
12238 | BB->addSuccessor(exitMBB); |
12239 | |
12240 | // exitMBB: |
12241 | // ... |
12242 | BB = exitMBB; |
12243 | } else if (MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 || |
12244 | MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) { |
12245 | // We must use 64-bit registers for addresses when targeting 64-bit, |
12246 | // since we're actually doing arithmetic on them. Other registers |
12247 | // can be 32-bit. |
12248 | bool is64bit = Subtarget.isPPC64(); |
12249 | bool isLittleEndian = Subtarget.isLittleEndian(); |
12250 | bool is8bit = MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I8; |
12251 | |
12252 | Register dest = MI.getOperand(0).getReg(); |
12253 | Register ptrA = MI.getOperand(1).getReg(); |
12254 | Register ptrB = MI.getOperand(2).getReg(); |
12255 | Register oldval = MI.getOperand(3).getReg(); |
12256 | Register newval = MI.getOperand(4).getReg(); |
12257 | DebugLoc dl = MI.getDebugLoc(); |
12258 | |
12259 | MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB); |
12260 | MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB); |
12261 | MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB); |
12262 | MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB); |
12263 | F->insert(It, loop1MBB); |
12264 | F->insert(It, loop2MBB); |
12265 | F->insert(It, midMBB); |
12266 | F->insert(It, exitMBB); |
12267 | exitMBB->splice(exitMBB->begin(), BB, |
12268 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); |
12269 | exitMBB->transferSuccessorsAndUpdatePHIs(BB); |
12270 | |
12271 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
12272 | const TargetRegisterClass *RC = |
12273 | is64bit ? &PPC::G8RCRegClass : &PPC::GPRCRegClass; |
12274 | const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; |
12275 | |
12276 | Register PtrReg = RegInfo.createVirtualRegister(RC); |
12277 | Register Shift1Reg = RegInfo.createVirtualRegister(GPRC); |
12278 | Register ShiftReg = |
12279 | isLittleEndian ? Shift1Reg : RegInfo.createVirtualRegister(GPRC); |
12280 | Register NewVal2Reg = RegInfo.createVirtualRegister(GPRC); |
12281 | Register NewVal3Reg = RegInfo.createVirtualRegister(GPRC); |
12282 | Register OldVal2Reg = RegInfo.createVirtualRegister(GPRC); |
12283 | Register OldVal3Reg = RegInfo.createVirtualRegister(GPRC); |
12284 | Register MaskReg = RegInfo.createVirtualRegister(GPRC); |
12285 | Register Mask2Reg = RegInfo.createVirtualRegister(GPRC); |
12286 | Register Mask3Reg = RegInfo.createVirtualRegister(GPRC); |
12287 | Register Tmp2Reg = RegInfo.createVirtualRegister(GPRC); |
12288 | Register Tmp4Reg = RegInfo.createVirtualRegister(GPRC); |
12289 | Register TmpDestReg = RegInfo.createVirtualRegister(GPRC); |
12290 | Register Ptr1Reg; |
12291 | Register TmpReg = RegInfo.createVirtualRegister(GPRC); |
12292 | Register ZeroReg = is64bit ? PPC::ZERO8 : PPC::ZERO; |
12293 | // thisMBB: |
12294 | // ... |
12295 | // fallthrough --> loopMBB |
12296 | BB->addSuccessor(loop1MBB); |
12297 | |
12298 | // The 4-byte load must be aligned, while a char or short may be |
12299 | // anywhere in the word. Hence all this nasty bookkeeping code. |
12300 | // add ptr1, ptrA, ptrB [copy if ptrA==0] |
12301 | // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27] |
12302 | // xori shift, shift1, 24 [16] |
12303 | // rlwinm ptr, ptr1, 0, 0, 29 |
12304 | // slw newval2, newval, shift |
12305 | // slw oldval2, oldval,shift |
12306 | // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535] |
12307 | // slw mask, mask2, shift |
12308 | // and newval3, newval2, mask |
12309 | // and oldval3, oldval2, mask |
12310 | // loop1MBB: |
12311 | // lwarx tmpDest, ptr |
12312 | // and tmp, tmpDest, mask |
12313 | // cmpw tmp, oldval3 |
12314 | // bne- midMBB |
12315 | // loop2MBB: |
12316 | // andc tmp2, tmpDest, mask |
12317 | // or tmp4, tmp2, newval3 |
12318 | // stwcx. tmp4, ptr |
12319 | // bne- loop1MBB |
12320 | // b exitBB |
12321 | // midMBB: |
12322 | // stwcx. tmpDest, ptr |
12323 | // exitBB: |
12324 | // srw dest, tmpDest, shift |
12325 | if (ptrA != ZeroReg) { |
12326 | Ptr1Reg = RegInfo.createVirtualRegister(RC); |
12327 | BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg) |
12328 | .addReg(ptrA) |
12329 | .addReg(ptrB); |
12330 | } else { |
12331 | Ptr1Reg = ptrB; |
12332 | } |
12333 | |
12334 | // We need use 32-bit subregister to avoid mismatch register class in 64-bit |
12335 | // mode. |
12336 | BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg) |
12337 | .addReg(Ptr1Reg, 0, is64bit ? PPC::sub_32 : 0) |
12338 | .addImm(3) |
12339 | .addImm(27) |
12340 | .addImm(is8bit ? 28 : 27); |
12341 | if (!isLittleEndian) |
12342 | BuildMI(BB, dl, TII->get(PPC::XORI), ShiftReg) |
12343 | .addReg(Shift1Reg) |
12344 | .addImm(is8bit ? 24 : 16); |
12345 | if (is64bit) |
12346 | BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg) |
12347 | .addReg(Ptr1Reg) |
12348 | .addImm(0) |
12349 | .addImm(61); |
12350 | else |
12351 | BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg) |
12352 | .addReg(Ptr1Reg) |
12353 | .addImm(0) |
12354 | .addImm(0) |
12355 | .addImm(29); |
12356 | BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg) |
12357 | .addReg(newval) |
12358 | .addReg(ShiftReg); |
12359 | BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg) |
12360 | .addReg(oldval) |
12361 | .addReg(ShiftReg); |
12362 | if (is8bit) |
12363 | BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255); |
12364 | else { |
12365 | BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0); |
12366 | BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg) |
12367 | .addReg(Mask3Reg) |
12368 | .addImm(65535); |
12369 | } |
12370 | BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg) |
12371 | .addReg(Mask2Reg) |
12372 | .addReg(ShiftReg); |
12373 | BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg) |
12374 | .addReg(NewVal2Reg) |
12375 | .addReg(MaskReg); |
12376 | BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg) |
12377 | .addReg(OldVal2Reg) |
12378 | .addReg(MaskReg); |
12379 | |
12380 | BB = loop1MBB; |
12381 | BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg) |
12382 | .addReg(ZeroReg) |
12383 | .addReg(PtrReg); |
12384 | BuildMI(BB, dl, TII->get(PPC::AND), TmpReg) |
12385 | .addReg(TmpDestReg) |
12386 | .addReg(MaskReg); |
12387 | BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0) |
12388 | .addReg(TmpReg) |
12389 | .addReg(OldVal3Reg); |
12390 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
12391 | .addImm(PPC::PRED_NE) |
12392 | .addReg(PPC::CR0) |
12393 | .addMBB(midMBB); |
12394 | BB->addSuccessor(loop2MBB); |
12395 | BB->addSuccessor(midMBB); |
12396 | |
12397 | BB = loop2MBB; |
12398 | BuildMI(BB, dl, TII->get(PPC::ANDC), Tmp2Reg) |
12399 | .addReg(TmpDestReg) |
12400 | .addReg(MaskReg); |
12401 | BuildMI(BB, dl, TII->get(PPC::OR), Tmp4Reg) |
12402 | .addReg(Tmp2Reg) |
12403 | .addReg(NewVal3Reg); |
12404 | BuildMI(BB, dl, TII->get(PPC::STWCX)) |
12405 | .addReg(Tmp4Reg) |
12406 | .addReg(ZeroReg) |
12407 | .addReg(PtrReg); |
12408 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
12409 | .addImm(PPC::PRED_NE) |
12410 | .addReg(PPC::CR0) |
12411 | .addMBB(loop1MBB); |
12412 | BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB); |
12413 | BB->addSuccessor(loop1MBB); |
12414 | BB->addSuccessor(exitMBB); |
12415 | |
12416 | BB = midMBB; |
12417 | BuildMI(BB, dl, TII->get(PPC::STWCX)) |
12418 | .addReg(TmpDestReg) |
12419 | .addReg(ZeroReg) |
12420 | .addReg(PtrReg); |
12421 | BB->addSuccessor(exitMBB); |
12422 | |
12423 | // exitMBB: |
12424 | // ... |
12425 | BB = exitMBB; |
12426 | BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest) |
12427 | .addReg(TmpReg) |
12428 | .addReg(ShiftReg); |
12429 | } else if (MI.getOpcode() == PPC::FADDrtz) { |
12430 | // This pseudo performs an FADD with rounding mode temporarily forced |
12431 | // to round-to-zero. We emit this via custom inserter since the FPSCR |
12432 | // is not modeled at the SelectionDAG level. |
12433 | Register Dest = MI.getOperand(0).getReg(); |
12434 | Register Src1 = MI.getOperand(1).getReg(); |
12435 | Register Src2 = MI.getOperand(2).getReg(); |
12436 | DebugLoc dl = MI.getDebugLoc(); |
12437 | |
12438 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
12439 | Register MFFSReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass); |
12440 | |
12441 | // Save FPSCR value. |
12442 | BuildMI(*BB, MI, dl, TII->get(PPC::MFFS), MFFSReg); |
12443 | |
12444 | // Set rounding mode to round-to-zero. |
12445 | BuildMI(*BB, MI, dl, TII->get(PPC::MTFSB1)) |
12446 | .addImm(31) |
12447 | .addReg(PPC::RM, RegState::ImplicitDefine); |
12448 | |
12449 | BuildMI(*BB, MI, dl, TII->get(PPC::MTFSB0)) |
12450 | .addImm(30) |
12451 | .addReg(PPC::RM, RegState::ImplicitDefine); |
12452 | |
12453 | // Perform addition. |
12454 | auto MIB = BuildMI(*BB, MI, dl, TII->get(PPC::FADD), Dest) |
12455 | .addReg(Src1) |
12456 | .addReg(Src2); |
12457 | if (MI.getFlag(MachineInstr::NoFPExcept)) |
12458 | MIB.setMIFlag(MachineInstr::NoFPExcept); |
12459 | |
12460 | // Restore FPSCR value. |
12461 | BuildMI(*BB, MI, dl, TII->get(PPC::MTFSFb)).addImm(1).addReg(MFFSReg); |
12462 | } else if (MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT || |
12463 | MI.getOpcode() == PPC::ANDI_rec_1_GT_BIT || |
12464 | MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT8 || |
12465 | MI.getOpcode() == PPC::ANDI_rec_1_GT_BIT8) { |
12466 | unsigned Opcode = (MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT8 || |
12467 | MI.getOpcode() == PPC::ANDI_rec_1_GT_BIT8) |
12468 | ? PPC::ANDI8_rec |
12469 | : PPC::ANDI_rec; |
12470 | bool IsEQ = (MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT || |
12471 | MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT8); |
12472 | |
12473 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
12474 | Register Dest = RegInfo.createVirtualRegister( |
12475 | Opcode == PPC::ANDI_rec ? &PPC::GPRCRegClass : &PPC::G8RCRegClass); |
12476 | |
12477 | DebugLoc Dl = MI.getDebugLoc(); |
12478 | BuildMI(*BB, MI, Dl, TII->get(Opcode), Dest) |
12479 | .addReg(MI.getOperand(1).getReg()) |
12480 | .addImm(1); |
12481 | BuildMI(*BB, MI, Dl, TII->get(TargetOpcode::COPY), |
12482 | MI.getOperand(0).getReg()) |
12483 | .addReg(IsEQ ? PPC::CR0EQ : PPC::CR0GT); |
12484 | } else if (MI.getOpcode() == PPC::TCHECK_RET) { |
12485 | DebugLoc Dl = MI.getDebugLoc(); |
12486 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
12487 | Register CRReg = RegInfo.createVirtualRegister(&PPC::CRRCRegClass); |
12488 | BuildMI(*BB, MI, Dl, TII->get(PPC::TCHECK), CRReg); |
12489 | BuildMI(*BB, MI, Dl, TII->get(TargetOpcode::COPY), |
12490 | MI.getOperand(0).getReg()) |
12491 | .addReg(CRReg); |
12492 | } else if (MI.getOpcode() == PPC::TBEGIN_RET) { |
12493 | DebugLoc Dl = MI.getDebugLoc(); |
12494 | unsigned Imm = MI.getOperand(1).getImm(); |
12495 | BuildMI(*BB, MI, Dl, TII->get(PPC::TBEGIN)).addImm(Imm); |
12496 | BuildMI(*BB, MI, Dl, TII->get(TargetOpcode::COPY), |
12497 | MI.getOperand(0).getReg()) |
12498 | .addReg(PPC::CR0EQ); |
12499 | } else if (MI.getOpcode() == PPC::SETRNDi) { |
12500 | DebugLoc dl = MI.getDebugLoc(); |
12501 | Register OldFPSCRReg = MI.getOperand(0).getReg(); |
12502 | |
12503 | // Save FPSCR value. |
12504 | BuildMI(*BB, MI, dl, TII->get(PPC::MFFS), OldFPSCRReg); |
12505 | |
12506 | // The floating point rounding mode is in the bits 62:63 of FPCSR, and has |
12507 | // the following settings: |
12508 | // 00 Round to nearest |
12509 | // 01 Round to 0 |
12510 | // 10 Round to +inf |
12511 | // 11 Round to -inf |
12512 | |
12513 | // When the operand is immediate, using the two least significant bits of |
12514 | // the immediate to set the bits 62:63 of FPSCR. |
12515 | unsigned Mode = MI.getOperand(1).getImm(); |
12516 | BuildMI(*BB, MI, dl, TII->get((Mode & 1) ? PPC::MTFSB1 : PPC::MTFSB0)) |
12517 | .addImm(31) |
12518 | .addReg(PPC::RM, RegState::ImplicitDefine); |
12519 | |
12520 | BuildMI(*BB, MI, dl, TII->get((Mode & 2) ? PPC::MTFSB1 : PPC::MTFSB0)) |
12521 | .addImm(30) |
12522 | .addReg(PPC::RM, RegState::ImplicitDefine); |
12523 | } else if (MI.getOpcode() == PPC::SETRND) { |
12524 | DebugLoc dl = MI.getDebugLoc(); |
12525 | |
12526 | // Copy register from F8RCRegClass::SrcReg to G8RCRegClass::DestReg |
12527 | // or copy register from G8RCRegClass::SrcReg to F8RCRegClass::DestReg. |
12528 | // If the target doesn't have DirectMove, we should use stack to do the |
12529 | // conversion, because the target doesn't have the instructions like mtvsrd |
12530 | // or mfvsrd to do this conversion directly. |
12531 | auto copyRegFromG8RCOrF8RC = [&] (unsigned DestReg, unsigned SrcReg) { |
12532 | if (Subtarget.hasDirectMove()) { |
12533 | BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), DestReg) |
12534 | .addReg(SrcReg); |
12535 | } else { |
12536 | // Use stack to do the register copy. |
12537 | unsigned StoreOp = PPC::STD, LoadOp = PPC::LFD; |
12538 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
12539 | const TargetRegisterClass *RC = RegInfo.getRegClass(SrcReg); |
12540 | if (RC == &PPC::F8RCRegClass) { |
12541 | // Copy register from F8RCRegClass to G8RCRegclass. |
12542 | assert((RegInfo.getRegClass(DestReg) == &PPC::G8RCRegClass) &&(((RegInfo.getRegClass(DestReg) == &PPC::G8RCRegClass) && "Unsupported RegClass.") ? static_cast<void> (0) : __assert_fail ("(RegInfo.getRegClass(DestReg) == &PPC::G8RCRegClass) && \"Unsupported RegClass.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12543, __PRETTY_FUNCTION__)) |
12543 | "Unsupported RegClass.")(((RegInfo.getRegClass(DestReg) == &PPC::G8RCRegClass) && "Unsupported RegClass.") ? static_cast<void> (0) : __assert_fail ("(RegInfo.getRegClass(DestReg) == &PPC::G8RCRegClass) && \"Unsupported RegClass.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12543, __PRETTY_FUNCTION__)); |
12544 | |
12545 | StoreOp = PPC::STFD; |
12546 | LoadOp = PPC::LD; |
12547 | } else { |
12548 | // Copy register from G8RCRegClass to F8RCRegclass. |
12549 | assert((RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) &&(((RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) && (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) && "Unsupported RegClass.") ? static_cast<void> (0) : __assert_fail ("(RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) && (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) && \"Unsupported RegClass.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12551, __PRETTY_FUNCTION__)) |
12550 | (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) &&(((RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) && (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) && "Unsupported RegClass.") ? static_cast<void> (0) : __assert_fail ("(RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) && (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) && \"Unsupported RegClass.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12551, __PRETTY_FUNCTION__)) |
12551 | "Unsupported RegClass.")(((RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) && (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) && "Unsupported RegClass.") ? static_cast<void> (0) : __assert_fail ("(RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) && (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) && \"Unsupported RegClass.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12551, __PRETTY_FUNCTION__)); |
12552 | } |
12553 | |
12554 | MachineFrameInfo &MFI = F->getFrameInfo(); |
12555 | int FrameIdx = MFI.CreateStackObject(8, Align(8), false); |
12556 | |
12557 | MachineMemOperand *MMOStore = F->getMachineMemOperand( |
12558 | MachinePointerInfo::getFixedStack(*F, FrameIdx, 0), |
12559 | MachineMemOperand::MOStore, MFI.getObjectSize(FrameIdx), |
12560 | MFI.getObjectAlign(FrameIdx)); |
12561 | |
12562 | // Store the SrcReg into the stack. |
12563 | BuildMI(*BB, MI, dl, TII->get(StoreOp)) |
12564 | .addReg(SrcReg) |
12565 | .addImm(0) |
12566 | .addFrameIndex(FrameIdx) |
12567 | .addMemOperand(MMOStore); |
12568 | |
12569 | MachineMemOperand *MMOLoad = F->getMachineMemOperand( |
12570 | MachinePointerInfo::getFixedStack(*F, FrameIdx, 0), |
12571 | MachineMemOperand::MOLoad, MFI.getObjectSize(FrameIdx), |
12572 | MFI.getObjectAlign(FrameIdx)); |
12573 | |
12574 | // Load from the stack where SrcReg is stored, and save to DestReg, |
12575 | // so we have done the RegClass conversion from RegClass::SrcReg to |
12576 | // RegClass::DestReg. |
12577 | BuildMI(*BB, MI, dl, TII->get(LoadOp), DestReg) |
12578 | .addImm(0) |
12579 | .addFrameIndex(FrameIdx) |
12580 | .addMemOperand(MMOLoad); |
12581 | } |
12582 | }; |
12583 | |
12584 | Register OldFPSCRReg = MI.getOperand(0).getReg(); |
12585 | |
12586 | // Save FPSCR value. |
12587 | BuildMI(*BB, MI, dl, TII->get(PPC::MFFS), OldFPSCRReg); |
12588 | |
12589 | // When the operand is gprc register, use two least significant bits of the |
12590 | // register and mtfsf instruction to set the bits 62:63 of FPSCR. |
12591 | // |
12592 | // copy OldFPSCRTmpReg, OldFPSCRReg |
12593 | // (INSERT_SUBREG ExtSrcReg, (IMPLICIT_DEF ImDefReg), SrcOp, 1) |
12594 | // rldimi NewFPSCRTmpReg, ExtSrcReg, OldFPSCRReg, 0, 62 |
12595 | // copy NewFPSCRReg, NewFPSCRTmpReg |
12596 | // mtfsf 255, NewFPSCRReg |
12597 | MachineOperand SrcOp = MI.getOperand(1); |
12598 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
12599 | Register OldFPSCRTmpReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass); |
12600 | |
12601 | copyRegFromG8RCOrF8RC(OldFPSCRTmpReg, OldFPSCRReg); |
12602 | |
12603 | Register ImDefReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass); |
12604 | Register ExtSrcReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass); |
12605 | |
12606 | // The first operand of INSERT_SUBREG should be a register which has |
12607 | // subregisters, we only care about its RegClass, so we should use an |
12608 | // IMPLICIT_DEF register. |
12609 | BuildMI(*BB, MI, dl, TII->get(TargetOpcode::IMPLICIT_DEF), ImDefReg); |
12610 | BuildMI(*BB, MI, dl, TII->get(PPC::INSERT_SUBREG), ExtSrcReg) |
12611 | .addReg(ImDefReg) |
12612 | .add(SrcOp) |
12613 | .addImm(1); |
12614 | |
12615 | Register NewFPSCRTmpReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass); |
12616 | BuildMI(*BB, MI, dl, TII->get(PPC::RLDIMI), NewFPSCRTmpReg) |
12617 | .addReg(OldFPSCRTmpReg) |
12618 | .addReg(ExtSrcReg) |
12619 | .addImm(0) |
12620 | .addImm(62); |
12621 | |
12622 | Register NewFPSCRReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass); |
12623 | copyRegFromG8RCOrF8RC(NewFPSCRReg, NewFPSCRTmpReg); |
12624 | |
12625 | // The mask 255 means that put the 32:63 bits of NewFPSCRReg to the 32:63 |
12626 | // bits of FPSCR. |
12627 | BuildMI(*BB, MI, dl, TII->get(PPC::MTFSF)) |
12628 | .addImm(255) |
12629 | .addReg(NewFPSCRReg) |
12630 | .addImm(0) |
12631 | .addImm(0); |
12632 | } else if (MI.getOpcode() == PPC::SETFLM) { |
12633 | DebugLoc Dl = MI.getDebugLoc(); |
12634 | |
12635 | // Result of setflm is previous FPSCR content, so we need to save it first. |
12636 | Register OldFPSCRReg = MI.getOperand(0).getReg(); |
12637 | BuildMI(*BB, MI, Dl, TII->get(PPC::MFFS), OldFPSCRReg); |
12638 | |
12639 | // Put bits in 32:63 to FPSCR. |
12640 | Register NewFPSCRReg = MI.getOperand(1).getReg(); |
12641 | BuildMI(*BB, MI, Dl, TII->get(PPC::MTFSF)) |
12642 | .addImm(255) |
12643 | .addReg(NewFPSCRReg) |
12644 | .addImm(0) |
12645 | .addImm(0); |
12646 | } else if (MI.getOpcode() == PPC::PROBED_ALLOCA_32 || |
12647 | MI.getOpcode() == PPC::PROBED_ALLOCA_64) { |
12648 | return emitProbedAlloca(MI, BB); |
12649 | } else { |
12650 | llvm_unreachable("Unexpected instr type to insert")::llvm::llvm_unreachable_internal("Unexpected instr type to insert" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12650); |
12651 | } |
12652 | |
12653 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
12654 | return BB; |
12655 | } |
12656 | |
12657 | //===----------------------------------------------------------------------===// |
12658 | // Target Optimization Hooks |
12659 | //===----------------------------------------------------------------------===// |
12660 | |
12661 | static int getEstimateRefinementSteps(EVT VT, const PPCSubtarget &Subtarget) { |
12662 | // For the estimates, convergence is quadratic, so we essentially double the |
12663 | // number of digits correct after every iteration. For both FRE and FRSQRTE, |
12664 | // the minimum architected relative accuracy is 2^-5. When hasRecipPrec(), |
12665 | // this is 2^-14. IEEE float has 23 digits and double has 52 digits. |
12666 | int RefinementSteps = Subtarget.hasRecipPrec() ? 1 : 3; |
12667 | if (VT.getScalarType() == MVT::f64) |
12668 | RefinementSteps++; |
12669 | return RefinementSteps; |
12670 | } |
12671 | |
12672 | SDValue PPCTargetLowering::getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, |
12673 | int Enabled, int &RefinementSteps, |
12674 | bool &UseOneConstNR, |
12675 | bool Reciprocal) const { |
12676 | EVT VT = Operand.getValueType(); |
12677 | if ((VT == MVT::f32 && Subtarget.hasFRSQRTES()) || |
12678 | (VT == MVT::f64 && Subtarget.hasFRSQRTE()) || |
12679 | (VT == MVT::v4f32 && Subtarget.hasAltivec()) || |
12680 | (VT == MVT::v2f64 && Subtarget.hasVSX())) { |
12681 | if (RefinementSteps == ReciprocalEstimate::Unspecified) |
12682 | RefinementSteps = getEstimateRefinementSteps(VT, Subtarget); |
12683 | |
12684 | // The Newton-Raphson computation with a single constant does not provide |
12685 | // enough accuracy on some CPUs. |
12686 | UseOneConstNR = !Subtarget.needsTwoConstNR(); |
12687 | return DAG.getNode(PPCISD::FRSQRTE, SDLoc(Operand), VT, Operand); |
12688 | } |
12689 | return SDValue(); |
12690 | } |
12691 | |
12692 | SDValue PPCTargetLowering::getRecipEstimate(SDValue Operand, SelectionDAG &DAG, |
12693 | int Enabled, |
12694 | int &RefinementSteps) const { |
12695 | EVT VT = Operand.getValueType(); |
12696 | if ((VT == MVT::f32 && Subtarget.hasFRES()) || |
12697 | (VT == MVT::f64 && Subtarget.hasFRE()) || |
12698 | (VT == MVT::v4f32 && Subtarget.hasAltivec()) || |
12699 | (VT == MVT::v2f64 && Subtarget.hasVSX())) { |
12700 | if (RefinementSteps == ReciprocalEstimate::Unspecified) |
12701 | RefinementSteps = getEstimateRefinementSteps(VT, Subtarget); |
12702 | return DAG.getNode(PPCISD::FRE, SDLoc(Operand), VT, Operand); |
12703 | } |
12704 | return SDValue(); |
12705 | } |
12706 | |
12707 | unsigned PPCTargetLowering::combineRepeatedFPDivisors() const { |
12708 | // Note: This functionality is used only when unsafe-fp-math is enabled, and |
12709 | // on cores with reciprocal estimates (which are used when unsafe-fp-math is |
12710 | // enabled for division), this functionality is redundant with the default |
12711 | // combiner logic (once the division -> reciprocal/multiply transformation |
12712 | // has taken place). As a result, this matters more for older cores than for |
12713 | // newer ones. |
12714 | |
12715 | // Combine multiple FDIVs with the same divisor into multiple FMULs by the |
12716 | // reciprocal if there are two or more FDIVs (for embedded cores with only |
12717 | // one FP pipeline) for three or more FDIVs (for generic OOO cores). |
12718 | switch (Subtarget.getCPUDirective()) { |
12719 | default: |
12720 | return 3; |
12721 | case PPC::DIR_440: |
12722 | case PPC::DIR_A2: |
12723 | case PPC::DIR_E500: |
12724 | case PPC::DIR_E500mc: |
12725 | case PPC::DIR_E5500: |
12726 | return 2; |
12727 | } |
12728 | } |
12729 | |
12730 | // isConsecutiveLSLoc needs to work even if all adds have not yet been |
12731 | // collapsed, and so we need to look through chains of them. |
12732 | static void getBaseWithConstantOffset(SDValue Loc, SDValue &Base, |
12733 | int64_t& Offset, SelectionDAG &DAG) { |
12734 | if (DAG.isBaseWithConstantOffset(Loc)) { |
12735 | Base = Loc.getOperand(0); |
12736 | Offset += cast<ConstantSDNode>(Loc.getOperand(1))->getSExtValue(); |
12737 | |
12738 | // The base might itself be a base plus an offset, and if so, accumulate |
12739 | // that as well. |
12740 | getBaseWithConstantOffset(Loc.getOperand(0), Base, Offset, DAG); |
12741 | } |
12742 | } |
12743 | |
12744 | static bool isConsecutiveLSLoc(SDValue Loc, EVT VT, LSBaseSDNode *Base, |
12745 | unsigned Bytes, int Dist, |
12746 | SelectionDAG &DAG) { |
12747 | if (VT.getSizeInBits() / 8 != Bytes) |
12748 | return false; |
12749 | |
12750 | SDValue BaseLoc = Base->getBasePtr(); |
12751 | if (Loc.getOpcode() == ISD::FrameIndex) { |
12752 | if (BaseLoc.getOpcode() != ISD::FrameIndex) |
12753 | return false; |
12754 | const MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
12755 | int FI = cast<FrameIndexSDNode>(Loc)->getIndex(); |
12756 | int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex(); |
12757 | int FS = MFI.getObjectSize(FI); |
12758 | int BFS = MFI.getObjectSize(BFI); |
12759 | if (FS != BFS || FS != (int)Bytes) return false; |
12760 | return MFI.getObjectOffset(FI) == (MFI.getObjectOffset(BFI) + Dist*Bytes); |
12761 | } |
12762 | |
12763 | SDValue Base1 = Loc, Base2 = BaseLoc; |
12764 | int64_t Offset1 = 0, Offset2 = 0; |
12765 | getBaseWithConstantOffset(Loc, Base1, Offset1, DAG); |
12766 | getBaseWithConstantOffset(BaseLoc, Base2, Offset2, DAG); |
12767 | if (Base1 == Base2 && Offset1 == (Offset2 + Dist * Bytes)) |
12768 | return true; |
12769 | |
12770 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
12771 | const GlobalValue *GV1 = nullptr; |
12772 | const GlobalValue *GV2 = nullptr; |
12773 | Offset1 = 0; |
12774 | Offset2 = 0; |
12775 | bool isGA1 = TLI.isGAPlusOffset(Loc.getNode(), GV1, Offset1); |
12776 | bool isGA2 = TLI.isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2); |
12777 | if (isGA1 && isGA2 && GV1 == GV2) |
12778 | return Offset1 == (Offset2 + Dist*Bytes); |
12779 | return false; |
12780 | } |
12781 | |
12782 | // Like SelectionDAG::isConsecutiveLoad, but also works for stores, and does |
12783 | // not enforce equality of the chain operands. |
12784 | static bool isConsecutiveLS(SDNode *N, LSBaseSDNode *Base, |
12785 | unsigned Bytes, int Dist, |
12786 | SelectionDAG &DAG) { |
12787 | if (LSBaseSDNode *LS = dyn_cast<LSBaseSDNode>(N)) { |
12788 | EVT VT = LS->getMemoryVT(); |
12789 | SDValue Loc = LS->getBasePtr(); |
12790 | return isConsecutiveLSLoc(Loc, VT, Base, Bytes, Dist, DAG); |
12791 | } |
12792 | |
12793 | if (N->getOpcode() == ISD::INTRINSIC_W_CHAIN) { |
12794 | EVT VT; |
12795 | switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) { |
12796 | default: return false; |
12797 | case Intrinsic::ppc_altivec_lvx: |
12798 | case Intrinsic::ppc_altivec_lvxl: |
12799 | case Intrinsic::ppc_vsx_lxvw4x: |
12800 | case Intrinsic::ppc_vsx_lxvw4x_be: |
12801 | VT = MVT::v4i32; |
12802 | break; |
12803 | case Intrinsic::ppc_vsx_lxvd2x: |
12804 | case Intrinsic::ppc_vsx_lxvd2x_be: |
12805 | VT = MVT::v2f64; |
12806 | break; |
12807 | case Intrinsic::ppc_altivec_lvebx: |
12808 | VT = MVT::i8; |
12809 | break; |
12810 | case Intrinsic::ppc_altivec_lvehx: |
12811 | VT = MVT::i16; |
12812 | break; |
12813 | case Intrinsic::ppc_altivec_lvewx: |
12814 | VT = MVT::i32; |
12815 | break; |
12816 | } |
12817 | |
12818 | return isConsecutiveLSLoc(N->getOperand(2), VT, Base, Bytes, Dist, DAG); |
12819 | } |
12820 | |
12821 | if (N->getOpcode() == ISD::INTRINSIC_VOID) { |
12822 | EVT VT; |
12823 | switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) { |
12824 | default: return false; |
12825 | case Intrinsic::ppc_altivec_stvx: |
12826 | case Intrinsic::ppc_altivec_stvxl: |
12827 | case Intrinsic::ppc_vsx_stxvw4x: |
12828 | VT = MVT::v4i32; |
12829 | break; |
12830 | case Intrinsic::ppc_vsx_stxvd2x: |
12831 | VT = MVT::v2f64; |
12832 | break; |
12833 | case Intrinsic::ppc_vsx_stxvw4x_be: |
12834 | VT = MVT::v4i32; |
12835 | break; |
12836 | case Intrinsic::ppc_vsx_stxvd2x_be: |
12837 | VT = MVT::v2f64; |
12838 | break; |
12839 | case Intrinsic::ppc_altivec_stvebx: |
12840 | VT = MVT::i8; |
12841 | break; |
12842 | case Intrinsic::ppc_altivec_stvehx: |
12843 | VT = MVT::i16; |
12844 | break; |
12845 | case Intrinsic::ppc_altivec_stvewx: |
12846 | VT = MVT::i32; |
12847 | break; |
12848 | } |
12849 | |
12850 | return isConsecutiveLSLoc(N->getOperand(3), VT, Base, Bytes, Dist, DAG); |
12851 | } |
12852 | |
12853 | return false; |
12854 | } |
12855 | |
12856 | // Return true is there is a nearyby consecutive load to the one provided |
12857 | // (regardless of alignment). We search up and down the chain, looking though |
12858 | // token factors and other loads (but nothing else). As a result, a true result |
12859 | // indicates that it is safe to create a new consecutive load adjacent to the |
12860 | // load provided. |
12861 | static bool findConsecutiveLoad(LoadSDNode *LD, SelectionDAG &DAG) { |
12862 | SDValue Chain = LD->getChain(); |
12863 | EVT VT = LD->getMemoryVT(); |
12864 | |
12865 | SmallSet<SDNode *, 16> LoadRoots; |
12866 | SmallVector<SDNode *, 8> Queue(1, Chain.getNode()); |
12867 | SmallSet<SDNode *, 16> Visited; |
12868 | |
12869 | // First, search up the chain, branching to follow all token-factor operands. |
12870 | // If we find a consecutive load, then we're done, otherwise, record all |
12871 | // nodes just above the top-level loads and token factors. |
12872 | while (!Queue.empty()) { |
12873 | SDNode *ChainNext = Queue.pop_back_val(); |
12874 | if (!Visited.insert(ChainNext).second) |
12875 | continue; |
12876 | |
12877 | if (MemSDNode *ChainLD = dyn_cast<MemSDNode>(ChainNext)) { |
12878 | if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG)) |
12879 | return true; |
12880 | |
12881 | if (!Visited.count(ChainLD->getChain().getNode())) |
12882 | Queue.push_back(ChainLD->getChain().getNode()); |
12883 | } else if (ChainNext->getOpcode() == ISD::TokenFactor) { |
12884 | for (const SDUse &O : ChainNext->ops()) |
12885 | if (!Visited.count(O.getNode())) |
12886 | Queue.push_back(O.getNode()); |
12887 | } else |
12888 | LoadRoots.insert(ChainNext); |
12889 | } |
12890 | |
12891 | // Second, search down the chain, starting from the top-level nodes recorded |
12892 | // in the first phase. These top-level nodes are the nodes just above all |
12893 | // loads and token factors. Starting with their uses, recursively look though |
12894 | // all loads (just the chain uses) and token factors to find a consecutive |
12895 | // load. |
12896 | Visited.clear(); |
12897 | Queue.clear(); |
12898 | |
12899 | for (SmallSet<SDNode *, 16>::iterator I = LoadRoots.begin(), |
12900 | IE = LoadRoots.end(); I != IE; ++I) { |
12901 | Queue.push_back(*I); |
12902 | |
12903 | while (!Queue.empty()) { |
12904 | SDNode *LoadRoot = Queue.pop_back_val(); |
12905 | if (!Visited.insert(LoadRoot).second) |
12906 | continue; |
12907 | |
12908 | if (MemSDNode *ChainLD = dyn_cast<MemSDNode>(LoadRoot)) |
12909 | if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG)) |
12910 | return true; |
12911 | |
12912 | for (SDNode::use_iterator UI = LoadRoot->use_begin(), |
12913 | UE = LoadRoot->use_end(); UI != UE; ++UI) |
12914 | if (((isa<MemSDNode>(*UI) && |
12915 | cast<MemSDNode>(*UI)->getChain().getNode() == LoadRoot) || |
12916 | UI->getOpcode() == ISD::TokenFactor) && !Visited.count(*UI)) |
12917 | Queue.push_back(*UI); |
12918 | } |
12919 | } |
12920 | |
12921 | return false; |
12922 | } |
12923 | |
12924 | /// This function is called when we have proved that a SETCC node can be replaced |
12925 | /// by subtraction (and other supporting instructions) so that the result of |
12926 | /// comparison is kept in a GPR instead of CR. This function is purely for |
12927 | /// codegen purposes and has some flags to guide the codegen process. |
12928 | static SDValue generateEquivalentSub(SDNode *N, int Size, bool Complement, |
12929 | bool Swap, SDLoc &DL, SelectionDAG &DAG) { |
12930 | assert(N->getOpcode() == ISD::SETCC && "ISD::SETCC Expected.")((N->getOpcode() == ISD::SETCC && "ISD::SETCC Expected." ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::SETCC && \"ISD::SETCC Expected.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12930, __PRETTY_FUNCTION__)); |
12931 | |
12932 | // Zero extend the operands to the largest legal integer. Originally, they |
12933 | // must be of a strictly smaller size. |
12934 | auto Op0 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, N->getOperand(0), |
12935 | DAG.getConstant(Size, DL, MVT::i32)); |
12936 | auto Op1 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, N->getOperand(1), |
12937 | DAG.getConstant(Size, DL, MVT::i32)); |
12938 | |
12939 | // Swap if needed. Depends on the condition code. |
12940 | if (Swap) |
12941 | std::swap(Op0, Op1); |
12942 | |
12943 | // Subtract extended integers. |
12944 | auto SubNode = DAG.getNode(ISD::SUB, DL, MVT::i64, Op0, Op1); |
12945 | |
12946 | // Move the sign bit to the least significant position and zero out the rest. |
12947 | // Now the least significant bit carries the result of original comparison. |
12948 | auto Shifted = DAG.getNode(ISD::SRL, DL, MVT::i64, SubNode, |
12949 | DAG.getConstant(Size - 1, DL, MVT::i32)); |
12950 | auto Final = Shifted; |
12951 | |
12952 | // Complement the result if needed. Based on the condition code. |
12953 | if (Complement) |
12954 | Final = DAG.getNode(ISD::XOR, DL, MVT::i64, Shifted, |
12955 | DAG.getConstant(1, DL, MVT::i64)); |
12956 | |
12957 | return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Final); |
12958 | } |
12959 | |
12960 | SDValue PPCTargetLowering::ConvertSETCCToSubtract(SDNode *N, |
12961 | DAGCombinerInfo &DCI) const { |
12962 | assert(N->getOpcode() == ISD::SETCC && "ISD::SETCC Expected.")((N->getOpcode() == ISD::SETCC && "ISD::SETCC Expected." ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::SETCC && \"ISD::SETCC Expected.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12962, __PRETTY_FUNCTION__)); |
12963 | |
12964 | SelectionDAG &DAG = DCI.DAG; |
12965 | SDLoc DL(N); |
12966 | |
12967 | // Size of integers being compared has a critical role in the following |
12968 | // analysis, so we prefer to do this when all types are legal. |
12969 | if (!DCI.isAfterLegalizeDAG()) |
12970 | return SDValue(); |
12971 | |
12972 | // If all users of SETCC extend its value to a legal integer type |
12973 | // then we replace SETCC with a subtraction |
12974 | for (SDNode::use_iterator UI = N->use_begin(), |
12975 | UE = N->use_end(); UI != UE; ++UI) { |
12976 | if (UI->getOpcode() != ISD::ZERO_EXTEND) |
12977 | return SDValue(); |
12978 | } |
12979 | |
12980 | ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get(); |
12981 | auto OpSize = N->getOperand(0).getValueSizeInBits(); |
12982 | |
12983 | unsigned Size = DAG.getDataLayout().getLargestLegalIntTypeSizeInBits(); |
12984 | |
12985 | if (OpSize < Size) { |
12986 | switch (CC) { |
12987 | default: break; |
12988 | case ISD::SETULT: |
12989 | return generateEquivalentSub(N, Size, false, false, DL, DAG); |
12990 | case ISD::SETULE: |
12991 | return generateEquivalentSub(N, Size, true, true, DL, DAG); |
12992 | case ISD::SETUGT: |
12993 | return generateEquivalentSub(N, Size, false, true, DL, DAG); |
12994 | case ISD::SETUGE: |
12995 | return generateEquivalentSub(N, Size, true, false, DL, DAG); |
12996 | } |
12997 | } |
12998 | |
12999 | return SDValue(); |
13000 | } |
13001 | |
13002 | SDValue PPCTargetLowering::DAGCombineTruncBoolExt(SDNode *N, |
13003 | DAGCombinerInfo &DCI) const { |
13004 | SelectionDAG &DAG = DCI.DAG; |
13005 | SDLoc dl(N); |
13006 | |
13007 | assert(Subtarget.useCRBits() && "Expecting to be tracking CR bits")((Subtarget.useCRBits() && "Expecting to be tracking CR bits" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.useCRBits() && \"Expecting to be tracking CR bits\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13007, __PRETTY_FUNCTION__)); |
13008 | // If we're tracking CR bits, we need to be careful that we don't have: |
13009 | // trunc(binary-ops(zext(x), zext(y))) |
13010 | // or |
13011 | // trunc(binary-ops(binary-ops(zext(x), zext(y)), ...) |
13012 | // such that we're unnecessarily moving things into GPRs when it would be |
13013 | // better to keep them in CR bits. |
13014 | |
13015 | // Note that trunc here can be an actual i1 trunc, or can be the effective |
13016 | // truncation that comes from a setcc or select_cc. |
13017 | if (N->getOpcode() == ISD::TRUNCATE && |
13018 | N->getValueType(0) != MVT::i1) |
13019 | return SDValue(); |
13020 | |
13021 | if (N->getOperand(0).getValueType() != MVT::i32 && |
13022 | N->getOperand(0).getValueType() != MVT::i64) |
13023 | return SDValue(); |
13024 | |
13025 | if (N->getOpcode() == ISD::SETCC || |
13026 | N->getOpcode() == ISD::SELECT_CC) { |
13027 | // If we're looking at a comparison, then we need to make sure that the |
13028 | // high bits (all except for the first) don't matter the result. |
13029 | ISD::CondCode CC = |
13030 | cast<CondCodeSDNode>(N->getOperand( |
13031 | N->getOpcode() == ISD::SETCC ? 2 : 4))->get(); |
13032 | unsigned OpBits = N->getOperand(0).getValueSizeInBits(); |
13033 | |
13034 | if (ISD::isSignedIntSetCC(CC)) { |
13035 | if (DAG.ComputeNumSignBits(N->getOperand(0)) != OpBits || |
13036 | DAG.ComputeNumSignBits(N->getOperand(1)) != OpBits) |
13037 | return SDValue(); |
13038 | } else if (ISD::isUnsignedIntSetCC(CC)) { |
13039 | if (!DAG.MaskedValueIsZero(N->getOperand(0), |
13040 | APInt::getHighBitsSet(OpBits, OpBits-1)) || |
13041 | !DAG.MaskedValueIsZero(N->getOperand(1), |
13042 | APInt::getHighBitsSet(OpBits, OpBits-1))) |
13043 | return (N->getOpcode() == ISD::SETCC ? ConvertSETCCToSubtract(N, DCI) |
13044 | : SDValue()); |
13045 | } else { |
13046 | // This is neither a signed nor an unsigned comparison, just make sure |
13047 | // that the high bits are equal. |
13048 | KnownBits Op1Known = DAG.computeKnownBits(N->getOperand(0)); |
13049 | KnownBits Op2Known = DAG.computeKnownBits(N->getOperand(1)); |
13050 | |
13051 | // We don't really care about what is known about the first bit (if |
13052 | // anything), so clear it in all masks prior to comparing them. |
13053 | Op1Known.Zero.clearBit(0); Op1Known.One.clearBit(0); |
13054 | Op2Known.Zero.clearBit(0); Op2Known.One.clearBit(0); |
13055 | |
13056 | if (Op1Known.Zero != Op2Known.Zero || Op1Known.One != Op2Known.One) |
13057 | return SDValue(); |
13058 | } |
13059 | } |
13060 | |
13061 | // We now know that the higher-order bits are irrelevant, we just need to |
13062 | // make sure that all of the intermediate operations are bit operations, and |
13063 | // all inputs are extensions. |
13064 | if (N->getOperand(0).getOpcode() != ISD::AND && |
13065 | N->getOperand(0).getOpcode() != ISD::OR && |
13066 | N->getOperand(0).getOpcode() != ISD::XOR && |
13067 | N->getOperand(0).getOpcode() != ISD::SELECT && |
13068 | N->getOperand(0).getOpcode() != ISD::SELECT_CC && |
13069 | N->getOperand(0).getOpcode() != ISD::TRUNCATE && |
13070 | N->getOperand(0).getOpcode() != ISD::SIGN_EXTEND && |
13071 | N->getOperand(0).getOpcode() != ISD::ZERO_EXTEND && |
13072 | N->getOperand(0).getOpcode() != ISD::ANY_EXTEND) |
13073 | return SDValue(); |
13074 | |
13075 | if ((N->getOpcode() == ISD::SETCC || N->getOpcode() == ISD::SELECT_CC) && |
13076 | N->getOperand(1).getOpcode() != ISD::AND && |
13077 | N->getOperand(1).getOpcode() != ISD::OR && |
13078 | N->getOperand(1).getOpcode() != ISD::XOR && |
13079 | N->getOperand(1).getOpcode() != ISD::SELECT && |
13080 | N->getOperand(1).getOpcode() != ISD::SELECT_CC && |
13081 | N->getOperand(1).getOpcode() != ISD::TRUNCATE && |
13082 | N->getOperand(1).getOpcode() != ISD::SIGN_EXTEND && |
13083 | N->getOperand(1).getOpcode() != ISD::ZERO_EXTEND && |
13084 | N->getOperand(1).getOpcode() != ISD::ANY_EXTEND) |
13085 | return SDValue(); |
13086 | |
13087 | SmallVector<SDValue, 4> Inputs; |
13088 | SmallVector<SDValue, 8> BinOps, PromOps; |
13089 | SmallPtrSet<SDNode *, 16> Visited; |
13090 | |
13091 | for (unsigned i = 0; i < 2; ++i) { |
13092 | if (((N->getOperand(i).getOpcode() == ISD::SIGN_EXTEND || |
13093 | N->getOperand(i).getOpcode() == ISD::ZERO_EXTEND || |
13094 | N->getOperand(i).getOpcode() == ISD::ANY_EXTEND) && |
13095 | N->getOperand(i).getOperand(0).getValueType() == MVT::i1) || |
13096 | isa<ConstantSDNode>(N->getOperand(i))) |
13097 | Inputs.push_back(N->getOperand(i)); |
13098 | else |
13099 | BinOps.push_back(N->getOperand(i)); |
13100 | |
13101 | if (N->getOpcode() == ISD::TRUNCATE) |
13102 | break; |
13103 | } |
13104 | |
13105 | // Visit all inputs, collect all binary operations (and, or, xor and |
13106 | // select) that are all fed by extensions. |
13107 | while (!BinOps.empty()) { |
13108 | SDValue BinOp = BinOps.back(); |
13109 | BinOps.pop_back(); |
13110 | |
13111 | if (!Visited.insert(BinOp.getNode()).second) |
13112 | continue; |
13113 | |
13114 | PromOps.push_back(BinOp); |
13115 | |
13116 | for (unsigned i = 0, ie = BinOp.getNumOperands(); i != ie; ++i) { |
13117 | // The condition of the select is not promoted. |
13118 | if (BinOp.getOpcode() == ISD::SELECT && i == 0) |
13119 | continue; |
13120 | if (BinOp.getOpcode() == ISD::SELECT_CC && i != 2 && i != 3) |
13121 | continue; |
13122 | |
13123 | if (((BinOp.getOperand(i).getOpcode() == ISD::SIGN_EXTEND || |
13124 | BinOp.getOperand(i).getOpcode() == ISD::ZERO_EXTEND || |
13125 | BinOp.getOperand(i).getOpcode() == ISD::ANY_EXTEND) && |
13126 | BinOp.getOperand(i).getOperand(0).getValueType() == MVT::i1) || |
13127 | isa<ConstantSDNode>(BinOp.getOperand(i))) { |
13128 | Inputs.push_back(BinOp.getOperand(i)); |
13129 | } else if (BinOp.getOperand(i).getOpcode() == ISD::AND || |
13130 | BinOp.getOperand(i).getOpcode() == ISD::OR || |
13131 | BinOp.getOperand(i).getOpcode() == ISD::XOR || |
13132 | BinOp.getOperand(i).getOpcode() == ISD::SELECT || |
13133 | BinOp.getOperand(i).getOpcode() == ISD::SELECT_CC || |
13134 | BinOp.getOperand(i).getOpcode() == ISD::TRUNCATE || |
13135 | BinOp.getOperand(i).getOpcode() == ISD::SIGN_EXTEND || |
13136 | BinOp.getOperand(i).getOpcode() == ISD::ZERO_EXTEND || |
13137 | BinOp.getOperand(i).getOpcode() == ISD::ANY_EXTEND) { |
13138 | BinOps.push_back(BinOp.getOperand(i)); |
13139 | } else { |
13140 | // We have an input that is not an extension or another binary |
13141 | // operation; we'll abort this transformation. |
13142 | return SDValue(); |
13143 | } |
13144 | } |
13145 | } |
13146 | |
13147 | // Make sure that this is a self-contained cluster of operations (which |
13148 | // is not quite the same thing as saying that everything has only one |
13149 | // use). |
13150 | for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) { |
13151 | if (isa<ConstantSDNode>(Inputs[i])) |
13152 | continue; |
13153 | |
13154 | for (SDNode::use_iterator UI = Inputs[i].getNode()->use_begin(), |
13155 | UE = Inputs[i].getNode()->use_end(); |
13156 | UI != UE; ++UI) { |
13157 | SDNode *User = *UI; |
13158 | if (User != N && !Visited.count(User)) |
13159 | return SDValue(); |
13160 | |
13161 | // Make sure that we're not going to promote the non-output-value |
13162 | // operand(s) or SELECT or SELECT_CC. |
13163 | // FIXME: Although we could sometimes handle this, and it does occur in |
13164 | // practice that one of the condition inputs to the select is also one of |
13165 | // the outputs, we currently can't deal with this. |
13166 | if (User->getOpcode() == ISD::SELECT) { |
13167 | if (User->getOperand(0) == Inputs[i]) |
13168 | return SDValue(); |
13169 | } else if (User->getOpcode() == ISD::SELECT_CC) { |
13170 | if (User->getOperand(0) == Inputs[i] || |
13171 | User->getOperand(1) == Inputs[i]) |
13172 | return SDValue(); |
13173 | } |
13174 | } |
13175 | } |
13176 | |
13177 | for (unsigned i = 0, ie = PromOps.size(); i != ie; ++i) { |
13178 | for (SDNode::use_iterator UI = PromOps[i].getNode()->use_begin(), |
13179 | UE = PromOps[i].getNode()->use_end(); |
13180 | UI != UE; ++UI) { |
13181 | SDNode *User = *UI; |
13182 | if (User != N && !Visited.count(User)) |
13183 | return SDValue(); |
13184 | |
13185 | // Make sure that we're not going to promote the non-output-value |
13186 | // operand(s) or SELECT or SELECT_CC. |
13187 | // FIXME: Although we could sometimes handle this, and it does occur in |
13188 | // practice that one of the condition inputs to the select is also one of |
13189 | // the outputs, we currently can't deal with this. |
13190 | if (User->getOpcode() == ISD::SELECT) { |
13191 | if (User->getOperand(0) == PromOps[i]) |
13192 | return SDValue(); |
13193 | } else if (User->getOpcode() == ISD::SELECT_CC) { |
13194 | if (User->getOperand(0) == PromOps[i] || |
13195 | User->getOperand(1) == PromOps[i]) |
13196 | return SDValue(); |
13197 | } |
13198 | } |
13199 | } |
13200 | |
13201 | // Replace all inputs with the extension operand. |
13202 | for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) { |
13203 | // Constants may have users outside the cluster of to-be-promoted nodes, |
13204 | // and so we need to replace those as we do the promotions. |
13205 | if (isa<ConstantSDNode>(Inputs[i])) |
13206 | continue; |
13207 | else |
13208 | DAG.ReplaceAllUsesOfValueWith(Inputs[i], Inputs[i].getOperand(0)); |
13209 | } |
13210 | |
13211 | std::list<HandleSDNode> PromOpHandles; |
13212 | for (auto &PromOp : PromOps) |
13213 | PromOpHandles.emplace_back(PromOp); |
13214 | |
13215 | // Replace all operations (these are all the same, but have a different |
13216 | // (i1) return type). DAG.getNode will validate that the types of |
13217 | // a binary operator match, so go through the list in reverse so that |
13218 | // we've likely promoted both operands first. Any intermediate truncations or |
13219 | // extensions disappear. |
13220 | while (!PromOpHandles.empty()) { |
13221 | SDValue PromOp = PromOpHandles.back().getValue(); |
13222 | PromOpHandles.pop_back(); |
13223 | |
13224 | if (PromOp.getOpcode() == ISD::TRUNCATE || |
13225 | PromOp.getOpcode() == ISD::SIGN_EXTEND || |
13226 | PromOp.getOpcode() == ISD::ZERO_EXTEND || |
13227 | PromOp.getOpcode() == ISD::ANY_EXTEND) { |
13228 | if (!isa<ConstantSDNode>(PromOp.getOperand(0)) && |
13229 | PromOp.getOperand(0).getValueType() != MVT::i1) { |
13230 | // The operand is not yet ready (see comment below). |
13231 | PromOpHandles.emplace_front(PromOp); |
13232 | continue; |
13233 | } |
13234 | |
13235 | SDValue RepValue = PromOp.getOperand(0); |
13236 | if (isa<ConstantSDNode>(RepValue)) |
13237 | RepValue = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, RepValue); |
13238 | |
13239 | DAG.ReplaceAllUsesOfValueWith(PromOp, RepValue); |
13240 | continue; |
13241 | } |
13242 | |
13243 | unsigned C; |
13244 | switch (PromOp.getOpcode()) { |
13245 | default: C = 0; break; |
13246 | case ISD::SELECT: C = 1; break; |
13247 | case ISD::SELECT_CC: C = 2; break; |
13248 | } |
13249 | |
13250 | if ((!isa<ConstantSDNode>(PromOp.getOperand(C)) && |
13251 | PromOp.getOperand(C).getValueType() != MVT::i1) || |
13252 | (!isa<ConstantSDNode>(PromOp.getOperand(C+1)) && |
13253 | PromOp.getOperand(C+1).getValueType() != MVT::i1)) { |
13254 | // The to-be-promoted operands of this node have not yet been |
13255 | // promoted (this should be rare because we're going through the |
13256 | // list backward, but if one of the operands has several users in |
13257 | // this cluster of to-be-promoted nodes, it is possible). |
13258 | PromOpHandles.emplace_front(PromOp); |
13259 | continue; |
13260 | } |
13261 | |
13262 | SmallVector<SDValue, 3> Ops(PromOp.getNode()->op_begin(), |
13263 | PromOp.getNode()->op_end()); |
13264 | |
13265 | // If there are any constant inputs, make sure they're replaced now. |
13266 | for (unsigned i = 0; i < 2; ++i) |
13267 | if (isa<ConstantSDNode>(Ops[C+i])) |
13268 | Ops[C+i] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ops[C+i]); |
13269 | |
13270 | DAG.ReplaceAllUsesOfValueWith(PromOp, |
13271 | DAG.getNode(PromOp.getOpcode(), dl, MVT::i1, Ops)); |
13272 | } |
13273 | |
13274 | // Now we're left with the initial truncation itself. |
13275 | if (N->getOpcode() == ISD::TRUNCATE) |
13276 | return N->getOperand(0); |
13277 | |
13278 | // Otherwise, this is a comparison. The operands to be compared have just |
13279 | // changed type (to i1), but everything else is the same. |
13280 | return SDValue(N, 0); |
13281 | } |
13282 | |
13283 | SDValue PPCTargetLowering::DAGCombineExtBoolTrunc(SDNode *N, |
13284 | DAGCombinerInfo &DCI) const { |
13285 | SelectionDAG &DAG = DCI.DAG; |
13286 | SDLoc dl(N); |
13287 | |
13288 | // If we're tracking CR bits, we need to be careful that we don't have: |
13289 | // zext(binary-ops(trunc(x), trunc(y))) |
13290 | // or |
13291 | // zext(binary-ops(binary-ops(trunc(x), trunc(y)), ...) |
13292 | // such that we're unnecessarily moving things into CR bits that can more |
13293 | // efficiently stay in GPRs. Note that if we're not certain that the high |
13294 | // bits are set as required by the final extension, we still may need to do |
13295 | // some masking to get the proper behavior. |
13296 | |
13297 | // This same functionality is important on PPC64 when dealing with |
13298 | // 32-to-64-bit extensions; these occur often when 32-bit values are used as |
13299 | // the return values of functions. Because it is so similar, it is handled |
13300 | // here as well. |
13301 | |
13302 | if (N->getValueType(0) != MVT::i32 && |
13303 | N->getValueType(0) != MVT::i64) |
13304 | return SDValue(); |
13305 | |
13306 | if (!((N->getOperand(0).getValueType() == MVT::i1 && Subtarget.useCRBits()) || |
13307 | (N->getOperand(0).getValueType() == MVT::i32 && Subtarget.isPPC64()))) |
13308 | return SDValue(); |
13309 | |
13310 | if (N->getOperand(0).getOpcode() != ISD::AND && |
13311 | N->getOperand(0).getOpcode() != ISD::OR && |
13312 | N->getOperand(0).getOpcode() != ISD::XOR && |
13313 | N->getOperand(0).getOpcode() != ISD::SELECT && |
13314 | N->getOperand(0).getOpcode() != ISD::SELECT_CC) |
13315 | return SDValue(); |
13316 | |
13317 | SmallVector<SDValue, 4> Inputs; |
13318 | SmallVector<SDValue, 8> BinOps(1, N->getOperand(0)), PromOps; |
13319 | SmallPtrSet<SDNode *, 16> Visited; |
13320 | |
13321 | // Visit all inputs, collect all binary operations (and, or, xor and |
13322 | // select) that are all fed by truncations. |
13323 | while (!BinOps.empty()) { |
13324 | SDValue BinOp = BinOps.back(); |
13325 | BinOps.pop_back(); |
13326 | |
13327 | if (!Visited.insert(BinOp.getNode()).second) |
13328 | continue; |
13329 | |
13330 | PromOps.push_back(BinOp); |
13331 | |
13332 | for (unsigned i = 0, ie = BinOp.getNumOperands(); i != ie; ++i) { |
13333 | // The condition of the select is not promoted. |
13334 | if (BinOp.getOpcode() == ISD::SELECT && i == 0) |
13335 | continue; |
13336 | if (BinOp.getOpcode() == ISD::SELECT_CC && i != 2 && i != 3) |
13337 | continue; |
13338 | |
13339 | if (BinOp.getOperand(i).getOpcode() == ISD::TRUNCATE || |
13340 | isa<ConstantSDNode>(BinOp.getOperand(i))) { |
13341 | Inputs.push_back(BinOp.getOperand(i)); |
13342 | } else if (BinOp.getOperand(i).getOpcode() == ISD::AND || |
13343 | BinOp.getOperand(i).getOpcode() == ISD::OR || |
13344 | BinOp.getOperand(i).getOpcode() == ISD::XOR || |
13345 | BinOp.getOperand(i).getOpcode() == ISD::SELECT || |
13346 | BinOp.getOperand(i).getOpcode() == ISD::SELECT_CC) { |
13347 | BinOps.push_back(BinOp.getOperand(i)); |
13348 | } else { |
13349 | // We have an input that is not a truncation or another binary |
13350 | // operation; we'll abort this transformation. |
13351 | return SDValue(); |
13352 | } |
13353 | } |
13354 | } |
13355 | |
13356 | // The operands of a select that must be truncated when the select is |
13357 | // promoted because the operand is actually part of the to-be-promoted set. |
13358 | DenseMap<SDNode *, EVT> SelectTruncOp[2]; |
13359 | |
13360 | // Make sure that this is a self-contained cluster of operations (which |
13361 | // is not quite the same thing as saying that everything has only one |
13362 | // use). |
13363 | for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) { |
13364 | if (isa<ConstantSDNode>(Inputs[i])) |
13365 | continue; |
13366 | |
13367 | for (SDNode::use_iterator UI = Inputs[i].getNode()->use_begin(), |
13368 | UE = Inputs[i].getNode()->use_end(); |
13369 | UI != UE; ++UI) { |
13370 | SDNode *User = *UI; |
13371 | if (User != N && !Visited.count(User)) |
13372 | return SDValue(); |
13373 | |
13374 | // If we're going to promote the non-output-value operand(s) or SELECT or |
13375 | // SELECT_CC, record them for truncation. |
13376 | if (User->getOpcode() == ISD::SELECT) { |
13377 | if (User->getOperand(0) == Inputs[i]) |
13378 | SelectTruncOp[0].insert(std::make_pair(User, |
13379 | User->getOperand(0).getValueType())); |
13380 | } else if (User->getOpcode() == ISD::SELECT_CC) { |
13381 | if (User->getOperand(0) == Inputs[i]) |
13382 | SelectTruncOp[0].insert(std::make_pair(User, |
13383 | User->getOperand(0).getValueType())); |
13384 | if (User->getOperand(1) == Inputs[i]) |
13385 | SelectTruncOp[1].insert(std::make_pair(User, |
13386 | User->getOperand(1).getValueType())); |
13387 | } |
13388 | } |
13389 | } |
13390 | |
13391 | for (unsigned i = 0, ie = PromOps.size(); i != ie; ++i) { |
13392 | for (SDNode::use_iterator UI = PromOps[i].getNode()->use_begin(), |
13393 | UE = PromOps[i].getNode()->use_end(); |
13394 | UI != UE; ++UI) { |
13395 | SDNode *User = *UI; |
13396 | if (User != N && !Visited.count(User)) |
13397 | return SDValue(); |
13398 | |
13399 | // If we're going to promote the non-output-value operand(s) or SELECT or |
13400 | // SELECT_CC, record them for truncation. |
13401 | if (User->getOpcode() == ISD::SELECT) { |
13402 | if (User->getOperand(0) == PromOps[i]) |
13403 | SelectTruncOp[0].insert(std::make_pair(User, |
13404 | User->getOperand(0).getValueType())); |
13405 | } else if (User->getOpcode() == ISD::SELECT_CC) { |
13406 | if (User->getOperand(0) == PromOps[i]) |
13407 | SelectTruncOp[0].insert(std::make_pair(User, |
13408 | User->getOperand(0).getValueType())); |
13409 | if (User->getOperand(1) == PromOps[i]) |
13410 | SelectTruncOp[1].insert(std::make_pair(User, |
13411 | User->getOperand(1).getValueType())); |
13412 | } |
13413 | } |
13414 | } |
13415 | |
13416 | unsigned PromBits = N->getOperand(0).getValueSizeInBits(); |
13417 | bool ReallyNeedsExt = false; |
13418 | if (N->getOpcode() != ISD::ANY_EXTEND) { |
13419 | // If all of the inputs are not already sign/zero extended, then |
13420 | // we'll still need to do that at the end. |
13421 | for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) { |
13422 | if (isa<ConstantSDNode>(Inputs[i])) |
13423 | continue; |
13424 | |
13425 | unsigned OpBits = |
13426 | Inputs[i].getOperand(0).getValueSizeInBits(); |
13427 | assert(PromBits < OpBits && "Truncation not to a smaller bit count?")((PromBits < OpBits && "Truncation not to a smaller bit count?" ) ? static_cast<void> (0) : __assert_fail ("PromBits < OpBits && \"Truncation not to a smaller bit count?\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13427, __PRETTY_FUNCTION__)); |
13428 | |
13429 | if ((N->getOpcode() == ISD::ZERO_EXTEND && |
13430 | !DAG.MaskedValueIsZero(Inputs[i].getOperand(0), |
13431 | APInt::getHighBitsSet(OpBits, |
13432 | OpBits-PromBits))) || |
13433 | (N->getOpcode() == ISD::SIGN_EXTEND && |
13434 | DAG.ComputeNumSignBits(Inputs[i].getOperand(0)) < |
13435 | (OpBits-(PromBits-1)))) { |
13436 | ReallyNeedsExt = true; |
13437 | break; |
13438 | } |
13439 | } |
13440 | } |
13441 | |
13442 | // Replace all inputs, either with the truncation operand, or a |
13443 | // truncation or extension to the final output type. |
13444 | for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) { |
13445 | // Constant inputs need to be replaced with the to-be-promoted nodes that |
13446 | // use them because they might have users outside of the cluster of |
13447 | // promoted nodes. |
13448 | if (isa<ConstantSDNode>(Inputs[i])) |
13449 | continue; |
13450 | |
13451 | SDValue InSrc = Inputs[i].getOperand(0); |
13452 | if (Inputs[i].getValueType() == N->getValueType(0)) |
13453 | DAG.ReplaceAllUsesOfValueWith(Inputs[i], InSrc); |
13454 | else if (N->getOpcode() == ISD::SIGN_EXTEND) |
13455 | DAG.ReplaceAllUsesOfValueWith(Inputs[i], |
13456 | DAG.getSExtOrTrunc(InSrc, dl, N->getValueType(0))); |
13457 | else if (N->getOpcode() == ISD::ZERO_EXTEND) |
13458 | DAG.ReplaceAllUsesOfValueWith(Inputs[i], |
13459 | DAG.getZExtOrTrunc(InSrc, dl, N->getValueType(0))); |
13460 | else |
13461 | DAG.ReplaceAllUsesOfValueWith(Inputs[i], |
13462 | DAG.getAnyExtOrTrunc(InSrc, dl, N->getValueType(0))); |
13463 | } |
13464 | |
13465 | std::list<HandleSDNode> PromOpHandles; |
13466 | for (auto &PromOp : PromOps) |
13467 | PromOpHandles.emplace_back(PromOp); |
13468 | |
13469 | // Replace all operations (these are all the same, but have a different |
13470 | // (promoted) return type). DAG.getNode will validate that the types of |
13471 | // a binary operator match, so go through the list in reverse so that |
13472 | // we've likely promoted both operands first. |
13473 | while (!PromOpHandles.empty()) { |
13474 | SDValue PromOp = PromOpHandles.back().getValue(); |
13475 | PromOpHandles.pop_back(); |
13476 | |
13477 | unsigned C; |
13478 | switch (PromOp.getOpcode()) { |
13479 | default: C = 0; break; |
13480 | case ISD::SELECT: C = 1; break; |
13481 | case ISD::SELECT_CC: C = 2; break; |
13482 | } |
13483 | |
13484 | if ((!isa<ConstantSDNode>(PromOp.getOperand(C)) && |
13485 | PromOp.getOperand(C).getValueType() != N->getValueType(0)) || |
13486 | (!isa<ConstantSDNode>(PromOp.getOperand(C+1)) && |
13487 | PromOp.getOperand(C+1).getValueType() != N->getValueType(0))) { |
13488 | // The to-be-promoted operands of this node have not yet been |
13489 | // promoted (this should be rare because we're going through the |
13490 | // list backward, but if one of the operands has several users in |
13491 | // this cluster of to-be-promoted nodes, it is possible). |
13492 | PromOpHandles.emplace_front(PromOp); |
13493 | continue; |
13494 | } |
13495 | |
13496 | // For SELECT and SELECT_CC nodes, we do a similar check for any |
13497 | // to-be-promoted comparison inputs. |
13498 | if (PromOp.getOpcode() == ISD::SELECT || |
13499 | PromOp.getOpcode() == ISD::SELECT_CC) { |
13500 | if ((SelectTruncOp[0].count(PromOp.getNode()) && |
13501 | PromOp.getOperand(0).getValueType() != N->getValueType(0)) || |
13502 | (SelectTruncOp[1].count(PromOp.getNode()) && |
13503 | PromOp.getOperand(1).getValueType() != N->getValueType(0))) { |
13504 | PromOpHandles.emplace_front(PromOp); |
13505 | continue; |
13506 | } |
13507 | } |
13508 | |
13509 | SmallVector<SDValue, 3> Ops(PromOp.getNode()->op_begin(), |
13510 | PromOp.getNode()->op_end()); |
13511 | |
13512 | // If this node has constant inputs, then they'll need to be promoted here. |
13513 | for (unsigned i = 0; i < 2; ++i) { |
13514 | if (!isa<ConstantSDNode>(Ops[C+i])) |
13515 | continue; |
13516 | if (Ops[C+i].getValueType() == N->getValueType(0)) |
13517 | continue; |
13518 | |
13519 | if (N->getOpcode() == ISD::SIGN_EXTEND) |
13520 | Ops[C+i] = DAG.getSExtOrTrunc(Ops[C+i], dl, N->getValueType(0)); |
13521 | else if (N->getOpcode() == ISD::ZERO_EXTEND) |
13522 | Ops[C+i] = DAG.getZExtOrTrunc(Ops[C+i], dl, N->getValueType(0)); |
13523 | else |
13524 | Ops[C+i] = DAG.getAnyExtOrTrunc(Ops[C+i], dl, N->getValueType(0)); |
13525 | } |
13526 | |
13527 | // If we've promoted the comparison inputs of a SELECT or SELECT_CC, |
13528 | // truncate them again to the original value type. |
13529 | if (PromOp.getOpcode() == ISD::SELECT || |
13530 | PromOp.getOpcode() == ISD::SELECT_CC) { |
13531 | auto SI0 = SelectTruncOp[0].find(PromOp.getNode()); |
13532 | if (SI0 != SelectTruncOp[0].end()) |
13533 | Ops[0] = DAG.getNode(ISD::TRUNCATE, dl, SI0->second, Ops[0]); |
13534 | auto SI1 = SelectTruncOp[1].find(PromOp.getNode()); |
13535 | if (SI1 != SelectTruncOp[1].end()) |
13536 | Ops[1] = DAG.getNode(ISD::TRUNCATE, dl, SI1->second, Ops[1]); |
13537 | } |
13538 | |
13539 | DAG.ReplaceAllUsesOfValueWith(PromOp, |
13540 | DAG.getNode(PromOp.getOpcode(), dl, N->getValueType(0), Ops)); |
13541 | } |
13542 | |
13543 | // Now we're left with the initial extension itself. |
13544 | if (!ReallyNeedsExt) |
13545 | return N->getOperand(0); |
13546 | |
13547 | // To zero extend, just mask off everything except for the first bit (in the |
13548 | // i1 case). |
13549 | if (N->getOpcode() == ISD::ZERO_EXTEND) |
13550 | return DAG.getNode(ISD::AND, dl, N->getValueType(0), N->getOperand(0), |
13551 | DAG.getConstant(APInt::getLowBitsSet( |
13552 | N->getValueSizeInBits(0), PromBits), |
13553 | dl, N->getValueType(0))); |
13554 | |
13555 | assert(N->getOpcode() == ISD::SIGN_EXTEND &&((N->getOpcode() == ISD::SIGN_EXTEND && "Invalid extension type" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::SIGN_EXTEND && \"Invalid extension type\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13556, __PRETTY_FUNCTION__)) |
13556 | "Invalid extension type")((N->getOpcode() == ISD::SIGN_EXTEND && "Invalid extension type" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::SIGN_EXTEND && \"Invalid extension type\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13556, __PRETTY_FUNCTION__)); |
13557 | EVT ShiftAmountTy = getShiftAmountTy(N->getValueType(0), DAG.getDataLayout()); |
13558 | SDValue ShiftCst = |
13559 | DAG.getConstant(N->getValueSizeInBits(0) - PromBits, dl, ShiftAmountTy); |
13560 | return DAG.getNode( |
13561 | ISD::SRA, dl, N->getValueType(0), |
13562 | DAG.getNode(ISD::SHL, dl, N->getValueType(0), N->getOperand(0), ShiftCst), |
13563 | ShiftCst); |
13564 | } |
13565 | |
13566 | SDValue PPCTargetLowering::combineSetCC(SDNode *N, |
13567 | DAGCombinerInfo &DCI) const { |
13568 | assert(N->getOpcode() == ISD::SETCC &&((N->getOpcode() == ISD::SETCC && "Should be called with a SETCC node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::SETCC && \"Should be called with a SETCC node\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13569, __PRETTY_FUNCTION__)) |
13569 | "Should be called with a SETCC node")((N->getOpcode() == ISD::SETCC && "Should be called with a SETCC node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::SETCC && \"Should be called with a SETCC node\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13569, __PRETTY_FUNCTION__)); |
13570 | |
13571 | ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get(); |
13572 | if (CC == ISD::SETNE || CC == ISD::SETEQ) { |
13573 | SDValue LHS = N->getOperand(0); |
13574 | SDValue RHS = N->getOperand(1); |
13575 | |
13576 | // If there is a '0 - y' pattern, canonicalize the pattern to the RHS. |
13577 | if (LHS.getOpcode() == ISD::SUB && isNullConstant(LHS.getOperand(0)) && |
13578 | LHS.hasOneUse()) |
13579 | std::swap(LHS, RHS); |
13580 | |
13581 | // x == 0-y --> x+y == 0 |
13582 | // x != 0-y --> x+y != 0 |
13583 | if (RHS.getOpcode() == ISD::SUB && isNullConstant(RHS.getOperand(0)) && |
13584 | RHS.hasOneUse()) { |
13585 | SDLoc DL(N); |
13586 | SelectionDAG &DAG = DCI.DAG; |
13587 | EVT VT = N->getValueType(0); |
13588 | EVT OpVT = LHS.getValueType(); |
13589 | SDValue Add = DAG.getNode(ISD::ADD, DL, OpVT, LHS, RHS.getOperand(1)); |
13590 | return DAG.getSetCC(DL, VT, Add, DAG.getConstant(0, DL, OpVT), CC); |
13591 | } |
13592 | } |
13593 | |
13594 | return DAGCombineTruncBoolExt(N, DCI); |
13595 | } |
13596 | |
13597 | // Is this an extending load from an f32 to an f64? |
13598 | static bool isFPExtLoad(SDValue Op) { |
13599 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op.getNode())) |
13600 | return LD->getExtensionType() == ISD::EXTLOAD && |
13601 | Op.getValueType() == MVT::f64; |
13602 | return false; |
13603 | } |
13604 | |
13605 | /// Reduces the number of fp-to-int conversion when building a vector. |
13606 | /// |
13607 | /// If this vector is built out of floating to integer conversions, |
13608 | /// transform it to a vector built out of floating point values followed by a |
13609 | /// single floating to integer conversion of the vector. |
13610 | /// Namely (build_vector (fptosi $A), (fptosi $B), ...) |
13611 | /// becomes (fptosi (build_vector ($A, $B, ...))) |
13612 | SDValue PPCTargetLowering:: |
13613 | combineElementTruncationToVectorTruncation(SDNode *N, |
13614 | DAGCombinerInfo &DCI) const { |
13615 | assert(N->getOpcode() == ISD::BUILD_VECTOR &&((N->getOpcode() == ISD::BUILD_VECTOR && "Should be called with a BUILD_VECTOR node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::BUILD_VECTOR && \"Should be called with a BUILD_VECTOR node\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13616, __PRETTY_FUNCTION__)) |
13616 | "Should be called with a BUILD_VECTOR node")((N->getOpcode() == ISD::BUILD_VECTOR && "Should be called with a BUILD_VECTOR node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::BUILD_VECTOR && \"Should be called with a BUILD_VECTOR node\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13616, __PRETTY_FUNCTION__)); |
13617 | |
13618 | SelectionDAG &DAG = DCI.DAG; |
13619 | SDLoc dl(N); |
13620 | |
13621 | SDValue FirstInput = N->getOperand(0); |
13622 | assert(FirstInput.getOpcode() == PPCISD::MFVSR &&((FirstInput.getOpcode() == PPCISD::MFVSR && "The input operand must be an fp-to-int conversion." ) ? static_cast<void> (0) : __assert_fail ("FirstInput.getOpcode() == PPCISD::MFVSR && \"The input operand must be an fp-to-int conversion.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13623, __PRETTY_FUNCTION__)) |
13623 | "The input operand must be an fp-to-int conversion.")((FirstInput.getOpcode() == PPCISD::MFVSR && "The input operand must be an fp-to-int conversion." ) ? static_cast<void> (0) : __assert_fail ("FirstInput.getOpcode() == PPCISD::MFVSR && \"The input operand must be an fp-to-int conversion.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13623, __PRETTY_FUNCTION__)); |
13624 | |
13625 | // This combine happens after legalization so the fp_to_[su]i nodes are |
13626 | // already converted to PPCSISD nodes. |
13627 | unsigned FirstConversion = FirstInput.getOperand(0).getOpcode(); |
13628 | if (FirstConversion == PPCISD::FCTIDZ || |
13629 | FirstConversion == PPCISD::FCTIDUZ || |
13630 | FirstConversion == PPCISD::FCTIWZ || |
13631 | FirstConversion == PPCISD::FCTIWUZ) { |
13632 | bool IsSplat = true; |
13633 | bool Is32Bit = FirstConversion == PPCISD::FCTIWZ || |
13634 | FirstConversion == PPCISD::FCTIWUZ; |
13635 | EVT SrcVT = FirstInput.getOperand(0).getValueType(); |
13636 | SmallVector<SDValue, 4> Ops; |
13637 | EVT TargetVT = N->getValueType(0); |
13638 | for (int i = 0, e = N->getNumOperands(); i < e; ++i) { |
13639 | SDValue NextOp = N->getOperand(i); |
13640 | if (NextOp.getOpcode() != PPCISD::MFVSR) |
13641 | return SDValue(); |
13642 | unsigned NextConversion = NextOp.getOperand(0).getOpcode(); |
13643 | if (NextConversion != FirstConversion) |
13644 | return SDValue(); |
13645 | // If we are converting to 32-bit integers, we need to add an FP_ROUND. |
13646 | // This is not valid if the input was originally double precision. It is |
13647 | // also not profitable to do unless this is an extending load in which |
13648 | // case doing this combine will allow us to combine consecutive loads. |
13649 | if (Is32Bit && !isFPExtLoad(NextOp.getOperand(0).getOperand(0))) |
13650 | return SDValue(); |
13651 | if (N->getOperand(i) != FirstInput) |
13652 | IsSplat = false; |
13653 | } |
13654 | |
13655 | // If this is a splat, we leave it as-is since there will be only a single |
13656 | // fp-to-int conversion followed by a splat of the integer. This is better |
13657 | // for 32-bit and smaller ints and neutral for 64-bit ints. |
13658 | if (IsSplat) |
13659 | return SDValue(); |
13660 | |
13661 | // Now that we know we have the right type of node, get its operands |
13662 | for (int i = 0, e = N->getNumOperands(); i < e; ++i) { |
13663 | SDValue In = N->getOperand(i).getOperand(0); |
13664 | if (Is32Bit) { |
13665 | // For 32-bit values, we need to add an FP_ROUND node (if we made it |
13666 | // here, we know that all inputs are extending loads so this is safe). |
13667 | if (In.isUndef()) |
13668 | Ops.push_back(DAG.getUNDEF(SrcVT)); |
13669 | else { |
13670 | SDValue Trunc = DAG.getNode(ISD::FP_ROUND, dl, |
13671 | MVT::f32, In.getOperand(0), |
13672 | DAG.getIntPtrConstant(1, dl)); |
13673 | Ops.push_back(Trunc); |
13674 | } |
13675 | } else |
13676 | Ops.push_back(In.isUndef() ? DAG.getUNDEF(SrcVT) : In.getOperand(0)); |
13677 | } |
13678 | |
13679 | unsigned Opcode; |
13680 | if (FirstConversion == PPCISD::FCTIDZ || |
13681 | FirstConversion == PPCISD::FCTIWZ) |
13682 | Opcode = ISD::FP_TO_SINT; |
13683 | else |
13684 | Opcode = ISD::FP_TO_UINT; |
13685 | |
13686 | EVT NewVT = TargetVT == MVT::v2i64 ? MVT::v2f64 : MVT::v4f32; |
13687 | SDValue BV = DAG.getBuildVector(NewVT, dl, Ops); |
13688 | return DAG.getNode(Opcode, dl, TargetVT, BV); |
13689 | } |
13690 | return SDValue(); |
13691 | } |
13692 | |
13693 | /// Reduce the number of loads when building a vector. |
13694 | /// |
13695 | /// Building a vector out of multiple loads can be converted to a load |
13696 | /// of the vector type if the loads are consecutive. If the loads are |
13697 | /// consecutive but in descending order, a shuffle is added at the end |
13698 | /// to reorder the vector. |
13699 | static SDValue combineBVOfConsecutiveLoads(SDNode *N, SelectionDAG &DAG) { |
13700 | assert(N->getOpcode() == ISD::BUILD_VECTOR &&((N->getOpcode() == ISD::BUILD_VECTOR && "Should be called with a BUILD_VECTOR node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::BUILD_VECTOR && \"Should be called with a BUILD_VECTOR node\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13701, __PRETTY_FUNCTION__)) |
13701 | "Should be called with a BUILD_VECTOR node")((N->getOpcode() == ISD::BUILD_VECTOR && "Should be called with a BUILD_VECTOR node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::BUILD_VECTOR && \"Should be called with a BUILD_VECTOR node\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13701, __PRETTY_FUNCTION__)); |
13702 | |
13703 | SDLoc dl(N); |
13704 | |
13705 | // Return early for non byte-sized type, as they can't be consecutive. |
13706 | if (!N->getValueType(0).getVectorElementType().isByteSized()) |
13707 | return SDValue(); |
13708 | |
13709 | bool InputsAreConsecutiveLoads = true; |
13710 | bool InputsAreReverseConsecutive = true; |
13711 | unsigned ElemSize = N->getValueType(0).getScalarType().getStoreSize(); |
13712 | SDValue FirstInput = N->getOperand(0); |
13713 | bool IsRoundOfExtLoad = false; |
13714 | |
13715 | if (FirstInput.getOpcode() == ISD::FP_ROUND && |
13716 | FirstInput.getOperand(0).getOpcode() == ISD::LOAD) { |
13717 | LoadSDNode *LD = dyn_cast<LoadSDNode>(FirstInput.getOperand(0)); |
13718 | IsRoundOfExtLoad = LD->getExtensionType() == ISD::EXTLOAD; |
13719 | } |
13720 | // Not a build vector of (possibly fp_rounded) loads. |
13721 | if ((!IsRoundOfExtLoad && FirstInput.getOpcode() != ISD::LOAD) || |
13722 | N->getNumOperands() == 1) |
13723 | return SDValue(); |
13724 | |
13725 | for (int i = 1, e = N->getNumOperands(); i < e; ++i) { |
13726 | // If any inputs are fp_round(extload), they all must be. |
13727 | if (IsRoundOfExtLoad && N->getOperand(i).getOpcode() != ISD::FP_ROUND) |
13728 | return SDValue(); |
13729 | |
13730 | SDValue NextInput = IsRoundOfExtLoad ? N->getOperand(i).getOperand(0) : |
13731 | N->getOperand(i); |
13732 | if (NextInput.getOpcode() != ISD::LOAD) |
13733 | return SDValue(); |
13734 | |
13735 | SDValue PreviousInput = |
13736 | IsRoundOfExtLoad ? N->getOperand(i-1).getOperand(0) : N->getOperand(i-1); |
13737 | LoadSDNode *LD1 = dyn_cast<LoadSDNode>(PreviousInput); |
13738 | LoadSDNode *LD2 = dyn_cast<LoadSDNode>(NextInput); |
13739 | |
13740 | // If any inputs are fp_round(extload), they all must be. |
13741 | if (IsRoundOfExtLoad && LD2->getExtensionType() != ISD::EXTLOAD) |
13742 | return SDValue(); |
13743 | |
13744 | if (!isConsecutiveLS(LD2, LD1, ElemSize, 1, DAG)) |
13745 | InputsAreConsecutiveLoads = false; |
13746 | if (!isConsecutiveLS(LD1, LD2, ElemSize, 1, DAG)) |
13747 | InputsAreReverseConsecutive = false; |
13748 | |
13749 | // Exit early if the loads are neither consecutive nor reverse consecutive. |
13750 | if (!InputsAreConsecutiveLoads && !InputsAreReverseConsecutive) |
13751 | return SDValue(); |
13752 | } |
13753 | |
13754 | assert(!(InputsAreConsecutiveLoads && InputsAreReverseConsecutive) &&((!(InputsAreConsecutiveLoads && InputsAreReverseConsecutive ) && "The loads cannot be both consecutive and reverse consecutive." ) ? static_cast<void> (0) : __assert_fail ("!(InputsAreConsecutiveLoads && InputsAreReverseConsecutive) && \"The loads cannot be both consecutive and reverse consecutive.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13755, __PRETTY_FUNCTION__)) |
13755 | "The loads cannot be both consecutive and reverse consecutive.")((!(InputsAreConsecutiveLoads && InputsAreReverseConsecutive ) && "The loads cannot be both consecutive and reverse consecutive." ) ? static_cast<void> (0) : __assert_fail ("!(InputsAreConsecutiveLoads && InputsAreReverseConsecutive) && \"The loads cannot be both consecutive and reverse consecutive.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13755, __PRETTY_FUNCTION__)); |
13756 | |
13757 | SDValue FirstLoadOp = |
13758 | IsRoundOfExtLoad ? FirstInput.getOperand(0) : FirstInput; |
13759 | SDValue LastLoadOp = |
13760 | IsRoundOfExtLoad ? N->getOperand(N->getNumOperands()-1).getOperand(0) : |
13761 | N->getOperand(N->getNumOperands()-1); |
13762 | |
13763 | LoadSDNode *LD1 = dyn_cast<LoadSDNode>(FirstLoadOp); |
13764 | LoadSDNode *LDL = dyn_cast<LoadSDNode>(LastLoadOp); |
13765 | if (InputsAreConsecutiveLoads) { |
13766 | assert(LD1 && "Input needs to be a LoadSDNode.")((LD1 && "Input needs to be a LoadSDNode.") ? static_cast <void> (0) : __assert_fail ("LD1 && \"Input needs to be a LoadSDNode.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13766, __PRETTY_FUNCTION__)); |
13767 | return DAG.getLoad(N->getValueType(0), dl, LD1->getChain(), |
13768 | LD1->getBasePtr(), LD1->getPointerInfo(), |
13769 | LD1->getAlignment()); |
13770 | } |
13771 | if (InputsAreReverseConsecutive) { |
13772 | assert(LDL && "Input needs to be a LoadSDNode.")((LDL && "Input needs to be a LoadSDNode.") ? static_cast <void> (0) : __assert_fail ("LDL && \"Input needs to be a LoadSDNode.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13772, __PRETTY_FUNCTION__)); |
13773 | SDValue Load = DAG.getLoad(N->getValueType(0), dl, LDL->getChain(), |
13774 | LDL->getBasePtr(), LDL->getPointerInfo(), |
13775 | LDL->getAlignment()); |
13776 | SmallVector<int, 16> Ops; |
13777 | for (int i = N->getNumOperands() - 1; i >= 0; i--) |
13778 | Ops.push_back(i); |
13779 | |
13780 | return DAG.getVectorShuffle(N->getValueType(0), dl, Load, |
13781 | DAG.getUNDEF(N->getValueType(0)), Ops); |
13782 | } |
13783 | return SDValue(); |
13784 | } |
13785 | |
13786 | // This function adds the required vector_shuffle needed to get |
13787 | // the elements of the vector extract in the correct position |
13788 | // as specified by the CorrectElems encoding. |
13789 | static SDValue addShuffleForVecExtend(SDNode *N, SelectionDAG &DAG, |
13790 | SDValue Input, uint64_t Elems, |
13791 | uint64_t CorrectElems) { |
13792 | SDLoc dl(N); |
13793 | |
13794 | unsigned NumElems = Input.getValueType().getVectorNumElements(); |
13795 | SmallVector<int, 16> ShuffleMask(NumElems, -1); |
13796 | |
13797 | // Knowing the element indices being extracted from the original |
13798 | // vector and the order in which they're being inserted, just put |
13799 | // them at element indices required for the instruction. |
13800 | for (unsigned i = 0; i < N->getNumOperands(); i++) { |
13801 | if (DAG.getDataLayout().isLittleEndian()) |
13802 | ShuffleMask[CorrectElems & 0xF] = Elems & 0xF; |
13803 | else |
13804 | ShuffleMask[(CorrectElems & 0xF0) >> 4] = (Elems & 0xF0) >> 4; |
13805 | CorrectElems = CorrectElems >> 8; |
13806 | Elems = Elems >> 8; |
13807 | } |
13808 | |
13809 | SDValue Shuffle = |
13810 | DAG.getVectorShuffle(Input.getValueType(), dl, Input, |
13811 | DAG.getUNDEF(Input.getValueType()), ShuffleMask); |
13812 | |
13813 | EVT VT = N->getValueType(0); |
13814 | SDValue Conv = DAG.getBitcast(VT, Shuffle); |
13815 | |
13816 | EVT ExtVT = EVT::getVectorVT(*DAG.getContext(), |
13817 | Input.getValueType().getVectorElementType(), |
13818 | VT.getVectorNumElements()); |
13819 | return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, VT, Conv, |
13820 | DAG.getValueType(ExtVT)); |
13821 | } |
13822 | |
13823 | // Look for build vector patterns where input operands come from sign |
13824 | // extended vector_extract elements of specific indices. If the correct indices |
13825 | // aren't used, add a vector shuffle to fix up the indices and create |
13826 | // SIGN_EXTEND_INREG node which selects the vector sign extend instructions |
13827 | // during instruction selection. |
13828 | static SDValue combineBVOfVecSExt(SDNode *N, SelectionDAG &DAG) { |
13829 | // This array encodes the indices that the vector sign extend instructions |
13830 | // extract from when extending from one type to another for both BE and LE. |
13831 | // The right nibble of each byte corresponds to the LE incides. |
13832 | // and the left nibble of each byte corresponds to the BE incides. |
13833 | // For example: 0x3074B8FC byte->word |
13834 | // For LE: the allowed indices are: 0x0,0x4,0x8,0xC |
13835 | // For BE: the allowed indices are: 0x3,0x7,0xB,0xF |
13836 | // For example: 0x000070F8 byte->double word |
13837 | // For LE: the allowed indices are: 0x0,0x8 |
13838 | // For BE: the allowed indices are: 0x7,0xF |
13839 | uint64_t TargetElems[] = { |
13840 | 0x3074B8FC, // b->w |
13841 | 0x000070F8, // b->d |
13842 | 0x10325476, // h->w |
13843 | 0x00003074, // h->d |
13844 | 0x00001032, // w->d |
13845 | }; |
13846 | |
13847 | uint64_t Elems = 0; |
13848 | int Index; |
13849 | SDValue Input; |
13850 | |
13851 | auto isSExtOfVecExtract = [&](SDValue Op) -> bool { |
13852 | if (!Op) |
13853 | return false; |
13854 | if (Op.getOpcode() != ISD::SIGN_EXTEND && |
13855 | Op.getOpcode() != ISD::SIGN_EXTEND_INREG) |
13856 | return false; |
13857 | |
13858 | // A SIGN_EXTEND_INREG might be fed by an ANY_EXTEND to produce a value |
13859 | // of the right width. |
13860 | SDValue Extract = Op.getOperand(0); |
13861 | if (Extract.getOpcode() == ISD::ANY_EXTEND) |
13862 | Extract = Extract.getOperand(0); |
13863 | if (Extract.getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
13864 | return false; |
13865 | |
13866 | ConstantSDNode *ExtOp = dyn_cast<ConstantSDNode>(Extract.getOperand(1)); |
13867 | if (!ExtOp) |
13868 | return false; |
13869 | |
13870 | Index = ExtOp->getZExtValue(); |
13871 | if (Input && Input != Extract.getOperand(0)) |
13872 | return false; |
13873 | |
13874 | if (!Input) |
13875 | Input = Extract.getOperand(0); |
13876 | |
13877 | Elems = Elems << 8; |
13878 | Index = DAG.getDataLayout().isLittleEndian() ? Index : Index << 4; |
13879 | Elems |= Index; |
13880 | |
13881 | return true; |
13882 | }; |
13883 | |
13884 | // If the build vector operands aren't sign extended vector extracts, |
13885 | // of the same input vector, then return. |
13886 | for (unsigned i = 0; i < N->getNumOperands(); i++) { |
13887 | if (!isSExtOfVecExtract(N->getOperand(i))) { |
13888 | return SDValue(); |
13889 | } |
13890 | } |
13891 | |
13892 | // If the vector extract indicies are not correct, add the appropriate |
13893 | // vector_shuffle. |
13894 | int TgtElemArrayIdx; |
13895 | int InputSize = Input.getValueType().getScalarSizeInBits(); |
13896 | int OutputSize = N->getValueType(0).getScalarSizeInBits(); |
13897 | if (InputSize + OutputSize == 40) |
13898 | TgtElemArrayIdx = 0; |
13899 | else if (InputSize + OutputSize == 72) |
13900 | TgtElemArrayIdx = 1; |
13901 | else if (InputSize + OutputSize == 48) |
13902 | TgtElemArrayIdx = 2; |
13903 | else if (InputSize + OutputSize == 80) |
13904 | TgtElemArrayIdx = 3; |
13905 | else if (InputSize + OutputSize == 96) |
13906 | TgtElemArrayIdx = 4; |
13907 | else |
13908 | return SDValue(); |
13909 | |
13910 | uint64_t CorrectElems = TargetElems[TgtElemArrayIdx]; |
13911 | CorrectElems = DAG.getDataLayout().isLittleEndian() |
13912 | ? CorrectElems & 0x0F0F0F0F0F0F0F0F |
13913 | : CorrectElems & 0xF0F0F0F0F0F0F0F0; |
13914 | if (Elems != CorrectElems) { |
13915 | return addShuffleForVecExtend(N, DAG, Input, Elems, CorrectElems); |
13916 | } |
13917 | |
13918 | // Regular lowering will catch cases where a shuffle is not needed. |
13919 | return SDValue(); |
13920 | } |
13921 | |
13922 | // Look for the pattern of a load from a narrow width to i128, feeding |
13923 | // into a BUILD_VECTOR of v1i128. Replace this sequence with a PPCISD node |
13924 | // (LXVRZX). This node represents a zero extending load that will be matched |
13925 | // to the Load VSX Vector Rightmost instructions. |
13926 | static SDValue combineBVZEXTLOAD(SDNode *N, SelectionDAG &DAG) { |
13927 | SDLoc DL(N); |
13928 | |
13929 | // This combine is only eligible for a BUILD_VECTOR of v1i128. |
13930 | if (N->getValueType(0) != MVT::v1i128) |
13931 | return SDValue(); |
13932 | |
13933 | SDValue Operand = N->getOperand(0); |
13934 | // Proceed with the transformation if the operand to the BUILD_VECTOR |
13935 | // is a load instruction. |
13936 | if (Operand.getOpcode() != ISD::LOAD) |
13937 | return SDValue(); |
13938 | |
13939 | LoadSDNode *LD = dyn_cast<LoadSDNode>(Operand); |
13940 | EVT MemoryType = LD->getMemoryVT(); |
13941 | |
13942 | // This transformation is only valid if the we are loading either a byte, |
13943 | // halfword, word, or doubleword. |
13944 | bool ValidLDType = MemoryType == MVT::i8 || MemoryType == MVT::i16 || |
13945 | MemoryType == MVT::i32 || MemoryType == MVT::i64; |
13946 | |
13947 | // Ensure that the load from the narrow width is being zero extended to i128. |
13948 | if (!ValidLDType || |
13949 | (LD->getExtensionType() != ISD::ZEXTLOAD && |
13950 | LD->getExtensionType() != ISD::EXTLOAD)) |
13951 | return SDValue(); |
13952 | |
13953 | SDValue LoadOps[] = { |
13954 | LD->getChain(), LD->getBasePtr(), |
13955 | DAG.getIntPtrConstant(MemoryType.getScalarSizeInBits(), DL)}; |
13956 | |
13957 | return DAG.getMemIntrinsicNode(PPCISD::LXVRZX, DL, |
13958 | DAG.getVTList(MVT::v1i128, MVT::Other), |
13959 | LoadOps, MemoryType, LD->getMemOperand()); |
13960 | } |
13961 | |
13962 | SDValue PPCTargetLowering::DAGCombineBuildVector(SDNode *N, |
13963 | DAGCombinerInfo &DCI) const { |
13964 | assert(N->getOpcode() == ISD::BUILD_VECTOR &&((N->getOpcode() == ISD::BUILD_VECTOR && "Should be called with a BUILD_VECTOR node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::BUILD_VECTOR && \"Should be called with a BUILD_VECTOR node\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13965, __PRETTY_FUNCTION__)) |
13965 | "Should be called with a BUILD_VECTOR node")((N->getOpcode() == ISD::BUILD_VECTOR && "Should be called with a BUILD_VECTOR node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::BUILD_VECTOR && \"Should be called with a BUILD_VECTOR node\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13965, __PRETTY_FUNCTION__)); |
13966 | |
13967 | SelectionDAG &DAG = DCI.DAG; |
13968 | SDLoc dl(N); |
13969 | |
13970 | if (!Subtarget.hasVSX()) |
13971 | return SDValue(); |
13972 | |
13973 | // The target independent DAG combiner will leave a build_vector of |
13974 | // float-to-int conversions intact. We can generate MUCH better code for |
13975 | // a float-to-int conversion of a vector of floats. |
13976 | SDValue FirstInput = N->getOperand(0); |
13977 | if (FirstInput.getOpcode() == PPCISD::MFVSR) { |
13978 | SDValue Reduced = combineElementTruncationToVectorTruncation(N, DCI); |
13979 | if (Reduced) |
13980 | return Reduced; |
13981 | } |
13982 | |
13983 | // If we're building a vector out of consecutive loads, just load that |
13984 | // vector type. |
13985 | SDValue Reduced = combineBVOfConsecutiveLoads(N, DAG); |
13986 | if (Reduced) |
13987 | return Reduced; |
13988 | |
13989 | // If we're building a vector out of extended elements from another vector |
13990 | // we have P9 vector integer extend instructions. The code assumes legal |
13991 | // input types (i.e. it can't handle things like v4i16) so do not run before |
13992 | // legalization. |
13993 | if (Subtarget.hasP9Altivec() && !DCI.isBeforeLegalize()) { |
13994 | Reduced = combineBVOfVecSExt(N, DAG); |
13995 | if (Reduced) |
13996 | return Reduced; |
13997 | } |
13998 | |
13999 | // On Power10, the Load VSX Vector Rightmost instructions can be utilized |
14000 | // if this is a BUILD_VECTOR of v1i128, and if the operand to the BUILD_VECTOR |
14001 | // is a load from <valid narrow width> to i128. |
14002 | if (Subtarget.isISA3_1()) { |
14003 | SDValue BVOfZLoad = combineBVZEXTLOAD(N, DAG); |
14004 | if (BVOfZLoad) |
14005 | return BVOfZLoad; |
14006 | } |
14007 | |
14008 | if (N->getValueType(0) != MVT::v2f64) |
14009 | return SDValue(); |
14010 | |
14011 | // Looking for: |
14012 | // (build_vector ([su]int_to_fp (extractelt 0)), [su]int_to_fp (extractelt 1)) |
14013 | if (FirstInput.getOpcode() != ISD::SINT_TO_FP && |
14014 | FirstInput.getOpcode() != ISD::UINT_TO_FP) |
14015 | return SDValue(); |
14016 | if (N->getOperand(1).getOpcode() != ISD::SINT_TO_FP && |
14017 | N->getOperand(1).getOpcode() != ISD::UINT_TO_FP) |
14018 | return SDValue(); |
14019 | if (FirstInput.getOpcode() != N->getOperand(1).getOpcode()) |
14020 | return SDValue(); |
14021 | |
14022 | SDValue Ext1 = FirstInput.getOperand(0); |
14023 | SDValue Ext2 = N->getOperand(1).getOperand(0); |
14024 | if(Ext1.getOpcode() != ISD::EXTRACT_VECTOR_ELT || |
14025 | Ext2.getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
14026 | return SDValue(); |
14027 | |
14028 | ConstantSDNode *Ext1Op = dyn_cast<ConstantSDNode>(Ext1.getOperand(1)); |
14029 | ConstantSDNode *Ext2Op = dyn_cast<ConstantSDNode>(Ext2.getOperand(1)); |
14030 | if (!Ext1Op || !Ext2Op) |
14031 | return SDValue(); |
14032 | if (Ext1.getOperand(0).getValueType() != MVT::v4i32 || |
14033 | Ext1.getOperand(0) != Ext2.getOperand(0)) |
14034 | return SDValue(); |
14035 | |
14036 | int FirstElem = Ext1Op->getZExtValue(); |
14037 | int SecondElem = Ext2Op->getZExtValue(); |
14038 | int SubvecIdx; |
14039 | if (FirstElem == 0 && SecondElem == 1) |
14040 | SubvecIdx = Subtarget.isLittleEndian() ? 1 : 0; |
14041 | else if (FirstElem == 2 && SecondElem == 3) |
14042 | SubvecIdx = Subtarget.isLittleEndian() ? 0 : 1; |
14043 | else |
14044 | return SDValue(); |
14045 | |
14046 | SDValue SrcVec = Ext1.getOperand(0); |
14047 | auto NodeType = (N->getOperand(1).getOpcode() == ISD::SINT_TO_FP) ? |
14048 | PPCISD::SINT_VEC_TO_FP : PPCISD::UINT_VEC_TO_FP; |
14049 | return DAG.getNode(NodeType, dl, MVT::v2f64, |
14050 | SrcVec, DAG.getIntPtrConstant(SubvecIdx, dl)); |
14051 | } |
14052 | |
14053 | SDValue PPCTargetLowering::combineFPToIntToFP(SDNode *N, |
14054 | DAGCombinerInfo &DCI) const { |
14055 | assert((N->getOpcode() == ISD::SINT_TO_FP ||(((N->getOpcode() == ISD::SINT_TO_FP || N->getOpcode() == ISD::UINT_TO_FP) && "Need an int -> FP conversion node here" ) ? static_cast<void> (0) : __assert_fail ("(N->getOpcode() == ISD::SINT_TO_FP || N->getOpcode() == ISD::UINT_TO_FP) && \"Need an int -> FP conversion node here\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14057, __PRETTY_FUNCTION__)) |
14056 | N->getOpcode() == ISD::UINT_TO_FP) &&(((N->getOpcode() == ISD::SINT_TO_FP || N->getOpcode() == ISD::UINT_TO_FP) && "Need an int -> FP conversion node here" ) ? static_cast<void> (0) : __assert_fail ("(N->getOpcode() == ISD::SINT_TO_FP || N->getOpcode() == ISD::UINT_TO_FP) && \"Need an int -> FP conversion node here\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14057, __PRETTY_FUNCTION__)) |
14057 | "Need an int -> FP conversion node here")(((N->getOpcode() == ISD::SINT_TO_FP || N->getOpcode() == ISD::UINT_TO_FP) && "Need an int -> FP conversion node here" ) ? static_cast<void> (0) : __assert_fail ("(N->getOpcode() == ISD::SINT_TO_FP || N->getOpcode() == ISD::UINT_TO_FP) && \"Need an int -> FP conversion node here\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14057, __PRETTY_FUNCTION__)); |
14058 | |
14059 | if (useSoftFloat() || !Subtarget.has64BitSupport()) |
14060 | return SDValue(); |
14061 | |
14062 | SelectionDAG &DAG = DCI.DAG; |
14063 | SDLoc dl(N); |
14064 | SDValue Op(N, 0); |
14065 | |
14066 | // Don't handle ppc_fp128 here or conversions that are out-of-range capable |
14067 | // from the hardware. |
14068 | if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64) |
14069 | return SDValue(); |
14070 | if (!Op.getOperand(0).getValueType().isSimple()) |
14071 | return SDValue(); |
14072 | if (Op.getOperand(0).getValueType().getSimpleVT() <= MVT(MVT::i1) || |
14073 | Op.getOperand(0).getValueType().getSimpleVT() > MVT(MVT::i64)) |
14074 | return SDValue(); |
14075 | |
14076 | SDValue FirstOperand(Op.getOperand(0)); |
14077 | bool SubWordLoad = FirstOperand.getOpcode() == ISD::LOAD && |
14078 | (FirstOperand.getValueType() == MVT::i8 || |
14079 | FirstOperand.getValueType() == MVT::i16); |
14080 | if (Subtarget.hasP9Vector() && Subtarget.hasP9Altivec() && SubWordLoad) { |
14081 | bool Signed = N->getOpcode() == ISD::SINT_TO_FP; |
14082 | bool DstDouble = Op.getValueType() == MVT::f64; |
14083 | unsigned ConvOp = Signed ? |
14084 | (DstDouble ? PPCISD::FCFID : PPCISD::FCFIDS) : |
14085 | (DstDouble ? PPCISD::FCFIDU : PPCISD::FCFIDUS); |
14086 | SDValue WidthConst = |
14087 | DAG.getIntPtrConstant(FirstOperand.getValueType() == MVT::i8 ? 1 : 2, |
14088 | dl, false); |
14089 | LoadSDNode *LDN = cast<LoadSDNode>(FirstOperand.getNode()); |
14090 | SDValue Ops[] = { LDN->getChain(), LDN->getBasePtr(), WidthConst }; |
14091 | SDValue Ld = DAG.getMemIntrinsicNode(PPCISD::LXSIZX, dl, |
14092 | DAG.getVTList(MVT::f64, MVT::Other), |
14093 | Ops, MVT::i8, LDN->getMemOperand()); |
14094 | |
14095 | // For signed conversion, we need to sign-extend the value in the VSR |
14096 | if (Signed) { |
14097 | SDValue ExtOps[] = { Ld, WidthConst }; |
14098 | SDValue Ext = DAG.getNode(PPCISD::VEXTS, dl, MVT::f64, ExtOps); |
14099 | return DAG.getNode(ConvOp, dl, DstDouble ? MVT::f64 : MVT::f32, Ext); |
14100 | } else |
14101 | return DAG.getNode(ConvOp, dl, DstDouble ? MVT::f64 : MVT::f32, Ld); |
14102 | } |
14103 | |
14104 | |
14105 | // For i32 intermediate values, unfortunately, the conversion functions |
14106 | // leave the upper 32 bits of the value are undefined. Within the set of |
14107 | // scalar instructions, we have no method for zero- or sign-extending the |
14108 | // value. Thus, we cannot handle i32 intermediate values here. |
14109 | if (Op.getOperand(0).getValueType() == MVT::i32) |
14110 | return SDValue(); |
14111 | |
14112 | assert((Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) &&(((Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) && "UINT_TO_FP is supported only with FPCVT") ? static_cast <void> (0) : __assert_fail ("(Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) && \"UINT_TO_FP is supported only with FPCVT\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14113, __PRETTY_FUNCTION__)) |
14113 | "UINT_TO_FP is supported only with FPCVT")(((Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) && "UINT_TO_FP is supported only with FPCVT") ? static_cast <void> (0) : __assert_fail ("(Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) && \"UINT_TO_FP is supported only with FPCVT\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14113, __PRETTY_FUNCTION__)); |
14114 | |
14115 | // If we have FCFIDS, then use it when converting to single-precision. |
14116 | // Otherwise, convert to double-precision and then round. |
14117 | unsigned FCFOp = (Subtarget.hasFPCVT() && Op.getValueType() == MVT::f32) |
14118 | ? (Op.getOpcode() == ISD::UINT_TO_FP ? PPCISD::FCFIDUS |
14119 | : PPCISD::FCFIDS) |
14120 | : (Op.getOpcode() == ISD::UINT_TO_FP ? PPCISD::FCFIDU |
14121 | : PPCISD::FCFID); |
14122 | MVT FCFTy = (Subtarget.hasFPCVT() && Op.getValueType() == MVT::f32) |
14123 | ? MVT::f32 |
14124 | : MVT::f64; |
14125 | |
14126 | // If we're converting from a float, to an int, and back to a float again, |
14127 | // then we don't need the store/load pair at all. |
14128 | if ((Op.getOperand(0).getOpcode() == ISD::FP_TO_UINT && |
14129 | Subtarget.hasFPCVT()) || |
14130 | (Op.getOperand(0).getOpcode() == ISD::FP_TO_SINT)) { |
14131 | SDValue Src = Op.getOperand(0).getOperand(0); |
14132 | if (Src.getValueType() == MVT::f32) { |
14133 | Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src); |
14134 | DCI.AddToWorklist(Src.getNode()); |
14135 | } else if (Src.getValueType() != MVT::f64) { |
14136 | // Make sure that we don't pick up a ppc_fp128 source value. |
14137 | return SDValue(); |
14138 | } |
14139 | |
14140 | unsigned FCTOp = |
14141 | Op.getOperand(0).getOpcode() == ISD::FP_TO_SINT ? PPCISD::FCTIDZ : |
14142 | PPCISD::FCTIDUZ; |
14143 | |
14144 | SDValue Tmp = DAG.getNode(FCTOp, dl, MVT::f64, Src); |
14145 | SDValue FP = DAG.getNode(FCFOp, dl, FCFTy, Tmp); |
14146 | |
14147 | if (Op.getValueType() == MVT::f32 && !Subtarget.hasFPCVT()) { |
14148 | FP = DAG.getNode(ISD::FP_ROUND, dl, |
14149 | MVT::f32, FP, DAG.getIntPtrConstant(0, dl)); |
14150 | DCI.AddToWorklist(FP.getNode()); |
14151 | } |
14152 | |
14153 | return FP; |
14154 | } |
14155 | |
14156 | return SDValue(); |
14157 | } |
14158 | |
14159 | // expandVSXLoadForLE - Convert VSX loads (which may be intrinsics for |
14160 | // builtins) into loads with swaps. |
14161 | SDValue PPCTargetLowering::expandVSXLoadForLE(SDNode *N, |
14162 | DAGCombinerInfo &DCI) const { |
14163 | SelectionDAG &DAG = DCI.DAG; |
14164 | SDLoc dl(N); |
14165 | SDValue Chain; |
14166 | SDValue Base; |
14167 | MachineMemOperand *MMO; |
14168 | |
14169 | switch (N->getOpcode()) { |
14170 | default: |
14171 | llvm_unreachable("Unexpected opcode for little endian VSX load")::llvm::llvm_unreachable_internal("Unexpected opcode for little endian VSX load" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14171); |
14172 | case ISD::LOAD: { |
14173 | LoadSDNode *LD = cast<LoadSDNode>(N); |
14174 | Chain = LD->getChain(); |
14175 | Base = LD->getBasePtr(); |
14176 | MMO = LD->getMemOperand(); |
14177 | // If the MMO suggests this isn't a load of a full vector, leave |
14178 | // things alone. For a built-in, we have to make the change for |
14179 | // correctness, so if there is a size problem that will be a bug. |
14180 | if (MMO->getSize() < 16) |
14181 | return SDValue(); |
14182 | break; |
14183 | } |
14184 | case ISD::INTRINSIC_W_CHAIN: { |
14185 | MemIntrinsicSDNode *Intrin = cast<MemIntrinsicSDNode>(N); |
14186 | Chain = Intrin->getChain(); |
14187 | // Similarly to the store case below, Intrin->getBasePtr() doesn't get |
14188 | // us what we want. Get operand 2 instead. |
14189 | Base = Intrin->getOperand(2); |
14190 | MMO = Intrin->getMemOperand(); |
14191 | break; |
14192 | } |
14193 | } |
14194 | |
14195 | MVT VecTy = N->getValueType(0).getSimpleVT(); |
14196 | |
14197 | // Do not expand to PPCISD::LXVD2X + PPCISD::XXSWAPD when the load is |
14198 | // aligned and the type is a vector with elements up to 4 bytes |
14199 | if (Subtarget.needsSwapsForVSXMemOps() && MMO->getAlign() >= Align(16) && |
14200 | VecTy.getScalarSizeInBits() <= 32) { |
14201 | return SDValue(); |
14202 | } |
14203 | |
14204 | SDValue LoadOps[] = { Chain, Base }; |
14205 | SDValue Load = DAG.getMemIntrinsicNode(PPCISD::LXVD2X, dl, |
14206 | DAG.getVTList(MVT::v2f64, MVT::Other), |
14207 | LoadOps, MVT::v2f64, MMO); |
14208 | |
14209 | DCI.AddToWorklist(Load.getNode()); |
14210 | Chain = Load.getValue(1); |
14211 | SDValue Swap = DAG.getNode( |
14212 | PPCISD::XXSWAPD, dl, DAG.getVTList(MVT::v2f64, MVT::Other), Chain, Load); |
14213 | DCI.AddToWorklist(Swap.getNode()); |
14214 | |
14215 | // Add a bitcast if the resulting load type doesn't match v2f64. |
14216 | if (VecTy != MVT::v2f64) { |
14217 | SDValue N = DAG.getNode(ISD::BITCAST, dl, VecTy, Swap); |
14218 | DCI.AddToWorklist(N.getNode()); |
14219 | // Package {bitcast value, swap's chain} to match Load's shape. |
14220 | return DAG.getNode(ISD::MERGE_VALUES, dl, DAG.getVTList(VecTy, MVT::Other), |
14221 | N, Swap.getValue(1)); |
14222 | } |
14223 | |
14224 | return Swap; |
14225 | } |
14226 | |
14227 | // expandVSXStoreForLE - Convert VSX stores (which may be intrinsics for |
14228 | // builtins) into stores with swaps. |
14229 | SDValue PPCTargetLowering::expandVSXStoreForLE(SDNode *N, |
14230 | DAGCombinerInfo &DCI) const { |
14231 | SelectionDAG &DAG = DCI.DAG; |
14232 | SDLoc dl(N); |
14233 | SDValue Chain; |
14234 | SDValue Base; |
14235 | unsigned SrcOpnd; |
14236 | MachineMemOperand *MMO; |
14237 | |
14238 | switch (N->getOpcode()) { |
14239 | default: |
14240 | llvm_unreachable("Unexpected opcode for little endian VSX store")::llvm::llvm_unreachable_internal("Unexpected opcode for little endian VSX store" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14240); |
14241 | case ISD::STORE: { |
14242 | StoreSDNode *ST = cast<StoreSDNode>(N); |
14243 | Chain = ST->getChain(); |
14244 | Base = ST->getBasePtr(); |
14245 | MMO = ST->getMemOperand(); |
14246 | SrcOpnd = 1; |
14247 | // If the MMO suggests this isn't a store of a full vector, leave |
14248 | // things alone. For a built-in, we have to make the change for |
14249 | // correctness, so if there is a size problem that will be a bug. |
14250 | if (MMO->getSize() < 16) |
14251 | return SDValue(); |
14252 | break; |
14253 | } |
14254 | case ISD::INTRINSIC_VOID: { |
14255 | MemIntrinsicSDNode *Intrin = cast<MemIntrinsicSDNode>(N); |
14256 | Chain = Intrin->getChain(); |
14257 | // Intrin->getBasePtr() oddly does not get what we want. |
14258 | Base = Intrin->getOperand(3); |
14259 | MMO = Intrin->getMemOperand(); |
14260 | SrcOpnd = 2; |
14261 | break; |
14262 | } |
14263 | } |
14264 | |
14265 | SDValue Src = N->getOperand(SrcOpnd); |
14266 | MVT VecTy = Src.getValueType().getSimpleVT(); |
14267 | |
14268 | // Do not expand to PPCISD::XXSWAPD and PPCISD::STXVD2X when the load is |
14269 | // aligned and the type is a vector with elements up to 4 bytes |
14270 | if (Subtarget.needsSwapsForVSXMemOps() && MMO->getAlign() >= Align(16) && |
14271 | VecTy.getScalarSizeInBits() <= 32) { |
14272 | return SDValue(); |
14273 | } |
14274 | |
14275 | // All stores are done as v2f64 and possible bit cast. |
14276 | if (VecTy != MVT::v2f64) { |
14277 | Src = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Src); |
14278 | DCI.AddToWorklist(Src.getNode()); |
14279 | } |
14280 | |
14281 | SDValue Swap = DAG.getNode(PPCISD::XXSWAPD, dl, |
14282 | DAG.getVTList(MVT::v2f64, MVT::Other), Chain, Src); |
14283 | DCI.AddToWorklist(Swap.getNode()); |
14284 | Chain = Swap.getValue(1); |
14285 | SDValue StoreOps[] = { Chain, Swap, Base }; |
14286 | SDValue Store = DAG.getMemIntrinsicNode(PPCISD::STXVD2X, dl, |
14287 | DAG.getVTList(MVT::Other), |
14288 | StoreOps, VecTy, MMO); |
14289 | DCI.AddToWorklist(Store.getNode()); |
14290 | return Store; |
14291 | } |
14292 | |
14293 | // Handle DAG combine for STORE (FP_TO_INT F). |
14294 | SDValue PPCTargetLowering::combineStoreFPToInt(SDNode *N, |
14295 | DAGCombinerInfo &DCI) const { |
14296 | |
14297 | SelectionDAG &DAG = DCI.DAG; |
14298 | SDLoc dl(N); |
14299 | unsigned Opcode = N->getOperand(1).getOpcode(); |
14300 | |
14301 | assert((Opcode == ISD::FP_TO_SINT || Opcode == ISD::FP_TO_UINT)(((Opcode == ISD::FP_TO_SINT || Opcode == ISD::FP_TO_UINT) && "Not a FP_TO_INT Instruction!") ? static_cast<void> (0 ) : __assert_fail ("(Opcode == ISD::FP_TO_SINT || Opcode == ISD::FP_TO_UINT) && \"Not a FP_TO_INT Instruction!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14302, __PRETTY_FUNCTION__)) |
14302 | && "Not a FP_TO_INT Instruction!")(((Opcode == ISD::FP_TO_SINT || Opcode == ISD::FP_TO_UINT) && "Not a FP_TO_INT Instruction!") ? static_cast<void> (0 ) : __assert_fail ("(Opcode == ISD::FP_TO_SINT || Opcode == ISD::FP_TO_UINT) && \"Not a FP_TO_INT Instruction!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14302, __PRETTY_FUNCTION__)); |
14303 | |
14304 | SDValue Val = N->getOperand(1).getOperand(0); |
14305 | EVT Op1VT = N->getOperand(1).getValueType(); |
14306 | EVT ResVT = Val.getValueType(); |
14307 | |
14308 | if (!isTypeLegal(ResVT)) |
14309 | return SDValue(); |
14310 | |
14311 | // Only perform combine for conversion to i64/i32 or power9 i16/i8. |
14312 | bool ValidTypeForStoreFltAsInt = |
14313 | (Op1VT == MVT::i32 || Op1VT == MVT::i64 || |
14314 | (Subtarget.hasP9Vector() && (Op1VT == MVT::i16 || Op1VT == MVT::i8))); |
14315 | |
14316 | if (ResVT == MVT::ppcf128 || !Subtarget.hasP8Vector() || |
14317 | cast<StoreSDNode>(N)->isTruncatingStore() || !ValidTypeForStoreFltAsInt) |
14318 | return SDValue(); |
14319 | |
14320 | // Extend f32 values to f64 |
14321 | if (ResVT.getScalarSizeInBits() == 32) { |
14322 | Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val); |
14323 | DCI.AddToWorklist(Val.getNode()); |
14324 | } |
14325 | |
14326 | // Set signed or unsigned conversion opcode. |
14327 | unsigned ConvOpcode = (Opcode == ISD::FP_TO_SINT) ? |
14328 | PPCISD::FP_TO_SINT_IN_VSR : |
14329 | PPCISD::FP_TO_UINT_IN_VSR; |
14330 | |
14331 | Val = DAG.getNode(ConvOpcode, |
14332 | dl, ResVT == MVT::f128 ? MVT::f128 : MVT::f64, Val); |
14333 | DCI.AddToWorklist(Val.getNode()); |
14334 | |
14335 | // Set number of bytes being converted. |
14336 | unsigned ByteSize = Op1VT.getScalarSizeInBits() / 8; |
14337 | SDValue Ops[] = { N->getOperand(0), Val, N->getOperand(2), |
14338 | DAG.getIntPtrConstant(ByteSize, dl, false), |
14339 | DAG.getValueType(Op1VT) }; |
14340 | |
14341 | Val = DAG.getMemIntrinsicNode(PPCISD::ST_VSR_SCAL_INT, dl, |
14342 | DAG.getVTList(MVT::Other), Ops, |
14343 | cast<StoreSDNode>(N)->getMemoryVT(), |
14344 | cast<StoreSDNode>(N)->getMemOperand()); |
14345 | |
14346 | DCI.AddToWorklist(Val.getNode()); |
14347 | return Val; |
14348 | } |
14349 | |
14350 | static bool isAlternatingShuffMask(const ArrayRef<int> &Mask, int NumElts) { |
14351 | // Check that the source of the element keeps flipping |
14352 | // (i.e. Mask[i] < NumElts -> Mask[i+i] >= NumElts). |
14353 | bool PrevElemFromFirstVec = Mask[0] < NumElts; |
14354 | for (int i = 1, e = Mask.size(); i < e; i++) { |
14355 | if (PrevElemFromFirstVec && Mask[i] < NumElts) |
14356 | return false; |
14357 | if (!PrevElemFromFirstVec && Mask[i] >= NumElts) |
14358 | return false; |
14359 | PrevElemFromFirstVec = !PrevElemFromFirstVec; |
14360 | } |
14361 | return true; |
14362 | } |
14363 | |
14364 | static bool isSplatBV(SDValue Op) { |
14365 | if (Op.getOpcode() != ISD::BUILD_VECTOR) |
14366 | return false; |
14367 | SDValue FirstOp; |
14368 | |
14369 | // Find first non-undef input. |
14370 | for (int i = 0, e = Op.getNumOperands(); i < e; i++) { |
14371 | FirstOp = Op.getOperand(i); |
14372 | if (!FirstOp.isUndef()) |
14373 | break; |
14374 | } |
14375 | |
14376 | // All inputs are undef or the same as the first non-undef input. |
14377 | for (int i = 1, e = Op.getNumOperands(); i < e; i++) |
14378 | if (Op.getOperand(i) != FirstOp && !Op.getOperand(i).isUndef()) |
14379 | return false; |
14380 | return true; |
14381 | } |
14382 | |
14383 | static SDValue isScalarToVec(SDValue Op) { |
14384 | if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR) |
14385 | return Op; |
14386 | if (Op.getOpcode() != ISD::BITCAST) |
14387 | return SDValue(); |
14388 | Op = Op.getOperand(0); |
14389 | if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR) |
14390 | return Op; |
14391 | return SDValue(); |
14392 | } |
14393 | |
14394 | static void fixupShuffleMaskForPermutedSToV(SmallVectorImpl<int> &ShuffV, |
14395 | int LHSMaxIdx, int RHSMinIdx, |
14396 | int RHSMaxIdx, int HalfVec) { |
14397 | for (int i = 0, e = ShuffV.size(); i < e; i++) { |
14398 | int Idx = ShuffV[i]; |
14399 | if ((Idx >= 0 && Idx < LHSMaxIdx) || (Idx >= RHSMinIdx && Idx < RHSMaxIdx)) |
14400 | ShuffV[i] += HalfVec; |
14401 | } |
14402 | return; |
14403 | } |
14404 | |
14405 | // Replace a SCALAR_TO_VECTOR with a SCALAR_TO_VECTOR_PERMUTED except if |
14406 | // the original is: |
14407 | // (<n x Ty> (scalar_to_vector (Ty (extract_elt <n x Ty> %a, C)))) |
14408 | // In such a case, just change the shuffle mask to extract the element |
14409 | // from the permuted index. |
14410 | static SDValue getSToVPermuted(SDValue OrigSToV, SelectionDAG &DAG) { |
14411 | SDLoc dl(OrigSToV); |
14412 | EVT VT = OrigSToV.getValueType(); |
14413 | assert(OrigSToV.getOpcode() == ISD::SCALAR_TO_VECTOR &&((OrigSToV.getOpcode() == ISD::SCALAR_TO_VECTOR && "Expecting a SCALAR_TO_VECTOR here" ) ? static_cast<void> (0) : __assert_fail ("OrigSToV.getOpcode() == ISD::SCALAR_TO_VECTOR && \"Expecting a SCALAR_TO_VECTOR here\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14414, __PRETTY_FUNCTION__)) |
14414 | "Expecting a SCALAR_TO_VECTOR here")((OrigSToV.getOpcode() == ISD::SCALAR_TO_VECTOR && "Expecting a SCALAR_TO_VECTOR here" ) ? static_cast<void> (0) : __assert_fail ("OrigSToV.getOpcode() == ISD::SCALAR_TO_VECTOR && \"Expecting a SCALAR_TO_VECTOR here\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14414, __PRETTY_FUNCTION__)); |
14415 | SDValue Input = OrigSToV.getOperand(0); |
14416 | |
14417 | if (Input.getOpcode() == ISD::EXTRACT_VECTOR_ELT) { |
14418 | ConstantSDNode *Idx = dyn_cast<ConstantSDNode>(Input.getOperand(1)); |
14419 | SDValue OrigVector = Input.getOperand(0); |
14420 | |
14421 | // Can't handle non-const element indices or different vector types |
14422 | // for the input to the extract and the output of the scalar_to_vector. |
14423 | if (Idx && VT == OrigVector.getValueType()) { |
14424 | SmallVector<int, 16> NewMask(VT.getVectorNumElements(), -1); |
14425 | NewMask[VT.getVectorNumElements() / 2] = Idx->getZExtValue(); |
14426 | return DAG.getVectorShuffle(VT, dl, OrigVector, OrigVector, NewMask); |
14427 | } |
14428 | } |
14429 | return DAG.getNode(PPCISD::SCALAR_TO_VECTOR_PERMUTED, dl, VT, |
14430 | OrigSToV.getOperand(0)); |
14431 | } |
14432 | |
14433 | // On little endian subtargets, combine shuffles such as: |
14434 | // vector_shuffle<16,1,17,3,18,5,19,7,20,9,21,11,22,13,23,15>, <zero>, %b |
14435 | // into: |
14436 | // vector_shuffle<16,0,17,1,18,2,19,3,20,4,21,5,22,6,23,7>, <zero>, %b |
14437 | // because the latter can be matched to a single instruction merge. |
14438 | // Furthermore, SCALAR_TO_VECTOR on little endian always involves a permute |
14439 | // to put the value into element zero. Adjust the shuffle mask so that the |
14440 | // vector can remain in permuted form (to prevent a swap prior to a shuffle). |
14441 | SDValue PPCTargetLowering::combineVectorShuffle(ShuffleVectorSDNode *SVN, |
14442 | SelectionDAG &DAG) const { |
14443 | SDValue LHS = SVN->getOperand(0); |
14444 | SDValue RHS = SVN->getOperand(1); |
14445 | auto Mask = SVN->getMask(); |
14446 | int NumElts = LHS.getValueType().getVectorNumElements(); |
14447 | SDValue Res(SVN, 0); |
14448 | SDLoc dl(SVN); |
14449 | |
14450 | // None of these combines are useful on big endian systems since the ISA |
14451 | // already has a big endian bias. |
14452 | if (!Subtarget.isLittleEndian() || !Subtarget.hasVSX()) |
14453 | return Res; |
14454 | |
14455 | // If this is not a shuffle of a shuffle and the first element comes from |
14456 | // the second vector, canonicalize to the commuted form. This will make it |
14457 | // more likely to match one of the single instruction patterns. |
14458 | if (Mask[0] >= NumElts && LHS.getOpcode() != ISD::VECTOR_SHUFFLE && |
14459 | RHS.getOpcode() != ISD::VECTOR_SHUFFLE) { |
14460 | std::swap(LHS, RHS); |
14461 | Res = DAG.getCommutedVectorShuffle(*SVN); |
14462 | Mask = cast<ShuffleVectorSDNode>(Res)->getMask(); |
14463 | } |
14464 | |
14465 | // Adjust the shuffle mask if either input vector comes from a |
14466 | // SCALAR_TO_VECTOR and keep the respective input vector in permuted |
14467 | // form (to prevent the need for a swap). |
14468 | SmallVector<int, 16> ShuffV(Mask.begin(), Mask.end()); |
14469 | SDValue SToVLHS = isScalarToVec(LHS); |
14470 | SDValue SToVRHS = isScalarToVec(RHS); |
14471 | if (SToVLHS || SToVRHS) { |
14472 | int NumEltsIn = SToVLHS ? SToVLHS.getValueType().getVectorNumElements() |
14473 | : SToVRHS.getValueType().getVectorNumElements(); |
14474 | int NumEltsOut = ShuffV.size(); |
14475 | |
14476 | // Initially assume that neither input is permuted. These will be adjusted |
14477 | // accordingly if either input is. |
14478 | int LHSMaxIdx = -1; |
14479 | int RHSMinIdx = -1; |
14480 | int RHSMaxIdx = -1; |
14481 | int HalfVec = LHS.getValueType().getVectorNumElements() / 2; |
14482 | |
14483 | // Get the permuted scalar to vector nodes for the source(s) that come from |
14484 | // ISD::SCALAR_TO_VECTOR. |
14485 | if (SToVLHS) { |
14486 | // Set up the values for the shuffle vector fixup. |
14487 | LHSMaxIdx = NumEltsOut / NumEltsIn; |
14488 | SToVLHS = getSToVPermuted(SToVLHS, DAG); |
14489 | if (SToVLHS.getValueType() != LHS.getValueType()) |
14490 | SToVLHS = DAG.getBitcast(LHS.getValueType(), SToVLHS); |
14491 | LHS = SToVLHS; |
14492 | } |
14493 | if (SToVRHS) { |
14494 | RHSMinIdx = NumEltsOut; |
14495 | RHSMaxIdx = NumEltsOut / NumEltsIn + RHSMinIdx; |
14496 | SToVRHS = getSToVPermuted(SToVRHS, DAG); |
14497 | if (SToVRHS.getValueType() != RHS.getValueType()) |
14498 | SToVRHS = DAG.getBitcast(RHS.getValueType(), SToVRHS); |
14499 | RHS = SToVRHS; |
14500 | } |
14501 | |
14502 | // Fix up the shuffle mask to reflect where the desired element actually is. |
14503 | // The minimum and maximum indices that correspond to element zero for both |
14504 | // the LHS and RHS are computed and will control which shuffle mask entries |
14505 | // are to be changed. For example, if the RHS is permuted, any shuffle mask |
14506 | // entries in the range [RHSMinIdx,RHSMaxIdx) will be incremented by |
14507 | // HalfVec to refer to the corresponding element in the permuted vector. |
14508 | fixupShuffleMaskForPermutedSToV(ShuffV, LHSMaxIdx, RHSMinIdx, RHSMaxIdx, |
14509 | HalfVec); |
14510 | Res = DAG.getVectorShuffle(SVN->getValueType(0), dl, LHS, RHS, ShuffV); |
14511 | |
14512 | // We may have simplified away the shuffle. We won't be able to do anything |
14513 | // further with it here. |
14514 | if (!isa<ShuffleVectorSDNode>(Res)) |
14515 | return Res; |
14516 | Mask = cast<ShuffleVectorSDNode>(Res)->getMask(); |
14517 | } |
14518 | |
14519 | // The common case after we commuted the shuffle is that the RHS is a splat |
14520 | // and we have elements coming in from the splat at indices that are not |
14521 | // conducive to using a merge. |
14522 | // Example: |
14523 | // vector_shuffle<0,17,1,19,2,21,3,23,4,25,5,27,6,29,7,31> t1, <zero> |
14524 | if (!isSplatBV(RHS)) |
14525 | return Res; |
14526 | |
14527 | // We are looking for a mask such that all even elements are from |
14528 | // one vector and all odd elements from the other. |
14529 | if (!isAlternatingShuffMask(Mask, NumElts)) |
14530 | return Res; |
14531 | |
14532 | // Adjust the mask so we are pulling in the same index from the splat |
14533 | // as the index from the interesting vector in consecutive elements. |
14534 | // Example (even elements from first vector): |
14535 | // vector_shuffle<0,16,1,17,2,18,3,19,4,20,5,21,6,22,7,23> t1, <zero> |
14536 | if (Mask[0] < NumElts) |
14537 | for (int i = 1, e = Mask.size(); i < e; i += 2) |
14538 | ShuffV[i] = (ShuffV[i - 1] + NumElts); |
14539 | // Example (odd elements from first vector): |
14540 | // vector_shuffle<16,0,17,1,18,2,19,3,20,4,21,5,22,6,23,7> t1, <zero> |
14541 | else |
14542 | for (int i = 0, e = Mask.size(); i < e; i += 2) |
14543 | ShuffV[i] = (ShuffV[i + 1] + NumElts); |
14544 | |
14545 | // If the RHS has undefs, we need to remove them since we may have created |
14546 | // a shuffle that adds those instead of the splat value. |
14547 | SDValue SplatVal = cast<BuildVectorSDNode>(RHS.getNode())->getSplatValue(); |
14548 | RHS = DAG.getSplatBuildVector(RHS.getValueType(), dl, SplatVal); |
14549 | |
14550 | Res = DAG.getVectorShuffle(SVN->getValueType(0), dl, LHS, RHS, ShuffV); |
14551 | return Res; |
14552 | } |
14553 | |
14554 | SDValue PPCTargetLowering::combineVReverseMemOP(ShuffleVectorSDNode *SVN, |
14555 | LSBaseSDNode *LSBase, |
14556 | DAGCombinerInfo &DCI) const { |
14557 | assert((ISD::isNormalLoad(LSBase) || ISD::isNormalStore(LSBase)) &&(((ISD::isNormalLoad(LSBase) || ISD::isNormalStore(LSBase)) && "Not a reverse memop pattern!") ? static_cast<void> (0 ) : __assert_fail ("(ISD::isNormalLoad(LSBase) || ISD::isNormalStore(LSBase)) && \"Not a reverse memop pattern!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14558, __PRETTY_FUNCTION__)) |
14558 | "Not a reverse memop pattern!")(((ISD::isNormalLoad(LSBase) || ISD::isNormalStore(LSBase)) && "Not a reverse memop pattern!") ? static_cast<void> (0 ) : __assert_fail ("(ISD::isNormalLoad(LSBase) || ISD::isNormalStore(LSBase)) && \"Not a reverse memop pattern!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14558, __PRETTY_FUNCTION__)); |
14559 | |
14560 | auto IsElementReverse = [](const ShuffleVectorSDNode *SVN) -> bool { |
14561 | auto Mask = SVN->getMask(); |
14562 | int i = 0; |
14563 | auto I = Mask.rbegin(); |
14564 | auto E = Mask.rend(); |
14565 | |
14566 | for (; I != E; ++I) { |
14567 | if (*I != i) |
14568 | return false; |
14569 | i++; |
14570 | } |
14571 | return true; |
14572 | }; |
14573 | |
14574 | SelectionDAG &DAG = DCI.DAG; |
14575 | EVT VT = SVN->getValueType(0); |
14576 | |
14577 | if (!isTypeLegal(VT) || !Subtarget.isLittleEndian() || !Subtarget.hasVSX()) |
14578 | return SDValue(); |
14579 | |
14580 | // Before P9, we have PPCVSXSwapRemoval pass to hack the element order. |
14581 | // See comment in PPCVSXSwapRemoval.cpp. |
14582 | // It is conflict with PPCVSXSwapRemoval opt. So we don't do it. |
14583 | if (!Subtarget.hasP9Vector()) |
14584 | return SDValue(); |
14585 | |
14586 | if(!IsElementReverse(SVN)) |
14587 | return SDValue(); |
14588 | |
14589 | if (LSBase->getOpcode() == ISD::LOAD) { |
14590 | SDLoc dl(SVN); |
14591 | SDValue LoadOps[] = {LSBase->getChain(), LSBase->getBasePtr()}; |
14592 | return DAG.getMemIntrinsicNode( |
14593 | PPCISD::LOAD_VEC_BE, dl, DAG.getVTList(VT, MVT::Other), LoadOps, |
14594 | LSBase->getMemoryVT(), LSBase->getMemOperand()); |
14595 | } |
14596 | |
14597 | if (LSBase->getOpcode() == ISD::STORE) { |
14598 | SDLoc dl(LSBase); |
14599 | SDValue StoreOps[] = {LSBase->getChain(), SVN->getOperand(0), |
14600 | LSBase->getBasePtr()}; |
14601 | return DAG.getMemIntrinsicNode( |
14602 | PPCISD::STORE_VEC_BE, dl, DAG.getVTList(MVT::Other), StoreOps, |
14603 | LSBase->getMemoryVT(), LSBase->getMemOperand()); |
14604 | } |
14605 | |
14606 | llvm_unreachable("Expected a load or store node here")::llvm::llvm_unreachable_internal("Expected a load or store node here" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14606); |
14607 | } |
14608 | |
14609 | SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N, |
14610 | DAGCombinerInfo &DCI) const { |
14611 | SelectionDAG &DAG = DCI.DAG; |
14612 | SDLoc dl(N); |
14613 | switch (N->getOpcode()) { |
14614 | default: break; |
14615 | case ISD::ADD: |
14616 | return combineADD(N, DCI); |
14617 | case ISD::SHL: |
14618 | return combineSHL(N, DCI); |
14619 | case ISD::SRA: |
14620 | return combineSRA(N, DCI); |
14621 | case ISD::SRL: |
14622 | return combineSRL(N, DCI); |
14623 | case ISD::MUL: |
14624 | return combineMUL(N, DCI); |
14625 | case ISD::FMA: |
14626 | case PPCISD::FNMSUB: |
14627 | return combineFMALike(N, DCI); |
14628 | case PPCISD::SHL: |
14629 | if (isNullConstant(N->getOperand(0))) // 0 << V -> 0. |
14630 | return N->getOperand(0); |
14631 | break; |
14632 | case PPCISD::SRL: |
14633 | if (isNullConstant(N->getOperand(0))) // 0 >>u V -> 0. |
14634 | return N->getOperand(0); |
14635 | break; |
14636 | case PPCISD::SRA: |
14637 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) { |
14638 | if (C->isNullValue() || // 0 >>s V -> 0. |
14639 | C->isAllOnesValue()) // -1 >>s V -> -1. |
14640 | return N->getOperand(0); |
14641 | } |
14642 | break; |
14643 | case ISD::SIGN_EXTEND: |
14644 | case ISD::ZERO_EXTEND: |
14645 | case ISD::ANY_EXTEND: |
14646 | return DAGCombineExtBoolTrunc(N, DCI); |
14647 | case ISD::TRUNCATE: |
14648 | return combineTRUNCATE(N, DCI); |
14649 | case ISD::SETCC: |
14650 | if (SDValue CSCC = combineSetCC(N, DCI)) |
14651 | return CSCC; |
14652 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
14653 | case ISD::SELECT_CC: |
14654 | return DAGCombineTruncBoolExt(N, DCI); |
14655 | case ISD::SINT_TO_FP: |
14656 | case ISD::UINT_TO_FP: |
14657 | return combineFPToIntToFP(N, DCI); |
14658 | case ISD::VECTOR_SHUFFLE: |
14659 | if (ISD::isNormalLoad(N->getOperand(0).getNode())) { |
14660 | LSBaseSDNode* LSBase = cast<LSBaseSDNode>(N->getOperand(0)); |
14661 | return combineVReverseMemOP(cast<ShuffleVectorSDNode>(N), LSBase, DCI); |
14662 | } |
14663 | return combineVectorShuffle(cast<ShuffleVectorSDNode>(N), DCI.DAG); |
14664 | case ISD::STORE: { |
14665 | |
14666 | EVT Op1VT = N->getOperand(1).getValueType(); |
14667 | unsigned Opcode = N->getOperand(1).getOpcode(); |
14668 | |
14669 | if (Opcode == ISD::FP_TO_SINT || Opcode == ISD::FP_TO_UINT) { |
14670 | SDValue Val= combineStoreFPToInt(N, DCI); |
14671 | if (Val) |
14672 | return Val; |
14673 | } |
14674 | |
14675 | if (Opcode == ISD::VECTOR_SHUFFLE && ISD::isNormalStore(N)) { |
14676 | ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N->getOperand(1)); |
14677 | SDValue Val= combineVReverseMemOP(SVN, cast<LSBaseSDNode>(N), DCI); |
14678 | if (Val) |
14679 | return Val; |
14680 | } |
14681 | |
14682 | // Turn STORE (BSWAP) -> sthbrx/stwbrx. |
14683 | if (cast<StoreSDNode>(N)->isUnindexed() && Opcode == ISD::BSWAP && |
14684 | N->getOperand(1).getNode()->hasOneUse() && |
14685 | (Op1VT == MVT::i32 || Op1VT == MVT::i16 || |
14686 | (Subtarget.hasLDBRX() && Subtarget.isPPC64() && Op1VT == MVT::i64))) { |
14687 | |
14688 | // STBRX can only handle simple types and it makes no sense to store less |
14689 | // two bytes in byte-reversed order. |
14690 | EVT mVT = cast<StoreSDNode>(N)->getMemoryVT(); |
14691 | if (mVT.isExtended() || mVT.getSizeInBits() < 16) |
14692 | break; |
14693 | |
14694 | SDValue BSwapOp = N->getOperand(1).getOperand(0); |
14695 | // Do an any-extend to 32-bits if this is a half-word input. |
14696 | if (BSwapOp.getValueType() == MVT::i16) |
14697 | BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp); |
14698 | |
14699 | // If the type of BSWAP operand is wider than stored memory width |
14700 | // it need to be shifted to the right side before STBRX. |
14701 | if (Op1VT.bitsGT(mVT)) { |
14702 | int Shift = Op1VT.getSizeInBits() - mVT.getSizeInBits(); |
14703 | BSwapOp = DAG.getNode(ISD::SRL, dl, Op1VT, BSwapOp, |
14704 | DAG.getConstant(Shift, dl, MVT::i32)); |
14705 | // Need to truncate if this is a bswap of i64 stored as i32/i16. |
14706 | if (Op1VT == MVT::i64) |
14707 | BSwapOp = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BSwapOp); |
14708 | } |
14709 | |
14710 | SDValue Ops[] = { |
14711 | N->getOperand(0), BSwapOp, N->getOperand(2), DAG.getValueType(mVT) |
14712 | }; |
14713 | return |
14714 | DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other), |
14715 | Ops, cast<StoreSDNode>(N)->getMemoryVT(), |
14716 | cast<StoreSDNode>(N)->getMemOperand()); |
14717 | } |
14718 | |
14719 | // STORE Constant:i32<0> -> STORE<trunc to i32> Constant:i64<0> |
14720 | // So it can increase the chance of CSE constant construction. |
14721 | if (Subtarget.isPPC64() && !DCI.isBeforeLegalize() && |
14722 | isa<ConstantSDNode>(N->getOperand(1)) && Op1VT == MVT::i32) { |
14723 | // Need to sign-extended to 64-bits to handle negative values. |
14724 | EVT MemVT = cast<StoreSDNode>(N)->getMemoryVT(); |
14725 | uint64_t Val64 = SignExtend64(N->getConstantOperandVal(1), |
14726 | MemVT.getSizeInBits()); |
14727 | SDValue Const64 = DAG.getConstant(Val64, dl, MVT::i64); |
14728 | |
14729 | // DAG.getTruncStore() can't be used here because it doesn't accept |
14730 | // the general (base + offset) addressing mode. |
14731 | // So we use UpdateNodeOperands and setTruncatingStore instead. |
14732 | DAG.UpdateNodeOperands(N, N->getOperand(0), Const64, N->getOperand(2), |
14733 | N->getOperand(3)); |
14734 | cast<StoreSDNode>(N)->setTruncatingStore(true); |
14735 | return SDValue(N, 0); |
14736 | } |
14737 | |
14738 | // For little endian, VSX stores require generating xxswapd/lxvd2x. |
14739 | // Not needed on ISA 3.0 based CPUs since we have a non-permuting store. |
14740 | if (Op1VT.isSimple()) { |
14741 | MVT StoreVT = Op1VT.getSimpleVT(); |
14742 | if (Subtarget.needsSwapsForVSXMemOps() && |
14743 | (StoreVT == MVT::v2f64 || StoreVT == MVT::v2i64 || |
14744 | StoreVT == MVT::v4f32 || StoreVT == MVT::v4i32)) |
14745 | return expandVSXStoreForLE(N, DCI); |
14746 | } |
14747 | break; |
14748 | } |
14749 | case ISD::LOAD: { |
14750 | LoadSDNode *LD = cast<LoadSDNode>(N); |
14751 | EVT VT = LD->getValueType(0); |
14752 | |
14753 | // For little endian, VSX loads require generating lxvd2x/xxswapd. |
14754 | // Not needed on ISA 3.0 based CPUs since we have a non-permuting load. |
14755 | if (VT.isSimple()) { |
14756 | MVT LoadVT = VT.getSimpleVT(); |
14757 | if (Subtarget.needsSwapsForVSXMemOps() && |
14758 | (LoadVT == MVT::v2f64 || LoadVT == MVT::v2i64 || |
14759 | LoadVT == MVT::v4f32 || LoadVT == MVT::v4i32)) |
14760 | return expandVSXLoadForLE(N, DCI); |
14761 | } |
14762 | |
14763 | // We sometimes end up with a 64-bit integer load, from which we extract |
14764 | // two single-precision floating-point numbers. This happens with |
14765 | // std::complex<float>, and other similar structures, because of the way we |
14766 | // canonicalize structure copies. However, if we lack direct moves, |
14767 | // then the final bitcasts from the extracted integer values to the |
14768 | // floating-point numbers turn into store/load pairs. Even with direct moves, |
14769 | // just loading the two floating-point numbers is likely better. |
14770 | auto ReplaceTwoFloatLoad = [&]() { |
14771 | if (VT != MVT::i64) |
14772 | return false; |
14773 | |
14774 | if (LD->getExtensionType() != ISD::NON_EXTLOAD || |
14775 | LD->isVolatile()) |
14776 | return false; |
14777 | |
14778 | // We're looking for a sequence like this: |
14779 | // t13: i64,ch = load<LD8[%ref.tmp]> t0, t6, undef:i64 |
14780 | // t16: i64 = srl t13, Constant:i32<32> |
14781 | // t17: i32 = truncate t16 |
14782 | // t18: f32 = bitcast t17 |
14783 | // t19: i32 = truncate t13 |
14784 | // t20: f32 = bitcast t19 |
14785 | |
14786 | if (!LD->hasNUsesOfValue(2, 0)) |
14787 | return false; |
14788 | |
14789 | auto UI = LD->use_begin(); |
14790 | while (UI.getUse().getResNo() != 0) ++UI; |
14791 | SDNode *Trunc = *UI++; |
14792 | while (UI.getUse().getResNo() != 0) ++UI; |
14793 | SDNode *RightShift = *UI; |
14794 | if (Trunc->getOpcode() != ISD::TRUNCATE) |
14795 | std::swap(Trunc, RightShift); |
14796 | |
14797 | if (Trunc->getOpcode() != ISD::TRUNCATE || |
14798 | Trunc->getValueType(0) != MVT::i32 || |
14799 | !Trunc->hasOneUse()) |
14800 | return false; |
14801 | if (RightShift->getOpcode() != ISD::SRL || |
14802 | !isa<ConstantSDNode>(RightShift->getOperand(1)) || |
14803 | RightShift->getConstantOperandVal(1) != 32 || |
14804 | !RightShift->hasOneUse()) |
14805 | return false; |
14806 | |
14807 | SDNode *Trunc2 = *RightShift->use_begin(); |
14808 | if (Trunc2->getOpcode() != ISD::TRUNCATE || |
14809 | Trunc2->getValueType(0) != MVT::i32 || |
14810 | !Trunc2->hasOneUse()) |
14811 | return false; |
14812 | |
14813 | SDNode *Bitcast = *Trunc->use_begin(); |
14814 | SDNode *Bitcast2 = *Trunc2->use_begin(); |
14815 | |
14816 | if (Bitcast->getOpcode() != ISD::BITCAST || |
14817 | Bitcast->getValueType(0) != MVT::f32) |
14818 | return false; |
14819 | if (Bitcast2->getOpcode() != ISD::BITCAST || |
14820 | Bitcast2->getValueType(0) != MVT::f32) |
14821 | return false; |
14822 | |
14823 | if (Subtarget.isLittleEndian()) |
14824 | std::swap(Bitcast, Bitcast2); |
14825 | |
14826 | // Bitcast has the second float (in memory-layout order) and Bitcast2 |
14827 | // has the first one. |
14828 | |
14829 | SDValue BasePtr = LD->getBasePtr(); |
14830 | if (LD->isIndexed()) { |
14831 | assert(LD->getAddressingMode() == ISD::PRE_INC &&((LD->getAddressingMode() == ISD::PRE_INC && "Non-pre-inc AM on PPC?" ) ? static_cast<void> (0) : __assert_fail ("LD->getAddressingMode() == ISD::PRE_INC && \"Non-pre-inc AM on PPC?\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14832, __PRETTY_FUNCTION__)) |
14832 | "Non-pre-inc AM on PPC?")((LD->getAddressingMode() == ISD::PRE_INC && "Non-pre-inc AM on PPC?" ) ? static_cast<void> (0) : __assert_fail ("LD->getAddressingMode() == ISD::PRE_INC && \"Non-pre-inc AM on PPC?\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14832, __PRETTY_FUNCTION__)); |
14833 | BasePtr = |
14834 | DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr, |
14835 | LD->getOffset()); |
14836 | } |
14837 | |
14838 | auto MMOFlags = |
14839 | LD->getMemOperand()->getFlags() & ~MachineMemOperand::MOVolatile; |
14840 | SDValue FloatLoad = DAG.getLoad(MVT::f32, dl, LD->getChain(), BasePtr, |
14841 | LD->getPointerInfo(), LD->getAlignment(), |
14842 | MMOFlags, LD->getAAInfo()); |
14843 | SDValue AddPtr = |
14844 | DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), |
14845 | BasePtr, DAG.getIntPtrConstant(4, dl)); |
14846 | SDValue FloatLoad2 = DAG.getLoad( |
14847 | MVT::f32, dl, SDValue(FloatLoad.getNode(), 1), AddPtr, |
14848 | LD->getPointerInfo().getWithOffset(4), |
14849 | MinAlign(LD->getAlignment(), 4), MMOFlags, LD->getAAInfo()); |
14850 | |
14851 | if (LD->isIndexed()) { |
14852 | // Note that DAGCombine should re-form any pre-increment load(s) from |
14853 | // what is produced here if that makes sense. |
14854 | DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), BasePtr); |
14855 | } |
14856 | |
14857 | DCI.CombineTo(Bitcast2, FloatLoad); |
14858 | DCI.CombineTo(Bitcast, FloatLoad2); |
14859 | |
14860 | DAG.ReplaceAllUsesOfValueWith(SDValue(LD, LD->isIndexed() ? 2 : 1), |
14861 | SDValue(FloatLoad2.getNode(), 1)); |
14862 | return true; |
14863 | }; |
14864 | |
14865 | if (ReplaceTwoFloatLoad()) |
14866 | return SDValue(N, 0); |
14867 | |
14868 | EVT MemVT = LD->getMemoryVT(); |
14869 | Type *Ty = MemVT.getTypeForEVT(*DAG.getContext()); |
14870 | Align ABIAlignment = DAG.getDataLayout().getABITypeAlign(Ty); |
14871 | if (LD->isUnindexed() && VT.isVector() && |
14872 | ((Subtarget.hasAltivec() && ISD::isNON_EXTLoad(N) && |
14873 | // P8 and later hardware should just use LOAD. |
14874 | !Subtarget.hasP8Vector() && |
14875 | (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 || |
14876 | VT == MVT::v4f32))) && |
14877 | LD->getAlign() < ABIAlignment) { |
14878 | // This is a type-legal unaligned Altivec load. |
14879 | SDValue Chain = LD->getChain(); |
14880 | SDValue Ptr = LD->getBasePtr(); |
14881 | bool isLittleEndian = Subtarget.isLittleEndian(); |
14882 | |
14883 | // This implements the loading of unaligned vectors as described in |
14884 | // the venerable Apple Velocity Engine overview. Specifically: |
14885 | // https://developer.apple.com/hardwaredrivers/ve/alignment.html |
14886 | // https://developer.apple.com/hardwaredrivers/ve/code_optimization.html |
14887 | // |
14888 | // The general idea is to expand a sequence of one or more unaligned |
14889 | // loads into an alignment-based permutation-control instruction (lvsl |
14890 | // or lvsr), a series of regular vector loads (which always truncate |
14891 | // their input address to an aligned address), and a series of |
14892 | // permutations. The results of these permutations are the requested |
14893 | // loaded values. The trick is that the last "extra" load is not taken |
14894 | // from the address you might suspect (sizeof(vector) bytes after the |
14895 | // last requested load), but rather sizeof(vector) - 1 bytes after the |
14896 | // last requested vector. The point of this is to avoid a page fault if |
14897 | // the base address happened to be aligned. This works because if the |
14898 | // base address is aligned, then adding less than a full vector length |
14899 | // will cause the last vector in the sequence to be (re)loaded. |
14900 | // Otherwise, the next vector will be fetched as you might suspect was |
14901 | // necessary. |
14902 | |
14903 | // We might be able to reuse the permutation generation from |
14904 | // a different base address offset from this one by an aligned amount. |
14905 | // The INTRINSIC_WO_CHAIN DAG combine will attempt to perform this |
14906 | // optimization later. |
14907 | Intrinsic::ID Intr, IntrLD, IntrPerm; |
14908 | MVT PermCntlTy, PermTy, LDTy; |
14909 | Intr = isLittleEndian ? Intrinsic::ppc_altivec_lvsr |
14910 | : Intrinsic::ppc_altivec_lvsl; |
14911 | IntrLD = Intrinsic::ppc_altivec_lvx; |
14912 | IntrPerm = Intrinsic::ppc_altivec_vperm; |
14913 | PermCntlTy = MVT::v16i8; |
14914 | PermTy = MVT::v4i32; |
14915 | LDTy = MVT::v4i32; |
14916 | |
14917 | SDValue PermCntl = BuildIntrinsicOp(Intr, Ptr, DAG, dl, PermCntlTy); |
14918 | |
14919 | // Create the new MMO for the new base load. It is like the original MMO, |
14920 | // but represents an area in memory almost twice the vector size centered |
14921 | // on the original address. If the address is unaligned, we might start |
14922 | // reading up to (sizeof(vector)-1) bytes below the address of the |
14923 | // original unaligned load. |
14924 | MachineFunction &MF = DAG.getMachineFunction(); |
14925 | MachineMemOperand *BaseMMO = |
14926 | MF.getMachineMemOperand(LD->getMemOperand(), |
14927 | -(long)MemVT.getStoreSize()+1, |
14928 | 2*MemVT.getStoreSize()-1); |
14929 | |
14930 | // Create the new base load. |
14931 | SDValue LDXIntID = |
14932 | DAG.getTargetConstant(IntrLD, dl, getPointerTy(MF.getDataLayout())); |
14933 | SDValue BaseLoadOps[] = { Chain, LDXIntID, Ptr }; |
14934 | SDValue BaseLoad = |
14935 | DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, dl, |
14936 | DAG.getVTList(PermTy, MVT::Other), |
14937 | BaseLoadOps, LDTy, BaseMMO); |
14938 | |
14939 | // Note that the value of IncOffset (which is provided to the next |
14940 | // load's pointer info offset value, and thus used to calculate the |
14941 | // alignment), and the value of IncValue (which is actually used to |
14942 | // increment the pointer value) are different! This is because we |
14943 | // require the next load to appear to be aligned, even though it |
14944 | // is actually offset from the base pointer by a lesser amount. |
14945 | int IncOffset = VT.getSizeInBits() / 8; |
14946 | int IncValue = IncOffset; |
14947 | |
14948 | // Walk (both up and down) the chain looking for another load at the real |
14949 | // (aligned) offset (the alignment of the other load does not matter in |
14950 | // this case). If found, then do not use the offset reduction trick, as |
14951 | // that will prevent the loads from being later combined (as they would |
14952 | // otherwise be duplicates). |
14953 | if (!findConsecutiveLoad(LD, DAG)) |
14954 | --IncValue; |
14955 | |
14956 | SDValue Increment = |
14957 | DAG.getConstant(IncValue, dl, getPointerTy(MF.getDataLayout())); |
14958 | Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment); |
14959 | |
14960 | MachineMemOperand *ExtraMMO = |
14961 | MF.getMachineMemOperand(LD->getMemOperand(), |
14962 | 1, 2*MemVT.getStoreSize()-1); |
14963 | SDValue ExtraLoadOps[] = { Chain, LDXIntID, Ptr }; |
14964 | SDValue ExtraLoad = |
14965 | DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, dl, |
14966 | DAG.getVTList(PermTy, MVT::Other), |
14967 | ExtraLoadOps, LDTy, ExtraMMO); |
14968 | |
14969 | SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, |
14970 | BaseLoad.getValue(1), ExtraLoad.getValue(1)); |
14971 | |
14972 | // Because vperm has a big-endian bias, we must reverse the order |
14973 | // of the input vectors and complement the permute control vector |
14974 | // when generating little endian code. We have already handled the |
14975 | // latter by using lvsr instead of lvsl, so just reverse BaseLoad |
14976 | // and ExtraLoad here. |
14977 | SDValue Perm; |
14978 | if (isLittleEndian) |
14979 | Perm = BuildIntrinsicOp(IntrPerm, |
14980 | ExtraLoad, BaseLoad, PermCntl, DAG, dl); |
14981 | else |
14982 | Perm = BuildIntrinsicOp(IntrPerm, |
14983 | BaseLoad, ExtraLoad, PermCntl, DAG, dl); |
14984 | |
14985 | if (VT != PermTy) |
14986 | Perm = Subtarget.hasAltivec() |
14987 | ? DAG.getNode(ISD::BITCAST, dl, VT, Perm) |
14988 | : DAG.getNode(ISD::FP_ROUND, dl, VT, Perm, |
14989 | DAG.getTargetConstant(1, dl, MVT::i64)); |
14990 | // second argument is 1 because this rounding |
14991 | // is always exact. |
14992 | |
14993 | // The output of the permutation is our loaded result, the TokenFactor is |
14994 | // our new chain. |
14995 | DCI.CombineTo(N, Perm, TF); |
14996 | return SDValue(N, 0); |
14997 | } |
14998 | } |
14999 | break; |
15000 | case ISD::INTRINSIC_WO_CHAIN: { |
15001 | bool isLittleEndian = Subtarget.isLittleEndian(); |
15002 | unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); |
15003 | Intrinsic::ID Intr = (isLittleEndian ? Intrinsic::ppc_altivec_lvsr |
15004 | : Intrinsic::ppc_altivec_lvsl); |
15005 | if (IID == Intr && N->getOperand(1)->getOpcode() == ISD::ADD) { |
15006 | SDValue Add = N->getOperand(1); |
15007 | |
15008 | int Bits = 4 /* 16 byte alignment */; |
15009 | |
15010 | if (DAG.MaskedValueIsZero(Add->getOperand(1), |
15011 | APInt::getAllOnesValue(Bits /* alignment */) |
15012 | .zext(Add.getScalarValueSizeInBits()))) { |
15013 | SDNode *BasePtr = Add->getOperand(0).getNode(); |
15014 | for (SDNode::use_iterator UI = BasePtr->use_begin(), |
15015 | UE = BasePtr->use_end(); |
15016 | UI != UE; ++UI) { |
15017 | if (UI->getOpcode() == ISD::INTRINSIC_WO_CHAIN && |
15018 | cast<ConstantSDNode>(UI->getOperand(0))->getZExtValue() == |
15019 | IID) { |
15020 | // We've found another LVSL/LVSR, and this address is an aligned |
15021 | // multiple of that one. The results will be the same, so use the |
15022 | // one we've just found instead. |
15023 | |
15024 | return SDValue(*UI, 0); |
15025 | } |
15026 | } |
15027 | } |
15028 | |
15029 | if (isa<ConstantSDNode>(Add->getOperand(1))) { |
15030 | SDNode *BasePtr = Add->getOperand(0).getNode(); |
15031 | for (SDNode::use_iterator UI = BasePtr->use_begin(), |
15032 | UE = BasePtr->use_end(); UI != UE; ++UI) { |
15033 | if (UI->getOpcode() == ISD::ADD && |
15034 | isa<ConstantSDNode>(UI->getOperand(1)) && |
15035 | (cast<ConstantSDNode>(Add->getOperand(1))->getZExtValue() - |
15036 | cast<ConstantSDNode>(UI->getOperand(1))->getZExtValue()) % |
15037 | (1ULL << Bits) == 0) { |
15038 | SDNode *OtherAdd = *UI; |
15039 | for (SDNode::use_iterator VI = OtherAdd->use_begin(), |
15040 | VE = OtherAdd->use_end(); VI != VE; ++VI) { |
15041 | if (VI->getOpcode() == ISD::INTRINSIC_WO_CHAIN && |
15042 | cast<ConstantSDNode>(VI->getOperand(0))->getZExtValue() == IID) { |
15043 | return SDValue(*VI, 0); |
15044 | } |
15045 | } |
15046 | } |
15047 | } |
15048 | } |
15049 | } |
15050 | |
15051 | // Combine vmaxsw/h/b(a, a's negation) to abs(a) |
15052 | // Expose the vabsduw/h/b opportunity for down stream |
15053 | if (!DCI.isAfterLegalizeDAG() && Subtarget.hasP9Altivec() && |
15054 | (IID == Intrinsic::ppc_altivec_vmaxsw || |
15055 | IID == Intrinsic::ppc_altivec_vmaxsh || |
15056 | IID == Intrinsic::ppc_altivec_vmaxsb)) { |
15057 | SDValue V1 = N->getOperand(1); |
15058 | SDValue V2 = N->getOperand(2); |
15059 | if ((V1.getSimpleValueType() == MVT::v4i32 || |
15060 | V1.getSimpleValueType() == MVT::v8i16 || |
15061 | V1.getSimpleValueType() == MVT::v16i8) && |
15062 | V1.getSimpleValueType() == V2.getSimpleValueType()) { |
15063 | // (0-a, a) |
15064 | if (V1.getOpcode() == ISD::SUB && |
15065 | ISD::isBuildVectorAllZeros(V1.getOperand(0).getNode()) && |
15066 | V1.getOperand(1) == V2) { |
15067 | return DAG.getNode(ISD::ABS, dl, V2.getValueType(), V2); |
15068 | } |
15069 | // (a, 0-a) |
15070 | if (V2.getOpcode() == ISD::SUB && |
15071 | ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()) && |
15072 | V2.getOperand(1) == V1) { |
15073 | return DAG.getNode(ISD::ABS, dl, V1.getValueType(), V1); |
15074 | } |
15075 | // (x-y, y-x) |
15076 | if (V1.getOpcode() == ISD::SUB && V2.getOpcode() == ISD::SUB && |
15077 | V1.getOperand(0) == V2.getOperand(1) && |
15078 | V1.getOperand(1) == V2.getOperand(0)) { |
15079 | return DAG.getNode(ISD::ABS, dl, V1.getValueType(), V1); |
15080 | } |
15081 | } |
15082 | } |
15083 | } |
15084 | |
15085 | break; |
15086 | case ISD::INTRINSIC_W_CHAIN: |
15087 | // For little endian, VSX loads require generating lxvd2x/xxswapd. |
15088 | // Not needed on ISA 3.0 based CPUs since we have a non-permuting load. |
15089 | if (Subtarget.needsSwapsForVSXMemOps()) { |
15090 | switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) { |
15091 | default: |
15092 | break; |
15093 | case Intrinsic::ppc_vsx_lxvw4x: |
15094 | case Intrinsic::ppc_vsx_lxvd2x: |
15095 | return expandVSXLoadForLE(N, DCI); |
15096 | } |
15097 | } |
15098 | break; |
15099 | case ISD::INTRINSIC_VOID: |
15100 | // For little endian, VSX stores require generating xxswapd/stxvd2x. |
15101 | // Not needed on ISA 3.0 based CPUs since we have a non-permuting store. |
15102 | if (Subtarget.needsSwapsForVSXMemOps()) { |
15103 | switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) { |
15104 | default: |
15105 | break; |
15106 | case Intrinsic::ppc_vsx_stxvw4x: |
15107 | case Intrinsic::ppc_vsx_stxvd2x: |
15108 | return expandVSXStoreForLE(N, DCI); |
15109 | } |
15110 | } |
15111 | break; |
15112 | case ISD::BSWAP: |
15113 | // Turn BSWAP (LOAD) -> lhbrx/lwbrx. |
15114 | if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) && |
15115 | N->getOperand(0).hasOneUse() && |
15116 | (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16 || |
15117 | (Subtarget.hasLDBRX() && Subtarget.isPPC64() && |
15118 | N->getValueType(0) == MVT::i64))) { |
15119 | SDValue Load = N->getOperand(0); |
15120 | LoadSDNode *LD = cast<LoadSDNode>(Load); |
15121 | // Create the byte-swapping load. |
15122 | SDValue Ops[] = { |
15123 | LD->getChain(), // Chain |
15124 | LD->getBasePtr(), // Ptr |
15125 | DAG.getValueType(N->getValueType(0)) // VT |
15126 | }; |
15127 | SDValue BSLoad = |
15128 | DAG.getMemIntrinsicNode(PPCISD::LBRX, dl, |
15129 | DAG.getVTList(N->getValueType(0) == MVT::i64 ? |
15130 | MVT::i64 : MVT::i32, MVT::Other), |
15131 | Ops, LD->getMemoryVT(), LD->getMemOperand()); |
15132 | |
15133 | // If this is an i16 load, insert the truncate. |
15134 | SDValue ResVal = BSLoad; |
15135 | if (N->getValueType(0) == MVT::i16) |
15136 | ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad); |
15137 | |
15138 | // First, combine the bswap away. This makes the value produced by the |
15139 | // load dead. |
15140 | DCI.CombineTo(N, ResVal); |
15141 | |
15142 | // Next, combine the load away, we give it a bogus result value but a real |
15143 | // chain result. The result value is dead because the bswap is dead. |
15144 | DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1)); |
15145 | |
15146 | // Return N so it doesn't get rechecked! |
15147 | return SDValue(N, 0); |
15148 | } |
15149 | break; |
15150 | case PPCISD::VCMP: |
15151 | // If a VCMPo node already exists with exactly the same operands as this |
15152 | // node, use its result instead of this node (VCMPo computes both a CR6 and |
15153 | // a normal output). |
15154 | // |
15155 | if (!N->getOperand(0).hasOneUse() && |
15156 | !N->getOperand(1).hasOneUse() && |
15157 | !N->getOperand(2).hasOneUse()) { |
15158 | |
15159 | // Scan all of the users of the LHS, looking for VCMPo's that match. |
15160 | SDNode *VCMPoNode = nullptr; |
15161 | |
15162 | SDNode *LHSN = N->getOperand(0).getNode(); |
15163 | for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end(); |
15164 | UI != E; ++UI) |
15165 | if (UI->getOpcode() == PPCISD::VCMPo && |
15166 | UI->getOperand(1) == N->getOperand(1) && |
15167 | UI->getOperand(2) == N->getOperand(2) && |
15168 | UI->getOperand(0) == N->getOperand(0)) { |
15169 | VCMPoNode = *UI; |
15170 | break; |
15171 | } |
15172 | |
15173 | // If there is no VCMPo node, or if the flag value has a single use, don't |
15174 | // transform this. |
15175 | if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1)) |
15176 | break; |
15177 | |
15178 | // Look at the (necessarily single) use of the flag value. If it has a |
15179 | // chain, this transformation is more complex. Note that multiple things |
15180 | // could use the value result, which we should ignore. |
15181 | SDNode *FlagUser = nullptr; |
15182 | for (SDNode::use_iterator UI = VCMPoNode->use_begin(); |
15183 | FlagUser == nullptr; ++UI) { |
15184 | assert(UI != VCMPoNode->use_end() && "Didn't find user!")((UI != VCMPoNode->use_end() && "Didn't find user!" ) ? static_cast<void> (0) : __assert_fail ("UI != VCMPoNode->use_end() && \"Didn't find user!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15184, __PRETTY_FUNCTION__)); |
15185 | SDNode *User = *UI; |
15186 | for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) { |
15187 | if (User->getOperand(i) == SDValue(VCMPoNode, 1)) { |
15188 | FlagUser = User; |
15189 | break; |
15190 | } |
15191 | } |
15192 | } |
15193 | |
15194 | // If the user is a MFOCRF instruction, we know this is safe. |
15195 | // Otherwise we give up for right now. |
15196 | if (FlagUser->getOpcode() == PPCISD::MFOCRF) |
15197 | return SDValue(VCMPoNode, 0); |
15198 | } |
15199 | break; |
15200 | case ISD::BRCOND: { |
15201 | SDValue Cond = N->getOperand(1); |
15202 | SDValue Target = N->getOperand(2); |
15203 | |
15204 | if (Cond.getOpcode() == ISD::INTRINSIC_W_CHAIN && |
15205 | cast<ConstantSDNode>(Cond.getOperand(1))->getZExtValue() == |
15206 | Intrinsic::loop_decrement) { |
15207 | |
15208 | // We now need to make the intrinsic dead (it cannot be instruction |
15209 | // selected). |
15210 | DAG.ReplaceAllUsesOfValueWith(Cond.getValue(1), Cond.getOperand(0)); |
15211 | assert(Cond.getNode()->hasOneUse() &&((Cond.getNode()->hasOneUse() && "Counter decrement has more than one use" ) ? static_cast<void> (0) : __assert_fail ("Cond.getNode()->hasOneUse() && \"Counter decrement has more than one use\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15212, __PRETTY_FUNCTION__)) |
15212 | "Counter decrement has more than one use")((Cond.getNode()->hasOneUse() && "Counter decrement has more than one use" ) ? static_cast<void> (0) : __assert_fail ("Cond.getNode()->hasOneUse() && \"Counter decrement has more than one use\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15212, __PRETTY_FUNCTION__)); |
15213 | |
15214 | return DAG.getNode(PPCISD::BDNZ, dl, MVT::Other, |
15215 | N->getOperand(0), Target); |
15216 | } |
15217 | } |
15218 | break; |
15219 | case ISD::BR_CC: { |
15220 | // If this is a branch on an altivec predicate comparison, lower this so |
15221 | // that we don't have to do a MFOCRF: instead, branch directly on CR6. This |
15222 | // lowering is done pre-legalize, because the legalizer lowers the predicate |
15223 | // compare down to code that is difficult to reassemble. |
15224 | ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get(); |
15225 | SDValue LHS = N->getOperand(2), RHS = N->getOperand(3); |
15226 | |
15227 | // Sometimes the promoted value of the intrinsic is ANDed by some non-zero |
15228 | // value. If so, pass-through the AND to get to the intrinsic. |
15229 | if (LHS.getOpcode() == ISD::AND && |
15230 | LHS.getOperand(0).getOpcode() == ISD::INTRINSIC_W_CHAIN && |
15231 | cast<ConstantSDNode>(LHS.getOperand(0).getOperand(1))->getZExtValue() == |
15232 | Intrinsic::loop_decrement && |
15233 | isa<ConstantSDNode>(LHS.getOperand(1)) && |
15234 | !isNullConstant(LHS.getOperand(1))) |
15235 | LHS = LHS.getOperand(0); |
15236 | |
15237 | if (LHS.getOpcode() == ISD::INTRINSIC_W_CHAIN && |
15238 | cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue() == |
15239 | Intrinsic::loop_decrement && |
15240 | isa<ConstantSDNode>(RHS)) { |
15241 | assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&(((CC == ISD::SETEQ || CC == ISD::SETNE) && "Counter decrement comparison is not EQ or NE" ) ? static_cast<void> (0) : __assert_fail ("(CC == ISD::SETEQ || CC == ISD::SETNE) && \"Counter decrement comparison is not EQ or NE\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15242, __PRETTY_FUNCTION__)) |
15242 | "Counter decrement comparison is not EQ or NE")(((CC == ISD::SETEQ || CC == ISD::SETNE) && "Counter decrement comparison is not EQ or NE" ) ? static_cast<void> (0) : __assert_fail ("(CC == ISD::SETEQ || CC == ISD::SETNE) && \"Counter decrement comparison is not EQ or NE\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15242, __PRETTY_FUNCTION__)); |
15243 | |
15244 | unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue(); |
15245 | bool isBDNZ = (CC == ISD::SETEQ && Val) || |
15246 | (CC == ISD::SETNE && !Val); |
15247 | |
15248 | // We now need to make the intrinsic dead (it cannot be instruction |
15249 | // selected). |
15250 | DAG.ReplaceAllUsesOfValueWith(LHS.getValue(1), LHS.getOperand(0)); |
15251 | assert(LHS.getNode()->hasOneUse() &&((LHS.getNode()->hasOneUse() && "Counter decrement has more than one use" ) ? static_cast<void> (0) : __assert_fail ("LHS.getNode()->hasOneUse() && \"Counter decrement has more than one use\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15252, __PRETTY_FUNCTION__)) |
15252 | "Counter decrement has more than one use")((LHS.getNode()->hasOneUse() && "Counter decrement has more than one use" ) ? static_cast<void> (0) : __assert_fail ("LHS.getNode()->hasOneUse() && \"Counter decrement has more than one use\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15252, __PRETTY_FUNCTION__)); |
15253 | |
15254 | return DAG.getNode(isBDNZ ? PPCISD::BDNZ : PPCISD::BDZ, dl, MVT::Other, |
15255 | N->getOperand(0), N->getOperand(4)); |
15256 | } |
15257 | |
15258 | int CompareOpc; |
15259 | bool isDot; |
15260 | |
15261 | if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN && |
15262 | isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) && |
15263 | getVectorCompareInfo(LHS, CompareOpc, isDot, Subtarget)) { |
15264 | assert(isDot && "Can't compare against a vector result!")((isDot && "Can't compare against a vector result!") ? static_cast<void> (0) : __assert_fail ("isDot && \"Can't compare against a vector result!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15264, __PRETTY_FUNCTION__)); |
15265 | |
15266 | // If this is a comparison against something other than 0/1, then we know |
15267 | // that the condition is never/always true. |
15268 | unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue(); |
15269 | if (Val != 0 && Val != 1) { |
15270 | if (CC == ISD::SETEQ) // Cond never true, remove branch. |
15271 | return N->getOperand(0); |
15272 | // Always !=, turn it into an unconditional branch. |
15273 | return DAG.getNode(ISD::BR, dl, MVT::Other, |
15274 | N->getOperand(0), N->getOperand(4)); |
15275 | } |
15276 | |
15277 | bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0); |
15278 | |
15279 | // Create the PPCISD altivec 'dot' comparison node. |
15280 | SDValue Ops[] = { |
15281 | LHS.getOperand(2), // LHS of compare |
15282 | LHS.getOperand(3), // RHS of compare |
15283 | DAG.getConstant(CompareOpc, dl, MVT::i32) |
15284 | }; |
15285 | EVT VTs[] = { LHS.getOperand(2).getValueType(), MVT::Glue }; |
15286 | SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops); |
15287 | |
15288 | // Unpack the result based on how the target uses it. |
15289 | PPC::Predicate CompOpc; |
15290 | switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) { |
15291 | default: // Can't happen, don't crash on invalid number though. |
15292 | case 0: // Branch on the value of the EQ bit of CR6. |
15293 | CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE; |
15294 | break; |
15295 | case 1: // Branch on the inverted value of the EQ bit of CR6. |
15296 | CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ; |
15297 | break; |
15298 | case 2: // Branch on the value of the LT bit of CR6. |
15299 | CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE; |
15300 | break; |
15301 | case 3: // Branch on the inverted value of the LT bit of CR6. |
15302 | CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT; |
15303 | break; |
15304 | } |
15305 | |
15306 | return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0), |
15307 | DAG.getConstant(CompOpc, dl, MVT::i32), |
15308 | DAG.getRegister(PPC::CR6, MVT::i32), |
15309 | N->getOperand(4), CompNode.getValue(1)); |
15310 | } |
15311 | break; |
15312 | } |
15313 | case ISD::BUILD_VECTOR: |
15314 | return DAGCombineBuildVector(N, DCI); |
15315 | case ISD::ABS: |
15316 | return combineABS(N, DCI); |
15317 | case ISD::VSELECT: |
15318 | return combineVSelect(N, DCI); |
15319 | } |
15320 | |
15321 | return SDValue(); |
15322 | } |
15323 | |
15324 | SDValue |
15325 | PPCTargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor, |
15326 | SelectionDAG &DAG, |
15327 | SmallVectorImpl<SDNode *> &Created) const { |
15328 | // fold (sdiv X, pow2) |
15329 | EVT VT = N->getValueType(0); |
15330 | if (VT == MVT::i64 && !Subtarget.isPPC64()) |
15331 | return SDValue(); |
15332 | if ((VT != MVT::i32 && VT != MVT::i64) || |
15333 | !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2())) |
15334 | return SDValue(); |
15335 | |
15336 | SDLoc DL(N); |
15337 | SDValue N0 = N->getOperand(0); |
15338 | |
15339 | bool IsNegPow2 = (-Divisor).isPowerOf2(); |
15340 | unsigned Lg2 = (IsNegPow2 ? -Divisor : Divisor).countTrailingZeros(); |
15341 | SDValue ShiftAmt = DAG.getConstant(Lg2, DL, VT); |
15342 | |
15343 | SDValue Op = DAG.getNode(PPCISD::SRA_ADDZE, DL, VT, N0, ShiftAmt); |
15344 | Created.push_back(Op.getNode()); |
15345 | |
15346 | if (IsNegPow2) { |
15347 | Op = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Op); |
15348 | Created.push_back(Op.getNode()); |
15349 | } |
15350 | |
15351 | return Op; |
15352 | } |
15353 | |
15354 | //===----------------------------------------------------------------------===// |
15355 | // Inline Assembly Support |
15356 | //===----------------------------------------------------------------------===// |
15357 | |
15358 | void PPCTargetLowering::computeKnownBitsForTargetNode(const SDValue Op, |
15359 | KnownBits &Known, |
15360 | const APInt &DemandedElts, |
15361 | const SelectionDAG &DAG, |
15362 | unsigned Depth) const { |
15363 | Known.resetAll(); |
15364 | switch (Op.getOpcode()) { |
15365 | default: break; |
15366 | case PPCISD::LBRX: { |
15367 | // lhbrx is known to have the top bits cleared out. |
15368 | if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16) |
15369 | Known.Zero = 0xFFFF0000; |
15370 | break; |
15371 | } |
15372 | case ISD::INTRINSIC_WO_CHAIN: { |
15373 | switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) { |
15374 | default: break; |
15375 | case Intrinsic::ppc_altivec_vcmpbfp_p: |
15376 | case Intrinsic::ppc_altivec_vcmpeqfp_p: |
15377 | case Intrinsic::ppc_altivec_vcmpequb_p: |
15378 | case Intrinsic::ppc_altivec_vcmpequh_p: |
15379 | case Intrinsic::ppc_altivec_vcmpequw_p: |
15380 | case Intrinsic::ppc_altivec_vcmpequd_p: |
15381 | case Intrinsic::ppc_altivec_vcmpequq_p: |
15382 | case Intrinsic::ppc_altivec_vcmpgefp_p: |
15383 | case Intrinsic::ppc_altivec_vcmpgtfp_p: |
15384 | case Intrinsic::ppc_altivec_vcmpgtsb_p: |
15385 | case Intrinsic::ppc_altivec_vcmpgtsh_p: |
15386 | case Intrinsic::ppc_altivec_vcmpgtsw_p: |
15387 | case Intrinsic::ppc_altivec_vcmpgtsd_p: |
15388 | case Intrinsic::ppc_altivec_vcmpgtsq_p: |
15389 | case Intrinsic::ppc_altivec_vcmpgtub_p: |
15390 | case Intrinsic::ppc_altivec_vcmpgtuh_p: |
15391 | case Intrinsic::ppc_altivec_vcmpgtuw_p: |
15392 | case Intrinsic::ppc_altivec_vcmpgtud_p: |
15393 | case Intrinsic::ppc_altivec_vcmpgtuq_p: |
15394 | Known.Zero = ~1U; // All bits but the low one are known to be zero. |
15395 | break; |
15396 | } |
15397 | } |
15398 | } |
15399 | } |
15400 | |
15401 | Align PPCTargetLowering::getPrefLoopAlignment(MachineLoop *ML) const { |
15402 | switch (Subtarget.getCPUDirective()) { |
15403 | default: break; |
15404 | case PPC::DIR_970: |
15405 | case PPC::DIR_PWR4: |
15406 | case PPC::DIR_PWR5: |
15407 | case PPC::DIR_PWR5X: |
15408 | case PPC::DIR_PWR6: |
15409 | case PPC::DIR_PWR6X: |
15410 | case PPC::DIR_PWR7: |
15411 | case PPC::DIR_PWR8: |
15412 | case PPC::DIR_PWR9: |
15413 | case PPC::DIR_PWR10: |
15414 | case PPC::DIR_PWR_FUTURE: { |
15415 | if (!ML) |
15416 | break; |
15417 | |
15418 | if (!DisableInnermostLoopAlign32) { |
15419 | // If the nested loop is an innermost loop, prefer to a 32-byte alignment, |
15420 | // so that we can decrease cache misses and branch-prediction misses. |
15421 | // Actual alignment of the loop will depend on the hotness check and other |
15422 | // logic in alignBlocks. |
15423 | if (ML->getLoopDepth() > 1 && ML->getSubLoops().empty()) |
15424 | return Align(32); |
15425 | } |
15426 | |
15427 | const PPCInstrInfo *TII = Subtarget.getInstrInfo(); |
15428 | |
15429 | // For small loops (between 5 and 8 instructions), align to a 32-byte |
15430 | // boundary so that the entire loop fits in one instruction-cache line. |
15431 | uint64_t LoopSize = 0; |
15432 | for (auto I = ML->block_begin(), IE = ML->block_end(); I != IE; ++I) |
15433 | for (auto J = (*I)->begin(), JE = (*I)->end(); J != JE; ++J) { |
15434 | LoopSize += TII->getInstSizeInBytes(*J); |
15435 | if (LoopSize > 32) |
15436 | break; |
15437 | } |
15438 | |
15439 | if (LoopSize > 16 && LoopSize <= 32) |
15440 | return Align(32); |
15441 | |
15442 | break; |
15443 | } |
15444 | } |
15445 | |
15446 | return TargetLowering::getPrefLoopAlignment(ML); |
15447 | } |
15448 | |
15449 | /// getConstraintType - Given a constraint, return the type of |
15450 | /// constraint it is for this target. |
15451 | PPCTargetLowering::ConstraintType |
15452 | PPCTargetLowering::getConstraintType(StringRef Constraint) const { |
15453 | if (Constraint.size() == 1) { |
15454 | switch (Constraint[0]) { |
15455 | default: break; |
15456 | case 'b': |
15457 | case 'r': |
15458 | case 'f': |
15459 | case 'd': |
15460 | case 'v': |
15461 | case 'y': |
15462 | return C_RegisterClass; |
15463 | case 'Z': |
15464 | // FIXME: While Z does indicate a memory constraint, it specifically |
15465 | // indicates an r+r address (used in conjunction with the 'y' modifier |
15466 | // in the replacement string). Currently, we're forcing the base |
15467 | // register to be r0 in the asm printer (which is interpreted as zero) |
15468 | // and forming the complete address in the second register. This is |
15469 | // suboptimal. |
15470 | return C_Memory; |
15471 | } |
15472 | } else if (Constraint == "wc") { // individual CR bits. |
15473 | return C_RegisterClass; |
15474 | } else if (Constraint == "wa" || Constraint == "wd" || |
15475 | Constraint == "wf" || Constraint == "ws" || |
15476 | Constraint == "wi" || Constraint == "ww") { |
15477 | return C_RegisterClass; // VSX registers. |
15478 | } |
15479 | return TargetLowering::getConstraintType(Constraint); |
15480 | } |
15481 | |
15482 | /// Examine constraint type and operand type and determine a weight value. |
15483 | /// This object must already have been set up with the operand type |
15484 | /// and the current alternative constraint selected. |
15485 | TargetLowering::ConstraintWeight |
15486 | PPCTargetLowering::getSingleConstraintMatchWeight( |
15487 | AsmOperandInfo &info, const char *constraint) const { |
15488 | ConstraintWeight weight = CW_Invalid; |
15489 | Value *CallOperandVal = info.CallOperandVal; |
15490 | // If we don't have a value, we can't do a match, |
15491 | // but allow it at the lowest weight. |
15492 | if (!CallOperandVal) |
15493 | return CW_Default; |
15494 | Type *type = CallOperandVal->getType(); |
15495 | |
15496 | // Look at the constraint type. |
15497 | if (StringRef(constraint) == "wc" && type->isIntegerTy(1)) |
15498 | return CW_Register; // an individual CR bit. |
15499 | else if ((StringRef(constraint) == "wa" || |
15500 | StringRef(constraint) == "wd" || |
15501 | StringRef(constraint) == "wf") && |
15502 | type->isVectorTy()) |
15503 | return CW_Register; |
15504 | else if (StringRef(constraint) == "wi" && type->isIntegerTy(64)) |
15505 | return CW_Register; // just hold 64-bit integers data. |
15506 | else if (StringRef(constraint) == "ws" && type->isDoubleTy()) |
15507 | return CW_Register; |
15508 | else if (StringRef(constraint) == "ww" && type->isFloatTy()) |
15509 | return CW_Register; |
15510 | |
15511 | switch (*constraint) { |
15512 | default: |
15513 | weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); |
15514 | break; |
15515 | case 'b': |
15516 | if (type->isIntegerTy()) |
15517 | weight = CW_Register; |
15518 | break; |
15519 | case 'f': |
15520 | if (type->isFloatTy()) |
15521 | weight = CW_Register; |
15522 | break; |
15523 | case 'd': |
15524 | if (type->isDoubleTy()) |
15525 | weight = CW_Register; |
15526 | break; |
15527 | case 'v': |
15528 | if (type->isVectorTy()) |
15529 | weight = CW_Register; |
15530 | break; |
15531 | case 'y': |
15532 | weight = CW_Register; |
15533 | break; |
15534 | case 'Z': |
15535 | weight = CW_Memory; |
15536 | break; |
15537 | } |
15538 | return weight; |
15539 | } |
15540 | |
15541 | std::pair<unsigned, const TargetRegisterClass *> |
15542 | PPCTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, |
15543 | StringRef Constraint, |
15544 | MVT VT) const { |
15545 | if (Constraint.size() == 1) { |
15546 | // GCC RS6000 Constraint Letters |
15547 | switch (Constraint[0]) { |
15548 | case 'b': // R1-R31 |
15549 | if (VT == MVT::i64 && Subtarget.isPPC64()) |
15550 | return std::make_pair(0U, &PPC::G8RC_NOX0RegClass); |
15551 | return std::make_pair(0U, &PPC::GPRC_NOR0RegClass); |
15552 | case 'r': // R0-R31 |
15553 | if (VT == MVT::i64 && Subtarget.isPPC64()) |
15554 | return std::make_pair(0U, &PPC::G8RCRegClass); |
15555 | return std::make_pair(0U, &PPC::GPRCRegClass); |
15556 | // 'd' and 'f' constraints are both defined to be "the floating point |
15557 | // registers", where one is for 32-bit and the other for 64-bit. We don't |
15558 | // really care overly much here so just give them all the same reg classes. |
15559 | case 'd': |
15560 | case 'f': |
15561 | if (Subtarget.hasSPE()) { |
15562 | if (VT == MVT::f32 || VT == MVT::i32) |
15563 | return std::make_pair(0U, &PPC::GPRCRegClass); |
15564 | if (VT == MVT::f64 || VT == MVT::i64) |
15565 | return std::make_pair(0U, &PPC::SPERCRegClass); |
15566 | } else { |
15567 | if (VT == MVT::f32 || VT == MVT::i32) |
15568 | return std::make_pair(0U, &PPC::F4RCRegClass); |
15569 | if (VT == MVT::f64 || VT == MVT::i64) |
15570 | return std::make_pair(0U, &PPC::F8RCRegClass); |
15571 | } |
15572 | break; |
15573 | case 'v': |
15574 | if (Subtarget.hasAltivec()) |
15575 | return std::make_pair(0U, &PPC::VRRCRegClass); |
15576 | break; |
15577 | case 'y': // crrc |
15578 | return std::make_pair(0U, &PPC::CRRCRegClass); |
15579 | } |
15580 | } else if (Constraint == "wc" && Subtarget.useCRBits()) { |
15581 | // An individual CR bit. |
15582 | return std::make_pair(0U, &PPC::CRBITRCRegClass); |
15583 | } else if ((Constraint == "wa" || Constraint == "wd" || |
15584 | Constraint == "wf" || Constraint == "wi") && |
15585 | Subtarget.hasVSX()) { |
15586 | return std::make_pair(0U, &PPC::VSRCRegClass); |
15587 | } else if ((Constraint == "ws" || Constraint == "ww") && Subtarget.hasVSX()) { |
15588 | if (VT == MVT::f32 && Subtarget.hasP8Vector()) |
15589 | return std::make_pair(0U, &PPC::VSSRCRegClass); |
15590 | else |
15591 | return std::make_pair(0U, &PPC::VSFRCRegClass); |
15592 | } |
15593 | |
15594 | // If we name a VSX register, we can't defer to the base class because it |
15595 | // will not recognize the correct register (their names will be VSL{0-31} |
15596 | // and V{0-31} so they won't match). So we match them here. |
15597 | if (Constraint.size() > 3 && Constraint[1] == 'v' && Constraint[2] == 's') { |
15598 | int VSNum = atoi(Constraint.data() + 3); |
15599 | assert(VSNum >= 0 && VSNum <= 63 &&((VSNum >= 0 && VSNum <= 63 && "Attempted to access a vsr out of range" ) ? static_cast<void> (0) : __assert_fail ("VSNum >= 0 && VSNum <= 63 && \"Attempted to access a vsr out of range\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15600, __PRETTY_FUNCTION__)) |
15600 | "Attempted to access a vsr out of range")((VSNum >= 0 && VSNum <= 63 && "Attempted to access a vsr out of range" ) ? static_cast<void> (0) : __assert_fail ("VSNum >= 0 && VSNum <= 63 && \"Attempted to access a vsr out of range\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15600, __PRETTY_FUNCTION__)); |
15601 | if (VSNum < 32) |
15602 | return std::make_pair(PPC::VSL0 + VSNum, &PPC::VSRCRegClass); |
15603 | return std::make_pair(PPC::V0 + VSNum - 32, &PPC::VSRCRegClass); |
15604 | } |
15605 | std::pair<unsigned, const TargetRegisterClass *> R = |
15606 | TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); |
15607 | |
15608 | // r[0-9]+ are used, on PPC64, to refer to the corresponding 64-bit registers |
15609 | // (which we call X[0-9]+). If a 64-bit value has been requested, and a |
15610 | // 32-bit GPR has been selected, then 'upgrade' it to the 64-bit parent |
15611 | // register. |
15612 | // FIXME: If TargetLowering::getRegForInlineAsmConstraint could somehow use |
15613 | // the AsmName field from *RegisterInfo.td, then this would not be necessary. |
15614 | if (R.first && VT == MVT::i64 && Subtarget.isPPC64() && |
15615 | PPC::GPRCRegClass.contains(R.first)) |
15616 | return std::make_pair(TRI->getMatchingSuperReg(R.first, |
15617 | PPC::sub_32, &PPC::G8RCRegClass), |
15618 | &PPC::G8RCRegClass); |
15619 | |
15620 | // GCC accepts 'cc' as an alias for 'cr0', and we need to do the same. |
15621 | if (!R.second && StringRef("{cc}").equals_lower(Constraint)) { |
15622 | R.first = PPC::CR0; |
15623 | R.second = &PPC::CRRCRegClass; |
15624 | } |
15625 | |
15626 | return R; |
15627 | } |
15628 | |
15629 | /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops |
15630 | /// vector. If it is invalid, don't add anything to Ops. |
15631 | void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op, |
15632 | std::string &Constraint, |
15633 | std::vector<SDValue>&Ops, |
15634 | SelectionDAG &DAG) const { |
15635 | SDValue Result; |
15636 | |
15637 | // Only support length 1 constraints. |
15638 | if (Constraint.length() > 1) return; |
15639 | |
15640 | char Letter = Constraint[0]; |
15641 | switch (Letter) { |
15642 | default: break; |
15643 | case 'I': |
15644 | case 'J': |
15645 | case 'K': |
15646 | case 'L': |
15647 | case 'M': |
15648 | case 'N': |
15649 | case 'O': |
15650 | case 'P': { |
15651 | ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op); |
15652 | if (!CST) return; // Must be an immediate to match. |
15653 | SDLoc dl(Op); |
15654 | int64_t Value = CST->getSExtValue(); |
15655 | EVT TCVT = MVT::i64; // All constants taken to be 64 bits so that negative |
15656 | // numbers are printed as such. |
15657 | switch (Letter) { |
15658 | default: llvm_unreachable("Unknown constraint letter!")::llvm::llvm_unreachable_internal("Unknown constraint letter!" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15658); |
15659 | case 'I': // "I" is a signed 16-bit constant. |
15660 | if (isInt<16>(Value)) |
15661 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15662 | break; |
15663 | case 'J': // "J" is a constant with only the high-order 16 bits nonzero. |
15664 | if (isShiftedUInt<16, 16>(Value)) |
15665 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15666 | break; |
15667 | case 'L': // "L" is a signed 16-bit constant shifted left 16 bits. |
15668 | if (isShiftedInt<16, 16>(Value)) |
15669 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15670 | break; |
15671 | case 'K': // "K" is a constant with only the low-order 16 bits nonzero. |
15672 | if (isUInt<16>(Value)) |
15673 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15674 | break; |
15675 | case 'M': // "M" is a constant that is greater than 31. |
15676 | if (Value > 31) |
15677 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15678 | break; |
15679 | case 'N': // "N" is a positive constant that is an exact power of two. |
15680 | if (Value > 0 && isPowerOf2_64(Value)) |
15681 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15682 | break; |
15683 | case 'O': // "O" is the constant zero. |
15684 | if (Value == 0) |
15685 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15686 | break; |
15687 | case 'P': // "P" is a constant whose negation is a signed 16-bit constant. |
15688 | if (isInt<16>(-Value)) |
15689 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15690 | break; |
15691 | } |
15692 | break; |
15693 | } |
15694 | } |
15695 | |
15696 | if (Result.getNode()) { |
15697 | Ops.push_back(Result); |
15698 | return; |
15699 | } |
15700 | |
15701 | // Handle standard constraint letters. |
15702 | TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); |
15703 | } |
15704 | |
15705 | // isLegalAddressingMode - Return true if the addressing mode represented |
15706 | // by AM is legal for this target, for a load/store of the specified type. |
15707 | bool PPCTargetLowering::isLegalAddressingMode(const DataLayout &DL, |
15708 | const AddrMode &AM, Type *Ty, |
15709 | unsigned AS, |
15710 | Instruction *I) const { |
15711 | // Vector type r+i form is supported since power9 as DQ form. We don't check |
15712 | // the offset matching DQ form requirement(off % 16 == 0), because on PowerPC, |
15713 | // imm form is preferred and the offset can be adjusted to use imm form later |
15714 | // in pass PPCLoopInstrFormPrep. Also in LSR, for one LSRUse, it uses min and |
15715 | // max offset to check legal addressing mode, we should be a little aggressive |
15716 | // to contain other offsets for that LSRUse. |
15717 | if (Ty->isVectorTy() && AM.BaseOffs != 0 && !Subtarget.hasP9Vector()) |
15718 | return false; |
15719 | |
15720 | // PPC allows a sign-extended 16-bit immediate field. |
15721 | if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) |
15722 | return false; |
15723 | |
15724 | // No global is ever allowed as a base. |
15725 | if (AM.BaseGV) |
15726 | return false; |
15727 | |
15728 | // PPC only support r+r, |
15729 | switch (AM.Scale) { |
15730 | case 0: // "r+i" or just "i", depending on HasBaseReg. |
15731 | break; |
15732 | case 1: |
15733 | if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. |
15734 | return false; |
15735 | // Otherwise we have r+r or r+i. |
15736 | break; |
15737 | case 2: |
15738 | if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. |
15739 | return false; |
15740 | // Allow 2*r as r+r. |
15741 | break; |
15742 | default: |
15743 | // No other scales are supported. |
15744 | return false; |
15745 | } |
15746 | |
15747 | return true; |
15748 | } |
15749 | |
15750 | SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op, |
15751 | SelectionDAG &DAG) const { |
15752 | MachineFunction &MF = DAG.getMachineFunction(); |
15753 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
15754 | MFI.setReturnAddressIsTaken(true); |
15755 | |
15756 | if (verifyReturnAddressArgumentIsConstant(Op, DAG)) |
15757 | return SDValue(); |
15758 | |
15759 | SDLoc dl(Op); |
15760 | unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); |
15761 | |
15762 | // Make sure the function does not optimize away the store of the RA to |
15763 | // the stack. |
15764 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
15765 | FuncInfo->setLRStoreRequired(); |
15766 | bool isPPC64 = Subtarget.isPPC64(); |
15767 | auto PtrVT = getPointerTy(MF.getDataLayout()); |
15768 | |
15769 | if (Depth > 0) { |
15770 | SDValue FrameAddr = LowerFRAMEADDR(Op, DAG); |
15771 | SDValue Offset = |
15772 | DAG.getConstant(Subtarget.getFrameLowering()->getReturnSaveOffset(), dl, |
15773 | isPPC64 ? MVT::i64 : MVT::i32); |
15774 | return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), |
15775 | DAG.getNode(ISD::ADD, dl, PtrVT, FrameAddr, Offset), |
15776 | MachinePointerInfo()); |
15777 | } |
15778 | |
15779 | // Just load the return address off the stack. |
15780 | SDValue RetAddrFI = getReturnAddrFrameIndex(DAG); |
15781 | return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), RetAddrFI, |
15782 | MachinePointerInfo()); |
15783 | } |
15784 | |
15785 | SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op, |
15786 | SelectionDAG &DAG) const { |
15787 | SDLoc dl(Op); |
15788 | unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); |
15789 | |
15790 | MachineFunction &MF = DAG.getMachineFunction(); |
15791 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
15792 | MFI.setFrameAddressIsTaken(true); |
15793 | |
15794 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
15795 | bool isPPC64 = PtrVT == MVT::i64; |
15796 | |
15797 | // Naked functions never have a frame pointer, and so we use r1. For all |
15798 | // other functions, this decision must be delayed until during PEI. |
15799 | unsigned FrameReg; |
15800 | if (MF.getFunction().hasFnAttribute(Attribute::Naked)) |
15801 | FrameReg = isPPC64 ? PPC::X1 : PPC::R1; |
15802 | else |
15803 | FrameReg = isPPC64 ? PPC::FP8 : PPC::FP; |
15804 | |
15805 | SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, |
15806 | PtrVT); |
15807 | while (Depth--) |
15808 | FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(), |
15809 | FrameAddr, MachinePointerInfo()); |
15810 | return FrameAddr; |
15811 | } |
15812 | |
15813 | // FIXME? Maybe this could be a TableGen attribute on some registers and |
15814 | // this table could be generated automatically from RegInfo. |
15815 | Register PPCTargetLowering::getRegisterByName(const char* RegName, LLT VT, |
15816 | const MachineFunction &MF) const { |
15817 | bool isPPC64 = Subtarget.isPPC64(); |
15818 | |
15819 | bool is64Bit = isPPC64 && VT == LLT::scalar(64); |
15820 | if (!is64Bit && VT != LLT::scalar(32)) |
15821 | report_fatal_error("Invalid register global variable type"); |
15822 | |
15823 | Register Reg = StringSwitch<Register>(RegName) |
15824 | .Case("r1", is64Bit ? PPC::X1 : PPC::R1) |
15825 | .Case("r2", isPPC64 ? Register() : PPC::R2) |
15826 | .Case("r13", (is64Bit ? PPC::X13 : PPC::R13)) |
15827 | .Default(Register()); |
15828 | |
15829 | if (Reg) |
15830 | return Reg; |
15831 | report_fatal_error("Invalid register name global variable"); |
15832 | } |
15833 | |
15834 | bool PPCTargetLowering::isAccessedAsGotIndirect(SDValue GA) const { |
15835 | // 32-bit SVR4 ABI access everything as got-indirect. |
15836 | if (Subtarget.is32BitELFABI()) |
15837 | return true; |
15838 | |
15839 | // AIX accesses everything indirectly through the TOC, which is similar to |
15840 | // the GOT. |
15841 | if (Subtarget.isAIXABI()) |
15842 | return true; |
15843 | |
15844 | CodeModel::Model CModel = getTargetMachine().getCodeModel(); |
15845 | // If it is small or large code model, module locals are accessed |
15846 | // indirectly by loading their address from .toc/.got. |
15847 | if (CModel == CodeModel::Small || CModel == CodeModel::Large) |
15848 | return true; |
15849 | |
15850 | // JumpTable and BlockAddress are accessed as got-indirect. |
15851 | if (isa<JumpTableSDNode>(GA) || isa<BlockAddressSDNode>(GA)) |
15852 | return true; |
15853 | |
15854 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA)) |
15855 | return Subtarget.isGVIndirectSymbol(G->getGlobal()); |
15856 | |
15857 | return false; |
15858 | } |
15859 | |
15860 | bool |
15861 | PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { |
15862 | // The PowerPC target isn't yet aware of offsets. |
15863 | return false; |
15864 | } |
15865 | |
15866 | bool PPCTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info, |
15867 | const CallInst &I, |
15868 | MachineFunction &MF, |
15869 | unsigned Intrinsic) const { |
15870 | switch (Intrinsic) { |
15871 | case Intrinsic::ppc_altivec_lvx: |
15872 | case Intrinsic::ppc_altivec_lvxl: |
15873 | case Intrinsic::ppc_altivec_lvebx: |
15874 | case Intrinsic::ppc_altivec_lvehx: |
15875 | case Intrinsic::ppc_altivec_lvewx: |
15876 | case Intrinsic::ppc_vsx_lxvd2x: |
15877 | case Intrinsic::ppc_vsx_lxvw4x: |
15878 | case Intrinsic::ppc_vsx_lxvd2x_be: |
15879 | case Intrinsic::ppc_vsx_lxvw4x_be: |
15880 | case Intrinsic::ppc_vsx_lxvl: |
15881 | case Intrinsic::ppc_vsx_lxvll: { |
15882 | EVT VT; |
15883 | switch (Intrinsic) { |
15884 | case Intrinsic::ppc_altivec_lvebx: |
15885 | VT = MVT::i8; |
15886 | break; |
15887 | case Intrinsic::ppc_altivec_lvehx: |
15888 | VT = MVT::i16; |
15889 | break; |
15890 | case Intrinsic::ppc_altivec_lvewx: |
15891 | VT = MVT::i32; |
15892 | break; |
15893 | case Intrinsic::ppc_vsx_lxvd2x: |
15894 | case Intrinsic::ppc_vsx_lxvd2x_be: |
15895 | VT = MVT::v2f64; |
15896 | break; |
15897 | default: |
15898 | VT = MVT::v4i32; |
15899 | break; |
15900 | } |
15901 | |
15902 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
15903 | Info.memVT = VT; |
15904 | Info.ptrVal = I.getArgOperand(0); |
15905 | Info.offset = -VT.getStoreSize()+1; |
15906 | Info.size = 2*VT.getStoreSize()-1; |
15907 | Info.align = Align(1); |
15908 | Info.flags = MachineMemOperand::MOLoad; |
15909 | return true; |
15910 | } |
15911 | case Intrinsic::ppc_altivec_stvx: |
15912 | case Intrinsic::ppc_altivec_stvxl: |
15913 | case Intrinsic::ppc_altivec_stvebx: |
15914 | case Intrinsic::ppc_altivec_stvehx: |
15915 | case Intrinsic::ppc_altivec_stvewx: |
15916 | case Intrinsic::ppc_vsx_stxvd2x: |
15917 | case Intrinsic::ppc_vsx_stxvw4x: |
15918 | case Intrinsic::ppc_vsx_stxvd2x_be: |
15919 | case Intrinsic::ppc_vsx_stxvw4x_be: |
15920 | case Intrinsic::ppc_vsx_stxvl: |
15921 | case Intrinsic::ppc_vsx_stxvll: { |
15922 | EVT VT; |
15923 | switch (Intrinsic) { |
15924 | case Intrinsic::ppc_altivec_stvebx: |
15925 | VT = MVT::i8; |
15926 | break; |
15927 | case Intrinsic::ppc_altivec_stvehx: |
15928 | VT = MVT::i16; |
15929 | break; |
15930 | case Intrinsic::ppc_altivec_stvewx: |
15931 | VT = MVT::i32; |
15932 | break; |
15933 | case Intrinsic::ppc_vsx_stxvd2x: |
15934 | case Intrinsic::ppc_vsx_stxvd2x_be: |
15935 | VT = MVT::v2f64; |
15936 | break; |
15937 | default: |
15938 | VT = MVT::v4i32; |
15939 | break; |
15940 | } |
15941 | |
15942 | Info.opc = ISD::INTRINSIC_VOID; |
15943 | Info.memVT = VT; |
15944 | Info.ptrVal = I.getArgOperand(1); |
15945 | Info.offset = -VT.getStoreSize()+1; |
15946 | Info.size = 2*VT.getStoreSize()-1; |
15947 | Info.align = Align(1); |
15948 | Info.flags = MachineMemOperand::MOStore; |
15949 | return true; |
15950 | } |
15951 | default: |
15952 | break; |
15953 | } |
15954 | |
15955 | return false; |
15956 | } |
15957 | |
15958 | /// It returns EVT::Other if the type should be determined using generic |
15959 | /// target-independent logic. |
15960 | EVT PPCTargetLowering::getOptimalMemOpType( |
15961 | const MemOp &Op, const AttributeList &FuncAttributes) const { |
15962 | if (getTargetMachine().getOptLevel() != CodeGenOpt::None) { |
15963 | // We should use Altivec/VSX loads and stores when available. For unaligned |
15964 | // addresses, unaligned VSX loads are only fast starting with the P8. |
15965 | if (Subtarget.hasAltivec() && Op.size() >= 16 && |
15966 | (Op.isAligned(Align(16)) || |
15967 | ((Op.isMemset() && Subtarget.hasVSX()) || Subtarget.hasP8Vector()))) |
15968 | return MVT::v4i32; |
15969 | } |
15970 | |
15971 | if (Subtarget.isPPC64()) { |
15972 | return MVT::i64; |
15973 | } |
15974 | |
15975 | return MVT::i32; |
15976 | } |
15977 | |
15978 | /// Returns true if it is beneficial to convert a load of a constant |
15979 | /// to just the constant itself. |
15980 | bool PPCTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm, |
15981 | Type *Ty) const { |
15982 | assert(Ty->isIntegerTy())((Ty->isIntegerTy()) ? static_cast<void> (0) : __assert_fail ("Ty->isIntegerTy()", "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15982, __PRETTY_FUNCTION__)); |
15983 | |
15984 | unsigned BitSize = Ty->getPrimitiveSizeInBits(); |
15985 | return !(BitSize == 0 || BitSize > 64); |
15986 | } |
15987 | |
15988 | bool PPCTargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const { |
15989 | if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy()) |
15990 | return false; |
15991 | unsigned NumBits1 = Ty1->getPrimitiveSizeInBits(); |
15992 | unsigned NumBits2 = Ty2->getPrimitiveSizeInBits(); |
15993 | return NumBits1 == 64 && NumBits2 == 32; |
15994 | } |
15995 | |
15996 | bool PPCTargetLowering::isTruncateFree(EVT VT1, EVT VT2) const { |
15997 | if (!VT1.isInteger() || !VT2.isInteger()) |
15998 | return false; |
15999 | unsigned NumBits1 = VT1.getSizeInBits(); |
16000 | unsigned NumBits2 = VT2.getSizeInBits(); |
16001 | return NumBits1 == 64 && NumBits2 == 32; |
16002 | } |
16003 | |
16004 | bool PPCTargetLowering::isZExtFree(SDValue Val, EVT VT2) const { |
16005 | // Generally speaking, zexts are not free, but they are free when they can be |
16006 | // folded with other operations. |
16007 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Val)) { |
16008 | EVT MemVT = LD->getMemoryVT(); |
16009 | if ((MemVT == MVT::i1 || MemVT == MVT::i8 || MemVT == MVT::i16 || |
16010 | (Subtarget.isPPC64() && MemVT == MVT::i32)) && |
16011 | (LD->getExtensionType() == ISD::NON_EXTLOAD || |
16012 | LD->getExtensionType() == ISD::ZEXTLOAD)) |
16013 | return true; |
16014 | } |
16015 | |
16016 | // FIXME: Add other cases... |
16017 | // - 32-bit shifts with a zext to i64 |
16018 | // - zext after ctlz, bswap, etc. |
16019 | // - zext after and by a constant mask |
16020 | |
16021 | return TargetLowering::isZExtFree(Val, VT2); |
16022 | } |
16023 | |
16024 | bool PPCTargetLowering::isFPExtFree(EVT DestVT, EVT SrcVT) const { |
16025 | assert(DestVT.isFloatingPoint() && SrcVT.isFloatingPoint() &&((DestVT.isFloatingPoint() && SrcVT.isFloatingPoint() && "invalid fpext types") ? static_cast<void> ( 0) : __assert_fail ("DestVT.isFloatingPoint() && SrcVT.isFloatingPoint() && \"invalid fpext types\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16026, __PRETTY_FUNCTION__)) |
16026 | "invalid fpext types")((DestVT.isFloatingPoint() && SrcVT.isFloatingPoint() && "invalid fpext types") ? static_cast<void> ( 0) : __assert_fail ("DestVT.isFloatingPoint() && SrcVT.isFloatingPoint() && \"invalid fpext types\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16026, __PRETTY_FUNCTION__)); |
16027 | // Extending to float128 is not free. |
16028 | if (DestVT == MVT::f128) |
16029 | return false; |
16030 | return true; |
16031 | } |
16032 | |
16033 | bool PPCTargetLowering::isLegalICmpImmediate(int64_t Imm) const { |
16034 | return isInt<16>(Imm) || isUInt<16>(Imm); |
16035 | } |
16036 | |
16037 | bool PPCTargetLowering::isLegalAddImmediate(int64_t Imm) const { |
16038 | return isInt<16>(Imm) || isUInt<16>(Imm); |
16039 | } |
16040 | |
16041 | bool PPCTargetLowering::allowsMisalignedMemoryAccesses(EVT VT, |
16042 | unsigned, |
16043 | unsigned, |
16044 | MachineMemOperand::Flags, |
16045 | bool *Fast) const { |
16046 | if (DisablePPCUnaligned) |
16047 | return false; |
16048 | |
16049 | // PowerPC supports unaligned memory access for simple non-vector types. |
16050 | // Although accessing unaligned addresses is not as efficient as accessing |
16051 | // aligned addresses, it is generally more efficient than manual expansion, |
16052 | // and generally only traps for software emulation when crossing page |
16053 | // boundaries. |
16054 | |
16055 | if (!VT.isSimple()) |
16056 | return false; |
16057 | |
16058 | if (VT.isFloatingPoint() && !VT.isVector() && |
16059 | !Subtarget.allowsUnalignedFPAccess()) |
16060 | return false; |
16061 | |
16062 | if (VT.getSimpleVT().isVector()) { |
16063 | if (Subtarget.hasVSX()) { |
16064 | if (VT != MVT::v2f64 && VT != MVT::v2i64 && |
16065 | VT != MVT::v4f32 && VT != MVT::v4i32) |
16066 | return false; |
16067 | } else { |
16068 | return false; |
16069 | } |
16070 | } |
16071 | |
16072 | if (VT == MVT::ppcf128) |
16073 | return false; |
16074 | |
16075 | if (Fast) |
16076 | *Fast = true; |
16077 | |
16078 | return true; |
16079 | } |
16080 | |
16081 | bool PPCTargetLowering::decomposeMulByConstant(LLVMContext &Context, EVT VT, |
16082 | SDValue C) const { |
16083 | // Check integral scalar types. |
16084 | if (!VT.isScalarInteger()) |
16085 | return false; |
16086 | if (auto *ConstNode = dyn_cast<ConstantSDNode>(C.getNode())) { |
16087 | if (!ConstNode->getAPIntValue().isSignedIntN(64)) |
16088 | return false; |
16089 | // This transformation will generate >= 2 operations. But the following |
16090 | // cases will generate <= 2 instructions during ISEL. So exclude them. |
16091 | // 1. If the constant multiplier fits 16 bits, it can be handled by one |
16092 | // HW instruction, ie. MULLI |
16093 | // 2. If the multiplier after shifted fits 16 bits, an extra shift |
16094 | // instruction is needed than case 1, ie. MULLI and RLDICR |
16095 | int64_t Imm = ConstNode->getSExtValue(); |
16096 | unsigned Shift = countTrailingZeros<uint64_t>(Imm); |
16097 | Imm >>= Shift; |
16098 | if (isInt<16>(Imm)) |
16099 | return false; |
16100 | uint64_t UImm = static_cast<uint64_t>(Imm); |
16101 | if (isPowerOf2_64(UImm + 1) || isPowerOf2_64(UImm - 1) || |
16102 | isPowerOf2_64(1 - UImm) || isPowerOf2_64(-1 - UImm)) |
16103 | return true; |
16104 | } |
16105 | return false; |
16106 | } |
16107 | |
16108 | bool PPCTargetLowering::isFMAFasterThanFMulAndFAdd(const MachineFunction &MF, |
16109 | EVT VT) const { |
16110 | return isFMAFasterThanFMulAndFAdd( |
16111 | MF.getFunction(), VT.getTypeForEVT(MF.getFunction().getContext())); |
16112 | } |
16113 | |
16114 | bool PPCTargetLowering::isFMAFasterThanFMulAndFAdd(const Function &F, |
16115 | Type *Ty) const { |
16116 | switch (Ty->getScalarType()->getTypeID()) { |
16117 | case Type::FloatTyID: |
16118 | case Type::DoubleTyID: |
16119 | return true; |
16120 | case Type::FP128TyID: |
16121 | return Subtarget.hasP9Vector(); |
16122 | default: |
16123 | return false; |
16124 | } |
16125 | } |
16126 | |
16127 | // FIXME: add more patterns which are not profitable to hoist. |
16128 | bool PPCTargetLowering::isProfitableToHoist(Instruction *I) const { |
16129 | if (!I->hasOneUse()) |
16130 | return true; |
16131 | |
16132 | Instruction *User = I->user_back(); |
16133 | assert(User && "A single use instruction with no uses.")((User && "A single use instruction with no uses.") ? static_cast<void> (0) : __assert_fail ("User && \"A single use instruction with no uses.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16133, __PRETTY_FUNCTION__)); |
16134 | |
16135 | switch (I->getOpcode()) { |
16136 | case Instruction::FMul: { |
16137 | // Don't break FMA, PowerPC prefers FMA. |
16138 | if (User->getOpcode() != Instruction::FSub && |
16139 | User->getOpcode() != Instruction::FAdd) |
16140 | return true; |
16141 | |
16142 | const TargetOptions &Options = getTargetMachine().Options; |
16143 | const Function *F = I->getFunction(); |
16144 | const DataLayout &DL = F->getParent()->getDataLayout(); |
16145 | Type *Ty = User->getOperand(0)->getType(); |
16146 | |
16147 | return !( |
16148 | isFMAFasterThanFMulAndFAdd(*F, Ty) && |
16149 | isOperationLegalOrCustom(ISD::FMA, getValueType(DL, Ty)) && |
16150 | (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath)); |
16151 | } |
16152 | case Instruction::Load: { |
16153 | // Don't break "store (load float*)" pattern, this pattern will be combined |
16154 | // to "store (load int32)" in later InstCombine pass. See function |
16155 | // combineLoadToOperationType. On PowerPC, loading a float point takes more |
16156 | // cycles than loading a 32 bit integer. |
16157 | LoadInst *LI = cast<LoadInst>(I); |
16158 | // For the loads that combineLoadToOperationType does nothing, like |
16159 | // ordered load, it should be profitable to hoist them. |
16160 | // For swifterror load, it can only be used for pointer to pointer type, so |
16161 | // later type check should get rid of this case. |
16162 | if (!LI->isUnordered()) |
16163 | return true; |
16164 | |
16165 | if (User->getOpcode() != Instruction::Store) |
16166 | return true; |
16167 | |
16168 | if (I->getType()->getTypeID() != Type::FloatTyID) |
16169 | return true; |
16170 | |
16171 | return false; |
16172 | } |
16173 | default: |
16174 | return true; |
16175 | } |
16176 | return true; |
16177 | } |
16178 | |
16179 | const MCPhysReg * |
16180 | PPCTargetLowering::getScratchRegisters(CallingConv::ID) const { |
16181 | // LR is a callee-save register, but we must treat it as clobbered by any call |
16182 | // site. Hence we include LR in the scratch registers, which are in turn added |
16183 | // as implicit-defs for stackmaps and patchpoints. The same reasoning applies |
16184 | // to CTR, which is used by any indirect call. |
16185 | static const MCPhysReg ScratchRegs[] = { |
16186 | PPC::X12, PPC::LR8, PPC::CTR8, 0 |
16187 | }; |
16188 | |
16189 | return ScratchRegs; |
16190 | } |
16191 | |
16192 | Register PPCTargetLowering::getExceptionPointerRegister( |
16193 | const Constant *PersonalityFn) const { |
16194 | return Subtarget.isPPC64() ? PPC::X3 : PPC::R3; |
16195 | } |
16196 | |
16197 | Register PPCTargetLowering::getExceptionSelectorRegister( |
16198 | const Constant *PersonalityFn) const { |
16199 | return Subtarget.isPPC64() ? PPC::X4 : PPC::R4; |
16200 | } |
16201 | |
16202 | bool |
16203 | PPCTargetLowering::shouldExpandBuildVectorWithShuffles( |
16204 | EVT VT , unsigned DefinedValues) const { |
16205 | if (VT == MVT::v2i64) |
16206 | return Subtarget.hasDirectMove(); // Don't need stack ops with direct moves |
16207 | |
16208 | if (Subtarget.hasVSX()) |
16209 | return true; |
16210 | |
16211 | return TargetLowering::shouldExpandBuildVectorWithShuffles(VT, DefinedValues); |
16212 | } |
16213 | |
16214 | Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const { |
16215 | if (DisableILPPref || Subtarget.enableMachineScheduler()) |
16216 | return TargetLowering::getSchedulingPreference(N); |
16217 | |
16218 | return Sched::ILP; |
16219 | } |
16220 | |
16221 | // Create a fast isel object. |
16222 | FastISel * |
16223 | PPCTargetLowering::createFastISel(FunctionLoweringInfo &FuncInfo, |
16224 | const TargetLibraryInfo *LibInfo) const { |
16225 | return PPC::createFastISel(FuncInfo, LibInfo); |
16226 | } |
16227 | |
16228 | // 'Inverted' means the FMA opcode after negating one multiplicand. |
16229 | // For example, (fma -a b c) = (fnmsub a b c) |
16230 | static unsigned invertFMAOpcode(unsigned Opc) { |
16231 | switch (Opc) { |
16232 | default: |
16233 | llvm_unreachable("Invalid FMA opcode for PowerPC!")::llvm::llvm_unreachable_internal("Invalid FMA opcode for PowerPC!" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16233); |
16234 | case ISD::FMA: |
16235 | return PPCISD::FNMSUB; |
16236 | case PPCISD::FNMSUB: |
16237 | return ISD::FMA; |
16238 | } |
16239 | } |
16240 | |
16241 | SDValue PPCTargetLowering::getNegatedExpression(SDValue Op, SelectionDAG &DAG, |
16242 | bool LegalOps, bool OptForSize, |
16243 | NegatibleCost &Cost, |
16244 | unsigned Depth) const { |
16245 | if (Depth > SelectionDAG::MaxRecursionDepth) |
16246 | return SDValue(); |
16247 | |
16248 | unsigned Opc = Op.getOpcode(); |
16249 | EVT VT = Op.getValueType(); |
16250 | SDNodeFlags Flags = Op.getNode()->getFlags(); |
16251 | |
16252 | switch (Opc) { |
16253 | case PPCISD::FNMSUB: |
16254 | if (!Op.hasOneUse() || !isTypeLegal(VT)) |
16255 | break; |
16256 | |
16257 | const TargetOptions &Options = getTargetMachine().Options; |
16258 | SDValue N0 = Op.getOperand(0); |
16259 | SDValue N1 = Op.getOperand(1); |
16260 | SDValue N2 = Op.getOperand(2); |
16261 | SDLoc Loc(Op); |
16262 | |
16263 | NegatibleCost N2Cost = NegatibleCost::Expensive; |
16264 | SDValue NegN2 = |
16265 | getNegatedExpression(N2, DAG, LegalOps, OptForSize, N2Cost, Depth + 1); |
16266 | |
16267 | if (!NegN2) |
16268 | return SDValue(); |
16269 | |
16270 | // (fneg (fnmsub a b c)) => (fnmsub (fneg a) b (fneg c)) |
16271 | // (fneg (fnmsub a b c)) => (fnmsub a (fneg b) (fneg c)) |
16272 | // These transformations may change sign of zeroes. For example, |
16273 | // -(-ab-(-c))=-0 while -(-(ab-c))=+0 when a=b=c=1. |
16274 | if (Flags.hasNoSignedZeros() || Options.NoSignedZerosFPMath) { |
16275 | // Try and choose the cheaper one to negate. |
16276 | NegatibleCost N0Cost = NegatibleCost::Expensive; |
16277 | SDValue NegN0 = getNegatedExpression(N0, DAG, LegalOps, OptForSize, |
16278 | N0Cost, Depth + 1); |
16279 | |
16280 | NegatibleCost N1Cost = NegatibleCost::Expensive; |
16281 | SDValue NegN1 = getNegatedExpression(N1, DAG, LegalOps, OptForSize, |
16282 | N1Cost, Depth + 1); |
16283 | |
16284 | if (NegN0 && N0Cost <= N1Cost) { |
16285 | Cost = std::min(N0Cost, N2Cost); |
16286 | return DAG.getNode(Opc, Loc, VT, NegN0, N1, NegN2, Flags); |
16287 | } else if (NegN1) { |
16288 | Cost = std::min(N1Cost, N2Cost); |
16289 | return DAG.getNode(Opc, Loc, VT, N0, NegN1, NegN2, Flags); |
16290 | } |
16291 | } |
16292 | |
16293 | // (fneg (fnmsub a b c)) => (fma a b (fneg c)) |
16294 | if (isOperationLegal(ISD::FMA, VT)) { |
16295 | Cost = N2Cost; |
16296 | return DAG.getNode(ISD::FMA, Loc, VT, N0, N1, NegN2, Flags); |
16297 | } |
16298 | |
16299 | break; |
16300 | } |
16301 | |
16302 | return TargetLowering::getNegatedExpression(Op, DAG, LegalOps, OptForSize, |
16303 | Cost, Depth); |
16304 | } |
16305 | |
16306 | // Override to enable LOAD_STACK_GUARD lowering on Linux. |
16307 | bool PPCTargetLowering::useLoadStackGuardNode() const { |
16308 | if (!Subtarget.isTargetLinux()) |
16309 | return TargetLowering::useLoadStackGuardNode(); |
16310 | return true; |
16311 | } |
16312 | |
16313 | // Override to disable global variable loading on Linux. |
16314 | void PPCTargetLowering::insertSSPDeclarations(Module &M) const { |
16315 | if (!Subtarget.isTargetLinux()) |
16316 | return TargetLowering::insertSSPDeclarations(M); |
16317 | } |
16318 | |
16319 | bool PPCTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT, |
16320 | bool ForCodeSize) const { |
16321 | if (!VT.isSimple() || !Subtarget.hasVSX()) |
16322 | return false; |
16323 | |
16324 | switch(VT.getSimpleVT().SimpleTy) { |
16325 | default: |
16326 | // For FP types that are currently not supported by PPC backend, return |
16327 | // false. Examples: f16, f80. |
16328 | return false; |
16329 | case MVT::f32: |
16330 | case MVT::f64: |
16331 | if (Subtarget.hasPrefixInstrs()) { |
16332 | // With prefixed instructions, we can materialize anything that can be |
16333 | // represented with a 32-bit immediate, not just positive zero. |
16334 | APFloat APFloatOfImm = Imm; |
16335 | return convertToNonDenormSingle(APFloatOfImm); |
16336 | } |
16337 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
16338 | case MVT::ppcf128: |
16339 | return Imm.isPosZero(); |
16340 | } |
16341 | } |
16342 | |
16343 | // For vector shift operation op, fold |
16344 | // (op x, (and y, ((1 << numbits(x)) - 1))) -> (target op x, y) |
16345 | static SDValue stripModuloOnShift(const TargetLowering &TLI, SDNode *N, |
16346 | SelectionDAG &DAG) { |
16347 | SDValue N0 = N->getOperand(0); |
16348 | SDValue N1 = N->getOperand(1); |
16349 | EVT VT = N0.getValueType(); |
16350 | unsigned OpSizeInBits = VT.getScalarSizeInBits(); |
16351 | unsigned Opcode = N->getOpcode(); |
16352 | unsigned TargetOpcode; |
16353 | |
16354 | switch (Opcode) { |
16355 | default: |
16356 | llvm_unreachable("Unexpected shift operation")::llvm::llvm_unreachable_internal("Unexpected shift operation" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16356); |
16357 | case ISD::SHL: |
16358 | TargetOpcode = PPCISD::SHL; |
16359 | break; |
16360 | case ISD::SRL: |
16361 | TargetOpcode = PPCISD::SRL; |
16362 | break; |
16363 | case ISD::SRA: |
16364 | TargetOpcode = PPCISD::SRA; |
16365 | break; |
16366 | } |
16367 | |
16368 | if (VT.isVector() && TLI.isOperationLegal(Opcode, VT) && |
16369 | N1->getOpcode() == ISD::AND) |
16370 | if (ConstantSDNode *Mask = isConstOrConstSplat(N1->getOperand(1))) |
16371 | if (Mask->getZExtValue() == OpSizeInBits - 1) |
16372 | return DAG.getNode(TargetOpcode, SDLoc(N), VT, N0, N1->getOperand(0)); |
16373 | |
16374 | return SDValue(); |
16375 | } |
16376 | |
16377 | SDValue PPCTargetLowering::combineSHL(SDNode *N, DAGCombinerInfo &DCI) const { |
16378 | if (auto Value = stripModuloOnShift(*this, N, DCI.DAG)) |
16379 | return Value; |
16380 | |
16381 | SDValue N0 = N->getOperand(0); |
16382 | ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(N->getOperand(1)); |
16383 | if (!Subtarget.isISA3_0() || !Subtarget.isPPC64() || |
16384 | N0.getOpcode() != ISD::SIGN_EXTEND || |
16385 | N0.getOperand(0).getValueType() != MVT::i32 || CN1 == nullptr || |
16386 | N->getValueType(0) != MVT::i64) |
16387 | return SDValue(); |
16388 | |
16389 | // We can't save an operation here if the value is already extended, and |
16390 | // the existing shift is easier to combine. |
16391 | SDValue ExtsSrc = N0.getOperand(0); |
16392 | if (ExtsSrc.getOpcode() == ISD::TRUNCATE && |
16393 | ExtsSrc.getOperand(0).getOpcode() == ISD::AssertSext) |
16394 | return SDValue(); |
16395 | |
16396 | SDLoc DL(N0); |
16397 | SDValue ShiftBy = SDValue(CN1, 0); |
16398 | // We want the shift amount to be i32 on the extswli, but the shift could |
16399 | // have an i64. |
16400 | if (ShiftBy.getValueType() == MVT::i64) |
16401 | ShiftBy = DCI.DAG.getConstant(CN1->getZExtValue(), DL, MVT::i32); |
16402 | |
16403 | return DCI.DAG.getNode(PPCISD::EXTSWSLI, DL, MVT::i64, N0->getOperand(0), |
16404 | ShiftBy); |
16405 | } |
16406 | |
16407 | SDValue PPCTargetLowering::combineSRA(SDNode *N, DAGCombinerInfo &DCI) const { |
16408 | if (auto Value = stripModuloOnShift(*this, N, DCI.DAG)) |
16409 | return Value; |
16410 | |
16411 | return SDValue(); |
16412 | } |
16413 | |
16414 | SDValue PPCTargetLowering::combineSRL(SDNode *N, DAGCombinerInfo &DCI) const { |
16415 | if (auto Value = stripModuloOnShift(*this, N, DCI.DAG)) |
16416 | return Value; |
16417 | |
16418 | return SDValue(); |
16419 | } |
16420 | |
16421 | // Transform (add X, (zext(setne Z, C))) -> (addze X, (addic (addi Z, -C), -1)) |
16422 | // Transform (add X, (zext(sete Z, C))) -> (addze X, (subfic (addi Z, -C), 0)) |
16423 | // When C is zero, the equation (addi Z, -C) can be simplified to Z |
16424 | // Requirement: -C in [-32768, 32767], X and Z are MVT::i64 types |
16425 | static SDValue combineADDToADDZE(SDNode *N, SelectionDAG &DAG, |
16426 | const PPCSubtarget &Subtarget) { |
16427 | if (!Subtarget.isPPC64()) |
16428 | return SDValue(); |
16429 | |
16430 | SDValue LHS = N->getOperand(0); |
16431 | SDValue RHS = N->getOperand(1); |
16432 | |
16433 | auto isZextOfCompareWithConstant = [](SDValue Op) { |
16434 | if (Op.getOpcode() != ISD::ZERO_EXTEND || !Op.hasOneUse() || |
16435 | Op.getValueType() != MVT::i64) |
16436 | return false; |
16437 | |
16438 | SDValue Cmp = Op.getOperand(0); |
16439 | if (Cmp.getOpcode() != ISD::SETCC || !Cmp.hasOneUse() || |
16440 | Cmp.getOperand(0).getValueType() != MVT::i64) |
16441 | return false; |
16442 | |
16443 | if (auto *Constant = dyn_cast<ConstantSDNode>(Cmp.getOperand(1))) { |
16444 | int64_t NegConstant = 0 - Constant->getSExtValue(); |
16445 | // Due to the limitations of the addi instruction, |
16446 | // -C is required to be [-32768, 32767]. |
16447 | return isInt<16>(NegConstant); |
16448 | } |
16449 | |
16450 | return false; |
16451 | }; |
16452 | |
16453 | bool LHSHasPattern = isZextOfCompareWithConstant(LHS); |
16454 | bool RHSHasPattern = isZextOfCompareWithConstant(RHS); |
16455 | |
16456 | // If there is a pattern, canonicalize a zext operand to the RHS. |
16457 | if (LHSHasPattern && !RHSHasPattern) |
16458 | std::swap(LHS, RHS); |
16459 | else if (!LHSHasPattern && !RHSHasPattern) |
16460 | return SDValue(); |
16461 | |
16462 | SDLoc DL(N); |
16463 | SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Glue); |
16464 | SDValue Cmp = RHS.getOperand(0); |
16465 | SDValue Z = Cmp.getOperand(0); |
16466 | auto *Constant = dyn_cast<ConstantSDNode>(Cmp.getOperand(1)); |
16467 | |
16468 | assert(Constant && "Constant Should not be a null pointer.")((Constant && "Constant Should not be a null pointer." ) ? static_cast<void> (0) : __assert_fail ("Constant && \"Constant Should not be a null pointer.\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16468, __PRETTY_FUNCTION__)); |
16469 | int64_t NegConstant = 0 - Constant->getSExtValue(); |
16470 | |
16471 | switch(cast<CondCodeSDNode>(Cmp.getOperand(2))->get()) { |
16472 | default: break; |
16473 | case ISD::SETNE: { |
16474 | // when C == 0 |
16475 | // --> addze X, (addic Z, -1).carry |
16476 | // / |
16477 | // add X, (zext(setne Z, C))-- |
16478 | // \ when -32768 <= -C <= 32767 && C != 0 |
16479 | // --> addze X, (addic (addi Z, -C), -1).carry |
16480 | SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Z, |
16481 | DAG.getConstant(NegConstant, DL, MVT::i64)); |
16482 | SDValue AddOrZ = NegConstant != 0 ? Add : Z; |
16483 | SDValue Addc = DAG.getNode(ISD::ADDC, DL, DAG.getVTList(MVT::i64, MVT::Glue), |
16484 | AddOrZ, DAG.getConstant(-1ULL, DL, MVT::i64)); |
16485 | return DAG.getNode(ISD::ADDE, DL, VTs, LHS, DAG.getConstant(0, DL, MVT::i64), |
16486 | SDValue(Addc.getNode(), 1)); |
16487 | } |
16488 | case ISD::SETEQ: { |
16489 | // when C == 0 |
16490 | // --> addze X, (subfic Z, 0).carry |
16491 | // / |
16492 | // add X, (zext(sete Z, C))-- |
16493 | // \ when -32768 <= -C <= 32767 && C != 0 |
16494 | // --> addze X, (subfic (addi Z, -C), 0).carry |
16495 | SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Z, |
16496 | DAG.getConstant(NegConstant, DL, MVT::i64)); |
16497 | SDValue AddOrZ = NegConstant != 0 ? Add : Z; |
16498 | SDValue Subc = DAG.getNode(ISD::SUBC, DL, DAG.getVTList(MVT::i64, MVT::Glue), |
16499 | DAG.getConstant(0, DL, MVT::i64), AddOrZ); |
16500 | return DAG.getNode(ISD::ADDE, DL, VTs, LHS, DAG.getConstant(0, DL, MVT::i64), |
16501 | SDValue(Subc.getNode(), 1)); |
16502 | } |
16503 | } |
16504 | |
16505 | return SDValue(); |
16506 | } |
16507 | |
16508 | // Transform |
16509 | // (add C1, (MAT_PCREL_ADDR GlobalAddr+C2)) to |
16510 | // (MAT_PCREL_ADDR GlobalAddr+(C1+C2)) |
16511 | // In this case both C1 and C2 must be known constants. |
16512 | // C1+C2 must fit into a 34 bit signed integer. |
16513 | static SDValue combineADDToMAT_PCREL_ADDR(SDNode *N, SelectionDAG &DAG, |
16514 | const PPCSubtarget &Subtarget) { |
16515 | if (!Subtarget.isUsingPCRelativeCalls()) |
16516 | return SDValue(); |
16517 | |
16518 | // Check both Operand 0 and Operand 1 of the ADD node for the PCRel node. |
16519 | // If we find that node try to cast the Global Address and the Constant. |
16520 | SDValue LHS = N->getOperand(0); |
16521 | SDValue RHS = N->getOperand(1); |
16522 | |
16523 | if (LHS.getOpcode() != PPCISD::MAT_PCREL_ADDR) |
16524 | std::swap(LHS, RHS); |
16525 | |
16526 | if (LHS.getOpcode() != PPCISD::MAT_PCREL_ADDR) |
16527 | return SDValue(); |
16528 | |
16529 | // Operand zero of PPCISD::MAT_PCREL_ADDR is the GA node. |
16530 | GlobalAddressSDNode *GSDN = dyn_cast<GlobalAddressSDNode>(LHS.getOperand(0)); |
16531 | ConstantSDNode* ConstNode = dyn_cast<ConstantSDNode>(RHS); |
16532 | |
16533 | // Check that both casts succeeded. |
16534 | if (!GSDN || !ConstNode) |
16535 | return SDValue(); |
16536 | |
16537 | int64_t NewOffset = GSDN->getOffset() + ConstNode->getSExtValue(); |
16538 | SDLoc DL(GSDN); |
16539 | |
16540 | // The signed int offset needs to fit in 34 bits. |
16541 | if (!isInt<34>(NewOffset)) |
16542 | return SDValue(); |
16543 | |
16544 | // The new global address is a copy of the old global address except |
16545 | // that it has the updated Offset. |
16546 | SDValue GA = |
16547 | DAG.getTargetGlobalAddress(GSDN->getGlobal(), DL, GSDN->getValueType(0), |
16548 | NewOffset, GSDN->getTargetFlags()); |
16549 | SDValue MatPCRel = |
16550 | DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, GSDN->getValueType(0), GA); |
16551 | return MatPCRel; |
16552 | } |
16553 | |
16554 | SDValue PPCTargetLowering::combineADD(SDNode *N, DAGCombinerInfo &DCI) const { |
16555 | if (auto Value = combineADDToADDZE(N, DCI.DAG, Subtarget)) |
16556 | return Value; |
16557 | |
16558 | if (auto Value = combineADDToMAT_PCREL_ADDR(N, DCI.DAG, Subtarget)) |
16559 | return Value; |
16560 | |
16561 | return SDValue(); |
16562 | } |
16563 | |
16564 | // Detect TRUNCATE operations on bitcasts of float128 values. |
16565 | // What we are looking for here is the situtation where we extract a subset |
16566 | // of bits from a 128 bit float. |
16567 | // This can be of two forms: |
16568 | // 1) BITCAST of f128 feeding TRUNCATE |
16569 | // 2) BITCAST of f128 feeding SRL (a shift) feeding TRUNCATE |
16570 | // The reason this is required is because we do not have a legal i128 type |
16571 | // and so we want to prevent having to store the f128 and then reload part |
16572 | // of it. |
16573 | SDValue PPCTargetLowering::combineTRUNCATE(SDNode *N, |
16574 | DAGCombinerInfo &DCI) const { |
16575 | // If we are using CRBits then try that first. |
16576 | if (Subtarget.useCRBits()) { |
16577 | // Check if CRBits did anything and return that if it did. |
16578 | if (SDValue CRTruncValue = DAGCombineTruncBoolExt(N, DCI)) |
16579 | return CRTruncValue; |
16580 | } |
16581 | |
16582 | SDLoc dl(N); |
16583 | SDValue Op0 = N->getOperand(0); |
16584 | |
16585 | // fold (truncate (abs (sub (zext a), (zext b)))) -> (vabsd a, b) |
16586 | if (Subtarget.hasP9Altivec() && Op0.getOpcode() == ISD::ABS) { |
16587 | EVT VT = N->getValueType(0); |
16588 | if (VT != MVT::v4i32 && VT != MVT::v8i16 && VT != MVT::v16i8) |
16589 | return SDValue(); |
16590 | SDValue Sub = Op0.getOperand(0); |
16591 | if (Sub.getOpcode() == ISD::SUB) { |
16592 | SDValue SubOp0 = Sub.getOperand(0); |
16593 | SDValue SubOp1 = Sub.getOperand(1); |
16594 | if ((SubOp0.getOpcode() == ISD::ZERO_EXTEND) && |
16595 | (SubOp1.getOpcode() == ISD::ZERO_EXTEND)) { |
16596 | return DCI.DAG.getNode(PPCISD::VABSD, dl, VT, SubOp0.getOperand(0), |
16597 | SubOp1.getOperand(0), |
16598 | DCI.DAG.getTargetConstant(0, dl, MVT::i32)); |
16599 | } |
16600 | } |
16601 | } |
16602 | |
16603 | // Looking for a truncate of i128 to i64. |
16604 | if (Op0.getValueType() != MVT::i128 || N->getValueType(0) != MVT::i64) |
16605 | return SDValue(); |
16606 | |
16607 | int EltToExtract = DCI.DAG.getDataLayout().isBigEndian() ? 1 : 0; |
16608 | |
16609 | // SRL feeding TRUNCATE. |
16610 | if (Op0.getOpcode() == ISD::SRL) { |
16611 | ConstantSDNode *ConstNode = dyn_cast<ConstantSDNode>(Op0.getOperand(1)); |
16612 | // The right shift has to be by 64 bits. |
16613 | if (!ConstNode || ConstNode->getZExtValue() != 64) |
16614 | return SDValue(); |
16615 | |
16616 | // Switch the element number to extract. |
16617 | EltToExtract = EltToExtract ? 0 : 1; |
16618 | // Update Op0 past the SRL. |
16619 | Op0 = Op0.getOperand(0); |
16620 | } |
16621 | |
16622 | // BITCAST feeding a TRUNCATE possibly via SRL. |
16623 | if (Op0.getOpcode() == ISD::BITCAST && |
16624 | Op0.getValueType() == MVT::i128 && |
16625 | Op0.getOperand(0).getValueType() == MVT::f128) { |
16626 | SDValue Bitcast = DCI.DAG.getBitcast(MVT::v2i64, Op0.getOperand(0)); |
16627 | return DCI.DAG.getNode( |
16628 | ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Bitcast, |
16629 | DCI.DAG.getTargetConstant(EltToExtract, dl, MVT::i32)); |
16630 | } |
16631 | return SDValue(); |
16632 | } |
16633 | |
16634 | SDValue PPCTargetLowering::combineMUL(SDNode *N, DAGCombinerInfo &DCI) const { |
16635 | SelectionDAG &DAG = DCI.DAG; |
16636 | |
16637 | ConstantSDNode *ConstOpOrElement = isConstOrConstSplat(N->getOperand(1)); |
16638 | if (!ConstOpOrElement) |
16639 | return SDValue(); |
16640 | |
16641 | // An imul is usually smaller than the alternative sequence for legal type. |
16642 | if (DAG.getMachineFunction().getFunction().hasMinSize() && |
16643 | isOperationLegal(ISD::MUL, N->getValueType(0))) |
16644 | return SDValue(); |
16645 | |
16646 | auto IsProfitable = [this](bool IsNeg, bool IsAddOne, EVT VT) -> bool { |
16647 | switch (this->Subtarget.getCPUDirective()) { |
16648 | default: |
16649 | // TODO: enhance the condition for subtarget before pwr8 |
16650 | return false; |
16651 | case PPC::DIR_PWR8: |
16652 | // type mul add shl |
16653 | // scalar 4 1 1 |
16654 | // vector 7 2 2 |
16655 | return true; |
16656 | case PPC::DIR_PWR9: |
16657 | case PPC::DIR_PWR10: |
16658 | case PPC::DIR_PWR_FUTURE: |
16659 | // type mul add shl |
16660 | // scalar 5 2 2 |
16661 | // vector 7 2 2 |
16662 | |
16663 | // The cycle RATIO of related operations are showed as a table above. |
16664 | // Because mul is 5(scalar)/7(vector), add/sub/shl are all 2 for both |
16665 | // scalar and vector type. For 2 instrs patterns, add/sub + shl |
16666 | // are 4, it is always profitable; but for 3 instrs patterns |
16667 | // (mul x, -(2^N + 1)) => -(add (shl x, N), x), sub + add + shl are 6. |
16668 | // So we should only do it for vector type. |
16669 | return IsAddOne && IsNeg ? VT.isVector() : true; |
16670 | } |
16671 | }; |
16672 | |
16673 | EVT VT = N->getValueType(0); |
16674 | SDLoc DL(N); |
16675 | |
16676 | const APInt &MulAmt = ConstOpOrElement->getAPIntValue(); |
16677 | bool IsNeg = MulAmt.isNegative(); |
16678 | APInt MulAmtAbs = MulAmt.abs(); |
16679 | |
16680 | if ((MulAmtAbs - 1).isPowerOf2()) { |
16681 | // (mul x, 2^N + 1) => (add (shl x, N), x) |
16682 | // (mul x, -(2^N + 1)) => -(add (shl x, N), x) |
16683 | |
16684 | if (!IsProfitable(IsNeg, true, VT)) |
16685 | return SDValue(); |
16686 | |
16687 | SDValue Op0 = N->getOperand(0); |
16688 | SDValue Op1 = |
16689 | DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0), |
16690 | DAG.getConstant((MulAmtAbs - 1).logBase2(), DL, VT)); |
16691 | SDValue Res = DAG.getNode(ISD::ADD, DL, VT, Op0, Op1); |
16692 | |
16693 | if (!IsNeg) |
16694 | return Res; |
16695 | |
16696 | return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Res); |
16697 | } else if ((MulAmtAbs + 1).isPowerOf2()) { |
16698 | // (mul x, 2^N - 1) => (sub (shl x, N), x) |
16699 | // (mul x, -(2^N - 1)) => (sub x, (shl x, N)) |
16700 | |
16701 | if (!IsProfitable(IsNeg, false, VT)) |
16702 | return SDValue(); |
16703 | |
16704 | SDValue Op0 = N->getOperand(0); |
16705 | SDValue Op1 = |
16706 | DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0), |
16707 | DAG.getConstant((MulAmtAbs + 1).logBase2(), DL, VT)); |
16708 | |
16709 | if (!IsNeg) |
16710 | return DAG.getNode(ISD::SUB, DL, VT, Op1, Op0); |
16711 | else |
16712 | return DAG.getNode(ISD::SUB, DL, VT, Op0, Op1); |
16713 | |
16714 | } else { |
16715 | return SDValue(); |
16716 | } |
16717 | } |
16718 | |
16719 | // Combine fma-like op (like fnmsub) with fnegs to appropriate op. Do this |
16720 | // in combiner since we need to check SD flags and other subtarget features. |
16721 | SDValue PPCTargetLowering::combineFMALike(SDNode *N, |
16722 | DAGCombinerInfo &DCI) const { |
16723 | SDValue N0 = N->getOperand(0); |
16724 | SDValue N1 = N->getOperand(1); |
16725 | SDValue N2 = N->getOperand(2); |
16726 | SDNodeFlags Flags = N->getFlags(); |
16727 | EVT VT = N->getValueType(0); |
16728 | SelectionDAG &DAG = DCI.DAG; |
16729 | const TargetOptions &Options = getTargetMachine().Options; |
16730 | unsigned Opc = N->getOpcode(); |
16731 | bool CodeSize = DAG.getMachineFunction().getFunction().hasOptSize(); |
16732 | bool LegalOps = !DCI.isBeforeLegalizeOps(); |
16733 | SDLoc Loc(N); |
16734 | |
16735 | if (!isOperationLegal(ISD::FMA, VT)) |
16736 | return SDValue(); |
16737 | |
16738 | // Allowing transformation to FNMSUB may change sign of zeroes when ab-c=0 |
16739 | // since (fnmsub a b c)=-0 while c-ab=+0. |
16740 | if (!Flags.hasNoSignedZeros() && !Options.NoSignedZerosFPMath) |
16741 | return SDValue(); |
16742 | |
16743 | // (fma (fneg a) b c) => (fnmsub a b c) |
16744 | // (fnmsub (fneg a) b c) => (fma a b c) |
16745 | if (SDValue NegN0 = getCheaperNegatedExpression(N0, DAG, LegalOps, CodeSize)) |
16746 | return DAG.getNode(invertFMAOpcode(Opc), Loc, VT, NegN0, N1, N2, Flags); |
16747 | |
16748 | // (fma a (fneg b) c) => (fnmsub a b c) |
16749 | // (fnmsub a (fneg b) c) => (fma a b c) |
16750 | if (SDValue NegN1 = getCheaperNegatedExpression(N1, DAG, LegalOps, CodeSize)) |
16751 | return DAG.getNode(invertFMAOpcode(Opc), Loc, VT, N0, NegN1, N2, Flags); |
16752 | |
16753 | return SDValue(); |
16754 | } |
16755 | |
16756 | bool PPCTargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const { |
16757 | // Only duplicate to increase tail-calls for the 64bit SysV ABIs. |
16758 | if (!Subtarget.is64BitELFABI()) |
16759 | return false; |
16760 | |
16761 | // If not a tail call then no need to proceed. |
16762 | if (!CI->isTailCall()) |
16763 | return false; |
16764 | |
16765 | // If sibling calls have been disabled and tail-calls aren't guaranteed |
16766 | // there is no reason to duplicate. |
16767 | auto &TM = getTargetMachine(); |
16768 | if (!TM.Options.GuaranteedTailCallOpt && DisableSCO) |
16769 | return false; |
16770 | |
16771 | // Can't tail call a function called indirectly, or if it has variadic args. |
16772 | const Function *Callee = CI->getCalledFunction(); |
16773 | if (!Callee || Callee->isVarArg()) |
16774 | return false; |
16775 | |
16776 | // Make sure the callee and caller calling conventions are eligible for tco. |
16777 | const Function *Caller = CI->getParent()->getParent(); |
16778 | if (!areCallingConvEligibleForTCO_64SVR4(Caller->getCallingConv(), |
16779 | CI->getCallingConv())) |
16780 | return false; |
16781 | |
16782 | // If the function is local then we have a good chance at tail-calling it |
16783 | return getTargetMachine().shouldAssumeDSOLocal(*Caller->getParent(), Callee); |
16784 | } |
16785 | |
16786 | bool PPCTargetLowering::hasBitPreservingFPLogic(EVT VT) const { |
16787 | if (!Subtarget.hasVSX()) |
16788 | return false; |
16789 | if (Subtarget.hasP9Vector() && VT == MVT::f128) |
16790 | return true; |
16791 | return VT == MVT::f32 || VT == MVT::f64 || |
16792 | VT == MVT::v4f32 || VT == MVT::v2f64; |
16793 | } |
16794 | |
16795 | bool PPCTargetLowering:: |
16796 | isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const { |
16797 | const Value *Mask = AndI.getOperand(1); |
16798 | // If the mask is suitable for andi. or andis. we should sink the and. |
16799 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(Mask)) { |
16800 | // Can't handle constants wider than 64-bits. |
16801 | if (CI->getBitWidth() > 64) |
16802 | return false; |
16803 | int64_t ConstVal = CI->getZExtValue(); |
16804 | return isUInt<16>(ConstVal) || |
16805 | (isUInt<16>(ConstVal >> 16) && !(ConstVal & 0xFFFF)); |
16806 | } |
16807 | |
16808 | // For non-constant masks, we can always use the record-form and. |
16809 | return true; |
16810 | } |
16811 | |
16812 | // Transform (abs (sub (zext a), (zext b))) to (vabsd a b 0) |
16813 | // Transform (abs (sub (zext a), (zext_invec b))) to (vabsd a b 0) |
16814 | // Transform (abs (sub (zext_invec a), (zext_invec b))) to (vabsd a b 0) |
16815 | // Transform (abs (sub (zext_invec a), (zext b))) to (vabsd a b 0) |
16816 | // Transform (abs (sub a, b) to (vabsd a b 1)) if a & b of type v4i32 |
16817 | SDValue PPCTargetLowering::combineABS(SDNode *N, DAGCombinerInfo &DCI) const { |
16818 | assert((N->getOpcode() == ISD::ABS) && "Need ABS node here")(((N->getOpcode() == ISD::ABS) && "Need ABS node here" ) ? static_cast<void> (0) : __assert_fail ("(N->getOpcode() == ISD::ABS) && \"Need ABS node here\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16818, __PRETTY_FUNCTION__)); |
16819 | assert(Subtarget.hasP9Altivec() &&((Subtarget.hasP9Altivec() && "Only combine this when P9 altivec supported!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasP9Altivec() && \"Only combine this when P9 altivec supported!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16820, __PRETTY_FUNCTION__)) |
16820 | "Only combine this when P9 altivec supported!")((Subtarget.hasP9Altivec() && "Only combine this when P9 altivec supported!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasP9Altivec() && \"Only combine this when P9 altivec supported!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16820, __PRETTY_FUNCTION__)); |
16821 | EVT VT = N->getValueType(0); |
16822 | if (VT != MVT::v4i32 && VT != MVT::v8i16 && VT != MVT::v16i8) |
16823 | return SDValue(); |
16824 | |
16825 | SelectionDAG &DAG = DCI.DAG; |
16826 | SDLoc dl(N); |
16827 | if (N->getOperand(0).getOpcode() == ISD::SUB) { |
16828 | // Even for signed integers, if it's known to be positive (as signed |
16829 | // integer) due to zero-extended inputs. |
16830 | unsigned SubOpcd0 = N->getOperand(0)->getOperand(0).getOpcode(); |
16831 | unsigned SubOpcd1 = N->getOperand(0)->getOperand(1).getOpcode(); |
16832 | if ((SubOpcd0 == ISD::ZERO_EXTEND || |
16833 | SubOpcd0 == ISD::ZERO_EXTEND_VECTOR_INREG) && |
16834 | (SubOpcd1 == ISD::ZERO_EXTEND || |
16835 | SubOpcd1 == ISD::ZERO_EXTEND_VECTOR_INREG)) { |
16836 | return DAG.getNode(PPCISD::VABSD, dl, N->getOperand(0).getValueType(), |
16837 | N->getOperand(0)->getOperand(0), |
16838 | N->getOperand(0)->getOperand(1), |
16839 | DAG.getTargetConstant(0, dl, MVT::i32)); |
16840 | } |
16841 | |
16842 | // For type v4i32, it can be optimized with xvnegsp + vabsduw |
16843 | if (N->getOperand(0).getValueType() == MVT::v4i32 && |
16844 | N->getOperand(0).hasOneUse()) { |
16845 | return DAG.getNode(PPCISD::VABSD, dl, N->getOperand(0).getValueType(), |
16846 | N->getOperand(0)->getOperand(0), |
16847 | N->getOperand(0)->getOperand(1), |
16848 | DAG.getTargetConstant(1, dl, MVT::i32)); |
16849 | } |
16850 | } |
16851 | |
16852 | return SDValue(); |
16853 | } |
16854 | |
16855 | // For type v4i32/v8ii16/v16i8, transform |
16856 | // from (vselect (setcc a, b, setugt), (sub a, b), (sub b, a)) to (vabsd a, b) |
16857 | // from (vselect (setcc a, b, setuge), (sub a, b), (sub b, a)) to (vabsd a, b) |
16858 | // from (vselect (setcc a, b, setult), (sub b, a), (sub a, b)) to (vabsd a, b) |
16859 | // from (vselect (setcc a, b, setule), (sub b, a), (sub a, b)) to (vabsd a, b) |
16860 | SDValue PPCTargetLowering::combineVSelect(SDNode *N, |
16861 | DAGCombinerInfo &DCI) const { |
16862 | assert((N->getOpcode() == ISD::VSELECT) && "Need VSELECT node here")(((N->getOpcode() == ISD::VSELECT) && "Need VSELECT node here" ) ? static_cast<void> (0) : __assert_fail ("(N->getOpcode() == ISD::VSELECT) && \"Need VSELECT node here\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16862, __PRETTY_FUNCTION__)); |
16863 | assert(Subtarget.hasP9Altivec() &&((Subtarget.hasP9Altivec() && "Only combine this when P9 altivec supported!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasP9Altivec() && \"Only combine this when P9 altivec supported!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16864, __PRETTY_FUNCTION__)) |
16864 | "Only combine this when P9 altivec supported!")((Subtarget.hasP9Altivec() && "Only combine this when P9 altivec supported!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasP9Altivec() && \"Only combine this when P9 altivec supported!\"" , "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16864, __PRETTY_FUNCTION__)); |
16865 | |
16866 | SelectionDAG &DAG = DCI.DAG; |
16867 | SDLoc dl(N); |
16868 | SDValue Cond = N->getOperand(0); |
16869 | SDValue TrueOpnd = N->getOperand(1); |
16870 | SDValue FalseOpnd = N->getOperand(2); |
16871 | EVT VT = N->getOperand(1).getValueType(); |
16872 | |
16873 | if (Cond.getOpcode() != ISD::SETCC || TrueOpnd.getOpcode() != ISD::SUB || |
16874 | FalseOpnd.getOpcode() != ISD::SUB) |
16875 | return SDValue(); |
16876 | |
16877 | // ABSD only available for type v4i32/v8i16/v16i8 |
16878 | if (VT != MVT::v4i32 && VT != MVT::v8i16 && VT != MVT::v16i8) |
16879 | return SDValue(); |
16880 | |
16881 | // At least to save one more dependent computation |
16882 | if (!(Cond.hasOneUse() || TrueOpnd.hasOneUse() || FalseOpnd.hasOneUse())) |
16883 | return SDValue(); |
16884 | |
16885 | ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get(); |
16886 | |
16887 | // Can only handle unsigned comparison here |
16888 | switch (CC) { |
16889 | default: |
16890 | return SDValue(); |
16891 | case ISD::SETUGT: |
16892 | case ISD::SETUGE: |
16893 | break; |
16894 | case ISD::SETULT: |
16895 | case ISD::SETULE: |
16896 | std::swap(TrueOpnd, FalseOpnd); |
16897 | break; |
16898 | } |
16899 | |
16900 | SDValue CmpOpnd1 = Cond.getOperand(0); |
16901 | SDValue CmpOpnd2 = Cond.getOperand(1); |
16902 | |
16903 | // SETCC CmpOpnd1 CmpOpnd2 cond |
16904 | // TrueOpnd = CmpOpnd1 - CmpOpnd2 |
16905 | // FalseOpnd = CmpOpnd2 - CmpOpnd1 |
16906 | if (TrueOpnd.getOperand(0) == CmpOpnd1 && |
16907 | TrueOpnd.getOperand(1) == CmpOpnd2 && |
16908 | FalseOpnd.getOperand(0) == CmpOpnd2 && |
16909 | FalseOpnd.getOperand(1) == CmpOpnd1) { |
16910 | return DAG.getNode(PPCISD::VABSD, dl, N->getOperand(1).getValueType(), |
16911 | CmpOpnd1, CmpOpnd2, |
16912 | DAG.getTargetConstant(0, dl, MVT::i32)); |
16913 | } |
16914 | |
16915 | return SDValue(); |
16916 | } |