1 //===-- X86FixupVectorConstants.cpp - optimize constant generation -------===// 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 examines all full size vector constant pool loads and attempts to 10 // replace them with smaller constant pool entries, including: 11 // * Converting AVX512 memory-fold instructions to their broadcast-fold form 12 // * Broadcasting of full width loads. 13 // * TODO: Sign/Zero extension of full width loads. 14 // 15 //===----------------------------------------------------------------------===// 16 17 #include "X86.h" 18 #include "X86InstrFoldTables.h" 19 #include "X86InstrInfo.h" 20 #include "X86Subtarget.h" 21 #include "llvm/ADT/Statistic.h" 22 #include "llvm/CodeGen/MachineConstantPool.h" 23 24 using namespace llvm; 25 26 #define DEBUG_TYPE "x86-fixup-vector-constants" 27 28 STATISTIC(NumInstChanges, "Number of instructions changes"); 29 30 namespace { 31 class X86FixupVectorConstantsPass : public MachineFunctionPass { 32 public: 33 static char ID; 34 35 X86FixupVectorConstantsPass() : MachineFunctionPass(ID) {} 36 37 StringRef getPassName() const override { 38 return "X86 Fixup Vector Constants"; 39 } 40 41 bool runOnMachineFunction(MachineFunction &MF) override; 42 bool processInstruction(MachineFunction &MF, MachineBasicBlock &MBB, 43 MachineInstr &MI); 44 45 // This pass runs after regalloc and doesn't support VReg operands. 46 MachineFunctionProperties getRequiredProperties() const override { 47 return MachineFunctionProperties().set( 48 MachineFunctionProperties::Property::NoVRegs); 49 } 50 51 private: 52 const X86InstrInfo *TII = nullptr; 53 const X86Subtarget *ST = nullptr; 54 const MCSchedModel *SM = nullptr; 55 }; 56 } // end anonymous namespace 57 58 char X86FixupVectorConstantsPass::ID = 0; 59 60 INITIALIZE_PASS(X86FixupVectorConstantsPass, DEBUG_TYPE, DEBUG_TYPE, false, false) 61 62 FunctionPass *llvm::createX86FixupVectorConstants() { 63 return new X86FixupVectorConstantsPass(); 64 } 65 66 // Attempt to extract the full width of bits data from the constant. 67 static std::optional<APInt> extractConstantBits(const Constant *C) { 68 unsigned NumBits = C->getType()->getPrimitiveSizeInBits(); 69 70 if (isa<UndefValue>(C)) 71 return APInt::getZero(NumBits); 72 73 if (auto *CInt = dyn_cast<ConstantInt>(C)) 74 return CInt->getValue(); 75 76 if (auto *CFP = dyn_cast<ConstantFP>(C)) 77 return CFP->getValue().bitcastToAPInt(); 78 79 if (auto *CV = dyn_cast<ConstantVector>(C)) { 80 if (auto *CVSplat = CV->getSplatValue(/*AllowUndefs*/ true)) { 81 if (std::optional<APInt> Bits = extractConstantBits(CVSplat)) { 82 assert((NumBits % Bits->getBitWidth()) == 0 && "Illegal splat"); 83 return APInt::getSplat(NumBits, *Bits); 84 } 85 } 86 87 APInt Bits = APInt::getZero(NumBits); 88 for (unsigned I = 0, E = CV->getNumOperands(); I != E; ++I) { 89 Constant *Elt = CV->getOperand(I); 90 std::optional<APInt> SubBits = extractConstantBits(Elt); 91 if (!SubBits) 92 return std::nullopt; 93 assert(NumBits == (E * SubBits->getBitWidth()) && 94 "Illegal vector element size"); 95 Bits.insertBits(*SubBits, I * SubBits->getBitWidth()); 96 } 97 return Bits; 98 } 99 100 if (auto *CDS = dyn_cast<ConstantDataSequential>(C)) { 101 bool IsInteger = CDS->getElementType()->isIntegerTy(); 102 bool IsFloat = CDS->getElementType()->isHalfTy() || 103 CDS->getElementType()->isBFloatTy() || 104 CDS->getElementType()->isFloatTy() || 105 CDS->getElementType()->isDoubleTy(); 106 if (IsInteger || IsFloat) { 107 APInt Bits = APInt::getZero(NumBits); 108 unsigned EltBits = CDS->getElementType()->getPrimitiveSizeInBits(); 109 for (unsigned I = 0, E = CDS->getNumElements(); I != E; ++I) { 110 if (IsInteger) 111 Bits.insertBits(CDS->getElementAsAPInt(I), I * EltBits); 112 else 113 Bits.insertBits(CDS->getElementAsAPFloat(I).bitcastToAPInt(), 114 I * EltBits); 115 } 116 return Bits; 117 } 118 } 119 120 return std::nullopt; 121 } 122 123 // Attempt to compute the splat width of bits data by normalizing the splat to 124 // remove undefs. 125 static std::optional<APInt> getSplatableConstant(const Constant *C, 126 unsigned SplatBitWidth) { 127 const Type *Ty = C->getType(); 128 assert((Ty->getPrimitiveSizeInBits() % SplatBitWidth) == 0 && 129 "Illegal splat width"); 130 131 if (std::optional<APInt> Bits = extractConstantBits(C)) 132 if (Bits->isSplat(SplatBitWidth)) 133 return Bits->trunc(SplatBitWidth); 134 135 // Detect general splats with undefs. 136 // TODO: Do we need to handle NumEltsBits > SplatBitWidth splitting? 137 if (auto *CV = dyn_cast<ConstantVector>(C)) { 138 unsigned NumOps = CV->getNumOperands(); 139 unsigned NumEltsBits = Ty->getScalarSizeInBits(); 140 unsigned NumScaleOps = SplatBitWidth / NumEltsBits; 141 if ((SplatBitWidth % NumEltsBits) == 0) { 142 // Collect the elements and ensure that within the repeated splat sequence 143 // they either match or are undef. 144 SmallVector<Constant *, 16> Sequence(NumScaleOps, nullptr); 145 for (unsigned Idx = 0; Idx != NumOps; ++Idx) { 146 if (Constant *Elt = CV->getAggregateElement(Idx)) { 147 if (isa<UndefValue>(Elt)) 148 continue; 149 unsigned SplatIdx = Idx % NumScaleOps; 150 if (!Sequence[SplatIdx] || Sequence[SplatIdx] == Elt) { 151 Sequence[SplatIdx] = Elt; 152 continue; 153 } 154 } 155 return std::nullopt; 156 } 157 // Extract the constant bits forming the splat and insert into the bits 158 // data, leave undef as zero. 159 APInt SplatBits = APInt::getZero(SplatBitWidth); 160 for (unsigned I = 0; I != NumScaleOps; ++I) { 161 if (!Sequence[I]) 162 continue; 163 if (std::optional<APInt> Bits = extractConstantBits(Sequence[I])) { 164 SplatBits.insertBits(*Bits, I * Bits->getBitWidth()); 165 continue; 166 } 167 return std::nullopt; 168 } 169 return SplatBits; 170 } 171 } 172 173 return std::nullopt; 174 } 175 176 // Split raw bits into a constant vector of elements of a specific bit width. 177 // NOTE: We don't always bother converting to scalars if the vector length is 1. 178 static Constant *rebuildConstant(LLVMContext &Ctx, Type *SclTy, 179 const APInt &Bits, unsigned NumSclBits) { 180 unsigned BitWidth = Bits.getBitWidth(); 181 182 if (NumSclBits == 8) { 183 SmallVector<uint8_t> RawBits; 184 for (unsigned I = 0; I != BitWidth; I += 8) 185 RawBits.push_back(Bits.extractBits(8, I).getZExtValue()); 186 return ConstantDataVector::get(Ctx, RawBits); 187 } 188 189 if (NumSclBits == 16) { 190 SmallVector<uint16_t> RawBits; 191 for (unsigned I = 0; I != BitWidth; I += 16) 192 RawBits.push_back(Bits.extractBits(16, I).getZExtValue()); 193 if (SclTy->is16bitFPTy()) 194 return ConstantDataVector::getFP(SclTy, RawBits); 195 return ConstantDataVector::get(Ctx, RawBits); 196 } 197 198 if (NumSclBits == 32) { 199 SmallVector<uint32_t> RawBits; 200 for (unsigned I = 0; I != BitWidth; I += 32) 201 RawBits.push_back(Bits.extractBits(32, I).getZExtValue()); 202 if (SclTy->isFloatTy()) 203 return ConstantDataVector::getFP(SclTy, RawBits); 204 return ConstantDataVector::get(Ctx, RawBits); 205 } 206 207 assert(NumSclBits == 64 && "Unhandled vector element width"); 208 209 SmallVector<uint64_t> RawBits; 210 for (unsigned I = 0; I != BitWidth; I += 64) 211 RawBits.push_back(Bits.extractBits(64, I).getZExtValue()); 212 if (SclTy->isDoubleTy()) 213 return ConstantDataVector::getFP(SclTy, RawBits); 214 return ConstantDataVector::get(Ctx, RawBits); 215 } 216 217 // Attempt to rebuild a normalized splat vector constant of the requested splat 218 // width, built up of potentially smaller scalar values. 219 static Constant *rebuildSplatCst(const Constant *C, unsigned /*NumElts*/, 220 unsigned SplatBitWidth) { 221 std::optional<APInt> Splat = getSplatableConstant(C, SplatBitWidth); 222 if (!Splat) 223 return nullptr; 224 225 // Determine scalar size to use for the constant splat vector, clamping as we 226 // might have found a splat smaller than the original constant data. 227 const Type *OriginalType = C->getType(); 228 Type *SclTy = OriginalType->getScalarType(); 229 unsigned NumSclBits = SclTy->getPrimitiveSizeInBits(); 230 NumSclBits = std::min<unsigned>(NumSclBits, SplatBitWidth); 231 232 // Fallback to i64 / double. 233 NumSclBits = (NumSclBits == 8 || NumSclBits == 16 || NumSclBits == 32) 234 ? NumSclBits 235 : 64; 236 237 // Extract per-element bits. 238 return rebuildConstant(OriginalType->getContext(), SclTy, *Splat, NumSclBits); 239 } 240 241 static Constant *rebuildZeroUpperCst(const Constant *C, unsigned /*NumElts*/, 242 unsigned ScalarBitWidth) { 243 Type *Ty = C->getType(); 244 Type *SclTy = Ty->getScalarType(); 245 unsigned NumBits = Ty->getPrimitiveSizeInBits(); 246 unsigned NumSclBits = SclTy->getPrimitiveSizeInBits(); 247 LLVMContext &Ctx = C->getContext(); 248 249 if (NumBits > ScalarBitWidth) { 250 // Determine if the upper bits are all zero. 251 if (std::optional<APInt> Bits = extractConstantBits(C)) { 252 if (Bits->countLeadingZeros() >= (NumBits - ScalarBitWidth)) { 253 // If the original constant was made of smaller elements, try to retain 254 // those types. 255 if (ScalarBitWidth > NumSclBits && (ScalarBitWidth % NumSclBits) == 0) 256 return rebuildConstant(Ctx, SclTy, *Bits, NumSclBits); 257 258 // Fallback to raw integer bits. 259 APInt RawBits = Bits->zextOrTrunc(ScalarBitWidth); 260 return ConstantInt::get(Ctx, RawBits); 261 } 262 } 263 } 264 265 return nullptr; 266 } 267 268 bool X86FixupVectorConstantsPass::processInstruction(MachineFunction &MF, 269 MachineBasicBlock &MBB, 270 MachineInstr &MI) { 271 unsigned Opc = MI.getOpcode(); 272 MachineConstantPool *CP = MI.getParent()->getParent()->getConstantPool(); 273 bool HasAVX2 = ST->hasAVX2(); 274 bool HasDQI = ST->hasDQI(); 275 bool HasBWI = ST->hasBWI(); 276 bool HasVLX = ST->hasVLX(); 277 278 struct FixupEntry { 279 int Op; 280 int NumCstElts; 281 int BitWidth; 282 std::function<Constant *(const Constant *, unsigned, unsigned)> 283 RebuildConstant; 284 }; 285 auto FixupConstant = [&](ArrayRef<FixupEntry> Fixups, unsigned OperandNo) { 286 #ifdef EXPENSIVE_CHECKS 287 assert(llvm::is_sorted(Fixups, 288 [](const FixupEntry &A, const FixupEntry &B) { 289 return (A.NumCstElts * A.BitWidth) < 290 (B.NumCstElts * B.BitWidth); 291 }) && 292 "Constant fixup table not sorted in ascending constant size"); 293 #endif 294 assert(MI.getNumOperands() >= (OperandNo + X86::AddrNumOperands) && 295 "Unexpected number of operands!"); 296 if (auto *C = X86::getConstantFromPool(MI, OperandNo)) { 297 for (const FixupEntry &Fixup : Fixups) { 298 if (Fixup.Op) { 299 // Construct a suitable constant and adjust the MI to use the new 300 // constant pool entry. 301 if (Constant *NewCst = 302 Fixup.RebuildConstant(C, Fixup.NumCstElts, Fixup.BitWidth)) { 303 unsigned NewCPI = 304 CP->getConstantPoolIndex(NewCst, Align(Fixup.BitWidth / 8)); 305 MI.setDesc(TII->get(Fixup.Op)); 306 MI.getOperand(OperandNo + X86::AddrDisp).setIndex(NewCPI); 307 return true; 308 } 309 } 310 } 311 } 312 return false; 313 }; 314 315 // Attempt to convert full width vector loads into broadcast/vzload loads. 316 switch (Opc) { 317 /* FP Loads */ 318 case X86::MOVAPDrm: 319 case X86::MOVAPSrm: 320 case X86::MOVUPDrm: 321 case X86::MOVUPSrm: 322 // TODO: SSE3 MOVDDUP Handling 323 return FixupConstant({{X86::MOVSSrm, 1, 32, rebuildZeroUpperCst}, 324 {X86::MOVSDrm, 1, 64, rebuildZeroUpperCst}}, 325 1); 326 case X86::VMOVAPDrm: 327 case X86::VMOVAPSrm: 328 case X86::VMOVUPDrm: 329 case X86::VMOVUPSrm: 330 return FixupConstant({{X86::VMOVSSrm, 1, 32, rebuildZeroUpperCst}, 331 {X86::VBROADCASTSSrm, 1, 32, rebuildSplatCst}, 332 {X86::VMOVSDrm, 1, 64, rebuildZeroUpperCst}, 333 {X86::VMOVDDUPrm, 1, 64, rebuildSplatCst}}, 334 1); 335 case X86::VMOVAPDYrm: 336 case X86::VMOVAPSYrm: 337 case X86::VMOVUPDYrm: 338 case X86::VMOVUPSYrm: 339 return FixupConstant({{X86::VBROADCASTSSYrm, 1, 32, rebuildSplatCst}, 340 {X86::VBROADCASTSDYrm, 1, 64, rebuildSplatCst}, 341 {X86::VBROADCASTF128rm, 1, 128, rebuildSplatCst}}, 342 1); 343 case X86::VMOVAPDZ128rm: 344 case X86::VMOVAPSZ128rm: 345 case X86::VMOVUPDZ128rm: 346 case X86::VMOVUPSZ128rm: 347 return FixupConstant({{X86::VMOVSSZrm, 1, 32, rebuildZeroUpperCst}, 348 {X86::VBROADCASTSSZ128rm, 1, 32, rebuildSplatCst}, 349 {X86::VMOVSDZrm, 1, 64, rebuildZeroUpperCst}, 350 {X86::VMOVDDUPZ128rm, 1, 64, rebuildSplatCst}}, 351 1); 352 case X86::VMOVAPDZ256rm: 353 case X86::VMOVAPSZ256rm: 354 case X86::VMOVUPDZ256rm: 355 case X86::VMOVUPSZ256rm: 356 return FixupConstant( 357 {{X86::VBROADCASTSSZ256rm, 1, 32, rebuildSplatCst}, 358 {X86::VBROADCASTSDZ256rm, 1, 64, rebuildSplatCst}, 359 {X86::VBROADCASTF32X4Z256rm, 1, 128, rebuildSplatCst}}, 360 1); 361 case X86::VMOVAPDZrm: 362 case X86::VMOVAPSZrm: 363 case X86::VMOVUPDZrm: 364 case X86::VMOVUPSZrm: 365 return FixupConstant({{X86::VBROADCASTSSZrm, 1, 32, rebuildSplatCst}, 366 {X86::VBROADCASTSDZrm, 1, 64, rebuildSplatCst}, 367 {X86::VBROADCASTF32X4rm, 1, 128, rebuildSplatCst}, 368 {X86::VBROADCASTF64X4rm, 1, 256, rebuildSplatCst}}, 369 1); 370 /* Integer Loads */ 371 case X86::MOVDQArm: 372 case X86::MOVDQUrm: { 373 return FixupConstant({{X86::MOVDI2PDIrm, 1, 32, rebuildZeroUpperCst}, 374 {X86::MOVQI2PQIrm, 1, 64, rebuildZeroUpperCst}}, 375 1); 376 } 377 case X86::VMOVDQArm: 378 case X86::VMOVDQUrm: { 379 FixupEntry Fixups[] = { 380 {HasAVX2 ? X86::VPBROADCASTBrm : 0, 1, 8, rebuildSplatCst}, 381 {HasAVX2 ? X86::VPBROADCASTWrm : 0, 1, 16, rebuildSplatCst}, 382 {X86::VMOVDI2PDIrm, 1, 32, rebuildZeroUpperCst}, 383 {HasAVX2 ? X86::VPBROADCASTDrm : X86::VBROADCASTSSrm, 1, 32, 384 rebuildSplatCst}, 385 {X86::VMOVQI2PQIrm, 1, 64, rebuildZeroUpperCst}, 386 {HasAVX2 ? X86::VPBROADCASTQrm : X86::VMOVDDUPrm, 1, 64, 387 rebuildSplatCst}, 388 }; 389 return FixupConstant(Fixups, 1); 390 } 391 case X86::VMOVDQAYrm: 392 case X86::VMOVDQUYrm: { 393 FixupEntry Fixups[] = { 394 {HasAVX2 ? X86::VPBROADCASTBYrm : 0, 1, 8, rebuildSplatCst}, 395 {HasAVX2 ? X86::VPBROADCASTWYrm : 0, 1, 16, rebuildSplatCst}, 396 {HasAVX2 ? X86::VPBROADCASTDYrm : X86::VBROADCASTSSYrm, 1, 32, 397 rebuildSplatCst}, 398 {HasAVX2 ? X86::VPBROADCASTQYrm : X86::VBROADCASTSDYrm, 1, 64, 399 rebuildSplatCst}, 400 {HasAVX2 ? X86::VBROADCASTI128rm : X86::VBROADCASTF128rm, 1, 128, 401 rebuildSplatCst}}; 402 return FixupConstant(Fixups, 1); 403 } 404 case X86::VMOVDQA32Z128rm: 405 case X86::VMOVDQA64Z128rm: 406 case X86::VMOVDQU32Z128rm: 407 case X86::VMOVDQU64Z128rm: { 408 FixupEntry Fixups[] = { 409 {HasBWI ? X86::VPBROADCASTBZ128rm : 0, 1, 8, rebuildSplatCst}, 410 {HasBWI ? X86::VPBROADCASTWZ128rm : 0, 1, 16, rebuildSplatCst}, 411 {X86::VMOVDI2PDIZrm, 1, 32, rebuildZeroUpperCst}, 412 {X86::VPBROADCASTDZ128rm, 1, 32, rebuildSplatCst}, 413 {X86::VMOVQI2PQIZrm, 1, 64, rebuildZeroUpperCst}, 414 {X86::VPBROADCASTQZ128rm, 1, 64, rebuildSplatCst}}; 415 return FixupConstant(Fixups, 1); 416 } 417 case X86::VMOVDQA32Z256rm: 418 case X86::VMOVDQA64Z256rm: 419 case X86::VMOVDQU32Z256rm: 420 case X86::VMOVDQU64Z256rm: { 421 FixupEntry Fixups[] = { 422 {HasBWI ? X86::VPBROADCASTBZ256rm : 0, 1, 8, rebuildSplatCst}, 423 {HasBWI ? X86::VPBROADCASTWZ256rm : 0, 1, 16, rebuildSplatCst}, 424 {X86::VPBROADCASTDZ256rm, 1, 32, rebuildSplatCst}, 425 {X86::VPBROADCASTQZ256rm, 1, 64, rebuildSplatCst}, 426 {X86::VBROADCASTI32X4Z256rm, 1, 128, rebuildSplatCst}}; 427 return FixupConstant(Fixups, 1); 428 } 429 case X86::VMOVDQA32Zrm: 430 case X86::VMOVDQA64Zrm: 431 case X86::VMOVDQU32Zrm: 432 case X86::VMOVDQU64Zrm: { 433 FixupEntry Fixups[] = { 434 {HasBWI ? X86::VPBROADCASTBZrm : 0, 1, 8, rebuildSplatCst}, 435 {HasBWI ? X86::VPBROADCASTWZrm : 0, 1, 16, rebuildSplatCst}, 436 {X86::VPBROADCASTDZrm, 1, 32, rebuildSplatCst}, 437 {X86::VPBROADCASTQZrm, 1, 64, rebuildSplatCst}, 438 {X86::VBROADCASTI32X4rm, 1, 128, rebuildSplatCst}, 439 {X86::VBROADCASTI64X4rm, 1, 256, rebuildSplatCst}}; 440 return FixupConstant(Fixups, 1); 441 } 442 } 443 444 auto ConvertToBroadcastAVX512 = [&](unsigned OpSrc32, unsigned OpSrc64) { 445 unsigned OpBcst32 = 0, OpBcst64 = 0; 446 unsigned OpNoBcst32 = 0, OpNoBcst64 = 0; 447 if (OpSrc32) { 448 if (const X86FoldTableEntry *Mem2Bcst = 449 llvm::lookupBroadcastFoldTableBySize(OpSrc32, 32)) { 450 OpBcst32 = Mem2Bcst->DstOp; 451 OpNoBcst32 = Mem2Bcst->Flags & TB_INDEX_MASK; 452 } 453 } 454 if (OpSrc64) { 455 if (const X86FoldTableEntry *Mem2Bcst = 456 llvm::lookupBroadcastFoldTableBySize(OpSrc64, 64)) { 457 OpBcst64 = Mem2Bcst->DstOp; 458 OpNoBcst64 = Mem2Bcst->Flags & TB_INDEX_MASK; 459 } 460 } 461 assert(((OpBcst32 == 0) || (OpBcst64 == 0) || (OpNoBcst32 == OpNoBcst64)) && 462 "OperandNo mismatch"); 463 464 if (OpBcst32 || OpBcst64) { 465 unsigned OpNo = OpBcst32 == 0 ? OpNoBcst64 : OpNoBcst32; 466 FixupEntry Fixups[] = {{(int)OpBcst32, 32, 32, rebuildSplatCst}, 467 {(int)OpBcst64, 64, 64, rebuildSplatCst}}; 468 return FixupConstant(Fixups, OpNo); 469 } 470 return false; 471 }; 472 473 // Attempt to find a AVX512 mapping from a full width memory-fold instruction 474 // to a broadcast-fold instruction variant. 475 if ((MI.getDesc().TSFlags & X86II::EncodingMask) == X86II::EVEX) 476 return ConvertToBroadcastAVX512(Opc, Opc); 477 478 // Reverse the X86InstrInfo::setExecutionDomainCustom EVEX->VEX logic 479 // conversion to see if we can convert to a broadcasted (integer) logic op. 480 if (HasVLX && !HasDQI) { 481 unsigned OpSrc32 = 0, OpSrc64 = 0; 482 switch (Opc) { 483 case X86::VANDPDrm: 484 case X86::VANDPSrm: 485 case X86::VPANDrm: 486 OpSrc32 = X86 ::VPANDDZ128rm; 487 OpSrc64 = X86 ::VPANDQZ128rm; 488 break; 489 case X86::VANDPDYrm: 490 case X86::VANDPSYrm: 491 case X86::VPANDYrm: 492 OpSrc32 = X86 ::VPANDDZ256rm; 493 OpSrc64 = X86 ::VPANDQZ256rm; 494 break; 495 case X86::VANDNPDrm: 496 case X86::VANDNPSrm: 497 case X86::VPANDNrm: 498 OpSrc32 = X86 ::VPANDNDZ128rm; 499 OpSrc64 = X86 ::VPANDNQZ128rm; 500 break; 501 case X86::VANDNPDYrm: 502 case X86::VANDNPSYrm: 503 case X86::VPANDNYrm: 504 OpSrc32 = X86 ::VPANDNDZ256rm; 505 OpSrc64 = X86 ::VPANDNQZ256rm; 506 break; 507 case X86::VORPDrm: 508 case X86::VORPSrm: 509 case X86::VPORrm: 510 OpSrc32 = X86 ::VPORDZ128rm; 511 OpSrc64 = X86 ::VPORQZ128rm; 512 break; 513 case X86::VORPDYrm: 514 case X86::VORPSYrm: 515 case X86::VPORYrm: 516 OpSrc32 = X86 ::VPORDZ256rm; 517 OpSrc64 = X86 ::VPORQZ256rm; 518 break; 519 case X86::VXORPDrm: 520 case X86::VXORPSrm: 521 case X86::VPXORrm: 522 OpSrc32 = X86 ::VPXORDZ128rm; 523 OpSrc64 = X86 ::VPXORQZ128rm; 524 break; 525 case X86::VXORPDYrm: 526 case X86::VXORPSYrm: 527 case X86::VPXORYrm: 528 OpSrc32 = X86 ::VPXORDZ256rm; 529 OpSrc64 = X86 ::VPXORQZ256rm; 530 break; 531 } 532 if (OpSrc32 || OpSrc64) 533 return ConvertToBroadcastAVX512(OpSrc32, OpSrc64); 534 } 535 536 return false; 537 } 538 539 bool X86FixupVectorConstantsPass::runOnMachineFunction(MachineFunction &MF) { 540 LLVM_DEBUG(dbgs() << "Start X86FixupVectorConstants\n";); 541 bool Changed = false; 542 ST = &MF.getSubtarget<X86Subtarget>(); 543 TII = ST->getInstrInfo(); 544 SM = &ST->getSchedModel(); 545 546 for (MachineBasicBlock &MBB : MF) { 547 for (MachineInstr &MI : MBB) { 548 if (processInstruction(MF, MBB, MI)) { 549 ++NumInstChanges; 550 Changed = true; 551 } 552 } 553 } 554 LLVM_DEBUG(dbgs() << "End X86FixupVectorConstants\n";); 555 return Changed; 556 } 557