1 //===- llvm/Analysis/TargetTransformInfo.cpp ------------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 10 #include "llvm/Analysis/TargetTransformInfo.h" 11 #include "llvm/Analysis/TargetTransformInfoImpl.h" 12 #include "llvm/IR/CallSite.h" 13 #include "llvm/IR/DataLayout.h" 14 #include "llvm/IR/Instruction.h" 15 #include "llvm/IR/Instructions.h" 16 #include "llvm/IR/IntrinsicInst.h" 17 #include "llvm/IR/Module.h" 18 #include "llvm/IR/Operator.h" 19 #include "llvm/IR/PatternMatch.h" 20 #include "llvm/Support/CommandLine.h" 21 #include "llvm/Support/ErrorHandling.h" 22 #include <utility> 23 24 using namespace llvm; 25 using namespace PatternMatch; 26 27 #define DEBUG_TYPE "tti" 28 29 static cl::opt<bool> EnableReduxCost("costmodel-reduxcost", cl::init(false), 30 cl::Hidden, 31 cl::desc("Recognize reduction patterns.")); 32 33 namespace { 34 /// No-op implementation of the TTI interface using the utility base 35 /// classes. 36 /// 37 /// This is used when no target specific information is available. 38 struct NoTTIImpl : TargetTransformInfoImplCRTPBase<NoTTIImpl> { 39 explicit NoTTIImpl(const DataLayout &DL) 40 : TargetTransformInfoImplCRTPBase<NoTTIImpl>(DL) {} 41 }; 42 } 43 44 TargetTransformInfo::TargetTransformInfo(const DataLayout &DL) 45 : TTIImpl(new Model<NoTTIImpl>(NoTTIImpl(DL))) {} 46 47 TargetTransformInfo::~TargetTransformInfo() {} 48 49 TargetTransformInfo::TargetTransformInfo(TargetTransformInfo &&Arg) 50 : TTIImpl(std::move(Arg.TTIImpl)) {} 51 52 TargetTransformInfo &TargetTransformInfo::operator=(TargetTransformInfo &&RHS) { 53 TTIImpl = std::move(RHS.TTIImpl); 54 return *this; 55 } 56 57 int TargetTransformInfo::getOperationCost(unsigned Opcode, Type *Ty, 58 Type *OpTy) const { 59 int Cost = TTIImpl->getOperationCost(Opcode, Ty, OpTy); 60 assert(Cost >= 0 && "TTI should not produce negative costs!"); 61 return Cost; 62 } 63 64 int TargetTransformInfo::getCallCost(FunctionType *FTy, int NumArgs) const { 65 int Cost = TTIImpl->getCallCost(FTy, NumArgs); 66 assert(Cost >= 0 && "TTI should not produce negative costs!"); 67 return Cost; 68 } 69 70 int TargetTransformInfo::getCallCost(const Function *F, 71 ArrayRef<const Value *> Arguments) const { 72 int Cost = TTIImpl->getCallCost(F, Arguments); 73 assert(Cost >= 0 && "TTI should not produce negative costs!"); 74 return Cost; 75 } 76 77 unsigned TargetTransformInfo::getInliningThresholdMultiplier() const { 78 return TTIImpl->getInliningThresholdMultiplier(); 79 } 80 81 int TargetTransformInfo::getGEPCost(Type *PointeeType, const Value *Ptr, 82 ArrayRef<const Value *> Operands) const { 83 return TTIImpl->getGEPCost(PointeeType, Ptr, Operands); 84 } 85 86 int TargetTransformInfo::getExtCost(const Instruction *I, 87 const Value *Src) const { 88 return TTIImpl->getExtCost(I, Src); 89 } 90 91 int TargetTransformInfo::getIntrinsicCost( 92 Intrinsic::ID IID, Type *RetTy, ArrayRef<const Value *> Arguments) const { 93 int Cost = TTIImpl->getIntrinsicCost(IID, RetTy, Arguments); 94 assert(Cost >= 0 && "TTI should not produce negative costs!"); 95 return Cost; 96 } 97 98 unsigned 99 TargetTransformInfo::getEstimatedNumberOfCaseClusters(const SwitchInst &SI, 100 unsigned &JTSize) const { 101 return TTIImpl->getEstimatedNumberOfCaseClusters(SI, JTSize); 102 } 103 104 int TargetTransformInfo::getUserCost(const User *U, 105 ArrayRef<const Value *> Operands) const { 106 int Cost = TTIImpl->getUserCost(U, Operands); 107 assert(Cost >= 0 && "TTI should not produce negative costs!"); 108 return Cost; 109 } 110 111 bool TargetTransformInfo::hasBranchDivergence() const { 112 return TTIImpl->hasBranchDivergence(); 113 } 114 115 bool TargetTransformInfo::isSourceOfDivergence(const Value *V) const { 116 return TTIImpl->isSourceOfDivergence(V); 117 } 118 119 bool llvm::TargetTransformInfo::isAlwaysUniform(const Value *V) const { 120 return TTIImpl->isAlwaysUniform(V); 121 } 122 123 unsigned TargetTransformInfo::getFlatAddressSpace() const { 124 return TTIImpl->getFlatAddressSpace(); 125 } 126 127 bool TargetTransformInfo::isLoweredToCall(const Function *F) const { 128 return TTIImpl->isLoweredToCall(F); 129 } 130 131 void TargetTransformInfo::getUnrollingPreferences( 132 Loop *L, ScalarEvolution &SE, UnrollingPreferences &UP) const { 133 return TTIImpl->getUnrollingPreferences(L, SE, UP); 134 } 135 136 bool TargetTransformInfo::isLegalAddImmediate(int64_t Imm) const { 137 return TTIImpl->isLegalAddImmediate(Imm); 138 } 139 140 bool TargetTransformInfo::isLegalICmpImmediate(int64_t Imm) const { 141 return TTIImpl->isLegalICmpImmediate(Imm); 142 } 143 144 bool TargetTransformInfo::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, 145 int64_t BaseOffset, 146 bool HasBaseReg, 147 int64_t Scale, 148 unsigned AddrSpace, 149 Instruction *I) const { 150 return TTIImpl->isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg, 151 Scale, AddrSpace, I); 152 } 153 154 bool TargetTransformInfo::isLSRCostLess(LSRCost &C1, LSRCost &C2) const { 155 return TTIImpl->isLSRCostLess(C1, C2); 156 } 157 158 bool TargetTransformInfo::canMacroFuseCmp() const { 159 return TTIImpl->canMacroFuseCmp(); 160 } 161 162 bool TargetTransformInfo::shouldFavorPostInc() const { 163 return TTIImpl->shouldFavorPostInc(); 164 } 165 166 bool TargetTransformInfo::isLegalMaskedStore(Type *DataType) const { 167 return TTIImpl->isLegalMaskedStore(DataType); 168 } 169 170 bool TargetTransformInfo::isLegalMaskedLoad(Type *DataType) const { 171 return TTIImpl->isLegalMaskedLoad(DataType); 172 } 173 174 bool TargetTransformInfo::isLegalMaskedGather(Type *DataType) const { 175 return TTIImpl->isLegalMaskedGather(DataType); 176 } 177 178 bool TargetTransformInfo::isLegalMaskedScatter(Type *DataType) const { 179 return TTIImpl->isLegalMaskedScatter(DataType); 180 } 181 182 bool TargetTransformInfo::hasDivRemOp(Type *DataType, bool IsSigned) const { 183 return TTIImpl->hasDivRemOp(DataType, IsSigned); 184 } 185 186 bool TargetTransformInfo::hasVolatileVariant(Instruction *I, 187 unsigned AddrSpace) const { 188 return TTIImpl->hasVolatileVariant(I, AddrSpace); 189 } 190 191 bool TargetTransformInfo::prefersVectorizedAddressing() const { 192 return TTIImpl->prefersVectorizedAddressing(); 193 } 194 195 int TargetTransformInfo::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, 196 int64_t BaseOffset, 197 bool HasBaseReg, 198 int64_t Scale, 199 unsigned AddrSpace) const { 200 int Cost = TTIImpl->getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg, 201 Scale, AddrSpace); 202 assert(Cost >= 0 && "TTI should not produce negative costs!"); 203 return Cost; 204 } 205 206 bool TargetTransformInfo::LSRWithInstrQueries() const { 207 return TTIImpl->LSRWithInstrQueries(); 208 } 209 210 bool TargetTransformInfo::isTruncateFree(Type *Ty1, Type *Ty2) const { 211 return TTIImpl->isTruncateFree(Ty1, Ty2); 212 } 213 214 bool TargetTransformInfo::isProfitableToHoist(Instruction *I) const { 215 return TTIImpl->isProfitableToHoist(I); 216 } 217 218 bool TargetTransformInfo::useAA() const { return TTIImpl->useAA(); } 219 220 bool TargetTransformInfo::isTypeLegal(Type *Ty) const { 221 return TTIImpl->isTypeLegal(Ty); 222 } 223 224 unsigned TargetTransformInfo::getJumpBufAlignment() const { 225 return TTIImpl->getJumpBufAlignment(); 226 } 227 228 unsigned TargetTransformInfo::getJumpBufSize() const { 229 return TTIImpl->getJumpBufSize(); 230 } 231 232 bool TargetTransformInfo::shouldBuildLookupTables() const { 233 return TTIImpl->shouldBuildLookupTables(); 234 } 235 bool TargetTransformInfo::shouldBuildLookupTablesForConstant(Constant *C) const { 236 return TTIImpl->shouldBuildLookupTablesForConstant(C); 237 } 238 239 bool TargetTransformInfo::useColdCCForColdCall(Function &F) const { 240 return TTIImpl->useColdCCForColdCall(F); 241 } 242 243 unsigned TargetTransformInfo:: 244 getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const { 245 return TTIImpl->getScalarizationOverhead(Ty, Insert, Extract); 246 } 247 248 unsigned TargetTransformInfo:: 249 getOperandsScalarizationOverhead(ArrayRef<const Value *> Args, 250 unsigned VF) const { 251 return TTIImpl->getOperandsScalarizationOverhead(Args, VF); 252 } 253 254 bool TargetTransformInfo::supportsEfficientVectorElementLoadStore() const { 255 return TTIImpl->supportsEfficientVectorElementLoadStore(); 256 } 257 258 bool TargetTransformInfo::enableAggressiveInterleaving(bool LoopHasReductions) const { 259 return TTIImpl->enableAggressiveInterleaving(LoopHasReductions); 260 } 261 262 const TargetTransformInfo::MemCmpExpansionOptions * 263 TargetTransformInfo::enableMemCmpExpansion(bool IsZeroCmp) const { 264 return TTIImpl->enableMemCmpExpansion(IsZeroCmp); 265 } 266 267 bool TargetTransformInfo::enableInterleavedAccessVectorization() const { 268 return TTIImpl->enableInterleavedAccessVectorization(); 269 } 270 271 bool TargetTransformInfo::isFPVectorizationPotentiallyUnsafe() const { 272 return TTIImpl->isFPVectorizationPotentiallyUnsafe(); 273 } 274 275 bool TargetTransformInfo::allowsMisalignedMemoryAccesses(LLVMContext &Context, 276 unsigned BitWidth, 277 unsigned AddressSpace, 278 unsigned Alignment, 279 bool *Fast) const { 280 return TTIImpl->allowsMisalignedMemoryAccesses(Context, BitWidth, AddressSpace, 281 Alignment, Fast); 282 } 283 284 TargetTransformInfo::PopcntSupportKind 285 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit) const { 286 return TTIImpl->getPopcntSupport(IntTyWidthInBit); 287 } 288 289 bool TargetTransformInfo::haveFastSqrt(Type *Ty) const { 290 return TTIImpl->haveFastSqrt(Ty); 291 } 292 293 bool TargetTransformInfo::isFCmpOrdCheaperThanFCmpZero(Type *Ty) const { 294 return TTIImpl->isFCmpOrdCheaperThanFCmpZero(Ty); 295 } 296 297 int TargetTransformInfo::getFPOpCost(Type *Ty) const { 298 int Cost = TTIImpl->getFPOpCost(Ty); 299 assert(Cost >= 0 && "TTI should not produce negative costs!"); 300 return Cost; 301 } 302 303 int TargetTransformInfo::getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, 304 const APInt &Imm, 305 Type *Ty) const { 306 int Cost = TTIImpl->getIntImmCodeSizeCost(Opcode, Idx, Imm, Ty); 307 assert(Cost >= 0 && "TTI should not produce negative costs!"); 308 return Cost; 309 } 310 311 int TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty) const { 312 int Cost = TTIImpl->getIntImmCost(Imm, Ty); 313 assert(Cost >= 0 && "TTI should not produce negative costs!"); 314 return Cost; 315 } 316 317 int TargetTransformInfo::getIntImmCost(unsigned Opcode, unsigned Idx, 318 const APInt &Imm, Type *Ty) const { 319 int Cost = TTIImpl->getIntImmCost(Opcode, Idx, Imm, Ty); 320 assert(Cost >= 0 && "TTI should not produce negative costs!"); 321 return Cost; 322 } 323 324 int TargetTransformInfo::getIntImmCost(Intrinsic::ID IID, unsigned Idx, 325 const APInt &Imm, Type *Ty) const { 326 int Cost = TTIImpl->getIntImmCost(IID, Idx, Imm, Ty); 327 assert(Cost >= 0 && "TTI should not produce negative costs!"); 328 return Cost; 329 } 330 331 unsigned TargetTransformInfo::getNumberOfRegisters(bool Vector) const { 332 return TTIImpl->getNumberOfRegisters(Vector); 333 } 334 335 unsigned TargetTransformInfo::getRegisterBitWidth(bool Vector) const { 336 return TTIImpl->getRegisterBitWidth(Vector); 337 } 338 339 unsigned TargetTransformInfo::getMinVectorRegisterBitWidth() const { 340 return TTIImpl->getMinVectorRegisterBitWidth(); 341 } 342 343 bool TargetTransformInfo::shouldMaximizeVectorBandwidth(bool OptSize) const { 344 return TTIImpl->shouldMaximizeVectorBandwidth(OptSize); 345 } 346 347 unsigned TargetTransformInfo::getMinimumVF(unsigned ElemWidth) const { 348 return TTIImpl->getMinimumVF(ElemWidth); 349 } 350 351 bool TargetTransformInfo::shouldConsiderAddressTypePromotion( 352 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const { 353 return TTIImpl->shouldConsiderAddressTypePromotion( 354 I, AllowPromotionWithoutCommonHeader); 355 } 356 357 unsigned TargetTransformInfo::getCacheLineSize() const { 358 return TTIImpl->getCacheLineSize(); 359 } 360 361 llvm::Optional<unsigned> TargetTransformInfo::getCacheSize(CacheLevel Level) 362 const { 363 return TTIImpl->getCacheSize(Level); 364 } 365 366 llvm::Optional<unsigned> TargetTransformInfo::getCacheAssociativity( 367 CacheLevel Level) const { 368 return TTIImpl->getCacheAssociativity(Level); 369 } 370 371 unsigned TargetTransformInfo::getPrefetchDistance() const { 372 return TTIImpl->getPrefetchDistance(); 373 } 374 375 unsigned TargetTransformInfo::getMinPrefetchStride() const { 376 return TTIImpl->getMinPrefetchStride(); 377 } 378 379 unsigned TargetTransformInfo::getMaxPrefetchIterationsAhead() const { 380 return TTIImpl->getMaxPrefetchIterationsAhead(); 381 } 382 383 unsigned TargetTransformInfo::getMaxInterleaveFactor(unsigned VF) const { 384 return TTIImpl->getMaxInterleaveFactor(VF); 385 } 386 387 int TargetTransformInfo::getArithmeticInstrCost( 388 unsigned Opcode, Type *Ty, OperandValueKind Opd1Info, 389 OperandValueKind Opd2Info, OperandValueProperties Opd1PropInfo, 390 OperandValueProperties Opd2PropInfo, 391 ArrayRef<const Value *> Args) const { 392 int Cost = TTIImpl->getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info, 393 Opd1PropInfo, Opd2PropInfo, Args); 394 assert(Cost >= 0 && "TTI should not produce negative costs!"); 395 return Cost; 396 } 397 398 int TargetTransformInfo::getShuffleCost(ShuffleKind Kind, Type *Ty, int Index, 399 Type *SubTp) const { 400 int Cost = TTIImpl->getShuffleCost(Kind, Ty, Index, SubTp); 401 assert(Cost >= 0 && "TTI should not produce negative costs!"); 402 return Cost; 403 } 404 405 int TargetTransformInfo::getCastInstrCost(unsigned Opcode, Type *Dst, 406 Type *Src, const Instruction *I) const { 407 assert ((I == nullptr || I->getOpcode() == Opcode) && 408 "Opcode should reflect passed instruction."); 409 int Cost = TTIImpl->getCastInstrCost(Opcode, Dst, Src, I); 410 assert(Cost >= 0 && "TTI should not produce negative costs!"); 411 return Cost; 412 } 413 414 int TargetTransformInfo::getExtractWithExtendCost(unsigned Opcode, Type *Dst, 415 VectorType *VecTy, 416 unsigned Index) const { 417 int Cost = TTIImpl->getExtractWithExtendCost(Opcode, Dst, VecTy, Index); 418 assert(Cost >= 0 && "TTI should not produce negative costs!"); 419 return Cost; 420 } 421 422 int TargetTransformInfo::getCFInstrCost(unsigned Opcode) const { 423 int Cost = TTIImpl->getCFInstrCost(Opcode); 424 assert(Cost >= 0 && "TTI should not produce negative costs!"); 425 return Cost; 426 } 427 428 int TargetTransformInfo::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 429 Type *CondTy, const Instruction *I) const { 430 assert ((I == nullptr || I->getOpcode() == Opcode) && 431 "Opcode should reflect passed instruction."); 432 int Cost = TTIImpl->getCmpSelInstrCost(Opcode, ValTy, CondTy, I); 433 assert(Cost >= 0 && "TTI should not produce negative costs!"); 434 return Cost; 435 } 436 437 int TargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val, 438 unsigned Index) const { 439 int Cost = TTIImpl->getVectorInstrCost(Opcode, Val, Index); 440 assert(Cost >= 0 && "TTI should not produce negative costs!"); 441 return Cost; 442 } 443 444 int TargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src, 445 unsigned Alignment, 446 unsigned AddressSpace, 447 const Instruction *I) const { 448 assert ((I == nullptr || I->getOpcode() == Opcode) && 449 "Opcode should reflect passed instruction."); 450 int Cost = TTIImpl->getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, I); 451 assert(Cost >= 0 && "TTI should not produce negative costs!"); 452 return Cost; 453 } 454 455 int TargetTransformInfo::getMaskedMemoryOpCost(unsigned Opcode, Type *Src, 456 unsigned Alignment, 457 unsigned AddressSpace) const { 458 int Cost = 459 TTIImpl->getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace); 460 assert(Cost >= 0 && "TTI should not produce negative costs!"); 461 return Cost; 462 } 463 464 int TargetTransformInfo::getGatherScatterOpCost(unsigned Opcode, Type *DataTy, 465 Value *Ptr, bool VariableMask, 466 unsigned Alignment) const { 467 int Cost = TTIImpl->getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask, 468 Alignment); 469 assert(Cost >= 0 && "TTI should not produce negative costs!"); 470 return Cost; 471 } 472 473 int TargetTransformInfo::getInterleavedMemoryOpCost( 474 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices, 475 unsigned Alignment, unsigned AddressSpace) const { 476 int Cost = TTIImpl->getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 477 Alignment, AddressSpace); 478 assert(Cost >= 0 && "TTI should not produce negative costs!"); 479 return Cost; 480 } 481 482 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 483 ArrayRef<Type *> Tys, FastMathFlags FMF, 484 unsigned ScalarizationCostPassed) const { 485 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Tys, FMF, 486 ScalarizationCostPassed); 487 assert(Cost >= 0 && "TTI should not produce negative costs!"); 488 return Cost; 489 } 490 491 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 492 ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) const { 493 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF); 494 assert(Cost >= 0 && "TTI should not produce negative costs!"); 495 return Cost; 496 } 497 498 int TargetTransformInfo::getCallInstrCost(Function *F, Type *RetTy, 499 ArrayRef<Type *> Tys) const { 500 int Cost = TTIImpl->getCallInstrCost(F, RetTy, Tys); 501 assert(Cost >= 0 && "TTI should not produce negative costs!"); 502 return Cost; 503 } 504 505 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const { 506 return TTIImpl->getNumberOfParts(Tp); 507 } 508 509 int TargetTransformInfo::getAddressComputationCost(Type *Tp, 510 ScalarEvolution *SE, 511 const SCEV *Ptr) const { 512 int Cost = TTIImpl->getAddressComputationCost(Tp, SE, Ptr); 513 assert(Cost >= 0 && "TTI should not produce negative costs!"); 514 return Cost; 515 } 516 517 int TargetTransformInfo::getArithmeticReductionCost(unsigned Opcode, Type *Ty, 518 bool IsPairwiseForm) const { 519 int Cost = TTIImpl->getArithmeticReductionCost(Opcode, Ty, IsPairwiseForm); 520 assert(Cost >= 0 && "TTI should not produce negative costs!"); 521 return Cost; 522 } 523 524 int TargetTransformInfo::getMinMaxReductionCost(Type *Ty, Type *CondTy, 525 bool IsPairwiseForm, 526 bool IsUnsigned) const { 527 int Cost = 528 TTIImpl->getMinMaxReductionCost(Ty, CondTy, IsPairwiseForm, IsUnsigned); 529 assert(Cost >= 0 && "TTI should not produce negative costs!"); 530 return Cost; 531 } 532 533 unsigned 534 TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const { 535 return TTIImpl->getCostOfKeepingLiveOverCall(Tys); 536 } 537 538 bool TargetTransformInfo::getTgtMemIntrinsic(IntrinsicInst *Inst, 539 MemIntrinsicInfo &Info) const { 540 return TTIImpl->getTgtMemIntrinsic(Inst, Info); 541 } 542 543 unsigned TargetTransformInfo::getAtomicMemIntrinsicMaxElementSize() const { 544 return TTIImpl->getAtomicMemIntrinsicMaxElementSize(); 545 } 546 547 Value *TargetTransformInfo::getOrCreateResultFromMemIntrinsic( 548 IntrinsicInst *Inst, Type *ExpectedType) const { 549 return TTIImpl->getOrCreateResultFromMemIntrinsic(Inst, ExpectedType); 550 } 551 552 Type *TargetTransformInfo::getMemcpyLoopLoweringType(LLVMContext &Context, 553 Value *Length, 554 unsigned SrcAlign, 555 unsigned DestAlign) const { 556 return TTIImpl->getMemcpyLoopLoweringType(Context, Length, SrcAlign, 557 DestAlign); 558 } 559 560 void TargetTransformInfo::getMemcpyLoopResidualLoweringType( 561 SmallVectorImpl<Type *> &OpsOut, LLVMContext &Context, 562 unsigned RemainingBytes, unsigned SrcAlign, unsigned DestAlign) const { 563 TTIImpl->getMemcpyLoopResidualLoweringType(OpsOut, Context, RemainingBytes, 564 SrcAlign, DestAlign); 565 } 566 567 bool TargetTransformInfo::areInlineCompatible(const Function *Caller, 568 const Function *Callee) const { 569 return TTIImpl->areInlineCompatible(Caller, Callee); 570 } 571 572 bool TargetTransformInfo::isIndexedLoadLegal(MemIndexedMode Mode, 573 Type *Ty) const { 574 return TTIImpl->isIndexedLoadLegal(Mode, Ty); 575 } 576 577 bool TargetTransformInfo::isIndexedStoreLegal(MemIndexedMode Mode, 578 Type *Ty) const { 579 return TTIImpl->isIndexedStoreLegal(Mode, Ty); 580 } 581 582 unsigned TargetTransformInfo::getLoadStoreVecRegBitWidth(unsigned AS) const { 583 return TTIImpl->getLoadStoreVecRegBitWidth(AS); 584 } 585 586 bool TargetTransformInfo::isLegalToVectorizeLoad(LoadInst *LI) const { 587 return TTIImpl->isLegalToVectorizeLoad(LI); 588 } 589 590 bool TargetTransformInfo::isLegalToVectorizeStore(StoreInst *SI) const { 591 return TTIImpl->isLegalToVectorizeStore(SI); 592 } 593 594 bool TargetTransformInfo::isLegalToVectorizeLoadChain( 595 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 596 return TTIImpl->isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment, 597 AddrSpace); 598 } 599 600 bool TargetTransformInfo::isLegalToVectorizeStoreChain( 601 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 602 return TTIImpl->isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment, 603 AddrSpace); 604 } 605 606 unsigned TargetTransformInfo::getLoadVectorFactor(unsigned VF, 607 unsigned LoadSize, 608 unsigned ChainSizeInBytes, 609 VectorType *VecTy) const { 610 return TTIImpl->getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy); 611 } 612 613 unsigned TargetTransformInfo::getStoreVectorFactor(unsigned VF, 614 unsigned StoreSize, 615 unsigned ChainSizeInBytes, 616 VectorType *VecTy) const { 617 return TTIImpl->getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy); 618 } 619 620 bool TargetTransformInfo::useReductionIntrinsic(unsigned Opcode, 621 Type *Ty, ReductionFlags Flags) const { 622 return TTIImpl->useReductionIntrinsic(Opcode, Ty, Flags); 623 } 624 625 bool TargetTransformInfo::shouldExpandReduction(const IntrinsicInst *II) const { 626 return TTIImpl->shouldExpandReduction(II); 627 } 628 629 int TargetTransformInfo::getInstructionLatency(const Instruction *I) const { 630 return TTIImpl->getInstructionLatency(I); 631 } 632 633 static bool isReverseVectorMask(ArrayRef<int> Mask) { 634 for (unsigned i = 0, MaskSize = Mask.size(); i < MaskSize; ++i) 635 if (Mask[i] >= 0 && Mask[i] != (int)(MaskSize - 1 - i)) 636 return false; 637 return true; 638 } 639 640 static bool isSingleSourceVectorMask(ArrayRef<int> Mask) { 641 bool Vec0 = false; 642 bool Vec1 = false; 643 for (unsigned i = 0, NumVecElts = Mask.size(); i < NumVecElts; ++i) { 644 if (Mask[i] >= 0) { 645 if ((unsigned)Mask[i] >= NumVecElts) 646 Vec1 = true; 647 else 648 Vec0 = true; 649 } 650 } 651 return !(Vec0 && Vec1); 652 } 653 654 static bool isZeroEltBroadcastVectorMask(ArrayRef<int> Mask) { 655 for (unsigned i = 0; i < Mask.size(); ++i) 656 if (Mask[i] > 0) 657 return false; 658 return true; 659 } 660 661 static bool isAlternateVectorMask(ArrayRef<int> Mask) { 662 bool isAlternate = true; 663 unsigned MaskSize = Mask.size(); 664 665 // Example: shufflevector A, B, <0,5,2,7> 666 for (unsigned i = 0; i < MaskSize && isAlternate; ++i) { 667 if (Mask[i] < 0) 668 continue; 669 isAlternate = Mask[i] == (int)((i & 1) ? MaskSize + i : i); 670 } 671 672 if (isAlternate) 673 return true; 674 675 isAlternate = true; 676 // Example: shufflevector A, B, <4,1,6,3> 677 for (unsigned i = 0; i < MaskSize && isAlternate; ++i) { 678 if (Mask[i] < 0) 679 continue; 680 isAlternate = Mask[i] == (int)((i & 1) ? i : MaskSize + i); 681 } 682 683 return isAlternate; 684 } 685 686 static bool isTransposeVectorMask(ArrayRef<int> Mask) { 687 // Transpose vector masks transpose a 2xn matrix. They read corresponding 688 // even- or odd-numbered vector elements from two n-dimensional source 689 // vectors and write each result into consecutive elements of an 690 // n-dimensional destination vector. Two shuffles are necessary to complete 691 // the transpose, one for the even elements and another for the odd elements. 692 // This description closely follows how the TRN1 and TRN2 AArch64 693 // instructions operate. 694 // 695 // For example, a simple 2x2 matrix can be transposed with: 696 // 697 // ; Original matrix 698 // m0 = <a, b> 699 // m1 = <c, d> 700 // 701 // ; Transposed matrix 702 // t0 = <a, c> = shufflevector m0, m1, <0, 2> 703 // t1 = <b, d> = shufflevector m0, m1, <1, 3> 704 // 705 // For matrices having greater than n columns, the resulting nx2 transposed 706 // matrix is stored in two result vectors such that one vector contains 707 // interleaved elements from all the even-numbered rows and the other vector 708 // contains interleaved elements from all the odd-numbered rows. For example, 709 // a 2x4 matrix can be transposed with: 710 // 711 // ; Original matrix 712 // m0 = <a, b, c, d> 713 // m1 = <e, f, g, h> 714 // 715 // ; Transposed matrix 716 // t0 = <a, e, c, g> = shufflevector m0, m1 <0, 4, 2, 6> 717 // t1 = <b, f, d, h> = shufflevector m0, m1 <1, 5, 3, 7> 718 // 719 // The above explanation places limitations on what valid transpose masks can 720 // look like. These limitations are defined by the checks below. 721 // 722 // 1. The number of elements in the mask must be a power of two. 723 if (!isPowerOf2_32(Mask.size())) 724 return false; 725 726 // 2. The first element of the mask must be either a zero (for the 727 // even-numbered vector elements) or a one (for the odd-numbered vector 728 // elements). 729 if (Mask[0] != 0 && Mask[0] != 1) 730 return false; 731 732 // 3. The difference between the first two elements must be equal to the 733 // number of elements in the mask. 734 if (Mask[1] - Mask[0] != (int)Mask.size()) 735 return false; 736 737 // 4. The difference between consecutive even-numbered and odd-numbered 738 // elements must be equal to two. 739 for (int I = 2; I < (int)Mask.size(); ++I) 740 if (Mask[I] - Mask[I - 2] != 2) 741 return false; 742 743 return true; 744 } 745 746 static TargetTransformInfo::OperandValueKind getOperandInfo(Value *V) { 747 TargetTransformInfo::OperandValueKind OpInfo = 748 TargetTransformInfo::OK_AnyValue; 749 750 // Check for a splat of a constant or for a non uniform vector of constants. 751 if (isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) { 752 OpInfo = TargetTransformInfo::OK_NonUniformConstantValue; 753 if (cast<Constant>(V)->getSplatValue() != nullptr) 754 OpInfo = TargetTransformInfo::OK_UniformConstantValue; 755 } 756 757 // Check for a splat of a uniform value. This is not loop aware, so return 758 // true only for the obviously uniform cases (argument, globalvalue) 759 const Value *Splat = getSplatValue(V); 760 if (Splat && (isa<Argument>(Splat) || isa<GlobalValue>(Splat))) 761 OpInfo = TargetTransformInfo::OK_UniformValue; 762 763 return OpInfo; 764 } 765 766 static bool matchPairwiseShuffleMask(ShuffleVectorInst *SI, bool IsLeft, 767 unsigned Level) { 768 // We don't need a shuffle if we just want to have element 0 in position 0 of 769 // the vector. 770 if (!SI && Level == 0 && IsLeft) 771 return true; 772 else if (!SI) 773 return false; 774 775 SmallVector<int, 32> Mask(SI->getType()->getVectorNumElements(), -1); 776 777 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right) depending on whether 778 // we look at the left or right side. 779 for (unsigned i = 0, e = (1 << Level), val = !IsLeft; i != e; ++i, val += 2) 780 Mask[i] = val; 781 782 SmallVector<int, 16> ActualMask = SI->getShuffleMask(); 783 return Mask == ActualMask; 784 } 785 786 namespace { 787 /// Kind of the reduction data. 788 enum ReductionKind { 789 RK_None, /// Not a reduction. 790 RK_Arithmetic, /// Binary reduction data. 791 RK_MinMax, /// Min/max reduction data. 792 RK_UnsignedMinMax, /// Unsigned min/max reduction data. 793 }; 794 /// Contains opcode + LHS/RHS parts of the reduction operations. 795 struct ReductionData { 796 ReductionData() = delete; 797 ReductionData(ReductionKind Kind, unsigned Opcode, Value *LHS, Value *RHS) 798 : Opcode(Opcode), LHS(LHS), RHS(RHS), Kind(Kind) { 799 assert(Kind != RK_None && "expected binary or min/max reduction only."); 800 } 801 unsigned Opcode = 0; 802 Value *LHS = nullptr; 803 Value *RHS = nullptr; 804 ReductionKind Kind = RK_None; 805 bool hasSameData(ReductionData &RD) const { 806 return Kind == RD.Kind && Opcode == RD.Opcode; 807 } 808 }; 809 } // namespace 810 811 static Optional<ReductionData> getReductionData(Instruction *I) { 812 Value *L, *R; 813 if (m_BinOp(m_Value(L), m_Value(R)).match(I)) 814 return ReductionData(RK_Arithmetic, I->getOpcode(), L, R); 815 if (auto *SI = dyn_cast<SelectInst>(I)) { 816 if (m_SMin(m_Value(L), m_Value(R)).match(SI) || 817 m_SMax(m_Value(L), m_Value(R)).match(SI) || 818 m_OrdFMin(m_Value(L), m_Value(R)).match(SI) || 819 m_OrdFMax(m_Value(L), m_Value(R)).match(SI) || 820 m_UnordFMin(m_Value(L), m_Value(R)).match(SI) || 821 m_UnordFMax(m_Value(L), m_Value(R)).match(SI)) { 822 auto *CI = cast<CmpInst>(SI->getCondition()); 823 return ReductionData(RK_MinMax, CI->getOpcode(), L, R); 824 } 825 if (m_UMin(m_Value(L), m_Value(R)).match(SI) || 826 m_UMax(m_Value(L), m_Value(R)).match(SI)) { 827 auto *CI = cast<CmpInst>(SI->getCondition()); 828 return ReductionData(RK_UnsignedMinMax, CI->getOpcode(), L, R); 829 } 830 } 831 return llvm::None; 832 } 833 834 static ReductionKind matchPairwiseReductionAtLevel(Instruction *I, 835 unsigned Level, 836 unsigned NumLevels) { 837 // Match one level of pairwise operations. 838 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 839 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 840 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 841 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 842 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 843 if (!I) 844 return RK_None; 845 846 assert(I->getType()->isVectorTy() && "Expecting a vector type"); 847 848 Optional<ReductionData> RD = getReductionData(I); 849 if (!RD) 850 return RK_None; 851 852 ShuffleVectorInst *LS = dyn_cast<ShuffleVectorInst>(RD->LHS); 853 if (!LS && Level) 854 return RK_None; 855 ShuffleVectorInst *RS = dyn_cast<ShuffleVectorInst>(RD->RHS); 856 if (!RS && Level) 857 return RK_None; 858 859 // On level 0 we can omit one shufflevector instruction. 860 if (!Level && !RS && !LS) 861 return RK_None; 862 863 // Shuffle inputs must match. 864 Value *NextLevelOpL = LS ? LS->getOperand(0) : nullptr; 865 Value *NextLevelOpR = RS ? RS->getOperand(0) : nullptr; 866 Value *NextLevelOp = nullptr; 867 if (NextLevelOpR && NextLevelOpL) { 868 // If we have two shuffles their operands must match. 869 if (NextLevelOpL != NextLevelOpR) 870 return RK_None; 871 872 NextLevelOp = NextLevelOpL; 873 } else if (Level == 0 && (NextLevelOpR || NextLevelOpL)) { 874 // On the first level we can omit the shufflevector <0, undef,...>. So the 875 // input to the other shufflevector <1, undef> must match with one of the 876 // inputs to the current binary operation. 877 // Example: 878 // %NextLevelOpL = shufflevector %R, <1, undef ...> 879 // %BinOp = fadd %NextLevelOpL, %R 880 if (NextLevelOpL && NextLevelOpL != RD->RHS) 881 return RK_None; 882 else if (NextLevelOpR && NextLevelOpR != RD->LHS) 883 return RK_None; 884 885 NextLevelOp = NextLevelOpL ? RD->RHS : RD->LHS; 886 } else 887 return RK_None; 888 889 // Check that the next levels binary operation exists and matches with the 890 // current one. 891 if (Level + 1 != NumLevels) { 892 Optional<ReductionData> NextLevelRD = 893 getReductionData(cast<Instruction>(NextLevelOp)); 894 if (!NextLevelRD || !RD->hasSameData(*NextLevelRD)) 895 return RK_None; 896 } 897 898 // Shuffle mask for pairwise operation must match. 899 if (matchPairwiseShuffleMask(LS, /*IsLeft=*/true, Level)) { 900 if (!matchPairwiseShuffleMask(RS, /*IsLeft=*/false, Level)) 901 return RK_None; 902 } else if (matchPairwiseShuffleMask(RS, /*IsLeft=*/true, Level)) { 903 if (!matchPairwiseShuffleMask(LS, /*IsLeft=*/false, Level)) 904 return RK_None; 905 } else { 906 return RK_None; 907 } 908 909 if (++Level == NumLevels) 910 return RD->Kind; 911 912 // Match next level. 913 return matchPairwiseReductionAtLevel(cast<Instruction>(NextLevelOp), Level, 914 NumLevels); 915 } 916 917 static ReductionKind matchPairwiseReduction(const ExtractElementInst *ReduxRoot, 918 unsigned &Opcode, Type *&Ty) { 919 if (!EnableReduxCost) 920 return RK_None; 921 922 // Need to extract the first element. 923 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 924 unsigned Idx = ~0u; 925 if (CI) 926 Idx = CI->getZExtValue(); 927 if (Idx != 0) 928 return RK_None; 929 930 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 931 if (!RdxStart) 932 return RK_None; 933 Optional<ReductionData> RD = getReductionData(RdxStart); 934 if (!RD) 935 return RK_None; 936 937 Type *VecTy = RdxStart->getType(); 938 unsigned NumVecElems = VecTy->getVectorNumElements(); 939 if (!isPowerOf2_32(NumVecElems)) 940 return RK_None; 941 942 // We look for a sequence of shuffle,shuffle,add triples like the following 943 // that builds a pairwise reduction tree. 944 // 945 // (X0, X1, X2, X3) 946 // (X0 + X1, X2 + X3, undef, undef) 947 // ((X0 + X1) + (X2 + X3), undef, undef, undef) 948 // 949 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 950 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 951 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 952 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 953 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 954 // %rdx.shuf.1.0 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 955 // <4 x i32> <i32 0, i32 undef, i32 undef, i32 undef> 956 // %rdx.shuf.1.1 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 957 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 958 // %bin.rdx8 = fadd <4 x float> %rdx.shuf.1.0, %rdx.shuf.1.1 959 // %r = extractelement <4 x float> %bin.rdx8, i32 0 960 if (matchPairwiseReductionAtLevel(RdxStart, 0, Log2_32(NumVecElems)) == 961 RK_None) 962 return RK_None; 963 964 Opcode = RD->Opcode; 965 Ty = VecTy; 966 967 return RD->Kind; 968 } 969 970 static std::pair<Value *, ShuffleVectorInst *> 971 getShuffleAndOtherOprd(Value *L, Value *R) { 972 ShuffleVectorInst *S = nullptr; 973 974 if ((S = dyn_cast<ShuffleVectorInst>(L))) 975 return std::make_pair(R, S); 976 977 S = dyn_cast<ShuffleVectorInst>(R); 978 return std::make_pair(L, S); 979 } 980 981 static ReductionKind 982 matchVectorSplittingReduction(const ExtractElementInst *ReduxRoot, 983 unsigned &Opcode, Type *&Ty) { 984 if (!EnableReduxCost) 985 return RK_None; 986 987 // Need to extract the first element. 988 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 989 unsigned Idx = ~0u; 990 if (CI) 991 Idx = CI->getZExtValue(); 992 if (Idx != 0) 993 return RK_None; 994 995 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 996 if (!RdxStart) 997 return RK_None; 998 Optional<ReductionData> RD = getReductionData(RdxStart); 999 if (!RD) 1000 return RK_None; 1001 1002 Type *VecTy = ReduxRoot->getOperand(0)->getType(); 1003 unsigned NumVecElems = VecTy->getVectorNumElements(); 1004 if (!isPowerOf2_32(NumVecElems)) 1005 return RK_None; 1006 1007 // We look for a sequence of shuffles and adds like the following matching one 1008 // fadd, shuffle vector pair at a time. 1009 // 1010 // %rdx.shuf = shufflevector <4 x float> %rdx, <4 x float> undef, 1011 // <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 1012 // %bin.rdx = fadd <4 x float> %rdx, %rdx.shuf 1013 // %rdx.shuf7 = shufflevector <4 x float> %bin.rdx, <4 x float> undef, 1014 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 1015 // %bin.rdx8 = fadd <4 x float> %bin.rdx, %rdx.shuf7 1016 // %r = extractelement <4 x float> %bin.rdx8, i32 0 1017 1018 unsigned MaskStart = 1; 1019 Instruction *RdxOp = RdxStart; 1020 SmallVector<int, 32> ShuffleMask(NumVecElems, 0); 1021 unsigned NumVecElemsRemain = NumVecElems; 1022 while (NumVecElemsRemain - 1) { 1023 // Check for the right reduction operation. 1024 if (!RdxOp) 1025 return RK_None; 1026 Optional<ReductionData> RDLevel = getReductionData(RdxOp); 1027 if (!RDLevel || !RDLevel->hasSameData(*RD)) 1028 return RK_None; 1029 1030 Value *NextRdxOp; 1031 ShuffleVectorInst *Shuffle; 1032 std::tie(NextRdxOp, Shuffle) = 1033 getShuffleAndOtherOprd(RDLevel->LHS, RDLevel->RHS); 1034 1035 // Check the current reduction operation and the shuffle use the same value. 1036 if (Shuffle == nullptr) 1037 return RK_None; 1038 if (Shuffle->getOperand(0) != NextRdxOp) 1039 return RK_None; 1040 1041 // Check that shuffle masks matches. 1042 for (unsigned j = 0; j != MaskStart; ++j) 1043 ShuffleMask[j] = MaskStart + j; 1044 // Fill the rest of the mask with -1 for undef. 1045 std::fill(&ShuffleMask[MaskStart], ShuffleMask.end(), -1); 1046 1047 SmallVector<int, 16> Mask = Shuffle->getShuffleMask(); 1048 if (ShuffleMask != Mask) 1049 return RK_None; 1050 1051 RdxOp = dyn_cast<Instruction>(NextRdxOp); 1052 NumVecElemsRemain /= 2; 1053 MaskStart *= 2; 1054 } 1055 1056 Opcode = RD->Opcode; 1057 Ty = VecTy; 1058 return RD->Kind; 1059 } 1060 1061 int TargetTransformInfo::getInstructionThroughput(const Instruction *I) const { 1062 switch (I->getOpcode()) { 1063 case Instruction::GetElementPtr: 1064 return getUserCost(I); 1065 1066 case Instruction::Ret: 1067 case Instruction::PHI: 1068 case Instruction::Br: { 1069 return getCFInstrCost(I->getOpcode()); 1070 } 1071 case Instruction::Add: 1072 case Instruction::FAdd: 1073 case Instruction::Sub: 1074 case Instruction::FSub: 1075 case Instruction::Mul: 1076 case Instruction::FMul: 1077 case Instruction::UDiv: 1078 case Instruction::SDiv: 1079 case Instruction::FDiv: 1080 case Instruction::URem: 1081 case Instruction::SRem: 1082 case Instruction::FRem: 1083 case Instruction::Shl: 1084 case Instruction::LShr: 1085 case Instruction::AShr: 1086 case Instruction::And: 1087 case Instruction::Or: 1088 case Instruction::Xor: { 1089 TargetTransformInfo::OperandValueKind Op1VK = 1090 getOperandInfo(I->getOperand(0)); 1091 TargetTransformInfo::OperandValueKind Op2VK = 1092 getOperandInfo(I->getOperand(1)); 1093 SmallVector<const Value*, 2> Operands(I->operand_values()); 1094 return getArithmeticInstrCost(I->getOpcode(), I->getType(), Op1VK, 1095 Op2VK, TargetTransformInfo::OP_None, 1096 TargetTransformInfo::OP_None, 1097 Operands); 1098 } 1099 case Instruction::Select: { 1100 const SelectInst *SI = cast<SelectInst>(I); 1101 Type *CondTy = SI->getCondition()->getType(); 1102 return getCmpSelInstrCost(I->getOpcode(), I->getType(), CondTy, I); 1103 } 1104 case Instruction::ICmp: 1105 case Instruction::FCmp: { 1106 Type *ValTy = I->getOperand(0)->getType(); 1107 return getCmpSelInstrCost(I->getOpcode(), ValTy, I->getType(), I); 1108 } 1109 case Instruction::Store: { 1110 const StoreInst *SI = cast<StoreInst>(I); 1111 Type *ValTy = SI->getValueOperand()->getType(); 1112 return getMemoryOpCost(I->getOpcode(), ValTy, 1113 SI->getAlignment(), 1114 SI->getPointerAddressSpace(), I); 1115 } 1116 case Instruction::Load: { 1117 const LoadInst *LI = cast<LoadInst>(I); 1118 return getMemoryOpCost(I->getOpcode(), I->getType(), 1119 LI->getAlignment(), 1120 LI->getPointerAddressSpace(), I); 1121 } 1122 case Instruction::ZExt: 1123 case Instruction::SExt: 1124 case Instruction::FPToUI: 1125 case Instruction::FPToSI: 1126 case Instruction::FPExt: 1127 case Instruction::PtrToInt: 1128 case Instruction::IntToPtr: 1129 case Instruction::SIToFP: 1130 case Instruction::UIToFP: 1131 case Instruction::Trunc: 1132 case Instruction::FPTrunc: 1133 case Instruction::BitCast: 1134 case Instruction::AddrSpaceCast: { 1135 Type *SrcTy = I->getOperand(0)->getType(); 1136 return getCastInstrCost(I->getOpcode(), I->getType(), SrcTy, I); 1137 } 1138 case Instruction::ExtractElement: { 1139 const ExtractElementInst * EEI = cast<ExtractElementInst>(I); 1140 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1141 unsigned Idx = -1; 1142 if (CI) 1143 Idx = CI->getZExtValue(); 1144 1145 // Try to match a reduction sequence (series of shufflevector and vector 1146 // adds followed by a extractelement). 1147 unsigned ReduxOpCode; 1148 Type *ReduxType; 1149 1150 switch (matchVectorSplittingReduction(EEI, ReduxOpCode, ReduxType)) { 1151 case RK_Arithmetic: 1152 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1153 /*IsPairwiseForm=*/false); 1154 case RK_MinMax: 1155 return getMinMaxReductionCost( 1156 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1157 /*IsPairwiseForm=*/false, /*IsUnsigned=*/false); 1158 case RK_UnsignedMinMax: 1159 return getMinMaxReductionCost( 1160 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1161 /*IsPairwiseForm=*/false, /*IsUnsigned=*/true); 1162 case RK_None: 1163 break; 1164 } 1165 1166 switch (matchPairwiseReduction(EEI, ReduxOpCode, ReduxType)) { 1167 case RK_Arithmetic: 1168 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1169 /*IsPairwiseForm=*/true); 1170 case RK_MinMax: 1171 return getMinMaxReductionCost( 1172 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1173 /*IsPairwiseForm=*/true, /*IsUnsigned=*/false); 1174 case RK_UnsignedMinMax: 1175 return getMinMaxReductionCost( 1176 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1177 /*IsPairwiseForm=*/true, /*IsUnsigned=*/true); 1178 case RK_None: 1179 break; 1180 } 1181 1182 return getVectorInstrCost(I->getOpcode(), 1183 EEI->getOperand(0)->getType(), Idx); 1184 } 1185 case Instruction::InsertElement: { 1186 const InsertElementInst * IE = cast<InsertElementInst>(I); 1187 ConstantInt *CI = dyn_cast<ConstantInt>(IE->getOperand(2)); 1188 unsigned Idx = -1; 1189 if (CI) 1190 Idx = CI->getZExtValue(); 1191 return getVectorInstrCost(I->getOpcode(), 1192 IE->getType(), Idx); 1193 } 1194 case Instruction::ShuffleVector: { 1195 const ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I); 1196 Type *VecTypOp0 = Shuffle->getOperand(0)->getType(); 1197 unsigned NumVecElems = VecTypOp0->getVectorNumElements(); 1198 SmallVector<int, 16> Mask = Shuffle->getShuffleMask(); 1199 1200 if (NumVecElems == Mask.size()) { 1201 if (isReverseVectorMask(Mask)) 1202 return TTIImpl->getShuffleCost(TargetTransformInfo::SK_Reverse, 1203 VecTypOp0, 0, nullptr); 1204 if (isAlternateVectorMask(Mask)) 1205 return TTIImpl->getShuffleCost(TargetTransformInfo::SK_Alternate, 1206 VecTypOp0, 0, nullptr); 1207 1208 if (isTransposeVectorMask(Mask)) 1209 return TTIImpl->getShuffleCost(TargetTransformInfo::SK_Transpose, 1210 VecTypOp0, 0, nullptr); 1211 1212 if (isZeroEltBroadcastVectorMask(Mask)) 1213 return TTIImpl->getShuffleCost(TargetTransformInfo::SK_Broadcast, 1214 VecTypOp0, 0, nullptr); 1215 1216 if (isSingleSourceVectorMask(Mask)) 1217 return TTIImpl->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, 1218 VecTypOp0, 0, nullptr); 1219 1220 return TTIImpl->getShuffleCost(TargetTransformInfo::SK_PermuteTwoSrc, 1221 VecTypOp0, 0, nullptr); 1222 } 1223 1224 return -1; 1225 } 1226 case Instruction::Call: 1227 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1228 SmallVector<Value *, 4> Args(II->arg_operands()); 1229 1230 FastMathFlags FMF; 1231 if (auto *FPMO = dyn_cast<FPMathOperator>(II)) 1232 FMF = FPMO->getFastMathFlags(); 1233 1234 return getIntrinsicInstrCost(II->getIntrinsicID(), II->getType(), 1235 Args, FMF); 1236 } 1237 return -1; 1238 default: 1239 // We don't have any information on this instruction. 1240 return -1; 1241 } 1242 } 1243 1244 TargetTransformInfo::Concept::~Concept() {} 1245 1246 TargetIRAnalysis::TargetIRAnalysis() : TTICallback(&getDefaultTTI) {} 1247 1248 TargetIRAnalysis::TargetIRAnalysis( 1249 std::function<Result(const Function &)> TTICallback) 1250 : TTICallback(std::move(TTICallback)) {} 1251 1252 TargetIRAnalysis::Result TargetIRAnalysis::run(const Function &F, 1253 FunctionAnalysisManager &) { 1254 return TTICallback(F); 1255 } 1256 1257 AnalysisKey TargetIRAnalysis::Key; 1258 1259 TargetIRAnalysis::Result TargetIRAnalysis::getDefaultTTI(const Function &F) { 1260 return Result(F.getParent()->getDataLayout()); 1261 } 1262 1263 // Register the basic pass. 1264 INITIALIZE_PASS(TargetTransformInfoWrapperPass, "tti", 1265 "Target Transform Information", false, true) 1266 char TargetTransformInfoWrapperPass::ID = 0; 1267 1268 void TargetTransformInfoWrapperPass::anchor() {} 1269 1270 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass() 1271 : ImmutablePass(ID) { 1272 initializeTargetTransformInfoWrapperPassPass( 1273 *PassRegistry::getPassRegistry()); 1274 } 1275 1276 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass( 1277 TargetIRAnalysis TIRA) 1278 : ImmutablePass(ID), TIRA(std::move(TIRA)) { 1279 initializeTargetTransformInfoWrapperPassPass( 1280 *PassRegistry::getPassRegistry()); 1281 } 1282 1283 TargetTransformInfo &TargetTransformInfoWrapperPass::getTTI(const Function &F) { 1284 FunctionAnalysisManager DummyFAM; 1285 TTI = TIRA.run(F, DummyFAM); 1286 return *TTI; 1287 } 1288 1289 ImmutablePass * 1290 llvm::createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA) { 1291 return new TargetTransformInfoWrapperPass(std::move(TIRA)); 1292 } 1293