1 //===- llvm/Analysis/TargetTransformInfo.cpp ------------------------------===// 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 #include "llvm/Analysis/TargetTransformInfo.h" 10 #include "llvm/Analysis/TargetTransformInfoImpl.h" 11 #include "llvm/IR/CallSite.h" 12 #include "llvm/IR/DataLayout.h" 13 #include "llvm/IR/Instruction.h" 14 #include "llvm/IR/Instructions.h" 15 #include "llvm/IR/IntrinsicInst.h" 16 #include "llvm/IR/Module.h" 17 #include "llvm/IR/Operator.h" 18 #include "llvm/IR/PatternMatch.h" 19 #include "llvm/Support/CommandLine.h" 20 #include "llvm/Support/ErrorHandling.h" 21 #include <utility> 22 23 using namespace llvm; 24 using namespace PatternMatch; 25 26 #define DEBUG_TYPE "tti" 27 28 static cl::opt<bool> EnableReduxCost("costmodel-reduxcost", cl::init(false), 29 cl::Hidden, 30 cl::desc("Recognize reduction patterns.")); 31 32 namespace { 33 /// No-op implementation of the TTI interface using the utility base 34 /// classes. 35 /// 36 /// This is used when no target specific information is available. 37 struct NoTTIImpl : TargetTransformInfoImplCRTPBase<NoTTIImpl> { 38 explicit NoTTIImpl(const DataLayout &DL) 39 : TargetTransformInfoImplCRTPBase<NoTTIImpl>(DL) {} 40 }; 41 } 42 43 TargetTransformInfo::TargetTransformInfo(const DataLayout &DL) 44 : TTIImpl(new Model<NoTTIImpl>(NoTTIImpl(DL))) {} 45 46 TargetTransformInfo::~TargetTransformInfo() {} 47 48 TargetTransformInfo::TargetTransformInfo(TargetTransformInfo &&Arg) 49 : TTIImpl(std::move(Arg.TTIImpl)) {} 50 51 TargetTransformInfo &TargetTransformInfo::operator=(TargetTransformInfo &&RHS) { 52 TTIImpl = std::move(RHS.TTIImpl); 53 return *this; 54 } 55 56 int TargetTransformInfo::getOperationCost(unsigned Opcode, Type *Ty, 57 Type *OpTy) const { 58 int Cost = TTIImpl->getOperationCost(Opcode, Ty, OpTy); 59 assert(Cost >= 0 && "TTI should not produce negative costs!"); 60 return Cost; 61 } 62 63 int TargetTransformInfo::getCallCost(FunctionType *FTy, int NumArgs) const { 64 int Cost = TTIImpl->getCallCost(FTy, NumArgs); 65 assert(Cost >= 0 && "TTI should not produce negative costs!"); 66 return Cost; 67 } 68 69 int TargetTransformInfo::getCallCost(const Function *F, 70 ArrayRef<const Value *> Arguments) const { 71 int Cost = TTIImpl->getCallCost(F, Arguments); 72 assert(Cost >= 0 && "TTI should not produce negative costs!"); 73 return Cost; 74 } 75 76 unsigned TargetTransformInfo::getInliningThresholdMultiplier() const { 77 return TTIImpl->getInliningThresholdMultiplier(); 78 } 79 80 int TargetTransformInfo::getGEPCost(Type *PointeeType, const Value *Ptr, 81 ArrayRef<const Value *> Operands) const { 82 return TTIImpl->getGEPCost(PointeeType, Ptr, Operands); 83 } 84 85 int TargetTransformInfo::getExtCost(const Instruction *I, 86 const Value *Src) const { 87 return TTIImpl->getExtCost(I, Src); 88 } 89 90 int TargetTransformInfo::getIntrinsicCost( 91 Intrinsic::ID IID, Type *RetTy, ArrayRef<const Value *> Arguments) const { 92 int Cost = TTIImpl->getIntrinsicCost(IID, RetTy, Arguments); 93 assert(Cost >= 0 && "TTI should not produce negative costs!"); 94 return Cost; 95 } 96 97 unsigned 98 TargetTransformInfo::getEstimatedNumberOfCaseClusters(const SwitchInst &SI, 99 unsigned &JTSize) const { 100 return TTIImpl->getEstimatedNumberOfCaseClusters(SI, JTSize); 101 } 102 103 int TargetTransformInfo::getUserCost(const User *U, 104 ArrayRef<const Value *> Operands) const { 105 int Cost = TTIImpl->getUserCost(U, Operands); 106 assert(Cost >= 0 && "TTI should not produce negative costs!"); 107 return Cost; 108 } 109 110 bool TargetTransformInfo::hasBranchDivergence() const { 111 return TTIImpl->hasBranchDivergence(); 112 } 113 114 bool TargetTransformInfo::isSourceOfDivergence(const Value *V) const { 115 return TTIImpl->isSourceOfDivergence(V); 116 } 117 118 bool llvm::TargetTransformInfo::isAlwaysUniform(const Value *V) const { 119 return TTIImpl->isAlwaysUniform(V); 120 } 121 122 unsigned TargetTransformInfo::getFlatAddressSpace() const { 123 return TTIImpl->getFlatAddressSpace(); 124 } 125 126 bool TargetTransformInfo::isLoweredToCall(const Function *F) const { 127 return TTIImpl->isLoweredToCall(F); 128 } 129 130 void TargetTransformInfo::getUnrollingPreferences( 131 Loop *L, ScalarEvolution &SE, UnrollingPreferences &UP) const { 132 return TTIImpl->getUnrollingPreferences(L, SE, UP); 133 } 134 135 bool TargetTransformInfo::isLegalAddImmediate(int64_t Imm) const { 136 return TTIImpl->isLegalAddImmediate(Imm); 137 } 138 139 bool TargetTransformInfo::isLegalICmpImmediate(int64_t Imm) const { 140 return TTIImpl->isLegalICmpImmediate(Imm); 141 } 142 143 bool TargetTransformInfo::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, 144 int64_t BaseOffset, 145 bool HasBaseReg, 146 int64_t Scale, 147 unsigned AddrSpace, 148 Instruction *I) const { 149 return TTIImpl->isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg, 150 Scale, AddrSpace, I); 151 } 152 153 bool TargetTransformInfo::isLSRCostLess(LSRCost &C1, LSRCost &C2) const { 154 return TTIImpl->isLSRCostLess(C1, C2); 155 } 156 157 bool TargetTransformInfo::canMacroFuseCmp() const { 158 return TTIImpl->canMacroFuseCmp(); 159 } 160 161 bool TargetTransformInfo::shouldFavorPostInc() const { 162 return TTIImpl->shouldFavorPostInc(); 163 } 164 165 bool TargetTransformInfo::shouldFavorBackedgeIndex(const Loop *L) const { 166 return TTIImpl->shouldFavorBackedgeIndex(L); 167 } 168 169 bool TargetTransformInfo::isLegalMaskedStore(Type *DataType) const { 170 return TTIImpl->isLegalMaskedStore(DataType); 171 } 172 173 bool TargetTransformInfo::isLegalMaskedLoad(Type *DataType) const { 174 return TTIImpl->isLegalMaskedLoad(DataType); 175 } 176 177 bool TargetTransformInfo::isLegalMaskedGather(Type *DataType) const { 178 return TTIImpl->isLegalMaskedGather(DataType); 179 } 180 181 bool TargetTransformInfo::isLegalMaskedScatter(Type *DataType) const { 182 return TTIImpl->isLegalMaskedScatter(DataType); 183 } 184 185 bool TargetTransformInfo::hasDivRemOp(Type *DataType, bool IsSigned) const { 186 return TTIImpl->hasDivRemOp(DataType, IsSigned); 187 } 188 189 bool TargetTransformInfo::hasVolatileVariant(Instruction *I, 190 unsigned AddrSpace) const { 191 return TTIImpl->hasVolatileVariant(I, AddrSpace); 192 } 193 194 bool TargetTransformInfo::prefersVectorizedAddressing() const { 195 return TTIImpl->prefersVectorizedAddressing(); 196 } 197 198 int TargetTransformInfo::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, 199 int64_t BaseOffset, 200 bool HasBaseReg, 201 int64_t Scale, 202 unsigned AddrSpace) const { 203 int Cost = TTIImpl->getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg, 204 Scale, AddrSpace); 205 assert(Cost >= 0 && "TTI should not produce negative costs!"); 206 return Cost; 207 } 208 209 bool TargetTransformInfo::LSRWithInstrQueries() const { 210 return TTIImpl->LSRWithInstrQueries(); 211 } 212 213 bool TargetTransformInfo::isTruncateFree(Type *Ty1, Type *Ty2) const { 214 return TTIImpl->isTruncateFree(Ty1, Ty2); 215 } 216 217 bool TargetTransformInfo::isProfitableToHoist(Instruction *I) const { 218 return TTIImpl->isProfitableToHoist(I); 219 } 220 221 bool TargetTransformInfo::useAA() const { return TTIImpl->useAA(); } 222 223 bool TargetTransformInfo::isTypeLegal(Type *Ty) const { 224 return TTIImpl->isTypeLegal(Ty); 225 } 226 227 unsigned TargetTransformInfo::getJumpBufAlignment() const { 228 return TTIImpl->getJumpBufAlignment(); 229 } 230 231 unsigned TargetTransformInfo::getJumpBufSize() const { 232 return TTIImpl->getJumpBufSize(); 233 } 234 235 bool TargetTransformInfo::shouldBuildLookupTables() const { 236 return TTIImpl->shouldBuildLookupTables(); 237 } 238 bool TargetTransformInfo::shouldBuildLookupTablesForConstant(Constant *C) const { 239 return TTIImpl->shouldBuildLookupTablesForConstant(C); 240 } 241 242 bool TargetTransformInfo::useColdCCForColdCall(Function &F) const { 243 return TTIImpl->useColdCCForColdCall(F); 244 } 245 246 unsigned TargetTransformInfo:: 247 getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const { 248 return TTIImpl->getScalarizationOverhead(Ty, Insert, Extract); 249 } 250 251 unsigned TargetTransformInfo:: 252 getOperandsScalarizationOverhead(ArrayRef<const Value *> Args, 253 unsigned VF) const { 254 return TTIImpl->getOperandsScalarizationOverhead(Args, VF); 255 } 256 257 bool TargetTransformInfo::supportsEfficientVectorElementLoadStore() const { 258 return TTIImpl->supportsEfficientVectorElementLoadStore(); 259 } 260 261 bool TargetTransformInfo::enableAggressiveInterleaving(bool LoopHasReductions) const { 262 return TTIImpl->enableAggressiveInterleaving(LoopHasReductions); 263 } 264 265 const TargetTransformInfo::MemCmpExpansionOptions * 266 TargetTransformInfo::enableMemCmpExpansion(bool IsZeroCmp) const { 267 return TTIImpl->enableMemCmpExpansion(IsZeroCmp); 268 } 269 270 bool TargetTransformInfo::enableInterleavedAccessVectorization() const { 271 return TTIImpl->enableInterleavedAccessVectorization(); 272 } 273 274 bool TargetTransformInfo::enableMaskedInterleavedAccessVectorization() const { 275 return TTIImpl->enableMaskedInterleavedAccessVectorization(); 276 } 277 278 bool TargetTransformInfo::isFPVectorizationPotentiallyUnsafe() const { 279 return TTIImpl->isFPVectorizationPotentiallyUnsafe(); 280 } 281 282 bool TargetTransformInfo::allowsMisalignedMemoryAccesses(LLVMContext &Context, 283 unsigned BitWidth, 284 unsigned AddressSpace, 285 unsigned Alignment, 286 bool *Fast) const { 287 return TTIImpl->allowsMisalignedMemoryAccesses(Context, BitWidth, AddressSpace, 288 Alignment, Fast); 289 } 290 291 TargetTransformInfo::PopcntSupportKind 292 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit) const { 293 return TTIImpl->getPopcntSupport(IntTyWidthInBit); 294 } 295 296 bool TargetTransformInfo::haveFastSqrt(Type *Ty) const { 297 return TTIImpl->haveFastSqrt(Ty); 298 } 299 300 bool TargetTransformInfo::isFCmpOrdCheaperThanFCmpZero(Type *Ty) const { 301 return TTIImpl->isFCmpOrdCheaperThanFCmpZero(Ty); 302 } 303 304 int TargetTransformInfo::getFPOpCost(Type *Ty) const { 305 int Cost = TTIImpl->getFPOpCost(Ty); 306 assert(Cost >= 0 && "TTI should not produce negative costs!"); 307 return Cost; 308 } 309 310 int TargetTransformInfo::getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, 311 const APInt &Imm, 312 Type *Ty) const { 313 int Cost = TTIImpl->getIntImmCodeSizeCost(Opcode, Idx, Imm, Ty); 314 assert(Cost >= 0 && "TTI should not produce negative costs!"); 315 return Cost; 316 } 317 318 int TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty) const { 319 int Cost = TTIImpl->getIntImmCost(Imm, Ty); 320 assert(Cost >= 0 && "TTI should not produce negative costs!"); 321 return Cost; 322 } 323 324 int TargetTransformInfo::getIntImmCost(unsigned Opcode, unsigned Idx, 325 const APInt &Imm, Type *Ty) const { 326 int Cost = TTIImpl->getIntImmCost(Opcode, Idx, Imm, Ty); 327 assert(Cost >= 0 && "TTI should not produce negative costs!"); 328 return Cost; 329 } 330 331 int TargetTransformInfo::getIntImmCost(Intrinsic::ID IID, unsigned Idx, 332 const APInt &Imm, Type *Ty) const { 333 int Cost = TTIImpl->getIntImmCost(IID, Idx, Imm, Ty); 334 assert(Cost >= 0 && "TTI should not produce negative costs!"); 335 return Cost; 336 } 337 338 unsigned TargetTransformInfo::getNumberOfRegisters(bool Vector) const { 339 return TTIImpl->getNumberOfRegisters(Vector); 340 } 341 342 unsigned TargetTransformInfo::getRegisterBitWidth(bool Vector) const { 343 return TTIImpl->getRegisterBitWidth(Vector); 344 } 345 346 unsigned TargetTransformInfo::getMinVectorRegisterBitWidth() const { 347 return TTIImpl->getMinVectorRegisterBitWidth(); 348 } 349 350 bool TargetTransformInfo::shouldMaximizeVectorBandwidth(bool OptSize) const { 351 return TTIImpl->shouldMaximizeVectorBandwidth(OptSize); 352 } 353 354 unsigned TargetTransformInfo::getMinimumVF(unsigned ElemWidth) const { 355 return TTIImpl->getMinimumVF(ElemWidth); 356 } 357 358 bool TargetTransformInfo::shouldConsiderAddressTypePromotion( 359 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const { 360 return TTIImpl->shouldConsiderAddressTypePromotion( 361 I, AllowPromotionWithoutCommonHeader); 362 } 363 364 unsigned TargetTransformInfo::getCacheLineSize() const { 365 return TTIImpl->getCacheLineSize(); 366 } 367 368 llvm::Optional<unsigned> TargetTransformInfo::getCacheSize(CacheLevel Level) 369 const { 370 return TTIImpl->getCacheSize(Level); 371 } 372 373 llvm::Optional<unsigned> TargetTransformInfo::getCacheAssociativity( 374 CacheLevel Level) const { 375 return TTIImpl->getCacheAssociativity(Level); 376 } 377 378 unsigned TargetTransformInfo::getPrefetchDistance() const { 379 return TTIImpl->getPrefetchDistance(); 380 } 381 382 unsigned TargetTransformInfo::getMinPrefetchStride() const { 383 return TTIImpl->getMinPrefetchStride(); 384 } 385 386 unsigned TargetTransformInfo::getMaxPrefetchIterationsAhead() const { 387 return TTIImpl->getMaxPrefetchIterationsAhead(); 388 } 389 390 unsigned TargetTransformInfo::getMaxInterleaveFactor(unsigned VF) const { 391 return TTIImpl->getMaxInterleaveFactor(VF); 392 } 393 394 TargetTransformInfo::OperandValueKind 395 TargetTransformInfo::getOperandInfo(Value *V, OperandValueProperties &OpProps) { 396 OperandValueKind OpInfo = OK_AnyValue; 397 OpProps = OP_None; 398 399 if (auto *CI = dyn_cast<ConstantInt>(V)) { 400 if (CI->getValue().isPowerOf2()) 401 OpProps = OP_PowerOf2; 402 return OK_UniformConstantValue; 403 } 404 405 // A broadcast shuffle creates a uniform value. 406 // TODO: Add support for non-zero index broadcasts. 407 // TODO: Add support for different source vector width. 408 if (auto *ShuffleInst = dyn_cast<ShuffleVectorInst>(V)) 409 if (ShuffleInst->isZeroEltSplat()) 410 OpInfo = OK_UniformValue; 411 412 const Value *Splat = getSplatValue(V); 413 414 // Check for a splat of a constant or for a non uniform vector of constants 415 // and check if the constant(s) are all powers of two. 416 if (isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) { 417 OpInfo = OK_NonUniformConstantValue; 418 if (Splat) { 419 OpInfo = OK_UniformConstantValue; 420 if (auto *CI = dyn_cast<ConstantInt>(Splat)) 421 if (CI->getValue().isPowerOf2()) 422 OpProps = OP_PowerOf2; 423 } else if (auto *CDS = dyn_cast<ConstantDataSequential>(V)) { 424 OpProps = OP_PowerOf2; 425 for (unsigned I = 0, E = CDS->getNumElements(); I != E; ++I) { 426 if (auto *CI = dyn_cast<ConstantInt>(CDS->getElementAsConstant(I))) 427 if (CI->getValue().isPowerOf2()) 428 continue; 429 OpProps = OP_None; 430 break; 431 } 432 } 433 } 434 435 // Check for a splat of a uniform value. This is not loop aware, so return 436 // true only for the obviously uniform cases (argument, globalvalue) 437 if (Splat && (isa<Argument>(Splat) || isa<GlobalValue>(Splat))) 438 OpInfo = OK_UniformValue; 439 440 return OpInfo; 441 } 442 443 int TargetTransformInfo::getArithmeticInstrCost( 444 unsigned Opcode, Type *Ty, OperandValueKind Opd1Info, 445 OperandValueKind Opd2Info, OperandValueProperties Opd1PropInfo, 446 OperandValueProperties Opd2PropInfo, 447 ArrayRef<const Value *> Args) const { 448 int Cost = TTIImpl->getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info, 449 Opd1PropInfo, Opd2PropInfo, Args); 450 assert(Cost >= 0 && "TTI should not produce negative costs!"); 451 return Cost; 452 } 453 454 int TargetTransformInfo::getShuffleCost(ShuffleKind Kind, Type *Ty, int Index, 455 Type *SubTp) const { 456 int Cost = TTIImpl->getShuffleCost(Kind, Ty, Index, SubTp); 457 assert(Cost >= 0 && "TTI should not produce negative costs!"); 458 return Cost; 459 } 460 461 int TargetTransformInfo::getCastInstrCost(unsigned Opcode, Type *Dst, 462 Type *Src, const Instruction *I) const { 463 assert ((I == nullptr || I->getOpcode() == Opcode) && 464 "Opcode should reflect passed instruction."); 465 int Cost = TTIImpl->getCastInstrCost(Opcode, Dst, Src, I); 466 assert(Cost >= 0 && "TTI should not produce negative costs!"); 467 return Cost; 468 } 469 470 int TargetTransformInfo::getExtractWithExtendCost(unsigned Opcode, Type *Dst, 471 VectorType *VecTy, 472 unsigned Index) const { 473 int Cost = TTIImpl->getExtractWithExtendCost(Opcode, Dst, VecTy, Index); 474 assert(Cost >= 0 && "TTI should not produce negative costs!"); 475 return Cost; 476 } 477 478 int TargetTransformInfo::getCFInstrCost(unsigned Opcode) const { 479 int Cost = TTIImpl->getCFInstrCost(Opcode); 480 assert(Cost >= 0 && "TTI should not produce negative costs!"); 481 return Cost; 482 } 483 484 int TargetTransformInfo::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 485 Type *CondTy, const Instruction *I) const { 486 assert ((I == nullptr || I->getOpcode() == Opcode) && 487 "Opcode should reflect passed instruction."); 488 int Cost = TTIImpl->getCmpSelInstrCost(Opcode, ValTy, CondTy, I); 489 assert(Cost >= 0 && "TTI should not produce negative costs!"); 490 return Cost; 491 } 492 493 int TargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val, 494 unsigned Index) const { 495 int Cost = TTIImpl->getVectorInstrCost(Opcode, Val, Index); 496 assert(Cost >= 0 && "TTI should not produce negative costs!"); 497 return Cost; 498 } 499 500 int TargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src, 501 unsigned Alignment, 502 unsigned AddressSpace, 503 const Instruction *I) const { 504 assert ((I == nullptr || I->getOpcode() == Opcode) && 505 "Opcode should reflect passed instruction."); 506 int Cost = TTIImpl->getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, I); 507 assert(Cost >= 0 && "TTI should not produce negative costs!"); 508 return Cost; 509 } 510 511 int TargetTransformInfo::getMaskedMemoryOpCost(unsigned Opcode, Type *Src, 512 unsigned Alignment, 513 unsigned AddressSpace) const { 514 int Cost = 515 TTIImpl->getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace); 516 assert(Cost >= 0 && "TTI should not produce negative costs!"); 517 return Cost; 518 } 519 520 int TargetTransformInfo::getGatherScatterOpCost(unsigned Opcode, Type *DataTy, 521 Value *Ptr, bool VariableMask, 522 unsigned Alignment) const { 523 int Cost = TTIImpl->getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask, 524 Alignment); 525 assert(Cost >= 0 && "TTI should not produce negative costs!"); 526 return Cost; 527 } 528 529 int TargetTransformInfo::getInterleavedMemoryOpCost( 530 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices, 531 unsigned Alignment, unsigned AddressSpace, bool UseMaskForCond, 532 bool UseMaskForGaps) const { 533 int Cost = TTIImpl->getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 534 Alignment, AddressSpace, 535 UseMaskForCond, 536 UseMaskForGaps); 537 assert(Cost >= 0 && "TTI should not produce negative costs!"); 538 return Cost; 539 } 540 541 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 542 ArrayRef<Type *> Tys, FastMathFlags FMF, 543 unsigned ScalarizationCostPassed) const { 544 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Tys, FMF, 545 ScalarizationCostPassed); 546 assert(Cost >= 0 && "TTI should not produce negative costs!"); 547 return Cost; 548 } 549 550 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 551 ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) const { 552 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF); 553 assert(Cost >= 0 && "TTI should not produce negative costs!"); 554 return Cost; 555 } 556 557 int TargetTransformInfo::getCallInstrCost(Function *F, Type *RetTy, 558 ArrayRef<Type *> Tys) const { 559 int Cost = TTIImpl->getCallInstrCost(F, RetTy, Tys); 560 assert(Cost >= 0 && "TTI should not produce negative costs!"); 561 return Cost; 562 } 563 564 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const { 565 return TTIImpl->getNumberOfParts(Tp); 566 } 567 568 int TargetTransformInfo::getAddressComputationCost(Type *Tp, 569 ScalarEvolution *SE, 570 const SCEV *Ptr) const { 571 int Cost = TTIImpl->getAddressComputationCost(Tp, SE, Ptr); 572 assert(Cost >= 0 && "TTI should not produce negative costs!"); 573 return Cost; 574 } 575 576 int TargetTransformInfo::getArithmeticReductionCost(unsigned Opcode, Type *Ty, 577 bool IsPairwiseForm) const { 578 int Cost = TTIImpl->getArithmeticReductionCost(Opcode, Ty, IsPairwiseForm); 579 assert(Cost >= 0 && "TTI should not produce negative costs!"); 580 return Cost; 581 } 582 583 int TargetTransformInfo::getMinMaxReductionCost(Type *Ty, Type *CondTy, 584 bool IsPairwiseForm, 585 bool IsUnsigned) const { 586 int Cost = 587 TTIImpl->getMinMaxReductionCost(Ty, CondTy, IsPairwiseForm, IsUnsigned); 588 assert(Cost >= 0 && "TTI should not produce negative costs!"); 589 return Cost; 590 } 591 592 unsigned 593 TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const { 594 return TTIImpl->getCostOfKeepingLiveOverCall(Tys); 595 } 596 597 bool TargetTransformInfo::getTgtMemIntrinsic(IntrinsicInst *Inst, 598 MemIntrinsicInfo &Info) const { 599 return TTIImpl->getTgtMemIntrinsic(Inst, Info); 600 } 601 602 unsigned TargetTransformInfo::getAtomicMemIntrinsicMaxElementSize() const { 603 return TTIImpl->getAtomicMemIntrinsicMaxElementSize(); 604 } 605 606 Value *TargetTransformInfo::getOrCreateResultFromMemIntrinsic( 607 IntrinsicInst *Inst, Type *ExpectedType) const { 608 return TTIImpl->getOrCreateResultFromMemIntrinsic(Inst, ExpectedType); 609 } 610 611 Type *TargetTransformInfo::getMemcpyLoopLoweringType(LLVMContext &Context, 612 Value *Length, 613 unsigned SrcAlign, 614 unsigned DestAlign) const { 615 return TTIImpl->getMemcpyLoopLoweringType(Context, Length, SrcAlign, 616 DestAlign); 617 } 618 619 void TargetTransformInfo::getMemcpyLoopResidualLoweringType( 620 SmallVectorImpl<Type *> &OpsOut, LLVMContext &Context, 621 unsigned RemainingBytes, unsigned SrcAlign, unsigned DestAlign) const { 622 TTIImpl->getMemcpyLoopResidualLoweringType(OpsOut, Context, RemainingBytes, 623 SrcAlign, DestAlign); 624 } 625 626 bool TargetTransformInfo::areInlineCompatible(const Function *Caller, 627 const Function *Callee) const { 628 return TTIImpl->areInlineCompatible(Caller, Callee); 629 } 630 631 bool TargetTransformInfo::areFunctionArgsABICompatible( 632 const Function *Caller, const Function *Callee, 633 SmallPtrSetImpl<Argument *> &Args) const { 634 return TTIImpl->areFunctionArgsABICompatible(Caller, Callee, Args); 635 } 636 637 bool TargetTransformInfo::isIndexedLoadLegal(MemIndexedMode Mode, 638 Type *Ty) const { 639 return TTIImpl->isIndexedLoadLegal(Mode, Ty); 640 } 641 642 bool TargetTransformInfo::isIndexedStoreLegal(MemIndexedMode Mode, 643 Type *Ty) const { 644 return TTIImpl->isIndexedStoreLegal(Mode, Ty); 645 } 646 647 unsigned TargetTransformInfo::getLoadStoreVecRegBitWidth(unsigned AS) const { 648 return TTIImpl->getLoadStoreVecRegBitWidth(AS); 649 } 650 651 bool TargetTransformInfo::isLegalToVectorizeLoad(LoadInst *LI) const { 652 return TTIImpl->isLegalToVectorizeLoad(LI); 653 } 654 655 bool TargetTransformInfo::isLegalToVectorizeStore(StoreInst *SI) const { 656 return TTIImpl->isLegalToVectorizeStore(SI); 657 } 658 659 bool TargetTransformInfo::isLegalToVectorizeLoadChain( 660 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 661 return TTIImpl->isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment, 662 AddrSpace); 663 } 664 665 bool TargetTransformInfo::isLegalToVectorizeStoreChain( 666 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 667 return TTIImpl->isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment, 668 AddrSpace); 669 } 670 671 unsigned TargetTransformInfo::getLoadVectorFactor(unsigned VF, 672 unsigned LoadSize, 673 unsigned ChainSizeInBytes, 674 VectorType *VecTy) const { 675 return TTIImpl->getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy); 676 } 677 678 unsigned TargetTransformInfo::getStoreVectorFactor(unsigned VF, 679 unsigned StoreSize, 680 unsigned ChainSizeInBytes, 681 VectorType *VecTy) const { 682 return TTIImpl->getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy); 683 } 684 685 bool TargetTransformInfo::useReductionIntrinsic(unsigned Opcode, 686 Type *Ty, ReductionFlags Flags) const { 687 return TTIImpl->useReductionIntrinsic(Opcode, Ty, Flags); 688 } 689 690 bool TargetTransformInfo::shouldExpandReduction(const IntrinsicInst *II) const { 691 return TTIImpl->shouldExpandReduction(II); 692 } 693 694 int TargetTransformInfo::getInstructionLatency(const Instruction *I) const { 695 return TTIImpl->getInstructionLatency(I); 696 } 697 698 static bool matchPairwiseShuffleMask(ShuffleVectorInst *SI, bool IsLeft, 699 unsigned Level) { 700 // We don't need a shuffle if we just want to have element 0 in position 0 of 701 // the vector. 702 if (!SI && Level == 0 && IsLeft) 703 return true; 704 else if (!SI) 705 return false; 706 707 SmallVector<int, 32> Mask(SI->getType()->getVectorNumElements(), -1); 708 709 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right) depending on whether 710 // we look at the left or right side. 711 for (unsigned i = 0, e = (1 << Level), val = !IsLeft; i != e; ++i, val += 2) 712 Mask[i] = val; 713 714 SmallVector<int, 16> ActualMask = SI->getShuffleMask(); 715 return Mask == ActualMask; 716 } 717 718 namespace { 719 /// Kind of the reduction data. 720 enum ReductionKind { 721 RK_None, /// Not a reduction. 722 RK_Arithmetic, /// Binary reduction data. 723 RK_MinMax, /// Min/max reduction data. 724 RK_UnsignedMinMax, /// Unsigned min/max reduction data. 725 }; 726 /// Contains opcode + LHS/RHS parts of the reduction operations. 727 struct ReductionData { 728 ReductionData() = delete; 729 ReductionData(ReductionKind Kind, unsigned Opcode, Value *LHS, Value *RHS) 730 : Opcode(Opcode), LHS(LHS), RHS(RHS), Kind(Kind) { 731 assert(Kind != RK_None && "expected binary or min/max reduction only."); 732 } 733 unsigned Opcode = 0; 734 Value *LHS = nullptr; 735 Value *RHS = nullptr; 736 ReductionKind Kind = RK_None; 737 bool hasSameData(ReductionData &RD) const { 738 return Kind == RD.Kind && Opcode == RD.Opcode; 739 } 740 }; 741 } // namespace 742 743 static Optional<ReductionData> getReductionData(Instruction *I) { 744 Value *L, *R; 745 if (m_BinOp(m_Value(L), m_Value(R)).match(I)) 746 return ReductionData(RK_Arithmetic, I->getOpcode(), L, R); 747 if (auto *SI = dyn_cast<SelectInst>(I)) { 748 if (m_SMin(m_Value(L), m_Value(R)).match(SI) || 749 m_SMax(m_Value(L), m_Value(R)).match(SI) || 750 m_OrdFMin(m_Value(L), m_Value(R)).match(SI) || 751 m_OrdFMax(m_Value(L), m_Value(R)).match(SI) || 752 m_UnordFMin(m_Value(L), m_Value(R)).match(SI) || 753 m_UnordFMax(m_Value(L), m_Value(R)).match(SI)) { 754 auto *CI = cast<CmpInst>(SI->getCondition()); 755 return ReductionData(RK_MinMax, CI->getOpcode(), L, R); 756 } 757 if (m_UMin(m_Value(L), m_Value(R)).match(SI) || 758 m_UMax(m_Value(L), m_Value(R)).match(SI)) { 759 auto *CI = cast<CmpInst>(SI->getCondition()); 760 return ReductionData(RK_UnsignedMinMax, CI->getOpcode(), L, R); 761 } 762 } 763 return llvm::None; 764 } 765 766 static ReductionKind matchPairwiseReductionAtLevel(Instruction *I, 767 unsigned Level, 768 unsigned NumLevels) { 769 // Match one level of pairwise operations. 770 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 771 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 772 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 773 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 774 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 775 if (!I) 776 return RK_None; 777 778 assert(I->getType()->isVectorTy() && "Expecting a vector type"); 779 780 Optional<ReductionData> RD = getReductionData(I); 781 if (!RD) 782 return RK_None; 783 784 ShuffleVectorInst *LS = dyn_cast<ShuffleVectorInst>(RD->LHS); 785 if (!LS && Level) 786 return RK_None; 787 ShuffleVectorInst *RS = dyn_cast<ShuffleVectorInst>(RD->RHS); 788 if (!RS && Level) 789 return RK_None; 790 791 // On level 0 we can omit one shufflevector instruction. 792 if (!Level && !RS && !LS) 793 return RK_None; 794 795 // Shuffle inputs must match. 796 Value *NextLevelOpL = LS ? LS->getOperand(0) : nullptr; 797 Value *NextLevelOpR = RS ? RS->getOperand(0) : nullptr; 798 Value *NextLevelOp = nullptr; 799 if (NextLevelOpR && NextLevelOpL) { 800 // If we have two shuffles their operands must match. 801 if (NextLevelOpL != NextLevelOpR) 802 return RK_None; 803 804 NextLevelOp = NextLevelOpL; 805 } else if (Level == 0 && (NextLevelOpR || NextLevelOpL)) { 806 // On the first level we can omit the shufflevector <0, undef,...>. So the 807 // input to the other shufflevector <1, undef> must match with one of the 808 // inputs to the current binary operation. 809 // Example: 810 // %NextLevelOpL = shufflevector %R, <1, undef ...> 811 // %BinOp = fadd %NextLevelOpL, %R 812 if (NextLevelOpL && NextLevelOpL != RD->RHS) 813 return RK_None; 814 else if (NextLevelOpR && NextLevelOpR != RD->LHS) 815 return RK_None; 816 817 NextLevelOp = NextLevelOpL ? RD->RHS : RD->LHS; 818 } else 819 return RK_None; 820 821 // Check that the next levels binary operation exists and matches with the 822 // current one. 823 if (Level + 1 != NumLevels) { 824 Optional<ReductionData> NextLevelRD = 825 getReductionData(cast<Instruction>(NextLevelOp)); 826 if (!NextLevelRD || !RD->hasSameData(*NextLevelRD)) 827 return RK_None; 828 } 829 830 // Shuffle mask for pairwise operation must match. 831 if (matchPairwiseShuffleMask(LS, /*IsLeft=*/true, Level)) { 832 if (!matchPairwiseShuffleMask(RS, /*IsLeft=*/false, Level)) 833 return RK_None; 834 } else if (matchPairwiseShuffleMask(RS, /*IsLeft=*/true, Level)) { 835 if (!matchPairwiseShuffleMask(LS, /*IsLeft=*/false, Level)) 836 return RK_None; 837 } else { 838 return RK_None; 839 } 840 841 if (++Level == NumLevels) 842 return RD->Kind; 843 844 // Match next level. 845 return matchPairwiseReductionAtLevel(cast<Instruction>(NextLevelOp), Level, 846 NumLevels); 847 } 848 849 static ReductionKind matchPairwiseReduction(const ExtractElementInst *ReduxRoot, 850 unsigned &Opcode, Type *&Ty) { 851 if (!EnableReduxCost) 852 return RK_None; 853 854 // Need to extract the first element. 855 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 856 unsigned Idx = ~0u; 857 if (CI) 858 Idx = CI->getZExtValue(); 859 if (Idx != 0) 860 return RK_None; 861 862 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 863 if (!RdxStart) 864 return RK_None; 865 Optional<ReductionData> RD = getReductionData(RdxStart); 866 if (!RD) 867 return RK_None; 868 869 Type *VecTy = RdxStart->getType(); 870 unsigned NumVecElems = VecTy->getVectorNumElements(); 871 if (!isPowerOf2_32(NumVecElems)) 872 return RK_None; 873 874 // We look for a sequence of shuffle,shuffle,add triples like the following 875 // that builds a pairwise reduction tree. 876 // 877 // (X0, X1, X2, X3) 878 // (X0 + X1, X2 + X3, undef, undef) 879 // ((X0 + X1) + (X2 + X3), undef, undef, undef) 880 // 881 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 882 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 883 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 884 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 885 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 886 // %rdx.shuf.1.0 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 887 // <4 x i32> <i32 0, i32 undef, i32 undef, i32 undef> 888 // %rdx.shuf.1.1 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 889 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 890 // %bin.rdx8 = fadd <4 x float> %rdx.shuf.1.0, %rdx.shuf.1.1 891 // %r = extractelement <4 x float> %bin.rdx8, i32 0 892 if (matchPairwiseReductionAtLevel(RdxStart, 0, Log2_32(NumVecElems)) == 893 RK_None) 894 return RK_None; 895 896 Opcode = RD->Opcode; 897 Ty = VecTy; 898 899 return RD->Kind; 900 } 901 902 static std::pair<Value *, ShuffleVectorInst *> 903 getShuffleAndOtherOprd(Value *L, Value *R) { 904 ShuffleVectorInst *S = nullptr; 905 906 if ((S = dyn_cast<ShuffleVectorInst>(L))) 907 return std::make_pair(R, S); 908 909 S = dyn_cast<ShuffleVectorInst>(R); 910 return std::make_pair(L, S); 911 } 912 913 static ReductionKind 914 matchVectorSplittingReduction(const ExtractElementInst *ReduxRoot, 915 unsigned &Opcode, Type *&Ty) { 916 if (!EnableReduxCost) 917 return RK_None; 918 919 // Need to extract the first element. 920 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 921 unsigned Idx = ~0u; 922 if (CI) 923 Idx = CI->getZExtValue(); 924 if (Idx != 0) 925 return RK_None; 926 927 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 928 if (!RdxStart) 929 return RK_None; 930 Optional<ReductionData> RD = getReductionData(RdxStart); 931 if (!RD) 932 return RK_None; 933 934 Type *VecTy = ReduxRoot->getOperand(0)->getType(); 935 unsigned NumVecElems = VecTy->getVectorNumElements(); 936 if (!isPowerOf2_32(NumVecElems)) 937 return RK_None; 938 939 // We look for a sequence of shuffles and adds like the following matching one 940 // fadd, shuffle vector pair at a time. 941 // 942 // %rdx.shuf = shufflevector <4 x float> %rdx, <4 x float> undef, 943 // <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 944 // %bin.rdx = fadd <4 x float> %rdx, %rdx.shuf 945 // %rdx.shuf7 = shufflevector <4 x float> %bin.rdx, <4 x float> undef, 946 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 947 // %bin.rdx8 = fadd <4 x float> %bin.rdx, %rdx.shuf7 948 // %r = extractelement <4 x float> %bin.rdx8, i32 0 949 950 unsigned MaskStart = 1; 951 Instruction *RdxOp = RdxStart; 952 SmallVector<int, 32> ShuffleMask(NumVecElems, 0); 953 unsigned NumVecElemsRemain = NumVecElems; 954 while (NumVecElemsRemain - 1) { 955 // Check for the right reduction operation. 956 if (!RdxOp) 957 return RK_None; 958 Optional<ReductionData> RDLevel = getReductionData(RdxOp); 959 if (!RDLevel || !RDLevel->hasSameData(*RD)) 960 return RK_None; 961 962 Value *NextRdxOp; 963 ShuffleVectorInst *Shuffle; 964 std::tie(NextRdxOp, Shuffle) = 965 getShuffleAndOtherOprd(RDLevel->LHS, RDLevel->RHS); 966 967 // Check the current reduction operation and the shuffle use the same value. 968 if (Shuffle == nullptr) 969 return RK_None; 970 if (Shuffle->getOperand(0) != NextRdxOp) 971 return RK_None; 972 973 // Check that shuffle masks matches. 974 for (unsigned j = 0; j != MaskStart; ++j) 975 ShuffleMask[j] = MaskStart + j; 976 // Fill the rest of the mask with -1 for undef. 977 std::fill(&ShuffleMask[MaskStart], ShuffleMask.end(), -1); 978 979 SmallVector<int, 16> Mask = Shuffle->getShuffleMask(); 980 if (ShuffleMask != Mask) 981 return RK_None; 982 983 RdxOp = dyn_cast<Instruction>(NextRdxOp); 984 NumVecElemsRemain /= 2; 985 MaskStart *= 2; 986 } 987 988 Opcode = RD->Opcode; 989 Ty = VecTy; 990 return RD->Kind; 991 } 992 993 int TargetTransformInfo::getInstructionThroughput(const Instruction *I) const { 994 switch (I->getOpcode()) { 995 case Instruction::GetElementPtr: 996 return getUserCost(I); 997 998 case Instruction::Ret: 999 case Instruction::PHI: 1000 case Instruction::Br: { 1001 return getCFInstrCost(I->getOpcode()); 1002 } 1003 case Instruction::Add: 1004 case Instruction::FAdd: 1005 case Instruction::Sub: 1006 case Instruction::FSub: 1007 case Instruction::Mul: 1008 case Instruction::FMul: 1009 case Instruction::UDiv: 1010 case Instruction::SDiv: 1011 case Instruction::FDiv: 1012 case Instruction::URem: 1013 case Instruction::SRem: 1014 case Instruction::FRem: 1015 case Instruction::Shl: 1016 case Instruction::LShr: 1017 case Instruction::AShr: 1018 case Instruction::And: 1019 case Instruction::Or: 1020 case Instruction::Xor: { 1021 TargetTransformInfo::OperandValueKind Op1VK, Op2VK; 1022 TargetTransformInfo::OperandValueProperties Op1VP, Op2VP; 1023 Op1VK = getOperandInfo(I->getOperand(0), Op1VP); 1024 Op2VK = getOperandInfo(I->getOperand(1), Op2VP); 1025 SmallVector<const Value *, 2> Operands(I->operand_values()); 1026 return getArithmeticInstrCost(I->getOpcode(), I->getType(), Op1VK, Op2VK, 1027 Op1VP, Op2VP, Operands); 1028 } 1029 case Instruction::Select: { 1030 const SelectInst *SI = cast<SelectInst>(I); 1031 Type *CondTy = SI->getCondition()->getType(); 1032 return getCmpSelInstrCost(I->getOpcode(), I->getType(), CondTy, I); 1033 } 1034 case Instruction::ICmp: 1035 case Instruction::FCmp: { 1036 Type *ValTy = I->getOperand(0)->getType(); 1037 return getCmpSelInstrCost(I->getOpcode(), ValTy, I->getType(), I); 1038 } 1039 case Instruction::Store: { 1040 const StoreInst *SI = cast<StoreInst>(I); 1041 Type *ValTy = SI->getValueOperand()->getType(); 1042 return getMemoryOpCost(I->getOpcode(), ValTy, 1043 SI->getAlignment(), 1044 SI->getPointerAddressSpace(), I); 1045 } 1046 case Instruction::Load: { 1047 const LoadInst *LI = cast<LoadInst>(I); 1048 return getMemoryOpCost(I->getOpcode(), I->getType(), 1049 LI->getAlignment(), 1050 LI->getPointerAddressSpace(), I); 1051 } 1052 case Instruction::ZExt: 1053 case Instruction::SExt: 1054 case Instruction::FPToUI: 1055 case Instruction::FPToSI: 1056 case Instruction::FPExt: 1057 case Instruction::PtrToInt: 1058 case Instruction::IntToPtr: 1059 case Instruction::SIToFP: 1060 case Instruction::UIToFP: 1061 case Instruction::Trunc: 1062 case Instruction::FPTrunc: 1063 case Instruction::BitCast: 1064 case Instruction::AddrSpaceCast: { 1065 Type *SrcTy = I->getOperand(0)->getType(); 1066 return getCastInstrCost(I->getOpcode(), I->getType(), SrcTy, I); 1067 } 1068 case Instruction::ExtractElement: { 1069 const ExtractElementInst * EEI = cast<ExtractElementInst>(I); 1070 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1071 unsigned Idx = -1; 1072 if (CI) 1073 Idx = CI->getZExtValue(); 1074 1075 // Try to match a reduction sequence (series of shufflevector and vector 1076 // adds followed by a extractelement). 1077 unsigned ReduxOpCode; 1078 Type *ReduxType; 1079 1080 switch (matchVectorSplittingReduction(EEI, ReduxOpCode, ReduxType)) { 1081 case RK_Arithmetic: 1082 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1083 /*IsPairwiseForm=*/false); 1084 case RK_MinMax: 1085 return getMinMaxReductionCost( 1086 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1087 /*IsPairwiseForm=*/false, /*IsUnsigned=*/false); 1088 case RK_UnsignedMinMax: 1089 return getMinMaxReductionCost( 1090 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1091 /*IsPairwiseForm=*/false, /*IsUnsigned=*/true); 1092 case RK_None: 1093 break; 1094 } 1095 1096 switch (matchPairwiseReduction(EEI, ReduxOpCode, ReduxType)) { 1097 case RK_Arithmetic: 1098 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1099 /*IsPairwiseForm=*/true); 1100 case RK_MinMax: 1101 return getMinMaxReductionCost( 1102 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1103 /*IsPairwiseForm=*/true, /*IsUnsigned=*/false); 1104 case RK_UnsignedMinMax: 1105 return getMinMaxReductionCost( 1106 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1107 /*IsPairwiseForm=*/true, /*IsUnsigned=*/true); 1108 case RK_None: 1109 break; 1110 } 1111 1112 return getVectorInstrCost(I->getOpcode(), 1113 EEI->getOperand(0)->getType(), Idx); 1114 } 1115 case Instruction::InsertElement: { 1116 const InsertElementInst * IE = cast<InsertElementInst>(I); 1117 ConstantInt *CI = dyn_cast<ConstantInt>(IE->getOperand(2)); 1118 unsigned Idx = -1; 1119 if (CI) 1120 Idx = CI->getZExtValue(); 1121 return getVectorInstrCost(I->getOpcode(), 1122 IE->getType(), Idx); 1123 } 1124 case Instruction::ShuffleVector: { 1125 const ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I); 1126 Type *Ty = Shuffle->getType(); 1127 Type *SrcTy = Shuffle->getOperand(0)->getType(); 1128 1129 // TODO: Identify and add costs for insert subvector, etc. 1130 int SubIndex; 1131 if (Shuffle->isExtractSubvectorMask(SubIndex)) 1132 return TTIImpl->getShuffleCost(SK_ExtractSubvector, SrcTy, SubIndex, Ty); 1133 1134 if (Shuffle->changesLength()) 1135 return -1; 1136 1137 if (Shuffle->isIdentity()) 1138 return 0; 1139 1140 if (Shuffle->isReverse()) 1141 return TTIImpl->getShuffleCost(SK_Reverse, Ty, 0, nullptr); 1142 1143 if (Shuffle->isSelect()) 1144 return TTIImpl->getShuffleCost(SK_Select, Ty, 0, nullptr); 1145 1146 if (Shuffle->isTranspose()) 1147 return TTIImpl->getShuffleCost(SK_Transpose, Ty, 0, nullptr); 1148 1149 if (Shuffle->isZeroEltSplat()) 1150 return TTIImpl->getShuffleCost(SK_Broadcast, Ty, 0, nullptr); 1151 1152 if (Shuffle->isSingleSource()) 1153 return TTIImpl->getShuffleCost(SK_PermuteSingleSrc, Ty, 0, nullptr); 1154 1155 return TTIImpl->getShuffleCost(SK_PermuteTwoSrc, Ty, 0, nullptr); 1156 } 1157 case Instruction::Call: 1158 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1159 SmallVector<Value *, 4> Args(II->arg_operands()); 1160 1161 FastMathFlags FMF; 1162 if (auto *FPMO = dyn_cast<FPMathOperator>(II)) 1163 FMF = FPMO->getFastMathFlags(); 1164 1165 return getIntrinsicInstrCost(II->getIntrinsicID(), II->getType(), 1166 Args, FMF); 1167 } 1168 return -1; 1169 default: 1170 // We don't have any information on this instruction. 1171 return -1; 1172 } 1173 } 1174 1175 TargetTransformInfo::Concept::~Concept() {} 1176 1177 TargetIRAnalysis::TargetIRAnalysis() : TTICallback(&getDefaultTTI) {} 1178 1179 TargetIRAnalysis::TargetIRAnalysis( 1180 std::function<Result(const Function &)> TTICallback) 1181 : TTICallback(std::move(TTICallback)) {} 1182 1183 TargetIRAnalysis::Result TargetIRAnalysis::run(const Function &F, 1184 FunctionAnalysisManager &) { 1185 return TTICallback(F); 1186 } 1187 1188 AnalysisKey TargetIRAnalysis::Key; 1189 1190 TargetIRAnalysis::Result TargetIRAnalysis::getDefaultTTI(const Function &F) { 1191 return Result(F.getParent()->getDataLayout()); 1192 } 1193 1194 // Register the basic pass. 1195 INITIALIZE_PASS(TargetTransformInfoWrapperPass, "tti", 1196 "Target Transform Information", false, true) 1197 char TargetTransformInfoWrapperPass::ID = 0; 1198 1199 void TargetTransformInfoWrapperPass::anchor() {} 1200 1201 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass() 1202 : ImmutablePass(ID) { 1203 initializeTargetTransformInfoWrapperPassPass( 1204 *PassRegistry::getPassRegistry()); 1205 } 1206 1207 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass( 1208 TargetIRAnalysis TIRA) 1209 : ImmutablePass(ID), TIRA(std::move(TIRA)) { 1210 initializeTargetTransformInfoWrapperPassPass( 1211 *PassRegistry::getPassRegistry()); 1212 } 1213 1214 TargetTransformInfo &TargetTransformInfoWrapperPass::getTTI(const Function &F) { 1215 FunctionAnalysisManager DummyFAM; 1216 TTI = TIRA.run(F, DummyFAM); 1217 return *TTI; 1218 } 1219 1220 ImmutablePass * 1221 llvm::createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA) { 1222 return new TargetTransformInfoWrapperPass(std::move(TIRA)); 1223 } 1224