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, 64 const User *U) const { 65 int Cost = TTIImpl->getCallCost(FTy, NumArgs, U); 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, 72 const User *U) const { 73 int Cost = TTIImpl->getCallCost(F, Arguments, U); 74 assert(Cost >= 0 && "TTI should not produce negative costs!"); 75 return Cost; 76 } 77 78 unsigned TargetTransformInfo::getInliningThresholdMultiplier() const { 79 return TTIImpl->getInliningThresholdMultiplier(); 80 } 81 82 int TargetTransformInfo::getGEPCost(Type *PointeeType, const Value *Ptr, 83 ArrayRef<const Value *> Operands) const { 84 return TTIImpl->getGEPCost(PointeeType, Ptr, Operands); 85 } 86 87 int TargetTransformInfo::getExtCost(const Instruction *I, 88 const Value *Src) const { 89 return TTIImpl->getExtCost(I, Src); 90 } 91 92 int TargetTransformInfo::getIntrinsicCost( 93 Intrinsic::ID IID, Type *RetTy, ArrayRef<const Value *> Arguments, 94 const User *U) const { 95 int Cost = TTIImpl->getIntrinsicCost(IID, RetTy, Arguments, U); 96 assert(Cost >= 0 && "TTI should not produce negative costs!"); 97 return Cost; 98 } 99 100 unsigned 101 TargetTransformInfo::getEstimatedNumberOfCaseClusters(const SwitchInst &SI, 102 unsigned &JTSize) const { 103 return TTIImpl->getEstimatedNumberOfCaseClusters(SI, JTSize); 104 } 105 106 int TargetTransformInfo::getUserCost(const User *U, 107 ArrayRef<const Value *> Operands) const { 108 int Cost = TTIImpl->getUserCost(U, Operands); 109 assert(Cost >= 0 && "TTI should not produce negative costs!"); 110 return Cost; 111 } 112 113 bool TargetTransformInfo::hasBranchDivergence() const { 114 return TTIImpl->hasBranchDivergence(); 115 } 116 117 bool TargetTransformInfo::isSourceOfDivergence(const Value *V) const { 118 return TTIImpl->isSourceOfDivergence(V); 119 } 120 121 bool llvm::TargetTransformInfo::isAlwaysUniform(const Value *V) const { 122 return TTIImpl->isAlwaysUniform(V); 123 } 124 125 unsigned TargetTransformInfo::getFlatAddressSpace() const { 126 return TTIImpl->getFlatAddressSpace(); 127 } 128 129 bool TargetTransformInfo::isLoweredToCall(const Function *F) const { 130 return TTIImpl->isLoweredToCall(F); 131 } 132 133 void TargetTransformInfo::getUnrollingPreferences( 134 Loop *L, ScalarEvolution &SE, UnrollingPreferences &UP) const { 135 return TTIImpl->getUnrollingPreferences(L, SE, UP); 136 } 137 138 bool TargetTransformInfo::isLegalAddImmediate(int64_t Imm) const { 139 return TTIImpl->isLegalAddImmediate(Imm); 140 } 141 142 bool TargetTransformInfo::isLegalICmpImmediate(int64_t Imm) const { 143 return TTIImpl->isLegalICmpImmediate(Imm); 144 } 145 146 bool TargetTransformInfo::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, 147 int64_t BaseOffset, 148 bool HasBaseReg, 149 int64_t Scale, 150 unsigned AddrSpace, 151 Instruction *I) const { 152 return TTIImpl->isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg, 153 Scale, AddrSpace, I); 154 } 155 156 bool TargetTransformInfo::isLSRCostLess(LSRCost &C1, LSRCost &C2) const { 157 return TTIImpl->isLSRCostLess(C1, C2); 158 } 159 160 bool TargetTransformInfo::canMacroFuseCmp() const { 161 return TTIImpl->canMacroFuseCmp(); 162 } 163 164 bool TargetTransformInfo::shouldFavorPostInc() const { 165 return TTIImpl->shouldFavorPostInc(); 166 } 167 168 bool TargetTransformInfo::shouldFavorBackedgeIndex(const Loop *L) const { 169 return TTIImpl->shouldFavorBackedgeIndex(L); 170 } 171 172 bool TargetTransformInfo::isLegalMaskedStore(Type *DataType) const { 173 return TTIImpl->isLegalMaskedStore(DataType); 174 } 175 176 bool TargetTransformInfo::isLegalMaskedLoad(Type *DataType) const { 177 return TTIImpl->isLegalMaskedLoad(DataType); 178 } 179 180 bool TargetTransformInfo::isLegalMaskedGather(Type *DataType) const { 181 return TTIImpl->isLegalMaskedGather(DataType); 182 } 183 184 bool TargetTransformInfo::isLegalMaskedScatter(Type *DataType) const { 185 return TTIImpl->isLegalMaskedScatter(DataType); 186 } 187 188 bool TargetTransformInfo::isLegalMaskedCompressStore(Type *DataType) const { 189 return TTIImpl->isLegalMaskedCompressStore(DataType); 190 } 191 192 bool TargetTransformInfo::isLegalMaskedExpandLoad(Type *DataType) const { 193 return TTIImpl->isLegalMaskedExpandLoad(DataType); 194 } 195 196 bool TargetTransformInfo::hasDivRemOp(Type *DataType, bool IsSigned) const { 197 return TTIImpl->hasDivRemOp(DataType, IsSigned); 198 } 199 200 bool TargetTransformInfo::hasVolatileVariant(Instruction *I, 201 unsigned AddrSpace) const { 202 return TTIImpl->hasVolatileVariant(I, AddrSpace); 203 } 204 205 bool TargetTransformInfo::prefersVectorizedAddressing() const { 206 return TTIImpl->prefersVectorizedAddressing(); 207 } 208 209 int TargetTransformInfo::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, 210 int64_t BaseOffset, 211 bool HasBaseReg, 212 int64_t Scale, 213 unsigned AddrSpace) const { 214 int Cost = TTIImpl->getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg, 215 Scale, AddrSpace); 216 assert(Cost >= 0 && "TTI should not produce negative costs!"); 217 return Cost; 218 } 219 220 bool TargetTransformInfo::LSRWithInstrQueries() const { 221 return TTIImpl->LSRWithInstrQueries(); 222 } 223 224 bool TargetTransformInfo::isTruncateFree(Type *Ty1, Type *Ty2) const { 225 return TTIImpl->isTruncateFree(Ty1, Ty2); 226 } 227 228 bool TargetTransformInfo::isProfitableToHoist(Instruction *I) const { 229 return TTIImpl->isProfitableToHoist(I); 230 } 231 232 bool TargetTransformInfo::useAA() const { return TTIImpl->useAA(); } 233 234 bool TargetTransformInfo::isTypeLegal(Type *Ty) const { 235 return TTIImpl->isTypeLegal(Ty); 236 } 237 238 unsigned TargetTransformInfo::getJumpBufAlignment() const { 239 return TTIImpl->getJumpBufAlignment(); 240 } 241 242 unsigned TargetTransformInfo::getJumpBufSize() const { 243 return TTIImpl->getJumpBufSize(); 244 } 245 246 bool TargetTransformInfo::shouldBuildLookupTables() const { 247 return TTIImpl->shouldBuildLookupTables(); 248 } 249 bool TargetTransformInfo::shouldBuildLookupTablesForConstant(Constant *C) const { 250 return TTIImpl->shouldBuildLookupTablesForConstant(C); 251 } 252 253 bool TargetTransformInfo::useColdCCForColdCall(Function &F) const { 254 return TTIImpl->useColdCCForColdCall(F); 255 } 256 257 unsigned TargetTransformInfo:: 258 getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const { 259 return TTIImpl->getScalarizationOverhead(Ty, Insert, Extract); 260 } 261 262 unsigned TargetTransformInfo:: 263 getOperandsScalarizationOverhead(ArrayRef<const Value *> Args, 264 unsigned VF) const { 265 return TTIImpl->getOperandsScalarizationOverhead(Args, VF); 266 } 267 268 bool TargetTransformInfo::supportsEfficientVectorElementLoadStore() const { 269 return TTIImpl->supportsEfficientVectorElementLoadStore(); 270 } 271 272 bool TargetTransformInfo::enableAggressiveInterleaving(bool LoopHasReductions) const { 273 return TTIImpl->enableAggressiveInterleaving(LoopHasReductions); 274 } 275 276 const TargetTransformInfo::MemCmpExpansionOptions * 277 TargetTransformInfo::enableMemCmpExpansion(bool IsZeroCmp) const { 278 return TTIImpl->enableMemCmpExpansion(IsZeroCmp); 279 } 280 281 bool TargetTransformInfo::enableInterleavedAccessVectorization() const { 282 return TTIImpl->enableInterleavedAccessVectorization(); 283 } 284 285 bool TargetTransformInfo::enableMaskedInterleavedAccessVectorization() const { 286 return TTIImpl->enableMaskedInterleavedAccessVectorization(); 287 } 288 289 bool TargetTransformInfo::isFPVectorizationPotentiallyUnsafe() const { 290 return TTIImpl->isFPVectorizationPotentiallyUnsafe(); 291 } 292 293 bool TargetTransformInfo::allowsMisalignedMemoryAccesses(LLVMContext &Context, 294 unsigned BitWidth, 295 unsigned AddressSpace, 296 unsigned Alignment, 297 bool *Fast) const { 298 return TTIImpl->allowsMisalignedMemoryAccesses(Context, BitWidth, AddressSpace, 299 Alignment, Fast); 300 } 301 302 TargetTransformInfo::PopcntSupportKind 303 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit) const { 304 return TTIImpl->getPopcntSupport(IntTyWidthInBit); 305 } 306 307 bool TargetTransformInfo::haveFastSqrt(Type *Ty) const { 308 return TTIImpl->haveFastSqrt(Ty); 309 } 310 311 bool TargetTransformInfo::isFCmpOrdCheaperThanFCmpZero(Type *Ty) const { 312 return TTIImpl->isFCmpOrdCheaperThanFCmpZero(Ty); 313 } 314 315 int TargetTransformInfo::getFPOpCost(Type *Ty) const { 316 int Cost = TTIImpl->getFPOpCost(Ty); 317 assert(Cost >= 0 && "TTI should not produce negative costs!"); 318 return Cost; 319 } 320 321 int TargetTransformInfo::getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, 322 const APInt &Imm, 323 Type *Ty) const { 324 int Cost = TTIImpl->getIntImmCodeSizeCost(Opcode, Idx, Imm, Ty); 325 assert(Cost >= 0 && "TTI should not produce negative costs!"); 326 return Cost; 327 } 328 329 int TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty) const { 330 int Cost = TTIImpl->getIntImmCost(Imm, Ty); 331 assert(Cost >= 0 && "TTI should not produce negative costs!"); 332 return Cost; 333 } 334 335 int TargetTransformInfo::getIntImmCost(unsigned Opcode, unsigned Idx, 336 const APInt &Imm, Type *Ty) const { 337 int Cost = TTIImpl->getIntImmCost(Opcode, Idx, Imm, Ty); 338 assert(Cost >= 0 && "TTI should not produce negative costs!"); 339 return Cost; 340 } 341 342 int TargetTransformInfo::getIntImmCost(Intrinsic::ID IID, unsigned Idx, 343 const APInt &Imm, Type *Ty) const { 344 int Cost = TTIImpl->getIntImmCost(IID, Idx, Imm, Ty); 345 assert(Cost >= 0 && "TTI should not produce negative costs!"); 346 return Cost; 347 } 348 349 unsigned TargetTransformInfo::getNumberOfRegisters(bool Vector) const { 350 return TTIImpl->getNumberOfRegisters(Vector); 351 } 352 353 unsigned TargetTransformInfo::getRegisterBitWidth(bool Vector) const { 354 return TTIImpl->getRegisterBitWidth(Vector); 355 } 356 357 unsigned TargetTransformInfo::getMinVectorRegisterBitWidth() const { 358 return TTIImpl->getMinVectorRegisterBitWidth(); 359 } 360 361 bool TargetTransformInfo::shouldMaximizeVectorBandwidth(bool OptSize) const { 362 return TTIImpl->shouldMaximizeVectorBandwidth(OptSize); 363 } 364 365 unsigned TargetTransformInfo::getMinimumVF(unsigned ElemWidth) const { 366 return TTIImpl->getMinimumVF(ElemWidth); 367 } 368 369 bool TargetTransformInfo::shouldConsiderAddressTypePromotion( 370 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const { 371 return TTIImpl->shouldConsiderAddressTypePromotion( 372 I, AllowPromotionWithoutCommonHeader); 373 } 374 375 unsigned TargetTransformInfo::getCacheLineSize() const { 376 return TTIImpl->getCacheLineSize(); 377 } 378 379 llvm::Optional<unsigned> TargetTransformInfo::getCacheSize(CacheLevel Level) 380 const { 381 return TTIImpl->getCacheSize(Level); 382 } 383 384 llvm::Optional<unsigned> TargetTransformInfo::getCacheAssociativity( 385 CacheLevel Level) const { 386 return TTIImpl->getCacheAssociativity(Level); 387 } 388 389 unsigned TargetTransformInfo::getPrefetchDistance() const { 390 return TTIImpl->getPrefetchDistance(); 391 } 392 393 unsigned TargetTransformInfo::getMinPrefetchStride() const { 394 return TTIImpl->getMinPrefetchStride(); 395 } 396 397 unsigned TargetTransformInfo::getMaxPrefetchIterationsAhead() const { 398 return TTIImpl->getMaxPrefetchIterationsAhead(); 399 } 400 401 unsigned TargetTransformInfo::getMaxInterleaveFactor(unsigned VF) const { 402 return TTIImpl->getMaxInterleaveFactor(VF); 403 } 404 405 TargetTransformInfo::OperandValueKind 406 TargetTransformInfo::getOperandInfo(Value *V, OperandValueProperties &OpProps) { 407 OperandValueKind OpInfo = OK_AnyValue; 408 OpProps = OP_None; 409 410 if (auto *CI = dyn_cast<ConstantInt>(V)) { 411 if (CI->getValue().isPowerOf2()) 412 OpProps = OP_PowerOf2; 413 return OK_UniformConstantValue; 414 } 415 416 // A broadcast shuffle creates a uniform value. 417 // TODO: Add support for non-zero index broadcasts. 418 // TODO: Add support for different source vector width. 419 if (auto *ShuffleInst = dyn_cast<ShuffleVectorInst>(V)) 420 if (ShuffleInst->isZeroEltSplat()) 421 OpInfo = OK_UniformValue; 422 423 const Value *Splat = getSplatValue(V); 424 425 // Check for a splat of a constant or for a non uniform vector of constants 426 // and check if the constant(s) are all powers of two. 427 if (isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) { 428 OpInfo = OK_NonUniformConstantValue; 429 if (Splat) { 430 OpInfo = OK_UniformConstantValue; 431 if (auto *CI = dyn_cast<ConstantInt>(Splat)) 432 if (CI->getValue().isPowerOf2()) 433 OpProps = OP_PowerOf2; 434 } else if (auto *CDS = dyn_cast<ConstantDataSequential>(V)) { 435 OpProps = OP_PowerOf2; 436 for (unsigned I = 0, E = CDS->getNumElements(); I != E; ++I) { 437 if (auto *CI = dyn_cast<ConstantInt>(CDS->getElementAsConstant(I))) 438 if (CI->getValue().isPowerOf2()) 439 continue; 440 OpProps = OP_None; 441 break; 442 } 443 } 444 } 445 446 // Check for a splat of a uniform value. This is not loop aware, so return 447 // true only for the obviously uniform cases (argument, globalvalue) 448 if (Splat && (isa<Argument>(Splat) || isa<GlobalValue>(Splat))) 449 OpInfo = OK_UniformValue; 450 451 return OpInfo; 452 } 453 454 int TargetTransformInfo::getArithmeticInstrCost( 455 unsigned Opcode, Type *Ty, OperandValueKind Opd1Info, 456 OperandValueKind Opd2Info, OperandValueProperties Opd1PropInfo, 457 OperandValueProperties Opd2PropInfo, 458 ArrayRef<const Value *> Args) const { 459 int Cost = TTIImpl->getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info, 460 Opd1PropInfo, Opd2PropInfo, Args); 461 assert(Cost >= 0 && "TTI should not produce negative costs!"); 462 return Cost; 463 } 464 465 int TargetTransformInfo::getShuffleCost(ShuffleKind Kind, Type *Ty, int Index, 466 Type *SubTp) const { 467 int Cost = TTIImpl->getShuffleCost(Kind, Ty, Index, SubTp); 468 assert(Cost >= 0 && "TTI should not produce negative costs!"); 469 return Cost; 470 } 471 472 int TargetTransformInfo::getCastInstrCost(unsigned Opcode, Type *Dst, 473 Type *Src, const Instruction *I) const { 474 assert ((I == nullptr || I->getOpcode() == Opcode) && 475 "Opcode should reflect passed instruction."); 476 int Cost = TTIImpl->getCastInstrCost(Opcode, Dst, Src, I); 477 assert(Cost >= 0 && "TTI should not produce negative costs!"); 478 return Cost; 479 } 480 481 int TargetTransformInfo::getExtractWithExtendCost(unsigned Opcode, Type *Dst, 482 VectorType *VecTy, 483 unsigned Index) const { 484 int Cost = TTIImpl->getExtractWithExtendCost(Opcode, Dst, VecTy, Index); 485 assert(Cost >= 0 && "TTI should not produce negative costs!"); 486 return Cost; 487 } 488 489 int TargetTransformInfo::getCFInstrCost(unsigned Opcode) const { 490 int Cost = TTIImpl->getCFInstrCost(Opcode); 491 assert(Cost >= 0 && "TTI should not produce negative costs!"); 492 return Cost; 493 } 494 495 int TargetTransformInfo::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 496 Type *CondTy, const Instruction *I) const { 497 assert ((I == nullptr || I->getOpcode() == Opcode) && 498 "Opcode should reflect passed instruction."); 499 int Cost = TTIImpl->getCmpSelInstrCost(Opcode, ValTy, CondTy, I); 500 assert(Cost >= 0 && "TTI should not produce negative costs!"); 501 return Cost; 502 } 503 504 int TargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val, 505 unsigned Index) const { 506 int Cost = TTIImpl->getVectorInstrCost(Opcode, Val, Index); 507 assert(Cost >= 0 && "TTI should not produce negative costs!"); 508 return Cost; 509 } 510 511 int TargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src, 512 unsigned Alignment, 513 unsigned AddressSpace, 514 const Instruction *I) const { 515 assert ((I == nullptr || I->getOpcode() == Opcode) && 516 "Opcode should reflect passed instruction."); 517 int Cost = TTIImpl->getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, I); 518 assert(Cost >= 0 && "TTI should not produce negative costs!"); 519 return Cost; 520 } 521 522 int TargetTransformInfo::getMaskedMemoryOpCost(unsigned Opcode, Type *Src, 523 unsigned Alignment, 524 unsigned AddressSpace) const { 525 int Cost = 526 TTIImpl->getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace); 527 assert(Cost >= 0 && "TTI should not produce negative costs!"); 528 return Cost; 529 } 530 531 int TargetTransformInfo::getGatherScatterOpCost(unsigned Opcode, Type *DataTy, 532 Value *Ptr, bool VariableMask, 533 unsigned Alignment) const { 534 int Cost = TTIImpl->getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask, 535 Alignment); 536 assert(Cost >= 0 && "TTI should not produce negative costs!"); 537 return Cost; 538 } 539 540 int TargetTransformInfo::getInterleavedMemoryOpCost( 541 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices, 542 unsigned Alignment, unsigned AddressSpace, bool UseMaskForCond, 543 bool UseMaskForGaps) const { 544 int Cost = TTIImpl->getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 545 Alignment, AddressSpace, 546 UseMaskForCond, 547 UseMaskForGaps); 548 assert(Cost >= 0 && "TTI should not produce negative costs!"); 549 return Cost; 550 } 551 552 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 553 ArrayRef<Type *> Tys, FastMathFlags FMF, 554 unsigned ScalarizationCostPassed) const { 555 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Tys, FMF, 556 ScalarizationCostPassed); 557 assert(Cost >= 0 && "TTI should not produce negative costs!"); 558 return Cost; 559 } 560 561 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 562 ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) const { 563 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF); 564 assert(Cost >= 0 && "TTI should not produce negative costs!"); 565 return Cost; 566 } 567 568 int TargetTransformInfo::getCallInstrCost(Function *F, Type *RetTy, 569 ArrayRef<Type *> Tys) const { 570 int Cost = TTIImpl->getCallInstrCost(F, RetTy, Tys); 571 assert(Cost >= 0 && "TTI should not produce negative costs!"); 572 return Cost; 573 } 574 575 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const { 576 return TTIImpl->getNumberOfParts(Tp); 577 } 578 579 int TargetTransformInfo::getAddressComputationCost(Type *Tp, 580 ScalarEvolution *SE, 581 const SCEV *Ptr) const { 582 int Cost = TTIImpl->getAddressComputationCost(Tp, SE, Ptr); 583 assert(Cost >= 0 && "TTI should not produce negative costs!"); 584 return Cost; 585 } 586 587 int TargetTransformInfo::getMemcpyCost(const Instruction *I) const { 588 int Cost = TTIImpl->getMemcpyCost(I); 589 assert(Cost >= 0 && "TTI should not produce negative costs!"); 590 return Cost; 591 } 592 593 int TargetTransformInfo::getArithmeticReductionCost(unsigned Opcode, Type *Ty, 594 bool IsPairwiseForm) const { 595 int Cost = TTIImpl->getArithmeticReductionCost(Opcode, Ty, IsPairwiseForm); 596 assert(Cost >= 0 && "TTI should not produce negative costs!"); 597 return Cost; 598 } 599 600 int TargetTransformInfo::getMinMaxReductionCost(Type *Ty, Type *CondTy, 601 bool IsPairwiseForm, 602 bool IsUnsigned) const { 603 int Cost = 604 TTIImpl->getMinMaxReductionCost(Ty, CondTy, IsPairwiseForm, IsUnsigned); 605 assert(Cost >= 0 && "TTI should not produce negative costs!"); 606 return Cost; 607 } 608 609 unsigned 610 TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const { 611 return TTIImpl->getCostOfKeepingLiveOverCall(Tys); 612 } 613 614 bool TargetTransformInfo::getTgtMemIntrinsic(IntrinsicInst *Inst, 615 MemIntrinsicInfo &Info) const { 616 return TTIImpl->getTgtMemIntrinsic(Inst, Info); 617 } 618 619 unsigned TargetTransformInfo::getAtomicMemIntrinsicMaxElementSize() const { 620 return TTIImpl->getAtomicMemIntrinsicMaxElementSize(); 621 } 622 623 Value *TargetTransformInfo::getOrCreateResultFromMemIntrinsic( 624 IntrinsicInst *Inst, Type *ExpectedType) const { 625 return TTIImpl->getOrCreateResultFromMemIntrinsic(Inst, ExpectedType); 626 } 627 628 Type *TargetTransformInfo::getMemcpyLoopLoweringType(LLVMContext &Context, 629 Value *Length, 630 unsigned SrcAlign, 631 unsigned DestAlign) const { 632 return TTIImpl->getMemcpyLoopLoweringType(Context, Length, SrcAlign, 633 DestAlign); 634 } 635 636 void TargetTransformInfo::getMemcpyLoopResidualLoweringType( 637 SmallVectorImpl<Type *> &OpsOut, LLVMContext &Context, 638 unsigned RemainingBytes, unsigned SrcAlign, unsigned DestAlign) const { 639 TTIImpl->getMemcpyLoopResidualLoweringType(OpsOut, Context, RemainingBytes, 640 SrcAlign, DestAlign); 641 } 642 643 bool TargetTransformInfo::areInlineCompatible(const Function *Caller, 644 const Function *Callee) const { 645 return TTIImpl->areInlineCompatible(Caller, Callee); 646 } 647 648 bool TargetTransformInfo::areFunctionArgsABICompatible( 649 const Function *Caller, const Function *Callee, 650 SmallPtrSetImpl<Argument *> &Args) const { 651 return TTIImpl->areFunctionArgsABICompatible(Caller, Callee, Args); 652 } 653 654 bool TargetTransformInfo::isIndexedLoadLegal(MemIndexedMode Mode, 655 Type *Ty) const { 656 return TTIImpl->isIndexedLoadLegal(Mode, Ty); 657 } 658 659 bool TargetTransformInfo::isIndexedStoreLegal(MemIndexedMode Mode, 660 Type *Ty) const { 661 return TTIImpl->isIndexedStoreLegal(Mode, Ty); 662 } 663 664 unsigned TargetTransformInfo::getLoadStoreVecRegBitWidth(unsigned AS) const { 665 return TTIImpl->getLoadStoreVecRegBitWidth(AS); 666 } 667 668 bool TargetTransformInfo::isLegalToVectorizeLoad(LoadInst *LI) const { 669 return TTIImpl->isLegalToVectorizeLoad(LI); 670 } 671 672 bool TargetTransformInfo::isLegalToVectorizeStore(StoreInst *SI) const { 673 return TTIImpl->isLegalToVectorizeStore(SI); 674 } 675 676 bool TargetTransformInfo::isLegalToVectorizeLoadChain( 677 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 678 return TTIImpl->isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment, 679 AddrSpace); 680 } 681 682 bool TargetTransformInfo::isLegalToVectorizeStoreChain( 683 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 684 return TTIImpl->isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment, 685 AddrSpace); 686 } 687 688 unsigned TargetTransformInfo::getLoadVectorFactor(unsigned VF, 689 unsigned LoadSize, 690 unsigned ChainSizeInBytes, 691 VectorType *VecTy) const { 692 return TTIImpl->getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy); 693 } 694 695 unsigned TargetTransformInfo::getStoreVectorFactor(unsigned VF, 696 unsigned StoreSize, 697 unsigned ChainSizeInBytes, 698 VectorType *VecTy) const { 699 return TTIImpl->getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy); 700 } 701 702 bool TargetTransformInfo::useReductionIntrinsic(unsigned Opcode, 703 Type *Ty, ReductionFlags Flags) const { 704 return TTIImpl->useReductionIntrinsic(Opcode, Ty, Flags); 705 } 706 707 bool TargetTransformInfo::shouldExpandReduction(const IntrinsicInst *II) const { 708 return TTIImpl->shouldExpandReduction(II); 709 } 710 711 int TargetTransformInfo::getInstructionLatency(const Instruction *I) const { 712 return TTIImpl->getInstructionLatency(I); 713 } 714 715 static bool matchPairwiseShuffleMask(ShuffleVectorInst *SI, bool IsLeft, 716 unsigned Level) { 717 // We don't need a shuffle if we just want to have element 0 in position 0 of 718 // the vector. 719 if (!SI && Level == 0 && IsLeft) 720 return true; 721 else if (!SI) 722 return false; 723 724 SmallVector<int, 32> Mask(SI->getType()->getVectorNumElements(), -1); 725 726 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right) depending on whether 727 // we look at the left or right side. 728 for (unsigned i = 0, e = (1 << Level), val = !IsLeft; i != e; ++i, val += 2) 729 Mask[i] = val; 730 731 SmallVector<int, 16> ActualMask = SI->getShuffleMask(); 732 return Mask == ActualMask; 733 } 734 735 namespace { 736 /// Kind of the reduction data. 737 enum ReductionKind { 738 RK_None, /// Not a reduction. 739 RK_Arithmetic, /// Binary reduction data. 740 RK_MinMax, /// Min/max reduction data. 741 RK_UnsignedMinMax, /// Unsigned min/max reduction data. 742 }; 743 /// Contains opcode + LHS/RHS parts of the reduction operations. 744 struct ReductionData { 745 ReductionData() = delete; 746 ReductionData(ReductionKind Kind, unsigned Opcode, Value *LHS, Value *RHS) 747 : Opcode(Opcode), LHS(LHS), RHS(RHS), Kind(Kind) { 748 assert(Kind != RK_None && "expected binary or min/max reduction only."); 749 } 750 unsigned Opcode = 0; 751 Value *LHS = nullptr; 752 Value *RHS = nullptr; 753 ReductionKind Kind = RK_None; 754 bool hasSameData(ReductionData &RD) const { 755 return Kind == RD.Kind && Opcode == RD.Opcode; 756 } 757 }; 758 } // namespace 759 760 static Optional<ReductionData> getReductionData(Instruction *I) { 761 Value *L, *R; 762 if (m_BinOp(m_Value(L), m_Value(R)).match(I)) 763 return ReductionData(RK_Arithmetic, I->getOpcode(), L, R); 764 if (auto *SI = dyn_cast<SelectInst>(I)) { 765 if (m_SMin(m_Value(L), m_Value(R)).match(SI) || 766 m_SMax(m_Value(L), m_Value(R)).match(SI) || 767 m_OrdFMin(m_Value(L), m_Value(R)).match(SI) || 768 m_OrdFMax(m_Value(L), m_Value(R)).match(SI) || 769 m_UnordFMin(m_Value(L), m_Value(R)).match(SI) || 770 m_UnordFMax(m_Value(L), m_Value(R)).match(SI)) { 771 auto *CI = cast<CmpInst>(SI->getCondition()); 772 return ReductionData(RK_MinMax, CI->getOpcode(), L, R); 773 } 774 if (m_UMin(m_Value(L), m_Value(R)).match(SI) || 775 m_UMax(m_Value(L), m_Value(R)).match(SI)) { 776 auto *CI = cast<CmpInst>(SI->getCondition()); 777 return ReductionData(RK_UnsignedMinMax, CI->getOpcode(), L, R); 778 } 779 } 780 return llvm::None; 781 } 782 783 static ReductionKind matchPairwiseReductionAtLevel(Instruction *I, 784 unsigned Level, 785 unsigned NumLevels) { 786 // Match one level of pairwise operations. 787 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 788 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 789 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 790 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 791 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 792 if (!I) 793 return RK_None; 794 795 assert(I->getType()->isVectorTy() && "Expecting a vector type"); 796 797 Optional<ReductionData> RD = getReductionData(I); 798 if (!RD) 799 return RK_None; 800 801 ShuffleVectorInst *LS = dyn_cast<ShuffleVectorInst>(RD->LHS); 802 if (!LS && Level) 803 return RK_None; 804 ShuffleVectorInst *RS = dyn_cast<ShuffleVectorInst>(RD->RHS); 805 if (!RS && Level) 806 return RK_None; 807 808 // On level 0 we can omit one shufflevector instruction. 809 if (!Level && !RS && !LS) 810 return RK_None; 811 812 // Shuffle inputs must match. 813 Value *NextLevelOpL = LS ? LS->getOperand(0) : nullptr; 814 Value *NextLevelOpR = RS ? RS->getOperand(0) : nullptr; 815 Value *NextLevelOp = nullptr; 816 if (NextLevelOpR && NextLevelOpL) { 817 // If we have two shuffles their operands must match. 818 if (NextLevelOpL != NextLevelOpR) 819 return RK_None; 820 821 NextLevelOp = NextLevelOpL; 822 } else if (Level == 0 && (NextLevelOpR || NextLevelOpL)) { 823 // On the first level we can omit the shufflevector <0, undef,...>. So the 824 // input to the other shufflevector <1, undef> must match with one of the 825 // inputs to the current binary operation. 826 // Example: 827 // %NextLevelOpL = shufflevector %R, <1, undef ...> 828 // %BinOp = fadd %NextLevelOpL, %R 829 if (NextLevelOpL && NextLevelOpL != RD->RHS) 830 return RK_None; 831 else if (NextLevelOpR && NextLevelOpR != RD->LHS) 832 return RK_None; 833 834 NextLevelOp = NextLevelOpL ? RD->RHS : RD->LHS; 835 } else 836 return RK_None; 837 838 // Check that the next levels binary operation exists and matches with the 839 // current one. 840 if (Level + 1 != NumLevels) { 841 Optional<ReductionData> NextLevelRD = 842 getReductionData(cast<Instruction>(NextLevelOp)); 843 if (!NextLevelRD || !RD->hasSameData(*NextLevelRD)) 844 return RK_None; 845 } 846 847 // Shuffle mask for pairwise operation must match. 848 if (matchPairwiseShuffleMask(LS, /*IsLeft=*/true, Level)) { 849 if (!matchPairwiseShuffleMask(RS, /*IsLeft=*/false, Level)) 850 return RK_None; 851 } else if (matchPairwiseShuffleMask(RS, /*IsLeft=*/true, Level)) { 852 if (!matchPairwiseShuffleMask(LS, /*IsLeft=*/false, Level)) 853 return RK_None; 854 } else { 855 return RK_None; 856 } 857 858 if (++Level == NumLevels) 859 return RD->Kind; 860 861 // Match next level. 862 return matchPairwiseReductionAtLevel(cast<Instruction>(NextLevelOp), Level, 863 NumLevels); 864 } 865 866 static ReductionKind matchPairwiseReduction(const ExtractElementInst *ReduxRoot, 867 unsigned &Opcode, Type *&Ty) { 868 if (!EnableReduxCost) 869 return RK_None; 870 871 // Need to extract the first element. 872 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 873 unsigned Idx = ~0u; 874 if (CI) 875 Idx = CI->getZExtValue(); 876 if (Idx != 0) 877 return RK_None; 878 879 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 880 if (!RdxStart) 881 return RK_None; 882 Optional<ReductionData> RD = getReductionData(RdxStart); 883 if (!RD) 884 return RK_None; 885 886 Type *VecTy = RdxStart->getType(); 887 unsigned NumVecElems = VecTy->getVectorNumElements(); 888 if (!isPowerOf2_32(NumVecElems)) 889 return RK_None; 890 891 // We look for a sequence of shuffle,shuffle,add triples like the following 892 // that builds a pairwise reduction tree. 893 // 894 // (X0, X1, X2, X3) 895 // (X0 + X1, X2 + X3, undef, undef) 896 // ((X0 + X1) + (X2 + X3), undef, undef, undef) 897 // 898 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 899 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 900 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 901 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 902 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 903 // %rdx.shuf.1.0 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 904 // <4 x i32> <i32 0, i32 undef, i32 undef, i32 undef> 905 // %rdx.shuf.1.1 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 906 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 907 // %bin.rdx8 = fadd <4 x float> %rdx.shuf.1.0, %rdx.shuf.1.1 908 // %r = extractelement <4 x float> %bin.rdx8, i32 0 909 if (matchPairwiseReductionAtLevel(RdxStart, 0, Log2_32(NumVecElems)) == 910 RK_None) 911 return RK_None; 912 913 Opcode = RD->Opcode; 914 Ty = VecTy; 915 916 return RD->Kind; 917 } 918 919 static std::pair<Value *, ShuffleVectorInst *> 920 getShuffleAndOtherOprd(Value *L, Value *R) { 921 ShuffleVectorInst *S = nullptr; 922 923 if ((S = dyn_cast<ShuffleVectorInst>(L))) 924 return std::make_pair(R, S); 925 926 S = dyn_cast<ShuffleVectorInst>(R); 927 return std::make_pair(L, S); 928 } 929 930 static ReductionKind 931 matchVectorSplittingReduction(const ExtractElementInst *ReduxRoot, 932 unsigned &Opcode, Type *&Ty) { 933 if (!EnableReduxCost) 934 return RK_None; 935 936 // Need to extract the first element. 937 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 938 unsigned Idx = ~0u; 939 if (CI) 940 Idx = CI->getZExtValue(); 941 if (Idx != 0) 942 return RK_None; 943 944 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 945 if (!RdxStart) 946 return RK_None; 947 Optional<ReductionData> RD = getReductionData(RdxStart); 948 if (!RD) 949 return RK_None; 950 951 Type *VecTy = ReduxRoot->getOperand(0)->getType(); 952 unsigned NumVecElems = VecTy->getVectorNumElements(); 953 if (!isPowerOf2_32(NumVecElems)) 954 return RK_None; 955 956 // We look for a sequence of shuffles and adds like the following matching one 957 // fadd, shuffle vector pair at a time. 958 // 959 // %rdx.shuf = shufflevector <4 x float> %rdx, <4 x float> undef, 960 // <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 961 // %bin.rdx = fadd <4 x float> %rdx, %rdx.shuf 962 // %rdx.shuf7 = shufflevector <4 x float> %bin.rdx, <4 x float> undef, 963 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 964 // %bin.rdx8 = fadd <4 x float> %bin.rdx, %rdx.shuf7 965 // %r = extractelement <4 x float> %bin.rdx8, i32 0 966 967 unsigned MaskStart = 1; 968 Instruction *RdxOp = RdxStart; 969 SmallVector<int, 32> ShuffleMask(NumVecElems, 0); 970 unsigned NumVecElemsRemain = NumVecElems; 971 while (NumVecElemsRemain - 1) { 972 // Check for the right reduction operation. 973 if (!RdxOp) 974 return RK_None; 975 Optional<ReductionData> RDLevel = getReductionData(RdxOp); 976 if (!RDLevel || !RDLevel->hasSameData(*RD)) 977 return RK_None; 978 979 Value *NextRdxOp; 980 ShuffleVectorInst *Shuffle; 981 std::tie(NextRdxOp, Shuffle) = 982 getShuffleAndOtherOprd(RDLevel->LHS, RDLevel->RHS); 983 984 // Check the current reduction operation and the shuffle use the same value. 985 if (Shuffle == nullptr) 986 return RK_None; 987 if (Shuffle->getOperand(0) != NextRdxOp) 988 return RK_None; 989 990 // Check that shuffle masks matches. 991 for (unsigned j = 0; j != MaskStart; ++j) 992 ShuffleMask[j] = MaskStart + j; 993 // Fill the rest of the mask with -1 for undef. 994 std::fill(&ShuffleMask[MaskStart], ShuffleMask.end(), -1); 995 996 SmallVector<int, 16> Mask = Shuffle->getShuffleMask(); 997 if (ShuffleMask != Mask) 998 return RK_None; 999 1000 RdxOp = dyn_cast<Instruction>(NextRdxOp); 1001 NumVecElemsRemain /= 2; 1002 MaskStart *= 2; 1003 } 1004 1005 Opcode = RD->Opcode; 1006 Ty = VecTy; 1007 return RD->Kind; 1008 } 1009 1010 int TargetTransformInfo::getInstructionThroughput(const Instruction *I) const { 1011 switch (I->getOpcode()) { 1012 case Instruction::GetElementPtr: 1013 return getUserCost(I); 1014 1015 case Instruction::Ret: 1016 case Instruction::PHI: 1017 case Instruction::Br: { 1018 return getCFInstrCost(I->getOpcode()); 1019 } 1020 case Instruction::Add: 1021 case Instruction::FAdd: 1022 case Instruction::Sub: 1023 case Instruction::FSub: 1024 case Instruction::Mul: 1025 case Instruction::FMul: 1026 case Instruction::UDiv: 1027 case Instruction::SDiv: 1028 case Instruction::FDiv: 1029 case Instruction::URem: 1030 case Instruction::SRem: 1031 case Instruction::FRem: 1032 case Instruction::Shl: 1033 case Instruction::LShr: 1034 case Instruction::AShr: 1035 case Instruction::And: 1036 case Instruction::Or: 1037 case Instruction::Xor: { 1038 TargetTransformInfo::OperandValueKind Op1VK, Op2VK; 1039 TargetTransformInfo::OperandValueProperties Op1VP, Op2VP; 1040 Op1VK = getOperandInfo(I->getOperand(0), Op1VP); 1041 Op2VK = getOperandInfo(I->getOperand(1), Op2VP); 1042 SmallVector<const Value *, 2> Operands(I->operand_values()); 1043 return getArithmeticInstrCost(I->getOpcode(), I->getType(), Op1VK, Op2VK, 1044 Op1VP, Op2VP, Operands); 1045 } 1046 case Instruction::Select: { 1047 const SelectInst *SI = cast<SelectInst>(I); 1048 Type *CondTy = SI->getCondition()->getType(); 1049 return getCmpSelInstrCost(I->getOpcode(), I->getType(), CondTy, I); 1050 } 1051 case Instruction::ICmp: 1052 case Instruction::FCmp: { 1053 Type *ValTy = I->getOperand(0)->getType(); 1054 return getCmpSelInstrCost(I->getOpcode(), ValTy, I->getType(), I); 1055 } 1056 case Instruction::Store: { 1057 const StoreInst *SI = cast<StoreInst>(I); 1058 Type *ValTy = SI->getValueOperand()->getType(); 1059 return getMemoryOpCost(I->getOpcode(), ValTy, 1060 SI->getAlignment(), 1061 SI->getPointerAddressSpace(), I); 1062 } 1063 case Instruction::Load: { 1064 const LoadInst *LI = cast<LoadInst>(I); 1065 return getMemoryOpCost(I->getOpcode(), I->getType(), 1066 LI->getAlignment(), 1067 LI->getPointerAddressSpace(), I); 1068 } 1069 case Instruction::ZExt: 1070 case Instruction::SExt: 1071 case Instruction::FPToUI: 1072 case Instruction::FPToSI: 1073 case Instruction::FPExt: 1074 case Instruction::PtrToInt: 1075 case Instruction::IntToPtr: 1076 case Instruction::SIToFP: 1077 case Instruction::UIToFP: 1078 case Instruction::Trunc: 1079 case Instruction::FPTrunc: 1080 case Instruction::BitCast: 1081 case Instruction::AddrSpaceCast: { 1082 Type *SrcTy = I->getOperand(0)->getType(); 1083 return getCastInstrCost(I->getOpcode(), I->getType(), SrcTy, I); 1084 } 1085 case Instruction::ExtractElement: { 1086 const ExtractElementInst * EEI = cast<ExtractElementInst>(I); 1087 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1088 unsigned Idx = -1; 1089 if (CI) 1090 Idx = CI->getZExtValue(); 1091 1092 // Try to match a reduction sequence (series of shufflevector and vector 1093 // adds followed by a extractelement). 1094 unsigned ReduxOpCode; 1095 Type *ReduxType; 1096 1097 switch (matchVectorSplittingReduction(EEI, ReduxOpCode, ReduxType)) { 1098 case RK_Arithmetic: 1099 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1100 /*IsPairwiseForm=*/false); 1101 case RK_MinMax: 1102 return getMinMaxReductionCost( 1103 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1104 /*IsPairwiseForm=*/false, /*IsUnsigned=*/false); 1105 case RK_UnsignedMinMax: 1106 return getMinMaxReductionCost( 1107 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1108 /*IsPairwiseForm=*/false, /*IsUnsigned=*/true); 1109 case RK_None: 1110 break; 1111 } 1112 1113 switch (matchPairwiseReduction(EEI, ReduxOpCode, ReduxType)) { 1114 case RK_Arithmetic: 1115 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1116 /*IsPairwiseForm=*/true); 1117 case RK_MinMax: 1118 return getMinMaxReductionCost( 1119 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1120 /*IsPairwiseForm=*/true, /*IsUnsigned=*/false); 1121 case RK_UnsignedMinMax: 1122 return getMinMaxReductionCost( 1123 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1124 /*IsPairwiseForm=*/true, /*IsUnsigned=*/true); 1125 case RK_None: 1126 break; 1127 } 1128 1129 return getVectorInstrCost(I->getOpcode(), 1130 EEI->getOperand(0)->getType(), Idx); 1131 } 1132 case Instruction::InsertElement: { 1133 const InsertElementInst * IE = cast<InsertElementInst>(I); 1134 ConstantInt *CI = dyn_cast<ConstantInt>(IE->getOperand(2)); 1135 unsigned Idx = -1; 1136 if (CI) 1137 Idx = CI->getZExtValue(); 1138 return getVectorInstrCost(I->getOpcode(), 1139 IE->getType(), Idx); 1140 } 1141 case Instruction::ShuffleVector: { 1142 const ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I); 1143 Type *Ty = Shuffle->getType(); 1144 Type *SrcTy = Shuffle->getOperand(0)->getType(); 1145 1146 // TODO: Identify and add costs for insert subvector, etc. 1147 int SubIndex; 1148 if (Shuffle->isExtractSubvectorMask(SubIndex)) 1149 return TTIImpl->getShuffleCost(SK_ExtractSubvector, SrcTy, SubIndex, Ty); 1150 1151 if (Shuffle->changesLength()) 1152 return -1; 1153 1154 if (Shuffle->isIdentity()) 1155 return 0; 1156 1157 if (Shuffle->isReverse()) 1158 return TTIImpl->getShuffleCost(SK_Reverse, Ty, 0, nullptr); 1159 1160 if (Shuffle->isSelect()) 1161 return TTIImpl->getShuffleCost(SK_Select, Ty, 0, nullptr); 1162 1163 if (Shuffle->isTranspose()) 1164 return TTIImpl->getShuffleCost(SK_Transpose, Ty, 0, nullptr); 1165 1166 if (Shuffle->isZeroEltSplat()) 1167 return TTIImpl->getShuffleCost(SK_Broadcast, Ty, 0, nullptr); 1168 1169 if (Shuffle->isSingleSource()) 1170 return TTIImpl->getShuffleCost(SK_PermuteSingleSrc, Ty, 0, nullptr); 1171 1172 return TTIImpl->getShuffleCost(SK_PermuteTwoSrc, Ty, 0, nullptr); 1173 } 1174 case Instruction::Call: 1175 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1176 SmallVector<Value *, 4> Args(II->arg_operands()); 1177 1178 FastMathFlags FMF; 1179 if (auto *FPMO = dyn_cast<FPMathOperator>(II)) 1180 FMF = FPMO->getFastMathFlags(); 1181 1182 return getIntrinsicInstrCost(II->getIntrinsicID(), II->getType(), 1183 Args, FMF); 1184 } 1185 return -1; 1186 default: 1187 // We don't have any information on this instruction. 1188 return -1; 1189 } 1190 } 1191 1192 TargetTransformInfo::Concept::~Concept() {} 1193 1194 TargetIRAnalysis::TargetIRAnalysis() : TTICallback(&getDefaultTTI) {} 1195 1196 TargetIRAnalysis::TargetIRAnalysis( 1197 std::function<Result(const Function &)> TTICallback) 1198 : TTICallback(std::move(TTICallback)) {} 1199 1200 TargetIRAnalysis::Result TargetIRAnalysis::run(const Function &F, 1201 FunctionAnalysisManager &) { 1202 return TTICallback(F); 1203 } 1204 1205 AnalysisKey TargetIRAnalysis::Key; 1206 1207 TargetIRAnalysis::Result TargetIRAnalysis::getDefaultTTI(const Function &F) { 1208 return Result(F.getParent()->getDataLayout()); 1209 } 1210 1211 // Register the basic pass. 1212 INITIALIZE_PASS(TargetTransformInfoWrapperPass, "tti", 1213 "Target Transform Information", false, true) 1214 char TargetTransformInfoWrapperPass::ID = 0; 1215 1216 void TargetTransformInfoWrapperPass::anchor() {} 1217 1218 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass() 1219 : ImmutablePass(ID) { 1220 initializeTargetTransformInfoWrapperPassPass( 1221 *PassRegistry::getPassRegistry()); 1222 } 1223 1224 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass( 1225 TargetIRAnalysis TIRA) 1226 : ImmutablePass(ID), TIRA(std::move(TIRA)) { 1227 initializeTargetTransformInfoWrapperPassPass( 1228 *PassRegistry::getPassRegistry()); 1229 } 1230 1231 TargetTransformInfo &TargetTransformInfoWrapperPass::getTTI(const Function &F) { 1232 FunctionAnalysisManager DummyFAM; 1233 TTI = TIRA.run(F, DummyFAM); 1234 return *TTI; 1235 } 1236 1237 ImmutablePass * 1238 llvm::createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA) { 1239 return new TargetTransformInfoWrapperPass(std::move(TIRA)); 1240 } 1241