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/CFG.h" 11 #include "llvm/Analysis/LoopIterator.h" 12 #include "llvm/Analysis/TargetTransformInfoImpl.h" 13 #include "llvm/IR/CFG.h" 14 #include "llvm/IR/CallSite.h" 15 #include "llvm/IR/DataLayout.h" 16 #include "llvm/IR/Instruction.h" 17 #include "llvm/IR/Instructions.h" 18 #include "llvm/IR/IntrinsicInst.h" 19 #include "llvm/IR/Module.h" 20 #include "llvm/IR/Operator.h" 21 #include "llvm/IR/PatternMatch.h" 22 #include "llvm/InitializePasses.h" 23 #include "llvm/Support/CommandLine.h" 24 #include "llvm/Support/ErrorHandling.h" 25 #include <utility> 26 27 using namespace llvm; 28 using namespace PatternMatch; 29 30 #define DEBUG_TYPE "tti" 31 32 static cl::opt<bool> EnableReduxCost("costmodel-reduxcost", cl::init(false), 33 cl::Hidden, 34 cl::desc("Recognize reduction patterns.")); 35 36 namespace { 37 /// No-op implementation of the TTI interface using the utility base 38 /// classes. 39 /// 40 /// This is used when no target specific information is available. 41 struct NoTTIImpl : TargetTransformInfoImplCRTPBase<NoTTIImpl> { 42 explicit NoTTIImpl(const DataLayout &DL) 43 : TargetTransformInfoImplCRTPBase<NoTTIImpl>(DL) {} 44 }; 45 } 46 47 bool HardwareLoopInfo::canAnalyze(LoopInfo &LI) { 48 // If the loop has irreducible control flow, it can not be converted to 49 // Hardware loop. 50 LoopBlocksRPO RPOT(L); 51 RPOT.perform(&LI); 52 if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI)) 53 return false; 54 return true; 55 } 56 57 bool HardwareLoopInfo::isHardwareLoopCandidate(ScalarEvolution &SE, 58 LoopInfo &LI, DominatorTree &DT, 59 bool ForceNestedLoop, 60 bool ForceHardwareLoopPHI) { 61 SmallVector<BasicBlock *, 4> ExitingBlocks; 62 L->getExitingBlocks(ExitingBlocks); 63 64 for (BasicBlock *BB : ExitingBlocks) { 65 // If we pass the updated counter back through a phi, we need to know 66 // which latch the updated value will be coming from. 67 if (!L->isLoopLatch(BB)) { 68 if (ForceHardwareLoopPHI || CounterInReg) 69 continue; 70 } 71 72 const SCEV *EC = SE.getExitCount(L, BB); 73 if (isa<SCEVCouldNotCompute>(EC)) 74 continue; 75 if (const SCEVConstant *ConstEC = dyn_cast<SCEVConstant>(EC)) { 76 if (ConstEC->getValue()->isZero()) 77 continue; 78 } else if (!SE.isLoopInvariant(EC, L)) 79 continue; 80 81 if (SE.getTypeSizeInBits(EC->getType()) > CountType->getBitWidth()) 82 continue; 83 84 // If this exiting block is contained in a nested loop, it is not eligible 85 // for insertion of the branch-and-decrement since the inner loop would 86 // end up messing up the value in the CTR. 87 if (!IsNestingLegal && LI.getLoopFor(BB) != L && !ForceNestedLoop) 88 continue; 89 90 // We now have a loop-invariant count of loop iterations (which is not the 91 // constant zero) for which we know that this loop will not exit via this 92 // existing block. 93 94 // We need to make sure that this block will run on every loop iteration. 95 // For this to be true, we must dominate all blocks with backedges. Such 96 // blocks are in-loop predecessors to the header block. 97 bool NotAlways = false; 98 for (BasicBlock *Pred : predecessors(L->getHeader())) { 99 if (!L->contains(Pred)) 100 continue; 101 102 if (!DT.dominates(BB, Pred)) { 103 NotAlways = true; 104 break; 105 } 106 } 107 108 if (NotAlways) 109 continue; 110 111 // Make sure this blocks ends with a conditional branch. 112 Instruction *TI = BB->getTerminator(); 113 if (!TI) 114 continue; 115 116 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 117 if (!BI->isConditional()) 118 continue; 119 120 ExitBranch = BI; 121 } else 122 continue; 123 124 // Note that this block may not be the loop latch block, even if the loop 125 // has a latch block. 126 ExitBlock = BB; 127 ExitCount = EC; 128 break; 129 } 130 131 if (!ExitBlock) 132 return false; 133 return true; 134 } 135 136 TargetTransformInfo::TargetTransformInfo(const DataLayout &DL) 137 : TTIImpl(new Model<NoTTIImpl>(NoTTIImpl(DL))) {} 138 139 TargetTransformInfo::~TargetTransformInfo() {} 140 141 TargetTransformInfo::TargetTransformInfo(TargetTransformInfo &&Arg) 142 : TTIImpl(std::move(Arg.TTIImpl)) {} 143 144 TargetTransformInfo &TargetTransformInfo::operator=(TargetTransformInfo &&RHS) { 145 TTIImpl = std::move(RHS.TTIImpl); 146 return *this; 147 } 148 149 int TargetTransformInfo::getOperationCost(unsigned Opcode, Type *Ty, 150 Type *OpTy) const { 151 int Cost = TTIImpl->getOperationCost(Opcode, Ty, OpTy); 152 assert(Cost >= 0 && "TTI should not produce negative costs!"); 153 return Cost; 154 } 155 156 unsigned TargetTransformInfo::getInliningThresholdMultiplier() const { 157 return TTIImpl->getInliningThresholdMultiplier(); 158 } 159 160 int TargetTransformInfo::getInlinerVectorBonusPercent() const { 161 return TTIImpl->getInlinerVectorBonusPercent(); 162 } 163 164 int TargetTransformInfo::getGEPCost(Type *PointeeType, const Value *Ptr, 165 ArrayRef<const Value *> Operands) const { 166 return TTIImpl->getGEPCost(PointeeType, Ptr, Operands); 167 } 168 169 int TargetTransformInfo::getExtCost(const Instruction *I, 170 const Value *Src) const { 171 return TTIImpl->getExtCost(I, Src); 172 } 173 174 int TargetTransformInfo::getIntrinsicCost( 175 Intrinsic::ID IID, Type *RetTy, ArrayRef<const Value *> Arguments, 176 const User *U) const { 177 int Cost = TTIImpl->getIntrinsicCost(IID, RetTy, Arguments, U); 178 assert(Cost >= 0 && "TTI should not produce negative costs!"); 179 return Cost; 180 } 181 182 unsigned 183 TargetTransformInfo::getEstimatedNumberOfCaseClusters( 184 const SwitchInst &SI, unsigned &JTSize, ProfileSummaryInfo *PSI, 185 BlockFrequencyInfo *BFI) const { 186 return TTIImpl->getEstimatedNumberOfCaseClusters(SI, JTSize, PSI, BFI); 187 } 188 189 int TargetTransformInfo::getUserCost(const User *U, 190 ArrayRef<const Value *> Operands) const { 191 int Cost = TTIImpl->getUserCost(U, Operands); 192 assert(Cost >= 0 && "TTI should not produce negative costs!"); 193 return Cost; 194 } 195 196 bool TargetTransformInfo::hasBranchDivergence() const { 197 return TTIImpl->hasBranchDivergence(); 198 } 199 200 bool TargetTransformInfo::useGPUDivergenceAnalysis() const { 201 return TTIImpl->useGPUDivergenceAnalysis(); 202 } 203 204 bool TargetTransformInfo::isSourceOfDivergence(const Value *V) const { 205 return TTIImpl->isSourceOfDivergence(V); 206 } 207 208 bool llvm::TargetTransformInfo::isAlwaysUniform(const Value *V) const { 209 return TTIImpl->isAlwaysUniform(V); 210 } 211 212 unsigned TargetTransformInfo::getFlatAddressSpace() const { 213 return TTIImpl->getFlatAddressSpace(); 214 } 215 216 bool TargetTransformInfo::collectFlatAddressOperands( 217 SmallVectorImpl<int> &OpIndexes, Intrinsic::ID IID) const { 218 return TTIImpl->collectFlatAddressOperands(OpIndexes, IID); 219 } 220 221 bool TargetTransformInfo::rewriteIntrinsicWithAddressSpace( 222 IntrinsicInst *II, Value *OldV, Value *NewV) const { 223 return TTIImpl->rewriteIntrinsicWithAddressSpace(II, OldV, NewV); 224 } 225 226 bool TargetTransformInfo::isLoweredToCall(const Function *F) const { 227 return TTIImpl->isLoweredToCall(F); 228 } 229 230 bool TargetTransformInfo::isHardwareLoopProfitable( 231 Loop *L, ScalarEvolution &SE, AssumptionCache &AC, 232 TargetLibraryInfo *LibInfo, HardwareLoopInfo &HWLoopInfo) const { 233 return TTIImpl->isHardwareLoopProfitable(L, SE, AC, LibInfo, HWLoopInfo); 234 } 235 236 bool TargetTransformInfo::preferPredicateOverEpilogue(Loop *L, LoopInfo *LI, 237 ScalarEvolution &SE, AssumptionCache &AC, TargetLibraryInfo *TLI, 238 DominatorTree *DT, const LoopAccessInfo *LAI) const { 239 return TTIImpl->preferPredicateOverEpilogue(L, LI, SE, AC, TLI, DT, LAI); 240 } 241 242 void TargetTransformInfo::getUnrollingPreferences( 243 Loop *L, ScalarEvolution &SE, UnrollingPreferences &UP) const { 244 return TTIImpl->getUnrollingPreferences(L, SE, UP); 245 } 246 247 bool TargetTransformInfo::isLegalAddImmediate(int64_t Imm) const { 248 return TTIImpl->isLegalAddImmediate(Imm); 249 } 250 251 bool TargetTransformInfo::isLegalICmpImmediate(int64_t Imm) const { 252 return TTIImpl->isLegalICmpImmediate(Imm); 253 } 254 255 bool TargetTransformInfo::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, 256 int64_t BaseOffset, 257 bool HasBaseReg, 258 int64_t Scale, 259 unsigned AddrSpace, 260 Instruction *I) const { 261 return TTIImpl->isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg, 262 Scale, AddrSpace, I); 263 } 264 265 bool TargetTransformInfo::isLSRCostLess(LSRCost &C1, LSRCost &C2) const { 266 return TTIImpl->isLSRCostLess(C1, C2); 267 } 268 269 bool TargetTransformInfo::canMacroFuseCmp() const { 270 return TTIImpl->canMacroFuseCmp(); 271 } 272 273 bool TargetTransformInfo::canSaveCmp(Loop *L, BranchInst **BI, 274 ScalarEvolution *SE, LoopInfo *LI, 275 DominatorTree *DT, AssumptionCache *AC, 276 TargetLibraryInfo *LibInfo) const { 277 return TTIImpl->canSaveCmp(L, BI, SE, LI, DT, AC, LibInfo); 278 } 279 280 bool TargetTransformInfo::shouldFavorPostInc() const { 281 return TTIImpl->shouldFavorPostInc(); 282 } 283 284 bool TargetTransformInfo::shouldFavorBackedgeIndex(const Loop *L) const { 285 return TTIImpl->shouldFavorBackedgeIndex(L); 286 } 287 288 bool TargetTransformInfo::isLegalMaskedStore(Type *DataType, 289 MaybeAlign Alignment) const { 290 return TTIImpl->isLegalMaskedStore(DataType, Alignment); 291 } 292 293 bool TargetTransformInfo::isLegalMaskedLoad(Type *DataType, 294 MaybeAlign Alignment) const { 295 return TTIImpl->isLegalMaskedLoad(DataType, Alignment); 296 } 297 298 bool TargetTransformInfo::isLegalNTStore(Type *DataType, 299 Align Alignment) const { 300 return TTIImpl->isLegalNTStore(DataType, Alignment); 301 } 302 303 bool TargetTransformInfo::isLegalNTLoad(Type *DataType, Align Alignment) const { 304 return TTIImpl->isLegalNTLoad(DataType, Alignment); 305 } 306 307 bool TargetTransformInfo::isLegalMaskedGather(Type *DataType, 308 MaybeAlign Alignment) const { 309 return TTIImpl->isLegalMaskedGather(DataType, Alignment); 310 } 311 312 bool TargetTransformInfo::isLegalMaskedScatter(Type *DataType, 313 MaybeAlign Alignment) const { 314 return TTIImpl->isLegalMaskedScatter(DataType, Alignment); 315 } 316 317 bool TargetTransformInfo::isLegalMaskedCompressStore(Type *DataType) const { 318 return TTIImpl->isLegalMaskedCompressStore(DataType); 319 } 320 321 bool TargetTransformInfo::isLegalMaskedExpandLoad(Type *DataType) const { 322 return TTIImpl->isLegalMaskedExpandLoad(DataType); 323 } 324 325 bool TargetTransformInfo::hasDivRemOp(Type *DataType, bool IsSigned) const { 326 return TTIImpl->hasDivRemOp(DataType, IsSigned); 327 } 328 329 bool TargetTransformInfo::hasVolatileVariant(Instruction *I, 330 unsigned AddrSpace) const { 331 return TTIImpl->hasVolatileVariant(I, AddrSpace); 332 } 333 334 bool TargetTransformInfo::prefersVectorizedAddressing() const { 335 return TTIImpl->prefersVectorizedAddressing(); 336 } 337 338 int TargetTransformInfo::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, 339 int64_t BaseOffset, 340 bool HasBaseReg, 341 int64_t Scale, 342 unsigned AddrSpace) const { 343 int Cost = TTIImpl->getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg, 344 Scale, AddrSpace); 345 assert(Cost >= 0 && "TTI should not produce negative costs!"); 346 return Cost; 347 } 348 349 bool TargetTransformInfo::LSRWithInstrQueries() const { 350 return TTIImpl->LSRWithInstrQueries(); 351 } 352 353 bool TargetTransformInfo::isTruncateFree(Type *Ty1, Type *Ty2) const { 354 return TTIImpl->isTruncateFree(Ty1, Ty2); 355 } 356 357 bool TargetTransformInfo::isProfitableToHoist(Instruction *I) const { 358 return TTIImpl->isProfitableToHoist(I); 359 } 360 361 bool TargetTransformInfo::useAA() const { return TTIImpl->useAA(); } 362 363 bool TargetTransformInfo::isTypeLegal(Type *Ty) const { 364 return TTIImpl->isTypeLegal(Ty); 365 } 366 367 bool TargetTransformInfo::shouldBuildLookupTables() const { 368 return TTIImpl->shouldBuildLookupTables(); 369 } 370 bool TargetTransformInfo::shouldBuildLookupTablesForConstant(Constant *C) const { 371 return TTIImpl->shouldBuildLookupTablesForConstant(C); 372 } 373 374 bool TargetTransformInfo::useColdCCForColdCall(Function &F) const { 375 return TTIImpl->useColdCCForColdCall(F); 376 } 377 378 unsigned TargetTransformInfo:: 379 getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const { 380 return TTIImpl->getScalarizationOverhead(Ty, Insert, Extract); 381 } 382 383 unsigned TargetTransformInfo:: 384 getOperandsScalarizationOverhead(ArrayRef<const Value *> Args, 385 unsigned VF) const { 386 return TTIImpl->getOperandsScalarizationOverhead(Args, VF); 387 } 388 389 bool TargetTransformInfo::supportsEfficientVectorElementLoadStore() const { 390 return TTIImpl->supportsEfficientVectorElementLoadStore(); 391 } 392 393 bool TargetTransformInfo::enableAggressiveInterleaving(bool LoopHasReductions) const { 394 return TTIImpl->enableAggressiveInterleaving(LoopHasReductions); 395 } 396 397 TargetTransformInfo::MemCmpExpansionOptions 398 TargetTransformInfo::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const { 399 return TTIImpl->enableMemCmpExpansion(OptSize, IsZeroCmp); 400 } 401 402 bool TargetTransformInfo::enableInterleavedAccessVectorization() const { 403 return TTIImpl->enableInterleavedAccessVectorization(); 404 } 405 406 bool TargetTransformInfo::enableMaskedInterleavedAccessVectorization() const { 407 return TTIImpl->enableMaskedInterleavedAccessVectorization(); 408 } 409 410 bool TargetTransformInfo::isFPVectorizationPotentiallyUnsafe() const { 411 return TTIImpl->isFPVectorizationPotentiallyUnsafe(); 412 } 413 414 bool TargetTransformInfo::allowsMisalignedMemoryAccesses(LLVMContext &Context, 415 unsigned BitWidth, 416 unsigned AddressSpace, 417 unsigned Alignment, 418 bool *Fast) const { 419 return TTIImpl->allowsMisalignedMemoryAccesses(Context, BitWidth, AddressSpace, 420 Alignment, Fast); 421 } 422 423 TargetTransformInfo::PopcntSupportKind 424 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit) const { 425 return TTIImpl->getPopcntSupport(IntTyWidthInBit); 426 } 427 428 bool TargetTransformInfo::haveFastSqrt(Type *Ty) const { 429 return TTIImpl->haveFastSqrt(Ty); 430 } 431 432 bool TargetTransformInfo::isFCmpOrdCheaperThanFCmpZero(Type *Ty) const { 433 return TTIImpl->isFCmpOrdCheaperThanFCmpZero(Ty); 434 } 435 436 int TargetTransformInfo::getFPOpCost(Type *Ty) const { 437 int Cost = TTIImpl->getFPOpCost(Ty); 438 assert(Cost >= 0 && "TTI should not produce negative costs!"); 439 return Cost; 440 } 441 442 int TargetTransformInfo::getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, 443 const APInt &Imm, 444 Type *Ty) const { 445 int Cost = TTIImpl->getIntImmCodeSizeCost(Opcode, Idx, Imm, Ty); 446 assert(Cost >= 0 && "TTI should not produce negative costs!"); 447 return Cost; 448 } 449 450 int TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty) const { 451 int Cost = TTIImpl->getIntImmCost(Imm, Ty); 452 assert(Cost >= 0 && "TTI should not produce negative costs!"); 453 return Cost; 454 } 455 456 int TargetTransformInfo::getIntImmCostInst(unsigned Opcode, unsigned Idx, 457 const APInt &Imm, Type *Ty) const { 458 int Cost = TTIImpl->getIntImmCostInst(Opcode, Idx, Imm, Ty); 459 assert(Cost >= 0 && "TTI should not produce negative costs!"); 460 return Cost; 461 } 462 463 int TargetTransformInfo::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx, 464 const APInt &Imm, Type *Ty) const { 465 int Cost = TTIImpl->getIntImmCostIntrin(IID, Idx, Imm, Ty); 466 assert(Cost >= 0 && "TTI should not produce negative costs!"); 467 return Cost; 468 } 469 470 unsigned TargetTransformInfo::getNumberOfRegisters(unsigned ClassID) const { 471 return TTIImpl->getNumberOfRegisters(ClassID); 472 } 473 474 unsigned TargetTransformInfo::getRegisterClassForType(bool Vector, Type *Ty) const { 475 return TTIImpl->getRegisterClassForType(Vector, Ty); 476 } 477 478 const char* TargetTransformInfo::getRegisterClassName(unsigned ClassID) const { 479 return TTIImpl->getRegisterClassName(ClassID); 480 } 481 482 unsigned TargetTransformInfo::getRegisterBitWidth(bool Vector) const { 483 return TTIImpl->getRegisterBitWidth(Vector); 484 } 485 486 unsigned TargetTransformInfo::getMinVectorRegisterBitWidth() const { 487 return TTIImpl->getMinVectorRegisterBitWidth(); 488 } 489 490 bool TargetTransformInfo::shouldMaximizeVectorBandwidth(bool OptSize) const { 491 return TTIImpl->shouldMaximizeVectorBandwidth(OptSize); 492 } 493 494 unsigned TargetTransformInfo::getMinimumVF(unsigned ElemWidth) const { 495 return TTIImpl->getMinimumVF(ElemWidth); 496 } 497 498 bool TargetTransformInfo::shouldConsiderAddressTypePromotion( 499 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const { 500 return TTIImpl->shouldConsiderAddressTypePromotion( 501 I, AllowPromotionWithoutCommonHeader); 502 } 503 504 unsigned TargetTransformInfo::getCacheLineSize() const { 505 return TTIImpl->getCacheLineSize(); 506 } 507 508 llvm::Optional<unsigned> TargetTransformInfo::getCacheSize(CacheLevel Level) 509 const { 510 return TTIImpl->getCacheSize(Level); 511 } 512 513 llvm::Optional<unsigned> TargetTransformInfo::getCacheAssociativity( 514 CacheLevel Level) const { 515 return TTIImpl->getCacheAssociativity(Level); 516 } 517 518 unsigned TargetTransformInfo::getPrefetchDistance() const { 519 return TTIImpl->getPrefetchDistance(); 520 } 521 522 unsigned TargetTransformInfo::getMinPrefetchStride() const { 523 return TTIImpl->getMinPrefetchStride(); 524 } 525 526 unsigned TargetTransformInfo::getMaxPrefetchIterationsAhead() const { 527 return TTIImpl->getMaxPrefetchIterationsAhead(); 528 } 529 530 unsigned TargetTransformInfo::getMaxInterleaveFactor(unsigned VF) const { 531 return TTIImpl->getMaxInterleaveFactor(VF); 532 } 533 534 TargetTransformInfo::OperandValueKind 535 TargetTransformInfo::getOperandInfo(Value *V, OperandValueProperties &OpProps) { 536 OperandValueKind OpInfo = OK_AnyValue; 537 OpProps = OP_None; 538 539 if (auto *CI = dyn_cast<ConstantInt>(V)) { 540 if (CI->getValue().isPowerOf2()) 541 OpProps = OP_PowerOf2; 542 return OK_UniformConstantValue; 543 } 544 545 // A broadcast shuffle creates a uniform value. 546 // TODO: Add support for non-zero index broadcasts. 547 // TODO: Add support for different source vector width. 548 if (auto *ShuffleInst = dyn_cast<ShuffleVectorInst>(V)) 549 if (ShuffleInst->isZeroEltSplat()) 550 OpInfo = OK_UniformValue; 551 552 const Value *Splat = getSplatValue(V); 553 554 // Check for a splat of a constant or for a non uniform vector of constants 555 // and check if the constant(s) are all powers of two. 556 if (isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) { 557 OpInfo = OK_NonUniformConstantValue; 558 if (Splat) { 559 OpInfo = OK_UniformConstantValue; 560 if (auto *CI = dyn_cast<ConstantInt>(Splat)) 561 if (CI->getValue().isPowerOf2()) 562 OpProps = OP_PowerOf2; 563 } else if (auto *CDS = dyn_cast<ConstantDataSequential>(V)) { 564 OpProps = OP_PowerOf2; 565 for (unsigned I = 0, E = CDS->getNumElements(); I != E; ++I) { 566 if (auto *CI = dyn_cast<ConstantInt>(CDS->getElementAsConstant(I))) 567 if (CI->getValue().isPowerOf2()) 568 continue; 569 OpProps = OP_None; 570 break; 571 } 572 } 573 } 574 575 // Check for a splat of a uniform value. This is not loop aware, so return 576 // true only for the obviously uniform cases (argument, globalvalue) 577 if (Splat && (isa<Argument>(Splat) || isa<GlobalValue>(Splat))) 578 OpInfo = OK_UniformValue; 579 580 return OpInfo; 581 } 582 583 int TargetTransformInfo::getArithmeticInstrCost( 584 unsigned Opcode, Type *Ty, OperandValueKind Opd1Info, 585 OperandValueKind Opd2Info, OperandValueProperties Opd1PropInfo, 586 OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args, 587 const Instruction *CxtI) const { 588 int Cost = TTIImpl->getArithmeticInstrCost( 589 Opcode, Ty, Opd1Info, Opd2Info, Opd1PropInfo, Opd2PropInfo, Args, CxtI); 590 assert(Cost >= 0 && "TTI should not produce negative costs!"); 591 return Cost; 592 } 593 594 int TargetTransformInfo::getShuffleCost(ShuffleKind Kind, Type *Ty, int Index, 595 Type *SubTp) const { 596 int Cost = TTIImpl->getShuffleCost(Kind, Ty, Index, SubTp); 597 assert(Cost >= 0 && "TTI should not produce negative costs!"); 598 return Cost; 599 } 600 601 int TargetTransformInfo::getCastInstrCost(unsigned Opcode, Type *Dst, 602 Type *Src, const Instruction *I) const { 603 assert ((I == nullptr || I->getOpcode() == Opcode) && 604 "Opcode should reflect passed instruction."); 605 int Cost = TTIImpl->getCastInstrCost(Opcode, Dst, Src, I); 606 assert(Cost >= 0 && "TTI should not produce negative costs!"); 607 return Cost; 608 } 609 610 int TargetTransformInfo::getExtractWithExtendCost(unsigned Opcode, Type *Dst, 611 VectorType *VecTy, 612 unsigned Index) const { 613 int Cost = TTIImpl->getExtractWithExtendCost(Opcode, Dst, VecTy, Index); 614 assert(Cost >= 0 && "TTI should not produce negative costs!"); 615 return Cost; 616 } 617 618 int TargetTransformInfo::getCFInstrCost(unsigned Opcode) const { 619 int Cost = TTIImpl->getCFInstrCost(Opcode); 620 assert(Cost >= 0 && "TTI should not produce negative costs!"); 621 return Cost; 622 } 623 624 int TargetTransformInfo::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 625 Type *CondTy, const Instruction *I) const { 626 assert ((I == nullptr || I->getOpcode() == Opcode) && 627 "Opcode should reflect passed instruction."); 628 int Cost = TTIImpl->getCmpSelInstrCost(Opcode, ValTy, CondTy, I); 629 assert(Cost >= 0 && "TTI should not produce negative costs!"); 630 return Cost; 631 } 632 633 int TargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val, 634 unsigned Index) const { 635 int Cost = TTIImpl->getVectorInstrCost(Opcode, Val, Index); 636 assert(Cost >= 0 && "TTI should not produce negative costs!"); 637 return Cost; 638 } 639 640 int TargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src, 641 MaybeAlign Alignment, 642 unsigned AddressSpace, 643 const Instruction *I) const { 644 assert ((I == nullptr || I->getOpcode() == Opcode) && 645 "Opcode should reflect passed instruction."); 646 int Cost = TTIImpl->getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, I); 647 assert(Cost >= 0 && "TTI should not produce negative costs!"); 648 return Cost; 649 } 650 651 int TargetTransformInfo::getMaskedMemoryOpCost(unsigned Opcode, Type *Src, 652 unsigned Alignment, 653 unsigned AddressSpace) const { 654 int Cost = 655 TTIImpl->getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace); 656 assert(Cost >= 0 && "TTI should not produce negative costs!"); 657 return Cost; 658 } 659 660 int TargetTransformInfo::getGatherScatterOpCost(unsigned Opcode, Type *DataTy, 661 Value *Ptr, bool VariableMask, 662 unsigned Alignment, 663 const Instruction *I) const { 664 int Cost = TTIImpl->getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask, 665 Alignment, I); 666 assert(Cost >= 0 && "TTI should not produce negative costs!"); 667 return Cost; 668 } 669 670 int TargetTransformInfo::getInterleavedMemoryOpCost( 671 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices, 672 unsigned Alignment, unsigned AddressSpace, bool UseMaskForCond, 673 bool UseMaskForGaps) const { 674 int Cost = TTIImpl->getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 675 Alignment, AddressSpace, 676 UseMaskForCond, 677 UseMaskForGaps); 678 assert(Cost >= 0 && "TTI should not produce negative costs!"); 679 return Cost; 680 } 681 682 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 683 ArrayRef<Type *> Tys, 684 FastMathFlags FMF, 685 unsigned ScalarizationCostPassed, 686 const Instruction *I) const { 687 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Tys, FMF, 688 ScalarizationCostPassed, I); 689 assert(Cost >= 0 && "TTI should not produce negative costs!"); 690 return Cost; 691 } 692 693 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 694 ArrayRef<Value *> Args, 695 FastMathFlags FMF, unsigned VF, 696 const Instruction *I) const { 697 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF, I); 698 assert(Cost >= 0 && "TTI should not produce negative costs!"); 699 return Cost; 700 } 701 702 int TargetTransformInfo::getCallInstrCost(Function *F, Type *RetTy, 703 ArrayRef<Type *> Tys) const { 704 int Cost = TTIImpl->getCallInstrCost(F, RetTy, Tys); 705 assert(Cost >= 0 && "TTI should not produce negative costs!"); 706 return Cost; 707 } 708 709 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const { 710 return TTIImpl->getNumberOfParts(Tp); 711 } 712 713 int TargetTransformInfo::getAddressComputationCost(Type *Tp, 714 ScalarEvolution *SE, 715 const SCEV *Ptr) const { 716 int Cost = TTIImpl->getAddressComputationCost(Tp, SE, Ptr); 717 assert(Cost >= 0 && "TTI should not produce negative costs!"); 718 return Cost; 719 } 720 721 int TargetTransformInfo::getMemcpyCost(const Instruction *I) const { 722 int Cost = TTIImpl->getMemcpyCost(I); 723 assert(Cost >= 0 && "TTI should not produce negative costs!"); 724 return Cost; 725 } 726 727 int TargetTransformInfo::getArithmeticReductionCost(unsigned Opcode, Type *Ty, 728 bool IsPairwiseForm) const { 729 int Cost = TTIImpl->getArithmeticReductionCost(Opcode, Ty, IsPairwiseForm); 730 assert(Cost >= 0 && "TTI should not produce negative costs!"); 731 return Cost; 732 } 733 734 int TargetTransformInfo::getMinMaxReductionCost(Type *Ty, Type *CondTy, 735 bool IsPairwiseForm, 736 bool IsUnsigned) const { 737 int Cost = 738 TTIImpl->getMinMaxReductionCost(Ty, CondTy, IsPairwiseForm, IsUnsigned); 739 assert(Cost >= 0 && "TTI should not produce negative costs!"); 740 return Cost; 741 } 742 743 unsigned 744 TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const { 745 return TTIImpl->getCostOfKeepingLiveOverCall(Tys); 746 } 747 748 bool TargetTransformInfo::getTgtMemIntrinsic(IntrinsicInst *Inst, 749 MemIntrinsicInfo &Info) const { 750 return TTIImpl->getTgtMemIntrinsic(Inst, Info); 751 } 752 753 unsigned TargetTransformInfo::getAtomicMemIntrinsicMaxElementSize() const { 754 return TTIImpl->getAtomicMemIntrinsicMaxElementSize(); 755 } 756 757 Value *TargetTransformInfo::getOrCreateResultFromMemIntrinsic( 758 IntrinsicInst *Inst, Type *ExpectedType) const { 759 return TTIImpl->getOrCreateResultFromMemIntrinsic(Inst, ExpectedType); 760 } 761 762 Type *TargetTransformInfo::getMemcpyLoopLoweringType(LLVMContext &Context, 763 Value *Length, 764 unsigned SrcAddrSpace, 765 unsigned DestAddrSpace, 766 unsigned SrcAlign, 767 unsigned DestAlign) const { 768 return TTIImpl->getMemcpyLoopLoweringType(Context, Length, SrcAddrSpace, 769 DestAddrSpace, SrcAlign, 770 DestAlign); 771 } 772 773 void TargetTransformInfo::getMemcpyLoopResidualLoweringType( 774 SmallVectorImpl<Type *> &OpsOut, LLVMContext &Context, 775 unsigned RemainingBytes, 776 unsigned SrcAddrSpace, 777 unsigned DestAddrSpace, 778 unsigned SrcAlign, unsigned DestAlign) const { 779 TTIImpl->getMemcpyLoopResidualLoweringType(OpsOut, Context, RemainingBytes, 780 SrcAddrSpace, DestAddrSpace, 781 SrcAlign, DestAlign); 782 } 783 784 bool TargetTransformInfo::areInlineCompatible(const Function *Caller, 785 const Function *Callee) const { 786 return TTIImpl->areInlineCompatible(Caller, Callee); 787 } 788 789 bool TargetTransformInfo::areFunctionArgsABICompatible( 790 const Function *Caller, const Function *Callee, 791 SmallPtrSetImpl<Argument *> &Args) const { 792 return TTIImpl->areFunctionArgsABICompatible(Caller, Callee, Args); 793 } 794 795 bool TargetTransformInfo::isIndexedLoadLegal(MemIndexedMode Mode, 796 Type *Ty) const { 797 return TTIImpl->isIndexedLoadLegal(Mode, Ty); 798 } 799 800 bool TargetTransformInfo::isIndexedStoreLegal(MemIndexedMode Mode, 801 Type *Ty) const { 802 return TTIImpl->isIndexedStoreLegal(Mode, Ty); 803 } 804 805 unsigned TargetTransformInfo::getLoadStoreVecRegBitWidth(unsigned AS) const { 806 return TTIImpl->getLoadStoreVecRegBitWidth(AS); 807 } 808 809 bool TargetTransformInfo::isLegalToVectorizeLoad(LoadInst *LI) const { 810 return TTIImpl->isLegalToVectorizeLoad(LI); 811 } 812 813 bool TargetTransformInfo::isLegalToVectorizeStore(StoreInst *SI) const { 814 return TTIImpl->isLegalToVectorizeStore(SI); 815 } 816 817 bool TargetTransformInfo::isLegalToVectorizeLoadChain( 818 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 819 return TTIImpl->isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment, 820 AddrSpace); 821 } 822 823 bool TargetTransformInfo::isLegalToVectorizeStoreChain( 824 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 825 return TTIImpl->isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment, 826 AddrSpace); 827 } 828 829 unsigned TargetTransformInfo::getLoadVectorFactor(unsigned VF, 830 unsigned LoadSize, 831 unsigned ChainSizeInBytes, 832 VectorType *VecTy) const { 833 return TTIImpl->getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy); 834 } 835 836 unsigned TargetTransformInfo::getStoreVectorFactor(unsigned VF, 837 unsigned StoreSize, 838 unsigned ChainSizeInBytes, 839 VectorType *VecTy) const { 840 return TTIImpl->getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy); 841 } 842 843 bool TargetTransformInfo::useReductionIntrinsic(unsigned Opcode, 844 Type *Ty, ReductionFlags Flags) const { 845 return TTIImpl->useReductionIntrinsic(Opcode, Ty, Flags); 846 } 847 848 bool TargetTransformInfo::shouldExpandReduction(const IntrinsicInst *II) const { 849 return TTIImpl->shouldExpandReduction(II); 850 } 851 852 unsigned TargetTransformInfo::getGISelRematGlobalCost() const { 853 return TTIImpl->getGISelRematGlobalCost(); 854 } 855 856 int TargetTransformInfo::getInstructionLatency(const Instruction *I) const { 857 return TTIImpl->getInstructionLatency(I); 858 } 859 860 static bool matchPairwiseShuffleMask(ShuffleVectorInst *SI, bool IsLeft, 861 unsigned Level) { 862 // We don't need a shuffle if we just want to have element 0 in position 0 of 863 // the vector. 864 if (!SI && Level == 0 && IsLeft) 865 return true; 866 else if (!SI) 867 return false; 868 869 SmallVector<int, 32> Mask(SI->getType()->getVectorNumElements(), -1); 870 871 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right) depending on whether 872 // we look at the left or right side. 873 for (unsigned i = 0, e = (1 << Level), val = !IsLeft; i != e; ++i, val += 2) 874 Mask[i] = val; 875 876 ArrayRef<int> ActualMask = SI->getShuffleMask(); 877 return Mask == ActualMask; 878 } 879 880 namespace { 881 /// Kind of the reduction data. 882 enum ReductionKind { 883 RK_None, /// Not a reduction. 884 RK_Arithmetic, /// Binary reduction data. 885 RK_MinMax, /// Min/max reduction data. 886 RK_UnsignedMinMax, /// Unsigned min/max reduction data. 887 }; 888 /// Contains opcode + LHS/RHS parts of the reduction operations. 889 struct ReductionData { 890 ReductionData() = delete; 891 ReductionData(ReductionKind Kind, unsigned Opcode, Value *LHS, Value *RHS) 892 : Opcode(Opcode), LHS(LHS), RHS(RHS), Kind(Kind) { 893 assert(Kind != RK_None && "expected binary or min/max reduction only."); 894 } 895 unsigned Opcode = 0; 896 Value *LHS = nullptr; 897 Value *RHS = nullptr; 898 ReductionKind Kind = RK_None; 899 bool hasSameData(ReductionData &RD) const { 900 return Kind == RD.Kind && Opcode == RD.Opcode; 901 } 902 }; 903 } // namespace 904 905 static Optional<ReductionData> getReductionData(Instruction *I) { 906 Value *L, *R; 907 if (m_BinOp(m_Value(L), m_Value(R)).match(I)) 908 return ReductionData(RK_Arithmetic, I->getOpcode(), L, R); 909 if (auto *SI = dyn_cast<SelectInst>(I)) { 910 if (m_SMin(m_Value(L), m_Value(R)).match(SI) || 911 m_SMax(m_Value(L), m_Value(R)).match(SI) || 912 m_OrdFMin(m_Value(L), m_Value(R)).match(SI) || 913 m_OrdFMax(m_Value(L), m_Value(R)).match(SI) || 914 m_UnordFMin(m_Value(L), m_Value(R)).match(SI) || 915 m_UnordFMax(m_Value(L), m_Value(R)).match(SI)) { 916 auto *CI = cast<CmpInst>(SI->getCondition()); 917 return ReductionData(RK_MinMax, CI->getOpcode(), L, R); 918 } 919 if (m_UMin(m_Value(L), m_Value(R)).match(SI) || 920 m_UMax(m_Value(L), m_Value(R)).match(SI)) { 921 auto *CI = cast<CmpInst>(SI->getCondition()); 922 return ReductionData(RK_UnsignedMinMax, CI->getOpcode(), L, R); 923 } 924 } 925 return llvm::None; 926 } 927 928 static ReductionKind matchPairwiseReductionAtLevel(Instruction *I, 929 unsigned Level, 930 unsigned NumLevels) { 931 // Match one level of pairwise operations. 932 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 933 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 934 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 935 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 936 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 937 if (!I) 938 return RK_None; 939 940 assert(I->getType()->isVectorTy() && "Expecting a vector type"); 941 942 Optional<ReductionData> RD = getReductionData(I); 943 if (!RD) 944 return RK_None; 945 946 ShuffleVectorInst *LS = dyn_cast<ShuffleVectorInst>(RD->LHS); 947 if (!LS && Level) 948 return RK_None; 949 ShuffleVectorInst *RS = dyn_cast<ShuffleVectorInst>(RD->RHS); 950 if (!RS && Level) 951 return RK_None; 952 953 // On level 0 we can omit one shufflevector instruction. 954 if (!Level && !RS && !LS) 955 return RK_None; 956 957 // Shuffle inputs must match. 958 Value *NextLevelOpL = LS ? LS->getOperand(0) : nullptr; 959 Value *NextLevelOpR = RS ? RS->getOperand(0) : nullptr; 960 Value *NextLevelOp = nullptr; 961 if (NextLevelOpR && NextLevelOpL) { 962 // If we have two shuffles their operands must match. 963 if (NextLevelOpL != NextLevelOpR) 964 return RK_None; 965 966 NextLevelOp = NextLevelOpL; 967 } else if (Level == 0 && (NextLevelOpR || NextLevelOpL)) { 968 // On the first level we can omit the shufflevector <0, undef,...>. So the 969 // input to the other shufflevector <1, undef> must match with one of the 970 // inputs to the current binary operation. 971 // Example: 972 // %NextLevelOpL = shufflevector %R, <1, undef ...> 973 // %BinOp = fadd %NextLevelOpL, %R 974 if (NextLevelOpL && NextLevelOpL != RD->RHS) 975 return RK_None; 976 else if (NextLevelOpR && NextLevelOpR != RD->LHS) 977 return RK_None; 978 979 NextLevelOp = NextLevelOpL ? RD->RHS : RD->LHS; 980 } else 981 return RK_None; 982 983 // Check that the next levels binary operation exists and matches with the 984 // current one. 985 if (Level + 1 != NumLevels) { 986 Optional<ReductionData> NextLevelRD = 987 getReductionData(cast<Instruction>(NextLevelOp)); 988 if (!NextLevelRD || !RD->hasSameData(*NextLevelRD)) 989 return RK_None; 990 } 991 992 // Shuffle mask for pairwise operation must match. 993 if (matchPairwiseShuffleMask(LS, /*IsLeft=*/true, Level)) { 994 if (!matchPairwiseShuffleMask(RS, /*IsLeft=*/false, Level)) 995 return RK_None; 996 } else if (matchPairwiseShuffleMask(RS, /*IsLeft=*/true, Level)) { 997 if (!matchPairwiseShuffleMask(LS, /*IsLeft=*/false, Level)) 998 return RK_None; 999 } else { 1000 return RK_None; 1001 } 1002 1003 if (++Level == NumLevels) 1004 return RD->Kind; 1005 1006 // Match next level. 1007 return matchPairwiseReductionAtLevel(cast<Instruction>(NextLevelOp), Level, 1008 NumLevels); 1009 } 1010 1011 static ReductionKind matchPairwiseReduction(const ExtractElementInst *ReduxRoot, 1012 unsigned &Opcode, Type *&Ty) { 1013 if (!EnableReduxCost) 1014 return RK_None; 1015 1016 // Need to extract the first element. 1017 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 1018 unsigned Idx = ~0u; 1019 if (CI) 1020 Idx = CI->getZExtValue(); 1021 if (Idx != 0) 1022 return RK_None; 1023 1024 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 1025 if (!RdxStart) 1026 return RK_None; 1027 Optional<ReductionData> RD = getReductionData(RdxStart); 1028 if (!RD) 1029 return RK_None; 1030 1031 Type *VecTy = RdxStart->getType(); 1032 unsigned NumVecElems = VecTy->getVectorNumElements(); 1033 if (!isPowerOf2_32(NumVecElems)) 1034 return RK_None; 1035 1036 // We look for a sequence of shuffle,shuffle,add triples like the following 1037 // that builds a pairwise reduction tree. 1038 // 1039 // (X0, X1, X2, X3) 1040 // (X0 + X1, X2 + X3, undef, undef) 1041 // ((X0 + X1) + (X2 + X3), undef, undef, undef) 1042 // 1043 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 1044 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 1045 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 1046 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 1047 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 1048 // %rdx.shuf.1.0 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 1049 // <4 x i32> <i32 0, i32 undef, i32 undef, i32 undef> 1050 // %rdx.shuf.1.1 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 1051 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 1052 // %bin.rdx8 = fadd <4 x float> %rdx.shuf.1.0, %rdx.shuf.1.1 1053 // %r = extractelement <4 x float> %bin.rdx8, i32 0 1054 if (matchPairwiseReductionAtLevel(RdxStart, 0, Log2_32(NumVecElems)) == 1055 RK_None) 1056 return RK_None; 1057 1058 Opcode = RD->Opcode; 1059 Ty = VecTy; 1060 1061 return RD->Kind; 1062 } 1063 1064 static std::pair<Value *, ShuffleVectorInst *> 1065 getShuffleAndOtherOprd(Value *L, Value *R) { 1066 ShuffleVectorInst *S = nullptr; 1067 1068 if ((S = dyn_cast<ShuffleVectorInst>(L))) 1069 return std::make_pair(R, S); 1070 1071 S = dyn_cast<ShuffleVectorInst>(R); 1072 return std::make_pair(L, S); 1073 } 1074 1075 static ReductionKind 1076 matchVectorSplittingReduction(const ExtractElementInst *ReduxRoot, 1077 unsigned &Opcode, Type *&Ty) { 1078 if (!EnableReduxCost) 1079 return RK_None; 1080 1081 // Need to extract the first element. 1082 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 1083 unsigned Idx = ~0u; 1084 if (CI) 1085 Idx = CI->getZExtValue(); 1086 if (Idx != 0) 1087 return RK_None; 1088 1089 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 1090 if (!RdxStart) 1091 return RK_None; 1092 Optional<ReductionData> RD = getReductionData(RdxStart); 1093 if (!RD) 1094 return RK_None; 1095 1096 Type *VecTy = ReduxRoot->getOperand(0)->getType(); 1097 unsigned NumVecElems = VecTy->getVectorNumElements(); 1098 if (!isPowerOf2_32(NumVecElems)) 1099 return RK_None; 1100 1101 // We look for a sequence of shuffles and adds like the following matching one 1102 // fadd, shuffle vector pair at a time. 1103 // 1104 // %rdx.shuf = shufflevector <4 x float> %rdx, <4 x float> undef, 1105 // <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 1106 // %bin.rdx = fadd <4 x float> %rdx, %rdx.shuf 1107 // %rdx.shuf7 = shufflevector <4 x float> %bin.rdx, <4 x float> undef, 1108 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 1109 // %bin.rdx8 = fadd <4 x float> %bin.rdx, %rdx.shuf7 1110 // %r = extractelement <4 x float> %bin.rdx8, i32 0 1111 1112 unsigned MaskStart = 1; 1113 Instruction *RdxOp = RdxStart; 1114 SmallVector<int, 32> ShuffleMask(NumVecElems, 0); 1115 unsigned NumVecElemsRemain = NumVecElems; 1116 while (NumVecElemsRemain - 1) { 1117 // Check for the right reduction operation. 1118 if (!RdxOp) 1119 return RK_None; 1120 Optional<ReductionData> RDLevel = getReductionData(RdxOp); 1121 if (!RDLevel || !RDLevel->hasSameData(*RD)) 1122 return RK_None; 1123 1124 Value *NextRdxOp; 1125 ShuffleVectorInst *Shuffle; 1126 std::tie(NextRdxOp, Shuffle) = 1127 getShuffleAndOtherOprd(RDLevel->LHS, RDLevel->RHS); 1128 1129 // Check the current reduction operation and the shuffle use the same value. 1130 if (Shuffle == nullptr) 1131 return RK_None; 1132 if (Shuffle->getOperand(0) != NextRdxOp) 1133 return RK_None; 1134 1135 // Check that shuffle masks matches. 1136 for (unsigned j = 0; j != MaskStart; ++j) 1137 ShuffleMask[j] = MaskStart + j; 1138 // Fill the rest of the mask with -1 for undef. 1139 std::fill(&ShuffleMask[MaskStart], ShuffleMask.end(), -1); 1140 1141 ArrayRef<int> Mask = Shuffle->getShuffleMask(); 1142 if (ShuffleMask != Mask) 1143 return RK_None; 1144 1145 RdxOp = dyn_cast<Instruction>(NextRdxOp); 1146 NumVecElemsRemain /= 2; 1147 MaskStart *= 2; 1148 } 1149 1150 Opcode = RD->Opcode; 1151 Ty = VecTy; 1152 return RD->Kind; 1153 } 1154 1155 int TargetTransformInfo::getInstructionThroughput(const Instruction *I) const { 1156 switch (I->getOpcode()) { 1157 case Instruction::GetElementPtr: 1158 return getUserCost(I); 1159 1160 case Instruction::Ret: 1161 case Instruction::PHI: 1162 case Instruction::Br: { 1163 return getCFInstrCost(I->getOpcode()); 1164 } 1165 case Instruction::Add: 1166 case Instruction::FAdd: 1167 case Instruction::Sub: 1168 case Instruction::FSub: 1169 case Instruction::Mul: 1170 case Instruction::FMul: 1171 case Instruction::UDiv: 1172 case Instruction::SDiv: 1173 case Instruction::FDiv: 1174 case Instruction::URem: 1175 case Instruction::SRem: 1176 case Instruction::FRem: 1177 case Instruction::Shl: 1178 case Instruction::LShr: 1179 case Instruction::AShr: 1180 case Instruction::And: 1181 case Instruction::Or: 1182 case Instruction::Xor: { 1183 TargetTransformInfo::OperandValueKind Op1VK, Op2VK; 1184 TargetTransformInfo::OperandValueProperties Op1VP, Op2VP; 1185 Op1VK = getOperandInfo(I->getOperand(0), Op1VP); 1186 Op2VK = getOperandInfo(I->getOperand(1), Op2VP); 1187 SmallVector<const Value *, 2> Operands(I->operand_values()); 1188 return getArithmeticInstrCost(I->getOpcode(), I->getType(), Op1VK, Op2VK, 1189 Op1VP, Op2VP, Operands, I); 1190 } 1191 case Instruction::FNeg: { 1192 TargetTransformInfo::OperandValueKind Op1VK, Op2VK; 1193 TargetTransformInfo::OperandValueProperties Op1VP, Op2VP; 1194 Op1VK = getOperandInfo(I->getOperand(0), Op1VP); 1195 Op2VK = OK_AnyValue; 1196 Op2VP = OP_None; 1197 SmallVector<const Value *, 2> Operands(I->operand_values()); 1198 return getArithmeticInstrCost(I->getOpcode(), I->getType(), Op1VK, Op2VK, 1199 Op1VP, Op2VP, Operands, I); 1200 } 1201 case Instruction::Select: { 1202 const SelectInst *SI = cast<SelectInst>(I); 1203 Type *CondTy = SI->getCondition()->getType(); 1204 return getCmpSelInstrCost(I->getOpcode(), I->getType(), CondTy, I); 1205 } 1206 case Instruction::ICmp: 1207 case Instruction::FCmp: { 1208 Type *ValTy = I->getOperand(0)->getType(); 1209 return getCmpSelInstrCost(I->getOpcode(), ValTy, I->getType(), I); 1210 } 1211 case Instruction::Store: { 1212 const StoreInst *SI = cast<StoreInst>(I); 1213 Type *ValTy = SI->getValueOperand()->getType(); 1214 return getMemoryOpCost(I->getOpcode(), ValTy, 1215 MaybeAlign(SI->getAlignment()), 1216 SI->getPointerAddressSpace(), I); 1217 } 1218 case Instruction::Load: { 1219 const LoadInst *LI = cast<LoadInst>(I); 1220 return getMemoryOpCost(I->getOpcode(), I->getType(), 1221 MaybeAlign(LI->getAlignment()), 1222 LI->getPointerAddressSpace(), I); 1223 } 1224 case Instruction::ZExt: 1225 case Instruction::SExt: 1226 case Instruction::FPToUI: 1227 case Instruction::FPToSI: 1228 case Instruction::FPExt: 1229 case Instruction::PtrToInt: 1230 case Instruction::IntToPtr: 1231 case Instruction::SIToFP: 1232 case Instruction::UIToFP: 1233 case Instruction::Trunc: 1234 case Instruction::FPTrunc: 1235 case Instruction::BitCast: 1236 case Instruction::AddrSpaceCast: { 1237 Type *SrcTy = I->getOperand(0)->getType(); 1238 return getCastInstrCost(I->getOpcode(), I->getType(), SrcTy, I); 1239 } 1240 case Instruction::ExtractElement: { 1241 const ExtractElementInst * EEI = cast<ExtractElementInst>(I); 1242 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1243 unsigned Idx = -1; 1244 if (CI) 1245 Idx = CI->getZExtValue(); 1246 1247 // Try to match a reduction sequence (series of shufflevector and vector 1248 // adds followed by a extractelement). 1249 unsigned ReduxOpCode; 1250 Type *ReduxType; 1251 1252 switch (matchVectorSplittingReduction(EEI, ReduxOpCode, ReduxType)) { 1253 case RK_Arithmetic: 1254 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1255 /*IsPairwiseForm=*/false); 1256 case RK_MinMax: 1257 return getMinMaxReductionCost( 1258 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1259 /*IsPairwiseForm=*/false, /*IsUnsigned=*/false); 1260 case RK_UnsignedMinMax: 1261 return getMinMaxReductionCost( 1262 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1263 /*IsPairwiseForm=*/false, /*IsUnsigned=*/true); 1264 case RK_None: 1265 break; 1266 } 1267 1268 switch (matchPairwiseReduction(EEI, ReduxOpCode, ReduxType)) { 1269 case RK_Arithmetic: 1270 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1271 /*IsPairwiseForm=*/true); 1272 case RK_MinMax: 1273 return getMinMaxReductionCost( 1274 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1275 /*IsPairwiseForm=*/true, /*IsUnsigned=*/false); 1276 case RK_UnsignedMinMax: 1277 return getMinMaxReductionCost( 1278 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1279 /*IsPairwiseForm=*/true, /*IsUnsigned=*/true); 1280 case RK_None: 1281 break; 1282 } 1283 1284 return getVectorInstrCost(I->getOpcode(), 1285 EEI->getOperand(0)->getType(), Idx); 1286 } 1287 case Instruction::InsertElement: { 1288 const InsertElementInst * IE = cast<InsertElementInst>(I); 1289 ConstantInt *CI = dyn_cast<ConstantInt>(IE->getOperand(2)); 1290 unsigned Idx = -1; 1291 if (CI) 1292 Idx = CI->getZExtValue(); 1293 return getVectorInstrCost(I->getOpcode(), 1294 IE->getType(), Idx); 1295 } 1296 case Instruction::ExtractValue: 1297 return 0; // Model all ExtractValue nodes as free. 1298 case Instruction::ShuffleVector: { 1299 const ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I); 1300 Type *Ty = Shuffle->getType(); 1301 Type *SrcTy = Shuffle->getOperand(0)->getType(); 1302 1303 // TODO: Identify and add costs for insert subvector, etc. 1304 int SubIndex; 1305 if (Shuffle->isExtractSubvectorMask(SubIndex)) 1306 return TTIImpl->getShuffleCost(SK_ExtractSubvector, SrcTy, SubIndex, Ty); 1307 1308 if (Shuffle->changesLength()) 1309 return -1; 1310 1311 if (Shuffle->isIdentity()) 1312 return 0; 1313 1314 if (Shuffle->isReverse()) 1315 return TTIImpl->getShuffleCost(SK_Reverse, Ty, 0, nullptr); 1316 1317 if (Shuffle->isSelect()) 1318 return TTIImpl->getShuffleCost(SK_Select, Ty, 0, nullptr); 1319 1320 if (Shuffle->isTranspose()) 1321 return TTIImpl->getShuffleCost(SK_Transpose, Ty, 0, nullptr); 1322 1323 if (Shuffle->isZeroEltSplat()) 1324 return TTIImpl->getShuffleCost(SK_Broadcast, Ty, 0, nullptr); 1325 1326 if (Shuffle->isSingleSource()) 1327 return TTIImpl->getShuffleCost(SK_PermuteSingleSrc, Ty, 0, nullptr); 1328 1329 return TTIImpl->getShuffleCost(SK_PermuteTwoSrc, Ty, 0, nullptr); 1330 } 1331 case Instruction::Call: 1332 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1333 SmallVector<Value *, 4> Args(II->arg_operands()); 1334 1335 FastMathFlags FMF; 1336 if (auto *FPMO = dyn_cast<FPMathOperator>(II)) 1337 FMF = FPMO->getFastMathFlags(); 1338 1339 return getIntrinsicInstrCost(II->getIntrinsicID(), II->getType(), Args, 1340 FMF, 1, II); 1341 } 1342 return -1; 1343 default: 1344 // We don't have any information on this instruction. 1345 return -1; 1346 } 1347 } 1348 1349 TargetTransformInfo::Concept::~Concept() {} 1350 1351 TargetIRAnalysis::TargetIRAnalysis() : TTICallback(&getDefaultTTI) {} 1352 1353 TargetIRAnalysis::TargetIRAnalysis( 1354 std::function<Result(const Function &)> TTICallback) 1355 : TTICallback(std::move(TTICallback)) {} 1356 1357 TargetIRAnalysis::Result TargetIRAnalysis::run(const Function &F, 1358 FunctionAnalysisManager &) { 1359 return TTICallback(F); 1360 } 1361 1362 AnalysisKey TargetIRAnalysis::Key; 1363 1364 TargetIRAnalysis::Result TargetIRAnalysis::getDefaultTTI(const Function &F) { 1365 return Result(F.getParent()->getDataLayout()); 1366 } 1367 1368 // Register the basic pass. 1369 INITIALIZE_PASS(TargetTransformInfoWrapperPass, "tti", 1370 "Target Transform Information", false, true) 1371 char TargetTransformInfoWrapperPass::ID = 0; 1372 1373 void TargetTransformInfoWrapperPass::anchor() {} 1374 1375 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass() 1376 : ImmutablePass(ID) { 1377 initializeTargetTransformInfoWrapperPassPass( 1378 *PassRegistry::getPassRegistry()); 1379 } 1380 1381 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass( 1382 TargetIRAnalysis TIRA) 1383 : ImmutablePass(ID), TIRA(std::move(TIRA)) { 1384 initializeTargetTransformInfoWrapperPassPass( 1385 *PassRegistry::getPassRegistry()); 1386 } 1387 1388 TargetTransformInfo &TargetTransformInfoWrapperPass::getTTI(const Function &F) { 1389 FunctionAnalysisManager DummyFAM; 1390 TTI = TIRA.run(F, DummyFAM); 1391 return *TTI; 1392 } 1393 1394 ImmutablePass * 1395 llvm::createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA) { 1396 return new TargetTransformInfoWrapperPass(std::move(TIRA)); 1397 } 1398