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 int TargetTransformInfo::getCallCost(FunctionType *FTy, int NumArgs, 157 const User *U) const { 158 int Cost = TTIImpl->getCallCost(FTy, NumArgs, U); 159 assert(Cost >= 0 && "TTI should not produce negative costs!"); 160 return Cost; 161 } 162 163 int TargetTransformInfo::getCallCost(const Function *F, 164 ArrayRef<const Value *> Arguments, 165 const User *U) const { 166 int Cost = TTIImpl->getCallCost(F, Arguments, U); 167 assert(Cost >= 0 && "TTI should not produce negative costs!"); 168 return Cost; 169 } 170 171 unsigned TargetTransformInfo::getInliningThresholdMultiplier() const { 172 return TTIImpl->getInliningThresholdMultiplier(); 173 } 174 175 int TargetTransformInfo::getInlinerVectorBonusPercent() const { 176 return TTIImpl->getInlinerVectorBonusPercent(); 177 } 178 179 int TargetTransformInfo::getGEPCost(Type *PointeeType, const Value *Ptr, 180 ArrayRef<const Value *> Operands) const { 181 return TTIImpl->getGEPCost(PointeeType, Ptr, Operands); 182 } 183 184 int TargetTransformInfo::getExtCost(const Instruction *I, 185 const Value *Src) const { 186 return TTIImpl->getExtCost(I, Src); 187 } 188 189 int TargetTransformInfo::getIntrinsicCost( 190 Intrinsic::ID IID, Type *RetTy, ArrayRef<const Value *> Arguments, 191 const User *U) const { 192 int Cost = TTIImpl->getIntrinsicCost(IID, RetTy, Arguments, U); 193 assert(Cost >= 0 && "TTI should not produce negative costs!"); 194 return Cost; 195 } 196 197 unsigned 198 TargetTransformInfo::getEstimatedNumberOfCaseClusters( 199 const SwitchInst &SI, unsigned &JTSize, ProfileSummaryInfo *PSI, 200 BlockFrequencyInfo *BFI) const { 201 return TTIImpl->getEstimatedNumberOfCaseClusters(SI, JTSize, PSI, BFI); 202 } 203 204 int TargetTransformInfo::getUserCost(const User *U, 205 ArrayRef<const Value *> Operands) const { 206 int Cost = TTIImpl->getUserCost(U, Operands); 207 assert(Cost >= 0 && "TTI should not produce negative costs!"); 208 return Cost; 209 } 210 211 bool TargetTransformInfo::hasBranchDivergence() const { 212 return TTIImpl->hasBranchDivergence(); 213 } 214 215 bool TargetTransformInfo::useGPUDivergenceAnalysis() const { 216 return TTIImpl->useGPUDivergenceAnalysis(); 217 } 218 219 bool TargetTransformInfo::isSourceOfDivergence(const Value *V) const { 220 return TTIImpl->isSourceOfDivergence(V); 221 } 222 223 bool llvm::TargetTransformInfo::isAlwaysUniform(const Value *V) const { 224 return TTIImpl->isAlwaysUniform(V); 225 } 226 227 unsigned TargetTransformInfo::getFlatAddressSpace() const { 228 return TTIImpl->getFlatAddressSpace(); 229 } 230 231 bool TargetTransformInfo::collectFlatAddressOperands( 232 SmallVectorImpl<int> &OpIndexes, Intrinsic::ID IID) const { 233 return TTIImpl->collectFlatAddressOperands(OpIndexes, IID); 234 } 235 236 bool TargetTransformInfo::rewriteIntrinsicWithAddressSpace( 237 IntrinsicInst *II, Value *OldV, Value *NewV) const { 238 return TTIImpl->rewriteIntrinsicWithAddressSpace(II, OldV, NewV); 239 } 240 241 bool TargetTransformInfo::isLoweredToCall(const Function *F) const { 242 return TTIImpl->isLoweredToCall(F); 243 } 244 245 bool TargetTransformInfo::isHardwareLoopProfitable( 246 Loop *L, ScalarEvolution &SE, AssumptionCache &AC, 247 TargetLibraryInfo *LibInfo, HardwareLoopInfo &HWLoopInfo) const { 248 return TTIImpl->isHardwareLoopProfitable(L, SE, AC, LibInfo, HWLoopInfo); 249 } 250 251 bool TargetTransformInfo::preferPredicateOverEpilogue(Loop *L, LoopInfo *LI, 252 ScalarEvolution &SE, AssumptionCache &AC, TargetLibraryInfo *TLI, 253 DominatorTree *DT, const LoopAccessInfo *LAI) const { 254 return TTIImpl->preferPredicateOverEpilogue(L, LI, SE, AC, TLI, DT, LAI); 255 } 256 257 void TargetTransformInfo::getUnrollingPreferences( 258 Loop *L, ScalarEvolution &SE, UnrollingPreferences &UP) const { 259 return TTIImpl->getUnrollingPreferences(L, SE, UP); 260 } 261 262 bool TargetTransformInfo::isLegalAddImmediate(int64_t Imm) const { 263 return TTIImpl->isLegalAddImmediate(Imm); 264 } 265 266 bool TargetTransformInfo::isLegalICmpImmediate(int64_t Imm) const { 267 return TTIImpl->isLegalICmpImmediate(Imm); 268 } 269 270 bool TargetTransformInfo::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, 271 int64_t BaseOffset, 272 bool HasBaseReg, 273 int64_t Scale, 274 unsigned AddrSpace, 275 Instruction *I) const { 276 return TTIImpl->isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg, 277 Scale, AddrSpace, I); 278 } 279 280 bool TargetTransformInfo::isLSRCostLess(LSRCost &C1, LSRCost &C2) const { 281 return TTIImpl->isLSRCostLess(C1, C2); 282 } 283 284 bool TargetTransformInfo::canMacroFuseCmp() const { 285 return TTIImpl->canMacroFuseCmp(); 286 } 287 288 bool TargetTransformInfo::canSaveCmp(Loop *L, BranchInst **BI, 289 ScalarEvolution *SE, LoopInfo *LI, 290 DominatorTree *DT, AssumptionCache *AC, 291 TargetLibraryInfo *LibInfo) const { 292 return TTIImpl->canSaveCmp(L, BI, SE, LI, DT, AC, LibInfo); 293 } 294 295 bool TargetTransformInfo::shouldFavorPostInc() const { 296 return TTIImpl->shouldFavorPostInc(); 297 } 298 299 bool TargetTransformInfo::shouldFavorBackedgeIndex(const Loop *L) const { 300 return TTIImpl->shouldFavorBackedgeIndex(L); 301 } 302 303 bool TargetTransformInfo::isLegalMaskedStore(Type *DataType, 304 MaybeAlign Alignment) const { 305 return TTIImpl->isLegalMaskedStore(DataType, Alignment); 306 } 307 308 bool TargetTransformInfo::isLegalMaskedLoad(Type *DataType, 309 MaybeAlign Alignment) const { 310 return TTIImpl->isLegalMaskedLoad(DataType, Alignment); 311 } 312 313 bool TargetTransformInfo::isLegalNTStore(Type *DataType, 314 Align Alignment) const { 315 return TTIImpl->isLegalNTStore(DataType, Alignment); 316 } 317 318 bool TargetTransformInfo::isLegalNTLoad(Type *DataType, Align Alignment) const { 319 return TTIImpl->isLegalNTLoad(DataType, Alignment); 320 } 321 322 bool TargetTransformInfo::isLegalMaskedGather(Type *DataType, 323 MaybeAlign Alignment) const { 324 return TTIImpl->isLegalMaskedGather(DataType, Alignment); 325 } 326 327 bool TargetTransformInfo::isLegalMaskedScatter(Type *DataType, 328 MaybeAlign Alignment) const { 329 return TTIImpl->isLegalMaskedScatter(DataType, Alignment); 330 } 331 332 bool TargetTransformInfo::isLegalMaskedCompressStore(Type *DataType) const { 333 return TTIImpl->isLegalMaskedCompressStore(DataType); 334 } 335 336 bool TargetTransformInfo::isLegalMaskedExpandLoad(Type *DataType) const { 337 return TTIImpl->isLegalMaskedExpandLoad(DataType); 338 } 339 340 bool TargetTransformInfo::hasDivRemOp(Type *DataType, bool IsSigned) const { 341 return TTIImpl->hasDivRemOp(DataType, IsSigned); 342 } 343 344 bool TargetTransformInfo::hasVolatileVariant(Instruction *I, 345 unsigned AddrSpace) const { 346 return TTIImpl->hasVolatileVariant(I, AddrSpace); 347 } 348 349 bool TargetTransformInfo::prefersVectorizedAddressing() const { 350 return TTIImpl->prefersVectorizedAddressing(); 351 } 352 353 int TargetTransformInfo::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, 354 int64_t BaseOffset, 355 bool HasBaseReg, 356 int64_t Scale, 357 unsigned AddrSpace) const { 358 int Cost = TTIImpl->getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg, 359 Scale, AddrSpace); 360 assert(Cost >= 0 && "TTI should not produce negative costs!"); 361 return Cost; 362 } 363 364 bool TargetTransformInfo::LSRWithInstrQueries() const { 365 return TTIImpl->LSRWithInstrQueries(); 366 } 367 368 bool TargetTransformInfo::isTruncateFree(Type *Ty1, Type *Ty2) const { 369 return TTIImpl->isTruncateFree(Ty1, Ty2); 370 } 371 372 bool TargetTransformInfo::isProfitableToHoist(Instruction *I) const { 373 return TTIImpl->isProfitableToHoist(I); 374 } 375 376 bool TargetTransformInfo::useAA() const { return TTIImpl->useAA(); } 377 378 bool TargetTransformInfo::isTypeLegal(Type *Ty) const { 379 return TTIImpl->isTypeLegal(Ty); 380 } 381 382 bool TargetTransformInfo::shouldBuildLookupTables() const { 383 return TTIImpl->shouldBuildLookupTables(); 384 } 385 bool TargetTransformInfo::shouldBuildLookupTablesForConstant(Constant *C) const { 386 return TTIImpl->shouldBuildLookupTablesForConstant(C); 387 } 388 389 bool TargetTransformInfo::useColdCCForColdCall(Function &F) const { 390 return TTIImpl->useColdCCForColdCall(F); 391 } 392 393 unsigned TargetTransformInfo:: 394 getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const { 395 return TTIImpl->getScalarizationOverhead(Ty, Insert, Extract); 396 } 397 398 unsigned TargetTransformInfo:: 399 getOperandsScalarizationOverhead(ArrayRef<const Value *> Args, 400 unsigned VF) const { 401 return TTIImpl->getOperandsScalarizationOverhead(Args, VF); 402 } 403 404 bool TargetTransformInfo::supportsEfficientVectorElementLoadStore() const { 405 return TTIImpl->supportsEfficientVectorElementLoadStore(); 406 } 407 408 bool TargetTransformInfo::enableAggressiveInterleaving(bool LoopHasReductions) const { 409 return TTIImpl->enableAggressiveInterleaving(LoopHasReductions); 410 } 411 412 TargetTransformInfo::MemCmpExpansionOptions 413 TargetTransformInfo::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const { 414 return TTIImpl->enableMemCmpExpansion(OptSize, IsZeroCmp); 415 } 416 417 bool TargetTransformInfo::enableInterleavedAccessVectorization() const { 418 return TTIImpl->enableInterleavedAccessVectorization(); 419 } 420 421 bool TargetTransformInfo::enableMaskedInterleavedAccessVectorization() const { 422 return TTIImpl->enableMaskedInterleavedAccessVectorization(); 423 } 424 425 bool TargetTransformInfo::isFPVectorizationPotentiallyUnsafe() const { 426 return TTIImpl->isFPVectorizationPotentiallyUnsafe(); 427 } 428 429 bool TargetTransformInfo::allowsMisalignedMemoryAccesses(LLVMContext &Context, 430 unsigned BitWidth, 431 unsigned AddressSpace, 432 unsigned Alignment, 433 bool *Fast) const { 434 return TTIImpl->allowsMisalignedMemoryAccesses(Context, BitWidth, AddressSpace, 435 Alignment, Fast); 436 } 437 438 TargetTransformInfo::PopcntSupportKind 439 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit) const { 440 return TTIImpl->getPopcntSupport(IntTyWidthInBit); 441 } 442 443 bool TargetTransformInfo::haveFastSqrt(Type *Ty) const { 444 return TTIImpl->haveFastSqrt(Ty); 445 } 446 447 bool TargetTransformInfo::isFCmpOrdCheaperThanFCmpZero(Type *Ty) const { 448 return TTIImpl->isFCmpOrdCheaperThanFCmpZero(Ty); 449 } 450 451 int TargetTransformInfo::getFPOpCost(Type *Ty) const { 452 int Cost = TTIImpl->getFPOpCost(Ty); 453 assert(Cost >= 0 && "TTI should not produce negative costs!"); 454 return Cost; 455 } 456 457 int TargetTransformInfo::getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, 458 const APInt &Imm, 459 Type *Ty) const { 460 int Cost = TTIImpl->getIntImmCodeSizeCost(Opcode, Idx, Imm, Ty); 461 assert(Cost >= 0 && "TTI should not produce negative costs!"); 462 return Cost; 463 } 464 465 int TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty) const { 466 int Cost = TTIImpl->getIntImmCost(Imm, Ty); 467 assert(Cost >= 0 && "TTI should not produce negative costs!"); 468 return Cost; 469 } 470 471 int TargetTransformInfo::getIntImmCostInst(unsigned Opcode, unsigned Idx, 472 const APInt &Imm, Type *Ty) const { 473 int Cost = TTIImpl->getIntImmCostInst(Opcode, Idx, Imm, Ty); 474 assert(Cost >= 0 && "TTI should not produce negative costs!"); 475 return Cost; 476 } 477 478 int TargetTransformInfo::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx, 479 const APInt &Imm, Type *Ty) const { 480 int Cost = TTIImpl->getIntImmCostIntrin(IID, Idx, Imm, Ty); 481 assert(Cost >= 0 && "TTI should not produce negative costs!"); 482 return Cost; 483 } 484 485 unsigned TargetTransformInfo::getNumberOfRegisters(unsigned ClassID) const { 486 return TTIImpl->getNumberOfRegisters(ClassID); 487 } 488 489 unsigned TargetTransformInfo::getRegisterClassForType(bool Vector, Type *Ty) const { 490 return TTIImpl->getRegisterClassForType(Vector, Ty); 491 } 492 493 const char* TargetTransformInfo::getRegisterClassName(unsigned ClassID) const { 494 return TTIImpl->getRegisterClassName(ClassID); 495 } 496 497 unsigned TargetTransformInfo::getRegisterBitWidth(bool Vector) const { 498 return TTIImpl->getRegisterBitWidth(Vector); 499 } 500 501 unsigned TargetTransformInfo::getMinVectorRegisterBitWidth() const { 502 return TTIImpl->getMinVectorRegisterBitWidth(); 503 } 504 505 bool TargetTransformInfo::shouldMaximizeVectorBandwidth(bool OptSize) const { 506 return TTIImpl->shouldMaximizeVectorBandwidth(OptSize); 507 } 508 509 unsigned TargetTransformInfo::getMinimumVF(unsigned ElemWidth) const { 510 return TTIImpl->getMinimumVF(ElemWidth); 511 } 512 513 bool TargetTransformInfo::shouldConsiderAddressTypePromotion( 514 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const { 515 return TTIImpl->shouldConsiderAddressTypePromotion( 516 I, AllowPromotionWithoutCommonHeader); 517 } 518 519 unsigned TargetTransformInfo::getCacheLineSize() const { 520 return TTIImpl->getCacheLineSize(); 521 } 522 523 llvm::Optional<unsigned> TargetTransformInfo::getCacheSize(CacheLevel Level) 524 const { 525 return TTIImpl->getCacheSize(Level); 526 } 527 528 llvm::Optional<unsigned> TargetTransformInfo::getCacheAssociativity( 529 CacheLevel Level) const { 530 return TTIImpl->getCacheAssociativity(Level); 531 } 532 533 unsigned TargetTransformInfo::getPrefetchDistance() const { 534 return TTIImpl->getPrefetchDistance(); 535 } 536 537 unsigned TargetTransformInfo::getMinPrefetchStride() const { 538 return TTIImpl->getMinPrefetchStride(); 539 } 540 541 unsigned TargetTransformInfo::getMaxPrefetchIterationsAhead() const { 542 return TTIImpl->getMaxPrefetchIterationsAhead(); 543 } 544 545 unsigned TargetTransformInfo::getMaxInterleaveFactor(unsigned VF) const { 546 return TTIImpl->getMaxInterleaveFactor(VF); 547 } 548 549 TargetTransformInfo::OperandValueKind 550 TargetTransformInfo::getOperandInfo(Value *V, OperandValueProperties &OpProps) { 551 OperandValueKind OpInfo = OK_AnyValue; 552 OpProps = OP_None; 553 554 if (auto *CI = dyn_cast<ConstantInt>(V)) { 555 if (CI->getValue().isPowerOf2()) 556 OpProps = OP_PowerOf2; 557 return OK_UniformConstantValue; 558 } 559 560 // A broadcast shuffle creates a uniform value. 561 // TODO: Add support for non-zero index broadcasts. 562 // TODO: Add support for different source vector width. 563 if (auto *ShuffleInst = dyn_cast<ShuffleVectorInst>(V)) 564 if (ShuffleInst->isZeroEltSplat()) 565 OpInfo = OK_UniformValue; 566 567 const Value *Splat = getSplatValue(V); 568 569 // Check for a splat of a constant or for a non uniform vector of constants 570 // and check if the constant(s) are all powers of two. 571 if (isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) { 572 OpInfo = OK_NonUniformConstantValue; 573 if (Splat) { 574 OpInfo = OK_UniformConstantValue; 575 if (auto *CI = dyn_cast<ConstantInt>(Splat)) 576 if (CI->getValue().isPowerOf2()) 577 OpProps = OP_PowerOf2; 578 } else if (auto *CDS = dyn_cast<ConstantDataSequential>(V)) { 579 OpProps = OP_PowerOf2; 580 for (unsigned I = 0, E = CDS->getNumElements(); I != E; ++I) { 581 if (auto *CI = dyn_cast<ConstantInt>(CDS->getElementAsConstant(I))) 582 if (CI->getValue().isPowerOf2()) 583 continue; 584 OpProps = OP_None; 585 break; 586 } 587 } 588 } 589 590 // Check for a splat of a uniform value. This is not loop aware, so return 591 // true only for the obviously uniform cases (argument, globalvalue) 592 if (Splat && (isa<Argument>(Splat) || isa<GlobalValue>(Splat))) 593 OpInfo = OK_UniformValue; 594 595 return OpInfo; 596 } 597 598 int TargetTransformInfo::getArithmeticInstrCost( 599 unsigned Opcode, Type *Ty, OperandValueKind Opd1Info, 600 OperandValueKind Opd2Info, OperandValueProperties Opd1PropInfo, 601 OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args, 602 const Instruction *CxtI) const { 603 int Cost = TTIImpl->getArithmeticInstrCost( 604 Opcode, Ty, Opd1Info, Opd2Info, Opd1PropInfo, Opd2PropInfo, Args, CxtI); 605 assert(Cost >= 0 && "TTI should not produce negative costs!"); 606 return Cost; 607 } 608 609 int TargetTransformInfo::getShuffleCost(ShuffleKind Kind, Type *Ty, int Index, 610 Type *SubTp) const { 611 int Cost = TTIImpl->getShuffleCost(Kind, Ty, Index, SubTp); 612 assert(Cost >= 0 && "TTI should not produce negative costs!"); 613 return Cost; 614 } 615 616 int TargetTransformInfo::getCastInstrCost(unsigned Opcode, Type *Dst, 617 Type *Src, const Instruction *I) const { 618 assert ((I == nullptr || I->getOpcode() == Opcode) && 619 "Opcode should reflect passed instruction."); 620 int Cost = TTIImpl->getCastInstrCost(Opcode, Dst, Src, I); 621 assert(Cost >= 0 && "TTI should not produce negative costs!"); 622 return Cost; 623 } 624 625 int TargetTransformInfo::getExtractWithExtendCost(unsigned Opcode, Type *Dst, 626 VectorType *VecTy, 627 unsigned Index) const { 628 int Cost = TTIImpl->getExtractWithExtendCost(Opcode, Dst, VecTy, Index); 629 assert(Cost >= 0 && "TTI should not produce negative costs!"); 630 return Cost; 631 } 632 633 int TargetTransformInfo::getCFInstrCost(unsigned Opcode) const { 634 int Cost = TTIImpl->getCFInstrCost(Opcode); 635 assert(Cost >= 0 && "TTI should not produce negative costs!"); 636 return Cost; 637 } 638 639 int TargetTransformInfo::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 640 Type *CondTy, const Instruction *I) const { 641 assert ((I == nullptr || I->getOpcode() == Opcode) && 642 "Opcode should reflect passed instruction."); 643 int Cost = TTIImpl->getCmpSelInstrCost(Opcode, ValTy, CondTy, I); 644 assert(Cost >= 0 && "TTI should not produce negative costs!"); 645 return Cost; 646 } 647 648 int TargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val, 649 unsigned Index) const { 650 int Cost = TTIImpl->getVectorInstrCost(Opcode, Val, Index); 651 assert(Cost >= 0 && "TTI should not produce negative costs!"); 652 return Cost; 653 } 654 655 int TargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src, 656 MaybeAlign Alignment, 657 unsigned AddressSpace, 658 const Instruction *I) const { 659 assert ((I == nullptr || I->getOpcode() == Opcode) && 660 "Opcode should reflect passed instruction."); 661 int Cost = TTIImpl->getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, I); 662 assert(Cost >= 0 && "TTI should not produce negative costs!"); 663 return Cost; 664 } 665 666 int TargetTransformInfo::getMaskedMemoryOpCost(unsigned Opcode, Type *Src, 667 unsigned Alignment, 668 unsigned AddressSpace) const { 669 int Cost = 670 TTIImpl->getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace); 671 assert(Cost >= 0 && "TTI should not produce negative costs!"); 672 return Cost; 673 } 674 675 int TargetTransformInfo::getGatherScatterOpCost(unsigned Opcode, Type *DataTy, 676 Value *Ptr, bool VariableMask, 677 unsigned Alignment, 678 const Instruction *I) const { 679 int Cost = TTIImpl->getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask, 680 Alignment, I); 681 assert(Cost >= 0 && "TTI should not produce negative costs!"); 682 return Cost; 683 } 684 685 int TargetTransformInfo::getInterleavedMemoryOpCost( 686 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices, 687 unsigned Alignment, unsigned AddressSpace, bool UseMaskForCond, 688 bool UseMaskForGaps) const { 689 int Cost = TTIImpl->getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 690 Alignment, AddressSpace, 691 UseMaskForCond, 692 UseMaskForGaps); 693 assert(Cost >= 0 && "TTI should not produce negative costs!"); 694 return Cost; 695 } 696 697 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 698 ArrayRef<Type *> Tys, 699 FastMathFlags FMF, 700 unsigned ScalarizationCostPassed, 701 const Instruction *I) const { 702 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Tys, FMF, 703 ScalarizationCostPassed, I); 704 assert(Cost >= 0 && "TTI should not produce negative costs!"); 705 return Cost; 706 } 707 708 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 709 ArrayRef<Value *> Args, 710 FastMathFlags FMF, unsigned VF, 711 const Instruction *I) const { 712 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF, I); 713 assert(Cost >= 0 && "TTI should not produce negative costs!"); 714 return Cost; 715 } 716 717 int TargetTransformInfo::getCallInstrCost(Function *F, Type *RetTy, 718 ArrayRef<Type *> Tys) const { 719 int Cost = TTIImpl->getCallInstrCost(F, RetTy, Tys); 720 assert(Cost >= 0 && "TTI should not produce negative costs!"); 721 return Cost; 722 } 723 724 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const { 725 return TTIImpl->getNumberOfParts(Tp); 726 } 727 728 int TargetTransformInfo::getAddressComputationCost(Type *Tp, 729 ScalarEvolution *SE, 730 const SCEV *Ptr) const { 731 int Cost = TTIImpl->getAddressComputationCost(Tp, SE, Ptr); 732 assert(Cost >= 0 && "TTI should not produce negative costs!"); 733 return Cost; 734 } 735 736 int TargetTransformInfo::getMemcpyCost(const Instruction *I) const { 737 int Cost = TTIImpl->getMemcpyCost(I); 738 assert(Cost >= 0 && "TTI should not produce negative costs!"); 739 return Cost; 740 } 741 742 int TargetTransformInfo::getArithmeticReductionCost(unsigned Opcode, Type *Ty, 743 bool IsPairwiseForm) const { 744 int Cost = TTIImpl->getArithmeticReductionCost(Opcode, Ty, IsPairwiseForm); 745 assert(Cost >= 0 && "TTI should not produce negative costs!"); 746 return Cost; 747 } 748 749 int TargetTransformInfo::getMinMaxReductionCost(Type *Ty, Type *CondTy, 750 bool IsPairwiseForm, 751 bool IsUnsigned) const { 752 int Cost = 753 TTIImpl->getMinMaxReductionCost(Ty, CondTy, IsPairwiseForm, IsUnsigned); 754 assert(Cost >= 0 && "TTI should not produce negative costs!"); 755 return Cost; 756 } 757 758 unsigned 759 TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const { 760 return TTIImpl->getCostOfKeepingLiveOverCall(Tys); 761 } 762 763 bool TargetTransformInfo::getTgtMemIntrinsic(IntrinsicInst *Inst, 764 MemIntrinsicInfo &Info) const { 765 return TTIImpl->getTgtMemIntrinsic(Inst, Info); 766 } 767 768 unsigned TargetTransformInfo::getAtomicMemIntrinsicMaxElementSize() const { 769 return TTIImpl->getAtomicMemIntrinsicMaxElementSize(); 770 } 771 772 Value *TargetTransformInfo::getOrCreateResultFromMemIntrinsic( 773 IntrinsicInst *Inst, Type *ExpectedType) const { 774 return TTIImpl->getOrCreateResultFromMemIntrinsic(Inst, ExpectedType); 775 } 776 777 Type *TargetTransformInfo::getMemcpyLoopLoweringType(LLVMContext &Context, 778 Value *Length, 779 unsigned SrcAddrSpace, 780 unsigned DestAddrSpace, 781 unsigned SrcAlign, 782 unsigned DestAlign) const { 783 return TTIImpl->getMemcpyLoopLoweringType(Context, Length, SrcAddrSpace, 784 DestAddrSpace, SrcAlign, 785 DestAlign); 786 } 787 788 void TargetTransformInfo::getMemcpyLoopResidualLoweringType( 789 SmallVectorImpl<Type *> &OpsOut, LLVMContext &Context, 790 unsigned RemainingBytes, 791 unsigned SrcAddrSpace, 792 unsigned DestAddrSpace, 793 unsigned SrcAlign, unsigned DestAlign) const { 794 TTIImpl->getMemcpyLoopResidualLoweringType(OpsOut, Context, RemainingBytes, 795 SrcAddrSpace, DestAddrSpace, 796 SrcAlign, DestAlign); 797 } 798 799 bool TargetTransformInfo::areInlineCompatible(const Function *Caller, 800 const Function *Callee) const { 801 return TTIImpl->areInlineCompatible(Caller, Callee); 802 } 803 804 bool TargetTransformInfo::areFunctionArgsABICompatible( 805 const Function *Caller, const Function *Callee, 806 SmallPtrSetImpl<Argument *> &Args) const { 807 return TTIImpl->areFunctionArgsABICompatible(Caller, Callee, Args); 808 } 809 810 bool TargetTransformInfo::isIndexedLoadLegal(MemIndexedMode Mode, 811 Type *Ty) const { 812 return TTIImpl->isIndexedLoadLegal(Mode, Ty); 813 } 814 815 bool TargetTransformInfo::isIndexedStoreLegal(MemIndexedMode Mode, 816 Type *Ty) const { 817 return TTIImpl->isIndexedStoreLegal(Mode, Ty); 818 } 819 820 unsigned TargetTransformInfo::getLoadStoreVecRegBitWidth(unsigned AS) const { 821 return TTIImpl->getLoadStoreVecRegBitWidth(AS); 822 } 823 824 bool TargetTransformInfo::isLegalToVectorizeLoad(LoadInst *LI) const { 825 return TTIImpl->isLegalToVectorizeLoad(LI); 826 } 827 828 bool TargetTransformInfo::isLegalToVectorizeStore(StoreInst *SI) const { 829 return TTIImpl->isLegalToVectorizeStore(SI); 830 } 831 832 bool TargetTransformInfo::isLegalToVectorizeLoadChain( 833 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 834 return TTIImpl->isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment, 835 AddrSpace); 836 } 837 838 bool TargetTransformInfo::isLegalToVectorizeStoreChain( 839 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 840 return TTIImpl->isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment, 841 AddrSpace); 842 } 843 844 unsigned TargetTransformInfo::getLoadVectorFactor(unsigned VF, 845 unsigned LoadSize, 846 unsigned ChainSizeInBytes, 847 VectorType *VecTy) const { 848 return TTIImpl->getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy); 849 } 850 851 unsigned TargetTransformInfo::getStoreVectorFactor(unsigned VF, 852 unsigned StoreSize, 853 unsigned ChainSizeInBytes, 854 VectorType *VecTy) const { 855 return TTIImpl->getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy); 856 } 857 858 bool TargetTransformInfo::useReductionIntrinsic(unsigned Opcode, 859 Type *Ty, ReductionFlags Flags) const { 860 return TTIImpl->useReductionIntrinsic(Opcode, Ty, Flags); 861 } 862 863 bool TargetTransformInfo::shouldExpandReduction(const IntrinsicInst *II) const { 864 return TTIImpl->shouldExpandReduction(II); 865 } 866 867 unsigned TargetTransformInfo::getGISelRematGlobalCost() const { 868 return TTIImpl->getGISelRematGlobalCost(); 869 } 870 871 int TargetTransformInfo::getInstructionLatency(const Instruction *I) const { 872 return TTIImpl->getInstructionLatency(I); 873 } 874 875 static bool matchPairwiseShuffleMask(ShuffleVectorInst *SI, bool IsLeft, 876 unsigned Level) { 877 // We don't need a shuffle if we just want to have element 0 in position 0 of 878 // the vector. 879 if (!SI && Level == 0 && IsLeft) 880 return true; 881 else if (!SI) 882 return false; 883 884 SmallVector<int, 32> Mask(SI->getType()->getVectorNumElements(), -1); 885 886 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right) depending on whether 887 // we look at the left or right side. 888 for (unsigned i = 0, e = (1 << Level), val = !IsLeft; i != e; ++i, val += 2) 889 Mask[i] = val; 890 891 SmallVector<int, 16> ActualMask = SI->getShuffleMask(); 892 return Mask == ActualMask; 893 } 894 895 namespace { 896 /// Kind of the reduction data. 897 enum ReductionKind { 898 RK_None, /// Not a reduction. 899 RK_Arithmetic, /// Binary reduction data. 900 RK_MinMax, /// Min/max reduction data. 901 RK_UnsignedMinMax, /// Unsigned min/max reduction data. 902 }; 903 /// Contains opcode + LHS/RHS parts of the reduction operations. 904 struct ReductionData { 905 ReductionData() = delete; 906 ReductionData(ReductionKind Kind, unsigned Opcode, Value *LHS, Value *RHS) 907 : Opcode(Opcode), LHS(LHS), RHS(RHS), Kind(Kind) { 908 assert(Kind != RK_None && "expected binary or min/max reduction only."); 909 } 910 unsigned Opcode = 0; 911 Value *LHS = nullptr; 912 Value *RHS = nullptr; 913 ReductionKind Kind = RK_None; 914 bool hasSameData(ReductionData &RD) const { 915 return Kind == RD.Kind && Opcode == RD.Opcode; 916 } 917 }; 918 } // namespace 919 920 static Optional<ReductionData> getReductionData(Instruction *I) { 921 Value *L, *R; 922 if (m_BinOp(m_Value(L), m_Value(R)).match(I)) 923 return ReductionData(RK_Arithmetic, I->getOpcode(), L, R); 924 if (auto *SI = dyn_cast<SelectInst>(I)) { 925 if (m_SMin(m_Value(L), m_Value(R)).match(SI) || 926 m_SMax(m_Value(L), m_Value(R)).match(SI) || 927 m_OrdFMin(m_Value(L), m_Value(R)).match(SI) || 928 m_OrdFMax(m_Value(L), m_Value(R)).match(SI) || 929 m_UnordFMin(m_Value(L), m_Value(R)).match(SI) || 930 m_UnordFMax(m_Value(L), m_Value(R)).match(SI)) { 931 auto *CI = cast<CmpInst>(SI->getCondition()); 932 return ReductionData(RK_MinMax, CI->getOpcode(), L, R); 933 } 934 if (m_UMin(m_Value(L), m_Value(R)).match(SI) || 935 m_UMax(m_Value(L), m_Value(R)).match(SI)) { 936 auto *CI = cast<CmpInst>(SI->getCondition()); 937 return ReductionData(RK_UnsignedMinMax, CI->getOpcode(), L, R); 938 } 939 } 940 return llvm::None; 941 } 942 943 static ReductionKind matchPairwiseReductionAtLevel(Instruction *I, 944 unsigned Level, 945 unsigned NumLevels) { 946 // Match one level of pairwise operations. 947 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 948 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 949 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 950 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 951 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 952 if (!I) 953 return RK_None; 954 955 assert(I->getType()->isVectorTy() && "Expecting a vector type"); 956 957 Optional<ReductionData> RD = getReductionData(I); 958 if (!RD) 959 return RK_None; 960 961 ShuffleVectorInst *LS = dyn_cast<ShuffleVectorInst>(RD->LHS); 962 if (!LS && Level) 963 return RK_None; 964 ShuffleVectorInst *RS = dyn_cast<ShuffleVectorInst>(RD->RHS); 965 if (!RS && Level) 966 return RK_None; 967 968 // On level 0 we can omit one shufflevector instruction. 969 if (!Level && !RS && !LS) 970 return RK_None; 971 972 // Shuffle inputs must match. 973 Value *NextLevelOpL = LS ? LS->getOperand(0) : nullptr; 974 Value *NextLevelOpR = RS ? RS->getOperand(0) : nullptr; 975 Value *NextLevelOp = nullptr; 976 if (NextLevelOpR && NextLevelOpL) { 977 // If we have two shuffles their operands must match. 978 if (NextLevelOpL != NextLevelOpR) 979 return RK_None; 980 981 NextLevelOp = NextLevelOpL; 982 } else if (Level == 0 && (NextLevelOpR || NextLevelOpL)) { 983 // On the first level we can omit the shufflevector <0, undef,...>. So the 984 // input to the other shufflevector <1, undef> must match with one of the 985 // inputs to the current binary operation. 986 // Example: 987 // %NextLevelOpL = shufflevector %R, <1, undef ...> 988 // %BinOp = fadd %NextLevelOpL, %R 989 if (NextLevelOpL && NextLevelOpL != RD->RHS) 990 return RK_None; 991 else if (NextLevelOpR && NextLevelOpR != RD->LHS) 992 return RK_None; 993 994 NextLevelOp = NextLevelOpL ? RD->RHS : RD->LHS; 995 } else 996 return RK_None; 997 998 // Check that the next levels binary operation exists and matches with the 999 // current one. 1000 if (Level + 1 != NumLevels) { 1001 Optional<ReductionData> NextLevelRD = 1002 getReductionData(cast<Instruction>(NextLevelOp)); 1003 if (!NextLevelRD || !RD->hasSameData(*NextLevelRD)) 1004 return RK_None; 1005 } 1006 1007 // Shuffle mask for pairwise operation must match. 1008 if (matchPairwiseShuffleMask(LS, /*IsLeft=*/true, Level)) { 1009 if (!matchPairwiseShuffleMask(RS, /*IsLeft=*/false, Level)) 1010 return RK_None; 1011 } else if (matchPairwiseShuffleMask(RS, /*IsLeft=*/true, Level)) { 1012 if (!matchPairwiseShuffleMask(LS, /*IsLeft=*/false, Level)) 1013 return RK_None; 1014 } else { 1015 return RK_None; 1016 } 1017 1018 if (++Level == NumLevels) 1019 return RD->Kind; 1020 1021 // Match next level. 1022 return matchPairwiseReductionAtLevel(cast<Instruction>(NextLevelOp), Level, 1023 NumLevels); 1024 } 1025 1026 static ReductionKind matchPairwiseReduction(const ExtractElementInst *ReduxRoot, 1027 unsigned &Opcode, Type *&Ty) { 1028 if (!EnableReduxCost) 1029 return RK_None; 1030 1031 // Need to extract the first element. 1032 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 1033 unsigned Idx = ~0u; 1034 if (CI) 1035 Idx = CI->getZExtValue(); 1036 if (Idx != 0) 1037 return RK_None; 1038 1039 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 1040 if (!RdxStart) 1041 return RK_None; 1042 Optional<ReductionData> RD = getReductionData(RdxStart); 1043 if (!RD) 1044 return RK_None; 1045 1046 Type *VecTy = RdxStart->getType(); 1047 unsigned NumVecElems = VecTy->getVectorNumElements(); 1048 if (!isPowerOf2_32(NumVecElems)) 1049 return RK_None; 1050 1051 // We look for a sequence of shuffle,shuffle,add triples like the following 1052 // that builds a pairwise reduction tree. 1053 // 1054 // (X0, X1, X2, X3) 1055 // (X0 + X1, X2 + X3, undef, undef) 1056 // ((X0 + X1) + (X2 + X3), undef, undef, undef) 1057 // 1058 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 1059 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 1060 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 1061 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 1062 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 1063 // %rdx.shuf.1.0 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 1064 // <4 x i32> <i32 0, i32 undef, i32 undef, i32 undef> 1065 // %rdx.shuf.1.1 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 1066 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 1067 // %bin.rdx8 = fadd <4 x float> %rdx.shuf.1.0, %rdx.shuf.1.1 1068 // %r = extractelement <4 x float> %bin.rdx8, i32 0 1069 if (matchPairwiseReductionAtLevel(RdxStart, 0, Log2_32(NumVecElems)) == 1070 RK_None) 1071 return RK_None; 1072 1073 Opcode = RD->Opcode; 1074 Ty = VecTy; 1075 1076 return RD->Kind; 1077 } 1078 1079 static std::pair<Value *, ShuffleVectorInst *> 1080 getShuffleAndOtherOprd(Value *L, Value *R) { 1081 ShuffleVectorInst *S = nullptr; 1082 1083 if ((S = dyn_cast<ShuffleVectorInst>(L))) 1084 return std::make_pair(R, S); 1085 1086 S = dyn_cast<ShuffleVectorInst>(R); 1087 return std::make_pair(L, S); 1088 } 1089 1090 static ReductionKind 1091 matchVectorSplittingReduction(const ExtractElementInst *ReduxRoot, 1092 unsigned &Opcode, Type *&Ty) { 1093 if (!EnableReduxCost) 1094 return RK_None; 1095 1096 // Need to extract the first element. 1097 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 1098 unsigned Idx = ~0u; 1099 if (CI) 1100 Idx = CI->getZExtValue(); 1101 if (Idx != 0) 1102 return RK_None; 1103 1104 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 1105 if (!RdxStart) 1106 return RK_None; 1107 Optional<ReductionData> RD = getReductionData(RdxStart); 1108 if (!RD) 1109 return RK_None; 1110 1111 Type *VecTy = ReduxRoot->getOperand(0)->getType(); 1112 unsigned NumVecElems = VecTy->getVectorNumElements(); 1113 if (!isPowerOf2_32(NumVecElems)) 1114 return RK_None; 1115 1116 // We look for a sequence of shuffles and adds like the following matching one 1117 // fadd, shuffle vector pair at a time. 1118 // 1119 // %rdx.shuf = shufflevector <4 x float> %rdx, <4 x float> undef, 1120 // <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 1121 // %bin.rdx = fadd <4 x float> %rdx, %rdx.shuf 1122 // %rdx.shuf7 = shufflevector <4 x float> %bin.rdx, <4 x float> undef, 1123 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 1124 // %bin.rdx8 = fadd <4 x float> %bin.rdx, %rdx.shuf7 1125 // %r = extractelement <4 x float> %bin.rdx8, i32 0 1126 1127 unsigned MaskStart = 1; 1128 Instruction *RdxOp = RdxStart; 1129 SmallVector<int, 32> ShuffleMask(NumVecElems, 0); 1130 unsigned NumVecElemsRemain = NumVecElems; 1131 while (NumVecElemsRemain - 1) { 1132 // Check for the right reduction operation. 1133 if (!RdxOp) 1134 return RK_None; 1135 Optional<ReductionData> RDLevel = getReductionData(RdxOp); 1136 if (!RDLevel || !RDLevel->hasSameData(*RD)) 1137 return RK_None; 1138 1139 Value *NextRdxOp; 1140 ShuffleVectorInst *Shuffle; 1141 std::tie(NextRdxOp, Shuffle) = 1142 getShuffleAndOtherOprd(RDLevel->LHS, RDLevel->RHS); 1143 1144 // Check the current reduction operation and the shuffle use the same value. 1145 if (Shuffle == nullptr) 1146 return RK_None; 1147 if (Shuffle->getOperand(0) != NextRdxOp) 1148 return RK_None; 1149 1150 // Check that shuffle masks matches. 1151 for (unsigned j = 0; j != MaskStart; ++j) 1152 ShuffleMask[j] = MaskStart + j; 1153 // Fill the rest of the mask with -1 for undef. 1154 std::fill(&ShuffleMask[MaskStart], ShuffleMask.end(), -1); 1155 1156 SmallVector<int, 16> Mask = Shuffle->getShuffleMask(); 1157 if (ShuffleMask != Mask) 1158 return RK_None; 1159 1160 RdxOp = dyn_cast<Instruction>(NextRdxOp); 1161 NumVecElemsRemain /= 2; 1162 MaskStart *= 2; 1163 } 1164 1165 Opcode = RD->Opcode; 1166 Ty = VecTy; 1167 return RD->Kind; 1168 } 1169 1170 int TargetTransformInfo::getInstructionThroughput(const Instruction *I) const { 1171 switch (I->getOpcode()) { 1172 case Instruction::GetElementPtr: 1173 return getUserCost(I); 1174 1175 case Instruction::Ret: 1176 case Instruction::PHI: 1177 case Instruction::Br: { 1178 return getCFInstrCost(I->getOpcode()); 1179 } 1180 case Instruction::Add: 1181 case Instruction::FAdd: 1182 case Instruction::Sub: 1183 case Instruction::FSub: 1184 case Instruction::Mul: 1185 case Instruction::FMul: 1186 case Instruction::UDiv: 1187 case Instruction::SDiv: 1188 case Instruction::FDiv: 1189 case Instruction::URem: 1190 case Instruction::SRem: 1191 case Instruction::FRem: 1192 case Instruction::Shl: 1193 case Instruction::LShr: 1194 case Instruction::AShr: 1195 case Instruction::And: 1196 case Instruction::Or: 1197 case Instruction::Xor: { 1198 TargetTransformInfo::OperandValueKind Op1VK, Op2VK; 1199 TargetTransformInfo::OperandValueProperties Op1VP, Op2VP; 1200 Op1VK = getOperandInfo(I->getOperand(0), Op1VP); 1201 Op2VK = getOperandInfo(I->getOperand(1), Op2VP); 1202 SmallVector<const Value *, 2> Operands(I->operand_values()); 1203 return getArithmeticInstrCost(I->getOpcode(), I->getType(), Op1VK, Op2VK, 1204 Op1VP, Op2VP, Operands, I); 1205 } 1206 case Instruction::FNeg: { 1207 TargetTransformInfo::OperandValueKind Op1VK, Op2VK; 1208 TargetTransformInfo::OperandValueProperties Op1VP, Op2VP; 1209 Op1VK = getOperandInfo(I->getOperand(0), Op1VP); 1210 Op2VK = OK_AnyValue; 1211 Op2VP = OP_None; 1212 SmallVector<const Value *, 2> Operands(I->operand_values()); 1213 return getArithmeticInstrCost(I->getOpcode(), I->getType(), Op1VK, Op2VK, 1214 Op1VP, Op2VP, Operands, I); 1215 } 1216 case Instruction::Select: { 1217 const SelectInst *SI = cast<SelectInst>(I); 1218 Type *CondTy = SI->getCondition()->getType(); 1219 return getCmpSelInstrCost(I->getOpcode(), I->getType(), CondTy, I); 1220 } 1221 case Instruction::ICmp: 1222 case Instruction::FCmp: { 1223 Type *ValTy = I->getOperand(0)->getType(); 1224 return getCmpSelInstrCost(I->getOpcode(), ValTy, I->getType(), I); 1225 } 1226 case Instruction::Store: { 1227 const StoreInst *SI = cast<StoreInst>(I); 1228 Type *ValTy = SI->getValueOperand()->getType(); 1229 return getMemoryOpCost(I->getOpcode(), ValTy, 1230 MaybeAlign(SI->getAlignment()), 1231 SI->getPointerAddressSpace(), I); 1232 } 1233 case Instruction::Load: { 1234 const LoadInst *LI = cast<LoadInst>(I); 1235 return getMemoryOpCost(I->getOpcode(), I->getType(), 1236 MaybeAlign(LI->getAlignment()), 1237 LI->getPointerAddressSpace(), I); 1238 } 1239 case Instruction::ZExt: 1240 case Instruction::SExt: 1241 case Instruction::FPToUI: 1242 case Instruction::FPToSI: 1243 case Instruction::FPExt: 1244 case Instruction::PtrToInt: 1245 case Instruction::IntToPtr: 1246 case Instruction::SIToFP: 1247 case Instruction::UIToFP: 1248 case Instruction::Trunc: 1249 case Instruction::FPTrunc: 1250 case Instruction::BitCast: 1251 case Instruction::AddrSpaceCast: { 1252 Type *SrcTy = I->getOperand(0)->getType(); 1253 return getCastInstrCost(I->getOpcode(), I->getType(), SrcTy, I); 1254 } 1255 case Instruction::ExtractElement: { 1256 const ExtractElementInst * EEI = cast<ExtractElementInst>(I); 1257 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1258 unsigned Idx = -1; 1259 if (CI) 1260 Idx = CI->getZExtValue(); 1261 1262 // Try to match a reduction sequence (series of shufflevector and vector 1263 // adds followed by a extractelement). 1264 unsigned ReduxOpCode; 1265 Type *ReduxType; 1266 1267 switch (matchVectorSplittingReduction(EEI, ReduxOpCode, ReduxType)) { 1268 case RK_Arithmetic: 1269 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1270 /*IsPairwiseForm=*/false); 1271 case RK_MinMax: 1272 return getMinMaxReductionCost( 1273 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1274 /*IsPairwiseForm=*/false, /*IsUnsigned=*/false); 1275 case RK_UnsignedMinMax: 1276 return getMinMaxReductionCost( 1277 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1278 /*IsPairwiseForm=*/false, /*IsUnsigned=*/true); 1279 case RK_None: 1280 break; 1281 } 1282 1283 switch (matchPairwiseReduction(EEI, ReduxOpCode, ReduxType)) { 1284 case RK_Arithmetic: 1285 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1286 /*IsPairwiseForm=*/true); 1287 case RK_MinMax: 1288 return getMinMaxReductionCost( 1289 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1290 /*IsPairwiseForm=*/true, /*IsUnsigned=*/false); 1291 case RK_UnsignedMinMax: 1292 return getMinMaxReductionCost( 1293 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1294 /*IsPairwiseForm=*/true, /*IsUnsigned=*/true); 1295 case RK_None: 1296 break; 1297 } 1298 1299 return getVectorInstrCost(I->getOpcode(), 1300 EEI->getOperand(0)->getType(), Idx); 1301 } 1302 case Instruction::InsertElement: { 1303 const InsertElementInst * IE = cast<InsertElementInst>(I); 1304 ConstantInt *CI = dyn_cast<ConstantInt>(IE->getOperand(2)); 1305 unsigned Idx = -1; 1306 if (CI) 1307 Idx = CI->getZExtValue(); 1308 return getVectorInstrCost(I->getOpcode(), 1309 IE->getType(), Idx); 1310 } 1311 case Instruction::ExtractValue: 1312 return 0; // Model all ExtractValue nodes as free. 1313 case Instruction::ShuffleVector: { 1314 const ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I); 1315 Type *Ty = Shuffle->getType(); 1316 Type *SrcTy = Shuffle->getOperand(0)->getType(); 1317 1318 // TODO: Identify and add costs for insert subvector, etc. 1319 int SubIndex; 1320 if (Shuffle->isExtractSubvectorMask(SubIndex)) 1321 return TTIImpl->getShuffleCost(SK_ExtractSubvector, SrcTy, SubIndex, Ty); 1322 1323 if (Shuffle->changesLength()) 1324 return -1; 1325 1326 if (Shuffle->isIdentity()) 1327 return 0; 1328 1329 if (Shuffle->isReverse()) 1330 return TTIImpl->getShuffleCost(SK_Reverse, Ty, 0, nullptr); 1331 1332 if (Shuffle->isSelect()) 1333 return TTIImpl->getShuffleCost(SK_Select, Ty, 0, nullptr); 1334 1335 if (Shuffle->isTranspose()) 1336 return TTIImpl->getShuffleCost(SK_Transpose, Ty, 0, nullptr); 1337 1338 if (Shuffle->isZeroEltSplat()) 1339 return TTIImpl->getShuffleCost(SK_Broadcast, Ty, 0, nullptr); 1340 1341 if (Shuffle->isSingleSource()) 1342 return TTIImpl->getShuffleCost(SK_PermuteSingleSrc, Ty, 0, nullptr); 1343 1344 return TTIImpl->getShuffleCost(SK_PermuteTwoSrc, Ty, 0, nullptr); 1345 } 1346 case Instruction::Call: 1347 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1348 SmallVector<Value *, 4> Args(II->arg_operands()); 1349 1350 FastMathFlags FMF; 1351 if (auto *FPMO = dyn_cast<FPMathOperator>(II)) 1352 FMF = FPMO->getFastMathFlags(); 1353 1354 return getIntrinsicInstrCost(II->getIntrinsicID(), II->getType(), Args, 1355 FMF, 1, II); 1356 } 1357 return -1; 1358 default: 1359 // We don't have any information on this instruction. 1360 return -1; 1361 } 1362 } 1363 1364 TargetTransformInfo::Concept::~Concept() {} 1365 1366 TargetIRAnalysis::TargetIRAnalysis() : TTICallback(&getDefaultTTI) {} 1367 1368 TargetIRAnalysis::TargetIRAnalysis( 1369 std::function<Result(const Function &)> TTICallback) 1370 : TTICallback(std::move(TTICallback)) {} 1371 1372 TargetIRAnalysis::Result TargetIRAnalysis::run(const Function &F, 1373 FunctionAnalysisManager &) { 1374 return TTICallback(F); 1375 } 1376 1377 AnalysisKey TargetIRAnalysis::Key; 1378 1379 TargetIRAnalysis::Result TargetIRAnalysis::getDefaultTTI(const Function &F) { 1380 return Result(F.getParent()->getDataLayout()); 1381 } 1382 1383 // Register the basic pass. 1384 INITIALIZE_PASS(TargetTransformInfoWrapperPass, "tti", 1385 "Target Transform Information", false, true) 1386 char TargetTransformInfoWrapperPass::ID = 0; 1387 1388 void TargetTransformInfoWrapperPass::anchor() {} 1389 1390 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass() 1391 : ImmutablePass(ID) { 1392 initializeTargetTransformInfoWrapperPassPass( 1393 *PassRegistry::getPassRegistry()); 1394 } 1395 1396 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass( 1397 TargetIRAnalysis TIRA) 1398 : ImmutablePass(ID), TIRA(std::move(TIRA)) { 1399 initializeTargetTransformInfoWrapperPassPass( 1400 *PassRegistry::getPassRegistry()); 1401 } 1402 1403 TargetTransformInfo &TargetTransformInfoWrapperPass::getTTI(const Function &F) { 1404 FunctionAnalysisManager DummyFAM; 1405 TTI = TIRA.run(F, DummyFAM); 1406 return *TTI; 1407 } 1408 1409 ImmutablePass * 1410 llvm::createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA) { 1411 return new TargetTransformInfoWrapperPass(std::move(TIRA)); 1412 } 1413