1//===- TargetTransformInfoImpl.h --------------------------------*- C++ -*-===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9/// \file 10/// This file provides helpers for the implementation of 11/// a TargetTransformInfo-conforming class. 12/// 13//===----------------------------------------------------------------------===// 14 15#ifndef LLVM_ANALYSIS_TARGETTRANSFORMINFOIMPL_H 16#define LLVM_ANALYSIS_TARGETTRANSFORMINFOIMPL_H 17 18#include "llvm/Analysis/ScalarEvolutionExpressions.h" 19#include "llvm/Analysis/TargetTransformInfo.h" 20#include "llvm/Analysis/VectorUtils.h" 21#include "llvm/IR/CallSite.h" 22#include "llvm/IR/DataLayout.h" 23#include "llvm/IR/Function.h" 24#include "llvm/IR/GetElementPtrTypeIterator.h" 25#include "llvm/IR/Operator.h" 26#include "llvm/IR/Type.h" 27 28namespace llvm { 29 30/// \brief Base class for use as a mix-in that aids implementing 31/// a TargetTransformInfo-compatible class. 32class TargetTransformInfoImplBase { 33protected: 34 typedef TargetTransformInfo TTI; 35 36 const DataLayout &DL; 37 38 explicit TargetTransformInfoImplBase(const DataLayout &DL) : DL(DL) {} 39 40public: 41 // Provide value semantics. MSVC requires that we spell all of these out. 42 TargetTransformInfoImplBase(const TargetTransformInfoImplBase &Arg) 43 : DL(Arg.DL) {} 44 TargetTransformInfoImplBase(TargetTransformInfoImplBase &&Arg) : DL(Arg.DL) {} 45 46 const DataLayout &getDataLayout() const { return DL; } 47 48 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) { 49 switch (Opcode) { 50 default: 51 // By default, just classify everything as 'basic'. 52 return TTI::TCC_Basic; 53 54 case Instruction::GetElementPtr: 55 llvm_unreachable("Use getGEPCost for GEP operations!"); 56 57 case Instruction::BitCast: 58 assert(OpTy && "Cast instructions must provide the operand type"); 59 if (Ty == OpTy || (Ty->isPointerTy() && OpTy->isPointerTy())) 60 // Identity and pointer-to-pointer casts are free. 61 return TTI::TCC_Free; 62 63 // Otherwise, the default basic cost is used. 64 return TTI::TCC_Basic; 65 66 case Instruction::FDiv: 67 case Instruction::FRem: 68 case Instruction::SDiv: 69 case Instruction::SRem: 70 case Instruction::UDiv: 71 case Instruction::URem: 72 return TTI::TCC_Expensive; 73 74 case Instruction::IntToPtr: { 75 // An inttoptr cast is free so long as the input is a legal integer type 76 // which doesn't contain values outside the range of a pointer. 77 unsigned OpSize = OpTy->getScalarSizeInBits(); 78 if (DL.isLegalInteger(OpSize) && 79 OpSize <= DL.getPointerTypeSizeInBits(Ty)) 80 return TTI::TCC_Free; 81 82 // Otherwise it's not a no-op. 83 return TTI::TCC_Basic; 84 } 85 case Instruction::PtrToInt: { 86 // A ptrtoint cast is free so long as the result is large enough to store 87 // the pointer, and a legal integer type. 88 unsigned DestSize = Ty->getScalarSizeInBits(); 89 if (DL.isLegalInteger(DestSize) && 90 DestSize >= DL.getPointerTypeSizeInBits(OpTy)) 91 return TTI::TCC_Free; 92 93 // Otherwise it's not a no-op. 94 return TTI::TCC_Basic; 95 } 96 case Instruction::Trunc: 97 // trunc to a native type is free (assuming the target has compare and 98 // shift-right of the same width). 99 if (DL.isLegalInteger(DL.getTypeSizeInBits(Ty))) 100 return TTI::TCC_Free; 101 102 return TTI::TCC_Basic; 103 } 104 } 105 106 int getGEPCost(Type *PointeeType, const Value *Ptr, 107 ArrayRef<const Value *> Operands) { 108 // In the basic model, we just assume that all-constant GEPs will be folded 109 // into their uses via addressing modes. 110 for (unsigned Idx = 0, Size = Operands.size(); Idx != Size; ++Idx) 111 if (!isa<Constant>(Operands[Idx])) 112 return TTI::TCC_Basic; 113 114 return TTI::TCC_Free; 115 } 116 117 unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI, 118 unsigned &JTSize) { 119 JTSize = 0; 120 return SI.getNumCases(); 121 } 122 123 unsigned getCallCost(FunctionType *FTy, int NumArgs) { 124 assert(FTy && "FunctionType must be provided to this routine."); 125 126 // The target-independent implementation just measures the size of the 127 // function by approximating that each argument will take on average one 128 // instruction to prepare. 129 130 if (NumArgs < 0) 131 // Set the argument number to the number of explicit arguments in the 132 // function. 133 NumArgs = FTy->getNumParams(); 134 135 return TTI::TCC_Basic * (NumArgs + 1); 136 } 137 138 unsigned getInliningThresholdMultiplier() { return 1; } 139 140 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy, 141 ArrayRef<Type *> ParamTys) { 142 switch (IID) { 143 default: 144 // Intrinsics rarely (if ever) have normal argument setup constraints. 145 // Model them as having a basic instruction cost. 146 // FIXME: This is wrong for libc intrinsics. 147 return TTI::TCC_Basic; 148 149 case Intrinsic::annotation: 150 case Intrinsic::assume: 151 case Intrinsic::dbg_declare: 152 case Intrinsic::dbg_value: 153 case Intrinsic::invariant_start: 154 case Intrinsic::invariant_end: 155 case Intrinsic::lifetime_start: 156 case Intrinsic::lifetime_end: 157 case Intrinsic::objectsize: 158 case Intrinsic::ptr_annotation: 159 case Intrinsic::var_annotation: 160 case Intrinsic::experimental_gc_result: 161 case Intrinsic::experimental_gc_relocate: 162 case Intrinsic::coro_alloc: 163 case Intrinsic::coro_begin: 164 case Intrinsic::coro_free: 165 case Intrinsic::coro_end: 166 case Intrinsic::coro_frame: 167 case Intrinsic::coro_size: 168 case Intrinsic::coro_suspend: 169 case Intrinsic::coro_param: 170 case Intrinsic::coro_subfn_addr: 171 // These intrinsics don't actually represent code after lowering. 172 return TTI::TCC_Free; 173 } 174 } 175 176 bool hasBranchDivergence() { return false; } 177 178 bool isSourceOfDivergence(const Value *V) { return false; } 179 180 bool isAlwaysUniform(const Value *V) { return false; } 181 182 unsigned getFlatAddressSpace () { 183 return -1; 184 } 185 186 bool isLoweredToCall(const Function *F) { 187 // FIXME: These should almost certainly not be handled here, and instead 188 // handled with the help of TLI or the target itself. This was largely 189 // ported from existing analysis heuristics here so that such refactorings 190 // can take place in the future. 191 192 if (F->isIntrinsic()) 193 return false; 194 195 if (F->hasLocalLinkage() || !F->hasName()) 196 return true; 197 198 StringRef Name = F->getName(); 199 200 // These will all likely lower to a single selection DAG node. 201 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" || 202 Name == "fabs" || Name == "fabsf" || Name == "fabsl" || Name == "sin" || 203 Name == "fmin" || Name == "fminf" || Name == "fminl" || 204 Name == "fmax" || Name == "fmaxf" || Name == "fmaxl" || 205 Name == "sinf" || Name == "sinl" || Name == "cos" || Name == "cosf" || 206 Name == "cosl" || Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl") 207 return false; 208 209 // These are all likely to be optimized into something smaller. 210 if (Name == "pow" || Name == "powf" || Name == "powl" || Name == "exp2" || 211 Name == "exp2l" || Name == "exp2f" || Name == "floor" || 212 Name == "floorf" || Name == "ceil" || Name == "round" || 213 Name == "ffs" || Name == "ffsl" || Name == "abs" || Name == "labs" || 214 Name == "llabs") 215 return false; 216 217 return true; 218 } 219 220 void getUnrollingPreferences(Loop *, TTI::UnrollingPreferences &) {} 221 222 bool isLegalAddImmediate(int64_t Imm) { return false; } 223 224 bool isLegalICmpImmediate(int64_t Imm) { return false; } 225 226 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset, 227 bool HasBaseReg, int64_t Scale, 228 unsigned AddrSpace) { 229 // Guess that only reg and reg+reg addressing is allowed. This heuristic is 230 // taken from the implementation of LSR. 231 return !BaseGV && BaseOffset == 0 && (Scale == 0 || Scale == 1); 232 } 233 234 bool isLSRCostLess(TTI::LSRCost &C1, TTI::LSRCost &C2) { 235 return std::tie(C1.NumRegs, C1.AddRecCost, C1.NumIVMuls, C1.NumBaseAdds, 236 C1.ScaleCost, C1.ImmCost, C1.SetupCost) < 237 std::tie(C2.NumRegs, C2.AddRecCost, C2.NumIVMuls, C2.NumBaseAdds, 238 C2.ScaleCost, C2.ImmCost, C2.SetupCost); 239 } 240 241 bool isLegalMaskedStore(Type *DataType) { return false; } 242 243 bool isLegalMaskedLoad(Type *DataType) { return false; } 244 245 bool isLegalMaskedScatter(Type *DataType) { return false; } 246 247 bool isLegalMaskedGather(Type *DataType) { return false; } 248 249 bool prefersVectorizedAddressing() { return true; } 250 251 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset, 252 bool HasBaseReg, int64_t Scale, unsigned AddrSpace) { 253 // Guess that all legal addressing mode are free. 254 if (isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg, 255 Scale, AddrSpace)) 256 return 0; 257 return -1; 258 } 259 260 bool isFoldableMemAccessOffset(Instruction *I, int64_t Offset) { return true; } 261 262 bool isTruncateFree(Type *Ty1, Type *Ty2) { return false; } 263 264 bool isProfitableToHoist(Instruction *I) { return true; } 265 266 bool isTypeLegal(Type *Ty) { return false; } 267 268 unsigned getJumpBufAlignment() { return 0; } 269 270 unsigned getJumpBufSize() { return 0; } 271 272 bool shouldBuildLookupTables() { return true; } 273 bool shouldBuildLookupTablesForConstant(Constant *C) { return true; } 274 275 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) { 276 return 0; 277 } 278 279 unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args, 280 unsigned VF) { return 0; } 281 282 bool supportsEfficientVectorElementLoadStore() { return false; } 283 284 bool enableAggressiveInterleaving(bool LoopHasReductions) { return false; } 285 286 bool expandMemCmp(Instruction *I, unsigned &MaxLoadSize) { return false; } 287 288 bool enableInterleavedAccessVectorization() { return false; } 289 290 bool isFPVectorizationPotentiallyUnsafe() { return false; } 291 292 bool allowsMisalignedMemoryAccesses(LLVMContext &Context, 293 unsigned BitWidth, 294 unsigned AddressSpace, 295 unsigned Alignment, 296 bool *Fast) { return false; } 297 298 TTI::PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) { 299 return TTI::PSK_Software; 300 } 301 302 bool haveFastSqrt(Type *Ty) { return false; } 303 304 unsigned getFPOpCost(Type *Ty) { return TargetTransformInfo::TCC_Basic; } 305 306 int getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, const APInt &Imm, 307 Type *Ty) { 308 return 0; 309 } 310 311 unsigned getIntImmCost(const APInt &Imm, Type *Ty) { return TTI::TCC_Basic; } 312 313 unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm, 314 Type *Ty) { 315 return TTI::TCC_Free; 316 } 317 318 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm, 319 Type *Ty) { 320 return TTI::TCC_Free; 321 } 322 323 unsigned getNumberOfRegisters(bool Vector) { return 8; } 324 325 unsigned getRegisterBitWidth(bool Vector) const { return 32; } 326 327 unsigned getMinVectorRegisterBitWidth() { return 128; } 328 329 bool 330 shouldConsiderAddressTypePromotion(const Instruction &I, 331 bool &AllowPromotionWithoutCommonHeader) { 332 AllowPromotionWithoutCommonHeader = false; 333 return false; 334 } 335 336 unsigned getCacheLineSize() { return 0; } 337 338 unsigned getPrefetchDistance() { return 0; } 339 340 unsigned getMinPrefetchStride() { return 1; } 341 342 unsigned getMaxPrefetchIterationsAhead() { return UINT_MAX; } 343 344 unsigned getMaxInterleaveFactor(unsigned VF) { return 1; } 345 346 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty, 347 TTI::OperandValueKind Opd1Info, 348 TTI::OperandValueKind Opd2Info, 349 TTI::OperandValueProperties Opd1PropInfo, 350 TTI::OperandValueProperties Opd2PropInfo, 351 ArrayRef<const Value *> Args) { 352 return 1; 353 } 354 355 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Ty, int Index, 356 Type *SubTp) { 357 return 1; 358 } 359 360 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src, 361 const Instruction *I) { return 1; } 362 363 unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst, 364 VectorType *VecTy, unsigned Index) { 365 return 1; 366 } 367 368 unsigned getCFInstrCost(unsigned Opcode) { return 1; } 369 370 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy, 371 const Instruction *I) { 372 return 1; 373 } 374 375 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) { 376 return 1; 377 } 378 379 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment, 380 unsigned AddressSpace, const Instruction *I) { 381 return 1; 382 } 383 384 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment, 385 unsigned AddressSpace) { 386 return 1; 387 } 388 389 unsigned getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr, 390 bool VariableMask, 391 unsigned Alignment) { 392 return 1; 393 } 394 395 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, 396 unsigned Factor, 397 ArrayRef<unsigned> Indices, 398 unsigned Alignment, 399 unsigned AddressSpace) { 400 return 1; 401 } 402 403 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 404 ArrayRef<Type *> Tys, FastMathFlags FMF, 405 unsigned ScalarizationCostPassed) { 406 return 1; 407 } 408 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 409 ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) { 410 return 1; 411 } 412 413 unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) { 414 return 1; 415 } 416 417 unsigned getNumberOfParts(Type *Tp) { return 0; } 418 419 unsigned getAddressComputationCost(Type *Tp, ScalarEvolution *, 420 const SCEV *) { 421 return 0; 422 } 423 424 unsigned getReductionCost(unsigned, Type *, bool) { return 1; } 425 426 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) { return 0; } 427 428 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) { 429 return false; 430 } 431 432 unsigned getAtomicMemIntrinsicMaxElementSize() const { 433 // Note for overrides: You must ensure for all element unordered-atomic 434 // memory intrinsics that all power-of-2 element sizes up to, and 435 // including, the return value of this method have a corresponding 436 // runtime lib call. These runtime lib call definitions can be found 437 // in RuntimeLibcalls.h 438 return 0; 439 } 440 441 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst, 442 Type *ExpectedType) { 443 return nullptr; 444 } 445 446 bool areInlineCompatible(const Function *Caller, 447 const Function *Callee) const { 448 return (Caller->getFnAttribute("target-cpu") == 449 Callee->getFnAttribute("target-cpu")) && 450 (Caller->getFnAttribute("target-features") == 451 Callee->getFnAttribute("target-features")); 452 } 453 454 unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const { return 128; } 455 456 bool isLegalToVectorizeLoad(LoadInst *LI) const { return true; } 457 458 bool isLegalToVectorizeStore(StoreInst *SI) const { return true; } 459 460 bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes, 461 unsigned Alignment, 462 unsigned AddrSpace) const { 463 return true; 464 } 465 466 bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes, 467 unsigned Alignment, 468 unsigned AddrSpace) const { 469 return true; 470 } 471 472 unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize, 473 unsigned ChainSizeInBytes, 474 VectorType *VecTy) const { 475 return VF; 476 } 477 478 unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize, 479 unsigned ChainSizeInBytes, 480 VectorType *VecTy) const { 481 return VF; 482 } 483 484 bool useReductionIntrinsic(unsigned Opcode, Type *Ty, 485 TTI::ReductionFlags Flags) const { 486 return false; 487 } 488 489 bool shouldExpandReduction(const IntrinsicInst *II) const { 490 return true; 491 } 492 493protected: 494 // Obtain the minimum required size to hold the value (without the sign) 495 // In case of a vector it returns the min required size for one element. 496 unsigned minRequiredElementSize(const Value* Val, bool &isSigned) { 497 if (isa<ConstantDataVector>(Val) || isa<ConstantVector>(Val)) { 498 const auto* VectorValue = cast<Constant>(Val); 499 500 // In case of a vector need to pick the max between the min 501 // required size for each element 502 auto *VT = cast<VectorType>(Val->getType()); 503 504 // Assume unsigned elements 505 isSigned = false; 506 507 // The max required size is the total vector width divided by num 508 // of elements in the vector 509 unsigned MaxRequiredSize = VT->getBitWidth() / VT->getNumElements(); 510 511 unsigned MinRequiredSize = 0; 512 for(unsigned i = 0, e = VT->getNumElements(); i < e; ++i) { 513 if (auto* IntElement = 514 dyn_cast<ConstantInt>(VectorValue->getAggregateElement(i))) { 515 bool signedElement = IntElement->getValue().isNegative(); 516 // Get the element min required size. 517 unsigned ElementMinRequiredSize = 518 IntElement->getValue().getMinSignedBits() - 1; 519 // In case one element is signed then all the vector is signed. 520 isSigned |= signedElement; 521 // Save the max required bit size between all the elements. 522 MinRequiredSize = std::max(MinRequiredSize, ElementMinRequiredSize); 523 } 524 else { 525 // not an int constant element 526 return MaxRequiredSize; 527 } 528 } 529 return MinRequiredSize; 530 } 531 532 if (const auto* CI = dyn_cast<ConstantInt>(Val)) { 533 isSigned = CI->getValue().isNegative(); 534 return CI->getValue().getMinSignedBits() - 1; 535 } 536 537 if (const auto* Cast = dyn_cast<SExtInst>(Val)) { 538 isSigned = true; 539 return Cast->getSrcTy()->getScalarSizeInBits() - 1; 540 } 541 542 if (const auto* Cast = dyn_cast<ZExtInst>(Val)) { 543 isSigned = false; 544 return Cast->getSrcTy()->getScalarSizeInBits(); 545 } 546 547 isSigned = false; 548 return Val->getType()->getScalarSizeInBits(); 549 } 550 551 bool isStridedAccess(const SCEV *Ptr) { 552 return Ptr && isa<SCEVAddRecExpr>(Ptr); 553 } 554 555 const SCEVConstant *getConstantStrideStep(ScalarEvolution *SE, 556 const SCEV *Ptr) { 557 if (!isStridedAccess(Ptr)) 558 return nullptr; 559 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ptr); 560 return dyn_cast<SCEVConstant>(AddRec->getStepRecurrence(*SE)); 561 } 562 563 bool isConstantStridedAccessLessThan(ScalarEvolution *SE, const SCEV *Ptr, 564 int64_t MergeDistance) { 565 const SCEVConstant *Step = getConstantStrideStep(SE, Ptr); 566 if (!Step) 567 return false; 568 APInt StrideVal = Step->getAPInt(); 569 if (StrideVal.getBitWidth() > 64) 570 return false; 571 // FIXME: need to take absolute value for negtive stride case 572 return StrideVal.getSExtValue() < MergeDistance; 573 } 574}; 575 576/// \brief CRTP base class for use as a mix-in that aids implementing 577/// a TargetTransformInfo-compatible class. 578template <typename T> 579class TargetTransformInfoImplCRTPBase : public TargetTransformInfoImplBase { 580private: 581 typedef TargetTransformInfoImplBase BaseT; 582 583protected: 584 explicit TargetTransformInfoImplCRTPBase(const DataLayout &DL) : BaseT(DL) {} 585 586public: 587 using BaseT::getCallCost; 588 589 unsigned getCallCost(const Function *F, int NumArgs) { 590 assert(F && "A concrete function must be provided to this routine."); 591 592 if (NumArgs < 0) 593 // Set the argument number to the number of explicit arguments in the 594 // function. 595 NumArgs = F->arg_size(); 596 597 if (Intrinsic::ID IID = F->getIntrinsicID()) { 598 FunctionType *FTy = F->getFunctionType(); 599 SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end()); 600 return static_cast<T *>(this) 601 ->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys); 602 } 603 604 if (!static_cast<T *>(this)->isLoweredToCall(F)) 605 return TTI::TCC_Basic; // Give a basic cost if it will be lowered 606 // directly. 607 608 return static_cast<T *>(this)->getCallCost(F->getFunctionType(), NumArgs); 609 } 610 611 unsigned getCallCost(const Function *F, ArrayRef<const Value *> Arguments) { 612 // Simply delegate to generic handling of the call. 613 // FIXME: We should use instsimplify or something else to catch calls which 614 // will constant fold with these arguments. 615 return static_cast<T *>(this)->getCallCost(F, Arguments.size()); 616 } 617 618 using BaseT::getGEPCost; 619 620 int getGEPCost(Type *PointeeType, const Value *Ptr, 621 ArrayRef<const Value *> Operands) { 622 const GlobalValue *BaseGV = nullptr; 623 if (Ptr != nullptr) { 624 // TODO: will remove this when pointers have an opaque type. 625 assert(Ptr->getType()->getScalarType()->getPointerElementType() == 626 PointeeType && 627 "explicit pointee type doesn't match operand's pointee type"); 628 BaseGV = dyn_cast<GlobalValue>(Ptr->stripPointerCasts()); 629 } 630 bool HasBaseReg = (BaseGV == nullptr); 631 int64_t BaseOffset = 0; 632 int64_t Scale = 0; 633 634 auto GTI = gep_type_begin(PointeeType, Operands); 635 Type *TargetType; 636 for (auto I = Operands.begin(); I != Operands.end(); ++I, ++GTI) { 637 TargetType = GTI.getIndexedType(); 638 // We assume that the cost of Scalar GEP with constant index and the 639 // cost of Vector GEP with splat constant index are the same. 640 const ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I); 641 if (!ConstIdx) 642 if (auto Splat = getSplatValue(*I)) 643 ConstIdx = dyn_cast<ConstantInt>(Splat); 644 if (StructType *STy = GTI.getStructTypeOrNull()) { 645 // For structures the index is always splat or scalar constant 646 assert(ConstIdx && "Unexpected GEP index"); 647 uint64_t Field = ConstIdx->getZExtValue(); 648 BaseOffset += DL.getStructLayout(STy)->getElementOffset(Field); 649 } else { 650 int64_t ElementSize = DL.getTypeAllocSize(GTI.getIndexedType()); 651 if (ConstIdx) 652 BaseOffset += ConstIdx->getSExtValue() * ElementSize; 653 else { 654 // Needs scale register. 655 if (Scale != 0) 656 // No addressing mode takes two scale registers. 657 return TTI::TCC_Basic; 658 Scale = ElementSize; 659 } 660 } 661 } 662 663 // Assumes the address space is 0 when Ptr is nullptr. 664 unsigned AS = 665 (Ptr == nullptr ? 0 : Ptr->getType()->getPointerAddressSpace()); 666 if (static_cast<T *>(this)->isLegalAddressingMode( 667 TargetType, const_cast<GlobalValue *>(BaseGV), BaseOffset, 668 HasBaseReg, Scale, AS)) 669 return TTI::TCC_Free; 670 return TTI::TCC_Basic; 671 } 672 673 using BaseT::getIntrinsicCost; 674 675 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy, 676 ArrayRef<const Value *> Arguments) { 677 // Delegate to the generic intrinsic handling code. This mostly provides an 678 // opportunity for targets to (for example) special case the cost of 679 // certain intrinsics based on constants used as arguments. 680 SmallVector<Type *, 8> ParamTys; 681 ParamTys.reserve(Arguments.size()); 682 for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx) 683 ParamTys.push_back(Arguments[Idx]->getType()); 684 return static_cast<T *>(this)->getIntrinsicCost(IID, RetTy, ParamTys); 685 } 686 687 unsigned getUserCost(const User *U) { 688 if (isa<PHINode>(U)) 689 return TTI::TCC_Free; // Model all PHI nodes as free. 690 691 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U)) { 692 SmallVector<Value *, 4> Indices(GEP->idx_begin(), GEP->idx_end()); 693 return static_cast<T *>(this)->getGEPCost( 694 GEP->getSourceElementType(), GEP->getPointerOperand(), Indices); 695 } 696 697 if (auto CS = ImmutableCallSite(U)) { 698 const Function *F = CS.getCalledFunction(); 699 if (!F) { 700 // Just use the called value type. 701 Type *FTy = CS.getCalledValue()->getType()->getPointerElementType(); 702 return static_cast<T *>(this) 703 ->getCallCost(cast<FunctionType>(FTy), CS.arg_size()); 704 } 705 706 SmallVector<const Value *, 8> Arguments(CS.arg_begin(), CS.arg_end()); 707 return static_cast<T *>(this)->getCallCost(F, Arguments); 708 } 709 710 if (const CastInst *CI = dyn_cast<CastInst>(U)) { 711 // Result of a cmp instruction is often extended (to be used by other 712 // cmp instructions, logical or return instructions). These are usually 713 // nop on most sane targets. 714 if (isa<CmpInst>(CI->getOperand(0))) 715 return TTI::TCC_Free; 716 } 717 718 return static_cast<T *>(this)->getOperationCost( 719 Operator::getOpcode(U), U->getType(), 720 U->getNumOperands() == 1 ? U->getOperand(0)->getType() : nullptr); 721 } 722}; 723} 724 725#endif 726