include/llvm/Analysis/TargetTransformInfoImpl.h

   1 //===- TargetTransformInfoImpl.h --------------------------------*- C++ -*-===//
   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 /// \file
   9 /// This file provides helpers for the implementation of
  10 /// a TargetTransformInfo-conforming class.
  11 ///
  12 //===----------------------------------------------------------------------===//
  13
  14 #ifndef LLVM_ANALYSIS_TARGETTRANSFORMINFOIMPL_H
  15 #define LLVM_ANALYSIS_TARGETTRANSFORMINFOIMPL_H
  16
  17 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
  18 #include "llvm/Analysis/TargetTransformInfo.h"
  19 #include "llvm/Analysis/VectorUtils.h"
  20 #include "llvm/IR/CallSite.h"
  21 #include "llvm/IR/DataLayout.h"
  22 #include "llvm/IR/Function.h"
  23 #include "llvm/IR/GetElementPtrTypeIterator.h"
  24 #include "llvm/IR/Operator.h"
  25 #include "llvm/IR/Type.h"
  26
  27 namespace llvm {
  28
  29 /// Base class for use as a mix-in that aids implementing
  30 /// a TargetTransformInfo-compatible class.
  31 class TargetTransformInfoImplBase {
  32 protected:
  33   typedef TargetTransformInfo TTI;
  34
  35   const DataLayout &DL;
  36
  37   explicit TargetTransformInfoImplBase(const DataLayout &DL) : DL(DL) {}
  38
  39 public:
  40   // Provide value semantics. MSVC requires that we spell all of these out.
  41   TargetTransformInfoImplBase(const TargetTransformInfoImplBase &Arg)
  42       : DL(Arg.DL) {}
  43   TargetTransformInfoImplBase(TargetTransformInfoImplBase &&Arg) : DL(Arg.DL) {}
  44
  45   const DataLayout &getDataLayout() const { return DL; }
  46
  47   unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) {
  48     switch (Opcode) {
  49     default:
  50       // By default, just classify everything as 'basic'.
  51       return TTI::TCC_Basic;
  52
  53     case Instruction::GetElementPtr:
  54       llvm_unreachable("Use getGEPCost for GEP operations!");
  55
  56     case Instruction::BitCast:
  57       assert(OpTy && "Cast instructions must provide the operand type");
  58       if (Ty == OpTy || (Ty->isPointerTy() && OpTy->isPointerTy()))
  59         // Identity and pointer-to-pointer casts are free.
  60         return TTI::TCC_Free;
  61
  62       // Otherwise, the default basic cost is used.
  63       return TTI::TCC_Basic;
  64
  65     case Instruction::FDiv:
  66     case Instruction::FRem:
  67     case Instruction::SDiv:
  68     case Instruction::SRem:
  69     case Instruction::UDiv:
  70     case Instruction::URem:
  71       return TTI::TCC_Expensive;
  72
  73     case Instruction::IntToPtr: {
  74       // An inttoptr cast is free so long as the input is a legal integer type
  75       // which doesn't contain values outside the range of a pointer.
  76       unsigned OpSize = OpTy->getScalarSizeInBits();
  77       if (DL.isLegalInteger(OpSize) &&
  78           OpSize <= DL.getPointerTypeSizeInBits(Ty))
  79         return TTI::TCC_Free;
  80
  81       // Otherwise it's not a no-op.
  82       return TTI::TCC_Basic;
  83     }
  84     case Instruction::PtrToInt: {
  85       // A ptrtoint cast is free so long as the result is large enough to store
  86       // the pointer, and a legal integer type.
  87       unsigned DestSize = Ty->getScalarSizeInBits();
  88       if (DL.isLegalInteger(DestSize) &&
  89           DestSize >= DL.getPointerTypeSizeInBits(OpTy))
  90         return TTI::TCC_Free;
  91
  92       // Otherwise it's not a no-op.
  93       return TTI::TCC_Basic;
  94     }
  95     case Instruction::Trunc:
  96       // trunc to a native type is free (assuming the target has compare and
  97       // shift-right of the same width).
  98       if (DL.isLegalInteger(DL.getTypeSizeInBits(Ty)))
  99         return TTI::TCC_Free;
 100
 101       return TTI::TCC_Basic;
 102     }
 103   }
 104
 105   int getGEPCost(Type *PointeeType, const Value *Ptr,
 106                  ArrayRef<const Value *> Operands) {
 107     // In the basic model, we just assume that all-constant GEPs will be folded
 108     // into their uses via addressing modes.
 109     for (unsigned Idx = 0, Size = Operands.size(); Idx != Size; ++Idx)
 110       if (!isa<Constant>(Operands[Idx]))
 111         return TTI::TCC_Basic;
 112
 113     return TTI::TCC_Free;
 114   }
 115
 116   unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
 117                                             unsigned &JTSize) {
 118     JTSize = 0;
 119     return SI.getNumCases();
 120   }
 121
 122   int getExtCost(const Instruction *I, const Value *Src) {
 123     return TTI::TCC_Basic;
 124   }
 125
 126   unsigned getCallCost(FunctionType *FTy, int NumArgs, const User *U) {
 127     assert(FTy && "FunctionType must be provided to this routine.");
 128
 129     // The target-independent implementation just measures the size of the
 130     // function by approximating that each argument will take on average one
 131     // instruction to prepare.
 132
 133     if (NumArgs < 0)
 134       // Set the argument number to the number of explicit arguments in the
 135       // function.
 136       NumArgs = FTy->getNumParams();
 137
 138     return TTI::TCC_Basic * (NumArgs + 1);
 139   }
 140
 141   unsigned getInliningThresholdMultiplier() { return 1; }
 142
 143   int getInlinerVectorBonusPercent() { return 150; }
 144
 145   unsigned getMemcpyCost(const Instruction *I) {
 146     return TTI::TCC_Expensive;
 147   }
 148
 149   bool hasBranchDivergence() { return false; }
 150
 151   bool isSourceOfDivergence(const Value *V) { return false; }
 152
 153   bool isAlwaysUniform(const Value *V) { return false; }
 154
 155   unsigned getFlatAddressSpace () {
 156     return -1;
 157   }
 158
 159   bool collectFlatAddressOperands(SmallVectorImpl<int> &OpIndexes,
 160                                   Intrinsic::ID IID) const {
 161     return false;
 162   }
 163
 164   bool rewriteIntrinsicWithAddressSpace(IntrinsicInst *II,
 165                                         Value *OldV, Value *NewV) const {
 166     return false;
 167   }
 168
 169   bool isLoweredToCall(const Function *F) {
 170     assert(F && "A concrete function must be provided to this routine.");
 171
 172     // FIXME: These should almost certainly not be handled here, and instead
 173     // handled with the help of TLI or the target itself. This was largely
 174     // ported from existing analysis heuristics here so that such refactorings
 175     // can take place in the future.
 176
 177     if (F->isIntrinsic())
 178       return false;
 179
 180     if (F->hasLocalLinkage() || !F->hasName())
 181       return true;
 182
 183     StringRef Name = F->getName();
 184
 185     // These will all likely lower to a single selection DAG node.
 186     if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
 187         Name == "fabs" || Name == "fabsf" || Name == "fabsl" || Name == "sin" ||
 188         Name == "fmin" || Name == "fminf" || Name == "fminl" ||
 189         Name == "fmax" || Name == "fmaxf" || Name == "fmaxl" ||
 190         Name == "sinf" || Name == "sinl" || Name == "cos" || Name == "cosf" ||
 191         Name == "cosl" || Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl")
 192       return false;
 193
 194     // These are all likely to be optimized into something smaller.
 195     if (Name == "pow" || Name == "powf" || Name == "powl" || Name == "exp2" ||
 196         Name == "exp2l" || Name == "exp2f" || Name == "floor" ||
 197         Name == "floorf" || Name == "ceil" || Name == "round" ||
 198         Name == "ffs" || Name == "ffsl" || Name == "abs" || Name == "labs" ||
 199         Name == "llabs")
 200       return false;
 201
 202     return true;
 203   }
 204
 205   bool isHardwareLoopProfitable(Loop *L, ScalarEvolution &SE,
 206                                 AssumptionCache &AC,
 207                                 TargetLibraryInfo *LibInfo,
 208                                 HardwareLoopInfo &HWLoopInfo) {
 209     return false;
 210   }
 211
 212   void getUnrollingPreferences(Loop *, ScalarEvolution &,
 213                                TTI::UnrollingPreferences &) {}
 214
 215   bool isLegalAddImmediate(int64_t Imm) { return false; }
 216
 217   bool isLegalICmpImmediate(int64_t Imm) { return false; }
 218
 219   bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
 220                              bool HasBaseReg, int64_t Scale,
 221                              unsigned AddrSpace, Instruction *I = nullptr) {
 222     // Guess that only reg and reg+reg addressing is allowed. This heuristic is
 223     // taken from the implementation of LSR.
 224     return !BaseGV && BaseOffset == 0 && (Scale == 0 || Scale == 1);
 225   }
 226
 227   bool isLSRCostLess(TTI::LSRCost &C1, TTI::LSRCost &C2) {
 228     return std::tie(C1.NumRegs, C1.AddRecCost, C1.NumIVMuls, C1.NumBaseAdds,
 229                     C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
 230            std::tie(C2.NumRegs, C2.AddRecCost, C2.NumIVMuls, C2.NumBaseAdds,
 231                     C2.ScaleCost, C2.ImmCost, C2.SetupCost);
 232   }
 233
 234   bool canMacroFuseCmp() { return false; }
 235
 236   bool canSaveCmp(Loop *L, BranchInst **BI, ScalarEvolution *SE, LoopInfo *LI,
 237                   DominatorTree *DT, AssumptionCache *AC,
 238                   TargetLibraryInfo *LibInfo) {
 239     return false;
 240   }
 241
 242   bool shouldFavorPostInc() const { return false; }
 243
 244   bool shouldFavorBackedgeIndex(const Loop *L) const { return false; }
 245
 246   bool isLegalMaskedStore(Type *DataType) { return false; }
 247
 248   bool isLegalMaskedLoad(Type *DataType) { return false; }
 249
 250   bool isLegalNTStore(Type *DataType, llvm::Align Alignment) {
 251     // By default, assume nontemporal memory stores are available for stores
 252     // that are aligned and have a size that is a power of 2.
 253     unsigned DataSize = DL.getTypeStoreSize(DataType);
 254     return Alignment >= DataSize && isPowerOf2_32(DataSize);
 255   }
 256
 257   bool isLegalNTLoad(Type *DataType, llvm::Align Alignment) {
 258     // By default, assume nontemporal memory loads are available for loads that
 259     // are aligned and have a size that is a power of 2.
 260     unsigned DataSize = DL.getTypeStoreSize(DataType);
 261     return Alignment >= DataSize && isPowerOf2_32(DataSize);
 262   }
 263
 264   bool isLegalMaskedScatter(Type *DataType) { return false; }
 265
 266   bool isLegalMaskedGather(Type *DataType) { return false; }
 267
 268   bool isLegalMaskedCompressStore(Type *DataType) { return false; }
 269
 270   bool isLegalMaskedExpandLoad(Type *DataType) { return false; }
 271
 272   bool hasDivRemOp(Type *DataType, bool IsSigned) { return false; }
 273
 274   bool hasVolatileVariant(Instruction *I, unsigned AddrSpace) { return false; }
 275
 276   bool prefersVectorizedAddressing() { return true; }
 277
 278   int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
 279                            bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
 280     // Guess that all legal addressing mode are free.
 281     if (isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
 282                               Scale, AddrSpace))
 283       return 0;
 284     return -1;
 285   }
 286
 287   bool LSRWithInstrQueries() { return false; }
 288
 289   bool isTruncateFree(Type *Ty1, Type *Ty2) { return false; }
 290
 291   bool isProfitableToHoist(Instruction *I) { return true; }
 292
 293   bool useAA() { return false; }
 294
 295   bool isTypeLegal(Type *Ty) { return false; }
 296
 297   unsigned getJumpBufAlignment() { return 0; }
 298
 299   unsigned getJumpBufSize() { return 0; }
 300
 301   bool shouldBuildLookupTables() { return true; }
 302   bool shouldBuildLookupTablesForConstant(Constant *C) { return true; }
 303
 304   bool useColdCCForColdCall(Function &F) { return false; }
 305
 306   unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
 307     return 0;
 308   }
 309
 310   unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args,
 311                                             unsigned VF) { return 0; }
 312
 313   bool supportsEfficientVectorElementLoadStore() { return false; }
 314
 315   bool enableAggressiveInterleaving(bool LoopHasReductions) { return false; }
 316
 317   TTI::MemCmpExpansionOptions enableMemCmpExpansion(bool OptSize,
 318                                                     bool IsZeroCmp) const {
 319     return {};
 320   }
 321
 322   bool enableInterleavedAccessVectorization() { return false; }
 323
 324   bool enableMaskedInterleavedAccessVectorization() { return false; }
 325
 326   bool isFPVectorizationPotentiallyUnsafe() { return false; }
 327
 328   bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
 329                                       unsigned BitWidth,
 330                                       unsigned AddressSpace,
 331                                       unsigned Alignment,
 332                                       bool *Fast) { return false; }
 333
 334   TTI::PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) {
 335     return TTI::PSK_Software;
 336   }
 337
 338   bool haveFastSqrt(Type *Ty) { return false; }
 339
 340   bool isFCmpOrdCheaperThanFCmpZero(Type *Ty) { return true; }
 341
 342   unsigned getFPOpCost(Type *Ty) { return TargetTransformInfo::TCC_Basic; }
 343
 344   int getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
 345                             Type *Ty) {
 346     return 0;
 347   }
 348
 349   unsigned getIntImmCost(const APInt &Imm, Type *Ty) { return TTI::TCC_Basic; }
 350
 351   unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
 352                          Type *Ty) {
 353     return TTI::TCC_Free;
 354   }
 355
 356   unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
 357                          Type *Ty) {
 358     return TTI::TCC_Free;
 359   }
 360
 361   unsigned getNumberOfRegisters(bool Vector) { return 8; }
 362
 363   unsigned getRegisterBitWidth(bool Vector) const { return 32; }
 364
 365   unsigned getMinVectorRegisterBitWidth() { return 128; }
 366
 367   bool shouldMaximizeVectorBandwidth(bool OptSize) const { return false; }
 368
 369   unsigned getMinimumVF(unsigned ElemWidth) const { return 0; }
 370
 371   bool
 372   shouldConsiderAddressTypePromotion(const Instruction &I,
 373                                      bool &AllowPromotionWithoutCommonHeader) {
 374     AllowPromotionWithoutCommonHeader = false;
 375     return false;
 376   }
 377
 378   unsigned getCacheLineSize() { return 0; }
 379
 380   llvm::Optional<unsigned> getCacheSize(TargetTransformInfo::CacheLevel Level) {
 381     switch (Level) {
 382     case TargetTransformInfo::CacheLevel::L1D:
 383       LLVM_FALLTHROUGH;
 384     case TargetTransformInfo::CacheLevel::L2D:
 385       return llvm::Optional<unsigned>();
 386     }
 387
 388     llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
 389   }
 390
 391   llvm::Optional<unsigned> getCacheAssociativity(
 392     TargetTransformInfo::CacheLevel Level) {
 393     switch (Level) {
 394     case TargetTransformInfo::CacheLevel::L1D:
 395       LLVM_FALLTHROUGH;
 396     case TargetTransformInfo::CacheLevel::L2D:
 397       return llvm::Optional<unsigned>();
 398     }
 399
 400     llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
 401   }
 402
 403   unsigned getPrefetchDistance() { return 0; }
 404
 405   unsigned getMinPrefetchStride() { return 1; }
 406
 407   unsigned getMaxPrefetchIterationsAhead() { return UINT_MAX; }
 408
 409   unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
 410
 411   unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
 412                                   TTI::OperandValueKind Opd1Info,
 413                                   TTI::OperandValueKind Opd2Info,
 414                                   TTI::OperandValueProperties Opd1PropInfo,
 415                                   TTI::OperandValueProperties Opd2PropInfo,
 416                                   ArrayRef<const Value *> Args) {
 417     return 1;
 418   }
 419
 420   unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Ty, int Index,
 421                           Type *SubTp) {
 422     return 1;
 423   }
 424
 425   unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
 426                             const Instruction *I) { return 1; }
 427
 428   unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst,
 429                                     VectorType *VecTy, unsigned Index) {
 430     return 1;
 431   }
 432
 433   unsigned getCFInstrCost(unsigned Opcode) { return 1; }
 434
 435   unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
 436                               const Instruction *I) {
 437     return 1;
 438   }
 439
 440   unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
 441     return 1;
 442   }
 443
 444   unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
 445                            unsigned AddressSpace, const Instruction *I) {
 446     return 1;
 447   }
 448
 449   unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
 450                                  unsigned AddressSpace) {
 451     return 1;
 452   }
 453
 454   unsigned getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
 455                                   bool VariableMask,
 456                                   unsigned Alignment) {
 457     return 1;
 458   }
 459
 460   unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
 461                                       unsigned Factor,
 462                                       ArrayRef<unsigned> Indices,
 463                                       unsigned Alignment, unsigned AddressSpace,
 464                                       bool UseMaskForCond = false,
 465                                       bool UseMaskForGaps = false) {
 466     return 1;
 467   }
 468
 469   unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
 470                                  ArrayRef<Type *> Tys, FastMathFlags FMF,
 471                                  unsigned ScalarizationCostPassed) {
 472     return 1;
 473   }
 474   unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
 475             ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) {
 476     return 1;
 477   }
 478
 479   unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
 480     return 1;
 481   }
 482
 483   unsigned getNumberOfParts(Type *Tp) { return 0; }
 484
 485   unsigned getAddressComputationCost(Type *Tp, ScalarEvolution *,
 486                                      const SCEV *) {
 487     return 0;
 488   }
 489
 490   unsigned getArithmeticReductionCost(unsigned, Type *, bool) { return 1; }
 491
 492   unsigned getMinMaxReductionCost(Type *, Type *, bool, bool) { return 1; }
 493
 494   unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) { return 0; }
 495
 496   bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) {
 497     return false;
 498   }
 499
 500   unsigned getAtomicMemIntrinsicMaxElementSize() const {
 501     // Note for overrides: You must ensure for all element unordered-atomic
 502     // memory intrinsics that all power-of-2 element sizes up to, and
 503     // including, the return value of this method have a corresponding
 504     // runtime lib call. These runtime lib call definitions can be found
 505     // in RuntimeLibcalls.h
 506     return 0;
 507   }
 508
 509   Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
 510                                            Type *ExpectedType) {
 511     return nullptr;
 512   }
 513
 514   Type *getMemcpyLoopLoweringType(LLVMContext &Context, Value *Length,
 515                                   unsigned SrcAlign, unsigned DestAlign) const {
 516     return Type::getInt8Ty(Context);
 517   }
 518
 519   void getMemcpyLoopResidualLoweringType(SmallVectorImpl<Type *> &OpsOut,
 520                                          LLVMContext &Context,
 521                                          unsigned RemainingBytes,
 522                                          unsigned SrcAlign,
 523                                          unsigned DestAlign) const {
 524     for (unsigned i = 0; i != RemainingBytes; ++i)
 525       OpsOut.push_back(Type::getInt8Ty(Context));
 526   }
 527
 528   bool areInlineCompatible(const Function *Caller,
 529                            const Function *Callee) const {
 530     return (Caller->getFnAttribute("target-cpu") ==
 531             Callee->getFnAttribute("target-cpu")) &&
 532            (Caller->getFnAttribute("target-features") ==
 533             Callee->getFnAttribute("target-features"));
 534   }
 535
 536   bool areFunctionArgsABICompatible(const Function *Caller, const Function *Callee,
 537                                     SmallPtrSetImpl<Argument *> &Args) const {
 538     return (Caller->getFnAttribute("target-cpu") ==
 539             Callee->getFnAttribute("target-cpu")) &&
 540            (Caller->getFnAttribute("target-features") ==
 541             Callee->getFnAttribute("target-features"));
 542   }
 543
 544   bool isIndexedLoadLegal(TTI::MemIndexedMode Mode, Type *Ty,
 545                           const DataLayout &DL) const {
 546     return false;
 547   }
 548
 549   bool isIndexedStoreLegal(TTI::MemIndexedMode Mode, Type *Ty,
 550                            const DataLayout &DL) const {
 551     return false;
 552   }
 553
 554   unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const { return 128; }
 555
 556   bool isLegalToVectorizeLoad(LoadInst *LI) const { return true; }
 557
 558   bool isLegalToVectorizeStore(StoreInst *SI) const { return true; }
 559
 560   bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
 561                                    unsigned Alignment,
 562                                    unsigned AddrSpace) const {
 563     return true;
 564   }
 565
 566   bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
 567                                     unsigned Alignment,
 568                                     unsigned AddrSpace) const {
 569     return true;
 570   }
 571
 572   unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
 573                                unsigned ChainSizeInBytes,
 574                                VectorType *VecTy) const {
 575     return VF;
 576   }
 577
 578   unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
 579                                 unsigned ChainSizeInBytes,
 580                                 VectorType *VecTy) const {
 581     return VF;
 582   }
 583
 584   bool useReductionIntrinsic(unsigned Opcode, Type *Ty,
 585                              TTI::ReductionFlags Flags) const {
 586     return false;
 587   }
 588
 589   bool shouldExpandReduction(const IntrinsicInst *II) const {
 590     return true;
 591   }
 592
 593   unsigned getGISelRematGlobalCost() const {
 594     return 1;
 595   }
 596
 597 protected:
 598   // Obtain the minimum required size to hold the value (without the sign)
 599   // In case of a vector it returns the min required size for one element.
 600   unsigned minRequiredElementSize(const Value* Val, bool &isSigned) {
 601     if (isa<ConstantDataVector>(Val) || isa<ConstantVector>(Val)) {
 602       const auto* VectorValue = cast<Constant>(Val);
 603
 604       // In case of a vector need to pick the max between the min
 605       // required size for each element
 606       auto *VT = cast<VectorType>(Val->getType());
 607
 608       // Assume unsigned elements
 609       isSigned = false;
 610
 611       // The max required size is the total vector width divided by num
 612       // of elements in the vector
 613       unsigned MaxRequiredSize = VT->getBitWidth() / VT->getNumElements();
 614
 615       unsigned MinRequiredSize = 0;
 616       for(unsigned i = 0, e = VT->getNumElements(); i < e; ++i) {
 617         if (auto* IntElement =
 618               dyn_cast<ConstantInt>(VectorValue->getAggregateElement(i))) {
 619           bool signedElement = IntElement->getValue().isNegative();
 620           // Get the element min required size.
 621           unsigned ElementMinRequiredSize =
 622             IntElement->getValue().getMinSignedBits() - 1;
 623           // In case one element is signed then all the vector is signed.
 624           isSigned |= signedElement;
 625           // Save the max required bit size between all the elements.
 626           MinRequiredSize = std::max(MinRequiredSize, ElementMinRequiredSize);
 627         }
 628         else {
 629           // not an int constant element
 630           return MaxRequiredSize;
 631         }
 632       }
 633       return MinRequiredSize;
 634     }
 635
 636     if (const auto* CI = dyn_cast<ConstantInt>(Val)) {
 637       isSigned = CI->getValue().isNegative();
 638       return CI->getValue().getMinSignedBits() - 1;
 639     }
 640
 641     if (const auto* Cast = dyn_cast<SExtInst>(Val)) {
 642       isSigned = true;
 643       return Cast->getSrcTy()->getScalarSizeInBits() - 1;
 644     }
 645
 646     if (const auto* Cast = dyn_cast<ZExtInst>(Val)) {
 647       isSigned = false;
 648       return Cast->getSrcTy()->getScalarSizeInBits();
 649     }
 650
 651     isSigned = false;
 652     return Val->getType()->getScalarSizeInBits();
 653   }
 654
 655   bool isStridedAccess(const SCEV *Ptr) {
 656     return Ptr && isa<SCEVAddRecExpr>(Ptr);
 657   }
 658
 659   const SCEVConstant *getConstantStrideStep(ScalarEvolution *SE,
 660                                             const SCEV *Ptr) {
 661     if (!isStridedAccess(Ptr))
 662       return nullptr;
 663     const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ptr);
 664     return dyn_cast<SCEVConstant>(AddRec->getStepRecurrence(*SE));
 665   }
 666
 667   bool isConstantStridedAccessLessThan(ScalarEvolution *SE, const SCEV *Ptr,
 668                                        int64_t MergeDistance) {
 669     const SCEVConstant *Step = getConstantStrideStep(SE, Ptr);
 670     if (!Step)
 671       return false;
 672     APInt StrideVal = Step->getAPInt();
 673     if (StrideVal.getBitWidth() > 64)
 674       return false;
 675     // FIXME: Need to take absolute value for negative stride case.
 676     return StrideVal.getSExtValue() < MergeDistance;
 677   }
 678 };
 679
 680 /// CRTP base class for use as a mix-in that aids implementing
 681 /// a TargetTransformInfo-compatible class.
 682 template <typename T>
 683 class TargetTransformInfoImplCRTPBase : public TargetTransformInfoImplBase {
 684 private:
 685   typedef TargetTransformInfoImplBase BaseT;
 686
 687 protected:
 688   explicit TargetTransformInfoImplCRTPBase(const DataLayout &DL) : BaseT(DL) {}
 689
 690 public:
 691   using BaseT::getCallCost;
 692
 693   unsigned getCallCost(const Function *F, int NumArgs, const User *U) {
 694     assert(F && "A concrete function must be provided to this routine.");
 695
 696     if (NumArgs < 0)
 697       // Set the argument number to the number of explicit arguments in the
 698       // function.
 699       NumArgs = F->arg_size();
 700
 701     if (Intrinsic::ID IID = F->getIntrinsicID()) {
 702       FunctionType *FTy = F->getFunctionType();
 703       SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end());
 704       return static_cast<T *>(this)
 705           ->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys, U);
 706     }
 707
 708     if (!static_cast<T *>(this)->isLoweredToCall(F))
 709       return TTI::TCC_Basic; // Give a basic cost if it will be lowered
 710                              // directly.
 711
 712     return static_cast<T *>(this)->getCallCost(F->getFunctionType(), NumArgs, U);
 713   }
 714
 715   unsigned getCallCost(const Function *F, ArrayRef<const Value *> Arguments,
 716                        const User *U) {
 717     // Simply delegate to generic handling of the call.
 718     // FIXME: We should use instsimplify or something else to catch calls which
 719     // will constant fold with these arguments.
 720     return static_cast<T *>(this)->getCallCost(F, Arguments.size(), U);
 721   }
 722
 723   using BaseT::getGEPCost;
 724
 725   int getGEPCost(Type *PointeeType, const Value *Ptr,
 726                  ArrayRef<const Value *> Operands) {
 727     assert(PointeeType && Ptr && "can't get GEPCost of nullptr");
 728     // TODO: will remove this when pointers have an opaque type.
 729     assert(Ptr->getType()->getScalarType()->getPointerElementType() ==
 730                PointeeType &&
 731            "explicit pointee type doesn't match operand's pointee type");
 732     auto *BaseGV = dyn_cast<GlobalValue>(Ptr->stripPointerCasts());
 733     bool HasBaseReg = (BaseGV == nullptr);
 734
 735     auto PtrSizeBits = DL.getPointerTypeSizeInBits(Ptr->getType());
 736     APInt BaseOffset(PtrSizeBits, 0);
 737     int64_t Scale = 0;
 738
 739     auto GTI = gep_type_begin(PointeeType, Operands);
 740     Type *TargetType = nullptr;
 741
 742     // Handle the case where the GEP instruction has a single operand,
 743     // the basis, therefore TargetType is a nullptr.
 744     if (Operands.empty())
 745       return !BaseGV ? TTI::TCC_Free : TTI::TCC_Basic;
 746
 747     for (auto I = Operands.begin(); I != Operands.end(); ++I, ++GTI) {
 748       TargetType = GTI.getIndexedType();
 749       // We assume that the cost of Scalar GEP with constant index and the
 750       // cost of Vector GEP with splat constant index are the same.
 751       const ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I);
 752       if (!ConstIdx)
 753         if (auto Splat = getSplatValue(*I))
 754           ConstIdx = dyn_cast<ConstantInt>(Splat);
 755       if (StructType *STy = GTI.getStructTypeOrNull()) {
 756         // For structures the index is always splat or scalar constant
 757         assert(ConstIdx && "Unexpected GEP index");
 758         uint64_t Field = ConstIdx->getZExtValue();
 759         BaseOffset += DL.getStructLayout(STy)->getElementOffset(Field);
 760       } else {
 761         int64_t ElementSize = DL.getTypeAllocSize(GTI.getIndexedType());
 762         if (ConstIdx) {
 763           BaseOffset +=
 764               ConstIdx->getValue().sextOrTrunc(PtrSizeBits) * ElementSize;
 765         } else {
 766           // Needs scale register.
 767           if (Scale != 0)
 768             // No addressing mode takes two scale registers.
 769             return TTI::TCC_Basic;
 770           Scale = ElementSize;
 771         }
 772       }
 773     }
 774
 775     if (static_cast<T *>(this)->isLegalAddressingMode(
 776             TargetType, const_cast<GlobalValue *>(BaseGV),
 777             BaseOffset.sextOrTrunc(64).getSExtValue(), HasBaseReg, Scale,
 778             Ptr->getType()->getPointerAddressSpace()))
 779       return TTI::TCC_Free;
 780     return TTI::TCC_Basic;
 781   }
 782
 783   unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
 784                             ArrayRef<Type *> ParamTys, const User *U) {
 785     switch (IID) {
 786     default:
 787       // Intrinsics rarely (if ever) have normal argument setup constraints.
 788       // Model them as having a basic instruction cost.
 789       return TTI::TCC_Basic;
 790
 791     // TODO: other libc intrinsics.
 792     case Intrinsic::memcpy:
 793       return static_cast<T *>(this)->getMemcpyCost(dyn_cast<Instruction>(U));
 794
 795     case Intrinsic::annotation:
 796     case Intrinsic::assume:
 797     case Intrinsic::sideeffect:
 798     case Intrinsic::dbg_declare:
 799     case Intrinsic::dbg_value:
 800     case Intrinsic::dbg_label:
 801     case Intrinsic::invariant_start:
 802     case Intrinsic::invariant_end:
 803     case Intrinsic::launder_invariant_group:
 804     case Intrinsic::strip_invariant_group:
 805     case Intrinsic::is_constant:
 806     case Intrinsic::lifetime_start:
 807     case Intrinsic::lifetime_end:
 808     case Intrinsic::objectsize:
 809     case Intrinsic::ptr_annotation:
 810     case Intrinsic::var_annotation:
 811     case Intrinsic::experimental_gc_result:
 812     case Intrinsic::experimental_gc_relocate:
 813     case Intrinsic::coro_alloc:
 814     case Intrinsic::coro_begin:
 815     case Intrinsic::coro_free:
 816     case Intrinsic::coro_end:
 817     case Intrinsic::coro_frame:
 818     case Intrinsic::coro_size:
 819     case Intrinsic::coro_suspend:
 820     case Intrinsic::coro_param:
 821     case Intrinsic::coro_subfn_addr:
 822       // These intrinsics don't actually represent code after lowering.
 823       return TTI::TCC_Free;
 824     }
 825   }
 826
 827   unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
 828                             ArrayRef<const Value *> Arguments, const User *U) {
 829     // Delegate to the generic intrinsic handling code. This mostly provides an
 830     // opportunity for targets to (for example) special case the cost of
 831     // certain intrinsics based on constants used as arguments.
 832     SmallVector<Type *, 8> ParamTys;
 833     ParamTys.reserve(Arguments.size());
 834     for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx)
 835       ParamTys.push_back(Arguments[Idx]->getType());
 836     return static_cast<T *>(this)->getIntrinsicCost(IID, RetTy, ParamTys, U);
 837   }
 838
 839   unsigned getUserCost(const User *U, ArrayRef<const Value *> Operands) {
 840     if (isa<PHINode>(U))
 841       return TTI::TCC_Free; // Model all PHI nodes as free.
 842
 843     if (isa<ExtractValueInst>(U))
 844       return TTI::TCC_Free; // Model all ExtractValue nodes as free.
 845
 846     // Static alloca doesn't generate target instructions.
 847     if (auto *A = dyn_cast<AllocaInst>(U))
 848       if (A->isStaticAlloca())
 849         return TTI::TCC_Free;
 850
 851     if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U)) {
 852       return static_cast<T *>(this)->getGEPCost(GEP->getSourceElementType(),
 853                                                 GEP->getPointerOperand(),
 854                                                 Operands.drop_front());
 855     }
 856
 857     if (auto CS = ImmutableCallSite(U)) {
 858       const Function *F = CS.getCalledFunction();
 859       if (!F) {
 860         // Just use the called value type.
 861         Type *FTy = CS.getCalledValue()->getType()->getPointerElementType();
 862         return static_cast<T *>(this)
 863             ->getCallCost(cast<FunctionType>(FTy), CS.arg_size(), U);
 864       }
 865
 866       SmallVector<const Value *, 8> Arguments(CS.arg_begin(), CS.arg_end());
 867       return static_cast<T *>(this)->getCallCost(F, Arguments, U);
 868     }
 869
 870     if (isa<SExtInst>(U) || isa<ZExtInst>(U) || isa<FPExtInst>(U))
 871       // The old behaviour of generally treating extensions of icmp to be free
 872       // has been removed. A target that needs it should override getUserCost().
 873       return static_cast<T *>(this)->getExtCost(cast<Instruction>(U),
 874                                                 Operands.back());
 875
 876     return static_cast<T *>(this)->getOperationCost(
 877         Operator::getOpcode(U), U->getType(),
 878         U->getNumOperands() == 1 ? U->getOperand(0)->getType() : nullptr);
 879   }
 880
 881   int getInstructionLatency(const Instruction *I) {
 882     SmallVector<const Value *, 4> Operands(I->value_op_begin(),
 883                                            I->value_op_end());
 884     if (getUserCost(I, Operands) == TTI::TCC_Free)
 885       return 0;
 886
 887     if (isa<LoadInst>(I))
 888       return 4;
 889
 890     Type *DstTy = I->getType();
 891
 892     // Usually an intrinsic is a simple instruction.
 893     // A real function call is much slower.
 894     if (auto *CI = dyn_cast<CallInst>(I)) {
 895       const Function *F = CI->getCalledFunction();
 896       if (!F || static_cast<T *>(this)->isLoweredToCall(F))
 897         return 40;
 898       // Some intrinsics return a value and a flag, we use the value type
 899       // to decide its latency.
 900       if (StructType* StructTy = dyn_cast<StructType>(DstTy))
 901         DstTy = StructTy->getElementType(0);
 902       // Fall through to simple instructions.
 903     }
 904
 905     if (VectorType *VectorTy = dyn_cast<VectorType>(DstTy))
 906       DstTy = VectorTy->getElementType();
 907     if (DstTy->isFloatingPointTy())
 908       return 3;
 909
 910     return 1;
 911   }
 912 };
 913 }
 914
 915 #endif