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