TargetTransformInfoImpl.h revision 360784 
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
27namespace llvm {
28
29/// Base class for use as a mix-in that aids implementing
30/// a TargetTransformInfo-compatible class.
31class TargetTransformInfoImplBase {
32protected:
33  typedef TargetTransformInfo TTI;
34
35  const DataLayout &DL;
36
37  explicit TargetTransformInfoImplBase(const DataLayout &DL) : DL(DL) {}
38
39public:
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                                            ProfileSummaryInfo *PSI,
119                                            BlockFrequencyInfo *BFI) {
120    (void)PSI;
121    (void)BFI;
122    JTSize = 0;
123    return SI.getNumCases();
124  }
125
126  int getExtCost(const Instruction *I, const Value *Src) {
127    return TTI::TCC_Basic;
128  }
129
130  unsigned getCallCost(FunctionType *FTy, int NumArgs, const User *U) {
131    assert(FTy && "FunctionType must be provided to this routine.");
132
133    // The target-independent implementation just measures the size of the
134    // function by approximating that each argument will take on average one
135    // instruction to prepare.
136
137    if (NumArgs < 0)
138      // Set the argument number to the number of explicit arguments in the
139      // function.
140      NumArgs = FTy->getNumParams();
141
142    return TTI::TCC_Basic * (NumArgs + 1);
143  }
144
145  unsigned getInliningThresholdMultiplier() { return 1; }
146
147  int getInlinerVectorBonusPercent() { return 150; }
148
149  unsigned getMemcpyCost(const Instruction *I) {
150    return TTI::TCC_Expensive;
151  }
152
153  bool hasBranchDivergence() { return false; }
154
155  bool isSourceOfDivergence(const Value *V) { return false; }
156
157  bool isAlwaysUniform(const Value *V) { return false; }
158
159  unsigned getFlatAddressSpace () {
160    return -1;
161  }
162
163  bool collectFlatAddressOperands(SmallVectorImpl<int> &OpIndexes,
164                                  Intrinsic::ID IID) const {
165    return false;
166  }
167
168  bool rewriteIntrinsicWithAddressSpace(IntrinsicInst *II,
169                                        Value *OldV, Value *NewV) const {
170    return false;
171  }
172
173  bool isLoweredToCall(const Function *F) {
174    assert(F && "A concrete function must be provided to this routine.");
175
176    // FIXME: These should almost certainly not be handled here, and instead
177    // handled with the help of TLI or the target itself. This was largely
178    // ported from existing analysis heuristics here so that such refactorings
179    // can take place in the future.
180
181    if (F->isIntrinsic())
182      return false;
183
184    if (F->hasLocalLinkage() || !F->hasName())
185      return true;
186
187    StringRef Name = F->getName();
188
189    // These will all likely lower to a single selection DAG node.
190    if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
191        Name == "fabs" || Name == "fabsf" || Name == "fabsl" || Name == "sin" ||
192        Name == "fmin" || Name == "fminf" || Name == "fminl" ||
193        Name == "fmax" || Name == "fmaxf" || Name == "fmaxl" ||
194        Name == "sinf" || Name == "sinl" || Name == "cos" || Name == "cosf" ||
195        Name == "cosl" || Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl")
196      return false;
197
198    // These are all likely to be optimized into something smaller.
199    if (Name == "pow" || Name == "powf" || Name == "powl" || Name == "exp2" ||
200        Name == "exp2l" || Name == "exp2f" || Name == "floor" ||
201        Name == "floorf" || Name == "ceil" || Name == "round" ||
202        Name == "ffs" || Name == "ffsl" || Name == "abs" || Name == "labs" ||
203        Name == "llabs")
204      return false;
205
206    return true;
207  }
208
209  bool isHardwareLoopProfitable(Loop *L, ScalarEvolution &SE,
210                                AssumptionCache &AC,
211                                TargetLibraryInfo *LibInfo,
212                                HardwareLoopInfo &HWLoopInfo) {
213    return false;
214  }
215
216  bool preferPredicateOverEpilogue(Loop *L, LoopInfo *LI, ScalarEvolution &SE,
217                                   AssumptionCache &AC, TargetLibraryInfo *TLI,
218                                   DominatorTree *DT,
219                                   const LoopAccessInfo *LAI) const {
220    return false;
221  }
222
223  void getUnrollingPreferences(Loop *, ScalarEvolution &,
224                               TTI::UnrollingPreferences &) {}
225
226  bool isLegalAddImmediate(int64_t Imm) { return false; }
227
228  bool isLegalICmpImmediate(int64_t Imm) { return false; }
229
230  bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
231                             bool HasBaseReg, int64_t Scale,
232                             unsigned AddrSpace, Instruction *I = nullptr) {
233    // Guess that only reg and reg+reg addressing is allowed. This heuristic is
234    // taken from the implementation of LSR.
235    return !BaseGV && BaseOffset == 0 && (Scale == 0 || Scale == 1);
236  }
237
238  bool isLSRCostLess(TTI::LSRCost &C1, TTI::LSRCost &C2) {
239    return std::tie(C1.NumRegs, C1.AddRecCost, C1.NumIVMuls, C1.NumBaseAdds,
240                    C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
241           std::tie(C2.NumRegs, C2.AddRecCost, C2.NumIVMuls, C2.NumBaseAdds,
242                    C2.ScaleCost, C2.ImmCost, C2.SetupCost);
243  }
244
245  bool canMacroFuseCmp() { return false; }
246
247  bool canSaveCmp(Loop *L, BranchInst **BI, ScalarEvolution *SE, LoopInfo *LI,
248                  DominatorTree *DT, AssumptionCache *AC,
249                  TargetLibraryInfo *LibInfo) {
250    return false;
251  }
252
253  bool shouldFavorPostInc() const { return false; }
254
255  bool shouldFavorBackedgeIndex(const Loop *L) const { return false; }
256
257  bool isLegalMaskedStore(Type *DataType, MaybeAlign Alignment) { return false; }
258
259  bool isLegalMaskedLoad(Type *DataType, MaybeAlign Alignment) { return false; }
260
261  bool isLegalNTStore(Type *DataType, Align Alignment) {
262    // By default, assume nontemporal memory stores are available for stores
263    // that are aligned and have a size that is a power of 2.
264    unsigned DataSize = DL.getTypeStoreSize(DataType);
265    return Alignment >= DataSize && isPowerOf2_32(DataSize);
266  }
267
268  bool isLegalNTLoad(Type *DataType, Align Alignment) {
269    // By default, assume nontemporal memory loads are available for loads that
270    // are aligned and have a size that is a power of 2.
271    unsigned DataSize = DL.getTypeStoreSize(DataType);
272    return Alignment >= DataSize && isPowerOf2_32(DataSize);
273  }
274
275  bool isLegalMaskedScatter(Type *DataType, MaybeAlign Alignment) {
276    return false;
277  }
278
279  bool isLegalMaskedGather(Type *DataType, MaybeAlign Alignment) {
280    return false;
281  }
282
283  bool isLegalMaskedCompressStore(Type *DataType) { return false; }
284
285  bool isLegalMaskedExpandLoad(Type *DataType) { return false; }
286
287  bool hasDivRemOp(Type *DataType, bool IsSigned) { return false; }
288
289  bool hasVolatileVariant(Instruction *I, unsigned AddrSpace) { return false; }
290
291  bool prefersVectorizedAddressing() { return true; }
292
293  int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
294                           bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
295    // Guess that all legal addressing mode are free.
296    if (isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
297                              Scale, AddrSpace))
298      return 0;
299    return -1;
300  }
301
302  bool LSRWithInstrQueries() { return false; }
303
304  bool isTruncateFree(Type *Ty1, Type *Ty2) { return false; }
305
306  bool isProfitableToHoist(Instruction *I) { return true; }
307
308  bool useAA() { return false; }
309
310  bool isTypeLegal(Type *Ty) { return false; }
311
312  bool shouldBuildLookupTables() { return true; }
313  bool shouldBuildLookupTablesForConstant(Constant *C) { return true; }
314
315  bool useColdCCForColdCall(Function &F) { return false; }
316
317  unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
318    return 0;
319  }
320
321  unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args,
322                                            unsigned VF) { return 0; }
323
324  bool supportsEfficientVectorElementLoadStore() { return false; }
325
326  bool enableAggressiveInterleaving(bool LoopHasReductions) { return false; }
327
328  TTI::MemCmpExpansionOptions enableMemCmpExpansion(bool OptSize,
329                                                    bool IsZeroCmp) const {
330    return {};
331  }
332
333  bool enableInterleavedAccessVectorization() { return false; }
334
335  bool enableMaskedInterleavedAccessVectorization() { return false; }
336
337  bool isFPVectorizationPotentiallyUnsafe() { return false; }
338
339  bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
340                                      unsigned BitWidth,
341                                      unsigned AddressSpace,
342                                      unsigned Alignment,
343                                      bool *Fast) { return false; }
344
345  TTI::PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) {
346    return TTI::PSK_Software;
347  }
348
349  bool haveFastSqrt(Type *Ty) { return false; }
350
351  bool isFCmpOrdCheaperThanFCmpZero(Type *Ty) { return true; }
352
353  unsigned getFPOpCost(Type *Ty) { return TargetTransformInfo::TCC_Basic; }
354
355  int getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
356                            Type *Ty) {
357    return 0;
358  }
359
360  unsigned getIntImmCost(const APInt &Imm, Type *Ty) { return TTI::TCC_Basic; }
361
362  unsigned getIntImmCostInst(unsigned Opcode, unsigned Idx, const APInt &Imm,
363                             Type *Ty) {
364    return TTI::TCC_Free;
365  }
366
367  unsigned getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx,
368                               const APInt &Imm, Type *Ty) {
369    return TTI::TCC_Free;
370  }
371
372  unsigned getNumberOfRegisters(unsigned ClassID) const { return 8; }
373
374  unsigned getRegisterClassForType(bool Vector, Type *Ty = nullptr) const {
375    return Vector ? 1 : 0;
376  };
377
378  const char* getRegisterClassName(unsigned ClassID) const {
379    switch (ClassID) {
380      default:
381        return "Generic::Unknown Register Class";
382      case 0: return "Generic::ScalarRC";
383      case 1: return "Generic::VectorRC";
384    }
385  }
386
387  unsigned getRegisterBitWidth(bool Vector) const { return 32; }
388
389  unsigned getMinVectorRegisterBitWidth() { return 128; }
390
391  bool shouldMaximizeVectorBandwidth(bool OptSize) const { return false; }
392
393  unsigned getMinimumVF(unsigned ElemWidth) const { return 0; }
394
395  bool
396  shouldConsiderAddressTypePromotion(const Instruction &I,
397                                     bool &AllowPromotionWithoutCommonHeader) {
398    AllowPromotionWithoutCommonHeader = false;
399    return false;
400  }
401
402  unsigned getCacheLineSize() const { return 0; }
403
404  llvm::Optional<unsigned> getCacheSize(TargetTransformInfo::CacheLevel Level) const {
405    switch (Level) {
406    case TargetTransformInfo::CacheLevel::L1D:
407      LLVM_FALLTHROUGH;
408    case TargetTransformInfo::CacheLevel::L2D:
409      return llvm::Optional<unsigned>();
410    }
411    llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
412  }
413
414  llvm::Optional<unsigned> getCacheAssociativity(
415    TargetTransformInfo::CacheLevel Level) const {
416    switch (Level) {
417    case TargetTransformInfo::CacheLevel::L1D:
418      LLVM_FALLTHROUGH;
419    case TargetTransformInfo::CacheLevel::L2D:
420      return llvm::Optional<unsigned>();
421    }
422
423    llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
424  }
425
426  unsigned getPrefetchDistance() const { return 0; }
427  unsigned getMinPrefetchStride() const { return 1; }
428  unsigned getMaxPrefetchIterationsAhead() const { return UINT_MAX; }
429
430  unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
431
432  unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
433                                  TTI::OperandValueKind Opd1Info,
434                                  TTI::OperandValueKind Opd2Info,
435                                  TTI::OperandValueProperties Opd1PropInfo,
436                                  TTI::OperandValueProperties Opd2PropInfo,
437                                  ArrayRef<const Value *> Args,
438                                  const Instruction *CxtI = nullptr) {
439    return 1;
440  }
441
442  unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Ty, int Index,
443                          Type *SubTp) {
444    return 1;
445  }
446
447  unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
448                            const Instruction *I) { return 1; }
449
450  unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst,
451                                    VectorType *VecTy, unsigned Index) {
452    return 1;
453  }
454
455  unsigned getCFInstrCost(unsigned Opcode) { return 1; }
456
457  unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
458                              const Instruction *I) {
459    return 1;
460  }
461
462  unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
463    return 1;
464  }
465
466  unsigned getMemoryOpCost(unsigned Opcode, Type *Src, MaybeAlign Alignment,
467                           unsigned AddressSpace, const Instruction *I) {
468    return 1;
469  }
470
471  unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
472                                 unsigned AddressSpace) {
473    return 1;
474  }
475
476  unsigned getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
477                                  bool VariableMask,
478                                  unsigned Alignment) {
479    return 1;
480  }
481
482  unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
483                                      unsigned Factor,
484                                      ArrayRef<unsigned> Indices,
485                                      unsigned Alignment, unsigned AddressSpace,
486                                      bool UseMaskForCond = false,
487                                      bool UseMaskForGaps = false) {
488    return 1;
489  }
490
491  unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
492                                 ArrayRef<Type *> Tys, FastMathFlags FMF,
493                                 unsigned ScalarizationCostPassed) {
494    return 1;
495  }
496  unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
497            ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) {
498    return 1;
499  }
500
501  unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
502    return 1;
503  }
504
505  unsigned getNumberOfParts(Type *Tp) { return 0; }
506
507  unsigned getAddressComputationCost(Type *Tp, ScalarEvolution *,
508                                     const SCEV *) {
509    return 0;
510  }
511
512  unsigned getArithmeticReductionCost(unsigned, Type *, bool) { return 1; }
513
514  unsigned getMinMaxReductionCost(Type *, Type *, bool, bool) { return 1; }
515
516  unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) { return 0; }
517
518  bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) {
519    return false;
520  }
521
522  unsigned getAtomicMemIntrinsicMaxElementSize() const {
523    // Note for overrides: You must ensure for all element unordered-atomic
524    // memory intrinsics that all power-of-2 element sizes up to, and
525    // including, the return value of this method have a corresponding
526    // runtime lib call. These runtime lib call definitions can be found
527    // in RuntimeLibcalls.h
528    return 0;
529  }
530
531  Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
532                                           Type *ExpectedType) {
533    return nullptr;
534  }
535
536  Type *getMemcpyLoopLoweringType(LLVMContext &Context, Value *Length,
537                                  unsigned SrcAlign, unsigned DestAlign) const {
538    return Type::getInt8Ty(Context);
539  }
540
541  void getMemcpyLoopResidualLoweringType(SmallVectorImpl<Type *> &OpsOut,
542                                         LLVMContext &Context,
543                                         unsigned RemainingBytes,
544                                         unsigned SrcAlign,
545                                         unsigned DestAlign) const {
546    for (unsigned i = 0; i != RemainingBytes; ++i)
547      OpsOut.push_back(Type::getInt8Ty(Context));
548  }
549
550  bool areInlineCompatible(const Function *Caller,
551                           const Function *Callee) const {
552    return (Caller->getFnAttribute("target-cpu") ==
553            Callee->getFnAttribute("target-cpu")) &&
554           (Caller->getFnAttribute("target-features") ==
555            Callee->getFnAttribute("target-features"));
556  }
557
558  bool areFunctionArgsABICompatible(const Function *Caller, const Function *Callee,
559                                    SmallPtrSetImpl<Argument *> &Args) const {
560    return (Caller->getFnAttribute("target-cpu") ==
561            Callee->getFnAttribute("target-cpu")) &&
562           (Caller->getFnAttribute("target-features") ==
563            Callee->getFnAttribute("target-features"));
564  }
565
566  bool isIndexedLoadLegal(TTI::MemIndexedMode Mode, Type *Ty,
567                          const DataLayout &DL) const {
568    return false;
569  }
570
571  bool isIndexedStoreLegal(TTI::MemIndexedMode Mode, Type *Ty,
572                           const DataLayout &DL) const {
573    return false;
574  }
575
576  unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const { return 128; }
577
578  bool isLegalToVectorizeLoad(LoadInst *LI) const { return true; }
579
580  bool isLegalToVectorizeStore(StoreInst *SI) const { return true; }
581
582  bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
583                                   unsigned Alignment,
584                                   unsigned AddrSpace) const {
585    return true;
586  }
587
588  bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
589                                    unsigned Alignment,
590                                    unsigned AddrSpace) const {
591    return true;
592  }
593
594  unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
595                               unsigned ChainSizeInBytes,
596                               VectorType *VecTy) const {
597    return VF;
598  }
599
600  unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
601                                unsigned ChainSizeInBytes,
602                                VectorType *VecTy) const {
603    return VF;
604  }
605
606  bool useReductionIntrinsic(unsigned Opcode, Type *Ty,
607                             TTI::ReductionFlags Flags) const {
608    return false;
609  }
610
611  bool shouldExpandReduction(const IntrinsicInst *II) const {
612    return true;
613  }
614
615  unsigned getGISelRematGlobalCost() const {
616    return 1;
617  }
618
619protected:
620  // Obtain the minimum required size to hold the value (without the sign)
621  // In case of a vector it returns the min required size for one element.
622  unsigned minRequiredElementSize(const Value* Val, bool &isSigned) {
623    if (isa<ConstantDataVector>(Val) || isa<ConstantVector>(Val)) {
624      const auto* VectorValue = cast<Constant>(Val);
625
626      // In case of a vector need to pick the max between the min
627      // required size for each element
628      auto *VT = cast<VectorType>(Val->getType());
629
630      // Assume unsigned elements
631      isSigned = false;
632
633      // The max required size is the total vector width divided by num
634      // of elements in the vector
635      unsigned MaxRequiredSize = VT->getBitWidth() / VT->getNumElements();
636
637      unsigned MinRequiredSize = 0;
638      for(unsigned i = 0, e = VT->getNumElements(); i < e; ++i) {
639        if (auto* IntElement =
640              dyn_cast<ConstantInt>(VectorValue->getAggregateElement(i))) {
641          bool signedElement = IntElement->getValue().isNegative();
642          // Get the element min required size.
643          unsigned ElementMinRequiredSize =
644            IntElement->getValue().getMinSignedBits() - 1;
645          // In case one element is signed then all the vector is signed.
646          isSigned |= signedElement;
647          // Save the max required bit size between all the elements.
648          MinRequiredSize = std::max(MinRequiredSize, ElementMinRequiredSize);
649        }
650        else {
651          // not an int constant element
652          return MaxRequiredSize;
653        }
654      }
655      return MinRequiredSize;
656    }
657
658    if (const auto* CI = dyn_cast<ConstantInt>(Val)) {
659      isSigned = CI->getValue().isNegative();
660      return CI->getValue().getMinSignedBits() - 1;
661    }
662
663    if (const auto* Cast = dyn_cast<SExtInst>(Val)) {
664      isSigned = true;
665      return Cast->getSrcTy()->getScalarSizeInBits() - 1;
666    }
667
668    if (const auto* Cast = dyn_cast<ZExtInst>(Val)) {
669      isSigned = false;
670      return Cast->getSrcTy()->getScalarSizeInBits();
671    }
672
673    isSigned = false;
674    return Val->getType()->getScalarSizeInBits();
675  }
676
677  bool isStridedAccess(const SCEV *Ptr) {
678    return Ptr && isa<SCEVAddRecExpr>(Ptr);
679  }
680
681  const SCEVConstant *getConstantStrideStep(ScalarEvolution *SE,
682                                            const SCEV *Ptr) {
683    if (!isStridedAccess(Ptr))
684      return nullptr;
685    const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ptr);
686    return dyn_cast<SCEVConstant>(AddRec->getStepRecurrence(*SE));
687  }
688
689  bool isConstantStridedAccessLessThan(ScalarEvolution *SE, const SCEV *Ptr,
690                                       int64_t MergeDistance) {
691    const SCEVConstant *Step = getConstantStrideStep(SE, Ptr);
692    if (!Step)
693      return false;
694    APInt StrideVal = Step->getAPInt();
695    if (StrideVal.getBitWidth() > 64)
696      return false;
697    // FIXME: Need to take absolute value for negative stride case.
698    return StrideVal.getSExtValue() < MergeDistance;
699  }
700};
701
702/// CRTP base class for use as a mix-in that aids implementing
703/// a TargetTransformInfo-compatible class.
704template <typename T>
705class TargetTransformInfoImplCRTPBase : public TargetTransformInfoImplBase {
706private:
707  typedef TargetTransformInfoImplBase BaseT;
708
709protected:
710  explicit TargetTransformInfoImplCRTPBase(const DataLayout &DL) : BaseT(DL) {}
711
712public:
713  using BaseT::getCallCost;
714
715  unsigned getCallCost(const Function *F, int NumArgs, const User *U) {
716    assert(F && "A concrete function must be provided to this routine.");
717
718    if (NumArgs < 0)
719      // Set the argument number to the number of explicit arguments in the
720      // function.
721      NumArgs = F->arg_size();
722
723    if (Intrinsic::ID IID = F->getIntrinsicID()) {
724      FunctionType *FTy = F->getFunctionType();
725      SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end());
726      return static_cast<T *>(this)
727          ->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys, U);
728    }
729
730    if (!static_cast<T *>(this)->isLoweredToCall(F))
731      return TTI::TCC_Basic; // Give a basic cost if it will be lowered
732                             // directly.
733
734    return static_cast<T *>(this)->getCallCost(F->getFunctionType(), NumArgs, U);
735  }
736
737  unsigned getCallCost(const Function *F, ArrayRef<const Value *> Arguments,
738                       const User *U) {
739    // Simply delegate to generic handling of the call.
740    // FIXME: We should use instsimplify or something else to catch calls which
741    // will constant fold with these arguments.
742    return static_cast<T *>(this)->getCallCost(F, Arguments.size(), U);
743  }
744
745  using BaseT::getGEPCost;
746
747  int getGEPCost(Type *PointeeType, const Value *Ptr,
748                 ArrayRef<const Value *> Operands) {
749    assert(PointeeType && Ptr && "can't get GEPCost of nullptr");
750    // TODO: will remove this when pointers have an opaque type.
751    assert(Ptr->getType()->getScalarType()->getPointerElementType() ==
752               PointeeType &&
753           "explicit pointee type doesn't match operand's pointee type");
754    auto *BaseGV = dyn_cast<GlobalValue>(Ptr->stripPointerCasts());
755    bool HasBaseReg = (BaseGV == nullptr);
756
757    auto PtrSizeBits = DL.getPointerTypeSizeInBits(Ptr->getType());
758    APInt BaseOffset(PtrSizeBits, 0);
759    int64_t Scale = 0;
760
761    auto GTI = gep_type_begin(PointeeType, Operands);
762    Type *TargetType = nullptr;
763
764    // Handle the case where the GEP instruction has a single operand,
765    // the basis, therefore TargetType is a nullptr.
766    if (Operands.empty())
767      return !BaseGV ? TTI::TCC_Free : TTI::TCC_Basic;
768
769    for (auto I = Operands.begin(); I != Operands.end(); ++I, ++GTI) {
770      TargetType = GTI.getIndexedType();
771      // We assume that the cost of Scalar GEP with constant index and the
772      // cost of Vector GEP with splat constant index are the same.
773      const ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I);
774      if (!ConstIdx)
775        if (auto Splat = getSplatValue(*I))
776          ConstIdx = dyn_cast<ConstantInt>(Splat);
777      if (StructType *STy = GTI.getStructTypeOrNull()) {
778        // For structures the index is always splat or scalar constant
779        assert(ConstIdx && "Unexpected GEP index");
780        uint64_t Field = ConstIdx->getZExtValue();
781        BaseOffset += DL.getStructLayout(STy)->getElementOffset(Field);
782      } else {
783        int64_t ElementSize = DL.getTypeAllocSize(GTI.getIndexedType());
784        if (ConstIdx) {
785          BaseOffset +=
786              ConstIdx->getValue().sextOrTrunc(PtrSizeBits) * ElementSize;
787        } else {
788          // Needs scale register.
789          if (Scale != 0)
790            // No addressing mode takes two scale registers.
791            return TTI::TCC_Basic;
792          Scale = ElementSize;
793        }
794      }
795    }
796
797    if (static_cast<T *>(this)->isLegalAddressingMode(
798            TargetType, const_cast<GlobalValue *>(BaseGV),
799            BaseOffset.sextOrTrunc(64).getSExtValue(), HasBaseReg, Scale,
800            Ptr->getType()->getPointerAddressSpace()))
801      return TTI::TCC_Free;
802    return TTI::TCC_Basic;
803  }
804
805  unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
806                            ArrayRef<Type *> ParamTys, const User *U) {
807    switch (IID) {
808    default:
809      // Intrinsics rarely (if ever) have normal argument setup constraints.
810      // Model them as having a basic instruction cost.
811      return TTI::TCC_Basic;
812
813    // TODO: other libc intrinsics.
814    case Intrinsic::memcpy:
815      return static_cast<T *>(this)->getMemcpyCost(dyn_cast<Instruction>(U));
816
817    case Intrinsic::annotation:
818    case Intrinsic::assume:
819    case Intrinsic::sideeffect:
820    case Intrinsic::dbg_declare:
821    case Intrinsic::dbg_value:
822    case Intrinsic::dbg_label:
823    case Intrinsic::invariant_start:
824    case Intrinsic::invariant_end:
825    case Intrinsic::launder_invariant_group:
826    case Intrinsic::strip_invariant_group:
827    case Intrinsic::is_constant:
828    case Intrinsic::lifetime_start:
829    case Intrinsic::lifetime_end:
830    case Intrinsic::objectsize:
831    case Intrinsic::ptr_annotation:
832    case Intrinsic::var_annotation:
833    case Intrinsic::experimental_gc_result:
834    case Intrinsic::experimental_gc_relocate:
835    case Intrinsic::coro_alloc:
836    case Intrinsic::coro_begin:
837    case Intrinsic::coro_free:
838    case Intrinsic::coro_end:
839    case Intrinsic::coro_frame:
840    case Intrinsic::coro_size:
841    case Intrinsic::coro_suspend:
842    case Intrinsic::coro_param:
843    case Intrinsic::coro_subfn_addr:
844      // These intrinsics don't actually represent code after lowering.
845      return TTI::TCC_Free;
846    }
847  }
848
849  unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
850                            ArrayRef<const Value *> Arguments, const User *U) {
851    // Delegate to the generic intrinsic handling code. This mostly provides an
852    // opportunity for targets to (for example) special case the cost of
853    // certain intrinsics based on constants used as arguments.
854    SmallVector<Type *, 8> ParamTys;
855    ParamTys.reserve(Arguments.size());
856    for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx)
857      ParamTys.push_back(Arguments[Idx]->getType());
858    return static_cast<T *>(this)->getIntrinsicCost(IID, RetTy, ParamTys, U);
859  }
860
861  unsigned getUserCost(const User *U, ArrayRef<const Value *> Operands) {
862    if (isa<PHINode>(U))
863      return TTI::TCC_Free; // Model all PHI nodes as free.
864
865    if (isa<ExtractValueInst>(U))
866      return TTI::TCC_Free; // Model all ExtractValue nodes as free.
867
868    // Static alloca doesn't generate target instructions.
869    if (auto *A = dyn_cast<AllocaInst>(U))
870      if (A->isStaticAlloca())
871        return TTI::TCC_Free;
872
873    if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U)) {
874      return static_cast<T *>(this)->getGEPCost(GEP->getSourceElementType(),
875                                                GEP->getPointerOperand(),
876                                                Operands.drop_front());
877    }
878
879    if (auto CS = ImmutableCallSite(U)) {
880      const Function *F = CS.getCalledFunction();
881      if (!F) {
882        // Just use the called value type.
883        Type *FTy = CS.getCalledValue()->getType()->getPointerElementType();
884        return static_cast<T *>(this)
885            ->getCallCost(cast<FunctionType>(FTy), CS.arg_size(), U);
886      }
887
888      SmallVector<const Value *, 8> Arguments(CS.arg_begin(), CS.arg_end());
889      return static_cast<T *>(this)->getCallCost(F, Arguments, U);
890    }
891
892    if (isa<SExtInst>(U) || isa<ZExtInst>(U) || isa<FPExtInst>(U))
893      // The old behaviour of generally treating extensions of icmp to be free
894      // has been removed. A target that needs it should override getUserCost().
895      return static_cast<T *>(this)->getExtCost(cast<Instruction>(U),
896                                                Operands.back());
897
898    return static_cast<T *>(this)->getOperationCost(
899        Operator::getOpcode(U), U->getType(),
900        U->getNumOperands() == 1 ? U->getOperand(0)->getType() : nullptr);
901  }
902
903  int getInstructionLatency(const Instruction *I) {
904    SmallVector<const Value *, 4> Operands(I->value_op_begin(),
905                                           I->value_op_end());
906    if (getUserCost(I, Operands) == TTI::TCC_Free)
907      return 0;
908
909    if (isa<LoadInst>(I))
910      return 4;
911
912    Type *DstTy = I->getType();
913
914    // Usually an intrinsic is a simple instruction.
915    // A real function call is much slower.
916    if (auto *CI = dyn_cast<CallInst>(I)) {
917      const Function *F = CI->getCalledFunction();
918      if (!F || static_cast<T *>(this)->isLoweredToCall(F))
919        return 40;
920      // Some intrinsics return a value and a flag, we use the value type
921      // to decide its latency.
922      if (StructType* StructTy = dyn_cast<StructType>(DstTy))
923        DstTy = StructTy->getElementType(0);
924      // Fall through to simple instructions.
925    }
926
927    if (VectorType *VectorTy = dyn_cast<VectorType>(DstTy))
928      DstTy = VectorTy->getElementType();
929    if (DstTy->isFloatingPointTy())
930      return 3;
931
932    return 1;
933  }
934};
935}
936
937#endif
938