LLVM  10.0.0svn
TargetTransformInfoImpl.h
Go to the documentation of this file.
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 
20 #include "llvm/IR/CallSite.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/Function.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.
32 protected:
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.
42  : 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,
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 
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 
160  Intrinsic::ID IID) const {
161  return false;
162  }
163 
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 
206  AssumptionCache &AC,
207  TargetLibraryInfo *LibInfo,
208  HardwareLoopInfo &HWLoopInfo) {
209  return false;
210  }
211 
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 
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 
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, MaybeAlign Alignment) { return false; }
247 
248  bool isLegalMaskedLoad(Type *DataType, MaybeAlign Alignment) { return false; }
249 
250  bool isLegalNTStore(Type *DataType, 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, 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  bool shouldBuildLookupTables() { return true; }
299 
300  bool useColdCCForColdCall(Function &F) { return false; }
301 
302  unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
303  return 0;
304  }
305 
307  unsigned VF) { return 0; }
308 
310 
311  bool enableAggressiveInterleaving(bool LoopHasReductions) { return false; }
312 
314  bool IsZeroCmp) const {
315  return {};
316  }
317 
318  bool enableInterleavedAccessVectorization() { return false; }
319 
321 
322  bool isFPVectorizationPotentiallyUnsafe() { return false; }
323 
325  unsigned BitWidth,
326  unsigned AddressSpace,
327  unsigned Alignment,
328  bool *Fast) { return false; }
329 
330  TTI::PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) {
331  return TTI::PSK_Software;
332  }
333 
334  bool haveFastSqrt(Type *Ty) { return false; }
335 
336  bool isFCmpOrdCheaperThanFCmpZero(Type *Ty) { return true; }
337 
339 
340  int getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
341  Type *Ty) {
342  return 0;
343  }
344 
345  unsigned getIntImmCost(const APInt &Imm, Type *Ty) { return TTI::TCC_Basic; }
346 
347  unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
348  Type *Ty) {
349  return TTI::TCC_Free;
350  }
351 
352  unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
353  Type *Ty) {
354  return TTI::TCC_Free;
355  }
356 
357  unsigned getNumberOfRegisters(unsigned ClassID) const { return 8; }
358 
359  unsigned getRegisterClassForType(bool Vector, Type *Ty = nullptr) const {
360  return Vector ? 1 : 0;
361  };
362 
363  const char* getRegisterClassName(unsigned ClassID) const {
364  switch (ClassID) {
365  default:
366  return "Generic::Unknown Register Class";
367  case 0: return "Generic::ScalarRC";
368  case 1: return "Generic::VectorRC";
369  }
370  }
371 
372  unsigned getRegisterBitWidth(bool Vector) const { return 32; }
373 
374  unsigned getMinVectorRegisterBitWidth() { return 128; }
375 
376  bool shouldMaximizeVectorBandwidth(bool OptSize) const { return false; }
377 
378  unsigned getMinimumVF(unsigned ElemWidth) const { return 0; }
379 
380  bool
382  bool &AllowPromotionWithoutCommonHeader) {
383  AllowPromotionWithoutCommonHeader = false;
384  return false;
385  }
386 
387  unsigned getCacheLineSize() const { return 0; }
388 
390  switch (Level) {
394  return llvm::Optional<unsigned>();
395  }
396  llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
397  }
398 
401  switch (Level) {
405  return llvm::Optional<unsigned>();
406  }
407 
408  llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
409  }
410 
411  unsigned getPrefetchDistance() const { return 0; }
412  unsigned getMinPrefetchStride() const { return 1; }
413  unsigned getMaxPrefetchIterationsAhead() const { return UINT_MAX; }
414 
415  unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
416 
417  unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
418  TTI::OperandValueKind Opd1Info,
419  TTI::OperandValueKind Opd2Info,
420  TTI::OperandValueProperties Opd1PropInfo,
421  TTI::OperandValueProperties Opd2PropInfo,
423  return 1;
424  }
425 
427  Type *SubTp) {
428  return 1;
429  }
430 
431  unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
432  const Instruction *I) { return 1; }
433 
434  unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst,
435  VectorType *VecTy, unsigned Index) {
436  return 1;
437  }
438 
439  unsigned getCFInstrCost(unsigned Opcode) { return 1; }
440 
441  unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
442  const Instruction *I) {
443  return 1;
444  }
445 
446  unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
447  return 1;
448  }
449 
450  unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
451  unsigned AddressSpace, const Instruction *I) {
452  return 1;
453  }
454 
455  unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
456  unsigned AddressSpace) {
457  return 1;
458  }
459 
460  unsigned getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
461  bool VariableMask,
462  unsigned Alignment) {
463  return 1;
464  }
465 
466  unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
467  unsigned Factor,
468  ArrayRef<unsigned> Indices,
469  unsigned Alignment, unsigned AddressSpace,
470  bool UseMaskForCond = false,
471  bool UseMaskForGaps = false) {
472  return 1;
473  }
474 
477  unsigned ScalarizationCostPassed) {
478  return 1;
479  }
481  ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) {
482  return 1;
483  }
484 
485  unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
486  return 1;
487  }
488 
489  unsigned getNumberOfParts(Type *Tp) { return 0; }
490 
492  const SCEV *) {
493  return 0;
494  }
495 
496  unsigned getArithmeticReductionCost(unsigned, Type *, bool) { return 1; }
497 
498  unsigned getMinMaxReductionCost(Type *, Type *, bool, bool) { return 1; }
499 
500  unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) { return 0; }
501 
503  return false;
504  }
505 
507  // Note for overrides: You must ensure for all element unordered-atomic
508  // memory intrinsics that all power-of-2 element sizes up to, and
509  // including, the return value of this method have a corresponding
510  // runtime lib call. These runtime lib call definitions can be found
511  // in RuntimeLibcalls.h
512  return 0;
513  }
514 
516  Type *ExpectedType) {
517  return nullptr;
518  }
519 
521  unsigned SrcAlign, unsigned DestAlign) const {
522  return Type::getInt8Ty(Context);
523  }
524 
527  unsigned RemainingBytes,
528  unsigned SrcAlign,
529  unsigned DestAlign) const {
530  for (unsigned i = 0; i != RemainingBytes; ++i)
531  OpsOut.push_back(Type::getInt8Ty(Context));
532  }
533 
534  bool areInlineCompatible(const Function *Caller,
535  const Function *Callee) const {
536  return (Caller->getFnAttribute("target-cpu") ==
537  Callee->getFnAttribute("target-cpu")) &&
538  (Caller->getFnAttribute("target-features") ==
539  Callee->getFnAttribute("target-features"));
540  }
541 
542  bool areFunctionArgsABICompatible(const Function *Caller, const Function *Callee,
544  return (Caller->getFnAttribute("target-cpu") ==
545  Callee->getFnAttribute("target-cpu")) &&
546  (Caller->getFnAttribute("target-features") ==
547  Callee->getFnAttribute("target-features"));
548  }
549 
551  const DataLayout &DL) const {
552  return false;
553  }
554 
556  const DataLayout &DL) const {
557  return false;
558  }
559 
560  unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const { return 128; }
561 
562  bool isLegalToVectorizeLoad(LoadInst *LI) const { return true; }
563 
564  bool isLegalToVectorizeStore(StoreInst *SI) const { return true; }
565 
566  bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
567  unsigned Alignment,
568  unsigned AddrSpace) const {
569  return true;
570  }
571 
572  bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
573  unsigned Alignment,
574  unsigned AddrSpace) const {
575  return true;
576  }
577 
578  unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
579  unsigned ChainSizeInBytes,
580  VectorType *VecTy) const {
581  return VF;
582  }
583 
584  unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
585  unsigned ChainSizeInBytes,
586  VectorType *VecTy) const {
587  return VF;
588  }
589 
590  bool useReductionIntrinsic(unsigned Opcode, Type *Ty,
591  TTI::ReductionFlags Flags) const {
592  return false;
593  }
594 
595  bool shouldExpandReduction(const IntrinsicInst *II) const {
596  return true;
597  }
598 
599  unsigned getGISelRematGlobalCost() const {
600  return 1;
601  }
602 
603 protected:
604  // Obtain the minimum required size to hold the value (without the sign)
605  // In case of a vector it returns the min required size for one element.
606  unsigned minRequiredElementSize(const Value* Val, bool &isSigned) {
607  if (isa<ConstantDataVector>(Val) || isa<ConstantVector>(Val)) {
608  const auto* VectorValue = cast<Constant>(Val);
609 
610  // In case of a vector need to pick the max between the min
611  // required size for each element
612  auto *VT = cast<VectorType>(Val->getType());
613 
614  // Assume unsigned elements
615  isSigned = false;
616 
617  // The max required size is the total vector width divided by num
618  // of elements in the vector
619  unsigned MaxRequiredSize = VT->getBitWidth() / VT->getNumElements();
620 
621  unsigned MinRequiredSize = 0;
622  for(unsigned i = 0, e = VT->getNumElements(); i < e; ++i) {
623  if (auto* IntElement =
624  dyn_cast<ConstantInt>(VectorValue->getAggregateElement(i))) {
625  bool signedElement = IntElement->getValue().isNegative();
626  // Get the element min required size.
627  unsigned ElementMinRequiredSize =
628  IntElement->getValue().getMinSignedBits() - 1;
629  // In case one element is signed then all the vector is signed.
630  isSigned |= signedElement;
631  // Save the max required bit size between all the elements.
632  MinRequiredSize = std::max(MinRequiredSize, ElementMinRequiredSize);
633  }
634  else {
635  // not an int constant element
636  return MaxRequiredSize;
637  }
638  }
639  return MinRequiredSize;
640  }
641 
642  if (const auto* CI = dyn_cast<ConstantInt>(Val)) {
643  isSigned = CI->getValue().isNegative();
644  return CI->getValue().getMinSignedBits() - 1;
645  }
646 
647  if (const auto* Cast = dyn_cast<SExtInst>(Val)) {
648  isSigned = true;
649  return Cast->getSrcTy()->getScalarSizeInBits() - 1;
650  }
651 
652  if (const auto* Cast = dyn_cast<ZExtInst>(Val)) {
653  isSigned = false;
654  return Cast->getSrcTy()->getScalarSizeInBits();
655  }
656 
657  isSigned = false;
658  return Val->getType()->getScalarSizeInBits();
659  }
660 
661  bool isStridedAccess(const SCEV *Ptr) {
662  return Ptr && isa<SCEVAddRecExpr>(Ptr);
663  }
664 
666  const SCEV *Ptr) {
667  if (!isStridedAccess(Ptr))
668  return nullptr;
669  const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ptr);
670  return dyn_cast<SCEVConstant>(AddRec->getStepRecurrence(*SE));
671  }
672 
674  int64_t MergeDistance) {
675  const SCEVConstant *Step = getConstantStrideStep(SE, Ptr);
676  if (!Step)
677  return false;
678  APInt StrideVal = Step->getAPInt();
679  if (StrideVal.getBitWidth() > 64)
680  return false;
681  // FIXME: Need to take absolute value for negative stride case.
682  return StrideVal.getSExtValue() < MergeDistance;
683  }
684 };
685 
686 /// CRTP base class for use as a mix-in that aids implementing
687 /// a TargetTransformInfo-compatible class.
688 template <typename T>
690 private:
692 
693 protected:
694  explicit TargetTransformInfoImplCRTPBase(const DataLayout &DL) : BaseT(DL) {}
695 
696 public:
697  using BaseT::getCallCost;
698 
699  unsigned getCallCost(const Function *F, int NumArgs, const User *U) {
700  assert(F && "A concrete function must be provided to this routine.");
701 
702  if (NumArgs < 0)
703  // Set the argument number to the number of explicit arguments in the
704  // function.
705  NumArgs = F->arg_size();
706 
707  if (Intrinsic::ID IID = F->getIntrinsicID()) {
708  FunctionType *FTy = F->getFunctionType();
709  SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end());
710  return static_cast<T *>(this)
711  ->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys, U);
712  }
713 
714  if (!static_cast<T *>(this)->isLoweredToCall(F))
715  return TTI::TCC_Basic; // Give a basic cost if it will be lowered
716  // directly.
717 
718  return static_cast<T *>(this)->getCallCost(F->getFunctionType(), NumArgs, U);
719  }
720 
722  const User *U) {
723  // Simply delegate to generic handling of the call.
724  // FIXME: We should use instsimplify or something else to catch calls which
725  // will constant fold with these arguments.
726  return static_cast<T *>(this)->getCallCost(F, Arguments.size(), U);
727  }
728 
729  using BaseT::getGEPCost;
730 
731  int getGEPCost(Type *PointeeType, const Value *Ptr,
733  assert(PointeeType && Ptr && "can't get GEPCost of nullptr");
734  // TODO: will remove this when pointers have an opaque type.
736  PointeeType &&
737  "explicit pointee type doesn't match operand's pointee type");
738  auto *BaseGV = dyn_cast<GlobalValue>(Ptr->stripPointerCasts());
739  bool HasBaseReg = (BaseGV == nullptr);
740 
741  auto PtrSizeBits = DL.getPointerTypeSizeInBits(Ptr->getType());
742  APInt BaseOffset(PtrSizeBits, 0);
743  int64_t Scale = 0;
744 
745  auto GTI = gep_type_begin(PointeeType, Operands);
746  Type *TargetType = nullptr;
747 
748  // Handle the case where the GEP instruction has a single operand,
749  // the basis, therefore TargetType is a nullptr.
750  if (Operands.empty())
751  return !BaseGV ? TTI::TCC_Free : TTI::TCC_Basic;
752 
753  for (auto I = Operands.begin(); I != Operands.end(); ++I, ++GTI) {
754  TargetType = GTI.getIndexedType();
755  // We assume that the cost of Scalar GEP with constant index and the
756  // cost of Vector GEP with splat constant index are the same.
757  const ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I);
758  if (!ConstIdx)
759  if (auto Splat = getSplatValue(*I))
760  ConstIdx = dyn_cast<ConstantInt>(Splat);
761  if (StructType *STy = GTI.getStructTypeOrNull()) {
762  // For structures the index is always splat or scalar constant
763  assert(ConstIdx && "Unexpected GEP index");
764  uint64_t Field = ConstIdx->getZExtValue();
765  BaseOffset += DL.getStructLayout(STy)->getElementOffset(Field);
766  } else {
767  int64_t ElementSize = DL.getTypeAllocSize(GTI.getIndexedType());
768  if (ConstIdx) {
769  BaseOffset +=
770  ConstIdx->getValue().sextOrTrunc(PtrSizeBits) * ElementSize;
771  } else {
772  // Needs scale register.
773  if (Scale != 0)
774  // No addressing mode takes two scale registers.
775  return TTI::TCC_Basic;
776  Scale = ElementSize;
777  }
778  }
779  }
780 
781  if (static_cast<T *>(this)->isLegalAddressingMode(
782  TargetType, const_cast<GlobalValue *>(BaseGV),
783  BaseOffset.sextOrTrunc(64).getSExtValue(), HasBaseReg, Scale,
784  Ptr->getType()->getPointerAddressSpace()))
785  return TTI::TCC_Free;
786  return TTI::TCC_Basic;
787  }
788 
789  unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
790  ArrayRef<Type *> ParamTys, const User *U) {
791  switch (IID) {
792  default:
793  // Intrinsics rarely (if ever) have normal argument setup constraints.
794  // Model them as having a basic instruction cost.
795  return TTI::TCC_Basic;
796 
797  // TODO: other libc intrinsics.
798  case Intrinsic::memcpy:
799  return static_cast<T *>(this)->getMemcpyCost(dyn_cast<Instruction>(U));
800 
801  case Intrinsic::annotation:
802  case Intrinsic::assume:
803  case Intrinsic::sideeffect:
804  case Intrinsic::dbg_declare:
805  case Intrinsic::dbg_value:
806  case Intrinsic::dbg_label:
807  case Intrinsic::invariant_start:
808  case Intrinsic::invariant_end:
809  case Intrinsic::launder_invariant_group:
810  case Intrinsic::strip_invariant_group:
811  case Intrinsic::is_constant:
812  case Intrinsic::lifetime_start:
813  case Intrinsic::lifetime_end:
814  case Intrinsic::objectsize:
815  case Intrinsic::ptr_annotation:
816  case Intrinsic::var_annotation:
817  case Intrinsic::experimental_gc_result:
818  case Intrinsic::experimental_gc_relocate:
819  case Intrinsic::coro_alloc:
820  case Intrinsic::coro_begin:
821  case Intrinsic::coro_free:
822  case Intrinsic::coro_end:
823  case Intrinsic::coro_frame:
824  case Intrinsic::coro_size:
825  case Intrinsic::coro_suspend:
826  case Intrinsic::coro_param:
827  case Intrinsic::coro_subfn_addr:
828  // These intrinsics don't actually represent code after lowering.
829  return TTI::TCC_Free;
830  }
831  }
832 
833  unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
835  // Delegate to the generic intrinsic handling code. This mostly provides an
836  // opportunity for targets to (for example) special case the cost of
837  // certain intrinsics based on constants used as arguments.
838  SmallVector<Type *, 8> ParamTys;
839  ParamTys.reserve(Arguments.size());
840  for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx)
841  ParamTys.push_back(Arguments[Idx]->getType());
842  return static_cast<T *>(this)->getIntrinsicCost(IID, RetTy, ParamTys, U);
843  }
844 
846  if (isa<PHINode>(U))
847  return TTI::TCC_Free; // Model all PHI nodes as free.
848 
849  if (isa<ExtractValueInst>(U))
850  return TTI::TCC_Free; // Model all ExtractValue nodes as free.
851 
852  // Static alloca doesn't generate target instructions.
853  if (auto *A = dyn_cast<AllocaInst>(U))
854  if (A->isStaticAlloca())
855  return TTI::TCC_Free;
856 
857  if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U)) {
858  return static_cast<T *>(this)->getGEPCost(GEP->getSourceElementType(),
859  GEP->getPointerOperand(),
860  Operands.drop_front());
861  }
862 
863  if (auto CS = ImmutableCallSite(U)) {
864  const Function *F = CS.getCalledFunction();
865  if (!F) {
866  // Just use the called value type.
867  Type *FTy = CS.getCalledValue()->getType()->getPointerElementType();
868  return static_cast<T *>(this)
869  ->getCallCost(cast<FunctionType>(FTy), CS.arg_size(), U);
870  }
871 
872  SmallVector<const Value *, 8> Arguments(CS.arg_begin(), CS.arg_end());
873  return static_cast<T *>(this)->getCallCost(F, Arguments, U);
874  }
875 
876  if (isa<SExtInst>(U) || isa<ZExtInst>(U) || isa<FPExtInst>(U))
877  // The old behaviour of generally treating extensions of icmp to be free
878  // has been removed. A target that needs it should override getUserCost().
879  return static_cast<T *>(this)->getExtCost(cast<Instruction>(U),
880  Operands.back());
881 
882  return static_cast<T *>(this)->getOperationCost(
883  Operator::getOpcode(U), U->getType(),
884  U->getNumOperands() == 1 ? U->getOperand(0)->getType() : nullptr);
885  }
886 
889  I->value_op_end());
890  if (getUserCost(I, Operands) == TTI::TCC_Free)
891  return 0;
892 
893  if (isa<LoadInst>(I))
894  return 4;
895 
896  Type *DstTy = I->getType();
897 
898  // Usually an intrinsic is a simple instruction.
899  // A real function call is much slower.
900  if (auto *CI = dyn_cast<CallInst>(I)) {
901  const Function *F = CI->getCalledFunction();
902  if (!F || static_cast<T *>(this)->isLoweredToCall(F))
903  return 40;
904  // Some intrinsics return a value and a flag, we use the value type
905  // to decide its latency.
906  if (StructType* StructTy = dyn_cast<StructType>(DstTy))
907  DstTy = StructTy->getElementType(0);
908  // Fall through to simple instructions.
909  }
910 
911  if (VectorType *VectorTy = dyn_cast<VectorType>(DstTy))
912  DstTy = VectorTy->getElementType();
913  if (DstTy->isFloatingPointTy())
914  return 3;
915 
916  return 1;
917  }
918 };
919 }
920 
921 #endif
uint64_t CallInst * C
llvm::Optional< unsigned > getCacheSize(TargetTransformInfo::CacheLevel Level) const
unsigned getNumCases() const
Return the number of &#39;cases&#39; in this switch instruction, excluding the default case.
bool isIntrinsic() const
isIntrinsic - Returns true if the function&#39;s name starts with "llvm.".
Definition: Function.h:198
Base class for use as a mix-in that aids implementing a TargetTransformInfo-compatible class...
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:112
bool areFunctionArgsABICompatible(const Function *Caller, const Function *Callee, SmallPtrSetImpl< Argument *> &Args) const
bool isConstantStridedAccessLessThan(ScalarEvolution *SE, const SCEV *Ptr, int64_t MergeDistance)
bool hasLocalLinkage() const
Definition: GlobalValue.h:445
unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, ArrayRef< Type *> Tys, FastMathFlags FMF, unsigned ScalarizationCostPassed)
LLVMContext & Context
const T & back() const
back - Get the last element.
Definition: ArrayRef.h:157
SI Whole Quad Mode
unsigned getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr, bool VariableMask, unsigned Alignment)
This class represents lattice values for constants.
Definition: AllocatorList.h:23
unsigned minRequiredElementSize(const Value *Val, bool &isSigned)
void getUnrollingPreferences(Loop *, ScalarEvolution &, TTI::UnrollingPreferences &)
unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index)
iterator begin() const
Definition: ArrayRef.h:136
amdgpu Simplify well known AMD library false FunctionCallee Value const Twine & Name
unsigned getCostOfKeepingLiveOverCall(ArrayRef< Type *> Tys)
const Value * getSplatValue(const Value *V)
Get splat value if the input is a splat vector or return nullptr.
const StructLayout * getStructLayout(StructType *Ty) const
Returns a StructLayout object, indicating the alignment of the struct, its size, and the offsets of i...
Definition: DataLayout.cpp:613
value_op_iterator value_op_begin()
Definition: User.h:255
The main scalar evolution driver.
MemIndexedMode
The type of load/store indexing.
unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty, TTI::OperandValueKind Opd1Info, TTI::OperandValueKind Opd2Info, TTI::OperandValueProperties Opd1PropInfo, TTI::OperandValueProperties Opd2PropInfo, ArrayRef< const Value *> Args)
A cache of @llvm.assume calls within a function.
unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy)
value_op_iterator value_op_end()
Definition: User.h:258
F(f)
unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src, const Instruction *I)
unsigned getPointerAddressSpace() const
Get the address space of this pointer or pointer vector type.
Definition: DerivedTypes.h:621
param_iterator param_end() const
Definition: DerivedTypes.h:134
An instruction for reading from memory.
Definition: Instructions.h:169
Hexagon Common GEP
TypeSize getTypeSizeInBits(Type *Ty) const
Size examples:
Definition: DataLayout.h:618
unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment, unsigned AddressSpace, const Instruction *I)
void reserve(size_type N)
Definition: SmallVector.h:369
unsigned getIntImmCost(const APInt &Imm, Type *Ty)
int getExtCost(const Instruction *I, const Value *Src)
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1517
A templated base class for SmallPtrSet which provides the typesafe interface that is common across al...
Definition: SmallPtrSet.h:343
bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const
bool isIndexedLoadLegal(TTI::MemIndexedMode Mode, Type *Ty, const DataLayout &DL) const
CRTP base class for use as a mix-in that aids implementing a TargetTransformInfo-compatible class...
bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info)
unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize, unsigned ChainSizeInBytes, VectorType *VecTy) const
unsigned getMemcpyCost(const Instruction *I)
unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy, ArrayRef< const Value *> Arguments, const User *U)
Type * getPointerElementType() const
Definition: Type.h:380
unsigned getPointerTypeSizeInBits(Type *) const
Layout pointer size, in bits, based on the type.
Definition: DataLayout.cpp:671
bool isFloatingPointTy() const
Return true if this is one of the six floating-point types.
Definition: Type.h:162
Class to represent struct types.
Definition: DerivedTypes.h:238
unsigned getCallCost(const Function *F, ArrayRef< const Value *> Arguments, const User *U)
unsigned getArithmeticReductionCost(unsigned, Type *, bool)
unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy, ArrayRef< Type *> ParamTys, const User *U)
mir Rename Register Operands
bool isLegalMaskedStore(Type *DataType, MaybeAlign Alignment)
const APInt & getAPInt() const
bool isLegalToVectorizeLoad(LoadInst *LI) const
Type * getMemcpyLoopLoweringType(LLVMContext &Context, Value *Length, unsigned SrcAlign, unsigned DestAlign) const
bool isTruncateFree(Type *Ty1, Type *Ty2)
bool enableAggressiveInterleaving(bool LoopHasReductions)
Class to represent function types.
Definition: DerivedTypes.h:108
unsigned getCallCost(FunctionType *FTy, int NumArgs, const User *U)
int64_t getSExtValue() const
Get sign extended value.
Definition: APInt.h:1583
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:246
This node represents a polynomial recurrence on the trip count of the specified loop.
void getMemcpyLoopResidualLoweringType(SmallVectorImpl< Type *> &OpsOut, LLVMContext &Context, unsigned RemainingBytes, unsigned SrcAlign, unsigned DestAlign) const
TypeSize getTypeStoreSize(Type *Ty) const
Returns the maximum number of bytes that may be overwritten by storing the specified type...
Definition: DataLayout.h:455
PopcntSupportKind
Flags indicating the kind of support for population count.
APInt sextOrTrunc(unsigned width) const
Sign extend or truncate to width.
Definition: APInt.cpp:938
bool isLegalNTLoad(Type *DataType, Align Alignment)
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:32
const APInt & getValue() const
Return the constant as an APInt value reference.
Definition: Constants.h:137
An instruction for storing to memory.
Definition: Instructions.h:325
const SCEV * getStepRecurrence(ScalarEvolution &SE) const
Constructs and returns the recurrence indicating how much this expression steps by.
unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst, VectorType *VecTy, unsigned Index)
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:144
bool canSaveCmp(Loop *L, BranchInst **BI, ScalarEvolution *SE, LoopInfo *LI, DominatorTree *DT, AssumptionCache *AC, TargetLibraryInfo *LibInfo)
Value * getOperand(unsigned i) const
Definition: User.h:169
Analysis containing CSE Info
Definition: CSEInfo.cpp:20
Type * getScalarType() const
If this is a vector type, return the element type, otherwise return &#39;this&#39;.
Definition: Type.h:307
bool isLegalToVectorizeStore(StoreInst *SI) const
unsigned getMinimumVF(unsigned ElemWidth) const
Returns options for expansion of memcmp. IsZeroCmp is.
uint64_t getZExtValue() const
Return the constant as a 64-bit unsigned integer value after it has been zero extended as appropriate...
Definition: Constants.h:148
unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Ty, int Index, Type *SubTp)
unsigned getRegisterClassForType(bool Vector, Type *Ty=nullptr) const
unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef< Type *> Tys)
constexpr bool isPowerOf2_32(uint32_t Value)
Return true if the argument is a power of two > 0.
Definition: MathExtras.h:465
bool hasName() const
Definition: Value.h:252
Flags describing the kind of vector reduction.
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
This is an important class for using LLVM in a threaded context.
Definition: LLVMContext.h:64
unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy, const Instruction *I)
Conditional or Unconditional Branch instruction.
TargetTransformInfoImplBase(const TargetTransformInfoImplBase &Arg)
bool shouldFavorBackedgeIndex(const Loop *L) const
size_t size() const
size - Get the array size.
Definition: ArrayRef.h:148
This is an important base class in LLVM.
Definition: Constant.h:41
bool isPointerTy() const
True if this is an instance of PointerType.
Definition: Type.h:224
bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const
TypeSize getTypeAllocSize(Type *Ty) const
Returns the offset in bytes between successive objects of the specified type, including alignment pad...
Definition: DataLayout.h:487
unsigned getNumParams() const
Return the number of fixed parameters this function type requires.
Definition: DerivedTypes.h:144
unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm, Type *Ty)
bool hasVolatileVariant(Instruction *I, unsigned AddrSpace)
Expected to fold away in lowering.
AMDGPU Lower Kernel Arguments
unsigned getUserCost(const User *U, ArrayRef< const Value *> Operands)
bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg, int64_t Scale, unsigned AddrSpace, Instruction *I=nullptr)
param_iterator param_begin() const
Definition: DerivedTypes.h:133
Fast - This calling convention attempts to make calls as fast as possible (e.g.
Definition: CallingConv.h:42
unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI, unsigned &JTSize)
constexpr double e
Definition: MathExtras.h:57
unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef< unsigned > Indices, unsigned Alignment, unsigned AddressSpace, bool UseMaskForCond=false, bool UseMaskForGaps=false)
amdgpu Simplify well known AMD library false FunctionCallee Value * Arg
size_t arg_size() const
Definition: Function.h:728
bool isLegalMaskedLoad(Type *DataType, MaybeAlign Alignment)
Attributes of a target dependent hardware loop.
bool isLegalNTStore(Type *DataType, Align Alignment)
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs and address space casts.
Definition: Value.cpp:529
llvm::Optional< unsigned > getCacheAssociativity(TargetTransformInfo::CacheLevel Level) const
unsigned getNumberOfRegisters(unsigned ClassID) const
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
This struct is a compact representation of a valid (non-zero power of two) alignment.
Definition: Alignment.h:40
const SCEVConstant * getConstantStrideStep(ScalarEvolution *SE, const SCEV *Ptr)
OperandValueProperties
Additional properties of an operand&#39;s values.
int getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, const APInt &Imm, Type *Ty)
bool collectFlatAddressOperands(SmallVectorImpl< int > &OpIndexes, Intrinsic::ID IID) const
const char * getRegisterClassName(unsigned ClassID) const
unsigned getNumOperands() const
Definition: User.h:191
This is the shared class of boolean and integer constants.
Definition: Constants.h:83
Align max(MaybeAlign Lhs, Align Rhs)
Definition: Alignment.h:390
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize, unsigned ChainSizeInBytes, VectorType *VecTy) const
unsigned getCallCost(const Function *F, int NumArgs, const User *U)
unsigned getScalarSizeInBits() const LLVM_READONLY
If this is a vector type, return the getPrimitiveSizeInBits value for the element type...
Definition: Type.cpp:134
This struct is a compact representation of a valid (power of two) or undefined (0) alignment...
Definition: Alignment.h:117
unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract)
TargetTransformInfoImplBase(const DataLayout &DL)
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:837
Provides information about what library functions are available for the current target.
AddressSpace
Definition: NVPTXBaseInfo.h:21
iterator end() const
Definition: ArrayRef.h:137
bool isLegalInteger(uint64_t Width) const
Returns true if the specified type is known to be a native integer type supported by the CPU...
Definition: DataLayout.h:255
Type * getReturnType() const
Definition: DerivedTypes.h:129
bool useReductionIntrinsic(unsigned Opcode, Type *Ty, TTI::ReductionFlags Flags) const
bool allowsMisalignedMemoryAccesses(LLVMContext &Context, unsigned BitWidth, unsigned AddressSpace, unsigned Alignment, bool *Fast)
Intrinsic::ID getIntrinsicID() const LLVM_READONLY
getIntrinsicID - This method returns the ID number of the specified function, or Intrinsic::not_intri...
Definition: Function.h:193
bool shouldConsiderAddressTypePromotion(const Instruction &I, bool &AllowPromotionWithoutCommonHeader)
FunctionType * getFunctionType() const
Returns the FunctionType for me.
Definition: Function.h:163
int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg, int64_t Scale, unsigned AddrSpace)
Class to represent vector types.
Definition: DerivedTypes.h:432
Class for arbitrary precision integers.
Definition: APInt.h:69
unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const
unsigned getMinMaxReductionCost(Type *, Type *, bool, bool)
bool isLSRCostLess(TTI::LSRCost &C1, TTI::LSRCost &C2)
unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment, unsigned AddressSpace)
unsigned getCFInstrCost(unsigned Opcode)
uint64_t getElementOffset(unsigned Idx) const
Definition: DataLayout.h:601
This class represents an analyzed expression in the program.
Represents a single loop in the control flow graph.
Definition: LoopInfo.h:509
ArrayRef< T > drop_front(size_t N=1) const
Drop the first N elements of the array.
Definition: ArrayRef.h:187
Parameters that control the generic loop unrolling transformation.
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:214
Establish a view to a call site for examination.
Definition: CallSite.h:906
#define I(x, y, z)
Definition: MD5.cpp:58
unsigned getOperandsScalarizationOverhead(ArrayRef< const Value *> Args, unsigned VF)
int getGEPCost(Type *PointeeType, const Value *Ptr, ArrayRef< const Value *> Operands)
bool isIndexedStoreLegal(TTI::MemIndexedMode Mode, Type *Ty, const DataLayout &DL) const
bool areInlineCompatible(const Function *Caller, const Function *Callee) const
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:332
uint32_t Size
Definition: Profile.cpp:46
unsigned getAddressComputationCost(Type *Tp, ScalarEvolution *, const SCEV *)
Multiway switch.
bool rewriteIntrinsicWithAddressSpace(IntrinsicInst *II, Value *OldV, Value *NewV) const
TargetTransformInfoImplBase(TargetTransformInfoImplBase &&Arg)
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
The cost of a typical &#39;add&#39; instruction.
bool isHardwareLoopProfitable(Loop *L, ScalarEvolution &SE, AssumptionCache &AC, TargetLibraryInfo *LibInfo, HardwareLoopInfo &HWLoopInfo)
LLVM Value Representation.
Definition: Value.h:74
unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm, Type *Ty)
unsigned getOpcode() const
Return the opcode for this Instruction or ConstantExpr.
Definition: Operator.h:40
#define LLVM_FALLTHROUGH
LLVM_FALLTHROUGH - Mark fallthrough cases in switch statements.
Definition: Compiler.h:273
bool shouldExpandReduction(const IntrinsicInst *II) const
Attribute getFnAttribute(Attribute::AttrKind Kind) const
Return the attribute for the given attribute kind.
Definition: Function.h:333
bool shouldMaximizeVectorBandwidth(bool OptSize) const
const DataLayout & getDataLayout() const
Convenience struct for specifying and reasoning about fast-math flags.
Definition: Operator.h:159
OperandValueKind
Additional information about an operand&#39;s possible values.
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:48
This pass exposes codegen information to IR-level passes.
CacheLevel
The possible cache levels.
unsigned getRegisterBitWidth(bool Vector) const
Information about a load/store intrinsic defined by the target.
The cost of a &#39;div&#39; instruction on x86.
static IntegerType * getInt8Ty(LLVMContext &C)
Definition: Type.cpp:178
int getGEPCost(Type *PointeeType, const Value *Ptr, ArrayRef< const Value *> Operands)
constexpr char Args[]
Key for Kernel::Metadata::mArgs.
bool hasDivRemOp(Type *DataType, bool IsSigned)
TTI::PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit)
Value * getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst, Type *ExpectedType)
bool empty() const
empty - Check if the array is empty.
Definition: ArrayRef.h:143
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:43
This class represents a constant integer value.
unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, ArrayRef< Value *> Args, FastMathFlags FMF, unsigned VF)
TTI::MemCmpExpansionOptions enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const
ShuffleKind
The various kinds of shuffle patterns for vector queries.
gep_type_iterator gep_type_begin(const User *GEP)