LLVM  9.0.0svn
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,
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 
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  unsigned getMemcpyCost(const Instruction *I) {
144  return TTI::TCC_Expensive;
145  }
146 
147  bool hasBranchDivergence() { return false; }
148 
149  bool isSourceOfDivergence(const Value *V) { return false; }
150 
151  bool isAlwaysUniform(const Value *V) { return false; }
152 
153  unsigned getFlatAddressSpace () {
154  return -1;
155  }
156 
157  bool isLoweredToCall(const Function *F) {
158  assert(F && "A concrete function must be provided to this routine.");
159 
160  // FIXME: These should almost certainly not be handled here, and instead
161  // handled with the help of TLI or the target itself. This was largely
162  // ported from existing analysis heuristics here so that such refactorings
163  // can take place in the future.
164 
165  if (F->isIntrinsic())
166  return false;
167 
168  if (F->hasLocalLinkage() || !F->hasName())
169  return true;
170 
171  StringRef Name = F->getName();
172 
173  // These will all likely lower to a single selection DAG node.
174  if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
175  Name == "fabs" || Name == "fabsf" || Name == "fabsl" || Name == "sin" ||
176  Name == "fmin" || Name == "fminf" || Name == "fminl" ||
177  Name == "fmax" || Name == "fmaxf" || Name == "fmaxl" ||
178  Name == "sinf" || Name == "sinl" || Name == "cos" || Name == "cosf" ||
179  Name == "cosl" || Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl")
180  return false;
181 
182  // These are all likely to be optimized into something smaller.
183  if (Name == "pow" || Name == "powf" || Name == "powl" || Name == "exp2" ||
184  Name == "exp2l" || Name == "exp2f" || Name == "floor" ||
185  Name == "floorf" || Name == "ceil" || Name == "round" ||
186  Name == "ffs" || Name == "ffsl" || Name == "abs" || Name == "labs" ||
187  Name == "llabs")
188  return false;
189 
190  return true;
191  }
192 
195 
196  bool isLegalAddImmediate(int64_t Imm) { return false; }
197 
198  bool isLegalICmpImmediate(int64_t Imm) { return false; }
199 
200  bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
201  bool HasBaseReg, int64_t Scale,
202  unsigned AddrSpace, Instruction *I = nullptr) {
203  // Guess that only reg and reg+reg addressing is allowed. This heuristic is
204  // taken from the implementation of LSR.
205  return !BaseGV && BaseOffset == 0 && (Scale == 0 || Scale == 1);
206  }
207 
209  return std::tie(C1.NumRegs, C1.AddRecCost, C1.NumIVMuls, C1.NumBaseAdds,
210  C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
211  std::tie(C2.NumRegs, C2.AddRecCost, C2.NumIVMuls, C2.NumBaseAdds,
212  C2.ScaleCost, C2.ImmCost, C2.SetupCost);
213  }
214 
215  bool canMacroFuseCmp() { return false; }
216 
217  bool shouldFavorPostInc() const { return false; }
218 
219  bool shouldFavorBackedgeIndex(const Loop *L) const { return false; }
220 
221  bool isLegalMaskedStore(Type *DataType) { return false; }
222 
223  bool isLegalMaskedLoad(Type *DataType) { return false; }
224 
225  bool isLegalMaskedScatter(Type *DataType) { return false; }
226 
227  bool isLegalMaskedGather(Type *DataType) { return false; }
228 
229  bool isLegalMaskedCompressStore(Type *DataType) { return false; }
230 
231  bool isLegalMaskedExpandLoad(Type *DataType) { return false; }
232 
233  bool hasDivRemOp(Type *DataType, bool IsSigned) { return false; }
234 
235  bool hasVolatileVariant(Instruction *I, unsigned AddrSpace) { return false; }
236 
237  bool prefersVectorizedAddressing() { return true; }
238 
239  int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
240  bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
241  // Guess that all legal addressing mode are free.
242  if (isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
243  Scale, AddrSpace))
244  return 0;
245  return -1;
246  }
247 
248  bool LSRWithInstrQueries() { return false; }
249 
250  bool isTruncateFree(Type *Ty1, Type *Ty2) { return false; }
251 
252  bool isProfitableToHoist(Instruction *I) { return true; }
253 
254  bool useAA() { return false; }
255 
256  bool isTypeLegal(Type *Ty) { return false; }
257 
258  unsigned getJumpBufAlignment() { return 0; }
259 
260  unsigned getJumpBufSize() { return 0; }
261 
262  bool shouldBuildLookupTables() { return true; }
264 
265  bool useColdCCForColdCall(Function &F) { return false; }
266 
267  unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
268  return 0;
269  }
270 
272  unsigned VF) { return 0; }
273 
275 
276  bool enableAggressiveInterleaving(bool LoopHasReductions) { return false; }
277 
279  bool IsZeroCmp) const {
280  return nullptr;
281  }
282 
283  bool enableInterleavedAccessVectorization() { return false; }
284 
286 
287  bool isFPVectorizationPotentiallyUnsafe() { return false; }
288 
290  unsigned BitWidth,
291  unsigned AddressSpace,
292  unsigned Alignment,
293  bool *Fast) { return false; }
294 
295  TTI::PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) {
296  return TTI::PSK_Software;
297  }
298 
299  bool haveFastSqrt(Type *Ty) { return false; }
300 
301  bool isFCmpOrdCheaperThanFCmpZero(Type *Ty) { return true; }
302 
304 
305  int getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
306  Type *Ty) {
307  return 0;
308  }
309 
310  unsigned getIntImmCost(const APInt &Imm, Type *Ty) { return TTI::TCC_Basic; }
311 
312  unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
313  Type *Ty) {
314  return TTI::TCC_Free;
315  }
316 
317  unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
318  Type *Ty) {
319  return TTI::TCC_Free;
320  }
321 
322  unsigned getNumberOfRegisters(bool Vector) { return 8; }
323 
324  unsigned getRegisterBitWidth(bool Vector) const { return 32; }
325 
326  unsigned getMinVectorRegisterBitWidth() { return 128; }
327 
328  bool shouldMaximizeVectorBandwidth(bool OptSize) const { return false; }
329 
330  unsigned getMinimumVF(unsigned ElemWidth) const { return 0; }
331 
332  bool
334  bool &AllowPromotionWithoutCommonHeader) {
335  AllowPromotionWithoutCommonHeader = false;
336  return false;
337  }
338 
339  unsigned getCacheLineSize() { return 0; }
340 
342  switch (Level) {
346  return llvm::Optional<unsigned>();
347  }
348 
349  llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
350  }
351 
354  switch (Level) {
358  return llvm::Optional<unsigned>();
359  }
360 
361  llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
362  }
363 
364  unsigned getPrefetchDistance() { return 0; }
365 
366  unsigned getMinPrefetchStride() { return 1; }
367 
368  unsigned getMaxPrefetchIterationsAhead() { return UINT_MAX; }
369 
370  unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
371 
372  unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
373  TTI::OperandValueKind Opd1Info,
374  TTI::OperandValueKind Opd2Info,
375  TTI::OperandValueProperties Opd1PropInfo,
376  TTI::OperandValueProperties Opd2PropInfo,
378  return 1;
379  }
380 
382  Type *SubTp) {
383  return 1;
384  }
385 
386  unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
387  const Instruction *I) { return 1; }
388 
389  unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst,
390  VectorType *VecTy, unsigned Index) {
391  return 1;
392  }
393 
394  unsigned getCFInstrCost(unsigned Opcode) { return 1; }
395 
396  unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
397  const Instruction *I) {
398  return 1;
399  }
400 
401  unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
402  return 1;
403  }
404 
405  unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
406  unsigned AddressSpace, const Instruction *I) {
407  return 1;
408  }
409 
410  unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
411  unsigned AddressSpace) {
412  return 1;
413  }
414 
415  unsigned getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
416  bool VariableMask,
417  unsigned Alignment) {
418  return 1;
419  }
420 
421  unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
422  unsigned Factor,
423  ArrayRef<unsigned> Indices,
424  unsigned Alignment, unsigned AddressSpace,
425  bool UseMaskForCond = false,
426  bool UseMaskForGaps = false) {
427  return 1;
428  }
429 
432  unsigned ScalarizationCostPassed) {
433  return 1;
434  }
436  ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) {
437  return 1;
438  }
439 
440  unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
441  return 1;
442  }
443 
444  unsigned getNumberOfParts(Type *Tp) { return 0; }
445 
447  const SCEV *) {
448  return 0;
449  }
450 
451  unsigned getArithmeticReductionCost(unsigned, Type *, bool) { return 1; }
452 
453  unsigned getMinMaxReductionCost(Type *, Type *, bool, bool) { return 1; }
454 
455  unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) { return 0; }
456 
458  return false;
459  }
460 
462  // Note for overrides: You must ensure for all element unordered-atomic
463  // memory intrinsics that all power-of-2 element sizes up to, and
464  // including, the return value of this method have a corresponding
465  // runtime lib call. These runtime lib call definitions can be found
466  // in RuntimeLibcalls.h
467  return 0;
468  }
469 
471  Type *ExpectedType) {
472  return nullptr;
473  }
474 
476  unsigned SrcAlign, unsigned DestAlign) const {
477  return Type::getInt8Ty(Context);
478  }
479 
482  unsigned RemainingBytes,
483  unsigned SrcAlign,
484  unsigned DestAlign) const {
485  for (unsigned i = 0; i != RemainingBytes; ++i)
486  OpsOut.push_back(Type::getInt8Ty(Context));
487  }
488 
489  bool areInlineCompatible(const Function *Caller,
490  const Function *Callee) const {
491  return (Caller->getFnAttribute("target-cpu") ==
492  Callee->getFnAttribute("target-cpu")) &&
493  (Caller->getFnAttribute("target-features") ==
494  Callee->getFnAttribute("target-features"));
495  }
496 
497  bool areFunctionArgsABICompatible(const Function *Caller, const Function *Callee,
499  return (Caller->getFnAttribute("target-cpu") ==
500  Callee->getFnAttribute("target-cpu")) &&
501  (Caller->getFnAttribute("target-features") ==
502  Callee->getFnAttribute("target-features"));
503  }
504 
506  const DataLayout &DL) const {
507  return false;
508  }
509 
511  const DataLayout &DL) const {
512  return false;
513  }
514 
515  unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const { return 128; }
516 
517  bool isLegalToVectorizeLoad(LoadInst *LI) const { return true; }
518 
519  bool isLegalToVectorizeStore(StoreInst *SI) const { return true; }
520 
521  bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
522  unsigned Alignment,
523  unsigned AddrSpace) const {
524  return true;
525  }
526 
527  bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
528  unsigned Alignment,
529  unsigned AddrSpace) const {
530  return true;
531  }
532 
533  unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
534  unsigned ChainSizeInBytes,
535  VectorType *VecTy) const {
536  return VF;
537  }
538 
539  unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
540  unsigned ChainSizeInBytes,
541  VectorType *VecTy) const {
542  return VF;
543  }
544 
545  bool useReductionIntrinsic(unsigned Opcode, Type *Ty,
546  TTI::ReductionFlags Flags) const {
547  return false;
548  }
549 
550  bool shouldExpandReduction(const IntrinsicInst *II) const {
551  return true;
552  }
553 
554 protected:
555  // Obtain the minimum required size to hold the value (without the sign)
556  // In case of a vector it returns the min required size for one element.
557  unsigned minRequiredElementSize(const Value* Val, bool &isSigned) {
558  if (isa<ConstantDataVector>(Val) || isa<ConstantVector>(Val)) {
559  const auto* VectorValue = cast<Constant>(Val);
560 
561  // In case of a vector need to pick the max between the min
562  // required size for each element
563  auto *VT = cast<VectorType>(Val->getType());
564 
565  // Assume unsigned elements
566  isSigned = false;
567 
568  // The max required size is the total vector width divided by num
569  // of elements in the vector
570  unsigned MaxRequiredSize = VT->getBitWidth() / VT->getNumElements();
571 
572  unsigned MinRequiredSize = 0;
573  for(unsigned i = 0, e = VT->getNumElements(); i < e; ++i) {
574  if (auto* IntElement =
575  dyn_cast<ConstantInt>(VectorValue->getAggregateElement(i))) {
576  bool signedElement = IntElement->getValue().isNegative();
577  // Get the element min required size.
578  unsigned ElementMinRequiredSize =
579  IntElement->getValue().getMinSignedBits() - 1;
580  // In case one element is signed then all the vector is signed.
581  isSigned |= signedElement;
582  // Save the max required bit size between all the elements.
583  MinRequiredSize = std::max(MinRequiredSize, ElementMinRequiredSize);
584  }
585  else {
586  // not an int constant element
587  return MaxRequiredSize;
588  }
589  }
590  return MinRequiredSize;
591  }
592 
593  if (const auto* CI = dyn_cast<ConstantInt>(Val)) {
594  isSigned = CI->getValue().isNegative();
595  return CI->getValue().getMinSignedBits() - 1;
596  }
597 
598  if (const auto* Cast = dyn_cast<SExtInst>(Val)) {
599  isSigned = true;
600  return Cast->getSrcTy()->getScalarSizeInBits() - 1;
601  }
602 
603  if (const auto* Cast = dyn_cast<ZExtInst>(Val)) {
604  isSigned = false;
605  return Cast->getSrcTy()->getScalarSizeInBits();
606  }
607 
608  isSigned = false;
609  return Val->getType()->getScalarSizeInBits();
610  }
611 
612  bool isStridedAccess(const SCEV *Ptr) {
613  return Ptr && isa<SCEVAddRecExpr>(Ptr);
614  }
615 
617  const SCEV *Ptr) {
618  if (!isStridedAccess(Ptr))
619  return nullptr;
620  const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ptr);
621  return dyn_cast<SCEVConstant>(AddRec->getStepRecurrence(*SE));
622  }
623 
625  int64_t MergeDistance) {
626  const SCEVConstant *Step = getConstantStrideStep(SE, Ptr);
627  if (!Step)
628  return false;
629  APInt StrideVal = Step->getAPInt();
630  if (StrideVal.getBitWidth() > 64)
631  return false;
632  // FIXME: Need to take absolute value for negative stride case.
633  return StrideVal.getSExtValue() < MergeDistance;
634  }
635 };
636 
637 /// CRTP base class for use as a mix-in that aids implementing
638 /// a TargetTransformInfo-compatible class.
639 template <typename T>
641 private:
643 
644 protected:
645  explicit TargetTransformInfoImplCRTPBase(const DataLayout &DL) : BaseT(DL) {}
646 
647 public:
648  using BaseT::getCallCost;
649 
650  unsigned getCallCost(const Function *F, int NumArgs, const User *U) {
651  assert(F && "A concrete function must be provided to this routine.");
652 
653  if (NumArgs < 0)
654  // Set the argument number to the number of explicit arguments in the
655  // function.
656  NumArgs = F->arg_size();
657 
658  if (Intrinsic::ID IID = F->getIntrinsicID()) {
659  FunctionType *FTy = F->getFunctionType();
660  SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end());
661  return static_cast<T *>(this)
662  ->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys, U);
663  }
664 
665  if (!static_cast<T *>(this)->isLoweredToCall(F))
666  return TTI::TCC_Basic; // Give a basic cost if it will be lowered
667  // directly.
668 
669  return static_cast<T *>(this)->getCallCost(F->getFunctionType(), NumArgs, U);
670  }
671 
673  const User *U) {
674  // Simply delegate to generic handling of the call.
675  // FIXME: We should use instsimplify or something else to catch calls which
676  // will constant fold with these arguments.
677  return static_cast<T *>(this)->getCallCost(F, Arguments.size(), U);
678  }
679 
680  using BaseT::getGEPCost;
681 
682  int getGEPCost(Type *PointeeType, const Value *Ptr,
683  ArrayRef<const Value *> Operands) {
684  const GlobalValue *BaseGV = nullptr;
685  if (Ptr != nullptr) {
686  // TODO: will remove this when pointers have an opaque type.
688  PointeeType &&
689  "explicit pointee type doesn't match operand's pointee type");
690  BaseGV = dyn_cast<GlobalValue>(Ptr->stripPointerCasts());
691  }
692  bool HasBaseReg = (BaseGV == nullptr);
693 
694  auto PtrSizeBits = DL.getPointerTypeSizeInBits(Ptr->getType());
695  APInt BaseOffset(PtrSizeBits, 0);
696  int64_t Scale = 0;
697 
698  auto GTI = gep_type_begin(PointeeType, Operands);
699  Type *TargetType = nullptr;
700 
701  // Handle the case where the GEP instruction has a single operand,
702  // the basis, therefore TargetType is a nullptr.
703  if (Operands.empty())
704  return !BaseGV ? TTI::TCC_Free : TTI::TCC_Basic;
705 
706  for (auto I = Operands.begin(); I != Operands.end(); ++I, ++GTI) {
707  TargetType = GTI.getIndexedType();
708  // We assume that the cost of Scalar GEP with constant index and the
709  // cost of Vector GEP with splat constant index are the same.
710  const ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I);
711  if (!ConstIdx)
712  if (auto Splat = getSplatValue(*I))
713  ConstIdx = dyn_cast<ConstantInt>(Splat);
714  if (StructType *STy = GTI.getStructTypeOrNull()) {
715  // For structures the index is always splat or scalar constant
716  assert(ConstIdx && "Unexpected GEP index");
717  uint64_t Field = ConstIdx->getZExtValue();
718  BaseOffset += DL.getStructLayout(STy)->getElementOffset(Field);
719  } else {
720  int64_t ElementSize = DL.getTypeAllocSize(GTI.getIndexedType());
721  if (ConstIdx) {
722  BaseOffset +=
723  ConstIdx->getValue().sextOrTrunc(PtrSizeBits) * ElementSize;
724  } else {
725  // Needs scale register.
726  if (Scale != 0)
727  // No addressing mode takes two scale registers.
728  return TTI::TCC_Basic;
729  Scale = ElementSize;
730  }
731  }
732  }
733 
734  // Assumes the address space is 0 when Ptr is nullptr.
735  unsigned AS =
736  (Ptr == nullptr ? 0 : Ptr->getType()->getPointerAddressSpace());
737 
738  if (static_cast<T *>(this)->isLegalAddressingMode(
739  TargetType, const_cast<GlobalValue *>(BaseGV),
740  BaseOffset.sextOrTrunc(64).getSExtValue(), HasBaseReg, Scale, AS))
741  return TTI::TCC_Free;
742  return TTI::TCC_Basic;
743  }
744 
745  unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
746  ArrayRef<Type *> ParamTys, const User *U) {
747  switch (IID) {
748  default:
749  // Intrinsics rarely (if ever) have normal argument setup constraints.
750  // Model them as having a basic instruction cost.
751  return TTI::TCC_Basic;
752 
753  // TODO: other libc intrinsics.
754  case Intrinsic::memcpy:
755  return static_cast<T *>(this)->getMemcpyCost(dyn_cast<Instruction>(U));
756 
757  case Intrinsic::annotation:
758  case Intrinsic::assume:
759  case Intrinsic::sideeffect:
760  case Intrinsic::dbg_declare:
761  case Intrinsic::dbg_value:
762  case Intrinsic::dbg_label:
763  case Intrinsic::invariant_start:
764  case Intrinsic::invariant_end:
765  case Intrinsic::launder_invariant_group:
766  case Intrinsic::strip_invariant_group:
767  case Intrinsic::is_constant:
768  case Intrinsic::lifetime_start:
769  case Intrinsic::lifetime_end:
770  case Intrinsic::objectsize:
771  case Intrinsic::ptr_annotation:
772  case Intrinsic::var_annotation:
773  case Intrinsic::experimental_gc_result:
774  case Intrinsic::experimental_gc_relocate:
775  case Intrinsic::coro_alloc:
776  case Intrinsic::coro_begin:
777  case Intrinsic::coro_free:
778  case Intrinsic::coro_end:
779  case Intrinsic::coro_frame:
780  case Intrinsic::coro_size:
781  case Intrinsic::coro_suspend:
782  case Intrinsic::coro_param:
783  case Intrinsic::coro_subfn_addr:
784  // These intrinsics don't actually represent code after lowering.
785  return TTI::TCC_Free;
786  }
787  }
788 
789  unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
791  // Delegate to the generic intrinsic handling code. This mostly provides an
792  // opportunity for targets to (for example) special case the cost of
793  // certain intrinsics based on constants used as arguments.
794  SmallVector<Type *, 8> ParamTys;
795  ParamTys.reserve(Arguments.size());
796  for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx)
797  ParamTys.push_back(Arguments[Idx]->getType());
798  return static_cast<T *>(this)->getIntrinsicCost(IID, RetTy, ParamTys, U);
799  }
800 
801  unsigned getUserCost(const User *U, ArrayRef<const Value *> Operands) {
802  if (isa<PHINode>(U))
803  return TTI::TCC_Free; // Model all PHI nodes as free.
804 
805  // Static alloca doesn't generate target instructions.
806  if (auto *A = dyn_cast<AllocaInst>(U))
807  if (A->isStaticAlloca())
808  return TTI::TCC_Free;
809 
810  if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U)) {
811  return static_cast<T *>(this)->getGEPCost(GEP->getSourceElementType(),
812  GEP->getPointerOperand(),
813  Operands.drop_front());
814  }
815 
816  if (auto CS = ImmutableCallSite(U)) {
817  const Function *F = CS.getCalledFunction();
818  if (!F) {
819  // Just use the called value type.
820  Type *FTy = CS.getCalledValue()->getType()->getPointerElementType();
821  return static_cast<T *>(this)
822  ->getCallCost(cast<FunctionType>(FTy), CS.arg_size(), U);
823  }
824 
825  SmallVector<const Value *, 8> Arguments(CS.arg_begin(), CS.arg_end());
826  return static_cast<T *>(this)->getCallCost(F, Arguments, U);
827  }
828 
829  if (const CastInst *CI = dyn_cast<CastInst>(U)) {
830  // Result of a cmp instruction is often extended (to be used by other
831  // cmp instructions, logical or return instructions). These are usually
832  // nop on most sane targets.
833  if (isa<CmpInst>(CI->getOperand(0)))
834  return TTI::TCC_Free;
835  if (isa<SExtInst>(CI) || isa<ZExtInst>(CI) || isa<FPExtInst>(CI))
836  return static_cast<T *>(this)->getExtCost(CI, Operands.back());
837  }
838 
839  return static_cast<T *>(this)->getOperationCost(
840  Operator::getOpcode(U), U->getType(),
841  U->getNumOperands() == 1 ? U->getOperand(0)->getType() : nullptr);
842  }
843 
846  I->value_op_end());
847  if (getUserCost(I, Operands) == TTI::TCC_Free)
848  return 0;
849 
850  if (isa<LoadInst>(I))
851  return 4;
852 
853  Type *DstTy = I->getType();
854 
855  // Usually an intrinsic is a simple instruction.
856  // A real function call is much slower.
857  if (auto *CI = dyn_cast<CallInst>(I)) {
858  const Function *F = CI->getCalledFunction();
859  if (!F || static_cast<T *>(this)->isLoweredToCall(F))
860  return 40;
861  // Some intrinsics return a value and a flag, we use the value type
862  // to decide its latency.
863  if (StructType* StructTy = dyn_cast<StructType>(DstTy))
864  DstTy = StructTy->getElementType(0);
865  // Fall through to simple instructions.
866  }
867 
868  if (VectorType *VectorTy = dyn_cast<VectorType>(DstTy))
869  DstTy = VectorTy->getElementType();
870  if (DstTy->isFloatingPointTy())
871  return 3;
872 
873  return 1;
874  }
875 };
876 }
877 
878 #endif
uint64_t CallInst * C
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:110
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:435
unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, ArrayRef< Type *> Tys, FastMathFlags FMF, unsigned ScalarizationCostPassed)
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
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:607
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)
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:534
param_iterator param_end() const
Definition: DerivedTypes.h:128
An instruction for reading from memory.
Definition: Instructions.h:167
Hexagon Common GEP
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:1508
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:375
This is the base class for all instructions that perform data casts.
Definition: InstrTypes.h:353
unsigned getPointerTypeSizeInBits(Type *) const
Layout pointer size, in bits, based on the type.
Definition: DataLayout.cpp:665
bool isFloatingPointTy() const
Return true if this is one of the six floating-point types.
Definition: Type.h:161
Class to represent struct types.
Definition: DerivedTypes.h:232
unsigned getCallCost(const Function *F, ArrayRef< const Value *> Arguments, const User *U)
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:41
unsigned getArithmeticReductionCost(unsigned, Type *, bool)
unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy, ArrayRef< Type *> ParamTys, const User *U)
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:102
unsigned getCallCost(FunctionType *FTy, int NumArgs, const User *U)
int64_t getSExtValue() const
Get sign extended value.
Definition: APInt.h:1574
llvm::Optional< unsigned > getCacheSize(TargetTransformInfo::CacheLevel Level)
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:244
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
PopcntSupportKind
Flags indicating the kind of support for population count.
APInt sextOrTrunc(unsigned width) const
Sign extend or truncate to width.
Definition: APInt.cpp:883
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:320
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)
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:303
bool isLegalToVectorizeStore(StoreInst *SI) const
unsigned getMinimumVF(unsigned ElemWidth) const
If not nullptr, enable inline expansion of memcmp.
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 getCallInstrCost(Function *F, Type *RetTy, ArrayRef< Type *> Tys)
bool hasName() const
Definition: Value.h:250
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:45
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)
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:223
bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const
unsigned getNumParams() const
Return the number of fixed parameters this function type requires.
Definition: DerivedTypes.h:138
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:127
const TTI::MemCmpExpansionOptions * enableMemCmpExpansion(bool IsZeroCmp) const
unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI, unsigned &JTSize)
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:700
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs, and aliases.
Definition: Value.cpp:529
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
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)
unsigned getNumOperands() const
Definition: User.h:191
This is the shared class of boolean and integer constants.
Definition: Constants.h:83
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:129
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:841
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:254
Type * getReturnType() const
Definition: DerivedTypes.h:123
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:424
Fast - This calling convention attempts to make calls as fast as possible (e.g.
Definition: CallingConv.h:42
Class for arbitrary precision integers.
Definition: APInt.h:69
unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const
unsigned getMinMaxReductionCost(Type *, Type *, bool, bool)
uint64_t getTypeSizeInBits(Type *Ty) const
Size examples:
Definition: DataLayout.h:593
bool isLSRCostLess(TTI::LSRCost &C1, TTI::LSRCost &C2)
uint64_t getTypeAllocSize(Type *Ty) const
Returns the offset in bytes between successive objects of the specified type, including alignment pad...
Definition: DataLayout.h:461
unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment, unsigned AddressSpace)
unsigned getCFInstrCost(unsigned Opcode)
uint64_t getElementOffset(unsigned Idx) const
Definition: DataLayout.h:576
This class represents an analyzed expression in the program.
Represents a single loop in the control flow graph.
Definition: LoopInfo.h:464
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:892
#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.
TargetTransformInfoImplBase(TargetTransformInfoImplBase &&Arg)
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
The cost of a typical &#39;add&#39; instruction.
LLVM Value Representation.
Definition: Value.h:72
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:250
bool shouldExpandReduction(const IntrinsicInst *II) const
Attribute getFnAttribute(Attribute::AttrKind Kind) const
Return the attribute for the given attribute kind.
Definition: Function.h:330
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.
llvm::Optional< unsigned > getCacheAssociativity(TargetTransformInfo::CacheLevel Level)
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:173
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)
ShuffleKind
The various kinds of shuffle patterns for vector queries.
gep_type_iterator gep_type_begin(const User *GEP)