1 //===- TargetTransformInfoImpl.h --------------------------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This file provides helpers for the implementation of
11 /// a TargetTransformInfo-conforming class.
13 //===----------------------------------------------------------------------===//
15 #ifndef LLVM_ANALYSIS_TARGETTRANSFORMINFOIMPL_H
16 #define LLVM_ANALYSIS_TARGETTRANSFORMINFOIMPL_H
18 #include "llvm/Analysis/TargetTransformInfo.h"
19 #include "llvm/IR/CallSite.h"
20 #include "llvm/IR/DataLayout.h"
21 #include "llvm/IR/Function.h"
22 #include "llvm/IR/GetElementPtrTypeIterator.h"
23 #include "llvm/IR/Operator.h"
24 #include "llvm/IR/Type.h"
28 /// \brief Base class for use as a mix-in that aids implementing
29 /// a TargetTransformInfo-compatible class.
30 class TargetTransformInfoImplBase {
32 typedef TargetTransformInfo TTI;
36 explicit TargetTransformInfoImplBase(const DataLayout &DL) : DL(DL) {}
39 // Provide value semantics. MSVC requires that we spell all of these out.
40 TargetTransformInfoImplBase(const TargetTransformInfoImplBase &Arg)
42 TargetTransformInfoImplBase(TargetTransformInfoImplBase &&Arg) : DL(Arg.DL) {}
44 const DataLayout &getDataLayout() const { return DL; }
46 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) {
49 // By default, just classify everything as 'basic'.
50 return TTI::TCC_Basic;
52 case Instruction::GetElementPtr:
53 llvm_unreachable("Use getGEPCost for GEP operations!");
55 case Instruction::BitCast:
56 assert(OpTy && "Cast instructions must provide the operand type");
57 if (Ty == OpTy || (Ty->isPointerTy() && OpTy->isPointerTy()))
58 // Identity and pointer-to-pointer casts are free.
61 // Otherwise, the default basic cost is used.
62 return TTI::TCC_Basic;
64 case Instruction::IntToPtr: {
65 // An inttoptr cast is free so long as the input is a legal integer type
66 // which doesn't contain values outside the range of a pointer.
67 unsigned OpSize = OpTy->getScalarSizeInBits();
68 if (DL.isLegalInteger(OpSize) &&
69 OpSize <= DL.getPointerTypeSizeInBits(Ty))
72 // Otherwise it's not a no-op.
73 return TTI::TCC_Basic;
75 case Instruction::PtrToInt: {
76 // A ptrtoint cast is free so long as the result is large enough to store
77 // the pointer, and a legal integer type.
78 unsigned DestSize = Ty->getScalarSizeInBits();
79 if (DL.isLegalInteger(DestSize) &&
80 DestSize >= DL.getPointerTypeSizeInBits(OpTy))
83 // Otherwise it's not a no-op.
84 return TTI::TCC_Basic;
86 case Instruction::Trunc:
87 // trunc to a native type is free (assuming the target has compare and
88 // shift-right of the same width).
89 if (DL.isLegalInteger(DL.getTypeSizeInBits(Ty)))
92 return TTI::TCC_Basic;
96 unsigned getGEPCost(Type *PointeeType, const Value *Ptr,
97 ArrayRef<const Value *> Operands) {
98 // In the basic model, we just assume that all-constant GEPs will be folded
99 // into their uses via addressing modes.
100 for (unsigned Idx = 0, Size = Operands.size(); Idx != Size; ++Idx)
101 if (!isa<Constant>(Operands[Idx]))
102 return TTI::TCC_Basic;
104 return TTI::TCC_Free;
107 unsigned getCallCost(FunctionType *FTy, int NumArgs) {
108 assert(FTy && "FunctionType must be provided to this routine.");
110 // The target-independent implementation just measures the size of the
111 // function by approximating that each argument will take on average one
112 // instruction to prepare.
115 // Set the argument number to the number of explicit arguments in the
117 NumArgs = FTy->getNumParams();
119 return TTI::TCC_Basic * (NumArgs + 1);
122 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
123 ArrayRef<Type *> ParamTys) {
126 // Intrinsics rarely (if ever) have normal argument setup constraints.
127 // Model them as having a basic instruction cost.
128 // FIXME: This is wrong for libc intrinsics.
129 return TTI::TCC_Basic;
131 case Intrinsic::annotation:
132 case Intrinsic::assume:
133 case Intrinsic::dbg_declare:
134 case Intrinsic::dbg_value:
135 case Intrinsic::invariant_start:
136 case Intrinsic::invariant_end:
137 case Intrinsic::lifetime_start:
138 case Intrinsic::lifetime_end:
139 case Intrinsic::objectsize:
140 case Intrinsic::ptr_annotation:
141 case Intrinsic::var_annotation:
142 case Intrinsic::experimental_gc_result_int:
143 case Intrinsic::experimental_gc_result_float:
144 case Intrinsic::experimental_gc_result_ptr:
145 case Intrinsic::experimental_gc_result:
146 case Intrinsic::experimental_gc_relocate:
147 // These intrinsics don't actually represent code after lowering.
148 return TTI::TCC_Free;
152 bool hasBranchDivergence() { return false; }
154 bool isSourceOfDivergence(const Value *V) { return false; }
156 bool isLoweredToCall(const Function *F) {
157 // FIXME: These should almost certainly not be handled here, and instead
158 // handled with the help of TLI or the target itself. This was largely
159 // ported from existing analysis heuristics here so that such refactorings
160 // can take place in the future.
162 if (F->isIntrinsic())
165 if (F->hasLocalLinkage() || !F->hasName())
168 StringRef Name = F->getName();
170 // These will all likely lower to a single selection DAG node.
171 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
172 Name == "fabs" || Name == "fabsf" || Name == "fabsl" || Name == "sin" ||
173 Name == "fmin" || Name == "fminf" || Name == "fminl" ||
174 Name == "fmax" || Name == "fmaxf" || Name == "fmaxl" ||
175 Name == "sinf" || Name == "sinl" || Name == "cos" || Name == "cosf" ||
176 Name == "cosl" || Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl")
179 // These are all likely to be optimized into something smaller.
180 if (Name == "pow" || Name == "powf" || Name == "powl" || Name == "exp2" ||
181 Name == "exp2l" || Name == "exp2f" || Name == "floor" ||
182 Name == "floorf" || Name == "ceil" || Name == "round" ||
183 Name == "ffs" || Name == "ffsl" || Name == "abs" || Name == "labs" ||
190 void getUnrollingPreferences(Loop *, TTI::UnrollingPreferences &) {}
192 bool isLegalAddImmediate(int64_t Imm) { return false; }
194 bool isLegalICmpImmediate(int64_t Imm) { return false; }
196 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
197 bool HasBaseReg, int64_t Scale,
198 unsigned AddrSpace) {
199 // Guess that only reg and reg+reg addressing is allowed. This heuristic is
200 // taken from the implementation of LSR.
201 return !BaseGV && BaseOffset == 0 && (Scale == 0 || Scale == 1);
204 bool isLegalMaskedStore(Type *DataType, int Consecutive) { return false; }
206 bool isLegalMaskedLoad(Type *DataType, int Consecutive) { return false; }
208 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
209 bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
210 // Guess that all legal addressing mode are free.
211 if (isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
217 bool isTruncateFree(Type *Ty1, Type *Ty2) { return false; }
219 bool isZExtFree(Type *Ty1, Type *Ty2) { return false; }
221 bool isProfitableToHoist(Instruction *I) { return true; }
223 bool isTypeLegal(Type *Ty) { return false; }
225 unsigned getJumpBufAlignment() { return 0; }
227 unsigned getJumpBufSize() { return 0; }
229 bool shouldBuildLookupTables() { return true; }
231 bool enableAggressiveInterleaving(bool LoopHasReductions) { return false; }
233 bool enableInterleavedAccessVectorization() { return false; }
235 TTI::PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) {
236 return TTI::PSK_Software;
239 bool haveFastSqrt(Type *Ty) { return false; }
241 unsigned getFPOpCost(Type *Ty) { return TargetTransformInfo::TCC_Basic; }
243 unsigned getIntImmCost(const APInt &Imm, Type *Ty) { return TTI::TCC_Basic; }
245 unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
247 return TTI::TCC_Free;
250 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
252 return TTI::TCC_Free;
255 unsigned getNumberOfRegisters(bool Vector) { return 8; }
257 unsigned getRegisterBitWidth(bool Vector) { return 32; }
259 unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
261 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
262 TTI::OperandValueKind Opd1Info,
263 TTI::OperandValueKind Opd2Info,
264 TTI::OperandValueProperties Opd1PropInfo,
265 TTI::OperandValueProperties Opd2PropInfo) {
269 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Ty, int Index,
274 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) { return 1; }
276 unsigned getCFInstrCost(unsigned Opcode) { return 1; }
278 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy) {
282 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
286 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
287 unsigned AddressSpace) {
291 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
292 unsigned AddressSpace) {
296 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
298 ArrayRef<unsigned> Indices,
300 unsigned AddressSpace) {
304 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
305 ArrayRef<Type *> Tys) {
309 unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
313 unsigned getNumberOfParts(Type *Tp) { return 0; }
315 unsigned getAddressComputationCost(Type *Tp, bool) { return 0; }
317 unsigned getReductionCost(unsigned, Type *, bool) { return 1; }
319 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) { return 0; }
321 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) {
325 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
326 Type *ExpectedType) {
330 bool areInlineCompatible(const Function *Caller,
331 const Function *Callee) const {
332 return (Caller->getFnAttribute("target-cpu") ==
333 Callee->getFnAttribute("target-cpu")) &&
334 (Caller->getFnAttribute("target-features") ==
335 Callee->getFnAttribute("target-features"));
339 /// \brief CRTP base class for use as a mix-in that aids implementing
340 /// a TargetTransformInfo-compatible class.
341 template <typename T>
342 class TargetTransformInfoImplCRTPBase : public TargetTransformInfoImplBase {
344 typedef TargetTransformInfoImplBase BaseT;
347 explicit TargetTransformInfoImplCRTPBase(const DataLayout &DL) : BaseT(DL) {}
350 // Provide value semantics. MSVC requires that we spell all of these out.
351 TargetTransformInfoImplCRTPBase(const TargetTransformInfoImplCRTPBase &Arg)
352 : BaseT(static_cast<const BaseT &>(Arg)) {}
353 TargetTransformInfoImplCRTPBase(TargetTransformInfoImplCRTPBase &&Arg)
354 : BaseT(std::move(static_cast<BaseT &>(Arg))) {}
356 using BaseT::getCallCost;
358 unsigned getCallCost(const Function *F, int NumArgs) {
359 assert(F && "A concrete function must be provided to this routine.");
362 // Set the argument number to the number of explicit arguments in the
364 NumArgs = F->arg_size();
366 if (Intrinsic::ID IID = F->getIntrinsicID()) {
367 FunctionType *FTy = F->getFunctionType();
368 SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end());
369 return static_cast<T *>(this)
370 ->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys);
373 if (!static_cast<T *>(this)->isLoweredToCall(F))
374 return TTI::TCC_Basic; // Give a basic cost if it will be lowered
377 return static_cast<T *>(this)->getCallCost(F->getFunctionType(), NumArgs);
380 unsigned getCallCost(const Function *F, ArrayRef<const Value *> Arguments) {
381 // Simply delegate to generic handling of the call.
382 // FIXME: We should use instsimplify or something else to catch calls which
383 // will constant fold with these arguments.
384 return static_cast<T *>(this)->getCallCost(F, Arguments.size());
387 using BaseT::getGEPCost;
389 unsigned getGEPCost(Type *PointeeType, const Value *Ptr,
390 ArrayRef<const Value *> Operands) {
391 const GlobalValue *BaseGV = nullptr;
392 if (Ptr != nullptr) {
393 // TODO: will remove this when pointers have an opaque type.
394 assert(Ptr->getType()->getScalarType()->getPointerElementType() ==
396 "explicit pointee type doesn't match operand's pointee type");
397 BaseGV = dyn_cast<GlobalValue>(Ptr->stripPointerCasts());
399 bool HasBaseReg = (BaseGV == nullptr);
400 int64_t BaseOffset = 0;
403 // Assumes the address space is 0 when Ptr is nullptr.
405 (Ptr == nullptr ? 0 : Ptr->getType()->getPointerAddressSpace());
406 auto GTI = gep_type_begin(PointerType::get(PointeeType, AS), Operands);
407 for (auto I = Operands.begin(); I != Operands.end(); ++I, ++GTI) {
408 if (isa<SequentialType>(*GTI)) {
409 int64_t ElementSize = DL.getTypeAllocSize(GTI.getIndexedType());
410 if (const ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I)) {
411 BaseOffset += ConstIdx->getSExtValue() * ElementSize;
413 // Needs scale register.
415 // No addressing mode takes two scale registers.
416 return TTI::TCC_Basic;
421 StructType *STy = cast<StructType>(*GTI);
422 uint64_t Field = cast<ConstantInt>(*I)->getZExtValue();
423 BaseOffset += DL.getStructLayout(STy)->getElementOffset(Field);
427 if (static_cast<T *>(this)->isLegalAddressingMode(
428 PointerType::get(*GTI, AS), const_cast<GlobalValue *>(BaseGV),
429 BaseOffset, HasBaseReg, Scale, AS)) {
430 return TTI::TCC_Free;
432 return TTI::TCC_Basic;
435 using BaseT::getIntrinsicCost;
437 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
438 ArrayRef<const Value *> Arguments) {
439 // Delegate to the generic intrinsic handling code. This mostly provides an
440 // opportunity for targets to (for example) special case the cost of
441 // certain intrinsics based on constants used as arguments.
442 SmallVector<Type *, 8> ParamTys;
443 ParamTys.reserve(Arguments.size());
444 for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx)
445 ParamTys.push_back(Arguments[Idx]->getType());
446 return static_cast<T *>(this)->getIntrinsicCost(IID, RetTy, ParamTys);
449 unsigned getUserCost(const User *U) {
451 return TTI::TCC_Free; // Model all PHI nodes as free.
453 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U)) {
454 SmallVector<Value *, 4> Indices(GEP->idx_begin(), GEP->idx_end());
455 return static_cast<T *>(this)->getGEPCost(
456 GEP->getSourceElementType(), GEP->getPointerOperand(), Indices);
459 if (auto CS = ImmutableCallSite(U)) {
460 const Function *F = CS.getCalledFunction();
462 // Just use the called value type.
463 Type *FTy = CS.getCalledValue()->getType()->getPointerElementType();
464 return static_cast<T *>(this)
465 ->getCallCost(cast<FunctionType>(FTy), CS.arg_size());
468 SmallVector<const Value *, 8> Arguments(CS.arg_begin(), CS.arg_end());
469 return static_cast<T *>(this)->getCallCost(F, Arguments);
472 if (const CastInst *CI = dyn_cast<CastInst>(U)) {
473 // Result of a cmp instruction is often extended (to be used by other
474 // cmp instructions, logical or return instructions). These are usually
475 // nop on most sane targets.
476 if (isa<CmpInst>(CI->getOperand(0)))
477 return TTI::TCC_Free;
480 return static_cast<T *>(this)->getOperationCost(
481 Operator::getOpcode(U), U->getType(),
482 U->getNumOperands() == 1 ? U->getOperand(0)->getType() : nullptr);