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/Operator.h"
23 #include "llvm/IR/Type.h"
27 /// \brief Base class for use as a mix-in that aids implementing
28 /// a TargetTransformInfo-compatible class.
29 class TargetTransformInfoImplBase {
31 typedef TargetTransformInfo TTI;
35 explicit TargetTransformInfoImplBase(const DataLayout &DL) : DL(DL) {}
38 // Provide value semantics. MSVC requires that we spell all of these out.
39 TargetTransformInfoImplBase(const TargetTransformInfoImplBase &Arg)
41 TargetTransformInfoImplBase(TargetTransformInfoImplBase &&Arg) : DL(Arg.DL) {}
43 const DataLayout &getDataLayout() const { return DL; }
45 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) {
48 // By default, just classify everything as 'basic'.
49 return TTI::TCC_Basic;
51 case Instruction::GetElementPtr:
52 llvm_unreachable("Use getGEPCost for GEP operations!");
54 case Instruction::BitCast:
55 assert(OpTy && "Cast instructions must provide the operand type");
56 if (Ty == OpTy || (Ty->isPointerTy() && OpTy->isPointerTy()))
57 // Identity and pointer-to-pointer casts are free.
60 // Otherwise, the default basic cost is used.
61 return TTI::TCC_Basic;
63 case Instruction::IntToPtr: {
64 // An inttoptr cast is free so long as the input is a legal integer type
65 // which doesn't contain values outside the range of a pointer.
66 unsigned OpSize = OpTy->getScalarSizeInBits();
67 if (DL.isLegalInteger(OpSize) &&
68 OpSize <= DL.getPointerTypeSizeInBits(Ty))
71 // Otherwise it's not a no-op.
72 return TTI::TCC_Basic;
74 case Instruction::PtrToInt: {
75 // A ptrtoint cast is free so long as the result is large enough to store
76 // the pointer, and a legal integer type.
77 unsigned DestSize = Ty->getScalarSizeInBits();
78 if (DL.isLegalInteger(DestSize) &&
79 DestSize >= DL.getPointerTypeSizeInBits(OpTy))
82 // Otherwise it's not a no-op.
83 return TTI::TCC_Basic;
85 case Instruction::Trunc:
86 // trunc to a native type is free (assuming the target has compare and
87 // shift-right of the same width).
88 if (DL.isLegalInteger(DL.getTypeSizeInBits(Ty)))
91 return TTI::TCC_Basic;
95 unsigned getGEPCost(const Value *Ptr, ArrayRef<const Value *> Operands) {
96 // In the basic model, we just assume that all-constant GEPs will be folded
97 // into their uses via addressing modes.
98 for (unsigned Idx = 0, Size = Operands.size(); Idx != Size; ++Idx)
99 if (!isa<Constant>(Operands[Idx]))
100 return TTI::TCC_Basic;
102 return TTI::TCC_Free;
105 unsigned getCallCost(FunctionType *FTy, int NumArgs) {
106 assert(FTy && "FunctionType must be provided to this routine.");
108 // The target-independent implementation just measures the size of the
109 // function by approximating that each argument will take on average one
110 // instruction to prepare.
113 // Set the argument number to the number of explicit arguments in the
115 NumArgs = FTy->getNumParams();
117 return TTI::TCC_Basic * (NumArgs + 1);
120 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
121 ArrayRef<Type *> ParamTys) {
124 // Intrinsics rarely (if ever) have normal argument setup constraints.
125 // Model them as having a basic instruction cost.
126 // FIXME: This is wrong for libc intrinsics.
127 return TTI::TCC_Basic;
129 case Intrinsic::annotation:
130 case Intrinsic::assume:
131 case Intrinsic::dbg_declare:
132 case Intrinsic::dbg_value:
133 case Intrinsic::invariant_start:
134 case Intrinsic::invariant_end:
135 case Intrinsic::lifetime_start:
136 case Intrinsic::lifetime_end:
137 case Intrinsic::objectsize:
138 case Intrinsic::ptr_annotation:
139 case Intrinsic::var_annotation:
140 case Intrinsic::experimental_gc_result_int:
141 case Intrinsic::experimental_gc_result_float:
142 case Intrinsic::experimental_gc_result_ptr:
143 case Intrinsic::experimental_gc_result:
144 case Intrinsic::experimental_gc_relocate:
145 // These intrinsics don't actually represent code after lowering.
146 return TTI::TCC_Free;
150 bool hasBranchDivergence() { return false; }
152 bool isSourceOfDivergence(const Value *V) { return false; }
154 bool isLoweredToCall(const Function *F) {
155 // FIXME: These should almost certainly not be handled here, and instead
156 // handled with the help of TLI or the target itself. This was largely
157 // ported from existing analysis heuristics here so that such refactorings
158 // can take place in the future.
160 if (F->isIntrinsic())
163 if (F->hasLocalLinkage() || !F->hasName())
166 StringRef Name = F->getName();
168 // These will all likely lower to a single selection DAG node.
169 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
170 Name == "fabs" || Name == "fabsf" || Name == "fabsl" || Name == "sin" ||
171 Name == "fmin" || Name == "fminf" || Name == "fminl" ||
172 Name == "fmax" || Name == "fmaxf" || Name == "fmaxl" ||
173 Name == "sinf" || Name == "sinl" || Name == "cos" || Name == "cosf" ||
174 Name == "cosl" || Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl")
177 // These are all likely to be optimized into something smaller.
178 if (Name == "pow" || Name == "powf" || Name == "powl" || Name == "exp2" ||
179 Name == "exp2l" || Name == "exp2f" || Name == "floor" ||
180 Name == "floorf" || Name == "ceil" || Name == "round" ||
181 Name == "ffs" || Name == "ffsl" || Name == "abs" || Name == "labs" ||
188 void getUnrollingPreferences(Loop *, TTI::UnrollingPreferences &) {}
190 bool isLegalAddImmediate(int64_t Imm) { return false; }
192 bool isLegalICmpImmediate(int64_t Imm) { return false; }
194 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
195 bool HasBaseReg, int64_t Scale,
196 unsigned AddrSpace) {
197 // Guess that only reg and reg+reg addressing is allowed. This heuristic is
198 // taken from the implementation of LSR.
199 return !BaseGV && BaseOffset == 0 && (Scale == 0 || Scale == 1);
202 bool isLegalMaskedStore(Type *DataType, int Consecutive) { return false; }
204 bool isLegalMaskedLoad(Type *DataType, int Consecutive) { return false; }
206 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
207 bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
208 // Guess that all legal addressing mode are free.
209 if (isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
215 bool isTruncateFree(Type *Ty1, Type *Ty2) { return false; }
217 bool isProfitableToHoist(Instruction *I) { return true; }
219 bool isTypeLegal(Type *Ty) { return false; }
221 unsigned getJumpBufAlignment() { return 0; }
223 unsigned getJumpBufSize() { return 0; }
225 bool shouldBuildLookupTables() { return true; }
227 bool enableAggressiveInterleaving(bool LoopHasReductions) { return false; }
229 TTI::PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) {
230 return TTI::PSK_Software;
233 bool haveFastSqrt(Type *Ty) { return false; }
235 unsigned getFPOpCost(Type *Ty) { return TargetTransformInfo::TCC_Basic; }
237 unsigned getIntImmCost(const APInt &Imm, Type *Ty) { return TTI::TCC_Basic; }
239 unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
241 return TTI::TCC_Free;
244 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
246 return TTI::TCC_Free;
249 unsigned getNumberOfRegisters(bool Vector) { return 8; }
251 unsigned getRegisterBitWidth(bool Vector) { return 32; }
253 unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
255 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
256 TTI::OperandValueKind Opd1Info,
257 TTI::OperandValueKind Opd2Info,
258 TTI::OperandValueProperties Opd1PropInfo,
259 TTI::OperandValueProperties Opd2PropInfo) {
263 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Ty, int Index,
268 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) { return 1; }
270 unsigned getCFInstrCost(unsigned Opcode) { return 1; }
272 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy) {
276 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
280 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
281 unsigned AddressSpace) {
285 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
286 unsigned AddressSpace) {
290 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
292 ArrayRef<unsigned> Indices,
294 unsigned AddressSpace) {
298 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
299 ArrayRef<Type *> Tys) {
303 unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
307 unsigned getNumberOfParts(Type *Tp) { return 0; }
309 unsigned getAddressComputationCost(Type *Tp, bool) { return 0; }
311 unsigned getReductionCost(unsigned, Type *, bool) { return 1; }
313 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) { return 0; }
315 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) {
319 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
320 Type *ExpectedType) {
324 bool hasCompatibleFunctionAttributes(const Function *Caller,
325 const Function *Callee) const {
326 return (Caller->getFnAttribute("target-cpu") ==
327 Callee->getFnAttribute("target-cpu")) &&
328 (Caller->getFnAttribute("target-features") ==
329 Callee->getFnAttribute("target-features"));
333 /// \brief CRTP base class for use as a mix-in that aids implementing
334 /// a TargetTransformInfo-compatible class.
335 template <typename T>
336 class TargetTransformInfoImplCRTPBase : public TargetTransformInfoImplBase {
338 typedef TargetTransformInfoImplBase BaseT;
341 explicit TargetTransformInfoImplCRTPBase(const DataLayout &DL) : BaseT(DL) {}
344 // Provide value semantics. MSVC requires that we spell all of these out.
345 TargetTransformInfoImplCRTPBase(const TargetTransformInfoImplCRTPBase &Arg)
346 : BaseT(static_cast<const BaseT &>(Arg)) {}
347 TargetTransformInfoImplCRTPBase(TargetTransformInfoImplCRTPBase &&Arg)
348 : BaseT(std::move(static_cast<BaseT &>(Arg))) {}
350 using BaseT::getCallCost;
352 unsigned getCallCost(const Function *F, int NumArgs) {
353 assert(F && "A concrete function must be provided to this routine.");
356 // Set the argument number to the number of explicit arguments in the
358 NumArgs = F->arg_size();
360 if (Intrinsic::ID IID = F->getIntrinsicID()) {
361 FunctionType *FTy = F->getFunctionType();
362 SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end());
363 return static_cast<T *>(this)
364 ->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys);
367 if (!static_cast<T *>(this)->isLoweredToCall(F))
368 return TTI::TCC_Basic; // Give a basic cost if it will be lowered
371 return static_cast<T *>(this)->getCallCost(F->getFunctionType(), NumArgs);
374 unsigned getCallCost(const Function *F, ArrayRef<const Value *> Arguments) {
375 // Simply delegate to generic handling of the call.
376 // FIXME: We should use instsimplify or something else to catch calls which
377 // will constant fold with these arguments.
378 return static_cast<T *>(this)->getCallCost(F, Arguments.size());
381 using BaseT::getIntrinsicCost;
383 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
384 ArrayRef<const Value *> Arguments) {
385 // Delegate to the generic intrinsic handling code. This mostly provides an
386 // opportunity for targets to (for example) special case the cost of
387 // certain intrinsics based on constants used as arguments.
388 SmallVector<Type *, 8> ParamTys;
389 ParamTys.reserve(Arguments.size());
390 for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx)
391 ParamTys.push_back(Arguments[Idx]->getType());
392 return static_cast<T *>(this)->getIntrinsicCost(IID, RetTy, ParamTys);
395 unsigned getUserCost(const User *U) {
397 return TTI::TCC_Free; // Model all PHI nodes as free.
399 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U)) {
400 SmallVector<const Value *, 4> Indices(GEP->idx_begin(), GEP->idx_end());
401 return static_cast<T *>(this)
402 ->getGEPCost(GEP->getPointerOperand(), Indices);
405 if (auto CS = ImmutableCallSite(U)) {
406 const Function *F = CS.getCalledFunction();
408 // Just use the called value type.
409 Type *FTy = CS.getCalledValue()->getType()->getPointerElementType();
410 return static_cast<T *>(this)
411 ->getCallCost(cast<FunctionType>(FTy), CS.arg_size());
414 SmallVector<const Value *, 8> Arguments(CS.arg_begin(), CS.arg_end());
415 return static_cast<T *>(this)->getCallCost(F, Arguments);
418 if (const CastInst *CI = dyn_cast<CastInst>(U)) {
419 // Result of a cmp instruction is often extended (to be used by other
420 // cmp instructions, logical or return instructions). These are usually
421 // nop on most sane targets.
422 if (isa<CmpInst>(CI->getOperand(0)))
423 return TTI::TCC_Free;
426 return static_cast<T *>(this)->getOperationCost(
427 Operator::getOpcode(U), U->getType(),
428 U->getNumOperands() == 1 ? U->getOperand(0)->getType() : nullptr);