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::FDiv:
65 case Instruction::FRem:
66 case Instruction::SDiv:
67 case Instruction::SRem:
68 case Instruction::UDiv:
69 case Instruction::URem:
70 return TTI::TCC_Expensive;
72 case Instruction::IntToPtr: {
73 // An inttoptr cast is free so long as the input is a legal integer type
74 // which doesn't contain values outside the range of a pointer.
75 unsigned OpSize = OpTy->getScalarSizeInBits();
76 if (DL.isLegalInteger(OpSize) &&
77 OpSize <= DL.getPointerTypeSizeInBits(Ty))
80 // Otherwise it's not a no-op.
81 return TTI::TCC_Basic;
83 case Instruction::PtrToInt: {
84 // A ptrtoint cast is free so long as the result is large enough to store
85 // the pointer, and a legal integer type.
86 unsigned DestSize = Ty->getScalarSizeInBits();
87 if (DL.isLegalInteger(DestSize) &&
88 DestSize >= DL.getPointerTypeSizeInBits(OpTy))
91 // Otherwise it's not a no-op.
92 return TTI::TCC_Basic;
94 case Instruction::Trunc:
95 // trunc to a native type is free (assuming the target has compare and
96 // shift-right of the same width).
97 if (DL.isLegalInteger(DL.getTypeSizeInBits(Ty)))
100 return TTI::TCC_Basic;
104 unsigned getGEPCost(Type *PointeeType, const Value *Ptr,
105 ArrayRef<const Value *> Operands) {
106 // In the basic model, we just assume that all-constant GEPs will be folded
107 // into their uses via addressing modes.
108 for (unsigned Idx = 0, Size = Operands.size(); Idx != Size; ++Idx)
109 if (!isa<Constant>(Operands[Idx]))
110 return TTI::TCC_Basic;
112 return TTI::TCC_Free;
115 unsigned getCallCost(FunctionType *FTy, int NumArgs) {
116 assert(FTy && "FunctionType must be provided to this routine.");
118 // The target-independent implementation just measures the size of the
119 // function by approximating that each argument will take on average one
120 // instruction to prepare.
123 // Set the argument number to the number of explicit arguments in the
125 NumArgs = FTy->getNumParams();
127 return TTI::TCC_Basic * (NumArgs + 1);
130 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
131 ArrayRef<Type *> ParamTys) {
134 // Intrinsics rarely (if ever) have normal argument setup constraints.
135 // Model them as having a basic instruction cost.
136 // FIXME: This is wrong for libc intrinsics.
137 return TTI::TCC_Basic;
139 case Intrinsic::annotation:
140 case Intrinsic::assume:
141 case Intrinsic::dbg_declare:
142 case Intrinsic::dbg_value:
143 case Intrinsic::invariant_start:
144 case Intrinsic::invariant_end:
145 case Intrinsic::lifetime_start:
146 case Intrinsic::lifetime_end:
147 case Intrinsic::objectsize:
148 case Intrinsic::ptr_annotation:
149 case Intrinsic::var_annotation:
150 case Intrinsic::experimental_gc_result_int:
151 case Intrinsic::experimental_gc_result_float:
152 case Intrinsic::experimental_gc_result_ptr:
153 case Intrinsic::experimental_gc_result:
154 case Intrinsic::experimental_gc_relocate:
155 // These intrinsics don't actually represent code after lowering.
156 return TTI::TCC_Free;
160 bool hasBranchDivergence() { return false; }
162 bool isSourceOfDivergence(const Value *V) { return false; }
164 bool isLoweredToCall(const Function *F) {
165 // FIXME: These should almost certainly not be handled here, and instead
166 // handled with the help of TLI or the target itself. This was largely
167 // ported from existing analysis heuristics here so that such refactorings
168 // can take place in the future.
170 if (F->isIntrinsic())
173 if (F->hasLocalLinkage() || !F->hasName())
176 StringRef Name = F->getName();
178 // These will all likely lower to a single selection DAG node.
179 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
180 Name == "fabs" || Name == "fabsf" || Name == "fabsl" || Name == "sin" ||
181 Name == "fmin" || Name == "fminf" || Name == "fminl" ||
182 Name == "fmax" || Name == "fmaxf" || Name == "fmaxl" ||
183 Name == "sinf" || Name == "sinl" || Name == "cos" || Name == "cosf" ||
184 Name == "cosl" || Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl")
187 // These are all likely to be optimized into something smaller.
188 if (Name == "pow" || Name == "powf" || Name == "powl" || Name == "exp2" ||
189 Name == "exp2l" || Name == "exp2f" || Name == "floor" ||
190 Name == "floorf" || Name == "ceil" || Name == "round" ||
191 Name == "ffs" || Name == "ffsl" || Name == "abs" || Name == "labs" ||
198 void getUnrollingPreferences(Loop *, TTI::UnrollingPreferences &) {}
200 bool isLegalAddImmediate(int64_t Imm) { return false; }
202 bool isLegalICmpImmediate(int64_t Imm) { return false; }
204 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
205 bool HasBaseReg, int64_t Scale,
206 unsigned AddrSpace) {
207 // Guess that only reg and reg+reg addressing is allowed. This heuristic is
208 // taken from the implementation of LSR.
209 return !BaseGV && BaseOffset == 0 && (Scale == 0 || Scale == 1);
212 bool isLegalMaskedStore(Type *DataType) { return false; }
214 bool isLegalMaskedLoad(Type *DataType) { return false; }
216 bool isLegalMaskedScatter(Type *DataType) { return false; }
218 bool isLegalMaskedGather(Type *DataType) { return false; }
220 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
221 bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
222 // Guess that all legal addressing mode are free.
223 if (isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
229 bool isTruncateFree(Type *Ty1, Type *Ty2) { return false; }
231 bool isProfitableToHoist(Instruction *I) { return true; }
233 bool isTypeLegal(Type *Ty) { return false; }
235 unsigned getJumpBufAlignment() { return 0; }
237 unsigned getJumpBufSize() { return 0; }
239 bool shouldBuildLookupTables() { return true; }
241 bool enableAggressiveInterleaving(bool LoopHasReductions) { return false; }
243 bool enableInterleavedAccessVectorization() { return false; }
245 TTI::PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) {
246 return TTI::PSK_Software;
249 bool haveFastSqrt(Type *Ty) { return false; }
251 unsigned getFPOpCost(Type *Ty) { return TargetTransformInfo::TCC_Basic; }
253 unsigned getIntImmCost(const APInt &Imm, Type *Ty) { return TTI::TCC_Basic; }
255 unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
257 return TTI::TCC_Free;
260 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
262 return TTI::TCC_Free;
265 unsigned getNumberOfRegisters(bool Vector) { return 8; }
267 unsigned getRegisterBitWidth(bool Vector) { return 32; }
269 unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
271 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
272 TTI::OperandValueKind Opd1Info,
273 TTI::OperandValueKind Opd2Info,
274 TTI::OperandValueProperties Opd1PropInfo,
275 TTI::OperandValueProperties Opd2PropInfo) {
279 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Ty, int Index,
284 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) { return 1; }
286 unsigned getCFInstrCost(unsigned Opcode) { return 1; }
288 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy) {
292 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
296 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
297 unsigned AddressSpace) {
301 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
302 unsigned AddressSpace) {
306 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
308 ArrayRef<unsigned> Indices,
310 unsigned AddressSpace) {
314 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
315 ArrayRef<Type *> Tys) {
319 unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
323 unsigned getNumberOfParts(Type *Tp) { return 0; }
325 unsigned getAddressComputationCost(Type *Tp, bool) { return 0; }
327 unsigned getReductionCost(unsigned, Type *, bool) { return 1; }
329 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) { return 0; }
331 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) {
335 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
336 Type *ExpectedType) {
340 bool areInlineCompatible(const Function *Caller,
341 const Function *Callee) const {
342 return (Caller->getFnAttribute("target-cpu") ==
343 Callee->getFnAttribute("target-cpu")) &&
344 (Caller->getFnAttribute("target-features") ==
345 Callee->getFnAttribute("target-features"));
349 /// \brief CRTP base class for use as a mix-in that aids implementing
350 /// a TargetTransformInfo-compatible class.
351 template <typename T>
352 class TargetTransformInfoImplCRTPBase : public TargetTransformInfoImplBase {
354 typedef TargetTransformInfoImplBase BaseT;
357 explicit TargetTransformInfoImplCRTPBase(const DataLayout &DL) : BaseT(DL) {}
360 // Provide value semantics. MSVC requires that we spell all of these out.
361 TargetTransformInfoImplCRTPBase(const TargetTransformInfoImplCRTPBase &Arg)
362 : BaseT(static_cast<const BaseT &>(Arg)) {}
363 TargetTransformInfoImplCRTPBase(TargetTransformInfoImplCRTPBase &&Arg)
364 : BaseT(std::move(static_cast<BaseT &>(Arg))) {}
366 using BaseT::getCallCost;
368 unsigned getCallCost(const Function *F, int NumArgs) {
369 assert(F && "A concrete function must be provided to this routine.");
372 // Set the argument number to the number of explicit arguments in the
374 NumArgs = F->arg_size();
376 if (Intrinsic::ID IID = F->getIntrinsicID()) {
377 FunctionType *FTy = F->getFunctionType();
378 SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end());
379 return static_cast<T *>(this)
380 ->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys);
383 if (!static_cast<T *>(this)->isLoweredToCall(F))
384 return TTI::TCC_Basic; // Give a basic cost if it will be lowered
387 return static_cast<T *>(this)->getCallCost(F->getFunctionType(), NumArgs);
390 unsigned getCallCost(const Function *F, ArrayRef<const Value *> Arguments) {
391 // Simply delegate to generic handling of the call.
392 // FIXME: We should use instsimplify or something else to catch calls which
393 // will constant fold with these arguments.
394 return static_cast<T *>(this)->getCallCost(F, Arguments.size());
397 using BaseT::getGEPCost;
399 unsigned getGEPCost(Type *PointeeType, const Value *Ptr,
400 ArrayRef<const Value *> Operands) {
401 const GlobalValue *BaseGV = nullptr;
402 if (Ptr != nullptr) {
403 // TODO: will remove this when pointers have an opaque type.
404 assert(Ptr->getType()->getScalarType()->getPointerElementType() ==
406 "explicit pointee type doesn't match operand's pointee type");
407 BaseGV = dyn_cast<GlobalValue>(Ptr->stripPointerCasts());
409 bool HasBaseReg = (BaseGV == nullptr);
410 int64_t BaseOffset = 0;
413 // Assumes the address space is 0 when Ptr is nullptr.
415 (Ptr == nullptr ? 0 : Ptr->getType()->getPointerAddressSpace());
416 auto GTI = gep_type_begin(PointerType::get(PointeeType, AS), Operands);
417 for (auto I = Operands.begin(); I != Operands.end(); ++I, ++GTI) {
418 if (isa<SequentialType>(*GTI)) {
419 int64_t ElementSize = DL.getTypeAllocSize(GTI.getIndexedType());
420 if (const ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I)) {
421 BaseOffset += ConstIdx->getSExtValue() * ElementSize;
423 // Needs scale register.
425 // No addressing mode takes two scale registers.
426 return TTI::TCC_Basic;
431 StructType *STy = cast<StructType>(*GTI);
432 uint64_t Field = cast<ConstantInt>(*I)->getZExtValue();
433 BaseOffset += DL.getStructLayout(STy)->getElementOffset(Field);
437 if (static_cast<T *>(this)->isLegalAddressingMode(
438 PointerType::get(*GTI, AS), const_cast<GlobalValue *>(BaseGV),
439 BaseOffset, HasBaseReg, Scale, AS)) {
440 return TTI::TCC_Free;
442 return TTI::TCC_Basic;
445 using BaseT::getIntrinsicCost;
447 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
448 ArrayRef<const Value *> Arguments) {
449 // Delegate to the generic intrinsic handling code. This mostly provides an
450 // opportunity for targets to (for example) special case the cost of
451 // certain intrinsics based on constants used as arguments.
452 SmallVector<Type *, 8> ParamTys;
453 ParamTys.reserve(Arguments.size());
454 for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx)
455 ParamTys.push_back(Arguments[Idx]->getType());
456 return static_cast<T *>(this)->getIntrinsicCost(IID, RetTy, ParamTys);
459 unsigned getUserCost(const User *U) {
461 return TTI::TCC_Free; // Model all PHI nodes as free.
463 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U)) {
464 SmallVector<Value *, 4> Indices(GEP->idx_begin(), GEP->idx_end());
465 return static_cast<T *>(this)->getGEPCost(
466 GEP->getSourceElementType(), GEP->getPointerOperand(), Indices);
469 if (auto CS = ImmutableCallSite(U)) {
470 const Function *F = CS.getCalledFunction();
472 // Just use the called value type.
473 Type *FTy = CS.getCalledValue()->getType()->getPointerElementType();
474 return static_cast<T *>(this)
475 ->getCallCost(cast<FunctionType>(FTy), CS.arg_size());
478 SmallVector<const Value *, 8> Arguments(CS.arg_begin(), CS.arg_end());
479 return static_cast<T *>(this)->getCallCost(F, Arguments);
482 if (const CastInst *CI = dyn_cast<CastInst>(U)) {
483 // Result of a cmp instruction is often extended (to be used by other
484 // cmp instructions, logical or return instructions). These are usually
485 // nop on most sane targets.
486 if (isa<CmpInst>(CI->getOperand(0)))
487 return TTI::TCC_Free;
490 return static_cast<T *>(this)->getOperationCost(
491 Operator::getOpcode(U), U->getType(),
492 U->getNumOperands() == 1 ? U->getOperand(0)->getType() : nullptr);