1 //===- llvm/Analysis/TargetTransformInfo.cpp ------------------------------===//
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 #include "llvm/Analysis/TargetTransformInfo.h"
11 #include "llvm/IR/CallSite.h"
12 #include "llvm/IR/DataLayout.h"
13 #include "llvm/IR/Instruction.h"
14 #include "llvm/IR/Instructions.h"
15 #include "llvm/IR/IntrinsicInst.h"
16 #include "llvm/IR/Operator.h"
17 #include "llvm/Support/ErrorHandling.h"
21 #define DEBUG_TYPE "tti"
23 // Setup the analysis group to manage the TargetTransformInfo passes.
24 INITIALIZE_ANALYSIS_GROUP(TargetTransformInfo, "Target Information", NoTTI)
25 char TargetTransformInfo::ID = 0;
27 TargetTransformInfo::~TargetTransformInfo() {
30 void TargetTransformInfo::pushTTIStack(Pass *P) {
32 PrevTTI = &P->getAnalysis<TargetTransformInfo>();
34 // Walk up the chain and update the top TTI pointer.
35 for (TargetTransformInfo *PTTI = PrevTTI; PTTI; PTTI = PTTI->PrevTTI)
39 void TargetTransformInfo::getAnalysisUsage(AnalysisUsage &AU) const {
40 AU.addRequired<TargetTransformInfo>();
43 unsigned TargetTransformInfo::getOperationCost(unsigned Opcode, Type *Ty,
45 return PrevTTI->getOperationCost(Opcode, Ty, OpTy);
48 unsigned TargetTransformInfo::getGEPCost(
49 const Value *Ptr, ArrayRef<const Value *> Operands) const {
50 return PrevTTI->getGEPCost(Ptr, Operands);
53 unsigned TargetTransformInfo::getCallCost(FunctionType *FTy,
55 return PrevTTI->getCallCost(FTy, NumArgs);
58 unsigned TargetTransformInfo::getCallCost(const Function *F,
60 return PrevTTI->getCallCost(F, NumArgs);
63 unsigned TargetTransformInfo::getCallCost(
64 const Function *F, ArrayRef<const Value *> Arguments) const {
65 return PrevTTI->getCallCost(F, Arguments);
68 unsigned TargetTransformInfo::getIntrinsicCost(
69 Intrinsic::ID IID, Type *RetTy, ArrayRef<Type *> ParamTys) const {
70 return PrevTTI->getIntrinsicCost(IID, RetTy, ParamTys);
73 unsigned TargetTransformInfo::getIntrinsicCost(
74 Intrinsic::ID IID, Type *RetTy, ArrayRef<const Value *> Arguments) const {
75 return PrevTTI->getIntrinsicCost(IID, RetTy, Arguments);
78 unsigned TargetTransformInfo::getUserCost(const User *U) const {
79 return PrevTTI->getUserCost(U);
82 bool TargetTransformInfo::hasBranchDivergence() const {
83 return PrevTTI->hasBranchDivergence();
86 bool TargetTransformInfo::isLoweredToCall(const Function *F) const {
87 return PrevTTI->isLoweredToCall(F);
91 TargetTransformInfo::getUnrollingPreferences(const Function *F, Loop *L,
92 UnrollingPreferences &UP) const {
93 PrevTTI->getUnrollingPreferences(F, L, UP);
96 bool TargetTransformInfo::isLegalAddImmediate(int64_t Imm) const {
97 return PrevTTI->isLegalAddImmediate(Imm);
100 bool TargetTransformInfo::isLegalICmpImmediate(int64_t Imm) const {
101 return PrevTTI->isLegalICmpImmediate(Imm);
104 bool TargetTransformInfo::isLegalMaskedLoad(Type *DataType,
105 int Consecutive) const {
109 bool TargetTransformInfo::isLegalMaskedStore(Type *DataType,
110 int Consecutive) const {
115 bool TargetTransformInfo::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
118 int64_t Scale) const {
119 return PrevTTI->isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
123 int TargetTransformInfo::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
126 int64_t Scale) const {
127 return PrevTTI->getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg,
131 bool TargetTransformInfo::isTruncateFree(Type *Ty1, Type *Ty2) const {
132 return PrevTTI->isTruncateFree(Ty1, Ty2);
135 bool TargetTransformInfo::isTypeLegal(Type *Ty) const {
136 return PrevTTI->isTypeLegal(Ty);
139 unsigned TargetTransformInfo::getJumpBufAlignment() const {
140 return PrevTTI->getJumpBufAlignment();
143 unsigned TargetTransformInfo::getJumpBufSize() const {
144 return PrevTTI->getJumpBufSize();
147 bool TargetTransformInfo::shouldBuildLookupTables() const {
148 return PrevTTI->shouldBuildLookupTables();
151 TargetTransformInfo::PopcntSupportKind
152 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit) const {
153 return PrevTTI->getPopcntSupport(IntTyWidthInBit);
156 bool TargetTransformInfo::haveFastSqrt(Type *Ty) const {
157 return PrevTTI->haveFastSqrt(Ty);
160 unsigned TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty) const {
161 return PrevTTI->getIntImmCost(Imm, Ty);
164 unsigned TargetTransformInfo::getIntImmCost(unsigned Opc, unsigned Idx,
165 const APInt &Imm, Type *Ty) const {
166 return PrevTTI->getIntImmCost(Opc, Idx, Imm, Ty);
169 unsigned TargetTransformInfo::getIntImmCost(Intrinsic::ID IID, unsigned Idx,
170 const APInt &Imm, Type *Ty) const {
171 return PrevTTI->getIntImmCost(IID, Idx, Imm, Ty);
174 unsigned TargetTransformInfo::getNumberOfRegisters(bool Vector) const {
175 return PrevTTI->getNumberOfRegisters(Vector);
178 unsigned TargetTransformInfo::getRegisterBitWidth(bool Vector) const {
179 return PrevTTI->getRegisterBitWidth(Vector);
182 unsigned TargetTransformInfo::getMaxInterleaveFactor() const {
183 return PrevTTI->getMaxInterleaveFactor();
186 unsigned TargetTransformInfo::getArithmeticInstrCost(
187 unsigned Opcode, Type *Ty, OperandValueKind Op1Info,
188 OperandValueKind Op2Info, OperandValueProperties Opd1PropInfo,
189 OperandValueProperties Opd2PropInfo) const {
190 return PrevTTI->getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info,
191 Opd1PropInfo, Opd2PropInfo);
194 unsigned TargetTransformInfo::getShuffleCost(ShuffleKind Kind, Type *Tp,
195 int Index, Type *SubTp) const {
196 return PrevTTI->getShuffleCost(Kind, Tp, Index, SubTp);
199 unsigned TargetTransformInfo::getCastInstrCost(unsigned Opcode, Type *Dst,
201 return PrevTTI->getCastInstrCost(Opcode, Dst, Src);
204 unsigned TargetTransformInfo::getCFInstrCost(unsigned Opcode) const {
205 return PrevTTI->getCFInstrCost(Opcode);
208 unsigned TargetTransformInfo::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
209 Type *CondTy) const {
210 return PrevTTI->getCmpSelInstrCost(Opcode, ValTy, CondTy);
213 unsigned TargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val,
214 unsigned Index) const {
215 return PrevTTI->getVectorInstrCost(Opcode, Val, Index);
218 unsigned TargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src,
220 unsigned AddressSpace) const {
221 return PrevTTI->getMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
225 TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID,
227 ArrayRef<Type *> Tys) const {
228 return PrevTTI->getIntrinsicInstrCost(ID, RetTy, Tys);
231 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const {
232 return PrevTTI->getNumberOfParts(Tp);
235 unsigned TargetTransformInfo::getAddressComputationCost(Type *Tp,
236 bool IsComplex) const {
237 return PrevTTI->getAddressComputationCost(Tp, IsComplex);
240 unsigned TargetTransformInfo::getReductionCost(unsigned Opcode, Type *Ty,
241 bool IsPairwise) const {
242 return PrevTTI->getReductionCost(Opcode, Ty, IsPairwise);
245 unsigned TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef<Type*> Tys)
247 return PrevTTI->getCostOfKeepingLiveOverCall(Tys);
252 struct NoTTI final : ImmutablePass, TargetTransformInfo {
253 const DataLayout *DL;
255 NoTTI() : ImmutablePass(ID), DL(nullptr) {
256 initializeNoTTIPass(*PassRegistry::getPassRegistry());
259 void initializePass() override {
260 // Note that this subclass is special, and must *not* call initializeTTI as
261 // it does not chain.
264 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
265 DL = DLP ? &DLP->getDataLayout() : nullptr;
268 void getAnalysisUsage(AnalysisUsage &AU) const override {
269 // Note that this subclass is special, and must *not* call
270 // TTI::getAnalysisUsage as it breaks the recursion.
273 /// Pass identification.
276 /// Provide necessary pointer adjustments for the two base classes.
277 void *getAdjustedAnalysisPointer(const void *ID) override {
278 if (ID == &TargetTransformInfo::ID)
279 return (TargetTransformInfo*)this;
283 unsigned getOperationCost(unsigned Opcode, Type *Ty,
284 Type *OpTy) const override {
287 // By default, just classify everything as 'basic'.
290 case Instruction::GetElementPtr:
291 llvm_unreachable("Use getGEPCost for GEP operations!");
293 case Instruction::BitCast:
294 assert(OpTy && "Cast instructions must provide the operand type");
295 if (Ty == OpTy || (Ty->isPointerTy() && OpTy->isPointerTy()))
296 // Identity and pointer-to-pointer casts are free.
299 // Otherwise, the default basic cost is used.
302 case Instruction::IntToPtr: {
306 // An inttoptr cast is free so long as the input is a legal integer type
307 // which doesn't contain values outside the range of a pointer.
308 unsigned OpSize = OpTy->getScalarSizeInBits();
309 if (DL->isLegalInteger(OpSize) &&
310 OpSize <= DL->getPointerTypeSizeInBits(Ty))
313 // Otherwise it's not a no-op.
316 case Instruction::PtrToInt: {
320 // A ptrtoint cast is free so long as the result is large enough to store
321 // the pointer, and a legal integer type.
322 unsigned DestSize = Ty->getScalarSizeInBits();
323 if (DL->isLegalInteger(DestSize) &&
324 DestSize >= DL->getPointerTypeSizeInBits(OpTy))
327 // Otherwise it's not a no-op.
330 case Instruction::Trunc:
331 // trunc to a native type is free (assuming the target has compare and
332 // shift-right of the same width).
333 if (DL && DL->isLegalInteger(DL->getTypeSizeInBits(Ty)))
340 unsigned getGEPCost(const Value *Ptr,
341 ArrayRef<const Value *> Operands) const override {
342 // In the basic model, we just assume that all-constant GEPs will be folded
343 // into their uses via addressing modes.
344 for (unsigned Idx = 0, Size = Operands.size(); Idx != Size; ++Idx)
345 if (!isa<Constant>(Operands[Idx]))
351 unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const override
353 assert(FTy && "FunctionType must be provided to this routine.");
355 // The target-independent implementation just measures the size of the
356 // function by approximating that each argument will take on average one
357 // instruction to prepare.
360 // Set the argument number to the number of explicit arguments in the
362 NumArgs = FTy->getNumParams();
364 return TCC_Basic * (NumArgs + 1);
367 unsigned getCallCost(const Function *F, int NumArgs = -1) const override
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 = (Intrinsic::ID)F->getIntrinsicID()) {
377 FunctionType *FTy = F->getFunctionType();
378 SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end());
379 return TopTTI->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys);
382 if (!TopTTI->isLoweredToCall(F))
383 return TCC_Basic; // Give a basic cost if it will be lowered directly.
385 return TopTTI->getCallCost(F->getFunctionType(), NumArgs);
388 unsigned getCallCost(const Function *F,
389 ArrayRef<const Value *> Arguments) const override {
390 // Simply delegate to generic handling of the call.
391 // FIXME: We should use instsimplify or something else to catch calls which
392 // will constant fold with these arguments.
393 return TopTTI->getCallCost(F, Arguments.size());
396 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
397 ArrayRef<Type *> ParamTys) const override {
400 // Intrinsics rarely (if ever) have normal argument setup constraints.
401 // Model them as having a basic instruction cost.
402 // FIXME: This is wrong for libc intrinsics.
405 case Intrinsic::annotation:
406 case Intrinsic::assume:
407 case Intrinsic::dbg_declare:
408 case Intrinsic::dbg_value:
409 case Intrinsic::invariant_start:
410 case Intrinsic::invariant_end:
411 case Intrinsic::lifetime_start:
412 case Intrinsic::lifetime_end:
413 case Intrinsic::objectsize:
414 case Intrinsic::ptr_annotation:
415 case Intrinsic::var_annotation:
416 case Intrinsic::experimental_gc_result_int:
417 case Intrinsic::experimental_gc_result_float:
418 case Intrinsic::experimental_gc_result_ptr:
419 case Intrinsic::experimental_gc_result:
420 case Intrinsic::experimental_gc_relocate:
421 // These intrinsics don't actually represent code after lowering.
427 getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
428 ArrayRef<const Value *> Arguments) const override {
429 // Delegate to the generic intrinsic handling code. This mostly provides an
430 // opportunity for targets to (for example) special case the cost of
431 // certain intrinsics based on constants used as arguments.
432 SmallVector<Type *, 8> ParamTys;
433 ParamTys.reserve(Arguments.size());
434 for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx)
435 ParamTys.push_back(Arguments[Idx]->getType());
436 return TopTTI->getIntrinsicCost(IID, RetTy, ParamTys);
439 unsigned getUserCost(const User *U) const override {
441 return TCC_Free; // Model all PHI nodes as free.
443 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U)) {
444 SmallVector<const Value *, 4> Indices(GEP->idx_begin(), GEP->idx_end());
445 return TopTTI->getGEPCost(GEP->getPointerOperand(), Indices);
448 if (ImmutableCallSite CS = U) {
449 const Function *F = CS.getCalledFunction();
451 // Just use the called value type.
452 Type *FTy = CS.getCalledValue()->getType()->getPointerElementType();
453 return TopTTI->getCallCost(cast<FunctionType>(FTy), CS.arg_size());
456 SmallVector<const Value *, 8> Arguments(CS.arg_begin(), CS.arg_end());
457 return TopTTI->getCallCost(F, Arguments);
460 if (const CastInst *CI = dyn_cast<CastInst>(U)) {
461 // Result of a cmp instruction is often extended (to be used by other
462 // cmp instructions, logical or return instructions). These are usually
463 // nop on most sane targets.
464 if (isa<CmpInst>(CI->getOperand(0)))
468 // Otherwise delegate to the fully generic implementations.
469 return getOperationCost(Operator::getOpcode(U), U->getType(),
470 U->getNumOperands() == 1 ?
471 U->getOperand(0)->getType() : nullptr);
474 bool hasBranchDivergence() const override { return false; }
476 bool isLoweredToCall(const Function *F) const override {
477 // FIXME: These should almost certainly not be handled here, and instead
478 // handled with the help of TLI or the target itself. This was largely
479 // ported from existing analysis heuristics here so that such refactorings
480 // can take place in the future.
482 if (F->isIntrinsic())
485 if (F->hasLocalLinkage() || !F->hasName())
488 StringRef Name = F->getName();
490 // These will all likely lower to a single selection DAG node.
491 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
492 Name == "fabs" || Name == "fabsf" || Name == "fabsl" || Name == "sin" ||
493 Name == "fmin" || Name == "fminf" || Name == "fminl" ||
494 Name == "fmax" || Name == "fmaxf" || Name == "fmaxl" ||
495 Name == "sinf" || Name == "sinl" || Name == "cos" || Name == "cosf" ||
496 Name == "cosl" || Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl")
499 // These are all likely to be optimized into something smaller.
500 if (Name == "pow" || Name == "powf" || Name == "powl" || Name == "exp2" ||
501 Name == "exp2l" || Name == "exp2f" || Name == "floor" || Name ==
502 "floorf" || Name == "ceil" || Name == "round" || Name == "ffs" ||
503 Name == "ffsl" || Name == "abs" || Name == "labs" || Name == "llabs")
509 void getUnrollingPreferences(const Function *, Loop *,
510 UnrollingPreferences &) const override {}
512 bool isLegalAddImmediate(int64_t Imm) const override {
516 bool isLegalICmpImmediate(int64_t Imm) const override {
520 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
521 bool HasBaseReg, int64_t Scale) const override
523 // Guess that reg+reg addressing is allowed. This heuristic is taken from
524 // the implementation of LSR.
525 return !BaseGV && BaseOffset == 0 && Scale <= 1;
528 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
529 bool HasBaseReg, int64_t Scale) const override {
530 // Guess that all legal addressing mode are free.
531 if(isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg, Scale))
536 bool isTruncateFree(Type *Ty1, Type *Ty2) const override {
540 bool isTypeLegal(Type *Ty) const override {
544 unsigned getJumpBufAlignment() const override {
548 unsigned getJumpBufSize() const override {
552 bool shouldBuildLookupTables() const override {
557 getPopcntSupport(unsigned IntTyWidthInBit) const override {
561 bool haveFastSqrt(Type *Ty) const override {
565 unsigned getIntImmCost(const APInt &Imm, Type *Ty) const override {
569 unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
570 Type *Ty) const override {
574 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
575 Type *Ty) const override {
579 unsigned getNumberOfRegisters(bool Vector) const override {
583 unsigned getRegisterBitWidth(bool Vector) const override {
587 unsigned getMaxInterleaveFactor() const override {
591 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind,
592 OperandValueKind, OperandValueProperties,
593 OperandValueProperties) const override {
597 unsigned getShuffleCost(ShuffleKind Kind, Type *Ty,
598 int Index = 0, Type *SubTp = nullptr) const override {
602 unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
603 Type *Src) const override {
607 unsigned getCFInstrCost(unsigned Opcode) const override {
611 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
612 Type *CondTy = nullptr) const override {
616 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
617 unsigned Index = -1) const override {
621 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
622 unsigned AddressSpace) const override {
626 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
627 ArrayRef<Type*> Tys) const override {
631 unsigned getNumberOfParts(Type *Tp) const override {
635 unsigned getAddressComputationCost(Type *Tp, bool) const override {
639 unsigned getReductionCost(unsigned, Type *, bool) const override {
643 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type*> Tys) const override {
649 } // end anonymous namespace
651 INITIALIZE_AG_PASS(NoTTI, TargetTransformInfo, "notti",
652 "No target information", true, true, true)
655 ImmutablePass *llvm::createNoTargetTransformInfoPass() {