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);
226 TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID,
228 ArrayRef<Type *> Tys) const {
229 return PrevTTI->getIntrinsicInstrCost(ID, RetTy, Tys);
232 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const {
233 return PrevTTI->getNumberOfParts(Tp);
236 unsigned TargetTransformInfo::getAddressComputationCost(Type *Tp,
237 bool IsComplex) const {
238 return PrevTTI->getAddressComputationCost(Tp, IsComplex);
241 unsigned TargetTransformInfo::getReductionCost(unsigned Opcode, Type *Ty,
242 bool IsPairwise) const {
243 return PrevTTI->getReductionCost(Opcode, Ty, IsPairwise);
246 unsigned TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef<Type*> Tys)
248 return PrevTTI->getCostOfKeepingLiveOverCall(Tys);
253 struct NoTTI final : ImmutablePass, TargetTransformInfo {
254 const DataLayout *DL;
256 NoTTI() : ImmutablePass(ID), DL(nullptr) {
257 initializeNoTTIPass(*PassRegistry::getPassRegistry());
260 void initializePass() override {
261 // Note that this subclass is special, and must *not* call initializeTTI as
262 // it does not chain.
265 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
266 DL = DLP ? &DLP->getDataLayout() : nullptr;
269 void getAnalysisUsage(AnalysisUsage &AU) const override {
270 // Note that this subclass is special, and must *not* call
271 // TTI::getAnalysisUsage as it breaks the recursion.
274 /// Pass identification.
277 /// Provide necessary pointer adjustments for the two base classes.
278 void *getAdjustedAnalysisPointer(const void *ID) override {
279 if (ID == &TargetTransformInfo::ID)
280 return (TargetTransformInfo*)this;
284 unsigned getOperationCost(unsigned Opcode, Type *Ty,
285 Type *OpTy) const override {
288 // By default, just classify everything as 'basic'.
291 case Instruction::GetElementPtr:
292 llvm_unreachable("Use getGEPCost for GEP operations!");
294 case Instruction::BitCast:
295 assert(OpTy && "Cast instructions must provide the operand type");
296 if (Ty == OpTy || (Ty->isPointerTy() && OpTy->isPointerTy()))
297 // Identity and pointer-to-pointer casts are free.
300 // Otherwise, the default basic cost is used.
303 case Instruction::IntToPtr: {
307 // An inttoptr cast is free so long as the input is a legal integer type
308 // which doesn't contain values outside the range of a pointer.
309 unsigned OpSize = OpTy->getScalarSizeInBits();
310 if (DL->isLegalInteger(OpSize) &&
311 OpSize <= DL->getPointerTypeSizeInBits(Ty))
314 // Otherwise it's not a no-op.
317 case Instruction::PtrToInt: {
321 // A ptrtoint cast is free so long as the result is large enough to store
322 // the pointer, and a legal integer type.
323 unsigned DestSize = Ty->getScalarSizeInBits();
324 if (DL->isLegalInteger(DestSize) &&
325 DestSize >= DL->getPointerTypeSizeInBits(OpTy))
328 // Otherwise it's not a no-op.
331 case Instruction::Trunc:
332 // trunc to a native type is free (assuming the target has compare and
333 // shift-right of the same width).
334 if (DL && DL->isLegalInteger(DL->getTypeSizeInBits(Ty)))
341 unsigned getGEPCost(const Value *Ptr,
342 ArrayRef<const Value *> Operands) const override {
343 // In the basic model, we just assume that all-constant GEPs will be folded
344 // into their uses via addressing modes.
345 for (unsigned Idx = 0, Size = Operands.size(); Idx != Size; ++Idx)
346 if (!isa<Constant>(Operands[Idx]))
352 unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const override
354 assert(FTy && "FunctionType must be provided to this routine.");
356 // The target-independent implementation just measures the size of the
357 // function by approximating that each argument will take on average one
358 // instruction to prepare.
361 // Set the argument number to the number of explicit arguments in the
363 NumArgs = FTy->getNumParams();
365 return TCC_Basic * (NumArgs + 1);
368 unsigned getCallCost(const Function *F, int NumArgs = -1) const override
370 assert(F && "A concrete function must be provided to this routine.");
373 // Set the argument number to the number of explicit arguments in the
375 NumArgs = F->arg_size();
377 if (Intrinsic::ID IID = (Intrinsic::ID)F->getIntrinsicID()) {
378 FunctionType *FTy = F->getFunctionType();
379 SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end());
380 return TopTTI->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys);
383 if (!TopTTI->isLoweredToCall(F))
384 return TCC_Basic; // Give a basic cost if it will be lowered directly.
386 return TopTTI->getCallCost(F->getFunctionType(), NumArgs);
389 unsigned getCallCost(const Function *F,
390 ArrayRef<const Value *> Arguments) const override {
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 TopTTI->getCallCost(F, Arguments.size());
397 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
398 ArrayRef<Type *> ParamTys) const override {
401 // Intrinsics rarely (if ever) have normal argument setup constraints.
402 // Model them as having a basic instruction cost.
403 // FIXME: This is wrong for libc intrinsics.
406 case Intrinsic::annotation:
407 case Intrinsic::assume:
408 case Intrinsic::dbg_declare:
409 case Intrinsic::dbg_value:
410 case Intrinsic::invariant_start:
411 case Intrinsic::invariant_end:
412 case Intrinsic::lifetime_start:
413 case Intrinsic::lifetime_end:
414 case Intrinsic::objectsize:
415 case Intrinsic::ptr_annotation:
416 case Intrinsic::var_annotation:
417 case Intrinsic::experimental_gc_result_int:
418 case Intrinsic::experimental_gc_result_float:
419 case Intrinsic::experimental_gc_result_ptr:
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() {