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 #define DEBUG_TYPE "tti"
11 #include "llvm/Analysis/TargetTransformInfo.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/CallSite.h"
18 #include "llvm/Support/ErrorHandling.h"
22 // Setup the analysis group to manage the TargetTransformInfo passes.
23 INITIALIZE_ANALYSIS_GROUP(TargetTransformInfo, "Target Information", NoTTI)
24 char TargetTransformInfo::ID = 0;
26 TargetTransformInfo::~TargetTransformInfo() {
29 void TargetTransformInfo::pushTTIStack(Pass *P) {
31 PrevTTI = &P->getAnalysis<TargetTransformInfo>();
33 // Walk up the chain and update the top TTI pointer.
34 for (TargetTransformInfo *PTTI = PrevTTI; PTTI; PTTI = PTTI->PrevTTI)
38 void TargetTransformInfo::popTTIStack() {
41 // Walk up the chain and update the top TTI pointer.
42 for (TargetTransformInfo *PTTI = PrevTTI; PTTI; PTTI = PTTI->PrevTTI)
43 PTTI->TopTTI = PrevTTI;
48 void TargetTransformInfo::getAnalysisUsage(AnalysisUsage &AU) const {
49 AU.addRequired<TargetTransformInfo>();
52 unsigned TargetTransformInfo::getOperationCost(unsigned Opcode, Type *Ty,
54 return PrevTTI->getOperationCost(Opcode, Ty, OpTy);
57 unsigned TargetTransformInfo::getGEPCost(
58 const Value *Ptr, ArrayRef<const Value *> Operands) const {
59 return PrevTTI->getGEPCost(Ptr, Operands);
62 unsigned TargetTransformInfo::getCallCost(FunctionType *FTy,
64 return PrevTTI->getCallCost(FTy, NumArgs);
67 unsigned TargetTransformInfo::getCallCost(const Function *F,
69 return PrevTTI->getCallCost(F, NumArgs);
72 unsigned TargetTransformInfo::getCallCost(
73 const Function *F, ArrayRef<const Value *> Arguments) const {
74 return PrevTTI->getCallCost(F, Arguments);
77 unsigned TargetTransformInfo::getIntrinsicCost(
78 Intrinsic::ID IID, Type *RetTy, ArrayRef<Type *> ParamTys) const {
79 return PrevTTI->getIntrinsicCost(IID, RetTy, ParamTys);
82 unsigned TargetTransformInfo::getIntrinsicCost(
83 Intrinsic::ID IID, Type *RetTy, ArrayRef<const Value *> Arguments) const {
84 return PrevTTI->getIntrinsicCost(IID, RetTy, Arguments);
87 unsigned TargetTransformInfo::getUserCost(const User *U) const {
88 return PrevTTI->getUserCost(U);
91 bool TargetTransformInfo::hasBranchDivergence() const {
92 return PrevTTI->hasBranchDivergence();
95 bool TargetTransformInfo::isLoweredToCall(const Function *F) const {
96 return PrevTTI->isLoweredToCall(F);
99 void TargetTransformInfo::getUnrollingPreferences(Loop *L,
100 UnrollingPreferences &UP) const {
101 PrevTTI->getUnrollingPreferences(L, UP);
104 bool TargetTransformInfo::isLegalAddImmediate(int64_t Imm) const {
105 return PrevTTI->isLegalAddImmediate(Imm);
108 bool TargetTransformInfo::isLegalICmpImmediate(int64_t Imm) const {
109 return PrevTTI->isLegalICmpImmediate(Imm);
112 bool TargetTransformInfo::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
115 int64_t Scale) const {
116 return PrevTTI->isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
120 int TargetTransformInfo::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
123 int64_t Scale) const {
124 return PrevTTI->getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg,
128 bool TargetTransformInfo::isTruncateFree(Type *Ty1, Type *Ty2) const {
129 return PrevTTI->isTruncateFree(Ty1, Ty2);
132 bool TargetTransformInfo::isTypeLegal(Type *Ty) const {
133 return PrevTTI->isTypeLegal(Ty);
136 unsigned TargetTransformInfo::getJumpBufAlignment() const {
137 return PrevTTI->getJumpBufAlignment();
140 unsigned TargetTransformInfo::getJumpBufSize() const {
141 return PrevTTI->getJumpBufSize();
144 bool TargetTransformInfo::shouldBuildLookupTables() const {
145 return PrevTTI->shouldBuildLookupTables();
148 TargetTransformInfo::PopcntSupportKind
149 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit) const {
150 return PrevTTI->getPopcntSupport(IntTyWidthInBit);
153 bool TargetTransformInfo::haveFastSqrt(Type *Ty) const {
154 return PrevTTI->haveFastSqrt(Ty);
157 unsigned TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty) const {
158 return PrevTTI->getIntImmCost(Imm, Ty);
161 unsigned TargetTransformInfo::getNumberOfRegisters(bool Vector) const {
162 return PrevTTI->getNumberOfRegisters(Vector);
165 unsigned TargetTransformInfo::getRegisterBitWidth(bool Vector) const {
166 return PrevTTI->getRegisterBitWidth(Vector);
169 unsigned TargetTransformInfo::getMaximumUnrollFactor() const {
170 return PrevTTI->getMaximumUnrollFactor();
173 unsigned TargetTransformInfo::getArithmeticInstrCost(unsigned Opcode,
175 OperandValueKind Op1Info,
176 OperandValueKind Op2Info) const {
177 return PrevTTI->getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info);
180 unsigned TargetTransformInfo::getShuffleCost(ShuffleKind Kind, Type *Tp,
181 int Index, Type *SubTp) const {
182 return PrevTTI->getShuffleCost(Kind, Tp, Index, SubTp);
185 unsigned TargetTransformInfo::getCastInstrCost(unsigned Opcode, Type *Dst,
187 return PrevTTI->getCastInstrCost(Opcode, Dst, Src);
190 unsigned TargetTransformInfo::getCFInstrCost(unsigned Opcode) const {
191 return PrevTTI->getCFInstrCost(Opcode);
194 unsigned TargetTransformInfo::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
195 Type *CondTy) const {
196 return PrevTTI->getCmpSelInstrCost(Opcode, ValTy, CondTy);
199 unsigned TargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val,
200 unsigned Index) const {
201 return PrevTTI->getVectorInstrCost(Opcode, Val, Index);
204 unsigned TargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src,
206 unsigned AddressSpace) const {
207 return PrevTTI->getMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
212 TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID,
214 ArrayRef<Type *> Tys) const {
215 return PrevTTI->getIntrinsicInstrCost(ID, RetTy, Tys);
218 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const {
219 return PrevTTI->getNumberOfParts(Tp);
222 unsigned TargetTransformInfo::getAddressComputationCost(Type *Tp,
223 bool IsComplex) const {
224 return PrevTTI->getAddressComputationCost(Tp, IsComplex);
227 unsigned TargetTransformInfo::getReductionCost(unsigned Opcode, Type *Ty,
228 bool IsPairwise) const {
229 return PrevTTI->getReductionCost(Opcode, Ty, IsPairwise);
234 struct NoTTI LLVM_FINAL : ImmutablePass, TargetTransformInfo {
235 const DataLayout *DL;
237 NoTTI() : ImmutablePass(ID), DL(0) {
238 initializeNoTTIPass(*PassRegistry::getPassRegistry());
241 virtual void initializePass() LLVM_OVERRIDE {
242 // Note that this subclass is special, and must *not* call initializeTTI as
243 // it does not chain.
246 DL = getAnalysisIfAvailable<DataLayout>();
249 virtual void getAnalysisUsage(AnalysisUsage &AU) const LLVM_OVERRIDE {
250 // Note that this subclass is special, and must *not* call
251 // TTI::getAnalysisUsage as it breaks the recursion.
254 /// Pass identification.
257 /// Provide necessary pointer adjustments for the two base classes.
258 virtual void *getAdjustedAnalysisPointer(const void *ID) LLVM_OVERRIDE {
259 if (ID == &TargetTransformInfo::ID)
260 return (TargetTransformInfo*)this;
264 unsigned getOperationCost(unsigned Opcode, Type *Ty,
265 Type *OpTy) const LLVM_OVERRIDE {
268 // By default, just classify everything as 'basic'.
271 case Instruction::GetElementPtr:
272 llvm_unreachable("Use getGEPCost for GEP operations!");
274 case Instruction::BitCast:
275 assert(OpTy && "Cast instructions must provide the operand type");
276 if (Ty == OpTy || (Ty->isPointerTy() && OpTy->isPointerTy()))
277 // Identity and pointer-to-pointer casts are free.
280 // Otherwise, the default basic cost is used.
283 case Instruction::IntToPtr: {
287 // An inttoptr cast is free so long as the input is a legal integer type
288 // which doesn't contain values outside the range of a pointer.
289 unsigned OpSize = OpTy->getScalarSizeInBits();
290 if (DL->isLegalInteger(OpSize) &&
291 OpSize <= DL->getPointerTypeSizeInBits(Ty))
294 // Otherwise it's not a no-op.
297 case Instruction::PtrToInt: {
301 // A ptrtoint cast is free so long as the result is large enough to store
302 // the pointer, and a legal integer type.
303 unsigned DestSize = Ty->getScalarSizeInBits();
304 if (DL->isLegalInteger(DestSize) &&
305 DestSize >= DL->getPointerTypeSizeInBits(OpTy))
308 // Otherwise it's not a no-op.
311 case Instruction::Trunc:
312 // trunc to a native type is free (assuming the target has compare and
313 // shift-right of the same width).
314 if (DL && DL->isLegalInteger(DL->getTypeSizeInBits(Ty)))
321 unsigned getGEPCost(const Value *Ptr,
322 ArrayRef<const Value *> Operands) const LLVM_OVERRIDE {
323 // In the basic model, we just assume that all-constant GEPs will be folded
324 // into their uses via addressing modes.
325 for (unsigned Idx = 0, Size = Operands.size(); Idx != Size; ++Idx)
326 if (!isa<Constant>(Operands[Idx]))
332 unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const LLVM_OVERRIDE
334 assert(FTy && "FunctionType must be provided to this routine.");
336 // The target-independent implementation just measures the size of the
337 // function by approximating that each argument will take on average one
338 // instruction to prepare.
341 // Set the argument number to the number of explicit arguments in the
343 NumArgs = FTy->getNumParams();
345 return TCC_Basic * (NumArgs + 1);
348 unsigned getCallCost(const Function *F, int NumArgs = -1) const LLVM_OVERRIDE
350 assert(F && "A concrete function must be provided to this routine.");
353 // Set the argument number to the number of explicit arguments in the
355 NumArgs = F->arg_size();
357 if (Intrinsic::ID IID = (Intrinsic::ID)F->getIntrinsicID()) {
358 FunctionType *FTy = F->getFunctionType();
359 SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end());
360 return TopTTI->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys);
363 if (!TopTTI->isLoweredToCall(F))
364 return TCC_Basic; // Give a basic cost if it will be lowered directly.
366 return TopTTI->getCallCost(F->getFunctionType(), NumArgs);
369 unsigned getCallCost(const Function *F,
370 ArrayRef<const Value *> Arguments) const LLVM_OVERRIDE {
371 // Simply delegate to generic handling of the call.
372 // FIXME: We should use instsimplify or something else to catch calls which
373 // will constant fold with these arguments.
374 return TopTTI->getCallCost(F, Arguments.size());
377 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
378 ArrayRef<Type *> ParamTys) const LLVM_OVERRIDE {
381 // Intrinsics rarely (if ever) have normal argument setup constraints.
382 // Model them as having a basic instruction cost.
383 // FIXME: This is wrong for libc intrinsics.
386 case Intrinsic::dbg_declare:
387 case Intrinsic::dbg_value:
388 case Intrinsic::invariant_start:
389 case Intrinsic::invariant_end:
390 case Intrinsic::lifetime_start:
391 case Intrinsic::lifetime_end:
392 case Intrinsic::objectsize:
393 case Intrinsic::ptr_annotation:
394 case Intrinsic::var_annotation:
395 // These intrinsics don't actually represent code after lowering.
401 getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
402 ArrayRef<const Value *> Arguments) const LLVM_OVERRIDE {
403 // Delegate to the generic intrinsic handling code. This mostly provides an
404 // opportunity for targets to (for example) special case the cost of
405 // certain intrinsics based on constants used as arguments.
406 SmallVector<Type *, 8> ParamTys;
407 ParamTys.reserve(Arguments.size());
408 for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx)
409 ParamTys.push_back(Arguments[Idx]->getType());
410 return TopTTI->getIntrinsicCost(IID, RetTy, ParamTys);
413 unsigned getUserCost(const User *U) const LLVM_OVERRIDE {
415 return TCC_Free; // Model all PHI nodes as free.
417 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U))
418 // In the basic model we just assume that all-constant GEPs will be
419 // folded into their uses via addressing modes.
420 return GEP->hasAllConstantIndices() ? TCC_Free : TCC_Basic;
422 if (ImmutableCallSite CS = U) {
423 const Function *F = CS.getCalledFunction();
425 // Just use the called value type.
426 Type *FTy = CS.getCalledValue()->getType()->getPointerElementType();
427 return TopTTI->getCallCost(cast<FunctionType>(FTy), CS.arg_size());
430 SmallVector<const Value *, 8> Arguments;
431 for (ImmutableCallSite::arg_iterator AI = CS.arg_begin(),
434 Arguments.push_back(*AI);
436 return TopTTI->getCallCost(F, Arguments);
439 if (const CastInst *CI = dyn_cast<CastInst>(U)) {
440 // Result of a cmp instruction is often extended (to be used by other
441 // cmp instructions, logical or return instructions). These are usually
442 // nop on most sane targets.
443 if (isa<CmpInst>(CI->getOperand(0)))
447 // Otherwise delegate to the fully generic implementations.
448 return getOperationCost(Operator::getOpcode(U), U->getType(),
449 U->getNumOperands() == 1 ?
450 U->getOperand(0)->getType() : 0);
453 bool hasBranchDivergence() const LLVM_OVERRIDE { return false; }
455 bool isLoweredToCall(const Function *F) const LLVM_OVERRIDE {
456 // FIXME: These should almost certainly not be handled here, and instead
457 // handled with the help of TLI or the target itself. This was largely
458 // ported from existing analysis heuristics here so that such refactorings
459 // can take place in the future.
461 if (F->isIntrinsic())
464 if (F->hasLocalLinkage() || !F->hasName())
467 StringRef Name = F->getName();
469 // These will all likely lower to a single selection DAG node.
470 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
471 Name == "fabs" || Name == "fabsf" || Name == "fabsl" || Name == "sin" ||
472 Name == "sinf" || Name == "sinl" || Name == "cos" || Name == "cosf" ||
473 Name == "cosl" || Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl")
476 // These are all likely to be optimized into something smaller.
477 if (Name == "pow" || Name == "powf" || Name == "powl" || Name == "exp2" ||
478 Name == "exp2l" || Name == "exp2f" || Name == "floor" || Name ==
479 "floorf" || Name == "ceil" || Name == "round" || Name == "ffs" ||
480 Name == "ffsl" || Name == "abs" || Name == "labs" || Name == "llabs")
486 void getUnrollingPreferences(Loop *,
487 UnrollingPreferences &) const LLVM_OVERRIDE
490 bool isLegalAddImmediate(int64_t Imm) const LLVM_OVERRIDE {
494 bool isLegalICmpImmediate(int64_t Imm) const LLVM_OVERRIDE {
498 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
499 bool HasBaseReg, int64_t Scale) const LLVM_OVERRIDE
501 // Guess that reg+reg addressing is allowed. This heuristic is taken from
502 // the implementation of LSR.
503 return !BaseGV && BaseOffset == 0 && Scale <= 1;
506 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
507 bool HasBaseReg, int64_t Scale) const LLVM_OVERRIDE {
508 // Guess that all legal addressing mode are free.
509 if(isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg, Scale))
514 bool isTruncateFree(Type *Ty1, Type *Ty2) const LLVM_OVERRIDE {
518 bool isTypeLegal(Type *Ty) const LLVM_OVERRIDE {
522 unsigned getJumpBufAlignment() const LLVM_OVERRIDE {
526 unsigned getJumpBufSize() const LLVM_OVERRIDE {
530 bool shouldBuildLookupTables() const LLVM_OVERRIDE {
535 getPopcntSupport(unsigned IntTyWidthInBit) const LLVM_OVERRIDE {
539 bool haveFastSqrt(Type *Ty) const LLVM_OVERRIDE {
543 unsigned getIntImmCost(const APInt &Imm, Type *Ty) const LLVM_OVERRIDE {
547 unsigned getNumberOfRegisters(bool Vector) const LLVM_OVERRIDE {
551 unsigned getRegisterBitWidth(bool Vector) const LLVM_OVERRIDE {
555 unsigned getMaximumUnrollFactor() const LLVM_OVERRIDE {
559 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind,
560 OperandValueKind) const LLVM_OVERRIDE {
564 unsigned getShuffleCost(ShuffleKind Kind, Type *Ty,
565 int Index = 0, Type *SubTp = 0) const LLVM_OVERRIDE {
569 unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
570 Type *Src) const LLVM_OVERRIDE {
574 unsigned getCFInstrCost(unsigned Opcode) const LLVM_OVERRIDE {
578 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
579 Type *CondTy = 0) const LLVM_OVERRIDE {
583 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
584 unsigned Index = -1) const LLVM_OVERRIDE {
588 unsigned getMemoryOpCost(unsigned Opcode,
591 unsigned AddressSpace) const LLVM_OVERRIDE {
595 unsigned getIntrinsicInstrCost(Intrinsic::ID ID,
597 ArrayRef<Type*> Tys) const LLVM_OVERRIDE {
601 unsigned getNumberOfParts(Type *Tp) const LLVM_OVERRIDE {
605 unsigned getAddressComputationCost(Type *Tp, bool) const LLVM_OVERRIDE {
609 unsigned getReductionCost(unsigned, Type *, bool) const LLVM_OVERRIDE {
614 } // end anonymous namespace
616 INITIALIZE_AG_PASS(NoTTI, TargetTransformInfo, "notti",
617 "No target information", true, true, true)
620 ImmutablePass *llvm::createNoTargetTransformInfoPass() {