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::getIntImmCost(unsigned Opcode, const APInt &Imm,
163 return PrevTTI->getIntImmCost(Opcode, Imm, Ty);
166 unsigned TargetTransformInfo::getIntImmCost(Intrinsic::ID IID, const APInt &Imm,
168 return PrevTTI->getIntImmCost(IID, Imm, Ty);
171 unsigned TargetTransformInfo::getNumberOfRegisters(bool Vector) const {
172 return PrevTTI->getNumberOfRegisters(Vector);
175 unsigned TargetTransformInfo::getRegisterBitWidth(bool Vector) const {
176 return PrevTTI->getRegisterBitWidth(Vector);
179 unsigned TargetTransformInfo::getMaximumUnrollFactor() const {
180 return PrevTTI->getMaximumUnrollFactor();
183 unsigned TargetTransformInfo::getArithmeticInstrCost(unsigned Opcode,
185 OperandValueKind Op1Info,
186 OperandValueKind Op2Info) const {
187 return PrevTTI->getArithmeticInstrCost(Opcode, Ty, Op1Info, Op2Info);
190 unsigned TargetTransformInfo::getShuffleCost(ShuffleKind Kind, Type *Tp,
191 int Index, Type *SubTp) const {
192 return PrevTTI->getShuffleCost(Kind, Tp, Index, SubTp);
195 unsigned TargetTransformInfo::getCastInstrCost(unsigned Opcode, Type *Dst,
197 return PrevTTI->getCastInstrCost(Opcode, Dst, Src);
200 unsigned TargetTransformInfo::getCFInstrCost(unsigned Opcode) const {
201 return PrevTTI->getCFInstrCost(Opcode);
204 unsigned TargetTransformInfo::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
205 Type *CondTy) const {
206 return PrevTTI->getCmpSelInstrCost(Opcode, ValTy, CondTy);
209 unsigned TargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val,
210 unsigned Index) const {
211 return PrevTTI->getVectorInstrCost(Opcode, Val, Index);
214 unsigned TargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src,
216 unsigned AddressSpace) const {
217 return PrevTTI->getMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
222 TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID,
224 ArrayRef<Type *> Tys) const {
225 return PrevTTI->getIntrinsicInstrCost(ID, RetTy, Tys);
228 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const {
229 return PrevTTI->getNumberOfParts(Tp);
232 unsigned TargetTransformInfo::getAddressComputationCost(Type *Tp,
233 bool IsComplex) const {
234 return PrevTTI->getAddressComputationCost(Tp, IsComplex);
237 unsigned TargetTransformInfo::getReductionCost(unsigned Opcode, Type *Ty,
238 bool IsPairwise) const {
239 return PrevTTI->getReductionCost(Opcode, Ty, IsPairwise);
244 struct NoTTI LLVM_FINAL : ImmutablePass, TargetTransformInfo {
245 const DataLayout *DL;
247 NoTTI() : ImmutablePass(ID), DL(0) {
248 initializeNoTTIPass(*PassRegistry::getPassRegistry());
251 virtual void initializePass() LLVM_OVERRIDE {
252 // Note that this subclass is special, and must *not* call initializeTTI as
253 // it does not chain.
256 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
257 DL = DLP ? &DLP->getDataLayout() : 0;
260 virtual void getAnalysisUsage(AnalysisUsage &AU) const LLVM_OVERRIDE {
261 // Note that this subclass is special, and must *not* call
262 // TTI::getAnalysisUsage as it breaks the recursion.
265 /// Pass identification.
268 /// Provide necessary pointer adjustments for the two base classes.
269 virtual void *getAdjustedAnalysisPointer(const void *ID) LLVM_OVERRIDE {
270 if (ID == &TargetTransformInfo::ID)
271 return (TargetTransformInfo*)this;
275 unsigned getOperationCost(unsigned Opcode, Type *Ty,
276 Type *OpTy) const LLVM_OVERRIDE {
279 // By default, just classify everything as 'basic'.
282 case Instruction::GetElementPtr:
283 llvm_unreachable("Use getGEPCost for GEP operations!");
285 case Instruction::BitCast:
286 assert(OpTy && "Cast instructions must provide the operand type");
287 if (Ty == OpTy || (Ty->isPointerTy() && OpTy->isPointerTy()))
288 // Identity and pointer-to-pointer casts are free.
291 // Otherwise, the default basic cost is used.
294 case Instruction::IntToPtr: {
298 // An inttoptr cast is free so long as the input is a legal integer type
299 // which doesn't contain values outside the range of a pointer.
300 unsigned OpSize = OpTy->getScalarSizeInBits();
301 if (DL->isLegalInteger(OpSize) &&
302 OpSize <= DL->getPointerTypeSizeInBits(Ty))
305 // Otherwise it's not a no-op.
308 case Instruction::PtrToInt: {
312 // A ptrtoint cast is free so long as the result is large enough to store
313 // the pointer, and a legal integer type.
314 unsigned DestSize = Ty->getScalarSizeInBits();
315 if (DL->isLegalInteger(DestSize) &&
316 DestSize >= DL->getPointerTypeSizeInBits(OpTy))
319 // Otherwise it's not a no-op.
322 case Instruction::Trunc:
323 // trunc to a native type is free (assuming the target has compare and
324 // shift-right of the same width).
325 if (DL && DL->isLegalInteger(DL->getTypeSizeInBits(Ty)))
332 unsigned getGEPCost(const Value *Ptr,
333 ArrayRef<const Value *> Operands) const LLVM_OVERRIDE {
334 // In the basic model, we just assume that all-constant GEPs will be folded
335 // into their uses via addressing modes.
336 for (unsigned Idx = 0, Size = Operands.size(); Idx != Size; ++Idx)
337 if (!isa<Constant>(Operands[Idx]))
343 unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const LLVM_OVERRIDE
345 assert(FTy && "FunctionType must be provided to this routine.");
347 // The target-independent implementation just measures the size of the
348 // function by approximating that each argument will take on average one
349 // instruction to prepare.
352 // Set the argument number to the number of explicit arguments in the
354 NumArgs = FTy->getNumParams();
356 return TCC_Basic * (NumArgs + 1);
359 unsigned getCallCost(const Function *F, int NumArgs = -1) const LLVM_OVERRIDE
361 assert(F && "A concrete function must be provided to this routine.");
364 // Set the argument number to the number of explicit arguments in the
366 NumArgs = F->arg_size();
368 if (Intrinsic::ID IID = (Intrinsic::ID)F->getIntrinsicID()) {
369 FunctionType *FTy = F->getFunctionType();
370 SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end());
371 return TopTTI->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys);
374 if (!TopTTI->isLoweredToCall(F))
375 return TCC_Basic; // Give a basic cost if it will be lowered directly.
377 return TopTTI->getCallCost(F->getFunctionType(), NumArgs);
380 unsigned getCallCost(const Function *F,
381 ArrayRef<const Value *> Arguments) const LLVM_OVERRIDE {
382 // Simply delegate to generic handling of the call.
383 // FIXME: We should use instsimplify or something else to catch calls which
384 // will constant fold with these arguments.
385 return TopTTI->getCallCost(F, Arguments.size());
388 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
389 ArrayRef<Type *> ParamTys) const LLVM_OVERRIDE {
392 // Intrinsics rarely (if ever) have normal argument setup constraints.
393 // Model them as having a basic instruction cost.
394 // FIXME: This is wrong for libc intrinsics.
397 case Intrinsic::dbg_declare:
398 case Intrinsic::dbg_value:
399 case Intrinsic::invariant_start:
400 case Intrinsic::invariant_end:
401 case Intrinsic::lifetime_start:
402 case Intrinsic::lifetime_end:
403 case Intrinsic::objectsize:
404 case Intrinsic::ptr_annotation:
405 case Intrinsic::var_annotation:
406 // These intrinsics don't actually represent code after lowering.
412 getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
413 ArrayRef<const Value *> Arguments) const LLVM_OVERRIDE {
414 // Delegate to the generic intrinsic handling code. This mostly provides an
415 // opportunity for targets to (for example) special case the cost of
416 // certain intrinsics based on constants used as arguments.
417 SmallVector<Type *, 8> ParamTys;
418 ParamTys.reserve(Arguments.size());
419 for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx)
420 ParamTys.push_back(Arguments[Idx]->getType());
421 return TopTTI->getIntrinsicCost(IID, RetTy, ParamTys);
424 unsigned getUserCost(const User *U) const LLVM_OVERRIDE {
426 return TCC_Free; // Model all PHI nodes as free.
428 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U))
429 // In the basic model we just assume that all-constant GEPs will be
430 // folded into their uses via addressing modes.
431 return GEP->hasAllConstantIndices() ? TCC_Free : TCC_Basic;
433 if (ImmutableCallSite CS = U) {
434 const Function *F = CS.getCalledFunction();
436 // Just use the called value type.
437 Type *FTy = CS.getCalledValue()->getType()->getPointerElementType();
438 return TopTTI->getCallCost(cast<FunctionType>(FTy), CS.arg_size());
441 SmallVector<const Value *, 8> Arguments(CS.arg_begin(), CS.arg_end());
442 return TopTTI->getCallCost(F, Arguments);
445 if (const CastInst *CI = dyn_cast<CastInst>(U)) {
446 // Result of a cmp instruction is often extended (to be used by other
447 // cmp instructions, logical or return instructions). These are usually
448 // nop on most sane targets.
449 if (isa<CmpInst>(CI->getOperand(0)))
453 // Otherwise delegate to the fully generic implementations.
454 return getOperationCost(Operator::getOpcode(U), U->getType(),
455 U->getNumOperands() == 1 ?
456 U->getOperand(0)->getType() : 0);
459 bool hasBranchDivergence() const LLVM_OVERRIDE { return false; }
461 bool isLoweredToCall(const Function *F) const LLVM_OVERRIDE {
462 // FIXME: These should almost certainly not be handled here, and instead
463 // handled with the help of TLI or the target itself. This was largely
464 // ported from existing analysis heuristics here so that such refactorings
465 // can take place in the future.
467 if (F->isIntrinsic())
470 if (F->hasLocalLinkage() || !F->hasName())
473 StringRef Name = F->getName();
475 // These will all likely lower to a single selection DAG node.
476 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
477 Name == "fabs" || Name == "fabsf" || Name == "fabsl" || Name == "sin" ||
478 Name == "sinf" || Name == "sinl" || Name == "cos" || Name == "cosf" ||
479 Name == "cosl" || Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl")
482 // These are all likely to be optimized into something smaller.
483 if (Name == "pow" || Name == "powf" || Name == "powl" || Name == "exp2" ||
484 Name == "exp2l" || Name == "exp2f" || Name == "floor" || Name ==
485 "floorf" || Name == "ceil" || Name == "round" || Name == "ffs" ||
486 Name == "ffsl" || Name == "abs" || Name == "labs" || Name == "llabs")
492 void getUnrollingPreferences(Loop *,
493 UnrollingPreferences &) const LLVM_OVERRIDE
496 bool isLegalAddImmediate(int64_t Imm) const LLVM_OVERRIDE {
500 bool isLegalICmpImmediate(int64_t Imm) const LLVM_OVERRIDE {
504 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
505 bool HasBaseReg, int64_t Scale) const LLVM_OVERRIDE
507 // Guess that reg+reg addressing is allowed. This heuristic is taken from
508 // the implementation of LSR.
509 return !BaseGV && BaseOffset == 0 && Scale <= 1;
512 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
513 bool HasBaseReg, int64_t Scale) const LLVM_OVERRIDE {
514 // Guess that all legal addressing mode are free.
515 if(isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg, Scale))
520 bool isTruncateFree(Type *Ty1, Type *Ty2) const LLVM_OVERRIDE {
524 bool isTypeLegal(Type *Ty) const LLVM_OVERRIDE {
528 unsigned getJumpBufAlignment() const LLVM_OVERRIDE {
532 unsigned getJumpBufSize() const LLVM_OVERRIDE {
536 bool shouldBuildLookupTables() const LLVM_OVERRIDE {
541 getPopcntSupport(unsigned IntTyWidthInBit) const LLVM_OVERRIDE {
545 bool haveFastSqrt(Type *Ty) const LLVM_OVERRIDE {
549 unsigned getIntImmCost(const APInt &Imm, Type *Ty) const LLVM_OVERRIDE {
553 unsigned getIntImmCost(unsigned Opcode, const APInt &Imm,
554 Type *Ty) const LLVM_OVERRIDE {
558 unsigned getIntImmCost(Intrinsic::ID IID, const APInt &Imm,
559 Type *Ty) const LLVM_OVERRIDE {
563 unsigned getNumberOfRegisters(bool Vector) const LLVM_OVERRIDE {
567 unsigned getRegisterBitWidth(bool Vector) const LLVM_OVERRIDE {
571 unsigned getMaximumUnrollFactor() const LLVM_OVERRIDE {
575 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind,
576 OperandValueKind) const LLVM_OVERRIDE {
580 unsigned getShuffleCost(ShuffleKind Kind, Type *Ty,
581 int Index = 0, Type *SubTp = 0) const LLVM_OVERRIDE {
585 unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
586 Type *Src) const LLVM_OVERRIDE {
590 unsigned getCFInstrCost(unsigned Opcode) const LLVM_OVERRIDE {
594 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
595 Type *CondTy = 0) const LLVM_OVERRIDE {
599 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
600 unsigned Index = -1) const LLVM_OVERRIDE {
604 unsigned getMemoryOpCost(unsigned Opcode,
607 unsigned AddressSpace) const LLVM_OVERRIDE {
611 unsigned getIntrinsicInstrCost(Intrinsic::ID ID,
613 ArrayRef<Type*> Tys) const LLVM_OVERRIDE {
617 unsigned getNumberOfParts(Type *Tp) const LLVM_OVERRIDE {
621 unsigned getAddressComputationCost(Type *Tp, bool) const LLVM_OVERRIDE {
625 unsigned getReductionCost(unsigned, Type *, bool) const LLVM_OVERRIDE {
630 } // end anonymous namespace
632 INITIALIZE_AG_PASS(NoTTI, TargetTransformInfo, "notti",
633 "No target information", true, true, true)
636 ImmutablePass *llvm::createNoTargetTransformInfoPass() {