1 //===- InlineCost.cpp - Cost analysis for inliner -------------------------===//
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 implements inline cost analysis.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Analysis/InlineCost.h"
15 #include "llvm/Support/CallSite.h"
16 #include "llvm/CallingConv.h"
17 #include "llvm/IntrinsicInst.h"
18 #include "llvm/Target/TargetData.h"
19 #include "llvm/ADT/SmallPtrSet.h"
23 /// callIsSmall - If a call is likely to lower to a single target instruction,
24 /// or is otherwise deemed small return true.
25 /// TODO: Perhaps calls like memcpy, strcpy, etc?
26 bool llvm::callIsSmall(const Function *F) {
29 if (F->hasLocalLinkage()) return false;
31 if (!F->hasName()) return false;
33 StringRef Name = F->getName();
35 // These will all likely lower to a single selection DAG node.
36 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
37 Name == "fabs" || Name == "fabsf" || Name == "fabsl" ||
38 Name == "sin" || Name == "sinf" || Name == "sinl" ||
39 Name == "cos" || Name == "cosf" || Name == "cosl" ||
40 Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl" )
43 // These are all likely to be optimized into something smaller.
44 if (Name == "pow" || Name == "powf" || Name == "powl" ||
45 Name == "exp2" || Name == "exp2l" || Name == "exp2f" ||
46 Name == "floor" || Name == "floorf" || Name == "ceil" ||
47 Name == "round" || Name == "ffs" || Name == "ffsl" ||
48 Name == "abs" || Name == "labs" || Name == "llabs")
54 /// analyzeBasicBlock - Fill in the current structure with information gleaned
55 /// from the specified block.
56 void CodeMetrics::analyzeBasicBlock(const BasicBlock *BB,
57 const TargetData *TD) {
59 unsigned NumInstsBeforeThisBB = NumInsts;
60 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
62 if (isa<PHINode>(II)) continue; // PHI nodes don't count.
64 // Special handling for calls.
65 if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
66 if (const IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(II)) {
67 switch (IntrinsicI->getIntrinsicID()) {
69 case Intrinsic::dbg_declare:
70 case Intrinsic::dbg_value:
71 case Intrinsic::invariant_start:
72 case Intrinsic::invariant_end:
73 case Intrinsic::lifetime_start:
74 case Intrinsic::lifetime_end:
75 case Intrinsic::objectsize:
76 case Intrinsic::ptr_annotation:
77 case Intrinsic::var_annotation:
78 // These intrinsics don't count as size.
83 ImmutableCallSite CS(cast<Instruction>(II));
85 if (const Function *F = CS.getCalledFunction()) {
86 // If a function is both internal and has a single use, then it is
87 // extremely likely to get inlined in the future (it was probably
88 // exposed by an interleaved devirtualization pass).
89 if (!CS.isNoInline() && F->hasInternalLinkage() && F->hasOneUse())
90 ++NumInlineCandidates;
92 // If this call is to function itself, then the function is recursive.
93 // Inlining it into other functions is a bad idea, because this is
94 // basically just a form of loop peeling, and our metrics aren't useful
96 if (F == BB->getParent())
100 if (!isa<IntrinsicInst>(II) && !callIsSmall(CS.getCalledFunction())) {
101 // Each argument to a call takes on average one instruction to set up.
102 NumInsts += CS.arg_size();
104 // We don't want inline asm to count as a call - that would prevent loop
105 // unrolling. The argument setup cost is still real, though.
106 if (!isa<InlineAsm>(CS.getCalledValue()))
111 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
112 if (!AI->isStaticAlloca())
113 this->usesDynamicAlloca = true;
116 if (isa<ExtractElementInst>(II) || II->getType()->isVectorTy())
119 if (const CastInst *CI = dyn_cast<CastInst>(II)) {
120 // Noop casts, including ptr <-> int, don't count.
121 if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
122 isa<PtrToIntInst>(CI))
124 // trunc to a native type is free (assuming the target has compare and
125 // shift-right of the same width).
126 if (isa<TruncInst>(CI) && TD &&
127 TD->isLegalInteger(TD->getTypeSizeInBits(CI->getType())))
129 // Result of a cmp instruction is often extended (to be used by other
130 // cmp instructions, logical or return instructions). These are usually
131 // nop on most sane targets.
132 if (isa<CmpInst>(CI->getOperand(0)))
134 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(II)){
135 // If a GEP has all constant indices, it will probably be folded with
137 if (GEPI->hasAllConstantIndices())
144 if (isa<ReturnInst>(BB->getTerminator()))
147 // We never want to inline functions that contain an indirectbr. This is
148 // incorrect because all the blockaddress's (in static global initializers
149 // for example) would be referring to the original function, and this indirect
150 // jump would jump from the inlined copy of the function into the original
151 // function which is extremely undefined behavior.
152 // FIXME: This logic isn't really right; we can safely inline functions
153 // with indirectbr's as long as no other function or global references the
154 // blockaddress of a block within the current function. And as a QOI issue,
155 // if someone is using a blockaddress without an indirectbr, and that
156 // reference somehow ends up in another function or global, we probably
157 // don't want to inline this function.
158 if (isa<IndirectBrInst>(BB->getTerminator()))
159 containsIndirectBr = true;
161 // Remember NumInsts for this BB.
162 NumBBInsts[BB] = NumInsts - NumInstsBeforeThisBB;
165 unsigned InlineCostAnalyzer::FunctionInfo::countCodeReductionForConstant(
166 const CodeMetrics &Metrics, Value *V) {
167 unsigned Reduction = 0;
168 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
170 if (isa<BranchInst>(U) || isa<SwitchInst>(U)) {
171 // We will be able to eliminate all but one of the successors.
172 const TerminatorInst &TI = cast<TerminatorInst>(*U);
173 const unsigned NumSucc = TI.getNumSuccessors();
175 for (unsigned I = 0; I != NumSucc; ++I)
176 Instrs += Metrics.NumBBInsts.lookup(TI.getSuccessor(I));
177 // We don't know which blocks will be eliminated, so use the average size.
178 Reduction += InlineConstants::InstrCost*Instrs*(NumSucc-1)/NumSucc;
180 // Figure out if this instruction will be removed due to simple constant
182 Instruction &Inst = cast<Instruction>(*U);
184 // We can't constant propagate instructions which have effects or
187 // FIXME: It would be nice to capture the fact that a load from a
188 // pointer-to-constant-global is actually a *really* good thing to zap.
189 // Unfortunately, we don't know the pointer that may get propagated here,
190 // so we can't make this decision.
191 if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
192 isa<AllocaInst>(Inst))
195 bool AllOperandsConstant = true;
196 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
197 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
198 AllOperandsConstant = false;
202 if (AllOperandsConstant) {
203 // We will get to remove this instruction...
204 Reduction += InlineConstants::InstrCost;
206 // And any other instructions that use it which become constants
208 Reduction += countCodeReductionForConstant(Metrics, &Inst);
215 static unsigned countCodeReductionForAllocaICmp(const CodeMetrics &Metrics,
217 unsigned Reduction = 0;
219 // Bail if this is comparing against a non-constant; there is nothing we can
221 if (!isa<Constant>(ICI->getOperand(1)))
224 // An icmp pred (alloca, C) becomes true if the predicate is true when
225 // equal and false otherwise.
226 bool Result = ICI->isTrueWhenEqual();
228 SmallVector<Instruction *, 4> Worklist;
229 Worklist.push_back(ICI);
231 Instruction *U = Worklist.pop_back_val();
232 Reduction += InlineConstants::InstrCost;
233 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
235 Instruction *I = dyn_cast<Instruction>(*UI);
236 if (!I || I->mayHaveSideEffects()) continue;
237 if (I->getNumOperands() == 1)
238 Worklist.push_back(I);
239 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
240 // If BO produces the same value as U, then the other operand is
241 // irrelevant and we can put it into the Worklist to continue
242 // deleting dead instructions. If BO produces the same value as the
243 // other operand, we can delete BO but that's it.
244 if (Result == true) {
245 if (BO->getOpcode() == Instruction::Or)
246 Worklist.push_back(I);
247 if (BO->getOpcode() == Instruction::And)
248 Reduction += InlineConstants::InstrCost;
250 if (BO->getOpcode() == Instruction::Or ||
251 BO->getOpcode() == Instruction::Xor)
252 Reduction += InlineConstants::InstrCost;
253 if (BO->getOpcode() == Instruction::And)
254 Worklist.push_back(I);
257 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
258 BasicBlock *BB = BI->getSuccessor(Result ? 0 : 1);
259 if (BB->getSinglePredecessor())
261 += InlineConstants::InstrCost * Metrics.NumBBInsts.lookup(BB);
264 } while (!Worklist.empty());
269 /// \brief Compute the reduction possible for a given instruction if we are able
270 /// to SROA an alloca.
272 /// The reduction for this instruction is added to the SROAReduction output
273 /// parameter. Returns false if this instruction is expected to defeat SROA in
275 static bool countCodeReductionForSROAInst(Instruction *I,
276 SmallVectorImpl<Value *> &Worklist,
277 unsigned &SROAReduction) {
278 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
281 SROAReduction += InlineConstants::InstrCost;
285 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
288 SROAReduction += InlineConstants::InstrCost;
292 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
293 // If the GEP has variable indices, we won't be able to do much with it.
294 if (!GEP->hasAllConstantIndices())
296 // A non-zero GEP will likely become a mask operation after SROA.
297 if (GEP->hasAllZeroIndices())
298 SROAReduction += InlineConstants::InstrCost;
299 Worklist.push_back(GEP);
303 if (BitCastInst *BCI = dyn_cast<BitCastInst>(I)) {
304 // Track pointer through bitcasts.
305 Worklist.push_back(BCI);
306 SROAReduction += InlineConstants::InstrCost;
310 // We just look for non-constant operands to ICmp instructions as those will
311 // defeat SROA. The actual reduction for these happens even without SROA.
312 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
313 return isa<Constant>(ICI->getOperand(1));
315 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
316 // SROA can handle a select of alloca iff all uses of the alloca are
317 // loads, and dereferenceable. We assume it's dereferenceable since
318 // we're told the input is an alloca.
319 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end();
321 LoadInst *LI = dyn_cast<LoadInst>(*UI);
322 if (LI == 0 || !LI->isSimple())
325 // We don't know whether we'll be deleting the rest of the chain of
326 // instructions from the SelectInst on, because we don't know whether
327 // the other side of the select is also an alloca or not.
331 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
332 switch (II->getIntrinsicID()) {
335 case Intrinsic::memset:
336 case Intrinsic::memcpy:
337 case Intrinsic::memmove:
338 case Intrinsic::lifetime_start:
339 case Intrinsic::lifetime_end:
340 // SROA can usually chew through these intrinsics.
341 SROAReduction += InlineConstants::InstrCost;
346 // If there is some other strange instruction, we're not going to be
347 // able to do much if we inline this.
351 unsigned InlineCostAnalyzer::FunctionInfo::countCodeReductionForAlloca(
352 const CodeMetrics &Metrics, Value *V) {
353 if (!V->getType()->isPointerTy()) return 0; // Not a pointer
354 unsigned Reduction = 0;
355 unsigned SROAReduction = 0;
356 bool CanSROAAlloca = true;
358 SmallVector<Value *, 4> Worklist;
359 Worklist.push_back(V);
361 Value *V = Worklist.pop_back_val();
362 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
364 Instruction *I = cast<Instruction>(*UI);
366 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
367 Reduction += countCodeReductionForAllocaICmp(Metrics, ICI);
370 CanSROAAlloca = countCodeReductionForSROAInst(I, Worklist,
373 } while (!Worklist.empty());
375 return Reduction + (CanSROAAlloca ? SROAReduction : 0);
378 /// analyzeFunction - Fill in the current structure with information gleaned
379 /// from the specified function.
380 void CodeMetrics::analyzeFunction(Function *F, const TargetData *TD) {
381 // If this function contains a call that "returns twice" (e.g., setjmp or
382 // _setjmp) and it isn't marked with "returns twice" itself, never inline it.
383 // This is a hack because we depend on the user marking their local variables
384 // as volatile if they are live across a setjmp call, and they probably
385 // won't do this in callers.
386 exposesReturnsTwice = F->callsFunctionThatReturnsTwice() &&
387 !F->hasFnAttr(Attribute::ReturnsTwice);
389 // Look at the size of the callee.
390 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
391 analyzeBasicBlock(&*BB, TD);
394 /// analyzeFunction - Fill in the current structure with information gleaned
395 /// from the specified function.
396 void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F,
397 const TargetData *TD) {
398 Metrics.analyzeFunction(F, TD);
400 // A function with exactly one return has it removed during the inlining
401 // process (see InlineFunction), so don't count it.
402 // FIXME: This knowledge should really be encoded outside of FunctionInfo.
403 if (Metrics.NumRets==1)
406 // Check out all of the arguments to the function, figuring out how much
407 // code can be eliminated if one of the arguments is a constant.
408 ArgumentWeights.reserve(F->arg_size());
409 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
410 ArgumentWeights.push_back(ArgInfo(countCodeReductionForConstant(Metrics, I),
411 countCodeReductionForAlloca(Metrics, I)));
414 /// NeverInline - returns true if the function should never be inlined into
416 bool InlineCostAnalyzer::FunctionInfo::NeverInline() {
417 return (Metrics.exposesReturnsTwice || Metrics.isRecursive ||
418 Metrics.containsIndirectBr);
420 // getSpecializationBonus - The heuristic used to determine the per-call
421 // performance boost for using a specialization of Callee with argument
422 // specializedArgNo replaced by a constant.
423 int InlineCostAnalyzer::getSpecializationBonus(Function *Callee,
424 SmallVectorImpl<unsigned> &SpecializedArgNos)
426 if (Callee->mayBeOverridden())
430 // If this function uses the coldcc calling convention, prefer not to
432 if (Callee->getCallingConv() == CallingConv::Cold)
433 Bonus -= InlineConstants::ColdccPenalty;
435 // Get information about the callee.
436 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
438 // If we haven't calculated this information yet, do so now.
439 if (CalleeFI->Metrics.NumBlocks == 0)
440 CalleeFI->analyzeFunction(Callee, TD);
444 for (Function::arg_iterator I = Callee->arg_begin(), E = Callee->arg_end();
445 I != E; ++I, ++ArgNo)
446 if (ArgNo == SpecializedArgNos[i]) {
448 Bonus += CountBonusForConstant(I);
451 // Calls usually take a long time, so they make the specialization gain
453 Bonus -= CalleeFI->Metrics.NumCalls * InlineConstants::CallPenalty;
458 // ConstantFunctionBonus - Figure out how much of a bonus we can get for
459 // possibly devirtualizing a function. We'll subtract the size of the function
460 // we may wish to inline from the indirect call bonus providing a limit on
461 // growth. Leave an upper limit of 0 for the bonus - we don't want to penalize
462 // inlining because we decide we don't want to give a bonus for
464 int InlineCostAnalyzer::ConstantFunctionBonus(CallSite CS, Constant *C) {
466 // This could just be NULL.
469 Function *F = dyn_cast<Function>(C);
472 int Bonus = InlineConstants::IndirectCallBonus + getInlineSize(CS, F);
473 return (Bonus > 0) ? 0 : Bonus;
476 // CountBonusForConstant - Figure out an approximation for how much per-call
477 // performance boost we can expect if the specified value is constant.
478 int InlineCostAnalyzer::CountBonusForConstant(Value *V, Constant *C) {
480 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
482 if (CallInst *CI = dyn_cast<CallInst>(U)) {
483 // Turning an indirect call into a direct call is a BIG win
484 if (CI->getCalledValue() == V)
485 Bonus += ConstantFunctionBonus(CallSite(CI), C);
486 } else if (InvokeInst *II = dyn_cast<InvokeInst>(U)) {
487 // Turning an indirect call into a direct call is a BIG win
488 if (II->getCalledValue() == V)
489 Bonus += ConstantFunctionBonus(CallSite(II), C);
491 // FIXME: Eliminating conditional branches and switches should
492 // also yield a per-call performance boost.
494 // Figure out the bonuses that wll accrue due to simple constant
496 Instruction &Inst = cast<Instruction>(*U);
498 // We can't constant propagate instructions which have effects or
501 // FIXME: It would be nice to capture the fact that a load from a
502 // pointer-to-constant-global is actually a *really* good thing to zap.
503 // Unfortunately, we don't know the pointer that may get propagated here,
504 // so we can't make this decision.
505 if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
506 isa<AllocaInst>(Inst))
509 bool AllOperandsConstant = true;
510 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
511 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
512 AllOperandsConstant = false;
516 if (AllOperandsConstant)
517 Bonus += CountBonusForConstant(&Inst);
524 int InlineCostAnalyzer::getInlineSize(CallSite CS, Function *Callee) {
525 // Get information about the callee.
526 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
528 // If we haven't calculated this information yet, do so now.
529 if (CalleeFI->Metrics.NumBlocks == 0)
530 CalleeFI->analyzeFunction(Callee, TD);
532 // InlineCost - This value measures how good of an inline candidate this call
533 // site is to inline. A lower inline cost make is more likely for the call to
534 // be inlined. This value may go negative.
538 // Compute any size reductions we can expect due to arguments being passed into
542 CallSite::arg_iterator I = CS.arg_begin();
543 for (Function::arg_iterator FI = Callee->arg_begin(), FE = Callee->arg_end();
544 FI != FE; ++I, ++FI, ++ArgNo) {
546 // If an alloca is passed in, inlining this function is likely to allow
547 // significant future optimization possibilities (like scalar promotion, and
548 // scalarization), so encourage the inlining of the function.
550 if (isa<AllocaInst>(I))
551 InlineCost -= CalleeFI->ArgumentWeights[ArgNo].AllocaWeight;
553 // If this is a constant being passed into the function, use the argument
554 // weights calculated for the callee to determine how much will be folded
555 // away with this information.
556 else if (isa<Constant>(I))
557 InlineCost -= CalleeFI->ArgumentWeights[ArgNo].ConstantWeight;
560 // Each argument passed in has a cost at both the caller and the callee
561 // sides. Measurements show that each argument costs about the same as an
563 InlineCost -= (CS.arg_size() * InlineConstants::InstrCost);
565 // Now that we have considered all of the factors that make the call site more
566 // likely to be inlined, look at factors that make us not want to inline it.
568 // Calls usually take a long time, so they make the inlining gain smaller.
569 InlineCost += CalleeFI->Metrics.NumCalls * InlineConstants::CallPenalty;
571 // Look at the size of the callee. Each instruction counts as 5.
572 InlineCost += CalleeFI->Metrics.NumInsts * InlineConstants::InstrCost;
577 int InlineCostAnalyzer::getInlineBonuses(CallSite CS, Function *Callee) {
578 // Get information about the callee.
579 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
581 // If we haven't calculated this information yet, do so now.
582 if (CalleeFI->Metrics.NumBlocks == 0)
583 CalleeFI->analyzeFunction(Callee, TD);
585 bool isDirectCall = CS.getCalledFunction() == Callee;
586 Instruction *TheCall = CS.getInstruction();
589 // If there is only one call of the function, and it has internal linkage,
590 // make it almost guaranteed to be inlined.
592 if (Callee->hasLocalLinkage() && Callee->hasOneUse() && isDirectCall)
593 Bonus += InlineConstants::LastCallToStaticBonus;
595 // If the instruction after the call, or if the normal destination of the
596 // invoke is an unreachable instruction, the function is noreturn. As such,
597 // there is little point in inlining this.
598 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
599 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
600 Bonus += InlineConstants::NoreturnPenalty;
601 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
602 Bonus += InlineConstants::NoreturnPenalty;
604 // If this function uses the coldcc calling convention, prefer not to inline
606 if (Callee->getCallingConv() == CallingConv::Cold)
607 Bonus += InlineConstants::ColdccPenalty;
609 // Add to the inline quality for properties that make the call valuable to
610 // inline. This includes factors that indicate that the result of inlining
611 // the function will be optimizable. Currently this just looks at arguments
612 // passed into the function.
614 CallSite::arg_iterator I = CS.arg_begin();
615 for (Function::arg_iterator FI = Callee->arg_begin(), FE = Callee->arg_end();
617 // Compute any constant bonus due to inlining we want to give here.
618 if (isa<Constant>(I))
619 Bonus += CountBonusForConstant(FI, cast<Constant>(I));
624 // getInlineCost - The heuristic used to determine if we should inline the
625 // function call or not.
627 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
628 SmallPtrSet<const Function*, 16> &NeverInline) {
629 return getInlineCost(CS, CS.getCalledFunction(), NeverInline);
632 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
634 SmallPtrSet<const Function*, 16> &NeverInline) {
635 Instruction *TheCall = CS.getInstruction();
636 Function *Caller = TheCall->getParent()->getParent();
638 // Don't inline functions which can be redefined at link-time to mean
639 // something else. Don't inline functions marked noinline or call sites
641 if (Callee->mayBeOverridden() ||
642 Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee) ||
644 return llvm::InlineCost::getNever();
646 // Get information about the callee.
647 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
649 // If we haven't calculated this information yet, do so now.
650 if (CalleeFI->Metrics.NumBlocks == 0)
651 CalleeFI->analyzeFunction(Callee, TD);
653 // If we should never inline this, return a huge cost.
654 if (CalleeFI->NeverInline())
655 return InlineCost::getNever();
657 // FIXME: It would be nice to kill off CalleeFI->NeverInline. Then we
658 // could move this up and avoid computing the FunctionInfo for
659 // things we are going to just return always inline for. This
660 // requires handling setjmp somewhere else, however.
661 if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
662 return InlineCost::getAlways();
664 if (CalleeFI->Metrics.usesDynamicAlloca) {
665 // Get information about the caller.
666 FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
668 // If we haven't calculated this information yet, do so now.
669 if (CallerFI.Metrics.NumBlocks == 0) {
670 CallerFI.analyzeFunction(Caller, TD);
672 // Recompute the CalleeFI pointer, getting Caller could have invalidated
674 CalleeFI = &CachedFunctionInfo[Callee];
677 // Don't inline a callee with dynamic alloca into a caller without them.
678 // Functions containing dynamic alloca's are inefficient in various ways;
679 // don't create more inefficiency.
680 if (!CallerFI.Metrics.usesDynamicAlloca)
681 return InlineCost::getNever();
684 // InlineCost - This value measures how good of an inline candidate this call
685 // site is to inline. A lower inline cost make is more likely for the call to
686 // be inlined. This value may go negative due to the fact that bonuses
687 // are negative numbers.
689 int InlineCost = getInlineSize(CS, Callee) + getInlineBonuses(CS, Callee);
690 return llvm::InlineCost::get(InlineCost);
693 // getSpecializationCost - The heuristic used to determine the code-size
694 // impact of creating a specialized version of Callee with argument
695 // SpecializedArgNo replaced by a constant.
696 InlineCost InlineCostAnalyzer::getSpecializationCost(Function *Callee,
697 SmallVectorImpl<unsigned> &SpecializedArgNos)
699 // Don't specialize functions which can be redefined at link-time to mean
701 if (Callee->mayBeOverridden())
702 return llvm::InlineCost::getNever();
704 // Get information about the callee.
705 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
707 // If we haven't calculated this information yet, do so now.
708 if (CalleeFI->Metrics.NumBlocks == 0)
709 CalleeFI->analyzeFunction(Callee, TD);
713 // Look at the original size of the callee. Each instruction counts as 5.
714 Cost += CalleeFI->Metrics.NumInsts * InlineConstants::InstrCost;
716 // Offset that with the amount of code that can be constant-folded
717 // away with the given arguments replaced by constants.
718 for (SmallVectorImpl<unsigned>::iterator an = SpecializedArgNos.begin(),
719 ae = SpecializedArgNos.end(); an != ae; ++an)
720 Cost -= CalleeFI->ArgumentWeights[*an].ConstantWeight;
722 return llvm::InlineCost::get(Cost);
725 // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
726 // higher threshold to determine if the function call should be inlined.
727 float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
728 Function *Callee = CS.getCalledFunction();
730 // Get information about the callee.
731 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
733 // If we haven't calculated this information yet, do so now.
734 if (CalleeFI.Metrics.NumBlocks == 0)
735 CalleeFI.analyzeFunction(Callee, TD);
738 // Single BB functions are often written to be inlined.
739 if (CalleeFI.Metrics.NumBlocks == 1)
742 // Be more aggressive if the function contains a good chunk (if it mades up
743 // at least 10% of the instructions) of vector instructions.
744 if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/2)
746 else if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/10)
751 /// growCachedCostInfo - update the cached cost info for Caller after Callee has
754 InlineCostAnalyzer::growCachedCostInfo(Function *Caller, Function *Callee) {
755 CodeMetrics &CallerMetrics = CachedFunctionInfo[Caller].Metrics;
757 // For small functions we prefer to recalculate the cost for better accuracy.
758 if (CallerMetrics.NumBlocks < 10 && CallerMetrics.NumInsts < 1000) {
759 resetCachedCostInfo(Caller);
763 // For large functions, we can save a lot of computation time by skipping
765 if (CallerMetrics.NumCalls > 0)
766 --CallerMetrics.NumCalls;
768 if (Callee == 0) return;
770 CodeMetrics &CalleeMetrics = CachedFunctionInfo[Callee].Metrics;
772 // If we don't have metrics for the callee, don't recalculate them just to
773 // update an approximation in the caller. Instead, just recalculate the
774 // caller info from scratch.
775 if (CalleeMetrics.NumBlocks == 0) {
776 resetCachedCostInfo(Caller);
780 // Since CalleeMetrics were already calculated, we know that the CallerMetrics
781 // reference isn't invalidated: both were in the DenseMap.
782 CallerMetrics.usesDynamicAlloca |= CalleeMetrics.usesDynamicAlloca;
784 // FIXME: If any of these three are true for the callee, the callee was
785 // not inlined into the caller, so I think they're redundant here.
786 CallerMetrics.exposesReturnsTwice |= CalleeMetrics.exposesReturnsTwice;
787 CallerMetrics.isRecursive |= CalleeMetrics.isRecursive;
788 CallerMetrics.containsIndirectBr |= CalleeMetrics.containsIndirectBr;
790 CallerMetrics.NumInsts += CalleeMetrics.NumInsts;
791 CallerMetrics.NumBlocks += CalleeMetrics.NumBlocks;
792 CallerMetrics.NumCalls += CalleeMetrics.NumCalls;
793 CallerMetrics.NumVectorInsts += CalleeMetrics.NumVectorInsts;
794 CallerMetrics.NumRets += CalleeMetrics.NumRets;
796 // analyzeBasicBlock counts each function argument as an inst.
797 if (CallerMetrics.NumInsts >= Callee->arg_size())
798 CallerMetrics.NumInsts -= Callee->arg_size();
800 CallerMetrics.NumInsts = 0;
802 // We are not updating the argument weights. We have already determined that
803 // Caller is a fairly large function, so we accept the loss of precision.
806 /// clear - empty the cache of inline costs
807 void InlineCostAnalyzer::clear() {
808 CachedFunctionInfo.clear();