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 // CountCodeReductionForConstant - Figure out an approximation for how many
166 // instructions will be constant folded if the specified value is constant.
168 unsigned CodeMetrics::CountCodeReductionForConstant(Value *V) {
169 unsigned Reduction = 0;
170 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
172 if (isa<BranchInst>(U) || isa<SwitchInst>(U)) {
173 // We will be able to eliminate all but one of the successors.
174 const TerminatorInst &TI = cast<TerminatorInst>(*U);
175 const unsigned NumSucc = TI.getNumSuccessors();
177 for (unsigned I = 0; I != NumSucc; ++I)
178 Instrs += NumBBInsts[TI.getSuccessor(I)];
179 // We don't know which blocks will be eliminated, so use the average size.
180 Reduction += InlineConstants::InstrCost*Instrs*(NumSucc-1)/NumSucc;
182 // Figure out if this instruction will be removed due to simple constant
184 Instruction &Inst = cast<Instruction>(*U);
186 // We can't constant propagate instructions which have effects or
189 // FIXME: It would be nice to capture the fact that a load from a
190 // pointer-to-constant-global is actually a *really* good thing to zap.
191 // Unfortunately, we don't know the pointer that may get propagated here,
192 // so we can't make this decision.
193 if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
194 isa<AllocaInst>(Inst))
197 bool AllOperandsConstant = true;
198 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
199 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
200 AllOperandsConstant = false;
204 if (AllOperandsConstant) {
205 // We will get to remove this instruction...
206 Reduction += InlineConstants::InstrCost;
208 // And any other instructions that use it which become constants
210 Reduction += CountCodeReductionForConstant(&Inst);
217 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
218 // the function will be if it is inlined into a context where an argument
219 // becomes an alloca.
221 unsigned CodeMetrics::CountCodeReductionForAlloca(Value *V) {
222 if (!V->getType()->isPointerTy()) return 0; // Not a pointer
223 unsigned Reduction = 0;
225 // Looking at ICmpInsts will never abort the analysis and return zero, and
226 // analyzing them is expensive, so save them for last so that we don't do
227 // extra work that we end up throwing out.
228 SmallVector<ICmpInst *, 4> ICmpInsts;
230 SmallVector<Value *, 4> Worklist;
231 Worklist.push_back(V);
233 Value *V = Worklist.pop_back_val();
234 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
236 Instruction *I = cast<Instruction>(*UI);
237 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
240 Reduction += InlineConstants::InstrCost;
241 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
244 Reduction += InlineConstants::InstrCost;
245 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
246 // If the GEP has variable indices, we won't be able to do much with it.
247 if (!GEP->hasAllConstantIndices())
249 // A non-zero GEP will likely become a mask operation after SROA.
250 if (GEP->hasAllZeroIndices())
251 Reduction += InlineConstants::InstrCost;
252 Worklist.push_back(GEP);
253 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(I)) {
254 // Track pointer through bitcasts.
255 Worklist.push_back(BCI);
256 Reduction += InlineConstants::InstrCost;
257 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
258 // SROA can handle a select of alloca iff all uses of the alloca are
259 // loads, and dereferenceable. We assume it's dereferenceable since
260 // we're told the input is an alloca.
261 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end();
263 LoadInst *LI = dyn_cast<LoadInst>(*UI);
264 if (LI == 0 || !LI->isSimple()) return 0;
266 // We don't know whether we'll be deleting the rest of the chain of
267 // instructions from the SelectInst on, because we don't know whether
268 // the other side of the select is also an alloca or not.
270 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
271 switch (II->getIntrinsicID()) {
274 case Intrinsic::memset:
275 case Intrinsic::memcpy:
276 case Intrinsic::memmove:
277 case Intrinsic::lifetime_start:
278 case Intrinsic::lifetime_end:
279 // SROA can usually chew through these intrinsics.
280 Reduction += InlineConstants::InstrCost;
283 } else if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
284 if (!isa<Constant>(ICI->getOperand(1)))
286 ICmpInsts.push_back(ICI);
288 // If there is some other strange instruction, we're not going to be
289 // able to do much if we inline this.
293 } while (!Worklist.empty());
295 while (!ICmpInsts.empty()) {
296 ICmpInst *ICI = ICmpInsts.pop_back_val();
298 // An icmp pred (alloca, C) becomes true if the predicate is true when
299 // equal and false otherwise.
300 bool Result = ICI->isTrueWhenEqual();
302 SmallVector<Instruction *, 4> Worklist;
303 Worklist.push_back(ICI);
305 Instruction *U = Worklist.pop_back_val();
306 Reduction += InlineConstants::InstrCost;
307 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
309 Instruction *I = dyn_cast<Instruction>(*UI);
310 if (!I || I->mayHaveSideEffects()) continue;
311 if (I->getNumOperands() == 1)
312 Worklist.push_back(I);
313 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
314 // If BO produces the same value as U, then the other operand is
315 // irrelevant and we can put it into the Worklist to continue
316 // deleting dead instructions. If BO produces the same value as the
317 // other operand, we can delete BO but that's it.
318 if (Result == true) {
319 if (BO->getOpcode() == Instruction::Or)
320 Worklist.push_back(I);
321 if (BO->getOpcode() == Instruction::And)
322 Reduction += InlineConstants::InstrCost;
324 if (BO->getOpcode() == Instruction::Or ||
325 BO->getOpcode() == Instruction::Xor)
326 Reduction += InlineConstants::InstrCost;
327 if (BO->getOpcode() == Instruction::And)
328 Worklist.push_back(I);
331 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
332 BasicBlock *BB = BI->getSuccessor(Result ? 0 : 1);
333 if (BB->getSinglePredecessor())
334 Reduction += InlineConstants::InstrCost * BB->size();
337 } while (!Worklist.empty());
343 /// analyzeFunction - Fill in the current structure with information gleaned
344 /// from the specified function.
345 void CodeMetrics::analyzeFunction(Function *F, const TargetData *TD) {
346 // If this function contains a call that "returns twice" (e.g., setjmp or
347 // _setjmp) and it isn't marked with "returns twice" itself, never inline it.
348 // This is a hack because we depend on the user marking their local variables
349 // as volatile if they are live across a setjmp call, and they probably
350 // won't do this in callers.
351 exposesReturnsTwice = F->callsFunctionThatReturnsTwice() &&
352 !F->hasFnAttr(Attribute::ReturnsTwice);
354 // Look at the size of the callee.
355 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
356 analyzeBasicBlock(&*BB, TD);
359 /// analyzeFunction - Fill in the current structure with information gleaned
360 /// from the specified function.
361 void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F,
362 const TargetData *TD) {
363 Metrics.analyzeFunction(F, TD);
365 // A function with exactly one return has it removed during the inlining
366 // process (see InlineFunction), so don't count it.
367 // FIXME: This knowledge should really be encoded outside of FunctionInfo.
368 if (Metrics.NumRets==1)
371 // Check out all of the arguments to the function, figuring out how much
372 // code can be eliminated if one of the arguments is a constant.
373 ArgumentWeights.reserve(F->arg_size());
374 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
375 ArgumentWeights.push_back(ArgInfo(Metrics.CountCodeReductionForConstant(I),
376 Metrics.CountCodeReductionForAlloca(I)));
379 /// NeverInline - returns true if the function should never be inlined into
381 bool InlineCostAnalyzer::FunctionInfo::NeverInline() {
382 return (Metrics.exposesReturnsTwice || Metrics.isRecursive ||
383 Metrics.containsIndirectBr);
385 // getSpecializationBonus - The heuristic used to determine the per-call
386 // performance boost for using a specialization of Callee with argument
387 // specializedArgNo replaced by a constant.
388 int InlineCostAnalyzer::getSpecializationBonus(Function *Callee,
389 SmallVectorImpl<unsigned> &SpecializedArgNos)
391 if (Callee->mayBeOverridden())
395 // If this function uses the coldcc calling convention, prefer not to
397 if (Callee->getCallingConv() == CallingConv::Cold)
398 Bonus -= InlineConstants::ColdccPenalty;
400 // Get information about the callee.
401 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
403 // If we haven't calculated this information yet, do so now.
404 if (CalleeFI->Metrics.NumBlocks == 0)
405 CalleeFI->analyzeFunction(Callee, TD);
409 for (Function::arg_iterator I = Callee->arg_begin(), E = Callee->arg_end();
410 I != E; ++I, ++ArgNo)
411 if (ArgNo == SpecializedArgNos[i]) {
413 Bonus += CountBonusForConstant(I);
416 // Calls usually take a long time, so they make the specialization gain
418 Bonus -= CalleeFI->Metrics.NumCalls * InlineConstants::CallPenalty;
423 // ConstantFunctionBonus - Figure out how much of a bonus we can get for
424 // possibly devirtualizing a function. We'll subtract the size of the function
425 // we may wish to inline from the indirect call bonus providing a limit on
426 // growth. Leave an upper limit of 0 for the bonus - we don't want to penalize
427 // inlining because we decide we don't want to give a bonus for
429 int InlineCostAnalyzer::ConstantFunctionBonus(CallSite CS, Constant *C) {
431 // This could just be NULL.
434 Function *F = dyn_cast<Function>(C);
437 int Bonus = InlineConstants::IndirectCallBonus + getInlineSize(CS, F);
438 return (Bonus > 0) ? 0 : Bonus;
441 // CountBonusForConstant - Figure out an approximation for how much per-call
442 // performance boost we can expect if the specified value is constant.
443 int InlineCostAnalyzer::CountBonusForConstant(Value *V, Constant *C) {
445 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
447 if (CallInst *CI = dyn_cast<CallInst>(U)) {
448 // Turning an indirect call into a direct call is a BIG win
449 if (CI->getCalledValue() == V)
450 Bonus += ConstantFunctionBonus(CallSite(CI), C);
451 } else if (InvokeInst *II = dyn_cast<InvokeInst>(U)) {
452 // Turning an indirect call into a direct call is a BIG win
453 if (II->getCalledValue() == V)
454 Bonus += ConstantFunctionBonus(CallSite(II), C);
456 // FIXME: Eliminating conditional branches and switches should
457 // also yield a per-call performance boost.
459 // Figure out the bonuses that wll accrue due to simple constant
461 Instruction &Inst = cast<Instruction>(*U);
463 // We can't constant propagate instructions which have effects or
466 // FIXME: It would be nice to capture the fact that a load from a
467 // pointer-to-constant-global is actually a *really* good thing to zap.
468 // Unfortunately, we don't know the pointer that may get propagated here,
469 // so we can't make this decision.
470 if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
471 isa<AllocaInst>(Inst))
474 bool AllOperandsConstant = true;
475 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
476 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
477 AllOperandsConstant = false;
481 if (AllOperandsConstant)
482 Bonus += CountBonusForConstant(&Inst);
489 int InlineCostAnalyzer::getInlineSize(CallSite CS, Function *Callee) {
490 // Get information about the callee.
491 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
493 // If we haven't calculated this information yet, do so now.
494 if (CalleeFI->Metrics.NumBlocks == 0)
495 CalleeFI->analyzeFunction(Callee, TD);
497 // InlineCost - This value measures how good of an inline candidate this call
498 // site is to inline. A lower inline cost make is more likely for the call to
499 // be inlined. This value may go negative.
503 // Compute any size reductions we can expect due to arguments being passed into
507 CallSite::arg_iterator I = CS.arg_begin();
508 for (Function::arg_iterator FI = Callee->arg_begin(), FE = Callee->arg_end();
509 FI != FE; ++I, ++FI, ++ArgNo) {
511 // If an alloca is passed in, inlining this function is likely to allow
512 // significant future optimization possibilities (like scalar promotion, and
513 // scalarization), so encourage the inlining of the function.
515 if (isa<AllocaInst>(I))
516 InlineCost -= CalleeFI->ArgumentWeights[ArgNo].AllocaWeight;
518 // If this is a constant being passed into the function, use the argument
519 // weights calculated for the callee to determine how much will be folded
520 // away with this information.
521 else if (isa<Constant>(I))
522 InlineCost -= CalleeFI->ArgumentWeights[ArgNo].ConstantWeight;
525 // Each argument passed in has a cost at both the caller and the callee
526 // sides. Measurements show that each argument costs about the same as an
528 InlineCost -= (CS.arg_size() * InlineConstants::InstrCost);
530 // Now that we have considered all of the factors that make the call site more
531 // likely to be inlined, look at factors that make us not want to inline it.
533 // Calls usually take a long time, so they make the inlining gain smaller.
534 InlineCost += CalleeFI->Metrics.NumCalls * InlineConstants::CallPenalty;
536 // Look at the size of the callee. Each instruction counts as 5.
537 InlineCost += CalleeFI->Metrics.NumInsts * InlineConstants::InstrCost;
542 int InlineCostAnalyzer::getInlineBonuses(CallSite CS, Function *Callee) {
543 // Get information about the callee.
544 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
546 // If we haven't calculated this information yet, do so now.
547 if (CalleeFI->Metrics.NumBlocks == 0)
548 CalleeFI->analyzeFunction(Callee, TD);
550 bool isDirectCall = CS.getCalledFunction() == Callee;
551 Instruction *TheCall = CS.getInstruction();
554 // If there is only one call of the function, and it has internal linkage,
555 // make it almost guaranteed to be inlined.
557 if (Callee->hasLocalLinkage() && Callee->hasOneUse() && isDirectCall)
558 Bonus += InlineConstants::LastCallToStaticBonus;
560 // If the instruction after the call, or if the normal destination of the
561 // invoke is an unreachable instruction, the function is noreturn. As such,
562 // there is little point in inlining this.
563 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
564 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
565 Bonus += InlineConstants::NoreturnPenalty;
566 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
567 Bonus += InlineConstants::NoreturnPenalty;
569 // If this function uses the coldcc calling convention, prefer not to inline
571 if (Callee->getCallingConv() == CallingConv::Cold)
572 Bonus += InlineConstants::ColdccPenalty;
574 // Add to the inline quality for properties that make the call valuable to
575 // inline. This includes factors that indicate that the result of inlining
576 // the function will be optimizable. Currently this just looks at arguments
577 // passed into the function.
579 CallSite::arg_iterator I = CS.arg_begin();
580 for (Function::arg_iterator FI = Callee->arg_begin(), FE = Callee->arg_end();
582 // Compute any constant bonus due to inlining we want to give here.
583 if (isa<Constant>(I))
584 Bonus += CountBonusForConstant(FI, cast<Constant>(I));
589 // getInlineCost - The heuristic used to determine if we should inline the
590 // function call or not.
592 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
593 SmallPtrSet<const Function*, 16> &NeverInline) {
594 return getInlineCost(CS, CS.getCalledFunction(), NeverInline);
597 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
599 SmallPtrSet<const Function*, 16> &NeverInline) {
600 Instruction *TheCall = CS.getInstruction();
601 Function *Caller = TheCall->getParent()->getParent();
603 // Don't inline functions which can be redefined at link-time to mean
604 // something else. Don't inline functions marked noinline or call sites
606 if (Callee->mayBeOverridden() ||
607 Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee) ||
609 return llvm::InlineCost::getNever();
611 // Get information about the callee.
612 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
614 // If we haven't calculated this information yet, do so now.
615 if (CalleeFI->Metrics.NumBlocks == 0)
616 CalleeFI->analyzeFunction(Callee, TD);
618 // If we should never inline this, return a huge cost.
619 if (CalleeFI->NeverInline())
620 return InlineCost::getNever();
622 // FIXME: It would be nice to kill off CalleeFI->NeverInline. Then we
623 // could move this up and avoid computing the FunctionInfo for
624 // things we are going to just return always inline for. This
625 // requires handling setjmp somewhere else, however.
626 if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
627 return InlineCost::getAlways();
629 if (CalleeFI->Metrics.usesDynamicAlloca) {
630 // Get information about the caller.
631 FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
633 // If we haven't calculated this information yet, do so now.
634 if (CallerFI.Metrics.NumBlocks == 0) {
635 CallerFI.analyzeFunction(Caller, TD);
637 // Recompute the CalleeFI pointer, getting Caller could have invalidated
639 CalleeFI = &CachedFunctionInfo[Callee];
642 // Don't inline a callee with dynamic alloca into a caller without them.
643 // Functions containing dynamic alloca's are inefficient in various ways;
644 // don't create more inefficiency.
645 if (!CallerFI.Metrics.usesDynamicAlloca)
646 return InlineCost::getNever();
649 // InlineCost - This value measures how good of an inline candidate this call
650 // site is to inline. A lower inline cost make is more likely for the call to
651 // be inlined. This value may go negative due to the fact that bonuses
652 // are negative numbers.
654 int InlineCost = getInlineSize(CS, Callee) + getInlineBonuses(CS, Callee);
655 return llvm::InlineCost::get(InlineCost);
658 // getSpecializationCost - The heuristic used to determine the code-size
659 // impact of creating a specialized version of Callee with argument
660 // SpecializedArgNo replaced by a constant.
661 InlineCost InlineCostAnalyzer::getSpecializationCost(Function *Callee,
662 SmallVectorImpl<unsigned> &SpecializedArgNos)
664 // Don't specialize functions which can be redefined at link-time to mean
666 if (Callee->mayBeOverridden())
667 return llvm::InlineCost::getNever();
669 // Get information about the callee.
670 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
672 // If we haven't calculated this information yet, do so now.
673 if (CalleeFI->Metrics.NumBlocks == 0)
674 CalleeFI->analyzeFunction(Callee, TD);
678 // Look at the original size of the callee. Each instruction counts as 5.
679 Cost += CalleeFI->Metrics.NumInsts * InlineConstants::InstrCost;
681 // Offset that with the amount of code that can be constant-folded
682 // away with the given arguments replaced by constants.
683 for (SmallVectorImpl<unsigned>::iterator an = SpecializedArgNos.begin(),
684 ae = SpecializedArgNos.end(); an != ae; ++an)
685 Cost -= CalleeFI->ArgumentWeights[*an].ConstantWeight;
687 return llvm::InlineCost::get(Cost);
690 // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
691 // higher threshold to determine if the function call should be inlined.
692 float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
693 Function *Callee = CS.getCalledFunction();
695 // Get information about the callee.
696 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
698 // If we haven't calculated this information yet, do so now.
699 if (CalleeFI.Metrics.NumBlocks == 0)
700 CalleeFI.analyzeFunction(Callee, TD);
703 // Single BB functions are often written to be inlined.
704 if (CalleeFI.Metrics.NumBlocks == 1)
707 // Be more aggressive if the function contains a good chunk (if it mades up
708 // at least 10% of the instructions) of vector instructions.
709 if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/2)
711 else if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/10)
716 /// growCachedCostInfo - update the cached cost info for Caller after Callee has
719 InlineCostAnalyzer::growCachedCostInfo(Function *Caller, Function *Callee) {
720 CodeMetrics &CallerMetrics = CachedFunctionInfo[Caller].Metrics;
722 // For small functions we prefer to recalculate the cost for better accuracy.
723 if (CallerMetrics.NumBlocks < 10 && CallerMetrics.NumInsts < 1000) {
724 resetCachedCostInfo(Caller);
728 // For large functions, we can save a lot of computation time by skipping
730 if (CallerMetrics.NumCalls > 0)
731 --CallerMetrics.NumCalls;
733 if (Callee == 0) return;
735 CodeMetrics &CalleeMetrics = CachedFunctionInfo[Callee].Metrics;
737 // If we don't have metrics for the callee, don't recalculate them just to
738 // update an approximation in the caller. Instead, just recalculate the
739 // caller info from scratch.
740 if (CalleeMetrics.NumBlocks == 0) {
741 resetCachedCostInfo(Caller);
745 // Since CalleeMetrics were already calculated, we know that the CallerMetrics
746 // reference isn't invalidated: both were in the DenseMap.
747 CallerMetrics.usesDynamicAlloca |= CalleeMetrics.usesDynamicAlloca;
749 // FIXME: If any of these three are true for the callee, the callee was
750 // not inlined into the caller, so I think they're redundant here.
751 CallerMetrics.exposesReturnsTwice |= CalleeMetrics.exposesReturnsTwice;
752 CallerMetrics.isRecursive |= CalleeMetrics.isRecursive;
753 CallerMetrics.containsIndirectBr |= CalleeMetrics.containsIndirectBr;
755 CallerMetrics.NumInsts += CalleeMetrics.NumInsts;
756 CallerMetrics.NumBlocks += CalleeMetrics.NumBlocks;
757 CallerMetrics.NumCalls += CalleeMetrics.NumCalls;
758 CallerMetrics.NumVectorInsts += CalleeMetrics.NumVectorInsts;
759 CallerMetrics.NumRets += CalleeMetrics.NumRets;
761 // analyzeBasicBlock counts each function argument as an inst.
762 if (CallerMetrics.NumInsts >= Callee->arg_size())
763 CallerMetrics.NumInsts -= Callee->arg_size();
765 CallerMetrics.NumInsts = 0;
767 // We are not updating the argument weights. We have already determined that
768 // Caller is a fairly large function, so we accept the loss of precision.
771 /// clear - empty the cache of inline costs
772 void InlineCostAnalyzer::clear() {
773 CachedFunctionInfo.clear();