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 SmallVector<Value *, 4> Worklist;
226 Worklist.push_back(V);
228 Value *V = Worklist.pop_back_val();
229 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
231 Instruction *I = cast<Instruction>(*UI);
232 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
235 Reduction += InlineConstants::InstrCost;
236 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
239 Reduction += InlineConstants::InstrCost;
240 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
241 // If the GEP has variable indices, we won't be able to do much with it.
242 if (!GEP->hasAllConstantIndices())
244 // A non-zero GEP will likely become a mask operation after SROA.
245 if (GEP->hasAllZeroIndices())
246 Reduction += InlineConstants::InstrCost;
247 Worklist.push_back(GEP);
248 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(I)) {
249 // Track pointer through bitcasts.
250 Worklist.push_back(BCI);
251 Reduction += InlineConstants::InstrCost;
252 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
253 // SROA can handle a select of alloca iff all uses of the alloca are
254 // loads, and dereferenceable. We assume it's dereferenceable since
255 // we're told the input is an alloca.
256 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end();
258 LoadInst *LI = dyn_cast<LoadInst>(*UI);
259 if (LI == 0 || !LI->isSimple()) return 0;
261 // We don't know whether we'll be deleting the rest of the chain of
262 // instructions from the SelectInst on, because we don't know whether
263 // the other side of the select is also an alloca or not.
265 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
266 switch (II->getIntrinsicID()) {
269 case Intrinsic::memset:
270 case Intrinsic::memcpy:
271 case Intrinsic::memmove:
272 case Intrinsic::lifetime_start:
273 case Intrinsic::lifetime_end:
274 // SROA can usually chew through these intrinsics.
275 Reduction += InlineConstants::InstrCost;
279 // If there is some other strange instruction, we're not going to be
280 // able to do much if we inline this.
284 } while (!Worklist.empty());
289 /// analyzeFunction - Fill in the current structure with information gleaned
290 /// from the specified function.
291 void CodeMetrics::analyzeFunction(Function *F, const TargetData *TD) {
292 // If this function contains a call that "returns twice" (e.g., setjmp or
293 // _setjmp) and it isn't marked with "returns twice" itself, never inline it.
294 // This is a hack because we depend on the user marking their local variables
295 // as volatile if they are live across a setjmp call, and they probably
296 // won't do this in callers.
297 exposesReturnsTwice = F->callsFunctionThatReturnsTwice() &&
298 !F->hasFnAttr(Attribute::ReturnsTwice);
300 // Look at the size of the callee.
301 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
302 analyzeBasicBlock(&*BB, TD);
305 /// analyzeFunction - Fill in the current structure with information gleaned
306 /// from the specified function.
307 void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F,
308 const TargetData *TD) {
309 Metrics.analyzeFunction(F, TD);
311 // A function with exactly one return has it removed during the inlining
312 // process (see InlineFunction), so don't count it.
313 // FIXME: This knowledge should really be encoded outside of FunctionInfo.
314 if (Metrics.NumRets==1)
317 // Check out all of the arguments to the function, figuring out how much
318 // code can be eliminated if one of the arguments is a constant.
319 ArgumentWeights.reserve(F->arg_size());
320 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
321 ArgumentWeights.push_back(ArgInfo(Metrics.CountCodeReductionForConstant(I),
322 Metrics.CountCodeReductionForAlloca(I)));
325 /// NeverInline - returns true if the function should never be inlined into
327 bool InlineCostAnalyzer::FunctionInfo::NeverInline() {
328 return (Metrics.exposesReturnsTwice || Metrics.isRecursive ||
329 Metrics.containsIndirectBr);
331 // getSpecializationBonus - The heuristic used to determine the per-call
332 // performance boost for using a specialization of Callee with argument
333 // specializedArgNo replaced by a constant.
334 int InlineCostAnalyzer::getSpecializationBonus(Function *Callee,
335 SmallVectorImpl<unsigned> &SpecializedArgNos)
337 if (Callee->mayBeOverridden())
341 // If this function uses the coldcc calling convention, prefer not to
343 if (Callee->getCallingConv() == CallingConv::Cold)
344 Bonus -= InlineConstants::ColdccPenalty;
346 // Get information about the callee.
347 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
349 // If we haven't calculated this information yet, do so now.
350 if (CalleeFI->Metrics.NumBlocks == 0)
351 CalleeFI->analyzeFunction(Callee, TD);
355 for (Function::arg_iterator I = Callee->arg_begin(), E = Callee->arg_end();
356 I != E; ++I, ++ArgNo)
357 if (ArgNo == SpecializedArgNos[i]) {
359 Bonus += CountBonusForConstant(I);
362 // Calls usually take a long time, so they make the specialization gain
364 Bonus -= CalleeFI->Metrics.NumCalls * InlineConstants::CallPenalty;
369 // ConstantFunctionBonus - Figure out how much of a bonus we can get for
370 // possibly devirtualizing a function. We'll subtract the size of the function
371 // we may wish to inline from the indirect call bonus providing a limit on
372 // growth. Leave an upper limit of 0 for the bonus - we don't want to penalize
373 // inlining because we decide we don't want to give a bonus for
375 int InlineCostAnalyzer::ConstantFunctionBonus(CallSite CS, Constant *C) {
377 // This could just be NULL.
380 Function *F = dyn_cast<Function>(C);
383 int Bonus = InlineConstants::IndirectCallBonus + getInlineSize(CS, F);
384 return (Bonus > 0) ? 0 : Bonus;
387 // CountBonusForConstant - Figure out an approximation for how much per-call
388 // performance boost we can expect if the specified value is constant.
389 int InlineCostAnalyzer::CountBonusForConstant(Value *V, Constant *C) {
391 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
393 if (CallInst *CI = dyn_cast<CallInst>(U)) {
394 // Turning an indirect call into a direct call is a BIG win
395 if (CI->getCalledValue() == V)
396 Bonus += ConstantFunctionBonus(CallSite(CI), C);
397 } else if (InvokeInst *II = dyn_cast<InvokeInst>(U)) {
398 // Turning an indirect call into a direct call is a BIG win
399 if (II->getCalledValue() == V)
400 Bonus += ConstantFunctionBonus(CallSite(II), C);
402 // FIXME: Eliminating conditional branches and switches should
403 // also yield a per-call performance boost.
405 // Figure out the bonuses that wll accrue due to simple constant
407 Instruction &Inst = cast<Instruction>(*U);
409 // We can't constant propagate instructions which have effects or
412 // FIXME: It would be nice to capture the fact that a load from a
413 // pointer-to-constant-global is actually a *really* good thing to zap.
414 // Unfortunately, we don't know the pointer that may get propagated here,
415 // so we can't make this decision.
416 if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
417 isa<AllocaInst>(Inst))
420 bool AllOperandsConstant = true;
421 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
422 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
423 AllOperandsConstant = false;
427 if (AllOperandsConstant)
428 Bonus += CountBonusForConstant(&Inst);
435 int InlineCostAnalyzer::getInlineSize(CallSite CS, Function *Callee) {
436 // Get information about the callee.
437 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
439 // If we haven't calculated this information yet, do so now.
440 if (CalleeFI->Metrics.NumBlocks == 0)
441 CalleeFI->analyzeFunction(Callee, TD);
443 // InlineCost - This value measures how good of an inline candidate this call
444 // site is to inline. A lower inline cost make is more likely for the call to
445 // be inlined. This value may go negative.
449 // Compute any size reductions we can expect due to arguments being passed into
453 CallSite::arg_iterator I = CS.arg_begin();
454 for (Function::arg_iterator FI = Callee->arg_begin(), FE = Callee->arg_end();
455 FI != FE; ++I, ++FI, ++ArgNo) {
457 // If an alloca is passed in, inlining this function is likely to allow
458 // significant future optimization possibilities (like scalar promotion, and
459 // scalarization), so encourage the inlining of the function.
461 if (isa<AllocaInst>(I))
462 InlineCost -= CalleeFI->ArgumentWeights[ArgNo].AllocaWeight;
464 // If this is a constant being passed into the function, use the argument
465 // weights calculated for the callee to determine how much will be folded
466 // away with this information.
467 else if (isa<Constant>(I))
468 InlineCost -= CalleeFI->ArgumentWeights[ArgNo].ConstantWeight;
471 // Each argument passed in has a cost at both the caller and the callee
472 // sides. Measurements show that each argument costs about the same as an
474 InlineCost -= (CS.arg_size() * InlineConstants::InstrCost);
476 // Now that we have considered all of the factors that make the call site more
477 // likely to be inlined, look at factors that make us not want to inline it.
479 // Calls usually take a long time, so they make the inlining gain smaller.
480 InlineCost += CalleeFI->Metrics.NumCalls * InlineConstants::CallPenalty;
482 // Look at the size of the callee. Each instruction counts as 5.
483 InlineCost += CalleeFI->Metrics.NumInsts * InlineConstants::InstrCost;
488 int InlineCostAnalyzer::getInlineBonuses(CallSite CS, Function *Callee) {
489 // Get information about the callee.
490 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
492 // If we haven't calculated this information yet, do so now.
493 if (CalleeFI->Metrics.NumBlocks == 0)
494 CalleeFI->analyzeFunction(Callee, TD);
496 bool isDirectCall = CS.getCalledFunction() == Callee;
497 Instruction *TheCall = CS.getInstruction();
500 // If there is only one call of the function, and it has internal linkage,
501 // make it almost guaranteed to be inlined.
503 if (Callee->hasLocalLinkage() && Callee->hasOneUse() && isDirectCall)
504 Bonus += InlineConstants::LastCallToStaticBonus;
506 // If the instruction after the call, or if the normal destination of the
507 // invoke is an unreachable instruction, the function is noreturn. As such,
508 // there is little point in inlining this.
509 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
510 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
511 Bonus += InlineConstants::NoreturnPenalty;
512 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
513 Bonus += InlineConstants::NoreturnPenalty;
515 // If this function uses the coldcc calling convention, prefer not to inline
517 if (Callee->getCallingConv() == CallingConv::Cold)
518 Bonus += InlineConstants::ColdccPenalty;
520 // Add to the inline quality for properties that make the call valuable to
521 // inline. This includes factors that indicate that the result of inlining
522 // the function will be optimizable. Currently this just looks at arguments
523 // passed into the function.
525 CallSite::arg_iterator I = CS.arg_begin();
526 for (Function::arg_iterator FI = Callee->arg_begin(), FE = Callee->arg_end();
528 // Compute any constant bonus due to inlining we want to give here.
529 if (isa<Constant>(I))
530 Bonus += CountBonusForConstant(FI, cast<Constant>(I));
535 // getInlineCost - The heuristic used to determine if we should inline the
536 // function call or not.
538 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
539 SmallPtrSet<const Function*, 16> &NeverInline) {
540 return getInlineCost(CS, CS.getCalledFunction(), NeverInline);
543 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
545 SmallPtrSet<const Function*, 16> &NeverInline) {
546 Instruction *TheCall = CS.getInstruction();
547 Function *Caller = TheCall->getParent()->getParent();
549 // Don't inline functions which can be redefined at link-time to mean
550 // something else. Don't inline functions marked noinline or call sites
552 if (Callee->mayBeOverridden() ||
553 Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee) ||
555 return llvm::InlineCost::getNever();
557 // Get information about the callee.
558 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
560 // If we haven't calculated this information yet, do so now.
561 if (CalleeFI->Metrics.NumBlocks == 0)
562 CalleeFI->analyzeFunction(Callee, TD);
564 // If we should never inline this, return a huge cost.
565 if (CalleeFI->NeverInline())
566 return InlineCost::getNever();
568 // FIXME: It would be nice to kill off CalleeFI->NeverInline. Then we
569 // could move this up and avoid computing the FunctionInfo for
570 // things we are going to just return always inline for. This
571 // requires handling setjmp somewhere else, however.
572 if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
573 return InlineCost::getAlways();
575 if (CalleeFI->Metrics.usesDynamicAlloca) {
576 // Get information about the caller.
577 FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
579 // If we haven't calculated this information yet, do so now.
580 if (CallerFI.Metrics.NumBlocks == 0) {
581 CallerFI.analyzeFunction(Caller, TD);
583 // Recompute the CalleeFI pointer, getting Caller could have invalidated
585 CalleeFI = &CachedFunctionInfo[Callee];
588 // Don't inline a callee with dynamic alloca into a caller without them.
589 // Functions containing dynamic alloca's are inefficient in various ways;
590 // don't create more inefficiency.
591 if (!CallerFI.Metrics.usesDynamicAlloca)
592 return InlineCost::getNever();
595 // InlineCost - This value measures how good of an inline candidate this call
596 // site is to inline. A lower inline cost make is more likely for the call to
597 // be inlined. This value may go negative due to the fact that bonuses
598 // are negative numbers.
600 int InlineCost = getInlineSize(CS, Callee) + getInlineBonuses(CS, Callee);
601 return llvm::InlineCost::get(InlineCost);
604 // getSpecializationCost - The heuristic used to determine the code-size
605 // impact of creating a specialized version of Callee with argument
606 // SpecializedArgNo replaced by a constant.
607 InlineCost InlineCostAnalyzer::getSpecializationCost(Function *Callee,
608 SmallVectorImpl<unsigned> &SpecializedArgNos)
610 // Don't specialize functions which can be redefined at link-time to mean
612 if (Callee->mayBeOverridden())
613 return llvm::InlineCost::getNever();
615 // Get information about the callee.
616 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
618 // If we haven't calculated this information yet, do so now.
619 if (CalleeFI->Metrics.NumBlocks == 0)
620 CalleeFI->analyzeFunction(Callee, TD);
624 // Look at the original size of the callee. Each instruction counts as 5.
625 Cost += CalleeFI->Metrics.NumInsts * InlineConstants::InstrCost;
627 // Offset that with the amount of code that can be constant-folded
628 // away with the given arguments replaced by constants.
629 for (SmallVectorImpl<unsigned>::iterator an = SpecializedArgNos.begin(),
630 ae = SpecializedArgNos.end(); an != ae; ++an)
631 Cost -= CalleeFI->ArgumentWeights[*an].ConstantWeight;
633 return llvm::InlineCost::get(Cost);
636 // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
637 // higher threshold to determine if the function call should be inlined.
638 float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
639 Function *Callee = CS.getCalledFunction();
641 // Get information about the callee.
642 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
644 // If we haven't calculated this information yet, do so now.
645 if (CalleeFI.Metrics.NumBlocks == 0)
646 CalleeFI.analyzeFunction(Callee, TD);
649 // Single BB functions are often written to be inlined.
650 if (CalleeFI.Metrics.NumBlocks == 1)
653 // Be more aggressive if the function contains a good chunk (if it mades up
654 // at least 10% of the instructions) of vector instructions.
655 if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/2)
657 else if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/10)
662 /// growCachedCostInfo - update the cached cost info for Caller after Callee has
665 InlineCostAnalyzer::growCachedCostInfo(Function *Caller, Function *Callee) {
666 CodeMetrics &CallerMetrics = CachedFunctionInfo[Caller].Metrics;
668 // For small functions we prefer to recalculate the cost for better accuracy.
669 if (CallerMetrics.NumBlocks < 10 && CallerMetrics.NumInsts < 1000) {
670 resetCachedCostInfo(Caller);
674 // For large functions, we can save a lot of computation time by skipping
676 if (CallerMetrics.NumCalls > 0)
677 --CallerMetrics.NumCalls;
679 if (Callee == 0) return;
681 CodeMetrics &CalleeMetrics = CachedFunctionInfo[Callee].Metrics;
683 // If we don't have metrics for the callee, don't recalculate them just to
684 // update an approximation in the caller. Instead, just recalculate the
685 // caller info from scratch.
686 if (CalleeMetrics.NumBlocks == 0) {
687 resetCachedCostInfo(Caller);
691 // Since CalleeMetrics were already calculated, we know that the CallerMetrics
692 // reference isn't invalidated: both were in the DenseMap.
693 CallerMetrics.usesDynamicAlloca |= CalleeMetrics.usesDynamicAlloca;
695 // FIXME: If any of these three are true for the callee, the callee was
696 // not inlined into the caller, so I think they're redundant here.
697 CallerMetrics.exposesReturnsTwice |= CalleeMetrics.exposesReturnsTwice;
698 CallerMetrics.isRecursive |= CalleeMetrics.isRecursive;
699 CallerMetrics.containsIndirectBr |= CalleeMetrics.containsIndirectBr;
701 CallerMetrics.NumInsts += CalleeMetrics.NumInsts;
702 CallerMetrics.NumBlocks += CalleeMetrics.NumBlocks;
703 CallerMetrics.NumCalls += CalleeMetrics.NumCalls;
704 CallerMetrics.NumVectorInsts += CalleeMetrics.NumVectorInsts;
705 CallerMetrics.NumRets += CalleeMetrics.NumRets;
707 // analyzeBasicBlock counts each function argument as an inst.
708 if (CallerMetrics.NumInsts >= Callee->arg_size())
709 CallerMetrics.NumInsts -= Callee->arg_size();
711 CallerMetrics.NumInsts = 0;
713 // We are not updating the argument weights. We have already determined that
714 // Caller is a fairly large function, so we accept the loss of precision.
717 /// clear - empty the cache of inline costs
718 void InlineCostAnalyzer::clear() {
719 CachedFunctionInfo.clear();