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/ADT/SmallPtrSet.h"
21 /// callIsSmall - If a call is likely to lower to a single target instruction,
22 /// or is otherwise deemed small return true.
23 /// TODO: Perhaps calls like memcpy, strcpy, etc?
24 bool llvm::callIsSmall(const Function *F) {
27 if (F->hasLocalLinkage()) return false;
29 if (!F->hasName()) return false;
31 StringRef Name = F->getName();
33 // These will all likely lower to a single selection DAG node.
34 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
35 Name == "fabs" || Name == "fabsf" || Name == "fabsl" ||
36 Name == "sin" || Name == "sinf" || Name == "sinl" ||
37 Name == "cos" || Name == "cosf" || Name == "cosl" ||
38 Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl" )
41 // These are all likely to be optimized into something smaller.
42 if (Name == "pow" || Name == "powf" || Name == "powl" ||
43 Name == "exp2" || Name == "exp2l" || Name == "exp2f" ||
44 Name == "floor" || Name == "floorf" || Name == "ceil" ||
45 Name == "round" || Name == "ffs" || Name == "ffsl" ||
46 Name == "abs" || Name == "labs" || Name == "llabs")
52 /// analyzeBasicBlock - Fill in the current structure with information gleaned
53 /// from the specified block.
54 void CodeMetrics::analyzeBasicBlock(const BasicBlock *BB) {
56 unsigned NumInstsBeforeThisBB = NumInsts;
57 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
59 if (isa<PHINode>(II)) continue; // PHI nodes don't count.
61 // Special handling for calls.
62 if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
63 if (isa<DbgInfoIntrinsic>(II))
64 continue; // Debug intrinsics don't count as size.
66 ImmutableCallSite CS(cast<Instruction>(II));
68 // If this function contains a call to setjmp or _setjmp, never inline
69 // it. This is a hack because we depend on the user marking their local
70 // variables as volatile if they are live across a setjmp call, and they
71 // probably won't do this in callers.
72 if (const Function *F = CS.getCalledFunction()) {
73 // If a function is both internal and has a single use, then it is
74 // extremely likely to get inlined in the future (it was probably
75 // exposed by an interleaved devirtualization pass).
76 if (F->hasInternalLinkage() && F->hasOneUse())
77 ++NumInlineCandidates;
79 if (F->isDeclaration() &&
80 (F->getName() == "setjmp" || F->getName() == "_setjmp"))
83 // If this call is to function itself, then the function is recursive.
84 // Inlining it into other functions is a bad idea, because this is
85 // basically just a form of loop peeling, and our metrics aren't useful
87 if (F == BB->getParent())
91 if (!isa<IntrinsicInst>(II) && !callIsSmall(CS.getCalledFunction())) {
92 // Each argument to a call takes on average one instruction to set up.
93 NumInsts += CS.arg_size();
95 // We don't want inline asm to count as a call - that would prevent loop
96 // unrolling. The argument setup cost is still real, though.
97 if (!isa<InlineAsm>(CS.getCalledValue()))
102 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
103 if (!AI->isStaticAlloca())
104 this->usesDynamicAlloca = true;
107 if (isa<ExtractElementInst>(II) || II->getType()->isVectorTy())
110 if (const CastInst *CI = dyn_cast<CastInst>(II)) {
111 // Noop casts, including ptr <-> int, don't count.
112 if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
113 isa<PtrToIntInst>(CI))
115 // Result of a cmp instruction is often extended (to be used by other
116 // cmp instructions, logical or return instructions). These are usually
117 // nop on most sane targets.
118 if (isa<CmpInst>(CI->getOperand(0)))
120 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(II)){
121 // If a GEP has all constant indices, it will probably be folded with
123 if (GEPI->hasAllConstantIndices())
130 if (isa<ReturnInst>(BB->getTerminator()))
133 // We never want to inline functions that contain an indirectbr. This is
134 // incorrect because all the blockaddress's (in static global initializers
135 // for example) would be referring to the original function, and this indirect
136 // jump would jump from the inlined copy of the function into the original
137 // function which is extremely undefined behavior.
138 if (isa<IndirectBrInst>(BB->getTerminator()))
139 containsIndirectBr = true;
141 // Remember NumInsts for this BB.
142 NumBBInsts[BB] = NumInsts - NumInstsBeforeThisBB;
145 // CountBonusForConstant - Figure out an approximation for how much per-call
146 // performance boost we can expect if the specified value is constant.
147 unsigned CodeMetrics::CountBonusForConstant(Value *V) {
149 bool indirectCallBonus = false;
150 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
152 if (CallInst *CI = dyn_cast<CallInst>(U)) {
153 // Turning an indirect call into a direct call is a BIG win
154 if (CI->getCalledValue() == V)
155 indirectCallBonus = true;
157 else if (InvokeInst *II = dyn_cast<InvokeInst>(U)) {
158 // Turning an indirect call into a direct call is a BIG win
159 if (II->getCalledValue() == V)
160 indirectCallBonus = true;
162 // FIXME: Eliminating conditional branches and switches should
163 // also yield a per-call performance boost.
165 // Figure out the bonuses that wll accrue due to simple constant
167 Instruction &Inst = cast<Instruction>(*U);
169 // We can't constant propagate instructions which have effects or
172 // FIXME: It would be nice to capture the fact that a load from a
173 // pointer-to-constant-global is actually a *really* good thing to zap.
174 // Unfortunately, we don't know the pointer that may get propagated here,
175 // so we can't make this decision.
176 if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
177 isa<AllocaInst>(Inst))
180 bool AllOperandsConstant = true;
181 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
182 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
183 AllOperandsConstant = false;
187 if (AllOperandsConstant)
188 Bonus += CountBonusForConstant(&Inst);
192 // FIXME: The only reason we're applying the bonus once is while it's great
193 // to devirtualize calls the magnitude of the bonus x number of call sites
194 // can lead to a huge code explosion when we prefer to inline 1000 instruction
195 // functions that have 10 call sites. This should be made a function of the
196 // estimated inline penalty/benefit + the indirect call bonus.
197 if (indirectCallBonus) Bonus += InlineConstants::IndirectCallBonus;
203 // CountCodeReductionForConstant - Figure out an approximation for how many
204 // instructions will be constant folded if the specified value is constant.
206 unsigned CodeMetrics::CountCodeReductionForConstant(Value *V) {
207 unsigned Reduction = 0;
208 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
210 if (isa<BranchInst>(U) || isa<SwitchInst>(U)) {
211 // We will be able to eliminate all but one of the successors.
212 const TerminatorInst &TI = cast<TerminatorInst>(*U);
213 const unsigned NumSucc = TI.getNumSuccessors();
215 for (unsigned I = 0; I != NumSucc; ++I)
216 Instrs += NumBBInsts[TI.getSuccessor(I)];
217 // We don't know which blocks will be eliminated, so use the average size.
218 Reduction += InlineConstants::InstrCost*Instrs*(NumSucc-1)/NumSucc;
220 // Figure out if this instruction will be removed due to simple constant
222 Instruction &Inst = cast<Instruction>(*U);
224 // We can't constant propagate instructions which have effects or
227 // FIXME: It would be nice to capture the fact that a load from a
228 // pointer-to-constant-global is actually a *really* good thing to zap.
229 // Unfortunately, we don't know the pointer that may get propagated here,
230 // so we can't make this decision.
231 if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
232 isa<AllocaInst>(Inst))
235 bool AllOperandsConstant = true;
236 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
237 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
238 AllOperandsConstant = false;
242 if (AllOperandsConstant) {
243 // We will get to remove this instruction...
244 Reduction += InlineConstants::InstrCost;
246 // And any other instructions that use it which become constants
248 Reduction += CountCodeReductionForConstant(&Inst);
255 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
256 // the function will be if it is inlined into a context where an argument
257 // becomes an alloca.
259 unsigned CodeMetrics::CountCodeReductionForAlloca(Value *V) {
260 if (!V->getType()->isPointerTy()) return 0; // Not a pointer
261 unsigned Reduction = 0;
262 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
263 Instruction *I = cast<Instruction>(*UI);
264 if (isa<LoadInst>(I) || isa<StoreInst>(I))
265 Reduction += InlineConstants::InstrCost;
266 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
267 // If the GEP has variable indices, we won't be able to do much with it.
268 if (GEP->hasAllConstantIndices())
269 Reduction += CountCodeReductionForAlloca(GEP);
270 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(I)) {
271 // Track pointer through bitcasts.
272 Reduction += CountCodeReductionForAlloca(BCI);
274 // If there is some other strange instruction, we're not going to be able
275 // to do much if we inline this.
283 /// analyzeFunction - Fill in the current structure with information gleaned
284 /// from the specified function.
285 void CodeMetrics::analyzeFunction(Function *F) {
286 // Look at the size of the callee.
287 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
288 analyzeBasicBlock(&*BB);
291 /// analyzeFunction - Fill in the current structure with information gleaned
292 /// from the specified function.
293 void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
294 Metrics.analyzeFunction(F);
296 // A function with exactly one return has it removed during the inlining
297 // process (see InlineFunction), so don't count it.
298 // FIXME: This knowledge should really be encoded outside of FunctionInfo.
299 if (Metrics.NumRets==1)
302 // Don't bother calculating argument weights if we are never going to inline
303 // the function anyway.
307 // Check out all of the arguments to the function, figuring out how much
308 // code can be eliminated if one of the arguments is a constant.
309 ArgumentWeights.reserve(F->arg_size());
310 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
311 ArgumentWeights.push_back(ArgInfo(Metrics.CountCodeReductionForConstant(I),
312 Metrics.CountCodeReductionForAlloca(I),
313 Metrics.CountBonusForConstant(I)));
316 /// NeverInline - returns true if the function should never be inlined into
318 bool InlineCostAnalyzer::FunctionInfo::NeverInline()
320 return (Metrics.callsSetJmp || Metrics.isRecursive ||
321 Metrics.containsIndirectBr);
324 // getSpecializationBonus - The heuristic used to determine the per-call
325 // performance boost for using a specialization of Callee with argument
326 // specializedArgNo replaced by a constant.
327 int InlineCostAnalyzer::getSpecializationBonus(Function *Callee,
328 SmallVectorImpl<unsigned> &SpecializedArgNos)
330 if (Callee->mayBeOverridden())
334 // If this function uses the coldcc calling convention, prefer not to
336 if (Callee->getCallingConv() == CallingConv::Cold)
337 Bonus -= InlineConstants::ColdccPenalty;
339 // Get information about the callee.
340 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
342 // If we haven't calculated this information yet, do so now.
343 if (CalleeFI->Metrics.NumBlocks == 0)
344 CalleeFI->analyzeFunction(Callee);
347 for (unsigned i = 0, s = SpecializedArgNos.size();
350 Bonus += CalleeFI->ArgumentWeights[SpecializedArgNos[i]].ConstantBonus;
352 // Calls usually take a long time, so they make the specialization gain
354 Bonus -= CalleeFI->Metrics.NumCalls * InlineConstants::CallPenalty;
360 // getInlineCost - The heuristic used to determine if we should inline the
361 // function call or not.
363 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
364 SmallPtrSet<const Function*, 16> &NeverInline) {
365 return getInlineCost(CS, CS.getCalledFunction(), NeverInline);
368 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
370 SmallPtrSet<const Function*, 16> &NeverInline) {
371 Instruction *TheCall = CS.getInstruction();
372 Function *Caller = TheCall->getParent()->getParent();
373 bool isDirectCall = CS.getCalledFunction() == Callee;
375 // Don't inline functions which can be redefined at link-time to mean
376 // something else. Don't inline functions marked noinline or call sites
378 if (Callee->mayBeOverridden() ||
379 Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee) ||
381 return llvm::InlineCost::getNever();
383 // InlineCost - This value measures how good of an inline candidate this call
384 // site is to inline. A lower inline cost make is more likely for the call to
385 // be inlined. This value may go negative.
389 // If there is only one call of the function, and it has internal linkage,
390 // make it almost guaranteed to be inlined.
392 if (Callee->hasLocalLinkage() && Callee->hasOneUse() && isDirectCall)
393 InlineCost += InlineConstants::LastCallToStaticBonus;
395 // If this function uses the coldcc calling convention, prefer not to inline
397 if (Callee->getCallingConv() == CallingConv::Cold)
398 InlineCost += InlineConstants::ColdccPenalty;
400 // If the instruction after the call, or if the normal destination of the
401 // invoke is an unreachable instruction, the function is noreturn. As such,
402 // there is little point in inlining this.
403 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
404 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
405 InlineCost += InlineConstants::NoreturnPenalty;
406 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
407 InlineCost += InlineConstants::NoreturnPenalty;
409 // Get information about the callee.
410 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
412 // If we haven't calculated this information yet, do so now.
413 if (CalleeFI->Metrics.NumBlocks == 0)
414 CalleeFI->analyzeFunction(Callee);
416 // If we should never inline this, return a huge cost.
417 if (CalleeFI->NeverInline())
418 return InlineCost::getNever();
420 // FIXME: It would be nice to kill off CalleeFI->NeverInline. Then we
421 // could move this up and avoid computing the FunctionInfo for
422 // things we are going to just return always inline for. This
423 // requires handling setjmp somewhere else, however.
424 if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
425 return InlineCost::getAlways();
427 if (CalleeFI->Metrics.usesDynamicAlloca) {
428 // Get infomation about the caller.
429 FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
431 // If we haven't calculated this information yet, do so now.
432 if (CallerFI.Metrics.NumBlocks == 0) {
433 CallerFI.analyzeFunction(Caller);
435 // Recompute the CalleeFI pointer, getting Caller could have invalidated
437 CalleeFI = &CachedFunctionInfo[Callee];
440 // Don't inline a callee with dynamic alloca into a caller without them.
441 // Functions containing dynamic alloca's are inefficient in various ways;
442 // don't create more inefficiency.
443 if (!CallerFI.Metrics.usesDynamicAlloca)
444 return InlineCost::getNever();
447 // Add to the inline quality for properties that make the call valuable to
448 // inline. This includes factors that indicate that the result of inlining
449 // the function will be optimizable. Currently this just looks at arguments
450 // passed into the function.
453 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
454 I != E; ++I, ++ArgNo) {
455 // Each argument passed in has a cost at both the caller and the callee
456 // sides. Measurements show that each argument costs about the same as an
458 InlineCost -= InlineConstants::InstrCost;
460 // If an alloca is passed in, inlining this function is likely to allow
461 // significant future optimization possibilities (like scalar promotion, and
462 // scalarization), so encourage the inlining of the function.
464 if (isa<AllocaInst>(I)) {
465 if (ArgNo < CalleeFI->ArgumentWeights.size())
466 InlineCost -= CalleeFI->ArgumentWeights[ArgNo].AllocaWeight;
468 // If this is a constant being passed into the function, use the argument
469 // weights calculated for the callee to determine how much will be folded
470 // away with this information.
471 } else if (isa<Constant>(I)) {
472 if (ArgNo < CalleeFI->ArgumentWeights.size())
473 InlineCost -= (CalleeFI->ArgumentWeights[ArgNo].ConstantWeight +
474 CalleeFI->ArgumentWeights[ArgNo].ConstantBonus);
478 // Now that we have considered all of the factors that make the call site more
479 // likely to be inlined, look at factors that make us not want to inline it.
481 // Calls usually take a long time, so they make the inlining gain smaller.
482 InlineCost += CalleeFI->Metrics.NumCalls * InlineConstants::CallPenalty;
484 // Look at the size of the callee. Each instruction counts as 5.
485 InlineCost += CalleeFI->Metrics.NumInsts*InlineConstants::InstrCost;
487 return llvm::InlineCost::get(InlineCost);
490 // getSpecializationCost - The heuristic used to determine the code-size
491 // impact of creating a specialized version of Callee with argument
492 // SpecializedArgNo replaced by a constant.
493 InlineCost InlineCostAnalyzer::getSpecializationCost(Function *Callee,
494 SmallVectorImpl<unsigned> &SpecializedArgNos)
496 // Don't specialize functions which can be redefined at link-time to mean
498 if (Callee->mayBeOverridden())
499 return llvm::InlineCost::getNever();
501 // Get information about the callee.
502 FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
504 // If we haven't calculated this information yet, do so now.
505 if (CalleeFI->Metrics.NumBlocks == 0)
506 CalleeFI->analyzeFunction(Callee);
510 // Look at the orginal size of the callee. Each instruction counts as 5.
511 Cost += CalleeFI->Metrics.NumInsts * InlineConstants::InstrCost;
513 // Offset that with the amount of code that can be constant-folded
514 // away with the given arguments replaced by constants.
515 for (SmallVectorImpl<unsigned>::iterator an = SpecializedArgNos.begin(),
516 ae = SpecializedArgNos.end(); an != ae; ++an)
518 Cost -= CalleeFI->ArgumentWeights[*an].ConstantWeight;
521 return llvm::InlineCost::get(Cost);
524 // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
525 // higher threshold to determine if the function call should be inlined.
526 float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
527 Function *Callee = CS.getCalledFunction();
529 // Get information about the callee.
530 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
532 // If we haven't calculated this information yet, do so now.
533 if (CalleeFI.Metrics.NumBlocks == 0)
534 CalleeFI.analyzeFunction(Callee);
537 // Single BB functions are often written to be inlined.
538 if (CalleeFI.Metrics.NumBlocks == 1)
541 // Be more aggressive if the function contains a good chunk (if it mades up
542 // at least 10% of the instructions) of vector instructions.
543 if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/2)
545 else if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/10)
550 /// growCachedCostInfo - update the cached cost info for Caller after Callee has
553 InlineCostAnalyzer::growCachedCostInfo(Function *Caller, Function *Callee) {
554 CodeMetrics &CallerMetrics = CachedFunctionInfo[Caller].Metrics;
556 // For small functions we prefer to recalculate the cost for better accuracy.
557 if (CallerMetrics.NumBlocks < 10 || CallerMetrics.NumInsts < 1000) {
558 resetCachedCostInfo(Caller);
562 // For large functions, we can save a lot of computation time by skipping
564 if (CallerMetrics.NumCalls > 0)
565 --CallerMetrics.NumCalls;
567 if (Callee == 0) return;
569 CodeMetrics &CalleeMetrics = CachedFunctionInfo[Callee].Metrics;
571 // If we don't have metrics for the callee, don't recalculate them just to
572 // update an approximation in the caller. Instead, just recalculate the
573 // caller info from scratch.
574 if (CalleeMetrics.NumBlocks == 0) {
575 resetCachedCostInfo(Caller);
579 // Since CalleeMetrics were already calculated, we know that the CallerMetrics
580 // reference isn't invalidated: both were in the DenseMap.
581 CallerMetrics.usesDynamicAlloca |= CalleeMetrics.usesDynamicAlloca;
583 // FIXME: If any of these three are true for the callee, the callee was
584 // not inlined into the caller, so I think they're redundant here.
585 CallerMetrics.callsSetJmp |= CalleeMetrics.callsSetJmp;
586 CallerMetrics.isRecursive |= CalleeMetrics.isRecursive;
587 CallerMetrics.containsIndirectBr |= CalleeMetrics.containsIndirectBr;
589 CallerMetrics.NumInsts += CalleeMetrics.NumInsts;
590 CallerMetrics.NumBlocks += CalleeMetrics.NumBlocks;
591 CallerMetrics.NumCalls += CalleeMetrics.NumCalls;
592 CallerMetrics.NumVectorInsts += CalleeMetrics.NumVectorInsts;
593 CallerMetrics.NumRets += CalleeMetrics.NumRets;
595 // analyzeBasicBlock counts each function argument as an inst.
596 if (CallerMetrics.NumInsts >= Callee->arg_size())
597 CallerMetrics.NumInsts -= Callee->arg_size();
599 CallerMetrics.NumInsts = 0;
601 // We are not updating the argument weights. We have already determined that
602 // Caller is a fairly large function, so we accept the loss of precision.
605 /// clear - empty the cache of inline costs
606 void InlineCostAnalyzer::clear() {
607 CachedFunctionInfo.clear();