1 //===- Inliner.cpp - Code common to all inliners --------------------------===//
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 the mechanics required to implement inlining without
11 // missing any calls and updating the call graph. The decisions of which calls
12 // are profitable to inline are implemented elsewhere.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/Transforms/IPO/InlinerPass.h"
17 #include "llvm/ADT/SmallPtrSet.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/CallGraph.h"
20 #include "llvm/Analysis/InlineCost.h"
21 #include "llvm/IR/CallSite.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/DiagnosticInfo.h"
24 #include "llvm/IR/Instructions.h"
25 #include "llvm/IR/IntrinsicInst.h"
26 #include "llvm/IR/Module.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/raw_ostream.h"
30 #include "llvm/Target/TargetLibraryInfo.h"
31 #include "llvm/Transforms/Utils/Cloning.h"
32 #include "llvm/Transforms/Utils/Local.h"
35 #define DEBUG_TYPE "inline"
37 STATISTIC(NumInlined, "Number of functions inlined");
38 STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined");
39 STATISTIC(NumDeleted, "Number of functions deleted because all callers found");
40 STATISTIC(NumMergedAllocas, "Number of allocas merged together");
42 // This weirdly named statistic tracks the number of times that, when attempting
43 // to inline a function A into B, we analyze the callers of B in order to see
44 // if those would be more profitable and blocked inline steps.
45 STATISTIC(NumCallerCallersAnalyzed, "Number of caller-callers analyzed");
48 InlineLimit("inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore,
49 cl::desc("Control the amount of inlining to perform (default = 225)"));
52 HintThreshold("inlinehint-threshold", cl::Hidden, cl::init(325),
53 cl::desc("Threshold for inlining functions with inline hint"));
55 // We instroduce this threshold to help performance of instrumentation based
56 // PGO before we actually hook up inliner with analysis passes such as BPI and
59 ColdThreshold("inlinecold-threshold", cl::Hidden, cl::init(225),
60 cl::desc("Threshold for inlining functions with cold attribute"));
62 // Threshold to use when optsize is specified (and there is no -inline-limit).
63 const int OptSizeThreshold = 75;
65 Inliner::Inliner(char &ID)
66 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit), InsertLifetime(true) {}
68 Inliner::Inliner(char &ID, int Threshold, bool InsertLifetime)
69 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit.getNumOccurrences() > 0 ?
70 InlineLimit : Threshold),
71 InsertLifetime(InsertLifetime) {}
73 /// getAnalysisUsage - For this class, we declare that we require and preserve
74 /// the call graph. If the derived class implements this method, it should
75 /// always explicitly call the implementation here.
76 void Inliner::getAnalysisUsage(AnalysisUsage &AU) const {
77 CallGraphSCCPass::getAnalysisUsage(AU);
81 typedef DenseMap<ArrayType*, std::vector<AllocaInst*> >
82 InlinedArrayAllocasTy;
84 /// \brief If the inlined function had a higher stack protection level than the
85 /// calling function, then bump up the caller's stack protection level.
86 static void AdjustCallerSSPLevel(Function *Caller, Function *Callee) {
87 // If upgrading the SSP attribute, clear out the old SSP Attributes first.
88 // Having multiple SSP attributes doesn't actually hurt, but it adds useless
91 B.addAttribute(Attribute::StackProtect)
92 .addAttribute(Attribute::StackProtectStrong);
93 AttributeSet OldSSPAttr = AttributeSet::get(Caller->getContext(),
94 AttributeSet::FunctionIndex,
96 AttributeSet CallerAttr = Caller->getAttributes(),
97 CalleeAttr = Callee->getAttributes();
99 if (CalleeAttr.hasAttribute(AttributeSet::FunctionIndex,
100 Attribute::StackProtectReq)) {
101 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
102 Caller->addFnAttr(Attribute::StackProtectReq);
103 } else if (CalleeAttr.hasAttribute(AttributeSet::FunctionIndex,
104 Attribute::StackProtectStrong) &&
105 !CallerAttr.hasAttribute(AttributeSet::FunctionIndex,
106 Attribute::StackProtectReq)) {
107 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
108 Caller->addFnAttr(Attribute::StackProtectStrong);
109 } else if (CalleeAttr.hasAttribute(AttributeSet::FunctionIndex,
110 Attribute::StackProtect) &&
111 !CallerAttr.hasAttribute(AttributeSet::FunctionIndex,
112 Attribute::StackProtectReq) &&
113 !CallerAttr.hasAttribute(AttributeSet::FunctionIndex,
114 Attribute::StackProtectStrong))
115 Caller->addFnAttr(Attribute::StackProtect);
118 /// InlineCallIfPossible - If it is possible to inline the specified call site,
119 /// do so and update the CallGraph for this operation.
121 /// This function also does some basic book-keeping to update the IR. The
122 /// InlinedArrayAllocas map keeps track of any allocas that are already
123 /// available from other functions inlined into the caller. If we are able to
124 /// inline this call site we attempt to reuse already available allocas or add
125 /// any new allocas to the set if not possible.
126 static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI,
127 InlinedArrayAllocasTy &InlinedArrayAllocas,
128 int InlineHistory, bool InsertLifetime,
129 const DataLayout *DL) {
130 Function *Callee = CS.getCalledFunction();
131 Function *Caller = CS.getCaller();
133 // Try to inline the function. Get the list of static allocas that were
135 if (!InlineFunction(CS, IFI, InsertLifetime))
138 AdjustCallerSSPLevel(Caller, Callee);
140 // Look at all of the allocas that we inlined through this call site. If we
141 // have already inlined other allocas through other calls into this function,
142 // then we know that they have disjoint lifetimes and that we can merge them.
144 // There are many heuristics possible for merging these allocas, and the
145 // different options have different tradeoffs. One thing that we *really*
146 // don't want to hurt is SRoA: once inlining happens, often allocas are no
147 // longer address taken and so they can be promoted.
149 // Our "solution" for that is to only merge allocas whose outermost type is an
150 // array type. These are usually not promoted because someone is using a
151 // variable index into them. These are also often the most important ones to
154 // A better solution would be to have real memory lifetime markers in the IR
155 // and not have the inliner do any merging of allocas at all. This would
156 // allow the backend to do proper stack slot coloring of all allocas that
157 // *actually make it to the backend*, which is really what we want.
159 // Because we don't have this information, we do this simple and useful hack.
161 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
163 // When processing our SCC, check to see if CS was inlined from some other
164 // call site. For example, if we're processing "A" in this code:
166 // B() { x = alloca ... C() }
167 // C() { y = alloca ... }
168 // Assume that C was not inlined into B initially, and so we're processing A
169 // and decide to inline B into A. Doing this makes an alloca available for
170 // reuse and makes a callsite (C) available for inlining. When we process
171 // the C call site we don't want to do any alloca merging between X and Y
172 // because their scopes are not disjoint. We could make this smarter by
173 // keeping track of the inline history for each alloca in the
174 // InlinedArrayAllocas but this isn't likely to be a significant win.
175 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
178 // Loop over all the allocas we have so far and see if they can be merged with
179 // a previously inlined alloca. If not, remember that we had it.
180 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
181 AllocaNo != e; ++AllocaNo) {
182 AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
184 // Don't bother trying to merge array allocations (they will usually be
185 // canonicalized to be an allocation *of* an array), or allocations whose
186 // type is not itself an array (because we're afraid of pessimizing SRoA).
187 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
188 if (!ATy || AI->isArrayAllocation())
191 // Get the list of all available allocas for this array type.
192 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
194 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
195 // that we have to be careful not to reuse the same "available" alloca for
196 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
197 // set to keep track of which "available" allocas are being used by this
198 // function. Also, AllocasForType can be empty of course!
199 bool MergedAwayAlloca = false;
200 for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) {
201 AllocaInst *AvailableAlloca = AllocasForType[i];
203 unsigned Align1 = AI->getAlignment(),
204 Align2 = AvailableAlloca->getAlignment();
205 // If we don't have data layout information, and only one alloca is using
206 // the target default, then we can't safely merge them because we can't
207 // pick the greater alignment.
208 if (!DL && (!Align1 || !Align2) && Align1 != Align2)
211 // The available alloca has to be in the right function, not in some other
212 // function in this SCC.
213 if (AvailableAlloca->getParent() != AI->getParent())
216 // If the inlined function already uses this alloca then we can't reuse
218 if (!UsedAllocas.insert(AvailableAlloca))
221 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
223 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: "
224 << *AvailableAlloca << '\n');
226 AI->replaceAllUsesWith(AvailableAlloca);
228 if (Align1 != Align2) {
229 if (!Align1 || !Align2) {
230 assert(DL && "DataLayout required to compare default alignments");
231 unsigned TypeAlign = DL->getABITypeAlignment(AI->getAllocatedType());
233 Align1 = Align1 ? Align1 : TypeAlign;
234 Align2 = Align2 ? Align2 : TypeAlign;
238 AvailableAlloca->setAlignment(AI->getAlignment());
241 AI->eraseFromParent();
242 MergedAwayAlloca = true;
244 IFI.StaticAllocas[AllocaNo] = nullptr;
248 // If we already nuked the alloca, we're done with it.
249 if (MergedAwayAlloca)
252 // If we were unable to merge away the alloca either because there are no
253 // allocas of the right type available or because we reused them all
254 // already, remember that this alloca came from an inlined function and mark
255 // it used so we don't reuse it for other allocas from this inline
257 AllocasForType.push_back(AI);
258 UsedAllocas.insert(AI);
264 unsigned Inliner::getInlineThreshold(CallSite CS) const {
265 int thres = InlineThreshold; // -inline-threshold or else selected by
268 // If -inline-threshold is not given, listen to the optsize attribute when it
269 // would decrease the threshold.
270 Function *Caller = CS.getCaller();
271 bool OptSize = Caller && !Caller->isDeclaration() &&
272 Caller->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
273 Attribute::OptimizeForSize);
274 if (!(InlineLimit.getNumOccurrences() > 0) && OptSize &&
275 OptSizeThreshold < thres)
276 thres = OptSizeThreshold;
278 // Listen to the inlinehint attribute when it would increase the threshold
279 // and the caller does not need to minimize its size.
280 Function *Callee = CS.getCalledFunction();
281 bool InlineHint = Callee && !Callee->isDeclaration() &&
282 Callee->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
283 Attribute::InlineHint);
284 if (InlineHint && HintThreshold > thres
285 && !Caller->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
287 thres = HintThreshold;
289 // Listen to the cold attribute when it would decrease the threshold.
290 bool ColdCallee = Callee && !Callee->isDeclaration() &&
291 Callee->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
293 // Command line argument for InlineLimit will override the default
294 // ColdThreshold. If we have -inline-threshold but no -inlinecold-threshold,
295 // do not use the default cold threshold even if it is smaller.
296 if ((InlineLimit.getNumOccurrences() == 0 ||
297 ColdThreshold.getNumOccurrences() > 0) && ColdCallee &&
298 ColdThreshold < thres)
299 thres = ColdThreshold;
304 static void emitAnalysis(CallSite CS, const Twine &Msg) {
305 Function *Caller = CS.getCaller();
306 LLVMContext &Ctx = Caller->getContext();
307 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
308 emitOptimizationRemarkAnalysis(Ctx, DEBUG_TYPE, *Caller, DLoc, Msg);
311 /// shouldInline - Return true if the inliner should attempt to inline
312 /// at the given CallSite.
313 bool Inliner::shouldInline(CallSite CS) {
314 InlineCost IC = getInlineCost(CS);
317 DEBUG(dbgs() << " Inlining: cost=always"
318 << ", Call: " << *CS.getInstruction() << "\n");
319 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName()) +
320 " should always be inlined (cost=always)");
325 DEBUG(dbgs() << " NOT Inlining: cost=never"
326 << ", Call: " << *CS.getInstruction() << "\n");
327 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() +
328 " should never be inlined (cost=never)"));
332 Function *Caller = CS.getCaller();
334 DEBUG(dbgs() << " NOT Inlining: cost=" << IC.getCost()
335 << ", thres=" << (IC.getCostDelta() + IC.getCost())
336 << ", Call: " << *CS.getInstruction() << "\n");
337 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() +
338 " too costly to inline (cost=") +
339 Twine(IC.getCost()) + ", threshold=" +
340 Twine(IC.getCostDelta() + IC.getCost()) + ")");
344 // Try to detect the case where the current inlining candidate caller (call
345 // it B) is a static or linkonce-ODR function and is an inlining candidate
346 // elsewhere, and the current candidate callee (call it C) is large enough
347 // that inlining it into B would make B too big to inline later. In these
348 // circumstances it may be best not to inline C into B, but to inline B into
351 // This only applies to static and linkonce-ODR functions because those are
352 // expected to be available for inlining in the translation units where they
353 // are used. Thus we will always have the opportunity to make local inlining
354 // decisions. Importantly the linkonce-ODR linkage covers inline functions
355 // and templates in C++.
357 // FIXME: All of this logic should be sunk into getInlineCost. It relies on
358 // the internal implementation of the inline cost metrics rather than
359 // treating them as truly abstract units etc.
360 if (Caller->hasLocalLinkage() || Caller->hasLinkOnceODRLinkage()) {
361 int TotalSecondaryCost = 0;
362 // The candidate cost to be imposed upon the current function.
363 int CandidateCost = IC.getCost() - (InlineConstants::CallPenalty + 1);
364 // This bool tracks what happens if we do NOT inline C into B.
365 bool callerWillBeRemoved = Caller->hasLocalLinkage();
366 // This bool tracks what happens if we DO inline C into B.
367 bool inliningPreventsSomeOuterInline = false;
368 for (User *U : Caller->users()) {
371 // If this isn't a call to Caller (it could be some other sort
372 // of reference) skip it. Such references will prevent the caller
373 // from being removed.
374 if (!CS2 || CS2.getCalledFunction() != Caller) {
375 callerWillBeRemoved = false;
379 InlineCost IC2 = getInlineCost(CS2);
380 ++NumCallerCallersAnalyzed;
382 callerWillBeRemoved = false;
388 // See if inlining or original callsite would erase the cost delta of
389 // this callsite. We subtract off the penalty for the call instruction,
390 // which we would be deleting.
391 if (IC2.getCostDelta() <= CandidateCost) {
392 inliningPreventsSomeOuterInline = true;
393 TotalSecondaryCost += IC2.getCost();
396 // If all outer calls to Caller would get inlined, the cost for the last
397 // one is set very low by getInlineCost, in anticipation that Caller will
398 // be removed entirely. We did not account for this above unless there
399 // is only one caller of Caller.
400 if (callerWillBeRemoved && !Caller->use_empty())
401 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus;
403 if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost()) {
404 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() <<
405 " Cost = " << IC.getCost() <<
406 ", outer Cost = " << TotalSecondaryCost << '\n');
408 CS, Twine("Not inlining. Cost of inlining " +
409 CS.getCalledFunction()->getName() +
410 " increases the cost of inlining " +
411 CS.getCaller()->getName() + " in other contexts"));
416 DEBUG(dbgs() << " Inlining: cost=" << IC.getCost()
417 << ", thres=" << (IC.getCostDelta() + IC.getCost())
418 << ", Call: " << *CS.getInstruction() << '\n');
420 CS, CS.getCalledFunction()->getName() + Twine(" can be inlined into ") +
421 CS.getCaller()->getName() + " with cost=" + Twine(IC.getCost()) +
422 " (threshold=" + Twine(IC.getCostDelta() + IC.getCost()) + ")");
426 /// InlineHistoryIncludes - Return true if the specified inline history ID
427 /// indicates an inline history that includes the specified function.
428 static bool InlineHistoryIncludes(Function *F, int InlineHistoryID,
429 const SmallVectorImpl<std::pair<Function*, int> > &InlineHistory) {
430 while (InlineHistoryID != -1) {
431 assert(unsigned(InlineHistoryID) < InlineHistory.size() &&
432 "Invalid inline history ID");
433 if (InlineHistory[InlineHistoryID].first == F)
435 InlineHistoryID = InlineHistory[InlineHistoryID].second;
440 bool Inliner::runOnSCC(CallGraphSCC &SCC) {
441 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
442 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
443 const DataLayout *DL = DLP ? &DLP->getDataLayout() : nullptr;
444 const TargetLibraryInfo *TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
446 SmallPtrSet<Function*, 8> SCCFunctions;
447 DEBUG(dbgs() << "Inliner visiting SCC:");
448 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
449 Function *F = (*I)->getFunction();
450 if (F) SCCFunctions.insert(F);
451 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
454 // Scan through and identify all call sites ahead of time so that we only
455 // inline call sites in the original functions, not call sites that result
456 // from inlining other functions.
457 SmallVector<std::pair<CallSite, int>, 16> CallSites;
459 // When inlining a callee produces new call sites, we want to keep track of
460 // the fact that they were inlined from the callee. This allows us to avoid
461 // infinite inlining in some obscure cases. To represent this, we use an
462 // index into the InlineHistory vector.
463 SmallVector<std::pair<Function*, int>, 8> InlineHistory;
465 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
466 Function *F = (*I)->getFunction();
469 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
470 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
471 CallSite CS(cast<Value>(I));
472 // If this isn't a call, or it is a call to an intrinsic, it can
474 if (!CS || isa<IntrinsicInst>(I))
477 // If this is a direct call to an external function, we can never inline
478 // it. If it is an indirect call, inlining may resolve it to be a
479 // direct call, so we keep it.
480 if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration())
483 CallSites.push_back(std::make_pair(CS, -1));
487 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
489 // If there are no calls in this function, exit early.
490 if (CallSites.empty())
493 // Now that we have all of the call sites, move the ones to functions in the
494 // current SCC to the end of the list.
495 unsigned FirstCallInSCC = CallSites.size();
496 for (unsigned i = 0; i < FirstCallInSCC; ++i)
497 if (Function *F = CallSites[i].first.getCalledFunction())
498 if (SCCFunctions.count(F))
499 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
502 InlinedArrayAllocasTy InlinedArrayAllocas;
503 InlineFunctionInfo InlineInfo(&CG, DL);
505 // Now that we have all of the call sites, loop over them and inline them if
506 // it looks profitable to do so.
507 bool Changed = false;
511 // Iterate over the outer loop because inlining functions can cause indirect
512 // calls to become direct calls.
513 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
514 CallSite CS = CallSites[CSi].first;
516 Function *Caller = CS.getCaller();
517 Function *Callee = CS.getCalledFunction();
519 // If this call site is dead and it is to a readonly function, we should
520 // just delete the call instead of trying to inline it, regardless of
521 // size. This happens because IPSCCP propagates the result out of the
522 // call and then we're left with the dead call.
523 if (isInstructionTriviallyDead(CS.getInstruction(), TLI)) {
524 DEBUG(dbgs() << " -> Deleting dead call: "
525 << *CS.getInstruction() << "\n");
526 // Update the call graph by deleting the edge from Callee to Caller.
527 CG[Caller]->removeCallEdgeFor(CS);
528 CS.getInstruction()->eraseFromParent();
531 // We can only inline direct calls to non-declarations.
532 if (!Callee || Callee->isDeclaration()) continue;
534 // If this call site was obtained by inlining another function, verify
535 // that the include path for the function did not include the callee
536 // itself. If so, we'd be recursively inlining the same function,
537 // which would provide the same callsites, which would cause us to
538 // infinitely inline.
539 int InlineHistoryID = CallSites[CSi].second;
540 if (InlineHistoryID != -1 &&
541 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory))
544 LLVMContext &CallerCtx = Caller->getContext();
546 // Get DebugLoc to report. CS will be invalid after Inliner.
547 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
549 // If the policy determines that we should inline this function,
551 if (!shouldInline(CS)) {
552 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
553 Twine(Callee->getName() +
554 " will not be inlined into " +
559 // Attempt to inline the function.
560 if (!InlineCallIfPossible(CS, InlineInfo, InlinedArrayAllocas,
561 InlineHistoryID, InsertLifetime, DL)) {
562 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
563 Twine(Callee->getName() +
564 " will not be inlined into " +
570 // Report the inline decision.
571 emitOptimizationRemark(
572 CallerCtx, DEBUG_TYPE, *Caller, DLoc,
573 Twine(Callee->getName() + " inlined into " + Caller->getName()));
575 // If inlining this function gave us any new call sites, throw them
576 // onto our worklist to process. They are useful inline candidates.
577 if (!InlineInfo.InlinedCalls.empty()) {
578 // Create a new inline history entry for this, so that we remember
579 // that these new callsites came about due to inlining Callee.
580 int NewHistoryID = InlineHistory.size();
581 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID));
583 for (unsigned i = 0, e = InlineInfo.InlinedCalls.size();
585 Value *Ptr = InlineInfo.InlinedCalls[i];
586 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID));
591 // If we inlined or deleted the last possible call site to the function,
592 // delete the function body now.
593 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
594 // TODO: Can remove if in SCC now.
595 !SCCFunctions.count(Callee) &&
597 // The function may be apparently dead, but if there are indirect
598 // callgraph references to the node, we cannot delete it yet, this
599 // could invalidate the CGSCC iterator.
600 CG[Callee]->getNumReferences() == 0) {
601 DEBUG(dbgs() << " -> Deleting dead function: "
602 << Callee->getName() << "\n");
603 CallGraphNode *CalleeNode = CG[Callee];
605 // Remove any call graph edges from the callee to its callees.
606 CalleeNode->removeAllCalledFunctions();
608 // Removing the node for callee from the call graph and delete it.
609 delete CG.removeFunctionFromModule(CalleeNode);
613 // Remove this call site from the list. If possible, use
614 // swap/pop_back for efficiency, but do not use it if doing so would
615 // move a call site to a function in this SCC before the
616 // 'FirstCallInSCC' barrier.
617 if (SCC.isSingular()) {
618 CallSites[CSi] = CallSites.back();
619 CallSites.pop_back();
621 CallSites.erase(CallSites.begin()+CSi);
628 } while (LocalChange);
633 // doFinalization - Remove now-dead linkonce functions at the end of
634 // processing to avoid breaking the SCC traversal.
635 bool Inliner::doFinalization(CallGraph &CG) {
636 return removeDeadFunctions(CG);
639 /// removeDeadFunctions - Remove dead functions that are not included in
640 /// DNR (Do Not Remove) list.
641 bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) {
642 SmallVector<CallGraphNode*, 16> FunctionsToRemove;
644 // Scan for all of the functions, looking for ones that should now be removed
645 // from the program. Insert the dead ones in the FunctionsToRemove set.
646 for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) {
647 CallGraphNode *CGN = I->second;
648 Function *F = CGN->getFunction();
649 if (!F || F->isDeclaration())
652 // Handle the case when this function is called and we only want to care
653 // about always-inline functions. This is a bit of a hack to share code
654 // between here and the InlineAlways pass.
655 if (AlwaysInlineOnly &&
656 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
657 Attribute::AlwaysInline))
660 // If the only remaining users of the function are dead constants, remove
662 F->removeDeadConstantUsers();
664 if (!F->isDefTriviallyDead())
667 // Remove any call graph edges from the function to its callees.
668 CGN->removeAllCalledFunctions();
670 // Remove any edges from the external node to the function's call graph
671 // node. These edges might have been made irrelegant due to
672 // optimization of the program.
673 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
675 // Removing the node for callee from the call graph and delete it.
676 FunctionsToRemove.push_back(CGN);
678 if (FunctionsToRemove.empty())
681 // Now that we know which functions to delete, do so. We didn't want to do
682 // this inline, because that would invalidate our CallGraph::iterator
685 // Note that it doesn't matter that we are iterating over a non-stable order
686 // here to do this, it doesn't matter which order the functions are deleted
688 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
689 FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(),
690 FunctionsToRemove.end()),
691 FunctionsToRemove.end());
692 for (SmallVectorImpl<CallGraphNode *>::iterator I = FunctionsToRemove.begin(),
693 E = FunctionsToRemove.end();
695 delete CG.removeFunctionFromModule(*I);