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/AliasAnalysis.h"
20 #include "llvm/Analysis/CallGraph.h"
21 #include "llvm/Analysis/InlineCost.h"
22 #include "llvm/IR/CallSite.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/DiagnosticInfo.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/Module.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/Target/TargetLibraryInfo.h"
32 #include "llvm/Transforms/Utils/Cloning.h"
33 #include "llvm/Transforms/Utils/Local.h"
36 #define DEBUG_TYPE "inline"
38 STATISTIC(NumInlined, "Number of functions inlined");
39 STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined");
40 STATISTIC(NumDeleted, "Number of functions deleted because all callers found");
41 STATISTIC(NumMergedAllocas, "Number of allocas merged together");
43 // This weirdly named statistic tracks the number of times that, when attempting
44 // to inline a function A into B, we analyze the callers of B in order to see
45 // if those would be more profitable and blocked inline steps.
46 STATISTIC(NumCallerCallersAnalyzed, "Number of caller-callers analyzed");
49 InlineLimit("inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore,
50 cl::desc("Control the amount of inlining to perform (default = 225)"));
53 HintThreshold("inlinehint-threshold", cl::Hidden, cl::init(325),
54 cl::desc("Threshold for inlining functions with inline hint"));
56 // We instroduce this threshold to help performance of instrumentation based
57 // PGO before we actually hook up inliner with analysis passes such as BPI and
60 ColdThreshold("inlinecold-threshold", cl::Hidden, cl::init(225),
61 cl::desc("Threshold for inlining functions with cold attribute"));
63 // Threshold to use when optsize is specified (and there is no -inline-limit).
64 const int OptSizeThreshold = 75;
66 Inliner::Inliner(char &ID)
67 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit), InsertLifetime(true) {}
69 Inliner::Inliner(char &ID, int Threshold, bool InsertLifetime)
70 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit.getNumOccurrences() > 0 ?
71 InlineLimit : Threshold),
72 InsertLifetime(InsertLifetime) {}
74 /// getAnalysisUsage - For this class, we declare that we require and preserve
75 /// the call graph. If the derived class implements this method, it should
76 /// always explicitly call the implementation here.
77 void Inliner::getAnalysisUsage(AnalysisUsage &AU) const {
78 AU.addRequired<AliasAnalysis>();
79 CallGraphSCCPass::getAnalysisUsage(AU);
83 typedef DenseMap<ArrayType*, std::vector<AllocaInst*> >
84 InlinedArrayAllocasTy;
86 /// \brief If the inlined function had a higher stack protection level than the
87 /// calling function, then bump up the caller's stack protection level.
88 static void AdjustCallerSSPLevel(Function *Caller, Function *Callee) {
89 // If upgrading the SSP attribute, clear out the old SSP Attributes first.
90 // Having multiple SSP attributes doesn't actually hurt, but it adds useless
93 B.addAttribute(Attribute::StackProtect)
94 .addAttribute(Attribute::StackProtectStrong);
95 AttributeSet OldSSPAttr = AttributeSet::get(Caller->getContext(),
96 AttributeSet::FunctionIndex,
98 AttributeSet CallerAttr = Caller->getAttributes(),
99 CalleeAttr = Callee->getAttributes();
101 if (CalleeAttr.hasAttribute(AttributeSet::FunctionIndex,
102 Attribute::StackProtectReq)) {
103 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
104 Caller->addFnAttr(Attribute::StackProtectReq);
105 } else if (CalleeAttr.hasAttribute(AttributeSet::FunctionIndex,
106 Attribute::StackProtectStrong) &&
107 !CallerAttr.hasAttribute(AttributeSet::FunctionIndex,
108 Attribute::StackProtectReq)) {
109 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
110 Caller->addFnAttr(Attribute::StackProtectStrong);
111 } else if (CalleeAttr.hasAttribute(AttributeSet::FunctionIndex,
112 Attribute::StackProtect) &&
113 !CallerAttr.hasAttribute(AttributeSet::FunctionIndex,
114 Attribute::StackProtectReq) &&
115 !CallerAttr.hasAttribute(AttributeSet::FunctionIndex,
116 Attribute::StackProtectStrong))
117 Caller->addFnAttr(Attribute::StackProtect);
120 /// InlineCallIfPossible - If it is possible to inline the specified call site,
121 /// do so and update the CallGraph for this operation.
123 /// This function also does some basic book-keeping to update the IR. The
124 /// InlinedArrayAllocas map keeps track of any allocas that are already
125 /// available from other functions inlined into the caller. If we are able to
126 /// inline this call site we attempt to reuse already available allocas or add
127 /// any new allocas to the set if not possible.
128 static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI,
129 InlinedArrayAllocasTy &InlinedArrayAllocas,
130 int InlineHistory, bool InsertLifetime,
131 const DataLayout *DL) {
132 Function *Callee = CS.getCalledFunction();
133 Function *Caller = CS.getCaller();
135 // Try to inline the function. Get the list of static allocas that were
137 if (!InlineFunction(CS, IFI, InsertLifetime))
140 AdjustCallerSSPLevel(Caller, Callee);
142 // Look at all of the allocas that we inlined through this call site. If we
143 // have already inlined other allocas through other calls into this function,
144 // then we know that they have disjoint lifetimes and that we can merge them.
146 // There are many heuristics possible for merging these allocas, and the
147 // different options have different tradeoffs. One thing that we *really*
148 // don't want to hurt is SRoA: once inlining happens, often allocas are no
149 // longer address taken and so they can be promoted.
151 // Our "solution" for that is to only merge allocas whose outermost type is an
152 // array type. These are usually not promoted because someone is using a
153 // variable index into them. These are also often the most important ones to
156 // A better solution would be to have real memory lifetime markers in the IR
157 // and not have the inliner do any merging of allocas at all. This would
158 // allow the backend to do proper stack slot coloring of all allocas that
159 // *actually make it to the backend*, which is really what we want.
161 // Because we don't have this information, we do this simple and useful hack.
163 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
165 // When processing our SCC, check to see if CS was inlined from some other
166 // call site. For example, if we're processing "A" in this code:
168 // B() { x = alloca ... C() }
169 // C() { y = alloca ... }
170 // Assume that C was not inlined into B initially, and so we're processing A
171 // and decide to inline B into A. Doing this makes an alloca available for
172 // reuse and makes a callsite (C) available for inlining. When we process
173 // the C call site we don't want to do any alloca merging between X and Y
174 // because their scopes are not disjoint. We could make this smarter by
175 // keeping track of the inline history for each alloca in the
176 // InlinedArrayAllocas but this isn't likely to be a significant win.
177 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
180 // Loop over all the allocas we have so far and see if they can be merged with
181 // a previously inlined alloca. If not, remember that we had it.
182 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
183 AllocaNo != e; ++AllocaNo) {
184 AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
186 // Don't bother trying to merge array allocations (they will usually be
187 // canonicalized to be an allocation *of* an array), or allocations whose
188 // type is not itself an array (because we're afraid of pessimizing SRoA).
189 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
190 if (!ATy || AI->isArrayAllocation())
193 // Get the list of all available allocas for this array type.
194 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
196 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
197 // that we have to be careful not to reuse the same "available" alloca for
198 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
199 // set to keep track of which "available" allocas are being used by this
200 // function. Also, AllocasForType can be empty of course!
201 bool MergedAwayAlloca = false;
202 for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) {
203 AllocaInst *AvailableAlloca = AllocasForType[i];
205 unsigned Align1 = AI->getAlignment(),
206 Align2 = AvailableAlloca->getAlignment();
207 // If we don't have data layout information, and only one alloca is using
208 // the target default, then we can't safely merge them because we can't
209 // pick the greater alignment.
210 if (!DL && (!Align1 || !Align2) && Align1 != Align2)
213 // The available alloca has to be in the right function, not in some other
214 // function in this SCC.
215 if (AvailableAlloca->getParent() != AI->getParent())
218 // If the inlined function already uses this alloca then we can't reuse
220 if (!UsedAllocas.insert(AvailableAlloca))
223 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
225 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: "
226 << *AvailableAlloca << '\n');
228 AI->replaceAllUsesWith(AvailableAlloca);
230 if (Align1 != Align2) {
231 if (!Align1 || !Align2) {
232 assert(DL && "DataLayout required to compare default alignments");
233 unsigned TypeAlign = DL->getABITypeAlignment(AI->getAllocatedType());
235 Align1 = Align1 ? Align1 : TypeAlign;
236 Align2 = Align2 ? Align2 : TypeAlign;
240 AvailableAlloca->setAlignment(AI->getAlignment());
243 AI->eraseFromParent();
244 MergedAwayAlloca = true;
246 IFI.StaticAllocas[AllocaNo] = nullptr;
250 // If we already nuked the alloca, we're done with it.
251 if (MergedAwayAlloca)
254 // If we were unable to merge away the alloca either because there are no
255 // allocas of the right type available or because we reused them all
256 // already, remember that this alloca came from an inlined function and mark
257 // it used so we don't reuse it for other allocas from this inline
259 AllocasForType.push_back(AI);
260 UsedAllocas.insert(AI);
266 unsigned Inliner::getInlineThreshold(CallSite CS) const {
267 int thres = InlineThreshold; // -inline-threshold or else selected by
270 // If -inline-threshold is not given, listen to the optsize attribute when it
271 // would decrease the threshold.
272 Function *Caller = CS.getCaller();
273 bool OptSize = Caller && !Caller->isDeclaration() &&
274 Caller->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
275 Attribute::OptimizeForSize);
276 if (!(InlineLimit.getNumOccurrences() > 0) && OptSize &&
277 OptSizeThreshold < thres)
278 thres = OptSizeThreshold;
280 // Listen to the inlinehint attribute when it would increase the threshold
281 // and the caller does not need to minimize its size.
282 Function *Callee = CS.getCalledFunction();
283 bool InlineHint = Callee && !Callee->isDeclaration() &&
284 Callee->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
285 Attribute::InlineHint);
286 if (InlineHint && HintThreshold > thres
287 && !Caller->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
289 thres = HintThreshold;
291 // Listen to the cold attribute when it would decrease the threshold.
292 bool ColdCallee = Callee && !Callee->isDeclaration() &&
293 Callee->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
295 // Command line argument for InlineLimit will override the default
296 // ColdThreshold. If we have -inline-threshold but no -inlinecold-threshold,
297 // do not use the default cold threshold even if it is smaller.
298 if ((InlineLimit.getNumOccurrences() == 0 ||
299 ColdThreshold.getNumOccurrences() > 0) && ColdCallee &&
300 ColdThreshold < thres)
301 thres = ColdThreshold;
306 static void emitAnalysis(CallSite CS, const Twine &Msg) {
307 Function *Caller = CS.getCaller();
308 LLVMContext &Ctx = Caller->getContext();
309 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
310 emitOptimizationRemarkAnalysis(Ctx, DEBUG_TYPE, *Caller, DLoc, Msg);
313 /// shouldInline - Return true if the inliner should attempt to inline
314 /// at the given CallSite.
315 bool Inliner::shouldInline(CallSite CS) {
316 InlineCost IC = getInlineCost(CS);
319 DEBUG(dbgs() << " Inlining: cost=always"
320 << ", Call: " << *CS.getInstruction() << "\n");
321 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName()) +
322 " should always be inlined (cost=always)");
327 DEBUG(dbgs() << " NOT Inlining: cost=never"
328 << ", Call: " << *CS.getInstruction() << "\n");
329 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() +
330 " should never be inlined (cost=never)"));
334 Function *Caller = CS.getCaller();
336 DEBUG(dbgs() << " NOT Inlining: cost=" << IC.getCost()
337 << ", thres=" << (IC.getCostDelta() + IC.getCost())
338 << ", Call: " << *CS.getInstruction() << "\n");
339 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() +
340 " too costly to inline (cost=") +
341 Twine(IC.getCost()) + ", threshold=" +
342 Twine(IC.getCostDelta() + IC.getCost()) + ")");
346 // Try to detect the case where the current inlining candidate caller (call
347 // it B) is a static or linkonce-ODR function and is an inlining candidate
348 // elsewhere, and the current candidate callee (call it C) is large enough
349 // that inlining it into B would make B too big to inline later. In these
350 // circumstances it may be best not to inline C into B, but to inline B into
353 // This only applies to static and linkonce-ODR functions because those are
354 // expected to be available for inlining in the translation units where they
355 // are used. Thus we will always have the opportunity to make local inlining
356 // decisions. Importantly the linkonce-ODR linkage covers inline functions
357 // and templates in C++.
359 // FIXME: All of this logic should be sunk into getInlineCost. It relies on
360 // the internal implementation of the inline cost metrics rather than
361 // treating them as truly abstract units etc.
362 if (Caller->hasLocalLinkage() || Caller->hasLinkOnceODRLinkage()) {
363 int TotalSecondaryCost = 0;
364 // The candidate cost to be imposed upon the current function.
365 int CandidateCost = IC.getCost() - (InlineConstants::CallPenalty + 1);
366 // This bool tracks what happens if we do NOT inline C into B.
367 bool callerWillBeRemoved = Caller->hasLocalLinkage();
368 // This bool tracks what happens if we DO inline C into B.
369 bool inliningPreventsSomeOuterInline = false;
370 for (User *U : Caller->users()) {
373 // If this isn't a call to Caller (it could be some other sort
374 // of reference) skip it. Such references will prevent the caller
375 // from being removed.
376 if (!CS2 || CS2.getCalledFunction() != Caller) {
377 callerWillBeRemoved = false;
381 InlineCost IC2 = getInlineCost(CS2);
382 ++NumCallerCallersAnalyzed;
384 callerWillBeRemoved = false;
390 // See if inlining or original callsite would erase the cost delta of
391 // this callsite. We subtract off the penalty for the call instruction,
392 // which we would be deleting.
393 if (IC2.getCostDelta() <= CandidateCost) {
394 inliningPreventsSomeOuterInline = true;
395 TotalSecondaryCost += IC2.getCost();
398 // If all outer calls to Caller would get inlined, the cost for the last
399 // one is set very low by getInlineCost, in anticipation that Caller will
400 // be removed entirely. We did not account for this above unless there
401 // is only one caller of Caller.
402 if (callerWillBeRemoved && !Caller->use_empty())
403 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus;
405 if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost()) {
406 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() <<
407 " Cost = " << IC.getCost() <<
408 ", outer Cost = " << TotalSecondaryCost << '\n');
410 CS, Twine("Not inlining. Cost of inlining " +
411 CS.getCalledFunction()->getName() +
412 " increases the cost of inlining " +
413 CS.getCaller()->getName() + " in other contexts"));
418 DEBUG(dbgs() << " Inlining: cost=" << IC.getCost()
419 << ", thres=" << (IC.getCostDelta() + IC.getCost())
420 << ", Call: " << *CS.getInstruction() << '\n');
422 CS, CS.getCalledFunction()->getName() + Twine(" can be inlined into ") +
423 CS.getCaller()->getName() + " with cost=" + Twine(IC.getCost()) +
424 " (threshold=" + Twine(IC.getCostDelta() + IC.getCost()) + ")");
428 /// InlineHistoryIncludes - Return true if the specified inline history ID
429 /// indicates an inline history that includes the specified function.
430 static bool InlineHistoryIncludes(Function *F, int InlineHistoryID,
431 const SmallVectorImpl<std::pair<Function*, int> > &InlineHistory) {
432 while (InlineHistoryID != -1) {
433 assert(unsigned(InlineHistoryID) < InlineHistory.size() &&
434 "Invalid inline history ID");
435 if (InlineHistory[InlineHistoryID].first == F)
437 InlineHistoryID = InlineHistory[InlineHistoryID].second;
442 bool Inliner::runOnSCC(CallGraphSCC &SCC) {
443 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
444 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
445 const DataLayout *DL = DLP ? &DLP->getDataLayout() : nullptr;
446 const TargetLibraryInfo *TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
447 AliasAnalysis *AA = &getAnalysis<AliasAnalysis>();
449 SmallPtrSet<Function*, 8> SCCFunctions;
450 DEBUG(dbgs() << "Inliner visiting SCC:");
451 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
452 Function *F = (*I)->getFunction();
453 if (F) SCCFunctions.insert(F);
454 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
457 // Scan through and identify all call sites ahead of time so that we only
458 // inline call sites in the original functions, not call sites that result
459 // from inlining other functions.
460 SmallVector<std::pair<CallSite, int>, 16> CallSites;
462 // When inlining a callee produces new call sites, we want to keep track of
463 // the fact that they were inlined from the callee. This allows us to avoid
464 // infinite inlining in some obscure cases. To represent this, we use an
465 // index into the InlineHistory vector.
466 SmallVector<std::pair<Function*, int>, 8> InlineHistory;
468 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
469 Function *F = (*I)->getFunction();
472 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
473 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
474 CallSite CS(cast<Value>(I));
475 // If this isn't a call, or it is a call to an intrinsic, it can
477 if (!CS || isa<IntrinsicInst>(I))
480 // If this is a direct call to an external function, we can never inline
481 // it. If it is an indirect call, inlining may resolve it to be a
482 // direct call, so we keep it.
483 if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration())
486 CallSites.push_back(std::make_pair(CS, -1));
490 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
492 // If there are no calls in this function, exit early.
493 if (CallSites.empty())
496 // Now that we have all of the call sites, move the ones to functions in the
497 // current SCC to the end of the list.
498 unsigned FirstCallInSCC = CallSites.size();
499 for (unsigned i = 0; i < FirstCallInSCC; ++i)
500 if (Function *F = CallSites[i].first.getCalledFunction())
501 if (SCCFunctions.count(F))
502 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
505 InlinedArrayAllocasTy InlinedArrayAllocas;
506 InlineFunctionInfo InlineInfo(&CG, DL, AA);
508 // Now that we have all of the call sites, loop over them and inline them if
509 // it looks profitable to do so.
510 bool Changed = false;
514 // Iterate over the outer loop because inlining functions can cause indirect
515 // calls to become direct calls.
516 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
517 CallSite CS = CallSites[CSi].first;
519 Function *Caller = CS.getCaller();
520 Function *Callee = CS.getCalledFunction();
522 // If this call site is dead and it is to a readonly function, we should
523 // just delete the call instead of trying to inline it, regardless of
524 // size. This happens because IPSCCP propagates the result out of the
525 // call and then we're left with the dead call.
526 if (isInstructionTriviallyDead(CS.getInstruction(), TLI)) {
527 DEBUG(dbgs() << " -> Deleting dead call: "
528 << *CS.getInstruction() << "\n");
529 // Update the call graph by deleting the edge from Callee to Caller.
530 CG[Caller]->removeCallEdgeFor(CS);
531 CS.getInstruction()->eraseFromParent();
534 // We can only inline direct calls to non-declarations.
535 if (!Callee || Callee->isDeclaration()) continue;
537 // If this call site was obtained by inlining another function, verify
538 // that the include path for the function did not include the callee
539 // itself. If so, we'd be recursively inlining the same function,
540 // which would provide the same callsites, which would cause us to
541 // infinitely inline.
542 int InlineHistoryID = CallSites[CSi].second;
543 if (InlineHistoryID != -1 &&
544 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory))
547 LLVMContext &CallerCtx = Caller->getContext();
549 // Get DebugLoc to report. CS will be invalid after Inliner.
550 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
552 // If the policy determines that we should inline this function,
554 if (!shouldInline(CS)) {
555 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
556 Twine(Callee->getName() +
557 " will not be inlined into " +
562 // Attempt to inline the function.
563 if (!InlineCallIfPossible(CS, InlineInfo, InlinedArrayAllocas,
564 InlineHistoryID, InsertLifetime, DL)) {
565 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
566 Twine(Callee->getName() +
567 " will not be inlined into " +
573 // Report the inline decision.
574 emitOptimizationRemark(
575 CallerCtx, DEBUG_TYPE, *Caller, DLoc,
576 Twine(Callee->getName() + " inlined into " + Caller->getName()));
578 // If inlining this function gave us any new call sites, throw them
579 // onto our worklist to process. They are useful inline candidates.
580 if (!InlineInfo.InlinedCalls.empty()) {
581 // Create a new inline history entry for this, so that we remember
582 // that these new callsites came about due to inlining Callee.
583 int NewHistoryID = InlineHistory.size();
584 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID));
586 for (unsigned i = 0, e = InlineInfo.InlinedCalls.size();
588 Value *Ptr = InlineInfo.InlinedCalls[i];
589 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID));
594 // If we inlined or deleted the last possible call site to the function,
595 // delete the function body now.
596 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
597 // TODO: Can remove if in SCC now.
598 !SCCFunctions.count(Callee) &&
600 // The function may be apparently dead, but if there are indirect
601 // callgraph references to the node, we cannot delete it yet, this
602 // could invalidate the CGSCC iterator.
603 CG[Callee]->getNumReferences() == 0) {
604 DEBUG(dbgs() << " -> Deleting dead function: "
605 << Callee->getName() << "\n");
606 CallGraphNode *CalleeNode = CG[Callee];
608 // Remove any call graph edges from the callee to its callees.
609 CalleeNode->removeAllCalledFunctions();
611 // Removing the node for callee from the call graph and delete it.
612 delete CG.removeFunctionFromModule(CalleeNode);
616 // Remove this call site from the list. If possible, use
617 // swap/pop_back for efficiency, but do not use it if doing so would
618 // move a call site to a function in this SCC before the
619 // 'FirstCallInSCC' barrier.
620 if (SCC.isSingular()) {
621 CallSites[CSi] = CallSites.back();
622 CallSites.pop_back();
624 CallSites.erase(CallSites.begin()+CSi);
631 } while (LocalChange);
636 // doFinalization - Remove now-dead linkonce functions at the end of
637 // processing to avoid breaking the SCC traversal.
638 bool Inliner::doFinalization(CallGraph &CG) {
639 return removeDeadFunctions(CG);
642 /// removeDeadFunctions - Remove dead functions that are not included in
643 /// DNR (Do Not Remove) list.
644 bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) {
645 SmallVector<CallGraphNode*, 16> FunctionsToRemove;
647 // Scan for all of the functions, looking for ones that should now be removed
648 // from the program. Insert the dead ones in the FunctionsToRemove set.
649 for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) {
650 CallGraphNode *CGN = I->second;
651 Function *F = CGN->getFunction();
652 if (!F || F->isDeclaration())
655 // Handle the case when this function is called and we only want to care
656 // about always-inline functions. This is a bit of a hack to share code
657 // between here and the InlineAlways pass.
658 if (AlwaysInlineOnly &&
659 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
660 Attribute::AlwaysInline))
663 // If the only remaining users of the function are dead constants, remove
665 F->removeDeadConstantUsers();
667 if (!F->isDefTriviallyDead())
670 // Remove any call graph edges from the function to its callees.
671 CGN->removeAllCalledFunctions();
673 // Remove any edges from the external node to the function's call graph
674 // node. These edges might have been made irrelegant due to
675 // optimization of the program.
676 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
678 // Removing the node for callee from the call graph and delete it.
679 FunctionsToRemove.push_back(CGN);
681 if (FunctionsToRemove.empty())
684 // Now that we know which functions to delete, do so. We didn't want to do
685 // this inline, because that would invalidate our CallGraph::iterator
688 // Note that it doesn't matter that we are iterating over a non-stable order
689 // here to do this, it doesn't matter which order the functions are deleted
691 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
692 FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(),
693 FunctionsToRemove.end()),
694 FunctionsToRemove.end());
695 for (SmallVectorImpl<CallGraphNode *>::iterator I = FunctionsToRemove.begin(),
696 E = FunctionsToRemove.end();
698 delete CG.removeFunctionFromModule(*I);