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/AssumptionTracker.h"
21 #include "llvm/Analysis/CallGraph.h"
22 #include "llvm/Analysis/InlineCost.h"
23 #include "llvm/IR/CallSite.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/DiagnosticInfo.h"
26 #include "llvm/IR/Instructions.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/IR/Module.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/raw_ostream.h"
32 #include "llvm/Target/TargetLibraryInfo.h"
33 #include "llvm/Transforms/Utils/Cloning.h"
34 #include "llvm/Transforms/Utils/Local.h"
37 #define DEBUG_TYPE "inline"
39 STATISTIC(NumInlined, "Number of functions inlined");
40 STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined");
41 STATISTIC(NumDeleted, "Number of functions deleted because all callers found");
42 STATISTIC(NumMergedAllocas, "Number of allocas merged together");
44 // This weirdly named statistic tracks the number of times that, when attempting
45 // to inline a function A into B, we analyze the callers of B in order to see
46 // if those would be more profitable and blocked inline steps.
47 STATISTIC(NumCallerCallersAnalyzed, "Number of caller-callers analyzed");
50 InlineLimit("inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore,
51 cl::desc("Control the amount of inlining to perform (default = 225)"));
54 HintThreshold("inlinehint-threshold", cl::Hidden, cl::init(325),
55 cl::desc("Threshold for inlining functions with inline hint"));
57 // We instroduce this threshold to help performance of instrumentation based
58 // PGO before we actually hook up inliner with analysis passes such as BPI and
61 ColdThreshold("inlinecold-threshold", cl::Hidden, cl::init(225),
62 cl::desc("Threshold for inlining functions with cold attribute"));
64 // Threshold to use when optsize is specified (and there is no -inline-limit).
65 const int OptSizeThreshold = 75;
67 Inliner::Inliner(char &ID)
68 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit), InsertLifetime(true) {}
70 Inliner::Inliner(char &ID, int Threshold, bool InsertLifetime)
71 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit.getNumOccurrences() > 0 ?
72 InlineLimit : Threshold),
73 InsertLifetime(InsertLifetime) {}
75 /// getAnalysisUsage - For this class, we declare that we require and preserve
76 /// the call graph. If the derived class implements this method, it should
77 /// always explicitly call the implementation here.
78 void Inliner::getAnalysisUsage(AnalysisUsage &AU) const {
79 AU.addRequired<AliasAnalysis>();
80 AU.addRequired<AssumptionTracker>();
81 CallGraphSCCPass::getAnalysisUsage(AU);
85 typedef DenseMap<ArrayType*, std::vector<AllocaInst*> >
86 InlinedArrayAllocasTy;
88 /// \brief If the inlined function had a higher stack protection level than the
89 /// calling function, then bump up the caller's stack protection level.
90 static void AdjustCallerSSPLevel(Function *Caller, Function *Callee) {
91 // If upgrading the SSP attribute, clear out the old SSP Attributes first.
92 // Having multiple SSP attributes doesn't actually hurt, but it adds useless
95 B.addAttribute(Attribute::StackProtect)
96 .addAttribute(Attribute::StackProtectStrong);
97 AttributeSet OldSSPAttr = AttributeSet::get(Caller->getContext(),
98 AttributeSet::FunctionIndex,
100 AttributeSet CallerAttr = Caller->getAttributes(),
101 CalleeAttr = Callee->getAttributes();
103 if (CalleeAttr.hasAttribute(AttributeSet::FunctionIndex,
104 Attribute::StackProtectReq)) {
105 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
106 Caller->addFnAttr(Attribute::StackProtectReq);
107 } else if (CalleeAttr.hasAttribute(AttributeSet::FunctionIndex,
108 Attribute::StackProtectStrong) &&
109 !CallerAttr.hasAttribute(AttributeSet::FunctionIndex,
110 Attribute::StackProtectReq)) {
111 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
112 Caller->addFnAttr(Attribute::StackProtectStrong);
113 } else if (CalleeAttr.hasAttribute(AttributeSet::FunctionIndex,
114 Attribute::StackProtect) &&
115 !CallerAttr.hasAttribute(AttributeSet::FunctionIndex,
116 Attribute::StackProtectReq) &&
117 !CallerAttr.hasAttribute(AttributeSet::FunctionIndex,
118 Attribute::StackProtectStrong))
119 Caller->addFnAttr(Attribute::StackProtect);
122 /// InlineCallIfPossible - If it is possible to inline the specified call site,
123 /// do so and update the CallGraph for this operation.
125 /// This function also does some basic book-keeping to update the IR. The
126 /// InlinedArrayAllocas map keeps track of any allocas that are already
127 /// available from other functions inlined into the caller. If we are able to
128 /// inline this call site we attempt to reuse already available allocas or add
129 /// any new allocas to the set if not possible.
130 static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI,
131 InlinedArrayAllocasTy &InlinedArrayAllocas,
132 int InlineHistory, bool InsertLifetime,
133 const DataLayout *DL) {
134 Function *Callee = CS.getCalledFunction();
135 Function *Caller = CS.getCaller();
137 // Try to inline the function. Get the list of static allocas that were
139 if (!InlineFunction(CS, IFI, InsertLifetime))
142 AdjustCallerSSPLevel(Caller, Callee);
144 // Look at all of the allocas that we inlined through this call site. If we
145 // have already inlined other allocas through other calls into this function,
146 // then we know that they have disjoint lifetimes and that we can merge them.
148 // There are many heuristics possible for merging these allocas, and the
149 // different options have different tradeoffs. One thing that we *really*
150 // don't want to hurt is SRoA: once inlining happens, often allocas are no
151 // longer address taken and so they can be promoted.
153 // Our "solution" for that is to only merge allocas whose outermost type is an
154 // array type. These are usually not promoted because someone is using a
155 // variable index into them. These are also often the most important ones to
158 // A better solution would be to have real memory lifetime markers in the IR
159 // and not have the inliner do any merging of allocas at all. This would
160 // allow the backend to do proper stack slot coloring of all allocas that
161 // *actually make it to the backend*, which is really what we want.
163 // Because we don't have this information, we do this simple and useful hack.
165 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
167 // When processing our SCC, check to see if CS was inlined from some other
168 // call site. For example, if we're processing "A" in this code:
170 // B() { x = alloca ... C() }
171 // C() { y = alloca ... }
172 // Assume that C was not inlined into B initially, and so we're processing A
173 // and decide to inline B into A. Doing this makes an alloca available for
174 // reuse and makes a callsite (C) available for inlining. When we process
175 // the C call site we don't want to do any alloca merging between X and Y
176 // because their scopes are not disjoint. We could make this smarter by
177 // keeping track of the inline history for each alloca in the
178 // InlinedArrayAllocas but this isn't likely to be a significant win.
179 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
182 // Loop over all the allocas we have so far and see if they can be merged with
183 // a previously inlined alloca. If not, remember that we had it.
184 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
185 AllocaNo != e; ++AllocaNo) {
186 AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
188 // Don't bother trying to merge array allocations (they will usually be
189 // canonicalized to be an allocation *of* an array), or allocations whose
190 // type is not itself an array (because we're afraid of pessimizing SRoA).
191 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
192 if (!ATy || AI->isArrayAllocation())
195 // Get the list of all available allocas for this array type.
196 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
198 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
199 // that we have to be careful not to reuse the same "available" alloca for
200 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
201 // set to keep track of which "available" allocas are being used by this
202 // function. Also, AllocasForType can be empty of course!
203 bool MergedAwayAlloca = false;
204 for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) {
205 AllocaInst *AvailableAlloca = AllocasForType[i];
207 unsigned Align1 = AI->getAlignment(),
208 Align2 = AvailableAlloca->getAlignment();
209 // If we don't have data layout information, and only one alloca is using
210 // the target default, then we can't safely merge them because we can't
211 // pick the greater alignment.
212 if (!DL && (!Align1 || !Align2) && Align1 != Align2)
215 // The available alloca has to be in the right function, not in some other
216 // function in this SCC.
217 if (AvailableAlloca->getParent() != AI->getParent())
220 // If the inlined function already uses this alloca then we can't reuse
222 if (!UsedAllocas.insert(AvailableAlloca))
225 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
227 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: "
228 << *AvailableAlloca << '\n');
230 AI->replaceAllUsesWith(AvailableAlloca);
232 if (Align1 != Align2) {
233 if (!Align1 || !Align2) {
234 assert(DL && "DataLayout required to compare default alignments");
235 unsigned TypeAlign = DL->getABITypeAlignment(AI->getAllocatedType());
237 Align1 = Align1 ? Align1 : TypeAlign;
238 Align2 = Align2 ? Align2 : TypeAlign;
242 AvailableAlloca->setAlignment(AI->getAlignment());
245 AI->eraseFromParent();
246 MergedAwayAlloca = true;
248 IFI.StaticAllocas[AllocaNo] = nullptr;
252 // If we already nuked the alloca, we're done with it.
253 if (MergedAwayAlloca)
256 // If we were unable to merge away the alloca either because there are no
257 // allocas of the right type available or because we reused them all
258 // already, remember that this alloca came from an inlined function and mark
259 // it used so we don't reuse it for other allocas from this inline
261 AllocasForType.push_back(AI);
262 UsedAllocas.insert(AI);
268 unsigned Inliner::getInlineThreshold(CallSite CS) const {
269 int thres = InlineThreshold; // -inline-threshold or else selected by
272 // If -inline-threshold is not given, listen to the optsize attribute when it
273 // would decrease the threshold.
274 Function *Caller = CS.getCaller();
275 bool OptSize = Caller && !Caller->isDeclaration() &&
276 Caller->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
277 Attribute::OptimizeForSize);
278 if (!(InlineLimit.getNumOccurrences() > 0) && OptSize &&
279 OptSizeThreshold < thres)
280 thres = OptSizeThreshold;
282 // Listen to the inlinehint attribute when it would increase the threshold
283 // and the caller does not need to minimize its size.
284 Function *Callee = CS.getCalledFunction();
285 bool InlineHint = Callee && !Callee->isDeclaration() &&
286 Callee->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
287 Attribute::InlineHint);
288 if (InlineHint && HintThreshold > thres
289 && !Caller->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
291 thres = HintThreshold;
293 // Listen to the cold attribute when it would decrease the threshold.
294 bool ColdCallee = Callee && !Callee->isDeclaration() &&
295 Callee->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
297 // Command line argument for InlineLimit will override the default
298 // ColdThreshold. If we have -inline-threshold but no -inlinecold-threshold,
299 // do not use the default cold threshold even if it is smaller.
300 if ((InlineLimit.getNumOccurrences() == 0 ||
301 ColdThreshold.getNumOccurrences() > 0) && ColdCallee &&
302 ColdThreshold < thres)
303 thres = ColdThreshold;
308 static void emitAnalysis(CallSite CS, const Twine &Msg) {
309 Function *Caller = CS.getCaller();
310 LLVMContext &Ctx = Caller->getContext();
311 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
312 emitOptimizationRemarkAnalysis(Ctx, DEBUG_TYPE, *Caller, DLoc, Msg);
315 /// shouldInline - Return true if the inliner should attempt to inline
316 /// at the given CallSite.
317 bool Inliner::shouldInline(CallSite CS) {
318 InlineCost IC = getInlineCost(CS);
321 DEBUG(dbgs() << " Inlining: cost=always"
322 << ", Call: " << *CS.getInstruction() << "\n");
323 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName()) +
324 " should always be inlined (cost=always)");
329 DEBUG(dbgs() << " NOT Inlining: cost=never"
330 << ", Call: " << *CS.getInstruction() << "\n");
331 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() +
332 " should never be inlined (cost=never)"));
336 Function *Caller = CS.getCaller();
338 DEBUG(dbgs() << " NOT Inlining: cost=" << IC.getCost()
339 << ", thres=" << (IC.getCostDelta() + IC.getCost())
340 << ", Call: " << *CS.getInstruction() << "\n");
341 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() +
342 " too costly to inline (cost=") +
343 Twine(IC.getCost()) + ", threshold=" +
344 Twine(IC.getCostDelta() + IC.getCost()) + ")");
348 // Try to detect the case where the current inlining candidate caller (call
349 // it B) is a static or linkonce-ODR function and is an inlining candidate
350 // elsewhere, and the current candidate callee (call it C) is large enough
351 // that inlining it into B would make B too big to inline later. In these
352 // circumstances it may be best not to inline C into B, but to inline B into
355 // This only applies to static and linkonce-ODR functions because those are
356 // expected to be available for inlining in the translation units where they
357 // are used. Thus we will always have the opportunity to make local inlining
358 // decisions. Importantly the linkonce-ODR linkage covers inline functions
359 // and templates in C++.
361 // FIXME: All of this logic should be sunk into getInlineCost. It relies on
362 // the internal implementation of the inline cost metrics rather than
363 // treating them as truly abstract units etc.
364 if (Caller->hasLocalLinkage() || Caller->hasLinkOnceODRLinkage()) {
365 int TotalSecondaryCost = 0;
366 // The candidate cost to be imposed upon the current function.
367 int CandidateCost = IC.getCost() - (InlineConstants::CallPenalty + 1);
368 // This bool tracks what happens if we do NOT inline C into B.
369 bool callerWillBeRemoved = Caller->hasLocalLinkage();
370 // This bool tracks what happens if we DO inline C into B.
371 bool inliningPreventsSomeOuterInline = false;
372 for (User *U : Caller->users()) {
375 // If this isn't a call to Caller (it could be some other sort
376 // of reference) skip it. Such references will prevent the caller
377 // from being removed.
378 if (!CS2 || CS2.getCalledFunction() != Caller) {
379 callerWillBeRemoved = false;
383 InlineCost IC2 = getInlineCost(CS2);
384 ++NumCallerCallersAnalyzed;
386 callerWillBeRemoved = false;
392 // See if inlining or original callsite would erase the cost delta of
393 // this callsite. We subtract off the penalty for the call instruction,
394 // which we would be deleting.
395 if (IC2.getCostDelta() <= CandidateCost) {
396 inliningPreventsSomeOuterInline = true;
397 TotalSecondaryCost += IC2.getCost();
400 // If all outer calls to Caller would get inlined, the cost for the last
401 // one is set very low by getInlineCost, in anticipation that Caller will
402 // be removed entirely. We did not account for this above unless there
403 // is only one caller of Caller.
404 if (callerWillBeRemoved && !Caller->use_empty())
405 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus;
407 if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost()) {
408 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() <<
409 " Cost = " << IC.getCost() <<
410 ", outer Cost = " << TotalSecondaryCost << '\n');
412 CS, Twine("Not inlining. Cost of inlining " +
413 CS.getCalledFunction()->getName() +
414 " increases the cost of inlining " +
415 CS.getCaller()->getName() + " in other contexts"));
420 DEBUG(dbgs() << " Inlining: cost=" << IC.getCost()
421 << ", thres=" << (IC.getCostDelta() + IC.getCost())
422 << ", Call: " << *CS.getInstruction() << '\n');
424 CS, CS.getCalledFunction()->getName() + Twine(" can be inlined into ") +
425 CS.getCaller()->getName() + " with cost=" + Twine(IC.getCost()) +
426 " (threshold=" + Twine(IC.getCostDelta() + IC.getCost()) + ")");
430 /// InlineHistoryIncludes - Return true if the specified inline history ID
431 /// indicates an inline history that includes the specified function.
432 static bool InlineHistoryIncludes(Function *F, int InlineHistoryID,
433 const SmallVectorImpl<std::pair<Function*, int> > &InlineHistory) {
434 while (InlineHistoryID != -1) {
435 assert(unsigned(InlineHistoryID) < InlineHistory.size() &&
436 "Invalid inline history ID");
437 if (InlineHistory[InlineHistoryID].first == F)
439 InlineHistoryID = InlineHistory[InlineHistoryID].second;
444 bool Inliner::runOnSCC(CallGraphSCC &SCC) {
445 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
446 AssumptionTracker *AT = &getAnalysis<AssumptionTracker>();
447 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
448 const DataLayout *DL = DLP ? &DLP->getDataLayout() : nullptr;
449 const TargetLibraryInfo *TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
450 AliasAnalysis *AA = &getAnalysis<AliasAnalysis>();
452 SmallPtrSet<Function*, 8> SCCFunctions;
453 DEBUG(dbgs() << "Inliner visiting SCC:");
454 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
455 Function *F = (*I)->getFunction();
456 if (F) SCCFunctions.insert(F);
457 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
460 // Scan through and identify all call sites ahead of time so that we only
461 // inline call sites in the original functions, not call sites that result
462 // from inlining other functions.
463 SmallVector<std::pair<CallSite, int>, 16> CallSites;
465 // When inlining a callee produces new call sites, we want to keep track of
466 // the fact that they were inlined from the callee. This allows us to avoid
467 // infinite inlining in some obscure cases. To represent this, we use an
468 // index into the InlineHistory vector.
469 SmallVector<std::pair<Function*, int>, 8> InlineHistory;
471 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
472 Function *F = (*I)->getFunction();
475 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
476 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
477 CallSite CS(cast<Value>(I));
478 // If this isn't a call, or it is a call to an intrinsic, it can
480 if (!CS || isa<IntrinsicInst>(I))
483 // If this is a direct call to an external function, we can never inline
484 // it. If it is an indirect call, inlining may resolve it to be a
485 // direct call, so we keep it.
486 if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration())
489 CallSites.push_back(std::make_pair(CS, -1));
493 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
495 // If there are no calls in this function, exit early.
496 if (CallSites.empty())
499 // Now that we have all of the call sites, move the ones to functions in the
500 // current SCC to the end of the list.
501 unsigned FirstCallInSCC = CallSites.size();
502 for (unsigned i = 0; i < FirstCallInSCC; ++i)
503 if (Function *F = CallSites[i].first.getCalledFunction())
504 if (SCCFunctions.count(F))
505 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
508 InlinedArrayAllocasTy InlinedArrayAllocas;
509 InlineFunctionInfo InlineInfo(&CG, DL, AA, AT);
511 // Now that we have all of the call sites, loop over them and inline them if
512 // it looks profitable to do so.
513 bool Changed = false;
517 // Iterate over the outer loop because inlining functions can cause indirect
518 // calls to become direct calls.
519 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
520 CallSite CS = CallSites[CSi].first;
522 Function *Caller = CS.getCaller();
523 Function *Callee = CS.getCalledFunction();
525 // If this call site is dead and it is to a readonly function, we should
526 // just delete the call instead of trying to inline it, regardless of
527 // size. This happens because IPSCCP propagates the result out of the
528 // call and then we're left with the dead call.
529 if (isInstructionTriviallyDead(CS.getInstruction(), TLI)) {
530 DEBUG(dbgs() << " -> Deleting dead call: "
531 << *CS.getInstruction() << "\n");
532 // Update the call graph by deleting the edge from Callee to Caller.
533 CG[Caller]->removeCallEdgeFor(CS);
534 CS.getInstruction()->eraseFromParent();
537 // We can only inline direct calls to non-declarations.
538 if (!Callee || Callee->isDeclaration()) continue;
540 // If this call site was obtained by inlining another function, verify
541 // that the include path for the function did not include the callee
542 // itself. If so, we'd be recursively inlining the same function,
543 // which would provide the same callsites, which would cause us to
544 // infinitely inline.
545 int InlineHistoryID = CallSites[CSi].second;
546 if (InlineHistoryID != -1 &&
547 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory))
550 LLVMContext &CallerCtx = Caller->getContext();
552 // Get DebugLoc to report. CS will be invalid after Inliner.
553 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
555 // If the policy determines that we should inline this function,
557 if (!shouldInline(CS)) {
558 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
559 Twine(Callee->getName() +
560 " will not be inlined into " +
565 // Attempt to inline the function.
566 if (!InlineCallIfPossible(CS, InlineInfo, InlinedArrayAllocas,
567 InlineHistoryID, InsertLifetime, DL)) {
568 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
569 Twine(Callee->getName() +
570 " will not be inlined into " +
576 // Report the inline decision.
577 emitOptimizationRemark(
578 CallerCtx, DEBUG_TYPE, *Caller, DLoc,
579 Twine(Callee->getName() + " inlined into " + Caller->getName()));
581 // If inlining this function gave us any new call sites, throw them
582 // onto our worklist to process. They are useful inline candidates.
583 if (!InlineInfo.InlinedCalls.empty()) {
584 // Create a new inline history entry for this, so that we remember
585 // that these new callsites came about due to inlining Callee.
586 int NewHistoryID = InlineHistory.size();
587 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID));
589 for (unsigned i = 0, e = InlineInfo.InlinedCalls.size();
591 Value *Ptr = InlineInfo.InlinedCalls[i];
592 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID));
597 // If we inlined or deleted the last possible call site to the function,
598 // delete the function body now.
599 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
600 // TODO: Can remove if in SCC now.
601 !SCCFunctions.count(Callee) &&
603 // The function may be apparently dead, but if there are indirect
604 // callgraph references to the node, we cannot delete it yet, this
605 // could invalidate the CGSCC iterator.
606 CG[Callee]->getNumReferences() == 0) {
607 DEBUG(dbgs() << " -> Deleting dead function: "
608 << Callee->getName() << "\n");
609 CallGraphNode *CalleeNode = CG[Callee];
611 // Remove any call graph edges from the callee to its callees.
612 CalleeNode->removeAllCalledFunctions();
614 // Removing the node for callee from the call graph and delete it.
615 delete CG.removeFunctionFromModule(CalleeNode);
619 // Remove this call site from the list. If possible, use
620 // swap/pop_back for efficiency, but do not use it if doing so would
621 // move a call site to a function in this SCC before the
622 // 'FirstCallInSCC' barrier.
623 if (SCC.isSingular()) {
624 CallSites[CSi] = CallSites.back();
625 CallSites.pop_back();
627 CallSites.erase(CallSites.begin()+CSi);
634 } while (LocalChange);
639 // doFinalization - Remove now-dead linkonce functions at the end of
640 // processing to avoid breaking the SCC traversal.
641 bool Inliner::doFinalization(CallGraph &CG) {
642 return removeDeadFunctions(CG);
645 /// removeDeadFunctions - Remove dead functions that are not included in
646 /// DNR (Do Not Remove) list.
647 bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) {
648 SmallVector<CallGraphNode*, 16> FunctionsToRemove;
650 // Scan for all of the functions, looking for ones that should now be removed
651 // from the program. Insert the dead ones in the FunctionsToRemove set.
652 for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) {
653 CallGraphNode *CGN = I->second;
654 Function *F = CGN->getFunction();
655 if (!F || F->isDeclaration())
658 // Handle the case when this function is called and we only want to care
659 // about always-inline functions. This is a bit of a hack to share code
660 // between here and the InlineAlways pass.
661 if (AlwaysInlineOnly &&
662 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
663 Attribute::AlwaysInline))
666 // If the only remaining users of the function are dead constants, remove
668 F->removeDeadConstantUsers();
670 if (!F->isDefTriviallyDead())
673 // Remove any call graph edges from the function to its callees.
674 CGN->removeAllCalledFunctions();
676 // Remove any edges from the external node to the function's call graph
677 // node. These edges might have been made irrelegant due to
678 // optimization of the program.
679 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
681 // Removing the node for callee from the call graph and delete it.
682 FunctionsToRemove.push_back(CGN);
684 if (FunctionsToRemove.empty())
687 // Now that we know which functions to delete, do so. We didn't want to do
688 // this inline, because that would invalidate our CallGraph::iterator
691 // Note that it doesn't matter that we are iterating over a non-stable order
692 // here to do this, it doesn't matter which order the functions are deleted
694 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
695 FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(),
696 FunctionsToRemove.end()),
697 FunctionsToRemove.end());
698 for (SmallVectorImpl<CallGraphNode *>::iterator I = FunctionsToRemove.begin(),
699 E = FunctionsToRemove.end();
701 delete CG.removeFunctionFromModule(*I);