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/AssumptionCache.h"
21 #include "llvm/Analysis/BasicAliasAnalysis.h"
22 #include "llvm/Analysis/CallGraph.h"
23 #include "llvm/Analysis/InlineCost.h"
24 #include "llvm/Analysis/TargetLibraryInfo.h"
25 #include "llvm/IR/CallSite.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DiagnosticInfo.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/Module.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Transforms/Utils/Cloning.h"
35 #include "llvm/Transforms/Utils/Local.h"
38 #define DEBUG_TYPE "inline"
40 STATISTIC(NumInlined, "Number of functions inlined");
41 STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined");
42 STATISTIC(NumDeleted, "Number of functions deleted because all callers found");
43 STATISTIC(NumMergedAllocas, "Number of allocas merged together");
45 // This weirdly named statistic tracks the number of times that, when attempting
46 // to inline a function A into B, we analyze the callers of B in order to see
47 // if those would be more profitable and blocked inline steps.
48 STATISTIC(NumCallerCallersAnalyzed, "Number of caller-callers analyzed");
51 InlineLimit("inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore,
52 cl::desc("Control the amount of inlining to perform (default = 225)"));
55 HintThreshold("inlinehint-threshold", cl::Hidden, cl::init(325),
56 cl::desc("Threshold for inlining functions with inline hint"));
58 // We instroduce this threshold to help performance of instrumentation based
59 // PGO before we actually hook up inliner with analysis passes such as BPI and
62 ColdThreshold("inlinecold-threshold", cl::Hidden, cl::init(225),
63 cl::desc("Threshold for inlining functions with cold attribute"));
65 // Threshold to use when optsize is specified (and there is no -inline-limit).
66 const int OptSizeThreshold = 75;
68 Inliner::Inliner(char &ID)
69 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit), InsertLifetime(true) {
72 Inliner::Inliner(char &ID, int Threshold, bool InsertLifetime)
73 : CallGraphSCCPass(ID),
74 InlineThreshold(InlineLimit.getNumOccurrences() > 0 ? InlineLimit
76 InsertLifetime(InsertLifetime) {}
78 /// For this class, we declare that we require and preserve the call graph.
79 /// If the derived class implements this method, it should
80 /// always explicitly call the implementation here.
81 void Inliner::getAnalysisUsage(AnalysisUsage &AU) const {
82 AU.addRequired<AssumptionCacheTracker>();
83 AU.addRequired<TargetLibraryInfoWrapperPass>();
84 CallGraphSCCPass::getAnalysisUsage(AU);
88 typedef DenseMap<ArrayType*, std::vector<AllocaInst*> >
89 InlinedArrayAllocasTy;
91 /// If it is possible to inline the specified call site,
92 /// do so and update the CallGraph for this operation.
94 /// This function also does some basic book-keeping to update the IR. The
95 /// InlinedArrayAllocas map keeps track of any allocas that are already
96 /// available from other functions inlined into the caller. If we are able to
97 /// inline this call site we attempt to reuse already available allocas or add
98 /// any new allocas to the set if not possible.
99 static bool InlineCallIfPossible(Pass &P, CallSite CS, InlineFunctionInfo &IFI,
100 InlinedArrayAllocasTy &InlinedArrayAllocas,
101 int InlineHistory, bool InsertLifetime) {
102 Function *Callee = CS.getCalledFunction();
103 Function *Caller = CS.getCaller();
105 // We need to manually construct BasicAA directly in order to disable
106 // its use of other function analyses.
107 BasicAAResult BAR(createLegacyPMBasicAAResult(P, *Callee));
109 // Construct our own AA results for this function. We do this manually to
110 // work around the limitations of the legacy pass manager.
111 AAResults AAR(createLegacyPMAAResults(P, *Callee, BAR));
113 // Try to inline the function. Get the list of static allocas that were
115 if (!InlineFunction(CS, IFI, &AAR, InsertLifetime))
118 AttributeFuncs::mergeAttributesForInlining(*Caller, *Callee);
120 // Look at all of the allocas that we inlined through this call site. If we
121 // have already inlined other allocas through other calls into this function,
122 // then we know that they have disjoint lifetimes and that we can merge them.
124 // There are many heuristics possible for merging these allocas, and the
125 // different options have different tradeoffs. One thing that we *really*
126 // don't want to hurt is SRoA: once inlining happens, often allocas are no
127 // longer address taken and so they can be promoted.
129 // Our "solution" for that is to only merge allocas whose outermost type is an
130 // array type. These are usually not promoted because someone is using a
131 // variable index into them. These are also often the most important ones to
134 // A better solution would be to have real memory lifetime markers in the IR
135 // and not have the inliner do any merging of allocas at all. This would
136 // allow the backend to do proper stack slot coloring of all allocas that
137 // *actually make it to the backend*, which is really what we want.
139 // Because we don't have this information, we do this simple and useful hack.
141 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
143 // When processing our SCC, check to see if CS was inlined from some other
144 // call site. For example, if we're processing "A" in this code:
146 // B() { x = alloca ... C() }
147 // C() { y = alloca ... }
148 // Assume that C was not inlined into B initially, and so we're processing A
149 // and decide to inline B into A. Doing this makes an alloca available for
150 // reuse and makes a callsite (C) available for inlining. When we process
151 // the C call site we don't want to do any alloca merging between X and Y
152 // because their scopes are not disjoint. We could make this smarter by
153 // keeping track of the inline history for each alloca in the
154 // InlinedArrayAllocas but this isn't likely to be a significant win.
155 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
158 // Loop over all the allocas we have so far and see if they can be merged with
159 // a previously inlined alloca. If not, remember that we had it.
160 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
161 AllocaNo != e; ++AllocaNo) {
162 AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
164 // Don't bother trying to merge array allocations (they will usually be
165 // canonicalized to be an allocation *of* an array), or allocations whose
166 // type is not itself an array (because we're afraid of pessimizing SRoA).
167 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
168 if (!ATy || AI->isArrayAllocation())
171 // Get the list of all available allocas for this array type.
172 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
174 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
175 // that we have to be careful not to reuse the same "available" alloca for
176 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
177 // set to keep track of which "available" allocas are being used by this
178 // function. Also, AllocasForType can be empty of course!
179 bool MergedAwayAlloca = false;
180 for (AllocaInst *AvailableAlloca : AllocasForType) {
182 unsigned Align1 = AI->getAlignment(),
183 Align2 = AvailableAlloca->getAlignment();
185 // The available alloca has to be in the right function, not in some other
186 // function in this SCC.
187 if (AvailableAlloca->getParent() != AI->getParent())
190 // If the inlined function already uses this alloca then we can't reuse
192 if (!UsedAllocas.insert(AvailableAlloca).second)
195 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
197 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: "
198 << *AvailableAlloca << '\n');
200 // Move affected dbg.declare calls immediately after the new alloca to
201 // avoid the situation when a dbg.declare preceeds its alloca.
202 if (auto *L = LocalAsMetadata::getIfExists(AI))
203 if (auto *MDV = MetadataAsValue::getIfExists(AI->getContext(), L))
204 for (User *U : MDV->users())
205 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
206 DDI->moveBefore(AvailableAlloca->getNextNode());
208 AI->replaceAllUsesWith(AvailableAlloca);
210 if (Align1 != Align2) {
211 if (!Align1 || !Align2) {
212 const DataLayout &DL = Caller->getParent()->getDataLayout();
213 unsigned TypeAlign = DL.getABITypeAlignment(AI->getAllocatedType());
215 Align1 = Align1 ? Align1 : TypeAlign;
216 Align2 = Align2 ? Align2 : TypeAlign;
220 AvailableAlloca->setAlignment(AI->getAlignment());
223 AI->eraseFromParent();
224 MergedAwayAlloca = true;
226 IFI.StaticAllocas[AllocaNo] = nullptr;
230 // If we already nuked the alloca, we're done with it.
231 if (MergedAwayAlloca)
234 // If we were unable to merge away the alloca either because there are no
235 // allocas of the right type available or because we reused them all
236 // already, remember that this alloca came from an inlined function and mark
237 // it used so we don't reuse it for other allocas from this inline
239 AllocasForType.push_back(AI);
240 UsedAllocas.insert(AI);
246 unsigned Inliner::getInlineThreshold(CallSite CS) const {
247 int Threshold = InlineThreshold; // -inline-threshold or else selected by
250 // If -inline-threshold is not given, listen to the optsize attribute when it
251 // would decrease the threshold.
252 Function *Caller = CS.getCaller();
253 bool OptSize = Caller && !Caller->isDeclaration() &&
254 // FIXME: Use Function::optForSize().
255 Caller->hasFnAttribute(Attribute::OptimizeForSize);
256 if (!(InlineLimit.getNumOccurrences() > 0) && OptSize &&
257 OptSizeThreshold < Threshold)
258 Threshold = OptSizeThreshold;
260 Function *Callee = CS.getCalledFunction();
261 if (!Callee || Callee->isDeclaration())
264 // If profile information is available, use that to adjust threshold of hot
265 // and cold functions.
266 // FIXME: The heuristic used below for determining hotness and coldness are
267 // based on preliminary SPEC tuning and may not be optimal. Replace this with
268 // a well-tuned heuristic based on *callsite* hotness and not callee hotness.
269 uint64_t FunctionCount = 0, MaxFunctionCount = 0;
270 bool HasPGOCounts = false;
271 if (Callee->getEntryCount() &&
272 Callee->getParent()->getMaximumFunctionCount()) {
274 FunctionCount = Callee->getEntryCount().getValue();
276 Callee->getParent()->getMaximumFunctionCount().getValue();
279 // Listen to the inlinehint attribute or profile based hotness information
280 // when it would increase the threshold and the caller does not need to
281 // minimize its size.
283 Callee->hasFnAttribute(Attribute::InlineHint) ||
285 FunctionCount >= (uint64_t)(0.3 * (double)MaxFunctionCount));
286 if (InlineHint && HintThreshold > Threshold &&
287 !Caller->hasFnAttribute(Attribute::MinSize))
288 Threshold = HintThreshold;
290 // Listen to the cold attribute or profile based coldness information
291 // when it would decrease the threshold.
293 Callee->hasFnAttribute(Attribute::Cold) ||
295 FunctionCount <= (uint64_t)(0.01 * (double)MaxFunctionCount));
296 // Command line argument for InlineLimit will override the default
297 // ColdThreshold. If we have -inline-threshold but no -inlinecold-threshold,
298 // do not use the default cold threshold even if it is smaller.
299 if ((InlineLimit.getNumOccurrences() == 0 ||
300 ColdThreshold.getNumOccurrences() > 0) && ColdCallee &&
301 ColdThreshold < Threshold)
302 Threshold = ColdThreshold;
307 static void emitAnalysis(CallSite CS, const Twine &Msg) {
308 Function *Caller = CS.getCaller();
309 LLVMContext &Ctx = Caller->getContext();
310 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
311 emitOptimizationRemarkAnalysis(Ctx, DEBUG_TYPE, *Caller, DLoc, Msg);
314 /// Return true if the inliner should attempt to inline 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 /// 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 ACT = &getAnalysis<AssumptionCacheTracker>();
445 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
447 SmallPtrSet<Function*, 8> SCCFunctions;
448 DEBUG(dbgs() << "Inliner visiting SCC:");
449 for (CallGraphNode *Node : SCC) {
450 Function *F = Node->getFunction();
451 if (F) SCCFunctions.insert(F);
452 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
455 // Scan through and identify all call sites ahead of time so that we only
456 // inline call sites in the original functions, not call sites that result
457 // from inlining other functions.
458 SmallVector<std::pair<CallSite, int>, 16> CallSites;
460 // When inlining a callee produces new call sites, we want to keep track of
461 // the fact that they were inlined from the callee. This allows us to avoid
462 // infinite inlining in some obscure cases. To represent this, we use an
463 // index into the InlineHistory vector.
464 SmallVector<std::pair<Function*, int>, 8> InlineHistory;
466 for (CallGraphNode *Node : SCC) {
467 Function *F = Node->getFunction();
470 for (BasicBlock &BB : *F)
471 for (Instruction &I : BB) {
472 CallSite CS(cast<Value>(&I));
473 // If this isn't a call, or it is a call to an intrinsic, it can
475 if (!CS || isa<IntrinsicInst>(I))
478 // If this is a direct call to an external function, we can never inline
479 // it. If it is an indirect call, inlining may resolve it to be a
480 // direct call, so we keep it.
481 if (Function *Callee = CS.getCalledFunction())
482 if (Callee->isDeclaration())
485 CallSites.push_back(std::make_pair(CS, -1));
489 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
491 // If there are no calls in this function, exit early.
492 if (CallSites.empty())
495 // Now that we have all of the call sites, move the ones to functions in the
496 // current SCC to the end of the list.
497 unsigned FirstCallInSCC = CallSites.size();
498 for (unsigned i = 0; i < FirstCallInSCC; ++i)
499 if (Function *F = CallSites[i].first.getCalledFunction())
500 if (SCCFunctions.count(F))
501 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
504 InlinedArrayAllocasTy InlinedArrayAllocas;
505 InlineFunctionInfo InlineInfo(&CG, ACT);
507 // Now that we have all of the call sites, loop over them and inline them if
508 // it looks profitable to do so.
509 bool Changed = false;
513 // Iterate over the outer loop because inlining functions can cause indirect
514 // calls to become direct calls.
515 // CallSites may be modified inside so ranged for loop can not be used.
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(*this, CS, InlineInfo, InlinedArrayAllocas,
564 InlineHistoryID, InsertLifetime)) {
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 (Value *Ptr : InlineInfo.InlinedCalls)
587 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 /// 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 /// Remove dead functions that are not included in DNR (Do Not Remove) list.
640 bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) {
641 SmallVector<CallGraphNode*, 16> FunctionsToRemove;
642 SmallVector<CallGraphNode *, 16> DeadFunctionsInComdats;
643 SmallDenseMap<const Comdat *, int, 16> ComdatEntriesAlive;
645 auto RemoveCGN = [&](CallGraphNode *CGN) {
646 // Remove any call graph edges from the function to its callees.
647 CGN->removeAllCalledFunctions();
649 // Remove any edges from the external node to the function's call graph
650 // node. These edges might have been made irrelegant due to
651 // optimization of the program.
652 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
654 // Removing the node for callee from the call graph and delete it.
655 FunctionsToRemove.push_back(CGN);
658 // Scan for all of the functions, looking for ones that should now be removed
659 // from the program. Insert the dead ones in the FunctionsToRemove set.
660 for (const auto &I : CG) {
661 CallGraphNode *CGN = I.second.get();
662 Function *F = CGN->getFunction();
663 if (!F || F->isDeclaration())
666 // Handle the case when this function is called and we only want to care
667 // about always-inline functions. This is a bit of a hack to share code
668 // between here and the InlineAlways pass.
669 if (AlwaysInlineOnly && !F->hasFnAttribute(Attribute::AlwaysInline))
672 // If the only remaining users of the function are dead constants, remove
674 F->removeDeadConstantUsers();
676 if (!F->isDefTriviallyDead())
679 // It is unsafe to drop a function with discardable linkage from a COMDAT
680 // without also dropping the other members of the COMDAT.
681 // The inliner doesn't visit non-function entities which are in COMDAT
682 // groups so it is unsafe to do so *unless* the linkage is local.
683 if (!F->hasLocalLinkage()) {
684 if (const Comdat *C = F->getComdat()) {
685 --ComdatEntriesAlive[C];
686 DeadFunctionsInComdats.push_back(CGN);
693 if (!DeadFunctionsInComdats.empty()) {
694 // Count up all the entities in COMDAT groups
695 auto ComdatGroupReferenced = [&](const Comdat *C) {
696 auto I = ComdatEntriesAlive.find(C);
697 if (I != ComdatEntriesAlive.end())
700 for (const Function &F : CG.getModule())
701 if (const Comdat *C = F.getComdat())
702 ComdatGroupReferenced(C);
703 for (const GlobalVariable &GV : CG.getModule().globals())
704 if (const Comdat *C = GV.getComdat())
705 ComdatGroupReferenced(C);
706 for (const GlobalAlias &GA : CG.getModule().aliases())
707 if (const Comdat *C = GA.getComdat())
708 ComdatGroupReferenced(C);
709 for (CallGraphNode *CGN : DeadFunctionsInComdats) {
710 Function *F = CGN->getFunction();
711 const Comdat *C = F->getComdat();
712 int NumAlive = ComdatEntriesAlive[C];
713 // We can remove functions in a COMDAT group if the entire group is dead.
714 assert(NumAlive >= 0);
722 if (FunctionsToRemove.empty())
725 // Now that we know which functions to delete, do so. We didn't want to do
726 // this inline, because that would invalidate our CallGraph::iterator
729 // Note that it doesn't matter that we are iterating over a non-stable order
730 // here to do this, it doesn't matter which order the functions are deleted
732 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
733 FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(),
734 FunctionsToRemove.end()),
735 FunctionsToRemove.end());
736 for (CallGraphNode *CGN : FunctionsToRemove) {
737 delete CG.removeFunctionFromModule(CGN);