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) {}
71 Inliner::Inliner(char &ID, int Threshold, bool InsertLifetime)
72 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit.getNumOccurrences() > 0 ?
73 InlineLimit : Threshold),
74 InsertLifetime(InsertLifetime) {}
76 /// For this class, we declare that we require and preserve the call graph.
77 /// If the derived class implements this method, it should
78 /// always explicitly call the implementation here.
79 void Inliner::getAnalysisUsage(AnalysisUsage &AU) const {
80 AU.addRequired<AssumptionCacheTracker>();
81 AU.addRequired<TargetLibraryInfoWrapperPass>();
82 CallGraphSCCPass::getAnalysisUsage(AU);
86 typedef DenseMap<ArrayType*, std::vector<AllocaInst*> >
87 InlinedArrayAllocasTy;
89 /// If it is possible to inline the specified call site,
90 /// do so and update the CallGraph for this operation.
92 /// This function also does some basic book-keeping to update the IR. The
93 /// InlinedArrayAllocas map keeps track of any allocas that are already
94 /// available from other functions inlined into the caller. If we are able to
95 /// inline this call site we attempt to reuse already available allocas or add
96 /// any new allocas to the set if not possible.
97 static bool InlineCallIfPossible(Pass &P, CallSite CS, InlineFunctionInfo &IFI,
98 InlinedArrayAllocasTy &InlinedArrayAllocas,
99 int InlineHistory, bool InsertLifetime) {
100 Function *Callee = CS.getCalledFunction();
101 Function *Caller = CS.getCaller();
103 // We need to manually construct BasicAA directly in order to disable
104 // its use of other function analyses.
105 BasicAAResult BAR(createLegacyPMBasicAAResult(P, *Callee));
107 // Construct our own AA results for this function. We do this manually to
108 // work around the limitations of the legacy pass manager.
109 AAResults AAR(createLegacyPMAAResults(P, *Callee, BAR));
111 // Try to inline the function. Get the list of static allocas that were
113 if (!InlineFunction(CS, IFI, &AAR, InsertLifetime))
116 AttributeFuncs::mergeAttributesForInlining(*Caller, *Callee);
118 // Look at all of the allocas that we inlined through this call site. If we
119 // have already inlined other allocas through other calls into this function,
120 // then we know that they have disjoint lifetimes and that we can merge them.
122 // There are many heuristics possible for merging these allocas, and the
123 // different options have different tradeoffs. One thing that we *really*
124 // don't want to hurt is SRoA: once inlining happens, often allocas are no
125 // longer address taken and so they can be promoted.
127 // Our "solution" for that is to only merge allocas whose outermost type is an
128 // array type. These are usually not promoted because someone is using a
129 // variable index into them. These are also often the most important ones to
132 // A better solution would be to have real memory lifetime markers in the IR
133 // and not have the inliner do any merging of allocas at all. This would
134 // allow the backend to do proper stack slot coloring of all allocas that
135 // *actually make it to the backend*, which is really what we want.
137 // Because we don't have this information, we do this simple and useful hack.
139 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
141 // When processing our SCC, check to see if CS was inlined from some other
142 // call site. For example, if we're processing "A" in this code:
144 // B() { x = alloca ... C() }
145 // C() { y = alloca ... }
146 // Assume that C was not inlined into B initially, and so we're processing A
147 // and decide to inline B into A. Doing this makes an alloca available for
148 // reuse and makes a callsite (C) available for inlining. When we process
149 // the C call site we don't want to do any alloca merging between X and Y
150 // because their scopes are not disjoint. We could make this smarter by
151 // keeping track of the inline history for each alloca in the
152 // InlinedArrayAllocas but this isn't likely to be a significant win.
153 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
156 // Loop over all the allocas we have so far and see if they can be merged with
157 // a previously inlined alloca. If not, remember that we had it.
158 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
159 AllocaNo != e; ++AllocaNo) {
160 AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
162 // Don't bother trying to merge array allocations (they will usually be
163 // canonicalized to be an allocation *of* an array), or allocations whose
164 // type is not itself an array (because we're afraid of pessimizing SRoA).
165 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
166 if (!ATy || AI->isArrayAllocation())
169 // Get the list of all available allocas for this array type.
170 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
172 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
173 // that we have to be careful not to reuse the same "available" alloca for
174 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
175 // set to keep track of which "available" allocas are being used by this
176 // function. Also, AllocasForType can be empty of course!
177 bool MergedAwayAlloca = false;
178 for (AllocaInst *AvailableAlloca : AllocasForType) {
180 unsigned Align1 = AI->getAlignment(),
181 Align2 = AvailableAlloca->getAlignment();
183 // The available alloca has to be in the right function, not in some other
184 // function in this SCC.
185 if (AvailableAlloca->getParent() != AI->getParent())
188 // If the inlined function already uses this alloca then we can't reuse
190 if (!UsedAllocas.insert(AvailableAlloca).second)
193 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
195 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: "
196 << *AvailableAlloca << '\n');
198 // Move affected dbg.declare calls immediately after the new alloca to
199 // avoid the situation when a dbg.declare preceeds its alloca.
200 if (auto *L = LocalAsMetadata::getIfExists(AI))
201 if (auto *MDV = MetadataAsValue::getIfExists(AI->getContext(), L))
202 for (User *U : MDV->users())
203 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
204 DDI->moveBefore(AvailableAlloca->getNextNode());
206 AI->replaceAllUsesWith(AvailableAlloca);
208 if (Align1 != Align2) {
209 if (!Align1 || !Align2) {
210 const DataLayout &DL = Caller->getParent()->getDataLayout();
211 unsigned TypeAlign = DL.getABITypeAlignment(AI->getAllocatedType());
213 Align1 = Align1 ? Align1 : TypeAlign;
214 Align2 = Align2 ? Align2 : TypeAlign;
218 AvailableAlloca->setAlignment(AI->getAlignment());
221 AI->eraseFromParent();
222 MergedAwayAlloca = true;
224 IFI.StaticAllocas[AllocaNo] = nullptr;
228 // If we already nuked the alloca, we're done with it.
229 if (MergedAwayAlloca)
232 // If we were unable to merge away the alloca either because there are no
233 // allocas of the right type available or because we reused them all
234 // already, remember that this alloca came from an inlined function and mark
235 // it used so we don't reuse it for other allocas from this inline
237 AllocasForType.push_back(AI);
238 UsedAllocas.insert(AI);
244 unsigned Inliner::getInlineThreshold(CallSite CS) const {
245 int Threshold = InlineThreshold; // -inline-threshold or else selected by
248 // If -inline-threshold is not given, listen to the optsize attribute when it
249 // would decrease the threshold.
250 Function *Caller = CS.getCaller();
251 bool OptSize = Caller && !Caller->isDeclaration() &&
252 // FIXME: Use Function::optForSize().
253 Caller->hasFnAttribute(Attribute::OptimizeForSize);
254 if (!(InlineLimit.getNumOccurrences() > 0) && OptSize &&
255 OptSizeThreshold < Threshold)
256 Threshold = OptSizeThreshold;
258 Function *Callee = CS.getCalledFunction();
259 if (!Callee || Callee->isDeclaration())
262 // If profile information is available, use that to adjust threshold of hot
263 // and cold functions.
264 // FIXME: The heuristic used below for determining hotness and coldness are
265 // based on preliminary SPEC tuning and may not be optimal. Replace this with
266 // a well-tuned heuristic based on *callsite* hotness and not callee hotness.
267 uint64_t FunctionCount = 0, MaxFunctionCount = 0;
268 bool HasPGOCounts = false;
269 if (Callee->getEntryCount() &&
270 Callee->getParent()->getMaximumFunctionCount()) {
272 FunctionCount = Callee->getEntryCount().getValue();
274 Callee->getParent()->getMaximumFunctionCount().getValue();
277 // Listen to the inlinehint attribute or profile based hotness information
278 // when it would increase the threshold and the caller does not need to
279 // minimize its size.
281 Callee->hasFnAttribute(Attribute::InlineHint) ||
283 FunctionCount >= (uint64_t)(0.3 * (double)MaxFunctionCount));
284 if (InlineHint && HintThreshold > Threshold &&
285 !Caller->hasFnAttribute(Attribute::MinSize))
286 Threshold = HintThreshold;
288 // Listen to the cold attribute or profile based coldness information
289 // when it would decrease the threshold.
291 Callee->hasFnAttribute(Attribute::Cold) ||
293 FunctionCount <= (uint64_t)(0.01 * (double)MaxFunctionCount));
294 // Command line argument for InlineLimit will override the default
295 // ColdThreshold. If we have -inline-threshold but no -inlinecold-threshold,
296 // do not use the default cold threshold even if it is smaller.
297 if ((InlineLimit.getNumOccurrences() == 0 ||
298 ColdThreshold.getNumOccurrences() > 0) && ColdCallee &&
299 ColdThreshold < Threshold)
300 Threshold = ColdThreshold;
305 static void emitAnalysis(CallSite CS, const Twine &Msg) {
306 Function *Caller = CS.getCaller();
307 LLVMContext &Ctx = Caller->getContext();
308 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
309 emitOptimizationRemarkAnalysis(Ctx, DEBUG_TYPE, *Caller, DLoc, Msg);
312 /// Return true if the inliner should attempt to inline 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 /// 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 AssumptionCacheTracker *ACT = &getAnalysis<AssumptionCacheTracker>();
443 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
445 SmallPtrSet<Function*, 8> SCCFunctions;
446 DEBUG(dbgs() << "Inliner visiting SCC:");
447 for (CallGraphNode *Node : SCC) {
448 Function *F = Node->getFunction();
449 if (F) SCCFunctions.insert(F);
450 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
453 // Scan through and identify all call sites ahead of time so that we only
454 // inline call sites in the original functions, not call sites that result
455 // from inlining other functions.
456 SmallVector<std::pair<CallSite, int>, 16> CallSites;
458 // When inlining a callee produces new call sites, we want to keep track of
459 // the fact that they were inlined from the callee. This allows us to avoid
460 // infinite inlining in some obscure cases. To represent this, we use an
461 // index into the InlineHistory vector.
462 SmallVector<std::pair<Function*, int>, 8> InlineHistory;
464 for (CallGraphNode *Node : SCC) {
465 Function *F = Node->getFunction();
468 for (BasicBlock &BB : *F)
469 for (Instruction &I : BB) {
470 CallSite CS(cast<Value>(&I));
471 // If this isn't a call, or it is a call to an intrinsic, it can
473 if (!CS || isa<IntrinsicInst>(I))
476 // If this is a direct call to an external function, we can never inline
477 // it. If it is an indirect call, inlining may resolve it to be a
478 // direct call, so we keep it.
479 if (Function *Callee = CS.getCalledFunction())
480 if (Callee->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, ACT);
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 // CallSites may be modified inside so ranged for loop can not be used.
514 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
515 CallSite CS = CallSites[CSi].first;
517 Function *Caller = CS.getCaller();
518 Function *Callee = CS.getCalledFunction();
520 // If this call site is dead and it is to a readonly function, we should
521 // just delete the call instead of trying to inline it, regardless of
522 // size. This happens because IPSCCP propagates the result out of the
523 // call and then we're left with the dead call.
524 if (isInstructionTriviallyDead(CS.getInstruction(), &TLI)) {
525 DEBUG(dbgs() << " -> Deleting dead call: "
526 << *CS.getInstruction() << "\n");
527 // Update the call graph by deleting the edge from Callee to Caller.
528 CG[Caller]->removeCallEdgeFor(CS);
529 CS.getInstruction()->eraseFromParent();
532 // We can only inline direct calls to non-declarations.
533 if (!Callee || Callee->isDeclaration()) continue;
535 // If this call site was obtained by inlining another function, verify
536 // that the include path for the function did not include the callee
537 // itself. If so, we'd be recursively inlining the same function,
538 // which would provide the same callsites, which would cause us to
539 // infinitely inline.
540 int InlineHistoryID = CallSites[CSi].second;
541 if (InlineHistoryID != -1 &&
542 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory))
545 LLVMContext &CallerCtx = Caller->getContext();
547 // Get DebugLoc to report. CS will be invalid after Inliner.
548 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
550 // If the policy determines that we should inline this function,
552 if (!shouldInline(CS)) {
553 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
554 Twine(Callee->getName() +
555 " will not be inlined into " +
560 // Attempt to inline the function.
561 if (!InlineCallIfPossible(*this, CS, InlineInfo, InlinedArrayAllocas,
562 InlineHistoryID, InsertLifetime)) {
563 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
564 Twine(Callee->getName() +
565 " will not be inlined into " +
571 // Report the inline decision.
572 emitOptimizationRemark(
573 CallerCtx, DEBUG_TYPE, *Caller, DLoc,
574 Twine(Callee->getName() + " inlined into " + Caller->getName()));
576 // If inlining this function gave us any new call sites, throw them
577 // onto our worklist to process. They are useful inline candidates.
578 if (!InlineInfo.InlinedCalls.empty()) {
579 // Create a new inline history entry for this, so that we remember
580 // that these new callsites came about due to inlining Callee.
581 int NewHistoryID = InlineHistory.size();
582 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID));
584 for (Value *Ptr : InlineInfo.InlinedCalls)
585 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID));
589 // If we inlined or deleted the last possible call site to the function,
590 // delete the function body now.
591 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
592 // TODO: Can remove if in SCC now.
593 !SCCFunctions.count(Callee) &&
595 // The function may be apparently dead, but if there are indirect
596 // callgraph references to the node, we cannot delete it yet, this
597 // could invalidate the CGSCC iterator.
598 CG[Callee]->getNumReferences() == 0) {
599 DEBUG(dbgs() << " -> Deleting dead function: "
600 << Callee->getName() << "\n");
601 CallGraphNode *CalleeNode = CG[Callee];
603 // Remove any call graph edges from the callee to its callees.
604 CalleeNode->removeAllCalledFunctions();
606 // Removing the node for callee from the call graph and delete it.
607 delete CG.removeFunctionFromModule(CalleeNode);
611 // Remove this call site from the list. If possible, use
612 // swap/pop_back for efficiency, but do not use it if doing so would
613 // move a call site to a function in this SCC before the
614 // 'FirstCallInSCC' barrier.
615 if (SCC.isSingular()) {
616 CallSites[CSi] = CallSites.back();
617 CallSites.pop_back();
619 CallSites.erase(CallSites.begin()+CSi);
626 } while (LocalChange);
631 /// Remove now-dead linkonce functions at the end of
632 /// processing to avoid breaking the SCC traversal.
633 bool Inliner::doFinalization(CallGraph &CG) {
634 return removeDeadFunctions(CG);
637 /// Remove dead functions that are not included in DNR (Do Not Remove) list.
638 bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) {
639 SmallVector<CallGraphNode*, 16> FunctionsToRemove;
640 SmallVector<CallGraphNode *, 16> DeadFunctionsInComdats;
641 SmallDenseMap<const Comdat *, int, 16> ComdatEntriesAlive;
643 auto RemoveCGN = [&](CallGraphNode *CGN) {
644 // Remove any call graph edges from the function to its callees.
645 CGN->removeAllCalledFunctions();
647 // Remove any edges from the external node to the function's call graph
648 // node. These edges might have been made irrelegant due to
649 // optimization of the program.
650 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
652 // Removing the node for callee from the call graph and delete it.
653 FunctionsToRemove.push_back(CGN);
656 // Scan for all of the functions, looking for ones that should now be removed
657 // from the program. Insert the dead ones in the FunctionsToRemove set.
658 for (const auto &I : CG) {
659 CallGraphNode *CGN = I.second.get();
660 Function *F = CGN->getFunction();
661 if (!F || F->isDeclaration())
664 // Handle the case when this function is called and we only want to care
665 // about always-inline functions. This is a bit of a hack to share code
666 // between here and the InlineAlways pass.
667 if (AlwaysInlineOnly && !F->hasFnAttribute(Attribute::AlwaysInline))
670 // If the only remaining users of the function are dead constants, remove
672 F->removeDeadConstantUsers();
674 if (!F->isDefTriviallyDead())
677 // It is unsafe to drop a function with discardable linkage from a COMDAT
678 // without also dropping the other members of the COMDAT.
679 // The inliner doesn't visit non-function entities which are in COMDAT
680 // groups so it is unsafe to do so *unless* the linkage is local.
681 if (!F->hasLocalLinkage()) {
682 if (const Comdat *C = F->getComdat()) {
683 --ComdatEntriesAlive[C];
684 DeadFunctionsInComdats.push_back(CGN);
691 if (!DeadFunctionsInComdats.empty()) {
692 // Count up all the entities in COMDAT groups
693 auto ComdatGroupReferenced = [&](const Comdat *C) {
694 auto I = ComdatEntriesAlive.find(C);
695 if (I != ComdatEntriesAlive.end())
698 for (const Function &F : CG.getModule())
699 if (const Comdat *C = F.getComdat())
700 ComdatGroupReferenced(C);
701 for (const GlobalVariable &GV : CG.getModule().globals())
702 if (const Comdat *C = GV.getComdat())
703 ComdatGroupReferenced(C);
704 for (const GlobalAlias &GA : CG.getModule().aliases())
705 if (const Comdat *C = GA.getComdat())
706 ComdatGroupReferenced(C);
707 for (CallGraphNode *CGN : DeadFunctionsInComdats) {
708 Function *F = CGN->getFunction();
709 const Comdat *C = F->getComdat();
710 int NumAlive = ComdatEntriesAlive[C];
711 // We can remove functions in a COMDAT group if the entire group is dead.
712 assert(NumAlive >= 0);
720 if (FunctionsToRemove.empty())
723 // Now that we know which functions to delete, do so. We didn't want to do
724 // this inline, because that would invalidate our CallGraph::iterator
727 // Note that it doesn't matter that we are iterating over a non-stable order
728 // here to do this, it doesn't matter which order the functions are deleted
730 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
731 FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(),
732 FunctionsToRemove.end()),
733 FunctionsToRemove.end());
734 for (CallGraphNode *CGN : FunctionsToRemove) {
735 delete CG.removeFunctionFromModule(CGN);