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 /// \brief If the inlined function had a higher stack protection level than the
90 /// calling function, then bump up the caller's stack protection level.
91 static void AdjustCallerSSPLevel(Function *Caller, Function *Callee) {
92 // If upgrading the SSP attribute, clear out the old SSP Attributes first.
93 // Having multiple SSP attributes doesn't actually hurt, but it adds useless
96 B.addAttribute(Attribute::StackProtect)
97 .addAttribute(Attribute::StackProtectStrong)
98 .addAttribute(Attribute::StackProtectReq);
99 AttributeSet OldSSPAttr = AttributeSet::get(Caller->getContext(),
100 AttributeSet::FunctionIndex,
103 if (Callee->hasFnAttribute(Attribute::SafeStack)) {
104 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
105 Caller->addFnAttr(Attribute::SafeStack);
106 } else if (Callee->hasFnAttribute(Attribute::StackProtectReq) &&
107 !Caller->hasFnAttribute(Attribute::SafeStack)) {
108 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
109 Caller->addFnAttr(Attribute::StackProtectReq);
110 } else if (Callee->hasFnAttribute(Attribute::StackProtectStrong) &&
111 !Caller->hasFnAttribute(Attribute::SafeStack) &&
112 !Caller->hasFnAttribute(Attribute::StackProtectReq)) {
113 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
114 Caller->addFnAttr(Attribute::StackProtectStrong);
115 } else if (Callee->hasFnAttribute(Attribute::StackProtect) &&
116 !Caller->hasFnAttribute(Attribute::SafeStack) &&
117 !Caller->hasFnAttribute(Attribute::StackProtectReq) &&
118 !Caller->hasFnAttribute(Attribute::StackProtectStrong))
119 Caller->addFnAttr(Attribute::StackProtect);
122 /// 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(Pass &P, CallSite CS, InlineFunctionInfo &IFI,
131 InlinedArrayAllocasTy &InlinedArrayAllocas,
132 int InlineHistory, bool InsertLifetime) {
133 Function *Callee = CS.getCalledFunction();
134 Function *Caller = CS.getCaller();
136 // We need to manually construct BasicAA directly in order to disable
137 // its use of other function analyses.
138 BasicAAResult BAR(createLegacyPMBasicAAResult(P, *Callee));
140 // Construct our own AA results for this function. We do this manually to
141 // work around the limitations of the legacy pass manager.
142 AAResults AAR(createLegacyPMAAResults(P, *Callee, BAR));
144 // Try to inline the function. Get the list of static allocas that were
146 if (!InlineFunction(CS, IFI, &AAR, InsertLifetime))
149 AdjustCallerSSPLevel(Caller, Callee);
151 // Look at all of the allocas that we inlined through this call site. If we
152 // have already inlined other allocas through other calls into this function,
153 // then we know that they have disjoint lifetimes and that we can merge them.
155 // There are many heuristics possible for merging these allocas, and the
156 // different options have different tradeoffs. One thing that we *really*
157 // don't want to hurt is SRoA: once inlining happens, often allocas are no
158 // longer address taken and so they can be promoted.
160 // Our "solution" for that is to only merge allocas whose outermost type is an
161 // array type. These are usually not promoted because someone is using a
162 // variable index into them. These are also often the most important ones to
165 // A better solution would be to have real memory lifetime markers in the IR
166 // and not have the inliner do any merging of allocas at all. This would
167 // allow the backend to do proper stack slot coloring of all allocas that
168 // *actually make it to the backend*, which is really what we want.
170 // Because we don't have this information, we do this simple and useful hack.
172 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
174 // When processing our SCC, check to see if CS was inlined from some other
175 // call site. For example, if we're processing "A" in this code:
177 // B() { x = alloca ... C() }
178 // C() { y = alloca ... }
179 // Assume that C was not inlined into B initially, and so we're processing A
180 // and decide to inline B into A. Doing this makes an alloca available for
181 // reuse and makes a callsite (C) available for inlining. When we process
182 // the C call site we don't want to do any alloca merging between X and Y
183 // because their scopes are not disjoint. We could make this smarter by
184 // keeping track of the inline history for each alloca in the
185 // InlinedArrayAllocas but this isn't likely to be a significant win.
186 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
189 // Loop over all the allocas we have so far and see if they can be merged with
190 // a previously inlined alloca. If not, remember that we had it.
191 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
192 AllocaNo != e; ++AllocaNo) {
193 AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
195 // Don't bother trying to merge array allocations (they will usually be
196 // canonicalized to be an allocation *of* an array), or allocations whose
197 // type is not itself an array (because we're afraid of pessimizing SRoA).
198 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
199 if (!ATy || AI->isArrayAllocation())
202 // Get the list of all available allocas for this array type.
203 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
205 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
206 // that we have to be careful not to reuse the same "available" alloca for
207 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
208 // set to keep track of which "available" allocas are being used by this
209 // function. Also, AllocasForType can be empty of course!
210 bool MergedAwayAlloca = false;
211 for (AllocaInst *AvailableAlloca : AllocasForType) {
213 unsigned Align1 = AI->getAlignment(),
214 Align2 = AvailableAlloca->getAlignment();
216 // The available alloca has to be in the right function, not in some other
217 // function in this SCC.
218 if (AvailableAlloca->getParent() != AI->getParent())
221 // If the inlined function already uses this alloca then we can't reuse
223 if (!UsedAllocas.insert(AvailableAlloca).second)
226 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
228 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: "
229 << *AvailableAlloca << '\n');
231 // Move affected dbg.declare calls immediately after the new alloca to
232 // avoid the situation when a dbg.declare preceeds its alloca.
233 if (auto *L = LocalAsMetadata::getIfExists(AI))
234 if (auto *MDV = MetadataAsValue::getIfExists(AI->getContext(), L))
235 for (User *U : MDV->users())
236 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
237 DDI->moveBefore(AvailableAlloca->getNextNode());
239 AI->replaceAllUsesWith(AvailableAlloca);
241 if (Align1 != Align2) {
242 if (!Align1 || !Align2) {
243 const DataLayout &DL = Caller->getParent()->getDataLayout();
244 unsigned TypeAlign = DL.getABITypeAlignment(AI->getAllocatedType());
246 Align1 = Align1 ? Align1 : TypeAlign;
247 Align2 = Align2 ? Align2 : TypeAlign;
251 AvailableAlloca->setAlignment(AI->getAlignment());
254 AI->eraseFromParent();
255 MergedAwayAlloca = true;
257 IFI.StaticAllocas[AllocaNo] = nullptr;
261 // If we already nuked the alloca, we're done with it.
262 if (MergedAwayAlloca)
265 // If we were unable to merge away the alloca either because there are no
266 // allocas of the right type available or because we reused them all
267 // already, remember that this alloca came from an inlined function and mark
268 // it used so we don't reuse it for other allocas from this inline
270 AllocasForType.push_back(AI);
271 UsedAllocas.insert(AI);
277 unsigned Inliner::getInlineThreshold(CallSite CS) const {
278 int Threshold = InlineThreshold; // -inline-threshold or else selected by
281 // If -inline-threshold is not given, listen to the optsize attribute when it
282 // would decrease the threshold.
283 Function *Caller = CS.getCaller();
284 bool OptSize = Caller && !Caller->isDeclaration() &&
285 // FIXME: Use Function::optForSize().
286 Caller->hasFnAttribute(Attribute::OptimizeForSize);
287 if (!(InlineLimit.getNumOccurrences() > 0) && OptSize &&
288 OptSizeThreshold < Threshold)
289 Threshold = OptSizeThreshold;
291 // Listen to the inlinehint attribute when it would increase the threshold
292 // and the caller does not need to minimize its size.
293 Function *Callee = CS.getCalledFunction();
294 bool InlineHint = Callee && !Callee->isDeclaration() &&
295 Callee->hasFnAttribute(Attribute::InlineHint);
296 if (InlineHint && HintThreshold > Threshold &&
297 !Caller->hasFnAttribute(Attribute::MinSize))
298 Threshold = HintThreshold;
300 // Listen to the cold attribute when it would decrease the threshold.
301 bool ColdCallee = Callee && !Callee->isDeclaration() &&
302 Callee->hasFnAttribute(Attribute::Cold);
303 // Command line argument for InlineLimit will override the default
304 // ColdThreshold. If we have -inline-threshold but no -inlinecold-threshold,
305 // do not use the default cold threshold even if it is smaller.
306 if ((InlineLimit.getNumOccurrences() == 0 ||
307 ColdThreshold.getNumOccurrences() > 0) && ColdCallee &&
308 ColdThreshold < Threshold)
309 Threshold = ColdThreshold;
314 static void emitAnalysis(CallSite CS, const Twine &Msg) {
315 Function *Caller = CS.getCaller();
316 LLVMContext &Ctx = Caller->getContext();
317 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
318 emitOptimizationRemarkAnalysis(Ctx, DEBUG_TYPE, *Caller, DLoc, Msg);
321 /// Return true if the inliner should attempt to inline at the given CallSite.
322 bool Inliner::shouldInline(CallSite CS) {
323 InlineCost IC = getInlineCost(CS);
326 DEBUG(dbgs() << " Inlining: cost=always"
327 << ", Call: " << *CS.getInstruction() << "\n");
328 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName()) +
329 " should always be inlined (cost=always)");
334 DEBUG(dbgs() << " NOT Inlining: cost=never"
335 << ", Call: " << *CS.getInstruction() << "\n");
336 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() +
337 " should never be inlined (cost=never)"));
341 Function *Caller = CS.getCaller();
343 DEBUG(dbgs() << " NOT Inlining: cost=" << IC.getCost()
344 << ", thres=" << (IC.getCostDelta() + IC.getCost())
345 << ", Call: " << *CS.getInstruction() << "\n");
346 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() +
347 " too costly to inline (cost=") +
348 Twine(IC.getCost()) + ", threshold=" +
349 Twine(IC.getCostDelta() + IC.getCost()) + ")");
353 // Try to detect the case where the current inlining candidate caller (call
354 // it B) is a static or linkonce-ODR function and is an inlining candidate
355 // elsewhere, and the current candidate callee (call it C) is large enough
356 // that inlining it into B would make B too big to inline later. In these
357 // circumstances it may be best not to inline C into B, but to inline B into
360 // This only applies to static and linkonce-ODR functions because those are
361 // expected to be available for inlining in the translation units where they
362 // are used. Thus we will always have the opportunity to make local inlining
363 // decisions. Importantly the linkonce-ODR linkage covers inline functions
364 // and templates in C++.
366 // FIXME: All of this logic should be sunk into getInlineCost. It relies on
367 // the internal implementation of the inline cost metrics rather than
368 // treating them as truly abstract units etc.
369 if (Caller->hasLocalLinkage() || Caller->hasLinkOnceODRLinkage()) {
370 int TotalSecondaryCost = 0;
371 // The candidate cost to be imposed upon the current function.
372 int CandidateCost = IC.getCost() - (InlineConstants::CallPenalty + 1);
373 // This bool tracks what happens if we do NOT inline C into B.
374 bool callerWillBeRemoved = Caller->hasLocalLinkage();
375 // This bool tracks what happens if we DO inline C into B.
376 bool inliningPreventsSomeOuterInline = false;
377 for (User *U : Caller->users()) {
380 // If this isn't a call to Caller (it could be some other sort
381 // of reference) skip it. Such references will prevent the caller
382 // from being removed.
383 if (!CS2 || CS2.getCalledFunction() != Caller) {
384 callerWillBeRemoved = false;
388 InlineCost IC2 = getInlineCost(CS2);
389 ++NumCallerCallersAnalyzed;
391 callerWillBeRemoved = false;
397 // See if inlining or original callsite would erase the cost delta of
398 // this callsite. We subtract off the penalty for the call instruction,
399 // which we would be deleting.
400 if (IC2.getCostDelta() <= CandidateCost) {
401 inliningPreventsSomeOuterInline = true;
402 TotalSecondaryCost += IC2.getCost();
405 // If all outer calls to Caller would get inlined, the cost for the last
406 // one is set very low by getInlineCost, in anticipation that Caller will
407 // be removed entirely. We did not account for this above unless there
408 // is only one caller of Caller.
409 if (callerWillBeRemoved && !Caller->use_empty())
410 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus;
412 if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost()) {
413 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() <<
414 " Cost = " << IC.getCost() <<
415 ", outer Cost = " << TotalSecondaryCost << '\n');
417 CS, Twine("Not inlining. Cost of inlining " +
418 CS.getCalledFunction()->getName() +
419 " increases the cost of inlining " +
420 CS.getCaller()->getName() + " in other contexts"));
425 DEBUG(dbgs() << " Inlining: cost=" << IC.getCost()
426 << ", thres=" << (IC.getCostDelta() + IC.getCost())
427 << ", Call: " << *CS.getInstruction() << '\n');
429 CS, CS.getCalledFunction()->getName() + Twine(" can be inlined into ") +
430 CS.getCaller()->getName() + " with cost=" + Twine(IC.getCost()) +
431 " (threshold=" + Twine(IC.getCostDelta() + IC.getCost()) + ")");
435 /// Return true if the specified inline history ID
436 /// indicates an inline history that includes the specified function.
437 static bool InlineHistoryIncludes(Function *F, int InlineHistoryID,
438 const SmallVectorImpl<std::pair<Function*, int> > &InlineHistory) {
439 while (InlineHistoryID != -1) {
440 assert(unsigned(InlineHistoryID) < InlineHistory.size() &&
441 "Invalid inline history ID");
442 if (InlineHistory[InlineHistoryID].first == F)
444 InlineHistoryID = InlineHistory[InlineHistoryID].second;
449 bool Inliner::runOnSCC(CallGraphSCC &SCC) {
450 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
451 AssumptionCacheTracker *ACT = &getAnalysis<AssumptionCacheTracker>();
452 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
454 SmallPtrSet<Function*, 8> SCCFunctions;
455 DEBUG(dbgs() << "Inliner visiting SCC:");
456 for (CallGraphNode *Node : SCC) {
457 Function *F = Node->getFunction();
458 if (F) SCCFunctions.insert(F);
459 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
462 // Scan through and identify all call sites ahead of time so that we only
463 // inline call sites in the original functions, not call sites that result
464 // from inlining other functions.
465 SmallVector<std::pair<CallSite, int>, 16> CallSites;
467 // When inlining a callee produces new call sites, we want to keep track of
468 // the fact that they were inlined from the callee. This allows us to avoid
469 // infinite inlining in some obscure cases. To represent this, we use an
470 // index into the InlineHistory vector.
471 SmallVector<std::pair<Function*, int>, 8> InlineHistory;
473 for (CallGraphNode *Node : SCC) {
474 Function *F = Node->getFunction();
477 for (BasicBlock &BB : *F)
478 for (Instruction &I : BB) {
479 CallSite CS(cast<Value>(&I));
480 // If this isn't a call, or it is a call to an intrinsic, it can
482 if (!CS || isa<IntrinsicInst>(I))
485 // If this is a direct call to an external function, we can never inline
486 // it. If it is an indirect call, inlining may resolve it to be a
487 // direct call, so we keep it.
488 if (Function *Callee = CS.getCalledFunction())
489 if (Callee->isDeclaration())
492 CallSites.push_back(std::make_pair(CS, -1));
496 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
498 // If there are no calls in this function, exit early.
499 if (CallSites.empty())
502 // Now that we have all of the call sites, move the ones to functions in the
503 // current SCC to the end of the list.
504 unsigned FirstCallInSCC = CallSites.size();
505 for (unsigned i = 0; i < FirstCallInSCC; ++i)
506 if (Function *F = CallSites[i].first.getCalledFunction())
507 if (SCCFunctions.count(F))
508 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
511 InlinedArrayAllocasTy InlinedArrayAllocas;
512 InlineFunctionInfo InlineInfo(&CG, ACT);
514 // Now that we have all of the call sites, loop over them and inline them if
515 // it looks profitable to do so.
516 bool Changed = false;
520 // Iterate over the outer loop because inlining functions can cause indirect
521 // calls to become direct calls.
522 // CallSites may be modified inside so ranged for loop can not be used.
523 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
524 CallSite CS = CallSites[CSi].first;
526 Function *Caller = CS.getCaller();
527 Function *Callee = CS.getCalledFunction();
529 // If this call site is dead and it is to a readonly function, we should
530 // just delete the call instead of trying to inline it, regardless of
531 // size. This happens because IPSCCP propagates the result out of the
532 // call and then we're left with the dead call.
533 if (isInstructionTriviallyDead(CS.getInstruction(), &TLI)) {
534 DEBUG(dbgs() << " -> Deleting dead call: "
535 << *CS.getInstruction() << "\n");
536 // Update the call graph by deleting the edge from Callee to Caller.
537 CG[Caller]->removeCallEdgeFor(CS);
538 CS.getInstruction()->eraseFromParent();
541 // We can only inline direct calls to non-declarations.
542 if (!Callee || Callee->isDeclaration()) continue;
544 // If this call site was obtained by inlining another function, verify
545 // that the include path for the function did not include the callee
546 // itself. If so, we'd be recursively inlining the same function,
547 // which would provide the same callsites, which would cause us to
548 // infinitely inline.
549 int InlineHistoryID = CallSites[CSi].second;
550 if (InlineHistoryID != -1 &&
551 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory))
554 LLVMContext &CallerCtx = Caller->getContext();
556 // Get DebugLoc to report. CS will be invalid after Inliner.
557 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
559 // If the policy determines that we should inline this function,
561 if (!shouldInline(CS)) {
562 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
563 Twine(Callee->getName() +
564 " will not be inlined into " +
569 // Attempt to inline the function.
570 if (!InlineCallIfPossible(*this, CS, InlineInfo, InlinedArrayAllocas,
571 InlineHistoryID, InsertLifetime)) {
572 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
573 Twine(Callee->getName() +
574 " will not be inlined into " +
580 // Report the inline decision.
581 emitOptimizationRemark(
582 CallerCtx, DEBUG_TYPE, *Caller, DLoc,
583 Twine(Callee->getName() + " inlined into " + Caller->getName()));
585 // If inlining this function gave us any new call sites, throw them
586 // onto our worklist to process. They are useful inline candidates.
587 if (!InlineInfo.InlinedCalls.empty()) {
588 // Create a new inline history entry for this, so that we remember
589 // that these new callsites came about due to inlining Callee.
590 int NewHistoryID = InlineHistory.size();
591 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID));
593 for (Value *Ptr : InlineInfo.InlinedCalls)
594 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID));
598 // If we inlined or deleted the last possible call site to the function,
599 // delete the function body now.
600 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
601 // TODO: Can remove if in SCC now.
602 !SCCFunctions.count(Callee) &&
604 // The function may be apparently dead, but if there are indirect
605 // callgraph references to the node, we cannot delete it yet, this
606 // could invalidate the CGSCC iterator.
607 CG[Callee]->getNumReferences() == 0) {
608 DEBUG(dbgs() << " -> Deleting dead function: "
609 << Callee->getName() << "\n");
610 CallGraphNode *CalleeNode = CG[Callee];
612 // Remove any call graph edges from the callee to its callees.
613 CalleeNode->removeAllCalledFunctions();
615 // Removing the node for callee from the call graph and delete it.
616 delete CG.removeFunctionFromModule(CalleeNode);
620 // Remove this call site from the list. If possible, use
621 // swap/pop_back for efficiency, but do not use it if doing so would
622 // move a call site to a function in this SCC before the
623 // 'FirstCallInSCC' barrier.
624 if (SCC.isSingular()) {
625 CallSites[CSi] = CallSites.back();
626 CallSites.pop_back();
628 CallSites.erase(CallSites.begin()+CSi);
635 } while (LocalChange);
640 /// Remove now-dead linkonce functions at the end of
641 /// processing to avoid breaking the SCC traversal.
642 bool Inliner::doFinalization(CallGraph &CG) {
643 return removeDeadFunctions(CG);
646 /// Remove dead functions that are not included in DNR (Do Not Remove) list.
647 bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) {
648 SmallVector<CallGraphNode*, 16> FunctionsToRemove;
649 SmallVector<CallGraphNode *, 16> DeadFunctionsInComdats;
650 SmallDenseMap<const Comdat *, int, 16> ComdatEntriesAlive;
652 auto RemoveCGN = [&](CallGraphNode *CGN) {
653 // Remove any call graph edges from the function to its callees.
654 CGN->removeAllCalledFunctions();
656 // Remove any edges from the external node to the function's call graph
657 // node. These edges might have been made irrelegant due to
658 // optimization of the program.
659 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
661 // Removing the node for callee from the call graph and delete it.
662 FunctionsToRemove.push_back(CGN);
665 // Scan for all of the functions, looking for ones that should now be removed
666 // from the program. Insert the dead ones in the FunctionsToRemove set.
667 for (const auto &I : CG) {
668 CallGraphNode *CGN = I.second.get();
669 Function *F = CGN->getFunction();
670 if (!F || F->isDeclaration())
673 // Handle the case when this function is called and we only want to care
674 // about always-inline functions. This is a bit of a hack to share code
675 // between here and the InlineAlways pass.
676 if (AlwaysInlineOnly && !F->hasFnAttribute(Attribute::AlwaysInline))
679 // If the only remaining users of the function are dead constants, remove
681 F->removeDeadConstantUsers();
683 if (!F->isDefTriviallyDead())
686 // It is unsafe to drop a function with discardable linkage from a COMDAT
687 // without also dropping the other members of the COMDAT.
688 // The inliner doesn't visit non-function entities which are in COMDAT
689 // groups so it is unsafe to do so *unless* the linkage is local.
690 if (!F->hasLocalLinkage()) {
691 if (const Comdat *C = F->getComdat()) {
692 --ComdatEntriesAlive[C];
693 DeadFunctionsInComdats.push_back(CGN);
700 if (!DeadFunctionsInComdats.empty()) {
701 // Count up all the entities in COMDAT groups
702 auto ComdatGroupReferenced = [&](const Comdat *C) {
703 auto I = ComdatEntriesAlive.find(C);
704 if (I != ComdatEntriesAlive.end())
707 for (const Function &F : CG.getModule())
708 if (const Comdat *C = F.getComdat())
709 ComdatGroupReferenced(C);
710 for (const GlobalVariable &GV : CG.getModule().globals())
711 if (const Comdat *C = GV.getComdat())
712 ComdatGroupReferenced(C);
713 for (const GlobalAlias &GA : CG.getModule().aliases())
714 if (const Comdat *C = GA.getComdat())
715 ComdatGroupReferenced(C);
716 for (CallGraphNode *CGN : DeadFunctionsInComdats) {
717 Function *F = CGN->getFunction();
718 const Comdat *C = F->getComdat();
719 int NumAlive = ComdatEntriesAlive[C];
720 // We can remove functions in a COMDAT group if the entire group is dead.
721 assert(NumAlive >= 0);
729 if (FunctionsToRemove.empty())
732 // Now that we know which functions to delete, do so. We didn't want to do
733 // this inline, because that would invalidate our CallGraph::iterator
736 // Note that it doesn't matter that we are iterating over a non-stable order
737 // here to do this, it doesn't matter which order the functions are deleted
739 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
740 FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(),
741 FunctionsToRemove.end()),
742 FunctionsToRemove.end());
743 for (CallGraphNode *CGN : FunctionsToRemove) {
744 delete CG.removeFunctionFromModule(CGN);