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/CallGraph.h"
22 #include "llvm/Analysis/InlineCost.h"
23 #include "llvm/Analysis/TargetLibraryInfo.h"
24 #include "llvm/IR/CallSite.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DiagnosticInfo.h"
27 #include "llvm/IR/Instructions.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/IR/Module.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/raw_ostream.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 /// For this class, we declare that we require and preserve the call graph.
76 /// 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<AssumptionCacheTracker>();
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 .addAttribute(Attribute::StackProtectReq);
98 AttributeSet OldSSPAttr = AttributeSet::get(Caller->getContext(),
99 AttributeSet::FunctionIndex,
102 if (Callee->hasFnAttribute(Attribute::SafeStack)) {
103 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
104 Caller->addFnAttr(Attribute::SafeStack);
105 } else if (Callee->hasFnAttribute(Attribute::StackProtectReq) &&
106 !Caller->hasFnAttribute(Attribute::SafeStack)) {
107 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
108 Caller->addFnAttr(Attribute::StackProtectReq);
109 } else if (Callee->hasFnAttribute(Attribute::StackProtectStrong) &&
110 !Caller->hasFnAttribute(Attribute::SafeStack) &&
111 !Caller->hasFnAttribute(Attribute::StackProtectReq)) {
112 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
113 Caller->addFnAttr(Attribute::StackProtectStrong);
114 } else if (Callee->hasFnAttribute(Attribute::StackProtect) &&
115 !Caller->hasFnAttribute(Attribute::SafeStack) &&
116 !Caller->hasFnAttribute(Attribute::StackProtectReq) &&
117 !Caller->hasFnAttribute(Attribute::StackProtectStrong))
118 Caller->addFnAttr(Attribute::StackProtect);
121 /// If it is possible to inline the specified call site,
122 /// do so and update the CallGraph for this operation.
124 /// This function also does some basic book-keeping to update the IR. The
125 /// InlinedArrayAllocas map keeps track of any allocas that are already
126 /// available from other functions inlined into the caller. If we are able to
127 /// inline this call site we attempt to reuse already available allocas or add
128 /// any new allocas to the set if not possible.
129 static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI,
130 InlinedArrayAllocasTy &InlinedArrayAllocas,
131 int InlineHistory, bool InsertLifetime) {
132 Function *Callee = CS.getCalledFunction();
133 Function *Caller = CS.getCaller();
135 // Try to inline the function. Get the list of static allocas that were
137 if (!InlineFunction(CS, IFI, InsertLifetime))
140 AdjustCallerSSPLevel(Caller, Callee);
142 // Look at all of the allocas that we inlined through this call site. If we
143 // have already inlined other allocas through other calls into this function,
144 // then we know that they have disjoint lifetimes and that we can merge them.
146 // There are many heuristics possible for merging these allocas, and the
147 // different options have different tradeoffs. One thing that we *really*
148 // don't want to hurt is SRoA: once inlining happens, often allocas are no
149 // longer address taken and so they can be promoted.
151 // Our "solution" for that is to only merge allocas whose outermost type is an
152 // array type. These are usually not promoted because someone is using a
153 // variable index into them. These are also often the most important ones to
156 // A better solution would be to have real memory lifetime markers in the IR
157 // and not have the inliner do any merging of allocas at all. This would
158 // allow the backend to do proper stack slot coloring of all allocas that
159 // *actually make it to the backend*, which is really what we want.
161 // Because we don't have this information, we do this simple and useful hack.
163 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
165 // When processing our SCC, check to see if CS was inlined from some other
166 // call site. For example, if we're processing "A" in this code:
168 // B() { x = alloca ... C() }
169 // C() { y = alloca ... }
170 // Assume that C was not inlined into B initially, and so we're processing A
171 // and decide to inline B into A. Doing this makes an alloca available for
172 // reuse and makes a callsite (C) available for inlining. When we process
173 // the C call site we don't want to do any alloca merging between X and Y
174 // because their scopes are not disjoint. We could make this smarter by
175 // keeping track of the inline history for each alloca in the
176 // InlinedArrayAllocas but this isn't likely to be a significant win.
177 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
180 // Loop over all the allocas we have so far and see if they can be merged with
181 // a previously inlined alloca. If not, remember that we had it.
182 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
183 AllocaNo != e; ++AllocaNo) {
184 AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
186 // Don't bother trying to merge array allocations (they will usually be
187 // canonicalized to be an allocation *of* an array), or allocations whose
188 // type is not itself an array (because we're afraid of pessimizing SRoA).
189 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
190 if (!ATy || AI->isArrayAllocation())
193 // Get the list of all available allocas for this array type.
194 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
196 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
197 // that we have to be careful not to reuse the same "available" alloca for
198 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
199 // set to keep track of which "available" allocas are being used by this
200 // function. Also, AllocasForType can be empty of course!
201 bool MergedAwayAlloca = false;
202 for (AllocaInst *AvailableAlloca : AllocasForType) {
204 unsigned Align1 = AI->getAlignment(),
205 Align2 = AvailableAlloca->getAlignment();
207 // The available alloca has to be in the right function, not in some other
208 // function in this SCC.
209 if (AvailableAlloca->getParent() != AI->getParent())
212 // If the inlined function already uses this alloca then we can't reuse
214 if (!UsedAllocas.insert(AvailableAlloca).second)
217 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
219 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: "
220 << *AvailableAlloca << '\n');
222 AI->replaceAllUsesWith(AvailableAlloca);
224 if (Align1 != Align2) {
225 if (!Align1 || !Align2) {
226 const DataLayout &DL = Caller->getParent()->getDataLayout();
227 unsigned TypeAlign = DL.getABITypeAlignment(AI->getAllocatedType());
229 Align1 = Align1 ? Align1 : TypeAlign;
230 Align2 = Align2 ? Align2 : TypeAlign;
234 AvailableAlloca->setAlignment(AI->getAlignment());
237 AI->eraseFromParent();
238 MergedAwayAlloca = true;
240 IFI.StaticAllocas[AllocaNo] = nullptr;
244 // If we already nuked the alloca, we're done with it.
245 if (MergedAwayAlloca)
248 // If we were unable to merge away the alloca either because there are no
249 // allocas of the right type available or because we reused them all
250 // already, remember that this alloca came from an inlined function and mark
251 // it used so we don't reuse it for other allocas from this inline
253 AllocasForType.push_back(AI);
254 UsedAllocas.insert(AI);
260 unsigned Inliner::getInlineThreshold(CallSite CS) const {
261 int thres = InlineThreshold; // -inline-threshold or else selected by
264 // If -inline-threshold is not given, listen to the optsize attribute when it
265 // would decrease the threshold.
266 Function *Caller = CS.getCaller();
267 bool OptSize = Caller && !Caller->isDeclaration() &&
268 Caller->hasFnAttribute(Attribute::OptimizeForSize);
269 if (!(InlineLimit.getNumOccurrences() > 0) && OptSize &&
270 OptSizeThreshold < thres)
271 thres = OptSizeThreshold;
273 // Listen to the inlinehint attribute when it would increase the threshold
274 // and the caller does not need to minimize its size.
275 Function *Callee = CS.getCalledFunction();
276 bool InlineHint = Callee && !Callee->isDeclaration() &&
277 Callee->hasFnAttribute(Attribute::InlineHint);
278 if (InlineHint && HintThreshold > thres &&
279 !Caller->hasFnAttribute(Attribute::MinSize))
280 thres = HintThreshold;
282 // Listen to the cold attribute when it would decrease the threshold.
283 bool ColdCallee = Callee && !Callee->isDeclaration() &&
284 Callee->hasFnAttribute(Attribute::Cold);
285 // Command line argument for InlineLimit will override the default
286 // ColdThreshold. If we have -inline-threshold but no -inlinecold-threshold,
287 // do not use the default cold threshold even if it is smaller.
288 if ((InlineLimit.getNumOccurrences() == 0 ||
289 ColdThreshold.getNumOccurrences() > 0) && ColdCallee &&
290 ColdThreshold < thres)
291 thres = ColdThreshold;
296 static void emitAnalysis(CallSite CS, const Twine &Msg) {
297 Function *Caller = CS.getCaller();
298 LLVMContext &Ctx = Caller->getContext();
299 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
300 emitOptimizationRemarkAnalysis(Ctx, DEBUG_TYPE, *Caller, DLoc, Msg);
303 /// Return true if the inliner should attempt to inline at the given CallSite.
304 bool Inliner::shouldInline(CallSite CS) {
305 InlineCost IC = getInlineCost(CS);
308 DEBUG(dbgs() << " Inlining: cost=always"
309 << ", Call: " << *CS.getInstruction() << "\n");
310 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName()) +
311 " should always be inlined (cost=always)");
316 DEBUG(dbgs() << " NOT Inlining: cost=never"
317 << ", Call: " << *CS.getInstruction() << "\n");
318 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() +
319 " should never be inlined (cost=never)"));
323 Function *Caller = CS.getCaller();
325 DEBUG(dbgs() << " NOT Inlining: cost=" << IC.getCost()
326 << ", thres=" << (IC.getCostDelta() + IC.getCost())
327 << ", Call: " << *CS.getInstruction() << "\n");
328 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() +
329 " too costly to inline (cost=") +
330 Twine(IC.getCost()) + ", threshold=" +
331 Twine(IC.getCostDelta() + IC.getCost()) + ")");
335 // Try to detect the case where the current inlining candidate caller (call
336 // it B) is a static or linkonce-ODR function and is an inlining candidate
337 // elsewhere, and the current candidate callee (call it C) is large enough
338 // that inlining it into B would make B too big to inline later. In these
339 // circumstances it may be best not to inline C into B, but to inline B into
342 // This only applies to static and linkonce-ODR functions because those are
343 // expected to be available for inlining in the translation units where they
344 // are used. Thus we will always have the opportunity to make local inlining
345 // decisions. Importantly the linkonce-ODR linkage covers inline functions
346 // and templates in C++.
348 // FIXME: All of this logic should be sunk into getInlineCost. It relies on
349 // the internal implementation of the inline cost metrics rather than
350 // treating them as truly abstract units etc.
351 if (Caller->hasLocalLinkage() || Caller->hasLinkOnceODRLinkage()) {
352 int TotalSecondaryCost = 0;
353 // The candidate cost to be imposed upon the current function.
354 int CandidateCost = IC.getCost() - (InlineConstants::CallPenalty + 1);
355 // This bool tracks what happens if we do NOT inline C into B.
356 bool callerWillBeRemoved = Caller->hasLocalLinkage();
357 // This bool tracks what happens if we DO inline C into B.
358 bool inliningPreventsSomeOuterInline = false;
359 for (User *U : Caller->users()) {
362 // If this isn't a call to Caller (it could be some other sort
363 // of reference) skip it. Such references will prevent the caller
364 // from being removed.
365 if (!CS2 || CS2.getCalledFunction() != Caller) {
366 callerWillBeRemoved = false;
370 InlineCost IC2 = getInlineCost(CS2);
371 ++NumCallerCallersAnalyzed;
373 callerWillBeRemoved = false;
379 // See if inlining or original callsite would erase the cost delta of
380 // this callsite. We subtract off the penalty for the call instruction,
381 // which we would be deleting.
382 if (IC2.getCostDelta() <= CandidateCost) {
383 inliningPreventsSomeOuterInline = true;
384 TotalSecondaryCost += IC2.getCost();
387 // If all outer calls to Caller would get inlined, the cost for the last
388 // one is set very low by getInlineCost, in anticipation that Caller will
389 // be removed entirely. We did not account for this above unless there
390 // is only one caller of Caller.
391 if (callerWillBeRemoved && !Caller->use_empty())
392 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus;
394 if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost()) {
395 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() <<
396 " Cost = " << IC.getCost() <<
397 ", outer Cost = " << TotalSecondaryCost << '\n');
399 CS, Twine("Not inlining. Cost of inlining " +
400 CS.getCalledFunction()->getName() +
401 " increases the cost of inlining " +
402 CS.getCaller()->getName() + " in other contexts"));
407 DEBUG(dbgs() << " Inlining: cost=" << IC.getCost()
408 << ", thres=" << (IC.getCostDelta() + IC.getCost())
409 << ", Call: " << *CS.getInstruction() << '\n');
411 CS, CS.getCalledFunction()->getName() + Twine(" can be inlined into ") +
412 CS.getCaller()->getName() + " with cost=" + Twine(IC.getCost()) +
413 " (threshold=" + Twine(IC.getCostDelta() + IC.getCost()) + ")");
417 /// Return true if the specified inline history ID
418 /// indicates an inline history that includes the specified function.
419 static bool InlineHistoryIncludes(Function *F, int InlineHistoryID,
420 const SmallVectorImpl<std::pair<Function*, int> > &InlineHistory) {
421 while (InlineHistoryID != -1) {
422 assert(unsigned(InlineHistoryID) < InlineHistory.size() &&
423 "Invalid inline history ID");
424 if (InlineHistory[InlineHistoryID].first == F)
426 InlineHistoryID = InlineHistory[InlineHistoryID].second;
431 bool Inliner::runOnSCC(CallGraphSCC &SCC) {
432 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
433 AssumptionCacheTracker *ACT = &getAnalysis<AssumptionCacheTracker>();
434 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
435 const TargetLibraryInfo *TLI = TLIP ? &TLIP->getTLI() : nullptr;
436 AliasAnalysis *AA = &getAnalysis<AliasAnalysis>();
438 SmallPtrSet<Function*, 8> SCCFunctions;
439 DEBUG(dbgs() << "Inliner visiting SCC:");
440 for (CallGraphNode *Node : SCC) {
441 Function *F = Node->getFunction();
442 if (F) SCCFunctions.insert(F);
443 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
446 // Scan through and identify all call sites ahead of time so that we only
447 // inline call sites in the original functions, not call sites that result
448 // from inlining other functions.
449 SmallVector<std::pair<CallSite, int>, 16> CallSites;
451 // When inlining a callee produces new call sites, we want to keep track of
452 // the fact that they were inlined from the callee. This allows us to avoid
453 // infinite inlining in some obscure cases. To represent this, we use an
454 // index into the InlineHistory vector.
455 SmallVector<std::pair<Function*, int>, 8> InlineHistory;
457 for (CallGraphNode *Node : SCC) {
458 Function *F = Node->getFunction();
461 for (BasicBlock &BB : *F)
462 for (Instruction &I : BB) {
463 CallSite CS(cast<Value>(&I));
464 // If this isn't a call, or it is a call to an intrinsic, it can
466 if (!CS || isa<IntrinsicInst>(I))
469 // If this is a direct call to an external function, we can never inline
470 // it. If it is an indirect call, inlining may resolve it to be a
471 // direct call, so we keep it.
472 if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration())
475 CallSites.push_back(std::make_pair(CS, -1));
479 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
481 // If there are no calls in this function, exit early.
482 if (CallSites.empty())
485 // Now that we have all of the call sites, move the ones to functions in the
486 // current SCC to the end of the list.
487 unsigned FirstCallInSCC = CallSites.size();
488 for (unsigned i = 0; i < FirstCallInSCC; ++i)
489 if (Function *F = CallSites[i].first.getCalledFunction())
490 if (SCCFunctions.count(F))
491 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
494 InlinedArrayAllocasTy InlinedArrayAllocas;
495 InlineFunctionInfo InlineInfo(&CG, AA, ACT);
497 // Now that we have all of the call sites, loop over them and inline them if
498 // it looks profitable to do so.
499 bool Changed = false;
503 // Iterate over the outer loop because inlining functions can cause indirect
504 // calls to become direct calls.
505 // CallSites may be modified inside so ranged for loop can not be used.
506 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
507 CallSite CS = CallSites[CSi].first;
509 Function *Caller = CS.getCaller();
510 Function *Callee = CS.getCalledFunction();
512 // If this call site is dead and it is to a readonly function, we should
513 // just delete the call instead of trying to inline it, regardless of
514 // size. This happens because IPSCCP propagates the result out of the
515 // call and then we're left with the dead call.
516 if (isInstructionTriviallyDead(CS.getInstruction(), TLI)) {
517 DEBUG(dbgs() << " -> Deleting dead call: "
518 << *CS.getInstruction() << "\n");
519 // Update the call graph by deleting the edge from Callee to Caller.
520 CG[Caller]->removeCallEdgeFor(CS);
521 CS.getInstruction()->eraseFromParent();
524 // We can only inline direct calls to non-declarations.
525 if (!Callee || Callee->isDeclaration()) continue;
527 // If this call site was obtained by inlining another function, verify
528 // that the include path for the function did not include the callee
529 // itself. If so, we'd be recursively inlining the same function,
530 // which would provide the same callsites, which would cause us to
531 // infinitely inline.
532 int InlineHistoryID = CallSites[CSi].second;
533 if (InlineHistoryID != -1 &&
534 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory))
537 LLVMContext &CallerCtx = Caller->getContext();
539 // Get DebugLoc to report. CS will be invalid after Inliner.
540 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
542 // If the policy determines that we should inline this function,
544 if (!shouldInline(CS)) {
545 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
546 Twine(Callee->getName() +
547 " will not be inlined into " +
552 // Attempt to inline the function.
553 if (!InlineCallIfPossible(CS, InlineInfo, InlinedArrayAllocas,
554 InlineHistoryID, InsertLifetime)) {
555 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
556 Twine(Callee->getName() +
557 " will not be inlined into " +
563 // Report the inline decision.
564 emitOptimizationRemark(
565 CallerCtx, DEBUG_TYPE, *Caller, DLoc,
566 Twine(Callee->getName() + " inlined into " + Caller->getName()));
568 // If inlining this function gave us any new call sites, throw them
569 // onto our worklist to process. They are useful inline candidates.
570 if (!InlineInfo.InlinedCalls.empty()) {
571 // Create a new inline history entry for this, so that we remember
572 // that these new callsites came about due to inlining Callee.
573 int NewHistoryID = InlineHistory.size();
574 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID));
576 for (Value *Ptr : InlineInfo.InlinedCalls)
577 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID));
581 // If we inlined or deleted the last possible call site to the function,
582 // delete the function body now.
583 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
584 // TODO: Can remove if in SCC now.
585 !SCCFunctions.count(Callee) &&
587 // The function may be apparently dead, but if there are indirect
588 // callgraph references to the node, we cannot delete it yet, this
589 // could invalidate the CGSCC iterator.
590 CG[Callee]->getNumReferences() == 0) {
591 DEBUG(dbgs() << " -> Deleting dead function: "
592 << Callee->getName() << "\n");
593 CallGraphNode *CalleeNode = CG[Callee];
595 // Remove any call graph edges from the callee to its callees.
596 CalleeNode->removeAllCalledFunctions();
598 // Removing the node for callee from the call graph and delete it.
599 delete CG.removeFunctionFromModule(CalleeNode);
603 // Remove this call site from the list. If possible, use
604 // swap/pop_back for efficiency, but do not use it if doing so would
605 // move a call site to a function in this SCC before the
606 // 'FirstCallInSCC' barrier.
607 if (SCC.isSingular()) {
608 CallSites[CSi] = CallSites.back();
609 CallSites.pop_back();
611 CallSites.erase(CallSites.begin()+CSi);
618 } while (LocalChange);
623 /// Remove now-dead linkonce functions at the end of
624 /// processing to avoid breaking the SCC traversal.
625 bool Inliner::doFinalization(CallGraph &CG) {
626 return removeDeadFunctions(CG);
629 /// Remove dead functions that are not included in DNR (Do Not Remove) list.
630 bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) {
631 SmallVector<CallGraphNode*, 16> FunctionsToRemove;
632 SmallVector<CallGraphNode *, 16> DeadFunctionsInComdats;
633 SmallDenseMap<const Comdat *, int, 16> ComdatEntriesAlive;
635 auto RemoveCGN = [&](CallGraphNode *CGN) {
636 // Remove any call graph edges from the function to its callees.
637 CGN->removeAllCalledFunctions();
639 // Remove any edges from the external node to the function's call graph
640 // node. These edges might have been made irrelegant due to
641 // optimization of the program.
642 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
644 // Removing the node for callee from the call graph and delete it.
645 FunctionsToRemove.push_back(CGN);
648 // Scan for all of the functions, looking for ones that should now be removed
649 // from the program. Insert the dead ones in the FunctionsToRemove set.
651 CallGraphNode *CGN = I.second;
652 Function *F = CGN->getFunction();
653 if (!F || F->isDeclaration())
656 // Handle the case when this function is called and we only want to care
657 // about always-inline functions. This is a bit of a hack to share code
658 // between here and the InlineAlways pass.
659 if (AlwaysInlineOnly && !F->hasFnAttribute(Attribute::AlwaysInline))
662 // If the only remaining users of the function are dead constants, remove
664 F->removeDeadConstantUsers();
666 if (!F->isDefTriviallyDead())
669 // It is unsafe to drop a function with discardable linkage from a COMDAT
670 // without also dropping the other members of the COMDAT.
671 // The inliner doesn't visit non-function entities which are in COMDAT
672 // groups so it is unsafe to do so *unless* the linkage is local.
673 if (!F->hasLocalLinkage()) {
674 if (const Comdat *C = F->getComdat()) {
675 --ComdatEntriesAlive[C];
676 DeadFunctionsInComdats.push_back(CGN);
683 if (!DeadFunctionsInComdats.empty()) {
684 // Count up all the entities in COMDAT groups
685 auto ComdatGroupReferenced = [&](const Comdat *C) {
686 auto I = ComdatEntriesAlive.find(C);
687 if (I != ComdatEntriesAlive.end())
690 for (const Function &F : CG.getModule())
691 if (const Comdat *C = F.getComdat())
692 ComdatGroupReferenced(C);
693 for (const GlobalVariable &GV : CG.getModule().globals())
694 if (const Comdat *C = GV.getComdat())
695 ComdatGroupReferenced(C);
696 for (const GlobalAlias &GA : CG.getModule().aliases())
697 if (const Comdat *C = GA.getComdat())
698 ComdatGroupReferenced(C);
699 for (CallGraphNode *CGN : DeadFunctionsInComdats) {
700 Function *F = CGN->getFunction();
701 const Comdat *C = F->getComdat();
702 int NumAlive = ComdatEntriesAlive[C];
703 // We can remove functions in a COMDAT group if the entire group is dead.
704 assert(NumAlive >= 0);
712 if (FunctionsToRemove.empty())
715 // Now that we know which functions to delete, do so. We didn't want to do
716 // this inline, because that would invalidate our CallGraph::iterator
719 // Note that it doesn't matter that we are iterating over a non-stable order
720 // here to do this, it doesn't matter which order the functions are deleted
722 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
723 FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(),
724 FunctionsToRemove.end()),
725 FunctionsToRemove.end());
726 for (CallGraphNode *CGN : FunctionsToRemove) {
727 delete CG.removeFunctionFromModule(CGN);