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/Inliner.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/BlockFrequencyInfo.h"
23 #include "llvm/Analysis/CallGraph.h"
24 #include "llvm/Analysis/InlineCost.h"
25 #include "llvm/Analysis/OptimizationDiagnosticInfo.h"
26 #include "llvm/Analysis/ProfileSummaryInfo.h"
27 #include "llvm/Analysis/TargetLibraryInfo.h"
28 #include "llvm/IR/CallSite.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/DiagnosticInfo.h"
31 #include "llvm/IR/InstIterator.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/Transforms/Utils/Cloning.h"
38 #include "llvm/Transforms/Utils/Local.h"
39 #include "llvm/Transforms/Utils/ModuleUtils.h"
42 #define DEBUG_TYPE "inline"
44 STATISTIC(NumInlined, "Number of functions inlined");
45 STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined");
46 STATISTIC(NumDeleted, "Number of functions deleted because all callers found");
47 STATISTIC(NumMergedAllocas, "Number of allocas merged together");
49 // This weirdly named statistic tracks the number of times that, when attempting
50 // to inline a function A into B, we analyze the callers of B in order to see
51 // if those would be more profitable and blocked inline steps.
52 STATISTIC(NumCallerCallersAnalyzed, "Number of caller-callers analyzed");
54 /// Flag to disable manual alloca merging.
56 /// Merging of allocas was originally done as a stack-size saving technique
57 /// prior to LLVM's code generator having support for stack coloring based on
58 /// lifetime markers. It is now in the process of being removed. To experiment
59 /// with disabling it and relying fully on lifetime marker based stack
60 /// coloring, you can pass this flag to LLVM.
62 DisableInlinedAllocaMerging("disable-inlined-alloca-merging",
63 cl::init(false), cl::Hidden);
66 enum class InlinerFunctionImportStatsOpts {
72 cl::opt<InlinerFunctionImportStatsOpts> InlinerFunctionImportStats(
73 "inliner-function-import-stats",
74 cl::init(InlinerFunctionImportStatsOpts::No),
75 cl::values(clEnumValN(InlinerFunctionImportStatsOpts::Basic, "basic",
77 clEnumValN(InlinerFunctionImportStatsOpts::Verbose, "verbose",
78 "printing of statistics for each inlined function")),
79 cl::Hidden, cl::desc("Enable inliner stats for imported functions"));
82 LegacyInlinerBase::LegacyInlinerBase(char &ID)
83 : CallGraphSCCPass(ID), InsertLifetime(true) {}
85 LegacyInlinerBase::LegacyInlinerBase(char &ID, bool InsertLifetime)
86 : CallGraphSCCPass(ID), InsertLifetime(InsertLifetime) {}
88 /// For this class, we declare that we require and preserve the call graph.
89 /// If the derived class implements this method, it should
90 /// always explicitly call the implementation here.
91 void LegacyInlinerBase::getAnalysisUsage(AnalysisUsage &AU) const {
92 AU.addRequired<AssumptionCacheTracker>();
93 AU.addRequired<ProfileSummaryInfoWrapperPass>();
94 AU.addRequired<TargetLibraryInfoWrapperPass>();
95 getAAResultsAnalysisUsage(AU);
96 CallGraphSCCPass::getAnalysisUsage(AU);
99 typedef DenseMap<ArrayType *, std::vector<AllocaInst *>> InlinedArrayAllocasTy;
101 /// Look at all of the allocas that we inlined through this call site. If we
102 /// have already inlined other allocas through other calls into this function,
103 /// then we know that they have disjoint lifetimes and that we can merge them.
105 /// There are many heuristics possible for merging these allocas, and the
106 /// different options have different tradeoffs. One thing that we *really*
107 /// don't want to hurt is SRoA: once inlining happens, often allocas are no
108 /// longer address taken and so they can be promoted.
110 /// Our "solution" for that is to only merge allocas whose outermost type is an
111 /// array type. These are usually not promoted because someone is using a
112 /// variable index into them. These are also often the most important ones to
115 /// A better solution would be to have real memory lifetime markers in the IR
116 /// and not have the inliner do any merging of allocas at all. This would
117 /// allow the backend to do proper stack slot coloring of all allocas that
118 /// *actually make it to the backend*, which is really what we want.
120 /// Because we don't have this information, we do this simple and useful hack.
121 static void mergeInlinedArrayAllocas(
122 Function *Caller, InlineFunctionInfo &IFI,
123 InlinedArrayAllocasTy &InlinedArrayAllocas, int InlineHistory) {
124 SmallPtrSet<AllocaInst *, 16> UsedAllocas;
126 // When processing our SCC, check to see if CS was inlined from some other
127 // call site. For example, if we're processing "A" in this code:
129 // B() { x = alloca ... C() }
130 // C() { y = alloca ... }
131 // Assume that C was not inlined into B initially, and so we're processing A
132 // and decide to inline B into A. Doing this makes an alloca available for
133 // reuse and makes a callsite (C) available for inlining. When we process
134 // the C call site we don't want to do any alloca merging between X and Y
135 // because their scopes are not disjoint. We could make this smarter by
136 // keeping track of the inline history for each alloca in the
137 // InlinedArrayAllocas but this isn't likely to be a significant win.
138 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
141 // Loop over all the allocas we have so far and see if they can be merged with
142 // a previously inlined alloca. If not, remember that we had it.
143 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size(); AllocaNo != e;
145 AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
147 // Don't bother trying to merge array allocations (they will usually be
148 // canonicalized to be an allocation *of* an array), or allocations whose
149 // type is not itself an array (because we're afraid of pessimizing SRoA).
150 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
151 if (!ATy || AI->isArrayAllocation())
154 // Get the list of all available allocas for this array type.
155 std::vector<AllocaInst *> &AllocasForType = InlinedArrayAllocas[ATy];
157 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
158 // that we have to be careful not to reuse the same "available" alloca for
159 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
160 // set to keep track of which "available" allocas are being used by this
161 // function. Also, AllocasForType can be empty of course!
162 bool MergedAwayAlloca = false;
163 for (AllocaInst *AvailableAlloca : AllocasForType) {
165 unsigned Align1 = AI->getAlignment(),
166 Align2 = AvailableAlloca->getAlignment();
168 // The available alloca has to be in the right function, not in some other
169 // function in this SCC.
170 if (AvailableAlloca->getParent() != AI->getParent())
173 // If the inlined function already uses this alloca then we can't reuse
175 if (!UsedAllocas.insert(AvailableAlloca).second)
178 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
180 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI
181 << "\n\t\tINTO: " << *AvailableAlloca << '\n');
183 // Move affected dbg.declare calls immediately after the new alloca to
184 // avoid the situation when a dbg.declare precedes its alloca.
185 if (auto *L = LocalAsMetadata::getIfExists(AI))
186 if (auto *MDV = MetadataAsValue::getIfExists(AI->getContext(), L))
187 for (User *U : MDV->users())
188 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
189 DDI->moveBefore(AvailableAlloca->getNextNode());
191 AI->replaceAllUsesWith(AvailableAlloca);
193 if (Align1 != Align2) {
194 if (!Align1 || !Align2) {
195 const DataLayout &DL = Caller->getParent()->getDataLayout();
196 unsigned TypeAlign = DL.getABITypeAlignment(AI->getAllocatedType());
198 Align1 = Align1 ? Align1 : TypeAlign;
199 Align2 = Align2 ? Align2 : TypeAlign;
203 AvailableAlloca->setAlignment(AI->getAlignment());
206 AI->eraseFromParent();
207 MergedAwayAlloca = true;
209 IFI.StaticAllocas[AllocaNo] = nullptr;
213 // If we already nuked the alloca, we're done with it.
214 if (MergedAwayAlloca)
217 // If we were unable to merge away the alloca either because there are no
218 // allocas of the right type available or because we reused them all
219 // already, remember that this alloca came from an inlined function and mark
220 // it used so we don't reuse it for other allocas from this inline
222 AllocasForType.push_back(AI);
223 UsedAllocas.insert(AI);
227 /// If it is possible to inline the specified call site,
228 /// do so and update the CallGraph for this operation.
230 /// This function also does some basic book-keeping to update the IR. The
231 /// InlinedArrayAllocas map keeps track of any allocas that are already
232 /// available from other functions inlined into the caller. If we are able to
233 /// inline this call site we attempt to reuse already available allocas or add
234 /// any new allocas to the set if not possible.
235 static bool InlineCallIfPossible(
236 CallSite CS, InlineFunctionInfo &IFI,
237 InlinedArrayAllocasTy &InlinedArrayAllocas, int InlineHistory,
238 bool InsertLifetime, function_ref<AAResults &(Function &)> &AARGetter,
239 ImportedFunctionsInliningStatistics &ImportedFunctionsStats) {
240 Function *Callee = CS.getCalledFunction();
241 Function *Caller = CS.getCaller();
243 AAResults &AAR = AARGetter(*Callee);
245 // Try to inline the function. Get the list of static allocas that were
247 if (!InlineFunction(CS, IFI, &AAR, InsertLifetime))
250 if (InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No)
251 ImportedFunctionsStats.recordInline(*Caller, *Callee);
253 AttributeFuncs::mergeAttributesForInlining(*Caller, *Callee);
255 if (!DisableInlinedAllocaMerging)
256 mergeInlinedArrayAllocas(Caller, IFI, InlinedArrayAllocas, InlineHistory);
261 /// Return true if inlining of CS can block the caller from being
262 /// inlined which is proved to be more beneficial. \p IC is the
263 /// estimated inline cost associated with callsite \p CS.
264 /// \p TotalSecondaryCost will be set to the estimated cost of inlining the
265 /// caller if \p CS is suppressed for inlining.
267 shouldBeDeferred(Function *Caller, CallSite CS, InlineCost IC,
268 int &TotalSecondaryCost,
269 function_ref<InlineCost(CallSite CS)> GetInlineCost) {
271 // For now we only handle local or inline functions.
272 if (!Caller->hasLocalLinkage() && !Caller->hasLinkOnceODRLinkage())
274 // Try to detect the case where the current inlining candidate caller (call
275 // it B) is a static or linkonce-ODR function and is an inlining candidate
276 // elsewhere, and the current candidate callee (call it C) is large enough
277 // that inlining it into B would make B too big to inline later. In these
278 // circumstances it may be best not to inline C into B, but to inline B into
281 // This only applies to static and linkonce-ODR functions because those are
282 // expected to be available for inlining in the translation units where they
283 // are used. Thus we will always have the opportunity to make local inlining
284 // decisions. Importantly the linkonce-ODR linkage covers inline functions
285 // and templates in C++.
287 // FIXME: All of this logic should be sunk into getInlineCost. It relies on
288 // the internal implementation of the inline cost metrics rather than
289 // treating them as truly abstract units etc.
290 TotalSecondaryCost = 0;
291 // The candidate cost to be imposed upon the current function.
292 int CandidateCost = IC.getCost() - 1;
293 // This bool tracks what happens if we do NOT inline C into B.
294 bool callerWillBeRemoved = Caller->hasLocalLinkage();
295 // This bool tracks what happens if we DO inline C into B.
296 bool inliningPreventsSomeOuterInline = false;
297 for (User *U : Caller->users()) {
300 // If this isn't a call to Caller (it could be some other sort
301 // of reference) skip it. Such references will prevent the caller
302 // from being removed.
303 if (!CS2 || CS2.getCalledFunction() != Caller) {
304 callerWillBeRemoved = false;
308 InlineCost IC2 = GetInlineCost(CS2);
309 ++NumCallerCallersAnalyzed;
311 callerWillBeRemoved = false;
317 // See if inlining of the original callsite would erase the cost delta of
318 // this callsite. We subtract off the penalty for the call instruction,
319 // which we would be deleting.
320 if (IC2.getCostDelta() <= CandidateCost) {
321 inliningPreventsSomeOuterInline = true;
322 TotalSecondaryCost += IC2.getCost();
325 // If all outer calls to Caller would get inlined, the cost for the last
326 // one is set very low by getInlineCost, in anticipation that Caller will
327 // be removed entirely. We did not account for this above unless there
328 // is only one caller of Caller.
329 if (callerWillBeRemoved && !Caller->hasOneUse())
330 TotalSecondaryCost -= InlineConstants::LastCallToStaticBonus;
332 if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost())
338 /// Return true if the inliner should attempt to inline at the given CallSite.
339 static bool shouldInline(CallSite CS,
340 function_ref<InlineCost(CallSite CS)> GetInlineCost,
341 OptimizationRemarkEmitter &ORE) {
343 InlineCost IC = GetInlineCost(CS);
344 Instruction *Call = CS.getInstruction();
345 Function *Callee = CS.getCalledFunction();
346 Function *Caller = CS.getCaller();
349 DEBUG(dbgs() << " Inlining: cost=always"
350 << ", Call: " << *CS.getInstruction() << "\n");
351 ORE.emit(OptimizationRemarkAnalysis(DEBUG_TYPE, "AlwaysInline", Call)
352 << NV("Callee", Callee)
353 << " should always be inlined (cost=always)");
358 DEBUG(dbgs() << " NOT Inlining: cost=never"
359 << ", Call: " << *CS.getInstruction() << "\n");
360 ORE.emit(OptimizationRemarkMissed(DEBUG_TYPE, "NeverInline", Call)
361 << NV("Callee", Callee) << " not inlined into "
362 << NV("Caller", Caller)
363 << " because it should never be inlined (cost=never)");
368 DEBUG(dbgs() << " NOT Inlining: cost=" << IC.getCost()
369 << ", thres=" << (IC.getCostDelta() + IC.getCost())
370 << ", Call: " << *CS.getInstruction() << "\n");
371 ORE.emit(OptimizationRemarkMissed(DEBUG_TYPE, "TooCostly", Call)
372 << NV("Callee", Callee) << " not inlined into "
373 << NV("Caller", Caller) << " because too costly to inline (cost="
374 << NV("Cost", IC.getCost()) << ", threshold="
375 << NV("Threshold", IC.getCostDelta() + IC.getCost()) << ")");
379 int TotalSecondaryCost = 0;
380 if (shouldBeDeferred(Caller, CS, IC, TotalSecondaryCost, GetInlineCost)) {
381 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction()
382 << " Cost = " << IC.getCost()
383 << ", outer Cost = " << TotalSecondaryCost << '\n');
384 ORE.emit(OptimizationRemarkMissed(DEBUG_TYPE, "IncreaseCostInOtherContexts",
386 << "Not inlining. Cost of inlining " << NV("Callee", Callee)
387 << " increases the cost of inlining " << NV("Caller", Caller)
388 << " in other contexts");
392 DEBUG(dbgs() << " Inlining: cost=" << IC.getCost()
393 << ", thres=" << (IC.getCostDelta() + IC.getCost())
394 << ", Call: " << *CS.getInstruction() << '\n');
395 ORE.emit(OptimizationRemarkAnalysis(DEBUG_TYPE, "CanBeInlined", Call)
396 << NV("Callee", Callee) << " can be inlined into "
397 << NV("Caller", Caller) << " with cost=" << NV("Cost", IC.getCost())
399 << NV("Threshold", IC.getCostDelta() + IC.getCost()) << ")");
403 /// Return true if the specified inline history ID
404 /// indicates an inline history that includes the specified function.
405 static bool InlineHistoryIncludes(
406 Function *F, int InlineHistoryID,
407 const SmallVectorImpl<std::pair<Function *, int>> &InlineHistory) {
408 while (InlineHistoryID != -1) {
409 assert(unsigned(InlineHistoryID) < InlineHistory.size() &&
410 "Invalid inline history ID");
411 if (InlineHistory[InlineHistoryID].first == F)
413 InlineHistoryID = InlineHistory[InlineHistoryID].second;
418 bool LegacyInlinerBase::doInitialization(CallGraph &CG) {
419 if (InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No)
420 ImportedFunctionsStats.setModuleInfo(CG.getModule());
421 return false; // No changes to CallGraph.
424 bool LegacyInlinerBase::runOnSCC(CallGraphSCC &SCC) {
427 return inlineCalls(SCC);
431 inlineCallsImpl(CallGraphSCC &SCC, CallGraph &CG,
432 std::function<AssumptionCache &(Function &)> GetAssumptionCache,
433 ProfileSummaryInfo *PSI, TargetLibraryInfo &TLI,
435 function_ref<InlineCost(CallSite CS)> GetInlineCost,
436 function_ref<AAResults &(Function &)> AARGetter,
437 ImportedFunctionsInliningStatistics &ImportedFunctionsStats) {
438 SmallPtrSet<Function *, 8> SCCFunctions;
439 DEBUG(dbgs() << "Inliner visiting SCC:");
440 for (CallGraphNode *Node : SCC) {
441 Function *F = Node->getFunction();
443 SCCFunctions.insert(F);
444 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
447 // Scan through and identify all call sites ahead of time so that we only
448 // inline call sites in the original functions, not call sites that result
449 // from inlining other functions.
450 SmallVector<std::pair<CallSite, int>, 16> CallSites;
452 // When inlining a callee produces new call sites, we want to keep track of
453 // the fact that they were inlined from the callee. This allows us to avoid
454 // infinite inlining in some obscure cases. To represent this, we use an
455 // index into the InlineHistory vector.
456 SmallVector<std::pair<Function *, int>, 8> InlineHistory;
458 for (CallGraphNode *Node : SCC) {
459 Function *F = Node->getFunction();
460 if (!F || F->isDeclaration())
463 OptimizationRemarkEmitter ORE(F);
464 for (BasicBlock &BB : *F)
465 for (Instruction &I : BB) {
466 CallSite CS(cast<Value>(&I));
467 // If this isn't a call, or it is a call to an intrinsic, it can
469 if (!CS || isa<IntrinsicInst>(I))
472 // If this is a direct call to an external function, we can never inline
473 // it. If it is an indirect call, inlining may resolve it to be a
474 // direct call, so we keep it.
475 if (Function *Callee = CS.getCalledFunction())
476 if (Callee->isDeclaration()) {
478 ORE.emit(OptimizationRemarkMissed(DEBUG_TYPE, "NoDefinition", &I)
479 << NV("Callee", Callee) << " will not be inlined into "
480 << NV("Caller", CS.getCaller())
481 << " because its definition is unavailable"
486 CallSites.push_back(std::make_pair(CS, -1));
490 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
492 // If there are no calls in this function, exit early.
493 if (CallSites.empty())
496 // Now that we have all of the call sites, move the ones to functions in the
497 // current SCC to the end of the list.
498 unsigned FirstCallInSCC = CallSites.size();
499 for (unsigned i = 0; i < FirstCallInSCC; ++i)
500 if (Function *F = CallSites[i].first.getCalledFunction())
501 if (SCCFunctions.count(F))
502 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
504 InlinedArrayAllocasTy InlinedArrayAllocas;
505 InlineFunctionInfo InlineInfo(&CG, &GetAssumptionCache);
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: " << *CS.getInstruction()
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())
538 // If this call site was obtained by inlining another function, verify
539 // that the include path for the function did not include the callee
540 // itself. If so, we'd be recursively inlining the same function,
541 // which would provide the same callsites, which would cause us to
542 // infinitely inline.
543 int InlineHistoryID = CallSites[CSi].second;
544 if (InlineHistoryID != -1 &&
545 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory))
548 // Get DebugLoc to report. CS will be invalid after Inliner.
549 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
550 BasicBlock *Block = CS.getParent();
551 // FIXME for new PM: because of the old PM we currently generate ORE and
552 // in turn BFI on demand. With the new PM, the ORE dependency should
553 // just become a regular analysis dependency.
554 OptimizationRemarkEmitter ORE(Caller);
556 // If the policy determines that we should inline this function,
558 if (!shouldInline(CS, GetInlineCost, ORE))
561 // Attempt to inline the function.
563 if (!InlineCallIfPossible(CS, InlineInfo, InlinedArrayAllocas,
564 InlineHistoryID, InsertLifetime, AARGetter,
565 ImportedFunctionsStats)) {
567 OptimizationRemarkMissed(DEBUG_TYPE, "NotInlined", DLoc, Block)
568 << NV("Callee", Callee) << " will not be inlined into "
569 << NV("Caller", Caller));
574 // Report the inline decision.
575 ORE.emit(OptimizationRemark(DEBUG_TYPE, "Inlined", DLoc, Block)
576 << NV("Callee", Callee) << " inlined into "
577 << NV("Caller", Caller));
579 // If inlining this function gave us any new call sites, throw them
580 // onto our worklist to process. They are useful inline candidates.
581 if (!InlineInfo.InlinedCalls.empty()) {
582 // Create a new inline history entry for this, so that we remember
583 // that these new callsites came about due to inlining Callee.
584 int NewHistoryID = InlineHistory.size();
585 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID));
587 for (Value *Ptr : InlineInfo.InlinedCalls)
588 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID));
592 // If we inlined or deleted the last possible call site to the function,
593 // delete the function body now.
594 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
595 // TODO: Can remove if in SCC now.
596 !SCCFunctions.count(Callee) &&
598 // The function may be apparently dead, but if there are indirect
599 // callgraph references to the node, we cannot delete it yet, this
600 // could invalidate the CGSCC iterator.
601 CG[Callee]->getNumReferences() == 0) {
602 DEBUG(dbgs() << " -> Deleting dead function: " << Callee->getName()
604 CallGraphNode *CalleeNode = CG[Callee];
606 // Remove any call graph edges from the callee to its callees.
607 CalleeNode->removeAllCalledFunctions();
609 // Removing the node for callee from the call graph and delete it.
610 delete CG.removeFunctionFromModule(CalleeNode);
614 // Remove this call site from the list. If possible, use
615 // swap/pop_back for efficiency, but do not use it if doing so would
616 // move a call site to a function in this SCC before the
617 // 'FirstCallInSCC' barrier.
618 if (SCC.isSingular()) {
619 CallSites[CSi] = CallSites.back();
620 CallSites.pop_back();
622 CallSites.erase(CallSites.begin() + CSi);
629 } while (LocalChange);
634 bool LegacyInlinerBase::inlineCalls(CallGraphSCC &SCC) {
635 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
636 ACT = &getAnalysis<AssumptionCacheTracker>();
637 PSI = getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
638 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
639 auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & {
640 return ACT->getAssumptionCache(F);
642 return inlineCallsImpl(SCC, CG, GetAssumptionCache, PSI, TLI, InsertLifetime,
643 [this](CallSite CS) { return getInlineCost(CS); },
644 LegacyAARGetter(*this), ImportedFunctionsStats);
647 /// Remove now-dead linkonce functions at the end of
648 /// processing to avoid breaking the SCC traversal.
649 bool LegacyInlinerBase::doFinalization(CallGraph &CG) {
650 if (InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No)
651 ImportedFunctionsStats.dump(InlinerFunctionImportStats ==
652 InlinerFunctionImportStatsOpts::Verbose);
653 return removeDeadFunctions(CG);
656 /// Remove dead functions that are not included in DNR (Do Not Remove) list.
657 bool LegacyInlinerBase::removeDeadFunctions(CallGraph &CG,
658 bool AlwaysInlineOnly) {
659 SmallVector<CallGraphNode *, 16> FunctionsToRemove;
660 SmallVector<Function *, 16> DeadFunctionsInComdats;
662 auto RemoveCGN = [&](CallGraphNode *CGN) {
663 // Remove any call graph edges from the function to its callees.
664 CGN->removeAllCalledFunctions();
666 // Remove any edges from the external node to the function's call graph
667 // node. These edges might have been made irrelegant due to
668 // optimization of the program.
669 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
671 // Removing the node for callee from the call graph and delete it.
672 FunctionsToRemove.push_back(CGN);
675 // Scan for all of the functions, looking for ones that should now be removed
676 // from the program. Insert the dead ones in the FunctionsToRemove set.
677 for (const auto &I : CG) {
678 CallGraphNode *CGN = I.second.get();
679 Function *F = CGN->getFunction();
680 if (!F || F->isDeclaration())
683 // Handle the case when this function is called and we only want to care
684 // about always-inline functions. This is a bit of a hack to share code
685 // between here and the InlineAlways pass.
686 if (AlwaysInlineOnly && !F->hasFnAttribute(Attribute::AlwaysInline))
689 // If the only remaining users of the function are dead constants, remove
691 F->removeDeadConstantUsers();
693 if (!F->isDefTriviallyDead())
696 // It is unsafe to drop a function with discardable linkage from a COMDAT
697 // without also dropping the other members of the COMDAT.
698 // The inliner doesn't visit non-function entities which are in COMDAT
699 // groups so it is unsafe to do so *unless* the linkage is local.
700 if (!F->hasLocalLinkage()) {
701 if (F->hasComdat()) {
702 DeadFunctionsInComdats.push_back(F);
709 if (!DeadFunctionsInComdats.empty()) {
710 // Filter out the functions whose comdats remain alive.
711 filterDeadComdatFunctions(CG.getModule(), DeadFunctionsInComdats);
713 for (Function *F : DeadFunctionsInComdats)
717 if (FunctionsToRemove.empty())
720 // Now that we know which functions to delete, do so. We didn't want to do
721 // this inline, because that would invalidate our CallGraph::iterator
724 // Note that it doesn't matter that we are iterating over a non-stable order
725 // here to do this, it doesn't matter which order the functions are deleted
727 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
728 FunctionsToRemove.erase(
729 std::unique(FunctionsToRemove.begin(), FunctionsToRemove.end()),
730 FunctionsToRemove.end());
731 for (CallGraphNode *CGN : FunctionsToRemove) {
732 delete CG.removeFunctionFromModule(CGN);
738 PreservedAnalyses InlinerPass::run(LazyCallGraph::SCC &InitialC,
739 CGSCCAnalysisManager &AM, LazyCallGraph &CG,
740 CGSCCUpdateResult &UR) {
741 const ModuleAnalysisManager &MAM =
742 AM.getResult<ModuleAnalysisManagerCGSCCProxy>(InitialC, CG).getManager();
743 bool Changed = false;
745 assert(InitialC.size() > 0 && "Cannot handle an empty SCC!");
746 Module &M = *InitialC.begin()->getFunction().getParent();
747 ProfileSummaryInfo *PSI = MAM.getCachedResult<ProfileSummaryAnalysis>(M);
749 // We use a single common worklist for calls across the entire SCC. We
750 // process these in-order and append new calls introduced during inlining to
753 // Note that this particular order of processing is actually critical to
754 // avoid very bad behaviors. Consider *highly connected* call graphs where
755 // each function contains a small amonut of code and a couple of calls to
756 // other functions. Because the LLVM inliner is fundamentally a bottom-up
757 // inliner, it can handle gracefully the fact that these all appear to be
758 // reasonable inlining candidates as it will flatten things until they become
759 // too big to inline, and then move on and flatten another batch.
761 // However, when processing call edges *within* an SCC we cannot rely on this
762 // bottom-up behavior. As a consequence, with heavily connected *SCCs* of
763 // functions we can end up incrementally inlining N calls into each of
764 // N functions because each incremental inlining decision looks good and we
765 // don't have a topological ordering to prevent explosions.
767 // To compensate for this, we don't process transitive edges made immediate
768 // by inlining until we've done one pass of inlining across the entire SCC.
769 // Large, highly connected SCCs still lead to some amount of code bloat in
770 // this model, but it is uniformly spread across all the functions in the SCC
771 // and eventually they all become too large to inline, rather than
772 // incrementally maknig a single function grow in a super linear fashion.
773 SmallVector<std::pair<CallSite, int>, 16> Calls;
775 // Populate the initial list of calls in this SCC.
776 for (auto &N : InitialC) {
777 // We want to generally process call sites top-down in order for
778 // simplifications stemming from replacing the call with the returned value
779 // after inlining to be visible to subsequent inlining decisions.
780 // FIXME: Using instructions sequence is a really bad way to do this.
781 // Instead we should do an actual RPO walk of the function body.
782 for (Instruction &I : instructions(N.getFunction()))
783 if (auto CS = CallSite(&I))
784 if (Function *Callee = CS.getCalledFunction())
785 if (!Callee->isDeclaration())
786 Calls.push_back({CS, -1});
789 return PreservedAnalyses::all();
791 // Capture updatable variables for the current SCC and RefSCC.
793 auto *RC = &C->getOuterRefSCC();
795 // When inlining a callee produces new call sites, we want to keep track of
796 // the fact that they were inlined from the callee. This allows us to avoid
797 // infinite inlining in some obscure cases. To represent this, we use an
798 // index into the InlineHistory vector.
799 SmallVector<std::pair<Function *, int>, 16> InlineHistory;
801 // Track a set vector of inlined callees so that we can augment the caller
802 // with all of their edges in the call graph before pruning out the ones that
803 // got simplified away.
804 SmallSetVector<Function *, 4> InlinedCallees;
806 // Track the dead functions to delete once finished with inlining calls. We
807 // defer deleting these to make it easier to handle the call graph updates.
808 SmallVector<Function *, 4> DeadFunctions;
810 // Loop forward over all of the calls. Note that we cannot cache the size as
811 // inlining can introduce new calls that need to be processed.
812 for (int i = 0; i < (int)Calls.size(); ++i) {
813 // We expect the calls to typically be batched with sequences of calls that
814 // have the same caller, so we first set up some shared infrastructure for
815 // this caller. We also do any pruning we can at this layer on the caller
817 Function &F = *Calls[i].first.getCaller();
818 LazyCallGraph::Node &N = *CG.lookup(F);
819 if (CG.lookupSCC(N) != C)
821 if (F.hasFnAttribute(Attribute::OptimizeNone))
824 DEBUG(dbgs() << "Inlining calls in: " << F.getName() << "\n");
826 // Get a FunctionAnalysisManager via a proxy for this particular node. We
827 // do this each time we visit a node as the SCC may have changed and as
828 // we're going to mutate this particular function we want to make sure the
829 // proxy is in place to forward any invalidation events. We can use the
830 // manager we get here for looking up results for functions other than this
831 // node however because those functions aren't going to be mutated by this
833 FunctionAnalysisManager &FAM =
834 AM.getResult<FunctionAnalysisManagerCGSCCProxy>(*C, CG)
836 std::function<AssumptionCache &(Function &)> GetAssumptionCache =
837 [&](Function &F) -> AssumptionCache & {
838 return FAM.getResult<AssumptionAnalysis>(F);
840 auto GetBFI = [&](Function &F) -> BlockFrequencyInfo & {
841 return FAM.getResult<BlockFrequencyAnalysis>(F);
844 auto GetInlineCost = [&](CallSite CS) {
845 Function &Callee = *CS.getCalledFunction();
846 auto &CalleeTTI = FAM.getResult<TargetIRAnalysis>(Callee);
847 return getInlineCost(CS, Params, CalleeTTI, GetAssumptionCache, {GetBFI},
851 // Get the remarks emission analysis for the caller.
852 auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(F);
854 // Now process as many calls as we have within this caller in the sequnece.
855 // We bail out as soon as the caller has to change so we can update the
856 // call graph and prepare the context of that new caller.
857 bool DidInline = false;
858 for (; i < (int)Calls.size() && Calls[i].first.getCaller() == &F; ++i) {
861 std::tie(CS, InlineHistoryID) = Calls[i];
862 Function &Callee = *CS.getCalledFunction();
864 if (InlineHistoryID != -1 &&
865 InlineHistoryIncludes(&Callee, InlineHistoryID, InlineHistory))
868 // Check whether we want to inline this callsite.
869 if (!shouldInline(CS, GetInlineCost, ORE))
872 // Setup the data structure used to plumb customization into the
873 // `InlineFunction` routine.
874 InlineFunctionInfo IFI(
875 /*cg=*/nullptr, &GetAssumptionCache,
876 &FAM.getResult<BlockFrequencyAnalysis>(*(CS.getCaller())),
877 &FAM.getResult<BlockFrequencyAnalysis>(Callee));
879 if (!InlineFunction(CS, IFI))
882 InlinedCallees.insert(&Callee);
884 // Add any new callsites to defined functions to the worklist.
885 if (!IFI.InlinedCallSites.empty()) {
886 int NewHistoryID = InlineHistory.size();
887 InlineHistory.push_back({&Callee, InlineHistoryID});
888 for (CallSite &CS : reverse(IFI.InlinedCallSites))
889 if (Function *NewCallee = CS.getCalledFunction())
890 if (!NewCallee->isDeclaration())
891 Calls.push_back({CS, NewHistoryID});
894 // Merge the attributes based on the inlining.
895 AttributeFuncs::mergeAttributesForInlining(F, Callee);
897 // For local functions, check whether this makes the callee trivially
898 // dead. In that case, we can drop the body of the function eagerly
899 // which may reduce the number of callers of other functions to one,
900 // changing inline cost thresholds.
901 if (Callee.hasLocalLinkage()) {
902 // To check this we also need to nuke any dead constant uses (perhaps
903 // made dead by this operation on other functions).
904 Callee.removeDeadConstantUsers();
905 if (Callee.use_empty()) {
907 std::remove_if(Calls.begin() + i + 1, Calls.end(),
908 [&Callee](const std::pair<CallSite, int> &Call) {
909 return Call.first.getCaller() == &Callee;
912 // Clear the body and queue the function itself for deletion when we
913 // finish inlining and call graph updates.
914 // Note that after this point, it is an error to do anything other
915 // than use the callee's address or delete it.
916 Callee.dropAllReferences();
917 assert(find(DeadFunctions, &Callee) == DeadFunctions.end() &&
918 "Cannot put cause a function to become dead twice!");
919 DeadFunctions.push_back(&Callee);
924 // Back the call index up by one to put us in a good position to go around
932 // Add all the inlined callees' edges as ref edges to the caller. These are
933 // by definition trivial edges as we always have *some* transitive ref edge
934 // chain. While in some cases these edges are direct calls inside the
935 // callee, they have to be modeled in the inliner as reference edges as
936 // there may be a reference edge anywhere along the chain from the current
937 // caller to the callee that causes the whole thing to appear like
938 // a (transitive) reference edge that will require promotion to a call edge
940 for (Function *InlinedCallee : InlinedCallees) {
941 LazyCallGraph::Node &CalleeN = *CG.lookup(*InlinedCallee);
942 for (LazyCallGraph::Edge &E : *CalleeN)
943 RC->insertTrivialRefEdge(N, E.getNode());
945 InlinedCallees.clear();
947 // At this point, since we have made changes we have at least removed
948 // a call instruction. However, in the process we do some incremental
949 // simplification of the surrounding code. This simplification can
950 // essentially do all of the same things as a function pass and we can
951 // re-use the exact same logic for updating the call graph to reflect the
953 C = &updateCGAndAnalysisManagerForFunctionPass(CG, *C, N, AM, UR);
954 DEBUG(dbgs() << "Updated inlining SCC: " << *C << "\n");
955 RC = &C->getOuterRefSCC();
958 // Now that we've finished inlining all of the calls across this SCC, delete
959 // all of the trivially dead functions, updating the call graph and the CGSCC
960 // pass manager in the process.
962 // Note that this walks a pointer set which has non-deterministic order but
963 // that is OK as all we do is delete things and add pointers to unordered
965 for (Function *DeadF : DeadFunctions) {
966 // Get the necessary information out of the call graph and nuke the
967 // function there. Also, cclear out any cached analyses.
968 auto &DeadC = *CG.lookupSCC(*CG.lookup(*DeadF));
969 FunctionAnalysisManager &FAM =
970 AM.getResult<FunctionAnalysisManagerCGSCCProxy>(DeadC, CG)
974 auto &DeadRC = DeadC.getOuterRefSCC();
975 CG.removeDeadFunction(*DeadF);
977 // Mark the relevant parts of the call graph as invalid so we don't visit
979 UR.InvalidatedSCCs.insert(&DeadC);
980 UR.InvalidatedRefSCCs.insert(&DeadRC);
982 // And delete the actual function from the module.
983 M.getFunctionList().erase(DeadF);
985 return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();