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/DenseMap.h"
18 #include "llvm/ADT/None.h"
19 #include "llvm/ADT/Optional.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringRef.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/Analysis/AssumptionCache.h"
28 #include "llvm/Analysis/BasicAliasAnalysis.h"
29 #include "llvm/Analysis/BlockFrequencyInfo.h"
30 #include "llvm/Analysis/CGSCCPassManager.h"
31 #include "llvm/Analysis/CallGraph.h"
32 #include "llvm/Analysis/InlineCost.h"
33 #include "llvm/Analysis/LazyCallGraph.h"
34 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
35 #include "llvm/Analysis/ProfileSummaryInfo.h"
36 #include "llvm/Analysis/TargetLibraryInfo.h"
37 #include "llvm/Analysis/TargetTransformInfo.h"
38 #include "llvm/IR/Attributes.h"
39 #include "llvm/IR/BasicBlock.h"
40 #include "llvm/IR/CallSite.h"
41 #include "llvm/IR/DataLayout.h"
42 #include "llvm/IR/DebugLoc.h"
43 #include "llvm/IR/DerivedTypes.h"
44 #include "llvm/IR/DiagnosticInfo.h"
45 #include "llvm/IR/Function.h"
46 #include "llvm/IR/InstIterator.h"
47 #include "llvm/IR/Instruction.h"
48 #include "llvm/IR/Instructions.h"
49 #include "llvm/IR/IntrinsicInst.h"
50 #include "llvm/IR/Metadata.h"
51 #include "llvm/IR/Module.h"
52 #include "llvm/IR/PassManager.h"
53 #include "llvm/IR/User.h"
54 #include "llvm/IR/Value.h"
55 #include "llvm/Pass.h"
56 #include "llvm/Support/Casting.h"
57 #include "llvm/Support/CommandLine.h"
58 #include "llvm/Support/Debug.h"
59 #include "llvm/Support/raw_ostream.h"
60 #include "llvm/Transforms/Utils/Cloning.h"
61 #include "llvm/Transforms/Utils/ImportedFunctionsInliningStatistics.h"
62 #include "llvm/Transforms/Utils/Local.h"
63 #include "llvm/Transforms/Utils/ModuleUtils.h"
73 #define DEBUG_TYPE "inline"
75 STATISTIC(NumInlined, "Number of functions inlined");
76 STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined");
77 STATISTIC(NumDeleted, "Number of functions deleted because all callers found");
78 STATISTIC(NumMergedAllocas, "Number of allocas merged together");
80 // This weirdly named statistic tracks the number of times that, when attempting
81 // to inline a function A into B, we analyze the callers of B in order to see
82 // if those would be more profitable and blocked inline steps.
83 STATISTIC(NumCallerCallersAnalyzed, "Number of caller-callers analyzed");
85 /// Flag to disable manual alloca merging.
87 /// Merging of allocas was originally done as a stack-size saving technique
88 /// prior to LLVM's code generator having support for stack coloring based on
89 /// lifetime markers. It is now in the process of being removed. To experiment
90 /// with disabling it and relying fully on lifetime marker based stack
91 /// coloring, you can pass this flag to LLVM.
93 DisableInlinedAllocaMerging("disable-inlined-alloca-merging",
94 cl::init(false), cl::Hidden);
98 enum class InlinerFunctionImportStatsOpts {
104 } // end anonymous namespace
106 static cl::opt<InlinerFunctionImportStatsOpts> InlinerFunctionImportStats(
107 "inliner-function-import-stats",
108 cl::init(InlinerFunctionImportStatsOpts::No),
109 cl::values(clEnumValN(InlinerFunctionImportStatsOpts::Basic, "basic",
111 clEnumValN(InlinerFunctionImportStatsOpts::Verbose, "verbose",
112 "printing of statistics for each inlined function")),
113 cl::Hidden, cl::desc("Enable inliner stats for imported functions"));
115 LegacyInlinerBase::LegacyInlinerBase(char &ID) : CallGraphSCCPass(ID) {}
117 LegacyInlinerBase::LegacyInlinerBase(char &ID, bool InsertLifetime)
118 : CallGraphSCCPass(ID), InsertLifetime(InsertLifetime) {}
120 /// For this class, we declare that we require and preserve the call graph.
121 /// If the derived class implements this method, it should
122 /// always explicitly call the implementation here.
123 void LegacyInlinerBase::getAnalysisUsage(AnalysisUsage &AU) const {
124 AU.addRequired<AssumptionCacheTracker>();
125 AU.addRequired<ProfileSummaryInfoWrapperPass>();
126 AU.addRequired<TargetLibraryInfoWrapperPass>();
127 getAAResultsAnalysisUsage(AU);
128 CallGraphSCCPass::getAnalysisUsage(AU);
131 using InlinedArrayAllocasTy = DenseMap<ArrayType *, std::vector<AllocaInst *>>;
133 /// Look at all of the allocas that we inlined through this call site. If we
134 /// have already inlined other allocas through other calls into this function,
135 /// then we know that they have disjoint lifetimes and that we can merge them.
137 /// There are many heuristics possible for merging these allocas, and the
138 /// different options have different tradeoffs. One thing that we *really*
139 /// don't want to hurt is SRoA: once inlining happens, often allocas are no
140 /// longer address taken and so they can be promoted.
142 /// Our "solution" for that is to only merge allocas whose outermost type is an
143 /// array type. These are usually not promoted because someone is using a
144 /// variable index into them. These are also often the most important ones to
147 /// A better solution would be to have real memory lifetime markers in the IR
148 /// and not have the inliner do any merging of allocas at all. This would
149 /// allow the backend to do proper stack slot coloring of all allocas that
150 /// *actually make it to the backend*, which is really what we want.
152 /// Because we don't have this information, we do this simple and useful hack.
153 static void mergeInlinedArrayAllocas(
154 Function *Caller, InlineFunctionInfo &IFI,
155 InlinedArrayAllocasTy &InlinedArrayAllocas, int InlineHistory) {
156 SmallPtrSet<AllocaInst *, 16> UsedAllocas;
158 // When processing our SCC, check to see if CS was inlined from some other
159 // call site. For example, if we're processing "A" in this code:
161 // B() { x = alloca ... C() }
162 // C() { y = alloca ... }
163 // Assume that C was not inlined into B initially, and so we're processing A
164 // and decide to inline B into A. Doing this makes an alloca available for
165 // reuse and makes a callsite (C) available for inlining. When we process
166 // the C call site we don't want to do any alloca merging between X and Y
167 // because their scopes are not disjoint. We could make this smarter by
168 // keeping track of the inline history for each alloca in the
169 // InlinedArrayAllocas but this isn't likely to be a significant win.
170 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
173 // Loop over all the allocas we have so far and see if they can be merged with
174 // a previously inlined alloca. If not, remember that we had it.
175 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size(); AllocaNo != e;
177 AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
179 // Don't bother trying to merge array allocations (they will usually be
180 // canonicalized to be an allocation *of* an array), or allocations whose
181 // type is not itself an array (because we're afraid of pessimizing SRoA).
182 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
183 if (!ATy || AI->isArrayAllocation())
186 // Get the list of all available allocas for this array type.
187 std::vector<AllocaInst *> &AllocasForType = InlinedArrayAllocas[ATy];
189 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
190 // that we have to be careful not to reuse the same "available" alloca for
191 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
192 // set to keep track of which "available" allocas are being used by this
193 // function. Also, AllocasForType can be empty of course!
194 bool MergedAwayAlloca = false;
195 for (AllocaInst *AvailableAlloca : AllocasForType) {
196 unsigned Align1 = AI->getAlignment(),
197 Align2 = AvailableAlloca->getAlignment();
199 // The available alloca has to be in the right function, not in some other
200 // function in this SCC.
201 if (AvailableAlloca->getParent() != AI->getParent())
204 // If the inlined function already uses this alloca then we can't reuse
206 if (!UsedAllocas.insert(AvailableAlloca).second)
209 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
211 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI
212 << "\n\t\tINTO: " << *AvailableAlloca << '\n');
214 // Move affected dbg.declare calls immediately after the new alloca to
215 // avoid the situation when a dbg.declare precedes its alloca.
216 if (auto *L = LocalAsMetadata::getIfExists(AI))
217 if (auto *MDV = MetadataAsValue::getIfExists(AI->getContext(), L))
218 for (User *U : MDV->users())
219 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
220 DDI->moveBefore(AvailableAlloca->getNextNode());
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);
258 /// If it is possible to inline the specified call site,
259 /// do so and update the CallGraph for this operation.
261 /// This function also does some basic book-keeping to update the IR. The
262 /// InlinedArrayAllocas map keeps track of any allocas that are already
263 /// available from other functions inlined into the caller. If we are able to
264 /// inline this call site we attempt to reuse already available allocas or add
265 /// any new allocas to the set if not possible.
266 static bool InlineCallIfPossible(
267 CallSite CS, InlineFunctionInfo &IFI,
268 InlinedArrayAllocasTy &InlinedArrayAllocas, int InlineHistory,
269 bool InsertLifetime, function_ref<AAResults &(Function &)> &AARGetter,
270 ImportedFunctionsInliningStatistics &ImportedFunctionsStats) {
271 Function *Callee = CS.getCalledFunction();
272 Function *Caller = CS.getCaller();
274 AAResults &AAR = AARGetter(*Callee);
276 // Try to inline the function. Get the list of static allocas that were
278 if (!InlineFunction(CS, IFI, &AAR, InsertLifetime))
281 if (InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No)
282 ImportedFunctionsStats.recordInline(*Caller, *Callee);
284 AttributeFuncs::mergeAttributesForInlining(*Caller, *Callee);
286 if (!DisableInlinedAllocaMerging)
287 mergeInlinedArrayAllocas(Caller, IFI, InlinedArrayAllocas, InlineHistory);
292 /// Return true if inlining of CS can block the caller from being
293 /// inlined which is proved to be more beneficial. \p IC is the
294 /// estimated inline cost associated with callsite \p CS.
295 /// \p TotalSecondaryCost will be set to the estimated cost of inlining the
296 /// caller if \p CS is suppressed for inlining.
298 shouldBeDeferred(Function *Caller, CallSite CS, InlineCost IC,
299 int &TotalSecondaryCost,
300 function_ref<InlineCost(CallSite CS)> GetInlineCost) {
301 // For now we only handle local or inline functions.
302 if (!Caller->hasLocalLinkage() && !Caller->hasLinkOnceODRLinkage())
304 // Try to detect the case where the current inlining candidate caller (call
305 // it B) is a static or linkonce-ODR function and is an inlining candidate
306 // elsewhere, and the current candidate callee (call it C) is large enough
307 // that inlining it into B would make B too big to inline later. In these
308 // circumstances it may be best not to inline C into B, but to inline B into
311 // This only applies to static and linkonce-ODR functions because those are
312 // expected to be available for inlining in the translation units where they
313 // are used. Thus we will always have the opportunity to make local inlining
314 // decisions. Importantly the linkonce-ODR linkage covers inline functions
315 // and templates in C++.
317 // FIXME: All of this logic should be sunk into getInlineCost. It relies on
318 // the internal implementation of the inline cost metrics rather than
319 // treating them as truly abstract units etc.
320 TotalSecondaryCost = 0;
321 // The candidate cost to be imposed upon the current function.
322 int CandidateCost = IC.getCost() - 1;
323 // This bool tracks what happens if we do NOT inline C into B.
324 bool callerWillBeRemoved = Caller->hasLocalLinkage();
325 // This bool tracks what happens if we DO inline C into B.
326 bool inliningPreventsSomeOuterInline = false;
327 for (User *U : Caller->users()) {
330 // If this isn't a call to Caller (it could be some other sort
331 // of reference) skip it. Such references will prevent the caller
332 // from being removed.
333 if (!CS2 || CS2.getCalledFunction() != Caller) {
334 callerWillBeRemoved = false;
338 InlineCost IC2 = GetInlineCost(CS2);
339 ++NumCallerCallersAnalyzed;
341 callerWillBeRemoved = false;
347 // See if inlining of the original callsite would erase the cost delta of
348 // this callsite. We subtract off the penalty for the call instruction,
349 // which we would be deleting.
350 if (IC2.getCostDelta() <= CandidateCost) {
351 inliningPreventsSomeOuterInline = true;
352 TotalSecondaryCost += IC2.getCost();
355 // If all outer calls to Caller would get inlined, the cost for the last
356 // one is set very low by getInlineCost, in anticipation that Caller will
357 // be removed entirely. We did not account for this above unless there
358 // is only one caller of Caller.
359 if (callerWillBeRemoved && !Caller->hasOneUse())
360 TotalSecondaryCost -= InlineConstants::LastCallToStaticBonus;
362 if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost())
368 /// Return the cost only if the inliner should attempt to inline at the given
369 /// CallSite. If we return the cost, we will emit an optimisation remark later
370 /// using that cost, so we won't do so from this function.
371 static Optional<InlineCost>
372 shouldInline(CallSite CS, function_ref<InlineCost(CallSite CS)> GetInlineCost,
373 OptimizationRemarkEmitter &ORE) {
376 InlineCost IC = GetInlineCost(CS);
377 Instruction *Call = CS.getInstruction();
378 Function *Callee = CS.getCalledFunction();
379 Function *Caller = CS.getCaller();
382 DEBUG(dbgs() << " Inlining: cost=always"
383 << ", Call: " << *CS.getInstruction() << "\n");
388 DEBUG(dbgs() << " NOT Inlining: cost=never"
389 << ", Call: " << *CS.getInstruction() << "\n");
391 return OptimizationRemarkMissed(DEBUG_TYPE, "NeverInline", Call)
392 << NV("Callee", Callee) << " not inlined into "
393 << NV("Caller", Caller)
394 << " because it should never be inlined (cost=never)";
400 DEBUG(dbgs() << " NOT Inlining: cost=" << IC.getCost()
401 << ", thres=" << IC.getThreshold()
402 << ", Call: " << *CS.getInstruction() << "\n");
404 return OptimizationRemarkMissed(DEBUG_TYPE, "TooCostly", Call)
405 << NV("Callee", Callee) << " not inlined into "
406 << NV("Caller", Caller) << " because too costly to inline (cost="
407 << NV("Cost", IC.getCost())
408 << ", threshold=" << NV("Threshold", IC.getThreshold()) << ")";
413 int TotalSecondaryCost = 0;
414 if (shouldBeDeferred(Caller, CS, IC, TotalSecondaryCost, GetInlineCost)) {
415 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction()
416 << " Cost = " << IC.getCost()
417 << ", outer Cost = " << TotalSecondaryCost << '\n');
419 return OptimizationRemarkMissed(DEBUG_TYPE, "IncreaseCostInOtherContexts",
421 << "Not inlining. Cost of inlining " << NV("Callee", Callee)
422 << " increases the cost of inlining " << NV("Caller", Caller)
423 << " in other contexts";
426 // IC does not bool() to false, so get an InlineCost that will.
427 // This will not be inspected to make an error message.
431 DEBUG(dbgs() << " Inlining: cost=" << IC.getCost()
432 << ", thres=" << IC.getThreshold()
433 << ", Call: " << *CS.getInstruction() << '\n');
437 /// Return true if the specified inline history ID
438 /// indicates an inline history that includes the specified function.
439 static bool InlineHistoryIncludes(
440 Function *F, int InlineHistoryID,
441 const SmallVectorImpl<std::pair<Function *, int>> &InlineHistory) {
442 while (InlineHistoryID != -1) {
443 assert(unsigned(InlineHistoryID) < InlineHistory.size() &&
444 "Invalid inline history ID");
445 if (InlineHistory[InlineHistoryID].first == F)
447 InlineHistoryID = InlineHistory[InlineHistoryID].second;
452 bool LegacyInlinerBase::doInitialization(CallGraph &CG) {
453 if (InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No)
454 ImportedFunctionsStats.setModuleInfo(CG.getModule());
455 return false; // No changes to CallGraph.
458 bool LegacyInlinerBase::runOnSCC(CallGraphSCC &SCC) {
461 return inlineCalls(SCC);
465 inlineCallsImpl(CallGraphSCC &SCC, CallGraph &CG,
466 std::function<AssumptionCache &(Function &)> GetAssumptionCache,
467 ProfileSummaryInfo *PSI, TargetLibraryInfo &TLI,
469 function_ref<InlineCost(CallSite CS)> GetInlineCost,
470 function_ref<AAResults &(Function &)> AARGetter,
471 ImportedFunctionsInliningStatistics &ImportedFunctionsStats) {
472 SmallPtrSet<Function *, 8> SCCFunctions;
473 DEBUG(dbgs() << "Inliner visiting SCC:");
474 for (CallGraphNode *Node : SCC) {
475 Function *F = Node->getFunction();
477 SCCFunctions.insert(F);
478 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
481 // Scan through and identify all call sites ahead of time so that we only
482 // inline call sites in the original functions, not call sites that result
483 // from inlining other functions.
484 SmallVector<std::pair<CallSite, int>, 16> CallSites;
486 // When inlining a callee produces new call sites, we want to keep track of
487 // the fact that they were inlined from the callee. This allows us to avoid
488 // infinite inlining in some obscure cases. To represent this, we use an
489 // index into the InlineHistory vector.
490 SmallVector<std::pair<Function *, int>, 8> InlineHistory;
492 for (CallGraphNode *Node : SCC) {
493 Function *F = Node->getFunction();
494 if (!F || F->isDeclaration())
497 OptimizationRemarkEmitter ORE(F);
498 for (BasicBlock &BB : *F)
499 for (Instruction &I : BB) {
500 CallSite CS(cast<Value>(&I));
501 // If this isn't a call, or it is a call to an intrinsic, it can
503 if (!CS || isa<IntrinsicInst>(I))
506 // If this is a direct call to an external function, we can never inline
507 // it. If it is an indirect call, inlining may resolve it to be a
508 // direct call, so we keep it.
509 if (Function *Callee = CS.getCalledFunction())
510 if (Callee->isDeclaration()) {
514 return OptimizationRemarkMissed(DEBUG_TYPE, "NoDefinition", &I)
515 << NV("Callee", Callee) << " will not be inlined into "
516 << NV("Caller", CS.getCaller())
517 << " because its definition is unavailable"
523 CallSites.push_back(std::make_pair(CS, -1));
527 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
529 // If there are no calls in this function, exit early.
530 if (CallSites.empty())
533 // Now that we have all of the call sites, move the ones to functions in the
534 // current SCC to the end of the list.
535 unsigned FirstCallInSCC = CallSites.size();
536 for (unsigned i = 0; i < FirstCallInSCC; ++i)
537 if (Function *F = CallSites[i].first.getCalledFunction())
538 if (SCCFunctions.count(F))
539 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
541 InlinedArrayAllocasTy InlinedArrayAllocas;
542 InlineFunctionInfo InlineInfo(&CG, &GetAssumptionCache, PSI);
544 // Now that we have all of the call sites, loop over them and inline them if
545 // it looks profitable to do so.
546 bool Changed = false;
550 // Iterate over the outer loop because inlining functions can cause indirect
551 // calls to become direct calls.
552 // CallSites may be modified inside so ranged for loop can not be used.
553 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
554 CallSite CS = CallSites[CSi].first;
556 Function *Caller = CS.getCaller();
557 Function *Callee = CS.getCalledFunction();
559 // We can only inline direct calls to non-declarations.
560 if (!Callee || Callee->isDeclaration())
563 Instruction *Instr = CS.getInstruction();
565 bool IsTriviallyDead = isInstructionTriviallyDead(Instr, &TLI);
568 if (!IsTriviallyDead) {
569 // If this call site was obtained by inlining another function, verify
570 // that the include path for the function did not include the callee
571 // itself. If so, we'd be recursively inlining the same function,
572 // which would provide the same callsites, which would cause us to
573 // infinitely inline.
574 InlineHistoryID = CallSites[CSi].second;
575 if (InlineHistoryID != -1 &&
576 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory))
580 // FIXME for new PM: because of the old PM we currently generate ORE and
581 // in turn BFI on demand. With the new PM, the ORE dependency should
582 // just become a regular analysis dependency.
583 OptimizationRemarkEmitter ORE(Caller);
585 Optional<InlineCost> OIC = shouldInline(CS, GetInlineCost, ORE);
586 // If the policy determines that we should inline this function,
587 // delete the call instead.
591 // If this call site is dead and it is to a readonly function, we should
592 // just delete the call instead of trying to inline it, regardless of
593 // size. This happens because IPSCCP propagates the result out of the
594 // call and then we're left with the dead call.
595 if (IsTriviallyDead) {
596 DEBUG(dbgs() << " -> Deleting dead call: " << *Instr << "\n");
597 // Update the call graph by deleting the edge from Callee to Caller.
598 CG[Caller]->removeCallEdgeFor(CS);
599 Instr->eraseFromParent();
602 // Get DebugLoc to report. CS will be invalid after Inliner.
603 DebugLoc DLoc = CS->getDebugLoc();
604 BasicBlock *Block = CS.getParent();
606 // Attempt to inline the function.
609 if (!InlineCallIfPossible(CS, InlineInfo, InlinedArrayAllocas,
610 InlineHistoryID, InsertLifetime, AARGetter,
611 ImportedFunctionsStats)) {
613 return OptimizationRemarkMissed(DEBUG_TYPE, "NotInlined", DLoc,
615 << NV("Callee", Callee) << " will not be inlined into "
616 << NV("Caller", Caller);
623 bool AlwaysInline = OIC->isAlways();
624 StringRef RemarkName = AlwaysInline ? "AlwaysInline" : "Inlined";
625 OptimizationRemark R(DEBUG_TYPE, RemarkName, DLoc, Block);
626 R << NV("Callee", Callee) << " inlined into ";
627 R << NV("Caller", Caller);
629 R << " with cost=always";
631 R << " with cost=" << NV("Cost", OIC->getCost());
632 R << " (threshold=" << NV("Threshold", OIC->getThreshold());
638 // If inlining this function gave us any new call sites, throw them
639 // onto our worklist to process. They are useful inline candidates.
640 if (!InlineInfo.InlinedCalls.empty()) {
641 // Create a new inline history entry for this, so that we remember
642 // that these new callsites came about due to inlining Callee.
643 int NewHistoryID = InlineHistory.size();
644 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID));
646 for (Value *Ptr : InlineInfo.InlinedCalls)
647 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID));
651 // If we inlined or deleted the last possible call site to the function,
652 // delete the function body now.
653 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
654 // TODO: Can remove if in SCC now.
655 !SCCFunctions.count(Callee) &&
656 // The function may be apparently dead, but if there are indirect
657 // callgraph references to the node, we cannot delete it yet, this
658 // could invalidate the CGSCC iterator.
659 CG[Callee]->getNumReferences() == 0) {
660 DEBUG(dbgs() << " -> Deleting dead function: " << Callee->getName()
662 CallGraphNode *CalleeNode = CG[Callee];
664 // Remove any call graph edges from the callee to its callees.
665 CalleeNode->removeAllCalledFunctions();
667 // Removing the node for callee from the call graph and delete it.
668 delete CG.removeFunctionFromModule(CalleeNode);
672 // Remove this call site from the list. If possible, use
673 // swap/pop_back for efficiency, but do not use it if doing so would
674 // move a call site to a function in this SCC before the
675 // 'FirstCallInSCC' barrier.
676 if (SCC.isSingular()) {
677 CallSites[CSi] = CallSites.back();
678 CallSites.pop_back();
680 CallSites.erase(CallSites.begin() + CSi);
687 } while (LocalChange);
692 bool LegacyInlinerBase::inlineCalls(CallGraphSCC &SCC) {
693 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
694 ACT = &getAnalysis<AssumptionCacheTracker>();
695 PSI = getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
696 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
697 auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & {
698 return ACT->getAssumptionCache(F);
700 return inlineCallsImpl(SCC, CG, GetAssumptionCache, PSI, TLI, InsertLifetime,
701 [this](CallSite CS) { return getInlineCost(CS); },
702 LegacyAARGetter(*this), ImportedFunctionsStats);
705 /// Remove now-dead linkonce functions at the end of
706 /// processing to avoid breaking the SCC traversal.
707 bool LegacyInlinerBase::doFinalization(CallGraph &CG) {
708 if (InlinerFunctionImportStats != InlinerFunctionImportStatsOpts::No)
709 ImportedFunctionsStats.dump(InlinerFunctionImportStats ==
710 InlinerFunctionImportStatsOpts::Verbose);
711 return removeDeadFunctions(CG);
714 /// Remove dead functions that are not included in DNR (Do Not Remove) list.
715 bool LegacyInlinerBase::removeDeadFunctions(CallGraph &CG,
716 bool AlwaysInlineOnly) {
717 SmallVector<CallGraphNode *, 16> FunctionsToRemove;
718 SmallVector<Function *, 16> DeadFunctionsInComdats;
720 auto RemoveCGN = [&](CallGraphNode *CGN) {
721 // Remove any call graph edges from the function to its callees.
722 CGN->removeAllCalledFunctions();
724 // Remove any edges from the external node to the function's call graph
725 // node. These edges might have been made irrelegant due to
726 // optimization of the program.
727 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
729 // Removing the node for callee from the call graph and delete it.
730 FunctionsToRemove.push_back(CGN);
733 // Scan for all of the functions, looking for ones that should now be removed
734 // from the program. Insert the dead ones in the FunctionsToRemove set.
735 for (const auto &I : CG) {
736 CallGraphNode *CGN = I.second.get();
737 Function *F = CGN->getFunction();
738 if (!F || F->isDeclaration())
741 // Handle the case when this function is called and we only want to care
742 // about always-inline functions. This is a bit of a hack to share code
743 // between here and the InlineAlways pass.
744 if (AlwaysInlineOnly && !F->hasFnAttribute(Attribute::AlwaysInline))
747 // If the only remaining users of the function are dead constants, remove
749 F->removeDeadConstantUsers();
751 if (!F->isDefTriviallyDead())
754 // It is unsafe to drop a function with discardable linkage from a COMDAT
755 // without also dropping the other members of the COMDAT.
756 // The inliner doesn't visit non-function entities which are in COMDAT
757 // groups so it is unsafe to do so *unless* the linkage is local.
758 if (!F->hasLocalLinkage()) {
759 if (F->hasComdat()) {
760 DeadFunctionsInComdats.push_back(F);
767 if (!DeadFunctionsInComdats.empty()) {
768 // Filter out the functions whose comdats remain alive.
769 filterDeadComdatFunctions(CG.getModule(), DeadFunctionsInComdats);
771 for (Function *F : DeadFunctionsInComdats)
775 if (FunctionsToRemove.empty())
778 // Now that we know which functions to delete, do so. We didn't want to do
779 // this inline, because that would invalidate our CallGraph::iterator
782 // Note that it doesn't matter that we are iterating over a non-stable order
783 // here to do this, it doesn't matter which order the functions are deleted
785 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
786 FunctionsToRemove.erase(
787 std::unique(FunctionsToRemove.begin(), FunctionsToRemove.end()),
788 FunctionsToRemove.end());
789 for (CallGraphNode *CGN : FunctionsToRemove) {
790 delete CG.removeFunctionFromModule(CGN);
796 PreservedAnalyses InlinerPass::run(LazyCallGraph::SCC &InitialC,
797 CGSCCAnalysisManager &AM, LazyCallGraph &CG,
798 CGSCCUpdateResult &UR) {
799 const ModuleAnalysisManager &MAM =
800 AM.getResult<ModuleAnalysisManagerCGSCCProxy>(InitialC, CG).getManager();
801 bool Changed = false;
803 assert(InitialC.size() > 0 && "Cannot handle an empty SCC!");
804 Module &M = *InitialC.begin()->getFunction().getParent();
805 ProfileSummaryInfo *PSI = MAM.getCachedResult<ProfileSummaryAnalysis>(M);
807 // We use a single common worklist for calls across the entire SCC. We
808 // process these in-order and append new calls introduced during inlining to
811 // Note that this particular order of processing is actually critical to
812 // avoid very bad behaviors. Consider *highly connected* call graphs where
813 // each function contains a small amonut of code and a couple of calls to
814 // other functions. Because the LLVM inliner is fundamentally a bottom-up
815 // inliner, it can handle gracefully the fact that these all appear to be
816 // reasonable inlining candidates as it will flatten things until they become
817 // too big to inline, and then move on and flatten another batch.
819 // However, when processing call edges *within* an SCC we cannot rely on this
820 // bottom-up behavior. As a consequence, with heavily connected *SCCs* of
821 // functions we can end up incrementally inlining N calls into each of
822 // N functions because each incremental inlining decision looks good and we
823 // don't have a topological ordering to prevent explosions.
825 // To compensate for this, we don't process transitive edges made immediate
826 // by inlining until we've done one pass of inlining across the entire SCC.
827 // Large, highly connected SCCs still lead to some amount of code bloat in
828 // this model, but it is uniformly spread across all the functions in the SCC
829 // and eventually they all become too large to inline, rather than
830 // incrementally maknig a single function grow in a super linear fashion.
831 SmallVector<std::pair<CallSite, int>, 16> Calls;
833 // Populate the initial list of calls in this SCC.
834 for (auto &N : InitialC) {
835 // We want to generally process call sites top-down in order for
836 // simplifications stemming from replacing the call with the returned value
837 // after inlining to be visible to subsequent inlining decisions.
838 // FIXME: Using instructions sequence is a really bad way to do this.
839 // Instead we should do an actual RPO walk of the function body.
840 for (Instruction &I : instructions(N.getFunction()))
841 if (auto CS = CallSite(&I))
842 if (Function *Callee = CS.getCalledFunction())
843 if (!Callee->isDeclaration())
844 Calls.push_back({CS, -1});
847 return PreservedAnalyses::all();
849 // Capture updatable variables for the current SCC and RefSCC.
851 auto *RC = &C->getOuterRefSCC();
853 // When inlining a callee produces new call sites, we want to keep track of
854 // the fact that they were inlined from the callee. This allows us to avoid
855 // infinite inlining in some obscure cases. To represent this, we use an
856 // index into the InlineHistory vector.
857 SmallVector<std::pair<Function *, int>, 16> InlineHistory;
859 // Track a set vector of inlined callees so that we can augment the caller
860 // with all of their edges in the call graph before pruning out the ones that
861 // got simplified away.
862 SmallSetVector<Function *, 4> InlinedCallees;
864 // Track the dead functions to delete once finished with inlining calls. We
865 // defer deleting these to make it easier to handle the call graph updates.
866 SmallVector<Function *, 4> DeadFunctions;
868 // Loop forward over all of the calls. Note that we cannot cache the size as
869 // inlining can introduce new calls that need to be processed.
870 for (int i = 0; i < (int)Calls.size(); ++i) {
871 // We expect the calls to typically be batched with sequences of calls that
872 // have the same caller, so we first set up some shared infrastructure for
873 // this caller. We also do any pruning we can at this layer on the caller
875 Function &F = *Calls[i].first.getCaller();
876 LazyCallGraph::Node &N = *CG.lookup(F);
877 if (CG.lookupSCC(N) != C)
879 if (F.hasFnAttribute(Attribute::OptimizeNone))
882 DEBUG(dbgs() << "Inlining calls in: " << F.getName() << "\n");
884 // Get a FunctionAnalysisManager via a proxy for this particular node. We
885 // do this each time we visit a node as the SCC may have changed and as
886 // we're going to mutate this particular function we want to make sure the
887 // proxy is in place to forward any invalidation events. We can use the
888 // manager we get here for looking up results for functions other than this
889 // node however because those functions aren't going to be mutated by this
891 FunctionAnalysisManager &FAM =
892 AM.getResult<FunctionAnalysisManagerCGSCCProxy>(*C, CG)
895 // Get the remarks emission analysis for the caller.
896 auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(F);
898 std::function<AssumptionCache &(Function &)> GetAssumptionCache =
899 [&](Function &F) -> AssumptionCache & {
900 return FAM.getResult<AssumptionAnalysis>(F);
902 auto GetBFI = [&](Function &F) -> BlockFrequencyInfo & {
903 return FAM.getResult<BlockFrequencyAnalysis>(F);
906 auto GetInlineCost = [&](CallSite CS) {
907 Function &Callee = *CS.getCalledFunction();
908 auto &CalleeTTI = FAM.getResult<TargetIRAnalysis>(Callee);
909 return getInlineCost(CS, Params, CalleeTTI, GetAssumptionCache, {GetBFI},
913 // Now process as many calls as we have within this caller in the sequnece.
914 // We bail out as soon as the caller has to change so we can update the
915 // call graph and prepare the context of that new caller.
916 bool DidInline = false;
917 for (; i < (int)Calls.size() && Calls[i].first.getCaller() == &F; ++i) {
920 std::tie(CS, InlineHistoryID) = Calls[i];
921 Function &Callee = *CS.getCalledFunction();
923 if (InlineHistoryID != -1 &&
924 InlineHistoryIncludes(&Callee, InlineHistoryID, InlineHistory))
927 // Check if this inlining may repeat breaking an SCC apart that has
928 // already been split once before. In that case, inlining here may
929 // trigger infinite inlining, much like is prevented within the inliner
930 // itself by the InlineHistory above, but spread across CGSCC iterations
931 // and thus hidden from the full inline history.
932 if (CG.lookupSCC(*CG.lookup(Callee)) == C &&
933 UR.InlinedInternalEdges.count({&N, C})) {
934 DEBUG(dbgs() << "Skipping inlining internal SCC edge from a node "
935 "previously split out of this SCC by inlining: "
936 << F.getName() << " -> " << Callee.getName() << "\n");
940 Optional<InlineCost> OIC = shouldInline(CS, GetInlineCost, ORE);
941 // Check whether we want to inline this callsite.
945 // Setup the data structure used to plumb customization into the
946 // `InlineFunction` routine.
947 InlineFunctionInfo IFI(
948 /*cg=*/nullptr, &GetAssumptionCache, PSI,
949 &FAM.getResult<BlockFrequencyAnalysis>(*(CS.getCaller())),
950 &FAM.getResult<BlockFrequencyAnalysis>(Callee));
952 // Get DebugLoc to report. CS will be invalid after Inliner.
953 DebugLoc DLoc = CS->getDebugLoc();
954 BasicBlock *Block = CS.getParent();
958 if (!InlineFunction(CS, IFI)) {
960 return OptimizationRemarkMissed(DEBUG_TYPE, "NotInlined", DLoc, Block)
961 << NV("Callee", &Callee) << " will not be inlined into "
967 InlinedCallees.insert(&Callee);
970 bool AlwaysInline = OIC->isAlways();
971 StringRef RemarkName = AlwaysInline ? "AlwaysInline" : "Inlined";
972 OptimizationRemark R(DEBUG_TYPE, RemarkName, DLoc, Block);
973 R << NV("Callee", &Callee) << " inlined into ";
974 R << NV("Caller", &F);
976 R << " with cost=always";
978 R << " with cost=" << NV("Cost", OIC->getCost());
979 R << " (threshold=" << NV("Threshold", OIC->getThreshold());
985 // Add any new callsites to defined functions to the worklist.
986 if (!IFI.InlinedCallSites.empty()) {
987 int NewHistoryID = InlineHistory.size();
988 InlineHistory.push_back({&Callee, InlineHistoryID});
989 for (CallSite &CS : reverse(IFI.InlinedCallSites))
990 if (Function *NewCallee = CS.getCalledFunction())
991 if (!NewCallee->isDeclaration())
992 Calls.push_back({CS, NewHistoryID});
995 // Merge the attributes based on the inlining.
996 AttributeFuncs::mergeAttributesForInlining(F, Callee);
998 // For local functions, check whether this makes the callee trivially
999 // dead. In that case, we can drop the body of the function eagerly
1000 // which may reduce the number of callers of other functions to one,
1001 // changing inline cost thresholds.
1002 if (Callee.hasLocalLinkage()) {
1003 // To check this we also need to nuke any dead constant uses (perhaps
1004 // made dead by this operation on other functions).
1005 Callee.removeDeadConstantUsers();
1006 if (Callee.use_empty() && !CG.isLibFunction(Callee)) {
1008 std::remove_if(Calls.begin() + i + 1, Calls.end(),
1009 [&Callee](const std::pair<CallSite, int> &Call) {
1010 return Call.first.getCaller() == &Callee;
1013 // Clear the body and queue the function itself for deletion when we
1014 // finish inlining and call graph updates.
1015 // Note that after this point, it is an error to do anything other
1016 // than use the callee's address or delete it.
1017 Callee.dropAllReferences();
1018 assert(find(DeadFunctions, &Callee) == DeadFunctions.end() &&
1019 "Cannot put cause a function to become dead twice!");
1020 DeadFunctions.push_back(&Callee);
1025 // Back the call index up by one to put us in a good position to go around
1033 // Add all the inlined callees' edges as ref edges to the caller. These are
1034 // by definition trivial edges as we always have *some* transitive ref edge
1035 // chain. While in some cases these edges are direct calls inside the
1036 // callee, they have to be modeled in the inliner as reference edges as
1037 // there may be a reference edge anywhere along the chain from the current
1038 // caller to the callee that causes the whole thing to appear like
1039 // a (transitive) reference edge that will require promotion to a call edge
1041 for (Function *InlinedCallee : InlinedCallees) {
1042 LazyCallGraph::Node &CalleeN = *CG.lookup(*InlinedCallee);
1043 for (LazyCallGraph::Edge &E : *CalleeN)
1044 RC->insertTrivialRefEdge(N, E.getNode());
1047 // At this point, since we have made changes we have at least removed
1048 // a call instruction. However, in the process we do some incremental
1049 // simplification of the surrounding code. This simplification can
1050 // essentially do all of the same things as a function pass and we can
1051 // re-use the exact same logic for updating the call graph to reflect the
1053 LazyCallGraph::SCC *OldC = C;
1054 C = &updateCGAndAnalysisManagerForFunctionPass(CG, *C, N, AM, UR);
1055 DEBUG(dbgs() << "Updated inlining SCC: " << *C << "\n");
1056 RC = &C->getOuterRefSCC();
1058 // If this causes an SCC to split apart into multiple smaller SCCs, there
1059 // is a subtle risk we need to prepare for. Other transformations may
1060 // expose an "infinite inlining" opportunity later, and because of the SCC
1061 // mutation, we will revisit this function and potentially re-inline. If we
1062 // do, and that re-inlining also has the potentially to mutate the SCC
1063 // structure, the infinite inlining problem can manifest through infinite
1064 // SCC splits and merges. To avoid this, we capture the originating caller
1065 // node and the SCC containing the call edge. This is a slight over
1066 // approximation of the possible inlining decisions that must be avoided,
1067 // but is relatively efficient to store.
1068 // FIXME: This seems like a very heavyweight way of retaining the inline
1069 // history, we should look for a more efficient way of tracking it.
1070 if (C != OldC && llvm::any_of(InlinedCallees, [&](Function *Callee) {
1071 return CG.lookupSCC(*CG.lookup(*Callee)) == OldC;
1073 DEBUG(dbgs() << "Inlined an internal call edge and split an SCC, "
1074 "retaining this to avoid infinite inlining.\n");
1075 UR.InlinedInternalEdges.insert({&N, OldC});
1077 InlinedCallees.clear();
1080 // Now that we've finished inlining all of the calls across this SCC, delete
1081 // all of the trivially dead functions, updating the call graph and the CGSCC
1082 // pass manager in the process.
1084 // Note that this walks a pointer set which has non-deterministic order but
1085 // that is OK as all we do is delete things and add pointers to unordered
1087 for (Function *DeadF : DeadFunctions) {
1088 // Get the necessary information out of the call graph and nuke the
1089 // function there. Also, cclear out any cached analyses.
1090 auto &DeadC = *CG.lookupSCC(*CG.lookup(*DeadF));
1091 FunctionAnalysisManager &FAM =
1092 AM.getResult<FunctionAnalysisManagerCGSCCProxy>(DeadC, CG)
1094 FAM.clear(*DeadF, DeadF->getName());
1095 AM.clear(DeadC, DeadC.getName());
1096 auto &DeadRC = DeadC.getOuterRefSCC();
1097 CG.removeDeadFunction(*DeadF);
1099 // Mark the relevant parts of the call graph as invalid so we don't visit
1101 UR.InvalidatedSCCs.insert(&DeadC);
1102 UR.InvalidatedRefSCCs.insert(&DeadRC);
1104 // And delete the actual function from the module.
1105 M.getFunctionList().erase(DeadF);
1109 return PreservedAnalyses::all();
1111 // Even if we change the IR, we update the core CGSCC data structures and so
1112 // can preserve the proxy to the function analysis manager.
1113 PreservedAnalyses PA;
1114 PA.preserve<FunctionAnalysisManagerCGSCCProxy>();