1 //===- Inliner.cpp - Code common to all inliners --------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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.
13 //
14 //===----------------------------------------------------------------------===//
15
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/Analysis/AliasAnalysis.h"
19 #include "llvm/Analysis/AssumptionCache.h"
20 #include "llvm/Analysis/BasicAliasAnalysis.h"
21 #include "llvm/Analysis/CallGraph.h"
22 #include "llvm/Analysis/InlineCost.h"
23 #include "llvm/Analysis/ProfileSummaryInfo.h"
24 #include "llvm/Analysis/TargetLibraryInfo.h"
25 #include "llvm/IR/CallSite.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DiagnosticInfo.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/Module.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/raw_ostream.h"
33 #include "llvm/Transforms/IPO/InlinerPass.h"
34 #include "llvm/Transforms/Utils/Cloning.h"
35 #include "llvm/Transforms/Utils/Local.h"
36 using namespace llvm;
37
38 #define DEBUG_TYPE "inline"
39
40 STATISTIC(NumInlined, "Number of functions inlined");
41 STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined");
42 STATISTIC(NumDeleted, "Number of functions deleted because all callers found");
43 STATISTIC(NumMergedAllocas, "Number of allocas merged together");
44
45 // This weirdly named statistic tracks the number of times that, when attempting
46 // to inline a function A into B, we analyze the callers of B in order to see
47 // if those would be more profitable and blocked inline steps.
48 STATISTIC(NumCallerCallersAnalyzed, "Number of caller-callers analyzed");
49
Inliner(char & ID)50 Inliner::Inliner(char &ID) : CallGraphSCCPass(ID), InsertLifetime(true) {}
51
Inliner(char & ID,bool InsertLifetime)52 Inliner::Inliner(char &ID, bool InsertLifetime)
53 : CallGraphSCCPass(ID), InsertLifetime(InsertLifetime) {}
54
55 /// For this class, we declare that we require and preserve the call graph.
56 /// If the derived class implements this method, it should
57 /// always explicitly call the implementation here.
getAnalysisUsage(AnalysisUsage & AU) const58 void Inliner::getAnalysisUsage(AnalysisUsage &AU) const {
59 AU.addRequired<AssumptionCacheTracker>();
60 AU.addRequired<ProfileSummaryInfoWrapperPass>();
61 AU.addRequired<TargetLibraryInfoWrapperPass>();
62 getAAResultsAnalysisUsage(AU);
63 CallGraphSCCPass::getAnalysisUsage(AU);
64 }
65
66
67 typedef DenseMap<ArrayType*, std::vector<AllocaInst*> >
68 InlinedArrayAllocasTy;
69
70 /// If it is possible to inline the specified call site,
71 /// do so and update the CallGraph for this operation.
72 ///
73 /// This function also does some basic book-keeping to update the IR. The
74 /// InlinedArrayAllocas map keeps track of any allocas that are already
75 /// available from other functions inlined into the caller. If we are able to
76 /// inline this call site we attempt to reuse already available allocas or add
77 /// any new allocas to the set if not possible.
InlineCallIfPossible(Pass & P,CallSite CS,InlineFunctionInfo & IFI,InlinedArrayAllocasTy & InlinedArrayAllocas,int InlineHistory,bool InsertLifetime)78 static bool InlineCallIfPossible(Pass &P, CallSite CS, InlineFunctionInfo &IFI,
79 InlinedArrayAllocasTy &InlinedArrayAllocas,
80 int InlineHistory, bool InsertLifetime) {
81 Function *Callee = CS.getCalledFunction();
82 Function *Caller = CS.getCaller();
83
84 // We need to manually construct BasicAA directly in order to disable
85 // its use of other function analyses.
86 BasicAAResult BAR(createLegacyPMBasicAAResult(P, *Callee));
87
88 // Construct our own AA results for this function. We do this manually to
89 // work around the limitations of the legacy pass manager.
90 AAResults AAR(createLegacyPMAAResults(P, *Callee, BAR));
91
92 // Try to inline the function. Get the list of static allocas that were
93 // inlined.
94 if (!InlineFunction(CS, IFI, &AAR, InsertLifetime))
95 return false;
96
97 AttributeFuncs::mergeAttributesForInlining(*Caller, *Callee);
98
99 // Look at all of the allocas that we inlined through this call site. If we
100 // have already inlined other allocas through other calls into this function,
101 // then we know that they have disjoint lifetimes and that we can merge them.
102 //
103 // There are many heuristics possible for merging these allocas, and the
104 // different options have different tradeoffs. One thing that we *really*
105 // don't want to hurt is SRoA: once inlining happens, often allocas are no
106 // longer address taken and so they can be promoted.
107 //
108 // Our "solution" for that is to only merge allocas whose outermost type is an
109 // array type. These are usually not promoted because someone is using a
110 // variable index into them. These are also often the most important ones to
111 // merge.
112 //
113 // A better solution would be to have real memory lifetime markers in the IR
114 // and not have the inliner do any merging of allocas at all. This would
115 // allow the backend to do proper stack slot coloring of all allocas that
116 // *actually make it to the backend*, which is really what we want.
117 //
118 // Because we don't have this information, we do this simple and useful hack.
119 //
120 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
121
122 // When processing our SCC, check to see if CS was inlined from some other
123 // call site. For example, if we're processing "A" in this code:
124 // A() { B() }
125 // B() { x = alloca ... C() }
126 // C() { y = alloca ... }
127 // Assume that C was not inlined into B initially, and so we're processing A
128 // and decide to inline B into A. Doing this makes an alloca available for
129 // reuse and makes a callsite (C) available for inlining. When we process
130 // the C call site we don't want to do any alloca merging between X and Y
131 // because their scopes are not disjoint. We could make this smarter by
132 // keeping track of the inline history for each alloca in the
133 // InlinedArrayAllocas but this isn't likely to be a significant win.
134 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
135 return true;
136
137 // Loop over all the allocas we have so far and see if they can be merged with
138 // a previously inlined alloca. If not, remember that we had it.
139 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
140 AllocaNo != e; ++AllocaNo) {
141 AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
142
143 // Don't bother trying to merge array allocations (they will usually be
144 // canonicalized to be an allocation *of* an array), or allocations whose
145 // type is not itself an array (because we're afraid of pessimizing SRoA).
146 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
147 if (!ATy || AI->isArrayAllocation())
148 continue;
149
150 // Get the list of all available allocas for this array type.
151 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
152
153 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
154 // that we have to be careful not to reuse the same "available" alloca for
155 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
156 // set to keep track of which "available" allocas are being used by this
157 // function. Also, AllocasForType can be empty of course!
158 bool MergedAwayAlloca = false;
159 for (AllocaInst *AvailableAlloca : AllocasForType) {
160
161 unsigned Align1 = AI->getAlignment(),
162 Align2 = AvailableAlloca->getAlignment();
163
164 // The available alloca has to be in the right function, not in some other
165 // function in this SCC.
166 if (AvailableAlloca->getParent() != AI->getParent())
167 continue;
168
169 // If the inlined function already uses this alloca then we can't reuse
170 // it.
171 if (!UsedAllocas.insert(AvailableAlloca).second)
172 continue;
173
174 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
175 // success!
176 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: "
177 << *AvailableAlloca << '\n');
178
179 // Move affected dbg.declare calls immediately after the new alloca to
180 // avoid the situation when a dbg.declare preceeds its alloca.
181 if (auto *L = LocalAsMetadata::getIfExists(AI))
182 if (auto *MDV = MetadataAsValue::getIfExists(AI->getContext(), L))
183 for (User *U : MDV->users())
184 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
185 DDI->moveBefore(AvailableAlloca->getNextNode());
186
187 AI->replaceAllUsesWith(AvailableAlloca);
188
189 if (Align1 != Align2) {
190 if (!Align1 || !Align2) {
191 const DataLayout &DL = Caller->getParent()->getDataLayout();
192 unsigned TypeAlign = DL.getABITypeAlignment(AI->getAllocatedType());
193
194 Align1 = Align1 ? Align1 : TypeAlign;
195 Align2 = Align2 ? Align2 : TypeAlign;
196 }
197
198 if (Align1 > Align2)
199 AvailableAlloca->setAlignment(AI->getAlignment());
200 }
201
202 AI->eraseFromParent();
203 MergedAwayAlloca = true;
204 ++NumMergedAllocas;
205 IFI.StaticAllocas[AllocaNo] = nullptr;
206 break;
207 }
208
209 // If we already nuked the alloca, we're done with it.
210 if (MergedAwayAlloca)
211 continue;
212
213 // If we were unable to merge away the alloca either because there are no
214 // allocas of the right type available or because we reused them all
215 // already, remember that this alloca came from an inlined function and mark
216 // it used so we don't reuse it for other allocas from this inline
217 // operation.
218 AllocasForType.push_back(AI);
219 UsedAllocas.insert(AI);
220 }
221
222 return true;
223 }
224
emitAnalysis(CallSite CS,const Twine & Msg)225 static void emitAnalysis(CallSite CS, const Twine &Msg) {
226 Function *Caller = CS.getCaller();
227 LLVMContext &Ctx = Caller->getContext();
228 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
229 emitOptimizationRemarkAnalysis(Ctx, DEBUG_TYPE, *Caller, DLoc, Msg);
230 }
231
shouldBeDeferred(Function * Caller,CallSite CS,InlineCost IC,int & TotalSecondaryCost)232 bool Inliner::shouldBeDeferred(Function *Caller, CallSite CS, InlineCost IC,
233 int &TotalSecondaryCost) {
234
235 // For now we only handle local or inline functions.
236 if (!Caller->hasLocalLinkage() && !Caller->hasLinkOnceODRLinkage())
237 return false;
238 // Try to detect the case where the current inlining candidate caller (call
239 // it B) is a static or linkonce-ODR function and is an inlining candidate
240 // elsewhere, and the current candidate callee (call it C) is large enough
241 // that inlining it into B would make B too big to inline later. In these
242 // circumstances it may be best not to inline C into B, but to inline B into
243 // its callers.
244 //
245 // This only applies to static and linkonce-ODR functions because those are
246 // expected to be available for inlining in the translation units where they
247 // are used. Thus we will always have the opportunity to make local inlining
248 // decisions. Importantly the linkonce-ODR linkage covers inline functions
249 // and templates in C++.
250 //
251 // FIXME: All of this logic should be sunk into getInlineCost. It relies on
252 // the internal implementation of the inline cost metrics rather than
253 // treating them as truly abstract units etc.
254 TotalSecondaryCost = 0;
255 // The candidate cost to be imposed upon the current function.
256 int CandidateCost = IC.getCost() - (InlineConstants::CallPenalty + 1);
257 // This bool tracks what happens if we do NOT inline C into B.
258 bool callerWillBeRemoved = Caller->hasLocalLinkage();
259 // This bool tracks what happens if we DO inline C into B.
260 bool inliningPreventsSomeOuterInline = false;
261 for (User *U : Caller->users()) {
262 CallSite CS2(U);
263
264 // If this isn't a call to Caller (it could be some other sort
265 // of reference) skip it. Such references will prevent the caller
266 // from being removed.
267 if (!CS2 || CS2.getCalledFunction() != Caller) {
268 callerWillBeRemoved = false;
269 continue;
270 }
271
272 InlineCost IC2 = getInlineCost(CS2);
273 ++NumCallerCallersAnalyzed;
274 if (!IC2) {
275 callerWillBeRemoved = false;
276 continue;
277 }
278 if (IC2.isAlways())
279 continue;
280
281 // See if inlining or original callsite would erase the cost delta of
282 // this callsite. We subtract off the penalty for the call instruction,
283 // which we would be deleting.
284 if (IC2.getCostDelta() <= CandidateCost) {
285 inliningPreventsSomeOuterInline = true;
286 TotalSecondaryCost += IC2.getCost();
287 }
288 }
289 // If all outer calls to Caller would get inlined, the cost for the last
290 // one is set very low by getInlineCost, in anticipation that Caller will
291 // be removed entirely. We did not account for this above unless there
292 // is only one caller of Caller.
293 if (callerWillBeRemoved && !Caller->use_empty())
294 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus;
295
296 if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost())
297 return true;
298
299 return false;
300 }
301
302 /// Return true if the inliner should attempt to inline at the given CallSite.
shouldInline(CallSite CS)303 bool Inliner::shouldInline(CallSite CS) {
304 InlineCost IC = getInlineCost(CS);
305
306 if (IC.isAlways()) {
307 DEBUG(dbgs() << " Inlining: cost=always"
308 << ", Call: " << *CS.getInstruction() << "\n");
309 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName()) +
310 " should always be inlined (cost=always)");
311 return true;
312 }
313
314 if (IC.isNever()) {
315 DEBUG(dbgs() << " NOT Inlining: cost=never"
316 << ", Call: " << *CS.getInstruction() << "\n");
317 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() +
318 " should never be inlined (cost=never)"));
319 return false;
320 }
321
322 Function *Caller = CS.getCaller();
323 if (!IC) {
324 DEBUG(dbgs() << " NOT Inlining: cost=" << IC.getCost()
325 << ", thres=" << (IC.getCostDelta() + IC.getCost())
326 << ", Call: " << *CS.getInstruction() << "\n");
327 emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() +
328 " too costly to inline (cost=") +
329 Twine(IC.getCost()) + ", threshold=" +
330 Twine(IC.getCostDelta() + IC.getCost()) + ")");
331 return false;
332 }
333
334 int TotalSecondaryCost = 0;
335 if (shouldBeDeferred(Caller, CS, IC, TotalSecondaryCost)) {
336 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction()
337 << " Cost = " << IC.getCost()
338 << ", outer Cost = " << TotalSecondaryCost << '\n');
339 emitAnalysis(CS, Twine("Not inlining. Cost of inlining " +
340 CS.getCalledFunction()->getName() +
341 " increases the cost of inlining " +
342 CS.getCaller()->getName() + " in other contexts"));
343 return false;
344 }
345
346 DEBUG(dbgs() << " Inlining: cost=" << IC.getCost()
347 << ", thres=" << (IC.getCostDelta() + IC.getCost())
348 << ", Call: " << *CS.getInstruction() << '\n');
349 emitAnalysis(
350 CS, CS.getCalledFunction()->getName() + Twine(" can be inlined into ") +
351 CS.getCaller()->getName() + " with cost=" + Twine(IC.getCost()) +
352 " (threshold=" + Twine(IC.getCostDelta() + IC.getCost()) + ")");
353 return true;
354 }
355
356 /// Return true if the specified inline history ID
357 /// indicates an inline history that includes the specified function.
InlineHistoryIncludes(Function * F,int InlineHistoryID,const SmallVectorImpl<std::pair<Function *,int>> & InlineHistory)358 static bool InlineHistoryIncludes(Function *F, int InlineHistoryID,
359 const SmallVectorImpl<std::pair<Function*, int> > &InlineHistory) {
360 while (InlineHistoryID != -1) {
361 assert(unsigned(InlineHistoryID) < InlineHistory.size() &&
362 "Invalid inline history ID");
363 if (InlineHistory[InlineHistoryID].first == F)
364 return true;
365 InlineHistoryID = InlineHistory[InlineHistoryID].second;
366 }
367 return false;
368 }
369
runOnSCC(CallGraphSCC & SCC)370 bool Inliner::runOnSCC(CallGraphSCC &SCC) {
371 if (skipSCC(SCC))
372 return false;
373 return inlineCalls(SCC);
374 }
375
inlineCalls(CallGraphSCC & SCC)376 bool Inliner::inlineCalls(CallGraphSCC &SCC) {
377 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
378 ACT = &getAnalysis<AssumptionCacheTracker>();
379 PSI = getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(CG.getModule());
380 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
381
382 SmallPtrSet<Function*, 8> SCCFunctions;
383 DEBUG(dbgs() << "Inliner visiting SCC:");
384 for (CallGraphNode *Node : SCC) {
385 Function *F = Node->getFunction();
386 if (F) SCCFunctions.insert(F);
387 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
388 }
389
390 // Scan through and identify all call sites ahead of time so that we only
391 // inline call sites in the original functions, not call sites that result
392 // from inlining other functions.
393 SmallVector<std::pair<CallSite, int>, 16> CallSites;
394
395 // When inlining a callee produces new call sites, we want to keep track of
396 // the fact that they were inlined from the callee. This allows us to avoid
397 // infinite inlining in some obscure cases. To represent this, we use an
398 // index into the InlineHistory vector.
399 SmallVector<std::pair<Function*, int>, 8> InlineHistory;
400
401 for (CallGraphNode *Node : SCC) {
402 Function *F = Node->getFunction();
403 if (!F) continue;
404
405 for (BasicBlock &BB : *F)
406 for (Instruction &I : BB) {
407 CallSite CS(cast<Value>(&I));
408 // If this isn't a call, or it is a call to an intrinsic, it can
409 // never be inlined.
410 if (!CS || isa<IntrinsicInst>(I))
411 continue;
412
413 // If this is a direct call to an external function, we can never inline
414 // it. If it is an indirect call, inlining may resolve it to be a
415 // direct call, so we keep it.
416 if (Function *Callee = CS.getCalledFunction())
417 if (Callee->isDeclaration())
418 continue;
419
420 CallSites.push_back(std::make_pair(CS, -1));
421 }
422 }
423
424 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
425
426 // If there are no calls in this function, exit early.
427 if (CallSites.empty())
428 return false;
429
430 // Now that we have all of the call sites, move the ones to functions in the
431 // current SCC to the end of the list.
432 unsigned FirstCallInSCC = CallSites.size();
433 for (unsigned i = 0; i < FirstCallInSCC; ++i)
434 if (Function *F = CallSites[i].first.getCalledFunction())
435 if (SCCFunctions.count(F))
436 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
437
438
439 InlinedArrayAllocasTy InlinedArrayAllocas;
440 InlineFunctionInfo InlineInfo(&CG, ACT);
441
442 // Now that we have all of the call sites, loop over them and inline them if
443 // it looks profitable to do so.
444 bool Changed = false;
445 bool LocalChange;
446 do {
447 LocalChange = false;
448 // Iterate over the outer loop because inlining functions can cause indirect
449 // calls to become direct calls.
450 // CallSites may be modified inside so ranged for loop can not be used.
451 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
452 CallSite CS = CallSites[CSi].first;
453
454 Function *Caller = CS.getCaller();
455 Function *Callee = CS.getCalledFunction();
456
457 // If this call site is dead and it is to a readonly function, we should
458 // just delete the call instead of trying to inline it, regardless of
459 // size. This happens because IPSCCP propagates the result out of the
460 // call and then we're left with the dead call.
461 if (isInstructionTriviallyDead(CS.getInstruction(), &TLI)) {
462 DEBUG(dbgs() << " -> Deleting dead call: "
463 << *CS.getInstruction() << "\n");
464 // Update the call graph by deleting the edge from Callee to Caller.
465 CG[Caller]->removeCallEdgeFor(CS);
466 CS.getInstruction()->eraseFromParent();
467 ++NumCallsDeleted;
468 } else {
469 // We can only inline direct calls to non-declarations.
470 if (!Callee || Callee->isDeclaration()) continue;
471
472 // If this call site was obtained by inlining another function, verify
473 // that the include path for the function did not include the callee
474 // itself. If so, we'd be recursively inlining the same function,
475 // which would provide the same callsites, which would cause us to
476 // infinitely inline.
477 int InlineHistoryID = CallSites[CSi].second;
478 if (InlineHistoryID != -1 &&
479 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory))
480 continue;
481
482 LLVMContext &CallerCtx = Caller->getContext();
483
484 // Get DebugLoc to report. CS will be invalid after Inliner.
485 DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
486
487 // If the policy determines that we should inline this function,
488 // try to do so.
489 if (!shouldInline(CS)) {
490 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
491 Twine(Callee->getName() +
492 " will not be inlined into " +
493 Caller->getName()));
494 continue;
495 }
496
497 // Attempt to inline the function.
498 if (!InlineCallIfPossible(*this, CS, InlineInfo, InlinedArrayAllocas,
499 InlineHistoryID, InsertLifetime)) {
500 emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
501 Twine(Callee->getName() +
502 " will not be inlined into " +
503 Caller->getName()));
504 continue;
505 }
506 ++NumInlined;
507
508 // Report the inline decision.
509 emitOptimizationRemark(
510 CallerCtx, DEBUG_TYPE, *Caller, DLoc,
511 Twine(Callee->getName() + " inlined into " + Caller->getName()));
512
513 // If inlining this function gave us any new call sites, throw them
514 // onto our worklist to process. They are useful inline candidates.
515 if (!InlineInfo.InlinedCalls.empty()) {
516 // Create a new inline history entry for this, so that we remember
517 // that these new callsites came about due to inlining Callee.
518 int NewHistoryID = InlineHistory.size();
519 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID));
520
521 for (Value *Ptr : InlineInfo.InlinedCalls)
522 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID));
523 }
524 }
525
526 // If we inlined or deleted the last possible call site to the function,
527 // delete the function body now.
528 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
529 // TODO: Can remove if in SCC now.
530 !SCCFunctions.count(Callee) &&
531
532 // The function may be apparently dead, but if there are indirect
533 // callgraph references to the node, we cannot delete it yet, this
534 // could invalidate the CGSCC iterator.
535 CG[Callee]->getNumReferences() == 0) {
536 DEBUG(dbgs() << " -> Deleting dead function: "
537 << Callee->getName() << "\n");
538 CallGraphNode *CalleeNode = CG[Callee];
539
540 // Remove any call graph edges from the callee to its callees.
541 CalleeNode->removeAllCalledFunctions();
542
543 // Removing the node for callee from the call graph and delete it.
544 delete CG.removeFunctionFromModule(CalleeNode);
545 ++NumDeleted;
546 }
547
548 // Remove this call site from the list. If possible, use
549 // swap/pop_back for efficiency, but do not use it if doing so would
550 // move a call site to a function in this SCC before the
551 // 'FirstCallInSCC' barrier.
552 if (SCC.isSingular()) {
553 CallSites[CSi] = CallSites.back();
554 CallSites.pop_back();
555 } else {
556 CallSites.erase(CallSites.begin()+CSi);
557 }
558 --CSi;
559
560 Changed = true;
561 LocalChange = true;
562 }
563 } while (LocalChange);
564
565 return Changed;
566 }
567
568 /// Remove now-dead linkonce functions at the end of
569 /// processing to avoid breaking the SCC traversal.
doFinalization(CallGraph & CG)570 bool Inliner::doFinalization(CallGraph &CG) {
571 return removeDeadFunctions(CG);
572 }
573
574 /// Remove dead functions that are not included in DNR (Do Not Remove) list.
removeDeadFunctions(CallGraph & CG,bool AlwaysInlineOnly)575 bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) {
576 SmallVector<CallGraphNode*, 16> FunctionsToRemove;
577 SmallVector<CallGraphNode *, 16> DeadFunctionsInComdats;
578 SmallDenseMap<const Comdat *, int, 16> ComdatEntriesAlive;
579
580 auto RemoveCGN = [&](CallGraphNode *CGN) {
581 // Remove any call graph edges from the function to its callees.
582 CGN->removeAllCalledFunctions();
583
584 // Remove any edges from the external node to the function's call graph
585 // node. These edges might have been made irrelegant due to
586 // optimization of the program.
587 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
588
589 // Removing the node for callee from the call graph and delete it.
590 FunctionsToRemove.push_back(CGN);
591 };
592
593 // Scan for all of the functions, looking for ones that should now be removed
594 // from the program. Insert the dead ones in the FunctionsToRemove set.
595 for (const auto &I : CG) {
596 CallGraphNode *CGN = I.second.get();
597 Function *F = CGN->getFunction();
598 if (!F || F->isDeclaration())
599 continue;
600
601 // Handle the case when this function is called and we only want to care
602 // about always-inline functions. This is a bit of a hack to share code
603 // between here and the InlineAlways pass.
604 if (AlwaysInlineOnly && !F->hasFnAttribute(Attribute::AlwaysInline))
605 continue;
606
607 // If the only remaining users of the function are dead constants, remove
608 // them.
609 F->removeDeadConstantUsers();
610
611 if (!F->isDefTriviallyDead())
612 continue;
613
614 // It is unsafe to drop a function with discardable linkage from a COMDAT
615 // without also dropping the other members of the COMDAT.
616 // The inliner doesn't visit non-function entities which are in COMDAT
617 // groups so it is unsafe to do so *unless* the linkage is local.
618 if (!F->hasLocalLinkage()) {
619 if (const Comdat *C = F->getComdat()) {
620 --ComdatEntriesAlive[C];
621 DeadFunctionsInComdats.push_back(CGN);
622 continue;
623 }
624 }
625
626 RemoveCGN(CGN);
627 }
628 if (!DeadFunctionsInComdats.empty()) {
629 // Count up all the entities in COMDAT groups
630 auto ComdatGroupReferenced = [&](const Comdat *C) {
631 auto I = ComdatEntriesAlive.find(C);
632 if (I != ComdatEntriesAlive.end())
633 ++(I->getSecond());
634 };
635 for (const Function &F : CG.getModule())
636 if (const Comdat *C = F.getComdat())
637 ComdatGroupReferenced(C);
638 for (const GlobalVariable &GV : CG.getModule().globals())
639 if (const Comdat *C = GV.getComdat())
640 ComdatGroupReferenced(C);
641 for (const GlobalAlias &GA : CG.getModule().aliases())
642 if (const Comdat *C = GA.getComdat())
643 ComdatGroupReferenced(C);
644 for (CallGraphNode *CGN : DeadFunctionsInComdats) {
645 Function *F = CGN->getFunction();
646 const Comdat *C = F->getComdat();
647 int NumAlive = ComdatEntriesAlive[C];
648 // We can remove functions in a COMDAT group if the entire group is dead.
649 assert(NumAlive >= 0);
650 if (NumAlive > 0)
651 continue;
652
653 RemoveCGN(CGN);
654 }
655 }
656
657 if (FunctionsToRemove.empty())
658 return false;
659
660 // Now that we know which functions to delete, do so. We didn't want to do
661 // this inline, because that would invalidate our CallGraph::iterator
662 // objects. :(
663 //
664 // Note that it doesn't matter that we are iterating over a non-stable order
665 // here to do this, it doesn't matter which order the functions are deleted
666 // in.
667 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
668 FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(),
669 FunctionsToRemove.end()),
670 FunctionsToRemove.end());
671 for (CallGraphNode *CGN : FunctionsToRemove) {
672 delete CG.removeFunctionFromModule(CGN);
673 ++NumDeleted;
674 }
675 return true;
676 }
677