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 #define DEBUG_TYPE "inline"
17 #include "llvm/Transforms/IPO/InlinerPass.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/CallGraph.h"
21 #include "llvm/Analysis/InlineCost.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/Instructions.h"
24 #include "llvm/IR/IntrinsicInst.h"
25 #include "llvm/IR/Module.h"
26 #include "llvm/Support/CallSite.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/raw_ostream.h"
30 #include "llvm/Target/TargetLibraryInfo.h"
31 #include "llvm/Transforms/Utils/Cloning.h"
32 #include "llvm/Transforms/Utils/Local.h"
35 STATISTIC(NumInlined, "Number of functions inlined");
36 STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined");
37 STATISTIC(NumDeleted, "Number of functions deleted because all callers found");
38 STATISTIC(NumMergedAllocas, "Number of allocas merged together");
40 // This weirdly named statistic tracks the number of times that, when attempting
41 // to inline a function A into B, we analyze the callers of B in order to see
42 // if those would be more profitable and blocked inline steps.
43 STATISTIC(NumCallerCallersAnalyzed, "Number of caller-callers analyzed");
46 InlineLimit("inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore,
47 cl::desc("Control the amount of inlining to perform (default = 225)"));
50 HintThreshold("inlinehint-threshold", cl::Hidden, cl::init(325),
51 cl::desc("Threshold for inlining functions with inline hint"));
53 // Threshold to use when optsize is specified (and there is no -inline-limit).
54 const int OptSizeThreshold = 75;
56 Inliner::Inliner(char &ID)
57 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit), InsertLifetime(true) {}
59 Inliner::Inliner(char &ID, int Threshold, bool InsertLifetime)
60 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit.getNumOccurrences() > 0 ?
61 InlineLimit : Threshold),
62 InsertLifetime(InsertLifetime) {}
64 /// getAnalysisUsage - For this class, we declare that we require and preserve
65 /// the call graph. If the derived class implements this method, it should
66 /// always explicitly call the implementation here.
67 void Inliner::getAnalysisUsage(AnalysisUsage &AU) const {
68 CallGraphSCCPass::getAnalysisUsage(AU);
72 typedef DenseMap<ArrayType*, std::vector<AllocaInst*> >
73 InlinedArrayAllocasTy;
75 /// \brief If the inlined function had a higher stack protection level than the
76 /// calling function, then bump up the caller's stack protection level.
77 static void AdjustCallerSSPLevel(Function *Caller, Function *Callee) {
78 // If upgrading the SSP attribute, clear out the old SSP Attributes first.
79 // Having multiple SSP attributes doesn't actually hurt, but it adds useless
82 B.addAttribute(Attribute::StackProtect)
83 .addAttribute(Attribute::StackProtectStrong);
84 AttributeSet OldSSPAttr = AttributeSet::get(Caller->getContext(),
85 AttributeSet::FunctionIndex,
87 AttributeSet CallerAttr = Caller->getAttributes(),
88 CalleeAttr = Callee->getAttributes();
90 if (CalleeAttr.hasAttribute(AttributeSet::FunctionIndex,
91 Attribute::StackProtectReq)) {
92 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
93 Caller->addFnAttr(Attribute::StackProtectReq);
94 } else if (CalleeAttr.hasAttribute(AttributeSet::FunctionIndex,
95 Attribute::StackProtectStrong) &&
96 !CallerAttr.hasAttribute(AttributeSet::FunctionIndex,
97 Attribute::StackProtectReq)) {
98 Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
99 Caller->addFnAttr(Attribute::StackProtectStrong);
100 } else if (CalleeAttr.hasAttribute(AttributeSet::FunctionIndex,
101 Attribute::StackProtect) &&
102 !CallerAttr.hasAttribute(AttributeSet::FunctionIndex,
103 Attribute::StackProtectReq) &&
104 !CallerAttr.hasAttribute(AttributeSet::FunctionIndex,
105 Attribute::StackProtectStrong))
106 Caller->addFnAttr(Attribute::StackProtect);
109 /// InlineCallIfPossible - If it is possible to inline the specified call site,
110 /// do so and update the CallGraph for this operation.
112 /// This function also does some basic book-keeping to update the IR. The
113 /// InlinedArrayAllocas map keeps track of any allocas that are already
114 /// available from other functions inlined into the caller. If we are able to
115 /// inline this call site we attempt to reuse already available allocas or add
116 /// any new allocas to the set if not possible.
117 static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI,
118 InlinedArrayAllocasTy &InlinedArrayAllocas,
119 int InlineHistory, bool InsertLifetime) {
120 Function *Callee = CS.getCalledFunction();
121 Function *Caller = CS.getCaller();
123 // Try to inline the function. Get the list of static allocas that were
125 if (!InlineFunction(CS, IFI, InsertLifetime))
128 AdjustCallerSSPLevel(Caller, Callee);
130 // Look at all of the allocas that we inlined through this call site. If we
131 // have already inlined other allocas through other calls into this function,
132 // then we know that they have disjoint lifetimes and that we can merge them.
134 // There are many heuristics possible for merging these allocas, and the
135 // different options have different tradeoffs. One thing that we *really*
136 // don't want to hurt is SRoA: once inlining happens, often allocas are no
137 // longer address taken and so they can be promoted.
139 // Our "solution" for that is to only merge allocas whose outermost type is an
140 // array type. These are usually not promoted because someone is using a
141 // variable index into them. These are also often the most important ones to
144 // A better solution would be to have real memory lifetime markers in the IR
145 // and not have the inliner do any merging of allocas at all. This would
146 // allow the backend to do proper stack slot coloring of all allocas that
147 // *actually make it to the backend*, which is really what we want.
149 // Because we don't have this information, we do this simple and useful hack.
151 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
153 // When processing our SCC, check to see if CS was inlined from some other
154 // call site. For example, if we're processing "A" in this code:
156 // B() { x = alloca ... C() }
157 // C() { y = alloca ... }
158 // Assume that C was not inlined into B initially, and so we're processing A
159 // and decide to inline B into A. Doing this makes an alloca available for
160 // reuse and makes a callsite (C) available for inlining. When we process
161 // the C call site we don't want to do any alloca merging between X and Y
162 // because their scopes are not disjoint. We could make this smarter by
163 // keeping track of the inline history for each alloca in the
164 // InlinedArrayAllocas but this isn't likely to be a significant win.
165 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
168 // Loop over all the allocas we have so far and see if they can be merged with
169 // a previously inlined alloca. If not, remember that we had it.
170 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
171 AllocaNo != e; ++AllocaNo) {
172 AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
174 // Don't bother trying to merge array allocations (they will usually be
175 // canonicalized to be an allocation *of* an array), or allocations whose
176 // type is not itself an array (because we're afraid of pessimizing SRoA).
177 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
178 if (ATy == 0 || AI->isArrayAllocation())
181 // Get the list of all available allocas for this array type.
182 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
184 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
185 // that we have to be careful not to reuse the same "available" alloca for
186 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
187 // set to keep track of which "available" allocas are being used by this
188 // function. Also, AllocasForType can be empty of course!
189 bool MergedAwayAlloca = false;
190 for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) {
191 AllocaInst *AvailableAlloca = AllocasForType[i];
193 // The available alloca has to be in the right function, not in some other
194 // function in this SCC.
195 if (AvailableAlloca->getParent() != AI->getParent())
198 // If the inlined function already uses this alloca then we can't reuse
200 if (!UsedAllocas.insert(AvailableAlloca))
203 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
205 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: "
206 << *AvailableAlloca << '\n');
208 AI->replaceAllUsesWith(AvailableAlloca);
209 AI->eraseFromParent();
210 MergedAwayAlloca = true;
212 IFI.StaticAllocas[AllocaNo] = 0;
216 // If we already nuked the alloca, we're done with it.
217 if (MergedAwayAlloca)
220 // If we were unable to merge away the alloca either because there are no
221 // allocas of the right type available or because we reused them all
222 // already, remember that this alloca came from an inlined function and mark
223 // it used so we don't reuse it for other allocas from this inline
225 AllocasForType.push_back(AI);
226 UsedAllocas.insert(AI);
232 unsigned Inliner::getInlineThreshold(CallSite CS) const {
233 int thres = InlineThreshold; // -inline-threshold or else selected by
236 // If -inline-threshold is not given, listen to the optsize attribute when it
237 // would decrease the threshold.
238 Function *Caller = CS.getCaller();
239 bool OptSize = Caller && !Caller->isDeclaration() &&
240 Caller->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
241 Attribute::OptimizeForSize);
242 if (!(InlineLimit.getNumOccurrences() > 0) && OptSize &&
243 OptSizeThreshold < thres)
244 thres = OptSizeThreshold;
246 // Listen to the inlinehint attribute when it would increase the threshold
247 // and the caller does not need to minimize its size.
248 Function *Callee = CS.getCalledFunction();
249 bool InlineHint = Callee && !Callee->isDeclaration() &&
250 Callee->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
251 Attribute::InlineHint);
252 if (InlineHint && HintThreshold > thres
253 && !Caller->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
255 thres = HintThreshold;
260 /// shouldInline - Return true if the inliner should attempt to inline
261 /// at the given CallSite.
262 bool Inliner::shouldInline(CallSite CS) {
263 InlineCost IC = getInlineCost(CS);
266 DEBUG(dbgs() << " Inlining: cost=always"
267 << ", Call: " << *CS.getInstruction() << "\n");
272 DEBUG(dbgs() << " NOT Inlining: cost=never"
273 << ", Call: " << *CS.getInstruction() << "\n");
277 Function *Caller = CS.getCaller();
279 DEBUG(dbgs() << " NOT Inlining: cost=" << IC.getCost()
280 << ", thres=" << (IC.getCostDelta() + IC.getCost())
281 << ", Call: " << *CS.getInstruction() << "\n");
285 // Try to detect the case where the current inlining candidate caller (call
286 // it B) is a static or linkonce-ODR function and is an inlining candidate
287 // elsewhere, and the current candidate callee (call it C) is large enough
288 // that inlining it into B would make B too big to inline later. In these
289 // circumstances it may be best not to inline C into B, but to inline B into
292 // This only applies to static and linkonce-ODR functions because those are
293 // expected to be available for inlining in the translation units where they
294 // are used. Thus we will always have the opportunity to make local inlining
295 // decisions. Importantly the linkonce-ODR linkage covers inline functions
296 // and templates in C++.
298 // FIXME: All of this logic should be sunk into getInlineCost. It relies on
299 // the internal implementation of the inline cost metrics rather than
300 // treating them as truly abstract units etc.
301 if (Caller->hasLocalLinkage() ||
302 Caller->getLinkage() == GlobalValue::LinkOnceODRLinkage) {
303 int TotalSecondaryCost = 0;
304 // The candidate cost to be imposed upon the current function.
305 int CandidateCost = IC.getCost() - (InlineConstants::CallPenalty + 1);
306 // This bool tracks what happens if we do NOT inline C into B.
307 bool callerWillBeRemoved = Caller->hasLocalLinkage();
308 // This bool tracks what happens if we DO inline C into B.
309 bool inliningPreventsSomeOuterInline = false;
310 for (Value::use_iterator I = Caller->use_begin(), E =Caller->use_end();
314 // If this isn't a call to Caller (it could be some other sort
315 // of reference) skip it. Such references will prevent the caller
316 // from being removed.
317 if (!CS2 || CS2.getCalledFunction() != Caller) {
318 callerWillBeRemoved = false;
322 InlineCost IC2 = getInlineCost(CS2);
323 ++NumCallerCallersAnalyzed;
325 callerWillBeRemoved = false;
331 // See if inlining or original callsite would erase the cost delta of
332 // this callsite. We subtract off the penalty for the call instruction,
333 // which we would be deleting.
334 if (IC2.getCostDelta() <= CandidateCost) {
335 inliningPreventsSomeOuterInline = true;
336 TotalSecondaryCost += IC2.getCost();
339 // If all outer calls to Caller would get inlined, the cost for the last
340 // one is set very low by getInlineCost, in anticipation that Caller will
341 // be removed entirely. We did not account for this above unless there
342 // is only one caller of Caller.
343 if (callerWillBeRemoved && Caller->use_begin() != Caller->use_end())
344 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus;
346 if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost()) {
347 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() <<
348 " Cost = " << IC.getCost() <<
349 ", outer Cost = " << TotalSecondaryCost << '\n');
354 DEBUG(dbgs() << " Inlining: cost=" << IC.getCost()
355 << ", thres=" << (IC.getCostDelta() + IC.getCost())
356 << ", Call: " << *CS.getInstruction() << '\n');
360 /// InlineHistoryIncludes - Return true if the specified inline history ID
361 /// indicates an inline history that includes the specified function.
362 static bool InlineHistoryIncludes(Function *F, int InlineHistoryID,
363 const SmallVectorImpl<std::pair<Function*, int> > &InlineHistory) {
364 while (InlineHistoryID != -1) {
365 assert(unsigned(InlineHistoryID) < InlineHistory.size() &&
366 "Invalid inline history ID");
367 if (InlineHistory[InlineHistoryID].first == F)
369 InlineHistoryID = InlineHistory[InlineHistoryID].second;
374 bool Inliner::runOnSCC(CallGraphSCC &SCC) {
375 CallGraph &CG = getAnalysis<CallGraph>();
376 const DataLayout *TD = getAnalysisIfAvailable<DataLayout>();
377 const TargetLibraryInfo *TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
379 SmallPtrSet<Function*, 8> SCCFunctions;
380 DEBUG(dbgs() << "Inliner visiting SCC:");
381 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
382 Function *F = (*I)->getFunction();
383 if (F) SCCFunctions.insert(F);
384 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
387 // Scan through and identify all call sites ahead of time so that we only
388 // inline call sites in the original functions, not call sites that result
389 // from inlining other functions.
390 SmallVector<std::pair<CallSite, int>, 16> CallSites;
392 // When inlining a callee produces new call sites, we want to keep track of
393 // the fact that they were inlined from the callee. This allows us to avoid
394 // infinite inlining in some obscure cases. To represent this, we use an
395 // index into the InlineHistory vector.
396 SmallVector<std::pair<Function*, int>, 8> InlineHistory;
398 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
399 Function *F = (*I)->getFunction();
402 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
403 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
404 CallSite CS(cast<Value>(I));
405 // If this isn't a call, or it is a call to an intrinsic, it can
407 if (!CS || isa<IntrinsicInst>(I))
410 // If this is a direct call to an external function, we can never inline
411 // it. If it is an indirect call, inlining may resolve it to be a
412 // direct call, so we keep it.
413 if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration())
416 CallSites.push_back(std::make_pair(CS, -1));
420 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
422 // If there are no calls in this function, exit early.
423 if (CallSites.empty())
426 // Now that we have all of the call sites, move the ones to functions in the
427 // current SCC to the end of the list.
428 unsigned FirstCallInSCC = CallSites.size();
429 for (unsigned i = 0; i < FirstCallInSCC; ++i)
430 if (Function *F = CallSites[i].first.getCalledFunction())
431 if (SCCFunctions.count(F))
432 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
435 InlinedArrayAllocasTy InlinedArrayAllocas;
436 InlineFunctionInfo InlineInfo(&CG, TD);
438 // Now that we have all of the call sites, loop over them and inline them if
439 // it looks profitable to do so.
440 bool Changed = false;
444 // Iterate over the outer loop because inlining functions can cause indirect
445 // calls to become direct calls.
446 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
447 CallSite CS = CallSites[CSi].first;
449 Function *Caller = CS.getCaller();
450 Function *Callee = CS.getCalledFunction();
452 // If this call site is dead and it is to a readonly function, we should
453 // just delete the call instead of trying to inline it, regardless of
454 // size. This happens because IPSCCP propagates the result out of the
455 // call and then we're left with the dead call.
456 if (isInstructionTriviallyDead(CS.getInstruction(), TLI)) {
457 DEBUG(dbgs() << " -> Deleting dead call: "
458 << *CS.getInstruction() << "\n");
459 // Update the call graph by deleting the edge from Callee to Caller.
460 CG[Caller]->removeCallEdgeFor(CS);
461 CS.getInstruction()->eraseFromParent();
464 // We can only inline direct calls to non-declarations.
465 if (Callee == 0 || Callee->isDeclaration()) continue;
467 // If this call site was obtained by inlining another function, verify
468 // that the include path for the function did not include the callee
469 // itself. If so, we'd be recursively inlining the same function,
470 // which would provide the same callsites, which would cause us to
471 // infinitely inline.
472 int InlineHistoryID = CallSites[CSi].second;
473 if (InlineHistoryID != -1 &&
474 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory))
478 // If the policy determines that we should inline this function,
480 if (!shouldInline(CS))
483 // Attempt to inline the function.
484 if (!InlineCallIfPossible(CS, InlineInfo, InlinedArrayAllocas,
485 InlineHistoryID, InsertLifetime))
489 // If inlining this function gave us any new call sites, throw them
490 // onto our worklist to process. They are useful inline candidates.
491 if (!InlineInfo.InlinedCalls.empty()) {
492 // Create a new inline history entry for this, so that we remember
493 // that these new callsites came about due to inlining Callee.
494 int NewHistoryID = InlineHistory.size();
495 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID));
497 for (unsigned i = 0, e = InlineInfo.InlinedCalls.size();
499 Value *Ptr = InlineInfo.InlinedCalls[i];
500 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID));
505 // If we inlined or deleted the last possible call site to the function,
506 // delete the function body now.
507 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
508 // TODO: Can remove if in SCC now.
509 !SCCFunctions.count(Callee) &&
511 // The function may be apparently dead, but if there are indirect
512 // callgraph references to the node, we cannot delete it yet, this
513 // could invalidate the CGSCC iterator.
514 CG[Callee]->getNumReferences() == 0) {
515 DEBUG(dbgs() << " -> Deleting dead function: "
516 << Callee->getName() << "\n");
517 CallGraphNode *CalleeNode = CG[Callee];
519 // Remove any call graph edges from the callee to its callees.
520 CalleeNode->removeAllCalledFunctions();
522 // Removing the node for callee from the call graph and delete it.
523 delete CG.removeFunctionFromModule(CalleeNode);
527 // Remove this call site from the list. If possible, use
528 // swap/pop_back for efficiency, but do not use it if doing so would
529 // move a call site to a function in this SCC before the
530 // 'FirstCallInSCC' barrier.
531 if (SCC.isSingular()) {
532 CallSites[CSi] = CallSites.back();
533 CallSites.pop_back();
535 CallSites.erase(CallSites.begin()+CSi);
542 } while (LocalChange);
547 // doFinalization - Remove now-dead linkonce functions at the end of
548 // processing to avoid breaking the SCC traversal.
549 bool Inliner::doFinalization(CallGraph &CG) {
550 return removeDeadFunctions(CG);
553 /// removeDeadFunctions - Remove dead functions that are not included in
554 /// DNR (Do Not Remove) list.
555 bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) {
556 SmallVector<CallGraphNode*, 16> FunctionsToRemove;
558 // Scan for all of the functions, looking for ones that should now be removed
559 // from the program. Insert the dead ones in the FunctionsToRemove set.
560 for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) {
561 CallGraphNode *CGN = I->second;
562 Function *F = CGN->getFunction();
563 if (!F || F->isDeclaration())
566 // Handle the case when this function is called and we only want to care
567 // about always-inline functions. This is a bit of a hack to share code
568 // between here and the InlineAlways pass.
569 if (AlwaysInlineOnly &&
570 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
571 Attribute::AlwaysInline))
574 // If the only remaining users of the function are dead constants, remove
576 F->removeDeadConstantUsers();
578 if (!F->isDefTriviallyDead())
581 // Remove any call graph edges from the function to its callees.
582 CGN->removeAllCalledFunctions();
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);
589 // Removing the node for callee from the call graph and delete it.
590 FunctionsToRemove.push_back(CGN);
592 if (FunctionsToRemove.empty())
595 // Now that we know which functions to delete, do so. We didn't want to do
596 // this inline, because that would invalidate our CallGraph::iterator
599 // Note that it doesn't matter that we are iterating over a non-stable order
600 // here to do this, it doesn't matter which order the functions are deleted
602 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
603 FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(),
604 FunctionsToRemove.end()),
605 FunctionsToRemove.end());
606 for (SmallVectorImpl<CallGraphNode *>::iterator I = FunctionsToRemove.begin(),
607 E = FunctionsToRemove.end();
609 delete CG.removeFunctionFromModule(*I);