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/Module.h"
18 #include "llvm/Instructions.h"
19 #include "llvm/IntrinsicInst.h"
20 #include "llvm/Analysis/CallGraph.h"
21 #include "llvm/Analysis/InlineCost.h"
22 #include "llvm/Target/TargetData.h"
23 #include "llvm/Transforms/IPO/InlinerPass.h"
24 #include "llvm/Transforms/Utils/Cloning.h"
25 #include "llvm/Transforms/Utils/Local.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/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/Statistic.h"
34 STATISTIC(NumInlined, "Number of functions inlined");
35 STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined");
36 STATISTIC(NumDeleted, "Number of functions deleted because all callers found");
37 STATISTIC(NumMergedAllocas, "Number of allocas merged together");
39 // This weirdly named statistic tracks the number of times that, when attemting
40 // to inline a function A into B, we analyze the callers of B in order to see
41 // if those would be more profitable and blocked inline steps.
42 STATISTIC(NumCallerCallersAnalyzed, "Number of caller-callers analyzed");
45 InlineLimit("inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore,
46 cl::desc("Control the amount of inlining to perform (default = 225)"));
49 HintThreshold("inlinehint-threshold", cl::Hidden, cl::init(325),
50 cl::desc("Threshold for inlining functions with inline hint"));
52 // Threshold to use when optsize is specified (and there is no -inline-limit).
53 const int OptSizeThreshold = 75;
55 Inliner::Inliner(char &ID)
56 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit), InsertLifetime(true) {}
58 Inliner::Inliner(char &ID, int Threshold, bool InsertLifetime)
59 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit.getNumOccurrences() > 0 ?
60 InlineLimit : Threshold),
61 InsertLifetime(InsertLifetime) {}
63 /// getAnalysisUsage - For this class, we declare that we require and preserve
64 /// the call graph. If the derived class implements this method, it should
65 /// always explicitly call the implementation here.
66 void Inliner::getAnalysisUsage(AnalysisUsage &Info) const {
67 CallGraphSCCPass::getAnalysisUsage(Info);
71 typedef DenseMap<ArrayType*, std::vector<AllocaInst*> >
72 InlinedArrayAllocasTy;
74 /// InlineCallIfPossible - If it is possible to inline the specified call site,
75 /// do so and update the CallGraph for this operation.
77 /// This function also does some basic book-keeping to update the IR. The
78 /// InlinedArrayAllocas map keeps track of any allocas that are already
79 /// available from other functions inlined into the caller. If we are able to
80 /// inline this call site we attempt to reuse already available allocas or add
81 /// any new allocas to the set if not possible.
82 static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI,
83 InlinedArrayAllocasTy &InlinedArrayAllocas,
84 int InlineHistory, bool InsertLifetime) {
85 Function *Callee = CS.getCalledFunction();
86 Function *Caller = CS.getCaller();
88 // Try to inline the function. Get the list of static allocas that were
90 if (!InlineFunction(CS, IFI, InsertLifetime))
93 // If the inlined function had a higher stack protection level than the
94 // calling function, then bump up the caller's stack protection level.
95 if (Callee->hasFnAttr(Attribute::StackProtectReq))
96 Caller->addFnAttr(Attribute::StackProtectReq);
97 else if (Callee->hasFnAttr(Attribute::StackProtect) &&
98 !Caller->hasFnAttr(Attribute::StackProtectReq))
99 Caller->addFnAttr(Attribute::StackProtect);
101 // Look at all of the allocas that we inlined through this call site. If we
102 // have already inlined other allocas through other calls into this function,
103 // then we know that they have disjoint lifetimes and that we can merge them.
105 // There are many heuristics possible for merging these allocas, and the
106 // different options have different tradeoffs. One thing that we *really*
107 // don't want to hurt is SRoA: once inlining happens, often allocas are no
108 // longer address taken and so they can be promoted.
110 // Our "solution" for that is to only merge allocas whose outermost type is an
111 // array type. These are usually not promoted because someone is using a
112 // variable index into them. These are also often the most important ones to
115 // A better solution would be to have real memory lifetime markers in the IR
116 // and not have the inliner do any merging of allocas at all. This would
117 // allow the backend to do proper stack slot coloring of all allocas that
118 // *actually make it to the backend*, which is really what we want.
120 // Because we don't have this information, we do this simple and useful hack.
122 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
124 // When processing our SCC, check to see if CS was inlined from some other
125 // call site. For example, if we're processing "A" in this code:
127 // B() { x = alloca ... C() }
128 // C() { y = alloca ... }
129 // Assume that C was not inlined into B initially, and so we're processing A
130 // and decide to inline B into A. Doing this makes an alloca available for
131 // reuse and makes a callsite (C) available for inlining. When we process
132 // the C call site we don't want to do any alloca merging between X and Y
133 // because their scopes are not disjoint. We could make this smarter by
134 // keeping track of the inline history for each alloca in the
135 // InlinedArrayAllocas but this isn't likely to be a significant win.
136 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
139 // Loop over all the allocas we have so far and see if they can be merged with
140 // a previously inlined alloca. If not, remember that we had it.
141 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
142 AllocaNo != e; ++AllocaNo) {
143 AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
145 // Don't bother trying to merge array allocations (they will usually be
146 // canonicalized to be an allocation *of* an array), or allocations whose
147 // type is not itself an array (because we're afraid of pessimizing SRoA).
148 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
149 if (ATy == 0 || AI->isArrayAllocation())
152 // Get the list of all available allocas for this array type.
153 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
155 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
156 // that we have to be careful not to reuse the same "available" alloca for
157 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
158 // set to keep track of which "available" allocas are being used by this
159 // function. Also, AllocasForType can be empty of course!
160 bool MergedAwayAlloca = false;
161 for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) {
162 AllocaInst *AvailableAlloca = AllocasForType[i];
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())
169 // If the inlined function already uses this alloca then we can't reuse
171 if (!UsedAllocas.insert(AvailableAlloca))
174 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
176 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: "
177 << *AvailableAlloca << '\n');
179 AI->replaceAllUsesWith(AvailableAlloca);
180 AI->eraseFromParent();
181 MergedAwayAlloca = true;
183 IFI.StaticAllocas[AllocaNo] = 0;
187 // If we already nuked the alloca, we're done with it.
188 if (MergedAwayAlloca)
191 // If we were unable to merge away the alloca either because there are no
192 // allocas of the right type available or because we reused them all
193 // already, remember that this alloca came from an inlined function and mark
194 // it used so we don't reuse it for other allocas from this inline
196 AllocasForType.push_back(AI);
197 UsedAllocas.insert(AI);
203 unsigned Inliner::getInlineThreshold(CallSite CS) const {
204 int thres = InlineThreshold; // -inline-threshold or else selected by
207 // If -inline-threshold is not given, listen to the optsize attribute when it
208 // would decrease the threshold.
209 Function *Caller = CS.getCaller();
210 bool OptSize = Caller && !Caller->isDeclaration() &&
211 Caller->hasFnAttr(Attribute::OptimizeForSize);
212 if (!(InlineLimit.getNumOccurrences() > 0) && OptSize && OptSizeThreshold < thres)
213 thres = OptSizeThreshold;
215 // Listen to the inlinehint attribute when it would increase the threshold.
216 Function *Callee = CS.getCalledFunction();
217 bool InlineHint = Callee && !Callee->isDeclaration() &&
218 Callee->hasFnAttr(Attribute::InlineHint);
219 if (InlineHint && HintThreshold > thres)
220 thres = HintThreshold;
225 /// shouldInline - Return true if the inliner should attempt to inline
226 /// at the given CallSite.
227 bool Inliner::shouldInline(CallSite CS) {
228 InlineCost IC = getInlineCost(CS);
231 DEBUG(dbgs() << " Inlining: cost=always"
232 << ", Call: " << *CS.getInstruction() << "\n");
237 DEBUG(dbgs() << " NOT Inlining: cost=never"
238 << ", Call: " << *CS.getInstruction() << "\n");
242 Function *Caller = CS.getCaller();
244 DEBUG(dbgs() << " NOT Inlining: cost=" << IC.getCost()
245 << ", thres=" << (IC.getCostDelta() + IC.getCost())
246 << ", Call: " << *CS.getInstruction() << "\n");
250 // Try to detect the case where the current inlining candidate caller (call
251 // it B) is a static or linkonce-ODR function and is an inlining candidate
252 // elsewhere, and the current candidate callee (call it C) is large enough
253 // that inlining it into B would make B too big to inline later. In these
254 // circumstances it may be best not to inline C into B, but to inline B into
257 // This only applies to static and linkonce-ODR functions because those are
258 // expected to be available for inlining in the translation units where they
259 // are used. Thus we will always have the opportunity to make local inlining
260 // decisions. Importantly the linkonce-ODR linkage covers inline functions
261 // and templates in C++.
263 // FIXME: All of this logic should be sunk into getInlineCost. It relies on
264 // the internal implementation of the inline cost metrics rather than
265 // treating them as truly abstract units etc.
266 if (Caller->hasLocalLinkage() ||
267 Caller->getLinkage() == GlobalValue::LinkOnceODRLinkage) {
268 int TotalSecondaryCost = 0;
269 // The candidate cost to be imposed upon the current function.
270 int CandidateCost = IC.getCost() - (InlineConstants::CallPenalty + 1);
271 // This bool tracks what happens if we do NOT inline C into B.
272 bool callerWillBeRemoved = Caller->hasLocalLinkage();
273 // This bool tracks what happens if we DO inline C into B.
274 bool inliningPreventsSomeOuterInline = false;
275 for (Value::use_iterator I = Caller->use_begin(), E =Caller->use_end();
279 // If this isn't a call to Caller (it could be some other sort
280 // of reference) skip it. Such references will prevent the caller
281 // from being removed.
282 if (!CS2 || CS2.getCalledFunction() != Caller) {
283 callerWillBeRemoved = false;
287 InlineCost IC2 = getInlineCost(CS2);
288 ++NumCallerCallersAnalyzed;
290 callerWillBeRemoved = false;
296 // See if inlining or original callsite would erase the cost delta of
297 // this callsite. We subtract off the penalty for the call instruction,
298 // which we would be deleting.
299 if (IC2.getCostDelta() <= CandidateCost) {
300 inliningPreventsSomeOuterInline = true;
301 TotalSecondaryCost += IC2.getCost();
304 // If all outer calls to Caller would get inlined, the cost for the last
305 // one is set very low by getInlineCost, in anticipation that Caller will
306 // be removed entirely. We did not account for this above unless there
307 // is only one caller of Caller.
308 if (callerWillBeRemoved && Caller->use_begin() != Caller->use_end())
309 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus;
311 if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost()) {
312 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() <<
313 " Cost = " << IC.getCost() <<
314 ", outer Cost = " << TotalSecondaryCost << '\n');
319 DEBUG(dbgs() << " Inlining: cost=" << IC.getCost()
320 << ", thres=" << (IC.getCostDelta() + IC.getCost())
321 << ", Call: " << *CS.getInstruction() << '\n');
325 /// InlineHistoryIncludes - Return true if the specified inline history ID
326 /// indicates an inline history that includes the specified function.
327 static bool InlineHistoryIncludes(Function *F, int InlineHistoryID,
328 const SmallVectorImpl<std::pair<Function*, int> > &InlineHistory) {
329 while (InlineHistoryID != -1) {
330 assert(unsigned(InlineHistoryID) < InlineHistory.size() &&
331 "Invalid inline history ID");
332 if (InlineHistory[InlineHistoryID].first == F)
334 InlineHistoryID = InlineHistory[InlineHistoryID].second;
339 bool Inliner::runOnSCC(CallGraphSCC &SCC) {
340 CallGraph &CG = getAnalysis<CallGraph>();
341 const TargetData *TD = getAnalysisIfAvailable<TargetData>();
343 SmallPtrSet<Function*, 8> SCCFunctions;
344 DEBUG(dbgs() << "Inliner visiting SCC:");
345 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
346 Function *F = (*I)->getFunction();
347 if (F) SCCFunctions.insert(F);
348 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
351 // Scan through and identify all call sites ahead of time so that we only
352 // inline call sites in the original functions, not call sites that result
353 // from inlining other functions.
354 SmallVector<std::pair<CallSite, int>, 16> CallSites;
356 // When inlining a callee produces new call sites, we want to keep track of
357 // the fact that they were inlined from the callee. This allows us to avoid
358 // infinite inlining in some obscure cases. To represent this, we use an
359 // index into the InlineHistory vector.
360 SmallVector<std::pair<Function*, int>, 8> InlineHistory;
362 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
363 Function *F = (*I)->getFunction();
366 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
367 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
368 CallSite CS(cast<Value>(I));
369 // If this isn't a call, or it is a call to an intrinsic, it can
371 if (!CS || isa<IntrinsicInst>(I))
374 // If this is a direct call to an external function, we can never inline
375 // it. If it is an indirect call, inlining may resolve it to be a
376 // direct call, so we keep it.
377 if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration())
380 CallSites.push_back(std::make_pair(CS, -1));
384 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
386 // If there are no calls in this function, exit early.
387 if (CallSites.empty())
390 // Now that we have all of the call sites, move the ones to functions in the
391 // current SCC to the end of the list.
392 unsigned FirstCallInSCC = CallSites.size();
393 for (unsigned i = 0; i < FirstCallInSCC; ++i)
394 if (Function *F = CallSites[i].first.getCalledFunction())
395 if (SCCFunctions.count(F))
396 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
399 InlinedArrayAllocasTy InlinedArrayAllocas;
400 InlineFunctionInfo InlineInfo(&CG, TD);
402 // Now that we have all of the call sites, loop over them and inline them if
403 // it looks profitable to do so.
404 bool Changed = false;
408 // Iterate over the outer loop because inlining functions can cause indirect
409 // calls to become direct calls.
410 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
411 CallSite CS = CallSites[CSi].first;
413 Function *Caller = CS.getCaller();
414 Function *Callee = CS.getCalledFunction();
416 // If this call site is dead and it is to a readonly function, we should
417 // just delete the call instead of trying to inline it, regardless of
418 // size. This happens because IPSCCP propagates the result out of the
419 // call and then we're left with the dead call.
420 if (isInstructionTriviallyDead(CS.getInstruction())) {
421 DEBUG(dbgs() << " -> Deleting dead call: "
422 << *CS.getInstruction() << "\n");
423 // Update the call graph by deleting the edge from Callee to Caller.
424 CG[Caller]->removeCallEdgeFor(CS);
425 CS.getInstruction()->eraseFromParent();
428 // We can only inline direct calls to non-declarations.
429 if (Callee == 0 || Callee->isDeclaration()) continue;
431 // If this call site was obtained by inlining another function, verify
432 // that the include path for the function did not include the callee
433 // itself. If so, we'd be recursively inlining the same function,
434 // which would provide the same callsites, which would cause us to
435 // infinitely inline.
436 int InlineHistoryID = CallSites[CSi].second;
437 if (InlineHistoryID != -1 &&
438 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory))
442 // If the policy determines that we should inline this function,
444 if (!shouldInline(CS))
447 // Attempt to inline the function.
448 if (!InlineCallIfPossible(CS, InlineInfo, InlinedArrayAllocas,
449 InlineHistoryID, InsertLifetime))
453 // If inlining this function gave us any new call sites, throw them
454 // onto our worklist to process. They are useful inline candidates.
455 if (!InlineInfo.InlinedCalls.empty()) {
456 // Create a new inline history entry for this, so that we remember
457 // that these new callsites came about due to inlining Callee.
458 int NewHistoryID = InlineHistory.size();
459 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID));
461 for (unsigned i = 0, e = InlineInfo.InlinedCalls.size();
463 Value *Ptr = InlineInfo.InlinedCalls[i];
464 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID));
469 // If we inlined or deleted the last possible call site to the function,
470 // delete the function body now.
471 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
472 // TODO: Can remove if in SCC now.
473 !SCCFunctions.count(Callee) &&
475 // The function may be apparently dead, but if there are indirect
476 // callgraph references to the node, we cannot delete it yet, this
477 // could invalidate the CGSCC iterator.
478 CG[Callee]->getNumReferences() == 0) {
479 DEBUG(dbgs() << " -> Deleting dead function: "
480 << Callee->getName() << "\n");
481 CallGraphNode *CalleeNode = CG[Callee];
483 // Remove any call graph edges from the callee to its callees.
484 CalleeNode->removeAllCalledFunctions();
486 // Removing the node for callee from the call graph and delete it.
487 delete CG.removeFunctionFromModule(CalleeNode);
491 // Remove this call site from the list. If possible, use
492 // swap/pop_back for efficiency, but do not use it if doing so would
493 // move a call site to a function in this SCC before the
494 // 'FirstCallInSCC' barrier.
495 if (SCC.isSingular()) {
496 CallSites[CSi] = CallSites.back();
497 CallSites.pop_back();
499 CallSites.erase(CallSites.begin()+CSi);
506 } while (LocalChange);
511 // doFinalization - Remove now-dead linkonce functions at the end of
512 // processing to avoid breaking the SCC traversal.
513 bool Inliner::doFinalization(CallGraph &CG) {
514 return removeDeadFunctions(CG);
517 /// removeDeadFunctions - Remove dead functions that are not included in
518 /// DNR (Do Not Remove) list.
519 bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) {
520 SmallVector<CallGraphNode*, 16> FunctionsToRemove;
522 // Scan for all of the functions, looking for ones that should now be removed
523 // from the program. Insert the dead ones in the FunctionsToRemove set.
524 for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) {
525 CallGraphNode *CGN = I->second;
526 Function *F = CGN->getFunction();
527 if (!F || F->isDeclaration())
530 // Handle the case when this function is called and we only want to care
531 // about always-inline functions. This is a bit of a hack to share code
532 // between here and the InlineAlways pass.
533 if (AlwaysInlineOnly && !F->hasFnAttr(Attribute::AlwaysInline))
536 // If the only remaining users of the function are dead constants, remove
538 F->removeDeadConstantUsers();
540 if (!F->isDefTriviallyDead())
543 // Remove any call graph edges from the function to its callees.
544 CGN->removeAllCalledFunctions();
546 // Remove any edges from the external node to the function's call graph
547 // node. These edges might have been made irrelegant due to
548 // optimization of the program.
549 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
551 // Removing the node for callee from the call graph and delete it.
552 FunctionsToRemove.push_back(CGN);
554 if (FunctionsToRemove.empty())
557 // Now that we know which functions to delete, do so. We didn't want to do
558 // this inline, because that would invalidate our CallGraph::iterator
561 // Note that it doesn't matter that we are iterating over a non-stable order
562 // here to do this, it doesn't matter which order the functions are deleted
564 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
565 FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(),
566 FunctionsToRemove.end()),
567 FunctionsToRemove.end());
568 for (SmallVectorImpl<CallGraphNode *>::iterator I = FunctionsToRemove.begin(),
569 E = FunctionsToRemove.end();
571 delete CG.removeFunctionFromModule(*I);