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/Target/TargetLibraryInfo.h"
24 #include "llvm/Transforms/IPO/InlinerPass.h"
25 #include "llvm/Transforms/Utils/Cloning.h"
26 #include "llvm/Transforms/Utils/Local.h"
27 #include "llvm/Support/CallSite.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/ADT/SmallPtrSet.h"
32 #include "llvm/ADT/Statistic.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 &Info) const {
68 CallGraphSCCPass::getAnalysisUsage(Info);
72 typedef DenseMap<ArrayType*, std::vector<AllocaInst*> >
73 InlinedArrayAllocasTy;
75 /// InlineCallIfPossible - If it is possible to inline the specified call site,
76 /// do so and update the CallGraph for this operation.
78 /// This function also does some basic book-keeping to update the IR. The
79 /// InlinedArrayAllocas map keeps track of any allocas that are already
80 /// available from other functions inlined into the caller. If we are able to
81 /// inline this call site we attempt to reuse already available allocas or add
82 /// any new allocas to the set if not possible.
83 static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI,
84 InlinedArrayAllocasTy &InlinedArrayAllocas,
85 int InlineHistory, bool InsertLifetime) {
86 Function *Callee = CS.getCalledFunction();
87 Function *Caller = CS.getCaller();
89 // Try to inline the function. Get the list of static allocas that were
91 if (!InlineFunction(CS, IFI, InsertLifetime))
94 // If the inlined function had a higher stack protection level than the
95 // calling function, then bump up the caller's stack protection level.
96 if (Callee->getFnAttributes().hasStackProtectReqAttr())
97 Caller->addFnAttr(Attribute::StackProtectReq);
98 else if (Callee->getFnAttributes().hasStackProtectAttr() &&
99 !Caller->getFnAttributes().hasStackProtectReqAttr())
100 Caller->addFnAttr(Attribute::StackProtect);
102 // Look at all of the allocas that we inlined through this call site. If we
103 // have already inlined other allocas through other calls into this function,
104 // then we know that they have disjoint lifetimes and that we can merge them.
106 // There are many heuristics possible for merging these allocas, and the
107 // different options have different tradeoffs. One thing that we *really*
108 // don't want to hurt is SRoA: once inlining happens, often allocas are no
109 // longer address taken and so they can be promoted.
111 // Our "solution" for that is to only merge allocas whose outermost type is an
112 // array type. These are usually not promoted because someone is using a
113 // variable index into them. These are also often the most important ones to
116 // A better solution would be to have real memory lifetime markers in the IR
117 // and not have the inliner do any merging of allocas at all. This would
118 // allow the backend to do proper stack slot coloring of all allocas that
119 // *actually make it to the backend*, which is really what we want.
121 // Because we don't have this information, we do this simple and useful hack.
123 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
125 // When processing our SCC, check to see if CS was inlined from some other
126 // call site. For example, if we're processing "A" in this code:
128 // B() { x = alloca ... C() }
129 // C() { y = alloca ... }
130 // Assume that C was not inlined into B initially, and so we're processing A
131 // and decide to inline B into A. Doing this makes an alloca available for
132 // reuse and makes a callsite (C) available for inlining. When we process
133 // the C call site we don't want to do any alloca merging between X and Y
134 // because their scopes are not disjoint. We could make this smarter by
135 // keeping track of the inline history for each alloca in the
136 // InlinedArrayAllocas but this isn't likely to be a significant win.
137 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
140 // Loop over all the allocas we have so far and see if they can be merged with
141 // a previously inlined alloca. If not, remember that we had it.
142 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
143 AllocaNo != e; ++AllocaNo) {
144 AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
146 // Don't bother trying to merge array allocations (they will usually be
147 // canonicalized to be an allocation *of* an array), or allocations whose
148 // type is not itself an array (because we're afraid of pessimizing SRoA).
149 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
150 if (ATy == 0 || AI->isArrayAllocation())
153 // Get the list of all available allocas for this array type.
154 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
156 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
157 // that we have to be careful not to reuse the same "available" alloca for
158 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
159 // set to keep track of which "available" allocas are being used by this
160 // function. Also, AllocasForType can be empty of course!
161 bool MergedAwayAlloca = false;
162 for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) {
163 AllocaInst *AvailableAlloca = AllocasForType[i];
165 // The available alloca has to be in the right function, not in some other
166 // function in this SCC.
167 if (AvailableAlloca->getParent() != AI->getParent())
170 // If the inlined function already uses this alloca then we can't reuse
172 if (!UsedAllocas.insert(AvailableAlloca))
175 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
177 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: "
178 << *AvailableAlloca << '\n');
180 AI->replaceAllUsesWith(AvailableAlloca);
181 AI->eraseFromParent();
182 MergedAwayAlloca = true;
184 IFI.StaticAllocas[AllocaNo] = 0;
188 // If we already nuked the alloca, we're done with it.
189 if (MergedAwayAlloca)
192 // If we were unable to merge away the alloca either because there are no
193 // allocas of the right type available or because we reused them all
194 // already, remember that this alloca came from an inlined function and mark
195 // it used so we don't reuse it for other allocas from this inline
197 AllocasForType.push_back(AI);
198 UsedAllocas.insert(AI);
204 unsigned Inliner::getInlineThreshold(CallSite CS) const {
205 int thres = InlineThreshold; // -inline-threshold or else selected by
208 // If -inline-threshold is not given, listen to the optsize attribute when it
209 // would decrease the threshold.
210 Function *Caller = CS.getCaller();
211 bool OptSize = Caller && !Caller->isDeclaration() &&
212 Caller->getFnAttributes().hasOptimizeForSizeAttr();
213 if (!(InlineLimit.getNumOccurrences() > 0) && OptSize &&
214 OptSizeThreshold < thres)
215 thres = OptSizeThreshold;
217 // Listen to the inlinehint attribute when it would increase the threshold.
218 Function *Callee = CS.getCalledFunction();
219 bool InlineHint = Callee && !Callee->isDeclaration() &&
220 Callee->getFnAttributes().hasInlineHintAttr();
221 if (InlineHint && HintThreshold > thres)
222 thres = HintThreshold;
227 /// shouldInline - Return true if the inliner should attempt to inline
228 /// at the given CallSite.
229 bool Inliner::shouldInline(CallSite CS) {
230 InlineCost IC = getInlineCost(CS);
233 DEBUG(dbgs() << " Inlining: cost=always"
234 << ", Call: " << *CS.getInstruction() << "\n");
239 DEBUG(dbgs() << " NOT Inlining: cost=never"
240 << ", Call: " << *CS.getInstruction() << "\n");
244 Function *Caller = CS.getCaller();
246 DEBUG(dbgs() << " NOT Inlining: cost=" << IC.getCost()
247 << ", thres=" << (IC.getCostDelta() + IC.getCost())
248 << ", Call: " << *CS.getInstruction() << "\n");
252 // Try to detect the case where the current inlining candidate caller (call
253 // it B) is a static or linkonce-ODR function and is an inlining candidate
254 // elsewhere, and the current candidate callee (call it C) is large enough
255 // that inlining it into B would make B too big to inline later. In these
256 // circumstances it may be best not to inline C into B, but to inline B into
259 // This only applies to static and linkonce-ODR functions because those are
260 // expected to be available for inlining in the translation units where they
261 // are used. Thus we will always have the opportunity to make local inlining
262 // decisions. Importantly the linkonce-ODR linkage covers inline functions
263 // and templates in C++.
265 // FIXME: All of this logic should be sunk into getInlineCost. It relies on
266 // the internal implementation of the inline cost metrics rather than
267 // treating them as truly abstract units etc.
268 if (Caller->hasLocalLinkage() ||
269 Caller->getLinkage() == GlobalValue::LinkOnceODRLinkage) {
270 int TotalSecondaryCost = 0;
271 // The candidate cost to be imposed upon the current function.
272 int CandidateCost = IC.getCost() - (InlineConstants::CallPenalty + 1);
273 // This bool tracks what happens if we do NOT inline C into B.
274 bool callerWillBeRemoved = Caller->hasLocalLinkage();
275 // This bool tracks what happens if we DO inline C into B.
276 bool inliningPreventsSomeOuterInline = false;
277 for (Value::use_iterator I = Caller->use_begin(), E =Caller->use_end();
281 // If this isn't a call to Caller (it could be some other sort
282 // of reference) skip it. Such references will prevent the caller
283 // from being removed.
284 if (!CS2 || CS2.getCalledFunction() != Caller) {
285 callerWillBeRemoved = false;
289 InlineCost IC2 = getInlineCost(CS2);
290 ++NumCallerCallersAnalyzed;
292 callerWillBeRemoved = false;
298 // See if inlining or original callsite would erase the cost delta of
299 // this callsite. We subtract off the penalty for the call instruction,
300 // which we would be deleting.
301 if (IC2.getCostDelta() <= CandidateCost) {
302 inliningPreventsSomeOuterInline = true;
303 TotalSecondaryCost += IC2.getCost();
306 // If all outer calls to Caller would get inlined, the cost for the last
307 // one is set very low by getInlineCost, in anticipation that Caller will
308 // be removed entirely. We did not account for this above unless there
309 // is only one caller of Caller.
310 if (callerWillBeRemoved && Caller->use_begin() != Caller->use_end())
311 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus;
313 if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost()) {
314 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() <<
315 " Cost = " << IC.getCost() <<
316 ", outer Cost = " << TotalSecondaryCost << '\n');
321 DEBUG(dbgs() << " Inlining: cost=" << IC.getCost()
322 << ", thres=" << (IC.getCostDelta() + IC.getCost())
323 << ", Call: " << *CS.getInstruction() << '\n');
327 /// InlineHistoryIncludes - Return true if the specified inline history ID
328 /// indicates an inline history that includes the specified function.
329 static bool InlineHistoryIncludes(Function *F, int InlineHistoryID,
330 const SmallVectorImpl<std::pair<Function*, int> > &InlineHistory) {
331 while (InlineHistoryID != -1) {
332 assert(unsigned(InlineHistoryID) < InlineHistory.size() &&
333 "Invalid inline history ID");
334 if (InlineHistory[InlineHistoryID].first == F)
336 InlineHistoryID = InlineHistory[InlineHistoryID].second;
341 bool Inliner::runOnSCC(CallGraphSCC &SCC) {
342 CallGraph &CG = getAnalysis<CallGraph>();
343 const TargetData *TD = getAnalysisIfAvailable<TargetData>();
344 const TargetLibraryInfo *TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
346 SmallPtrSet<Function*, 8> SCCFunctions;
347 DEBUG(dbgs() << "Inliner visiting SCC:");
348 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
349 Function *F = (*I)->getFunction();
350 if (F) SCCFunctions.insert(F);
351 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
354 // Scan through and identify all call sites ahead of time so that we only
355 // inline call sites in the original functions, not call sites that result
356 // from inlining other functions.
357 SmallVector<std::pair<CallSite, int>, 16> CallSites;
359 // When inlining a callee produces new call sites, we want to keep track of
360 // the fact that they were inlined from the callee. This allows us to avoid
361 // infinite inlining in some obscure cases. To represent this, we use an
362 // index into the InlineHistory vector.
363 SmallVector<std::pair<Function*, int>, 8> InlineHistory;
365 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
366 Function *F = (*I)->getFunction();
369 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
370 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
371 CallSite CS(cast<Value>(I));
372 // If this isn't a call, or it is a call to an intrinsic, it can
374 if (!CS || isa<IntrinsicInst>(I))
377 // If this is a direct call to an external function, we can never inline
378 // it. If it is an indirect call, inlining may resolve it to be a
379 // direct call, so we keep it.
380 if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration())
383 CallSites.push_back(std::make_pair(CS, -1));
387 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
389 // If there are no calls in this function, exit early.
390 if (CallSites.empty())
393 // Now that we have all of the call sites, move the ones to functions in the
394 // current SCC to the end of the list.
395 unsigned FirstCallInSCC = CallSites.size();
396 for (unsigned i = 0; i < FirstCallInSCC; ++i)
397 if (Function *F = CallSites[i].first.getCalledFunction())
398 if (SCCFunctions.count(F))
399 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
402 InlinedArrayAllocasTy InlinedArrayAllocas;
403 InlineFunctionInfo InlineInfo(&CG, TD);
405 // Now that we have all of the call sites, loop over them and inline them if
406 // it looks profitable to do so.
407 bool Changed = false;
411 // Iterate over the outer loop because inlining functions can cause indirect
412 // calls to become direct calls.
413 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
414 CallSite CS = CallSites[CSi].first;
416 Function *Caller = CS.getCaller();
417 Function *Callee = CS.getCalledFunction();
419 // If this call site is dead and it is to a readonly function, we should
420 // just delete the call instead of trying to inline it, regardless of
421 // size. This happens because IPSCCP propagates the result out of the
422 // call and then we're left with the dead call.
423 if (isInstructionTriviallyDead(CS.getInstruction(), TLI)) {
424 DEBUG(dbgs() << " -> Deleting dead call: "
425 << *CS.getInstruction() << "\n");
426 // Update the call graph by deleting the edge from Callee to Caller.
427 CG[Caller]->removeCallEdgeFor(CS);
428 CS.getInstruction()->eraseFromParent();
431 // We can only inline direct calls to non-declarations.
432 if (Callee == 0 || Callee->isDeclaration()) continue;
434 // If this call site was obtained by inlining another function, verify
435 // that the include path for the function did not include the callee
436 // itself. If so, we'd be recursively inlining the same function,
437 // which would provide the same callsites, which would cause us to
438 // infinitely inline.
439 int InlineHistoryID = CallSites[CSi].second;
440 if (InlineHistoryID != -1 &&
441 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory))
445 // If the policy determines that we should inline this function,
447 if (!shouldInline(CS))
450 // Attempt to inline the function.
451 if (!InlineCallIfPossible(CS, InlineInfo, InlinedArrayAllocas,
452 InlineHistoryID, InsertLifetime))
456 // If inlining this function gave us any new call sites, throw them
457 // onto our worklist to process. They are useful inline candidates.
458 if (!InlineInfo.InlinedCalls.empty()) {
459 // Create a new inline history entry for this, so that we remember
460 // that these new callsites came about due to inlining Callee.
461 int NewHistoryID = InlineHistory.size();
462 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID));
464 for (unsigned i = 0, e = InlineInfo.InlinedCalls.size();
466 Value *Ptr = InlineInfo.InlinedCalls[i];
467 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID));
472 // If we inlined or deleted the last possible call site to the function,
473 // delete the function body now.
474 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
475 // TODO: Can remove if in SCC now.
476 !SCCFunctions.count(Callee) &&
478 // The function may be apparently dead, but if there are indirect
479 // callgraph references to the node, we cannot delete it yet, this
480 // could invalidate the CGSCC iterator.
481 CG[Callee]->getNumReferences() == 0) {
482 DEBUG(dbgs() << " -> Deleting dead function: "
483 << Callee->getName() << "\n");
484 CallGraphNode *CalleeNode = CG[Callee];
486 // Remove any call graph edges from the callee to its callees.
487 CalleeNode->removeAllCalledFunctions();
489 // Removing the node for callee from the call graph and delete it.
490 delete CG.removeFunctionFromModule(CalleeNode);
494 // Remove this call site from the list. If possible, use
495 // swap/pop_back for efficiency, but do not use it if doing so would
496 // move a call site to a function in this SCC before the
497 // 'FirstCallInSCC' barrier.
498 if (SCC.isSingular()) {
499 CallSites[CSi] = CallSites.back();
500 CallSites.pop_back();
502 CallSites.erase(CallSites.begin()+CSi);
509 } while (LocalChange);
514 // doFinalization - Remove now-dead linkonce functions at the end of
515 // processing to avoid breaking the SCC traversal.
516 bool Inliner::doFinalization(CallGraph &CG) {
517 return removeDeadFunctions(CG);
520 /// removeDeadFunctions - Remove dead functions that are not included in
521 /// DNR (Do Not Remove) list.
522 bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) {
523 SmallVector<CallGraphNode*, 16> FunctionsToRemove;
525 // Scan for all of the functions, looking for ones that should now be removed
526 // from the program. Insert the dead ones in the FunctionsToRemove set.
527 for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) {
528 CallGraphNode *CGN = I->second;
529 Function *F = CGN->getFunction();
530 if (!F || F->isDeclaration())
533 // Handle the case when this function is called and we only want to care
534 // about always-inline functions. This is a bit of a hack to share code
535 // between here and the InlineAlways pass.
536 if (AlwaysInlineOnly && !F->getFnAttributes().hasAlwaysInlineAttr())
539 // If the only remaining users of the function are dead constants, remove
541 F->removeDeadConstantUsers();
543 if (!F->isDefTriviallyDead())
546 // Remove any call graph edges from the function to its callees.
547 CGN->removeAllCalledFunctions();
549 // Remove any edges from the external node to the function's call graph
550 // node. These edges might have been made irrelegant due to
551 // optimization of the program.
552 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
554 // Removing the node for callee from the call graph and delete it.
555 FunctionsToRemove.push_back(CGN);
557 if (FunctionsToRemove.empty())
560 // Now that we know which functions to delete, do so. We didn't want to do
561 // this inline, because that would invalidate our CallGraph::iterator
564 // Note that it doesn't matter that we are iterating over a non-stable order
565 // here to do this, it doesn't matter which order the functions are deleted
567 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
568 FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(),
569 FunctionsToRemove.end()),
570 FunctionsToRemove.end());
571 for (SmallVectorImpl<CallGraphNode *>::iterator I = FunctionsToRemove.begin(),
572 E = FunctionsToRemove.end();
574 delete CG.removeFunctionFromModule(*I);