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");
40 InlineLimit("inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore,
41 cl::desc("Control the amount of inlining to perform (default = 225)"));
44 HintThreshold("inlinehint-threshold", cl::Hidden, cl::init(325),
45 cl::desc("Threshold for inlining functions with inline hint"));
47 // Threshold to use when optsize is specified (and there is no -inline-limit).
48 const int OptSizeThreshold = 75;
50 Inliner::Inliner(char &ID)
51 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit), InsertLifetime(true) {}
53 Inliner::Inliner(char &ID, int Threshold, bool InsertLifetime)
54 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit.getNumOccurrences() > 0 ?
55 InlineLimit : Threshold),
56 InsertLifetime(InsertLifetime) {}
58 /// getAnalysisUsage - For this class, we declare that we require and preserve
59 /// the call graph. If the derived class implements this method, it should
60 /// always explicitly call the implementation here.
61 void Inliner::getAnalysisUsage(AnalysisUsage &Info) const {
62 CallGraphSCCPass::getAnalysisUsage(Info);
66 typedef DenseMap<ArrayType*, std::vector<AllocaInst*> >
67 InlinedArrayAllocasTy;
69 /// InlineCallIfPossible - If it is possible to inline the specified call site,
70 /// do so and update the CallGraph for this operation.
72 /// This function also does some basic book-keeping to update the IR. The
73 /// InlinedArrayAllocas map keeps track of any allocas that are already
74 /// available from other functions inlined into the caller. If we are able to
75 /// inline this call site we attempt to reuse already available allocas or add
76 /// any new allocas to the set if not possible.
77 static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI,
78 InlinedArrayAllocasTy &InlinedArrayAllocas,
79 int InlineHistory, bool InsertLifetime) {
80 Function *Callee = CS.getCalledFunction();
81 Function *Caller = CS.getCaller();
83 // Try to inline the function. Get the list of static allocas that were
85 if (!InlineFunction(CS, IFI, InsertLifetime))
88 // If the inlined function had a higher stack protection level than the
89 // calling function, then bump up the caller's stack protection level.
90 if (Callee->hasFnAttr(Attribute::StackProtectReq))
91 Caller->addFnAttr(Attribute::StackProtectReq);
92 else if (Callee->hasFnAttr(Attribute::StackProtect) &&
93 !Caller->hasFnAttr(Attribute::StackProtectReq))
94 Caller->addFnAttr(Attribute::StackProtect);
96 // Look at all of the allocas that we inlined through this call site. If we
97 // have already inlined other allocas through other calls into this function,
98 // then we know that they have disjoint lifetimes and that we can merge them.
100 // There are many heuristics possible for merging these allocas, and the
101 // different options have different tradeoffs. One thing that we *really*
102 // don't want to hurt is SRoA: once inlining happens, often allocas are no
103 // longer address taken and so they can be promoted.
105 // Our "solution" for that is to only merge allocas whose outermost type is an
106 // array type. These are usually not promoted because someone is using a
107 // variable index into them. These are also often the most important ones to
110 // A better solution would be to have real memory lifetime markers in the IR
111 // and not have the inliner do any merging of allocas at all. This would
112 // allow the backend to do proper stack slot coloring of all allocas that
113 // *actually make it to the backend*, which is really what we want.
115 // Because we don't have this information, we do this simple and useful hack.
117 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
119 // When processing our SCC, check to see if CS was inlined from some other
120 // call site. For example, if we're processing "A" in this code:
122 // B() { x = alloca ... C() }
123 // C() { y = alloca ... }
124 // Assume that C was not inlined into B initially, and so we're processing A
125 // and decide to inline B into A. Doing this makes an alloca available for
126 // reuse and makes a callsite (C) available for inlining. When we process
127 // the C call site we don't want to do any alloca merging between X and Y
128 // because their scopes are not disjoint. We could make this smarter by
129 // keeping track of the inline history for each alloca in the
130 // InlinedArrayAllocas but this isn't likely to be a significant win.
131 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
134 // Loop over all the allocas we have so far and see if they can be merged with
135 // a previously inlined alloca. If not, remember that we had it.
136 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
137 AllocaNo != e; ++AllocaNo) {
138 AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
140 // Don't bother trying to merge array allocations (they will usually be
141 // canonicalized to be an allocation *of* an array), or allocations whose
142 // type is not itself an array (because we're afraid of pessimizing SRoA).
143 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
144 if (ATy == 0 || AI->isArrayAllocation())
147 // Get the list of all available allocas for this array type.
148 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
150 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
151 // that we have to be careful not to reuse the same "available" alloca for
152 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
153 // set to keep track of which "available" allocas are being used by this
154 // function. Also, AllocasForType can be empty of course!
155 bool MergedAwayAlloca = false;
156 for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) {
157 AllocaInst *AvailableAlloca = AllocasForType[i];
159 // The available alloca has to be in the right function, not in some other
160 // function in this SCC.
161 if (AvailableAlloca->getParent() != AI->getParent())
164 // If the inlined function already uses this alloca then we can't reuse
166 if (!UsedAllocas.insert(AvailableAlloca))
169 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
171 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: "
172 << *AvailableAlloca << '\n');
174 AI->replaceAllUsesWith(AvailableAlloca);
175 AI->eraseFromParent();
176 MergedAwayAlloca = true;
178 IFI.StaticAllocas[AllocaNo] = 0;
182 // If we already nuked the alloca, we're done with it.
183 if (MergedAwayAlloca)
186 // If we were unable to merge away the alloca either because there are no
187 // allocas of the right type available or because we reused them all
188 // already, remember that this alloca came from an inlined function and mark
189 // it used so we don't reuse it for other allocas from this inline
191 AllocasForType.push_back(AI);
192 UsedAllocas.insert(AI);
198 unsigned Inliner::getInlineThreshold(CallSite CS) const {
199 int thres = InlineThreshold;
201 // Listen to optsize when -inline-limit is not given.
202 Function *Caller = CS.getCaller();
203 if (Caller && !Caller->isDeclaration() &&
204 Caller->hasFnAttr(Attribute::OptimizeForSize) &&
205 InlineLimit.getNumOccurrences() == 0)
206 thres = OptSizeThreshold;
208 // Listen to inlinehint when it would increase the threshold.
209 Function *Callee = CS.getCalledFunction();
210 if (HintThreshold > thres && Callee && !Callee->isDeclaration() &&
211 Callee->hasFnAttr(Attribute::InlineHint))
212 thres = HintThreshold;
217 /// shouldInline - Return true if the inliner should attempt to inline
218 /// at the given CallSite.
219 bool Inliner::shouldInline(CallSite CS) {
220 InlineCost IC = getInlineCost(CS);
223 DEBUG(dbgs() << " Inlining: cost=always"
224 << ", Call: " << *CS.getInstruction() << "\n");
229 DEBUG(dbgs() << " NOT Inlining: cost=never"
230 << ", Call: " << *CS.getInstruction() << "\n");
234 int Cost = IC.getValue();
235 Function *Caller = CS.getCaller();
236 int CurrentThreshold = getInlineThreshold(CS);
237 float FudgeFactor = getInlineFudgeFactor(CS);
238 int AdjThreshold = (int)(CurrentThreshold * FudgeFactor);
239 if (Cost >= AdjThreshold) {
240 DEBUG(dbgs() << " NOT Inlining: cost=" << Cost
241 << ", thres=" << AdjThreshold
242 << ", Call: " << *CS.getInstruction() << "\n");
246 // Try to detect the case where the current inlining candidate caller (call
247 // it B) is a static or linkonce-ODR function and is an inlining candidate
248 // elsewhere, and the current candidate callee (call it C) is large enough
249 // that inlining it into B would make B too big to inline later. In these
250 // circumstances it may be best not to inline C into B, but to inline B into
253 // This only applies to static and linkonce-ODR functions because those are
254 // expected to be available for inlining in the translation units where they
255 // are used. Thus we will always have the opportunity to make local inlining
256 // decisions. Importantly the linkonce-ODR linkage covers inline functions
257 // and templates in C++.
258 if (Caller->hasLocalLinkage() ||
259 Caller->getLinkage() == GlobalValue::LinkOnceODRLinkage) {
260 int TotalSecondaryCost = 0;
261 bool outerCallsFound = false;
262 // This bool tracks what happens if we do NOT inline C into B.
263 bool callerWillBeRemoved = Caller->hasLocalLinkage();
264 // This bool tracks what happens if we DO inline C into B.
265 bool inliningPreventsSomeOuterInline = false;
266 for (Value::use_iterator I = Caller->use_begin(), E =Caller->use_end();
270 // If this isn't a call to Caller (it could be some other sort
271 // of reference) skip it. Such references will prevent the caller
272 // from being removed.
273 if (!CS2 || CS2.getCalledFunction() != Caller) {
274 callerWillBeRemoved = false;
278 InlineCost IC2 = getInlineCost(CS2);
280 callerWillBeRemoved = false;
281 if (IC2.isAlways() || IC2.isNever())
284 outerCallsFound = true;
285 int Cost2 = IC2.getValue();
286 int CurrentThreshold2 = getInlineThreshold(CS2);
287 float FudgeFactor2 = getInlineFudgeFactor(CS2);
289 if (Cost2 >= (int)(CurrentThreshold2 * FudgeFactor2))
290 callerWillBeRemoved = false;
292 // See if we have this case. We subtract off the penalty
293 // for the call instruction, which we would be deleting.
294 if (Cost2 < (int)(CurrentThreshold2 * FudgeFactor2) &&
295 Cost2 + Cost - (InlineConstants::CallPenalty + 1) >=
296 (int)(CurrentThreshold2 * FudgeFactor2)) {
297 inliningPreventsSomeOuterInline = true;
298 TotalSecondaryCost += Cost2;
301 // If all outer calls to Caller would get inlined, the cost for the last
302 // one is set very low by getInlineCost, in anticipation that Caller will
303 // be removed entirely. We did not account for this above unless there
304 // is only one caller of Caller.
305 if (callerWillBeRemoved && Caller->use_begin() != Caller->use_end())
306 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus;
308 if (outerCallsFound && inliningPreventsSomeOuterInline &&
309 TotalSecondaryCost < Cost) {
310 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() <<
311 " Cost = " << Cost <<
312 ", outer Cost = " << TotalSecondaryCost << '\n');
317 DEBUG(dbgs() << " Inlining: cost=" << Cost
318 << ", thres=" << AdjThreshold
319 << ", Call: " << *CS.getInstruction() << '\n');
323 /// InlineHistoryIncludes - Return true if the specified inline history ID
324 /// indicates an inline history that includes the specified function.
325 static bool InlineHistoryIncludes(Function *F, int InlineHistoryID,
326 const SmallVectorImpl<std::pair<Function*, int> > &InlineHistory) {
327 while (InlineHistoryID != -1) {
328 assert(unsigned(InlineHistoryID) < InlineHistory.size() &&
329 "Invalid inline history ID");
330 if (InlineHistory[InlineHistoryID].first == F)
332 InlineHistoryID = InlineHistory[InlineHistoryID].second;
337 bool Inliner::runOnSCC(CallGraphSCC &SCC) {
338 CallGraph &CG = getAnalysis<CallGraph>();
339 const TargetData *TD = getAnalysisIfAvailable<TargetData>();
341 SmallPtrSet<Function*, 8> SCCFunctions;
342 DEBUG(dbgs() << "Inliner visiting SCC:");
343 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
344 Function *F = (*I)->getFunction();
345 if (F) SCCFunctions.insert(F);
346 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
349 // Scan through and identify all call sites ahead of time so that we only
350 // inline call sites in the original functions, not call sites that result
351 // from inlining other functions.
352 SmallVector<std::pair<CallSite, int>, 16> CallSites;
354 // When inlining a callee produces new call sites, we want to keep track of
355 // the fact that they were inlined from the callee. This allows us to avoid
356 // infinite inlining in some obscure cases. To represent this, we use an
357 // index into the InlineHistory vector.
358 SmallVector<std::pair<Function*, int>, 8> InlineHistory;
360 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
361 Function *F = (*I)->getFunction();
364 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
365 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
366 CallSite CS(cast<Value>(I));
367 // If this isn't a call, or it is a call to an intrinsic, it can
369 if (!CS || isa<IntrinsicInst>(I))
372 // If this is a direct call to an external function, we can never inline
373 // it. If it is an indirect call, inlining may resolve it to be a
374 // direct call, so we keep it.
375 if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration())
378 CallSites.push_back(std::make_pair(CS, -1));
382 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
384 // If there are no calls in this function, exit early.
385 if (CallSites.empty())
388 // Now that we have all of the call sites, move the ones to functions in the
389 // current SCC to the end of the list.
390 unsigned FirstCallInSCC = CallSites.size();
391 for (unsigned i = 0; i < FirstCallInSCC; ++i)
392 if (Function *F = CallSites[i].first.getCalledFunction())
393 if (SCCFunctions.count(F))
394 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
397 InlinedArrayAllocasTy InlinedArrayAllocas;
398 InlineFunctionInfo InlineInfo(&CG, TD);
400 // Now that we have all of the call sites, loop over them and inline them if
401 // it looks profitable to do so.
402 bool Changed = false;
406 // Iterate over the outer loop because inlining functions can cause indirect
407 // calls to become direct calls.
408 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
409 CallSite CS = CallSites[CSi].first;
411 Function *Caller = CS.getCaller();
412 Function *Callee = CS.getCalledFunction();
414 // If this call site is dead and it is to a readonly function, we should
415 // just delete the call instead of trying to inline it, regardless of
416 // size. This happens because IPSCCP propagates the result out of the
417 // call and then we're left with the dead call.
418 if (isInstructionTriviallyDead(CS.getInstruction())) {
419 DEBUG(dbgs() << " -> Deleting dead call: "
420 << *CS.getInstruction() << "\n");
421 // Update the call graph by deleting the edge from Callee to Caller.
422 CG[Caller]->removeCallEdgeFor(CS);
423 CS.getInstruction()->eraseFromParent();
425 // Update the cached cost info with the missing call
426 growCachedCostInfo(Caller, NULL);
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));
468 // Update the cached cost info with the inlined call.
469 growCachedCostInfo(Caller, Callee);
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 resetCachedCostInfo(Callee);
491 // Removing the node for callee from the call graph and delete it.
492 delete CG.removeFunctionFromModule(CalleeNode);
496 // Remove this call site from the list. If possible, use
497 // swap/pop_back for efficiency, but do not use it if doing so would
498 // move a call site to a function in this SCC before the
499 // 'FirstCallInSCC' barrier.
500 if (SCC.isSingular()) {
501 CallSites[CSi] = CallSites.back();
502 CallSites.pop_back();
504 CallSites.erase(CallSites.begin()+CSi);
511 } while (LocalChange);
516 // doFinalization - Remove now-dead linkonce functions at the end of
517 // processing to avoid breaking the SCC traversal.
518 bool Inliner::doFinalization(CallGraph &CG) {
519 return removeDeadFunctions(CG);
522 /// removeDeadFunctions - Remove dead functions that are not included in
523 /// DNR (Do Not Remove) list.
524 bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) {
525 SmallVector<CallGraphNode*, 16> FunctionsToRemove;
527 // Scan for all of the functions, looking for ones that should now be removed
528 // from the program. Insert the dead ones in the FunctionsToRemove set.
529 for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) {
530 CallGraphNode *CGN = I->second;
531 Function *F = CGN->getFunction();
532 if (!F || F->isDeclaration())
535 // Handle the case when this function is called and we only want to care
536 // about always-inline functions. This is a bit of a hack to share code
537 // between here and the InlineAlways pass.
538 if (AlwaysInlineOnly && !F->hasFnAttr(Attribute::AlwaysInline))
541 // If the only remaining users of the function are dead constants, remove
543 F->removeDeadConstantUsers();
545 if (!F->isDefTriviallyDead())
548 // Remove any call graph edges from the function to its callees.
549 CGN->removeAllCalledFunctions();
551 // Remove any edges from the external node to the function's call graph
552 // node. These edges might have been made irrelegant due to
553 // optimization of the program.
554 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
556 // Removing the node for callee from the call graph and delete it.
557 FunctionsToRemove.push_back(CGN);
559 if (FunctionsToRemove.empty())
562 // Now that we know which functions to delete, do so. We didn't want to do
563 // this inline, because that would invalidate our CallGraph::iterator
566 // Note that it doesn't matter that we are iterating over a non-stable order
567 // here to do this, it doesn't matter which order the functions are deleted
569 std::sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
570 FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(),
571 FunctionsToRemove.end()),
572 FunctionsToRemove.end());
573 for (SmallVectorImpl<CallGraphNode *>::iterator I = FunctionsToRemove.begin(),
574 E = FunctionsToRemove.end();
576 resetCachedCostInfo((*I)->getFunction());
577 delete CG.removeFunctionFromModule(*I);