LLVM API Documentation

Inliner.cpp
Go to the documentation of this file.
00001 //===- Inliner.cpp - Code common to all inliners --------------------------===//
00002 //
00003 //                     The LLVM Compiler Infrastructure
00004 //
00005 // This file is distributed under the University of Illinois Open Source
00006 // License. See LICENSE.TXT for details.
00007 //
00008 //===----------------------------------------------------------------------===//
00009 //
00010 // This file implements the mechanics required to implement inlining without
00011 // missing any calls and updating the call graph.  The decisions of which calls
00012 // are profitable to inline are implemented elsewhere.
00013 //
00014 //===----------------------------------------------------------------------===//
00015 
00016 #include "llvm/Transforms/IPO/InlinerPass.h"
00017 #include "llvm/ADT/SmallPtrSet.h"
00018 #include "llvm/ADT/Statistic.h"
00019 #include "llvm/Analysis/AliasAnalysis.h"
00020 #include "llvm/Analysis/AssumptionCache.h"
00021 #include "llvm/Analysis/CallGraph.h"
00022 #include "llvm/Analysis/InlineCost.h"
00023 #include "llvm/IR/CallSite.h"
00024 #include "llvm/IR/DataLayout.h"
00025 #include "llvm/IR/DiagnosticInfo.h"
00026 #include "llvm/IR/Instructions.h"
00027 #include "llvm/IR/IntrinsicInst.h"
00028 #include "llvm/IR/Module.h"
00029 #include "llvm/Support/CommandLine.h"
00030 #include "llvm/Support/Debug.h"
00031 #include "llvm/Support/raw_ostream.h"
00032 #include "llvm/Analysis/TargetLibraryInfo.h"
00033 #include "llvm/Transforms/Utils/Cloning.h"
00034 #include "llvm/Transforms/Utils/Local.h"
00035 using namespace llvm;
00036 
00037 #define DEBUG_TYPE "inline"
00038 
00039 STATISTIC(NumInlined, "Number of functions inlined");
00040 STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined");
00041 STATISTIC(NumDeleted, "Number of functions deleted because all callers found");
00042 STATISTIC(NumMergedAllocas, "Number of allocas merged together");
00043 
00044 // This weirdly named statistic tracks the number of times that, when attempting
00045 // to inline a function A into B, we analyze the callers of B in order to see
00046 // if those would be more profitable and blocked inline steps.
00047 STATISTIC(NumCallerCallersAnalyzed, "Number of caller-callers analyzed");
00048 
00049 static cl::opt<int>
00050 InlineLimit("inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore,
00051         cl::desc("Control the amount of inlining to perform (default = 225)"));
00052 
00053 static cl::opt<int>
00054 HintThreshold("inlinehint-threshold", cl::Hidden, cl::init(325),
00055               cl::desc("Threshold for inlining functions with inline hint"));
00056 
00057 // We instroduce this threshold to help performance of instrumentation based
00058 // PGO before we actually hook up inliner with analysis passes such as BPI and
00059 // BFI.
00060 static cl::opt<int>
00061 ColdThreshold("inlinecold-threshold", cl::Hidden, cl::init(225),
00062               cl::desc("Threshold for inlining functions with cold attribute"));
00063 
00064 // Threshold to use when optsize is specified (and there is no -inline-limit).
00065 const int OptSizeThreshold = 75;
00066 
00067 Inliner::Inliner(char &ID) 
00068   : CallGraphSCCPass(ID), InlineThreshold(InlineLimit), InsertLifetime(true) {}
00069 
00070 Inliner::Inliner(char &ID, int Threshold, bool InsertLifetime)
00071   : CallGraphSCCPass(ID), InlineThreshold(InlineLimit.getNumOccurrences() > 0 ?
00072                                           InlineLimit : Threshold),
00073     InsertLifetime(InsertLifetime) {}
00074 
00075 /// getAnalysisUsage - For this class, we declare that we require and preserve
00076 /// the call graph.  If the derived class implements this method, it should
00077 /// always explicitly call the implementation here.
00078 void Inliner::getAnalysisUsage(AnalysisUsage &AU) const {
00079   AU.addRequired<AliasAnalysis>();
00080   AU.addRequired<AssumptionCacheTracker>();
00081   CallGraphSCCPass::getAnalysisUsage(AU);
00082 }
00083 
00084 
00085 typedef DenseMap<ArrayType*, std::vector<AllocaInst*> >
00086 InlinedArrayAllocasTy;
00087 
00088 /// \brief If the inlined function had a higher stack protection level than the
00089 /// calling function, then bump up the caller's stack protection level.
00090 static void AdjustCallerSSPLevel(Function *Caller, Function *Callee) {
00091   // If upgrading the SSP attribute, clear out the old SSP Attributes first.
00092   // Having multiple SSP attributes doesn't actually hurt, but it adds useless
00093   // clutter to the IR.
00094   AttrBuilder B;
00095   B.addAttribute(Attribute::StackProtect)
00096     .addAttribute(Attribute::StackProtectStrong);
00097   AttributeSet OldSSPAttr = AttributeSet::get(Caller->getContext(),
00098                                               AttributeSet::FunctionIndex,
00099                                               B);
00100 
00101   if (Callee->hasFnAttribute(Attribute::StackProtectReq)) {
00102     Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
00103     Caller->addFnAttr(Attribute::StackProtectReq);
00104   } else if (Callee->hasFnAttribute(Attribute::StackProtectStrong) &&
00105              !Caller->hasFnAttribute(Attribute::StackProtectReq)) {
00106     Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
00107     Caller->addFnAttr(Attribute::StackProtectStrong);
00108   } else if (Callee->hasFnAttribute(Attribute::StackProtect) &&
00109              !Caller->hasFnAttribute(Attribute::StackProtectReq) &&
00110              !Caller->hasFnAttribute(Attribute::StackProtectStrong))
00111     Caller->addFnAttr(Attribute::StackProtect);
00112 }
00113 
00114 /// InlineCallIfPossible - If it is possible to inline the specified call site,
00115 /// do so and update the CallGraph for this operation.
00116 ///
00117 /// This function also does some basic book-keeping to update the IR.  The
00118 /// InlinedArrayAllocas map keeps track of any allocas that are already
00119 /// available from other  functions inlined into the caller.  If we are able to
00120 /// inline this call site we attempt to reuse already available allocas or add
00121 /// any new allocas to the set if not possible.
00122 static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI,
00123                                  InlinedArrayAllocasTy &InlinedArrayAllocas,
00124                                  int InlineHistory, bool InsertLifetime,
00125                                  const DataLayout *DL) {
00126   Function *Callee = CS.getCalledFunction();
00127   Function *Caller = CS.getCaller();
00128 
00129   // Try to inline the function.  Get the list of static allocas that were
00130   // inlined.
00131   if (!InlineFunction(CS, IFI, InsertLifetime))
00132     return false;
00133 
00134   AdjustCallerSSPLevel(Caller, Callee);
00135 
00136   // Look at all of the allocas that we inlined through this call site.  If we
00137   // have already inlined other allocas through other calls into this function,
00138   // then we know that they have disjoint lifetimes and that we can merge them.
00139   //
00140   // There are many heuristics possible for merging these allocas, and the
00141   // different options have different tradeoffs.  One thing that we *really*
00142   // don't want to hurt is SRoA: once inlining happens, often allocas are no
00143   // longer address taken and so they can be promoted.
00144   //
00145   // Our "solution" for that is to only merge allocas whose outermost type is an
00146   // array type.  These are usually not promoted because someone is using a
00147   // variable index into them.  These are also often the most important ones to
00148   // merge.
00149   //
00150   // A better solution would be to have real memory lifetime markers in the IR
00151   // and not have the inliner do any merging of allocas at all.  This would
00152   // allow the backend to do proper stack slot coloring of all allocas that
00153   // *actually make it to the backend*, which is really what we want.
00154   //
00155   // Because we don't have this information, we do this simple and useful hack.
00156   //
00157   SmallPtrSet<AllocaInst*, 16> UsedAllocas;
00158   
00159   // When processing our SCC, check to see if CS was inlined from some other
00160   // call site.  For example, if we're processing "A" in this code:
00161   //   A() { B() }
00162   //   B() { x = alloca ... C() }
00163   //   C() { y = alloca ... }
00164   // Assume that C was not inlined into B initially, and so we're processing A
00165   // and decide to inline B into A.  Doing this makes an alloca available for
00166   // reuse and makes a callsite (C) available for inlining.  When we process
00167   // the C call site we don't want to do any alloca merging between X and Y
00168   // because their scopes are not disjoint.  We could make this smarter by
00169   // keeping track of the inline history for each alloca in the
00170   // InlinedArrayAllocas but this isn't likely to be a significant win.
00171   if (InlineHistory != -1)  // Only do merging for top-level call sites in SCC.
00172     return true;
00173   
00174   // Loop over all the allocas we have so far and see if they can be merged with
00175   // a previously inlined alloca.  If not, remember that we had it.
00176   for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
00177        AllocaNo != e; ++AllocaNo) {
00178     AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
00179     
00180     // Don't bother trying to merge array allocations (they will usually be
00181     // canonicalized to be an allocation *of* an array), or allocations whose
00182     // type is not itself an array (because we're afraid of pessimizing SRoA).
00183     ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
00184     if (!ATy || AI->isArrayAllocation())
00185       continue;
00186     
00187     // Get the list of all available allocas for this array type.
00188     std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
00189     
00190     // Loop over the allocas in AllocasForType to see if we can reuse one.  Note
00191     // that we have to be careful not to reuse the same "available" alloca for
00192     // multiple different allocas that we just inlined, we use the 'UsedAllocas'
00193     // set to keep track of which "available" allocas are being used by this
00194     // function.  Also, AllocasForType can be empty of course!
00195     bool MergedAwayAlloca = false;
00196     for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) {
00197       AllocaInst *AvailableAlloca = AllocasForType[i];
00198 
00199       unsigned Align1 = AI->getAlignment(),
00200                Align2 = AvailableAlloca->getAlignment();
00201       // If we don't have data layout information, and only one alloca is using
00202       // the target default, then we can't safely merge them because we can't
00203       // pick the greater alignment.
00204       if (!DL && (!Align1 || !Align2) && Align1 != Align2)
00205         continue;
00206       
00207       // The available alloca has to be in the right function, not in some other
00208       // function in this SCC.
00209       if (AvailableAlloca->getParent() != AI->getParent())
00210         continue;
00211       
00212       // If the inlined function already uses this alloca then we can't reuse
00213       // it.
00214       if (!UsedAllocas.insert(AvailableAlloca).second)
00215         continue;
00216       
00217       // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
00218       // success!
00219       DEBUG(dbgs() << "    ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: "
00220                    << *AvailableAlloca << '\n');
00221       
00222       AI->replaceAllUsesWith(AvailableAlloca);
00223 
00224       if (Align1 != Align2) {
00225         if (!Align1 || !Align2) {
00226           assert(DL && "DataLayout required to compare default alignments");
00227           unsigned TypeAlign = DL->getABITypeAlignment(AI->getAllocatedType());
00228 
00229           Align1 = Align1 ? Align1 : TypeAlign;
00230           Align2 = Align2 ? Align2 : TypeAlign;
00231         }
00232 
00233         if (Align1 > Align2)
00234           AvailableAlloca->setAlignment(AI->getAlignment());
00235       }
00236 
00237       AI->eraseFromParent();
00238       MergedAwayAlloca = true;
00239       ++NumMergedAllocas;
00240       IFI.StaticAllocas[AllocaNo] = nullptr;
00241       break;
00242     }
00243 
00244     // If we already nuked the alloca, we're done with it.
00245     if (MergedAwayAlloca)
00246       continue;
00247     
00248     // If we were unable to merge away the alloca either because there are no
00249     // allocas of the right type available or because we reused them all
00250     // already, remember that this alloca came from an inlined function and mark
00251     // it used so we don't reuse it for other allocas from this inline
00252     // operation.
00253     AllocasForType.push_back(AI);
00254     UsedAllocas.insert(AI);
00255   }
00256   
00257   return true;
00258 }
00259 
00260 unsigned Inliner::getInlineThreshold(CallSite CS) const {
00261   int thres = InlineThreshold; // -inline-threshold or else selected by
00262                                // overall opt level
00263 
00264   // If -inline-threshold is not given, listen to the optsize attribute when it
00265   // would decrease the threshold.
00266   Function *Caller = CS.getCaller();
00267   bool OptSize = Caller && !Caller->isDeclaration() &&
00268                  Caller->hasFnAttribute(Attribute::OptimizeForSize);
00269   if (!(InlineLimit.getNumOccurrences() > 0) && OptSize &&
00270       OptSizeThreshold < thres)
00271     thres = OptSizeThreshold;
00272 
00273   // Listen to the inlinehint attribute when it would increase the threshold
00274   // and the caller does not need to minimize its size.
00275   Function *Callee = CS.getCalledFunction();
00276   bool InlineHint = Callee && !Callee->isDeclaration() &&
00277                     Callee->hasFnAttribute(Attribute::InlineHint);
00278   if (InlineHint && HintThreshold > thres &&
00279       !Caller->hasFnAttribute(Attribute::MinSize))
00280     thres = HintThreshold;
00281 
00282   // Listen to the cold attribute when it would decrease the threshold.
00283   bool ColdCallee = Callee && !Callee->isDeclaration() &&
00284                     Callee->hasFnAttribute(Attribute::Cold);
00285   // Command line argument for InlineLimit will override the default
00286   // ColdThreshold. If we have -inline-threshold but no -inlinecold-threshold,
00287   // do not use the default cold threshold even if it is smaller.
00288   if ((InlineLimit.getNumOccurrences() == 0 ||
00289        ColdThreshold.getNumOccurrences() > 0) && ColdCallee &&
00290       ColdThreshold < thres)
00291     thres = ColdThreshold;
00292 
00293   return thres;
00294 }
00295 
00296 static void emitAnalysis(CallSite CS, const Twine &Msg) {
00297   Function *Caller = CS.getCaller();
00298   LLVMContext &Ctx = Caller->getContext();
00299   DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
00300   emitOptimizationRemarkAnalysis(Ctx, DEBUG_TYPE, *Caller, DLoc, Msg);
00301 }
00302 
00303 /// shouldInline - Return true if the inliner should attempt to inline
00304 /// at the given CallSite.
00305 bool Inliner::shouldInline(CallSite CS) {
00306   InlineCost IC = getInlineCost(CS);
00307   
00308   if (IC.isAlways()) {
00309     DEBUG(dbgs() << "    Inlining: cost=always"
00310           << ", Call: " << *CS.getInstruction() << "\n");
00311     emitAnalysis(CS, Twine(CS.getCalledFunction()->getName()) +
00312                          " should always be inlined (cost=always)");
00313     return true;
00314   }
00315   
00316   if (IC.isNever()) {
00317     DEBUG(dbgs() << "    NOT Inlining: cost=never"
00318           << ", Call: " << *CS.getInstruction() << "\n");
00319     emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() +
00320                            " should never be inlined (cost=never)"));
00321     return false;
00322   }
00323   
00324   Function *Caller = CS.getCaller();
00325   if (!IC) {
00326     DEBUG(dbgs() << "    NOT Inlining: cost=" << IC.getCost()
00327           << ", thres=" << (IC.getCostDelta() + IC.getCost())
00328           << ", Call: " << *CS.getInstruction() << "\n");
00329     emitAnalysis(CS, Twine(CS.getCalledFunction()->getName() +
00330                            " too costly to inline (cost=") +
00331                          Twine(IC.getCost()) + ", threshold=" +
00332                          Twine(IC.getCostDelta() + IC.getCost()) + ")");
00333     return false;
00334   }
00335   
00336   // Try to detect the case where the current inlining candidate caller (call
00337   // it B) is a static or linkonce-ODR function and is an inlining candidate
00338   // elsewhere, and the current candidate callee (call it C) is large enough
00339   // that inlining it into B would make B too big to inline later. In these
00340   // circumstances it may be best not to inline C into B, but to inline B into
00341   // its callers.
00342   //
00343   // This only applies to static and linkonce-ODR functions because those are
00344   // expected to be available for inlining in the translation units where they
00345   // are used. Thus we will always have the opportunity to make local inlining
00346   // decisions. Importantly the linkonce-ODR linkage covers inline functions
00347   // and templates in C++.
00348   //
00349   // FIXME: All of this logic should be sunk into getInlineCost. It relies on
00350   // the internal implementation of the inline cost metrics rather than
00351   // treating them as truly abstract units etc.
00352   if (Caller->hasLocalLinkage() || Caller->hasLinkOnceODRLinkage()) {
00353     int TotalSecondaryCost = 0;
00354     // The candidate cost to be imposed upon the current function.
00355     int CandidateCost = IC.getCost() - (InlineConstants::CallPenalty + 1);
00356     // This bool tracks what happens if we do NOT inline C into B.
00357     bool callerWillBeRemoved = Caller->hasLocalLinkage();
00358     // This bool tracks what happens if we DO inline C into B.
00359     bool inliningPreventsSomeOuterInline = false;
00360     for (User *U : Caller->users()) {
00361       CallSite CS2(U);
00362 
00363       // If this isn't a call to Caller (it could be some other sort
00364       // of reference) skip it.  Such references will prevent the caller
00365       // from being removed.
00366       if (!CS2 || CS2.getCalledFunction() != Caller) {
00367         callerWillBeRemoved = false;
00368         continue;
00369       }
00370 
00371       InlineCost IC2 = getInlineCost(CS2);
00372       ++NumCallerCallersAnalyzed;
00373       if (!IC2) {
00374         callerWillBeRemoved = false;
00375         continue;
00376       }
00377       if (IC2.isAlways())
00378         continue;
00379 
00380       // See if inlining or original callsite would erase the cost delta of
00381       // this callsite. We subtract off the penalty for the call instruction,
00382       // which we would be deleting.
00383       if (IC2.getCostDelta() <= CandidateCost) {
00384         inliningPreventsSomeOuterInline = true;
00385         TotalSecondaryCost += IC2.getCost();
00386       }
00387     }
00388     // If all outer calls to Caller would get inlined, the cost for the last
00389     // one is set very low by getInlineCost, in anticipation that Caller will
00390     // be removed entirely.  We did not account for this above unless there
00391     // is only one caller of Caller.
00392     if (callerWillBeRemoved && !Caller->use_empty())
00393       TotalSecondaryCost += InlineConstants::LastCallToStaticBonus;
00394 
00395     if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost()) {
00396       DEBUG(dbgs() << "    NOT Inlining: " << *CS.getInstruction() <<
00397            " Cost = " << IC.getCost() <<
00398            ", outer Cost = " << TotalSecondaryCost << '\n');
00399       emitAnalysis(
00400           CS, Twine("Not inlining. Cost of inlining " +
00401                     CS.getCalledFunction()->getName() +
00402                     " increases the cost of inlining " +
00403                     CS.getCaller()->getName() + " in other contexts"));
00404       return false;
00405     }
00406   }
00407 
00408   DEBUG(dbgs() << "    Inlining: cost=" << IC.getCost()
00409         << ", thres=" << (IC.getCostDelta() + IC.getCost())
00410         << ", Call: " << *CS.getInstruction() << '\n');
00411   emitAnalysis(
00412       CS, CS.getCalledFunction()->getName() + Twine(" can be inlined into ") +
00413               CS.getCaller()->getName() + " with cost=" + Twine(IC.getCost()) +
00414               " (threshold=" + Twine(IC.getCostDelta() + IC.getCost()) + ")");
00415   return true;
00416 }
00417 
00418 /// InlineHistoryIncludes - Return true if the specified inline history ID
00419 /// indicates an inline history that includes the specified function.
00420 static bool InlineHistoryIncludes(Function *F, int InlineHistoryID,
00421             const SmallVectorImpl<std::pair<Function*, int> > &InlineHistory) {
00422   while (InlineHistoryID != -1) {
00423     assert(unsigned(InlineHistoryID) < InlineHistory.size() &&
00424            "Invalid inline history ID");
00425     if (InlineHistory[InlineHistoryID].first == F)
00426       return true;
00427     InlineHistoryID = InlineHistory[InlineHistoryID].second;
00428   }
00429   return false;
00430 }
00431 
00432 bool Inliner::runOnSCC(CallGraphSCC &SCC) {
00433   CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
00434   AssumptionCacheTracker *ACT = &getAnalysis<AssumptionCacheTracker>();
00435   DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
00436   const DataLayout *DL = DLP ? &DLP->getDataLayout() : nullptr;
00437   auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
00438   const TargetLibraryInfo *TLI = TLIP ? &TLIP->getTLI() : nullptr;
00439   AliasAnalysis *AA = &getAnalysis<AliasAnalysis>();
00440 
00441   SmallPtrSet<Function*, 8> SCCFunctions;
00442   DEBUG(dbgs() << "Inliner visiting SCC:");
00443   for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
00444     Function *F = (*I)->getFunction();
00445     if (F) SCCFunctions.insert(F);
00446     DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
00447   }
00448 
00449   // Scan through and identify all call sites ahead of time so that we only
00450   // inline call sites in the original functions, not call sites that result
00451   // from inlining other functions.
00452   SmallVector<std::pair<CallSite, int>, 16> CallSites;
00453   
00454   // When inlining a callee produces new call sites, we want to keep track of
00455   // the fact that they were inlined from the callee.  This allows us to avoid
00456   // infinite inlining in some obscure cases.  To represent this, we use an
00457   // index into the InlineHistory vector.
00458   SmallVector<std::pair<Function*, int>, 8> InlineHistory;
00459 
00460   for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
00461     Function *F = (*I)->getFunction();
00462     if (!F) continue;
00463     
00464     for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
00465       for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
00466         CallSite CS(cast<Value>(I));
00467         // If this isn't a call, or it is a call to an intrinsic, it can
00468         // never be inlined.
00469         if (!CS || isa<IntrinsicInst>(I))
00470           continue;
00471         
00472         // If this is a direct call to an external function, we can never inline
00473         // it.  If it is an indirect call, inlining may resolve it to be a
00474         // direct call, so we keep it.
00475         if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration())
00476           continue;
00477         
00478         CallSites.push_back(std::make_pair(CS, -1));
00479       }
00480   }
00481 
00482   DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
00483 
00484   // If there are no calls in this function, exit early.
00485   if (CallSites.empty())
00486     return false;
00487   
00488   // Now that we have all of the call sites, move the ones to functions in the
00489   // current SCC to the end of the list.
00490   unsigned FirstCallInSCC = CallSites.size();
00491   for (unsigned i = 0; i < FirstCallInSCC; ++i)
00492     if (Function *F = CallSites[i].first.getCalledFunction())
00493       if (SCCFunctions.count(F))
00494         std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
00495 
00496   
00497   InlinedArrayAllocasTy InlinedArrayAllocas;
00498   InlineFunctionInfo InlineInfo(&CG, DL, AA, ACT);
00499 
00500   // Now that we have all of the call sites, loop over them and inline them if
00501   // it looks profitable to do so.
00502   bool Changed = false;
00503   bool LocalChange;
00504   do {
00505     LocalChange = false;
00506     // Iterate over the outer loop because inlining functions can cause indirect
00507     // calls to become direct calls.
00508     for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
00509       CallSite CS = CallSites[CSi].first;
00510       
00511       Function *Caller = CS.getCaller();
00512       Function *Callee = CS.getCalledFunction();
00513 
00514       // If this call site is dead and it is to a readonly function, we should
00515       // just delete the call instead of trying to inline it, regardless of
00516       // size.  This happens because IPSCCP propagates the result out of the
00517       // call and then we're left with the dead call.
00518       if (isInstructionTriviallyDead(CS.getInstruction(), TLI)) {
00519         DEBUG(dbgs() << "    -> Deleting dead call: "
00520                      << *CS.getInstruction() << "\n");
00521         // Update the call graph by deleting the edge from Callee to Caller.
00522         CG[Caller]->removeCallEdgeFor(CS);
00523         CS.getInstruction()->eraseFromParent();
00524         ++NumCallsDeleted;
00525       } else {
00526         // We can only inline direct calls to non-declarations.
00527         if (!Callee || Callee->isDeclaration()) continue;
00528       
00529         // If this call site was obtained by inlining another function, verify
00530         // that the include path for the function did not include the callee
00531         // itself.  If so, we'd be recursively inlining the same function,
00532         // which would provide the same callsites, which would cause us to
00533         // infinitely inline.
00534         int InlineHistoryID = CallSites[CSi].second;
00535         if (InlineHistoryID != -1 &&
00536             InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory))
00537           continue;
00538         
00539         LLVMContext &CallerCtx = Caller->getContext();
00540 
00541         // Get DebugLoc to report. CS will be invalid after Inliner.
00542         DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
00543 
00544         // If the policy determines that we should inline this function,
00545         // try to do so.
00546         if (!shouldInline(CS)) {
00547           emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
00548                                        Twine(Callee->getName() +
00549                                              " will not be inlined into " +
00550                                              Caller->getName()));
00551           continue;
00552         }
00553 
00554         // Attempt to inline the function.
00555         if (!InlineCallIfPossible(CS, InlineInfo, InlinedArrayAllocas,
00556                                   InlineHistoryID, InsertLifetime, DL)) {
00557           emitOptimizationRemarkMissed(CallerCtx, DEBUG_TYPE, *Caller, DLoc,
00558                                        Twine(Callee->getName() +
00559                                              " will not be inlined into " +
00560                                              Caller->getName()));
00561           continue;
00562         }
00563         ++NumInlined;
00564 
00565         // Report the inline decision.
00566         emitOptimizationRemark(
00567             CallerCtx, DEBUG_TYPE, *Caller, DLoc,
00568             Twine(Callee->getName() + " inlined into " + Caller->getName()));
00569 
00570         // If inlining this function gave us any new call sites, throw them
00571         // onto our worklist to process.  They are useful inline candidates.
00572         if (!InlineInfo.InlinedCalls.empty()) {
00573           // Create a new inline history entry for this, so that we remember
00574           // that these new callsites came about due to inlining Callee.
00575           int NewHistoryID = InlineHistory.size();
00576           InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID));
00577 
00578           for (unsigned i = 0, e = InlineInfo.InlinedCalls.size();
00579                i != e; ++i) {
00580             Value *Ptr = InlineInfo.InlinedCalls[i];
00581             CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID));
00582           }
00583         }
00584       }
00585       
00586       // If we inlined or deleted the last possible call site to the function,
00587       // delete the function body now.
00588       if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
00589           // TODO: Can remove if in SCC now.
00590           !SCCFunctions.count(Callee) &&
00591           
00592           // The function may be apparently dead, but if there are indirect
00593           // callgraph references to the node, we cannot delete it yet, this
00594           // could invalidate the CGSCC iterator.
00595           CG[Callee]->getNumReferences() == 0) {
00596         DEBUG(dbgs() << "    -> Deleting dead function: "
00597               << Callee->getName() << "\n");
00598         CallGraphNode *CalleeNode = CG[Callee];
00599         
00600         // Remove any call graph edges from the callee to its callees.
00601         CalleeNode->removeAllCalledFunctions();
00602         
00603         // Removing the node for callee from the call graph and delete it.
00604         delete CG.removeFunctionFromModule(CalleeNode);
00605         ++NumDeleted;
00606       }
00607 
00608       // Remove this call site from the list.  If possible, use 
00609       // swap/pop_back for efficiency, but do not use it if doing so would
00610       // move a call site to a function in this SCC before the
00611       // 'FirstCallInSCC' barrier.
00612       if (SCC.isSingular()) {
00613         CallSites[CSi] = CallSites.back();
00614         CallSites.pop_back();
00615       } else {
00616         CallSites.erase(CallSites.begin()+CSi);
00617       }
00618       --CSi;
00619 
00620       Changed = true;
00621       LocalChange = true;
00622     }
00623   } while (LocalChange);
00624 
00625   return Changed;
00626 }
00627 
00628 // doFinalization - Remove now-dead linkonce functions at the end of
00629 // processing to avoid breaking the SCC traversal.
00630 bool Inliner::doFinalization(CallGraph &CG) {
00631   return removeDeadFunctions(CG);
00632 }
00633 
00634 /// removeDeadFunctions - Remove dead functions that are not included in
00635 /// DNR (Do Not Remove) list.
00636 bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) {
00637   SmallVector<CallGraphNode*, 16> FunctionsToRemove;
00638 
00639   // Scan for all of the functions, looking for ones that should now be removed
00640   // from the program.  Insert the dead ones in the FunctionsToRemove set.
00641   for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) {
00642     CallGraphNode *CGN = I->second;
00643     Function *F = CGN->getFunction();
00644     if (!F || F->isDeclaration())
00645       continue;
00646 
00647     // Handle the case when this function is called and we only want to care
00648     // about always-inline functions. This is a bit of a hack to share code
00649     // between here and the InlineAlways pass.
00650     if (AlwaysInlineOnly && !F->hasFnAttribute(Attribute::AlwaysInline))
00651       continue;
00652 
00653     // If the only remaining users of the function are dead constants, remove
00654     // them.
00655     F->removeDeadConstantUsers();
00656 
00657     if (!F->isDefTriviallyDead())
00658       continue;
00659 
00660     // It is unsafe to drop a function with discardable linkage from a COMDAT
00661     // without also dropping the other members of the COMDAT.
00662     // The inliner doesn't visit non-function entities which are in COMDAT
00663     // groups so it is unsafe to do so *unless* the linkage is local.
00664     if (!F->hasLocalLinkage() && F->hasComdat())
00665       continue;
00666     
00667     // Remove any call graph edges from the function to its callees.
00668     CGN->removeAllCalledFunctions();
00669 
00670     // Remove any edges from the external node to the function's call graph
00671     // node.  These edges might have been made irrelegant due to
00672     // optimization of the program.
00673     CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
00674 
00675     // Removing the node for callee from the call graph and delete it.
00676     FunctionsToRemove.push_back(CGN);
00677   }
00678   if (FunctionsToRemove.empty())
00679     return false;
00680 
00681   // Now that we know which functions to delete, do so.  We didn't want to do
00682   // this inline, because that would invalidate our CallGraph::iterator
00683   // objects. :(
00684   //
00685   // Note that it doesn't matter that we are iterating over a non-stable order
00686   // here to do this, it doesn't matter which order the functions are deleted
00687   // in.
00688   array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
00689   FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(),
00690                                       FunctionsToRemove.end()),
00691                           FunctionsToRemove.end());
00692   for (SmallVectorImpl<CallGraphNode *>::iterator I = FunctionsToRemove.begin(),
00693                                                   E = FunctionsToRemove.end();
00694        I != E; ++I) {
00695     delete CG.removeFunctionFromModule(*I);
00696     ++NumDeleted;
00697   }
00698   return true;
00699 }