Home | History | Annotate | Download | only in IPO
      1 //===- Inliner.cpp - Code common to all inliners --------------------------===//
      2 //
      3 //                     The LLVM Compiler Infrastructure
      4 //
      5 // This file is distributed under the University of Illinois Open Source
      6 // License. See LICENSE.TXT for details.
      7 //
      8 //===----------------------------------------------------------------------===//
      9 //
     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.
     13 //
     14 //===----------------------------------------------------------------------===//
     15 
     16 #define DEBUG_TYPE "inline"
     17 #include "llvm/Transforms/IPO/InlinerPass.h"
     18 #include "llvm/ADT/SmallPtrSet.h"
     19 #include "llvm/ADT/Statistic.h"
     20 #include "llvm/Analysis/CallGraph.h"
     21 #include "llvm/Analysis/InlineCost.h"
     22 #include "llvm/IR/DataLayout.h"
     23 #include "llvm/IR/Instructions.h"
     24 #include "llvm/IR/IntrinsicInst.h"
     25 #include "llvm/IR/Module.h"
     26 #include "llvm/Support/CallSite.h"
     27 #include "llvm/Support/CommandLine.h"
     28 #include "llvm/Support/Debug.h"
     29 #include "llvm/Support/raw_ostream.h"
     30 #include "llvm/Target/TargetLibraryInfo.h"
     31 #include "llvm/Transforms/Utils/Cloning.h"
     32 #include "llvm/Transforms/Utils/Local.h"
     33 using namespace llvm;
     34 
     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");
     39 
     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");
     44 
     45 static cl::opt<int>
     46 InlineLimit("inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore,
     47         cl::desc("Control the amount of inlining to perform (default = 225)"));
     48 
     49 static cl::opt<int>
     50 HintThreshold("inlinehint-threshold", cl::Hidden, cl::init(325),
     51               cl::desc("Threshold for inlining functions with inline hint"));
     52 
     53 // Threshold to use when optsize is specified (and there is no -inline-limit).
     54 const int OptSizeThreshold = 75;
     55 
     56 Inliner::Inliner(char &ID)
     57   : CallGraphSCCPass(ID), InlineThreshold(InlineLimit), InsertLifetime(true) {}
     58 
     59 Inliner::Inliner(char &ID, int Threshold, bool InsertLifetime)
     60   : CallGraphSCCPass(ID), InlineThreshold(InlineLimit.getNumOccurrences() > 0 ?
     61                                           InlineLimit : Threshold),
     62     InsertLifetime(InsertLifetime) {}
     63 
     64 /// getAnalysisUsage - For this class, we declare that we require and preserve
     65 /// the call graph.  If the derived class implements this method, it should
     66 /// always explicitly call the implementation here.
     67 void Inliner::getAnalysisUsage(AnalysisUsage &AU) const {
     68   CallGraphSCCPass::getAnalysisUsage(AU);
     69 }
     70 
     71 
     72 typedef DenseMap<ArrayType*, std::vector<AllocaInst*> >
     73 InlinedArrayAllocasTy;
     74 
     75 /// \brief If the inlined function had a higher stack protection level than the
     76 /// calling function, then bump up the caller's stack protection level.
     77 static void AdjustCallerSSPLevel(Function *Caller, Function *Callee) {
     78   // If upgrading the SSP attribute, clear out the old SSP Attributes first.
     79   // Having multiple SSP attributes doesn't actually hurt, but it adds useless
     80   // clutter to the IR.
     81   AttrBuilder B;
     82   B.addAttribute(Attribute::StackProtect)
     83     .addAttribute(Attribute::StackProtectStrong);
     84   AttributeSet OldSSPAttr = AttributeSet::get(Caller->getContext(),
     85                                               AttributeSet::FunctionIndex,
     86                                               B);
     87   AttributeSet CallerAttr = Caller->getAttributes(),
     88                CalleeAttr = Callee->getAttributes();
     89 
     90   if (CalleeAttr.hasAttribute(AttributeSet::FunctionIndex,
     91                               Attribute::StackProtectReq)) {
     92     Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
     93     Caller->addFnAttr(Attribute::StackProtectReq);
     94   } else if (CalleeAttr.hasAttribute(AttributeSet::FunctionIndex,
     95                                      Attribute::StackProtectStrong) &&
     96              !CallerAttr.hasAttribute(AttributeSet::FunctionIndex,
     97                                       Attribute::StackProtectReq)) {
     98     Caller->removeAttributes(AttributeSet::FunctionIndex, OldSSPAttr);
     99     Caller->addFnAttr(Attribute::StackProtectStrong);
    100   } else if (CalleeAttr.hasAttribute(AttributeSet::FunctionIndex,
    101                                      Attribute::StackProtect) &&
    102            !CallerAttr.hasAttribute(AttributeSet::FunctionIndex,
    103                                     Attribute::StackProtectReq) &&
    104            !CallerAttr.hasAttribute(AttributeSet::FunctionIndex,
    105                                     Attribute::StackProtectStrong))
    106     Caller->addFnAttr(Attribute::StackProtect);
    107 }
    108 
    109 /// InlineCallIfPossible - If it is possible to inline the specified call site,
    110 /// do so and update the CallGraph for this operation.
    111 ///
    112 /// This function also does some basic book-keeping to update the IR.  The
    113 /// InlinedArrayAllocas map keeps track of any allocas that are already
    114 /// available from other  functions inlined into the caller.  If we are able to
    115 /// inline this call site we attempt to reuse already available allocas or add
    116 /// any new allocas to the set if not possible.
    117 static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI,
    118                                  InlinedArrayAllocasTy &InlinedArrayAllocas,
    119                                  int InlineHistory, bool InsertLifetime,
    120                                  const DataLayout *TD) {
    121   Function *Callee = CS.getCalledFunction();
    122   Function *Caller = CS.getCaller();
    123 
    124   // Try to inline the function.  Get the list of static allocas that were
    125   // inlined.
    126   if (!InlineFunction(CS, IFI, InsertLifetime))
    127     return false;
    128 
    129   AdjustCallerSSPLevel(Caller, Callee);
    130 
    131   // Look at all of the allocas that we inlined through this call site.  If we
    132   // have already inlined other allocas through other calls into this function,
    133   // then we know that they have disjoint lifetimes and that we can merge them.
    134   //
    135   // There are many heuristics possible for merging these allocas, and the
    136   // different options have different tradeoffs.  One thing that we *really*
    137   // don't want to hurt is SRoA: once inlining happens, often allocas are no
    138   // longer address taken and so they can be promoted.
    139   //
    140   // Our "solution" for that is to only merge allocas whose outermost type is an
    141   // array type.  These are usually not promoted because someone is using a
    142   // variable index into them.  These are also often the most important ones to
    143   // merge.
    144   //
    145   // A better solution would be to have real memory lifetime markers in the IR
    146   // and not have the inliner do any merging of allocas at all.  This would
    147   // allow the backend to do proper stack slot coloring of all allocas that
    148   // *actually make it to the backend*, which is really what we want.
    149   //
    150   // Because we don't have this information, we do this simple and useful hack.
    151   //
    152   SmallPtrSet<AllocaInst*, 16> UsedAllocas;
    153 
    154   // When processing our SCC, check to see if CS was inlined from some other
    155   // call site.  For example, if we're processing "A" in this code:
    156   //   A() { B() }
    157   //   B() { x = alloca ... C() }
    158   //   C() { y = alloca ... }
    159   // Assume that C was not inlined into B initially, and so we're processing A
    160   // and decide to inline B into A.  Doing this makes an alloca available for
    161   // reuse and makes a callsite (C) available for inlining.  When we process
    162   // the C call site we don't want to do any alloca merging between X and Y
    163   // because their scopes are not disjoint.  We could make this smarter by
    164   // keeping track of the inline history for each alloca in the
    165   // InlinedArrayAllocas but this isn't likely to be a significant win.
    166   if (InlineHistory != -1)  // Only do merging for top-level call sites in SCC.
    167     return true;
    168 
    169   // Loop over all the allocas we have so far and see if they can be merged with
    170   // a previously inlined alloca.  If not, remember that we had it.
    171   for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
    172        AllocaNo != e; ++AllocaNo) {
    173     AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
    174 
    175     // Don't bother trying to merge array allocations (they will usually be
    176     // canonicalized to be an allocation *of* an array), or allocations whose
    177     // type is not itself an array (because we're afraid of pessimizing SRoA).
    178     ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
    179     if (ATy == 0 || AI->isArrayAllocation())
    180       continue;
    181 
    182     // Get the list of all available allocas for this array type.
    183     std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
    184 
    185     // Loop over the allocas in AllocasForType to see if we can reuse one.  Note
    186     // that we have to be careful not to reuse the same "available" alloca for
    187     // multiple different allocas that we just inlined, we use the 'UsedAllocas'
    188     // set to keep track of which "available" allocas are being used by this
    189     // function.  Also, AllocasForType can be empty of course!
    190     bool MergedAwayAlloca = false;
    191     for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) {
    192       AllocaInst *AvailableAlloca = AllocasForType[i];
    193 
    194       unsigned Align1 = AI->getAlignment(),
    195                Align2 = AvailableAlloca->getAlignment();
    196       // If we don't have data layout information, and only one alloca is using
    197       // the target default, then we can't safely merge them because we can't
    198       // pick the greater alignment.
    199       if (!TD && (!Align1 || !Align2) && Align1 != Align2)
    200         continue;
    201 
    202       // The available alloca has to be in the right function, not in some other
    203       // function in this SCC.
    204       if (AvailableAlloca->getParent() != AI->getParent())
    205         continue;
    206 
    207       // If the inlined function already uses this alloca then we can't reuse
    208       // it.
    209       if (!UsedAllocas.insert(AvailableAlloca))
    210         continue;
    211 
    212       // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
    213       // success!
    214       DEBUG(dbgs() << "    ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: "
    215                    << *AvailableAlloca << '\n');
    216 
    217       AI->replaceAllUsesWith(AvailableAlloca);
    218 
    219       if (Align1 != Align2) {
    220         if (!Align1 || !Align2) {
    221           assert(TD && "DataLayout required to compare default alignments");
    222           unsigned TypeAlign = TD->getABITypeAlignment(AI->getAllocatedType());
    223 
    224           Align1 = Align1 ? Align1 : TypeAlign;
    225           Align2 = Align2 ? Align2 : TypeAlign;
    226         }
    227 
    228         if (Align1 > Align2)
    229           AvailableAlloca->setAlignment(AI->getAlignment());
    230       }
    231 
    232       AI->eraseFromParent();
    233       MergedAwayAlloca = true;
    234       ++NumMergedAllocas;
    235       IFI.StaticAllocas[AllocaNo] = 0;
    236       break;
    237     }
    238 
    239     // If we already nuked the alloca, we're done with it.
    240     if (MergedAwayAlloca)
    241       continue;
    242 
    243     // If we were unable to merge away the alloca either because there are no
    244     // allocas of the right type available or because we reused them all
    245     // already, remember that this alloca came from an inlined function and mark
    246     // it used so we don't reuse it for other allocas from this inline
    247     // operation.
    248     AllocasForType.push_back(AI);
    249     UsedAllocas.insert(AI);
    250   }
    251 
    252   return true;
    253 }
    254 
    255 unsigned Inliner::getInlineThreshold(CallSite CS) const {
    256   int thres = InlineThreshold; // -inline-threshold or else selected by
    257                                // overall opt level
    258 
    259   // If -inline-threshold is not given, listen to the optsize attribute when it
    260   // would decrease the threshold.
    261   Function *Caller = CS.getCaller();
    262   bool OptSize = Caller && !Caller->isDeclaration() &&
    263     Caller->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
    264                                          Attribute::OptimizeForSize);
    265   if (!(InlineLimit.getNumOccurrences() > 0) && OptSize &&
    266       OptSizeThreshold < thres)
    267     thres = OptSizeThreshold;
    268 
    269   // Listen to the inlinehint attribute when it would increase the threshold
    270   // and the caller does not need to minimize its size.
    271   Function *Callee = CS.getCalledFunction();
    272   bool InlineHint = Callee && !Callee->isDeclaration() &&
    273     Callee->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
    274                                          Attribute::InlineHint);
    275   if (InlineHint && HintThreshold > thres
    276       && !Caller->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
    277                                                Attribute::MinSize))
    278     thres = HintThreshold;
    279 
    280   return thres;
    281 }
    282 
    283 /// shouldInline - Return true if the inliner should attempt to inline
    284 /// at the given CallSite.
    285 bool Inliner::shouldInline(CallSite CS) {
    286   InlineCost IC = getInlineCost(CS);
    287 
    288   if (IC.isAlways()) {
    289     DEBUG(dbgs() << "    Inlining: cost=always"
    290           << ", Call: " << *CS.getInstruction() << "\n");
    291     return true;
    292   }
    293 
    294   if (IC.isNever()) {
    295     DEBUG(dbgs() << "    NOT Inlining: cost=never"
    296           << ", Call: " << *CS.getInstruction() << "\n");
    297     return false;
    298   }
    299 
    300   Function *Caller = CS.getCaller();
    301   if (!IC) {
    302     DEBUG(dbgs() << "    NOT Inlining: cost=" << IC.getCost()
    303           << ", thres=" << (IC.getCostDelta() + IC.getCost())
    304           << ", Call: " << *CS.getInstruction() << "\n");
    305     return false;
    306   }
    307 
    308   // Try to detect the case where the current inlining candidate caller (call
    309   // it B) is a static or linkonce-ODR function and is an inlining candidate
    310   // elsewhere, and the current candidate callee (call it C) is large enough
    311   // that inlining it into B would make B too big to inline later. In these
    312   // circumstances it may be best not to inline C into B, but to inline B into
    313   // its callers.
    314   //
    315   // This only applies to static and linkonce-ODR functions because those are
    316   // expected to be available for inlining in the translation units where they
    317   // are used. Thus we will always have the opportunity to make local inlining
    318   // decisions. Importantly the linkonce-ODR linkage covers inline functions
    319   // and templates in C++.
    320   //
    321   // FIXME: All of this logic should be sunk into getInlineCost. It relies on
    322   // the internal implementation of the inline cost metrics rather than
    323   // treating them as truly abstract units etc.
    324   if (Caller->hasLocalLinkage() ||
    325       Caller->getLinkage() == GlobalValue::LinkOnceODRLinkage) {
    326     int TotalSecondaryCost = 0;
    327     // The candidate cost to be imposed upon the current function.
    328     int CandidateCost = IC.getCost() - (InlineConstants::CallPenalty + 1);
    329     // This bool tracks what happens if we do NOT inline C into B.
    330     bool callerWillBeRemoved = Caller->hasLocalLinkage();
    331     // This bool tracks what happens if we DO inline C into B.
    332     bool inliningPreventsSomeOuterInline = false;
    333     for (Value::use_iterator I = Caller->use_begin(), E =Caller->use_end();
    334          I != E; ++I) {
    335       CallSite CS2(*I);
    336 
    337       // If this isn't a call to Caller (it could be some other sort
    338       // of reference) skip it.  Such references will prevent the caller
    339       // from being removed.
    340       if (!CS2 || CS2.getCalledFunction() != Caller) {
    341         callerWillBeRemoved = false;
    342         continue;
    343       }
    344 
    345       InlineCost IC2 = getInlineCost(CS2);
    346       ++NumCallerCallersAnalyzed;
    347       if (!IC2) {
    348         callerWillBeRemoved = false;
    349         continue;
    350       }
    351       if (IC2.isAlways())
    352         continue;
    353 
    354       // See if inlining or original callsite would erase the cost delta of
    355       // this callsite. We subtract off the penalty for the call instruction,
    356       // which we would be deleting.
    357       if (IC2.getCostDelta() <= CandidateCost) {
    358         inliningPreventsSomeOuterInline = true;
    359         TotalSecondaryCost += IC2.getCost();
    360       }
    361     }
    362     // If all outer calls to Caller would get inlined, the cost for the last
    363     // one is set very low by getInlineCost, in anticipation that Caller will
    364     // be removed entirely.  We did not account for this above unless there
    365     // is only one caller of Caller.
    366     if (callerWillBeRemoved && Caller->use_begin() != Caller->use_end())
    367       TotalSecondaryCost += InlineConstants::LastCallToStaticBonus;
    368 
    369     if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost()) {
    370       DEBUG(dbgs() << "    NOT Inlining: " << *CS.getInstruction() <<
    371            " Cost = " << IC.getCost() <<
    372            ", outer Cost = " << TotalSecondaryCost << '\n');
    373       return false;
    374     }
    375   }
    376 
    377   DEBUG(dbgs() << "    Inlining: cost=" << IC.getCost()
    378         << ", thres=" << (IC.getCostDelta() + IC.getCost())
    379         << ", Call: " << *CS.getInstruction() << '\n');
    380   return true;
    381 }
    382 
    383 /// InlineHistoryIncludes - Return true if the specified inline history ID
    384 /// indicates an inline history that includes the specified function.
    385 static bool InlineHistoryIncludes(Function *F, int InlineHistoryID,
    386             const SmallVectorImpl<std::pair<Function*, int> > &InlineHistory) {
    387   while (InlineHistoryID != -1) {
    388     assert(unsigned(InlineHistoryID) < InlineHistory.size() &&
    389            "Invalid inline history ID");
    390     if (InlineHistory[InlineHistoryID].first == F)
    391       return true;
    392     InlineHistoryID = InlineHistory[InlineHistoryID].second;
    393   }
    394   return false;
    395 }
    396 
    397 bool Inliner::runOnSCC(CallGraphSCC &SCC) {
    398   CallGraph &CG = getAnalysis<CallGraph>();
    399   const DataLayout *TD = getAnalysisIfAvailable<DataLayout>();
    400   const TargetLibraryInfo *TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
    401 
    402   SmallPtrSet<Function*, 8> SCCFunctions;
    403   DEBUG(dbgs() << "Inliner visiting SCC:");
    404   for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
    405     Function *F = (*I)->getFunction();
    406     if (F) SCCFunctions.insert(F);
    407     DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
    408   }
    409 
    410   // Scan through and identify all call sites ahead of time so that we only
    411   // inline call sites in the original functions, not call sites that result
    412   // from inlining other functions.
    413   SmallVector<std::pair<CallSite, int>, 16> CallSites;
    414 
    415   // When inlining a callee produces new call sites, we want to keep track of
    416   // the fact that they were inlined from the callee.  This allows us to avoid
    417   // infinite inlining in some obscure cases.  To represent this, we use an
    418   // index into the InlineHistory vector.
    419   SmallVector<std::pair<Function*, int>, 8> InlineHistory;
    420 
    421   for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
    422     Function *F = (*I)->getFunction();
    423     if (!F) continue;
    424 
    425     for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
    426       for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
    427         CallSite CS(cast<Value>(I));
    428         // If this isn't a call, or it is a call to an intrinsic, it can
    429         // never be inlined.
    430         if (!CS || isa<IntrinsicInst>(I))
    431           continue;
    432 
    433         // If this is a direct call to an external function, we can never inline
    434         // it.  If it is an indirect call, inlining may resolve it to be a
    435         // direct call, so we keep it.
    436         if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration())
    437           continue;
    438 
    439         CallSites.push_back(std::make_pair(CS, -1));
    440       }
    441   }
    442 
    443   DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
    444 
    445   // If there are no calls in this function, exit early.
    446   if (CallSites.empty())
    447     return false;
    448 
    449   // Now that we have all of the call sites, move the ones to functions in the
    450   // current SCC to the end of the list.
    451   unsigned FirstCallInSCC = CallSites.size();
    452   for (unsigned i = 0; i < FirstCallInSCC; ++i)
    453     if (Function *F = CallSites[i].first.getCalledFunction())
    454       if (SCCFunctions.count(F))
    455         std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
    456 
    457 
    458   InlinedArrayAllocasTy InlinedArrayAllocas;
    459   InlineFunctionInfo InlineInfo(&CG, TD);
    460 
    461   // Now that we have all of the call sites, loop over them and inline them if
    462   // it looks profitable to do so.
    463   bool Changed = false;
    464   bool LocalChange;
    465   do {
    466     LocalChange = false;
    467     // Iterate over the outer loop because inlining functions can cause indirect
    468     // calls to become direct calls.
    469     for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
    470       CallSite CS = CallSites[CSi].first;
    471 
    472       Function *Caller = CS.getCaller();
    473       Function *Callee = CS.getCalledFunction();
    474 
    475       // If this call site is dead and it is to a readonly function, we should
    476       // just delete the call instead of trying to inline it, regardless of
    477       // size.  This happens because IPSCCP propagates the result out of the
    478       // call and then we're left with the dead call.
    479       if (isInstructionTriviallyDead(CS.getInstruction(), TLI)) {
    480         DEBUG(dbgs() << "    -> Deleting dead call: "
    481                      << *CS.getInstruction() << "\n");
    482         // Update the call graph by deleting the edge from Callee to Caller.
    483         CG[Caller]->removeCallEdgeFor(CS);
    484         CS.getInstruction()->eraseFromParent();
    485         ++NumCallsDeleted;
    486       } else {
    487         // We can only inline direct calls to non-declarations.
    488         if (Callee == 0 || Callee->isDeclaration()) continue;
    489 
    490         // If this call site was obtained by inlining another function, verify
    491         // that the include path for the function did not include the callee
    492         // itself.  If so, we'd be recursively inlining the same function,
    493         // which would provide the same callsites, which would cause us to
    494         // infinitely inline.
    495         int InlineHistoryID = CallSites[CSi].second;
    496         if (InlineHistoryID != -1 &&
    497             InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory))
    498           continue;
    499 
    500 
    501         // If the policy determines that we should inline this function,
    502         // try to do so.
    503         if (!shouldInline(CS))
    504           continue;
    505 
    506         // Attempt to inline the function.
    507         if (!InlineCallIfPossible(CS, InlineInfo, InlinedArrayAllocas,
    508                                   InlineHistoryID, InsertLifetime, TD))
    509           continue;
    510         ++NumInlined;
    511 
    512         // If inlining this function gave us any new call sites, throw them
    513         // onto our worklist to process.  They are useful inline candidates.
    514         if (!InlineInfo.InlinedCalls.empty()) {
    515           // Create a new inline history entry for this, so that we remember
    516           // that these new callsites came about due to inlining Callee.
    517           int NewHistoryID = InlineHistory.size();
    518           InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID));
    519 
    520           for (unsigned i = 0, e = InlineInfo.InlinedCalls.size();
    521                i != e; ++i) {
    522             Value *Ptr = InlineInfo.InlinedCalls[i];
    523             CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID));
    524           }
    525         }
    526       }
    527 
    528       // If we inlined or deleted the last possible call site to the function,
    529       // delete the function body now.
    530       if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
    531           // TODO: Can remove if in SCC now.
    532           !SCCFunctions.count(Callee) &&
    533 
    534           // The function may be apparently dead, but if there are indirect
    535           // callgraph references to the node, we cannot delete it yet, this
    536           // could invalidate the CGSCC iterator.
    537           CG[Callee]->getNumReferences() == 0) {
    538         DEBUG(dbgs() << "    -> Deleting dead function: "
    539               << Callee->getName() << "\n");
    540         CallGraphNode *CalleeNode = CG[Callee];
    541 
    542         // Remove any call graph edges from the callee to its callees.
    543         CalleeNode->removeAllCalledFunctions();
    544 
    545         // Removing the node for callee from the call graph and delete it.
    546         delete CG.removeFunctionFromModule(CalleeNode);
    547         ++NumDeleted;
    548       }
    549 
    550       // Remove this call site from the list.  If possible, use
    551       // swap/pop_back for efficiency, but do not use it if doing so would
    552       // move a call site to a function in this SCC before the
    553       // 'FirstCallInSCC' barrier.
    554       if (SCC.isSingular()) {
    555         CallSites[CSi] = CallSites.back();
    556         CallSites.pop_back();
    557       } else {
    558         CallSites.erase(CallSites.begin()+CSi);
    559       }
    560       --CSi;
    561 
    562       Changed = true;
    563       LocalChange = true;
    564     }
    565   } while (LocalChange);
    566 
    567   return Changed;
    568 }
    569 
    570 // doFinalization - Remove now-dead linkonce functions at the end of
    571 // processing to avoid breaking the SCC traversal.
    572 bool Inliner::doFinalization(CallGraph &CG) {
    573   return removeDeadFunctions(CG);
    574 }
    575 
    576 /// removeDeadFunctions - Remove dead functions that are not included in
    577 /// DNR (Do Not Remove) list.
    578 bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) {
    579   SmallVector<CallGraphNode*, 16> FunctionsToRemove;
    580 
    581   // Scan for all of the functions, looking for ones that should now be removed
    582   // from the program.  Insert the dead ones in the FunctionsToRemove set.
    583   for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) {
    584     CallGraphNode *CGN = I->second;
    585     Function *F = CGN->getFunction();
    586     if (!F || F->isDeclaration())
    587       continue;
    588 
    589     // Handle the case when this function is called and we only want to care
    590     // about always-inline functions. This is a bit of a hack to share code
    591     // between here and the InlineAlways pass.
    592     if (AlwaysInlineOnly &&
    593         !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
    594                                          Attribute::AlwaysInline))
    595       continue;
    596 
    597     // If the only remaining users of the function are dead constants, remove
    598     // them.
    599     F->removeDeadConstantUsers();
    600 
    601     if (!F->isDefTriviallyDead())
    602       continue;
    603 
    604     // Remove any call graph edges from the function to its callees.
    605     CGN->removeAllCalledFunctions();
    606 
    607     // Remove any edges from the external node to the function's call graph
    608     // node.  These edges might have been made irrelegant due to
    609     // optimization of the program.
    610     CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
    611 
    612     // Removing the node for callee from the call graph and delete it.
    613     FunctionsToRemove.push_back(CGN);
    614   }
    615   if (FunctionsToRemove.empty())
    616     return false;
    617 
    618   // Now that we know which functions to delete, do so.  We didn't want to do
    619   // this inline, because that would invalidate our CallGraph::iterator
    620   // objects. :(
    621   //
    622   // Note that it doesn't matter that we are iterating over a non-stable order
    623   // here to do this, it doesn't matter which order the functions are deleted
    624   // in.
    625   array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end());
    626   FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(),
    627                                       FunctionsToRemove.end()),
    628                           FunctionsToRemove.end());
    629   for (SmallVectorImpl<CallGraphNode *>::iterator I = FunctionsToRemove.begin(),
    630                                                   E = FunctionsToRemove.end();
    631        I != E; ++I) {
    632     delete CG.removeFunctionFromModule(*I);
    633     ++NumDeleted;
    634   }
    635   return true;
    636 }
    637