LLVM API Documentation

InlineFunction.cpp
Go to the documentation of this file.
00001 //===- InlineFunction.cpp - Code to perform function inlining -------------===//
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 inlining of a function into a call site, resolving
00011 // parameters and the return value as appropriate.
00012 //
00013 //===----------------------------------------------------------------------===//
00014 
00015 #include "llvm/Transforms/Utils/Cloning.h"
00016 #include "llvm/ADT/SmallSet.h"
00017 #include "llvm/ADT/SmallVector.h"
00018 #include "llvm/ADT/SetVector.h"
00019 #include "llvm/ADT/StringExtras.h"
00020 #include "llvm/Analysis/AliasAnalysis.h"
00021 #include "llvm/Analysis/AssumptionCache.h"
00022 #include "llvm/Analysis/CallGraph.h"
00023 #include "llvm/Analysis/CaptureTracking.h"
00024 #include "llvm/Analysis/InstructionSimplify.h"
00025 #include "llvm/Analysis/ValueTracking.h"
00026 #include "llvm/IR/Attributes.h"
00027 #include "llvm/IR/CallSite.h"
00028 #include "llvm/IR/CFG.h"
00029 #include "llvm/IR/Constants.h"
00030 #include "llvm/IR/DataLayout.h"
00031 #include "llvm/IR/DebugInfo.h"
00032 #include "llvm/IR/DerivedTypes.h"
00033 #include "llvm/IR/Dominators.h"
00034 #include "llvm/IR/IRBuilder.h"
00035 #include "llvm/IR/Instructions.h"
00036 #include "llvm/IR/IntrinsicInst.h"
00037 #include "llvm/IR/Intrinsics.h"
00038 #include "llvm/IR/MDBuilder.h"
00039 #include "llvm/IR/Module.h"
00040 #include "llvm/Transforms/Utils/Local.h"
00041 #include "llvm/Support/CommandLine.h"
00042 #include <algorithm>
00043 using namespace llvm;
00044 
00045 static cl::opt<bool>
00046 EnableNoAliasConversion("enable-noalias-to-md-conversion", cl::init(true),
00047   cl::Hidden,
00048   cl::desc("Convert noalias attributes to metadata during inlining."));
00049 
00050 static cl::opt<bool>
00051 PreserveAlignmentAssumptions("preserve-alignment-assumptions-during-inlining",
00052   cl::init(true), cl::Hidden,
00053   cl::desc("Convert align attributes to assumptions during inlining."));
00054 
00055 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI,
00056                           bool InsertLifetime) {
00057   return InlineFunction(CallSite(CI), IFI, InsertLifetime);
00058 }
00059 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI,
00060                           bool InsertLifetime) {
00061   return InlineFunction(CallSite(II), IFI, InsertLifetime);
00062 }
00063 
00064 namespace {
00065   /// A class for recording information about inlining through an invoke.
00066   class InvokeInliningInfo {
00067     BasicBlock *OuterResumeDest; ///< Destination of the invoke's unwind.
00068     BasicBlock *InnerResumeDest; ///< Destination for the callee's resume.
00069     LandingPadInst *CallerLPad;  ///< LandingPadInst associated with the invoke.
00070     PHINode *InnerEHValuesPHI;   ///< PHI for EH values from landingpad insts.
00071     SmallVector<Value*, 8> UnwindDestPHIValues;
00072 
00073   public:
00074     InvokeInliningInfo(InvokeInst *II)
00075       : OuterResumeDest(II->getUnwindDest()), InnerResumeDest(nullptr),
00076         CallerLPad(nullptr), InnerEHValuesPHI(nullptr) {
00077       // If there are PHI nodes in the unwind destination block, we need to keep
00078       // track of which values came into them from the invoke before removing
00079       // the edge from this block.
00080       llvm::BasicBlock *InvokeBB = II->getParent();
00081       BasicBlock::iterator I = OuterResumeDest->begin();
00082       for (; isa<PHINode>(I); ++I) {
00083         // Save the value to use for this edge.
00084         PHINode *PHI = cast<PHINode>(I);
00085         UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
00086       }
00087 
00088       CallerLPad = cast<LandingPadInst>(I);
00089     }
00090 
00091     /// getOuterResumeDest - The outer unwind destination is the target of
00092     /// unwind edges introduced for calls within the inlined function.
00093     BasicBlock *getOuterResumeDest() const {
00094       return OuterResumeDest;
00095     }
00096 
00097     BasicBlock *getInnerResumeDest();
00098 
00099     LandingPadInst *getLandingPadInst() const { return CallerLPad; }
00100 
00101     /// forwardResume - Forward the 'resume' instruction to the caller's landing
00102     /// pad block. When the landing pad block has only one predecessor, this is
00103     /// a simple branch. When there is more than one predecessor, we need to
00104     /// split the landing pad block after the landingpad instruction and jump
00105     /// to there.
00106     void forwardResume(ResumeInst *RI,
00107                        SmallPtrSetImpl<LandingPadInst*> &InlinedLPads);
00108 
00109     /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind
00110     /// destination block for the given basic block, using the values for the
00111     /// original invoke's source block.
00112     void addIncomingPHIValuesFor(BasicBlock *BB) const {
00113       addIncomingPHIValuesForInto(BB, OuterResumeDest);
00114     }
00115 
00116     void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const {
00117       BasicBlock::iterator I = dest->begin();
00118       for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
00119         PHINode *phi = cast<PHINode>(I);
00120         phi->addIncoming(UnwindDestPHIValues[i], src);
00121       }
00122     }
00123   };
00124 }
00125 
00126 /// getInnerResumeDest - Get or create a target for the branch from ResumeInsts.
00127 BasicBlock *InvokeInliningInfo::getInnerResumeDest() {
00128   if (InnerResumeDest) return InnerResumeDest;
00129 
00130   // Split the landing pad.
00131   BasicBlock::iterator SplitPoint = CallerLPad; ++SplitPoint;
00132   InnerResumeDest =
00133     OuterResumeDest->splitBasicBlock(SplitPoint,
00134                                      OuterResumeDest->getName() + ".body");
00135 
00136   // The number of incoming edges we expect to the inner landing pad.
00137   const unsigned PHICapacity = 2;
00138 
00139   // Create corresponding new PHIs for all the PHIs in the outer landing pad.
00140   BasicBlock::iterator InsertPoint = InnerResumeDest->begin();
00141   BasicBlock::iterator I = OuterResumeDest->begin();
00142   for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
00143     PHINode *OuterPHI = cast<PHINode>(I);
00144     PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity,
00145                                         OuterPHI->getName() + ".lpad-body",
00146                                         InsertPoint);
00147     OuterPHI->replaceAllUsesWith(InnerPHI);
00148     InnerPHI->addIncoming(OuterPHI, OuterResumeDest);
00149   }
00150 
00151   // Create a PHI for the exception values.
00152   InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity,
00153                                      "eh.lpad-body", InsertPoint);
00154   CallerLPad->replaceAllUsesWith(InnerEHValuesPHI);
00155   InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest);
00156 
00157   // All done.
00158   return InnerResumeDest;
00159 }
00160 
00161 /// forwardResume - Forward the 'resume' instruction to the caller's landing pad
00162 /// block. When the landing pad block has only one predecessor, this is a simple
00163 /// branch. When there is more than one predecessor, we need to split the
00164 /// landing pad block after the landingpad instruction and jump to there.
00165 void InvokeInliningInfo::forwardResume(ResumeInst *RI,
00166                                SmallPtrSetImpl<LandingPadInst*> &InlinedLPads) {
00167   BasicBlock *Dest = getInnerResumeDest();
00168   BasicBlock *Src = RI->getParent();
00169 
00170   BranchInst::Create(Dest, Src);
00171 
00172   // Update the PHIs in the destination. They were inserted in an order which
00173   // makes this work.
00174   addIncomingPHIValuesForInto(Src, Dest);
00175 
00176   InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src);
00177   RI->eraseFromParent();
00178 }
00179 
00180 /// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
00181 /// an invoke, we have to turn all of the calls that can throw into
00182 /// invokes.  This function analyze BB to see if there are any calls, and if so,
00183 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
00184 /// nodes in that block with the values specified in InvokeDestPHIValues.
00185 static void HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
00186                                                    InvokeInliningInfo &Invoke) {
00187   for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
00188     Instruction *I = BBI++;
00189 
00190     // We only need to check for function calls: inlined invoke
00191     // instructions require no special handling.
00192     CallInst *CI = dyn_cast<CallInst>(I);
00193 
00194     // If this call cannot unwind, don't convert it to an invoke.
00195     // Inline asm calls cannot throw.
00196     if (!CI || CI->doesNotThrow() || isa<InlineAsm>(CI->getCalledValue()))
00197       continue;
00198 
00199     // Convert this function call into an invoke instruction.  First, split the
00200     // basic block.
00201     BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
00202 
00203     // Delete the unconditional branch inserted by splitBasicBlock
00204     BB->getInstList().pop_back();
00205 
00206     // Create the new invoke instruction.
00207     ImmutableCallSite CS(CI);
00208     SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
00209     InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split,
00210                                         Invoke.getOuterResumeDest(),
00211                                         InvokeArgs, CI->getName(), BB);
00212     II->setDebugLoc(CI->getDebugLoc());
00213     II->setCallingConv(CI->getCallingConv());
00214     II->setAttributes(CI->getAttributes());
00215     
00216     // Make sure that anything using the call now uses the invoke!  This also
00217     // updates the CallGraph if present, because it uses a WeakVH.
00218     CI->replaceAllUsesWith(II);
00219 
00220     // Delete the original call
00221     Split->getInstList().pop_front();
00222 
00223     // Update any PHI nodes in the exceptional block to indicate that there is
00224     // now a new entry in them.
00225     Invoke.addIncomingPHIValuesFor(BB);
00226     return;
00227   }
00228 }
00229 
00230 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
00231 /// in the body of the inlined function into invokes.
00232 ///
00233 /// II is the invoke instruction being inlined.  FirstNewBlock is the first
00234 /// block of the inlined code (the last block is the end of the function),
00235 /// and InlineCodeInfo is information about the code that got inlined.
00236 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
00237                                 ClonedCodeInfo &InlinedCodeInfo) {
00238   BasicBlock *InvokeDest = II->getUnwindDest();
00239 
00240   Function *Caller = FirstNewBlock->getParent();
00241 
00242   // The inlined code is currently at the end of the function, scan from the
00243   // start of the inlined code to its end, checking for stuff we need to
00244   // rewrite.
00245   InvokeInliningInfo Invoke(II);
00246 
00247   // Get all of the inlined landing pad instructions.
00248   SmallPtrSet<LandingPadInst*, 16> InlinedLPads;
00249   for (Function::iterator I = FirstNewBlock, E = Caller->end(); I != E; ++I)
00250     if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator()))
00251       InlinedLPads.insert(II->getLandingPadInst());
00252 
00253   // Append the clauses from the outer landing pad instruction into the inlined
00254   // landing pad instructions.
00255   LandingPadInst *OuterLPad = Invoke.getLandingPadInst();
00256   for (LandingPadInst *InlinedLPad : InlinedLPads) {
00257     unsigned OuterNum = OuterLPad->getNumClauses();
00258     InlinedLPad->reserveClauses(OuterNum);
00259     for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx)
00260       InlinedLPad->addClause(OuterLPad->getClause(OuterIdx));
00261     if (OuterLPad->isCleanup())
00262       InlinedLPad->setCleanup(true);
00263   }
00264 
00265   for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
00266     if (InlinedCodeInfo.ContainsCalls)
00267       HandleCallsInBlockInlinedThroughInvoke(BB, Invoke);
00268 
00269     // Forward any resumes that are remaining here.
00270     if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator()))
00271       Invoke.forwardResume(RI, InlinedLPads);
00272   }
00273 
00274   // Now that everything is happy, we have one final detail.  The PHI nodes in
00275   // the exception destination block still have entries due to the original
00276   // invoke instruction. Eliminate these entries (which might even delete the
00277   // PHI node) now.
00278   InvokeDest->removePredecessor(II->getParent());
00279 }
00280 
00281 /// CloneAliasScopeMetadata - When inlining a function that contains noalias
00282 /// scope metadata, this metadata needs to be cloned so that the inlined blocks
00283 /// have different "unqiue scopes" at every call site. Were this not done, then
00284 /// aliasing scopes from a function inlined into a caller multiple times could
00285 /// not be differentiated (and this would lead to miscompiles because the
00286 /// non-aliasing property communicated by the metadata could have
00287 /// call-site-specific control dependencies).
00288 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) {
00289   const Function *CalledFunc = CS.getCalledFunction();
00290   SetVector<const MDNode *> MD;
00291 
00292   // Note: We could only clone the metadata if it is already used in the
00293   // caller. I'm omitting that check here because it might confuse
00294   // inter-procedural alias analysis passes. We can revisit this if it becomes
00295   // an efficiency or overhead problem.
00296 
00297   for (Function::const_iterator I = CalledFunc->begin(), IE = CalledFunc->end();
00298        I != IE; ++I)
00299     for (BasicBlock::const_iterator J = I->begin(), JE = I->end(); J != JE; ++J) {
00300       if (const MDNode *M = J->getMetadata(LLVMContext::MD_alias_scope))
00301         MD.insert(M);
00302       if (const MDNode *M = J->getMetadata(LLVMContext::MD_noalias))
00303         MD.insert(M);
00304     }
00305 
00306   if (MD.empty())
00307     return;
00308 
00309   // Walk the existing metadata, adding the complete (perhaps cyclic) chain to
00310   // the set.
00311   SmallVector<const Metadata *, 16> Queue(MD.begin(), MD.end());
00312   while (!Queue.empty()) {
00313     const MDNode *M = cast<MDNode>(Queue.pop_back_val());
00314     for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i)
00315       if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i)))
00316         if (MD.insert(M1))
00317           Queue.push_back(M1);
00318   }
00319 
00320   // Now we have a complete set of all metadata in the chains used to specify
00321   // the noalias scopes and the lists of those scopes.
00322   SmallVector<TempMDTuple, 16> DummyNodes;
00323   DenseMap<const MDNode *, TrackingMDNodeRef> MDMap;
00324   for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
00325        I != IE; ++I) {
00326     DummyNodes.push_back(MDTuple::getTemporary(CalledFunc->getContext(), None));
00327     MDMap[*I].reset(DummyNodes.back().get());
00328   }
00329 
00330   // Create new metadata nodes to replace the dummy nodes, replacing old
00331   // metadata references with either a dummy node or an already-created new
00332   // node.
00333   for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
00334        I != IE; ++I) {
00335     SmallVector<Metadata *, 4> NewOps;
00336     for (unsigned i = 0, ie = (*I)->getNumOperands(); i != ie; ++i) {
00337       const Metadata *V = (*I)->getOperand(i);
00338       if (const MDNode *M = dyn_cast<MDNode>(V))
00339         NewOps.push_back(MDMap[M]);
00340       else
00341         NewOps.push_back(const_cast<Metadata *>(V));
00342     }
00343 
00344     MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps);
00345     MDTuple *TempM = cast<MDTuple>(MDMap[*I]);
00346     assert(TempM->isTemporary() && "Expected temporary node");
00347 
00348     TempM->replaceAllUsesWith(NewM);
00349   }
00350 
00351   // Now replace the metadata in the new inlined instructions with the
00352   // repacements from the map.
00353   for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
00354        VMI != VMIE; ++VMI) {
00355     if (!VMI->second)
00356       continue;
00357 
00358     Instruction *NI = dyn_cast<Instruction>(VMI->second);
00359     if (!NI)
00360       continue;
00361 
00362     if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope)) {
00363       MDNode *NewMD = MDMap[M];
00364       // If the call site also had alias scope metadata (a list of scopes to
00365       // which instructions inside it might belong), propagate those scopes to
00366       // the inlined instructions.
00367       if (MDNode *CSM =
00368               CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
00369         NewMD = MDNode::concatenate(NewMD, CSM);
00370       NI->setMetadata(LLVMContext::MD_alias_scope, NewMD);
00371     } else if (NI->mayReadOrWriteMemory()) {
00372       if (MDNode *M =
00373               CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
00374         NI->setMetadata(LLVMContext::MD_alias_scope, M);
00375     }
00376 
00377     if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias)) {
00378       MDNode *NewMD = MDMap[M];
00379       // If the call site also had noalias metadata (a list of scopes with
00380       // which instructions inside it don't alias), propagate those scopes to
00381       // the inlined instructions.
00382       if (MDNode *CSM =
00383               CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
00384         NewMD = MDNode::concatenate(NewMD, CSM);
00385       NI->setMetadata(LLVMContext::MD_noalias, NewMD);
00386     } else if (NI->mayReadOrWriteMemory()) {
00387       if (MDNode *M = CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
00388         NI->setMetadata(LLVMContext::MD_noalias, M);
00389     }
00390   }
00391 }
00392 
00393 /// AddAliasScopeMetadata - If the inlined function has noalias arguments, then
00394 /// add new alias scopes for each noalias argument, tag the mapped noalias
00395 /// parameters with noalias metadata specifying the new scope, and tag all
00396 /// non-derived loads, stores and memory intrinsics with the new alias scopes.
00397 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
00398                                   const DataLayout *DL, AliasAnalysis *AA) {
00399   if (!EnableNoAliasConversion)
00400     return;
00401 
00402   const Function *CalledFunc = CS.getCalledFunction();
00403   SmallVector<const Argument *, 4> NoAliasArgs;
00404 
00405   for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
00406        E = CalledFunc->arg_end(); I != E; ++I) {
00407     if (I->hasNoAliasAttr() && !I->hasNUses(0))
00408       NoAliasArgs.push_back(I);
00409   }
00410 
00411   if (NoAliasArgs.empty())
00412     return;
00413 
00414   // To do a good job, if a noalias variable is captured, we need to know if
00415   // the capture point dominates the particular use we're considering.
00416   DominatorTree DT;
00417   DT.recalculate(const_cast<Function&>(*CalledFunc));
00418 
00419   // noalias indicates that pointer values based on the argument do not alias
00420   // pointer values which are not based on it. So we add a new "scope" for each
00421   // noalias function argument. Accesses using pointers based on that argument
00422   // become part of that alias scope, accesses using pointers not based on that
00423   // argument are tagged as noalias with that scope.
00424 
00425   DenseMap<const Argument *, MDNode *> NewScopes;
00426   MDBuilder MDB(CalledFunc->getContext());
00427 
00428   // Create a new scope domain for this function.
00429   MDNode *NewDomain =
00430     MDB.createAnonymousAliasScopeDomain(CalledFunc->getName());
00431   for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) {
00432     const Argument *A = NoAliasArgs[i];
00433 
00434     std::string Name = CalledFunc->getName();
00435     if (A->hasName()) {
00436       Name += ": %";
00437       Name += A->getName();
00438     } else {
00439       Name += ": argument ";
00440       Name += utostr(i);
00441     }
00442 
00443     // Note: We always create a new anonymous root here. This is true regardless
00444     // of the linkage of the callee because the aliasing "scope" is not just a
00445     // property of the callee, but also all control dependencies in the caller.
00446     MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
00447     NewScopes.insert(std::make_pair(A, NewScope));
00448   }
00449 
00450   // Iterate over all new instructions in the map; for all memory-access
00451   // instructions, add the alias scope metadata.
00452   for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
00453        VMI != VMIE; ++VMI) {
00454     if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) {
00455       if (!VMI->second)
00456         continue;
00457 
00458       Instruction *NI = dyn_cast<Instruction>(VMI->second);
00459       if (!NI)
00460         continue;
00461 
00462       bool IsArgMemOnlyCall = false, IsFuncCall = false;
00463       SmallVector<const Value *, 2> PtrArgs;
00464 
00465       if (const LoadInst *LI = dyn_cast<LoadInst>(I))
00466         PtrArgs.push_back(LI->getPointerOperand());
00467       else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
00468         PtrArgs.push_back(SI->getPointerOperand());
00469       else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
00470         PtrArgs.push_back(VAAI->getPointerOperand());
00471       else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
00472         PtrArgs.push_back(CXI->getPointerOperand());
00473       else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
00474         PtrArgs.push_back(RMWI->getPointerOperand());
00475       else if (ImmutableCallSite ICS = ImmutableCallSite(I)) {
00476         // If we know that the call does not access memory, then we'll still
00477         // know that about the inlined clone of this call site, and we don't
00478         // need to add metadata.
00479         if (ICS.doesNotAccessMemory())
00480           continue;
00481 
00482         IsFuncCall = true;
00483         if (AA) {
00484           AliasAnalysis::ModRefBehavior MRB = AA->getModRefBehavior(ICS);
00485           if (MRB == AliasAnalysis::OnlyAccessesArgumentPointees ||
00486               MRB == AliasAnalysis::OnlyReadsArgumentPointees)
00487             IsArgMemOnlyCall = true;
00488         }
00489 
00490         for (ImmutableCallSite::arg_iterator AI = ICS.arg_begin(),
00491              AE = ICS.arg_end(); AI != AE; ++AI) {
00492           // We need to check the underlying objects of all arguments, not just
00493           // the pointer arguments, because we might be passing pointers as
00494           // integers, etc.
00495           // However, if we know that the call only accesses pointer arguments,
00496           // then we only need to check the pointer arguments.
00497           if (IsArgMemOnlyCall && !(*AI)->getType()->isPointerTy())
00498             continue;
00499 
00500           PtrArgs.push_back(*AI);
00501         }
00502       }
00503 
00504       // If we found no pointers, then this instruction is not suitable for
00505       // pairing with an instruction to receive aliasing metadata.
00506       // However, if this is a call, this we might just alias with none of the
00507       // noalias arguments.
00508       if (PtrArgs.empty() && !IsFuncCall)
00509         continue;
00510 
00511       // It is possible that there is only one underlying object, but you
00512       // need to go through several PHIs to see it, and thus could be
00513       // repeated in the Objects list.
00514       SmallPtrSet<const Value *, 4> ObjSet;
00515       SmallVector<Metadata *, 4> Scopes, NoAliases;
00516 
00517       SmallSetVector<const Argument *, 4> NAPtrArgs;
00518       for (unsigned i = 0, ie = PtrArgs.size(); i != ie; ++i) {
00519         SmallVector<Value *, 4> Objects;
00520         GetUnderlyingObjects(const_cast<Value*>(PtrArgs[i]),
00521                              Objects, DL, /* MaxLookup = */ 0);
00522 
00523         for (Value *O : Objects)
00524           ObjSet.insert(O);
00525       }
00526 
00527       // Figure out if we're derived from anything that is not a noalias
00528       // argument.
00529       bool CanDeriveViaCapture = false, UsesAliasingPtr = false;
00530       for (const Value *V : ObjSet) {
00531         // Is this value a constant that cannot be derived from any pointer
00532         // value (we need to exclude constant expressions, for example, that
00533         // are formed from arithmetic on global symbols).
00534         bool IsNonPtrConst = isa<ConstantInt>(V) || isa<ConstantFP>(V) ||
00535                              isa<ConstantPointerNull>(V) ||
00536                              isa<ConstantDataVector>(V) || isa<UndefValue>(V);
00537         if (IsNonPtrConst)
00538           continue;
00539 
00540         // If this is anything other than a noalias argument, then we cannot
00541         // completely describe the aliasing properties using alias.scope
00542         // metadata (and, thus, won't add any).
00543         if (const Argument *A = dyn_cast<Argument>(V)) {
00544           if (!A->hasNoAliasAttr())
00545             UsesAliasingPtr = true;
00546         } else {
00547           UsesAliasingPtr = true;
00548         }
00549 
00550         // If this is not some identified function-local object (which cannot
00551         // directly alias a noalias argument), or some other argument (which,
00552         // by definition, also cannot alias a noalias argument), then we could
00553         // alias a noalias argument that has been captured).
00554         if (!isa<Argument>(V) &&
00555             !isIdentifiedFunctionLocal(const_cast<Value*>(V)))
00556           CanDeriveViaCapture = true;
00557       }
00558 
00559       // A function call can always get captured noalias pointers (via other
00560       // parameters, globals, etc.).
00561       if (IsFuncCall && !IsArgMemOnlyCall)
00562         CanDeriveViaCapture = true;
00563 
00564       // First, we want to figure out all of the sets with which we definitely
00565       // don't alias. Iterate over all noalias set, and add those for which:
00566       //   1. The noalias argument is not in the set of objects from which we
00567       //      definitely derive.
00568       //   2. The noalias argument has not yet been captured.
00569       // An arbitrary function that might load pointers could see captured
00570       // noalias arguments via other noalias arguments or globals, and so we
00571       // must always check for prior capture.
00572       for (const Argument *A : NoAliasArgs) {
00573         if (!ObjSet.count(A) && (!CanDeriveViaCapture ||
00574                                  // It might be tempting to skip the
00575                                  // PointerMayBeCapturedBefore check if
00576                                  // A->hasNoCaptureAttr() is true, but this is
00577                                  // incorrect because nocapture only guarantees
00578                                  // that no copies outlive the function, not
00579                                  // that the value cannot be locally captured.
00580                                  !PointerMayBeCapturedBefore(A,
00581                                    /* ReturnCaptures */ false,
00582                                    /* StoreCaptures */ false, I, &DT)))
00583           NoAliases.push_back(NewScopes[A]);
00584       }
00585 
00586       if (!NoAliases.empty())
00587         NI->setMetadata(LLVMContext::MD_noalias,
00588                         MDNode::concatenate(
00589                             NI->getMetadata(LLVMContext::MD_noalias),
00590                             MDNode::get(CalledFunc->getContext(), NoAliases)));
00591 
00592       // Next, we want to figure out all of the sets to which we might belong.
00593       // We might belong to a set if the noalias argument is in the set of
00594       // underlying objects. If there is some non-noalias argument in our list
00595       // of underlying objects, then we cannot add a scope because the fact
00596       // that some access does not alias with any set of our noalias arguments
00597       // cannot itself guarantee that it does not alias with this access
00598       // (because there is some pointer of unknown origin involved and the
00599       // other access might also depend on this pointer). We also cannot add
00600       // scopes to arbitrary functions unless we know they don't access any
00601       // non-parameter pointer-values.
00602       bool CanAddScopes = !UsesAliasingPtr;
00603       if (CanAddScopes && IsFuncCall)
00604         CanAddScopes = IsArgMemOnlyCall;
00605 
00606       if (CanAddScopes)
00607         for (const Argument *A : NoAliasArgs) {
00608           if (ObjSet.count(A))
00609             Scopes.push_back(NewScopes[A]);
00610         }
00611 
00612       if (!Scopes.empty())
00613         NI->setMetadata(
00614             LLVMContext::MD_alias_scope,
00615             MDNode::concatenate(NI->getMetadata(LLVMContext::MD_alias_scope),
00616                                 MDNode::get(CalledFunc->getContext(), Scopes)));
00617     }
00618   }
00619 }
00620 
00621 /// If the inlined function has non-byval align arguments, then
00622 /// add @llvm.assume-based alignment assumptions to preserve this information.
00623 static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) {
00624   if (!PreserveAlignmentAssumptions || !IFI.DL)
00625     return;
00626 
00627   // To avoid inserting redundant assumptions, we should check for assumptions
00628   // already in the caller. To do this, we might need a DT of the caller.
00629   DominatorTree DT;
00630   bool DTCalculated = false;
00631 
00632   Function *CalledFunc = CS.getCalledFunction();
00633   for (Function::arg_iterator I = CalledFunc->arg_begin(),
00634                               E = CalledFunc->arg_end();
00635        I != E; ++I) {
00636     unsigned Align = I->getType()->isPointerTy() ? I->getParamAlignment() : 0;
00637     if (Align && !I->hasByValOrInAllocaAttr() && !I->hasNUses(0)) {
00638       if (!DTCalculated) {
00639         DT.recalculate(const_cast<Function&>(*CS.getInstruction()->getParent()
00640                                                ->getParent()));
00641         DTCalculated = true;
00642       }
00643 
00644       // If we can already prove the asserted alignment in the context of the
00645       // caller, then don't bother inserting the assumption.
00646       Value *Arg = CS.getArgument(I->getArgNo());
00647       if (getKnownAlignment(Arg, IFI.DL,
00648                             &IFI.ACT->getAssumptionCache(*CalledFunc),
00649                             CS.getInstruction(), &DT) >= Align)
00650         continue;
00651 
00652       IRBuilder<>(CS.getInstruction()).CreateAlignmentAssumption(*IFI.DL, Arg,
00653                                                                  Align);
00654     }
00655   }
00656 }
00657 
00658 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
00659 /// into the caller, update the specified callgraph to reflect the changes we
00660 /// made.  Note that it's possible that not all code was copied over, so only
00661 /// some edges of the callgraph may remain.
00662 static void UpdateCallGraphAfterInlining(CallSite CS,
00663                                          Function::iterator FirstNewBlock,
00664                                          ValueToValueMapTy &VMap,
00665                                          InlineFunctionInfo &IFI) {
00666   CallGraph &CG = *IFI.CG;
00667   const Function *Caller = CS.getInstruction()->getParent()->getParent();
00668   const Function *Callee = CS.getCalledFunction();
00669   CallGraphNode *CalleeNode = CG[Callee];
00670   CallGraphNode *CallerNode = CG[Caller];
00671 
00672   // Since we inlined some uninlined call sites in the callee into the caller,
00673   // add edges from the caller to all of the callees of the callee.
00674   CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
00675 
00676   // Consider the case where CalleeNode == CallerNode.
00677   CallGraphNode::CalledFunctionsVector CallCache;
00678   if (CalleeNode == CallerNode) {
00679     CallCache.assign(I, E);
00680     I = CallCache.begin();
00681     E = CallCache.end();
00682   }
00683 
00684   for (; I != E; ++I) {
00685     const Value *OrigCall = I->first;
00686 
00687     ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
00688     // Only copy the edge if the call was inlined!
00689     if (VMI == VMap.end() || VMI->second == nullptr)
00690       continue;
00691     
00692     // If the call was inlined, but then constant folded, there is no edge to
00693     // add.  Check for this case.
00694     Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
00695     if (!NewCall) continue;
00696 
00697     // Remember that this call site got inlined for the client of
00698     // InlineFunction.
00699     IFI.InlinedCalls.push_back(NewCall);
00700 
00701     // It's possible that inlining the callsite will cause it to go from an
00702     // indirect to a direct call by resolving a function pointer.  If this
00703     // happens, set the callee of the new call site to a more precise
00704     // destination.  This can also happen if the call graph node of the caller
00705     // was just unnecessarily imprecise.
00706     if (!I->second->getFunction())
00707       if (Function *F = CallSite(NewCall).getCalledFunction()) {
00708         // Indirect call site resolved to direct call.
00709         CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
00710 
00711         continue;
00712       }
00713 
00714     CallerNode->addCalledFunction(CallSite(NewCall), I->second);
00715   }
00716   
00717   // Update the call graph by deleting the edge from Callee to Caller.  We must
00718   // do this after the loop above in case Caller and Callee are the same.
00719   CallerNode->removeCallEdgeFor(CS);
00720 }
00721 
00722 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
00723                                     BasicBlock *InsertBlock,
00724                                     InlineFunctionInfo &IFI) {
00725   Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
00726   IRBuilder<> Builder(InsertBlock->begin());
00727 
00728   Value *Size;
00729   if (IFI.DL == nullptr)
00730     Size = ConstantExpr::getSizeOf(AggTy);
00731   else
00732     Size = Builder.getInt64(IFI.DL->getTypeStoreSize(AggTy));
00733 
00734   // Always generate a memcpy of alignment 1 here because we don't know
00735   // the alignment of the src pointer.  Other optimizations can infer
00736   // better alignment.
00737   Builder.CreateMemCpy(Dst, Src, Size, /*Align=*/1);
00738 }
00739 
00740 /// HandleByValArgument - When inlining a call site that has a byval argument,
00741 /// we have to make the implicit memcpy explicit by adding it.
00742 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
00743                                   const Function *CalledFunc,
00744                                   InlineFunctionInfo &IFI,
00745                                   unsigned ByValAlignment) {
00746   PointerType *ArgTy = cast<PointerType>(Arg->getType());
00747   Type *AggTy = ArgTy->getElementType();
00748 
00749   Function *Caller = TheCall->getParent()->getParent();
00750 
00751   // If the called function is readonly, then it could not mutate the caller's
00752   // copy of the byval'd memory.  In this case, it is safe to elide the copy and
00753   // temporary.
00754   if (CalledFunc->onlyReadsMemory()) {
00755     // If the byval argument has a specified alignment that is greater than the
00756     // passed in pointer, then we either have to round up the input pointer or
00757     // give up on this transformation.
00758     if (ByValAlignment <= 1)  // 0 = unspecified, 1 = no particular alignment.
00759       return Arg;
00760 
00761     // If the pointer is already known to be sufficiently aligned, or if we can
00762     // round it up to a larger alignment, then we don't need a temporary.
00763     if (getOrEnforceKnownAlignment(Arg, ByValAlignment, IFI.DL,
00764                                    &IFI.ACT->getAssumptionCache(*Caller),
00765                                    TheCall) >= ByValAlignment)
00766       return Arg;
00767     
00768     // Otherwise, we have to make a memcpy to get a safe alignment.  This is bad
00769     // for code quality, but rarely happens and is required for correctness.
00770   }
00771 
00772   // Create the alloca.  If we have DataLayout, use nice alignment.
00773   unsigned Align = 1;
00774   if (IFI.DL)
00775     Align = IFI.DL->getPrefTypeAlignment(AggTy);
00776   
00777   // If the byval had an alignment specified, we *must* use at least that
00778   // alignment, as it is required by the byval argument (and uses of the
00779   // pointer inside the callee).
00780   Align = std::max(Align, ByValAlignment);
00781   
00782   Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(), 
00783                                     &*Caller->begin()->begin());
00784   IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
00785   
00786   // Uses of the argument in the function should use our new alloca
00787   // instead.
00788   return NewAlloca;
00789 }
00790 
00791 // isUsedByLifetimeMarker - Check whether this Value is used by a lifetime
00792 // intrinsic.
00793 static bool isUsedByLifetimeMarker(Value *V) {
00794   for (User *U : V->users()) {
00795     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
00796       switch (II->getIntrinsicID()) {
00797       default: break;
00798       case Intrinsic::lifetime_start:
00799       case Intrinsic::lifetime_end:
00800         return true;
00801       }
00802     }
00803   }
00804   return false;
00805 }
00806 
00807 // hasLifetimeMarkers - Check whether the given alloca already has
00808 // lifetime.start or lifetime.end intrinsics.
00809 static bool hasLifetimeMarkers(AllocaInst *AI) {
00810   Type *Ty = AI->getType();
00811   Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
00812                                        Ty->getPointerAddressSpace());
00813   if (Ty == Int8PtrTy)
00814     return isUsedByLifetimeMarker(AI);
00815 
00816   // Do a scan to find all the casts to i8*.
00817   for (User *U : AI->users()) {
00818     if (U->getType() != Int8PtrTy) continue;
00819     if (U->stripPointerCasts() != AI) continue;
00820     if (isUsedByLifetimeMarker(U))
00821       return true;
00822   }
00823   return false;
00824 }
00825 
00826 /// Rebuild the entire inlined-at chain for this instruction so that the top of
00827 /// the chain now is inlined-at the new call site.
00828 static DebugLoc
00829 updateInlinedAtInfo(DebugLoc DL, MDLocation *InlinedAtNode,
00830                     LLVMContext &Ctx,
00831                     DenseMap<const MDLocation *, MDLocation *> &IANodes) {
00832   SmallVector<MDLocation*, 3> InlinedAtLocations;
00833   MDLocation *Last = InlinedAtNode;
00834   DebugLoc CurInlinedAt = DL;
00835 
00836   // Gather all the inlined-at nodes
00837   while (MDLocation *IA =
00838              cast_or_null<MDLocation>(CurInlinedAt.getInlinedAt(Ctx))) {
00839     // Skip any we've already built nodes for
00840     if (MDLocation *Found = IANodes[IA]) {
00841       Last = Found;
00842       break;
00843     }
00844 
00845     InlinedAtLocations.push_back(IA);
00846     CurInlinedAt = DebugLoc::getFromDILocation(IA);
00847   }
00848 
00849   // Starting from the top, rebuild the nodes to point to the new inlined-at
00850   // location (then rebuilding the rest of the chain behind it) and update the
00851   // map of already-constructed inlined-at nodes.
00852   for (auto I = InlinedAtLocations.rbegin(), E = InlinedAtLocations.rend();
00853        I != E; ++I) {
00854     const MDLocation *MD = *I;
00855     Last = IANodes[MD] = MDLocation::getDistinct(
00856         Ctx, MD->getLine(), MD->getColumn(), MD->getScope(), Last);
00857   }
00858 
00859   // And finally create the normal location for this instruction, referring to
00860   // the new inlined-at chain.
00861   return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), Last);
00862 }
00863 
00864 /// fixupLineNumbers - Update inlined instructions' line numbers to 
00865 /// to encode location where these instructions are inlined.
00866 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
00867                              Instruction *TheCall) {
00868   DebugLoc TheCallDL = TheCall->getDebugLoc();
00869   if (TheCallDL.isUnknown())
00870     return;
00871 
00872   auto &Ctx = Fn->getContext();
00873   auto *InlinedAtNode = cast<MDLocation>(TheCallDL.getAsMDNode(Ctx));
00874 
00875   // Create a unique call site, not to be confused with any other call from the
00876   // same location.
00877   InlinedAtNode = MDLocation::getDistinct(
00878       Ctx, InlinedAtNode->getLine(), InlinedAtNode->getColumn(),
00879       InlinedAtNode->getScope(), InlinedAtNode->getInlinedAt());
00880 
00881   // Cache the inlined-at nodes as they're built so they are reused, without
00882   // this every instruction's inlined-at chain would become distinct from each
00883   // other.
00884   DenseMap<const MDLocation *, MDLocation *> IANodes;
00885 
00886   for (; FI != Fn->end(); ++FI) {
00887     for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
00888          BI != BE; ++BI) {
00889       DebugLoc DL = BI->getDebugLoc();
00890       if (DL.isUnknown()) {
00891         // If the inlined instruction has no line number, make it look as if it
00892         // originates from the call location. This is important for
00893         // ((__always_inline__, __nodebug__)) functions which must use caller
00894         // location for all instructions in their function body.
00895 
00896         // Don't update static allocas, as they may get moved later.
00897         if (auto *AI = dyn_cast<AllocaInst>(BI))
00898           if (isa<Constant>(AI->getArraySize()))
00899             continue;
00900 
00901         BI->setDebugLoc(TheCallDL);
00902       } else {
00903         BI->setDebugLoc(updateInlinedAtInfo(DL, InlinedAtNode, BI->getContext(), IANodes));
00904         if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) {
00905           LLVMContext &Ctx = BI->getContext();
00906           MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
00907           DVI->setOperand(2, MetadataAsValue::get(
00908                                  Ctx, createInlinedVariable(DVI->getVariable(),
00909                                                             InlinedAt, Ctx)));
00910         } else if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(BI)) {
00911           LLVMContext &Ctx = BI->getContext();
00912           MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
00913           DDI->setOperand(1, MetadataAsValue::get(
00914                                  Ctx, createInlinedVariable(DDI->getVariable(),
00915                                                             InlinedAt, Ctx)));
00916         }
00917       }
00918     }
00919   }
00920 }
00921 
00922 /// InlineFunction - This function inlines the called function into the basic
00923 /// block of the caller.  This returns false if it is not possible to inline
00924 /// this call.  The program is still in a well defined state if this occurs
00925 /// though.
00926 ///
00927 /// Note that this only does one level of inlining.  For example, if the
00928 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
00929 /// exists in the instruction stream.  Similarly this will inline a recursive
00930 /// function by one level.
00931 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
00932                           bool InsertLifetime) {
00933   Instruction *TheCall = CS.getInstruction();
00934   assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
00935          "Instruction not in function!");
00936 
00937   // If IFI has any state in it, zap it before we fill it in.
00938   IFI.reset();
00939   
00940   const Function *CalledFunc = CS.getCalledFunction();
00941   if (!CalledFunc ||              // Can't inline external function or indirect
00942       CalledFunc->isDeclaration() || // call, or call to a vararg function!
00943       CalledFunc->getFunctionType()->isVarArg()) return false;
00944 
00945   // If the call to the callee cannot throw, set the 'nounwind' flag on any
00946   // calls that we inline.
00947   bool MarkNoUnwind = CS.doesNotThrow();
00948 
00949   BasicBlock *OrigBB = TheCall->getParent();
00950   Function *Caller = OrigBB->getParent();
00951 
00952   // GC poses two hazards to inlining, which only occur when the callee has GC:
00953   //  1. If the caller has no GC, then the callee's GC must be propagated to the
00954   //     caller.
00955   //  2. If the caller has a differing GC, it is invalid to inline.
00956   if (CalledFunc->hasGC()) {
00957     if (!Caller->hasGC())
00958       Caller->setGC(CalledFunc->getGC());
00959     else if (CalledFunc->getGC() != Caller->getGC())
00960       return false;
00961   }
00962 
00963   // Get the personality function from the callee if it contains a landing pad.
00964   Value *CalleePersonality = nullptr;
00965   for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end();
00966        I != E; ++I)
00967     if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
00968       const BasicBlock *BB = II->getUnwindDest();
00969       const LandingPadInst *LP = BB->getLandingPadInst();
00970       CalleePersonality = LP->getPersonalityFn();
00971       break;
00972     }
00973 
00974   // Find the personality function used by the landing pads of the caller. If it
00975   // exists, then check to see that it matches the personality function used in
00976   // the callee.
00977   if (CalleePersonality) {
00978     for (Function::const_iterator I = Caller->begin(), E = Caller->end();
00979          I != E; ++I)
00980       if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
00981         const BasicBlock *BB = II->getUnwindDest();
00982         const LandingPadInst *LP = BB->getLandingPadInst();
00983 
00984         // If the personality functions match, then we can perform the
00985         // inlining. Otherwise, we can't inline.
00986         // TODO: This isn't 100% true. Some personality functions are proper
00987         //       supersets of others and can be used in place of the other.
00988         if (LP->getPersonalityFn() != CalleePersonality)
00989           return false;
00990 
00991         break;
00992       }
00993   }
00994 
00995   // Get an iterator to the last basic block in the function, which will have
00996   // the new function inlined after it.
00997   Function::iterator LastBlock = &Caller->back();
00998 
00999   // Make sure to capture all of the return instructions from the cloned
01000   // function.
01001   SmallVector<ReturnInst*, 8> Returns;
01002   ClonedCodeInfo InlinedFunctionInfo;
01003   Function::iterator FirstNewBlock;
01004 
01005   { // Scope to destroy VMap after cloning.
01006     ValueToValueMapTy VMap;
01007     // Keep a list of pair (dst, src) to emit byval initializations.
01008     SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
01009 
01010     assert(CalledFunc->arg_size() == CS.arg_size() &&
01011            "No varargs calls can be inlined!");
01012 
01013     // Calculate the vector of arguments to pass into the function cloner, which
01014     // matches up the formal to the actual argument values.
01015     CallSite::arg_iterator AI = CS.arg_begin();
01016     unsigned ArgNo = 0;
01017     for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
01018          E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
01019       Value *ActualArg = *AI;
01020 
01021       // When byval arguments actually inlined, we need to make the copy implied
01022       // by them explicit.  However, we don't do this if the callee is readonly
01023       // or readnone, because the copy would be unneeded: the callee doesn't
01024       // modify the struct.
01025       if (CS.isByValArgument(ArgNo)) {
01026         ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
01027                                         CalledFunc->getParamAlignment(ArgNo+1));
01028         if (ActualArg != *AI)
01029           ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
01030       }
01031 
01032       VMap[I] = ActualArg;
01033     }
01034 
01035     // Add alignment assumptions if necessary. We do this before the inlined
01036     // instructions are actually cloned into the caller so that we can easily
01037     // check what will be known at the start of the inlined code.
01038     AddAlignmentAssumptions(CS, IFI);
01039 
01040     // We want the inliner to prune the code as it copies.  We would LOVE to
01041     // have no dead or constant instructions leftover after inlining occurs
01042     // (which can happen, e.g., because an argument was constant), but we'll be
01043     // happy with whatever the cloner can do.
01044     CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 
01045                               /*ModuleLevelChanges=*/false, Returns, ".i",
01046                               &InlinedFunctionInfo, IFI.DL, TheCall);
01047 
01048     // Remember the first block that is newly cloned over.
01049     FirstNewBlock = LastBlock; ++FirstNewBlock;
01050 
01051     // Inject byval arguments initialization.
01052     for (std::pair<Value*, Value*> &Init : ByValInit)
01053       HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
01054                               FirstNewBlock, IFI);
01055 
01056     // Update the callgraph if requested.
01057     if (IFI.CG)
01058       UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
01059 
01060     // Update inlined instructions' line number information.
01061     fixupLineNumbers(Caller, FirstNewBlock, TheCall);
01062 
01063     // Clone existing noalias metadata if necessary.
01064     CloneAliasScopeMetadata(CS, VMap);
01065 
01066     // Add noalias metadata if necessary.
01067     AddAliasScopeMetadata(CS, VMap, IFI.DL, IFI.AA);
01068 
01069     // FIXME: We could register any cloned assumptions instead of clearing the
01070     // whole function's cache.
01071     if (IFI.ACT)
01072       IFI.ACT->getAssumptionCache(*Caller).clear();
01073   }
01074 
01075   // If there are any alloca instructions in the block that used to be the entry
01076   // block for the callee, move them to the entry block of the caller.  First
01077   // calculate which instruction they should be inserted before.  We insert the
01078   // instructions at the end of the current alloca list.
01079   {
01080     BasicBlock::iterator InsertPoint = Caller->begin()->begin();
01081     for (BasicBlock::iterator I = FirstNewBlock->begin(),
01082          E = FirstNewBlock->end(); I != E; ) {
01083       AllocaInst *AI = dyn_cast<AllocaInst>(I++);
01084       if (!AI) continue;
01085       
01086       // If the alloca is now dead, remove it.  This often occurs due to code
01087       // specialization.
01088       if (AI->use_empty()) {
01089         AI->eraseFromParent();
01090         continue;
01091       }
01092 
01093       if (!isa<Constant>(AI->getArraySize()))
01094         continue;
01095       
01096       // Keep track of the static allocas that we inline into the caller.
01097       IFI.StaticAllocas.push_back(AI);
01098       
01099       // Scan for the block of allocas that we can move over, and move them
01100       // all at once.
01101       while (isa<AllocaInst>(I) &&
01102              isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
01103         IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
01104         ++I;
01105       }
01106 
01107       // Transfer all of the allocas over in a block.  Using splice means
01108       // that the instructions aren't removed from the symbol table, then
01109       // reinserted.
01110       Caller->getEntryBlock().getInstList().splice(InsertPoint,
01111                                                    FirstNewBlock->getInstList(),
01112                                                    AI, I);
01113     }
01114     // Move any dbg.declares describing the allocas into the entry basic block.
01115     for (auto &I : IFI.StaticAllocas)
01116       if (auto AI = dyn_cast<AllocaInst>(I))
01117         if (auto *DDI = FindAllocaDbgDeclare(AI))
01118           if (DDI->getParent() != Caller->begin())
01119             Caller->getEntryBlock().getInstList()
01120               .splice(AI->getNextNode(), FirstNewBlock->getInstList(),
01121                       DDI, DDI->getNextNode());
01122   }
01123 
01124   bool InlinedMustTailCalls = false;
01125   if (InlinedFunctionInfo.ContainsCalls) {
01126     CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
01127     if (CallInst *CI = dyn_cast<CallInst>(TheCall))
01128       CallSiteTailKind = CI->getTailCallKind();
01129 
01130     for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
01131          ++BB) {
01132       for (Instruction &I : *BB) {
01133         CallInst *CI = dyn_cast<CallInst>(&I);
01134         if (!CI)
01135           continue;
01136 
01137         // We need to reduce the strength of any inlined tail calls.  For
01138         // musttail, we have to avoid introducing potential unbounded stack
01139         // growth.  For example, if functions 'f' and 'g' are mutually recursive
01140         // with musttail, we can inline 'g' into 'f' so long as we preserve
01141         // musttail on the cloned call to 'f'.  If either the inlined call site
01142         // or the cloned call site is *not* musttail, the program already has
01143         // one frame of stack growth, so it's safe to remove musttail.  Here is
01144         // a table of example transformations:
01145         //
01146         //    f -> musttail g -> musttail f  ==>  f -> musttail f
01147         //    f -> musttail g ->     tail f  ==>  f ->     tail f
01148         //    f ->          g -> musttail f  ==>  f ->          f
01149         //    f ->          g ->     tail f  ==>  f ->          f
01150         CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
01151         ChildTCK = std::min(CallSiteTailKind, ChildTCK);
01152         CI->setTailCallKind(ChildTCK);
01153         InlinedMustTailCalls |= CI->isMustTailCall();
01154 
01155         // Calls inlined through a 'nounwind' call site should be marked
01156         // 'nounwind'.
01157         if (MarkNoUnwind)
01158           CI->setDoesNotThrow();
01159       }
01160     }
01161   }
01162 
01163   // Leave lifetime markers for the static alloca's, scoping them to the
01164   // function we just inlined.
01165   if (InsertLifetime && !IFI.StaticAllocas.empty()) {
01166     IRBuilder<> builder(FirstNewBlock->begin());
01167     for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
01168       AllocaInst *AI = IFI.StaticAllocas[ai];
01169 
01170       // If the alloca is already scoped to something smaller than the whole
01171       // function then there's no need to add redundant, less accurate markers.
01172       if (hasLifetimeMarkers(AI))
01173         continue;
01174 
01175       // Try to determine the size of the allocation.
01176       ConstantInt *AllocaSize = nullptr;
01177       if (ConstantInt *AIArraySize =
01178           dyn_cast<ConstantInt>(AI->getArraySize())) {
01179         if (IFI.DL) {
01180           Type *AllocaType = AI->getAllocatedType();
01181           uint64_t AllocaTypeSize = IFI.DL->getTypeAllocSize(AllocaType);
01182           uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
01183           assert(AllocaArraySize > 0 && "array size of AllocaInst is zero");
01184           // Check that array size doesn't saturate uint64_t and doesn't
01185           // overflow when it's multiplied by type size.
01186           if (AllocaArraySize != ~0ULL &&
01187               UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
01188             AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
01189                                           AllocaArraySize * AllocaTypeSize);
01190           }
01191         }
01192       }
01193 
01194       builder.CreateLifetimeStart(AI, AllocaSize);
01195       for (ReturnInst *RI : Returns) {
01196         // Don't insert llvm.lifetime.end calls between a musttail call and a
01197         // return.  The return kills all local allocas.
01198         if (InlinedMustTailCalls &&
01199             RI->getParent()->getTerminatingMustTailCall())
01200           continue;
01201         IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
01202       }
01203     }
01204   }
01205 
01206   // If the inlined code contained dynamic alloca instructions, wrap the inlined
01207   // code with llvm.stacksave/llvm.stackrestore intrinsics.
01208   if (InlinedFunctionInfo.ContainsDynamicAllocas) {
01209     Module *M = Caller->getParent();
01210     // Get the two intrinsics we care about.
01211     Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
01212     Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
01213 
01214     // Insert the llvm.stacksave.
01215     CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
01216       .CreateCall(StackSave, "savedstack");
01217 
01218     // Insert a call to llvm.stackrestore before any return instructions in the
01219     // inlined function.
01220     for (ReturnInst *RI : Returns) {
01221       // Don't insert llvm.stackrestore calls between a musttail call and a
01222       // return.  The return will restore the stack pointer.
01223       if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
01224         continue;
01225       IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
01226     }
01227   }
01228 
01229   // If we are inlining for an invoke instruction, we must make sure to rewrite
01230   // any call instructions into invoke instructions.
01231   if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
01232     HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
01233 
01234   // Handle any inlined musttail call sites.  In order for a new call site to be
01235   // musttail, the source of the clone and the inlined call site must have been
01236   // musttail.  Therefore it's safe to return without merging control into the
01237   // phi below.
01238   if (InlinedMustTailCalls) {
01239     // Check if we need to bitcast the result of any musttail calls.
01240     Type *NewRetTy = Caller->getReturnType();
01241     bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
01242 
01243     // Handle the returns preceded by musttail calls separately.
01244     SmallVector<ReturnInst *, 8> NormalReturns;
01245     for (ReturnInst *RI : Returns) {
01246       CallInst *ReturnedMustTail =
01247           RI->getParent()->getTerminatingMustTailCall();
01248       if (!ReturnedMustTail) {
01249         NormalReturns.push_back(RI);
01250         continue;
01251       }
01252       if (!NeedBitCast)
01253         continue;
01254 
01255       // Delete the old return and any preceding bitcast.
01256       BasicBlock *CurBB = RI->getParent();
01257       auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
01258       RI->eraseFromParent();
01259       if (OldCast)
01260         OldCast->eraseFromParent();
01261 
01262       // Insert a new bitcast and return with the right type.
01263       IRBuilder<> Builder(CurBB);
01264       Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
01265     }
01266 
01267     // Leave behind the normal returns so we can merge control flow.
01268     std::swap(Returns, NormalReturns);
01269   }
01270 
01271   // If we cloned in _exactly one_ basic block, and if that block ends in a
01272   // return instruction, we splice the body of the inlined callee directly into
01273   // the calling basic block.
01274   if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
01275     // Move all of the instructions right before the call.
01276     OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
01277                                  FirstNewBlock->begin(), FirstNewBlock->end());
01278     // Remove the cloned basic block.
01279     Caller->getBasicBlockList().pop_back();
01280 
01281     // If the call site was an invoke instruction, add a branch to the normal
01282     // destination.
01283     if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
01284       BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
01285       NewBr->setDebugLoc(Returns[0]->getDebugLoc());
01286     }
01287 
01288     // If the return instruction returned a value, replace uses of the call with
01289     // uses of the returned value.
01290     if (!TheCall->use_empty()) {
01291       ReturnInst *R = Returns[0];
01292       if (TheCall == R->getReturnValue())
01293         TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
01294       else
01295         TheCall->replaceAllUsesWith(R->getReturnValue());
01296     }
01297     // Since we are now done with the Call/Invoke, we can delete it.
01298     TheCall->eraseFromParent();
01299 
01300     // Since we are now done with the return instruction, delete it also.
01301     Returns[0]->eraseFromParent();
01302 
01303     // We are now done with the inlining.
01304     return true;
01305   }
01306 
01307   // Otherwise, we have the normal case, of more than one block to inline or
01308   // multiple return sites.
01309 
01310   // We want to clone the entire callee function into the hole between the
01311   // "starter" and "ender" blocks.  How we accomplish this depends on whether
01312   // this is an invoke instruction or a call instruction.
01313   BasicBlock *AfterCallBB;
01314   BranchInst *CreatedBranchToNormalDest = nullptr;
01315   if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
01316 
01317     // Add an unconditional branch to make this look like the CallInst case...
01318     CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
01319 
01320     // Split the basic block.  This guarantees that no PHI nodes will have to be
01321     // updated due to new incoming edges, and make the invoke case more
01322     // symmetric to the call case.
01323     AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest,
01324                                           CalledFunc->getName()+".exit");
01325 
01326   } else {  // It's a call
01327     // If this is a call instruction, we need to split the basic block that
01328     // the call lives in.
01329     //
01330     AfterCallBB = OrigBB->splitBasicBlock(TheCall,
01331                                           CalledFunc->getName()+".exit");
01332   }
01333 
01334   // Change the branch that used to go to AfterCallBB to branch to the first
01335   // basic block of the inlined function.
01336   //
01337   TerminatorInst *Br = OrigBB->getTerminator();
01338   assert(Br && Br->getOpcode() == Instruction::Br &&
01339          "splitBasicBlock broken!");
01340   Br->setOperand(0, FirstNewBlock);
01341 
01342 
01343   // Now that the function is correct, make it a little bit nicer.  In
01344   // particular, move the basic blocks inserted from the end of the function
01345   // into the space made by splitting the source basic block.
01346   Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
01347                                      FirstNewBlock, Caller->end());
01348 
01349   // Handle all of the return instructions that we just cloned in, and eliminate
01350   // any users of the original call/invoke instruction.
01351   Type *RTy = CalledFunc->getReturnType();
01352 
01353   PHINode *PHI = nullptr;
01354   if (Returns.size() > 1) {
01355     // The PHI node should go at the front of the new basic block to merge all
01356     // possible incoming values.
01357     if (!TheCall->use_empty()) {
01358       PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
01359                             AfterCallBB->begin());
01360       // Anything that used the result of the function call should now use the
01361       // PHI node as their operand.
01362       TheCall->replaceAllUsesWith(PHI);
01363     }
01364 
01365     // Loop over all of the return instructions adding entries to the PHI node
01366     // as appropriate.
01367     if (PHI) {
01368       for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
01369         ReturnInst *RI = Returns[i];
01370         assert(RI->getReturnValue()->getType() == PHI->getType() &&
01371                "Ret value not consistent in function!");
01372         PHI->addIncoming(RI->getReturnValue(), RI->getParent());
01373       }
01374     }
01375 
01376 
01377     // Add a branch to the merge points and remove return instructions.
01378     DebugLoc Loc;
01379     for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
01380       ReturnInst *RI = Returns[i];
01381       BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
01382       Loc = RI->getDebugLoc();
01383       BI->setDebugLoc(Loc);
01384       RI->eraseFromParent();
01385     }
01386     // We need to set the debug location to *somewhere* inside the
01387     // inlined function. The line number may be nonsensical, but the
01388     // instruction will at least be associated with the right
01389     // function.
01390     if (CreatedBranchToNormalDest)
01391       CreatedBranchToNormalDest->setDebugLoc(Loc);
01392   } else if (!Returns.empty()) {
01393     // Otherwise, if there is exactly one return value, just replace anything
01394     // using the return value of the call with the computed value.
01395     if (!TheCall->use_empty()) {
01396       if (TheCall == Returns[0]->getReturnValue())
01397         TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
01398       else
01399         TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
01400     }
01401 
01402     // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
01403     BasicBlock *ReturnBB = Returns[0]->getParent();
01404     ReturnBB->replaceAllUsesWith(AfterCallBB);
01405 
01406     // Splice the code from the return block into the block that it will return
01407     // to, which contains the code that was after the call.
01408     AfterCallBB->getInstList().splice(AfterCallBB->begin(),
01409                                       ReturnBB->getInstList());
01410 
01411     if (CreatedBranchToNormalDest)
01412       CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
01413 
01414     // Delete the return instruction now and empty ReturnBB now.
01415     Returns[0]->eraseFromParent();
01416     ReturnBB->eraseFromParent();
01417   } else if (!TheCall->use_empty()) {
01418     // No returns, but something is using the return value of the call.  Just
01419     // nuke the result.
01420     TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
01421   }
01422 
01423   // Since we are now done with the Call/Invoke, we can delete it.
01424   TheCall->eraseFromParent();
01425 
01426   // If we inlined any musttail calls and the original return is now
01427   // unreachable, delete it.  It can only contain a bitcast and ret.
01428   if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB))
01429     AfterCallBB->eraseFromParent();
01430 
01431   // We should always be able to fold the entry block of the function into the
01432   // single predecessor of the block...
01433   assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
01434   BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
01435 
01436   // Splice the code entry block into calling block, right before the
01437   // unconditional branch.
01438   CalleeEntry->replaceAllUsesWith(OrigBB);  // Update PHI nodes
01439   OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
01440 
01441   // Remove the unconditional branch.
01442   OrigBB->getInstList().erase(Br);
01443 
01444   // Now we can remove the CalleeEntry block, which is now empty.
01445   Caller->getBasicBlockList().erase(CalleeEntry);
01446 
01447   // If we inserted a phi node, check to see if it has a single value (e.g. all
01448   // the entries are the same or undef).  If so, remove the PHI so it doesn't
01449   // block other optimizations.
01450   if (PHI) {
01451     if (Value *V = SimplifyInstruction(PHI, IFI.DL, nullptr, nullptr,
01452                                        &IFI.ACT->getAssumptionCache(*Caller))) {
01453       PHI->replaceAllUsesWith(V);
01454       PHI->eraseFromParent();
01455     }
01456   }
01457 
01458   return true;
01459 }