LLVM  mainline
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/DIBuilder.h"
00034 #include "llvm/IR/Dominators.h"
00035 #include "llvm/IR/IRBuilder.h"
00036 #include "llvm/IR/Instructions.h"
00037 #include "llvm/IR/IntrinsicInst.h"
00038 #include "llvm/IR/Intrinsics.h"
00039 #include "llvm/IR/MDBuilder.h"
00040 #include "llvm/IR/Module.h"
00041 #include "llvm/Transforms/Utils/Local.h"
00042 #include "llvm/Support/CommandLine.h"
00043 #include <algorithm>
00044 using namespace llvm;
00045 
00046 static cl::opt<bool>
00047 EnableNoAliasConversion("enable-noalias-to-md-conversion", cl::init(true),
00048   cl::Hidden,
00049   cl::desc("Convert noalias attributes to metadata during inlining."));
00050 
00051 static cl::opt<bool>
00052 PreserveAlignmentAssumptions("preserve-alignment-assumptions-during-inlining",
00053   cl::init(true), cl::Hidden,
00054   cl::desc("Convert align attributes to assumptions during inlining."));
00055 
00056 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI,
00057                           bool InsertLifetime) {
00058   return InlineFunction(CallSite(CI), IFI, InsertLifetime);
00059 }
00060 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI,
00061                           bool InsertLifetime) {
00062   return InlineFunction(CallSite(II), IFI, InsertLifetime);
00063 }
00064 
00065 namespace {
00066   /// A class for recording information about inlining through an invoke.
00067   class InvokeInliningInfo {
00068     BasicBlock *OuterResumeDest; ///< Destination of the invoke's unwind.
00069     BasicBlock *InnerResumeDest; ///< Destination for the callee's resume.
00070     LandingPadInst *CallerLPad;  ///< LandingPadInst associated with the invoke.
00071     PHINode *InnerEHValuesPHI;   ///< PHI for EH values from landingpad insts.
00072     SmallVector<Value*, 8> UnwindDestPHIValues;
00073 
00074   public:
00075     InvokeInliningInfo(InvokeInst *II)
00076       : OuterResumeDest(II->getUnwindDest()), InnerResumeDest(nullptr),
00077         CallerLPad(nullptr), InnerEHValuesPHI(nullptr) {
00078       // If there are PHI nodes in the unwind destination block, we need to keep
00079       // track of which values came into them from the invoke before removing
00080       // the edge from this block.
00081       llvm::BasicBlock *InvokeBB = II->getParent();
00082       BasicBlock::iterator I = OuterResumeDest->begin();
00083       for (; isa<PHINode>(I); ++I) {
00084         // Save the value to use for this edge.
00085         PHINode *PHI = cast<PHINode>(I);
00086         UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
00087       }
00088 
00089       CallerLPad = cast<LandingPadInst>(I);
00090     }
00091 
00092     /// The outer unwind destination is the target of
00093     /// unwind edges introduced for calls within the inlined function.
00094     BasicBlock *getOuterResumeDest() const {
00095       return OuterResumeDest;
00096     }
00097 
00098     BasicBlock *getInnerResumeDest();
00099 
00100     LandingPadInst *getLandingPadInst() const { return CallerLPad; }
00101 
00102     /// Forward the 'resume' instruction to the caller's landing pad block.
00103     /// When the landing pad block has only one predecessor, this is
00104     /// a simple branch. When there is more than one predecessor, we need to
00105     /// split the landing pad block after the landingpad instruction and jump
00106     /// to there.
00107     void forwardResume(ResumeInst *RI,
00108                        SmallPtrSetImpl<LandingPadInst*> &InlinedLPads);
00109 
00110     /// Add incoming-PHI values to the unwind destination block for the given
00111     /// basic block, using the values for the 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 /// 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 /// Forward the 'resume' instruction to the caller's landing pad block.
00162 /// 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 /// When we inline a basic block into an invoke,
00181 /// we have to turn all of the calls that can throw into invokes.
00182 /// 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 /// 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 /// When inlining a function that contains noalias scope metadata,
00282 /// 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 /// If the inlined function has noalias arguments,
00394 /// then 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)
00625     return;
00626   auto &DL = CS.getCaller()->getParent()->getDataLayout();
00627 
00628   // To avoid inserting redundant assumptions, we should check for assumptions
00629   // already in the caller. To do this, we might need a DT of the caller.
00630   DominatorTree DT;
00631   bool DTCalculated = false;
00632 
00633   Function *CalledFunc = CS.getCalledFunction();
00634   for (Function::arg_iterator I = CalledFunc->arg_begin(),
00635                               E = CalledFunc->arg_end();
00636        I != E; ++I) {
00637     unsigned Align = I->getType()->isPointerTy() ? I->getParamAlignment() : 0;
00638     if (Align && !I->hasByValOrInAllocaAttr() && !I->hasNUses(0)) {
00639       if (!DTCalculated) {
00640         DT.recalculate(const_cast<Function&>(*CS.getInstruction()->getParent()
00641                                                ->getParent()));
00642         DTCalculated = true;
00643       }
00644 
00645       // If we can already prove the asserted alignment in the context of the
00646       // caller, then don't bother inserting the assumption.
00647       Value *Arg = CS.getArgument(I->getArgNo());
00648       if (getKnownAlignment(Arg, DL, CS.getInstruction(),
00649                             &IFI.ACT->getAssumptionCache(*CalledFunc),
00650                             &DT) >= Align)
00651         continue;
00652 
00653       IRBuilder<>(CS.getInstruction())
00654           .CreateAlignmentAssumption(DL, Arg, Align);
00655     }
00656   }
00657 }
00658 
00659 /// Once we have cloned code over from a callee into the caller,
00660 /// update the specified callgraph to reflect the changes we made.
00661 /// Note that it's possible that not all code was copied over, so only
00662 /// some edges of the callgraph may remain.
00663 static void UpdateCallGraphAfterInlining(CallSite CS,
00664                                          Function::iterator FirstNewBlock,
00665                                          ValueToValueMapTy &VMap,
00666                                          InlineFunctionInfo &IFI) {
00667   CallGraph &CG = *IFI.CG;
00668   const Function *Caller = CS.getInstruction()->getParent()->getParent();
00669   const Function *Callee = CS.getCalledFunction();
00670   CallGraphNode *CalleeNode = CG[Callee];
00671   CallGraphNode *CallerNode = CG[Caller];
00672 
00673   // Since we inlined some uninlined call sites in the callee into the caller,
00674   // add edges from the caller to all of the callees of the callee.
00675   CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
00676 
00677   // Consider the case where CalleeNode == CallerNode.
00678   CallGraphNode::CalledFunctionsVector CallCache;
00679   if (CalleeNode == CallerNode) {
00680     CallCache.assign(I, E);
00681     I = CallCache.begin();
00682     E = CallCache.end();
00683   }
00684 
00685   for (; I != E; ++I) {
00686     const Value *OrigCall = I->first;
00687 
00688     ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
00689     // Only copy the edge if the call was inlined!
00690     if (VMI == VMap.end() || VMI->second == nullptr)
00691       continue;
00692     
00693     // If the call was inlined, but then constant folded, there is no edge to
00694     // add.  Check for this case.
00695     Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
00696     if (!NewCall)
00697       continue;
00698 
00699     // We do not treat intrinsic calls like real function calls because we
00700     // expect them to become inline code; do not add an edge for an intrinsic.
00701     CallSite CS = CallSite(NewCall);
00702     if (CS && CS.getCalledFunction() && CS.getCalledFunction()->isIntrinsic())
00703       continue;
00704     
00705     // Remember that this call site got inlined for the client of
00706     // InlineFunction.
00707     IFI.InlinedCalls.push_back(NewCall);
00708 
00709     // It's possible that inlining the callsite will cause it to go from an
00710     // indirect to a direct call by resolving a function pointer.  If this
00711     // happens, set the callee of the new call site to a more precise
00712     // destination.  This can also happen if the call graph node of the caller
00713     // was just unnecessarily imprecise.
00714     if (!I->second->getFunction())
00715       if (Function *F = CallSite(NewCall).getCalledFunction()) {
00716         // Indirect call site resolved to direct call.
00717         CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
00718 
00719         continue;
00720       }
00721 
00722     CallerNode->addCalledFunction(CallSite(NewCall), I->second);
00723   }
00724   
00725   // Update the call graph by deleting the edge from Callee to Caller.  We must
00726   // do this after the loop above in case Caller and Callee are the same.
00727   CallerNode->removeCallEdgeFor(CS);
00728 }
00729 
00730 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
00731                                     BasicBlock *InsertBlock,
00732                                     InlineFunctionInfo &IFI) {
00733   Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
00734   IRBuilder<> Builder(InsertBlock->begin());
00735 
00736   Value *Size = Builder.getInt64(M->getDataLayout().getTypeStoreSize(AggTy));
00737 
00738   // Always generate a memcpy of alignment 1 here because we don't know
00739   // the alignment of the src pointer.  Other optimizations can infer
00740   // better alignment.
00741   Builder.CreateMemCpy(Dst, Src, Size, /*Align=*/1);
00742 }
00743 
00744 /// When inlining a call site that has a byval argument,
00745 /// we have to make the implicit memcpy explicit by adding it.
00746 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
00747                                   const Function *CalledFunc,
00748                                   InlineFunctionInfo &IFI,
00749                                   unsigned ByValAlignment) {
00750   PointerType *ArgTy = cast<PointerType>(Arg->getType());
00751   Type *AggTy = ArgTy->getElementType();
00752 
00753   Function *Caller = TheCall->getParent()->getParent();
00754 
00755   // If the called function is readonly, then it could not mutate the caller's
00756   // copy of the byval'd memory.  In this case, it is safe to elide the copy and
00757   // temporary.
00758   if (CalledFunc->onlyReadsMemory()) {
00759     // If the byval argument has a specified alignment that is greater than the
00760     // passed in pointer, then we either have to round up the input pointer or
00761     // give up on this transformation.
00762     if (ByValAlignment <= 1)  // 0 = unspecified, 1 = no particular alignment.
00763       return Arg;
00764 
00765     const DataLayout &DL = Caller->getParent()->getDataLayout();
00766 
00767     // If the pointer is already known to be sufficiently aligned, or if we can
00768     // round it up to a larger alignment, then we don't need a temporary.
00769     if (getOrEnforceKnownAlignment(Arg, ByValAlignment, DL, TheCall,
00770                                    &IFI.ACT->getAssumptionCache(*Caller)) >=
00771         ByValAlignment)
00772       return Arg;
00773     
00774     // Otherwise, we have to make a memcpy to get a safe alignment.  This is bad
00775     // for code quality, but rarely happens and is required for correctness.
00776   }
00777 
00778   // Create the alloca.  If we have DataLayout, use nice alignment.
00779   unsigned Align =
00780       Caller->getParent()->getDataLayout().getPrefTypeAlignment(AggTy);
00781 
00782   // If the byval had an alignment specified, we *must* use at least that
00783   // alignment, as it is required by the byval argument (and uses of the
00784   // pointer inside the callee).
00785   Align = std::max(Align, ByValAlignment);
00786   
00787   Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(), 
00788                                     &*Caller->begin()->begin());
00789   IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
00790   
00791   // Uses of the argument in the function should use our new alloca
00792   // instead.
00793   return NewAlloca;
00794 }
00795 
00796 // Check whether this Value is used by a lifetime intrinsic.
00797 static bool isUsedByLifetimeMarker(Value *V) {
00798   for (User *U : V->users()) {
00799     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
00800       switch (II->getIntrinsicID()) {
00801       default: break;
00802       case Intrinsic::lifetime_start:
00803       case Intrinsic::lifetime_end:
00804         return true;
00805       }
00806     }
00807   }
00808   return false;
00809 }
00810 
00811 // Check whether the given alloca already has
00812 // lifetime.start or lifetime.end intrinsics.
00813 static bool hasLifetimeMarkers(AllocaInst *AI) {
00814   Type *Ty = AI->getType();
00815   Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
00816                                        Ty->getPointerAddressSpace());
00817   if (Ty == Int8PtrTy)
00818     return isUsedByLifetimeMarker(AI);
00819 
00820   // Do a scan to find all the casts to i8*.
00821   for (User *U : AI->users()) {
00822     if (U->getType() != Int8PtrTy) continue;
00823     if (U->stripPointerCasts() != AI) continue;
00824     if (isUsedByLifetimeMarker(U))
00825       return true;
00826   }
00827   return false;
00828 }
00829 
00830 /// Rebuild the entire inlined-at chain for this instruction so that the top of
00831 /// the chain now is inlined-at the new call site.
00832 static DebugLoc
00833 updateInlinedAtInfo(DebugLoc DL, DILocation *InlinedAtNode, LLVMContext &Ctx,
00834                     DenseMap<const DILocation *, DILocation *> &IANodes) {
00835   SmallVector<DILocation *, 3> InlinedAtLocations;
00836   DILocation *Last = InlinedAtNode;
00837   DILocation *CurInlinedAt = DL;
00838 
00839   // Gather all the inlined-at nodes
00840   while (DILocation *IA = CurInlinedAt->getInlinedAt()) {
00841     // Skip any we've already built nodes for
00842     if (DILocation *Found = IANodes[IA]) {
00843       Last = Found;
00844       break;
00845     }
00846 
00847     InlinedAtLocations.push_back(IA);
00848     CurInlinedAt = IA;
00849   }
00850 
00851   // Starting from the top, rebuild the nodes to point to the new inlined-at
00852   // location (then rebuilding the rest of the chain behind it) and update the
00853   // map of already-constructed inlined-at nodes.
00854   for (auto I = InlinedAtLocations.rbegin(), E = InlinedAtLocations.rend();
00855        I != E; ++I) {
00856     const DILocation *MD = *I;
00857     Last = IANodes[MD] = DILocation::getDistinct(
00858         Ctx, MD->getLine(), MD->getColumn(), MD->getScope(), Last);
00859   }
00860 
00861   // And finally create the normal location for this instruction, referring to
00862   // the new inlined-at chain.
00863   return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(), Last);
00864 }
00865 
00866 /// Update inlined instructions' line numbers to
00867 /// to encode location where these instructions are inlined.
00868 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
00869                              Instruction *TheCall) {
00870   DebugLoc TheCallDL = TheCall->getDebugLoc();
00871   if (!TheCallDL)
00872     return;
00873 
00874   auto &Ctx = Fn->getContext();
00875   DILocation *InlinedAtNode = TheCallDL;
00876 
00877   // Create a unique call site, not to be confused with any other call from the
00878   // same location.
00879   InlinedAtNode = DILocation::getDistinct(
00880       Ctx, InlinedAtNode->getLine(), InlinedAtNode->getColumn(),
00881       InlinedAtNode->getScope(), InlinedAtNode->getInlinedAt());
00882 
00883   // Cache the inlined-at nodes as they're built so they are reused, without
00884   // this every instruction's inlined-at chain would become distinct from each
00885   // other.
00886   DenseMap<const DILocation *, DILocation *> IANodes;
00887 
00888   for (; FI != Fn->end(); ++FI) {
00889     for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
00890          BI != BE; ++BI) {
00891       DebugLoc DL = BI->getDebugLoc();
00892       if (!DL) {
00893         // If the inlined instruction has no line number, make it look as if it
00894         // originates from the call location. This is important for
00895         // ((__always_inline__, __nodebug__)) functions which must use caller
00896         // location for all instructions in their function body.
00897 
00898         // Don't update static allocas, as they may get moved later.
00899         if (auto *AI = dyn_cast<AllocaInst>(BI))
00900           if (isa<Constant>(AI->getArraySize()))
00901             continue;
00902 
00903         BI->setDebugLoc(TheCallDL);
00904       } else {
00905         BI->setDebugLoc(updateInlinedAtInfo(DL, InlinedAtNode, BI->getContext(), IANodes));
00906       }
00907     }
00908   }
00909 }
00910 
00911 /// This function inlines the called function into the basic block of the
00912 /// caller. This returns false if it is not possible to inline this call.
00913 /// The program is still in a well defined state if this occurs though.
00914 ///
00915 /// Note that this only does one level of inlining.  For example, if the
00916 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
00917 /// exists in the instruction stream.  Similarly this will inline a recursive
00918 /// function by one level.
00919 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
00920                           bool InsertLifetime) {
00921   Instruction *TheCall = CS.getInstruction();
00922   assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
00923          "Instruction not in function!");
00924 
00925   // If IFI has any state in it, zap it before we fill it in.
00926   IFI.reset();
00927   
00928   const Function *CalledFunc = CS.getCalledFunction();
00929   if (!CalledFunc ||              // Can't inline external function or indirect
00930       CalledFunc->isDeclaration() || // call, or call to a vararg function!
00931       CalledFunc->getFunctionType()->isVarArg()) return false;
00932 
00933   // If the call to the callee cannot throw, set the 'nounwind' flag on any
00934   // calls that we inline.
00935   bool MarkNoUnwind = CS.doesNotThrow();
00936 
00937   BasicBlock *OrigBB = TheCall->getParent();
00938   Function *Caller = OrigBB->getParent();
00939 
00940   // GC poses two hazards to inlining, which only occur when the callee has GC:
00941   //  1. If the caller has no GC, then the callee's GC must be propagated to the
00942   //     caller.
00943   //  2. If the caller has a differing GC, it is invalid to inline.
00944   if (CalledFunc->hasGC()) {
00945     if (!Caller->hasGC())
00946       Caller->setGC(CalledFunc->getGC());
00947     else if (CalledFunc->getGC() != Caller->getGC())
00948       return false;
00949   }
00950 
00951   // Get the personality function from the callee if it contains a landing pad.
00952   Constant *CalledPersonality =
00953       CalledFunc->hasPersonalityFn() ? CalledFunc->getPersonalityFn() : nullptr;
00954 
00955   // Find the personality function used by the landing pads of the caller. If it
00956   // exists, then check to see that it matches the personality function used in
00957   // the callee.
00958   Constant *CallerPersonality =
00959       Caller->hasPersonalityFn() ? Caller->getPersonalityFn() : nullptr;
00960   if (CalledPersonality) {
00961     if (!CallerPersonality)
00962       Caller->setPersonalityFn(CalledPersonality);
00963     // If the personality functions match, then we can perform the
00964     // inlining. Otherwise, we can't inline.
00965     // TODO: This isn't 100% true. Some personality functions are proper
00966     //       supersets of others and can be used in place of the other.
00967     else if (CalledPersonality != CallerPersonality)
00968       return false;
00969   }
00970 
00971   // Get an iterator to the last basic block in the function, which will have
00972   // the new function inlined after it.
00973   Function::iterator LastBlock = &Caller->back();
00974 
00975   // Make sure to capture all of the return instructions from the cloned
00976   // function.
00977   SmallVector<ReturnInst*, 8> Returns;
00978   ClonedCodeInfo InlinedFunctionInfo;
00979   Function::iterator FirstNewBlock;
00980 
00981   { // Scope to destroy VMap after cloning.
00982     ValueToValueMapTy VMap;
00983     // Keep a list of pair (dst, src) to emit byval initializations.
00984     SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
00985 
00986     auto &DL = Caller->getParent()->getDataLayout();
00987 
00988     assert(CalledFunc->arg_size() == CS.arg_size() &&
00989            "No varargs calls can be inlined!");
00990 
00991     // Calculate the vector of arguments to pass into the function cloner, which
00992     // matches up the formal to the actual argument values.
00993     CallSite::arg_iterator AI = CS.arg_begin();
00994     unsigned ArgNo = 0;
00995     for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
00996          E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
00997       Value *ActualArg = *AI;
00998 
00999       // When byval arguments actually inlined, we need to make the copy implied
01000       // by them explicit.  However, we don't do this if the callee is readonly
01001       // or readnone, because the copy would be unneeded: the callee doesn't
01002       // modify the struct.
01003       if (CS.isByValArgument(ArgNo)) {
01004         ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
01005                                         CalledFunc->getParamAlignment(ArgNo+1));
01006         if (ActualArg != *AI)
01007           ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
01008       }
01009 
01010       VMap[I] = ActualArg;
01011     }
01012 
01013     // Add alignment assumptions if necessary. We do this before the inlined
01014     // instructions are actually cloned into the caller so that we can easily
01015     // check what will be known at the start of the inlined code.
01016     AddAlignmentAssumptions(CS, IFI);
01017 
01018     // We want the inliner to prune the code as it copies.  We would LOVE to
01019     // have no dead or constant instructions leftover after inlining occurs
01020     // (which can happen, e.g., because an argument was constant), but we'll be
01021     // happy with whatever the cloner can do.
01022     CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
01023                               /*ModuleLevelChanges=*/false, Returns, ".i",
01024                               &InlinedFunctionInfo, TheCall);
01025 
01026     // Remember the first block that is newly cloned over.
01027     FirstNewBlock = LastBlock; ++FirstNewBlock;
01028 
01029     // Inject byval arguments initialization.
01030     for (std::pair<Value*, Value*> &Init : ByValInit)
01031       HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
01032                               FirstNewBlock, IFI);
01033 
01034     // Update the callgraph if requested.
01035     if (IFI.CG)
01036       UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
01037 
01038     // Update inlined instructions' line number information.
01039     fixupLineNumbers(Caller, FirstNewBlock, TheCall);
01040 
01041     // Clone existing noalias metadata if necessary.
01042     CloneAliasScopeMetadata(CS, VMap);
01043 
01044     // Add noalias metadata if necessary.
01045     AddAliasScopeMetadata(CS, VMap, DL, IFI.AA);
01046 
01047     // FIXME: We could register any cloned assumptions instead of clearing the
01048     // whole function's cache.
01049     if (IFI.ACT)
01050       IFI.ACT->getAssumptionCache(*Caller).clear();
01051   }
01052 
01053   // If there are any alloca instructions in the block that used to be the entry
01054   // block for the callee, move them to the entry block of the caller.  First
01055   // calculate which instruction they should be inserted before.  We insert the
01056   // instructions at the end of the current alloca list.
01057   {
01058     BasicBlock::iterator InsertPoint = Caller->begin()->begin();
01059     for (BasicBlock::iterator I = FirstNewBlock->begin(),
01060          E = FirstNewBlock->end(); I != E; ) {
01061       AllocaInst *AI = dyn_cast<AllocaInst>(I++);
01062       if (!AI) continue;
01063       
01064       // If the alloca is now dead, remove it.  This often occurs due to code
01065       // specialization.
01066       if (AI->use_empty()) {
01067         AI->eraseFromParent();
01068         continue;
01069       }
01070 
01071       if (!isa<Constant>(AI->getArraySize()))
01072         continue;
01073       
01074       // Keep track of the static allocas that we inline into the caller.
01075       IFI.StaticAllocas.push_back(AI);
01076       
01077       // Scan for the block of allocas that we can move over, and move them
01078       // all at once.
01079       while (isa<AllocaInst>(I) &&
01080              isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
01081         IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
01082         ++I;
01083       }
01084 
01085       // Transfer all of the allocas over in a block.  Using splice means
01086       // that the instructions aren't removed from the symbol table, then
01087       // reinserted.
01088       Caller->getEntryBlock().getInstList().splice(InsertPoint,
01089                                                    FirstNewBlock->getInstList(),
01090                                                    AI, I);
01091     }
01092     // Move any dbg.declares describing the allocas into the entry basic block.
01093     DIBuilder DIB(*Caller->getParent());
01094     for (auto &AI : IFI.StaticAllocas)
01095       replaceDbgDeclareForAlloca(AI, AI, DIB, /*Deref=*/false);
01096   }
01097 
01098   bool InlinedMustTailCalls = false;
01099   if (InlinedFunctionInfo.ContainsCalls) {
01100     CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
01101     if (CallInst *CI = dyn_cast<CallInst>(TheCall))
01102       CallSiteTailKind = CI->getTailCallKind();
01103 
01104     for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
01105          ++BB) {
01106       for (Instruction &I : *BB) {
01107         CallInst *CI = dyn_cast<CallInst>(&I);
01108         if (!CI)
01109           continue;
01110 
01111         // We need to reduce the strength of any inlined tail calls.  For
01112         // musttail, we have to avoid introducing potential unbounded stack
01113         // growth.  For example, if functions 'f' and 'g' are mutually recursive
01114         // with musttail, we can inline 'g' into 'f' so long as we preserve
01115         // musttail on the cloned call to 'f'.  If either the inlined call site
01116         // or the cloned call site is *not* musttail, the program already has
01117         // one frame of stack growth, so it's safe to remove musttail.  Here is
01118         // a table of example transformations:
01119         //
01120         //    f -> musttail g -> musttail f  ==>  f -> musttail f
01121         //    f -> musttail g ->     tail f  ==>  f ->     tail f
01122         //    f ->          g -> musttail f  ==>  f ->          f
01123         //    f ->          g ->     tail f  ==>  f ->          f
01124         CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
01125         ChildTCK = std::min(CallSiteTailKind, ChildTCK);
01126         CI->setTailCallKind(ChildTCK);
01127         InlinedMustTailCalls |= CI->isMustTailCall();
01128 
01129         // Calls inlined through a 'nounwind' call site should be marked
01130         // 'nounwind'.
01131         if (MarkNoUnwind)
01132           CI->setDoesNotThrow();
01133       }
01134     }
01135   }
01136 
01137   // Leave lifetime markers for the static alloca's, scoping them to the
01138   // function we just inlined.
01139   if (InsertLifetime && !IFI.StaticAllocas.empty()) {
01140     IRBuilder<> builder(FirstNewBlock->begin());
01141     for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
01142       AllocaInst *AI = IFI.StaticAllocas[ai];
01143 
01144       // If the alloca is already scoped to something smaller than the whole
01145       // function then there's no need to add redundant, less accurate markers.
01146       if (hasLifetimeMarkers(AI))
01147         continue;
01148 
01149       // Try to determine the size of the allocation.
01150       ConstantInt *AllocaSize = nullptr;
01151       if (ConstantInt *AIArraySize =
01152           dyn_cast<ConstantInt>(AI->getArraySize())) {
01153         auto &DL = Caller->getParent()->getDataLayout();
01154         Type *AllocaType = AI->getAllocatedType();
01155         uint64_t AllocaTypeSize = DL.getTypeAllocSize(AllocaType);
01156         uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
01157 
01158         // Don't add markers for zero-sized allocas.
01159         if (AllocaArraySize == 0)
01160           continue;
01161 
01162         // Check that array size doesn't saturate uint64_t and doesn't
01163         // overflow when it's multiplied by type size.
01164         if (AllocaArraySize != ~0ULL &&
01165             UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
01166           AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
01167                                         AllocaArraySize * AllocaTypeSize);
01168         }
01169       }
01170 
01171       builder.CreateLifetimeStart(AI, AllocaSize);
01172       for (ReturnInst *RI : Returns) {
01173         // Don't insert llvm.lifetime.end calls between a musttail call and a
01174         // return.  The return kills all local allocas.
01175         if (InlinedMustTailCalls &&
01176             RI->getParent()->getTerminatingMustTailCall())
01177           continue;
01178         IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
01179       }
01180     }
01181   }
01182 
01183   // If the inlined code contained dynamic alloca instructions, wrap the inlined
01184   // code with llvm.stacksave/llvm.stackrestore intrinsics.
01185   if (InlinedFunctionInfo.ContainsDynamicAllocas) {
01186     Module *M = Caller->getParent();
01187     // Get the two intrinsics we care about.
01188     Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
01189     Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
01190 
01191     // Insert the llvm.stacksave.
01192     CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
01193                              .CreateCall(StackSave, {}, "savedstack");
01194 
01195     // Insert a call to llvm.stackrestore before any return instructions in the
01196     // inlined function.
01197     for (ReturnInst *RI : Returns) {
01198       // Don't insert llvm.stackrestore calls between a musttail call and a
01199       // return.  The return will restore the stack pointer.
01200       if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
01201         continue;
01202       IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
01203     }
01204   }
01205 
01206   // If we are inlining for an invoke instruction, we must make sure to rewrite
01207   // any call instructions into invoke instructions.
01208   if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
01209     HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
01210 
01211   // Handle any inlined musttail call sites.  In order for a new call site to be
01212   // musttail, the source of the clone and the inlined call site must have been
01213   // musttail.  Therefore it's safe to return without merging control into the
01214   // phi below.
01215   if (InlinedMustTailCalls) {
01216     // Check if we need to bitcast the result of any musttail calls.
01217     Type *NewRetTy = Caller->getReturnType();
01218     bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
01219 
01220     // Handle the returns preceded by musttail calls separately.
01221     SmallVector<ReturnInst *, 8> NormalReturns;
01222     for (ReturnInst *RI : Returns) {
01223       CallInst *ReturnedMustTail =
01224           RI->getParent()->getTerminatingMustTailCall();
01225       if (!ReturnedMustTail) {
01226         NormalReturns.push_back(RI);
01227         continue;
01228       }
01229       if (!NeedBitCast)
01230         continue;
01231 
01232       // Delete the old return and any preceding bitcast.
01233       BasicBlock *CurBB = RI->getParent();
01234       auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
01235       RI->eraseFromParent();
01236       if (OldCast)
01237         OldCast->eraseFromParent();
01238 
01239       // Insert a new bitcast and return with the right type.
01240       IRBuilder<> Builder(CurBB);
01241       Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
01242     }
01243 
01244     // Leave behind the normal returns so we can merge control flow.
01245     std::swap(Returns, NormalReturns);
01246   }
01247 
01248   // If we cloned in _exactly one_ basic block, and if that block ends in a
01249   // return instruction, we splice the body of the inlined callee directly into
01250   // the calling basic block.
01251   if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
01252     // Move all of the instructions right before the call.
01253     OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
01254                                  FirstNewBlock->begin(), FirstNewBlock->end());
01255     // Remove the cloned basic block.
01256     Caller->getBasicBlockList().pop_back();
01257 
01258     // If the call site was an invoke instruction, add a branch to the normal
01259     // destination.
01260     if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
01261       BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
01262       NewBr->setDebugLoc(Returns[0]->getDebugLoc());
01263     }
01264 
01265     // If the return instruction returned a value, replace uses of the call with
01266     // uses of the returned value.
01267     if (!TheCall->use_empty()) {
01268       ReturnInst *R = Returns[0];
01269       if (TheCall == R->getReturnValue())
01270         TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
01271       else
01272         TheCall->replaceAllUsesWith(R->getReturnValue());
01273     }
01274     // Since we are now done with the Call/Invoke, we can delete it.
01275     TheCall->eraseFromParent();
01276 
01277     // Since we are now done with the return instruction, delete it also.
01278     Returns[0]->eraseFromParent();
01279 
01280     // We are now done with the inlining.
01281     return true;
01282   }
01283 
01284   // Otherwise, we have the normal case, of more than one block to inline or
01285   // multiple return sites.
01286 
01287   // We want to clone the entire callee function into the hole between the
01288   // "starter" and "ender" blocks.  How we accomplish this depends on whether
01289   // this is an invoke instruction or a call instruction.
01290   BasicBlock *AfterCallBB;
01291   BranchInst *CreatedBranchToNormalDest = nullptr;
01292   if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
01293 
01294     // Add an unconditional branch to make this look like the CallInst case...
01295     CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
01296 
01297     // Split the basic block.  This guarantees that no PHI nodes will have to be
01298     // updated due to new incoming edges, and make the invoke case more
01299     // symmetric to the call case.
01300     AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest,
01301                                           CalledFunc->getName()+".exit");
01302 
01303   } else {  // It's a call
01304     // If this is a call instruction, we need to split the basic block that
01305     // the call lives in.
01306     //
01307     AfterCallBB = OrigBB->splitBasicBlock(TheCall,
01308                                           CalledFunc->getName()+".exit");
01309   }
01310 
01311   // Change the branch that used to go to AfterCallBB to branch to the first
01312   // basic block of the inlined function.
01313   //
01314   TerminatorInst *Br = OrigBB->getTerminator();
01315   assert(Br && Br->getOpcode() == Instruction::Br &&
01316          "splitBasicBlock broken!");
01317   Br->setOperand(0, FirstNewBlock);
01318 
01319 
01320   // Now that the function is correct, make it a little bit nicer.  In
01321   // particular, move the basic blocks inserted from the end of the function
01322   // into the space made by splitting the source basic block.
01323   Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
01324                                      FirstNewBlock, Caller->end());
01325 
01326   // Handle all of the return instructions that we just cloned in, and eliminate
01327   // any users of the original call/invoke instruction.
01328   Type *RTy = CalledFunc->getReturnType();
01329 
01330   PHINode *PHI = nullptr;
01331   if (Returns.size() > 1) {
01332     // The PHI node should go at the front of the new basic block to merge all
01333     // possible incoming values.
01334     if (!TheCall->use_empty()) {
01335       PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
01336                             AfterCallBB->begin());
01337       // Anything that used the result of the function call should now use the
01338       // PHI node as their operand.
01339       TheCall->replaceAllUsesWith(PHI);
01340     }
01341 
01342     // Loop over all of the return instructions adding entries to the PHI node
01343     // as appropriate.
01344     if (PHI) {
01345       for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
01346         ReturnInst *RI = Returns[i];
01347         assert(RI->getReturnValue()->getType() == PHI->getType() &&
01348                "Ret value not consistent in function!");
01349         PHI->addIncoming(RI->getReturnValue(), RI->getParent());
01350       }
01351     }
01352 
01353 
01354     // Add a branch to the merge points and remove return instructions.
01355     DebugLoc Loc;
01356     for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
01357       ReturnInst *RI = Returns[i];
01358       BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
01359       Loc = RI->getDebugLoc();
01360       BI->setDebugLoc(Loc);
01361       RI->eraseFromParent();
01362     }
01363     // We need to set the debug location to *somewhere* inside the
01364     // inlined function. The line number may be nonsensical, but the
01365     // instruction will at least be associated with the right
01366     // function.
01367     if (CreatedBranchToNormalDest)
01368       CreatedBranchToNormalDest->setDebugLoc(Loc);
01369   } else if (!Returns.empty()) {
01370     // Otherwise, if there is exactly one return value, just replace anything
01371     // using the return value of the call with the computed value.
01372     if (!TheCall->use_empty()) {
01373       if (TheCall == Returns[0]->getReturnValue())
01374         TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
01375       else
01376         TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
01377     }
01378 
01379     // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
01380     BasicBlock *ReturnBB = Returns[0]->getParent();
01381     ReturnBB->replaceAllUsesWith(AfterCallBB);
01382 
01383     // Splice the code from the return block into the block that it will return
01384     // to, which contains the code that was after the call.
01385     AfterCallBB->getInstList().splice(AfterCallBB->begin(),
01386                                       ReturnBB->getInstList());
01387 
01388     if (CreatedBranchToNormalDest)
01389       CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
01390 
01391     // Delete the return instruction now and empty ReturnBB now.
01392     Returns[0]->eraseFromParent();
01393     ReturnBB->eraseFromParent();
01394   } else if (!TheCall->use_empty()) {
01395     // No returns, but something is using the return value of the call.  Just
01396     // nuke the result.
01397     TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
01398   }
01399 
01400   // Since we are now done with the Call/Invoke, we can delete it.
01401   TheCall->eraseFromParent();
01402 
01403   // If we inlined any musttail calls and the original return is now
01404   // unreachable, delete it.  It can only contain a bitcast and ret.
01405   if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB))
01406     AfterCallBB->eraseFromParent();
01407 
01408   // We should always be able to fold the entry block of the function into the
01409   // single predecessor of the block...
01410   assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
01411   BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
01412 
01413   // Splice the code entry block into calling block, right before the
01414   // unconditional branch.
01415   CalleeEntry->replaceAllUsesWith(OrigBB);  // Update PHI nodes
01416   OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
01417 
01418   // Remove the unconditional branch.
01419   OrigBB->getInstList().erase(Br);
01420 
01421   // Now we can remove the CalleeEntry block, which is now empty.
01422   Caller->getBasicBlockList().erase(CalleeEntry);
01423 
01424   // If we inserted a phi node, check to see if it has a single value (e.g. all
01425   // the entries are the same or undef).  If so, remove the PHI so it doesn't
01426   // block other optimizations.
01427   if (PHI) {
01428     auto &DL = Caller->getParent()->getDataLayout();
01429     if (Value *V = SimplifyInstruction(PHI, DL, nullptr, nullptr,
01430                                        &IFI.ACT->getAssumptionCache(*Caller))) {
01431       PHI->replaceAllUsesWith(V);
01432       PHI->eraseFromParent();
01433     }
01434   }
01435 
01436   return true;
01437 }