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/AssumptionTracker.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 Value *, 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<MDNode *, 16> DummyNodes;
00323   DenseMap<const MDNode *, TrackingVH<MDNode> > MDMap;
00324   for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
00325        I != IE; ++I) {
00326     MDNode *Dummy = MDNode::getTemporary(CalledFunc->getContext(), None);
00327     DummyNodes.push_back(Dummy);
00328     MDMap[*I] = Dummy;
00329   }
00330 
00331   // Create new metadata nodes to replace the dummy nodes, replacing old
00332   // metadata references with either a dummy node or an already-created new
00333   // node.
00334   for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
00335        I != IE; ++I) {
00336     SmallVector<Value *, 4> NewOps;
00337     for (unsigned i = 0, ie = (*I)->getNumOperands(); i != ie; ++i) {
00338       const Value *V = (*I)->getOperand(i);
00339       if (const MDNode *M = dyn_cast<MDNode>(V))
00340         NewOps.push_back(MDMap[M]);
00341       else
00342         NewOps.push_back(const_cast<Value *>(V));
00343     }
00344 
00345     MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps),
00346            *TempM = MDMap[*I];
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 =
00388           CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
00389         NI->setMetadata(LLVMContext::MD_noalias, M);
00390     }
00391   }
00392 
00393   // Now that everything has been replaced, delete the dummy nodes.
00394   for (unsigned i = 0, ie = DummyNodes.size(); i != ie; ++i)
00395     MDNode::deleteTemporary(DummyNodes[i]);
00396 }
00397 
00398 /// AddAliasScopeMetadata - If the inlined function has noalias arguments, then
00399 /// add new alias scopes for each noalias argument, tag the mapped noalias
00400 /// parameters with noalias metadata specifying the new scope, and tag all
00401 /// non-derived loads, stores and memory intrinsics with the new alias scopes.
00402 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
00403                                   const DataLayout *DL, AliasAnalysis *AA) {
00404   if (!EnableNoAliasConversion)
00405     return;
00406 
00407   const Function *CalledFunc = CS.getCalledFunction();
00408   SmallVector<const Argument *, 4> NoAliasArgs;
00409 
00410   for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
00411        E = CalledFunc->arg_end(); I != E; ++I) {
00412     if (I->hasNoAliasAttr() && !I->hasNUses(0))
00413       NoAliasArgs.push_back(I);
00414   }
00415 
00416   if (NoAliasArgs.empty())
00417     return;
00418 
00419   // To do a good job, if a noalias variable is captured, we need to know if
00420   // the capture point dominates the particular use we're considering.
00421   DominatorTree DT;
00422   DT.recalculate(const_cast<Function&>(*CalledFunc));
00423 
00424   // noalias indicates that pointer values based on the argument do not alias
00425   // pointer values which are not based on it. So we add a new "scope" for each
00426   // noalias function argument. Accesses using pointers based on that argument
00427   // become part of that alias scope, accesses using pointers not based on that
00428   // argument are tagged as noalias with that scope.
00429 
00430   DenseMap<const Argument *, MDNode *> NewScopes;
00431   MDBuilder MDB(CalledFunc->getContext());
00432 
00433   // Create a new scope domain for this function.
00434   MDNode *NewDomain =
00435     MDB.createAnonymousAliasScopeDomain(CalledFunc->getName());
00436   for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) {
00437     const Argument *A = NoAliasArgs[i];
00438 
00439     std::string Name = CalledFunc->getName();
00440     if (A->hasName()) {
00441       Name += ": %";
00442       Name += A->getName();
00443     } else {
00444       Name += ": argument ";
00445       Name += utostr(i);
00446     }
00447 
00448     // Note: We always create a new anonymous root here. This is true regardless
00449     // of the linkage of the callee because the aliasing "scope" is not just a
00450     // property of the callee, but also all control dependencies in the caller.
00451     MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
00452     NewScopes.insert(std::make_pair(A, NewScope));
00453   }
00454 
00455   // Iterate over all new instructions in the map; for all memory-access
00456   // instructions, add the alias scope metadata.
00457   for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
00458        VMI != VMIE; ++VMI) {
00459     if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) {
00460       if (!VMI->second)
00461         continue;
00462 
00463       Instruction *NI = dyn_cast<Instruction>(VMI->second);
00464       if (!NI)
00465         continue;
00466 
00467       bool IsArgMemOnlyCall = false, IsFuncCall = false;
00468       SmallVector<const Value *, 2> PtrArgs;
00469 
00470       if (const LoadInst *LI = dyn_cast<LoadInst>(I))
00471         PtrArgs.push_back(LI->getPointerOperand());
00472       else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
00473         PtrArgs.push_back(SI->getPointerOperand());
00474       else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
00475         PtrArgs.push_back(VAAI->getPointerOperand());
00476       else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
00477         PtrArgs.push_back(CXI->getPointerOperand());
00478       else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
00479         PtrArgs.push_back(RMWI->getPointerOperand());
00480       else if (ImmutableCallSite ICS = ImmutableCallSite(I)) {
00481         // If we know that the call does not access memory, then we'll still
00482         // know that about the inlined clone of this call site, and we don't
00483         // need to add metadata.
00484         if (ICS.doesNotAccessMemory())
00485           continue;
00486 
00487         IsFuncCall = true;
00488         if (AA) {
00489           AliasAnalysis::ModRefBehavior MRB = AA->getModRefBehavior(ICS);
00490           if (MRB == AliasAnalysis::OnlyAccessesArgumentPointees ||
00491               MRB == AliasAnalysis::OnlyReadsArgumentPointees)
00492             IsArgMemOnlyCall = true;
00493         }
00494 
00495         for (ImmutableCallSite::arg_iterator AI = ICS.arg_begin(),
00496              AE = ICS.arg_end(); AI != AE; ++AI) {
00497           // We need to check the underlying objects of all arguments, not just
00498           // the pointer arguments, because we might be passing pointers as
00499           // integers, etc.
00500           // However, if we know that the call only accesses pointer arguments,
00501           // then we only need to check the pointer arguments.
00502           if (IsArgMemOnlyCall && !(*AI)->getType()->isPointerTy())
00503             continue;
00504 
00505           PtrArgs.push_back(*AI);
00506         }
00507       }
00508 
00509       // If we found no pointers, then this instruction is not suitable for
00510       // pairing with an instruction to receive aliasing metadata.
00511       // However, if this is a call, this we might just alias with none of the
00512       // noalias arguments.
00513       if (PtrArgs.empty() && !IsFuncCall)
00514         continue;
00515 
00516       // It is possible that there is only one underlying object, but you
00517       // need to go through several PHIs to see it, and thus could be
00518       // repeated in the Objects list.
00519       SmallPtrSet<const Value *, 4> ObjSet;
00520       SmallVector<Value *, 4> Scopes, NoAliases;
00521 
00522       SmallSetVector<const Argument *, 4> NAPtrArgs;
00523       for (unsigned i = 0, ie = PtrArgs.size(); i != ie; ++i) {
00524         SmallVector<Value *, 4> Objects;
00525         GetUnderlyingObjects(const_cast<Value*>(PtrArgs[i]),
00526                              Objects, DL, /* MaxLookup = */ 0);
00527 
00528         for (Value *O : Objects)
00529           ObjSet.insert(O);
00530       }
00531 
00532       // Figure out if we're derived from anything that is not a noalias
00533       // argument.
00534       bool CanDeriveViaCapture = false, UsesAliasingPtr = false;
00535       for (const Value *V : ObjSet) {
00536         // Is this value a constant that cannot be derived from any pointer
00537         // value (we need to exclude constant expressions, for example, that
00538         // are formed from arithmetic on global symbols).
00539         bool IsNonPtrConst = isa<ConstantInt>(V) || isa<ConstantFP>(V) ||
00540                              isa<ConstantPointerNull>(V) ||
00541                              isa<ConstantDataVector>(V) || isa<UndefValue>(V);
00542         if (IsNonPtrConst)
00543           continue;
00544 
00545         // If this is anything other than a noalias argument, then we cannot
00546         // completely describe the aliasing properties using alias.scope
00547         // metadata (and, thus, won't add any).
00548         if (const Argument *A = dyn_cast<Argument>(V)) {
00549           if (!A->hasNoAliasAttr())
00550             UsesAliasingPtr = true;
00551         } else {
00552           UsesAliasingPtr = true;
00553         }
00554 
00555         // If this is not some identified function-local object (which cannot
00556         // directly alias a noalias argument), or some other argument (which,
00557         // by definition, also cannot alias a noalias argument), then we could
00558         // alias a noalias argument that has been captured).
00559         if (!isa<Argument>(V) &&
00560             !isIdentifiedFunctionLocal(const_cast<Value*>(V)))
00561           CanDeriveViaCapture = true;
00562       }
00563 
00564       // A function call can always get captured noalias pointers (via other
00565       // parameters, globals, etc.).
00566       if (IsFuncCall && !IsArgMemOnlyCall)
00567         CanDeriveViaCapture = true;
00568 
00569       // First, we want to figure out all of the sets with which we definitely
00570       // don't alias. Iterate over all noalias set, and add those for which:
00571       //   1. The noalias argument is not in the set of objects from which we
00572       //      definitely derive.
00573       //   2. The noalias argument has not yet been captured.
00574       // An arbitrary function that might load pointers could see captured
00575       // noalias arguments via other noalias arguments or globals, and so we
00576       // must always check for prior capture.
00577       for (const Argument *A : NoAliasArgs) {
00578         if (!ObjSet.count(A) && (!CanDeriveViaCapture ||
00579                                  // It might be tempting to skip the
00580                                  // PointerMayBeCapturedBefore check if
00581                                  // A->hasNoCaptureAttr() is true, but this is
00582                                  // incorrect because nocapture only guarantees
00583                                  // that no copies outlive the function, not
00584                                  // that the value cannot be locally captured.
00585                                  !PointerMayBeCapturedBefore(A,
00586                                    /* ReturnCaptures */ false,
00587                                    /* StoreCaptures */ false, I, &DT)))
00588           NoAliases.push_back(NewScopes[A]);
00589       }
00590 
00591       if (!NoAliases.empty())
00592         NI->setMetadata(LLVMContext::MD_noalias, MDNode::concatenate(
00593           NI->getMetadata(LLVMContext::MD_noalias),
00594             MDNode::get(CalledFunc->getContext(), NoAliases)));
00595 
00596       // Next, we want to figure out all of the sets to which we might belong.
00597       // We might belong to a set if the noalias argument is in the set of
00598       // underlying objects. If there is some non-noalias argument in our list
00599       // of underlying objects, then we cannot add a scope because the fact
00600       // that some access does not alias with any set of our noalias arguments
00601       // cannot itself guarantee that it does not alias with this access
00602       // (because there is some pointer of unknown origin involved and the
00603       // other access might also depend on this pointer). We also cannot add
00604       // scopes to arbitrary functions unless we know they don't access any
00605       // non-parameter pointer-values.
00606       bool CanAddScopes = !UsesAliasingPtr;
00607       if (CanAddScopes && IsFuncCall)
00608         CanAddScopes = IsArgMemOnlyCall;
00609 
00610       if (CanAddScopes)
00611         for (const Argument *A : NoAliasArgs) {
00612           if (ObjSet.count(A))
00613             Scopes.push_back(NewScopes[A]);
00614         }
00615 
00616       if (!Scopes.empty())
00617         NI->setMetadata(LLVMContext::MD_alias_scope, MDNode::concatenate(
00618           NI->getMetadata(LLVMContext::MD_alias_scope),
00619             MDNode::get(CalledFunc->getContext(), Scopes)));
00620     }
00621   }
00622 }
00623 
00624 /// If the inlined function has non-byval align arguments, then
00625 /// add @llvm.assume-based alignment assumptions to preserve this information.
00626 static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) {
00627   if (!PreserveAlignmentAssumptions || !IFI.DL)
00628     return;
00629 
00630   // To avoid inserting redundant assumptions, we should check for assumptions
00631   // already in the caller. To do this, we might need a DT of the caller.
00632   DominatorTree DT;
00633   bool DTCalculated = false;
00634 
00635   const Function *CalledFunc = CS.getCalledFunction();
00636   for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
00637        E = CalledFunc->arg_end(); I != E; ++I) {
00638     unsigned Align = I->getType()->isPointerTy() ? I->getParamAlignment() : 0;
00639     if (Align && !I->hasByValOrInAllocaAttr() && !I->hasNUses(0)) {
00640       if (!DTCalculated) {
00641         DT.recalculate(const_cast<Function&>(*CS.getInstruction()->getParent()
00642                                                ->getParent()));
00643         DTCalculated = true;
00644       }
00645 
00646       // If we can already prove the asserted alignment in the context of the
00647       // caller, then don't bother inserting the assumption.
00648       Value *Arg = CS.getArgument(I->getArgNo());
00649       if (getKnownAlignment(Arg, IFI.DL, IFI.AT, CS.getInstruction(),
00650                             &DT) >= Align)
00651         continue;
00652 
00653       IRBuilder<>(CS.getInstruction()).CreateAlignmentAssumption(*IFI.DL, Arg,
00654                                                                  Align);
00655     }
00656   }
00657 }
00658 
00659 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
00660 /// into the caller, update the specified callgraph to reflect the changes we
00661 /// made.  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) continue;
00697 
00698     // Remember that this call site got inlined for the client of
00699     // InlineFunction.
00700     IFI.InlinedCalls.push_back(NewCall);
00701 
00702     // It's possible that inlining the callsite will cause it to go from an
00703     // indirect to a direct call by resolving a function pointer.  If this
00704     // happens, set the callee of the new call site to a more precise
00705     // destination.  This can also happen if the call graph node of the caller
00706     // was just unnecessarily imprecise.
00707     if (!I->second->getFunction())
00708       if (Function *F = CallSite(NewCall).getCalledFunction()) {
00709         // Indirect call site resolved to direct call.
00710         CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
00711 
00712         continue;
00713       }
00714 
00715     CallerNode->addCalledFunction(CallSite(NewCall), I->second);
00716   }
00717   
00718   // Update the call graph by deleting the edge from Callee to Caller.  We must
00719   // do this after the loop above in case Caller and Callee are the same.
00720   CallerNode->removeCallEdgeFor(CS);
00721 }
00722 
00723 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
00724                                     BasicBlock *InsertBlock,
00725                                     InlineFunctionInfo &IFI) {
00726   Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
00727   IRBuilder<> Builder(InsertBlock->begin());
00728 
00729   Value *Size;
00730   if (IFI.DL == nullptr)
00731     Size = ConstantExpr::getSizeOf(AggTy);
00732   else
00733     Size = Builder.getInt64(IFI.DL->getTypeStoreSize(AggTy));
00734 
00735   // Always generate a memcpy of alignment 1 here because we don't know
00736   // the alignment of the src pointer.  Other optimizations can infer
00737   // better alignment.
00738   Builder.CreateMemCpy(Dst, Src, Size, /*Align=*/1);
00739 }
00740 
00741 /// HandleByValArgument - When inlining a call site that has a byval argument,
00742 /// we have to make the implicit memcpy explicit by adding it.
00743 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
00744                                   const Function *CalledFunc,
00745                                   InlineFunctionInfo &IFI,
00746                                   unsigned ByValAlignment,
00747                                   bool &AddedNewAllocas) {
00748   PointerType *ArgTy = cast<PointerType>(Arg->getType());
00749   Type *AggTy = ArgTy->getElementType();
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,
00764                                    IFI.DL, IFI.AT, TheCall) >= ByValAlignment)
00765       return Arg;
00766     
00767     // Otherwise, we have to make a memcpy to get a safe alignment.  This is bad
00768     // for code quality, but rarely happens and is required for correctness.
00769   }
00770 
00771   // Create the alloca.  If we have DataLayout, use nice alignment.
00772   unsigned Align = 1;
00773   if (IFI.DL)
00774     Align = IFI.DL->getPrefTypeAlignment(AggTy);
00775   
00776   // If the byval had an alignment specified, we *must* use at least that
00777   // alignment, as it is required by the byval argument (and uses of the
00778   // pointer inside the callee).
00779   Align = std::max(Align, ByValAlignment);
00780   
00781   Function *Caller = TheCall->getParent()->getParent(); 
00782   
00783   Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(), 
00784                                     &*Caller->begin()->begin());
00785   IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
00786   
00787   // Uses of the argument in the function should use our new alloca
00788   // instead.
00789   AddedNewAllocas = true;
00790   return NewAlloca;
00791 }
00792 
00793 // isUsedByLifetimeMarker - Check whether this Value is used by a lifetime
00794 // intrinsic.
00795 static bool isUsedByLifetimeMarker(Value *V) {
00796   for (User *U : V->users()) {
00797     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
00798       switch (II->getIntrinsicID()) {
00799       default: break;
00800       case Intrinsic::lifetime_start:
00801       case Intrinsic::lifetime_end:
00802         return true;
00803       }
00804     }
00805   }
00806   return false;
00807 }
00808 
00809 // hasLifetimeMarkers - Check whether the given alloca already has
00810 // lifetime.start or lifetime.end intrinsics.
00811 static bool hasLifetimeMarkers(AllocaInst *AI) {
00812   Type *Ty = AI->getType();
00813   Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
00814                                        Ty->getPointerAddressSpace());
00815   if (Ty == Int8PtrTy)
00816     return isUsedByLifetimeMarker(AI);
00817 
00818   // Do a scan to find all the casts to i8*.
00819   for (User *U : AI->users()) {
00820     if (U->getType() != Int8PtrTy) continue;
00821     if (U->stripPointerCasts() != AI) continue;
00822     if (isUsedByLifetimeMarker(U))
00823       return true;
00824   }
00825   return false;
00826 }
00827 
00828 /// updateInlinedAtInfo - Helper function used by fixupLineNumbers to
00829 /// recursively update InlinedAtEntry of a DebugLoc.
00830 static DebugLoc updateInlinedAtInfo(const DebugLoc &DL, 
00831                                     const DebugLoc &InlinedAtDL,
00832                                     LLVMContext &Ctx) {
00833   if (MDNode *IA = DL.getInlinedAt(Ctx)) {
00834     DebugLoc NewInlinedAtDL 
00835       = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx);
00836     return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
00837                          NewInlinedAtDL.getAsMDNode(Ctx));
00838   }
00839 
00840   return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
00841                        InlinedAtDL.getAsMDNode(Ctx));
00842 }
00843 
00844 /// fixupLineNumbers - Update inlined instructions' line numbers to 
00845 /// to encode location where these instructions are inlined.
00846 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
00847                              Instruction *TheCall) {
00848   DebugLoc TheCallDL = TheCall->getDebugLoc();
00849   if (TheCallDL.isUnknown())
00850     return;
00851 
00852   for (; FI != Fn->end(); ++FI) {
00853     for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
00854          BI != BE; ++BI) {
00855       DebugLoc DL = BI->getDebugLoc();
00856       if (DL.isUnknown()) {
00857         // If the inlined instruction has no line number, make it look as if it
00858         // originates from the call location. This is important for
00859         // ((__always_inline__, __nodebug__)) functions which must use caller
00860         // location for all instructions in their function body.
00861 
00862         // Don't update static allocas, as they may get moved later.
00863         if (auto *AI = dyn_cast<AllocaInst>(BI))
00864           if (isa<Constant>(AI->getArraySize()))
00865             continue;
00866 
00867         BI->setDebugLoc(TheCallDL);
00868       } else {
00869         BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext()));
00870         if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) {
00871           LLVMContext &Ctx = BI->getContext();
00872           MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
00873           DVI->setOperand(2, createInlinedVariable(DVI->getVariable(), 
00874                                                    InlinedAt, Ctx));
00875         }
00876       }
00877     }
00878   }
00879 }
00880 
00881 /// InlineFunction - This function inlines the called function into the basic
00882 /// block of the caller.  This returns false if it is not possible to inline
00883 /// this call.  The program is still in a well defined state if this occurs
00884 /// though.
00885 ///
00886 /// Note that this only does one level of inlining.  For example, if the
00887 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
00888 /// exists in the instruction stream.  Similarly this will inline a recursive
00889 /// function by one level.
00890 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
00891                           bool InsertLifetime) {
00892   Instruction *TheCall = CS.getInstruction();
00893   assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
00894          "Instruction not in function!");
00895 
00896   // If IFI has any state in it, zap it before we fill it in.
00897   IFI.reset();
00898   
00899   const Function *CalledFunc = CS.getCalledFunction();
00900   if (!CalledFunc ||              // Can't inline external function or indirect
00901       CalledFunc->isDeclaration() || // call, or call to a vararg function!
00902       CalledFunc->getFunctionType()->isVarArg()) return false;
00903 
00904   // If the call to the callee cannot throw, set the 'nounwind' flag on any
00905   // calls that we inline.
00906   bool MarkNoUnwind = CS.doesNotThrow();
00907 
00908   BasicBlock *OrigBB = TheCall->getParent();
00909   Function *Caller = OrigBB->getParent();
00910 
00911   // GC poses two hazards to inlining, which only occur when the callee has GC:
00912   //  1. If the caller has no GC, then the callee's GC must be propagated to the
00913   //     caller.
00914   //  2. If the caller has a differing GC, it is invalid to inline.
00915   if (CalledFunc->hasGC()) {
00916     if (!Caller->hasGC())
00917       Caller->setGC(CalledFunc->getGC());
00918     else if (CalledFunc->getGC() != Caller->getGC())
00919       return false;
00920   }
00921 
00922   // Get the personality function from the callee if it contains a landing pad.
00923   Value *CalleePersonality = nullptr;
00924   for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end();
00925        I != E; ++I)
00926     if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
00927       const BasicBlock *BB = II->getUnwindDest();
00928       const LandingPadInst *LP = BB->getLandingPadInst();
00929       CalleePersonality = LP->getPersonalityFn();
00930       break;
00931     }
00932 
00933   // Find the personality function used by the landing pads of the caller. If it
00934   // exists, then check to see that it matches the personality function used in
00935   // the callee.
00936   if (CalleePersonality) {
00937     for (Function::const_iterator I = Caller->begin(), E = Caller->end();
00938          I != E; ++I)
00939       if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
00940         const BasicBlock *BB = II->getUnwindDest();
00941         const LandingPadInst *LP = BB->getLandingPadInst();
00942 
00943         // If the personality functions match, then we can perform the
00944         // inlining. Otherwise, we can't inline.
00945         // TODO: This isn't 100% true. Some personality functions are proper
00946         //       supersets of others and can be used in place of the other.
00947         if (LP->getPersonalityFn() != CalleePersonality)
00948           return false;
00949 
00950         break;
00951       }
00952   }
00953 
00954   // Get an iterator to the last basic block in the function, which will have
00955   // the new function inlined after it.
00956   Function::iterator LastBlock = &Caller->back();
00957 
00958   // Make sure to capture all of the return instructions from the cloned
00959   // function.
00960   SmallVector<ReturnInst*, 8> Returns;
00961   ClonedCodeInfo InlinedFunctionInfo;
00962   Function::iterator FirstNewBlock;
00963   bool AddedNewAllocas = false;
00964 
00965   { // Scope to destroy VMap after cloning.
00966     ValueToValueMapTy VMap;
00967     // Keep a list of pair (dst, src) to emit byval initializations.
00968     SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
00969 
00970     assert(CalledFunc->arg_size() == CS.arg_size() &&
00971            "No varargs calls can be inlined!");
00972 
00973     // Calculate the vector of arguments to pass into the function cloner, which
00974     // matches up the formal to the actual argument values.
00975     CallSite::arg_iterator AI = CS.arg_begin();
00976     unsigned ArgNo = 0;
00977     for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
00978          E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
00979       Value *ActualArg = *AI;
00980 
00981       // When byval arguments actually inlined, we need to make the copy implied
00982       // by them explicit.  However, we don't do this if the callee is readonly
00983       // or readnone, because the copy would be unneeded: the callee doesn't
00984       // modify the struct.
00985       if (CS.isByValArgument(ArgNo)) {
00986         ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
00987                                         CalledFunc->getParamAlignment(ArgNo+1),
00988                                         AddedNewAllocas);
00989         if (ActualArg != *AI)
00990           ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
00991       }
00992 
00993       VMap[I] = ActualArg;
00994     }
00995 
00996     // Add alignment assumptions if necessary. We do this before the inlined
00997     // instructions are actually cloned into the caller so that we can easily
00998     // check what will be known at the start of the inlined code.
00999     AddAlignmentAssumptions(CS, IFI);
01000 
01001     // We want the inliner to prune the code as it copies.  We would LOVE to
01002     // have no dead or constant instructions leftover after inlining occurs
01003     // (which can happen, e.g., because an argument was constant), but we'll be
01004     // happy with whatever the cloner can do.
01005     CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 
01006                               /*ModuleLevelChanges=*/false, Returns, ".i",
01007                               &InlinedFunctionInfo, IFI.DL, TheCall);
01008 
01009     // Remember the first block that is newly cloned over.
01010     FirstNewBlock = LastBlock; ++FirstNewBlock;
01011 
01012     // Inject byval arguments initialization.
01013     for (std::pair<Value*, Value*> &Init : ByValInit)
01014       HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
01015                               FirstNewBlock, IFI);
01016 
01017     // Update the callgraph if requested.
01018     if (IFI.CG)
01019       UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
01020 
01021     // Update inlined instructions' line number information.
01022     fixupLineNumbers(Caller, FirstNewBlock, TheCall);
01023 
01024     // Clone existing noalias metadata if necessary.
01025     CloneAliasScopeMetadata(CS, VMap);
01026 
01027     // Add noalias metadata if necessary.
01028     AddAliasScopeMetadata(CS, VMap, IFI.DL, IFI.AA);
01029 
01030     // FIXME: We could register any cloned assumptions instead of clearing the
01031     // whole function's cache.
01032     if (IFI.AT)
01033       IFI.AT->forgetCachedAssumptions(Caller);
01034   }
01035 
01036   // If there are any alloca instructions in the block that used to be the entry
01037   // block for the callee, move them to the entry block of the caller.  First
01038   // calculate which instruction they should be inserted before.  We insert the
01039   // instructions at the end of the current alloca list.
01040   {
01041     BasicBlock::iterator InsertPoint = Caller->begin()->begin();
01042     for (BasicBlock::iterator I = FirstNewBlock->begin(),
01043          E = FirstNewBlock->end(); I != E; ) {
01044       AllocaInst *AI = dyn_cast<AllocaInst>(I++);
01045       if (!AI) continue;
01046       
01047       // If the alloca is now dead, remove it.  This often occurs due to code
01048       // specialization.
01049       if (AI->use_empty()) {
01050         AI->eraseFromParent();
01051         continue;
01052       }
01053 
01054       if (!isa<Constant>(AI->getArraySize()))
01055         continue;
01056       
01057       // Keep track of the static allocas that we inline into the caller.
01058       IFI.StaticAllocas.push_back(AI);
01059       
01060       // Scan for the block of allocas that we can move over, and move them
01061       // all at once.
01062       while (isa<AllocaInst>(I) &&
01063              isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
01064         IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
01065         ++I;
01066       }
01067 
01068       // Transfer all of the allocas over in a block.  Using splice means
01069       // that the instructions aren't removed from the symbol table, then
01070       // reinserted.
01071       Caller->getEntryBlock().getInstList().splice(InsertPoint,
01072                                                    FirstNewBlock->getInstList(),
01073                                                    AI, I);
01074     }
01075   }
01076 
01077   bool InlinedMustTailCalls = false;
01078   if (InlinedFunctionInfo.ContainsCalls) {
01079     CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
01080     if (CallInst *CI = dyn_cast<CallInst>(TheCall))
01081       CallSiteTailKind = CI->getTailCallKind();
01082 
01083     for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
01084          ++BB) {
01085       for (Instruction &I : *BB) {
01086         CallInst *CI = dyn_cast<CallInst>(&I);
01087         if (!CI)
01088           continue;
01089 
01090         // We need to reduce the strength of any inlined tail calls.  For
01091         // musttail, we have to avoid introducing potential unbounded stack
01092         // growth.  For example, if functions 'f' and 'g' are mutually recursive
01093         // with musttail, we can inline 'g' into 'f' so long as we preserve
01094         // musttail on the cloned call to 'f'.  If either the inlined call site
01095         // or the cloned call site is *not* musttail, the program already has
01096         // one frame of stack growth, so it's safe to remove musttail.  Here is
01097         // a table of example transformations:
01098         //
01099         //    f -> musttail g -> musttail f  ==>  f -> musttail f
01100         //    f -> musttail g ->     tail f  ==>  f ->     tail f
01101         //    f ->          g -> musttail f  ==>  f ->          f
01102         //    f ->          g ->     tail f  ==>  f ->          f
01103         //
01104         // If an alloca was introduced in the frame due to a byval parameter
01105         // being passed to a subsequent call, tail calls must have the tail
01106         // stripped as they may not access variables in the caller's stack.
01107         // A single alloca ripples through out as the alloca may be aliased by
01108         // bitcasts or may escape and be mutated outside of the function.
01109         CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
01110         ChildTCK = std::min(CallSiteTailKind, ChildTCK);
01111         if (AddedNewAllocas)
01112           CI->setTailCallKind(CallInst::TCK_None);
01113         else
01114           CI->setTailCallKind(ChildTCK);
01115         InlinedMustTailCalls |= CI->isMustTailCall();
01116 
01117         // Calls inlined through a 'nounwind' call site should be marked
01118         // 'nounwind'.
01119         if (MarkNoUnwind)
01120           CI->setDoesNotThrow();
01121       }
01122     }
01123   }
01124 
01125   // Leave lifetime markers for the static alloca's, scoping them to the
01126   // function we just inlined.
01127   if (InsertLifetime && !IFI.StaticAllocas.empty()) {
01128     IRBuilder<> builder(FirstNewBlock->begin());
01129     for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
01130       AllocaInst *AI = IFI.StaticAllocas[ai];
01131 
01132       // If the alloca is already scoped to something smaller than the whole
01133       // function then there's no need to add redundant, less accurate markers.
01134       if (hasLifetimeMarkers(AI))
01135         continue;
01136 
01137       // Try to determine the size of the allocation.
01138       ConstantInt *AllocaSize = nullptr;
01139       if (ConstantInt *AIArraySize =
01140           dyn_cast<ConstantInt>(AI->getArraySize())) {
01141         if (IFI.DL) {
01142           Type *AllocaType = AI->getAllocatedType();
01143           uint64_t AllocaTypeSize = IFI.DL->getTypeAllocSize(AllocaType);
01144           uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
01145           assert(AllocaArraySize > 0 && "array size of AllocaInst is zero");
01146           // Check that array size doesn't saturate uint64_t and doesn't
01147           // overflow when it's multiplied by type size.
01148           if (AllocaArraySize != ~0ULL &&
01149               UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
01150             AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
01151                                           AllocaArraySize * AllocaTypeSize);
01152           }
01153         }
01154       }
01155 
01156       builder.CreateLifetimeStart(AI, AllocaSize);
01157       for (ReturnInst *RI : Returns) {
01158         // Don't insert llvm.lifetime.end calls between a musttail call and a
01159         // return.  The return kills all local allocas.
01160         if (InlinedMustTailCalls &&
01161             RI->getParent()->getTerminatingMustTailCall())
01162           continue;
01163         IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
01164       }
01165     }
01166   }
01167 
01168   // If the inlined code contained dynamic alloca instructions, wrap the inlined
01169   // code with llvm.stacksave/llvm.stackrestore intrinsics.
01170   if (InlinedFunctionInfo.ContainsDynamicAllocas) {
01171     Module *M = Caller->getParent();
01172     // Get the two intrinsics we care about.
01173     Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
01174     Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
01175 
01176     // Insert the llvm.stacksave.
01177     CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
01178       .CreateCall(StackSave, "savedstack");
01179 
01180     // Insert a call to llvm.stackrestore before any return instructions in the
01181     // inlined function.
01182     for (ReturnInst *RI : Returns) {
01183       // Don't insert llvm.stackrestore calls between a musttail call and a
01184       // return.  The return will restore the stack pointer.
01185       if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
01186         continue;
01187       IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
01188     }
01189   }
01190 
01191   // If we are inlining for an invoke instruction, we must make sure to rewrite
01192   // any call instructions into invoke instructions.
01193   if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
01194     HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
01195 
01196   // Handle any inlined musttail call sites.  In order for a new call site to be
01197   // musttail, the source of the clone and the inlined call site must have been
01198   // musttail.  Therefore it's safe to return without merging control into the
01199   // phi below.
01200   if (InlinedMustTailCalls) {
01201     // Check if we need to bitcast the result of any musttail calls.
01202     Type *NewRetTy = Caller->getReturnType();
01203     bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
01204 
01205     // Handle the returns preceded by musttail calls separately.
01206     SmallVector<ReturnInst *, 8> NormalReturns;
01207     for (ReturnInst *RI : Returns) {
01208       CallInst *ReturnedMustTail =
01209           RI->getParent()->getTerminatingMustTailCall();
01210       if (!ReturnedMustTail) {
01211         NormalReturns.push_back(RI);
01212         continue;
01213       }
01214       if (!NeedBitCast)
01215         continue;
01216 
01217       // Delete the old return and any preceding bitcast.
01218       BasicBlock *CurBB = RI->getParent();
01219       auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
01220       RI->eraseFromParent();
01221       if (OldCast)
01222         OldCast->eraseFromParent();
01223 
01224       // Insert a new bitcast and return with the right type.
01225       IRBuilder<> Builder(CurBB);
01226       Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
01227     }
01228 
01229     // Leave behind the normal returns so we can merge control flow.
01230     std::swap(Returns, NormalReturns);
01231   }
01232 
01233   // If we cloned in _exactly one_ basic block, and if that block ends in a
01234   // return instruction, we splice the body of the inlined callee directly into
01235   // the calling basic block.
01236   if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
01237     // Move all of the instructions right before the call.
01238     OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
01239                                  FirstNewBlock->begin(), FirstNewBlock->end());
01240     // Remove the cloned basic block.
01241     Caller->getBasicBlockList().pop_back();
01242 
01243     // If the call site was an invoke instruction, add a branch to the normal
01244     // destination.
01245     if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
01246       BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
01247       NewBr->setDebugLoc(Returns[0]->getDebugLoc());
01248     }
01249 
01250     // If the return instruction returned a value, replace uses of the call with
01251     // uses of the returned value.
01252     if (!TheCall->use_empty()) {
01253       ReturnInst *R = Returns[0];
01254       if (TheCall == R->getReturnValue())
01255         TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
01256       else
01257         TheCall->replaceAllUsesWith(R->getReturnValue());
01258     }
01259     // Since we are now done with the Call/Invoke, we can delete it.
01260     TheCall->eraseFromParent();
01261 
01262     // Since we are now done with the return instruction, delete it also.
01263     Returns[0]->eraseFromParent();
01264 
01265     // We are now done with the inlining.
01266     return true;
01267   }
01268 
01269   // Otherwise, we have the normal case, of more than one block to inline or
01270   // multiple return sites.
01271 
01272   // We want to clone the entire callee function into the hole between the
01273   // "starter" and "ender" blocks.  How we accomplish this depends on whether
01274   // this is an invoke instruction or a call instruction.
01275   BasicBlock *AfterCallBB;
01276   BranchInst *CreatedBranchToNormalDest = nullptr;
01277   if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
01278 
01279     // Add an unconditional branch to make this look like the CallInst case...
01280     CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
01281 
01282     // Split the basic block.  This guarantees that no PHI nodes will have to be
01283     // updated due to new incoming edges, and make the invoke case more
01284     // symmetric to the call case.
01285     AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest,
01286                                           CalledFunc->getName()+".exit");
01287 
01288   } else {  // It's a call
01289     // If this is a call instruction, we need to split the basic block that
01290     // the call lives in.
01291     //
01292     AfterCallBB = OrigBB->splitBasicBlock(TheCall,
01293                                           CalledFunc->getName()+".exit");
01294   }
01295 
01296   // Change the branch that used to go to AfterCallBB to branch to the first
01297   // basic block of the inlined function.
01298   //
01299   TerminatorInst *Br = OrigBB->getTerminator();
01300   assert(Br && Br->getOpcode() == Instruction::Br &&
01301          "splitBasicBlock broken!");
01302   Br->setOperand(0, FirstNewBlock);
01303 
01304 
01305   // Now that the function is correct, make it a little bit nicer.  In
01306   // particular, move the basic blocks inserted from the end of the function
01307   // into the space made by splitting the source basic block.
01308   Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
01309                                      FirstNewBlock, Caller->end());
01310 
01311   // Handle all of the return instructions that we just cloned in, and eliminate
01312   // any users of the original call/invoke instruction.
01313   Type *RTy = CalledFunc->getReturnType();
01314 
01315   PHINode *PHI = nullptr;
01316   if (Returns.size() > 1) {
01317     // The PHI node should go at the front of the new basic block to merge all
01318     // possible incoming values.
01319     if (!TheCall->use_empty()) {
01320       PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
01321                             AfterCallBB->begin());
01322       // Anything that used the result of the function call should now use the
01323       // PHI node as their operand.
01324       TheCall->replaceAllUsesWith(PHI);
01325     }
01326 
01327     // Loop over all of the return instructions adding entries to the PHI node
01328     // as appropriate.
01329     if (PHI) {
01330       for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
01331         ReturnInst *RI = Returns[i];
01332         assert(RI->getReturnValue()->getType() == PHI->getType() &&
01333                "Ret value not consistent in function!");
01334         PHI->addIncoming(RI->getReturnValue(), RI->getParent());
01335       }
01336     }
01337 
01338 
01339     // Add a branch to the merge points and remove return instructions.
01340     DebugLoc Loc;
01341     for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
01342       ReturnInst *RI = Returns[i];
01343       BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
01344       Loc = RI->getDebugLoc();
01345       BI->setDebugLoc(Loc);
01346       RI->eraseFromParent();
01347     }
01348     // We need to set the debug location to *somewhere* inside the
01349     // inlined function. The line number may be nonsensical, but the
01350     // instruction will at least be associated with the right
01351     // function.
01352     if (CreatedBranchToNormalDest)
01353       CreatedBranchToNormalDest->setDebugLoc(Loc);
01354   } else if (!Returns.empty()) {
01355     // Otherwise, if there is exactly one return value, just replace anything
01356     // using the return value of the call with the computed value.
01357     if (!TheCall->use_empty()) {
01358       if (TheCall == Returns[0]->getReturnValue())
01359         TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
01360       else
01361         TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
01362     }
01363 
01364     // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
01365     BasicBlock *ReturnBB = Returns[0]->getParent();
01366     ReturnBB->replaceAllUsesWith(AfterCallBB);
01367 
01368     // Splice the code from the return block into the block that it will return
01369     // to, which contains the code that was after the call.
01370     AfterCallBB->getInstList().splice(AfterCallBB->begin(),
01371                                       ReturnBB->getInstList());
01372 
01373     if (CreatedBranchToNormalDest)
01374       CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
01375 
01376     // Delete the return instruction now and empty ReturnBB now.
01377     Returns[0]->eraseFromParent();
01378     ReturnBB->eraseFromParent();
01379   } else if (!TheCall->use_empty()) {
01380     // No returns, but something is using the return value of the call.  Just
01381     // nuke the result.
01382     TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
01383   }
01384 
01385   // Since we are now done with the Call/Invoke, we can delete it.
01386   TheCall->eraseFromParent();
01387 
01388   // If we inlined any musttail calls and the original return is now
01389   // unreachable, delete it.  It can only contain a bitcast and ret.
01390   if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB))
01391     AfterCallBB->eraseFromParent();
01392 
01393   // We should always be able to fold the entry block of the function into the
01394   // single predecessor of the block...
01395   assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
01396   BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
01397 
01398   // Splice the code entry block into calling block, right before the
01399   // unconditional branch.
01400   CalleeEntry->replaceAllUsesWith(OrigBB);  // Update PHI nodes
01401   OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
01402 
01403   // Remove the unconditional branch.
01404   OrigBB->getInstList().erase(Br);
01405 
01406   // Now we can remove the CalleeEntry block, which is now empty.
01407   Caller->getBasicBlockList().erase(CalleeEntry);
01408 
01409   // If we inserted a phi node, check to see if it has a single value (e.g. all
01410   // the entries are the same or undef).  If so, remove the PHI so it doesn't
01411   // block other optimizations.
01412   if (PHI) {
01413     if (Value *V = SimplifyInstruction(PHI, IFI.DL, nullptr, nullptr, IFI.AT)) {
01414       PHI->replaceAllUsesWith(V);
01415       PHI->eraseFromParent();
01416     }
01417   }
01418 
01419   return true;
01420 }