1 //===- InlineFunction.cpp - Code to perform function inlining -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements inlining of a function into a call site, resolving
11 // parameters and the return value as appropriate.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Utils/Cloning.h"
16 #include "llvm/ADT/SmallSet.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/StringExtras.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/CallGraph.h"
22 #include "llvm/Analysis/CaptureTracking.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/Attributes.h"
26 #include "llvm/IR/CallSite.h"
27 #include "llvm/IR/CFG.h"
28 #include "llvm/IR/Constants.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/DebugInfo.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/Dominators.h"
33 #include "llvm/IR/IRBuilder.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/IntrinsicInst.h"
36 #include "llvm/IR/Intrinsics.h"
37 #include "llvm/IR/MDBuilder.h"
38 #include "llvm/IR/Module.h"
39 #include "llvm/Transforms/Utils/Local.h"
40 #include "llvm/Support/CommandLine.h"
45 EnableNoAliasConversion("enable-noalias-to-md-conversion", cl::init(false),
47 cl::desc("Convert noalias attributes to metadata during inlining."));
49 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI,
50 bool InsertLifetime) {
51 return InlineFunction(CallSite(CI), IFI, InsertLifetime);
53 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI,
54 bool InsertLifetime) {
55 return InlineFunction(CallSite(II), IFI, InsertLifetime);
59 /// A class for recording information about inlining through an invoke.
60 class InvokeInliningInfo {
61 BasicBlock *OuterResumeDest; ///< Destination of the invoke's unwind.
62 BasicBlock *InnerResumeDest; ///< Destination for the callee's resume.
63 LandingPadInst *CallerLPad; ///< LandingPadInst associated with the invoke.
64 PHINode *InnerEHValuesPHI; ///< PHI for EH values from landingpad insts.
65 SmallVector<Value*, 8> UnwindDestPHIValues;
68 InvokeInliningInfo(InvokeInst *II)
69 : OuterResumeDest(II->getUnwindDest()), InnerResumeDest(nullptr),
70 CallerLPad(nullptr), InnerEHValuesPHI(nullptr) {
71 // If there are PHI nodes in the unwind destination block, we need to keep
72 // track of which values came into them from the invoke before removing
73 // the edge from this block.
74 llvm::BasicBlock *InvokeBB = II->getParent();
75 BasicBlock::iterator I = OuterResumeDest->begin();
76 for (; isa<PHINode>(I); ++I) {
77 // Save the value to use for this edge.
78 PHINode *PHI = cast<PHINode>(I);
79 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
82 CallerLPad = cast<LandingPadInst>(I);
85 /// getOuterResumeDest - The outer unwind destination is the target of
86 /// unwind edges introduced for calls within the inlined function.
87 BasicBlock *getOuterResumeDest() const {
88 return OuterResumeDest;
91 BasicBlock *getInnerResumeDest();
93 LandingPadInst *getLandingPadInst() const { return CallerLPad; }
95 /// forwardResume - Forward the 'resume' instruction to the caller's landing
96 /// pad block. When the landing pad block has only one predecessor, this is
97 /// a simple branch. When there is more than one predecessor, we need to
98 /// split the landing pad block after the landingpad instruction and jump
100 void forwardResume(ResumeInst *RI,
101 SmallPtrSet<LandingPadInst*, 16> &InlinedLPads);
103 /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind
104 /// destination block for the given basic block, using the values for the
105 /// original invoke's source block.
106 void addIncomingPHIValuesFor(BasicBlock *BB) const {
107 addIncomingPHIValuesForInto(BB, OuterResumeDest);
110 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const {
111 BasicBlock::iterator I = dest->begin();
112 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
113 PHINode *phi = cast<PHINode>(I);
114 phi->addIncoming(UnwindDestPHIValues[i], src);
120 /// getInnerResumeDest - Get or create a target for the branch from ResumeInsts.
121 BasicBlock *InvokeInliningInfo::getInnerResumeDest() {
122 if (InnerResumeDest) return InnerResumeDest;
124 // Split the landing pad.
125 BasicBlock::iterator SplitPoint = CallerLPad; ++SplitPoint;
127 OuterResumeDest->splitBasicBlock(SplitPoint,
128 OuterResumeDest->getName() + ".body");
130 // The number of incoming edges we expect to the inner landing pad.
131 const unsigned PHICapacity = 2;
133 // Create corresponding new PHIs for all the PHIs in the outer landing pad.
134 BasicBlock::iterator InsertPoint = InnerResumeDest->begin();
135 BasicBlock::iterator I = OuterResumeDest->begin();
136 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
137 PHINode *OuterPHI = cast<PHINode>(I);
138 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity,
139 OuterPHI->getName() + ".lpad-body",
141 OuterPHI->replaceAllUsesWith(InnerPHI);
142 InnerPHI->addIncoming(OuterPHI, OuterResumeDest);
145 // Create a PHI for the exception values.
146 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity,
147 "eh.lpad-body", InsertPoint);
148 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI);
149 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest);
152 return InnerResumeDest;
155 /// forwardResume - Forward the 'resume' instruction to the caller's landing pad
156 /// block. When the landing pad block has only one predecessor, this is a simple
157 /// branch. When there is more than one predecessor, we need to split the
158 /// landing pad block after the landingpad instruction and jump to there.
159 void InvokeInliningInfo::forwardResume(ResumeInst *RI,
160 SmallPtrSet<LandingPadInst*, 16> &InlinedLPads) {
161 BasicBlock *Dest = getInnerResumeDest();
162 BasicBlock *Src = RI->getParent();
164 BranchInst::Create(Dest, Src);
166 // Update the PHIs in the destination. They were inserted in an order which
168 addIncomingPHIValuesForInto(Src, Dest);
170 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src);
171 RI->eraseFromParent();
174 /// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
175 /// an invoke, we have to turn all of the calls that can throw into
176 /// invokes. This function analyze BB to see if there are any calls, and if so,
177 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
178 /// nodes in that block with the values specified in InvokeDestPHIValues.
179 static void HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
180 InvokeInliningInfo &Invoke) {
181 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
182 Instruction *I = BBI++;
184 // We only need to check for function calls: inlined invoke
185 // instructions require no special handling.
186 CallInst *CI = dyn_cast<CallInst>(I);
188 // If this call cannot unwind, don't convert it to an invoke.
189 // Inline asm calls cannot throw.
190 if (!CI || CI->doesNotThrow() || isa<InlineAsm>(CI->getCalledValue()))
193 // Convert this function call into an invoke instruction. First, split the
195 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
197 // Delete the unconditional branch inserted by splitBasicBlock
198 BB->getInstList().pop_back();
200 // Create the new invoke instruction.
201 ImmutableCallSite CS(CI);
202 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
203 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split,
204 Invoke.getOuterResumeDest(),
205 InvokeArgs, CI->getName(), BB);
206 II->setDebugLoc(CI->getDebugLoc());
207 II->setCallingConv(CI->getCallingConv());
208 II->setAttributes(CI->getAttributes());
210 // Make sure that anything using the call now uses the invoke! This also
211 // updates the CallGraph if present, because it uses a WeakVH.
212 CI->replaceAllUsesWith(II);
214 // Delete the original call
215 Split->getInstList().pop_front();
217 // Update any PHI nodes in the exceptional block to indicate that there is
218 // now a new entry in them.
219 Invoke.addIncomingPHIValuesFor(BB);
224 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
225 /// in the body of the inlined function into invokes.
227 /// II is the invoke instruction being inlined. FirstNewBlock is the first
228 /// block of the inlined code (the last block is the end of the function),
229 /// and InlineCodeInfo is information about the code that got inlined.
230 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
231 ClonedCodeInfo &InlinedCodeInfo) {
232 BasicBlock *InvokeDest = II->getUnwindDest();
234 Function *Caller = FirstNewBlock->getParent();
236 // The inlined code is currently at the end of the function, scan from the
237 // start of the inlined code to its end, checking for stuff we need to
239 InvokeInliningInfo Invoke(II);
241 // Get all of the inlined landing pad instructions.
242 SmallPtrSet<LandingPadInst*, 16> InlinedLPads;
243 for (Function::iterator I = FirstNewBlock, E = Caller->end(); I != E; ++I)
244 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator()))
245 InlinedLPads.insert(II->getLandingPadInst());
247 // Append the clauses from the outer landing pad instruction into the inlined
248 // landing pad instructions.
249 LandingPadInst *OuterLPad = Invoke.getLandingPadInst();
250 for (SmallPtrSet<LandingPadInst*, 16>::iterator I = InlinedLPads.begin(),
251 E = InlinedLPads.end(); I != E; ++I) {
252 LandingPadInst *InlinedLPad = *I;
253 unsigned OuterNum = OuterLPad->getNumClauses();
254 InlinedLPad->reserveClauses(OuterNum);
255 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx)
256 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx));
257 if (OuterLPad->isCleanup())
258 InlinedLPad->setCleanup(true);
261 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
262 if (InlinedCodeInfo.ContainsCalls)
263 HandleCallsInBlockInlinedThroughInvoke(BB, Invoke);
265 // Forward any resumes that are remaining here.
266 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator()))
267 Invoke.forwardResume(RI, InlinedLPads);
270 // Now that everything is happy, we have one final detail. The PHI nodes in
271 // the exception destination block still have entries due to the original
272 // invoke instruction. Eliminate these entries (which might even delete the
274 InvokeDest->removePredecessor(II->getParent());
277 /// CloneAliasScopeMetadata - When inlining a function that contains noalias
278 /// scope metadata, this metadata needs to be cloned so that the inlined blocks
279 /// have different "unqiue scopes" at every call site. Were this not done, then
280 /// aliasing scopes from a function inlined into a caller multiple times could
281 /// not be differentiated (and this would lead to miscompiles because the
282 /// non-aliasing property communicated by the metadata could have
283 /// call-site-specific control dependencies).
284 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) {
285 const Function *CalledFunc = CS.getCalledFunction();
286 SetVector<const MDNode *> MD;
288 // Note: We could only clone the metadata if it is already used in the
289 // caller. I'm omitting that check here because it might confuse
290 // inter-procedural alias analysis passes. We can revisit this if it becomes
291 // an efficiency or overhead problem.
293 for (Function::const_iterator I = CalledFunc->begin(), IE = CalledFunc->end();
295 for (BasicBlock::const_iterator J = I->begin(), JE = I->end(); J != JE; ++J) {
296 if (const MDNode *M = J->getMetadata(LLVMContext::MD_alias_scope))
298 if (const MDNode *M = J->getMetadata(LLVMContext::MD_noalias))
305 // Walk the existing metadata, adding the complete (perhaps cyclic) chain to
307 SmallVector<const Value *, 16> Queue(MD.begin(), MD.end());
308 while (!Queue.empty()) {
309 const MDNode *M = cast<MDNode>(Queue.pop_back_val());
310 for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i)
311 if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i)))
316 // Now we have a complete set of all metadata in the chains used to specify
317 // the noalias scopes and the lists of those scopes.
318 SmallVector<MDNode *, 16> DummyNodes;
319 DenseMap<const MDNode *, TrackingVH<MDNode> > MDMap;
320 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
322 MDNode *Dummy = MDNode::getTemporary(CalledFunc->getContext(),
324 DummyNodes.push_back(Dummy);
328 // Create new metadata nodes to replace the dummy nodes, replacing old
329 // metadata references with either a dummy node or an already-created new
331 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
333 SmallVector<Value *, 4> NewOps;
334 for (unsigned i = 0, ie = (*I)->getNumOperands(); i != ie; ++i) {
335 const Value *V = (*I)->getOperand(i);
336 if (const MDNode *M = dyn_cast<MDNode>(V))
337 NewOps.push_back(MDMap[M]);
339 NewOps.push_back(const_cast<Value *>(V));
342 MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps),
345 TempM->replaceAllUsesWith(NewM);
348 // Now replace the metadata in the new inlined instructions with the
349 // repacements from the map.
350 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
351 VMI != VMIE; ++VMI) {
355 Instruction *NI = dyn_cast<Instruction>(VMI->second);
359 if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope)) {
360 MDNode *NewMD = MDMap[M];
361 // If the call site also had alias scope metadata (a list of scopes to
362 // which instructions inside it might belong), propagate those scopes to
363 // the inlined instructions.
365 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
366 NewMD = MDNode::concatenate(NewMD, CSM);
367 NI->setMetadata(LLVMContext::MD_alias_scope, NewMD);
368 } else if (NI->mayReadOrWriteMemory()) {
370 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
371 NI->setMetadata(LLVMContext::MD_alias_scope, M);
374 if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias)) {
375 MDNode *NewMD = MDMap[M];
376 // If the call site also had noalias metadata (a list of scopes with
377 // which instructions inside it don't alias), propagate those scopes to
378 // the inlined instructions.
380 CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
381 NewMD = MDNode::concatenate(NewMD, CSM);
382 NI->setMetadata(LLVMContext::MD_noalias, NewMD);
383 } else if (NI->mayReadOrWriteMemory()) {
385 CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
386 NI->setMetadata(LLVMContext::MD_noalias, M);
390 // Now that everything has been replaced, delete the dummy nodes.
391 for (unsigned i = 0, ie = DummyNodes.size(); i != ie; ++i)
392 MDNode::deleteTemporary(DummyNodes[i]);
395 /// AddAliasScopeMetadata - If the inlined function has noalias arguments, then
396 /// add new alias scopes for each noalias argument, tag the mapped noalias
397 /// parameters with noalias metadata specifying the new scope, and tag all
398 /// non-derived loads, stores and memory intrinsics with the new alias scopes.
399 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
400 const DataLayout *DL) {
401 if (!EnableNoAliasConversion)
404 const Function *CalledFunc = CS.getCalledFunction();
405 SmallVector<const Argument *, 4> NoAliasArgs;
407 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
408 E = CalledFunc->arg_end(); I != E; ++I) {
409 if (I->hasNoAliasAttr() && !I->hasNUses(0))
410 NoAliasArgs.push_back(I);
413 if (NoAliasArgs.empty())
416 // To do a good job, if a noalias variable is captured, we need to know if
417 // the capture point dominates the particular use we're considering.
419 DT.recalculate(const_cast<Function&>(*CalledFunc));
421 // noalias indicates that pointer values based on the argument do not alias
422 // pointer values which are not based on it. So we add a new "scope" for each
423 // noalias function argument. Accesses using pointers based on that argument
424 // become part of that alias scope, accesses using pointers not based on that
425 // argument are tagged as noalias with that scope.
427 DenseMap<const Argument *, MDNode *> NewScopes;
428 MDBuilder MDB(CalledFunc->getContext());
430 // Create a new scope domain for this function.
432 MDB.createAnonymousAliasScopeDomain(CalledFunc->getName());
433 for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) {
434 const Argument *A = NoAliasArgs[i];
436 std::string Name = CalledFunc->getName();
439 Name += A->getName();
441 Name += ": argument ";
445 // Note: We always create a new anonymous root here. This is true regardless
446 // of the linkage of the callee because the aliasing "scope" is not just a
447 // property of the callee, but also all control dependencies in the caller.
448 MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
449 NewScopes.insert(std::make_pair(A, NewScope));
452 // Iterate over all new instructions in the map; for all memory-access
453 // instructions, add the alias scope metadata.
454 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
455 VMI != VMIE; ++VMI) {
456 if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) {
460 Instruction *NI = dyn_cast<Instruction>(VMI->second);
464 SmallVector<const Value *, 2> PtrArgs;
466 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
467 PtrArgs.push_back(LI->getPointerOperand());
468 else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
469 PtrArgs.push_back(SI->getPointerOperand());
470 else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
471 PtrArgs.push_back(VAAI->getPointerOperand());
472 else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
473 PtrArgs.push_back(CXI->getPointerOperand());
474 else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
475 PtrArgs.push_back(RMWI->getPointerOperand());
476 else if (ImmutableCallSite ICS = ImmutableCallSite(I)) {
477 // If we know that the call does not access memory, then we'll still
478 // know that about the inlined clone of this call site, and we don't
479 // need to add metadata.
480 if (ICS.doesNotAccessMemory())
483 for (ImmutableCallSite::arg_iterator AI = ICS.arg_begin(),
484 AE = ICS.arg_end(); AI != AE; ++AI)
485 // We need to check the underlying objects of all arguments, not just
486 // the pointer arguments, because we might be passing pointers as
488 // FIXME: If we know that the call only accesses pointer arguments,
489 // then we only need to check the pointer arguments.
490 PtrArgs.push_back(*AI);
493 // If we found no pointers, then this instruction is not suitable for
494 // pairing with an instruction to receive aliasing metadata.
495 // However, if this is a call, this we might just alias with none of the
496 // noalias arguments.
497 if (PtrArgs.empty() && !isa<CallInst>(I) && !isa<InvokeInst>(I))
500 // It is possible that there is only one underlying object, but you
501 // need to go through several PHIs to see it, and thus could be
502 // repeated in the Objects list.
503 SmallPtrSet<const Value *, 4> ObjSet;
504 SmallVector<Value *, 4> Scopes, NoAliases;
506 SmallSetVector<const Argument *, 4> NAPtrArgs;
507 for (unsigned i = 0, ie = PtrArgs.size(); i != ie; ++i) {
508 SmallVector<Value *, 4> Objects;
509 GetUnderlyingObjects(const_cast<Value*>(PtrArgs[i]),
510 Objects, DL, /* MaxLookup = */ 0);
512 for (Value *O : Objects)
516 // Figure out if we're derived from anyhing that is not a noalias
518 bool CanDeriveViaCapture = false;
519 for (const Value *V : ObjSet)
520 if (!isIdentifiedFunctionLocal(const_cast<Value*>(V))) {
521 CanDeriveViaCapture = true;
525 // First, we want to figure out all of the sets with which we definitely
526 // don't alias. Iterate over all noalias set, and add those for which:
527 // 1. The noalias argument is not in the set of objects from which we
528 // definitely derive.
529 // 2. The noalias argument has not yet been captured.
530 for (const Argument *A : NoAliasArgs) {
531 if (!ObjSet.count(A) && (!CanDeriveViaCapture ||
532 A->hasNoCaptureAttr() ||
533 !PointerMayBeCapturedBefore(A,
534 /* ReturnCaptures */ false,
535 /* StoreCaptures */ false, I, &DT)))
536 NoAliases.push_back(NewScopes[A]);
539 if (!NoAliases.empty())
540 NI->setMetadata(LLVMContext::MD_noalias, MDNode::concatenate(
541 NI->getMetadata(LLVMContext::MD_noalias),
542 MDNode::get(CalledFunc->getContext(), NoAliases)));
543 // Next, we want to figure out all of the sets to which we might belong.
544 // We might below to a set if:
545 // 1. The noalias argument is in the set of underlying objects
547 // 2. There is some non-noalias argument in our list and the no-alias
548 // argument has been captured.
550 for (const Argument *A : NoAliasArgs) {
551 if (ObjSet.count(A) || (CanDeriveViaCapture &&
552 PointerMayBeCapturedBefore(A,
553 /* ReturnCaptures */ false,
554 /* StoreCaptures */ false,
556 Scopes.push_back(NewScopes[A]);
560 NI->setMetadata(LLVMContext::MD_alias_scope, MDNode::concatenate(
561 NI->getMetadata(LLVMContext::MD_alias_scope),
562 MDNode::get(CalledFunc->getContext(), Scopes)));
567 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
568 /// into the caller, update the specified callgraph to reflect the changes we
569 /// made. Note that it's possible that not all code was copied over, so only
570 /// some edges of the callgraph may remain.
571 static void UpdateCallGraphAfterInlining(CallSite CS,
572 Function::iterator FirstNewBlock,
573 ValueToValueMapTy &VMap,
574 InlineFunctionInfo &IFI) {
575 CallGraph &CG = *IFI.CG;
576 const Function *Caller = CS.getInstruction()->getParent()->getParent();
577 const Function *Callee = CS.getCalledFunction();
578 CallGraphNode *CalleeNode = CG[Callee];
579 CallGraphNode *CallerNode = CG[Caller];
581 // Since we inlined some uninlined call sites in the callee into the caller,
582 // add edges from the caller to all of the callees of the callee.
583 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
585 // Consider the case where CalleeNode == CallerNode.
586 CallGraphNode::CalledFunctionsVector CallCache;
587 if (CalleeNode == CallerNode) {
588 CallCache.assign(I, E);
589 I = CallCache.begin();
593 for (; I != E; ++I) {
594 const Value *OrigCall = I->first;
596 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
597 // Only copy the edge if the call was inlined!
598 if (VMI == VMap.end() || VMI->second == nullptr)
601 // If the call was inlined, but then constant folded, there is no edge to
602 // add. Check for this case.
603 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
604 if (!NewCall) continue;
606 // Remember that this call site got inlined for the client of
608 IFI.InlinedCalls.push_back(NewCall);
610 // It's possible that inlining the callsite will cause it to go from an
611 // indirect to a direct call by resolving a function pointer. If this
612 // happens, set the callee of the new call site to a more precise
613 // destination. This can also happen if the call graph node of the caller
614 // was just unnecessarily imprecise.
615 if (!I->second->getFunction())
616 if (Function *F = CallSite(NewCall).getCalledFunction()) {
617 // Indirect call site resolved to direct call.
618 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
623 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
626 // Update the call graph by deleting the edge from Callee to Caller. We must
627 // do this after the loop above in case Caller and Callee are the same.
628 CallerNode->removeCallEdgeFor(CS);
631 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
632 BasicBlock *InsertBlock,
633 InlineFunctionInfo &IFI) {
634 LLVMContext &Context = Src->getContext();
635 Type *VoidPtrTy = Type::getInt8PtrTy(Context);
636 Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
637 Type *Tys[3] = { VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context) };
638 Function *MemCpyFn = Intrinsic::getDeclaration(M, Intrinsic::memcpy, Tys);
639 IRBuilder<> builder(InsertBlock->begin());
640 Value *DstCast = builder.CreateBitCast(Dst, VoidPtrTy, "tmp");
641 Value *SrcCast = builder.CreateBitCast(Src, VoidPtrTy, "tmp");
644 if (IFI.DL == nullptr)
645 Size = ConstantExpr::getSizeOf(AggTy);
647 Size = ConstantInt::get(Type::getInt64Ty(Context),
648 IFI.DL->getTypeStoreSize(AggTy));
650 // Always generate a memcpy of alignment 1 here because we don't know
651 // the alignment of the src pointer. Other optimizations can infer
653 Value *CallArgs[] = {
654 DstCast, SrcCast, Size,
655 ConstantInt::get(Type::getInt32Ty(Context), 1),
656 ConstantInt::getFalse(Context) // isVolatile
658 builder.CreateCall(MemCpyFn, CallArgs);
661 /// HandleByValArgument - When inlining a call site that has a byval argument,
662 /// we have to make the implicit memcpy explicit by adding it.
663 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
664 const Function *CalledFunc,
665 InlineFunctionInfo &IFI,
666 unsigned ByValAlignment) {
667 PointerType *ArgTy = cast<PointerType>(Arg->getType());
668 Type *AggTy = ArgTy->getElementType();
670 // If the called function is readonly, then it could not mutate the caller's
671 // copy of the byval'd memory. In this case, it is safe to elide the copy and
673 if (CalledFunc->onlyReadsMemory()) {
674 // If the byval argument has a specified alignment that is greater than the
675 // passed in pointer, then we either have to round up the input pointer or
676 // give up on this transformation.
677 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
680 // If the pointer is already known to be sufficiently aligned, or if we can
681 // round it up to a larger alignment, then we don't need a temporary.
682 if (getOrEnforceKnownAlignment(Arg, ByValAlignment,
683 IFI.DL) >= ByValAlignment)
686 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
687 // for code quality, but rarely happens and is required for correctness.
690 // Create the alloca. If we have DataLayout, use nice alignment.
693 Align = IFI.DL->getPrefTypeAlignment(AggTy);
695 // If the byval had an alignment specified, we *must* use at least that
696 // alignment, as it is required by the byval argument (and uses of the
697 // pointer inside the callee).
698 Align = std::max(Align, ByValAlignment);
700 Function *Caller = TheCall->getParent()->getParent();
702 Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(),
703 &*Caller->begin()->begin());
704 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
706 // Uses of the argument in the function should use our new alloca
711 // isUsedByLifetimeMarker - Check whether this Value is used by a lifetime
713 static bool isUsedByLifetimeMarker(Value *V) {
714 for (User *U : V->users()) {
715 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
716 switch (II->getIntrinsicID()) {
718 case Intrinsic::lifetime_start:
719 case Intrinsic::lifetime_end:
727 // hasLifetimeMarkers - Check whether the given alloca already has
728 // lifetime.start or lifetime.end intrinsics.
729 static bool hasLifetimeMarkers(AllocaInst *AI) {
730 Type *Ty = AI->getType();
731 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
732 Ty->getPointerAddressSpace());
734 return isUsedByLifetimeMarker(AI);
736 // Do a scan to find all the casts to i8*.
737 for (User *U : AI->users()) {
738 if (U->getType() != Int8PtrTy) continue;
739 if (U->stripPointerCasts() != AI) continue;
740 if (isUsedByLifetimeMarker(U))
746 /// updateInlinedAtInfo - Helper function used by fixupLineNumbers to
747 /// recursively update InlinedAtEntry of a DebugLoc.
748 static DebugLoc updateInlinedAtInfo(const DebugLoc &DL,
749 const DebugLoc &InlinedAtDL,
751 if (MDNode *IA = DL.getInlinedAt(Ctx)) {
752 DebugLoc NewInlinedAtDL
753 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx);
754 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
755 NewInlinedAtDL.getAsMDNode(Ctx));
758 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
759 InlinedAtDL.getAsMDNode(Ctx));
762 /// fixupLineNumbers - Update inlined instructions' line numbers to
763 /// to encode location where these instructions are inlined.
764 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
765 Instruction *TheCall) {
766 DebugLoc TheCallDL = TheCall->getDebugLoc();
767 if (TheCallDL.isUnknown())
770 for (; FI != Fn->end(); ++FI) {
771 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
773 DebugLoc DL = BI->getDebugLoc();
774 if (DL.isUnknown()) {
775 // If the inlined instruction has no line number, make it look as if it
776 // originates from the call location. This is important for
777 // ((__always_inline__, __nodebug__)) functions which must use caller
778 // location for all instructions in their function body.
779 BI->setDebugLoc(TheCallDL);
781 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext()));
782 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) {
783 LLVMContext &Ctx = BI->getContext();
784 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
785 DVI->setOperand(2, createInlinedVariable(DVI->getVariable(),
793 /// InlineFunction - This function inlines the called function into the basic
794 /// block of the caller. This returns false if it is not possible to inline
795 /// this call. The program is still in a well defined state if this occurs
798 /// Note that this only does one level of inlining. For example, if the
799 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
800 /// exists in the instruction stream. Similarly this will inline a recursive
801 /// function by one level.
802 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
803 bool InsertLifetime) {
804 Instruction *TheCall = CS.getInstruction();
805 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
806 "Instruction not in function!");
808 // If IFI has any state in it, zap it before we fill it in.
811 const Function *CalledFunc = CS.getCalledFunction();
812 if (!CalledFunc || // Can't inline external function or indirect
813 CalledFunc->isDeclaration() || // call, or call to a vararg function!
814 CalledFunc->getFunctionType()->isVarArg()) return false;
816 // If the call to the callee cannot throw, set the 'nounwind' flag on any
817 // calls that we inline.
818 bool MarkNoUnwind = CS.doesNotThrow();
820 BasicBlock *OrigBB = TheCall->getParent();
821 Function *Caller = OrigBB->getParent();
823 // GC poses two hazards to inlining, which only occur when the callee has GC:
824 // 1. If the caller has no GC, then the callee's GC must be propagated to the
826 // 2. If the caller has a differing GC, it is invalid to inline.
827 if (CalledFunc->hasGC()) {
828 if (!Caller->hasGC())
829 Caller->setGC(CalledFunc->getGC());
830 else if (CalledFunc->getGC() != Caller->getGC())
834 // Get the personality function from the callee if it contains a landing pad.
835 Value *CalleePersonality = nullptr;
836 for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end();
838 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
839 const BasicBlock *BB = II->getUnwindDest();
840 const LandingPadInst *LP = BB->getLandingPadInst();
841 CalleePersonality = LP->getPersonalityFn();
845 // Find the personality function used by the landing pads of the caller. If it
846 // exists, then check to see that it matches the personality function used in
848 if (CalleePersonality) {
849 for (Function::const_iterator I = Caller->begin(), E = Caller->end();
851 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
852 const BasicBlock *BB = II->getUnwindDest();
853 const LandingPadInst *LP = BB->getLandingPadInst();
855 // If the personality functions match, then we can perform the
856 // inlining. Otherwise, we can't inline.
857 // TODO: This isn't 100% true. Some personality functions are proper
858 // supersets of others and can be used in place of the other.
859 if (LP->getPersonalityFn() != CalleePersonality)
866 // Get an iterator to the last basic block in the function, which will have
867 // the new function inlined after it.
868 Function::iterator LastBlock = &Caller->back();
870 // Make sure to capture all of the return instructions from the cloned
872 SmallVector<ReturnInst*, 8> Returns;
873 ClonedCodeInfo InlinedFunctionInfo;
874 Function::iterator FirstNewBlock;
876 { // Scope to destroy VMap after cloning.
877 ValueToValueMapTy VMap;
878 // Keep a list of pair (dst, src) to emit byval initializations.
879 SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
881 assert(CalledFunc->arg_size() == CS.arg_size() &&
882 "No varargs calls can be inlined!");
884 // Calculate the vector of arguments to pass into the function cloner, which
885 // matches up the formal to the actual argument values.
886 CallSite::arg_iterator AI = CS.arg_begin();
888 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
889 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
890 Value *ActualArg = *AI;
892 // When byval arguments actually inlined, we need to make the copy implied
893 // by them explicit. However, we don't do this if the callee is readonly
894 // or readnone, because the copy would be unneeded: the callee doesn't
895 // modify the struct.
896 if (CS.isByValArgument(ArgNo)) {
897 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
898 CalledFunc->getParamAlignment(ArgNo+1));
899 if (ActualArg != *AI)
900 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
906 // We want the inliner to prune the code as it copies. We would LOVE to
907 // have no dead or constant instructions leftover after inlining occurs
908 // (which can happen, e.g., because an argument was constant), but we'll be
909 // happy with whatever the cloner can do.
910 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
911 /*ModuleLevelChanges=*/false, Returns, ".i",
912 &InlinedFunctionInfo, IFI.DL, TheCall);
914 // Remember the first block that is newly cloned over.
915 FirstNewBlock = LastBlock; ++FirstNewBlock;
917 // Inject byval arguments initialization.
918 for (std::pair<Value*, Value*> &Init : ByValInit)
919 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
922 // Update the callgraph if requested.
924 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
926 // Update inlined instructions' line number information.
927 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
929 // Clone existing noalias metadata if necessary.
930 CloneAliasScopeMetadata(CS, VMap);
932 // Add noalias metadata if necessary.
933 AddAliasScopeMetadata(CS, VMap, IFI.DL);
936 // If there are any alloca instructions in the block that used to be the entry
937 // block for the callee, move them to the entry block of the caller. First
938 // calculate which instruction they should be inserted before. We insert the
939 // instructions at the end of the current alloca list.
941 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
942 for (BasicBlock::iterator I = FirstNewBlock->begin(),
943 E = FirstNewBlock->end(); I != E; ) {
944 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
947 // If the alloca is now dead, remove it. This often occurs due to code
949 if (AI->use_empty()) {
950 AI->eraseFromParent();
954 if (!isa<Constant>(AI->getArraySize()))
957 // Keep track of the static allocas that we inline into the caller.
958 IFI.StaticAllocas.push_back(AI);
960 // Scan for the block of allocas that we can move over, and move them
962 while (isa<AllocaInst>(I) &&
963 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
964 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
968 // Transfer all of the allocas over in a block. Using splice means
969 // that the instructions aren't removed from the symbol table, then
971 Caller->getEntryBlock().getInstList().splice(InsertPoint,
972 FirstNewBlock->getInstList(),
977 bool InlinedMustTailCalls = false;
978 if (InlinedFunctionInfo.ContainsCalls) {
979 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
980 if (CallInst *CI = dyn_cast<CallInst>(TheCall))
981 CallSiteTailKind = CI->getTailCallKind();
983 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
985 for (Instruction &I : *BB) {
986 CallInst *CI = dyn_cast<CallInst>(&I);
990 // We need to reduce the strength of any inlined tail calls. For
991 // musttail, we have to avoid introducing potential unbounded stack
992 // growth. For example, if functions 'f' and 'g' are mutually recursive
993 // with musttail, we can inline 'g' into 'f' so long as we preserve
994 // musttail on the cloned call to 'f'. If either the inlined call site
995 // or the cloned call site is *not* musttail, the program already has
996 // one frame of stack growth, so it's safe to remove musttail. Here is
997 // a table of example transformations:
999 // f -> musttail g -> musttail f ==> f -> musttail f
1000 // f -> musttail g -> tail f ==> f -> tail f
1001 // f -> g -> musttail f ==> f -> f
1002 // f -> g -> tail f ==> f -> f
1003 CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
1004 ChildTCK = std::min(CallSiteTailKind, ChildTCK);
1005 CI->setTailCallKind(ChildTCK);
1006 InlinedMustTailCalls |= CI->isMustTailCall();
1008 // Calls inlined through a 'nounwind' call site should be marked
1011 CI->setDoesNotThrow();
1016 // Leave lifetime markers for the static alloca's, scoping them to the
1017 // function we just inlined.
1018 if (InsertLifetime && !IFI.StaticAllocas.empty()) {
1019 IRBuilder<> builder(FirstNewBlock->begin());
1020 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
1021 AllocaInst *AI = IFI.StaticAllocas[ai];
1023 // If the alloca is already scoped to something smaller than the whole
1024 // function then there's no need to add redundant, less accurate markers.
1025 if (hasLifetimeMarkers(AI))
1028 // Try to determine the size of the allocation.
1029 ConstantInt *AllocaSize = nullptr;
1030 if (ConstantInt *AIArraySize =
1031 dyn_cast<ConstantInt>(AI->getArraySize())) {
1033 Type *AllocaType = AI->getAllocatedType();
1034 uint64_t AllocaTypeSize = IFI.DL->getTypeAllocSize(AllocaType);
1035 uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
1036 assert(AllocaArraySize > 0 && "array size of AllocaInst is zero");
1037 // Check that array size doesn't saturate uint64_t and doesn't
1038 // overflow when it's multiplied by type size.
1039 if (AllocaArraySize != ~0ULL &&
1040 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
1041 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
1042 AllocaArraySize * AllocaTypeSize);
1047 builder.CreateLifetimeStart(AI, AllocaSize);
1048 for (ReturnInst *RI : Returns) {
1049 // Don't insert llvm.lifetime.end calls between a musttail call and a
1050 // return. The return kills all local allocas.
1051 if (InlinedMustTailCalls &&
1052 RI->getParent()->getTerminatingMustTailCall())
1054 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
1059 // If the inlined code contained dynamic alloca instructions, wrap the inlined
1060 // code with llvm.stacksave/llvm.stackrestore intrinsics.
1061 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
1062 Module *M = Caller->getParent();
1063 // Get the two intrinsics we care about.
1064 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
1065 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
1067 // Insert the llvm.stacksave.
1068 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
1069 .CreateCall(StackSave, "savedstack");
1071 // Insert a call to llvm.stackrestore before any return instructions in the
1072 // inlined function.
1073 for (ReturnInst *RI : Returns) {
1074 // Don't insert llvm.stackrestore calls between a musttail call and a
1075 // return. The return will restore the stack pointer.
1076 if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
1078 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
1082 // If we are inlining for an invoke instruction, we must make sure to rewrite
1083 // any call instructions into invoke instructions.
1084 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
1085 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
1087 // Handle any inlined musttail call sites. In order for a new call site to be
1088 // musttail, the source of the clone and the inlined call site must have been
1089 // musttail. Therefore it's safe to return without merging control into the
1091 if (InlinedMustTailCalls) {
1092 // Check if we need to bitcast the result of any musttail calls.
1093 Type *NewRetTy = Caller->getReturnType();
1094 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
1096 // Handle the returns preceded by musttail calls separately.
1097 SmallVector<ReturnInst *, 8> NormalReturns;
1098 for (ReturnInst *RI : Returns) {
1099 CallInst *ReturnedMustTail =
1100 RI->getParent()->getTerminatingMustTailCall();
1101 if (!ReturnedMustTail) {
1102 NormalReturns.push_back(RI);
1108 // Delete the old return and any preceding bitcast.
1109 BasicBlock *CurBB = RI->getParent();
1110 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
1111 RI->eraseFromParent();
1113 OldCast->eraseFromParent();
1115 // Insert a new bitcast and return with the right type.
1116 IRBuilder<> Builder(CurBB);
1117 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
1120 // Leave behind the normal returns so we can merge control flow.
1121 std::swap(Returns, NormalReturns);
1124 // If we cloned in _exactly one_ basic block, and if that block ends in a
1125 // return instruction, we splice the body of the inlined callee directly into
1126 // the calling basic block.
1127 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
1128 // Move all of the instructions right before the call.
1129 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
1130 FirstNewBlock->begin(), FirstNewBlock->end());
1131 // Remove the cloned basic block.
1132 Caller->getBasicBlockList().pop_back();
1134 // If the call site was an invoke instruction, add a branch to the normal
1136 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1137 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
1138 NewBr->setDebugLoc(Returns[0]->getDebugLoc());
1141 // If the return instruction returned a value, replace uses of the call with
1142 // uses of the returned value.
1143 if (!TheCall->use_empty()) {
1144 ReturnInst *R = Returns[0];
1145 if (TheCall == R->getReturnValue())
1146 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1148 TheCall->replaceAllUsesWith(R->getReturnValue());
1150 // Since we are now done with the Call/Invoke, we can delete it.
1151 TheCall->eraseFromParent();
1153 // Since we are now done with the return instruction, delete it also.
1154 Returns[0]->eraseFromParent();
1156 // We are now done with the inlining.
1160 // Otherwise, we have the normal case, of more than one block to inline or
1161 // multiple return sites.
1163 // We want to clone the entire callee function into the hole between the
1164 // "starter" and "ender" blocks. How we accomplish this depends on whether
1165 // this is an invoke instruction or a call instruction.
1166 BasicBlock *AfterCallBB;
1167 BranchInst *CreatedBranchToNormalDest = nullptr;
1168 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1170 // Add an unconditional branch to make this look like the CallInst case...
1171 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
1173 // Split the basic block. This guarantees that no PHI nodes will have to be
1174 // updated due to new incoming edges, and make the invoke case more
1175 // symmetric to the call case.
1176 AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest,
1177 CalledFunc->getName()+".exit");
1179 } else { // It's a call
1180 // If this is a call instruction, we need to split the basic block that
1181 // the call lives in.
1183 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
1184 CalledFunc->getName()+".exit");
1187 // Change the branch that used to go to AfterCallBB to branch to the first
1188 // basic block of the inlined function.
1190 TerminatorInst *Br = OrigBB->getTerminator();
1191 assert(Br && Br->getOpcode() == Instruction::Br &&
1192 "splitBasicBlock broken!");
1193 Br->setOperand(0, FirstNewBlock);
1196 // Now that the function is correct, make it a little bit nicer. In
1197 // particular, move the basic blocks inserted from the end of the function
1198 // into the space made by splitting the source basic block.
1199 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
1200 FirstNewBlock, Caller->end());
1202 // Handle all of the return instructions that we just cloned in, and eliminate
1203 // any users of the original call/invoke instruction.
1204 Type *RTy = CalledFunc->getReturnType();
1206 PHINode *PHI = nullptr;
1207 if (Returns.size() > 1) {
1208 // The PHI node should go at the front of the new basic block to merge all
1209 // possible incoming values.
1210 if (!TheCall->use_empty()) {
1211 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
1212 AfterCallBB->begin());
1213 // Anything that used the result of the function call should now use the
1214 // PHI node as their operand.
1215 TheCall->replaceAllUsesWith(PHI);
1218 // Loop over all of the return instructions adding entries to the PHI node
1221 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1222 ReturnInst *RI = Returns[i];
1223 assert(RI->getReturnValue()->getType() == PHI->getType() &&
1224 "Ret value not consistent in function!");
1225 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
1230 // Add a branch to the merge points and remove return instructions.
1232 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1233 ReturnInst *RI = Returns[i];
1234 BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
1235 Loc = RI->getDebugLoc();
1236 BI->setDebugLoc(Loc);
1237 RI->eraseFromParent();
1239 // We need to set the debug location to *somewhere* inside the
1240 // inlined function. The line number may be nonsensical, but the
1241 // instruction will at least be associated with the right
1243 if (CreatedBranchToNormalDest)
1244 CreatedBranchToNormalDest->setDebugLoc(Loc);
1245 } else if (!Returns.empty()) {
1246 // Otherwise, if there is exactly one return value, just replace anything
1247 // using the return value of the call with the computed value.
1248 if (!TheCall->use_empty()) {
1249 if (TheCall == Returns[0]->getReturnValue())
1250 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1252 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
1255 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
1256 BasicBlock *ReturnBB = Returns[0]->getParent();
1257 ReturnBB->replaceAllUsesWith(AfterCallBB);
1259 // Splice the code from the return block into the block that it will return
1260 // to, which contains the code that was after the call.
1261 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
1262 ReturnBB->getInstList());
1264 if (CreatedBranchToNormalDest)
1265 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
1267 // Delete the return instruction now and empty ReturnBB now.
1268 Returns[0]->eraseFromParent();
1269 ReturnBB->eraseFromParent();
1270 } else if (!TheCall->use_empty()) {
1271 // No returns, but something is using the return value of the call. Just
1273 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1276 // Since we are now done with the Call/Invoke, we can delete it.
1277 TheCall->eraseFromParent();
1279 // If we inlined any musttail calls and the original return is now
1280 // unreachable, delete it. It can only contain a bitcast and ret.
1281 if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB))
1282 AfterCallBB->eraseFromParent();
1284 // We should always be able to fold the entry block of the function into the
1285 // single predecessor of the block...
1286 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
1287 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
1289 // Splice the code entry block into calling block, right before the
1290 // unconditional branch.
1291 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
1292 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
1294 // Remove the unconditional branch.
1295 OrigBB->getInstList().erase(Br);
1297 // Now we can remove the CalleeEntry block, which is now empty.
1298 Caller->getBasicBlockList().erase(CalleeEntry);
1300 // If we inserted a phi node, check to see if it has a single value (e.g. all
1301 // the entries are the same or undef). If so, remove the PHI so it doesn't
1302 // block other optimizations.
1304 if (Value *V = SimplifyInstruction(PHI, IFI.DL)) {
1305 PHI->replaceAllUsesWith(V);
1306 PHI->eraseFromParent();