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 SmallPtrSetImpl<LandingPadInst*> &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 SmallPtrSetImpl<LandingPadInst*> &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 (LandingPadInst *InlinedLPad : InlinedLPads) {
251 unsigned OuterNum = OuterLPad->getNumClauses();
252 InlinedLPad->reserveClauses(OuterNum);
253 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx)
254 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx));
255 if (OuterLPad->isCleanup())
256 InlinedLPad->setCleanup(true);
259 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
260 if (InlinedCodeInfo.ContainsCalls)
261 HandleCallsInBlockInlinedThroughInvoke(BB, Invoke);
263 // Forward any resumes that are remaining here.
264 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator()))
265 Invoke.forwardResume(RI, InlinedLPads);
268 // Now that everything is happy, we have one final detail. The PHI nodes in
269 // the exception destination block still have entries due to the original
270 // invoke instruction. Eliminate these entries (which might even delete the
272 InvokeDest->removePredecessor(II->getParent());
275 /// CloneAliasScopeMetadata - When inlining a function that contains noalias
276 /// scope metadata, this metadata needs to be cloned so that the inlined blocks
277 /// have different "unqiue scopes" at every call site. Were this not done, then
278 /// aliasing scopes from a function inlined into a caller multiple times could
279 /// not be differentiated (and this would lead to miscompiles because the
280 /// non-aliasing property communicated by the metadata could have
281 /// call-site-specific control dependencies).
282 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) {
283 const Function *CalledFunc = CS.getCalledFunction();
284 SetVector<const MDNode *> MD;
286 // Note: We could only clone the metadata if it is already used in the
287 // caller. I'm omitting that check here because it might confuse
288 // inter-procedural alias analysis passes. We can revisit this if it becomes
289 // an efficiency or overhead problem.
291 for (Function::const_iterator I = CalledFunc->begin(), IE = CalledFunc->end();
293 for (BasicBlock::const_iterator J = I->begin(), JE = I->end(); J != JE; ++J) {
294 if (const MDNode *M = J->getMetadata(LLVMContext::MD_alias_scope))
296 if (const MDNode *M = J->getMetadata(LLVMContext::MD_noalias))
303 // Walk the existing metadata, adding the complete (perhaps cyclic) chain to
305 SmallVector<const Value *, 16> Queue(MD.begin(), MD.end());
306 while (!Queue.empty()) {
307 const MDNode *M = cast<MDNode>(Queue.pop_back_val());
308 for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i)
309 if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i)))
314 // Now we have a complete set of all metadata in the chains used to specify
315 // the noalias scopes and the lists of those scopes.
316 SmallVector<MDNode *, 16> DummyNodes;
317 DenseMap<const MDNode *, TrackingVH<MDNode> > MDMap;
318 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
320 MDNode *Dummy = MDNode::getTemporary(CalledFunc->getContext(),
322 DummyNodes.push_back(Dummy);
326 // Create new metadata nodes to replace the dummy nodes, replacing old
327 // metadata references with either a dummy node or an already-created new
329 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
331 SmallVector<Value *, 4> NewOps;
332 for (unsigned i = 0, ie = (*I)->getNumOperands(); i != ie; ++i) {
333 const Value *V = (*I)->getOperand(i);
334 if (const MDNode *M = dyn_cast<MDNode>(V))
335 NewOps.push_back(MDMap[M]);
337 NewOps.push_back(const_cast<Value *>(V));
340 MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps),
343 TempM->replaceAllUsesWith(NewM);
346 // Now replace the metadata in the new inlined instructions with the
347 // repacements from the map.
348 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
349 VMI != VMIE; ++VMI) {
353 Instruction *NI = dyn_cast<Instruction>(VMI->second);
357 if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope)) {
358 MDNode *NewMD = MDMap[M];
359 // If the call site also had alias scope metadata (a list of scopes to
360 // which instructions inside it might belong), propagate those scopes to
361 // the inlined instructions.
363 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
364 NewMD = MDNode::concatenate(NewMD, CSM);
365 NI->setMetadata(LLVMContext::MD_alias_scope, NewMD);
366 } else if (NI->mayReadOrWriteMemory()) {
368 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
369 NI->setMetadata(LLVMContext::MD_alias_scope, M);
372 if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias)) {
373 MDNode *NewMD = MDMap[M];
374 // If the call site also had noalias metadata (a list of scopes with
375 // which instructions inside it don't alias), propagate those scopes to
376 // the inlined instructions.
378 CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
379 NewMD = MDNode::concatenate(NewMD, CSM);
380 NI->setMetadata(LLVMContext::MD_noalias, NewMD);
381 } else if (NI->mayReadOrWriteMemory()) {
383 CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
384 NI->setMetadata(LLVMContext::MD_noalias, M);
388 // Now that everything has been replaced, delete the dummy nodes.
389 for (unsigned i = 0, ie = DummyNodes.size(); i != ie; ++i)
390 MDNode::deleteTemporary(DummyNodes[i]);
393 /// AddAliasScopeMetadata - If the inlined function has noalias arguments, then
394 /// add new alias scopes for each noalias argument, tag the mapped noalias
395 /// parameters with noalias metadata specifying the new scope, and tag all
396 /// non-derived loads, stores and memory intrinsics with the new alias scopes.
397 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
398 const DataLayout *DL) {
399 if (!EnableNoAliasConversion)
402 const Function *CalledFunc = CS.getCalledFunction();
403 SmallVector<const Argument *, 4> NoAliasArgs;
405 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
406 E = CalledFunc->arg_end(); I != E; ++I) {
407 if (I->hasNoAliasAttr() && !I->hasNUses(0))
408 NoAliasArgs.push_back(I);
411 if (NoAliasArgs.empty())
414 // To do a good job, if a noalias variable is captured, we need to know if
415 // the capture point dominates the particular use we're considering.
417 DT.recalculate(const_cast<Function&>(*CalledFunc));
419 // noalias indicates that pointer values based on the argument do not alias
420 // pointer values which are not based on it. So we add a new "scope" for each
421 // noalias function argument. Accesses using pointers based on that argument
422 // become part of that alias scope, accesses using pointers not based on that
423 // argument are tagged as noalias with that scope.
425 DenseMap<const Argument *, MDNode *> NewScopes;
426 MDBuilder MDB(CalledFunc->getContext());
428 // Create a new scope domain for this function.
430 MDB.createAnonymousAliasScopeDomain(CalledFunc->getName());
431 for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) {
432 const Argument *A = NoAliasArgs[i];
434 std::string Name = CalledFunc->getName();
437 Name += A->getName();
439 Name += ": argument ";
443 // Note: We always create a new anonymous root here. This is true regardless
444 // of the linkage of the callee because the aliasing "scope" is not just a
445 // property of the callee, but also all control dependencies in the caller.
446 MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
447 NewScopes.insert(std::make_pair(A, NewScope));
450 // Iterate over all new instructions in the map; for all memory-access
451 // instructions, add the alias scope metadata.
452 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
453 VMI != VMIE; ++VMI) {
454 if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) {
458 Instruction *NI = dyn_cast<Instruction>(VMI->second);
462 SmallVector<const Value *, 2> PtrArgs;
464 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
465 PtrArgs.push_back(LI->getPointerOperand());
466 else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
467 PtrArgs.push_back(SI->getPointerOperand());
468 else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
469 PtrArgs.push_back(VAAI->getPointerOperand());
470 else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
471 PtrArgs.push_back(CXI->getPointerOperand());
472 else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
473 PtrArgs.push_back(RMWI->getPointerOperand());
474 else if (ImmutableCallSite ICS = ImmutableCallSite(I)) {
475 // If we know that the call does not access memory, then we'll still
476 // know that about the inlined clone of this call site, and we don't
477 // need to add metadata.
478 if (ICS.doesNotAccessMemory())
481 for (ImmutableCallSite::arg_iterator AI = ICS.arg_begin(),
482 AE = ICS.arg_end(); AI != AE; ++AI)
483 // We need to check the underlying objects of all arguments, not just
484 // the pointer arguments, because we might be passing pointers as
486 // FIXME: If we know that the call only accesses pointer arguments,
487 // then we only need to check the pointer arguments.
488 PtrArgs.push_back(*AI);
491 // If we found no pointers, then this instruction is not suitable for
492 // pairing with an instruction to receive aliasing metadata.
493 // However, if this is a call, this we might just alias with none of the
494 // noalias arguments.
495 if (PtrArgs.empty() && !isa<CallInst>(I) && !isa<InvokeInst>(I))
498 // It is possible that there is only one underlying object, but you
499 // need to go through several PHIs to see it, and thus could be
500 // repeated in the Objects list.
501 SmallPtrSet<const Value *, 4> ObjSet;
502 SmallVector<Value *, 4> Scopes, NoAliases;
504 SmallSetVector<const Argument *, 4> NAPtrArgs;
505 for (unsigned i = 0, ie = PtrArgs.size(); i != ie; ++i) {
506 SmallVector<Value *, 4> Objects;
507 GetUnderlyingObjects(const_cast<Value*>(PtrArgs[i]),
508 Objects, DL, /* MaxLookup = */ 0);
510 for (Value *O : Objects)
514 // Figure out if we're derived from anyhing that is not a noalias
516 bool CanDeriveViaCapture = false;
517 for (const Value *V : ObjSet)
518 if (!isIdentifiedFunctionLocal(const_cast<Value*>(V))) {
519 CanDeriveViaCapture = true;
523 // First, we want to figure out all of the sets with which we definitely
524 // don't alias. Iterate over all noalias set, and add those for which:
525 // 1. The noalias argument is not in the set of objects from which we
526 // definitely derive.
527 // 2. The noalias argument has not yet been captured.
528 for (const Argument *A : NoAliasArgs) {
529 if (!ObjSet.count(A) && (!CanDeriveViaCapture ||
530 A->hasNoCaptureAttr() ||
531 !PointerMayBeCapturedBefore(A,
532 /* ReturnCaptures */ false,
533 /* StoreCaptures */ false, I, &DT)))
534 NoAliases.push_back(NewScopes[A]);
537 if (!NoAliases.empty())
538 NI->setMetadata(LLVMContext::MD_noalias, MDNode::concatenate(
539 NI->getMetadata(LLVMContext::MD_noalias),
540 MDNode::get(CalledFunc->getContext(), NoAliases)));
541 // Next, we want to figure out all of the sets to which we might belong.
542 // We might below to a set if:
543 // 1. The noalias argument is in the set of underlying objects
545 // 2. There is some non-noalias argument in our list and the no-alias
546 // argument has been captured.
548 for (const Argument *A : NoAliasArgs) {
549 if (ObjSet.count(A) || (CanDeriveViaCapture &&
550 PointerMayBeCapturedBefore(A,
551 /* ReturnCaptures */ false,
552 /* StoreCaptures */ false,
554 Scopes.push_back(NewScopes[A]);
558 NI->setMetadata(LLVMContext::MD_alias_scope, MDNode::concatenate(
559 NI->getMetadata(LLVMContext::MD_alias_scope),
560 MDNode::get(CalledFunc->getContext(), Scopes)));
565 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
566 /// into the caller, update the specified callgraph to reflect the changes we
567 /// made. Note that it's possible that not all code was copied over, so only
568 /// some edges of the callgraph may remain.
569 static void UpdateCallGraphAfterInlining(CallSite CS,
570 Function::iterator FirstNewBlock,
571 ValueToValueMapTy &VMap,
572 InlineFunctionInfo &IFI) {
573 CallGraph &CG = *IFI.CG;
574 const Function *Caller = CS.getInstruction()->getParent()->getParent();
575 const Function *Callee = CS.getCalledFunction();
576 CallGraphNode *CalleeNode = CG[Callee];
577 CallGraphNode *CallerNode = CG[Caller];
579 // Since we inlined some uninlined call sites in the callee into the caller,
580 // add edges from the caller to all of the callees of the callee.
581 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
583 // Consider the case where CalleeNode == CallerNode.
584 CallGraphNode::CalledFunctionsVector CallCache;
585 if (CalleeNode == CallerNode) {
586 CallCache.assign(I, E);
587 I = CallCache.begin();
591 for (; I != E; ++I) {
592 const Value *OrigCall = I->first;
594 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
595 // Only copy the edge if the call was inlined!
596 if (VMI == VMap.end() || VMI->second == nullptr)
599 // If the call was inlined, but then constant folded, there is no edge to
600 // add. Check for this case.
601 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
602 if (!NewCall) continue;
604 // Remember that this call site got inlined for the client of
606 IFI.InlinedCalls.push_back(NewCall);
608 // It's possible that inlining the callsite will cause it to go from an
609 // indirect to a direct call by resolving a function pointer. If this
610 // happens, set the callee of the new call site to a more precise
611 // destination. This can also happen if the call graph node of the caller
612 // was just unnecessarily imprecise.
613 if (!I->second->getFunction())
614 if (Function *F = CallSite(NewCall).getCalledFunction()) {
615 // Indirect call site resolved to direct call.
616 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
621 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
624 // Update the call graph by deleting the edge from Callee to Caller. We must
625 // do this after the loop above in case Caller and Callee are the same.
626 CallerNode->removeCallEdgeFor(CS);
629 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
630 BasicBlock *InsertBlock,
631 InlineFunctionInfo &IFI) {
632 LLVMContext &Context = Src->getContext();
633 Type *VoidPtrTy = Type::getInt8PtrTy(Context);
634 Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
635 Type *Tys[3] = { VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context) };
636 Function *MemCpyFn = Intrinsic::getDeclaration(M, Intrinsic::memcpy, Tys);
637 IRBuilder<> builder(InsertBlock->begin());
638 Value *DstCast = builder.CreateBitCast(Dst, VoidPtrTy, "tmp");
639 Value *SrcCast = builder.CreateBitCast(Src, VoidPtrTy, "tmp");
642 if (IFI.DL == nullptr)
643 Size = ConstantExpr::getSizeOf(AggTy);
645 Size = ConstantInt::get(Type::getInt64Ty(Context),
646 IFI.DL->getTypeStoreSize(AggTy));
648 // Always generate a memcpy of alignment 1 here because we don't know
649 // the alignment of the src pointer. Other optimizations can infer
651 Value *CallArgs[] = {
652 DstCast, SrcCast, Size,
653 ConstantInt::get(Type::getInt32Ty(Context), 1),
654 ConstantInt::getFalse(Context) // isVolatile
656 builder.CreateCall(MemCpyFn, CallArgs);
659 /// HandleByValArgument - When inlining a call site that has a byval argument,
660 /// we have to make the implicit memcpy explicit by adding it.
661 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
662 const Function *CalledFunc,
663 InlineFunctionInfo &IFI,
664 unsigned ByValAlignment) {
665 PointerType *ArgTy = cast<PointerType>(Arg->getType());
666 Type *AggTy = ArgTy->getElementType();
668 // If the called function is readonly, then it could not mutate the caller's
669 // copy of the byval'd memory. In this case, it is safe to elide the copy and
671 if (CalledFunc->onlyReadsMemory()) {
672 // If the byval argument has a specified alignment that is greater than the
673 // passed in pointer, then we either have to round up the input pointer or
674 // give up on this transformation.
675 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
678 // If the pointer is already known to be sufficiently aligned, or if we can
679 // round it up to a larger alignment, then we don't need a temporary.
680 if (getOrEnforceKnownAlignment(Arg, ByValAlignment,
681 IFI.DL) >= ByValAlignment)
684 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
685 // for code quality, but rarely happens and is required for correctness.
688 // Create the alloca. If we have DataLayout, use nice alignment.
691 Align = IFI.DL->getPrefTypeAlignment(AggTy);
693 // If the byval had an alignment specified, we *must* use at least that
694 // alignment, as it is required by the byval argument (and uses of the
695 // pointer inside the callee).
696 Align = std::max(Align, ByValAlignment);
698 Function *Caller = TheCall->getParent()->getParent();
700 Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(),
701 &*Caller->begin()->begin());
702 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
704 // Uses of the argument in the function should use our new alloca
709 // isUsedByLifetimeMarker - Check whether this Value is used by a lifetime
711 static bool isUsedByLifetimeMarker(Value *V) {
712 for (User *U : V->users()) {
713 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
714 switch (II->getIntrinsicID()) {
716 case Intrinsic::lifetime_start:
717 case Intrinsic::lifetime_end:
725 // hasLifetimeMarkers - Check whether the given alloca already has
726 // lifetime.start or lifetime.end intrinsics.
727 static bool hasLifetimeMarkers(AllocaInst *AI) {
728 Type *Ty = AI->getType();
729 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
730 Ty->getPointerAddressSpace());
732 return isUsedByLifetimeMarker(AI);
734 // Do a scan to find all the casts to i8*.
735 for (User *U : AI->users()) {
736 if (U->getType() != Int8PtrTy) continue;
737 if (U->stripPointerCasts() != AI) continue;
738 if (isUsedByLifetimeMarker(U))
744 /// updateInlinedAtInfo - Helper function used by fixupLineNumbers to
745 /// recursively update InlinedAtEntry of a DebugLoc.
746 static DebugLoc updateInlinedAtInfo(const DebugLoc &DL,
747 const DebugLoc &InlinedAtDL,
749 if (MDNode *IA = DL.getInlinedAt(Ctx)) {
750 DebugLoc NewInlinedAtDL
751 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx);
752 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
753 NewInlinedAtDL.getAsMDNode(Ctx));
756 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
757 InlinedAtDL.getAsMDNode(Ctx));
760 /// fixupLineNumbers - Update inlined instructions' line numbers to
761 /// to encode location where these instructions are inlined.
762 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
763 Instruction *TheCall) {
764 DebugLoc TheCallDL = TheCall->getDebugLoc();
765 if (TheCallDL.isUnknown())
768 for (; FI != Fn->end(); ++FI) {
769 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
771 DebugLoc DL = BI->getDebugLoc();
772 if (DL.isUnknown()) {
773 // If the inlined instruction has no line number, make it look as if it
774 // originates from the call location. This is important for
775 // ((__always_inline__, __nodebug__)) functions which must use caller
776 // location for all instructions in their function body.
777 BI->setDebugLoc(TheCallDL);
779 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext()));
780 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) {
781 LLVMContext &Ctx = BI->getContext();
782 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
783 DVI->setOperand(2, createInlinedVariable(DVI->getVariable(),
791 /// InlineFunction - This function inlines the called function into the basic
792 /// block of the caller. This returns false if it is not possible to inline
793 /// this call. The program is still in a well defined state if this occurs
796 /// Note that this only does one level of inlining. For example, if the
797 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
798 /// exists in the instruction stream. Similarly this will inline a recursive
799 /// function by one level.
800 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
801 bool InsertLifetime) {
802 Instruction *TheCall = CS.getInstruction();
803 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
804 "Instruction not in function!");
806 // If IFI has any state in it, zap it before we fill it in.
809 const Function *CalledFunc = CS.getCalledFunction();
810 if (!CalledFunc || // Can't inline external function or indirect
811 CalledFunc->isDeclaration() || // call, or call to a vararg function!
812 CalledFunc->getFunctionType()->isVarArg()) return false;
814 // If the call to the callee cannot throw, set the 'nounwind' flag on any
815 // calls that we inline.
816 bool MarkNoUnwind = CS.doesNotThrow();
818 BasicBlock *OrigBB = TheCall->getParent();
819 Function *Caller = OrigBB->getParent();
821 // GC poses two hazards to inlining, which only occur when the callee has GC:
822 // 1. If the caller has no GC, then the callee's GC must be propagated to the
824 // 2. If the caller has a differing GC, it is invalid to inline.
825 if (CalledFunc->hasGC()) {
826 if (!Caller->hasGC())
827 Caller->setGC(CalledFunc->getGC());
828 else if (CalledFunc->getGC() != Caller->getGC())
832 // Get the personality function from the callee if it contains a landing pad.
833 Value *CalleePersonality = nullptr;
834 for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end();
836 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
837 const BasicBlock *BB = II->getUnwindDest();
838 const LandingPadInst *LP = BB->getLandingPadInst();
839 CalleePersonality = LP->getPersonalityFn();
843 // Find the personality function used by the landing pads of the caller. If it
844 // exists, then check to see that it matches the personality function used in
846 if (CalleePersonality) {
847 for (Function::const_iterator I = Caller->begin(), E = Caller->end();
849 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
850 const BasicBlock *BB = II->getUnwindDest();
851 const LandingPadInst *LP = BB->getLandingPadInst();
853 // If the personality functions match, then we can perform the
854 // inlining. Otherwise, we can't inline.
855 // TODO: This isn't 100% true. Some personality functions are proper
856 // supersets of others and can be used in place of the other.
857 if (LP->getPersonalityFn() != CalleePersonality)
864 // Get an iterator to the last basic block in the function, which will have
865 // the new function inlined after it.
866 Function::iterator LastBlock = &Caller->back();
868 // Make sure to capture all of the return instructions from the cloned
870 SmallVector<ReturnInst*, 8> Returns;
871 ClonedCodeInfo InlinedFunctionInfo;
872 Function::iterator FirstNewBlock;
874 { // Scope to destroy VMap after cloning.
875 ValueToValueMapTy VMap;
876 // Keep a list of pair (dst, src) to emit byval initializations.
877 SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
879 assert(CalledFunc->arg_size() == CS.arg_size() &&
880 "No varargs calls can be inlined!");
882 // Calculate the vector of arguments to pass into the function cloner, which
883 // matches up the formal to the actual argument values.
884 CallSite::arg_iterator AI = CS.arg_begin();
886 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
887 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
888 Value *ActualArg = *AI;
890 // When byval arguments actually inlined, we need to make the copy implied
891 // by them explicit. However, we don't do this if the callee is readonly
892 // or readnone, because the copy would be unneeded: the callee doesn't
893 // modify the struct.
894 if (CS.isByValArgument(ArgNo)) {
895 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
896 CalledFunc->getParamAlignment(ArgNo+1));
897 if (ActualArg != *AI)
898 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
904 // We want the inliner to prune the code as it copies. We would LOVE to
905 // have no dead or constant instructions leftover after inlining occurs
906 // (which can happen, e.g., because an argument was constant), but we'll be
907 // happy with whatever the cloner can do.
908 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
909 /*ModuleLevelChanges=*/false, Returns, ".i",
910 &InlinedFunctionInfo, IFI.DL, TheCall);
912 // Remember the first block that is newly cloned over.
913 FirstNewBlock = LastBlock; ++FirstNewBlock;
915 // Inject byval arguments initialization.
916 for (std::pair<Value*, Value*> &Init : ByValInit)
917 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
920 // Update the callgraph if requested.
922 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
924 // Update inlined instructions' line number information.
925 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
927 // Clone existing noalias metadata if necessary.
928 CloneAliasScopeMetadata(CS, VMap);
930 // Add noalias metadata if necessary.
931 AddAliasScopeMetadata(CS, VMap, IFI.DL);
934 // If there are any alloca instructions in the block that used to be the entry
935 // block for the callee, move them to the entry block of the caller. First
936 // calculate which instruction they should be inserted before. We insert the
937 // instructions at the end of the current alloca list.
939 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
940 for (BasicBlock::iterator I = FirstNewBlock->begin(),
941 E = FirstNewBlock->end(); I != E; ) {
942 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
945 // If the alloca is now dead, remove it. This often occurs due to code
947 if (AI->use_empty()) {
948 AI->eraseFromParent();
952 if (!isa<Constant>(AI->getArraySize()))
955 // Keep track of the static allocas that we inline into the caller.
956 IFI.StaticAllocas.push_back(AI);
958 // Scan for the block of allocas that we can move over, and move them
960 while (isa<AllocaInst>(I) &&
961 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
962 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
966 // Transfer all of the allocas over in a block. Using splice means
967 // that the instructions aren't removed from the symbol table, then
969 Caller->getEntryBlock().getInstList().splice(InsertPoint,
970 FirstNewBlock->getInstList(),
975 bool InlinedMustTailCalls = false;
976 if (InlinedFunctionInfo.ContainsCalls) {
977 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
978 if (CallInst *CI = dyn_cast<CallInst>(TheCall))
979 CallSiteTailKind = CI->getTailCallKind();
981 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
983 for (Instruction &I : *BB) {
984 CallInst *CI = dyn_cast<CallInst>(&I);
988 // We need to reduce the strength of any inlined tail calls. For
989 // musttail, we have to avoid introducing potential unbounded stack
990 // growth. For example, if functions 'f' and 'g' are mutually recursive
991 // with musttail, we can inline 'g' into 'f' so long as we preserve
992 // musttail on the cloned call to 'f'. If either the inlined call site
993 // or the cloned call site is *not* musttail, the program already has
994 // one frame of stack growth, so it's safe to remove musttail. Here is
995 // a table of example transformations:
997 // f -> musttail g -> musttail f ==> f -> musttail f
998 // f -> musttail g -> tail f ==> f -> tail f
999 // f -> g -> musttail f ==> f -> f
1000 // f -> g -> tail f ==> f -> f
1001 CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
1002 ChildTCK = std::min(CallSiteTailKind, ChildTCK);
1003 CI->setTailCallKind(ChildTCK);
1004 InlinedMustTailCalls |= CI->isMustTailCall();
1006 // Calls inlined through a 'nounwind' call site should be marked
1009 CI->setDoesNotThrow();
1014 // Leave lifetime markers for the static alloca's, scoping them to the
1015 // function we just inlined.
1016 if (InsertLifetime && !IFI.StaticAllocas.empty()) {
1017 IRBuilder<> builder(FirstNewBlock->begin());
1018 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
1019 AllocaInst *AI = IFI.StaticAllocas[ai];
1021 // If the alloca is already scoped to something smaller than the whole
1022 // function then there's no need to add redundant, less accurate markers.
1023 if (hasLifetimeMarkers(AI))
1026 // Try to determine the size of the allocation.
1027 ConstantInt *AllocaSize = nullptr;
1028 if (ConstantInt *AIArraySize =
1029 dyn_cast<ConstantInt>(AI->getArraySize())) {
1031 Type *AllocaType = AI->getAllocatedType();
1032 uint64_t AllocaTypeSize = IFI.DL->getTypeAllocSize(AllocaType);
1033 uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
1034 assert(AllocaArraySize > 0 && "array size of AllocaInst is zero");
1035 // Check that array size doesn't saturate uint64_t and doesn't
1036 // overflow when it's multiplied by type size.
1037 if (AllocaArraySize != ~0ULL &&
1038 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
1039 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
1040 AllocaArraySize * AllocaTypeSize);
1045 builder.CreateLifetimeStart(AI, AllocaSize);
1046 for (ReturnInst *RI : Returns) {
1047 // Don't insert llvm.lifetime.end calls between a musttail call and a
1048 // return. The return kills all local allocas.
1049 if (InlinedMustTailCalls &&
1050 RI->getParent()->getTerminatingMustTailCall())
1052 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
1057 // If the inlined code contained dynamic alloca instructions, wrap the inlined
1058 // code with llvm.stacksave/llvm.stackrestore intrinsics.
1059 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
1060 Module *M = Caller->getParent();
1061 // Get the two intrinsics we care about.
1062 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
1063 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
1065 // Insert the llvm.stacksave.
1066 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
1067 .CreateCall(StackSave, "savedstack");
1069 // Insert a call to llvm.stackrestore before any return instructions in the
1070 // inlined function.
1071 for (ReturnInst *RI : Returns) {
1072 // Don't insert llvm.stackrestore calls between a musttail call and a
1073 // return. The return will restore the stack pointer.
1074 if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
1076 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
1080 // If we are inlining for an invoke instruction, we must make sure to rewrite
1081 // any call instructions into invoke instructions.
1082 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
1083 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
1085 // Handle any inlined musttail call sites. In order for a new call site to be
1086 // musttail, the source of the clone and the inlined call site must have been
1087 // musttail. Therefore it's safe to return without merging control into the
1089 if (InlinedMustTailCalls) {
1090 // Check if we need to bitcast the result of any musttail calls.
1091 Type *NewRetTy = Caller->getReturnType();
1092 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
1094 // Handle the returns preceded by musttail calls separately.
1095 SmallVector<ReturnInst *, 8> NormalReturns;
1096 for (ReturnInst *RI : Returns) {
1097 CallInst *ReturnedMustTail =
1098 RI->getParent()->getTerminatingMustTailCall();
1099 if (!ReturnedMustTail) {
1100 NormalReturns.push_back(RI);
1106 // Delete the old return and any preceding bitcast.
1107 BasicBlock *CurBB = RI->getParent();
1108 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
1109 RI->eraseFromParent();
1111 OldCast->eraseFromParent();
1113 // Insert a new bitcast and return with the right type.
1114 IRBuilder<> Builder(CurBB);
1115 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
1118 // Leave behind the normal returns so we can merge control flow.
1119 std::swap(Returns, NormalReturns);
1122 // If we cloned in _exactly one_ basic block, and if that block ends in a
1123 // return instruction, we splice the body of the inlined callee directly into
1124 // the calling basic block.
1125 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
1126 // Move all of the instructions right before the call.
1127 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
1128 FirstNewBlock->begin(), FirstNewBlock->end());
1129 // Remove the cloned basic block.
1130 Caller->getBasicBlockList().pop_back();
1132 // If the call site was an invoke instruction, add a branch to the normal
1134 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1135 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
1136 NewBr->setDebugLoc(Returns[0]->getDebugLoc());
1139 // If the return instruction returned a value, replace uses of the call with
1140 // uses of the returned value.
1141 if (!TheCall->use_empty()) {
1142 ReturnInst *R = Returns[0];
1143 if (TheCall == R->getReturnValue())
1144 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1146 TheCall->replaceAllUsesWith(R->getReturnValue());
1148 // Since we are now done with the Call/Invoke, we can delete it.
1149 TheCall->eraseFromParent();
1151 // Since we are now done with the return instruction, delete it also.
1152 Returns[0]->eraseFromParent();
1154 // We are now done with the inlining.
1158 // Otherwise, we have the normal case, of more than one block to inline or
1159 // multiple return sites.
1161 // We want to clone the entire callee function into the hole between the
1162 // "starter" and "ender" blocks. How we accomplish this depends on whether
1163 // this is an invoke instruction or a call instruction.
1164 BasicBlock *AfterCallBB;
1165 BranchInst *CreatedBranchToNormalDest = nullptr;
1166 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1168 // Add an unconditional branch to make this look like the CallInst case...
1169 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
1171 // Split the basic block. This guarantees that no PHI nodes will have to be
1172 // updated due to new incoming edges, and make the invoke case more
1173 // symmetric to the call case.
1174 AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest,
1175 CalledFunc->getName()+".exit");
1177 } else { // It's a call
1178 // If this is a call instruction, we need to split the basic block that
1179 // the call lives in.
1181 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
1182 CalledFunc->getName()+".exit");
1185 // Change the branch that used to go to AfterCallBB to branch to the first
1186 // basic block of the inlined function.
1188 TerminatorInst *Br = OrigBB->getTerminator();
1189 assert(Br && Br->getOpcode() == Instruction::Br &&
1190 "splitBasicBlock broken!");
1191 Br->setOperand(0, FirstNewBlock);
1194 // Now that the function is correct, make it a little bit nicer. In
1195 // particular, move the basic blocks inserted from the end of the function
1196 // into the space made by splitting the source basic block.
1197 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
1198 FirstNewBlock, Caller->end());
1200 // Handle all of the return instructions that we just cloned in, and eliminate
1201 // any users of the original call/invoke instruction.
1202 Type *RTy = CalledFunc->getReturnType();
1204 PHINode *PHI = nullptr;
1205 if (Returns.size() > 1) {
1206 // The PHI node should go at the front of the new basic block to merge all
1207 // possible incoming values.
1208 if (!TheCall->use_empty()) {
1209 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
1210 AfterCallBB->begin());
1211 // Anything that used the result of the function call should now use the
1212 // PHI node as their operand.
1213 TheCall->replaceAllUsesWith(PHI);
1216 // Loop over all of the return instructions adding entries to the PHI node
1219 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1220 ReturnInst *RI = Returns[i];
1221 assert(RI->getReturnValue()->getType() == PHI->getType() &&
1222 "Ret value not consistent in function!");
1223 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
1228 // Add a branch to the merge points and remove return instructions.
1230 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1231 ReturnInst *RI = Returns[i];
1232 BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
1233 Loc = RI->getDebugLoc();
1234 BI->setDebugLoc(Loc);
1235 RI->eraseFromParent();
1237 // We need to set the debug location to *somewhere* inside the
1238 // inlined function. The line number may be nonsensical, but the
1239 // instruction will at least be associated with the right
1241 if (CreatedBranchToNormalDest)
1242 CreatedBranchToNormalDest->setDebugLoc(Loc);
1243 } else if (!Returns.empty()) {
1244 // Otherwise, if there is exactly one return value, just replace anything
1245 // using the return value of the call with the computed value.
1246 if (!TheCall->use_empty()) {
1247 if (TheCall == Returns[0]->getReturnValue())
1248 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1250 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
1253 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
1254 BasicBlock *ReturnBB = Returns[0]->getParent();
1255 ReturnBB->replaceAllUsesWith(AfterCallBB);
1257 // Splice the code from the return block into the block that it will return
1258 // to, which contains the code that was after the call.
1259 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
1260 ReturnBB->getInstList());
1262 if (CreatedBranchToNormalDest)
1263 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
1265 // Delete the return instruction now and empty ReturnBB now.
1266 Returns[0]->eraseFromParent();
1267 ReturnBB->eraseFromParent();
1268 } else if (!TheCall->use_empty()) {
1269 // No returns, but something is using the return value of the call. Just
1271 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1274 // Since we are now done with the Call/Invoke, we can delete it.
1275 TheCall->eraseFromParent();
1277 // If we inlined any musttail calls and the original return is now
1278 // unreachable, delete it. It can only contain a bitcast and ret.
1279 if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB))
1280 AfterCallBB->eraseFromParent();
1282 // We should always be able to fold the entry block of the function into the
1283 // single predecessor of the block...
1284 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
1285 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
1287 // Splice the code entry block into calling block, right before the
1288 // unconditional branch.
1289 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
1290 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
1292 // Remove the unconditional branch.
1293 OrigBB->getInstList().erase(Br);
1295 // Now we can remove the CalleeEntry block, which is now empty.
1296 Caller->getBasicBlockList().erase(CalleeEntry);
1298 // If we inserted a phi node, check to see if it has a single value (e.g. all
1299 // the entries are the same or undef). If so, remove the PHI so it doesn't
1300 // block other optimizations.
1302 if (Value *V = SimplifyInstruction(PHI, IFI.DL)) {
1303 PHI->replaceAllUsesWith(V);
1304 PHI->eraseFromParent();