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/AssumptionCache.h"
22 #include "llvm/Analysis/CallGraph.h"
23 #include "llvm/Analysis/CaptureTracking.h"
24 #include "llvm/Analysis/InstructionSimplify.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/IR/Attributes.h"
27 #include "llvm/IR/CallSite.h"
28 #include "llvm/IR/CFG.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DebugInfo.h"
32 #include "llvm/IR/DerivedTypes.h"
33 #include "llvm/IR/DIBuilder.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/IRBuilder.h"
36 #include "llvm/IR/Instructions.h"
37 #include "llvm/IR/IntrinsicInst.h"
38 #include "llvm/IR/Intrinsics.h"
39 #include "llvm/IR/MDBuilder.h"
40 #include "llvm/IR/Module.h"
41 #include "llvm/Transforms/Utils/Local.h"
42 #include "llvm/Support/CommandLine.h"
47 EnableNoAliasConversion("enable-noalias-to-md-conversion", cl::init(true),
49 cl::desc("Convert noalias attributes to metadata during inlining."));
52 PreserveAlignmentAssumptions("preserve-alignment-assumptions-during-inlining",
53 cl::init(true), cl::Hidden,
54 cl::desc("Convert align attributes to assumptions during inlining."));
56 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI,
57 bool InsertLifetime) {
58 return InlineFunction(CallSite(CI), IFI, InsertLifetime);
60 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI,
61 bool InsertLifetime) {
62 return InlineFunction(CallSite(II), IFI, InsertLifetime);
66 /// A class for recording information about inlining through an invoke.
67 class InvokeInliningInfo {
68 BasicBlock *OuterResumeDest; ///< Destination of the invoke's unwind.
69 BasicBlock *InnerResumeDest; ///< Destination for the callee's resume.
70 LandingPadInst *CallerLPad; ///< LandingPadInst associated with the invoke.
71 PHINode *InnerEHValuesPHI; ///< PHI for EH values from landingpad insts.
72 SmallVector<Value*, 8> UnwindDestPHIValues;
75 InvokeInliningInfo(InvokeInst *II)
76 : OuterResumeDest(II->getUnwindDest()), InnerResumeDest(nullptr),
77 CallerLPad(nullptr), InnerEHValuesPHI(nullptr) {
78 // If there are PHI nodes in the unwind destination block, we need to keep
79 // track of which values came into them from the invoke before removing
80 // the edge from this block.
81 llvm::BasicBlock *InvokeBB = II->getParent();
82 BasicBlock::iterator I = OuterResumeDest->begin();
83 for (; isa<PHINode>(I); ++I) {
84 // Save the value to use for this edge.
85 PHINode *PHI = cast<PHINode>(I);
86 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
89 CallerLPad = cast<LandingPadInst>(I);
92 /// getOuterResumeDest - The outer unwind destination is the target of
93 /// unwind edges introduced for calls within the inlined function.
94 BasicBlock *getOuterResumeDest() const {
95 return OuterResumeDest;
98 BasicBlock *getInnerResumeDest();
100 LandingPadInst *getLandingPadInst() const { return CallerLPad; }
102 /// forwardResume - Forward the 'resume' instruction to the caller's landing
103 /// pad block. When the landing pad block has only one predecessor, this is
104 /// a simple branch. When there is more than one predecessor, we need to
105 /// split the landing pad block after the landingpad instruction and jump
107 void forwardResume(ResumeInst *RI,
108 SmallPtrSetImpl<LandingPadInst*> &InlinedLPads);
110 /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind
111 /// destination block for the given basic block, using the values for the
112 /// original invoke's source block.
113 void addIncomingPHIValuesFor(BasicBlock *BB) const {
114 addIncomingPHIValuesForInto(BB, OuterResumeDest);
117 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const {
118 BasicBlock::iterator I = dest->begin();
119 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
120 PHINode *phi = cast<PHINode>(I);
121 phi->addIncoming(UnwindDestPHIValues[i], src);
127 /// getInnerResumeDest - Get or create a target for the branch from ResumeInsts.
128 BasicBlock *InvokeInliningInfo::getInnerResumeDest() {
129 if (InnerResumeDest) return InnerResumeDest;
131 // Split the landing pad.
132 BasicBlock::iterator SplitPoint = CallerLPad; ++SplitPoint;
134 OuterResumeDest->splitBasicBlock(SplitPoint,
135 OuterResumeDest->getName() + ".body");
137 // The number of incoming edges we expect to the inner landing pad.
138 const unsigned PHICapacity = 2;
140 // Create corresponding new PHIs for all the PHIs in the outer landing pad.
141 BasicBlock::iterator InsertPoint = InnerResumeDest->begin();
142 BasicBlock::iterator I = OuterResumeDest->begin();
143 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
144 PHINode *OuterPHI = cast<PHINode>(I);
145 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity,
146 OuterPHI->getName() + ".lpad-body",
148 OuterPHI->replaceAllUsesWith(InnerPHI);
149 InnerPHI->addIncoming(OuterPHI, OuterResumeDest);
152 // Create a PHI for the exception values.
153 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity,
154 "eh.lpad-body", InsertPoint);
155 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI);
156 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest);
159 return InnerResumeDest;
162 /// forwardResume - Forward the 'resume' instruction to the caller's landing pad
163 /// block. When the landing pad block has only one predecessor, this is a simple
164 /// branch. When there is more than one predecessor, we need to split the
165 /// landing pad block after the landingpad instruction and jump to there.
166 void InvokeInliningInfo::forwardResume(ResumeInst *RI,
167 SmallPtrSetImpl<LandingPadInst*> &InlinedLPads) {
168 BasicBlock *Dest = getInnerResumeDest();
169 BasicBlock *Src = RI->getParent();
171 BranchInst::Create(Dest, Src);
173 // Update the PHIs in the destination. They were inserted in an order which
175 addIncomingPHIValuesForInto(Src, Dest);
177 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src);
178 RI->eraseFromParent();
181 /// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
182 /// an invoke, we have to turn all of the calls that can throw into
183 /// invokes. This function analyze BB to see if there are any calls, and if so,
184 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
185 /// nodes in that block with the values specified in InvokeDestPHIValues.
186 static void HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
187 InvokeInliningInfo &Invoke) {
188 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
189 Instruction *I = BBI++;
191 // We only need to check for function calls: inlined invoke
192 // instructions require no special handling.
193 CallInst *CI = dyn_cast<CallInst>(I);
195 // If this call cannot unwind, don't convert it to an invoke.
196 // Inline asm calls cannot throw.
197 if (!CI || CI->doesNotThrow() || isa<InlineAsm>(CI->getCalledValue()))
200 // Convert this function call into an invoke instruction. First, split the
202 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
204 // Delete the unconditional branch inserted by splitBasicBlock
205 BB->getInstList().pop_back();
207 // Create the new invoke instruction.
208 ImmutableCallSite CS(CI);
209 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
210 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split,
211 Invoke.getOuterResumeDest(),
212 InvokeArgs, CI->getName(), BB);
213 II->setDebugLoc(CI->getDebugLoc());
214 II->setCallingConv(CI->getCallingConv());
215 II->setAttributes(CI->getAttributes());
217 // Make sure that anything using the call now uses the invoke! This also
218 // updates the CallGraph if present, because it uses a WeakVH.
219 CI->replaceAllUsesWith(II);
221 // Delete the original call
222 Split->getInstList().pop_front();
224 // Update any PHI nodes in the exceptional block to indicate that there is
225 // now a new entry in them.
226 Invoke.addIncomingPHIValuesFor(BB);
231 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
232 /// in the body of the inlined function into invokes.
234 /// II is the invoke instruction being inlined. FirstNewBlock is the first
235 /// block of the inlined code (the last block is the end of the function),
236 /// and InlineCodeInfo is information about the code that got inlined.
237 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
238 ClonedCodeInfo &InlinedCodeInfo) {
239 BasicBlock *InvokeDest = II->getUnwindDest();
241 Function *Caller = FirstNewBlock->getParent();
243 // The inlined code is currently at the end of the function, scan from the
244 // start of the inlined code to its end, checking for stuff we need to
246 InvokeInliningInfo Invoke(II);
248 // Get all of the inlined landing pad instructions.
249 SmallPtrSet<LandingPadInst*, 16> InlinedLPads;
250 for (Function::iterator I = FirstNewBlock, E = Caller->end(); I != E; ++I)
251 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator()))
252 InlinedLPads.insert(II->getLandingPadInst());
254 // Append the clauses from the outer landing pad instruction into the inlined
255 // landing pad instructions.
256 LandingPadInst *OuterLPad = Invoke.getLandingPadInst();
257 for (LandingPadInst *InlinedLPad : InlinedLPads) {
258 unsigned OuterNum = OuterLPad->getNumClauses();
259 InlinedLPad->reserveClauses(OuterNum);
260 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx)
261 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx));
262 if (OuterLPad->isCleanup())
263 InlinedLPad->setCleanup(true);
266 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
267 if (InlinedCodeInfo.ContainsCalls)
268 HandleCallsInBlockInlinedThroughInvoke(BB, Invoke);
270 // Forward any resumes that are remaining here.
271 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator()))
272 Invoke.forwardResume(RI, InlinedLPads);
275 // Now that everything is happy, we have one final detail. The PHI nodes in
276 // the exception destination block still have entries due to the original
277 // invoke instruction. Eliminate these entries (which might even delete the
279 InvokeDest->removePredecessor(II->getParent());
282 /// CloneAliasScopeMetadata - When inlining a function that contains noalias
283 /// scope metadata, this metadata needs to be cloned so that the inlined blocks
284 /// have different "unqiue scopes" at every call site. Were this not done, then
285 /// aliasing scopes from a function inlined into a caller multiple times could
286 /// not be differentiated (and this would lead to miscompiles because the
287 /// non-aliasing property communicated by the metadata could have
288 /// call-site-specific control dependencies).
289 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) {
290 const Function *CalledFunc = CS.getCalledFunction();
291 SetVector<const MDNode *> MD;
293 // Note: We could only clone the metadata if it is already used in the
294 // caller. I'm omitting that check here because it might confuse
295 // inter-procedural alias analysis passes. We can revisit this if it becomes
296 // an efficiency or overhead problem.
298 for (Function::const_iterator I = CalledFunc->begin(), IE = CalledFunc->end();
300 for (BasicBlock::const_iterator J = I->begin(), JE = I->end(); J != JE; ++J) {
301 if (const MDNode *M = J->getMetadata(LLVMContext::MD_alias_scope))
303 if (const MDNode *M = J->getMetadata(LLVMContext::MD_noalias))
310 // Walk the existing metadata, adding the complete (perhaps cyclic) chain to
312 SmallVector<const Metadata *, 16> Queue(MD.begin(), MD.end());
313 while (!Queue.empty()) {
314 const MDNode *M = cast<MDNode>(Queue.pop_back_val());
315 for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i)
316 if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i)))
321 // Now we have a complete set of all metadata in the chains used to specify
322 // the noalias scopes and the lists of those scopes.
323 SmallVector<TempMDTuple, 16> DummyNodes;
324 DenseMap<const MDNode *, TrackingMDNodeRef> MDMap;
325 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
327 DummyNodes.push_back(MDTuple::getTemporary(CalledFunc->getContext(), None));
328 MDMap[*I].reset(DummyNodes.back().get());
331 // Create new metadata nodes to replace the dummy nodes, replacing old
332 // metadata references with either a dummy node or an already-created new
334 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
336 SmallVector<Metadata *, 4> NewOps;
337 for (unsigned i = 0, ie = (*I)->getNumOperands(); i != ie; ++i) {
338 const Metadata *V = (*I)->getOperand(i);
339 if (const MDNode *M = dyn_cast<MDNode>(V))
340 NewOps.push_back(MDMap[M]);
342 NewOps.push_back(const_cast<Metadata *>(V));
345 MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps);
346 MDTuple *TempM = cast<MDTuple>(MDMap[*I]);
347 assert(TempM->isTemporary() && "Expected temporary node");
349 TempM->replaceAllUsesWith(NewM);
352 // Now replace the metadata in the new inlined instructions with the
353 // repacements from the map.
354 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
355 VMI != VMIE; ++VMI) {
359 Instruction *NI = dyn_cast<Instruction>(VMI->second);
363 if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope)) {
364 MDNode *NewMD = MDMap[M];
365 // If the call site also had alias scope metadata (a list of scopes to
366 // which instructions inside it might belong), propagate those scopes to
367 // the inlined instructions.
369 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
370 NewMD = MDNode::concatenate(NewMD, CSM);
371 NI->setMetadata(LLVMContext::MD_alias_scope, NewMD);
372 } else if (NI->mayReadOrWriteMemory()) {
374 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
375 NI->setMetadata(LLVMContext::MD_alias_scope, M);
378 if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias)) {
379 MDNode *NewMD = MDMap[M];
380 // If the call site also had noalias metadata (a list of scopes with
381 // which instructions inside it don't alias), propagate those scopes to
382 // the inlined instructions.
384 CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
385 NewMD = MDNode::concatenate(NewMD, CSM);
386 NI->setMetadata(LLVMContext::MD_noalias, NewMD);
387 } else if (NI->mayReadOrWriteMemory()) {
388 if (MDNode *M = CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
389 NI->setMetadata(LLVMContext::MD_noalias, M);
394 /// AddAliasScopeMetadata - If the inlined function has noalias arguments, then
395 /// add new alias scopes for each noalias argument, tag the mapped noalias
396 /// parameters with noalias metadata specifying the new scope, and tag all
397 /// non-derived loads, stores and memory intrinsics with the new alias scopes.
398 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
399 const DataLayout *DL, AliasAnalysis *AA) {
400 if (!EnableNoAliasConversion)
403 const Function *CalledFunc = CS.getCalledFunction();
404 SmallVector<const Argument *, 4> NoAliasArgs;
406 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
407 E = CalledFunc->arg_end(); I != E; ++I) {
408 if (I->hasNoAliasAttr() && !I->hasNUses(0))
409 NoAliasArgs.push_back(I);
412 if (NoAliasArgs.empty())
415 // To do a good job, if a noalias variable is captured, we need to know if
416 // the capture point dominates the particular use we're considering.
418 DT.recalculate(const_cast<Function&>(*CalledFunc));
420 // noalias indicates that pointer values based on the argument do not alias
421 // pointer values which are not based on it. So we add a new "scope" for each
422 // noalias function argument. Accesses using pointers based on that argument
423 // become part of that alias scope, accesses using pointers not based on that
424 // argument are tagged as noalias with that scope.
426 DenseMap<const Argument *, MDNode *> NewScopes;
427 MDBuilder MDB(CalledFunc->getContext());
429 // Create a new scope domain for this function.
431 MDB.createAnonymousAliasScopeDomain(CalledFunc->getName());
432 for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) {
433 const Argument *A = NoAliasArgs[i];
435 std::string Name = CalledFunc->getName();
438 Name += A->getName();
440 Name += ": argument ";
444 // Note: We always create a new anonymous root here. This is true regardless
445 // of the linkage of the callee because the aliasing "scope" is not just a
446 // property of the callee, but also all control dependencies in the caller.
447 MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
448 NewScopes.insert(std::make_pair(A, NewScope));
451 // Iterate over all new instructions in the map; for all memory-access
452 // instructions, add the alias scope metadata.
453 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
454 VMI != VMIE; ++VMI) {
455 if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) {
459 Instruction *NI = dyn_cast<Instruction>(VMI->second);
463 bool IsArgMemOnlyCall = false, IsFuncCall = false;
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())
485 AliasAnalysis::ModRefBehavior MRB = AA->getModRefBehavior(ICS);
486 if (MRB == AliasAnalysis::OnlyAccessesArgumentPointees ||
487 MRB == AliasAnalysis::OnlyReadsArgumentPointees)
488 IsArgMemOnlyCall = true;
491 for (ImmutableCallSite::arg_iterator AI = ICS.arg_begin(),
492 AE = ICS.arg_end(); AI != AE; ++AI) {
493 // We need to check the underlying objects of all arguments, not just
494 // the pointer arguments, because we might be passing pointers as
496 // However, if we know that the call only accesses pointer arguments,
497 // then we only need to check the pointer arguments.
498 if (IsArgMemOnlyCall && !(*AI)->getType()->isPointerTy())
501 PtrArgs.push_back(*AI);
505 // If we found no pointers, then this instruction is not suitable for
506 // pairing with an instruction to receive aliasing metadata.
507 // However, if this is a call, this we might just alias with none of the
508 // noalias arguments.
509 if (PtrArgs.empty() && !IsFuncCall)
512 // It is possible that there is only one underlying object, but you
513 // need to go through several PHIs to see it, and thus could be
514 // repeated in the Objects list.
515 SmallPtrSet<const Value *, 4> ObjSet;
516 SmallVector<Metadata *, 4> Scopes, NoAliases;
518 SmallSetVector<const Argument *, 4> NAPtrArgs;
519 for (unsigned i = 0, ie = PtrArgs.size(); i != ie; ++i) {
520 SmallVector<Value *, 4> Objects;
521 GetUnderlyingObjects(const_cast<Value*>(PtrArgs[i]),
522 Objects, DL, /* MaxLookup = */ 0);
524 for (Value *O : Objects)
528 // Figure out if we're derived from anything that is not a noalias
530 bool CanDeriveViaCapture = false, UsesAliasingPtr = false;
531 for (const Value *V : ObjSet) {
532 // Is this value a constant that cannot be derived from any pointer
533 // value (we need to exclude constant expressions, for example, that
534 // are formed from arithmetic on global symbols).
535 bool IsNonPtrConst = isa<ConstantInt>(V) || isa<ConstantFP>(V) ||
536 isa<ConstantPointerNull>(V) ||
537 isa<ConstantDataVector>(V) || isa<UndefValue>(V);
541 // If this is anything other than a noalias argument, then we cannot
542 // completely describe the aliasing properties using alias.scope
543 // metadata (and, thus, won't add any).
544 if (const Argument *A = dyn_cast<Argument>(V)) {
545 if (!A->hasNoAliasAttr())
546 UsesAliasingPtr = true;
548 UsesAliasingPtr = true;
551 // If this is not some identified function-local object (which cannot
552 // directly alias a noalias argument), or some other argument (which,
553 // by definition, also cannot alias a noalias argument), then we could
554 // alias a noalias argument that has been captured).
555 if (!isa<Argument>(V) &&
556 !isIdentifiedFunctionLocal(const_cast<Value*>(V)))
557 CanDeriveViaCapture = true;
560 // A function call can always get captured noalias pointers (via other
561 // parameters, globals, etc.).
562 if (IsFuncCall && !IsArgMemOnlyCall)
563 CanDeriveViaCapture = true;
565 // First, we want to figure out all of the sets with which we definitely
566 // don't alias. Iterate over all noalias set, and add those for which:
567 // 1. The noalias argument is not in the set of objects from which we
568 // definitely derive.
569 // 2. The noalias argument has not yet been captured.
570 // An arbitrary function that might load pointers could see captured
571 // noalias arguments via other noalias arguments or globals, and so we
572 // must always check for prior capture.
573 for (const Argument *A : NoAliasArgs) {
574 if (!ObjSet.count(A) && (!CanDeriveViaCapture ||
575 // It might be tempting to skip the
576 // PointerMayBeCapturedBefore check if
577 // A->hasNoCaptureAttr() is true, but this is
578 // incorrect because nocapture only guarantees
579 // that no copies outlive the function, not
580 // that the value cannot be locally captured.
581 !PointerMayBeCapturedBefore(A,
582 /* ReturnCaptures */ false,
583 /* StoreCaptures */ false, I, &DT)))
584 NoAliases.push_back(NewScopes[A]);
587 if (!NoAliases.empty())
588 NI->setMetadata(LLVMContext::MD_noalias,
590 NI->getMetadata(LLVMContext::MD_noalias),
591 MDNode::get(CalledFunc->getContext(), NoAliases)));
593 // Next, we want to figure out all of the sets to which we might belong.
594 // We might belong to a set if the noalias argument is in the set of
595 // underlying objects. If there is some non-noalias argument in our list
596 // of underlying objects, then we cannot add a scope because the fact
597 // that some access does not alias with any set of our noalias arguments
598 // cannot itself guarantee that it does not alias with this access
599 // (because there is some pointer of unknown origin involved and the
600 // other access might also depend on this pointer). We also cannot add
601 // scopes to arbitrary functions unless we know they don't access any
602 // non-parameter pointer-values.
603 bool CanAddScopes = !UsesAliasingPtr;
604 if (CanAddScopes && IsFuncCall)
605 CanAddScopes = IsArgMemOnlyCall;
608 for (const Argument *A : NoAliasArgs) {
610 Scopes.push_back(NewScopes[A]);
615 LLVMContext::MD_alias_scope,
616 MDNode::concatenate(NI->getMetadata(LLVMContext::MD_alias_scope),
617 MDNode::get(CalledFunc->getContext(), Scopes)));
622 /// If the inlined function has non-byval align arguments, then
623 /// add @llvm.assume-based alignment assumptions to preserve this information.
624 static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) {
625 if (!PreserveAlignmentAssumptions)
627 auto &DL = CS.getCaller()->getParent()->getDataLayout();
629 // To avoid inserting redundant assumptions, we should check for assumptions
630 // already in the caller. To do this, we might need a DT of the caller.
632 bool DTCalculated = false;
634 Function *CalledFunc = CS.getCalledFunction();
635 for (Function::arg_iterator I = CalledFunc->arg_begin(),
636 E = CalledFunc->arg_end();
638 unsigned Align = I->getType()->isPointerTy() ? I->getParamAlignment() : 0;
639 if (Align && !I->hasByValOrInAllocaAttr() && !I->hasNUses(0)) {
641 DT.recalculate(const_cast<Function&>(*CS.getInstruction()->getParent()
646 // If we can already prove the asserted alignment in the context of the
647 // caller, then don't bother inserting the assumption.
648 Value *Arg = CS.getArgument(I->getArgNo());
649 if (getKnownAlignment(Arg, &DL, &IFI.ACT->getAssumptionCache(*CalledFunc),
650 CS.getInstruction(), &DT) >= Align)
653 IRBuilder<>(CS.getInstruction())
654 .CreateAlignmentAssumption(DL, Arg, Align);
659 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
660 /// into the caller, update the specified callgraph to reflect the changes we
661 /// made. Note that it's possible that not all code was copied over, so only
662 /// some edges of the callgraph may remain.
663 static void UpdateCallGraphAfterInlining(CallSite CS,
664 Function::iterator FirstNewBlock,
665 ValueToValueMapTy &VMap,
666 InlineFunctionInfo &IFI) {
667 CallGraph &CG = *IFI.CG;
668 const Function *Caller = CS.getInstruction()->getParent()->getParent();
669 const Function *Callee = CS.getCalledFunction();
670 CallGraphNode *CalleeNode = CG[Callee];
671 CallGraphNode *CallerNode = CG[Caller];
673 // Since we inlined some uninlined call sites in the callee into the caller,
674 // add edges from the caller to all of the callees of the callee.
675 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
677 // Consider the case where CalleeNode == CallerNode.
678 CallGraphNode::CalledFunctionsVector CallCache;
679 if (CalleeNode == CallerNode) {
680 CallCache.assign(I, E);
681 I = CallCache.begin();
685 for (; I != E; ++I) {
686 const Value *OrigCall = I->first;
688 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
689 // Only copy the edge if the call was inlined!
690 if (VMI == VMap.end() || VMI->second == nullptr)
693 // If the call was inlined, but then constant folded, there is no edge to
694 // add. Check for this case.
695 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
696 if (!NewCall) continue;
698 // Remember that this call site got inlined for the client of
700 IFI.InlinedCalls.push_back(NewCall);
702 // It's possible that inlining the callsite will cause it to go from an
703 // indirect to a direct call by resolving a function pointer. If this
704 // happens, set the callee of the new call site to a more precise
705 // destination. This can also happen if the call graph node of the caller
706 // was just unnecessarily imprecise.
707 if (!I->second->getFunction())
708 if (Function *F = CallSite(NewCall).getCalledFunction()) {
709 // Indirect call site resolved to direct call.
710 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
715 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
718 // Update the call graph by deleting the edge from Callee to Caller. We must
719 // do this after the loop above in case Caller and Callee are the same.
720 CallerNode->removeCallEdgeFor(CS);
723 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
724 BasicBlock *InsertBlock,
725 InlineFunctionInfo &IFI) {
726 Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
727 IRBuilder<> Builder(InsertBlock->begin());
729 Value *Size = Builder.getInt64(M->getDataLayout().getTypeStoreSize(AggTy));
731 // Always generate a memcpy of alignment 1 here because we don't know
732 // the alignment of the src pointer. Other optimizations can infer
734 Builder.CreateMemCpy(Dst, Src, Size, /*Align=*/1);
737 /// HandleByValArgument - When inlining a call site that has a byval argument,
738 /// we have to make the implicit memcpy explicit by adding it.
739 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
740 const Function *CalledFunc,
741 InlineFunctionInfo &IFI,
742 unsigned ByValAlignment) {
743 PointerType *ArgTy = cast<PointerType>(Arg->getType());
744 Type *AggTy = ArgTy->getElementType();
746 Function *Caller = TheCall->getParent()->getParent();
748 // If the called function is readonly, then it could not mutate the caller's
749 // copy of the byval'd memory. In this case, it is safe to elide the copy and
751 if (CalledFunc->onlyReadsMemory()) {
752 // If the byval argument has a specified alignment that is greater than the
753 // passed in pointer, then we either have to round up the input pointer or
754 // give up on this transformation.
755 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
758 // If the pointer is already known to be sufficiently aligned, or if we can
759 // round it up to a larger alignment, then we don't need a temporary.
760 auto &DL = Caller->getParent()->getDataLayout();
761 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, &DL,
762 &IFI.ACT->getAssumptionCache(*Caller),
763 TheCall) >= ByValAlignment)
766 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
767 // for code quality, but rarely happens and is required for correctness.
770 // Create the alloca. If we have DataLayout, use nice alignment.
772 Caller->getParent()->getDataLayout().getPrefTypeAlignment(AggTy);
774 // If the byval had an alignment specified, we *must* use at least that
775 // alignment, as it is required by the byval argument (and uses of the
776 // pointer inside the callee).
777 Align = std::max(Align, ByValAlignment);
779 Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(),
780 &*Caller->begin()->begin());
781 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
783 // Uses of the argument in the function should use our new alloca
788 // isUsedByLifetimeMarker - Check whether this Value is used by a lifetime
790 static bool isUsedByLifetimeMarker(Value *V) {
791 for (User *U : V->users()) {
792 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
793 switch (II->getIntrinsicID()) {
795 case Intrinsic::lifetime_start:
796 case Intrinsic::lifetime_end:
804 // hasLifetimeMarkers - Check whether the given alloca already has
805 // lifetime.start or lifetime.end intrinsics.
806 static bool hasLifetimeMarkers(AllocaInst *AI) {
807 Type *Ty = AI->getType();
808 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
809 Ty->getPointerAddressSpace());
811 return isUsedByLifetimeMarker(AI);
813 // Do a scan to find all the casts to i8*.
814 for (User *U : AI->users()) {
815 if (U->getType() != Int8PtrTy) continue;
816 if (U->stripPointerCasts() != AI) continue;
817 if (isUsedByLifetimeMarker(U))
823 /// Rebuild the entire inlined-at chain for this instruction so that the top of
824 /// the chain now is inlined-at the new call site.
826 updateInlinedAtInfo(DebugLoc DL, MDLocation *InlinedAtNode,
828 DenseMap<const MDLocation *, MDLocation *> &IANodes) {
829 SmallVector<MDLocation*, 3> InlinedAtLocations;
830 MDLocation *Last = InlinedAtNode;
831 DebugLoc CurInlinedAt = DL;
833 // Gather all the inlined-at nodes
834 while (MDLocation *IA =
835 cast_or_null<MDLocation>(CurInlinedAt.getInlinedAt(Ctx))) {
836 // Skip any we've already built nodes for
837 if (MDLocation *Found = IANodes[IA]) {
842 InlinedAtLocations.push_back(IA);
843 CurInlinedAt = DebugLoc::getFromDILocation(IA);
846 // Starting from the top, rebuild the nodes to point to the new inlined-at
847 // location (then rebuilding the rest of the chain behind it) and update the
848 // map of already-constructed inlined-at nodes.
849 for (auto I = InlinedAtLocations.rbegin(), E = InlinedAtLocations.rend();
851 const MDLocation *MD = *I;
852 Last = IANodes[MD] = MDLocation::getDistinct(
853 Ctx, MD->getLine(), MD->getColumn(), MD->getScope(), Last);
856 // And finally create the normal location for this instruction, referring to
857 // the new inlined-at chain.
858 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), Last);
861 /// fixupLineNumbers - Update inlined instructions' line numbers to
862 /// to encode location where these instructions are inlined.
863 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
864 Instruction *TheCall) {
865 DebugLoc TheCallDL = TheCall->getDebugLoc();
866 if (TheCallDL.isUnknown())
869 auto &Ctx = Fn->getContext();
870 auto *InlinedAtNode = cast<MDLocation>(TheCallDL.getAsMDNode(Ctx));
872 // Create a unique call site, not to be confused with any other call from the
874 InlinedAtNode = MDLocation::getDistinct(
875 Ctx, InlinedAtNode->getLine(), InlinedAtNode->getColumn(),
876 InlinedAtNode->getScope(), InlinedAtNode->getInlinedAt());
878 // Cache the inlined-at nodes as they're built so they are reused, without
879 // this every instruction's inlined-at chain would become distinct from each
881 DenseMap<const MDLocation *, MDLocation *> IANodes;
883 for (; FI != Fn->end(); ++FI) {
884 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
886 DebugLoc DL = BI->getDebugLoc();
887 if (DL.isUnknown()) {
888 // If the inlined instruction has no line number, make it look as if it
889 // originates from the call location. This is important for
890 // ((__always_inline__, __nodebug__)) functions which must use caller
891 // location for all instructions in their function body.
893 // Don't update static allocas, as they may get moved later.
894 if (auto *AI = dyn_cast<AllocaInst>(BI))
895 if (isa<Constant>(AI->getArraySize()))
898 BI->setDebugLoc(TheCallDL);
900 BI->setDebugLoc(updateInlinedAtInfo(DL, InlinedAtNode, BI->getContext(), IANodes));
901 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) {
902 LLVMContext &Ctx = BI->getContext();
903 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
904 DVI->setOperand(2, MetadataAsValue::get(
905 Ctx, createInlinedVariable(DVI->getVariable(),
907 } else if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(BI)) {
908 LLVMContext &Ctx = BI->getContext();
909 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
910 DDI->setOperand(1, MetadataAsValue::get(
911 Ctx, createInlinedVariable(DDI->getVariable(),
919 /// InlineFunction - This function inlines the called function into the basic
920 /// block of the caller. This returns false if it is not possible to inline
921 /// this call. The program is still in a well defined state if this occurs
924 /// Note that this only does one level of inlining. For example, if the
925 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
926 /// exists in the instruction stream. Similarly this will inline a recursive
927 /// function by one level.
928 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
929 bool InsertLifetime) {
930 Instruction *TheCall = CS.getInstruction();
931 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
932 "Instruction not in function!");
934 // If IFI has any state in it, zap it before we fill it in.
937 const Function *CalledFunc = CS.getCalledFunction();
938 if (!CalledFunc || // Can't inline external function or indirect
939 CalledFunc->isDeclaration() || // call, or call to a vararg function!
940 CalledFunc->getFunctionType()->isVarArg()) return false;
942 // If the call to the callee cannot throw, set the 'nounwind' flag on any
943 // calls that we inline.
944 bool MarkNoUnwind = CS.doesNotThrow();
946 BasicBlock *OrigBB = TheCall->getParent();
947 Function *Caller = OrigBB->getParent();
949 // GC poses two hazards to inlining, which only occur when the callee has GC:
950 // 1. If the caller has no GC, then the callee's GC must be propagated to the
952 // 2. If the caller has a differing GC, it is invalid to inline.
953 if (CalledFunc->hasGC()) {
954 if (!Caller->hasGC())
955 Caller->setGC(CalledFunc->getGC());
956 else if (CalledFunc->getGC() != Caller->getGC())
960 // Get the personality function from the callee if it contains a landing pad.
961 Value *CalleePersonality = nullptr;
962 for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end();
964 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
965 const BasicBlock *BB = II->getUnwindDest();
966 const LandingPadInst *LP = BB->getLandingPadInst();
967 CalleePersonality = LP->getPersonalityFn();
971 // Find the personality function used by the landing pads of the caller. If it
972 // exists, then check to see that it matches the personality function used in
974 if (CalleePersonality) {
975 for (Function::const_iterator I = Caller->begin(), E = Caller->end();
977 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
978 const BasicBlock *BB = II->getUnwindDest();
979 const LandingPadInst *LP = BB->getLandingPadInst();
981 // If the personality functions match, then we can perform the
982 // inlining. Otherwise, we can't inline.
983 // TODO: This isn't 100% true. Some personality functions are proper
984 // supersets of others and can be used in place of the other.
985 if (LP->getPersonalityFn() != CalleePersonality)
992 // Get an iterator to the last basic block in the function, which will have
993 // the new function inlined after it.
994 Function::iterator LastBlock = &Caller->back();
996 // Make sure to capture all of the return instructions from the cloned
998 SmallVector<ReturnInst*, 8> Returns;
999 ClonedCodeInfo InlinedFunctionInfo;
1000 Function::iterator FirstNewBlock;
1002 { // Scope to destroy VMap after cloning.
1003 ValueToValueMapTy VMap;
1004 // Keep a list of pair (dst, src) to emit byval initializations.
1005 SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
1007 auto &DL = Caller->getParent()->getDataLayout();
1009 assert(CalledFunc->arg_size() == CS.arg_size() &&
1010 "No varargs calls can be inlined!");
1012 // Calculate the vector of arguments to pass into the function cloner, which
1013 // matches up the formal to the actual argument values.
1014 CallSite::arg_iterator AI = CS.arg_begin();
1016 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
1017 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
1018 Value *ActualArg = *AI;
1020 // When byval arguments actually inlined, we need to make the copy implied
1021 // by them explicit. However, we don't do this if the callee is readonly
1022 // or readnone, because the copy would be unneeded: the callee doesn't
1023 // modify the struct.
1024 if (CS.isByValArgument(ArgNo)) {
1025 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
1026 CalledFunc->getParamAlignment(ArgNo+1));
1027 if (ActualArg != *AI)
1028 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
1031 VMap[I] = ActualArg;
1034 // Add alignment assumptions if necessary. We do this before the inlined
1035 // instructions are actually cloned into the caller so that we can easily
1036 // check what will be known at the start of the inlined code.
1037 AddAlignmentAssumptions(CS, IFI);
1039 // We want the inliner to prune the code as it copies. We would LOVE to
1040 // have no dead or constant instructions leftover after inlining occurs
1041 // (which can happen, e.g., because an argument was constant), but we'll be
1042 // happy with whatever the cloner can do.
1043 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
1044 /*ModuleLevelChanges=*/false, Returns, ".i",
1045 &InlinedFunctionInfo, &DL, TheCall);
1047 // Remember the first block that is newly cloned over.
1048 FirstNewBlock = LastBlock; ++FirstNewBlock;
1050 // Inject byval arguments initialization.
1051 for (std::pair<Value*, Value*> &Init : ByValInit)
1052 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
1053 FirstNewBlock, IFI);
1055 // Update the callgraph if requested.
1057 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
1059 // Update inlined instructions' line number information.
1060 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
1062 // Clone existing noalias metadata if necessary.
1063 CloneAliasScopeMetadata(CS, VMap);
1065 // Add noalias metadata if necessary.
1066 AddAliasScopeMetadata(CS, VMap, &DL, IFI.AA);
1068 // FIXME: We could register any cloned assumptions instead of clearing the
1069 // whole function's cache.
1071 IFI.ACT->getAssumptionCache(*Caller).clear();
1074 // If there are any alloca instructions in the block that used to be the entry
1075 // block for the callee, move them to the entry block of the caller. First
1076 // calculate which instruction they should be inserted before. We insert the
1077 // instructions at the end of the current alloca list.
1079 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
1080 for (BasicBlock::iterator I = FirstNewBlock->begin(),
1081 E = FirstNewBlock->end(); I != E; ) {
1082 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
1085 // If the alloca is now dead, remove it. This often occurs due to code
1087 if (AI->use_empty()) {
1088 AI->eraseFromParent();
1092 if (!isa<Constant>(AI->getArraySize()))
1095 // Keep track of the static allocas that we inline into the caller.
1096 IFI.StaticAllocas.push_back(AI);
1098 // Scan for the block of allocas that we can move over, and move them
1100 while (isa<AllocaInst>(I) &&
1101 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
1102 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
1106 // Transfer all of the allocas over in a block. Using splice means
1107 // that the instructions aren't removed from the symbol table, then
1109 Caller->getEntryBlock().getInstList().splice(InsertPoint,
1110 FirstNewBlock->getInstList(),
1113 // Move any dbg.declares describing the allocas into the entry basic block.
1114 DIBuilder DIB(*Caller->getParent());
1115 for (auto &AI : IFI.StaticAllocas)
1116 replaceDbgDeclareForAlloca(AI, AI, DIB, /*Deref=*/false);
1119 bool InlinedMustTailCalls = false;
1120 if (InlinedFunctionInfo.ContainsCalls) {
1121 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
1122 if (CallInst *CI = dyn_cast<CallInst>(TheCall))
1123 CallSiteTailKind = CI->getTailCallKind();
1125 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
1127 for (Instruction &I : *BB) {
1128 CallInst *CI = dyn_cast<CallInst>(&I);
1132 // We need to reduce the strength of any inlined tail calls. For
1133 // musttail, we have to avoid introducing potential unbounded stack
1134 // growth. For example, if functions 'f' and 'g' are mutually recursive
1135 // with musttail, we can inline 'g' into 'f' so long as we preserve
1136 // musttail on the cloned call to 'f'. If either the inlined call site
1137 // or the cloned call site is *not* musttail, the program already has
1138 // one frame of stack growth, so it's safe to remove musttail. Here is
1139 // a table of example transformations:
1141 // f -> musttail g -> musttail f ==> f -> musttail f
1142 // f -> musttail g -> tail f ==> f -> tail f
1143 // f -> g -> musttail f ==> f -> f
1144 // f -> g -> tail f ==> f -> f
1145 CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
1146 ChildTCK = std::min(CallSiteTailKind, ChildTCK);
1147 CI->setTailCallKind(ChildTCK);
1148 InlinedMustTailCalls |= CI->isMustTailCall();
1150 // Calls inlined through a 'nounwind' call site should be marked
1153 CI->setDoesNotThrow();
1158 // Leave lifetime markers for the static alloca's, scoping them to the
1159 // function we just inlined.
1160 if (InsertLifetime && !IFI.StaticAllocas.empty()) {
1161 IRBuilder<> builder(FirstNewBlock->begin());
1162 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
1163 AllocaInst *AI = IFI.StaticAllocas[ai];
1165 // If the alloca is already scoped to something smaller than the whole
1166 // function then there's no need to add redundant, less accurate markers.
1167 if (hasLifetimeMarkers(AI))
1170 // Try to determine the size of the allocation.
1171 ConstantInt *AllocaSize = nullptr;
1172 if (ConstantInt *AIArraySize =
1173 dyn_cast<ConstantInt>(AI->getArraySize())) {
1174 auto &DL = Caller->getParent()->getDataLayout();
1175 Type *AllocaType = AI->getAllocatedType();
1176 uint64_t AllocaTypeSize = DL.getTypeAllocSize(AllocaType);
1177 uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
1178 assert(AllocaArraySize > 0 && "array size of AllocaInst is zero");
1179 // Check that array size doesn't saturate uint64_t and doesn't
1180 // overflow when it's multiplied by type size.
1181 if (AllocaArraySize != ~0ULL &&
1182 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
1183 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
1184 AllocaArraySize * AllocaTypeSize);
1188 builder.CreateLifetimeStart(AI, AllocaSize);
1189 for (ReturnInst *RI : Returns) {
1190 // Don't insert llvm.lifetime.end calls between a musttail call and a
1191 // return. The return kills all local allocas.
1192 if (InlinedMustTailCalls &&
1193 RI->getParent()->getTerminatingMustTailCall())
1195 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
1200 // If the inlined code contained dynamic alloca instructions, wrap the inlined
1201 // code with llvm.stacksave/llvm.stackrestore intrinsics.
1202 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
1203 Module *M = Caller->getParent();
1204 // Get the two intrinsics we care about.
1205 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
1206 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
1208 // Insert the llvm.stacksave.
1209 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
1210 .CreateCall(StackSave, "savedstack");
1212 // Insert a call to llvm.stackrestore before any return instructions in the
1213 // inlined function.
1214 for (ReturnInst *RI : Returns) {
1215 // Don't insert llvm.stackrestore calls between a musttail call and a
1216 // return. The return will restore the stack pointer.
1217 if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
1219 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
1223 // If we are inlining for an invoke instruction, we must make sure to rewrite
1224 // any call instructions into invoke instructions.
1225 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
1226 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
1228 // Handle any inlined musttail call sites. In order for a new call site to be
1229 // musttail, the source of the clone and the inlined call site must have been
1230 // musttail. Therefore it's safe to return without merging control into the
1232 if (InlinedMustTailCalls) {
1233 // Check if we need to bitcast the result of any musttail calls.
1234 Type *NewRetTy = Caller->getReturnType();
1235 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
1237 // Handle the returns preceded by musttail calls separately.
1238 SmallVector<ReturnInst *, 8> NormalReturns;
1239 for (ReturnInst *RI : Returns) {
1240 CallInst *ReturnedMustTail =
1241 RI->getParent()->getTerminatingMustTailCall();
1242 if (!ReturnedMustTail) {
1243 NormalReturns.push_back(RI);
1249 // Delete the old return and any preceding bitcast.
1250 BasicBlock *CurBB = RI->getParent();
1251 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
1252 RI->eraseFromParent();
1254 OldCast->eraseFromParent();
1256 // Insert a new bitcast and return with the right type.
1257 IRBuilder<> Builder(CurBB);
1258 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
1261 // Leave behind the normal returns so we can merge control flow.
1262 std::swap(Returns, NormalReturns);
1265 // If we cloned in _exactly one_ basic block, and if that block ends in a
1266 // return instruction, we splice the body of the inlined callee directly into
1267 // the calling basic block.
1268 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
1269 // Move all of the instructions right before the call.
1270 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
1271 FirstNewBlock->begin(), FirstNewBlock->end());
1272 // Remove the cloned basic block.
1273 Caller->getBasicBlockList().pop_back();
1275 // If the call site was an invoke instruction, add a branch to the normal
1277 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1278 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
1279 NewBr->setDebugLoc(Returns[0]->getDebugLoc());
1282 // If the return instruction returned a value, replace uses of the call with
1283 // uses of the returned value.
1284 if (!TheCall->use_empty()) {
1285 ReturnInst *R = Returns[0];
1286 if (TheCall == R->getReturnValue())
1287 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1289 TheCall->replaceAllUsesWith(R->getReturnValue());
1291 // Since we are now done with the Call/Invoke, we can delete it.
1292 TheCall->eraseFromParent();
1294 // Since we are now done with the return instruction, delete it also.
1295 Returns[0]->eraseFromParent();
1297 // We are now done with the inlining.
1301 // Otherwise, we have the normal case, of more than one block to inline or
1302 // multiple return sites.
1304 // We want to clone the entire callee function into the hole between the
1305 // "starter" and "ender" blocks. How we accomplish this depends on whether
1306 // this is an invoke instruction or a call instruction.
1307 BasicBlock *AfterCallBB;
1308 BranchInst *CreatedBranchToNormalDest = nullptr;
1309 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1311 // Add an unconditional branch to make this look like the CallInst case...
1312 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
1314 // Split the basic block. This guarantees that no PHI nodes will have to be
1315 // updated due to new incoming edges, and make the invoke case more
1316 // symmetric to the call case.
1317 AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest,
1318 CalledFunc->getName()+".exit");
1320 } else { // It's a call
1321 // If this is a call instruction, we need to split the basic block that
1322 // the call lives in.
1324 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
1325 CalledFunc->getName()+".exit");
1328 // Change the branch that used to go to AfterCallBB to branch to the first
1329 // basic block of the inlined function.
1331 TerminatorInst *Br = OrigBB->getTerminator();
1332 assert(Br && Br->getOpcode() == Instruction::Br &&
1333 "splitBasicBlock broken!");
1334 Br->setOperand(0, FirstNewBlock);
1337 // Now that the function is correct, make it a little bit nicer. In
1338 // particular, move the basic blocks inserted from the end of the function
1339 // into the space made by splitting the source basic block.
1340 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
1341 FirstNewBlock, Caller->end());
1343 // Handle all of the return instructions that we just cloned in, and eliminate
1344 // any users of the original call/invoke instruction.
1345 Type *RTy = CalledFunc->getReturnType();
1347 PHINode *PHI = nullptr;
1348 if (Returns.size() > 1) {
1349 // The PHI node should go at the front of the new basic block to merge all
1350 // possible incoming values.
1351 if (!TheCall->use_empty()) {
1352 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
1353 AfterCallBB->begin());
1354 // Anything that used the result of the function call should now use the
1355 // PHI node as their operand.
1356 TheCall->replaceAllUsesWith(PHI);
1359 // Loop over all of the return instructions adding entries to the PHI node
1362 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1363 ReturnInst *RI = Returns[i];
1364 assert(RI->getReturnValue()->getType() == PHI->getType() &&
1365 "Ret value not consistent in function!");
1366 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
1371 // Add a branch to the merge points and remove return instructions.
1373 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1374 ReturnInst *RI = Returns[i];
1375 BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
1376 Loc = RI->getDebugLoc();
1377 BI->setDebugLoc(Loc);
1378 RI->eraseFromParent();
1380 // We need to set the debug location to *somewhere* inside the
1381 // inlined function. The line number may be nonsensical, but the
1382 // instruction will at least be associated with the right
1384 if (CreatedBranchToNormalDest)
1385 CreatedBranchToNormalDest->setDebugLoc(Loc);
1386 } else if (!Returns.empty()) {
1387 // Otherwise, if there is exactly one return value, just replace anything
1388 // using the return value of the call with the computed value.
1389 if (!TheCall->use_empty()) {
1390 if (TheCall == Returns[0]->getReturnValue())
1391 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1393 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
1396 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
1397 BasicBlock *ReturnBB = Returns[0]->getParent();
1398 ReturnBB->replaceAllUsesWith(AfterCallBB);
1400 // Splice the code from the return block into the block that it will return
1401 // to, which contains the code that was after the call.
1402 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
1403 ReturnBB->getInstList());
1405 if (CreatedBranchToNormalDest)
1406 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
1408 // Delete the return instruction now and empty ReturnBB now.
1409 Returns[0]->eraseFromParent();
1410 ReturnBB->eraseFromParent();
1411 } else if (!TheCall->use_empty()) {
1412 // No returns, but something is using the return value of the call. Just
1414 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1417 // Since we are now done with the Call/Invoke, we can delete it.
1418 TheCall->eraseFromParent();
1420 // If we inlined any musttail calls and the original return is now
1421 // unreachable, delete it. It can only contain a bitcast and ret.
1422 if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB))
1423 AfterCallBB->eraseFromParent();
1425 // We should always be able to fold the entry block of the function into the
1426 // single predecessor of the block...
1427 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
1428 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
1430 // Splice the code entry block into calling block, right before the
1431 // unconditional branch.
1432 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
1433 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
1435 // Remove the unconditional branch.
1436 OrigBB->getInstList().erase(Br);
1438 // Now we can remove the CalleeEntry block, which is now empty.
1439 Caller->getBasicBlockList().erase(CalleeEntry);
1441 // If we inserted a phi node, check to see if it has a single value (e.g. all
1442 // the entries are the same or undef). If so, remove the PHI so it doesn't
1443 // block other optimizations.
1445 auto &DL = Caller->getParent()->getDataLayout();
1446 if (Value *V = SimplifyInstruction(PHI, &DL, nullptr, nullptr,
1447 &IFI.ACT->getAssumptionCache(*Caller))) {
1448 PHI->replaceAllUsesWith(V);
1449 PHI->eraseFromParent();