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/Dominators.h"
34 #include "llvm/IR/IRBuilder.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/IntrinsicInst.h"
37 #include "llvm/IR/Intrinsics.h"
38 #include "llvm/IR/MDBuilder.h"
39 #include "llvm/IR/Module.h"
40 #include "llvm/Transforms/Utils/Local.h"
41 #include "llvm/Support/CommandLine.h"
46 EnableNoAliasConversion("enable-noalias-to-md-conversion", cl::init(true),
48 cl::desc("Convert noalias attributes to metadata during inlining."));
51 PreserveAlignmentAssumptions("preserve-alignment-assumptions-during-inlining",
52 cl::init(true), cl::Hidden,
53 cl::desc("Convert align attributes to assumptions during inlining."));
55 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI,
56 bool InsertLifetime) {
57 return InlineFunction(CallSite(CI), IFI, InsertLifetime);
59 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI,
60 bool InsertLifetime) {
61 return InlineFunction(CallSite(II), IFI, InsertLifetime);
65 /// A class for recording information about inlining through an invoke.
66 class InvokeInliningInfo {
67 BasicBlock *OuterResumeDest; ///< Destination of the invoke's unwind.
68 BasicBlock *InnerResumeDest; ///< Destination for the callee's resume.
69 LandingPadInst *CallerLPad; ///< LandingPadInst associated with the invoke.
70 PHINode *InnerEHValuesPHI; ///< PHI for EH values from landingpad insts.
71 SmallVector<Value*, 8> UnwindDestPHIValues;
74 InvokeInliningInfo(InvokeInst *II)
75 : OuterResumeDest(II->getUnwindDest()), InnerResumeDest(nullptr),
76 CallerLPad(nullptr), InnerEHValuesPHI(nullptr) {
77 // If there are PHI nodes in the unwind destination block, we need to keep
78 // track of which values came into them from the invoke before removing
79 // the edge from this block.
80 llvm::BasicBlock *InvokeBB = II->getParent();
81 BasicBlock::iterator I = OuterResumeDest->begin();
82 for (; isa<PHINode>(I); ++I) {
83 // Save the value to use for this edge.
84 PHINode *PHI = cast<PHINode>(I);
85 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
88 CallerLPad = cast<LandingPadInst>(I);
91 /// getOuterResumeDest - The outer unwind destination is the target of
92 /// unwind edges introduced for calls within the inlined function.
93 BasicBlock *getOuterResumeDest() const {
94 return OuterResumeDest;
97 BasicBlock *getInnerResumeDest();
99 LandingPadInst *getLandingPadInst() const { return CallerLPad; }
101 /// forwardResume - Forward the 'resume' instruction to the caller's landing
102 /// pad block. When the landing pad block has only one predecessor, this is
103 /// a simple branch. When there is more than one predecessor, we need to
104 /// split the landing pad block after the landingpad instruction and jump
106 void forwardResume(ResumeInst *RI,
107 SmallPtrSetImpl<LandingPadInst*> &InlinedLPads);
109 /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind
110 /// destination block for the given basic block, using the values for the
111 /// original invoke's source block.
112 void addIncomingPHIValuesFor(BasicBlock *BB) const {
113 addIncomingPHIValuesForInto(BB, OuterResumeDest);
116 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const {
117 BasicBlock::iterator I = dest->begin();
118 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
119 PHINode *phi = cast<PHINode>(I);
120 phi->addIncoming(UnwindDestPHIValues[i], src);
126 /// getInnerResumeDest - Get or create a target for the branch from ResumeInsts.
127 BasicBlock *InvokeInliningInfo::getInnerResumeDest() {
128 if (InnerResumeDest) return InnerResumeDest;
130 // Split the landing pad.
131 BasicBlock::iterator SplitPoint = CallerLPad; ++SplitPoint;
133 OuterResumeDest->splitBasicBlock(SplitPoint,
134 OuterResumeDest->getName() + ".body");
136 // The number of incoming edges we expect to the inner landing pad.
137 const unsigned PHICapacity = 2;
139 // Create corresponding new PHIs for all the PHIs in the outer landing pad.
140 BasicBlock::iterator InsertPoint = InnerResumeDest->begin();
141 BasicBlock::iterator I = OuterResumeDest->begin();
142 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
143 PHINode *OuterPHI = cast<PHINode>(I);
144 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity,
145 OuterPHI->getName() + ".lpad-body",
147 OuterPHI->replaceAllUsesWith(InnerPHI);
148 InnerPHI->addIncoming(OuterPHI, OuterResumeDest);
151 // Create a PHI for the exception values.
152 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity,
153 "eh.lpad-body", InsertPoint);
154 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI);
155 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest);
158 return InnerResumeDest;
161 /// forwardResume - Forward the 'resume' instruction to the caller's landing pad
162 /// block. When the landing pad block has only one predecessor, this is a simple
163 /// branch. When there is more than one predecessor, we need to split the
164 /// landing pad block after the landingpad instruction and jump to there.
165 void InvokeInliningInfo::forwardResume(ResumeInst *RI,
166 SmallPtrSetImpl<LandingPadInst*> &InlinedLPads) {
167 BasicBlock *Dest = getInnerResumeDest();
168 BasicBlock *Src = RI->getParent();
170 BranchInst::Create(Dest, Src);
172 // Update the PHIs in the destination. They were inserted in an order which
174 addIncomingPHIValuesForInto(Src, Dest);
176 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src);
177 RI->eraseFromParent();
180 /// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
181 /// an invoke, we have to turn all of the calls that can throw into
182 /// invokes. This function analyze BB to see if there are any calls, and if so,
183 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
184 /// nodes in that block with the values specified in InvokeDestPHIValues.
185 static void HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
186 InvokeInliningInfo &Invoke) {
187 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
188 Instruction *I = BBI++;
190 // We only need to check for function calls: inlined invoke
191 // instructions require no special handling.
192 CallInst *CI = dyn_cast<CallInst>(I);
194 // If this call cannot unwind, don't convert it to an invoke.
195 // Inline asm calls cannot throw.
196 if (!CI || CI->doesNotThrow() || isa<InlineAsm>(CI->getCalledValue()))
199 // Convert this function call into an invoke instruction. First, split the
201 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
203 // Delete the unconditional branch inserted by splitBasicBlock
204 BB->getInstList().pop_back();
206 // Create the new invoke instruction.
207 ImmutableCallSite CS(CI);
208 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
209 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split,
210 Invoke.getOuterResumeDest(),
211 InvokeArgs, CI->getName(), BB);
212 II->setDebugLoc(CI->getDebugLoc());
213 II->setCallingConv(CI->getCallingConv());
214 II->setAttributes(CI->getAttributes());
216 // Make sure that anything using the call now uses the invoke! This also
217 // updates the CallGraph if present, because it uses a WeakVH.
218 CI->replaceAllUsesWith(II);
220 // Delete the original call
221 Split->getInstList().pop_front();
223 // Update any PHI nodes in the exceptional block to indicate that there is
224 // now a new entry in them.
225 Invoke.addIncomingPHIValuesFor(BB);
230 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
231 /// in the body of the inlined function into invokes.
233 /// II is the invoke instruction being inlined. FirstNewBlock is the first
234 /// block of the inlined code (the last block is the end of the function),
235 /// and InlineCodeInfo is information about the code that got inlined.
236 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
237 ClonedCodeInfo &InlinedCodeInfo) {
238 BasicBlock *InvokeDest = II->getUnwindDest();
240 Function *Caller = FirstNewBlock->getParent();
242 // The inlined code is currently at the end of the function, scan from the
243 // start of the inlined code to its end, checking for stuff we need to
245 InvokeInliningInfo Invoke(II);
247 // Get all of the inlined landing pad instructions.
248 SmallPtrSet<LandingPadInst*, 16> InlinedLPads;
249 for (Function::iterator I = FirstNewBlock, E = Caller->end(); I != E; ++I)
250 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator()))
251 InlinedLPads.insert(II->getLandingPadInst());
253 // Append the clauses from the outer landing pad instruction into the inlined
254 // landing pad instructions.
255 LandingPadInst *OuterLPad = Invoke.getLandingPadInst();
256 for (LandingPadInst *InlinedLPad : InlinedLPads) {
257 unsigned OuterNum = OuterLPad->getNumClauses();
258 InlinedLPad->reserveClauses(OuterNum);
259 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx)
260 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx));
261 if (OuterLPad->isCleanup())
262 InlinedLPad->setCleanup(true);
265 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
266 if (InlinedCodeInfo.ContainsCalls)
267 HandleCallsInBlockInlinedThroughInvoke(BB, Invoke);
269 // Forward any resumes that are remaining here.
270 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator()))
271 Invoke.forwardResume(RI, InlinedLPads);
274 // Now that everything is happy, we have one final detail. The PHI nodes in
275 // the exception destination block still have entries due to the original
276 // invoke instruction. Eliminate these entries (which might even delete the
278 InvokeDest->removePredecessor(II->getParent());
281 /// CloneAliasScopeMetadata - When inlining a function that contains noalias
282 /// scope metadata, this metadata needs to be cloned so that the inlined blocks
283 /// have different "unqiue scopes" at every call site. Were this not done, then
284 /// aliasing scopes from a function inlined into a caller multiple times could
285 /// not be differentiated (and this would lead to miscompiles because the
286 /// non-aliasing property communicated by the metadata could have
287 /// call-site-specific control dependencies).
288 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) {
289 const Function *CalledFunc = CS.getCalledFunction();
290 SetVector<const MDNode *> MD;
292 // Note: We could only clone the metadata if it is already used in the
293 // caller. I'm omitting that check here because it might confuse
294 // inter-procedural alias analysis passes. We can revisit this if it becomes
295 // an efficiency or overhead problem.
297 for (Function::const_iterator I = CalledFunc->begin(), IE = CalledFunc->end();
299 for (BasicBlock::const_iterator J = I->begin(), JE = I->end(); J != JE; ++J) {
300 if (const MDNode *M = J->getMetadata(LLVMContext::MD_alias_scope))
302 if (const MDNode *M = J->getMetadata(LLVMContext::MD_noalias))
309 // Walk the existing metadata, adding the complete (perhaps cyclic) chain to
311 SmallVector<const Metadata *, 16> Queue(MD.begin(), MD.end());
312 while (!Queue.empty()) {
313 const MDNode *M = cast<MDNode>(Queue.pop_back_val());
314 for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i)
315 if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i)))
320 // Now we have a complete set of all metadata in the chains used to specify
321 // the noalias scopes and the lists of those scopes.
322 SmallVector<TempMDTuple, 16> DummyNodes;
323 DenseMap<const MDNode *, TrackingMDNodeRef> MDMap;
324 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
326 DummyNodes.push_back(MDTuple::getTemporary(CalledFunc->getContext(), None));
327 MDMap[*I].reset(DummyNodes.back().get());
330 // Create new metadata nodes to replace the dummy nodes, replacing old
331 // metadata references with either a dummy node or an already-created new
333 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
335 SmallVector<Metadata *, 4> NewOps;
336 for (unsigned i = 0, ie = (*I)->getNumOperands(); i != ie; ++i) {
337 const Metadata *V = (*I)->getOperand(i);
338 if (const MDNode *M = dyn_cast<MDNode>(V))
339 NewOps.push_back(MDMap[M]);
341 NewOps.push_back(const_cast<Metadata *>(V));
344 MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps);
345 MDTuple *TempM = cast<MDTuple>(MDMap[*I]);
346 assert(TempM->isTemporary() && "Expected temporary node");
348 TempM->replaceAllUsesWith(NewM);
351 // Now replace the metadata in the new inlined instructions with the
352 // repacements from the map.
353 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
354 VMI != VMIE; ++VMI) {
358 Instruction *NI = dyn_cast<Instruction>(VMI->second);
362 if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope)) {
363 MDNode *NewMD = MDMap[M];
364 // If the call site also had alias scope metadata (a list of scopes to
365 // which instructions inside it might belong), propagate those scopes to
366 // the inlined instructions.
368 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
369 NewMD = MDNode::concatenate(NewMD, CSM);
370 NI->setMetadata(LLVMContext::MD_alias_scope, NewMD);
371 } else if (NI->mayReadOrWriteMemory()) {
373 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
374 NI->setMetadata(LLVMContext::MD_alias_scope, M);
377 if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias)) {
378 MDNode *NewMD = MDMap[M];
379 // If the call site also had noalias metadata (a list of scopes with
380 // which instructions inside it don't alias), propagate those scopes to
381 // the inlined instructions.
383 CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
384 NewMD = MDNode::concatenate(NewMD, CSM);
385 NI->setMetadata(LLVMContext::MD_noalias, NewMD);
386 } else if (NI->mayReadOrWriteMemory()) {
387 if (MDNode *M = CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
388 NI->setMetadata(LLVMContext::MD_noalias, M);
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, AliasAnalysis *AA) {
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 bool IsArgMemOnlyCall = false, IsFuncCall = false;
463 SmallVector<const Value *, 2> PtrArgs;
465 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
466 PtrArgs.push_back(LI->getPointerOperand());
467 else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
468 PtrArgs.push_back(SI->getPointerOperand());
469 else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
470 PtrArgs.push_back(VAAI->getPointerOperand());
471 else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
472 PtrArgs.push_back(CXI->getPointerOperand());
473 else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
474 PtrArgs.push_back(RMWI->getPointerOperand());
475 else if (ImmutableCallSite ICS = ImmutableCallSite(I)) {
476 // If we know that the call does not access memory, then we'll still
477 // know that about the inlined clone of this call site, and we don't
478 // need to add metadata.
479 if (ICS.doesNotAccessMemory())
484 AliasAnalysis::ModRefBehavior MRB = AA->getModRefBehavior(ICS);
485 if (MRB == AliasAnalysis::OnlyAccessesArgumentPointees ||
486 MRB == AliasAnalysis::OnlyReadsArgumentPointees)
487 IsArgMemOnlyCall = true;
490 for (ImmutableCallSite::arg_iterator AI = ICS.arg_begin(),
491 AE = ICS.arg_end(); AI != AE; ++AI) {
492 // We need to check the underlying objects of all arguments, not just
493 // the pointer arguments, because we might be passing pointers as
495 // However, if we know that the call only accesses pointer arguments,
496 // then we only need to check the pointer arguments.
497 if (IsArgMemOnlyCall && !(*AI)->getType()->isPointerTy())
500 PtrArgs.push_back(*AI);
504 // If we found no pointers, then this instruction is not suitable for
505 // pairing with an instruction to receive aliasing metadata.
506 // However, if this is a call, this we might just alias with none of the
507 // noalias arguments.
508 if (PtrArgs.empty() && !IsFuncCall)
511 // It is possible that there is only one underlying object, but you
512 // need to go through several PHIs to see it, and thus could be
513 // repeated in the Objects list.
514 SmallPtrSet<const Value *, 4> ObjSet;
515 SmallVector<Metadata *, 4> Scopes, NoAliases;
517 SmallSetVector<const Argument *, 4> NAPtrArgs;
518 for (unsigned i = 0, ie = PtrArgs.size(); i != ie; ++i) {
519 SmallVector<Value *, 4> Objects;
520 GetUnderlyingObjects(const_cast<Value*>(PtrArgs[i]),
521 Objects, DL, /* MaxLookup = */ 0);
523 for (Value *O : Objects)
527 // Figure out if we're derived from anything that is not a noalias
529 bool CanDeriveViaCapture = false, UsesAliasingPtr = false;
530 for (const Value *V : ObjSet) {
531 // Is this value a constant that cannot be derived from any pointer
532 // value (we need to exclude constant expressions, for example, that
533 // are formed from arithmetic on global symbols).
534 bool IsNonPtrConst = isa<ConstantInt>(V) || isa<ConstantFP>(V) ||
535 isa<ConstantPointerNull>(V) ||
536 isa<ConstantDataVector>(V) || isa<UndefValue>(V);
540 // If this is anything other than a noalias argument, then we cannot
541 // completely describe the aliasing properties using alias.scope
542 // metadata (and, thus, won't add any).
543 if (const Argument *A = dyn_cast<Argument>(V)) {
544 if (!A->hasNoAliasAttr())
545 UsesAliasingPtr = true;
547 UsesAliasingPtr = true;
550 // If this is not some identified function-local object (which cannot
551 // directly alias a noalias argument), or some other argument (which,
552 // by definition, also cannot alias a noalias argument), then we could
553 // alias a noalias argument that has been captured).
554 if (!isa<Argument>(V) &&
555 !isIdentifiedFunctionLocal(const_cast<Value*>(V)))
556 CanDeriveViaCapture = true;
559 // A function call can always get captured noalias pointers (via other
560 // parameters, globals, etc.).
561 if (IsFuncCall && !IsArgMemOnlyCall)
562 CanDeriveViaCapture = true;
564 // First, we want to figure out all of the sets with which we definitely
565 // don't alias. Iterate over all noalias set, and add those for which:
566 // 1. The noalias argument is not in the set of objects from which we
567 // definitely derive.
568 // 2. The noalias argument has not yet been captured.
569 // An arbitrary function that might load pointers could see captured
570 // noalias arguments via other noalias arguments or globals, and so we
571 // must always check for prior capture.
572 for (const Argument *A : NoAliasArgs) {
573 if (!ObjSet.count(A) && (!CanDeriveViaCapture ||
574 // It might be tempting to skip the
575 // PointerMayBeCapturedBefore check if
576 // A->hasNoCaptureAttr() is true, but this is
577 // incorrect because nocapture only guarantees
578 // that no copies outlive the function, not
579 // that the value cannot be locally captured.
580 !PointerMayBeCapturedBefore(A,
581 /* ReturnCaptures */ false,
582 /* StoreCaptures */ false, I, &DT)))
583 NoAliases.push_back(NewScopes[A]);
586 if (!NoAliases.empty())
587 NI->setMetadata(LLVMContext::MD_noalias,
589 NI->getMetadata(LLVMContext::MD_noalias),
590 MDNode::get(CalledFunc->getContext(), NoAliases)));
592 // Next, we want to figure out all of the sets to which we might belong.
593 // We might belong to a set if the noalias argument is in the set of
594 // underlying objects. If there is some non-noalias argument in our list
595 // of underlying objects, then we cannot add a scope because the fact
596 // that some access does not alias with any set of our noalias arguments
597 // cannot itself guarantee that it does not alias with this access
598 // (because there is some pointer of unknown origin involved and the
599 // other access might also depend on this pointer). We also cannot add
600 // scopes to arbitrary functions unless we know they don't access any
601 // non-parameter pointer-values.
602 bool CanAddScopes = !UsesAliasingPtr;
603 if (CanAddScopes && IsFuncCall)
604 CanAddScopes = IsArgMemOnlyCall;
607 for (const Argument *A : NoAliasArgs) {
609 Scopes.push_back(NewScopes[A]);
614 LLVMContext::MD_alias_scope,
615 MDNode::concatenate(NI->getMetadata(LLVMContext::MD_alias_scope),
616 MDNode::get(CalledFunc->getContext(), Scopes)));
621 /// If the inlined function has non-byval align arguments, then
622 /// add @llvm.assume-based alignment assumptions to preserve this information.
623 static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) {
624 if (!PreserveAlignmentAssumptions || !IFI.DL)
627 // To avoid inserting redundant assumptions, we should check for assumptions
628 // already in the caller. To do this, we might need a DT of the caller.
630 bool DTCalculated = false;
632 Function *CalledFunc = CS.getCalledFunction();
633 for (Function::arg_iterator I = CalledFunc->arg_begin(),
634 E = CalledFunc->arg_end();
636 unsigned Align = I->getType()->isPointerTy() ? I->getParamAlignment() : 0;
637 if (Align && !I->hasByValOrInAllocaAttr() && !I->hasNUses(0)) {
639 DT.recalculate(const_cast<Function&>(*CS.getInstruction()->getParent()
644 // If we can already prove the asserted alignment in the context of the
645 // caller, then don't bother inserting the assumption.
646 Value *Arg = CS.getArgument(I->getArgNo());
647 if (getKnownAlignment(Arg, IFI.DL,
648 &IFI.ACT->getAssumptionCache(*CalledFunc),
649 CS.getInstruction(), &DT) >= Align)
652 IRBuilder<>(CS.getInstruction()).CreateAlignmentAssumption(*IFI.DL, Arg,
658 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
659 /// into the caller, update the specified callgraph to reflect the changes we
660 /// made. Note that it's possible that not all code was copied over, so only
661 /// some edges of the callgraph may remain.
662 static void UpdateCallGraphAfterInlining(CallSite CS,
663 Function::iterator FirstNewBlock,
664 ValueToValueMapTy &VMap,
665 InlineFunctionInfo &IFI) {
666 CallGraph &CG = *IFI.CG;
667 const Function *Caller = CS.getInstruction()->getParent()->getParent();
668 const Function *Callee = CS.getCalledFunction();
669 CallGraphNode *CalleeNode = CG[Callee];
670 CallGraphNode *CallerNode = CG[Caller];
672 // Since we inlined some uninlined call sites in the callee into the caller,
673 // add edges from the caller to all of the callees of the callee.
674 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
676 // Consider the case where CalleeNode == CallerNode.
677 CallGraphNode::CalledFunctionsVector CallCache;
678 if (CalleeNode == CallerNode) {
679 CallCache.assign(I, E);
680 I = CallCache.begin();
684 for (; I != E; ++I) {
685 const Value *OrigCall = I->first;
687 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
688 // Only copy the edge if the call was inlined!
689 if (VMI == VMap.end() || VMI->second == nullptr)
692 // If the call was inlined, but then constant folded, there is no edge to
693 // add. Check for this case.
694 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
695 if (!NewCall) continue;
697 // Remember that this call site got inlined for the client of
699 IFI.InlinedCalls.push_back(NewCall);
701 // It's possible that inlining the callsite will cause it to go from an
702 // indirect to a direct call by resolving a function pointer. If this
703 // happens, set the callee of the new call site to a more precise
704 // destination. This can also happen if the call graph node of the caller
705 // was just unnecessarily imprecise.
706 if (!I->second->getFunction())
707 if (Function *F = CallSite(NewCall).getCalledFunction()) {
708 // Indirect call site resolved to direct call.
709 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
714 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
717 // Update the call graph by deleting the edge from Callee to Caller. We must
718 // do this after the loop above in case Caller and Callee are the same.
719 CallerNode->removeCallEdgeFor(CS);
722 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
723 BasicBlock *InsertBlock,
724 InlineFunctionInfo &IFI) {
725 Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
726 IRBuilder<> Builder(InsertBlock->begin());
729 if (IFI.DL == nullptr)
730 Size = ConstantExpr::getSizeOf(AggTy);
732 Size = Builder.getInt64(IFI.DL->getTypeStoreSize(AggTy));
734 // Always generate a memcpy of alignment 1 here because we don't know
735 // the alignment of the src pointer. Other optimizations can infer
737 Builder.CreateMemCpy(Dst, Src, Size, /*Align=*/1);
740 /// HandleByValArgument - When inlining a call site that has a byval argument,
741 /// we have to make the implicit memcpy explicit by adding it.
742 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
743 const Function *CalledFunc,
744 InlineFunctionInfo &IFI,
745 unsigned ByValAlignment) {
746 PointerType *ArgTy = cast<PointerType>(Arg->getType());
747 Type *AggTy = ArgTy->getElementType();
749 Function *Caller = TheCall->getParent()->getParent();
751 // If the called function is readonly, then it could not mutate the caller's
752 // copy of the byval'd memory. In this case, it is safe to elide the copy and
754 if (CalledFunc->onlyReadsMemory()) {
755 // If the byval argument has a specified alignment that is greater than the
756 // passed in pointer, then we either have to round up the input pointer or
757 // give up on this transformation.
758 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
761 // If the pointer is already known to be sufficiently aligned, or if we can
762 // round it up to a larger alignment, then we don't need a temporary.
763 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, IFI.DL,
764 &IFI.ACT->getAssumptionCache(*Caller),
765 TheCall) >= ByValAlignment)
768 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
769 // for code quality, but rarely happens and is required for correctness.
772 // Create the alloca. If we have DataLayout, use nice alignment.
775 Align = IFI.DL->getPrefTypeAlignment(AggTy);
777 // If the byval had an alignment specified, we *must* use at least that
778 // alignment, as it is required by the byval argument (and uses of the
779 // pointer inside the callee).
780 Align = std::max(Align, ByValAlignment);
782 Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(),
783 &*Caller->begin()->begin());
784 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
786 // Uses of the argument in the function should use our new alloca
791 // isUsedByLifetimeMarker - Check whether this Value is used by a lifetime
793 static bool isUsedByLifetimeMarker(Value *V) {
794 for (User *U : V->users()) {
795 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
796 switch (II->getIntrinsicID()) {
798 case Intrinsic::lifetime_start:
799 case Intrinsic::lifetime_end:
807 // hasLifetimeMarkers - Check whether the given alloca already has
808 // lifetime.start or lifetime.end intrinsics.
809 static bool hasLifetimeMarkers(AllocaInst *AI) {
810 Type *Ty = AI->getType();
811 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
812 Ty->getPointerAddressSpace());
814 return isUsedByLifetimeMarker(AI);
816 // Do a scan to find all the casts to i8*.
817 for (User *U : AI->users()) {
818 if (U->getType() != Int8PtrTy) continue;
819 if (U->stripPointerCasts() != AI) continue;
820 if (isUsedByLifetimeMarker(U))
826 /// updateInlinedAtInfo - Helper function used by fixupLineNumbers to
827 /// recursively update InlinedAtEntry of a DebugLoc.
828 static DebugLoc updateInlinedAtInfo(const DebugLoc &DL,
829 const DebugLoc &InlinedAtDL,
831 if (MDNode *IA = DL.getInlinedAt(Ctx)) {
832 DebugLoc NewInlinedAtDL
833 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx);
834 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
835 NewInlinedAtDL.getAsMDNode(Ctx));
838 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
839 InlinedAtDL.getAsMDNode(Ctx));
842 /// fixupLineNumbers - Update inlined instructions' line numbers to
843 /// to encode location where these instructions are inlined.
844 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
845 Instruction *TheCall) {
846 DebugLoc TheCallDL = TheCall->getDebugLoc();
847 if (TheCallDL.isUnknown())
850 for (; FI != Fn->end(); ++FI) {
851 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
853 DebugLoc DL = BI->getDebugLoc();
854 if (DL.isUnknown()) {
855 // If the inlined instruction has no line number, make it look as if it
856 // originates from the call location. This is important for
857 // ((__always_inline__, __nodebug__)) functions which must use caller
858 // location for all instructions in their function body.
860 // Don't update static allocas, as they may get moved later.
861 if (auto *AI = dyn_cast<AllocaInst>(BI))
862 if (isa<Constant>(AI->getArraySize()))
865 BI->setDebugLoc(TheCallDL);
867 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext()));
868 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) {
869 LLVMContext &Ctx = BI->getContext();
870 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
871 DVI->setOperand(2, MetadataAsValue::get(
872 Ctx, createInlinedVariable(DVI->getVariable(),
874 } else if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(BI)) {
875 LLVMContext &Ctx = BI->getContext();
876 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
877 DDI->setOperand(1, MetadataAsValue::get(
878 Ctx, createInlinedVariable(DDI->getVariable(),
886 /// InlineFunction - This function inlines the called function into the basic
887 /// block of the caller. This returns false if it is not possible to inline
888 /// this call. The program is still in a well defined state if this occurs
891 /// Note that this only does one level of inlining. For example, if the
892 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
893 /// exists in the instruction stream. Similarly this will inline a recursive
894 /// function by one level.
895 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
896 bool InsertLifetime) {
897 Instruction *TheCall = CS.getInstruction();
898 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
899 "Instruction not in function!");
901 // If IFI has any state in it, zap it before we fill it in.
904 const Function *CalledFunc = CS.getCalledFunction();
905 if (!CalledFunc || // Can't inline external function or indirect
906 CalledFunc->isDeclaration() || // call, or call to a vararg function!
907 CalledFunc->getFunctionType()->isVarArg()) return false;
909 // If the call to the callee cannot throw, set the 'nounwind' flag on any
910 // calls that we inline.
911 bool MarkNoUnwind = CS.doesNotThrow();
913 BasicBlock *OrigBB = TheCall->getParent();
914 Function *Caller = OrigBB->getParent();
916 // GC poses two hazards to inlining, which only occur when the callee has GC:
917 // 1. If the caller has no GC, then the callee's GC must be propagated to the
919 // 2. If the caller has a differing GC, it is invalid to inline.
920 if (CalledFunc->hasGC()) {
921 if (!Caller->hasGC())
922 Caller->setGC(CalledFunc->getGC());
923 else if (CalledFunc->getGC() != Caller->getGC())
927 // Get the personality function from the callee if it contains a landing pad.
928 Value *CalleePersonality = nullptr;
929 for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end();
931 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
932 const BasicBlock *BB = II->getUnwindDest();
933 const LandingPadInst *LP = BB->getLandingPadInst();
934 CalleePersonality = LP->getPersonalityFn();
938 // Find the personality function used by the landing pads of the caller. If it
939 // exists, then check to see that it matches the personality function used in
941 if (CalleePersonality) {
942 for (Function::const_iterator I = Caller->begin(), E = Caller->end();
944 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
945 const BasicBlock *BB = II->getUnwindDest();
946 const LandingPadInst *LP = BB->getLandingPadInst();
948 // If the personality functions match, then we can perform the
949 // inlining. Otherwise, we can't inline.
950 // TODO: This isn't 100% true. Some personality functions are proper
951 // supersets of others and can be used in place of the other.
952 if (LP->getPersonalityFn() != CalleePersonality)
959 // Get an iterator to the last basic block in the function, which will have
960 // the new function inlined after it.
961 Function::iterator LastBlock = &Caller->back();
963 // Make sure to capture all of the return instructions from the cloned
965 SmallVector<ReturnInst*, 8> Returns;
966 ClonedCodeInfo InlinedFunctionInfo;
967 Function::iterator FirstNewBlock;
969 { // Scope to destroy VMap after cloning.
970 ValueToValueMapTy VMap;
971 // Keep a list of pair (dst, src) to emit byval initializations.
972 SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
974 assert(CalledFunc->arg_size() == CS.arg_size() &&
975 "No varargs calls can be inlined!");
977 // Calculate the vector of arguments to pass into the function cloner, which
978 // matches up the formal to the actual argument values.
979 CallSite::arg_iterator AI = CS.arg_begin();
981 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
982 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
983 Value *ActualArg = *AI;
985 // When byval arguments actually inlined, we need to make the copy implied
986 // by them explicit. However, we don't do this if the callee is readonly
987 // or readnone, because the copy would be unneeded: the callee doesn't
988 // modify the struct.
989 if (CS.isByValArgument(ArgNo)) {
990 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
991 CalledFunc->getParamAlignment(ArgNo+1));
992 if (ActualArg != *AI)
993 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
999 // Add alignment assumptions if necessary. We do this before the inlined
1000 // instructions are actually cloned into the caller so that we can easily
1001 // check what will be known at the start of the inlined code.
1002 AddAlignmentAssumptions(CS, IFI);
1004 // We want the inliner to prune the code as it copies. We would LOVE to
1005 // have no dead or constant instructions leftover after inlining occurs
1006 // (which can happen, e.g., because an argument was constant), but we'll be
1007 // happy with whatever the cloner can do.
1008 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
1009 /*ModuleLevelChanges=*/false, Returns, ".i",
1010 &InlinedFunctionInfo, IFI.DL, TheCall);
1012 // Remember the first block that is newly cloned over.
1013 FirstNewBlock = LastBlock; ++FirstNewBlock;
1015 // Inject byval arguments initialization.
1016 for (std::pair<Value*, Value*> &Init : ByValInit)
1017 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
1018 FirstNewBlock, IFI);
1020 // Update the callgraph if requested.
1022 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
1024 // Update inlined instructions' line number information.
1025 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
1027 // Clone existing noalias metadata if necessary.
1028 CloneAliasScopeMetadata(CS, VMap);
1030 // Add noalias metadata if necessary.
1031 AddAliasScopeMetadata(CS, VMap, IFI.DL, IFI.AA);
1033 // FIXME: We could register any cloned assumptions instead of clearing the
1034 // whole function's cache.
1036 IFI.ACT->getAssumptionCache(*Caller).clear();
1039 // If there are any alloca instructions in the block that used to be the entry
1040 // block for the callee, move them to the entry block of the caller. First
1041 // calculate which instruction they should be inserted before. We insert the
1042 // instructions at the end of the current alloca list.
1044 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
1045 for (BasicBlock::iterator I = FirstNewBlock->begin(),
1046 E = FirstNewBlock->end(); I != E; ) {
1047 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
1050 // If the alloca is now dead, remove it. This often occurs due to code
1052 if (AI->use_empty()) {
1053 AI->eraseFromParent();
1057 if (!isa<Constant>(AI->getArraySize()))
1060 // Keep track of the static allocas that we inline into the caller.
1061 IFI.StaticAllocas.push_back(AI);
1063 // Scan for the block of allocas that we can move over, and move them
1065 while (isa<AllocaInst>(I) &&
1066 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
1067 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
1071 // Transfer all of the allocas over in a block. Using splice means
1072 // that the instructions aren't removed from the symbol table, then
1074 Caller->getEntryBlock().getInstList().splice(InsertPoint,
1075 FirstNewBlock->getInstList(),
1080 bool InlinedMustTailCalls = false;
1081 if (InlinedFunctionInfo.ContainsCalls) {
1082 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
1083 if (CallInst *CI = dyn_cast<CallInst>(TheCall))
1084 CallSiteTailKind = CI->getTailCallKind();
1086 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
1088 for (Instruction &I : *BB) {
1089 CallInst *CI = dyn_cast<CallInst>(&I);
1093 // We need to reduce the strength of any inlined tail calls. For
1094 // musttail, we have to avoid introducing potential unbounded stack
1095 // growth. For example, if functions 'f' and 'g' are mutually recursive
1096 // with musttail, we can inline 'g' into 'f' so long as we preserve
1097 // musttail on the cloned call to 'f'. If either the inlined call site
1098 // or the cloned call site is *not* musttail, the program already has
1099 // one frame of stack growth, so it's safe to remove musttail. Here is
1100 // a table of example transformations:
1102 // f -> musttail g -> musttail f ==> f -> musttail f
1103 // f -> musttail g -> tail f ==> f -> tail f
1104 // f -> g -> musttail f ==> f -> f
1105 // f -> g -> tail f ==> f -> f
1106 CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
1107 ChildTCK = std::min(CallSiteTailKind, ChildTCK);
1108 CI->setTailCallKind(ChildTCK);
1109 InlinedMustTailCalls |= CI->isMustTailCall();
1111 // Calls inlined through a 'nounwind' call site should be marked
1114 CI->setDoesNotThrow();
1119 // Leave lifetime markers for the static alloca's, scoping them to the
1120 // function we just inlined.
1121 if (InsertLifetime && !IFI.StaticAllocas.empty()) {
1122 IRBuilder<> builder(FirstNewBlock->begin());
1123 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
1124 AllocaInst *AI = IFI.StaticAllocas[ai];
1126 // If the alloca is already scoped to something smaller than the whole
1127 // function then there's no need to add redundant, less accurate markers.
1128 if (hasLifetimeMarkers(AI))
1131 // Try to determine the size of the allocation.
1132 ConstantInt *AllocaSize = nullptr;
1133 if (ConstantInt *AIArraySize =
1134 dyn_cast<ConstantInt>(AI->getArraySize())) {
1136 Type *AllocaType = AI->getAllocatedType();
1137 uint64_t AllocaTypeSize = IFI.DL->getTypeAllocSize(AllocaType);
1138 uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
1139 assert(AllocaArraySize > 0 && "array size of AllocaInst is zero");
1140 // Check that array size doesn't saturate uint64_t and doesn't
1141 // overflow when it's multiplied by type size.
1142 if (AllocaArraySize != ~0ULL &&
1143 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
1144 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
1145 AllocaArraySize * AllocaTypeSize);
1150 builder.CreateLifetimeStart(AI, AllocaSize);
1151 for (ReturnInst *RI : Returns) {
1152 // Don't insert llvm.lifetime.end calls between a musttail call and a
1153 // return. The return kills all local allocas.
1154 if (InlinedMustTailCalls &&
1155 RI->getParent()->getTerminatingMustTailCall())
1157 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
1162 // If the inlined code contained dynamic alloca instructions, wrap the inlined
1163 // code with llvm.stacksave/llvm.stackrestore intrinsics.
1164 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
1165 Module *M = Caller->getParent();
1166 // Get the two intrinsics we care about.
1167 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
1168 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
1170 // Insert the llvm.stacksave.
1171 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
1172 .CreateCall(StackSave, "savedstack");
1174 // Insert a call to llvm.stackrestore before any return instructions in the
1175 // inlined function.
1176 for (ReturnInst *RI : Returns) {
1177 // Don't insert llvm.stackrestore calls between a musttail call and a
1178 // return. The return will restore the stack pointer.
1179 if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
1181 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
1185 // If we are inlining for an invoke instruction, we must make sure to rewrite
1186 // any call instructions into invoke instructions.
1187 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
1188 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
1190 // Handle any inlined musttail call sites. In order for a new call site to be
1191 // musttail, the source of the clone and the inlined call site must have been
1192 // musttail. Therefore it's safe to return without merging control into the
1194 if (InlinedMustTailCalls) {
1195 // Check if we need to bitcast the result of any musttail calls.
1196 Type *NewRetTy = Caller->getReturnType();
1197 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
1199 // Handle the returns preceded by musttail calls separately.
1200 SmallVector<ReturnInst *, 8> NormalReturns;
1201 for (ReturnInst *RI : Returns) {
1202 CallInst *ReturnedMustTail =
1203 RI->getParent()->getTerminatingMustTailCall();
1204 if (!ReturnedMustTail) {
1205 NormalReturns.push_back(RI);
1211 // Delete the old return and any preceding bitcast.
1212 BasicBlock *CurBB = RI->getParent();
1213 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
1214 RI->eraseFromParent();
1216 OldCast->eraseFromParent();
1218 // Insert a new bitcast and return with the right type.
1219 IRBuilder<> Builder(CurBB);
1220 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
1223 // Leave behind the normal returns so we can merge control flow.
1224 std::swap(Returns, NormalReturns);
1227 // If we cloned in _exactly one_ basic block, and if that block ends in a
1228 // return instruction, we splice the body of the inlined callee directly into
1229 // the calling basic block.
1230 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
1231 // Move all of the instructions right before the call.
1232 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
1233 FirstNewBlock->begin(), FirstNewBlock->end());
1234 // Remove the cloned basic block.
1235 Caller->getBasicBlockList().pop_back();
1237 // If the call site was an invoke instruction, add a branch to the normal
1239 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1240 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
1241 NewBr->setDebugLoc(Returns[0]->getDebugLoc());
1244 // If the return instruction returned a value, replace uses of the call with
1245 // uses of the returned value.
1246 if (!TheCall->use_empty()) {
1247 ReturnInst *R = Returns[0];
1248 if (TheCall == R->getReturnValue())
1249 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1251 TheCall->replaceAllUsesWith(R->getReturnValue());
1253 // Since we are now done with the Call/Invoke, we can delete it.
1254 TheCall->eraseFromParent();
1256 // Since we are now done with the return instruction, delete it also.
1257 Returns[0]->eraseFromParent();
1259 // We are now done with the inlining.
1263 // Otherwise, we have the normal case, of more than one block to inline or
1264 // multiple return sites.
1266 // We want to clone the entire callee function into the hole between the
1267 // "starter" and "ender" blocks. How we accomplish this depends on whether
1268 // this is an invoke instruction or a call instruction.
1269 BasicBlock *AfterCallBB;
1270 BranchInst *CreatedBranchToNormalDest = nullptr;
1271 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1273 // Add an unconditional branch to make this look like the CallInst case...
1274 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
1276 // Split the basic block. This guarantees that no PHI nodes will have to be
1277 // updated due to new incoming edges, and make the invoke case more
1278 // symmetric to the call case.
1279 AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest,
1280 CalledFunc->getName()+".exit");
1282 } else { // It's a call
1283 // If this is a call instruction, we need to split the basic block that
1284 // the call lives in.
1286 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
1287 CalledFunc->getName()+".exit");
1290 // Change the branch that used to go to AfterCallBB to branch to the first
1291 // basic block of the inlined function.
1293 TerminatorInst *Br = OrigBB->getTerminator();
1294 assert(Br && Br->getOpcode() == Instruction::Br &&
1295 "splitBasicBlock broken!");
1296 Br->setOperand(0, FirstNewBlock);
1299 // Now that the function is correct, make it a little bit nicer. In
1300 // particular, move the basic blocks inserted from the end of the function
1301 // into the space made by splitting the source basic block.
1302 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
1303 FirstNewBlock, Caller->end());
1305 // Handle all of the return instructions that we just cloned in, and eliminate
1306 // any users of the original call/invoke instruction.
1307 Type *RTy = CalledFunc->getReturnType();
1309 PHINode *PHI = nullptr;
1310 if (Returns.size() > 1) {
1311 // The PHI node should go at the front of the new basic block to merge all
1312 // possible incoming values.
1313 if (!TheCall->use_empty()) {
1314 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
1315 AfterCallBB->begin());
1316 // Anything that used the result of the function call should now use the
1317 // PHI node as their operand.
1318 TheCall->replaceAllUsesWith(PHI);
1321 // Loop over all of the return instructions adding entries to the PHI node
1324 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1325 ReturnInst *RI = Returns[i];
1326 assert(RI->getReturnValue()->getType() == PHI->getType() &&
1327 "Ret value not consistent in function!");
1328 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
1333 // Add a branch to the merge points and remove return instructions.
1335 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1336 ReturnInst *RI = Returns[i];
1337 BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
1338 Loc = RI->getDebugLoc();
1339 BI->setDebugLoc(Loc);
1340 RI->eraseFromParent();
1342 // We need to set the debug location to *somewhere* inside the
1343 // inlined function. The line number may be nonsensical, but the
1344 // instruction will at least be associated with the right
1346 if (CreatedBranchToNormalDest)
1347 CreatedBranchToNormalDest->setDebugLoc(Loc);
1348 } else if (!Returns.empty()) {
1349 // Otherwise, if there is exactly one return value, just replace anything
1350 // using the return value of the call with the computed value.
1351 if (!TheCall->use_empty()) {
1352 if (TheCall == Returns[0]->getReturnValue())
1353 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1355 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
1358 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
1359 BasicBlock *ReturnBB = Returns[0]->getParent();
1360 ReturnBB->replaceAllUsesWith(AfterCallBB);
1362 // Splice the code from the return block into the block that it will return
1363 // to, which contains the code that was after the call.
1364 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
1365 ReturnBB->getInstList());
1367 if (CreatedBranchToNormalDest)
1368 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
1370 // Delete the return instruction now and empty ReturnBB now.
1371 Returns[0]->eraseFromParent();
1372 ReturnBB->eraseFromParent();
1373 } else if (!TheCall->use_empty()) {
1374 // No returns, but something is using the return value of the call. Just
1376 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1379 // Since we are now done with the Call/Invoke, we can delete it.
1380 TheCall->eraseFromParent();
1382 // If we inlined any musttail calls and the original return is now
1383 // unreachable, delete it. It can only contain a bitcast and ret.
1384 if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB))
1385 AfterCallBB->eraseFromParent();
1387 // We should always be able to fold the entry block of the function into the
1388 // single predecessor of the block...
1389 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
1390 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
1392 // Splice the code entry block into calling block, right before the
1393 // unconditional branch.
1394 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
1395 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
1397 // Remove the unconditional branch.
1398 OrigBB->getInstList().erase(Br);
1400 // Now we can remove the CalleeEntry block, which is now empty.
1401 Caller->getBasicBlockList().erase(CalleeEntry);
1403 // If we inserted a phi node, check to see if it has a single value (e.g. all
1404 // the entries are the same or undef). If so, remove the PHI so it doesn't
1405 // block other optimizations.
1407 if (Value *V = SimplifyInstruction(PHI, IFI.DL, nullptr, nullptr,
1408 &IFI.ACT->getAssumptionCache(*Caller))) {
1409 PHI->replaceAllUsesWith(V);
1410 PHI->eraseFromParent();