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"
48 EnableNoAliasConversion("enable-noalias-to-md-conversion", cl::init(true),
50 cl::desc("Convert noalias attributes to metadata during inlining."));
53 PreserveAlignmentAssumptions("preserve-alignment-assumptions-during-inlining",
54 cl::init(true), cl::Hidden,
55 cl::desc("Convert align attributes to assumptions during inlining."));
57 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI,
58 AAResults *CalleeAAR, bool InsertLifetime) {
59 return InlineFunction(CallSite(CI), IFI, CalleeAAR, InsertLifetime);
61 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI,
62 AAResults *CalleeAAR, bool InsertLifetime) {
63 return InlineFunction(CallSite(II), IFI, CalleeAAR, InsertLifetime);
67 /// A class for recording information about inlining a landing pad.
68 class LandingPadInliningInfo {
69 BasicBlock *OuterResumeDest; ///< Destination of the invoke's unwind.
70 BasicBlock *InnerResumeDest; ///< Destination for the callee's resume.
71 LandingPadInst *CallerLPad; ///< LandingPadInst associated with the invoke.
72 PHINode *InnerEHValuesPHI; ///< PHI for EH values from landingpad insts.
73 SmallVector<Value*, 8> UnwindDestPHIValues;
76 LandingPadInliningInfo(InvokeInst *II)
77 : OuterResumeDest(II->getUnwindDest()), InnerResumeDest(nullptr),
78 CallerLPad(nullptr), InnerEHValuesPHI(nullptr) {
79 // If there are PHI nodes in the unwind destination block, we need to keep
80 // track of which values came into them from the invoke before removing
81 // the edge from this block.
82 llvm::BasicBlock *InvokeBB = II->getParent();
83 BasicBlock::iterator I = OuterResumeDest->begin();
84 for (; isa<PHINode>(I); ++I) {
85 // Save the value to use for this edge.
86 PHINode *PHI = cast<PHINode>(I);
87 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
90 CallerLPad = cast<LandingPadInst>(I);
93 /// The outer unwind destination is the target of
94 /// unwind edges introduced for calls within the inlined function.
95 BasicBlock *getOuterResumeDest() const {
96 return OuterResumeDest;
99 BasicBlock *getInnerResumeDest();
101 LandingPadInst *getLandingPadInst() const { return CallerLPad; }
103 /// Forward the 'resume' instruction to the caller's landing pad block.
104 /// When the landing pad block has only one predecessor, this is
105 /// a simple branch. When there is more than one predecessor, we need to
106 /// split the landing pad block after the landingpad instruction and jump
108 void forwardResume(ResumeInst *RI,
109 SmallPtrSetImpl<LandingPadInst*> &InlinedLPads);
111 /// Add incoming-PHI values to the unwind destination block for the given
112 /// basic block, using the values for the 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);
125 } // anonymous namespace
127 /// Get or create a target for the branch from ResumeInsts.
128 BasicBlock *LandingPadInliningInfo::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 /// Forward the 'resume' instruction to the caller's landing pad block.
163 /// 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 LandingPadInliningInfo::forwardResume(
167 ResumeInst *RI, 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 /// When we inline a basic block into an invoke,
182 /// we have to turn all of the calls that can throw into invokes.
183 /// 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.
187 HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB, BasicBlock *UnwindEdge) {
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, UnwindEdge,
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();
227 /// If we inlined an invoke site, we need to convert calls
228 /// in the body of the inlined function into invokes.
230 /// II is the invoke instruction being inlined. FirstNewBlock is the first
231 /// block of the inlined code (the last block is the end of the function),
232 /// and InlineCodeInfo is information about the code that got inlined.
233 static void HandleInlinedLandingPad(InvokeInst *II, BasicBlock *FirstNewBlock,
234 ClonedCodeInfo &InlinedCodeInfo) {
235 BasicBlock *InvokeDest = II->getUnwindDest();
237 Function *Caller = FirstNewBlock->getParent();
239 // The inlined code is currently at the end of the function, scan from the
240 // start of the inlined code to its end, checking for stuff we need to
242 LandingPadInliningInfo Invoke(II);
244 // Get all of the inlined landing pad instructions.
245 SmallPtrSet<LandingPadInst*, 16> InlinedLPads;
246 for (Function::iterator I = FirstNewBlock, E = Caller->end(); I != E; ++I)
247 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator()))
248 InlinedLPads.insert(II->getLandingPadInst());
250 // Append the clauses from the outer landing pad instruction into the inlined
251 // landing pad instructions.
252 LandingPadInst *OuterLPad = Invoke.getLandingPadInst();
253 for (LandingPadInst *InlinedLPad : InlinedLPads) {
254 unsigned OuterNum = OuterLPad->getNumClauses();
255 InlinedLPad->reserveClauses(OuterNum);
256 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx)
257 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx));
258 if (OuterLPad->isCleanup())
259 InlinedLPad->setCleanup(true);
262 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
263 if (InlinedCodeInfo.ContainsCalls)
264 if (BasicBlock *NewBB = HandleCallsInBlockInlinedThroughInvoke(
265 BB, Invoke.getOuterResumeDest()))
266 // Update any PHI nodes in the exceptional block to indicate that there
267 // is now a new entry in them.
268 Invoke.addIncomingPHIValuesFor(NewBB);
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 /// If we inlined an invoke site, we need to convert calls
283 /// in the body of the inlined function into invokes.
285 /// II is the invoke instruction being inlined. FirstNewBlock is the first
286 /// block of the inlined code (the last block is the end of the function),
287 /// and InlineCodeInfo is information about the code that got inlined.
288 static void HandleInlinedEHPad(InvokeInst *II, BasicBlock *FirstNewBlock,
289 ClonedCodeInfo &InlinedCodeInfo) {
290 BasicBlock *UnwindDest = II->getUnwindDest();
291 Function *Caller = FirstNewBlock->getParent();
293 assert(UnwindDest->getFirstNonPHI()->isEHPad() && "unexpected BasicBlock!");
295 // If there are PHI nodes in the unwind destination block, we need to keep
296 // track of which values came into them from the invoke before removing the
297 // edge from this block.
298 SmallVector<Value *, 8> UnwindDestPHIValues;
299 llvm::BasicBlock *InvokeBB = II->getParent();
300 for (Instruction &I : *UnwindDest) {
301 // Save the value to use for this edge.
302 PHINode *PHI = dyn_cast<PHINode>(&I);
305 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
308 // Add incoming-PHI values to the unwind destination block for the given basic
309 // block, using the values for the original invoke's source block.
310 auto UpdatePHINodes = [&](BasicBlock *Src) {
311 BasicBlock::iterator I = UnwindDest->begin();
312 for (Value *V : UnwindDestPHIValues) {
313 PHINode *PHI = cast<PHINode>(I);
314 PHI->addIncoming(V, Src);
319 // Forward EH terminator instructions to the caller's invoke destination.
320 // This is as simple as connect all the instructions which 'unwind to caller'
321 // to the invoke destination.
322 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
324 Instruction *I = BB->getFirstNonPHI();
326 if (auto *CEPI = dyn_cast<CatchEndPadInst>(I)) {
327 if (CEPI->unwindsToCaller()) {
328 CatchEndPadInst::Create(CEPI->getContext(), UnwindDest, CEPI);
329 CEPI->eraseFromParent();
332 } else if (auto *CEPI = dyn_cast<CleanupEndPadInst>(I)) {
333 if (CEPI->unwindsToCaller()) {
334 CleanupEndPadInst::Create(CEPI->getCleanupPad(), UnwindDest, CEPI);
335 CEPI->eraseFromParent();
338 } else if (auto *TPI = dyn_cast<TerminatePadInst>(I)) {
339 if (TPI->unwindsToCaller()) {
340 SmallVector<Value *, 3> TerminatePadArgs;
341 for (Value *Operand : TPI->operands())
342 TerminatePadArgs.push_back(Operand);
343 TerminatePadInst::Create(TPI->getContext(), UnwindDest, TPI);
344 TPI->eraseFromParent();
348 assert(isa<CatchPadInst>(I) || isa<CleanupPadInst>(I));
352 if (auto *CRI = dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
353 if (CRI->unwindsToCaller()) {
354 CleanupReturnInst::Create(CRI->getCleanupPad(), UnwindDest, CRI);
355 CRI->eraseFromParent();
361 if (InlinedCodeInfo.ContainsCalls)
362 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
364 if (BasicBlock *NewBB =
365 HandleCallsInBlockInlinedThroughInvoke(BB, UnwindDest))
366 // Update any PHI nodes in the exceptional block to indicate that there
367 // is now a new entry in them.
368 UpdatePHINodes(NewBB);
370 // Now that everything is happy, we have one final detail. The PHI nodes in
371 // the exception destination block still have entries due to the original
372 // invoke instruction. Eliminate these entries (which might even delete the
374 UnwindDest->removePredecessor(InvokeBB);
377 /// When inlining a function that contains noalias scope metadata,
378 /// this metadata needs to be cloned so that the inlined blocks
379 /// have different "unqiue scopes" at every call site. Were this not done, then
380 /// aliasing scopes from a function inlined into a caller multiple times could
381 /// not be differentiated (and this would lead to miscompiles because the
382 /// non-aliasing property communicated by the metadata could have
383 /// call-site-specific control dependencies).
384 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) {
385 const Function *CalledFunc = CS.getCalledFunction();
386 SetVector<const MDNode *> MD;
388 // Note: We could only clone the metadata if it is already used in the
389 // caller. I'm omitting that check here because it might confuse
390 // inter-procedural alias analysis passes. We can revisit this if it becomes
391 // an efficiency or overhead problem.
393 for (Function::const_iterator I = CalledFunc->begin(), IE = CalledFunc->end();
395 for (BasicBlock::const_iterator J = I->begin(), JE = I->end(); J != JE; ++J) {
396 if (const MDNode *M = J->getMetadata(LLVMContext::MD_alias_scope))
398 if (const MDNode *M = J->getMetadata(LLVMContext::MD_noalias))
405 // Walk the existing metadata, adding the complete (perhaps cyclic) chain to
407 SmallVector<const Metadata *, 16> Queue(MD.begin(), MD.end());
408 while (!Queue.empty()) {
409 const MDNode *M = cast<MDNode>(Queue.pop_back_val());
410 for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i)
411 if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i)))
416 // Now we have a complete set of all metadata in the chains used to specify
417 // the noalias scopes and the lists of those scopes.
418 SmallVector<TempMDTuple, 16> DummyNodes;
419 DenseMap<const MDNode *, TrackingMDNodeRef> MDMap;
420 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
422 DummyNodes.push_back(MDTuple::getTemporary(CalledFunc->getContext(), None));
423 MDMap[*I].reset(DummyNodes.back().get());
426 // Create new metadata nodes to replace the dummy nodes, replacing old
427 // metadata references with either a dummy node or an already-created new
429 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
431 SmallVector<Metadata *, 4> NewOps;
432 for (unsigned i = 0, ie = (*I)->getNumOperands(); i != ie; ++i) {
433 const Metadata *V = (*I)->getOperand(i);
434 if (const MDNode *M = dyn_cast<MDNode>(V))
435 NewOps.push_back(MDMap[M]);
437 NewOps.push_back(const_cast<Metadata *>(V));
440 MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps);
441 MDTuple *TempM = cast<MDTuple>(MDMap[*I]);
442 assert(TempM->isTemporary() && "Expected temporary node");
444 TempM->replaceAllUsesWith(NewM);
447 // Now replace the metadata in the new inlined instructions with the
448 // repacements from the map.
449 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
450 VMI != VMIE; ++VMI) {
454 Instruction *NI = dyn_cast<Instruction>(VMI->second);
458 if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope)) {
459 MDNode *NewMD = MDMap[M];
460 // If the call site also had alias scope metadata (a list of scopes to
461 // which instructions inside it might belong), propagate those scopes to
462 // the inlined instructions.
464 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
465 NewMD = MDNode::concatenate(NewMD, CSM);
466 NI->setMetadata(LLVMContext::MD_alias_scope, NewMD);
467 } else if (NI->mayReadOrWriteMemory()) {
469 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
470 NI->setMetadata(LLVMContext::MD_alias_scope, M);
473 if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias)) {
474 MDNode *NewMD = MDMap[M];
475 // If the call site also had noalias metadata (a list of scopes with
476 // which instructions inside it don't alias), propagate those scopes to
477 // the inlined instructions.
479 CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
480 NewMD = MDNode::concatenate(NewMD, CSM);
481 NI->setMetadata(LLVMContext::MD_noalias, NewMD);
482 } else if (NI->mayReadOrWriteMemory()) {
483 if (MDNode *M = CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
484 NI->setMetadata(LLVMContext::MD_noalias, M);
489 /// If the inlined function has noalias arguments,
490 /// then add new alias scopes for each noalias argument, tag the mapped noalias
491 /// parameters with noalias metadata specifying the new scope, and tag all
492 /// non-derived loads, stores and memory intrinsics with the new alias scopes.
493 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
494 const DataLayout &DL, AAResults *CalleeAAR) {
495 if (!EnableNoAliasConversion)
498 const Function *CalledFunc = CS.getCalledFunction();
499 SmallVector<const Argument *, 4> NoAliasArgs;
501 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
502 E = CalledFunc->arg_end(); I != E; ++I) {
503 if (I->hasNoAliasAttr() && !I->hasNUses(0))
504 NoAliasArgs.push_back(I);
507 if (NoAliasArgs.empty())
510 // To do a good job, if a noalias variable is captured, we need to know if
511 // the capture point dominates the particular use we're considering.
513 DT.recalculate(const_cast<Function&>(*CalledFunc));
515 // noalias indicates that pointer values based on the argument do not alias
516 // pointer values which are not based on it. So we add a new "scope" for each
517 // noalias function argument. Accesses using pointers based on that argument
518 // become part of that alias scope, accesses using pointers not based on that
519 // argument are tagged as noalias with that scope.
521 DenseMap<const Argument *, MDNode *> NewScopes;
522 MDBuilder MDB(CalledFunc->getContext());
524 // Create a new scope domain for this function.
526 MDB.createAnonymousAliasScopeDomain(CalledFunc->getName());
527 for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) {
528 const Argument *A = NoAliasArgs[i];
530 std::string Name = CalledFunc->getName();
533 Name += A->getName();
535 Name += ": argument ";
539 // Note: We always create a new anonymous root here. This is true regardless
540 // of the linkage of the callee because the aliasing "scope" is not just a
541 // property of the callee, but also all control dependencies in the caller.
542 MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
543 NewScopes.insert(std::make_pair(A, NewScope));
546 // Iterate over all new instructions in the map; for all memory-access
547 // instructions, add the alias scope metadata.
548 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
549 VMI != VMIE; ++VMI) {
550 if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) {
554 Instruction *NI = dyn_cast<Instruction>(VMI->second);
558 bool IsArgMemOnlyCall = false, IsFuncCall = false;
559 SmallVector<const Value *, 2> PtrArgs;
561 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
562 PtrArgs.push_back(LI->getPointerOperand());
563 else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
564 PtrArgs.push_back(SI->getPointerOperand());
565 else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
566 PtrArgs.push_back(VAAI->getPointerOperand());
567 else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
568 PtrArgs.push_back(CXI->getPointerOperand());
569 else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
570 PtrArgs.push_back(RMWI->getPointerOperand());
571 else if (ImmutableCallSite ICS = ImmutableCallSite(I)) {
572 // If we know that the call does not access memory, then we'll still
573 // know that about the inlined clone of this call site, and we don't
574 // need to add metadata.
575 if (ICS.doesNotAccessMemory())
580 FunctionModRefBehavior MRB = CalleeAAR->getModRefBehavior(ICS);
581 if (MRB == FMRB_OnlyAccessesArgumentPointees ||
582 MRB == FMRB_OnlyReadsArgumentPointees)
583 IsArgMemOnlyCall = true;
586 for (ImmutableCallSite::arg_iterator AI = ICS.arg_begin(),
587 AE = ICS.arg_end(); AI != AE; ++AI) {
588 // We need to check the underlying objects of all arguments, not just
589 // the pointer arguments, because we might be passing pointers as
591 // However, if we know that the call only accesses pointer arguments,
592 // then we only need to check the pointer arguments.
593 if (IsArgMemOnlyCall && !(*AI)->getType()->isPointerTy())
596 PtrArgs.push_back(*AI);
600 // If we found no pointers, then this instruction is not suitable for
601 // pairing with an instruction to receive aliasing metadata.
602 // However, if this is a call, this we might just alias with none of the
603 // noalias arguments.
604 if (PtrArgs.empty() && !IsFuncCall)
607 // It is possible that there is only one underlying object, but you
608 // need to go through several PHIs to see it, and thus could be
609 // repeated in the Objects list.
610 SmallPtrSet<const Value *, 4> ObjSet;
611 SmallVector<Metadata *, 4> Scopes, NoAliases;
613 SmallSetVector<const Argument *, 4> NAPtrArgs;
614 for (unsigned i = 0, ie = PtrArgs.size(); i != ie; ++i) {
615 SmallVector<Value *, 4> Objects;
616 GetUnderlyingObjects(const_cast<Value*>(PtrArgs[i]),
617 Objects, DL, /* LI = */ nullptr);
619 for (Value *O : Objects)
623 // Figure out if we're derived from anything that is not a noalias
625 bool CanDeriveViaCapture = false, UsesAliasingPtr = false;
626 for (const Value *V : ObjSet) {
627 // Is this value a constant that cannot be derived from any pointer
628 // value (we need to exclude constant expressions, for example, that
629 // are formed from arithmetic on global symbols).
630 bool IsNonPtrConst = isa<ConstantInt>(V) || isa<ConstantFP>(V) ||
631 isa<ConstantPointerNull>(V) ||
632 isa<ConstantDataVector>(V) || isa<UndefValue>(V);
636 // If this is anything other than a noalias argument, then we cannot
637 // completely describe the aliasing properties using alias.scope
638 // metadata (and, thus, won't add any).
639 if (const Argument *A = dyn_cast<Argument>(V)) {
640 if (!A->hasNoAliasAttr())
641 UsesAliasingPtr = true;
643 UsesAliasingPtr = true;
646 // If this is not some identified function-local object (which cannot
647 // directly alias a noalias argument), or some other argument (which,
648 // by definition, also cannot alias a noalias argument), then we could
649 // alias a noalias argument that has been captured).
650 if (!isa<Argument>(V) &&
651 !isIdentifiedFunctionLocal(const_cast<Value*>(V)))
652 CanDeriveViaCapture = true;
655 // A function call can always get captured noalias pointers (via other
656 // parameters, globals, etc.).
657 if (IsFuncCall && !IsArgMemOnlyCall)
658 CanDeriveViaCapture = true;
660 // First, we want to figure out all of the sets with which we definitely
661 // don't alias. Iterate over all noalias set, and add those for which:
662 // 1. The noalias argument is not in the set of objects from which we
663 // definitely derive.
664 // 2. The noalias argument has not yet been captured.
665 // An arbitrary function that might load pointers could see captured
666 // noalias arguments via other noalias arguments or globals, and so we
667 // must always check for prior capture.
668 for (const Argument *A : NoAliasArgs) {
669 if (!ObjSet.count(A) && (!CanDeriveViaCapture ||
670 // It might be tempting to skip the
671 // PointerMayBeCapturedBefore check if
672 // A->hasNoCaptureAttr() is true, but this is
673 // incorrect because nocapture only guarantees
674 // that no copies outlive the function, not
675 // that the value cannot be locally captured.
676 !PointerMayBeCapturedBefore(A,
677 /* ReturnCaptures */ false,
678 /* StoreCaptures */ false, I, &DT)))
679 NoAliases.push_back(NewScopes[A]);
682 if (!NoAliases.empty())
683 NI->setMetadata(LLVMContext::MD_noalias,
685 NI->getMetadata(LLVMContext::MD_noalias),
686 MDNode::get(CalledFunc->getContext(), NoAliases)));
688 // Next, we want to figure out all of the sets to which we might belong.
689 // We might belong to a set if the noalias argument is in the set of
690 // underlying objects. If there is some non-noalias argument in our list
691 // of underlying objects, then we cannot add a scope because the fact
692 // that some access does not alias with any set of our noalias arguments
693 // cannot itself guarantee that it does not alias with this access
694 // (because there is some pointer of unknown origin involved and the
695 // other access might also depend on this pointer). We also cannot add
696 // scopes to arbitrary functions unless we know they don't access any
697 // non-parameter pointer-values.
698 bool CanAddScopes = !UsesAliasingPtr;
699 if (CanAddScopes && IsFuncCall)
700 CanAddScopes = IsArgMemOnlyCall;
703 for (const Argument *A : NoAliasArgs) {
705 Scopes.push_back(NewScopes[A]);
710 LLVMContext::MD_alias_scope,
711 MDNode::concatenate(NI->getMetadata(LLVMContext::MD_alias_scope),
712 MDNode::get(CalledFunc->getContext(), Scopes)));
717 /// If the inlined function has non-byval align arguments, then
718 /// add @llvm.assume-based alignment assumptions to preserve this information.
719 static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) {
720 if (!PreserveAlignmentAssumptions)
722 auto &DL = CS.getCaller()->getParent()->getDataLayout();
724 // To avoid inserting redundant assumptions, we should check for assumptions
725 // already in the caller. To do this, we might need a DT of the caller.
727 bool DTCalculated = false;
729 Function *CalledFunc = CS.getCalledFunction();
730 for (Function::arg_iterator I = CalledFunc->arg_begin(),
731 E = CalledFunc->arg_end();
733 unsigned Align = I->getType()->isPointerTy() ? I->getParamAlignment() : 0;
734 if (Align && !I->hasByValOrInAllocaAttr() && !I->hasNUses(0)) {
736 DT.recalculate(const_cast<Function&>(*CS.getInstruction()->getParent()
741 // If we can already prove the asserted alignment in the context of the
742 // caller, then don't bother inserting the assumption.
743 Value *Arg = CS.getArgument(I->getArgNo());
744 if (getKnownAlignment(Arg, DL, CS.getInstruction(),
745 &IFI.ACT->getAssumptionCache(*CS.getCaller()),
749 IRBuilder<>(CS.getInstruction())
750 .CreateAlignmentAssumption(DL, Arg, Align);
755 /// Once we have cloned code over from a callee into the caller,
756 /// update the specified callgraph to reflect the changes we made.
757 /// Note that it's possible that not all code was copied over, so only
758 /// some edges of the callgraph may remain.
759 static void UpdateCallGraphAfterInlining(CallSite CS,
760 Function::iterator FirstNewBlock,
761 ValueToValueMapTy &VMap,
762 InlineFunctionInfo &IFI) {
763 CallGraph &CG = *IFI.CG;
764 const Function *Caller = CS.getInstruction()->getParent()->getParent();
765 const Function *Callee = CS.getCalledFunction();
766 CallGraphNode *CalleeNode = CG[Callee];
767 CallGraphNode *CallerNode = CG[Caller];
769 // Since we inlined some uninlined call sites in the callee into the caller,
770 // add edges from the caller to all of the callees of the callee.
771 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
773 // Consider the case where CalleeNode == CallerNode.
774 CallGraphNode::CalledFunctionsVector CallCache;
775 if (CalleeNode == CallerNode) {
776 CallCache.assign(I, E);
777 I = CallCache.begin();
781 for (; I != E; ++I) {
782 const Value *OrigCall = I->first;
784 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
785 // Only copy the edge if the call was inlined!
786 if (VMI == VMap.end() || VMI->second == nullptr)
789 // If the call was inlined, but then constant folded, there is no edge to
790 // add. Check for this case.
791 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
795 // We do not treat intrinsic calls like real function calls because we
796 // expect them to become inline code; do not add an edge for an intrinsic.
797 CallSite CS = CallSite(NewCall);
798 if (CS && CS.getCalledFunction() && CS.getCalledFunction()->isIntrinsic())
801 // Remember that this call site got inlined for the client of
803 IFI.InlinedCalls.push_back(NewCall);
805 // It's possible that inlining the callsite will cause it to go from an
806 // indirect to a direct call by resolving a function pointer. If this
807 // happens, set the callee of the new call site to a more precise
808 // destination. This can also happen if the call graph node of the caller
809 // was just unnecessarily imprecise.
810 if (!I->second->getFunction())
811 if (Function *F = CallSite(NewCall).getCalledFunction()) {
812 // Indirect call site resolved to direct call.
813 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
818 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
821 // Update the call graph by deleting the edge from Callee to Caller. We must
822 // do this after the loop above in case Caller and Callee are the same.
823 CallerNode->removeCallEdgeFor(CS);
826 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
827 BasicBlock *InsertBlock,
828 InlineFunctionInfo &IFI) {
829 Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
830 IRBuilder<> Builder(InsertBlock->begin());
832 Value *Size = Builder.getInt64(M->getDataLayout().getTypeStoreSize(AggTy));
834 // Always generate a memcpy of alignment 1 here because we don't know
835 // the alignment of the src pointer. Other optimizations can infer
837 Builder.CreateMemCpy(Dst, Src, Size, /*Align=*/1);
840 /// When inlining a call site that has a byval argument,
841 /// we have to make the implicit memcpy explicit by adding it.
842 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
843 const Function *CalledFunc,
844 InlineFunctionInfo &IFI,
845 unsigned ByValAlignment) {
846 PointerType *ArgTy = cast<PointerType>(Arg->getType());
847 Type *AggTy = ArgTy->getElementType();
849 Function *Caller = TheCall->getParent()->getParent();
851 // If the called function is readonly, then it could not mutate the caller's
852 // copy of the byval'd memory. In this case, it is safe to elide the copy and
854 if (CalledFunc->onlyReadsMemory()) {
855 // If the byval argument has a specified alignment that is greater than the
856 // passed in pointer, then we either have to round up the input pointer or
857 // give up on this transformation.
858 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
861 const DataLayout &DL = Caller->getParent()->getDataLayout();
863 // If the pointer is already known to be sufficiently aligned, or if we can
864 // round it up to a larger alignment, then we don't need a temporary.
865 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, DL, TheCall,
866 &IFI.ACT->getAssumptionCache(*Caller)) >=
870 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
871 // for code quality, but rarely happens and is required for correctness.
874 // Create the alloca. If we have DataLayout, use nice alignment.
876 Caller->getParent()->getDataLayout().getPrefTypeAlignment(AggTy);
878 // If the byval had an alignment specified, we *must* use at least that
879 // alignment, as it is required by the byval argument (and uses of the
880 // pointer inside the callee).
881 Align = std::max(Align, ByValAlignment);
883 Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(),
884 &*Caller->begin()->begin());
885 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
887 // Uses of the argument in the function should use our new alloca
892 // Check whether this Value is used by a lifetime intrinsic.
893 static bool isUsedByLifetimeMarker(Value *V) {
894 for (User *U : V->users()) {
895 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
896 switch (II->getIntrinsicID()) {
898 case Intrinsic::lifetime_start:
899 case Intrinsic::lifetime_end:
907 // Check whether the given alloca already has
908 // lifetime.start or lifetime.end intrinsics.
909 static bool hasLifetimeMarkers(AllocaInst *AI) {
910 Type *Ty = AI->getType();
911 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
912 Ty->getPointerAddressSpace());
914 return isUsedByLifetimeMarker(AI);
916 // Do a scan to find all the casts to i8*.
917 for (User *U : AI->users()) {
918 if (U->getType() != Int8PtrTy) continue;
919 if (U->stripPointerCasts() != AI) continue;
920 if (isUsedByLifetimeMarker(U))
926 /// Rebuild the entire inlined-at chain for this instruction so that the top of
927 /// the chain now is inlined-at the new call site.
929 updateInlinedAtInfo(DebugLoc DL, DILocation *InlinedAtNode, LLVMContext &Ctx,
930 DenseMap<const DILocation *, DILocation *> &IANodes) {
931 SmallVector<DILocation *, 3> InlinedAtLocations;
932 DILocation *Last = InlinedAtNode;
933 DILocation *CurInlinedAt = DL;
935 // Gather all the inlined-at nodes
936 while (DILocation *IA = CurInlinedAt->getInlinedAt()) {
937 // Skip any we've already built nodes for
938 if (DILocation *Found = IANodes[IA]) {
943 InlinedAtLocations.push_back(IA);
947 // Starting from the top, rebuild the nodes to point to the new inlined-at
948 // location (then rebuilding the rest of the chain behind it) and update the
949 // map of already-constructed inlined-at nodes.
950 for (const DILocation *MD : make_range(InlinedAtLocations.rbegin(),
951 InlinedAtLocations.rend())) {
952 Last = IANodes[MD] = DILocation::getDistinct(
953 Ctx, MD->getLine(), MD->getColumn(), MD->getScope(), Last);
956 // And finally create the normal location for this instruction, referring to
957 // the new inlined-at chain.
958 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(), Last);
961 /// Update inlined instructions' line numbers to
962 /// to encode location where these instructions are inlined.
963 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
964 Instruction *TheCall) {
965 DebugLoc TheCallDL = TheCall->getDebugLoc();
969 auto &Ctx = Fn->getContext();
970 DILocation *InlinedAtNode = TheCallDL;
972 // Create a unique call site, not to be confused with any other call from the
974 InlinedAtNode = DILocation::getDistinct(
975 Ctx, InlinedAtNode->getLine(), InlinedAtNode->getColumn(),
976 InlinedAtNode->getScope(), InlinedAtNode->getInlinedAt());
978 // Cache the inlined-at nodes as they're built so they are reused, without
979 // this every instruction's inlined-at chain would become distinct from each
981 DenseMap<const DILocation *, DILocation *> IANodes;
983 for (; FI != Fn->end(); ++FI) {
984 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
986 DebugLoc DL = BI->getDebugLoc();
988 // If the inlined instruction has no line number, make it look as if it
989 // originates from the call location. This is important for
990 // ((__always_inline__, __nodebug__)) functions which must use caller
991 // location for all instructions in their function body.
993 // Don't update static allocas, as they may get moved later.
994 if (auto *AI = dyn_cast<AllocaInst>(BI))
995 if (isa<Constant>(AI->getArraySize()))
998 BI->setDebugLoc(TheCallDL);
1000 BI->setDebugLoc(updateInlinedAtInfo(DL, InlinedAtNode, BI->getContext(), IANodes));
1006 /// This function inlines the called function into the basic block of the
1007 /// caller. This returns false if it is not possible to inline this call.
1008 /// The program is still in a well defined state if this occurs though.
1010 /// Note that this only does one level of inlining. For example, if the
1011 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
1012 /// exists in the instruction stream. Similarly this will inline a recursive
1013 /// function by one level.
1014 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
1015 AAResults *CalleeAAR, bool InsertLifetime) {
1016 Instruction *TheCall = CS.getInstruction();
1017 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
1018 "Instruction not in function!");
1020 // If IFI has any state in it, zap it before we fill it in.
1023 const Function *CalledFunc = CS.getCalledFunction();
1024 if (!CalledFunc || // Can't inline external function or indirect
1025 CalledFunc->isDeclaration() || // call, or call to a vararg function!
1026 CalledFunc->getFunctionType()->isVarArg()) return false;
1028 // If the call to the callee cannot throw, set the 'nounwind' flag on any
1029 // calls that we inline.
1030 bool MarkNoUnwind = CS.doesNotThrow();
1032 BasicBlock *OrigBB = TheCall->getParent();
1033 Function *Caller = OrigBB->getParent();
1035 // GC poses two hazards to inlining, which only occur when the callee has GC:
1036 // 1. If the caller has no GC, then the callee's GC must be propagated to the
1038 // 2. If the caller has a differing GC, it is invalid to inline.
1039 if (CalledFunc->hasGC()) {
1040 if (!Caller->hasGC())
1041 Caller->setGC(CalledFunc->getGC());
1042 else if (CalledFunc->getGC() != Caller->getGC())
1046 // Get the personality function from the callee if it contains a landing pad.
1047 Constant *CalledPersonality =
1048 CalledFunc->hasPersonalityFn() ? CalledFunc->getPersonalityFn() : nullptr;
1050 // Find the personality function used by the landing pads of the caller. If it
1051 // exists, then check to see that it matches the personality function used in
1053 Constant *CallerPersonality =
1054 Caller->hasPersonalityFn() ? Caller->getPersonalityFn() : nullptr;
1055 if (CalledPersonality) {
1056 if (!CallerPersonality)
1057 Caller->setPersonalityFn(CalledPersonality);
1058 // If the personality functions match, then we can perform the
1059 // inlining. Otherwise, we can't inline.
1060 // TODO: This isn't 100% true. Some personality functions are proper
1061 // supersets of others and can be used in place of the other.
1062 else if (CalledPersonality != CallerPersonality)
1066 // Get an iterator to the last basic block in the function, which will have
1067 // the new function inlined after it.
1068 Function::iterator LastBlock = &Caller->back();
1070 // Make sure to capture all of the return instructions from the cloned
1072 SmallVector<ReturnInst*, 8> Returns;
1073 ClonedCodeInfo InlinedFunctionInfo;
1074 Function::iterator FirstNewBlock;
1076 { // Scope to destroy VMap after cloning.
1077 ValueToValueMapTy VMap;
1078 // Keep a list of pair (dst, src) to emit byval initializations.
1079 SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
1081 auto &DL = Caller->getParent()->getDataLayout();
1083 assert(CalledFunc->arg_size() == CS.arg_size() &&
1084 "No varargs calls can be inlined!");
1086 // Calculate the vector of arguments to pass into the function cloner, which
1087 // matches up the formal to the actual argument values.
1088 CallSite::arg_iterator AI = CS.arg_begin();
1090 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
1091 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
1092 Value *ActualArg = *AI;
1094 // When byval arguments actually inlined, we need to make the copy implied
1095 // by them explicit. However, we don't do this if the callee is readonly
1096 // or readnone, because the copy would be unneeded: the callee doesn't
1097 // modify the struct.
1098 if (CS.isByValArgument(ArgNo)) {
1099 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
1100 CalledFunc->getParamAlignment(ArgNo+1));
1101 if (ActualArg != *AI)
1102 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
1105 VMap[I] = ActualArg;
1108 // Add alignment assumptions if necessary. We do this before the inlined
1109 // instructions are actually cloned into the caller so that we can easily
1110 // check what will be known at the start of the inlined code.
1111 AddAlignmentAssumptions(CS, IFI);
1113 // We want the inliner to prune the code as it copies. We would LOVE to
1114 // have no dead or constant instructions leftover after inlining occurs
1115 // (which can happen, e.g., because an argument was constant), but we'll be
1116 // happy with whatever the cloner can do.
1117 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
1118 /*ModuleLevelChanges=*/false, Returns, ".i",
1119 &InlinedFunctionInfo, TheCall);
1121 // Remember the first block that is newly cloned over.
1122 FirstNewBlock = LastBlock; ++FirstNewBlock;
1124 // Inject byval arguments initialization.
1125 for (std::pair<Value*, Value*> &Init : ByValInit)
1126 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
1127 FirstNewBlock, IFI);
1129 // Update the callgraph if requested.
1131 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
1133 // Update inlined instructions' line number information.
1134 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
1136 // Clone existing noalias metadata if necessary.
1137 CloneAliasScopeMetadata(CS, VMap);
1139 // Add noalias metadata if necessary.
1140 AddAliasScopeMetadata(CS, VMap, DL, CalleeAAR);
1142 // FIXME: We could register any cloned assumptions instead of clearing the
1143 // whole function's cache.
1145 IFI.ACT->getAssumptionCache(*Caller).clear();
1148 // If there are any alloca instructions in the block that used to be the entry
1149 // block for the callee, move them to the entry block of the caller. First
1150 // calculate which instruction they should be inserted before. We insert the
1151 // instructions at the end of the current alloca list.
1153 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
1154 for (BasicBlock::iterator I = FirstNewBlock->begin(),
1155 E = FirstNewBlock->end(); I != E; ) {
1156 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
1159 // If the alloca is now dead, remove it. This often occurs due to code
1161 if (AI->use_empty()) {
1162 AI->eraseFromParent();
1166 if (!isa<Constant>(AI->getArraySize()))
1169 // Keep track of the static allocas that we inline into the caller.
1170 IFI.StaticAllocas.push_back(AI);
1172 // Scan for the block of allocas that we can move over, and move them
1174 while (isa<AllocaInst>(I) &&
1175 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
1176 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
1180 // Transfer all of the allocas over in a block. Using splice means
1181 // that the instructions aren't removed from the symbol table, then
1183 Caller->getEntryBlock().getInstList().splice(InsertPoint,
1184 FirstNewBlock->getInstList(),
1187 // Move any dbg.declares describing the allocas into the entry basic block.
1188 DIBuilder DIB(*Caller->getParent());
1189 for (auto &AI : IFI.StaticAllocas)
1190 replaceDbgDeclareForAlloca(AI, AI, DIB, /*Deref=*/false);
1193 bool InlinedMustTailCalls = false;
1194 if (InlinedFunctionInfo.ContainsCalls) {
1195 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
1196 if (CallInst *CI = dyn_cast<CallInst>(TheCall))
1197 CallSiteTailKind = CI->getTailCallKind();
1199 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
1201 for (Instruction &I : *BB) {
1202 CallInst *CI = dyn_cast<CallInst>(&I);
1206 // We need to reduce the strength of any inlined tail calls. For
1207 // musttail, we have to avoid introducing potential unbounded stack
1208 // growth. For example, if functions 'f' and 'g' are mutually recursive
1209 // with musttail, we can inline 'g' into 'f' so long as we preserve
1210 // musttail on the cloned call to 'f'. If either the inlined call site
1211 // or the cloned call site is *not* musttail, the program already has
1212 // one frame of stack growth, so it's safe to remove musttail. Here is
1213 // a table of example transformations:
1215 // f -> musttail g -> musttail f ==> f -> musttail f
1216 // f -> musttail g -> tail f ==> f -> tail f
1217 // f -> g -> musttail f ==> f -> f
1218 // f -> g -> tail f ==> f -> f
1219 CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
1220 ChildTCK = std::min(CallSiteTailKind, ChildTCK);
1221 CI->setTailCallKind(ChildTCK);
1222 InlinedMustTailCalls |= CI->isMustTailCall();
1224 // Calls inlined through a 'nounwind' call site should be marked
1227 CI->setDoesNotThrow();
1232 // Leave lifetime markers for the static alloca's, scoping them to the
1233 // function we just inlined.
1234 if (InsertLifetime && !IFI.StaticAllocas.empty()) {
1235 IRBuilder<> builder(FirstNewBlock->begin());
1236 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
1237 AllocaInst *AI = IFI.StaticAllocas[ai];
1239 // If the alloca is already scoped to something smaller than the whole
1240 // function then there's no need to add redundant, less accurate markers.
1241 if (hasLifetimeMarkers(AI))
1244 // Try to determine the size of the allocation.
1245 ConstantInt *AllocaSize = nullptr;
1246 if (ConstantInt *AIArraySize =
1247 dyn_cast<ConstantInt>(AI->getArraySize())) {
1248 auto &DL = Caller->getParent()->getDataLayout();
1249 Type *AllocaType = AI->getAllocatedType();
1250 uint64_t AllocaTypeSize = DL.getTypeAllocSize(AllocaType);
1251 uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
1253 // Don't add markers for zero-sized allocas.
1254 if (AllocaArraySize == 0)
1257 // Check that array size doesn't saturate uint64_t and doesn't
1258 // overflow when it's multiplied by type size.
1259 if (AllocaArraySize != ~0ULL &&
1260 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
1261 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
1262 AllocaArraySize * AllocaTypeSize);
1266 builder.CreateLifetimeStart(AI, AllocaSize);
1267 for (ReturnInst *RI : Returns) {
1268 // Don't insert llvm.lifetime.end calls between a musttail call and a
1269 // return. The return kills all local allocas.
1270 if (InlinedMustTailCalls &&
1271 RI->getParent()->getTerminatingMustTailCall())
1273 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
1278 // If the inlined code contained dynamic alloca instructions, wrap the inlined
1279 // code with llvm.stacksave/llvm.stackrestore intrinsics.
1280 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
1281 Module *M = Caller->getParent();
1282 // Get the two intrinsics we care about.
1283 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
1284 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
1286 // Insert the llvm.stacksave.
1287 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
1288 .CreateCall(StackSave, {}, "savedstack");
1290 // Insert a call to llvm.stackrestore before any return instructions in the
1291 // inlined function.
1292 for (ReturnInst *RI : Returns) {
1293 // Don't insert llvm.stackrestore calls between a musttail call and a
1294 // return. The return will restore the stack pointer.
1295 if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
1297 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
1301 // If we are inlining for an invoke instruction, we must make sure to rewrite
1302 // any call instructions into invoke instructions.
1303 if (auto *II = dyn_cast<InvokeInst>(TheCall)) {
1304 BasicBlock *UnwindDest = II->getUnwindDest();
1305 Instruction *FirstNonPHI = UnwindDest->getFirstNonPHI();
1306 if (isa<LandingPadInst>(FirstNonPHI)) {
1307 HandleInlinedLandingPad(II, FirstNewBlock, InlinedFunctionInfo);
1309 HandleInlinedEHPad(II, FirstNewBlock, InlinedFunctionInfo);
1313 // Handle any inlined musttail call sites. In order for a new call site to be
1314 // musttail, the source of the clone and the inlined call site must have been
1315 // musttail. Therefore it's safe to return without merging control into the
1317 if (InlinedMustTailCalls) {
1318 // Check if we need to bitcast the result of any musttail calls.
1319 Type *NewRetTy = Caller->getReturnType();
1320 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
1322 // Handle the returns preceded by musttail calls separately.
1323 SmallVector<ReturnInst *, 8> NormalReturns;
1324 for (ReturnInst *RI : Returns) {
1325 CallInst *ReturnedMustTail =
1326 RI->getParent()->getTerminatingMustTailCall();
1327 if (!ReturnedMustTail) {
1328 NormalReturns.push_back(RI);
1334 // Delete the old return and any preceding bitcast.
1335 BasicBlock *CurBB = RI->getParent();
1336 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
1337 RI->eraseFromParent();
1339 OldCast->eraseFromParent();
1341 // Insert a new bitcast and return with the right type.
1342 IRBuilder<> Builder(CurBB);
1343 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
1346 // Leave behind the normal returns so we can merge control flow.
1347 std::swap(Returns, NormalReturns);
1350 // If we cloned in _exactly one_ basic block, and if that block ends in a
1351 // return instruction, we splice the body of the inlined callee directly into
1352 // the calling basic block.
1353 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
1354 // Move all of the instructions right before the call.
1355 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
1356 FirstNewBlock->begin(), FirstNewBlock->end());
1357 // Remove the cloned basic block.
1358 Caller->getBasicBlockList().pop_back();
1360 // If the call site was an invoke instruction, add a branch to the normal
1362 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1363 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
1364 NewBr->setDebugLoc(Returns[0]->getDebugLoc());
1367 // If the return instruction returned a value, replace uses of the call with
1368 // uses of the returned value.
1369 if (!TheCall->use_empty()) {
1370 ReturnInst *R = Returns[0];
1371 if (TheCall == R->getReturnValue())
1372 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1374 TheCall->replaceAllUsesWith(R->getReturnValue());
1376 // Since we are now done with the Call/Invoke, we can delete it.
1377 TheCall->eraseFromParent();
1379 // Since we are now done with the return instruction, delete it also.
1380 Returns[0]->eraseFromParent();
1382 // We are now done with the inlining.
1386 // Otherwise, we have the normal case, of more than one block to inline or
1387 // multiple return sites.
1389 // We want to clone the entire callee function into the hole between the
1390 // "starter" and "ender" blocks. How we accomplish this depends on whether
1391 // this is an invoke instruction or a call instruction.
1392 BasicBlock *AfterCallBB;
1393 BranchInst *CreatedBranchToNormalDest = nullptr;
1394 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1396 // Add an unconditional branch to make this look like the CallInst case...
1397 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
1399 // Split the basic block. This guarantees that no PHI nodes will have to be
1400 // updated due to new incoming edges, and make the invoke case more
1401 // symmetric to the call case.
1402 AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest,
1403 CalledFunc->getName()+".exit");
1405 } else { // It's a call
1406 // If this is a call instruction, we need to split the basic block that
1407 // the call lives in.
1409 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
1410 CalledFunc->getName()+".exit");
1413 // Change the branch that used to go to AfterCallBB to branch to the first
1414 // basic block of the inlined function.
1416 TerminatorInst *Br = OrigBB->getTerminator();
1417 assert(Br && Br->getOpcode() == Instruction::Br &&
1418 "splitBasicBlock broken!");
1419 Br->setOperand(0, FirstNewBlock);
1422 // Now that the function is correct, make it a little bit nicer. In
1423 // particular, move the basic blocks inserted from the end of the function
1424 // into the space made by splitting the source basic block.
1425 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
1426 FirstNewBlock, Caller->end());
1428 // Handle all of the return instructions that we just cloned in, and eliminate
1429 // any users of the original call/invoke instruction.
1430 Type *RTy = CalledFunc->getReturnType();
1432 PHINode *PHI = nullptr;
1433 if (Returns.size() > 1) {
1434 // The PHI node should go at the front of the new basic block to merge all
1435 // possible incoming values.
1436 if (!TheCall->use_empty()) {
1437 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
1438 AfterCallBB->begin());
1439 // Anything that used the result of the function call should now use the
1440 // PHI node as their operand.
1441 TheCall->replaceAllUsesWith(PHI);
1444 // Loop over all of the return instructions adding entries to the PHI node
1447 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1448 ReturnInst *RI = Returns[i];
1449 assert(RI->getReturnValue()->getType() == PHI->getType() &&
1450 "Ret value not consistent in function!");
1451 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
1455 // Add a branch to the merge points and remove return instructions.
1457 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1458 ReturnInst *RI = Returns[i];
1459 BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
1460 Loc = RI->getDebugLoc();
1461 BI->setDebugLoc(Loc);
1462 RI->eraseFromParent();
1464 // We need to set the debug location to *somewhere* inside the
1465 // inlined function. The line number may be nonsensical, but the
1466 // instruction will at least be associated with the right
1468 if (CreatedBranchToNormalDest)
1469 CreatedBranchToNormalDest->setDebugLoc(Loc);
1470 } else if (!Returns.empty()) {
1471 // Otherwise, if there is exactly one return value, just replace anything
1472 // using the return value of the call with the computed value.
1473 if (!TheCall->use_empty()) {
1474 if (TheCall == Returns[0]->getReturnValue())
1475 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1477 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
1480 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
1481 BasicBlock *ReturnBB = Returns[0]->getParent();
1482 ReturnBB->replaceAllUsesWith(AfterCallBB);
1484 // Splice the code from the return block into the block that it will return
1485 // to, which contains the code that was after the call.
1486 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
1487 ReturnBB->getInstList());
1489 if (CreatedBranchToNormalDest)
1490 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
1492 // Delete the return instruction now and empty ReturnBB now.
1493 Returns[0]->eraseFromParent();
1494 ReturnBB->eraseFromParent();
1495 } else if (!TheCall->use_empty()) {
1496 // No returns, but something is using the return value of the call. Just
1498 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1501 // Since we are now done with the Call/Invoke, we can delete it.
1502 TheCall->eraseFromParent();
1504 // If we inlined any musttail calls and the original return is now
1505 // unreachable, delete it. It can only contain a bitcast and ret.
1506 if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB))
1507 AfterCallBB->eraseFromParent();
1509 // We should always be able to fold the entry block of the function into the
1510 // single predecessor of the block...
1511 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
1512 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
1514 // Splice the code entry block into calling block, right before the
1515 // unconditional branch.
1516 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
1517 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
1519 // Remove the unconditional branch.
1520 OrigBB->getInstList().erase(Br);
1522 // Now we can remove the CalleeEntry block, which is now empty.
1523 Caller->getBasicBlockList().erase(CalleeEntry);
1525 // If we inserted a phi node, check to see if it has a single value (e.g. all
1526 // the entries are the same or undef). If so, remove the PHI so it doesn't
1527 // block other optimizations.
1529 auto &DL = Caller->getParent()->getDataLayout();
1530 if (Value *V = SimplifyInstruction(PHI, DL, nullptr, nullptr,
1531 &IFI.ACT->getAssumptionCache(*Caller))) {
1532 PHI->replaceAllUsesWith(V);
1533 PHI->eraseFromParent();