1 //===- InlineFunction.cpp - Code to perform function inlining -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements inlining of a function into a call site, resolving
11 // parameters and the return value as appropriate.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Utils/Cloning.h"
16 #include "llvm/ADT/SmallSet.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/StringExtras.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/AssumptionCache.h"
22 #include "llvm/Analysis/CallGraph.h"
23 #include "llvm/Analysis/CaptureTracking.h"
24 #include "llvm/Analysis/InstructionSimplify.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/IR/Attributes.h"
27 #include "llvm/IR/CallSite.h"
28 #include "llvm/IR/CFG.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DebugInfo.h"
32 #include "llvm/IR/DerivedTypes.h"
33 #include "llvm/IR/DIBuilder.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/IRBuilder.h"
36 #include "llvm/IR/Instructions.h"
37 #include "llvm/IR/IntrinsicInst.h"
38 #include "llvm/IR/Intrinsics.h"
39 #include "llvm/IR/MDBuilder.h"
40 #include "llvm/IR/Module.h"
41 #include "llvm/Transforms/Utils/Local.h"
42 #include "llvm/Support/CommandLine.h"
47 EnableNoAliasConversion("enable-noalias-to-md-conversion", cl::init(true),
49 cl::desc("Convert noalias attributes to metadata during inlining."));
52 PreserveAlignmentAssumptions("preserve-alignment-assumptions-during-inlining",
53 cl::init(true), cl::Hidden,
54 cl::desc("Convert align attributes to assumptions during inlining."));
56 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI,
57 AAResults *CalleeAAR, bool InsertLifetime) {
58 return InlineFunction(CallSite(CI), IFI, CalleeAAR, InsertLifetime);
60 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI,
61 AAResults *CalleeAAR, bool InsertLifetime) {
62 return InlineFunction(CallSite(II), IFI, CalleeAAR, InsertLifetime);
66 /// A class for recording information about inlining a landing pad.
67 class LandingPadInliningInfo {
68 BasicBlock *OuterResumeDest; ///< Destination of the invoke's unwind.
69 BasicBlock *InnerResumeDest; ///< Destination for the callee's resume.
70 LandingPadInst *CallerLPad; ///< LandingPadInst associated with the invoke.
71 PHINode *InnerEHValuesPHI; ///< PHI for EH values from landingpad insts.
72 SmallVector<Value*, 8> UnwindDestPHIValues;
75 LandingPadInliningInfo(InvokeInst *II)
76 : OuterResumeDest(II->getUnwindDest()), InnerResumeDest(nullptr),
77 CallerLPad(nullptr), InnerEHValuesPHI(nullptr) {
78 // If there are PHI nodes in the unwind destination block, we need to keep
79 // track of which values came into them from the invoke before removing
80 // the edge from this block.
81 llvm::BasicBlock *InvokeBB = II->getParent();
82 BasicBlock::iterator I = OuterResumeDest->begin();
83 for (; isa<PHINode>(I); ++I) {
84 // Save the value to use for this edge.
85 PHINode *PHI = cast<PHINode>(I);
86 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
89 CallerLPad = cast<LandingPadInst>(I);
92 /// The outer unwind destination is the target of
93 /// unwind edges introduced for calls within the inlined function.
94 BasicBlock *getOuterResumeDest() const {
95 return OuterResumeDest;
98 BasicBlock *getInnerResumeDest();
100 LandingPadInst *getLandingPadInst() const { return CallerLPad; }
102 /// Forward the 'resume' instruction to the caller's landing pad block.
103 /// When the landing pad block has only one predecessor, this is
104 /// a simple branch. When there is more than one predecessor, we need to
105 /// split the landing pad block after the landingpad instruction and jump
107 void forwardResume(ResumeInst *RI,
108 SmallPtrSetImpl<LandingPadInst*> &InlinedLPads);
110 /// Add incoming-PHI values to the unwind destination block for the given
111 /// basic block, using the values for the 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 /// Get or create a target for the branch from ResumeInsts.
127 BasicBlock *LandingPadInliningInfo::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 /// Forward the 'resume' instruction to the caller's landing pad block.
162 /// 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 LandingPadInliningInfo::forwardResume(
166 ResumeInst *RI, 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 /// When we inline a basic block into an invoke,
181 /// we have to turn all of the calls that can throw into invokes.
182 /// 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.
186 HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB, BasicBlock *UnwindEdge) {
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, UnwindEdge,
210 InvokeArgs, CI->getName(), BB);
211 II->setDebugLoc(CI->getDebugLoc());
212 II->setCallingConv(CI->getCallingConv());
213 II->setAttributes(CI->getAttributes());
215 // Make sure that anything using the call now uses the invoke! This also
216 // updates the CallGraph if present, because it uses a WeakVH.
217 CI->replaceAllUsesWith(II);
219 // Delete the original call
220 Split->getInstList().pop_front();
226 /// If we inlined an invoke site, we need to convert calls
227 /// in the body of the inlined function into invokes.
229 /// II is the invoke instruction being inlined. FirstNewBlock is the first
230 /// block of the inlined code (the last block is the end of the function),
231 /// and InlineCodeInfo is information about the code that got inlined.
232 static void HandleInlinedLandingPad(InvokeInst *II, BasicBlock *FirstNewBlock,
233 ClonedCodeInfo &InlinedCodeInfo) {
234 BasicBlock *InvokeDest = II->getUnwindDest();
236 Function *Caller = FirstNewBlock->getParent();
238 // The inlined code is currently at the end of the function, scan from the
239 // start of the inlined code to its end, checking for stuff we need to
241 LandingPadInliningInfo Invoke(II);
243 // Get all of the inlined landing pad instructions.
244 SmallPtrSet<LandingPadInst*, 16> InlinedLPads;
245 for (Function::iterator I = FirstNewBlock, E = Caller->end(); I != E; ++I)
246 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator()))
247 InlinedLPads.insert(II->getLandingPadInst());
249 // Append the clauses from the outer landing pad instruction into the inlined
250 // landing pad instructions.
251 LandingPadInst *OuterLPad = Invoke.getLandingPadInst();
252 for (LandingPadInst *InlinedLPad : InlinedLPads) {
253 unsigned OuterNum = OuterLPad->getNumClauses();
254 InlinedLPad->reserveClauses(OuterNum);
255 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx)
256 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx));
257 if (OuterLPad->isCleanup())
258 InlinedLPad->setCleanup(true);
261 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
262 if (InlinedCodeInfo.ContainsCalls)
263 if (BasicBlock *NewBB = HandleCallsInBlockInlinedThroughInvoke(
264 BB, Invoke.getOuterResumeDest()))
265 // Update any PHI nodes in the exceptional block to indicate that there
266 // is now a new entry in them.
267 Invoke.addIncomingPHIValuesFor(NewBB);
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 /// If we inlined an invoke site, we need to convert calls
282 /// in the body of the inlined function into invokes.
284 /// II is the invoke instruction being inlined. FirstNewBlock is the first
285 /// block of the inlined code (the last block is the end of the function),
286 /// and InlineCodeInfo is information about the code that got inlined.
287 static void HandleInlinedEHPad(InvokeInst *II, BasicBlock *FirstNewBlock,
288 ClonedCodeInfo &InlinedCodeInfo) {
289 BasicBlock *UnwindDest = II->getUnwindDest();
290 Function *Caller = FirstNewBlock->getParent();
292 assert(UnwindDest->getFirstNonPHI()->isEHPad() && "unexpected BasicBlock!");
294 // If there are PHI nodes in the unwind destination block, we need to keep
295 // track of which values came into them from the invoke before removing the
296 // edge from this block.
297 SmallVector<Value *, 8> UnwindDestPHIValues;
298 llvm::BasicBlock *InvokeBB = II->getParent();
299 for (Instruction &I : *UnwindDest) {
300 // Save the value to use for this edge.
301 PHINode *PHI = dyn_cast<PHINode>(&I);
304 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
307 // Add incoming-PHI values to the unwind destination block for the given basic
308 // block, using the values for the original invoke's source block.
309 auto UpdatePHINodes = [&](BasicBlock *Src) {
310 BasicBlock::iterator I = UnwindDest->begin();
311 for (Value *V : UnwindDestPHIValues) {
312 PHINode *PHI = cast<PHINode>(I);
313 PHI->addIncoming(V, Src);
318 // Forward EH terminator instructions to the caller's invoke destination.
319 // This is as simple as connect all the instructions which 'unwind to caller'
320 // to the invoke destination.
321 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
323 Instruction *I = BB->getFirstNonPHI();
325 if (auto *CEPI = dyn_cast<CatchEndPadInst>(I)) {
326 if (CEPI->unwindsToCaller()) {
327 CatchEndPadInst::Create(CEPI->getContext(), UnwindDest, CEPI);
328 CEPI->eraseFromParent();
331 } else if (auto *CEPI = dyn_cast<CleanupEndPadInst>(I)) {
332 if (CEPI->unwindsToCaller()) {
333 CleanupEndPadInst::Create(CEPI->getCleanupPad(), UnwindDest, CEPI);
334 CEPI->eraseFromParent();
337 } else if (auto *TPI = dyn_cast<TerminatePadInst>(I)) {
338 if (TPI->unwindsToCaller()) {
339 SmallVector<Value *, 3> TerminatePadArgs;
340 for (Value *Operand : TPI->operands())
341 TerminatePadArgs.push_back(Operand);
342 TerminatePadInst::Create(TPI->getContext(), UnwindDest, TPI);
343 TPI->eraseFromParent();
347 assert(isa<CatchPadInst>(I) || isa<CleanupPadInst>(I));
351 if (auto *CRI = dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
352 if (CRI->unwindsToCaller()) {
353 CleanupReturnInst::Create(CRI->getCleanupPad(), UnwindDest, CRI);
354 CRI->eraseFromParent();
360 if (InlinedCodeInfo.ContainsCalls)
361 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
363 if (BasicBlock *NewBB =
364 HandleCallsInBlockInlinedThroughInvoke(BB, UnwindDest))
365 // Update any PHI nodes in the exceptional block to indicate that there
366 // is now a new entry in them.
367 UpdatePHINodes(NewBB);
369 // Now that everything is happy, we have one final detail. The PHI nodes in
370 // the exception destination block still have entries due to the original
371 // invoke instruction. Eliminate these entries (which might even delete the
373 UnwindDest->removePredecessor(InvokeBB);
376 /// When inlining a function that contains noalias scope metadata,
377 /// this metadata needs to be cloned so that the inlined blocks
378 /// have different "unqiue scopes" at every call site. Were this not done, then
379 /// aliasing scopes from a function inlined into a caller multiple times could
380 /// not be differentiated (and this would lead to miscompiles because the
381 /// non-aliasing property communicated by the metadata could have
382 /// call-site-specific control dependencies).
383 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) {
384 const Function *CalledFunc = CS.getCalledFunction();
385 SetVector<const MDNode *> MD;
387 // Note: We could only clone the metadata if it is already used in the
388 // caller. I'm omitting that check here because it might confuse
389 // inter-procedural alias analysis passes. We can revisit this if it becomes
390 // an efficiency or overhead problem.
392 for (Function::const_iterator I = CalledFunc->begin(), IE = CalledFunc->end();
394 for (BasicBlock::const_iterator J = I->begin(), JE = I->end(); J != JE; ++J) {
395 if (const MDNode *M = J->getMetadata(LLVMContext::MD_alias_scope))
397 if (const MDNode *M = J->getMetadata(LLVMContext::MD_noalias))
404 // Walk the existing metadata, adding the complete (perhaps cyclic) chain to
406 SmallVector<const Metadata *, 16> Queue(MD.begin(), MD.end());
407 while (!Queue.empty()) {
408 const MDNode *M = cast<MDNode>(Queue.pop_back_val());
409 for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i)
410 if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i)))
415 // Now we have a complete set of all metadata in the chains used to specify
416 // the noalias scopes and the lists of those scopes.
417 SmallVector<TempMDTuple, 16> DummyNodes;
418 DenseMap<const MDNode *, TrackingMDNodeRef> MDMap;
419 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
421 DummyNodes.push_back(MDTuple::getTemporary(CalledFunc->getContext(), None));
422 MDMap[*I].reset(DummyNodes.back().get());
425 // Create new metadata nodes to replace the dummy nodes, replacing old
426 // metadata references with either a dummy node or an already-created new
428 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
430 SmallVector<Metadata *, 4> NewOps;
431 for (unsigned i = 0, ie = (*I)->getNumOperands(); i != ie; ++i) {
432 const Metadata *V = (*I)->getOperand(i);
433 if (const MDNode *M = dyn_cast<MDNode>(V))
434 NewOps.push_back(MDMap[M]);
436 NewOps.push_back(const_cast<Metadata *>(V));
439 MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps);
440 MDTuple *TempM = cast<MDTuple>(MDMap[*I]);
441 assert(TempM->isTemporary() && "Expected temporary node");
443 TempM->replaceAllUsesWith(NewM);
446 // Now replace the metadata in the new inlined instructions with the
447 // repacements from the map.
448 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
449 VMI != VMIE; ++VMI) {
453 Instruction *NI = dyn_cast<Instruction>(VMI->second);
457 if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope)) {
458 MDNode *NewMD = MDMap[M];
459 // If the call site also had alias scope metadata (a list of scopes to
460 // which instructions inside it might belong), propagate those scopes to
461 // the inlined instructions.
463 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
464 NewMD = MDNode::concatenate(NewMD, CSM);
465 NI->setMetadata(LLVMContext::MD_alias_scope, NewMD);
466 } else if (NI->mayReadOrWriteMemory()) {
468 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
469 NI->setMetadata(LLVMContext::MD_alias_scope, M);
472 if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias)) {
473 MDNode *NewMD = MDMap[M];
474 // If the call site also had noalias metadata (a list of scopes with
475 // which instructions inside it don't alias), propagate those scopes to
476 // the inlined instructions.
478 CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
479 NewMD = MDNode::concatenate(NewMD, CSM);
480 NI->setMetadata(LLVMContext::MD_noalias, NewMD);
481 } else if (NI->mayReadOrWriteMemory()) {
482 if (MDNode *M = CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
483 NI->setMetadata(LLVMContext::MD_noalias, M);
488 /// If the inlined function has noalias arguments,
489 /// then add new alias scopes for each noalias argument, tag the mapped noalias
490 /// parameters with noalias metadata specifying the new scope, and tag all
491 /// non-derived loads, stores and memory intrinsics with the new alias scopes.
492 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
493 const DataLayout &DL, AAResults *CalleeAAR) {
494 if (!EnableNoAliasConversion)
497 const Function *CalledFunc = CS.getCalledFunction();
498 SmallVector<const Argument *, 4> NoAliasArgs;
500 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
501 E = CalledFunc->arg_end(); I != E; ++I) {
502 if (I->hasNoAliasAttr() && !I->hasNUses(0))
503 NoAliasArgs.push_back(I);
506 if (NoAliasArgs.empty())
509 // To do a good job, if a noalias variable is captured, we need to know if
510 // the capture point dominates the particular use we're considering.
512 DT.recalculate(const_cast<Function&>(*CalledFunc));
514 // noalias indicates that pointer values based on the argument do not alias
515 // pointer values which are not based on it. So we add a new "scope" for each
516 // noalias function argument. Accesses using pointers based on that argument
517 // become part of that alias scope, accesses using pointers not based on that
518 // argument are tagged as noalias with that scope.
520 DenseMap<const Argument *, MDNode *> NewScopes;
521 MDBuilder MDB(CalledFunc->getContext());
523 // Create a new scope domain for this function.
525 MDB.createAnonymousAliasScopeDomain(CalledFunc->getName());
526 for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) {
527 const Argument *A = NoAliasArgs[i];
529 std::string Name = CalledFunc->getName();
532 Name += A->getName();
534 Name += ": argument ";
538 // Note: We always create a new anonymous root here. This is true regardless
539 // of the linkage of the callee because the aliasing "scope" is not just a
540 // property of the callee, but also all control dependencies in the caller.
541 MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
542 NewScopes.insert(std::make_pair(A, NewScope));
545 // Iterate over all new instructions in the map; for all memory-access
546 // instructions, add the alias scope metadata.
547 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
548 VMI != VMIE; ++VMI) {
549 if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) {
553 Instruction *NI = dyn_cast<Instruction>(VMI->second);
557 bool IsArgMemOnlyCall = false, IsFuncCall = false;
558 SmallVector<const Value *, 2> PtrArgs;
560 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
561 PtrArgs.push_back(LI->getPointerOperand());
562 else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
563 PtrArgs.push_back(SI->getPointerOperand());
564 else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
565 PtrArgs.push_back(VAAI->getPointerOperand());
566 else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
567 PtrArgs.push_back(CXI->getPointerOperand());
568 else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
569 PtrArgs.push_back(RMWI->getPointerOperand());
570 else if (ImmutableCallSite ICS = ImmutableCallSite(I)) {
571 // If we know that the call does not access memory, then we'll still
572 // know that about the inlined clone of this call site, and we don't
573 // need to add metadata.
574 if (ICS.doesNotAccessMemory())
579 FunctionModRefBehavior MRB = CalleeAAR->getModRefBehavior(ICS);
580 if (MRB == FMRB_OnlyAccessesArgumentPointees ||
581 MRB == FMRB_OnlyReadsArgumentPointees)
582 IsArgMemOnlyCall = true;
585 for (ImmutableCallSite::arg_iterator AI = ICS.arg_begin(),
586 AE = ICS.arg_end(); AI != AE; ++AI) {
587 // We need to check the underlying objects of all arguments, not just
588 // the pointer arguments, because we might be passing pointers as
590 // However, if we know that the call only accesses pointer arguments,
591 // then we only need to check the pointer arguments.
592 if (IsArgMemOnlyCall && !(*AI)->getType()->isPointerTy())
595 PtrArgs.push_back(*AI);
599 // If we found no pointers, then this instruction is not suitable for
600 // pairing with an instruction to receive aliasing metadata.
601 // However, if this is a call, this we might just alias with none of the
602 // noalias arguments.
603 if (PtrArgs.empty() && !IsFuncCall)
606 // It is possible that there is only one underlying object, but you
607 // need to go through several PHIs to see it, and thus could be
608 // repeated in the Objects list.
609 SmallPtrSet<const Value *, 4> ObjSet;
610 SmallVector<Metadata *, 4> Scopes, NoAliases;
612 SmallSetVector<const Argument *, 4> NAPtrArgs;
613 for (unsigned i = 0, ie = PtrArgs.size(); i != ie; ++i) {
614 SmallVector<Value *, 4> Objects;
615 GetUnderlyingObjects(const_cast<Value*>(PtrArgs[i]),
616 Objects, DL, /* MaxLookup = */ 0);
618 for (Value *O : Objects)
622 // Figure out if we're derived from anything that is not a noalias
624 bool CanDeriveViaCapture = false, UsesAliasingPtr = false;
625 for (const Value *V : ObjSet) {
626 // Is this value a constant that cannot be derived from any pointer
627 // value (we need to exclude constant expressions, for example, that
628 // are formed from arithmetic on global symbols).
629 bool IsNonPtrConst = isa<ConstantInt>(V) || isa<ConstantFP>(V) ||
630 isa<ConstantPointerNull>(V) ||
631 isa<ConstantDataVector>(V) || isa<UndefValue>(V);
635 // If this is anything other than a noalias argument, then we cannot
636 // completely describe the aliasing properties using alias.scope
637 // metadata (and, thus, won't add any).
638 if (const Argument *A = dyn_cast<Argument>(V)) {
639 if (!A->hasNoAliasAttr())
640 UsesAliasingPtr = true;
642 UsesAliasingPtr = true;
645 // If this is not some identified function-local object (which cannot
646 // directly alias a noalias argument), or some other argument (which,
647 // by definition, also cannot alias a noalias argument), then we could
648 // alias a noalias argument that has been captured).
649 if (!isa<Argument>(V) &&
650 !isIdentifiedFunctionLocal(const_cast<Value*>(V)))
651 CanDeriveViaCapture = true;
654 // A function call can always get captured noalias pointers (via other
655 // parameters, globals, etc.).
656 if (IsFuncCall && !IsArgMemOnlyCall)
657 CanDeriveViaCapture = true;
659 // First, we want to figure out all of the sets with which we definitely
660 // don't alias. Iterate over all noalias set, and add those for which:
661 // 1. The noalias argument is not in the set of objects from which we
662 // definitely derive.
663 // 2. The noalias argument has not yet been captured.
664 // An arbitrary function that might load pointers could see captured
665 // noalias arguments via other noalias arguments or globals, and so we
666 // must always check for prior capture.
667 for (const Argument *A : NoAliasArgs) {
668 if (!ObjSet.count(A) && (!CanDeriveViaCapture ||
669 // It might be tempting to skip the
670 // PointerMayBeCapturedBefore check if
671 // A->hasNoCaptureAttr() is true, but this is
672 // incorrect because nocapture only guarantees
673 // that no copies outlive the function, not
674 // that the value cannot be locally captured.
675 !PointerMayBeCapturedBefore(A,
676 /* ReturnCaptures */ false,
677 /* StoreCaptures */ false, I, &DT)))
678 NoAliases.push_back(NewScopes[A]);
681 if (!NoAliases.empty())
682 NI->setMetadata(LLVMContext::MD_noalias,
684 NI->getMetadata(LLVMContext::MD_noalias),
685 MDNode::get(CalledFunc->getContext(), NoAliases)));
687 // Next, we want to figure out all of the sets to which we might belong.
688 // We might belong to a set if the noalias argument is in the set of
689 // underlying objects. If there is some non-noalias argument in our list
690 // of underlying objects, then we cannot add a scope because the fact
691 // that some access does not alias with any set of our noalias arguments
692 // cannot itself guarantee that it does not alias with this access
693 // (because there is some pointer of unknown origin involved and the
694 // other access might also depend on this pointer). We also cannot add
695 // scopes to arbitrary functions unless we know they don't access any
696 // non-parameter pointer-values.
697 bool CanAddScopes = !UsesAliasingPtr;
698 if (CanAddScopes && IsFuncCall)
699 CanAddScopes = IsArgMemOnlyCall;
702 for (const Argument *A : NoAliasArgs) {
704 Scopes.push_back(NewScopes[A]);
709 LLVMContext::MD_alias_scope,
710 MDNode::concatenate(NI->getMetadata(LLVMContext::MD_alias_scope),
711 MDNode::get(CalledFunc->getContext(), Scopes)));
716 /// If the inlined function has non-byval align arguments, then
717 /// add @llvm.assume-based alignment assumptions to preserve this information.
718 static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) {
719 if (!PreserveAlignmentAssumptions)
721 auto &DL = CS.getCaller()->getParent()->getDataLayout();
723 // To avoid inserting redundant assumptions, we should check for assumptions
724 // already in the caller. To do this, we might need a DT of the caller.
726 bool DTCalculated = false;
728 Function *CalledFunc = CS.getCalledFunction();
729 for (Function::arg_iterator I = CalledFunc->arg_begin(),
730 E = CalledFunc->arg_end();
732 unsigned Align = I->getType()->isPointerTy() ? I->getParamAlignment() : 0;
733 if (Align && !I->hasByValOrInAllocaAttr() && !I->hasNUses(0)) {
735 DT.recalculate(const_cast<Function&>(*CS.getInstruction()->getParent()
740 // If we can already prove the asserted alignment in the context of the
741 // caller, then don't bother inserting the assumption.
742 Value *Arg = CS.getArgument(I->getArgNo());
743 if (getKnownAlignment(Arg, DL, CS.getInstruction(),
744 &IFI.ACT->getAssumptionCache(*CS.getCaller()),
748 IRBuilder<>(CS.getInstruction())
749 .CreateAlignmentAssumption(DL, Arg, Align);
754 /// Once we have cloned code over from a callee into the caller,
755 /// update the specified callgraph to reflect the changes we made.
756 /// Note that it's possible that not all code was copied over, so only
757 /// some edges of the callgraph may remain.
758 static void UpdateCallGraphAfterInlining(CallSite CS,
759 Function::iterator FirstNewBlock,
760 ValueToValueMapTy &VMap,
761 InlineFunctionInfo &IFI) {
762 CallGraph &CG = *IFI.CG;
763 const Function *Caller = CS.getInstruction()->getParent()->getParent();
764 const Function *Callee = CS.getCalledFunction();
765 CallGraphNode *CalleeNode = CG[Callee];
766 CallGraphNode *CallerNode = CG[Caller];
768 // Since we inlined some uninlined call sites in the callee into the caller,
769 // add edges from the caller to all of the callees of the callee.
770 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
772 // Consider the case where CalleeNode == CallerNode.
773 CallGraphNode::CalledFunctionsVector CallCache;
774 if (CalleeNode == CallerNode) {
775 CallCache.assign(I, E);
776 I = CallCache.begin();
780 for (; I != E; ++I) {
781 const Value *OrigCall = I->first;
783 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
784 // Only copy the edge if the call was inlined!
785 if (VMI == VMap.end() || VMI->second == nullptr)
788 // If the call was inlined, but then constant folded, there is no edge to
789 // add. Check for this case.
790 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
794 // We do not treat intrinsic calls like real function calls because we
795 // expect them to become inline code; do not add an edge for an intrinsic.
796 CallSite CS = CallSite(NewCall);
797 if (CS && CS.getCalledFunction() && CS.getCalledFunction()->isIntrinsic())
800 // Remember that this call site got inlined for the client of
802 IFI.InlinedCalls.push_back(NewCall);
804 // It's possible that inlining the callsite will cause it to go from an
805 // indirect to a direct call by resolving a function pointer. If this
806 // happens, set the callee of the new call site to a more precise
807 // destination. This can also happen if the call graph node of the caller
808 // was just unnecessarily imprecise.
809 if (!I->second->getFunction())
810 if (Function *F = CallSite(NewCall).getCalledFunction()) {
811 // Indirect call site resolved to direct call.
812 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
817 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
820 // Update the call graph by deleting the edge from Callee to Caller. We must
821 // do this after the loop above in case Caller and Callee are the same.
822 CallerNode->removeCallEdgeFor(CS);
825 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
826 BasicBlock *InsertBlock,
827 InlineFunctionInfo &IFI) {
828 Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
829 IRBuilder<> Builder(InsertBlock->begin());
831 Value *Size = Builder.getInt64(M->getDataLayout().getTypeStoreSize(AggTy));
833 // Always generate a memcpy of alignment 1 here because we don't know
834 // the alignment of the src pointer. Other optimizations can infer
836 Builder.CreateMemCpy(Dst, Src, Size, /*Align=*/1);
839 /// When inlining a call site that has a byval argument,
840 /// we have to make the implicit memcpy explicit by adding it.
841 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
842 const Function *CalledFunc,
843 InlineFunctionInfo &IFI,
844 unsigned ByValAlignment) {
845 PointerType *ArgTy = cast<PointerType>(Arg->getType());
846 Type *AggTy = ArgTy->getElementType();
848 Function *Caller = TheCall->getParent()->getParent();
850 // If the called function is readonly, then it could not mutate the caller's
851 // copy of the byval'd memory. In this case, it is safe to elide the copy and
853 if (CalledFunc->onlyReadsMemory()) {
854 // If the byval argument has a specified alignment that is greater than the
855 // passed in pointer, then we either have to round up the input pointer or
856 // give up on this transformation.
857 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
860 const DataLayout &DL = Caller->getParent()->getDataLayout();
862 // If the pointer is already known to be sufficiently aligned, or if we can
863 // round it up to a larger alignment, then we don't need a temporary.
864 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, DL, TheCall,
865 &IFI.ACT->getAssumptionCache(*Caller)) >=
869 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
870 // for code quality, but rarely happens and is required for correctness.
873 // Create the alloca. If we have DataLayout, use nice alignment.
875 Caller->getParent()->getDataLayout().getPrefTypeAlignment(AggTy);
877 // If the byval had an alignment specified, we *must* use at least that
878 // alignment, as it is required by the byval argument (and uses of the
879 // pointer inside the callee).
880 Align = std::max(Align, ByValAlignment);
882 Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(),
883 &*Caller->begin()->begin());
884 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
886 // Uses of the argument in the function should use our new alloca
891 // Check whether this Value is used by a lifetime intrinsic.
892 static bool isUsedByLifetimeMarker(Value *V) {
893 for (User *U : V->users()) {
894 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
895 switch (II->getIntrinsicID()) {
897 case Intrinsic::lifetime_start:
898 case Intrinsic::lifetime_end:
906 // Check whether the given alloca already has
907 // lifetime.start or lifetime.end intrinsics.
908 static bool hasLifetimeMarkers(AllocaInst *AI) {
909 Type *Ty = AI->getType();
910 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
911 Ty->getPointerAddressSpace());
913 return isUsedByLifetimeMarker(AI);
915 // Do a scan to find all the casts to i8*.
916 for (User *U : AI->users()) {
917 if (U->getType() != Int8PtrTy) continue;
918 if (U->stripPointerCasts() != AI) continue;
919 if (isUsedByLifetimeMarker(U))
925 /// Rebuild the entire inlined-at chain for this instruction so that the top of
926 /// the chain now is inlined-at the new call site.
928 updateInlinedAtInfo(DebugLoc DL, DILocation *InlinedAtNode, LLVMContext &Ctx,
929 DenseMap<const DILocation *, DILocation *> &IANodes) {
930 SmallVector<DILocation *, 3> InlinedAtLocations;
931 DILocation *Last = InlinedAtNode;
932 DILocation *CurInlinedAt = DL;
934 // Gather all the inlined-at nodes
935 while (DILocation *IA = CurInlinedAt->getInlinedAt()) {
936 // Skip any we've already built nodes for
937 if (DILocation *Found = IANodes[IA]) {
942 InlinedAtLocations.push_back(IA);
946 // Starting from the top, rebuild the nodes to point to the new inlined-at
947 // location (then rebuilding the rest of the chain behind it) and update the
948 // map of already-constructed inlined-at nodes.
949 for (const DILocation *MD : make_range(InlinedAtLocations.rbegin(),
950 InlinedAtLocations.rend())) {
951 Last = IANodes[MD] = DILocation::getDistinct(
952 Ctx, MD->getLine(), MD->getColumn(), MD->getScope(), Last);
955 // And finally create the normal location for this instruction, referring to
956 // the new inlined-at chain.
957 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(), Last);
960 /// Update inlined instructions' line numbers to
961 /// to encode location where these instructions are inlined.
962 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
963 Instruction *TheCall) {
964 DebugLoc TheCallDL = TheCall->getDebugLoc();
968 auto &Ctx = Fn->getContext();
969 DILocation *InlinedAtNode = TheCallDL;
971 // Create a unique call site, not to be confused with any other call from the
973 InlinedAtNode = DILocation::getDistinct(
974 Ctx, InlinedAtNode->getLine(), InlinedAtNode->getColumn(),
975 InlinedAtNode->getScope(), InlinedAtNode->getInlinedAt());
977 // Cache the inlined-at nodes as they're built so they are reused, without
978 // this every instruction's inlined-at chain would become distinct from each
980 DenseMap<const DILocation *, DILocation *> IANodes;
982 for (; FI != Fn->end(); ++FI) {
983 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
985 DebugLoc DL = BI->getDebugLoc();
987 // If the inlined instruction has no line number, make it look as if it
988 // originates from the call location. This is important for
989 // ((__always_inline__, __nodebug__)) functions which must use caller
990 // location for all instructions in their function body.
992 // Don't update static allocas, as they may get moved later.
993 if (auto *AI = dyn_cast<AllocaInst>(BI))
994 if (isa<Constant>(AI->getArraySize()))
997 BI->setDebugLoc(TheCallDL);
999 BI->setDebugLoc(updateInlinedAtInfo(DL, InlinedAtNode, BI->getContext(), IANodes));
1005 /// This function inlines the called function into the basic block of the
1006 /// caller. This returns false if it is not possible to inline this call.
1007 /// The program is still in a well defined state if this occurs though.
1009 /// Note that this only does one level of inlining. For example, if the
1010 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
1011 /// exists in the instruction stream. Similarly this will inline a recursive
1012 /// function by one level.
1013 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
1014 AAResults *CalleeAAR, bool InsertLifetime) {
1015 Instruction *TheCall = CS.getInstruction();
1016 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
1017 "Instruction not in function!");
1019 // If IFI has any state in it, zap it before we fill it in.
1022 const Function *CalledFunc = CS.getCalledFunction();
1023 if (!CalledFunc || // Can't inline external function or indirect
1024 CalledFunc->isDeclaration() || // call, or call to a vararg function!
1025 CalledFunc->getFunctionType()->isVarArg()) return false;
1027 // If the call to the callee cannot throw, set the 'nounwind' flag on any
1028 // calls that we inline.
1029 bool MarkNoUnwind = CS.doesNotThrow();
1031 BasicBlock *OrigBB = TheCall->getParent();
1032 Function *Caller = OrigBB->getParent();
1034 // GC poses two hazards to inlining, which only occur when the callee has GC:
1035 // 1. If the caller has no GC, then the callee's GC must be propagated to the
1037 // 2. If the caller has a differing GC, it is invalid to inline.
1038 if (CalledFunc->hasGC()) {
1039 if (!Caller->hasGC())
1040 Caller->setGC(CalledFunc->getGC());
1041 else if (CalledFunc->getGC() != Caller->getGC())
1045 // Get the personality function from the callee if it contains a landing pad.
1046 Constant *CalledPersonality =
1047 CalledFunc->hasPersonalityFn() ? CalledFunc->getPersonalityFn() : nullptr;
1049 // Find the personality function used by the landing pads of the caller. If it
1050 // exists, then check to see that it matches the personality function used in
1052 Constant *CallerPersonality =
1053 Caller->hasPersonalityFn() ? Caller->getPersonalityFn() : nullptr;
1054 if (CalledPersonality) {
1055 if (!CallerPersonality)
1056 Caller->setPersonalityFn(CalledPersonality);
1057 // If the personality functions match, then we can perform the
1058 // inlining. Otherwise, we can't inline.
1059 // TODO: This isn't 100% true. Some personality functions are proper
1060 // supersets of others and can be used in place of the other.
1061 else if (CalledPersonality != CallerPersonality)
1065 // Get an iterator to the last basic block in the function, which will have
1066 // the new function inlined after it.
1067 Function::iterator LastBlock = &Caller->back();
1069 // Make sure to capture all of the return instructions from the cloned
1071 SmallVector<ReturnInst*, 8> Returns;
1072 ClonedCodeInfo InlinedFunctionInfo;
1073 Function::iterator FirstNewBlock;
1075 { // Scope to destroy VMap after cloning.
1076 ValueToValueMapTy VMap;
1077 // Keep a list of pair (dst, src) to emit byval initializations.
1078 SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
1080 auto &DL = Caller->getParent()->getDataLayout();
1082 assert(CalledFunc->arg_size() == CS.arg_size() &&
1083 "No varargs calls can be inlined!");
1085 // Calculate the vector of arguments to pass into the function cloner, which
1086 // matches up the formal to the actual argument values.
1087 CallSite::arg_iterator AI = CS.arg_begin();
1089 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
1090 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
1091 Value *ActualArg = *AI;
1093 // When byval arguments actually inlined, we need to make the copy implied
1094 // by them explicit. However, we don't do this if the callee is readonly
1095 // or readnone, because the copy would be unneeded: the callee doesn't
1096 // modify the struct.
1097 if (CS.isByValArgument(ArgNo)) {
1098 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
1099 CalledFunc->getParamAlignment(ArgNo+1));
1100 if (ActualArg != *AI)
1101 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
1104 VMap[I] = ActualArg;
1107 // Add alignment assumptions if necessary. We do this before the inlined
1108 // instructions are actually cloned into the caller so that we can easily
1109 // check what will be known at the start of the inlined code.
1110 AddAlignmentAssumptions(CS, IFI);
1112 // We want the inliner to prune the code as it copies. We would LOVE to
1113 // have no dead or constant instructions leftover after inlining occurs
1114 // (which can happen, e.g., because an argument was constant), but we'll be
1115 // happy with whatever the cloner can do.
1116 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
1117 /*ModuleLevelChanges=*/false, Returns, ".i",
1118 &InlinedFunctionInfo, TheCall);
1120 // Remember the first block that is newly cloned over.
1121 FirstNewBlock = LastBlock; ++FirstNewBlock;
1123 // Inject byval arguments initialization.
1124 for (std::pair<Value*, Value*> &Init : ByValInit)
1125 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
1126 FirstNewBlock, IFI);
1128 // Update the callgraph if requested.
1130 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
1132 // Update inlined instructions' line number information.
1133 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
1135 // Clone existing noalias metadata if necessary.
1136 CloneAliasScopeMetadata(CS, VMap);
1138 // Add noalias metadata if necessary.
1139 AddAliasScopeMetadata(CS, VMap, DL, CalleeAAR);
1141 // FIXME: We could register any cloned assumptions instead of clearing the
1142 // whole function's cache.
1144 IFI.ACT->getAssumptionCache(*Caller).clear();
1147 // If there are any alloca instructions in the block that used to be the entry
1148 // block for the callee, move them to the entry block of the caller. First
1149 // calculate which instruction they should be inserted before. We insert the
1150 // instructions at the end of the current alloca list.
1152 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
1153 for (BasicBlock::iterator I = FirstNewBlock->begin(),
1154 E = FirstNewBlock->end(); I != E; ) {
1155 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
1158 // If the alloca is now dead, remove it. This often occurs due to code
1160 if (AI->use_empty()) {
1161 AI->eraseFromParent();
1165 if (!isa<Constant>(AI->getArraySize()))
1168 // Keep track of the static allocas that we inline into the caller.
1169 IFI.StaticAllocas.push_back(AI);
1171 // Scan for the block of allocas that we can move over, and move them
1173 while (isa<AllocaInst>(I) &&
1174 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
1175 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
1179 // Transfer all of the allocas over in a block. Using splice means
1180 // that the instructions aren't removed from the symbol table, then
1182 Caller->getEntryBlock().getInstList().splice(InsertPoint,
1183 FirstNewBlock->getInstList(),
1186 // Move any dbg.declares describing the allocas into the entry basic block.
1187 DIBuilder DIB(*Caller->getParent());
1188 for (auto &AI : IFI.StaticAllocas)
1189 replaceDbgDeclareForAlloca(AI, AI, DIB, /*Deref=*/false);
1192 bool InlinedMustTailCalls = false;
1193 if (InlinedFunctionInfo.ContainsCalls) {
1194 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
1195 if (CallInst *CI = dyn_cast<CallInst>(TheCall))
1196 CallSiteTailKind = CI->getTailCallKind();
1198 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
1200 for (Instruction &I : *BB) {
1201 CallInst *CI = dyn_cast<CallInst>(&I);
1205 // We need to reduce the strength of any inlined tail calls. For
1206 // musttail, we have to avoid introducing potential unbounded stack
1207 // growth. For example, if functions 'f' and 'g' are mutually recursive
1208 // with musttail, we can inline 'g' into 'f' so long as we preserve
1209 // musttail on the cloned call to 'f'. If either the inlined call site
1210 // or the cloned call site is *not* musttail, the program already has
1211 // one frame of stack growth, so it's safe to remove musttail. Here is
1212 // a table of example transformations:
1214 // f -> musttail g -> musttail f ==> f -> musttail f
1215 // f -> musttail g -> tail f ==> f -> tail f
1216 // f -> g -> musttail f ==> f -> f
1217 // f -> g -> tail f ==> f -> f
1218 CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
1219 ChildTCK = std::min(CallSiteTailKind, ChildTCK);
1220 CI->setTailCallKind(ChildTCK);
1221 InlinedMustTailCalls |= CI->isMustTailCall();
1223 // Calls inlined through a 'nounwind' call site should be marked
1226 CI->setDoesNotThrow();
1231 // Leave lifetime markers for the static alloca's, scoping them to the
1232 // function we just inlined.
1233 if (InsertLifetime && !IFI.StaticAllocas.empty()) {
1234 IRBuilder<> builder(FirstNewBlock->begin());
1235 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
1236 AllocaInst *AI = IFI.StaticAllocas[ai];
1238 // If the alloca is already scoped to something smaller than the whole
1239 // function then there's no need to add redundant, less accurate markers.
1240 if (hasLifetimeMarkers(AI))
1243 // Try to determine the size of the allocation.
1244 ConstantInt *AllocaSize = nullptr;
1245 if (ConstantInt *AIArraySize =
1246 dyn_cast<ConstantInt>(AI->getArraySize())) {
1247 auto &DL = Caller->getParent()->getDataLayout();
1248 Type *AllocaType = AI->getAllocatedType();
1249 uint64_t AllocaTypeSize = DL.getTypeAllocSize(AllocaType);
1250 uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
1252 // Don't add markers for zero-sized allocas.
1253 if (AllocaArraySize == 0)
1256 // Check that array size doesn't saturate uint64_t and doesn't
1257 // overflow when it's multiplied by type size.
1258 if (AllocaArraySize != ~0ULL &&
1259 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
1260 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
1261 AllocaArraySize * AllocaTypeSize);
1265 builder.CreateLifetimeStart(AI, AllocaSize);
1266 for (ReturnInst *RI : Returns) {
1267 // Don't insert llvm.lifetime.end calls between a musttail call and a
1268 // return. The return kills all local allocas.
1269 if (InlinedMustTailCalls &&
1270 RI->getParent()->getTerminatingMustTailCall())
1272 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
1277 // If the inlined code contained dynamic alloca instructions, wrap the inlined
1278 // code with llvm.stacksave/llvm.stackrestore intrinsics.
1279 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
1280 Module *M = Caller->getParent();
1281 // Get the two intrinsics we care about.
1282 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
1283 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
1285 // Insert the llvm.stacksave.
1286 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
1287 .CreateCall(StackSave, {}, "savedstack");
1289 // Insert a call to llvm.stackrestore before any return instructions in the
1290 // inlined function.
1291 for (ReturnInst *RI : Returns) {
1292 // Don't insert llvm.stackrestore calls between a musttail call and a
1293 // return. The return will restore the stack pointer.
1294 if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
1296 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
1300 // If we are inlining for an invoke instruction, we must make sure to rewrite
1301 // any call instructions into invoke instructions.
1302 if (auto *II = dyn_cast<InvokeInst>(TheCall)) {
1303 BasicBlock *UnwindDest = II->getUnwindDest();
1304 Instruction *FirstNonPHI = UnwindDest->getFirstNonPHI();
1305 if (isa<LandingPadInst>(FirstNonPHI)) {
1306 HandleInlinedLandingPad(II, FirstNewBlock, InlinedFunctionInfo);
1308 HandleInlinedEHPad(II, FirstNewBlock, InlinedFunctionInfo);
1312 // Handle any inlined musttail call sites. In order for a new call site to be
1313 // musttail, the source of the clone and the inlined call site must have been
1314 // musttail. Therefore it's safe to return without merging control into the
1316 if (InlinedMustTailCalls) {
1317 // Check if we need to bitcast the result of any musttail calls.
1318 Type *NewRetTy = Caller->getReturnType();
1319 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
1321 // Handle the returns preceded by musttail calls separately.
1322 SmallVector<ReturnInst *, 8> NormalReturns;
1323 for (ReturnInst *RI : Returns) {
1324 CallInst *ReturnedMustTail =
1325 RI->getParent()->getTerminatingMustTailCall();
1326 if (!ReturnedMustTail) {
1327 NormalReturns.push_back(RI);
1333 // Delete the old return and any preceding bitcast.
1334 BasicBlock *CurBB = RI->getParent();
1335 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
1336 RI->eraseFromParent();
1338 OldCast->eraseFromParent();
1340 // Insert a new bitcast and return with the right type.
1341 IRBuilder<> Builder(CurBB);
1342 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
1345 // Leave behind the normal returns so we can merge control flow.
1346 std::swap(Returns, NormalReturns);
1349 // If we cloned in _exactly one_ basic block, and if that block ends in a
1350 // return instruction, we splice the body of the inlined callee directly into
1351 // the calling basic block.
1352 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
1353 // Move all of the instructions right before the call.
1354 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
1355 FirstNewBlock->begin(), FirstNewBlock->end());
1356 // Remove the cloned basic block.
1357 Caller->getBasicBlockList().pop_back();
1359 // If the call site was an invoke instruction, add a branch to the normal
1361 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1362 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
1363 NewBr->setDebugLoc(Returns[0]->getDebugLoc());
1366 // If the return instruction returned a value, replace uses of the call with
1367 // uses of the returned value.
1368 if (!TheCall->use_empty()) {
1369 ReturnInst *R = Returns[0];
1370 if (TheCall == R->getReturnValue())
1371 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1373 TheCall->replaceAllUsesWith(R->getReturnValue());
1375 // Since we are now done with the Call/Invoke, we can delete it.
1376 TheCall->eraseFromParent();
1378 // Since we are now done with the return instruction, delete it also.
1379 Returns[0]->eraseFromParent();
1381 // We are now done with the inlining.
1385 // Otherwise, we have the normal case, of more than one block to inline or
1386 // multiple return sites.
1388 // We want to clone the entire callee function into the hole between the
1389 // "starter" and "ender" blocks. How we accomplish this depends on whether
1390 // this is an invoke instruction or a call instruction.
1391 BasicBlock *AfterCallBB;
1392 BranchInst *CreatedBranchToNormalDest = nullptr;
1393 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1395 // Add an unconditional branch to make this look like the CallInst case...
1396 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
1398 // Split the basic block. This guarantees that no PHI nodes will have to be
1399 // updated due to new incoming edges, and make the invoke case more
1400 // symmetric to the call case.
1401 AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest,
1402 CalledFunc->getName()+".exit");
1404 } else { // It's a call
1405 // If this is a call instruction, we need to split the basic block that
1406 // the call lives in.
1408 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
1409 CalledFunc->getName()+".exit");
1412 // Change the branch that used to go to AfterCallBB to branch to the first
1413 // basic block of the inlined function.
1415 TerminatorInst *Br = OrigBB->getTerminator();
1416 assert(Br && Br->getOpcode() == Instruction::Br &&
1417 "splitBasicBlock broken!");
1418 Br->setOperand(0, FirstNewBlock);
1421 // Now that the function is correct, make it a little bit nicer. In
1422 // particular, move the basic blocks inserted from the end of the function
1423 // into the space made by splitting the source basic block.
1424 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
1425 FirstNewBlock, Caller->end());
1427 // Handle all of the return instructions that we just cloned in, and eliminate
1428 // any users of the original call/invoke instruction.
1429 Type *RTy = CalledFunc->getReturnType();
1431 PHINode *PHI = nullptr;
1432 if (Returns.size() > 1) {
1433 // The PHI node should go at the front of the new basic block to merge all
1434 // possible incoming values.
1435 if (!TheCall->use_empty()) {
1436 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
1437 AfterCallBB->begin());
1438 // Anything that used the result of the function call should now use the
1439 // PHI node as their operand.
1440 TheCall->replaceAllUsesWith(PHI);
1443 // Loop over all of the return instructions adding entries to the PHI node
1446 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1447 ReturnInst *RI = Returns[i];
1448 assert(RI->getReturnValue()->getType() == PHI->getType() &&
1449 "Ret value not consistent in function!");
1450 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();