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->getIterator();
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 Instruction *InsertPoint = &InnerResumeDest->front();
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
203 BB->splitBasicBlock(CI->getIterator(), CI->getName() + ".noexc");
205 // Delete the unconditional branch inserted by splitBasicBlock
206 BB->getInstList().pop_back();
208 // Create the new invoke instruction.
209 ImmutableCallSite CS(CI);
210 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
211 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split, UnwindEdge,
212 InvokeArgs, CI->getName(), BB);
213 II->setDebugLoc(CI->getDebugLoc());
214 II->setCallingConv(CI->getCallingConv());
215 II->setAttributes(CI->getAttributes());
217 // Make sure that anything using the call now uses the invoke! This also
218 // updates the CallGraph if present, because it uses a WeakVH.
219 CI->replaceAllUsesWith(II);
221 // Delete the original call
222 Split->getInstList().pop_front();
228 /// If we inlined an invoke site, we need to convert calls
229 /// in the body of the inlined function into invokes.
231 /// II is the invoke instruction being inlined. FirstNewBlock is the first
232 /// block of the inlined code (the last block is the end of the function),
233 /// and InlineCodeInfo is information about the code that got inlined.
234 static void HandleInlinedLandingPad(InvokeInst *II, BasicBlock *FirstNewBlock,
235 ClonedCodeInfo &InlinedCodeInfo) {
236 BasicBlock *InvokeDest = II->getUnwindDest();
238 Function *Caller = FirstNewBlock->getParent();
240 // The inlined code is currently at the end of the function, scan from the
241 // start of the inlined code to its end, checking for stuff we need to
243 LandingPadInliningInfo Invoke(II);
245 // Get all of the inlined landing pad instructions.
246 SmallPtrSet<LandingPadInst*, 16> InlinedLPads;
247 for (Function::iterator I = FirstNewBlock->getIterator(), E = Caller->end();
249 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator()))
250 InlinedLPads.insert(II->getLandingPadInst());
252 // Append the clauses from the outer landing pad instruction into the inlined
253 // landing pad instructions.
254 LandingPadInst *OuterLPad = Invoke.getLandingPadInst();
255 for (LandingPadInst *InlinedLPad : InlinedLPads) {
256 unsigned OuterNum = OuterLPad->getNumClauses();
257 InlinedLPad->reserveClauses(OuterNum);
258 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx)
259 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx));
260 if (OuterLPad->isCleanup())
261 InlinedLPad->setCleanup(true);
264 for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end();
266 if (InlinedCodeInfo.ContainsCalls)
267 if (BasicBlock *NewBB = HandleCallsInBlockInlinedThroughInvoke(
268 &*BB, Invoke.getOuterResumeDest()))
269 // Update any PHI nodes in the exceptional block to indicate that there
270 // is now a new entry in them.
271 Invoke.addIncomingPHIValuesFor(NewBB);
273 // Forward any resumes that are remaining here.
274 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator()))
275 Invoke.forwardResume(RI, InlinedLPads);
278 // Now that everything is happy, we have one final detail. The PHI nodes in
279 // the exception destination block still have entries due to the original
280 // invoke instruction. Eliminate these entries (which might even delete the
282 InvokeDest->removePredecessor(II->getParent());
285 /// If we inlined an invoke site, we need to convert calls
286 /// in the body of the inlined function into invokes.
288 /// II is the invoke instruction being inlined. FirstNewBlock is the first
289 /// block of the inlined code (the last block is the end of the function),
290 /// and InlineCodeInfo is information about the code that got inlined.
291 static void HandleInlinedEHPad(InvokeInst *II, BasicBlock *FirstNewBlock,
292 ClonedCodeInfo &InlinedCodeInfo) {
293 BasicBlock *UnwindDest = II->getUnwindDest();
294 Function *Caller = FirstNewBlock->getParent();
296 assert(UnwindDest->getFirstNonPHI()->isEHPad() && "unexpected BasicBlock!");
298 // If there are PHI nodes in the unwind destination block, we need to keep
299 // track of which values came into them from the invoke before removing the
300 // edge from this block.
301 SmallVector<Value *, 8> UnwindDestPHIValues;
302 llvm::BasicBlock *InvokeBB = II->getParent();
303 for (Instruction &I : *UnwindDest) {
304 // Save the value to use for this edge.
305 PHINode *PHI = dyn_cast<PHINode>(&I);
308 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
311 // Add incoming-PHI values to the unwind destination block for the given basic
312 // block, using the values for the original invoke's source block.
313 auto UpdatePHINodes = [&](BasicBlock *Src) {
314 BasicBlock::iterator I = UnwindDest->begin();
315 for (Value *V : UnwindDestPHIValues) {
316 PHINode *PHI = cast<PHINode>(I);
317 PHI->addIncoming(V, Src);
322 // Forward EH terminator instructions to the caller's invoke destination.
323 // This is as simple as connect all the instructions which 'unwind to caller'
324 // to the invoke destination.
325 for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end();
327 Instruction *I = BB->getFirstNonPHI();
329 if (auto *CEPI = dyn_cast<CatchEndPadInst>(I)) {
330 if (CEPI->unwindsToCaller()) {
331 CatchEndPadInst::Create(CEPI->getContext(), UnwindDest, CEPI);
332 CEPI->eraseFromParent();
333 UpdatePHINodes(&*BB);
335 } else if (auto *CEPI = dyn_cast<CleanupEndPadInst>(I)) {
336 if (CEPI->unwindsToCaller()) {
337 CleanupEndPadInst::Create(CEPI->getCleanupPad(), UnwindDest, CEPI);
338 CEPI->eraseFromParent();
339 UpdatePHINodes(&*BB);
341 } else if (auto *TPI = dyn_cast<TerminatePadInst>(I)) {
342 if (TPI->unwindsToCaller()) {
343 SmallVector<Value *, 3> TerminatePadArgs;
344 for (Value *ArgOperand : TPI->arg_operands())
345 TerminatePadArgs.push_back(ArgOperand);
346 TerminatePadInst::Create(TPI->getContext(), UnwindDest,
347 TerminatePadArgs, TPI);
348 TPI->eraseFromParent();
349 UpdatePHINodes(&*BB);
352 assert(isa<CatchPadInst>(I) || isa<CleanupPadInst>(I));
356 if (auto *CRI = dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
357 if (CRI->unwindsToCaller()) {
358 CleanupReturnInst::Create(CRI->getCleanupPad(), UnwindDest, CRI);
359 CRI->eraseFromParent();
360 UpdatePHINodes(&*BB);
365 if (InlinedCodeInfo.ContainsCalls)
366 for (Function::iterator BB = FirstNewBlock->getIterator(),
369 if (BasicBlock *NewBB =
370 HandleCallsInBlockInlinedThroughInvoke(&*BB, UnwindDest))
371 // Update any PHI nodes in the exceptional block to indicate that there
372 // is now a new entry in them.
373 UpdatePHINodes(NewBB);
375 // Now that everything is happy, we have one final detail. The PHI nodes in
376 // the exception destination block still have entries due to the original
377 // invoke instruction. Eliminate these entries (which might even delete the
379 UnwindDest->removePredecessor(InvokeBB);
382 /// When inlining a function that contains noalias scope metadata,
383 /// this metadata needs to be cloned so that the inlined blocks
384 /// have different "unqiue scopes" at every call site. Were this not done, then
385 /// aliasing scopes from a function inlined into a caller multiple times could
386 /// not be differentiated (and this would lead to miscompiles because the
387 /// non-aliasing property communicated by the metadata could have
388 /// call-site-specific control dependencies).
389 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) {
390 const Function *CalledFunc = CS.getCalledFunction();
391 SetVector<const MDNode *> MD;
393 // Note: We could only clone the metadata if it is already used in the
394 // caller. I'm omitting that check here because it might confuse
395 // inter-procedural alias analysis passes. We can revisit this if it becomes
396 // an efficiency or overhead problem.
398 for (Function::const_iterator I = CalledFunc->begin(), IE = CalledFunc->end();
400 for (BasicBlock::const_iterator J = I->begin(), JE = I->end(); J != JE; ++J) {
401 if (const MDNode *M = J->getMetadata(LLVMContext::MD_alias_scope))
403 if (const MDNode *M = J->getMetadata(LLVMContext::MD_noalias))
410 // Walk the existing metadata, adding the complete (perhaps cyclic) chain to
412 SmallVector<const Metadata *, 16> Queue(MD.begin(), MD.end());
413 while (!Queue.empty()) {
414 const MDNode *M = cast<MDNode>(Queue.pop_back_val());
415 for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i)
416 if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i)))
421 // Now we have a complete set of all metadata in the chains used to specify
422 // the noalias scopes and the lists of those scopes.
423 SmallVector<TempMDTuple, 16> DummyNodes;
424 DenseMap<const MDNode *, TrackingMDNodeRef> MDMap;
425 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
427 DummyNodes.push_back(MDTuple::getTemporary(CalledFunc->getContext(), None));
428 MDMap[*I].reset(DummyNodes.back().get());
431 // Create new metadata nodes to replace the dummy nodes, replacing old
432 // metadata references with either a dummy node or an already-created new
434 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
436 SmallVector<Metadata *, 4> NewOps;
437 for (unsigned i = 0, ie = (*I)->getNumOperands(); i != ie; ++i) {
438 const Metadata *V = (*I)->getOperand(i);
439 if (const MDNode *M = dyn_cast<MDNode>(V))
440 NewOps.push_back(MDMap[M]);
442 NewOps.push_back(const_cast<Metadata *>(V));
445 MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps);
446 MDTuple *TempM = cast<MDTuple>(MDMap[*I]);
447 assert(TempM->isTemporary() && "Expected temporary node");
449 TempM->replaceAllUsesWith(NewM);
452 // Now replace the metadata in the new inlined instructions with the
453 // repacements from the map.
454 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
455 VMI != VMIE; ++VMI) {
459 Instruction *NI = dyn_cast<Instruction>(VMI->second);
463 if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope)) {
464 MDNode *NewMD = MDMap[M];
465 // If the call site also had alias scope metadata (a list of scopes to
466 // which instructions inside it might belong), propagate those scopes to
467 // the inlined instructions.
469 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
470 NewMD = MDNode::concatenate(NewMD, CSM);
471 NI->setMetadata(LLVMContext::MD_alias_scope, NewMD);
472 } else if (NI->mayReadOrWriteMemory()) {
474 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
475 NI->setMetadata(LLVMContext::MD_alias_scope, M);
478 if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias)) {
479 MDNode *NewMD = MDMap[M];
480 // If the call site also had noalias metadata (a list of scopes with
481 // which instructions inside it don't alias), propagate those scopes to
482 // the inlined instructions.
484 CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
485 NewMD = MDNode::concatenate(NewMD, CSM);
486 NI->setMetadata(LLVMContext::MD_noalias, NewMD);
487 } else if (NI->mayReadOrWriteMemory()) {
488 if (MDNode *M = CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
489 NI->setMetadata(LLVMContext::MD_noalias, M);
494 /// If the inlined function has noalias arguments,
495 /// then add new alias scopes for each noalias argument, tag the mapped noalias
496 /// parameters with noalias metadata specifying the new scope, and tag all
497 /// non-derived loads, stores and memory intrinsics with the new alias scopes.
498 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
499 const DataLayout &DL, AAResults *CalleeAAR) {
500 if (!EnableNoAliasConversion)
503 const Function *CalledFunc = CS.getCalledFunction();
504 SmallVector<const Argument *, 4> NoAliasArgs;
506 for (const Argument &I : CalledFunc->args()) {
507 if (I.hasNoAliasAttr() && !I.hasNUses(0))
508 NoAliasArgs.push_back(&I);
511 if (NoAliasArgs.empty())
514 // To do a good job, if a noalias variable is captured, we need to know if
515 // the capture point dominates the particular use we're considering.
517 DT.recalculate(const_cast<Function&>(*CalledFunc));
519 // noalias indicates that pointer values based on the argument do not alias
520 // pointer values which are not based on it. So we add a new "scope" for each
521 // noalias function argument. Accesses using pointers based on that argument
522 // become part of that alias scope, accesses using pointers not based on that
523 // argument are tagged as noalias with that scope.
525 DenseMap<const Argument *, MDNode *> NewScopes;
526 MDBuilder MDB(CalledFunc->getContext());
528 // Create a new scope domain for this function.
530 MDB.createAnonymousAliasScopeDomain(CalledFunc->getName());
531 for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) {
532 const Argument *A = NoAliasArgs[i];
534 std::string Name = CalledFunc->getName();
537 Name += A->getName();
539 Name += ": argument ";
543 // Note: We always create a new anonymous root here. This is true regardless
544 // of the linkage of the callee because the aliasing "scope" is not just a
545 // property of the callee, but also all control dependencies in the caller.
546 MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
547 NewScopes.insert(std::make_pair(A, NewScope));
550 // Iterate over all new instructions in the map; for all memory-access
551 // instructions, add the alias scope metadata.
552 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
553 VMI != VMIE; ++VMI) {
554 if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) {
558 Instruction *NI = dyn_cast<Instruction>(VMI->second);
562 bool IsArgMemOnlyCall = false, IsFuncCall = false;
563 SmallVector<const Value *, 2> PtrArgs;
565 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
566 PtrArgs.push_back(LI->getPointerOperand());
567 else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
568 PtrArgs.push_back(SI->getPointerOperand());
569 else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
570 PtrArgs.push_back(VAAI->getPointerOperand());
571 else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
572 PtrArgs.push_back(CXI->getPointerOperand());
573 else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
574 PtrArgs.push_back(RMWI->getPointerOperand());
575 else if (ImmutableCallSite ICS = ImmutableCallSite(I)) {
576 // If we know that the call does not access memory, then we'll still
577 // know that about the inlined clone of this call site, and we don't
578 // need to add metadata.
579 if (ICS.doesNotAccessMemory())
584 FunctionModRefBehavior MRB = CalleeAAR->getModRefBehavior(ICS);
585 if (MRB == FMRB_OnlyAccessesArgumentPointees ||
586 MRB == FMRB_OnlyReadsArgumentPointees)
587 IsArgMemOnlyCall = true;
590 for (ImmutableCallSite::arg_iterator AI = ICS.arg_begin(),
591 AE = ICS.arg_end(); AI != AE; ++AI) {
592 // We need to check the underlying objects of all arguments, not just
593 // the pointer arguments, because we might be passing pointers as
595 // However, if we know that the call only accesses pointer arguments,
596 // then we only need to check the pointer arguments.
597 if (IsArgMemOnlyCall && !(*AI)->getType()->isPointerTy())
600 PtrArgs.push_back(*AI);
604 // If we found no pointers, then this instruction is not suitable for
605 // pairing with an instruction to receive aliasing metadata.
606 // However, if this is a call, this we might just alias with none of the
607 // noalias arguments.
608 if (PtrArgs.empty() && !IsFuncCall)
611 // It is possible that there is only one underlying object, but you
612 // need to go through several PHIs to see it, and thus could be
613 // repeated in the Objects list.
614 SmallPtrSet<const Value *, 4> ObjSet;
615 SmallVector<Metadata *, 4> Scopes, NoAliases;
617 SmallSetVector<const Argument *, 4> NAPtrArgs;
618 for (unsigned i = 0, ie = PtrArgs.size(); i != ie; ++i) {
619 SmallVector<Value *, 4> Objects;
620 GetUnderlyingObjects(const_cast<Value*>(PtrArgs[i]),
621 Objects, DL, /* LI = */ nullptr);
623 for (Value *O : Objects)
627 // Figure out if we're derived from anything that is not a noalias
629 bool CanDeriveViaCapture = false, UsesAliasingPtr = false;
630 for (const Value *V : ObjSet) {
631 // Is this value a constant that cannot be derived from any pointer
632 // value (we need to exclude constant expressions, for example, that
633 // are formed from arithmetic on global symbols).
634 bool IsNonPtrConst = isa<ConstantInt>(V) || isa<ConstantFP>(V) ||
635 isa<ConstantPointerNull>(V) ||
636 isa<ConstantDataVector>(V) || isa<UndefValue>(V);
640 // If this is anything other than a noalias argument, then we cannot
641 // completely describe the aliasing properties using alias.scope
642 // metadata (and, thus, won't add any).
643 if (const Argument *A = dyn_cast<Argument>(V)) {
644 if (!A->hasNoAliasAttr())
645 UsesAliasingPtr = true;
647 UsesAliasingPtr = true;
650 // If this is not some identified function-local object (which cannot
651 // directly alias a noalias argument), or some other argument (which,
652 // by definition, also cannot alias a noalias argument), then we could
653 // alias a noalias argument that has been captured).
654 if (!isa<Argument>(V) &&
655 !isIdentifiedFunctionLocal(const_cast<Value*>(V)))
656 CanDeriveViaCapture = true;
659 // A function call can always get captured noalias pointers (via other
660 // parameters, globals, etc.).
661 if (IsFuncCall && !IsArgMemOnlyCall)
662 CanDeriveViaCapture = true;
664 // First, we want to figure out all of the sets with which we definitely
665 // don't alias. Iterate over all noalias set, and add those for which:
666 // 1. The noalias argument is not in the set of objects from which we
667 // definitely derive.
668 // 2. The noalias argument has not yet been captured.
669 // An arbitrary function that might load pointers could see captured
670 // noalias arguments via other noalias arguments or globals, and so we
671 // must always check for prior capture.
672 for (const Argument *A : NoAliasArgs) {
673 if (!ObjSet.count(A) && (!CanDeriveViaCapture ||
674 // It might be tempting to skip the
675 // PointerMayBeCapturedBefore check if
676 // A->hasNoCaptureAttr() is true, but this is
677 // incorrect because nocapture only guarantees
678 // that no copies outlive the function, not
679 // that the value cannot be locally captured.
680 !PointerMayBeCapturedBefore(A,
681 /* ReturnCaptures */ false,
682 /* StoreCaptures */ false, I, &DT)))
683 NoAliases.push_back(NewScopes[A]);
686 if (!NoAliases.empty())
687 NI->setMetadata(LLVMContext::MD_noalias,
689 NI->getMetadata(LLVMContext::MD_noalias),
690 MDNode::get(CalledFunc->getContext(), NoAliases)));
692 // Next, we want to figure out all of the sets to which we might belong.
693 // We might belong to a set if the noalias argument is in the set of
694 // underlying objects. If there is some non-noalias argument in our list
695 // of underlying objects, then we cannot add a scope because the fact
696 // that some access does not alias with any set of our noalias arguments
697 // cannot itself guarantee that it does not alias with this access
698 // (because there is some pointer of unknown origin involved and the
699 // other access might also depend on this pointer). We also cannot add
700 // scopes to arbitrary functions unless we know they don't access any
701 // non-parameter pointer-values.
702 bool CanAddScopes = !UsesAliasingPtr;
703 if (CanAddScopes && IsFuncCall)
704 CanAddScopes = IsArgMemOnlyCall;
707 for (const Argument *A : NoAliasArgs) {
709 Scopes.push_back(NewScopes[A]);
714 LLVMContext::MD_alias_scope,
715 MDNode::concatenate(NI->getMetadata(LLVMContext::MD_alias_scope),
716 MDNode::get(CalledFunc->getContext(), Scopes)));
721 /// If the inlined function has non-byval align arguments, then
722 /// add @llvm.assume-based alignment assumptions to preserve this information.
723 static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) {
724 if (!PreserveAlignmentAssumptions)
726 auto &DL = CS.getCaller()->getParent()->getDataLayout();
728 // To avoid inserting redundant assumptions, we should check for assumptions
729 // already in the caller. To do this, we might need a DT of the caller.
731 bool DTCalculated = false;
733 Function *CalledFunc = CS.getCalledFunction();
734 for (Function::arg_iterator I = CalledFunc->arg_begin(),
735 E = CalledFunc->arg_end();
737 unsigned Align = I->getType()->isPointerTy() ? I->getParamAlignment() : 0;
738 if (Align && !I->hasByValOrInAllocaAttr() && !I->hasNUses(0)) {
740 DT.recalculate(const_cast<Function&>(*CS.getInstruction()->getParent()
745 // If we can already prove the asserted alignment in the context of the
746 // caller, then don't bother inserting the assumption.
747 Value *Arg = CS.getArgument(I->getArgNo());
748 if (getKnownAlignment(Arg, DL, CS.getInstruction(),
749 &IFI.ACT->getAssumptionCache(*CS.getCaller()),
753 IRBuilder<>(CS.getInstruction())
754 .CreateAlignmentAssumption(DL, Arg, Align);
759 /// Once we have cloned code over from a callee into the caller,
760 /// update the specified callgraph to reflect the changes we made.
761 /// Note that it's possible that not all code was copied over, so only
762 /// some edges of the callgraph may remain.
763 static void UpdateCallGraphAfterInlining(CallSite CS,
764 Function::iterator FirstNewBlock,
765 ValueToValueMapTy &VMap,
766 InlineFunctionInfo &IFI) {
767 CallGraph &CG = *IFI.CG;
768 const Function *Caller = CS.getInstruction()->getParent()->getParent();
769 const Function *Callee = CS.getCalledFunction();
770 CallGraphNode *CalleeNode = CG[Callee];
771 CallGraphNode *CallerNode = CG[Caller];
773 // Since we inlined some uninlined call sites in the callee into the caller,
774 // add edges from the caller to all of the callees of the callee.
775 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
777 // Consider the case where CalleeNode == CallerNode.
778 CallGraphNode::CalledFunctionsVector CallCache;
779 if (CalleeNode == CallerNode) {
780 CallCache.assign(I, E);
781 I = CallCache.begin();
785 for (; I != E; ++I) {
786 const Value *OrigCall = I->first;
788 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
789 // Only copy the edge if the call was inlined!
790 if (VMI == VMap.end() || VMI->second == nullptr)
793 // If the call was inlined, but then constant folded, there is no edge to
794 // add. Check for this case.
795 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
799 // We do not treat intrinsic calls like real function calls because we
800 // expect them to become inline code; do not add an edge for an intrinsic.
801 CallSite CS = CallSite(NewCall);
802 if (CS && CS.getCalledFunction() && CS.getCalledFunction()->isIntrinsic())
805 // Remember that this call site got inlined for the client of
807 IFI.InlinedCalls.push_back(NewCall);
809 // It's possible that inlining the callsite will cause it to go from an
810 // indirect to a direct call by resolving a function pointer. If this
811 // happens, set the callee of the new call site to a more precise
812 // destination. This can also happen if the call graph node of the caller
813 // was just unnecessarily imprecise.
814 if (!I->second->getFunction())
815 if (Function *F = CallSite(NewCall).getCalledFunction()) {
816 // Indirect call site resolved to direct call.
817 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
822 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
825 // Update the call graph by deleting the edge from Callee to Caller. We must
826 // do this after the loop above in case Caller and Callee are the same.
827 CallerNode->removeCallEdgeFor(CS);
830 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
831 BasicBlock *InsertBlock,
832 InlineFunctionInfo &IFI) {
833 Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
834 IRBuilder<> Builder(InsertBlock, InsertBlock->begin());
836 Value *Size = Builder.getInt64(M->getDataLayout().getTypeStoreSize(AggTy));
838 // Always generate a memcpy of alignment 1 here because we don't know
839 // the alignment of the src pointer. Other optimizations can infer
841 Builder.CreateMemCpy(Dst, Src, Size, /*Align=*/1);
844 /// When inlining a call site that has a byval argument,
845 /// we have to make the implicit memcpy explicit by adding it.
846 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
847 const Function *CalledFunc,
848 InlineFunctionInfo &IFI,
849 unsigned ByValAlignment) {
850 PointerType *ArgTy = cast<PointerType>(Arg->getType());
851 Type *AggTy = ArgTy->getElementType();
853 Function *Caller = TheCall->getParent()->getParent();
855 // If the called function is readonly, then it could not mutate the caller's
856 // copy of the byval'd memory. In this case, it is safe to elide the copy and
858 if (CalledFunc->onlyReadsMemory()) {
859 // If the byval argument has a specified alignment that is greater than the
860 // passed in pointer, then we either have to round up the input pointer or
861 // give up on this transformation.
862 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
865 const DataLayout &DL = Caller->getParent()->getDataLayout();
867 // If the pointer is already known to be sufficiently aligned, or if we can
868 // round it up to a larger alignment, then we don't need a temporary.
869 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, DL, TheCall,
870 &IFI.ACT->getAssumptionCache(*Caller)) >=
874 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
875 // for code quality, but rarely happens and is required for correctness.
878 // Create the alloca. If we have DataLayout, use nice alignment.
880 Caller->getParent()->getDataLayout().getPrefTypeAlignment(AggTy);
882 // If the byval had an alignment specified, we *must* use at least that
883 // alignment, as it is required by the byval argument (and uses of the
884 // pointer inside the callee).
885 Align = std::max(Align, ByValAlignment);
887 Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(),
888 &*Caller->begin()->begin());
889 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
891 // Uses of the argument in the function should use our new alloca
896 // Check whether this Value is used by a lifetime intrinsic.
897 static bool isUsedByLifetimeMarker(Value *V) {
898 for (User *U : V->users()) {
899 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
900 switch (II->getIntrinsicID()) {
902 case Intrinsic::lifetime_start:
903 case Intrinsic::lifetime_end:
911 // Check whether the given alloca already has
912 // lifetime.start or lifetime.end intrinsics.
913 static bool hasLifetimeMarkers(AllocaInst *AI) {
914 Type *Ty = AI->getType();
915 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
916 Ty->getPointerAddressSpace());
918 return isUsedByLifetimeMarker(AI);
920 // Do a scan to find all the casts to i8*.
921 for (User *U : AI->users()) {
922 if (U->getType() != Int8PtrTy) continue;
923 if (U->stripPointerCasts() != AI) continue;
924 if (isUsedByLifetimeMarker(U))
930 /// Rebuild the entire inlined-at chain for this instruction so that the top of
931 /// the chain now is inlined-at the new call site.
933 updateInlinedAtInfo(DebugLoc DL, DILocation *InlinedAtNode, LLVMContext &Ctx,
934 DenseMap<const DILocation *, DILocation *> &IANodes) {
935 SmallVector<DILocation *, 3> InlinedAtLocations;
936 DILocation *Last = InlinedAtNode;
937 DILocation *CurInlinedAt = DL;
939 // Gather all the inlined-at nodes
940 while (DILocation *IA = CurInlinedAt->getInlinedAt()) {
941 // Skip any we've already built nodes for
942 if (DILocation *Found = IANodes[IA]) {
947 InlinedAtLocations.push_back(IA);
951 // Starting from the top, rebuild the nodes to point to the new inlined-at
952 // location (then rebuilding the rest of the chain behind it) and update the
953 // map of already-constructed inlined-at nodes.
954 for (const DILocation *MD : make_range(InlinedAtLocations.rbegin(),
955 InlinedAtLocations.rend())) {
956 Last = IANodes[MD] = DILocation::getDistinct(
957 Ctx, MD->getLine(), MD->getColumn(), MD->getScope(), Last);
960 // And finally create the normal location for this instruction, referring to
961 // the new inlined-at chain.
962 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(), Last);
965 /// Update inlined instructions' line numbers to
966 /// to encode location where these instructions are inlined.
967 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
968 Instruction *TheCall) {
969 DebugLoc TheCallDL = TheCall->getDebugLoc();
973 auto &Ctx = Fn->getContext();
974 DILocation *InlinedAtNode = TheCallDL;
976 // Create a unique call site, not to be confused with any other call from the
978 InlinedAtNode = DILocation::getDistinct(
979 Ctx, InlinedAtNode->getLine(), InlinedAtNode->getColumn(),
980 InlinedAtNode->getScope(), InlinedAtNode->getInlinedAt());
982 // Cache the inlined-at nodes as they're built so they are reused, without
983 // this every instruction's inlined-at chain would become distinct from each
985 DenseMap<const DILocation *, DILocation *> IANodes;
987 for (; FI != Fn->end(); ++FI) {
988 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
990 DebugLoc DL = BI->getDebugLoc();
992 // If the inlined instruction has no line number, make it look as if it
993 // originates from the call location. This is important for
994 // ((__always_inline__, __nodebug__)) functions which must use caller
995 // location for all instructions in their function body.
997 // Don't update static allocas, as they may get moved later.
998 if (auto *AI = dyn_cast<AllocaInst>(BI))
999 if (isa<Constant>(AI->getArraySize()))
1002 BI->setDebugLoc(TheCallDL);
1004 BI->setDebugLoc(updateInlinedAtInfo(DL, InlinedAtNode, BI->getContext(), IANodes));
1010 /// This function inlines the called function into the basic block of the
1011 /// caller. This returns false if it is not possible to inline this call.
1012 /// The program is still in a well defined state if this occurs though.
1014 /// Note that this only does one level of inlining. For example, if the
1015 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
1016 /// exists in the instruction stream. Similarly this will inline a recursive
1017 /// function by one level.
1018 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
1019 AAResults *CalleeAAR, bool InsertLifetime) {
1020 Instruction *TheCall = CS.getInstruction();
1021 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
1022 "Instruction not in function!");
1024 // If IFI has any state in it, zap it before we fill it in.
1027 const Function *CalledFunc = CS.getCalledFunction();
1028 if (!CalledFunc || // Can't inline external function or indirect
1029 CalledFunc->isDeclaration() || // call, or call to a vararg function!
1030 CalledFunc->getFunctionType()->isVarArg()) return false;
1032 // The inliner does not know how to inline through calls with operand bundles.
1033 if (CS.hasOperandBundles())
1036 // If the call to the callee cannot throw, set the 'nounwind' flag on any
1037 // calls that we inline.
1038 bool MarkNoUnwind = CS.doesNotThrow();
1040 BasicBlock *OrigBB = TheCall->getParent();
1041 Function *Caller = OrigBB->getParent();
1043 // GC poses two hazards to inlining, which only occur when the callee has GC:
1044 // 1. If the caller has no GC, then the callee's GC must be propagated to the
1046 // 2. If the caller has a differing GC, it is invalid to inline.
1047 if (CalledFunc->hasGC()) {
1048 if (!Caller->hasGC())
1049 Caller->setGC(CalledFunc->getGC());
1050 else if (CalledFunc->getGC() != Caller->getGC())
1054 // Get the personality function from the callee if it contains a landing pad.
1055 Constant *CalledPersonality =
1056 CalledFunc->hasPersonalityFn()
1057 ? CalledFunc->getPersonalityFn()->stripPointerCasts()
1060 // Find the personality function used by the landing pads of the caller. If it
1061 // exists, then check to see that it matches the personality function used in
1063 Constant *CallerPersonality =
1064 Caller->hasPersonalityFn()
1065 ? Caller->getPersonalityFn()->stripPointerCasts()
1067 if (CalledPersonality) {
1068 if (!CallerPersonality)
1069 Caller->setPersonalityFn(CalledPersonality);
1070 // If the personality functions match, then we can perform the
1071 // inlining. Otherwise, we can't inline.
1072 // TODO: This isn't 100% true. Some personality functions are proper
1073 // supersets of others and can be used in place of the other.
1074 else if (CalledPersonality != CallerPersonality)
1078 // Get an iterator to the last basic block in the function, which will have
1079 // the new function inlined after it.
1080 Function::iterator LastBlock = --Caller->end();
1082 // Make sure to capture all of the return instructions from the cloned
1084 SmallVector<ReturnInst*, 8> Returns;
1085 ClonedCodeInfo InlinedFunctionInfo;
1086 Function::iterator FirstNewBlock;
1088 { // Scope to destroy VMap after cloning.
1089 ValueToValueMapTy VMap;
1090 // Keep a list of pair (dst, src) to emit byval initializations.
1091 SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
1093 auto &DL = Caller->getParent()->getDataLayout();
1095 assert(CalledFunc->arg_size() == CS.arg_size() &&
1096 "No varargs calls can be inlined!");
1098 // Calculate the vector of arguments to pass into the function cloner, which
1099 // matches up the formal to the actual argument values.
1100 CallSite::arg_iterator AI = CS.arg_begin();
1102 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
1103 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
1104 Value *ActualArg = *AI;
1106 // When byval arguments actually inlined, we need to make the copy implied
1107 // by them explicit. However, we don't do this if the callee is readonly
1108 // or readnone, because the copy would be unneeded: the callee doesn't
1109 // modify the struct.
1110 if (CS.isByValArgument(ArgNo)) {
1111 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
1112 CalledFunc->getParamAlignment(ArgNo+1));
1113 if (ActualArg != *AI)
1114 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
1117 VMap[&*I] = ActualArg;
1120 // Add alignment assumptions if necessary. We do this before the inlined
1121 // instructions are actually cloned into the caller so that we can easily
1122 // check what will be known at the start of the inlined code.
1123 AddAlignmentAssumptions(CS, IFI);
1125 // We want the inliner to prune the code as it copies. We would LOVE to
1126 // have no dead or constant instructions leftover after inlining occurs
1127 // (which can happen, e.g., because an argument was constant), but we'll be
1128 // happy with whatever the cloner can do.
1129 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
1130 /*ModuleLevelChanges=*/false, Returns, ".i",
1131 &InlinedFunctionInfo, TheCall);
1133 // Remember the first block that is newly cloned over.
1134 FirstNewBlock = LastBlock; ++FirstNewBlock;
1136 // Inject byval arguments initialization.
1137 for (std::pair<Value*, Value*> &Init : ByValInit)
1138 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
1139 &*FirstNewBlock, IFI);
1141 // Update the callgraph if requested.
1143 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
1145 // Update inlined instructions' line number information.
1146 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
1148 // Clone existing noalias metadata if necessary.
1149 CloneAliasScopeMetadata(CS, VMap);
1151 // Add noalias metadata if necessary.
1152 AddAliasScopeMetadata(CS, VMap, DL, CalleeAAR);
1154 // FIXME: We could register any cloned assumptions instead of clearing the
1155 // whole function's cache.
1157 IFI.ACT->getAssumptionCache(*Caller).clear();
1160 // If there are any alloca instructions in the block that used to be the entry
1161 // block for the callee, move them to the entry block of the caller. First
1162 // calculate which instruction they should be inserted before. We insert the
1163 // instructions at the end of the current alloca list.
1165 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
1166 for (BasicBlock::iterator I = FirstNewBlock->begin(),
1167 E = FirstNewBlock->end(); I != E; ) {
1168 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
1171 // If the alloca is now dead, remove it. This often occurs due to code
1173 if (AI->use_empty()) {
1174 AI->eraseFromParent();
1178 if (!isa<Constant>(AI->getArraySize()))
1181 // Keep track of the static allocas that we inline into the caller.
1182 IFI.StaticAllocas.push_back(AI);
1184 // Scan for the block of allocas that we can move over, and move them
1186 while (isa<AllocaInst>(I) &&
1187 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
1188 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
1192 // Transfer all of the allocas over in a block. Using splice means
1193 // that the instructions aren't removed from the symbol table, then
1195 Caller->getEntryBlock().getInstList().splice(
1196 InsertPoint, FirstNewBlock->getInstList(), AI->getIterator(), I);
1198 // Move any dbg.declares describing the allocas into the entry basic block.
1199 DIBuilder DIB(*Caller->getParent());
1200 for (auto &AI : IFI.StaticAllocas)
1201 replaceDbgDeclareForAlloca(AI, AI, DIB, /*Deref=*/false);
1204 bool InlinedMustTailCalls = false;
1205 if (InlinedFunctionInfo.ContainsCalls) {
1206 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
1207 if (CallInst *CI = dyn_cast<CallInst>(TheCall))
1208 CallSiteTailKind = CI->getTailCallKind();
1210 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
1212 for (Instruction &I : *BB) {
1213 CallInst *CI = dyn_cast<CallInst>(&I);
1217 // We need to reduce the strength of any inlined tail calls. For
1218 // musttail, we have to avoid introducing potential unbounded stack
1219 // growth. For example, if functions 'f' and 'g' are mutually recursive
1220 // with musttail, we can inline 'g' into 'f' so long as we preserve
1221 // musttail on the cloned call to 'f'. If either the inlined call site
1222 // or the cloned call site is *not* musttail, the program already has
1223 // one frame of stack growth, so it's safe to remove musttail. Here is
1224 // a table of example transformations:
1226 // f -> musttail g -> musttail f ==> f -> musttail f
1227 // f -> musttail g -> tail f ==> f -> tail f
1228 // f -> g -> musttail f ==> f -> f
1229 // f -> g -> tail f ==> f -> f
1230 CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
1231 ChildTCK = std::min(CallSiteTailKind, ChildTCK);
1232 CI->setTailCallKind(ChildTCK);
1233 InlinedMustTailCalls |= CI->isMustTailCall();
1235 // Calls inlined through a 'nounwind' call site should be marked
1238 CI->setDoesNotThrow();
1243 // Leave lifetime markers for the static alloca's, scoping them to the
1244 // function we just inlined.
1245 if (InsertLifetime && !IFI.StaticAllocas.empty()) {
1246 IRBuilder<> builder(&FirstNewBlock->front());
1247 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
1248 AllocaInst *AI = IFI.StaticAllocas[ai];
1250 // If the alloca is already scoped to something smaller than the whole
1251 // function then there's no need to add redundant, less accurate markers.
1252 if (hasLifetimeMarkers(AI))
1255 // Try to determine the size of the allocation.
1256 ConstantInt *AllocaSize = nullptr;
1257 if (ConstantInt *AIArraySize =
1258 dyn_cast<ConstantInt>(AI->getArraySize())) {
1259 auto &DL = Caller->getParent()->getDataLayout();
1260 Type *AllocaType = AI->getAllocatedType();
1261 uint64_t AllocaTypeSize = DL.getTypeAllocSize(AllocaType);
1262 uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
1264 // Don't add markers for zero-sized allocas.
1265 if (AllocaArraySize == 0)
1268 // Check that array size doesn't saturate uint64_t and doesn't
1269 // overflow when it's multiplied by type size.
1270 if (AllocaArraySize != ~0ULL &&
1271 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
1272 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
1273 AllocaArraySize * AllocaTypeSize);
1277 builder.CreateLifetimeStart(AI, AllocaSize);
1278 for (ReturnInst *RI : Returns) {
1279 // Don't insert llvm.lifetime.end calls between a musttail call and a
1280 // return. The return kills all local allocas.
1281 if (InlinedMustTailCalls &&
1282 RI->getParent()->getTerminatingMustTailCall())
1284 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
1289 // If the inlined code contained dynamic alloca instructions, wrap the inlined
1290 // code with llvm.stacksave/llvm.stackrestore intrinsics.
1291 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
1292 Module *M = Caller->getParent();
1293 // Get the two intrinsics we care about.
1294 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
1295 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
1297 // Insert the llvm.stacksave.
1298 CallInst *SavedPtr = IRBuilder<>(&*FirstNewBlock, FirstNewBlock->begin())
1299 .CreateCall(StackSave, {}, "savedstack");
1301 // Insert a call to llvm.stackrestore before any return instructions in the
1302 // inlined function.
1303 for (ReturnInst *RI : Returns) {
1304 // Don't insert llvm.stackrestore calls between a musttail call and a
1305 // return. The return will restore the stack pointer.
1306 if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
1308 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
1312 // If we are inlining for an invoke instruction, we must make sure to rewrite
1313 // any call instructions into invoke instructions.
1314 if (auto *II = dyn_cast<InvokeInst>(TheCall)) {
1315 BasicBlock *UnwindDest = II->getUnwindDest();
1316 Instruction *FirstNonPHI = UnwindDest->getFirstNonPHI();
1317 if (isa<LandingPadInst>(FirstNonPHI)) {
1318 HandleInlinedLandingPad(II, &*FirstNewBlock, InlinedFunctionInfo);
1320 HandleInlinedEHPad(II, &*FirstNewBlock, InlinedFunctionInfo);
1324 // Handle any inlined musttail call sites. In order for a new call site to be
1325 // musttail, the source of the clone and the inlined call site must have been
1326 // musttail. Therefore it's safe to return without merging control into the
1328 if (InlinedMustTailCalls) {
1329 // Check if we need to bitcast the result of any musttail calls.
1330 Type *NewRetTy = Caller->getReturnType();
1331 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
1333 // Handle the returns preceded by musttail calls separately.
1334 SmallVector<ReturnInst *, 8> NormalReturns;
1335 for (ReturnInst *RI : Returns) {
1336 CallInst *ReturnedMustTail =
1337 RI->getParent()->getTerminatingMustTailCall();
1338 if (!ReturnedMustTail) {
1339 NormalReturns.push_back(RI);
1345 // Delete the old return and any preceding bitcast.
1346 BasicBlock *CurBB = RI->getParent();
1347 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
1348 RI->eraseFromParent();
1350 OldCast->eraseFromParent();
1352 // Insert a new bitcast and return with the right type.
1353 IRBuilder<> Builder(CurBB);
1354 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
1357 // Leave behind the normal returns so we can merge control flow.
1358 std::swap(Returns, NormalReturns);
1361 // If we cloned in _exactly one_ basic block, and if that block ends in a
1362 // return instruction, we splice the body of the inlined callee directly into
1363 // the calling basic block.
1364 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
1365 // Move all of the instructions right before the call.
1366 OrigBB->getInstList().splice(TheCall->getIterator(),
1367 FirstNewBlock->getInstList(),
1368 FirstNewBlock->begin(), FirstNewBlock->end());
1369 // Remove the cloned basic block.
1370 Caller->getBasicBlockList().pop_back();
1372 // If the call site was an invoke instruction, add a branch to the normal
1374 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1375 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
1376 NewBr->setDebugLoc(Returns[0]->getDebugLoc());
1379 // If the return instruction returned a value, replace uses of the call with
1380 // uses of the returned value.
1381 if (!TheCall->use_empty()) {
1382 ReturnInst *R = Returns[0];
1383 if (TheCall == R->getReturnValue())
1384 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1386 TheCall->replaceAllUsesWith(R->getReturnValue());
1388 // Since we are now done with the Call/Invoke, we can delete it.
1389 TheCall->eraseFromParent();
1391 // Since we are now done with the return instruction, delete it also.
1392 Returns[0]->eraseFromParent();
1394 // We are now done with the inlining.
1398 // Otherwise, we have the normal case, of more than one block to inline or
1399 // multiple return sites.
1401 // We want to clone the entire callee function into the hole between the
1402 // "starter" and "ender" blocks. How we accomplish this depends on whether
1403 // this is an invoke instruction or a call instruction.
1404 BasicBlock *AfterCallBB;
1405 BranchInst *CreatedBranchToNormalDest = nullptr;
1406 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1408 // Add an unconditional branch to make this look like the CallInst case...
1409 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
1411 // Split the basic block. This guarantees that no PHI nodes will have to be
1412 // updated due to new incoming edges, and make the invoke case more
1413 // symmetric to the call case.
1415 OrigBB->splitBasicBlock(CreatedBranchToNormalDest->getIterator(),
1416 CalledFunc->getName() + ".exit");
1418 } else { // It's a call
1419 // If this is a call instruction, we need to split the basic block that
1420 // the call lives in.
1422 AfterCallBB = OrigBB->splitBasicBlock(TheCall->getIterator(),
1423 CalledFunc->getName() + ".exit");
1426 // Change the branch that used to go to AfterCallBB to branch to the first
1427 // basic block of the inlined function.
1429 TerminatorInst *Br = OrigBB->getTerminator();
1430 assert(Br && Br->getOpcode() == Instruction::Br &&
1431 "splitBasicBlock broken!");
1432 Br->setOperand(0, &*FirstNewBlock);
1434 // Now that the function is correct, make it a little bit nicer. In
1435 // particular, move the basic blocks inserted from the end of the function
1436 // into the space made by splitting the source basic block.
1437 Caller->getBasicBlockList().splice(AfterCallBB->getIterator(),
1438 Caller->getBasicBlockList(), FirstNewBlock,
1441 // Handle all of the return instructions that we just cloned in, and eliminate
1442 // any users of the original call/invoke instruction.
1443 Type *RTy = CalledFunc->getReturnType();
1445 PHINode *PHI = nullptr;
1446 if (Returns.size() > 1) {
1447 // The PHI node should go at the front of the new basic block to merge all
1448 // possible incoming values.
1449 if (!TheCall->use_empty()) {
1450 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
1451 &AfterCallBB->front());
1452 // Anything that used the result of the function call should now use the
1453 // PHI node as their operand.
1454 TheCall->replaceAllUsesWith(PHI);
1457 // Loop over all of the return instructions adding entries to the PHI node
1460 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1461 ReturnInst *RI = Returns[i];
1462 assert(RI->getReturnValue()->getType() == PHI->getType() &&
1463 "Ret value not consistent in function!");
1464 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
1468 // Add a branch to the merge points and remove return instructions.
1470 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1471 ReturnInst *RI = Returns[i];
1472 BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
1473 Loc = RI->getDebugLoc();
1474 BI->setDebugLoc(Loc);
1475 RI->eraseFromParent();
1477 // We need to set the debug location to *somewhere* inside the
1478 // inlined function. The line number may be nonsensical, but the
1479 // instruction will at least be associated with the right
1481 if (CreatedBranchToNormalDest)
1482 CreatedBranchToNormalDest->setDebugLoc(Loc);
1483 } else if (!Returns.empty()) {
1484 // Otherwise, if there is exactly one return value, just replace anything
1485 // using the return value of the call with the computed value.
1486 if (!TheCall->use_empty()) {
1487 if (TheCall == Returns[0]->getReturnValue())
1488 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1490 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
1493 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
1494 BasicBlock *ReturnBB = Returns[0]->getParent();
1495 ReturnBB->replaceAllUsesWith(AfterCallBB);
1497 // Splice the code from the return block into the block that it will return
1498 // to, which contains the code that was after the call.
1499 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
1500 ReturnBB->getInstList());
1502 if (CreatedBranchToNormalDest)
1503 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
1505 // Delete the return instruction now and empty ReturnBB now.
1506 Returns[0]->eraseFromParent();
1507 ReturnBB->eraseFromParent();
1508 } else if (!TheCall->use_empty()) {
1509 // No returns, but something is using the return value of the call. Just
1511 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1514 // Since we are now done with the Call/Invoke, we can delete it.
1515 TheCall->eraseFromParent();
1517 // If we inlined any musttail calls and the original return is now
1518 // unreachable, delete it. It can only contain a bitcast and ret.
1519 if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB))
1520 AfterCallBB->eraseFromParent();
1522 // We should always be able to fold the entry block of the function into the
1523 // single predecessor of the block...
1524 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
1525 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
1527 // Splice the code entry block into calling block, right before the
1528 // unconditional branch.
1529 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
1530 OrigBB->getInstList().splice(Br->getIterator(), CalleeEntry->getInstList());
1532 // Remove the unconditional branch.
1533 OrigBB->getInstList().erase(Br);
1535 // Now we can remove the CalleeEntry block, which is now empty.
1536 Caller->getBasicBlockList().erase(CalleeEntry);
1538 // If we inserted a phi node, check to see if it has a single value (e.g. all
1539 // the entries are the same or undef). If so, remove the PHI so it doesn't
1540 // block other optimizations.
1542 auto &DL = Caller->getParent()->getDataLayout();
1543 if (Value *V = SimplifyInstruction(PHI, DL, nullptr, nullptr,
1544 &IFI.ACT->getAssumptionCache(*Caller))) {
1545 PHI->replaceAllUsesWith(V);
1546 PHI->eraseFromParent();