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/SetVector.h"
17 #include "llvm/ADT/SmallSet.h"
18 #include "llvm/ADT/SmallVector.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 SmallVector<OperandBundleDef, 1> OpBundles;
213 CS.getOperandBundlesAsDefs(OpBundles);
215 // Note: we're round tripping operand bundles through memory here, and that
216 // can potentially be avoided with a cleverer API design that we do not have
220 InvokeInst::Create(CI->getCalledValue(), Split, UnwindEdge, InvokeArgs,
221 OpBundles, CI->getName(), BB);
222 II->setDebugLoc(CI->getDebugLoc());
223 II->setCallingConv(CI->getCallingConv());
224 II->setAttributes(CI->getAttributes());
226 // Make sure that anything using the call now uses the invoke! This also
227 // updates the CallGraph if present, because it uses a WeakVH.
228 CI->replaceAllUsesWith(II);
230 // Delete the original call
231 Split->getInstList().pop_front();
237 /// If we inlined an invoke site, we need to convert calls
238 /// in the body of the inlined function into invokes.
240 /// II is the invoke instruction being inlined. FirstNewBlock is the first
241 /// block of the inlined code (the last block is the end of the function),
242 /// and InlineCodeInfo is information about the code that got inlined.
243 static void HandleInlinedLandingPad(InvokeInst *II, BasicBlock *FirstNewBlock,
244 ClonedCodeInfo &InlinedCodeInfo) {
245 BasicBlock *InvokeDest = II->getUnwindDest();
247 Function *Caller = FirstNewBlock->getParent();
249 // The inlined code is currently at the end of the function, scan from the
250 // start of the inlined code to its end, checking for stuff we need to
252 LandingPadInliningInfo Invoke(II);
254 // Get all of the inlined landing pad instructions.
255 SmallPtrSet<LandingPadInst*, 16> InlinedLPads;
256 for (Function::iterator I = FirstNewBlock->getIterator(), E = Caller->end();
258 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator()))
259 InlinedLPads.insert(II->getLandingPadInst());
261 // Append the clauses from the outer landing pad instruction into the inlined
262 // landing pad instructions.
263 LandingPadInst *OuterLPad = Invoke.getLandingPadInst();
264 for (LandingPadInst *InlinedLPad : InlinedLPads) {
265 unsigned OuterNum = OuterLPad->getNumClauses();
266 InlinedLPad->reserveClauses(OuterNum);
267 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx)
268 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx));
269 if (OuterLPad->isCleanup())
270 InlinedLPad->setCleanup(true);
273 for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end();
275 if (InlinedCodeInfo.ContainsCalls)
276 if (BasicBlock *NewBB = HandleCallsInBlockInlinedThroughInvoke(
277 &*BB, Invoke.getOuterResumeDest()))
278 // Update any PHI nodes in the exceptional block to indicate that there
279 // is now a new entry in them.
280 Invoke.addIncomingPHIValuesFor(NewBB);
282 // Forward any resumes that are remaining here.
283 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator()))
284 Invoke.forwardResume(RI, InlinedLPads);
287 // Now that everything is happy, we have one final detail. The PHI nodes in
288 // the exception destination block still have entries due to the original
289 // invoke instruction. Eliminate these entries (which might even delete the
291 InvokeDest->removePredecessor(II->getParent());
294 /// If we inlined an invoke site, we need to convert calls
295 /// in the body of the inlined function into invokes.
297 /// II is the invoke instruction being inlined. FirstNewBlock is the first
298 /// block of the inlined code (the last block is the end of the function),
299 /// and InlineCodeInfo is information about the code that got inlined.
300 static void HandleInlinedEHPad(InvokeInst *II, BasicBlock *FirstNewBlock,
301 ClonedCodeInfo &InlinedCodeInfo) {
302 BasicBlock *UnwindDest = II->getUnwindDest();
303 Function *Caller = FirstNewBlock->getParent();
305 assert(UnwindDest->getFirstNonPHI()->isEHPad() && "unexpected BasicBlock!");
307 // If there are PHI nodes in the unwind destination block, we need to keep
308 // track of which values came into them from the invoke before removing the
309 // edge from this block.
310 SmallVector<Value *, 8> UnwindDestPHIValues;
311 llvm::BasicBlock *InvokeBB = II->getParent();
312 for (Instruction &I : *UnwindDest) {
313 // Save the value to use for this edge.
314 PHINode *PHI = dyn_cast<PHINode>(&I);
317 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
320 // Add incoming-PHI values to the unwind destination block for the given basic
321 // block, using the values for the original invoke's source block.
322 auto UpdatePHINodes = [&](BasicBlock *Src) {
323 BasicBlock::iterator I = UnwindDest->begin();
324 for (Value *V : UnwindDestPHIValues) {
325 PHINode *PHI = cast<PHINode>(I);
326 PHI->addIncoming(V, Src);
331 // Forward EH terminator instructions to the caller's invoke destination.
332 // This is as simple as connect all the instructions which 'unwind to caller'
333 // to the invoke destination.
334 for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end();
336 Instruction *I = BB->getFirstNonPHI();
338 if (auto *CEPI = dyn_cast<CatchEndPadInst>(I)) {
339 if (CEPI->unwindsToCaller()) {
340 CatchEndPadInst::Create(CEPI->getContext(), UnwindDest, CEPI);
341 CEPI->eraseFromParent();
342 UpdatePHINodes(&*BB);
344 } else if (auto *CEPI = dyn_cast<CleanupEndPadInst>(I)) {
345 if (CEPI->unwindsToCaller()) {
346 CleanupEndPadInst::Create(CEPI->getCleanupPad(), UnwindDest, CEPI);
347 CEPI->eraseFromParent();
348 UpdatePHINodes(&*BB);
350 } else if (auto *TPI = dyn_cast<TerminatePadInst>(I)) {
351 if (TPI->unwindsToCaller()) {
352 SmallVector<Value *, 3> TerminatePadArgs;
353 for (Value *ArgOperand : TPI->arg_operands())
354 TerminatePadArgs.push_back(ArgOperand);
355 TerminatePadInst::Create(TPI->getContext(), UnwindDest,
356 TerminatePadArgs, TPI);
357 TPI->eraseFromParent();
358 UpdatePHINodes(&*BB);
361 assert(isa<CatchPadInst>(I) || isa<CleanupPadInst>(I));
365 if (auto *CRI = dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
366 if (CRI->unwindsToCaller()) {
367 CleanupReturnInst::Create(CRI->getCleanupPad(), UnwindDest, CRI);
368 CRI->eraseFromParent();
369 UpdatePHINodes(&*BB);
374 if (InlinedCodeInfo.ContainsCalls)
375 for (Function::iterator BB = FirstNewBlock->getIterator(),
378 if (BasicBlock *NewBB =
379 HandleCallsInBlockInlinedThroughInvoke(&*BB, UnwindDest))
380 // Update any PHI nodes in the exceptional block to indicate that there
381 // is now a new entry in them.
382 UpdatePHINodes(NewBB);
384 // Now that everything is happy, we have one final detail. The PHI nodes in
385 // the exception destination block still have entries due to the original
386 // invoke instruction. Eliminate these entries (which might even delete the
388 UnwindDest->removePredecessor(InvokeBB);
391 /// When inlining a function that contains noalias scope metadata,
392 /// this metadata needs to be cloned so that the inlined blocks
393 /// have different "unqiue scopes" at every call site. Were this not done, then
394 /// aliasing scopes from a function inlined into a caller multiple times could
395 /// not be differentiated (and this would lead to miscompiles because the
396 /// non-aliasing property communicated by the metadata could have
397 /// call-site-specific control dependencies).
398 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) {
399 const Function *CalledFunc = CS.getCalledFunction();
400 SetVector<const MDNode *> MD;
402 // Note: We could only clone the metadata if it is already used in the
403 // caller. I'm omitting that check here because it might confuse
404 // inter-procedural alias analysis passes. We can revisit this if it becomes
405 // an efficiency or overhead problem.
407 for (Function::const_iterator I = CalledFunc->begin(), IE = CalledFunc->end();
409 for (BasicBlock::const_iterator J = I->begin(), JE = I->end(); J != JE; ++J) {
410 if (const MDNode *M = J->getMetadata(LLVMContext::MD_alias_scope))
412 if (const MDNode *M = J->getMetadata(LLVMContext::MD_noalias))
419 // Walk the existing metadata, adding the complete (perhaps cyclic) chain to
421 SmallVector<const Metadata *, 16> Queue(MD.begin(), MD.end());
422 while (!Queue.empty()) {
423 const MDNode *M = cast<MDNode>(Queue.pop_back_val());
424 for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i)
425 if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i)))
430 // Now we have a complete set of all metadata in the chains used to specify
431 // the noalias scopes and the lists of those scopes.
432 SmallVector<TempMDTuple, 16> DummyNodes;
433 DenseMap<const MDNode *, TrackingMDNodeRef> MDMap;
434 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
436 DummyNodes.push_back(MDTuple::getTemporary(CalledFunc->getContext(), None));
437 MDMap[*I].reset(DummyNodes.back().get());
440 // Create new metadata nodes to replace the dummy nodes, replacing old
441 // metadata references with either a dummy node or an already-created new
443 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
445 SmallVector<Metadata *, 4> NewOps;
446 for (unsigned i = 0, ie = (*I)->getNumOperands(); i != ie; ++i) {
447 const Metadata *V = (*I)->getOperand(i);
448 if (const MDNode *M = dyn_cast<MDNode>(V))
449 NewOps.push_back(MDMap[M]);
451 NewOps.push_back(const_cast<Metadata *>(V));
454 MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps);
455 MDTuple *TempM = cast<MDTuple>(MDMap[*I]);
456 assert(TempM->isTemporary() && "Expected temporary node");
458 TempM->replaceAllUsesWith(NewM);
461 // Now replace the metadata in the new inlined instructions with the
462 // repacements from the map.
463 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
464 VMI != VMIE; ++VMI) {
468 Instruction *NI = dyn_cast<Instruction>(VMI->second);
472 if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope)) {
473 MDNode *NewMD = MDMap[M];
474 // If the call site also had alias scope metadata (a list of scopes to
475 // which instructions inside it might belong), propagate those scopes to
476 // the inlined instructions.
478 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
479 NewMD = MDNode::concatenate(NewMD, CSM);
480 NI->setMetadata(LLVMContext::MD_alias_scope, NewMD);
481 } else if (NI->mayReadOrWriteMemory()) {
483 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
484 NI->setMetadata(LLVMContext::MD_alias_scope, M);
487 if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias)) {
488 MDNode *NewMD = MDMap[M];
489 // If the call site also had noalias metadata (a list of scopes with
490 // which instructions inside it don't alias), propagate those scopes to
491 // the inlined instructions.
493 CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
494 NewMD = MDNode::concatenate(NewMD, CSM);
495 NI->setMetadata(LLVMContext::MD_noalias, NewMD);
496 } else if (NI->mayReadOrWriteMemory()) {
497 if (MDNode *M = CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
498 NI->setMetadata(LLVMContext::MD_noalias, M);
503 /// If the inlined function has noalias arguments,
504 /// then add new alias scopes for each noalias argument, tag the mapped noalias
505 /// parameters with noalias metadata specifying the new scope, and tag all
506 /// non-derived loads, stores and memory intrinsics with the new alias scopes.
507 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
508 const DataLayout &DL, AAResults *CalleeAAR) {
509 if (!EnableNoAliasConversion)
512 const Function *CalledFunc = CS.getCalledFunction();
513 SmallVector<const Argument *, 4> NoAliasArgs;
515 for (const Argument &I : CalledFunc->args()) {
516 if (I.hasNoAliasAttr() && !I.hasNUses(0))
517 NoAliasArgs.push_back(&I);
520 if (NoAliasArgs.empty())
523 // To do a good job, if a noalias variable is captured, we need to know if
524 // the capture point dominates the particular use we're considering.
526 DT.recalculate(const_cast<Function&>(*CalledFunc));
528 // noalias indicates that pointer values based on the argument do not alias
529 // pointer values which are not based on it. So we add a new "scope" for each
530 // noalias function argument. Accesses using pointers based on that argument
531 // become part of that alias scope, accesses using pointers not based on that
532 // argument are tagged as noalias with that scope.
534 DenseMap<const Argument *, MDNode *> NewScopes;
535 MDBuilder MDB(CalledFunc->getContext());
537 // Create a new scope domain for this function.
539 MDB.createAnonymousAliasScopeDomain(CalledFunc->getName());
540 for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) {
541 const Argument *A = NoAliasArgs[i];
543 std::string Name = CalledFunc->getName();
546 Name += A->getName();
548 Name += ": argument ";
552 // Note: We always create a new anonymous root here. This is true regardless
553 // of the linkage of the callee because the aliasing "scope" is not just a
554 // property of the callee, but also all control dependencies in the caller.
555 MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
556 NewScopes.insert(std::make_pair(A, NewScope));
559 // Iterate over all new instructions in the map; for all memory-access
560 // instructions, add the alias scope metadata.
561 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
562 VMI != VMIE; ++VMI) {
563 if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) {
567 Instruction *NI = dyn_cast<Instruction>(VMI->second);
571 bool IsArgMemOnlyCall = false, IsFuncCall = false;
572 SmallVector<const Value *, 2> PtrArgs;
574 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
575 PtrArgs.push_back(LI->getPointerOperand());
576 else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
577 PtrArgs.push_back(SI->getPointerOperand());
578 else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
579 PtrArgs.push_back(VAAI->getPointerOperand());
580 else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
581 PtrArgs.push_back(CXI->getPointerOperand());
582 else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
583 PtrArgs.push_back(RMWI->getPointerOperand());
584 else if (ImmutableCallSite ICS = ImmutableCallSite(I)) {
585 // If we know that the call does not access memory, then we'll still
586 // know that about the inlined clone of this call site, and we don't
587 // need to add metadata.
588 if (ICS.doesNotAccessMemory())
593 FunctionModRefBehavior MRB = CalleeAAR->getModRefBehavior(ICS);
594 if (MRB == FMRB_OnlyAccessesArgumentPointees ||
595 MRB == FMRB_OnlyReadsArgumentPointees)
596 IsArgMemOnlyCall = true;
599 for (ImmutableCallSite::arg_iterator AI = ICS.arg_begin(),
600 AE = ICS.arg_end(); AI != AE; ++AI) {
601 // We need to check the underlying objects of all arguments, not just
602 // the pointer arguments, because we might be passing pointers as
604 // However, if we know that the call only accesses pointer arguments,
605 // then we only need to check the pointer arguments.
606 if (IsArgMemOnlyCall && !(*AI)->getType()->isPointerTy())
609 PtrArgs.push_back(*AI);
613 // If we found no pointers, then this instruction is not suitable for
614 // pairing with an instruction to receive aliasing metadata.
615 // However, if this is a call, this we might just alias with none of the
616 // noalias arguments.
617 if (PtrArgs.empty() && !IsFuncCall)
620 // It is possible that there is only one underlying object, but you
621 // need to go through several PHIs to see it, and thus could be
622 // repeated in the Objects list.
623 SmallPtrSet<const Value *, 4> ObjSet;
624 SmallVector<Metadata *, 4> Scopes, NoAliases;
626 SmallSetVector<const Argument *, 4> NAPtrArgs;
627 for (unsigned i = 0, ie = PtrArgs.size(); i != ie; ++i) {
628 SmallVector<Value *, 4> Objects;
629 GetUnderlyingObjects(const_cast<Value*>(PtrArgs[i]),
630 Objects, DL, /* LI = */ nullptr);
632 for (Value *O : Objects)
636 // Figure out if we're derived from anything that is not a noalias
638 bool CanDeriveViaCapture = false, UsesAliasingPtr = false;
639 for (const Value *V : ObjSet) {
640 // Is this value a constant that cannot be derived from any pointer
641 // value (we need to exclude constant expressions, for example, that
642 // are formed from arithmetic on global symbols).
643 bool IsNonPtrConst = isa<ConstantInt>(V) || isa<ConstantFP>(V) ||
644 isa<ConstantPointerNull>(V) ||
645 isa<ConstantDataVector>(V) || isa<UndefValue>(V);
649 // If this is anything other than a noalias argument, then we cannot
650 // completely describe the aliasing properties using alias.scope
651 // metadata (and, thus, won't add any).
652 if (const Argument *A = dyn_cast<Argument>(V)) {
653 if (!A->hasNoAliasAttr())
654 UsesAliasingPtr = true;
656 UsesAliasingPtr = true;
659 // If this is not some identified function-local object (which cannot
660 // directly alias a noalias argument), or some other argument (which,
661 // by definition, also cannot alias a noalias argument), then we could
662 // alias a noalias argument that has been captured).
663 if (!isa<Argument>(V) &&
664 !isIdentifiedFunctionLocal(const_cast<Value*>(V)))
665 CanDeriveViaCapture = true;
668 // A function call can always get captured noalias pointers (via other
669 // parameters, globals, etc.).
670 if (IsFuncCall && !IsArgMemOnlyCall)
671 CanDeriveViaCapture = true;
673 // First, we want to figure out all of the sets with which we definitely
674 // don't alias. Iterate over all noalias set, and add those for which:
675 // 1. The noalias argument is not in the set of objects from which we
676 // definitely derive.
677 // 2. The noalias argument has not yet been captured.
678 // An arbitrary function that might load pointers could see captured
679 // noalias arguments via other noalias arguments or globals, and so we
680 // must always check for prior capture.
681 for (const Argument *A : NoAliasArgs) {
682 if (!ObjSet.count(A) && (!CanDeriveViaCapture ||
683 // It might be tempting to skip the
684 // PointerMayBeCapturedBefore check if
685 // A->hasNoCaptureAttr() is true, but this is
686 // incorrect because nocapture only guarantees
687 // that no copies outlive the function, not
688 // that the value cannot be locally captured.
689 !PointerMayBeCapturedBefore(A,
690 /* ReturnCaptures */ false,
691 /* StoreCaptures */ false, I, &DT)))
692 NoAliases.push_back(NewScopes[A]);
695 if (!NoAliases.empty())
696 NI->setMetadata(LLVMContext::MD_noalias,
698 NI->getMetadata(LLVMContext::MD_noalias),
699 MDNode::get(CalledFunc->getContext(), NoAliases)));
701 // Next, we want to figure out all of the sets to which we might belong.
702 // We might belong to a set if the noalias argument is in the set of
703 // underlying objects. If there is some non-noalias argument in our list
704 // of underlying objects, then we cannot add a scope because the fact
705 // that some access does not alias with any set of our noalias arguments
706 // cannot itself guarantee that it does not alias with this access
707 // (because there is some pointer of unknown origin involved and the
708 // other access might also depend on this pointer). We also cannot add
709 // scopes to arbitrary functions unless we know they don't access any
710 // non-parameter pointer-values.
711 bool CanAddScopes = !UsesAliasingPtr;
712 if (CanAddScopes && IsFuncCall)
713 CanAddScopes = IsArgMemOnlyCall;
716 for (const Argument *A : NoAliasArgs) {
718 Scopes.push_back(NewScopes[A]);
723 LLVMContext::MD_alias_scope,
724 MDNode::concatenate(NI->getMetadata(LLVMContext::MD_alias_scope),
725 MDNode::get(CalledFunc->getContext(), Scopes)));
730 /// If the inlined function has non-byval align arguments, then
731 /// add @llvm.assume-based alignment assumptions to preserve this information.
732 static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) {
733 if (!PreserveAlignmentAssumptions)
735 auto &DL = CS.getCaller()->getParent()->getDataLayout();
737 // To avoid inserting redundant assumptions, we should check for assumptions
738 // already in the caller. To do this, we might need a DT of the caller.
740 bool DTCalculated = false;
742 Function *CalledFunc = CS.getCalledFunction();
743 for (Function::arg_iterator I = CalledFunc->arg_begin(),
744 E = CalledFunc->arg_end();
746 unsigned Align = I->getType()->isPointerTy() ? I->getParamAlignment() : 0;
747 if (Align && !I->hasByValOrInAllocaAttr() && !I->hasNUses(0)) {
749 DT.recalculate(const_cast<Function&>(*CS.getInstruction()->getParent()
754 // If we can already prove the asserted alignment in the context of the
755 // caller, then don't bother inserting the assumption.
756 Value *Arg = CS.getArgument(I->getArgNo());
757 if (getKnownAlignment(Arg, DL, CS.getInstruction(),
758 &IFI.ACT->getAssumptionCache(*CS.getCaller()),
762 IRBuilder<>(CS.getInstruction())
763 .CreateAlignmentAssumption(DL, Arg, Align);
768 /// Once we have cloned code over from a callee into the caller,
769 /// update the specified callgraph to reflect the changes we made.
770 /// Note that it's possible that not all code was copied over, so only
771 /// some edges of the callgraph may remain.
772 static void UpdateCallGraphAfterInlining(CallSite CS,
773 Function::iterator FirstNewBlock,
774 ValueToValueMapTy &VMap,
775 InlineFunctionInfo &IFI) {
776 CallGraph &CG = *IFI.CG;
777 const Function *Caller = CS.getInstruction()->getParent()->getParent();
778 const Function *Callee = CS.getCalledFunction();
779 CallGraphNode *CalleeNode = CG[Callee];
780 CallGraphNode *CallerNode = CG[Caller];
782 // Since we inlined some uninlined call sites in the callee into the caller,
783 // add edges from the caller to all of the callees of the callee.
784 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
786 // Consider the case where CalleeNode == CallerNode.
787 CallGraphNode::CalledFunctionsVector CallCache;
788 if (CalleeNode == CallerNode) {
789 CallCache.assign(I, E);
790 I = CallCache.begin();
794 for (; I != E; ++I) {
795 const Value *OrigCall = I->first;
797 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
798 // Only copy the edge if the call was inlined!
799 if (VMI == VMap.end() || VMI->second == nullptr)
802 // If the call was inlined, but then constant folded, there is no edge to
803 // add. Check for this case.
804 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
808 // We do not treat intrinsic calls like real function calls because we
809 // expect them to become inline code; do not add an edge for an intrinsic.
810 CallSite CS = CallSite(NewCall);
811 if (CS && CS.getCalledFunction() && CS.getCalledFunction()->isIntrinsic())
814 // Remember that this call site got inlined for the client of
816 IFI.InlinedCalls.push_back(NewCall);
818 // It's possible that inlining the callsite will cause it to go from an
819 // indirect to a direct call by resolving a function pointer. If this
820 // happens, set the callee of the new call site to a more precise
821 // destination. This can also happen if the call graph node of the caller
822 // was just unnecessarily imprecise.
823 if (!I->second->getFunction())
824 if (Function *F = CallSite(NewCall).getCalledFunction()) {
825 // Indirect call site resolved to direct call.
826 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
831 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
834 // Update the call graph by deleting the edge from Callee to Caller. We must
835 // do this after the loop above in case Caller and Callee are the same.
836 CallerNode->removeCallEdgeFor(CS);
839 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
840 BasicBlock *InsertBlock,
841 InlineFunctionInfo &IFI) {
842 Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
843 IRBuilder<> Builder(InsertBlock, InsertBlock->begin());
845 Value *Size = Builder.getInt64(M->getDataLayout().getTypeStoreSize(AggTy));
847 // Always generate a memcpy of alignment 1 here because we don't know
848 // the alignment of the src pointer. Other optimizations can infer
850 Builder.CreateMemCpy(Dst, Src, Size, /*Align=*/1);
853 /// When inlining a call site that has a byval argument,
854 /// we have to make the implicit memcpy explicit by adding it.
855 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
856 const Function *CalledFunc,
857 InlineFunctionInfo &IFI,
858 unsigned ByValAlignment) {
859 PointerType *ArgTy = cast<PointerType>(Arg->getType());
860 Type *AggTy = ArgTy->getElementType();
862 Function *Caller = TheCall->getParent()->getParent();
864 // If the called function is readonly, then it could not mutate the caller's
865 // copy of the byval'd memory. In this case, it is safe to elide the copy and
867 if (CalledFunc->onlyReadsMemory()) {
868 // If the byval argument has a specified alignment that is greater than the
869 // passed in pointer, then we either have to round up the input pointer or
870 // give up on this transformation.
871 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
874 const DataLayout &DL = Caller->getParent()->getDataLayout();
876 // If the pointer is already known to be sufficiently aligned, or if we can
877 // round it up to a larger alignment, then we don't need a temporary.
878 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, DL, TheCall,
879 &IFI.ACT->getAssumptionCache(*Caller)) >=
883 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
884 // for code quality, but rarely happens and is required for correctness.
887 // Create the alloca. If we have DataLayout, use nice alignment.
889 Caller->getParent()->getDataLayout().getPrefTypeAlignment(AggTy);
891 // If the byval had an alignment specified, we *must* use at least that
892 // alignment, as it is required by the byval argument (and uses of the
893 // pointer inside the callee).
894 Align = std::max(Align, ByValAlignment);
896 Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(),
897 &*Caller->begin()->begin());
898 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
900 // Uses of the argument in the function should use our new alloca
905 // Check whether this Value is used by a lifetime intrinsic.
906 static bool isUsedByLifetimeMarker(Value *V) {
907 for (User *U : V->users()) {
908 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
909 switch (II->getIntrinsicID()) {
911 case Intrinsic::lifetime_start:
912 case Intrinsic::lifetime_end:
920 // Check whether the given alloca already has
921 // lifetime.start or lifetime.end intrinsics.
922 static bool hasLifetimeMarkers(AllocaInst *AI) {
923 Type *Ty = AI->getType();
924 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
925 Ty->getPointerAddressSpace());
927 return isUsedByLifetimeMarker(AI);
929 // Do a scan to find all the casts to i8*.
930 for (User *U : AI->users()) {
931 if (U->getType() != Int8PtrTy) continue;
932 if (U->stripPointerCasts() != AI) continue;
933 if (isUsedByLifetimeMarker(U))
939 /// Rebuild the entire inlined-at chain for this instruction so that the top of
940 /// the chain now is inlined-at the new call site.
942 updateInlinedAtInfo(DebugLoc DL, DILocation *InlinedAtNode, LLVMContext &Ctx,
943 DenseMap<const DILocation *, DILocation *> &IANodes) {
944 SmallVector<DILocation *, 3> InlinedAtLocations;
945 DILocation *Last = InlinedAtNode;
946 DILocation *CurInlinedAt = DL;
948 // Gather all the inlined-at nodes
949 while (DILocation *IA = CurInlinedAt->getInlinedAt()) {
950 // Skip any we've already built nodes for
951 if (DILocation *Found = IANodes[IA]) {
956 InlinedAtLocations.push_back(IA);
960 // Starting from the top, rebuild the nodes to point to the new inlined-at
961 // location (then rebuilding the rest of the chain behind it) and update the
962 // map of already-constructed inlined-at nodes.
963 for (const DILocation *MD : make_range(InlinedAtLocations.rbegin(),
964 InlinedAtLocations.rend())) {
965 Last = IANodes[MD] = DILocation::getDistinct(
966 Ctx, MD->getLine(), MD->getColumn(), MD->getScope(), Last);
969 // And finally create the normal location for this instruction, referring to
970 // the new inlined-at chain.
971 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(), Last);
974 /// Update inlined instructions' line numbers to
975 /// to encode location where these instructions are inlined.
976 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
977 Instruction *TheCall) {
978 DebugLoc TheCallDL = TheCall->getDebugLoc();
982 auto &Ctx = Fn->getContext();
983 DILocation *InlinedAtNode = TheCallDL;
985 // Create a unique call site, not to be confused with any other call from the
987 InlinedAtNode = DILocation::getDistinct(
988 Ctx, InlinedAtNode->getLine(), InlinedAtNode->getColumn(),
989 InlinedAtNode->getScope(), InlinedAtNode->getInlinedAt());
991 // Cache the inlined-at nodes as they're built so they are reused, without
992 // this every instruction's inlined-at chain would become distinct from each
994 DenseMap<const DILocation *, DILocation *> IANodes;
996 for (; FI != Fn->end(); ++FI) {
997 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
999 DebugLoc DL = BI->getDebugLoc();
1001 // If the inlined instruction has no line number, make it look as if it
1002 // originates from the call location. This is important for
1003 // ((__always_inline__, __nodebug__)) functions which must use caller
1004 // location for all instructions in their function body.
1006 // Don't update static allocas, as they may get moved later.
1007 if (auto *AI = dyn_cast<AllocaInst>(BI))
1008 if (isa<Constant>(AI->getArraySize()))
1011 BI->setDebugLoc(TheCallDL);
1013 BI->setDebugLoc(updateInlinedAtInfo(DL, InlinedAtNode, BI->getContext(), IANodes));
1019 /// This function inlines the called function into the basic block of the
1020 /// caller. This returns false if it is not possible to inline this call.
1021 /// The program is still in a well defined state if this occurs though.
1023 /// Note that this only does one level of inlining. For example, if the
1024 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
1025 /// exists in the instruction stream. Similarly this will inline a recursive
1026 /// function by one level.
1027 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
1028 AAResults *CalleeAAR, bool InsertLifetime) {
1029 Instruction *TheCall = CS.getInstruction();
1030 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
1031 "Instruction not in function!");
1033 // If IFI has any state in it, zap it before we fill it in.
1036 const Function *CalledFunc = CS.getCalledFunction();
1037 if (!CalledFunc || // Can't inline external function or indirect
1038 CalledFunc->isDeclaration() || // call, or call to a vararg function!
1039 CalledFunc->getFunctionType()->isVarArg()) return false;
1041 // The inliner does not know how to inline through calls with operand bundles
1043 if (CS.hasOperandBundles()) {
1044 // ... but it knows how to inline through "deopt" operand bundles.
1046 CS.getNumOperandBundles() == 1 &&
1047 CS.getOperandBundleAt(0).getTagID() == LLVMContext::OB_deopt;
1052 // If the call to the callee cannot throw, set the 'nounwind' flag on any
1053 // calls that we inline.
1054 bool MarkNoUnwind = CS.doesNotThrow();
1056 BasicBlock *OrigBB = TheCall->getParent();
1057 Function *Caller = OrigBB->getParent();
1059 // GC poses two hazards to inlining, which only occur when the callee has GC:
1060 // 1. If the caller has no GC, then the callee's GC must be propagated to the
1062 // 2. If the caller has a differing GC, it is invalid to inline.
1063 if (CalledFunc->hasGC()) {
1064 if (!Caller->hasGC())
1065 Caller->setGC(CalledFunc->getGC());
1066 else if (CalledFunc->getGC() != Caller->getGC())
1070 // Get the personality function from the callee if it contains a landing pad.
1071 Constant *CalledPersonality =
1072 CalledFunc->hasPersonalityFn()
1073 ? CalledFunc->getPersonalityFn()->stripPointerCasts()
1076 // Find the personality function used by the landing pads of the caller. If it
1077 // exists, then check to see that it matches the personality function used in
1079 Constant *CallerPersonality =
1080 Caller->hasPersonalityFn()
1081 ? Caller->getPersonalityFn()->stripPointerCasts()
1083 if (CalledPersonality) {
1084 if (!CallerPersonality)
1085 Caller->setPersonalityFn(CalledPersonality);
1086 // If the personality functions match, then we can perform the
1087 // inlining. Otherwise, we can't inline.
1088 // TODO: This isn't 100% true. Some personality functions are proper
1089 // supersets of others and can be used in place of the other.
1090 else if (CalledPersonality != CallerPersonality)
1094 // Get an iterator to the last basic block in the function, which will have
1095 // the new function inlined after it.
1096 Function::iterator LastBlock = --Caller->end();
1098 // Make sure to capture all of the return instructions from the cloned
1100 SmallVector<ReturnInst*, 8> Returns;
1101 ClonedCodeInfo InlinedFunctionInfo;
1102 Function::iterator FirstNewBlock;
1104 { // Scope to destroy VMap after cloning.
1105 ValueToValueMapTy VMap;
1106 // Keep a list of pair (dst, src) to emit byval initializations.
1107 SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
1109 auto &DL = Caller->getParent()->getDataLayout();
1111 assert(CalledFunc->arg_size() == CS.arg_size() &&
1112 "No varargs calls can be inlined!");
1114 // Calculate the vector of arguments to pass into the function cloner, which
1115 // matches up the formal to the actual argument values.
1116 CallSite::arg_iterator AI = CS.arg_begin();
1118 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
1119 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
1120 Value *ActualArg = *AI;
1122 // When byval arguments actually inlined, we need to make the copy implied
1123 // by them explicit. However, we don't do this if the callee is readonly
1124 // or readnone, because the copy would be unneeded: the callee doesn't
1125 // modify the struct.
1126 if (CS.isByValArgument(ArgNo)) {
1127 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
1128 CalledFunc->getParamAlignment(ArgNo+1));
1129 if (ActualArg != *AI)
1130 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
1133 VMap[&*I] = ActualArg;
1136 // Add alignment assumptions if necessary. We do this before the inlined
1137 // instructions are actually cloned into the caller so that we can easily
1138 // check what will be known at the start of the inlined code.
1139 AddAlignmentAssumptions(CS, IFI);
1141 // We want the inliner to prune the code as it copies. We would LOVE to
1142 // have no dead or constant instructions leftover after inlining occurs
1143 // (which can happen, e.g., because an argument was constant), but we'll be
1144 // happy with whatever the cloner can do.
1145 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
1146 /*ModuleLevelChanges=*/false, Returns, ".i",
1147 &InlinedFunctionInfo, TheCall);
1149 // Remember the first block that is newly cloned over.
1150 FirstNewBlock = LastBlock; ++FirstNewBlock;
1152 // Inject byval arguments initialization.
1153 for (std::pair<Value*, Value*> &Init : ByValInit)
1154 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
1155 &*FirstNewBlock, IFI);
1157 if (CS.hasOperandBundles()) {
1158 auto ParentDeopt = CS.getOperandBundleAt(0);
1159 assert(ParentDeopt.getTagID() == LLVMContext::OB_deopt &&
1160 "Checked on entry!");
1162 SmallVector<OperandBundleDef, 2> OpDefs;
1164 for (auto &VH : InlinedFunctionInfo.OperandBundleCallSites) {
1165 Instruction *I = VH;
1170 OpDefs.reserve(ICS.getNumOperandBundles());
1172 for (unsigned i = 0, e = ICS.getNumOperandBundles(); i < e; ++i) {
1173 auto ChildOB = ICS.getOperandBundleAt(i);
1174 if (ChildOB.getTagID() != LLVMContext::OB_deopt) {
1175 // If the inlined call has other operand bundles, let them be
1176 OpDefs.emplace_back(ChildOB);
1180 // It may be useful to separate this logic (of handling operand
1181 // bundles) out to a separate "policy" component if this gets crowded.
1182 // Prepend the parent's deoptimization continuation to the newly
1183 // inlined call's deoptimization continuation.
1184 std::vector<Value *> MergedDeoptArgs;
1185 MergedDeoptArgs.reserve(ParentDeopt.Inputs.size() +
1186 ChildOB.Inputs.size());
1188 MergedDeoptArgs.insert(MergedDeoptArgs.end(),
1189 ParentDeopt.Inputs.begin(),
1190 ParentDeopt.Inputs.end());
1191 MergedDeoptArgs.insert(MergedDeoptArgs.end(), ChildOB.Inputs.begin(),
1192 ChildOB.Inputs.end());
1194 OpDefs.emplace_back(StringRef("deopt"), std::move(MergedDeoptArgs));
1197 Instruction *NewI = nullptr;
1198 if (isa<CallInst>(I))
1199 NewI = CallInst::Create(cast<CallInst>(I), OpDefs, I);
1201 NewI = InvokeInst::Create(cast<InvokeInst>(I), OpDefs, I);
1203 // Note: the RAUW does the appropriate fixup in VMap, so we need to do
1204 // this even if the call returns void.
1205 I->replaceAllUsesWith(NewI);
1208 I->eraseFromParent();
1212 // Update the callgraph if requested.
1214 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
1216 // Update inlined instructions' line number information.
1217 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
1219 // Clone existing noalias metadata if necessary.
1220 CloneAliasScopeMetadata(CS, VMap);
1222 // Add noalias metadata if necessary.
1223 AddAliasScopeMetadata(CS, VMap, DL, CalleeAAR);
1225 // FIXME: We could register any cloned assumptions instead of clearing the
1226 // whole function's cache.
1228 IFI.ACT->getAssumptionCache(*Caller).clear();
1231 // If there are any alloca instructions in the block that used to be the entry
1232 // block for the callee, move them to the entry block of the caller. First
1233 // calculate which instruction they should be inserted before. We insert the
1234 // instructions at the end of the current alloca list.
1236 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
1237 for (BasicBlock::iterator I = FirstNewBlock->begin(),
1238 E = FirstNewBlock->end(); I != E; ) {
1239 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
1242 // If the alloca is now dead, remove it. This often occurs due to code
1244 if (AI->use_empty()) {
1245 AI->eraseFromParent();
1249 if (!isa<Constant>(AI->getArraySize()))
1252 // Keep track of the static allocas that we inline into the caller.
1253 IFI.StaticAllocas.push_back(AI);
1255 // Scan for the block of allocas that we can move over, and move them
1257 while (isa<AllocaInst>(I) &&
1258 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
1259 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
1263 // Transfer all of the allocas over in a block. Using splice means
1264 // that the instructions aren't removed from the symbol table, then
1266 Caller->getEntryBlock().getInstList().splice(
1267 InsertPoint, FirstNewBlock->getInstList(), AI->getIterator(), I);
1269 // Move any dbg.declares describing the allocas into the entry basic block.
1270 DIBuilder DIB(*Caller->getParent());
1271 for (auto &AI : IFI.StaticAllocas)
1272 replaceDbgDeclareForAlloca(AI, AI, DIB, /*Deref=*/false);
1275 bool InlinedMustTailCalls = false;
1276 if (InlinedFunctionInfo.ContainsCalls) {
1277 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
1278 if (CallInst *CI = dyn_cast<CallInst>(TheCall))
1279 CallSiteTailKind = CI->getTailCallKind();
1281 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
1283 for (Instruction &I : *BB) {
1284 CallInst *CI = dyn_cast<CallInst>(&I);
1288 // We need to reduce the strength of any inlined tail calls. For
1289 // musttail, we have to avoid introducing potential unbounded stack
1290 // growth. For example, if functions 'f' and 'g' are mutually recursive
1291 // with musttail, we can inline 'g' into 'f' so long as we preserve
1292 // musttail on the cloned call to 'f'. If either the inlined call site
1293 // or the cloned call site is *not* musttail, the program already has
1294 // one frame of stack growth, so it's safe to remove musttail. Here is
1295 // a table of example transformations:
1297 // f -> musttail g -> musttail f ==> f -> musttail f
1298 // f -> musttail g -> tail f ==> f -> tail f
1299 // f -> g -> musttail f ==> f -> f
1300 // f -> g -> tail f ==> f -> f
1301 CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
1302 ChildTCK = std::min(CallSiteTailKind, ChildTCK);
1303 CI->setTailCallKind(ChildTCK);
1304 InlinedMustTailCalls |= CI->isMustTailCall();
1306 // Calls inlined through a 'nounwind' call site should be marked
1309 CI->setDoesNotThrow();
1314 // Leave lifetime markers for the static alloca's, scoping them to the
1315 // function we just inlined.
1316 if (InsertLifetime && !IFI.StaticAllocas.empty()) {
1317 IRBuilder<> builder(&FirstNewBlock->front());
1318 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
1319 AllocaInst *AI = IFI.StaticAllocas[ai];
1321 // If the alloca is already scoped to something smaller than the whole
1322 // function then there's no need to add redundant, less accurate markers.
1323 if (hasLifetimeMarkers(AI))
1326 // Try to determine the size of the allocation.
1327 ConstantInt *AllocaSize = nullptr;
1328 if (ConstantInt *AIArraySize =
1329 dyn_cast<ConstantInt>(AI->getArraySize())) {
1330 auto &DL = Caller->getParent()->getDataLayout();
1331 Type *AllocaType = AI->getAllocatedType();
1332 uint64_t AllocaTypeSize = DL.getTypeAllocSize(AllocaType);
1333 uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
1335 // Don't add markers for zero-sized allocas.
1336 if (AllocaArraySize == 0)
1339 // Check that array size doesn't saturate uint64_t and doesn't
1340 // overflow when it's multiplied by type size.
1341 if (AllocaArraySize != ~0ULL &&
1342 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
1343 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
1344 AllocaArraySize * AllocaTypeSize);
1348 builder.CreateLifetimeStart(AI, AllocaSize);
1349 for (ReturnInst *RI : Returns) {
1350 // Don't insert llvm.lifetime.end calls between a musttail call and a
1351 // return. The return kills all local allocas.
1352 if (InlinedMustTailCalls &&
1353 RI->getParent()->getTerminatingMustTailCall())
1355 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
1360 // If the inlined code contained dynamic alloca instructions, wrap the inlined
1361 // code with llvm.stacksave/llvm.stackrestore intrinsics.
1362 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
1363 Module *M = Caller->getParent();
1364 // Get the two intrinsics we care about.
1365 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
1366 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
1368 // Insert the llvm.stacksave.
1369 CallInst *SavedPtr = IRBuilder<>(&*FirstNewBlock, FirstNewBlock->begin())
1370 .CreateCall(StackSave, {}, "savedstack");
1372 // Insert a call to llvm.stackrestore before any return instructions in the
1373 // inlined function.
1374 for (ReturnInst *RI : Returns) {
1375 // Don't insert llvm.stackrestore calls between a musttail call and a
1376 // return. The return will restore the stack pointer.
1377 if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
1379 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
1383 // If we are inlining for an invoke instruction, we must make sure to rewrite
1384 // any call instructions into invoke instructions.
1385 if (auto *II = dyn_cast<InvokeInst>(TheCall)) {
1386 BasicBlock *UnwindDest = II->getUnwindDest();
1387 Instruction *FirstNonPHI = UnwindDest->getFirstNonPHI();
1388 if (isa<LandingPadInst>(FirstNonPHI)) {
1389 HandleInlinedLandingPad(II, &*FirstNewBlock, InlinedFunctionInfo);
1391 HandleInlinedEHPad(II, &*FirstNewBlock, InlinedFunctionInfo);
1395 // Handle any inlined musttail call sites. In order for a new call site to be
1396 // musttail, the source of the clone and the inlined call site must have been
1397 // musttail. Therefore it's safe to return without merging control into the
1399 if (InlinedMustTailCalls) {
1400 // Check if we need to bitcast the result of any musttail calls.
1401 Type *NewRetTy = Caller->getReturnType();
1402 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
1404 // Handle the returns preceded by musttail calls separately.
1405 SmallVector<ReturnInst *, 8> NormalReturns;
1406 for (ReturnInst *RI : Returns) {
1407 CallInst *ReturnedMustTail =
1408 RI->getParent()->getTerminatingMustTailCall();
1409 if (!ReturnedMustTail) {
1410 NormalReturns.push_back(RI);
1416 // Delete the old return and any preceding bitcast.
1417 BasicBlock *CurBB = RI->getParent();
1418 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
1419 RI->eraseFromParent();
1421 OldCast->eraseFromParent();
1423 // Insert a new bitcast and return with the right type.
1424 IRBuilder<> Builder(CurBB);
1425 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
1428 // Leave behind the normal returns so we can merge control flow.
1429 std::swap(Returns, NormalReturns);
1432 // If we cloned in _exactly one_ basic block, and if that block ends in a
1433 // return instruction, we splice the body of the inlined callee directly into
1434 // the calling basic block.
1435 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
1436 // Move all of the instructions right before the call.
1437 OrigBB->getInstList().splice(TheCall->getIterator(),
1438 FirstNewBlock->getInstList(),
1439 FirstNewBlock->begin(), FirstNewBlock->end());
1440 // Remove the cloned basic block.
1441 Caller->getBasicBlockList().pop_back();
1443 // If the call site was an invoke instruction, add a branch to the normal
1445 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1446 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
1447 NewBr->setDebugLoc(Returns[0]->getDebugLoc());
1450 // If the return instruction returned a value, replace uses of the call with
1451 // uses of the returned value.
1452 if (!TheCall->use_empty()) {
1453 ReturnInst *R = Returns[0];
1454 if (TheCall == R->getReturnValue())
1455 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1457 TheCall->replaceAllUsesWith(R->getReturnValue());
1459 // Since we are now done with the Call/Invoke, we can delete it.
1460 TheCall->eraseFromParent();
1462 // Since we are now done with the return instruction, delete it also.
1463 Returns[0]->eraseFromParent();
1465 // We are now done with the inlining.
1469 // Otherwise, we have the normal case, of more than one block to inline or
1470 // multiple return sites.
1472 // We want to clone the entire callee function into the hole between the
1473 // "starter" and "ender" blocks. How we accomplish this depends on whether
1474 // this is an invoke instruction or a call instruction.
1475 BasicBlock *AfterCallBB;
1476 BranchInst *CreatedBranchToNormalDest = nullptr;
1477 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1479 // Add an unconditional branch to make this look like the CallInst case...
1480 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
1482 // Split the basic block. This guarantees that no PHI nodes will have to be
1483 // updated due to new incoming edges, and make the invoke case more
1484 // symmetric to the call case.
1486 OrigBB->splitBasicBlock(CreatedBranchToNormalDest->getIterator(),
1487 CalledFunc->getName() + ".exit");
1489 } else { // It's a call
1490 // If this is a call instruction, we need to split the basic block that
1491 // the call lives in.
1493 AfterCallBB = OrigBB->splitBasicBlock(TheCall->getIterator(),
1494 CalledFunc->getName() + ".exit");
1497 // Change the branch that used to go to AfterCallBB to branch to the first
1498 // basic block of the inlined function.
1500 TerminatorInst *Br = OrigBB->getTerminator();
1501 assert(Br && Br->getOpcode() == Instruction::Br &&
1502 "splitBasicBlock broken!");
1503 Br->setOperand(0, &*FirstNewBlock);
1505 // Now that the function is correct, make it a little bit nicer. In
1506 // particular, move the basic blocks inserted from the end of the function
1507 // into the space made by splitting the source basic block.
1508 Caller->getBasicBlockList().splice(AfterCallBB->getIterator(),
1509 Caller->getBasicBlockList(), FirstNewBlock,
1512 // Handle all of the return instructions that we just cloned in, and eliminate
1513 // any users of the original call/invoke instruction.
1514 Type *RTy = CalledFunc->getReturnType();
1516 PHINode *PHI = nullptr;
1517 if (Returns.size() > 1) {
1518 // The PHI node should go at the front of the new basic block to merge all
1519 // possible incoming values.
1520 if (!TheCall->use_empty()) {
1521 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
1522 &AfterCallBB->front());
1523 // Anything that used the result of the function call should now use the
1524 // PHI node as their operand.
1525 TheCall->replaceAllUsesWith(PHI);
1528 // Loop over all of the return instructions adding entries to the PHI node
1531 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1532 ReturnInst *RI = Returns[i];
1533 assert(RI->getReturnValue()->getType() == PHI->getType() &&
1534 "Ret value not consistent in function!");
1535 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
1539 // Add a branch to the merge points and remove return instructions.
1541 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1542 ReturnInst *RI = Returns[i];
1543 BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
1544 Loc = RI->getDebugLoc();
1545 BI->setDebugLoc(Loc);
1546 RI->eraseFromParent();
1548 // We need to set the debug location to *somewhere* inside the
1549 // inlined function. The line number may be nonsensical, but the
1550 // instruction will at least be associated with the right
1552 if (CreatedBranchToNormalDest)
1553 CreatedBranchToNormalDest->setDebugLoc(Loc);
1554 } else if (!Returns.empty()) {
1555 // Otherwise, if there is exactly one return value, just replace anything
1556 // using the return value of the call with the computed value.
1557 if (!TheCall->use_empty()) {
1558 if (TheCall == Returns[0]->getReturnValue())
1559 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1561 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
1564 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
1565 BasicBlock *ReturnBB = Returns[0]->getParent();
1566 ReturnBB->replaceAllUsesWith(AfterCallBB);
1568 // Splice the code from the return block into the block that it will return
1569 // to, which contains the code that was after the call.
1570 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
1571 ReturnBB->getInstList());
1573 if (CreatedBranchToNormalDest)
1574 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
1576 // Delete the return instruction now and empty ReturnBB now.
1577 Returns[0]->eraseFromParent();
1578 ReturnBB->eraseFromParent();
1579 } else if (!TheCall->use_empty()) {
1580 // No returns, but something is using the return value of the call. Just
1582 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1585 // Since we are now done with the Call/Invoke, we can delete it.
1586 TheCall->eraseFromParent();
1588 // If we inlined any musttail calls and the original return is now
1589 // unreachable, delete it. It can only contain a bitcast and ret.
1590 if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB))
1591 AfterCallBB->eraseFromParent();
1593 // We should always be able to fold the entry block of the function into the
1594 // single predecessor of the block...
1595 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
1596 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
1598 // Splice the code entry block into calling block, right before the
1599 // unconditional branch.
1600 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
1601 OrigBB->getInstList().splice(Br->getIterator(), CalleeEntry->getInstList());
1603 // Remove the unconditional branch.
1604 OrigBB->getInstList().erase(Br);
1606 // Now we can remove the CalleeEntry block, which is now empty.
1607 Caller->getBasicBlockList().erase(CalleeEntry);
1609 // If we inserted a phi node, check to see if it has a single value (e.g. all
1610 // the entries are the same or undef). If so, remove the PHI so it doesn't
1611 // block other optimizations.
1613 auto &DL = Caller->getParent()->getDataLayout();
1614 if (Value *V = SimplifyInstruction(PHI, DL, nullptr, nullptr,
1615 &IFI.ACT->getAssumptionCache(*Caller))) {
1616 PHI->replaceAllUsesWith(V);
1617 PHI->eraseFromParent();