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/EHPersonalities.h"
25 #include "llvm/Analysis/InstructionSimplify.h"
26 #include "llvm/Analysis/ValueTracking.h"
27 #include "llvm/IR/Attributes.h"
28 #include "llvm/IR/CallSite.h"
29 #include "llvm/IR/CFG.h"
30 #include "llvm/IR/Constants.h"
31 #include "llvm/IR/DataLayout.h"
32 #include "llvm/IR/DebugInfo.h"
33 #include "llvm/IR/DerivedTypes.h"
34 #include "llvm/IR/DIBuilder.h"
35 #include "llvm/IR/Dominators.h"
36 #include "llvm/IR/IRBuilder.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/IntrinsicInst.h"
39 #include "llvm/IR/Intrinsics.h"
40 #include "llvm/IR/MDBuilder.h"
41 #include "llvm/IR/Module.h"
42 #include "llvm/Transforms/Utils/Local.h"
43 #include "llvm/Support/CommandLine.h"
49 EnableNoAliasConversion("enable-noalias-to-md-conversion", cl::init(true),
51 cl::desc("Convert noalias attributes to metadata during inlining."));
54 PreserveAlignmentAssumptions("preserve-alignment-assumptions-during-inlining",
55 cl::init(true), cl::Hidden,
56 cl::desc("Convert align attributes to assumptions during inlining."));
58 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI,
59 AAResults *CalleeAAR, bool InsertLifetime) {
60 return InlineFunction(CallSite(CI), IFI, CalleeAAR, InsertLifetime);
62 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI,
63 AAResults *CalleeAAR, bool InsertLifetime) {
64 return InlineFunction(CallSite(II), IFI, CalleeAAR, InsertLifetime);
68 /// A class for recording information about inlining a landing pad.
69 class LandingPadInliningInfo {
70 BasicBlock *OuterResumeDest; ///< Destination of the invoke's unwind.
71 BasicBlock *InnerResumeDest; ///< Destination for the callee's resume.
72 LandingPadInst *CallerLPad; ///< LandingPadInst associated with the invoke.
73 PHINode *InnerEHValuesPHI; ///< PHI for EH values from landingpad insts.
74 SmallVector<Value*, 8> UnwindDestPHIValues;
77 LandingPadInliningInfo(InvokeInst *II)
78 : OuterResumeDest(II->getUnwindDest()), InnerResumeDest(nullptr),
79 CallerLPad(nullptr), InnerEHValuesPHI(nullptr) {
80 // If there are PHI nodes in the unwind destination block, we need to keep
81 // track of which values came into them from the invoke before removing
82 // the edge from this block.
83 llvm::BasicBlock *InvokeBB = II->getParent();
84 BasicBlock::iterator I = OuterResumeDest->begin();
85 for (; isa<PHINode>(I); ++I) {
86 // Save the value to use for this edge.
87 PHINode *PHI = cast<PHINode>(I);
88 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
91 CallerLPad = cast<LandingPadInst>(I);
94 /// The outer unwind destination is the target of
95 /// unwind edges introduced for calls within the inlined function.
96 BasicBlock *getOuterResumeDest() const {
97 return OuterResumeDest;
100 BasicBlock *getInnerResumeDest();
102 LandingPadInst *getLandingPadInst() const { return CallerLPad; }
104 /// Forward the 'resume' instruction to the caller's landing pad block.
105 /// When the landing pad block has only one predecessor, this is
106 /// a simple branch. When there is more than one predecessor, we need to
107 /// split the landing pad block after the landingpad instruction and jump
109 void forwardResume(ResumeInst *RI,
110 SmallPtrSetImpl<LandingPadInst*> &InlinedLPads);
112 /// Add incoming-PHI values to the unwind destination block for the given
113 /// basic block, using the values for the original invoke's source block.
114 void addIncomingPHIValuesFor(BasicBlock *BB) const {
115 addIncomingPHIValuesForInto(BB, OuterResumeDest);
118 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const {
119 BasicBlock::iterator I = dest->begin();
120 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
121 PHINode *phi = cast<PHINode>(I);
122 phi->addIncoming(UnwindDestPHIValues[i], src);
126 } // anonymous namespace
128 /// Get or create a target for the branch from ResumeInsts.
129 BasicBlock *LandingPadInliningInfo::getInnerResumeDest() {
130 if (InnerResumeDest) return InnerResumeDest;
132 // Split the landing pad.
133 BasicBlock::iterator SplitPoint = ++CallerLPad->getIterator();
135 OuterResumeDest->splitBasicBlock(SplitPoint,
136 OuterResumeDest->getName() + ".body");
138 // The number of incoming edges we expect to the inner landing pad.
139 const unsigned PHICapacity = 2;
141 // Create corresponding new PHIs for all the PHIs in the outer landing pad.
142 Instruction *InsertPoint = &InnerResumeDest->front();
143 BasicBlock::iterator I = OuterResumeDest->begin();
144 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
145 PHINode *OuterPHI = cast<PHINode>(I);
146 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity,
147 OuterPHI->getName() + ".lpad-body",
149 OuterPHI->replaceAllUsesWith(InnerPHI);
150 InnerPHI->addIncoming(OuterPHI, OuterResumeDest);
153 // Create a PHI for the exception values.
154 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity,
155 "eh.lpad-body", InsertPoint);
156 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI);
157 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest);
160 return InnerResumeDest;
163 /// Forward the 'resume' instruction to the caller's landing pad block.
164 /// When the landing pad block has only one predecessor, this is a simple
165 /// branch. When there is more than one predecessor, we need to split the
166 /// landing pad block after the landingpad instruction and jump to there.
167 void LandingPadInliningInfo::forwardResume(
168 ResumeInst *RI, SmallPtrSetImpl<LandingPadInst *> &InlinedLPads) {
169 BasicBlock *Dest = getInnerResumeDest();
170 BasicBlock *Src = RI->getParent();
172 BranchInst::Create(Dest, Src);
174 // Update the PHIs in the destination. They were inserted in an order which
176 addIncomingPHIValuesForInto(Src, Dest);
178 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src);
179 RI->eraseFromParent();
182 /// When we inline a basic block into an invoke,
183 /// we have to turn all of the calls that can throw into invokes.
184 /// This function analyze BB to see if there are any calls, and if so,
185 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
186 /// nodes in that block with the values specified in InvokeDestPHIValues.
188 HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB, BasicBlock *UnwindEdge) {
189 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
190 Instruction *I = &*BBI++;
192 // We only need to check for function calls: inlined invoke
193 // instructions require no special handling.
194 CallInst *CI = dyn_cast<CallInst>(I);
196 if (!CI || CI->doesNotThrow() || isa<InlineAsm>(CI->getCalledValue()))
199 // Convert this function call into an invoke instruction. First, split the
202 BB->splitBasicBlock(CI->getIterator(), CI->getName() + ".noexc");
204 // Delete the unconditional branch inserted by splitBasicBlock
205 BB->getInstList().pop_back();
207 // Create the new invoke instruction.
208 SmallVector<Value*, 8> InvokeArgs(CI->arg_begin(), CI->arg_end());
209 SmallVector<OperandBundleDef, 1> OpBundles;
211 CI->getOperandBundlesAsDefs(OpBundles);
213 // Note: we're round tripping operand bundles through memory here, and that
214 // can potentially be avoided with a cleverer API design that we do not have
218 InvokeInst::Create(CI->getCalledValue(), Split, UnwindEdge, InvokeArgs,
219 OpBundles, CI->getName(), BB);
220 II->setDebugLoc(CI->getDebugLoc());
221 II->setCallingConv(CI->getCallingConv());
222 II->setAttributes(CI->getAttributes());
224 // Make sure that anything using the call now uses the invoke! This also
225 // updates the CallGraph if present, because it uses a WeakVH.
226 CI->replaceAllUsesWith(II);
228 // Delete the original call
229 Split->getInstList().pop_front();
235 /// If we inlined an invoke site, we need to convert calls
236 /// in the body of the inlined function into invokes.
238 /// II is the invoke instruction being inlined. FirstNewBlock is the first
239 /// block of the inlined code (the last block is the end of the function),
240 /// and InlineCodeInfo is information about the code that got inlined.
241 static void HandleInlinedLandingPad(InvokeInst *II, BasicBlock *FirstNewBlock,
242 ClonedCodeInfo &InlinedCodeInfo) {
243 BasicBlock *InvokeDest = II->getUnwindDest();
245 Function *Caller = FirstNewBlock->getParent();
247 // The inlined code is currently at the end of the function, scan from the
248 // start of the inlined code to its end, checking for stuff we need to
250 LandingPadInliningInfo Invoke(II);
252 // Get all of the inlined landing pad instructions.
253 SmallPtrSet<LandingPadInst*, 16> InlinedLPads;
254 for (Function::iterator I = FirstNewBlock->getIterator(), E = Caller->end();
256 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator()))
257 InlinedLPads.insert(II->getLandingPadInst());
259 // Append the clauses from the outer landing pad instruction into the inlined
260 // landing pad instructions.
261 LandingPadInst *OuterLPad = Invoke.getLandingPadInst();
262 for (LandingPadInst *InlinedLPad : InlinedLPads) {
263 unsigned OuterNum = OuterLPad->getNumClauses();
264 InlinedLPad->reserveClauses(OuterNum);
265 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx)
266 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx));
267 if (OuterLPad->isCleanup())
268 InlinedLPad->setCleanup(true);
271 for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end();
273 if (InlinedCodeInfo.ContainsCalls)
274 if (BasicBlock *NewBB = HandleCallsInBlockInlinedThroughInvoke(
275 &*BB, Invoke.getOuterResumeDest()))
276 // Update any PHI nodes in the exceptional block to indicate that there
277 // is now a new entry in them.
278 Invoke.addIncomingPHIValuesFor(NewBB);
280 // Forward any resumes that are remaining here.
281 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator()))
282 Invoke.forwardResume(RI, InlinedLPads);
285 // Now that everything is happy, we have one final detail. The PHI nodes in
286 // the exception destination block still have entries due to the original
287 // invoke instruction. Eliminate these entries (which might even delete the
289 InvokeDest->removePredecessor(II->getParent());
292 /// If we inlined an invoke site, we need to convert calls
293 /// in the body of the inlined function into invokes.
295 /// II is the invoke instruction being inlined. FirstNewBlock is the first
296 /// block of the inlined code (the last block is the end of the function),
297 /// and InlineCodeInfo is information about the code that got inlined.
298 static void HandleInlinedEHPad(InvokeInst *II, BasicBlock *FirstNewBlock,
299 ClonedCodeInfo &InlinedCodeInfo) {
300 BasicBlock *UnwindDest = II->getUnwindDest();
301 Function *Caller = FirstNewBlock->getParent();
303 assert(UnwindDest->getFirstNonPHI()->isEHPad() && "unexpected BasicBlock!");
305 // If there are PHI nodes in the unwind destination block, we need to keep
306 // track of which values came into them from the invoke before removing the
307 // edge from this block.
308 SmallVector<Value *, 8> UnwindDestPHIValues;
309 llvm::BasicBlock *InvokeBB = II->getParent();
310 for (Instruction &I : *UnwindDest) {
311 // Save the value to use for this edge.
312 PHINode *PHI = dyn_cast<PHINode>(&I);
315 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
318 // Add incoming-PHI values to the unwind destination block for the given basic
319 // block, using the values for the original invoke's source block.
320 auto UpdatePHINodes = [&](BasicBlock *Src) {
321 BasicBlock::iterator I = UnwindDest->begin();
322 for (Value *V : UnwindDestPHIValues) {
323 PHINode *PHI = cast<PHINode>(I);
324 PHI->addIncoming(V, Src);
329 // This connects all the instructions which 'unwind to caller' to the invoke
331 for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end();
333 if (auto *CRI = dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
334 if (CRI->unwindsToCaller()) {
335 CleanupReturnInst::Create(CRI->getCleanupPad(), UnwindDest, CRI);
336 CRI->eraseFromParent();
337 UpdatePHINodes(&*BB);
341 Instruction *I = BB->getFirstNonPHI();
345 Instruction *Replacement = nullptr;
346 if (auto *TPI = dyn_cast<TerminatePadInst>(I)) {
347 if (TPI->unwindsToCaller()) {
348 SmallVector<Value *, 3> TerminatePadArgs;
349 for (Value *ArgOperand : TPI->arg_operands())
350 TerminatePadArgs.push_back(ArgOperand);
351 Replacement = TerminatePadInst::Create(TPI->getParentPad(), UnwindDest,
352 TerminatePadArgs, TPI);
354 } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I)) {
355 if (CatchSwitch->unwindsToCaller()) {
356 auto *NewCatchSwitch = CatchSwitchInst::Create(
357 CatchSwitch->getParentPad(), UnwindDest,
358 CatchSwitch->getNumHandlers(), CatchSwitch->getName(),
360 for (BasicBlock *PadBB : CatchSwitch->handlers())
361 NewCatchSwitch->addHandler(PadBB);
362 Replacement = NewCatchSwitch;
364 } else if (!isa<FuncletPadInst>(I)) {
365 llvm_unreachable("unexpected EHPad!");
369 Replacement->takeName(I);
370 I->replaceAllUsesWith(Replacement);
371 I->eraseFromParent();
372 UpdatePHINodes(&*BB);
376 if (InlinedCodeInfo.ContainsCalls)
377 for (Function::iterator BB = FirstNewBlock->getIterator(),
380 if (BasicBlock *NewBB =
381 HandleCallsInBlockInlinedThroughInvoke(&*BB, UnwindDest))
382 // Update any PHI nodes in the exceptional block to indicate that there
383 // is now a new entry in them.
384 UpdatePHINodes(NewBB);
386 // Now that everything is happy, we have one final detail. The PHI nodes in
387 // the exception destination block still have entries due to the original
388 // invoke instruction. Eliminate these entries (which might even delete the
390 UnwindDest->removePredecessor(InvokeBB);
393 /// When inlining a function that contains noalias scope metadata,
394 /// this metadata needs to be cloned so that the inlined blocks
395 /// have different "unqiue scopes" at every call site. Were this not done, then
396 /// aliasing scopes from a function inlined into a caller multiple times could
397 /// not be differentiated (and this would lead to miscompiles because the
398 /// non-aliasing property communicated by the metadata could have
399 /// call-site-specific control dependencies).
400 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) {
401 const Function *CalledFunc = CS.getCalledFunction();
402 SetVector<const MDNode *> MD;
404 // Note: We could only clone the metadata if it is already used in the
405 // caller. I'm omitting that check here because it might confuse
406 // inter-procedural alias analysis passes. We can revisit this if it becomes
407 // an efficiency or overhead problem.
409 for (Function::const_iterator I = CalledFunc->begin(), IE = CalledFunc->end();
411 for (BasicBlock::const_iterator J = I->begin(), JE = I->end(); J != JE; ++J) {
412 if (const MDNode *M = J->getMetadata(LLVMContext::MD_alias_scope))
414 if (const MDNode *M = J->getMetadata(LLVMContext::MD_noalias))
421 // Walk the existing metadata, adding the complete (perhaps cyclic) chain to
423 SmallVector<const Metadata *, 16> Queue(MD.begin(), MD.end());
424 while (!Queue.empty()) {
425 const MDNode *M = cast<MDNode>(Queue.pop_back_val());
426 for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i)
427 if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i)))
432 // Now we have a complete set of all metadata in the chains used to specify
433 // the noalias scopes and the lists of those scopes.
434 SmallVector<TempMDTuple, 16> DummyNodes;
435 DenseMap<const MDNode *, TrackingMDNodeRef> MDMap;
436 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
438 DummyNodes.push_back(MDTuple::getTemporary(CalledFunc->getContext(), None));
439 MDMap[*I].reset(DummyNodes.back().get());
442 // Create new metadata nodes to replace the dummy nodes, replacing old
443 // metadata references with either a dummy node or an already-created new
445 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
447 SmallVector<Metadata *, 4> NewOps;
448 for (unsigned i = 0, ie = (*I)->getNumOperands(); i != ie; ++i) {
449 const Metadata *V = (*I)->getOperand(i);
450 if (const MDNode *M = dyn_cast<MDNode>(V))
451 NewOps.push_back(MDMap[M]);
453 NewOps.push_back(const_cast<Metadata *>(V));
456 MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps);
457 MDTuple *TempM = cast<MDTuple>(MDMap[*I]);
458 assert(TempM->isTemporary() && "Expected temporary node");
460 TempM->replaceAllUsesWith(NewM);
463 // Now replace the metadata in the new inlined instructions with the
464 // repacements from the map.
465 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
466 VMI != VMIE; ++VMI) {
470 Instruction *NI = dyn_cast<Instruction>(VMI->second);
474 if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope)) {
475 MDNode *NewMD = MDMap[M];
476 // If the call site also had alias scope metadata (a list of scopes to
477 // which instructions inside it might belong), propagate those scopes to
478 // the inlined instructions.
480 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
481 NewMD = MDNode::concatenate(NewMD, CSM);
482 NI->setMetadata(LLVMContext::MD_alias_scope, NewMD);
483 } else if (NI->mayReadOrWriteMemory()) {
485 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
486 NI->setMetadata(LLVMContext::MD_alias_scope, M);
489 if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias)) {
490 MDNode *NewMD = MDMap[M];
491 // If the call site also had noalias metadata (a list of scopes with
492 // which instructions inside it don't alias), propagate those scopes to
493 // the inlined instructions.
495 CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
496 NewMD = MDNode::concatenate(NewMD, CSM);
497 NI->setMetadata(LLVMContext::MD_noalias, NewMD);
498 } else if (NI->mayReadOrWriteMemory()) {
499 if (MDNode *M = CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
500 NI->setMetadata(LLVMContext::MD_noalias, M);
505 /// If the inlined function has noalias arguments,
506 /// then add new alias scopes for each noalias argument, tag the mapped noalias
507 /// parameters with noalias metadata specifying the new scope, and tag all
508 /// non-derived loads, stores and memory intrinsics with the new alias scopes.
509 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
510 const DataLayout &DL, AAResults *CalleeAAR) {
511 if (!EnableNoAliasConversion)
514 const Function *CalledFunc = CS.getCalledFunction();
515 SmallVector<const Argument *, 4> NoAliasArgs;
517 for (const Argument &I : CalledFunc->args()) {
518 if (I.hasNoAliasAttr() && !I.hasNUses(0))
519 NoAliasArgs.push_back(&I);
522 if (NoAliasArgs.empty())
525 // To do a good job, if a noalias variable is captured, we need to know if
526 // the capture point dominates the particular use we're considering.
528 DT.recalculate(const_cast<Function&>(*CalledFunc));
530 // noalias indicates that pointer values based on the argument do not alias
531 // pointer values which are not based on it. So we add a new "scope" for each
532 // noalias function argument. Accesses using pointers based on that argument
533 // become part of that alias scope, accesses using pointers not based on that
534 // argument are tagged as noalias with that scope.
536 DenseMap<const Argument *, MDNode *> NewScopes;
537 MDBuilder MDB(CalledFunc->getContext());
539 // Create a new scope domain for this function.
541 MDB.createAnonymousAliasScopeDomain(CalledFunc->getName());
542 for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) {
543 const Argument *A = NoAliasArgs[i];
545 std::string Name = CalledFunc->getName();
548 Name += A->getName();
550 Name += ": argument ";
554 // Note: We always create a new anonymous root here. This is true regardless
555 // of the linkage of the callee because the aliasing "scope" is not just a
556 // property of the callee, but also all control dependencies in the caller.
557 MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
558 NewScopes.insert(std::make_pair(A, NewScope));
561 // Iterate over all new instructions in the map; for all memory-access
562 // instructions, add the alias scope metadata.
563 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
564 VMI != VMIE; ++VMI) {
565 if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) {
569 Instruction *NI = dyn_cast<Instruction>(VMI->second);
573 bool IsArgMemOnlyCall = false, IsFuncCall = false;
574 SmallVector<const Value *, 2> PtrArgs;
576 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
577 PtrArgs.push_back(LI->getPointerOperand());
578 else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
579 PtrArgs.push_back(SI->getPointerOperand());
580 else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
581 PtrArgs.push_back(VAAI->getPointerOperand());
582 else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
583 PtrArgs.push_back(CXI->getPointerOperand());
584 else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
585 PtrArgs.push_back(RMWI->getPointerOperand());
586 else if (ImmutableCallSite ICS = ImmutableCallSite(I)) {
587 // If we know that the call does not access memory, then we'll still
588 // know that about the inlined clone of this call site, and we don't
589 // need to add metadata.
590 if (ICS.doesNotAccessMemory())
595 FunctionModRefBehavior MRB = CalleeAAR->getModRefBehavior(ICS);
596 if (MRB == FMRB_OnlyAccessesArgumentPointees ||
597 MRB == FMRB_OnlyReadsArgumentPointees)
598 IsArgMemOnlyCall = true;
601 for (ImmutableCallSite::arg_iterator AI = ICS.arg_begin(),
602 AE = ICS.arg_end(); AI != AE; ++AI) {
603 // We need to check the underlying objects of all arguments, not just
604 // the pointer arguments, because we might be passing pointers as
606 // However, if we know that the call only accesses pointer arguments,
607 // then we only need to check the pointer arguments.
608 if (IsArgMemOnlyCall && !(*AI)->getType()->isPointerTy())
611 PtrArgs.push_back(*AI);
615 // If we found no pointers, then this instruction is not suitable for
616 // pairing with an instruction to receive aliasing metadata.
617 // However, if this is a call, this we might just alias with none of the
618 // noalias arguments.
619 if (PtrArgs.empty() && !IsFuncCall)
622 // It is possible that there is only one underlying object, but you
623 // need to go through several PHIs to see it, and thus could be
624 // repeated in the Objects list.
625 SmallPtrSet<const Value *, 4> ObjSet;
626 SmallVector<Metadata *, 4> Scopes, NoAliases;
628 SmallSetVector<const Argument *, 4> NAPtrArgs;
629 for (unsigned i = 0, ie = PtrArgs.size(); i != ie; ++i) {
630 SmallVector<Value *, 4> Objects;
631 GetUnderlyingObjects(const_cast<Value*>(PtrArgs[i]),
632 Objects, DL, /* LI = */ nullptr);
634 for (Value *O : Objects)
638 // Figure out if we're derived from anything that is not a noalias
640 bool CanDeriveViaCapture = false, UsesAliasingPtr = false;
641 for (const Value *V : ObjSet) {
642 // Is this value a constant that cannot be derived from any pointer
643 // value (we need to exclude constant expressions, for example, that
644 // are formed from arithmetic on global symbols).
645 bool IsNonPtrConst = isa<ConstantInt>(V) || isa<ConstantFP>(V) ||
646 isa<ConstantPointerNull>(V) ||
647 isa<ConstantDataVector>(V) || isa<UndefValue>(V);
651 // If this is anything other than a noalias argument, then we cannot
652 // completely describe the aliasing properties using alias.scope
653 // metadata (and, thus, won't add any).
654 if (const Argument *A = dyn_cast<Argument>(V)) {
655 if (!A->hasNoAliasAttr())
656 UsesAliasingPtr = true;
658 UsesAliasingPtr = true;
661 // If this is not some identified function-local object (which cannot
662 // directly alias a noalias argument), or some other argument (which,
663 // by definition, also cannot alias a noalias argument), then we could
664 // alias a noalias argument that has been captured).
665 if (!isa<Argument>(V) &&
666 !isIdentifiedFunctionLocal(const_cast<Value*>(V)))
667 CanDeriveViaCapture = true;
670 // A function call can always get captured noalias pointers (via other
671 // parameters, globals, etc.).
672 if (IsFuncCall && !IsArgMemOnlyCall)
673 CanDeriveViaCapture = true;
675 // First, we want to figure out all of the sets with which we definitely
676 // don't alias. Iterate over all noalias set, and add those for which:
677 // 1. The noalias argument is not in the set of objects from which we
678 // definitely derive.
679 // 2. The noalias argument has not yet been captured.
680 // An arbitrary function that might load pointers could see captured
681 // noalias arguments via other noalias arguments or globals, and so we
682 // must always check for prior capture.
683 for (const Argument *A : NoAliasArgs) {
684 if (!ObjSet.count(A) && (!CanDeriveViaCapture ||
685 // It might be tempting to skip the
686 // PointerMayBeCapturedBefore check if
687 // A->hasNoCaptureAttr() is true, but this is
688 // incorrect because nocapture only guarantees
689 // that no copies outlive the function, not
690 // that the value cannot be locally captured.
691 !PointerMayBeCapturedBefore(A,
692 /* ReturnCaptures */ false,
693 /* StoreCaptures */ false, I, &DT)))
694 NoAliases.push_back(NewScopes[A]);
697 if (!NoAliases.empty())
698 NI->setMetadata(LLVMContext::MD_noalias,
700 NI->getMetadata(LLVMContext::MD_noalias),
701 MDNode::get(CalledFunc->getContext(), NoAliases)));
703 // Next, we want to figure out all of the sets to which we might belong.
704 // We might belong to a set if the noalias argument is in the set of
705 // underlying objects. If there is some non-noalias argument in our list
706 // of underlying objects, then we cannot add a scope because the fact
707 // that some access does not alias with any set of our noalias arguments
708 // cannot itself guarantee that it does not alias with this access
709 // (because there is some pointer of unknown origin involved and the
710 // other access might also depend on this pointer). We also cannot add
711 // scopes to arbitrary functions unless we know they don't access any
712 // non-parameter pointer-values.
713 bool CanAddScopes = !UsesAliasingPtr;
714 if (CanAddScopes && IsFuncCall)
715 CanAddScopes = IsArgMemOnlyCall;
718 for (const Argument *A : NoAliasArgs) {
720 Scopes.push_back(NewScopes[A]);
725 LLVMContext::MD_alias_scope,
726 MDNode::concatenate(NI->getMetadata(LLVMContext::MD_alias_scope),
727 MDNode::get(CalledFunc->getContext(), Scopes)));
732 /// If the inlined function has non-byval align arguments, then
733 /// add @llvm.assume-based alignment assumptions to preserve this information.
734 static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) {
735 if (!PreserveAlignmentAssumptions)
737 auto &DL = CS.getCaller()->getParent()->getDataLayout();
739 // To avoid inserting redundant assumptions, we should check for assumptions
740 // already in the caller. To do this, we might need a DT of the caller.
742 bool DTCalculated = false;
744 Function *CalledFunc = CS.getCalledFunction();
745 for (Function::arg_iterator I = CalledFunc->arg_begin(),
746 E = CalledFunc->arg_end();
748 unsigned Align = I->getType()->isPointerTy() ? I->getParamAlignment() : 0;
749 if (Align && !I->hasByValOrInAllocaAttr() && !I->hasNUses(0)) {
751 DT.recalculate(const_cast<Function&>(*CS.getInstruction()->getParent()
756 // If we can already prove the asserted alignment in the context of the
757 // caller, then don't bother inserting the assumption.
758 Value *Arg = CS.getArgument(I->getArgNo());
759 if (getKnownAlignment(Arg, DL, CS.getInstruction(),
760 &IFI.ACT->getAssumptionCache(*CS.getCaller()),
764 IRBuilder<>(CS.getInstruction())
765 .CreateAlignmentAssumption(DL, Arg, Align);
770 /// Once we have cloned code over from a callee into the caller,
771 /// update the specified callgraph to reflect the changes we made.
772 /// Note that it's possible that not all code was copied over, so only
773 /// some edges of the callgraph may remain.
774 static void UpdateCallGraphAfterInlining(CallSite CS,
775 Function::iterator FirstNewBlock,
776 ValueToValueMapTy &VMap,
777 InlineFunctionInfo &IFI) {
778 CallGraph &CG = *IFI.CG;
779 const Function *Caller = CS.getInstruction()->getParent()->getParent();
780 const Function *Callee = CS.getCalledFunction();
781 CallGraphNode *CalleeNode = CG[Callee];
782 CallGraphNode *CallerNode = CG[Caller];
784 // Since we inlined some uninlined call sites in the callee into the caller,
785 // add edges from the caller to all of the callees of the callee.
786 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
788 // Consider the case where CalleeNode == CallerNode.
789 CallGraphNode::CalledFunctionsVector CallCache;
790 if (CalleeNode == CallerNode) {
791 CallCache.assign(I, E);
792 I = CallCache.begin();
796 for (; I != E; ++I) {
797 const Value *OrigCall = I->first;
799 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
800 // Only copy the edge if the call was inlined!
801 if (VMI == VMap.end() || VMI->second == nullptr)
804 // If the call was inlined, but then constant folded, there is no edge to
805 // add. Check for this case.
806 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
810 // We do not treat intrinsic calls like real function calls because we
811 // expect them to become inline code; do not add an edge for an intrinsic.
812 CallSite CS = CallSite(NewCall);
813 if (CS && CS.getCalledFunction() && CS.getCalledFunction()->isIntrinsic())
816 // Remember that this call site got inlined for the client of
818 IFI.InlinedCalls.push_back(NewCall);
820 // It's possible that inlining the callsite will cause it to go from an
821 // indirect to a direct call by resolving a function pointer. If this
822 // happens, set the callee of the new call site to a more precise
823 // destination. This can also happen if the call graph node of the caller
824 // was just unnecessarily imprecise.
825 if (!I->second->getFunction())
826 if (Function *F = CallSite(NewCall).getCalledFunction()) {
827 // Indirect call site resolved to direct call.
828 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
833 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
836 // Update the call graph by deleting the edge from Callee to Caller. We must
837 // do this after the loop above in case Caller and Callee are the same.
838 CallerNode->removeCallEdgeFor(CS);
841 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
842 BasicBlock *InsertBlock,
843 InlineFunctionInfo &IFI) {
844 Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
845 IRBuilder<> Builder(InsertBlock, InsertBlock->begin());
847 Value *Size = Builder.getInt64(M->getDataLayout().getTypeStoreSize(AggTy));
849 // Always generate a memcpy of alignment 1 here because we don't know
850 // the alignment of the src pointer. Other optimizations can infer
852 Builder.CreateMemCpy(Dst, Src, Size, /*Align=*/1);
855 /// When inlining a call site that has a byval argument,
856 /// we have to make the implicit memcpy explicit by adding it.
857 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
858 const Function *CalledFunc,
859 InlineFunctionInfo &IFI,
860 unsigned ByValAlignment) {
861 PointerType *ArgTy = cast<PointerType>(Arg->getType());
862 Type *AggTy = ArgTy->getElementType();
864 Function *Caller = TheCall->getParent()->getParent();
866 // If the called function is readonly, then it could not mutate the caller's
867 // copy of the byval'd memory. In this case, it is safe to elide the copy and
869 if (CalledFunc->onlyReadsMemory()) {
870 // If the byval argument has a specified alignment that is greater than the
871 // passed in pointer, then we either have to round up the input pointer or
872 // give up on this transformation.
873 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
876 const DataLayout &DL = Caller->getParent()->getDataLayout();
878 // If the pointer is already known to be sufficiently aligned, or if we can
879 // round it up to a larger alignment, then we don't need a temporary.
880 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, DL, TheCall,
881 &IFI.ACT->getAssumptionCache(*Caller)) >=
885 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
886 // for code quality, but rarely happens and is required for correctness.
889 // Create the alloca. If we have DataLayout, use nice alignment.
891 Caller->getParent()->getDataLayout().getPrefTypeAlignment(AggTy);
893 // If the byval had an alignment specified, we *must* use at least that
894 // alignment, as it is required by the byval argument (and uses of the
895 // pointer inside the callee).
896 Align = std::max(Align, ByValAlignment);
898 Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(),
899 &*Caller->begin()->begin());
900 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
902 // Uses of the argument in the function should use our new alloca
907 // Check whether this Value is used by a lifetime intrinsic.
908 static bool isUsedByLifetimeMarker(Value *V) {
909 for (User *U : V->users()) {
910 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
911 switch (II->getIntrinsicID()) {
913 case Intrinsic::lifetime_start:
914 case Intrinsic::lifetime_end:
922 // Check whether the given alloca already has
923 // lifetime.start or lifetime.end intrinsics.
924 static bool hasLifetimeMarkers(AllocaInst *AI) {
925 Type *Ty = AI->getType();
926 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
927 Ty->getPointerAddressSpace());
929 return isUsedByLifetimeMarker(AI);
931 // Do a scan to find all the casts to i8*.
932 for (User *U : AI->users()) {
933 if (U->getType() != Int8PtrTy) continue;
934 if (U->stripPointerCasts() != AI) continue;
935 if (isUsedByLifetimeMarker(U))
941 /// Rebuild the entire inlined-at chain for this instruction so that the top of
942 /// the chain now is inlined-at the new call site.
944 updateInlinedAtInfo(DebugLoc DL, DILocation *InlinedAtNode, LLVMContext &Ctx,
945 DenseMap<const DILocation *, DILocation *> &IANodes) {
946 SmallVector<DILocation *, 3> InlinedAtLocations;
947 DILocation *Last = InlinedAtNode;
948 DILocation *CurInlinedAt = DL;
950 // Gather all the inlined-at nodes
951 while (DILocation *IA = CurInlinedAt->getInlinedAt()) {
952 // Skip any we've already built nodes for
953 if (DILocation *Found = IANodes[IA]) {
958 InlinedAtLocations.push_back(IA);
962 // Starting from the top, rebuild the nodes to point to the new inlined-at
963 // location (then rebuilding the rest of the chain behind it) and update the
964 // map of already-constructed inlined-at nodes.
965 for (const DILocation *MD : make_range(InlinedAtLocations.rbegin(),
966 InlinedAtLocations.rend())) {
967 Last = IANodes[MD] = DILocation::getDistinct(
968 Ctx, MD->getLine(), MD->getColumn(), MD->getScope(), Last);
971 // And finally create the normal location for this instruction, referring to
972 // the new inlined-at chain.
973 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(), Last);
976 /// Update inlined instructions' line numbers to
977 /// to encode location where these instructions are inlined.
978 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
979 Instruction *TheCall) {
980 DebugLoc TheCallDL = TheCall->getDebugLoc();
984 auto &Ctx = Fn->getContext();
985 DILocation *InlinedAtNode = TheCallDL;
987 // Create a unique call site, not to be confused with any other call from the
989 InlinedAtNode = DILocation::getDistinct(
990 Ctx, InlinedAtNode->getLine(), InlinedAtNode->getColumn(),
991 InlinedAtNode->getScope(), InlinedAtNode->getInlinedAt());
993 // Cache the inlined-at nodes as they're built so they are reused, without
994 // this every instruction's inlined-at chain would become distinct from each
996 DenseMap<const DILocation *, DILocation *> IANodes;
998 for (; FI != Fn->end(); ++FI) {
999 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
1001 DebugLoc DL = BI->getDebugLoc();
1003 // If the inlined instruction has no line number, make it look as if it
1004 // originates from the call location. This is important for
1005 // ((__always_inline__, __nodebug__)) functions which must use caller
1006 // location for all instructions in their function body.
1008 // Don't update static allocas, as they may get moved later.
1009 if (auto *AI = dyn_cast<AllocaInst>(BI))
1010 if (isa<Constant>(AI->getArraySize()))
1013 BI->setDebugLoc(TheCallDL);
1015 BI->setDebugLoc(updateInlinedAtInfo(DL, InlinedAtNode, BI->getContext(), IANodes));
1021 /// This function inlines the called function into the basic block of the
1022 /// caller. This returns false if it is not possible to inline this call.
1023 /// The program is still in a well defined state if this occurs though.
1025 /// Note that this only does one level of inlining. For example, if the
1026 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
1027 /// exists in the instruction stream. Similarly this will inline a recursive
1028 /// function by one level.
1029 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
1030 AAResults *CalleeAAR, bool InsertLifetime) {
1031 Instruction *TheCall = CS.getInstruction();
1032 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
1033 "Instruction not in function!");
1035 // If IFI has any state in it, zap it before we fill it in.
1038 const Function *CalledFunc = CS.getCalledFunction();
1039 if (!CalledFunc || // Can't inline external function or indirect
1040 CalledFunc->isDeclaration() || // call, or call to a vararg function!
1041 CalledFunc->getFunctionType()->isVarArg()) return false;
1043 // The inliner does not know how to inline through calls with operand bundles
1045 if (CS.hasOperandBundles()) {
1046 // ... but it knows how to inline through "deopt" operand bundles.
1048 CS.getNumOperandBundles() == 1 &&
1049 CS.getOperandBundleAt(0).getTagID() == LLVMContext::OB_deopt;
1054 // If the call to the callee cannot throw, set the 'nounwind' flag on any
1055 // calls that we inline.
1056 bool MarkNoUnwind = CS.doesNotThrow();
1058 BasicBlock *OrigBB = TheCall->getParent();
1059 Function *Caller = OrigBB->getParent();
1061 // GC poses two hazards to inlining, which only occur when the callee has GC:
1062 // 1. If the caller has no GC, then the callee's GC must be propagated to the
1064 // 2. If the caller has a differing GC, it is invalid to inline.
1065 if (CalledFunc->hasGC()) {
1066 if (!Caller->hasGC())
1067 Caller->setGC(CalledFunc->getGC());
1068 else if (CalledFunc->getGC() != Caller->getGC())
1072 // Get the personality function from the callee if it contains a landing pad.
1073 Constant *CalledPersonality =
1074 CalledFunc->hasPersonalityFn()
1075 ? CalledFunc->getPersonalityFn()->stripPointerCasts()
1078 // Find the personality function used by the landing pads of the caller. If it
1079 // exists, then check to see that it matches the personality function used in
1081 Constant *CallerPersonality =
1082 Caller->hasPersonalityFn()
1083 ? Caller->getPersonalityFn()->stripPointerCasts()
1085 if (CalledPersonality) {
1086 if (!CallerPersonality)
1087 Caller->setPersonalityFn(CalledPersonality);
1088 // If the personality functions match, then we can perform the
1089 // inlining. Otherwise, we can't inline.
1090 // TODO: This isn't 100% true. Some personality functions are proper
1091 // supersets of others and can be used in place of the other.
1092 else if (CalledPersonality != CallerPersonality)
1096 // We need to figure out which funclet the callsite was in so that we may
1097 // properly nest the callee.
1098 Instruction *CallSiteEHPad = nullptr;
1099 if (CalledPersonality && CallerPersonality) {
1100 EHPersonality Personality = classifyEHPersonality(CalledPersonality);
1101 if (isFuncletEHPersonality(Personality)) {
1102 DenseMap<BasicBlock *, ColorVector> CallerBlockColors =
1103 colorEHFunclets(*Caller);
1104 ColorVector &CallSiteColors = CallerBlockColors[OrigBB];
1105 size_t NumColors = CallSiteColors.size();
1106 // There is no single parent, inlining will not succeed.
1109 if (NumColors == 1) {
1110 BasicBlock *CallSiteFuncletBB = CallSiteColors.front();
1111 if (CallSiteFuncletBB != Caller->begin()) {
1112 CallSiteEHPad = CallSiteFuncletBB->getFirstNonPHI();
1113 assert(CallSiteEHPad->isEHPad() && "Expected an EHPad!");
1117 // OK, the inlining site is legal. What about the target function?
1119 if (CallSiteEHPad) {
1120 if (Personality == EHPersonality::MSVC_CXX) {
1121 // The MSVC personality cannot tolerate catches getting inlined into
1122 // cleanup funclets.
1123 if (isa<CleanupPadInst>(CallSiteEHPad)) {
1124 // Ok, the call site is within a cleanuppad. Let's check the callee
1126 for (const BasicBlock &CalledBB : *CalledFunc) {
1127 if (isa<CatchPadInst>(CalledBB.getFirstNonPHI()))
1131 } else if (isAsynchronousEHPersonality(Personality)) {
1132 // SEH is even less tolerant, there may not be any sort of exceptional
1133 // funclet in the callee.
1134 for (const BasicBlock &CalledBB : *CalledFunc) {
1135 if (CalledBB.isEHPad())
1143 // Get an iterator to the last basic block in the function, which will have
1144 // the new function inlined after it.
1145 Function::iterator LastBlock = --Caller->end();
1147 // Make sure to capture all of the return instructions from the cloned
1149 SmallVector<ReturnInst*, 8> Returns;
1150 ClonedCodeInfo InlinedFunctionInfo;
1151 Function::iterator FirstNewBlock;
1153 { // Scope to destroy VMap after cloning.
1154 ValueToValueMapTy VMap;
1155 // Keep a list of pair (dst, src) to emit byval initializations.
1156 SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
1158 auto &DL = Caller->getParent()->getDataLayout();
1160 assert(CalledFunc->arg_size() == CS.arg_size() &&
1161 "No varargs calls can be inlined!");
1163 // Calculate the vector of arguments to pass into the function cloner, which
1164 // matches up the formal to the actual argument values.
1165 CallSite::arg_iterator AI = CS.arg_begin();
1167 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
1168 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
1169 Value *ActualArg = *AI;
1171 // When byval arguments actually inlined, we need to make the copy implied
1172 // by them explicit. However, we don't do this if the callee is readonly
1173 // or readnone, because the copy would be unneeded: the callee doesn't
1174 // modify the struct.
1175 if (CS.isByValArgument(ArgNo)) {
1176 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
1177 CalledFunc->getParamAlignment(ArgNo+1));
1178 if (ActualArg != *AI)
1179 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
1182 VMap[&*I] = ActualArg;
1185 // Add alignment assumptions if necessary. We do this before the inlined
1186 // instructions are actually cloned into the caller so that we can easily
1187 // check what will be known at the start of the inlined code.
1188 AddAlignmentAssumptions(CS, IFI);
1190 // We want the inliner to prune the code as it copies. We would LOVE to
1191 // have no dead or constant instructions leftover after inlining occurs
1192 // (which can happen, e.g., because an argument was constant), but we'll be
1193 // happy with whatever the cloner can do.
1194 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
1195 /*ModuleLevelChanges=*/false, Returns, ".i",
1196 &InlinedFunctionInfo, TheCall);
1198 // Remember the first block that is newly cloned over.
1199 FirstNewBlock = LastBlock; ++FirstNewBlock;
1201 // Inject byval arguments initialization.
1202 for (std::pair<Value*, Value*> &Init : ByValInit)
1203 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
1204 &*FirstNewBlock, IFI);
1206 if (CS.hasOperandBundles()) {
1207 auto ParentDeopt = CS.getOperandBundleAt(0);
1208 assert(ParentDeopt.getTagID() == LLVMContext::OB_deopt &&
1209 "Checked on entry!");
1211 SmallVector<OperandBundleDef, 2> OpDefs;
1213 for (auto &VH : InlinedFunctionInfo.OperandBundleCallSites) {
1214 if (!VH) continue; // instruction was DCE'd after being cloned
1216 Instruction *I = cast<Instruction>(VH);
1221 OpDefs.reserve(ICS.getNumOperandBundles());
1223 for (unsigned i = 0, e = ICS.getNumOperandBundles(); i < e; ++i) {
1224 auto ChildOB = ICS.getOperandBundleAt(i);
1225 if (ChildOB.getTagID() != LLVMContext::OB_deopt) {
1226 // If the inlined call has other operand bundles, let them be
1227 OpDefs.emplace_back(ChildOB);
1231 // It may be useful to separate this logic (of handling operand
1232 // bundles) out to a separate "policy" component if this gets crowded.
1233 // Prepend the parent's deoptimization continuation to the newly
1234 // inlined call's deoptimization continuation.
1235 std::vector<Value *> MergedDeoptArgs;
1236 MergedDeoptArgs.reserve(ParentDeopt.Inputs.size() +
1237 ChildOB.Inputs.size());
1239 MergedDeoptArgs.insert(MergedDeoptArgs.end(),
1240 ParentDeopt.Inputs.begin(),
1241 ParentDeopt.Inputs.end());
1242 MergedDeoptArgs.insert(MergedDeoptArgs.end(), ChildOB.Inputs.begin(),
1243 ChildOB.Inputs.end());
1245 OpDefs.emplace_back("deopt", std::move(MergedDeoptArgs));
1248 Instruction *NewI = nullptr;
1249 if (isa<CallInst>(I))
1250 NewI = CallInst::Create(cast<CallInst>(I), OpDefs, I);
1252 NewI = InvokeInst::Create(cast<InvokeInst>(I), OpDefs, I);
1254 // Note: the RAUW does the appropriate fixup in VMap, so we need to do
1255 // this even if the call returns void.
1256 I->replaceAllUsesWith(NewI);
1259 I->eraseFromParent();
1263 // Update the callgraph if requested.
1265 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
1267 // Update inlined instructions' line number information.
1268 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
1270 // Clone existing noalias metadata if necessary.
1271 CloneAliasScopeMetadata(CS, VMap);
1273 // Add noalias metadata if necessary.
1274 AddAliasScopeMetadata(CS, VMap, DL, CalleeAAR);
1276 // FIXME: We could register any cloned assumptions instead of clearing the
1277 // whole function's cache.
1279 IFI.ACT->getAssumptionCache(*Caller).clear();
1282 // If there are any alloca instructions in the block that used to be the entry
1283 // block for the callee, move them to the entry block of the caller. First
1284 // calculate which instruction they should be inserted before. We insert the
1285 // instructions at the end of the current alloca list.
1287 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
1288 for (BasicBlock::iterator I = FirstNewBlock->begin(),
1289 E = FirstNewBlock->end(); I != E; ) {
1290 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
1293 // If the alloca is now dead, remove it. This often occurs due to code
1295 if (AI->use_empty()) {
1296 AI->eraseFromParent();
1300 if (!isa<Constant>(AI->getArraySize()))
1303 // Keep track of the static allocas that we inline into the caller.
1304 IFI.StaticAllocas.push_back(AI);
1306 // Scan for the block of allocas that we can move over, and move them
1308 while (isa<AllocaInst>(I) &&
1309 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
1310 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
1314 // Transfer all of the allocas over in a block. Using splice means
1315 // that the instructions aren't removed from the symbol table, then
1317 Caller->getEntryBlock().getInstList().splice(
1318 InsertPoint, FirstNewBlock->getInstList(), AI->getIterator(), I);
1320 // Move any dbg.declares describing the allocas into the entry basic block.
1321 DIBuilder DIB(*Caller->getParent());
1322 for (auto &AI : IFI.StaticAllocas)
1323 replaceDbgDeclareForAlloca(AI, AI, DIB, /*Deref=*/false);
1326 bool InlinedMustTailCalls = false;
1327 if (InlinedFunctionInfo.ContainsCalls) {
1328 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
1329 if (CallInst *CI = dyn_cast<CallInst>(TheCall))
1330 CallSiteTailKind = CI->getTailCallKind();
1332 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
1334 for (Instruction &I : *BB) {
1335 CallInst *CI = dyn_cast<CallInst>(&I);
1339 // We need to reduce the strength of any inlined tail calls. For
1340 // musttail, we have to avoid introducing potential unbounded stack
1341 // growth. For example, if functions 'f' and 'g' are mutually recursive
1342 // with musttail, we can inline 'g' into 'f' so long as we preserve
1343 // musttail on the cloned call to 'f'. If either the inlined call site
1344 // or the cloned call site is *not* musttail, the program already has
1345 // one frame of stack growth, so it's safe to remove musttail. Here is
1346 // a table of example transformations:
1348 // f -> musttail g -> musttail f ==> f -> musttail f
1349 // f -> musttail g -> tail f ==> f -> tail f
1350 // f -> g -> musttail f ==> f -> f
1351 // f -> g -> tail f ==> f -> f
1352 CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
1353 ChildTCK = std::min(CallSiteTailKind, ChildTCK);
1354 CI->setTailCallKind(ChildTCK);
1355 InlinedMustTailCalls |= CI->isMustTailCall();
1357 // Calls inlined through a 'nounwind' call site should be marked
1360 CI->setDoesNotThrow();
1365 // Leave lifetime markers for the static alloca's, scoping them to the
1366 // function we just inlined.
1367 if (InsertLifetime && !IFI.StaticAllocas.empty()) {
1368 IRBuilder<> builder(&FirstNewBlock->front());
1369 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
1370 AllocaInst *AI = IFI.StaticAllocas[ai];
1372 // If the alloca is already scoped to something smaller than the whole
1373 // function then there's no need to add redundant, less accurate markers.
1374 if (hasLifetimeMarkers(AI))
1377 // Try to determine the size of the allocation.
1378 ConstantInt *AllocaSize = nullptr;
1379 if (ConstantInt *AIArraySize =
1380 dyn_cast<ConstantInt>(AI->getArraySize())) {
1381 auto &DL = Caller->getParent()->getDataLayout();
1382 Type *AllocaType = AI->getAllocatedType();
1383 uint64_t AllocaTypeSize = DL.getTypeAllocSize(AllocaType);
1384 uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
1386 // Don't add markers for zero-sized allocas.
1387 if (AllocaArraySize == 0)
1390 // Check that array size doesn't saturate uint64_t and doesn't
1391 // overflow when it's multiplied by type size.
1392 if (AllocaArraySize != ~0ULL &&
1393 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
1394 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
1395 AllocaArraySize * AllocaTypeSize);
1399 builder.CreateLifetimeStart(AI, AllocaSize);
1400 for (ReturnInst *RI : Returns) {
1401 // Don't insert llvm.lifetime.end calls between a musttail call and a
1402 // return. The return kills all local allocas.
1403 if (InlinedMustTailCalls &&
1404 RI->getParent()->getTerminatingMustTailCall())
1406 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
1411 // If the inlined code contained dynamic alloca instructions, wrap the inlined
1412 // code with llvm.stacksave/llvm.stackrestore intrinsics.
1413 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
1414 Module *M = Caller->getParent();
1415 // Get the two intrinsics we care about.
1416 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
1417 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
1419 // Insert the llvm.stacksave.
1420 CallInst *SavedPtr = IRBuilder<>(&*FirstNewBlock, FirstNewBlock->begin())
1421 .CreateCall(StackSave, {}, "savedstack");
1423 // Insert a call to llvm.stackrestore before any return instructions in the
1424 // inlined function.
1425 for (ReturnInst *RI : Returns) {
1426 // Don't insert llvm.stackrestore calls between a musttail call and a
1427 // return. The return will restore the stack pointer.
1428 if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
1430 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
1434 // Update the lexical scopes of the new funclets. Anything that had 'none' as
1435 // its parent is now nested inside the callsite's EHPad.
1436 if (CallSiteEHPad) {
1437 for (Function::iterator BB = FirstNewBlock->getIterator(),
1440 Instruction *I = BB->getFirstNonPHI();
1444 if (auto *TPI = dyn_cast<TerminatePadInst>(I)) {
1445 if (isa<ConstantTokenNone>(TPI->getParentPad()))
1446 TPI->setParentPad(CallSiteEHPad);
1447 } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I)) {
1448 if (isa<ConstantTokenNone>(CatchSwitch->getParentPad()))
1449 CatchSwitch->setParentPad(CallSiteEHPad);
1451 auto *FPI = cast<FuncletPadInst>(I);
1452 if (isa<ConstantTokenNone>(FPI->getParentPad()))
1453 FPI->setParentPad(CallSiteEHPad);
1458 // If we are inlining for an invoke instruction, we must make sure to rewrite
1459 // any call instructions into invoke instructions.
1460 if (auto *II = dyn_cast<InvokeInst>(TheCall)) {
1461 BasicBlock *UnwindDest = II->getUnwindDest();
1462 Instruction *FirstNonPHI = UnwindDest->getFirstNonPHI();
1463 if (isa<LandingPadInst>(FirstNonPHI)) {
1464 HandleInlinedLandingPad(II, &*FirstNewBlock, InlinedFunctionInfo);
1466 HandleInlinedEHPad(II, &*FirstNewBlock, InlinedFunctionInfo);
1470 // Handle any inlined musttail call sites. In order for a new call site to be
1471 // musttail, the source of the clone and the inlined call site must have been
1472 // musttail. Therefore it's safe to return without merging control into the
1474 if (InlinedMustTailCalls) {
1475 // Check if we need to bitcast the result of any musttail calls.
1476 Type *NewRetTy = Caller->getReturnType();
1477 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
1479 // Handle the returns preceded by musttail calls separately.
1480 SmallVector<ReturnInst *, 8> NormalReturns;
1481 for (ReturnInst *RI : Returns) {
1482 CallInst *ReturnedMustTail =
1483 RI->getParent()->getTerminatingMustTailCall();
1484 if (!ReturnedMustTail) {
1485 NormalReturns.push_back(RI);
1491 // Delete the old return and any preceding bitcast.
1492 BasicBlock *CurBB = RI->getParent();
1493 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
1494 RI->eraseFromParent();
1496 OldCast->eraseFromParent();
1498 // Insert a new bitcast and return with the right type.
1499 IRBuilder<> Builder(CurBB);
1500 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
1503 // Leave behind the normal returns so we can merge control flow.
1504 std::swap(Returns, NormalReturns);
1507 // If we cloned in _exactly one_ basic block, and if that block ends in a
1508 // return instruction, we splice the body of the inlined callee directly into
1509 // the calling basic block.
1510 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
1511 // Move all of the instructions right before the call.
1512 OrigBB->getInstList().splice(TheCall->getIterator(),
1513 FirstNewBlock->getInstList(),
1514 FirstNewBlock->begin(), FirstNewBlock->end());
1515 // Remove the cloned basic block.
1516 Caller->getBasicBlockList().pop_back();
1518 // If the call site was an invoke instruction, add a branch to the normal
1520 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1521 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
1522 NewBr->setDebugLoc(Returns[0]->getDebugLoc());
1525 // If the return instruction returned a value, replace uses of the call with
1526 // uses of the returned value.
1527 if (!TheCall->use_empty()) {
1528 ReturnInst *R = Returns[0];
1529 if (TheCall == R->getReturnValue())
1530 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1532 TheCall->replaceAllUsesWith(R->getReturnValue());
1534 // Since we are now done with the Call/Invoke, we can delete it.
1535 TheCall->eraseFromParent();
1537 // Since we are now done with the return instruction, delete it also.
1538 Returns[0]->eraseFromParent();
1540 // We are now done with the inlining.
1544 // Otherwise, we have the normal case, of more than one block to inline or
1545 // multiple return sites.
1547 // We want to clone the entire callee function into the hole between the
1548 // "starter" and "ender" blocks. How we accomplish this depends on whether
1549 // this is an invoke instruction or a call instruction.
1550 BasicBlock *AfterCallBB;
1551 BranchInst *CreatedBranchToNormalDest = nullptr;
1552 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1554 // Add an unconditional branch to make this look like the CallInst case...
1555 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
1557 // Split the basic block. This guarantees that no PHI nodes will have to be
1558 // updated due to new incoming edges, and make the invoke case more
1559 // symmetric to the call case.
1561 OrigBB->splitBasicBlock(CreatedBranchToNormalDest->getIterator(),
1562 CalledFunc->getName() + ".exit");
1564 } else { // It's a call
1565 // If this is a call instruction, we need to split the basic block that
1566 // the call lives in.
1568 AfterCallBB = OrigBB->splitBasicBlock(TheCall->getIterator(),
1569 CalledFunc->getName() + ".exit");
1572 // Change the branch that used to go to AfterCallBB to branch to the first
1573 // basic block of the inlined function.
1575 TerminatorInst *Br = OrigBB->getTerminator();
1576 assert(Br && Br->getOpcode() == Instruction::Br &&
1577 "splitBasicBlock broken!");
1578 Br->setOperand(0, &*FirstNewBlock);
1580 // Now that the function is correct, make it a little bit nicer. In
1581 // particular, move the basic blocks inserted from the end of the function
1582 // into the space made by splitting the source basic block.
1583 Caller->getBasicBlockList().splice(AfterCallBB->getIterator(),
1584 Caller->getBasicBlockList(), FirstNewBlock,
1587 // Handle all of the return instructions that we just cloned in, and eliminate
1588 // any users of the original call/invoke instruction.
1589 Type *RTy = CalledFunc->getReturnType();
1591 PHINode *PHI = nullptr;
1592 if (Returns.size() > 1) {
1593 // The PHI node should go at the front of the new basic block to merge all
1594 // possible incoming values.
1595 if (!TheCall->use_empty()) {
1596 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
1597 &AfterCallBB->front());
1598 // Anything that used the result of the function call should now use the
1599 // PHI node as their operand.
1600 TheCall->replaceAllUsesWith(PHI);
1603 // Loop over all of the return instructions adding entries to the PHI node
1606 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1607 ReturnInst *RI = Returns[i];
1608 assert(RI->getReturnValue()->getType() == PHI->getType() &&
1609 "Ret value not consistent in function!");
1610 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
1614 // Add a branch to the merge points and remove return instructions.
1616 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1617 ReturnInst *RI = Returns[i];
1618 BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
1619 Loc = RI->getDebugLoc();
1620 BI->setDebugLoc(Loc);
1621 RI->eraseFromParent();
1623 // We need to set the debug location to *somewhere* inside the
1624 // inlined function. The line number may be nonsensical, but the
1625 // instruction will at least be associated with the right
1627 if (CreatedBranchToNormalDest)
1628 CreatedBranchToNormalDest->setDebugLoc(Loc);
1629 } else if (!Returns.empty()) {
1630 // Otherwise, if there is exactly one return value, just replace anything
1631 // using the return value of the call with the computed value.
1632 if (!TheCall->use_empty()) {
1633 if (TheCall == Returns[0]->getReturnValue())
1634 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1636 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
1639 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
1640 BasicBlock *ReturnBB = Returns[0]->getParent();
1641 ReturnBB->replaceAllUsesWith(AfterCallBB);
1643 // Splice the code from the return block into the block that it will return
1644 // to, which contains the code that was after the call.
1645 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
1646 ReturnBB->getInstList());
1648 if (CreatedBranchToNormalDest)
1649 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
1651 // Delete the return instruction now and empty ReturnBB now.
1652 Returns[0]->eraseFromParent();
1653 ReturnBB->eraseFromParent();
1654 } else if (!TheCall->use_empty()) {
1655 // No returns, but something is using the return value of the call. Just
1657 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1660 // Since we are now done with the Call/Invoke, we can delete it.
1661 TheCall->eraseFromParent();
1663 // If we inlined any musttail calls and the original return is now
1664 // unreachable, delete it. It can only contain a bitcast and ret.
1665 if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB))
1666 AfterCallBB->eraseFromParent();
1668 // We should always be able to fold the entry block of the function into the
1669 // single predecessor of the block...
1670 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
1671 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
1673 // Splice the code entry block into calling block, right before the
1674 // unconditional branch.
1675 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
1676 OrigBB->getInstList().splice(Br->getIterator(), CalleeEntry->getInstList());
1678 // Remove the unconditional branch.
1679 OrigBB->getInstList().erase(Br);
1681 // Now we can remove the CalleeEntry block, which is now empty.
1682 Caller->getBasicBlockList().erase(CalleeEntry);
1684 // If we inserted a phi node, check to see if it has a single value (e.g. all
1685 // the entries are the same or undef). If so, remove the PHI so it doesn't
1686 // block other optimizations.
1688 auto &DL = Caller->getParent()->getDataLayout();
1689 if (Value *V = SimplifyInstruction(PHI, DL, nullptr, nullptr,
1690 &IFI.ACT->getAssumptionCache(*Caller))) {
1691 PHI->replaceAllUsesWith(V);
1692 PHI->eraseFromParent();