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 SmallVector<Value*, 8> InvokeArgs(CI->arg_begin(), CI->arg_end());
210 SmallVector<OperandBundleDef, 1> OpBundles;
212 CI->getOperandBundlesAsDefs(OpBundles);
214 // Note: we're round tripping operand bundles through memory here, and that
215 // can potentially be avoided with a cleverer API design that we do not have
219 InvokeInst::Create(CI->getCalledValue(), Split, UnwindEdge, InvokeArgs,
220 OpBundles, CI->getName(), BB);
221 II->setDebugLoc(CI->getDebugLoc());
222 II->setCallingConv(CI->getCallingConv());
223 II->setAttributes(CI->getAttributes());
225 // Make sure that anything using the call now uses the invoke! This also
226 // updates the CallGraph if present, because it uses a WeakVH.
227 CI->replaceAllUsesWith(II);
229 // Delete the original call
230 Split->getInstList().pop_front();
236 /// If we inlined an invoke site, we need to convert calls
237 /// in the body of the inlined function into invokes.
239 /// II is the invoke instruction being inlined. FirstNewBlock is the first
240 /// block of the inlined code (the last block is the end of the function),
241 /// and InlineCodeInfo is information about the code that got inlined.
242 static void HandleInlinedLandingPad(InvokeInst *II, BasicBlock *FirstNewBlock,
243 ClonedCodeInfo &InlinedCodeInfo) {
244 BasicBlock *InvokeDest = II->getUnwindDest();
246 Function *Caller = FirstNewBlock->getParent();
248 // The inlined code is currently at the end of the function, scan from the
249 // start of the inlined code to its end, checking for stuff we need to
251 LandingPadInliningInfo Invoke(II);
253 // Get all of the inlined landing pad instructions.
254 SmallPtrSet<LandingPadInst*, 16> InlinedLPads;
255 for (Function::iterator I = FirstNewBlock->getIterator(), E = Caller->end();
257 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator()))
258 InlinedLPads.insert(II->getLandingPadInst());
260 // Append the clauses from the outer landing pad instruction into the inlined
261 // landing pad instructions.
262 LandingPadInst *OuterLPad = Invoke.getLandingPadInst();
263 for (LandingPadInst *InlinedLPad : InlinedLPads) {
264 unsigned OuterNum = OuterLPad->getNumClauses();
265 InlinedLPad->reserveClauses(OuterNum);
266 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx)
267 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx));
268 if (OuterLPad->isCleanup())
269 InlinedLPad->setCleanup(true);
272 for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end();
274 if (InlinedCodeInfo.ContainsCalls)
275 if (BasicBlock *NewBB = HandleCallsInBlockInlinedThroughInvoke(
276 &*BB, Invoke.getOuterResumeDest()))
277 // Update any PHI nodes in the exceptional block to indicate that there
278 // is now a new entry in them.
279 Invoke.addIncomingPHIValuesFor(NewBB);
281 // Forward any resumes that are remaining here.
282 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator()))
283 Invoke.forwardResume(RI, InlinedLPads);
286 // Now that everything is happy, we have one final detail. The PHI nodes in
287 // the exception destination block still have entries due to the original
288 // invoke instruction. Eliminate these entries (which might even delete the
290 InvokeDest->removePredecessor(II->getParent());
293 /// If we inlined an invoke site, we need to convert calls
294 /// in the body of the inlined function into invokes.
296 /// II is the invoke instruction being inlined. FirstNewBlock is the first
297 /// block of the inlined code (the last block is the end of the function),
298 /// and InlineCodeInfo is information about the code that got inlined.
299 static void HandleInlinedEHPad(InvokeInst *II, BasicBlock *FirstNewBlock,
300 ClonedCodeInfo &InlinedCodeInfo) {
301 BasicBlock *UnwindDest = II->getUnwindDest();
302 Function *Caller = FirstNewBlock->getParent();
304 assert(UnwindDest->getFirstNonPHI()->isEHPad() && "unexpected BasicBlock!");
306 // If there are PHI nodes in the unwind destination block, we need to keep
307 // track of which values came into them from the invoke before removing the
308 // edge from this block.
309 SmallVector<Value *, 8> UnwindDestPHIValues;
310 llvm::BasicBlock *InvokeBB = II->getParent();
311 for (Instruction &I : *UnwindDest) {
312 // Save the value to use for this edge.
313 PHINode *PHI = dyn_cast<PHINode>(&I);
316 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
319 // Add incoming-PHI values to the unwind destination block for the given basic
320 // block, using the values for the original invoke's source block.
321 auto UpdatePHINodes = [&](BasicBlock *Src) {
322 BasicBlock::iterator I = UnwindDest->begin();
323 for (Value *V : UnwindDestPHIValues) {
324 PHINode *PHI = cast<PHINode>(I);
325 PHI->addIncoming(V, Src);
330 // Forward EH terminator instructions to the caller's invoke destination.
331 // This is as simple as connect all the instructions which 'unwind to caller'
332 // to the invoke destination.
333 for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end();
335 Instruction *I = BB->getFirstNonPHI();
337 if (auto *CEPI = dyn_cast<CatchEndPadInst>(I)) {
338 if (CEPI->unwindsToCaller()) {
339 CatchEndPadInst::Create(CEPI->getContext(), UnwindDest, CEPI);
340 CEPI->eraseFromParent();
341 UpdatePHINodes(&*BB);
343 } else if (auto *CEPI = dyn_cast<CleanupEndPadInst>(I)) {
344 if (CEPI->unwindsToCaller()) {
345 CleanupEndPadInst::Create(CEPI->getCleanupPad(), UnwindDest, CEPI);
346 CEPI->eraseFromParent();
347 UpdatePHINodes(&*BB);
349 } else if (auto *TPI = dyn_cast<TerminatePadInst>(I)) {
350 if (TPI->unwindsToCaller()) {
351 SmallVector<Value *, 3> TerminatePadArgs;
352 for (Value *ArgOperand : TPI->arg_operands())
353 TerminatePadArgs.push_back(ArgOperand);
354 TerminatePadInst::Create(TPI->getContext(), UnwindDest,
355 TerminatePadArgs, TPI);
356 TPI->eraseFromParent();
357 UpdatePHINodes(&*BB);
360 assert(isa<CatchPadInst>(I) || isa<CleanupPadInst>(I));
364 if (auto *CRI = dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
365 if (CRI->unwindsToCaller()) {
366 CleanupReturnInst::Create(CRI->getCleanupPad(), UnwindDest, CRI);
367 CRI->eraseFromParent();
368 UpdatePHINodes(&*BB);
373 if (InlinedCodeInfo.ContainsCalls)
374 for (Function::iterator BB = FirstNewBlock->getIterator(),
377 if (BasicBlock *NewBB =
378 HandleCallsInBlockInlinedThroughInvoke(&*BB, UnwindDest))
379 // Update any PHI nodes in the exceptional block to indicate that there
380 // is now a new entry in them.
381 UpdatePHINodes(NewBB);
383 // Now that everything is happy, we have one final detail. The PHI nodes in
384 // the exception destination block still have entries due to the original
385 // invoke instruction. Eliminate these entries (which might even delete the
387 UnwindDest->removePredecessor(InvokeBB);
390 /// When inlining a function that contains noalias scope metadata,
391 /// this metadata needs to be cloned so that the inlined blocks
392 /// have different "unqiue scopes" at every call site. Were this not done, then
393 /// aliasing scopes from a function inlined into a caller multiple times could
394 /// not be differentiated (and this would lead to miscompiles because the
395 /// non-aliasing property communicated by the metadata could have
396 /// call-site-specific control dependencies).
397 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) {
398 const Function *CalledFunc = CS.getCalledFunction();
399 SetVector<const MDNode *> MD;
401 // Note: We could only clone the metadata if it is already used in the
402 // caller. I'm omitting that check here because it might confuse
403 // inter-procedural alias analysis passes. We can revisit this if it becomes
404 // an efficiency or overhead problem.
406 for (Function::const_iterator I = CalledFunc->begin(), IE = CalledFunc->end();
408 for (BasicBlock::const_iterator J = I->begin(), JE = I->end(); J != JE; ++J) {
409 if (const MDNode *M = J->getMetadata(LLVMContext::MD_alias_scope))
411 if (const MDNode *M = J->getMetadata(LLVMContext::MD_noalias))
418 // Walk the existing metadata, adding the complete (perhaps cyclic) chain to
420 SmallVector<const Metadata *, 16> Queue(MD.begin(), MD.end());
421 while (!Queue.empty()) {
422 const MDNode *M = cast<MDNode>(Queue.pop_back_val());
423 for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i)
424 if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i)))
429 // Now we have a complete set of all metadata in the chains used to specify
430 // the noalias scopes and the lists of those scopes.
431 SmallVector<TempMDTuple, 16> DummyNodes;
432 DenseMap<const MDNode *, TrackingMDNodeRef> MDMap;
433 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
435 DummyNodes.push_back(MDTuple::getTemporary(CalledFunc->getContext(), None));
436 MDMap[*I].reset(DummyNodes.back().get());
439 // Create new metadata nodes to replace the dummy nodes, replacing old
440 // metadata references with either a dummy node or an already-created new
442 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
444 SmallVector<Metadata *, 4> NewOps;
445 for (unsigned i = 0, ie = (*I)->getNumOperands(); i != ie; ++i) {
446 const Metadata *V = (*I)->getOperand(i);
447 if (const MDNode *M = dyn_cast<MDNode>(V))
448 NewOps.push_back(MDMap[M]);
450 NewOps.push_back(const_cast<Metadata *>(V));
453 MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps);
454 MDTuple *TempM = cast<MDTuple>(MDMap[*I]);
455 assert(TempM->isTemporary() && "Expected temporary node");
457 TempM->replaceAllUsesWith(NewM);
460 // Now replace the metadata in the new inlined instructions with the
461 // repacements from the map.
462 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
463 VMI != VMIE; ++VMI) {
467 Instruction *NI = dyn_cast<Instruction>(VMI->second);
471 if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope)) {
472 MDNode *NewMD = MDMap[M];
473 // If the call site also had alias scope metadata (a list of scopes to
474 // which instructions inside it might belong), propagate those scopes to
475 // the inlined instructions.
477 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
478 NewMD = MDNode::concatenate(NewMD, CSM);
479 NI->setMetadata(LLVMContext::MD_alias_scope, NewMD);
480 } else if (NI->mayReadOrWriteMemory()) {
482 CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
483 NI->setMetadata(LLVMContext::MD_alias_scope, M);
486 if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias)) {
487 MDNode *NewMD = MDMap[M];
488 // If the call site also had noalias metadata (a list of scopes with
489 // which instructions inside it don't alias), propagate those scopes to
490 // the inlined instructions.
492 CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
493 NewMD = MDNode::concatenate(NewMD, CSM);
494 NI->setMetadata(LLVMContext::MD_noalias, NewMD);
495 } else if (NI->mayReadOrWriteMemory()) {
496 if (MDNode *M = CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
497 NI->setMetadata(LLVMContext::MD_noalias, M);
502 /// If the inlined function has noalias arguments,
503 /// then add new alias scopes for each noalias argument, tag the mapped noalias
504 /// parameters with noalias metadata specifying the new scope, and tag all
505 /// non-derived loads, stores and memory intrinsics with the new alias scopes.
506 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
507 const DataLayout &DL, AAResults *CalleeAAR) {
508 if (!EnableNoAliasConversion)
511 const Function *CalledFunc = CS.getCalledFunction();
512 SmallVector<const Argument *, 4> NoAliasArgs;
514 for (const Argument &I : CalledFunc->args()) {
515 if (I.hasNoAliasAttr() && !I.hasNUses(0))
516 NoAliasArgs.push_back(&I);
519 if (NoAliasArgs.empty())
522 // To do a good job, if a noalias variable is captured, we need to know if
523 // the capture point dominates the particular use we're considering.
525 DT.recalculate(const_cast<Function&>(*CalledFunc));
527 // noalias indicates that pointer values based on the argument do not alias
528 // pointer values which are not based on it. So we add a new "scope" for each
529 // noalias function argument. Accesses using pointers based on that argument
530 // become part of that alias scope, accesses using pointers not based on that
531 // argument are tagged as noalias with that scope.
533 DenseMap<const Argument *, MDNode *> NewScopes;
534 MDBuilder MDB(CalledFunc->getContext());
536 // Create a new scope domain for this function.
538 MDB.createAnonymousAliasScopeDomain(CalledFunc->getName());
539 for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) {
540 const Argument *A = NoAliasArgs[i];
542 std::string Name = CalledFunc->getName();
545 Name += A->getName();
547 Name += ": argument ";
551 // Note: We always create a new anonymous root here. This is true regardless
552 // of the linkage of the callee because the aliasing "scope" is not just a
553 // property of the callee, but also all control dependencies in the caller.
554 MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
555 NewScopes.insert(std::make_pair(A, NewScope));
558 // Iterate over all new instructions in the map; for all memory-access
559 // instructions, add the alias scope metadata.
560 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
561 VMI != VMIE; ++VMI) {
562 if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) {
566 Instruction *NI = dyn_cast<Instruction>(VMI->second);
570 bool IsArgMemOnlyCall = false, IsFuncCall = false;
571 SmallVector<const Value *, 2> PtrArgs;
573 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
574 PtrArgs.push_back(LI->getPointerOperand());
575 else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
576 PtrArgs.push_back(SI->getPointerOperand());
577 else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
578 PtrArgs.push_back(VAAI->getPointerOperand());
579 else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
580 PtrArgs.push_back(CXI->getPointerOperand());
581 else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
582 PtrArgs.push_back(RMWI->getPointerOperand());
583 else if (ImmutableCallSite ICS = ImmutableCallSite(I)) {
584 // If we know that the call does not access memory, then we'll still
585 // know that about the inlined clone of this call site, and we don't
586 // need to add metadata.
587 if (ICS.doesNotAccessMemory())
592 FunctionModRefBehavior MRB = CalleeAAR->getModRefBehavior(ICS);
593 if (MRB == FMRB_OnlyAccessesArgumentPointees ||
594 MRB == FMRB_OnlyReadsArgumentPointees)
595 IsArgMemOnlyCall = true;
598 for (ImmutableCallSite::arg_iterator AI = ICS.arg_begin(),
599 AE = ICS.arg_end(); AI != AE; ++AI) {
600 // We need to check the underlying objects of all arguments, not just
601 // the pointer arguments, because we might be passing pointers as
603 // However, if we know that the call only accesses pointer arguments,
604 // then we only need to check the pointer arguments.
605 if (IsArgMemOnlyCall && !(*AI)->getType()->isPointerTy())
608 PtrArgs.push_back(*AI);
612 // If we found no pointers, then this instruction is not suitable for
613 // pairing with an instruction to receive aliasing metadata.
614 // However, if this is a call, this we might just alias with none of the
615 // noalias arguments.
616 if (PtrArgs.empty() && !IsFuncCall)
619 // It is possible that there is only one underlying object, but you
620 // need to go through several PHIs to see it, and thus could be
621 // repeated in the Objects list.
622 SmallPtrSet<const Value *, 4> ObjSet;
623 SmallVector<Metadata *, 4> Scopes, NoAliases;
625 SmallSetVector<const Argument *, 4> NAPtrArgs;
626 for (unsigned i = 0, ie = PtrArgs.size(); i != ie; ++i) {
627 SmallVector<Value *, 4> Objects;
628 GetUnderlyingObjects(const_cast<Value*>(PtrArgs[i]),
629 Objects, DL, /* LI = */ nullptr);
631 for (Value *O : Objects)
635 // Figure out if we're derived from anything that is not a noalias
637 bool CanDeriveViaCapture = false, UsesAliasingPtr = false;
638 for (const Value *V : ObjSet) {
639 // Is this value a constant that cannot be derived from any pointer
640 // value (we need to exclude constant expressions, for example, that
641 // are formed from arithmetic on global symbols).
642 bool IsNonPtrConst = isa<ConstantInt>(V) || isa<ConstantFP>(V) ||
643 isa<ConstantPointerNull>(V) ||
644 isa<ConstantDataVector>(V) || isa<UndefValue>(V);
648 // If this is anything other than a noalias argument, then we cannot
649 // completely describe the aliasing properties using alias.scope
650 // metadata (and, thus, won't add any).
651 if (const Argument *A = dyn_cast<Argument>(V)) {
652 if (!A->hasNoAliasAttr())
653 UsesAliasingPtr = true;
655 UsesAliasingPtr = true;
658 // If this is not some identified function-local object (which cannot
659 // directly alias a noalias argument), or some other argument (which,
660 // by definition, also cannot alias a noalias argument), then we could
661 // alias a noalias argument that has been captured).
662 if (!isa<Argument>(V) &&
663 !isIdentifiedFunctionLocal(const_cast<Value*>(V)))
664 CanDeriveViaCapture = true;
667 // A function call can always get captured noalias pointers (via other
668 // parameters, globals, etc.).
669 if (IsFuncCall && !IsArgMemOnlyCall)
670 CanDeriveViaCapture = true;
672 // First, we want to figure out all of the sets with which we definitely
673 // don't alias. Iterate over all noalias set, and add those for which:
674 // 1. The noalias argument is not in the set of objects from which we
675 // definitely derive.
676 // 2. The noalias argument has not yet been captured.
677 // An arbitrary function that might load pointers could see captured
678 // noalias arguments via other noalias arguments or globals, and so we
679 // must always check for prior capture.
680 for (const Argument *A : NoAliasArgs) {
681 if (!ObjSet.count(A) && (!CanDeriveViaCapture ||
682 // It might be tempting to skip the
683 // PointerMayBeCapturedBefore check if
684 // A->hasNoCaptureAttr() is true, but this is
685 // incorrect because nocapture only guarantees
686 // that no copies outlive the function, not
687 // that the value cannot be locally captured.
688 !PointerMayBeCapturedBefore(A,
689 /* ReturnCaptures */ false,
690 /* StoreCaptures */ false, I, &DT)))
691 NoAliases.push_back(NewScopes[A]);
694 if (!NoAliases.empty())
695 NI->setMetadata(LLVMContext::MD_noalias,
697 NI->getMetadata(LLVMContext::MD_noalias),
698 MDNode::get(CalledFunc->getContext(), NoAliases)));
700 // Next, we want to figure out all of the sets to which we might belong.
701 // We might belong to a set if the noalias argument is in the set of
702 // underlying objects. If there is some non-noalias argument in our list
703 // of underlying objects, then we cannot add a scope because the fact
704 // that some access does not alias with any set of our noalias arguments
705 // cannot itself guarantee that it does not alias with this access
706 // (because there is some pointer of unknown origin involved and the
707 // other access might also depend on this pointer). We also cannot add
708 // scopes to arbitrary functions unless we know they don't access any
709 // non-parameter pointer-values.
710 bool CanAddScopes = !UsesAliasingPtr;
711 if (CanAddScopes && IsFuncCall)
712 CanAddScopes = IsArgMemOnlyCall;
715 for (const Argument *A : NoAliasArgs) {
717 Scopes.push_back(NewScopes[A]);
722 LLVMContext::MD_alias_scope,
723 MDNode::concatenate(NI->getMetadata(LLVMContext::MD_alias_scope),
724 MDNode::get(CalledFunc->getContext(), Scopes)));
729 /// If the inlined function has non-byval align arguments, then
730 /// add @llvm.assume-based alignment assumptions to preserve this information.
731 static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) {
732 if (!PreserveAlignmentAssumptions)
734 auto &DL = CS.getCaller()->getParent()->getDataLayout();
736 // To avoid inserting redundant assumptions, we should check for assumptions
737 // already in the caller. To do this, we might need a DT of the caller.
739 bool DTCalculated = false;
741 Function *CalledFunc = CS.getCalledFunction();
742 for (Function::arg_iterator I = CalledFunc->arg_begin(),
743 E = CalledFunc->arg_end();
745 unsigned Align = I->getType()->isPointerTy() ? I->getParamAlignment() : 0;
746 if (Align && !I->hasByValOrInAllocaAttr() && !I->hasNUses(0)) {
748 DT.recalculate(const_cast<Function&>(*CS.getInstruction()->getParent()
753 // If we can already prove the asserted alignment in the context of the
754 // caller, then don't bother inserting the assumption.
755 Value *Arg = CS.getArgument(I->getArgNo());
756 if (getKnownAlignment(Arg, DL, CS.getInstruction(),
757 &IFI.ACT->getAssumptionCache(*CS.getCaller()),
761 IRBuilder<>(CS.getInstruction())
762 .CreateAlignmentAssumption(DL, Arg, Align);
767 /// Once we have cloned code over from a callee into the caller,
768 /// update the specified callgraph to reflect the changes we made.
769 /// Note that it's possible that not all code was copied over, so only
770 /// some edges of the callgraph may remain.
771 static void UpdateCallGraphAfterInlining(CallSite CS,
772 Function::iterator FirstNewBlock,
773 ValueToValueMapTy &VMap,
774 InlineFunctionInfo &IFI) {
775 CallGraph &CG = *IFI.CG;
776 const Function *Caller = CS.getInstruction()->getParent()->getParent();
777 const Function *Callee = CS.getCalledFunction();
778 CallGraphNode *CalleeNode = CG[Callee];
779 CallGraphNode *CallerNode = CG[Caller];
781 // Since we inlined some uninlined call sites in the callee into the caller,
782 // add edges from the caller to all of the callees of the callee.
783 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
785 // Consider the case where CalleeNode == CallerNode.
786 CallGraphNode::CalledFunctionsVector CallCache;
787 if (CalleeNode == CallerNode) {
788 CallCache.assign(I, E);
789 I = CallCache.begin();
793 for (; I != E; ++I) {
794 const Value *OrigCall = I->first;
796 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
797 // Only copy the edge if the call was inlined!
798 if (VMI == VMap.end() || VMI->second == nullptr)
801 // If the call was inlined, but then constant folded, there is no edge to
802 // add. Check for this case.
803 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
807 // We do not treat intrinsic calls like real function calls because we
808 // expect them to become inline code; do not add an edge for an intrinsic.
809 CallSite CS = CallSite(NewCall);
810 if (CS && CS.getCalledFunction() && CS.getCalledFunction()->isIntrinsic())
813 // Remember that this call site got inlined for the client of
815 IFI.InlinedCalls.push_back(NewCall);
817 // It's possible that inlining the callsite will cause it to go from an
818 // indirect to a direct call by resolving a function pointer. If this
819 // happens, set the callee of the new call site to a more precise
820 // destination. This can also happen if the call graph node of the caller
821 // was just unnecessarily imprecise.
822 if (!I->second->getFunction())
823 if (Function *F = CallSite(NewCall).getCalledFunction()) {
824 // Indirect call site resolved to direct call.
825 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
830 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
833 // Update the call graph by deleting the edge from Callee to Caller. We must
834 // do this after the loop above in case Caller and Callee are the same.
835 CallerNode->removeCallEdgeFor(CS);
838 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
839 BasicBlock *InsertBlock,
840 InlineFunctionInfo &IFI) {
841 Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
842 IRBuilder<> Builder(InsertBlock, InsertBlock->begin());
844 Value *Size = Builder.getInt64(M->getDataLayout().getTypeStoreSize(AggTy));
846 // Always generate a memcpy of alignment 1 here because we don't know
847 // the alignment of the src pointer. Other optimizations can infer
849 Builder.CreateMemCpy(Dst, Src, Size, /*Align=*/1);
852 /// When inlining a call site that has a byval argument,
853 /// we have to make the implicit memcpy explicit by adding it.
854 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
855 const Function *CalledFunc,
856 InlineFunctionInfo &IFI,
857 unsigned ByValAlignment) {
858 PointerType *ArgTy = cast<PointerType>(Arg->getType());
859 Type *AggTy = ArgTy->getElementType();
861 Function *Caller = TheCall->getParent()->getParent();
863 // If the called function is readonly, then it could not mutate the caller's
864 // copy of the byval'd memory. In this case, it is safe to elide the copy and
866 if (CalledFunc->onlyReadsMemory()) {
867 // If the byval argument has a specified alignment that is greater than the
868 // passed in pointer, then we either have to round up the input pointer or
869 // give up on this transformation.
870 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
873 const DataLayout &DL = Caller->getParent()->getDataLayout();
875 // If the pointer is already known to be sufficiently aligned, or if we can
876 // round it up to a larger alignment, then we don't need a temporary.
877 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, DL, TheCall,
878 &IFI.ACT->getAssumptionCache(*Caller)) >=
882 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
883 // for code quality, but rarely happens and is required for correctness.
886 // Create the alloca. If we have DataLayout, use nice alignment.
888 Caller->getParent()->getDataLayout().getPrefTypeAlignment(AggTy);
890 // If the byval had an alignment specified, we *must* use at least that
891 // alignment, as it is required by the byval argument (and uses of the
892 // pointer inside the callee).
893 Align = std::max(Align, ByValAlignment);
895 Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(),
896 &*Caller->begin()->begin());
897 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
899 // Uses of the argument in the function should use our new alloca
904 // Check whether this Value is used by a lifetime intrinsic.
905 static bool isUsedByLifetimeMarker(Value *V) {
906 for (User *U : V->users()) {
907 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
908 switch (II->getIntrinsicID()) {
910 case Intrinsic::lifetime_start:
911 case Intrinsic::lifetime_end:
919 // Check whether the given alloca already has
920 // lifetime.start or lifetime.end intrinsics.
921 static bool hasLifetimeMarkers(AllocaInst *AI) {
922 Type *Ty = AI->getType();
923 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
924 Ty->getPointerAddressSpace());
926 return isUsedByLifetimeMarker(AI);
928 // Do a scan to find all the casts to i8*.
929 for (User *U : AI->users()) {
930 if (U->getType() != Int8PtrTy) continue;
931 if (U->stripPointerCasts() != AI) continue;
932 if (isUsedByLifetimeMarker(U))
938 /// Rebuild the entire inlined-at chain for this instruction so that the top of
939 /// the chain now is inlined-at the new call site.
941 updateInlinedAtInfo(DebugLoc DL, DILocation *InlinedAtNode, LLVMContext &Ctx,
942 DenseMap<const DILocation *, DILocation *> &IANodes) {
943 SmallVector<DILocation *, 3> InlinedAtLocations;
944 DILocation *Last = InlinedAtNode;
945 DILocation *CurInlinedAt = DL;
947 // Gather all the inlined-at nodes
948 while (DILocation *IA = CurInlinedAt->getInlinedAt()) {
949 // Skip any we've already built nodes for
950 if (DILocation *Found = IANodes[IA]) {
955 InlinedAtLocations.push_back(IA);
959 // Starting from the top, rebuild the nodes to point to the new inlined-at
960 // location (then rebuilding the rest of the chain behind it) and update the
961 // map of already-constructed inlined-at nodes.
962 for (const DILocation *MD : make_range(InlinedAtLocations.rbegin(),
963 InlinedAtLocations.rend())) {
964 Last = IANodes[MD] = DILocation::getDistinct(
965 Ctx, MD->getLine(), MD->getColumn(), MD->getScope(), Last);
968 // And finally create the normal location for this instruction, referring to
969 // the new inlined-at chain.
970 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(), Last);
973 /// Update inlined instructions' line numbers to
974 /// to encode location where these instructions are inlined.
975 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
976 Instruction *TheCall) {
977 DebugLoc TheCallDL = TheCall->getDebugLoc();
981 auto &Ctx = Fn->getContext();
982 DILocation *InlinedAtNode = TheCallDL;
984 // Create a unique call site, not to be confused with any other call from the
986 InlinedAtNode = DILocation::getDistinct(
987 Ctx, InlinedAtNode->getLine(), InlinedAtNode->getColumn(),
988 InlinedAtNode->getScope(), InlinedAtNode->getInlinedAt());
990 // Cache the inlined-at nodes as they're built so they are reused, without
991 // this every instruction's inlined-at chain would become distinct from each
993 DenseMap<const DILocation *, DILocation *> IANodes;
995 for (; FI != Fn->end(); ++FI) {
996 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
998 DebugLoc DL = BI->getDebugLoc();
1000 // If the inlined instruction has no line number, make it look as if it
1001 // originates from the call location. This is important for
1002 // ((__always_inline__, __nodebug__)) functions which must use caller
1003 // location for all instructions in their function body.
1005 // Don't update static allocas, as they may get moved later.
1006 if (auto *AI = dyn_cast<AllocaInst>(BI))
1007 if (isa<Constant>(AI->getArraySize()))
1010 BI->setDebugLoc(TheCallDL);
1012 BI->setDebugLoc(updateInlinedAtInfo(DL, InlinedAtNode, BI->getContext(), IANodes));
1018 /// This function inlines the called function into the basic block of the
1019 /// caller. This returns false if it is not possible to inline this call.
1020 /// The program is still in a well defined state if this occurs though.
1022 /// Note that this only does one level of inlining. For example, if the
1023 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
1024 /// exists in the instruction stream. Similarly this will inline a recursive
1025 /// function by one level.
1026 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
1027 AAResults *CalleeAAR, bool InsertLifetime) {
1028 Instruction *TheCall = CS.getInstruction();
1029 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
1030 "Instruction not in function!");
1032 // If IFI has any state in it, zap it before we fill it in.
1035 const Function *CalledFunc = CS.getCalledFunction();
1036 if (!CalledFunc || // Can't inline external function or indirect
1037 CalledFunc->isDeclaration() || // call, or call to a vararg function!
1038 CalledFunc->getFunctionType()->isVarArg()) return false;
1040 // The inliner does not know how to inline through calls with operand bundles
1042 if (CS.hasOperandBundles()) {
1043 // ... but it knows how to inline through "deopt" operand bundles.
1045 CS.getNumOperandBundles() == 1 &&
1046 CS.getOperandBundleAt(0).getTagID() == LLVMContext::OB_deopt;
1051 // If the call to the callee cannot throw, set the 'nounwind' flag on any
1052 // calls that we inline.
1053 bool MarkNoUnwind = CS.doesNotThrow();
1055 BasicBlock *OrigBB = TheCall->getParent();
1056 Function *Caller = OrigBB->getParent();
1058 // GC poses two hazards to inlining, which only occur when the callee has GC:
1059 // 1. If the caller has no GC, then the callee's GC must be propagated to the
1061 // 2. If the caller has a differing GC, it is invalid to inline.
1062 if (CalledFunc->hasGC()) {
1063 if (!Caller->hasGC())
1064 Caller->setGC(CalledFunc->getGC());
1065 else if (CalledFunc->getGC() != Caller->getGC())
1069 // Get the personality function from the callee if it contains a landing pad.
1070 Constant *CalledPersonality =
1071 CalledFunc->hasPersonalityFn()
1072 ? CalledFunc->getPersonalityFn()->stripPointerCasts()
1075 // Find the personality function used by the landing pads of the caller. If it
1076 // exists, then check to see that it matches the personality function used in
1078 Constant *CallerPersonality =
1079 Caller->hasPersonalityFn()
1080 ? Caller->getPersonalityFn()->stripPointerCasts()
1082 if (CalledPersonality) {
1083 if (!CallerPersonality)
1084 Caller->setPersonalityFn(CalledPersonality);
1085 // If the personality functions match, then we can perform the
1086 // inlining. Otherwise, we can't inline.
1087 // TODO: This isn't 100% true. Some personality functions are proper
1088 // supersets of others and can be used in place of the other.
1089 else if (CalledPersonality != CallerPersonality)
1093 // Get an iterator to the last basic block in the function, which will have
1094 // the new function inlined after it.
1095 Function::iterator LastBlock = --Caller->end();
1097 // Make sure to capture all of the return instructions from the cloned
1099 SmallVector<ReturnInst*, 8> Returns;
1100 ClonedCodeInfo InlinedFunctionInfo;
1101 Function::iterator FirstNewBlock;
1103 { // Scope to destroy VMap after cloning.
1104 ValueToValueMapTy VMap;
1105 // Keep a list of pair (dst, src) to emit byval initializations.
1106 SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
1108 auto &DL = Caller->getParent()->getDataLayout();
1110 assert(CalledFunc->arg_size() == CS.arg_size() &&
1111 "No varargs calls can be inlined!");
1113 // Calculate the vector of arguments to pass into the function cloner, which
1114 // matches up the formal to the actual argument values.
1115 CallSite::arg_iterator AI = CS.arg_begin();
1117 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
1118 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
1119 Value *ActualArg = *AI;
1121 // When byval arguments actually inlined, we need to make the copy implied
1122 // by them explicit. However, we don't do this if the callee is readonly
1123 // or readnone, because the copy would be unneeded: the callee doesn't
1124 // modify the struct.
1125 if (CS.isByValArgument(ArgNo)) {
1126 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
1127 CalledFunc->getParamAlignment(ArgNo+1));
1128 if (ActualArg != *AI)
1129 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
1132 VMap[&*I] = ActualArg;
1135 // Add alignment assumptions if necessary. We do this before the inlined
1136 // instructions are actually cloned into the caller so that we can easily
1137 // check what will be known at the start of the inlined code.
1138 AddAlignmentAssumptions(CS, IFI);
1140 // We want the inliner to prune the code as it copies. We would LOVE to
1141 // have no dead or constant instructions leftover after inlining occurs
1142 // (which can happen, e.g., because an argument was constant), but we'll be
1143 // happy with whatever the cloner can do.
1144 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
1145 /*ModuleLevelChanges=*/false, Returns, ".i",
1146 &InlinedFunctionInfo, TheCall);
1148 // Remember the first block that is newly cloned over.
1149 FirstNewBlock = LastBlock; ++FirstNewBlock;
1151 // Inject byval arguments initialization.
1152 for (std::pair<Value*, Value*> &Init : ByValInit)
1153 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
1154 &*FirstNewBlock, IFI);
1156 if (CS.hasOperandBundles()) {
1157 auto ParentDeopt = CS.getOperandBundleAt(0);
1158 assert(ParentDeopt.getTagID() == LLVMContext::OB_deopt &&
1159 "Checked on entry!");
1161 SmallVector<OperandBundleDef, 2> OpDefs;
1163 for (auto &VH : InlinedFunctionInfo.OperandBundleCallSites) {
1164 if (!VH) continue; // instruction was DCE'd after being cloned
1166 Instruction *I = cast<Instruction>(VH);
1171 OpDefs.reserve(ICS.getNumOperandBundles());
1173 for (unsigned i = 0, e = ICS.getNumOperandBundles(); i < e; ++i) {
1174 auto ChildOB = ICS.getOperandBundleAt(i);
1175 if (ChildOB.getTagID() != LLVMContext::OB_deopt) {
1176 // If the inlined call has other operand bundles, let them be
1177 OpDefs.emplace_back(ChildOB);
1181 // It may be useful to separate this logic (of handling operand
1182 // bundles) out to a separate "policy" component if this gets crowded.
1183 // Prepend the parent's deoptimization continuation to the newly
1184 // inlined call's deoptimization continuation.
1185 std::vector<Value *> MergedDeoptArgs;
1186 MergedDeoptArgs.reserve(ParentDeopt.Inputs.size() +
1187 ChildOB.Inputs.size());
1189 MergedDeoptArgs.insert(MergedDeoptArgs.end(),
1190 ParentDeopt.Inputs.begin(),
1191 ParentDeopt.Inputs.end());
1192 MergedDeoptArgs.insert(MergedDeoptArgs.end(), ChildOB.Inputs.begin(),
1193 ChildOB.Inputs.end());
1195 OpDefs.emplace_back("deopt", std::move(MergedDeoptArgs));
1198 Instruction *NewI = nullptr;
1199 if (isa<CallInst>(I))
1200 NewI = CallInst::Create(cast<CallInst>(I), OpDefs, I);
1202 NewI = InvokeInst::Create(cast<InvokeInst>(I), OpDefs, I);
1204 // Note: the RAUW does the appropriate fixup in VMap, so we need to do
1205 // this even if the call returns void.
1206 I->replaceAllUsesWith(NewI);
1209 I->eraseFromParent();
1213 // Update the callgraph if requested.
1215 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
1217 // Update inlined instructions' line number information.
1218 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
1220 // Clone existing noalias metadata if necessary.
1221 CloneAliasScopeMetadata(CS, VMap);
1223 // Add noalias metadata if necessary.
1224 AddAliasScopeMetadata(CS, VMap, DL, CalleeAAR);
1226 // FIXME: We could register any cloned assumptions instead of clearing the
1227 // whole function's cache.
1229 IFI.ACT->getAssumptionCache(*Caller).clear();
1232 // If there are any alloca instructions in the block that used to be the entry
1233 // block for the callee, move them to the entry block of the caller. First
1234 // calculate which instruction they should be inserted before. We insert the
1235 // instructions at the end of the current alloca list.
1237 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
1238 for (BasicBlock::iterator I = FirstNewBlock->begin(),
1239 E = FirstNewBlock->end(); I != E; ) {
1240 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
1243 // If the alloca is now dead, remove it. This often occurs due to code
1245 if (AI->use_empty()) {
1246 AI->eraseFromParent();
1250 if (!isa<Constant>(AI->getArraySize()))
1253 // Keep track of the static allocas that we inline into the caller.
1254 IFI.StaticAllocas.push_back(AI);
1256 // Scan for the block of allocas that we can move over, and move them
1258 while (isa<AllocaInst>(I) &&
1259 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
1260 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
1264 // Transfer all of the allocas over in a block. Using splice means
1265 // that the instructions aren't removed from the symbol table, then
1267 Caller->getEntryBlock().getInstList().splice(
1268 InsertPoint, FirstNewBlock->getInstList(), AI->getIterator(), I);
1270 // Move any dbg.declares describing the allocas into the entry basic block.
1271 DIBuilder DIB(*Caller->getParent());
1272 for (auto &AI : IFI.StaticAllocas)
1273 replaceDbgDeclareForAlloca(AI, AI, DIB, /*Deref=*/false);
1276 bool InlinedMustTailCalls = false;
1277 if (InlinedFunctionInfo.ContainsCalls) {
1278 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
1279 if (CallInst *CI = dyn_cast<CallInst>(TheCall))
1280 CallSiteTailKind = CI->getTailCallKind();
1282 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
1284 for (Instruction &I : *BB) {
1285 CallInst *CI = dyn_cast<CallInst>(&I);
1289 // We need to reduce the strength of any inlined tail calls. For
1290 // musttail, we have to avoid introducing potential unbounded stack
1291 // growth. For example, if functions 'f' and 'g' are mutually recursive
1292 // with musttail, we can inline 'g' into 'f' so long as we preserve
1293 // musttail on the cloned call to 'f'. If either the inlined call site
1294 // or the cloned call site is *not* musttail, the program already has
1295 // one frame of stack growth, so it's safe to remove musttail. Here is
1296 // a table of example transformations:
1298 // f -> musttail g -> musttail f ==> f -> musttail f
1299 // f -> musttail g -> tail f ==> f -> tail f
1300 // f -> g -> musttail f ==> f -> f
1301 // f -> g -> tail f ==> f -> f
1302 CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
1303 ChildTCK = std::min(CallSiteTailKind, ChildTCK);
1304 CI->setTailCallKind(ChildTCK);
1305 InlinedMustTailCalls |= CI->isMustTailCall();
1307 // Calls inlined through a 'nounwind' call site should be marked
1310 CI->setDoesNotThrow();
1315 // Leave lifetime markers for the static alloca's, scoping them to the
1316 // function we just inlined.
1317 if (InsertLifetime && !IFI.StaticAllocas.empty()) {
1318 IRBuilder<> builder(&FirstNewBlock->front());
1319 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
1320 AllocaInst *AI = IFI.StaticAllocas[ai];
1322 // If the alloca is already scoped to something smaller than the whole
1323 // function then there's no need to add redundant, less accurate markers.
1324 if (hasLifetimeMarkers(AI))
1327 // Try to determine the size of the allocation.
1328 ConstantInt *AllocaSize = nullptr;
1329 if (ConstantInt *AIArraySize =
1330 dyn_cast<ConstantInt>(AI->getArraySize())) {
1331 auto &DL = Caller->getParent()->getDataLayout();
1332 Type *AllocaType = AI->getAllocatedType();
1333 uint64_t AllocaTypeSize = DL.getTypeAllocSize(AllocaType);
1334 uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
1336 // Don't add markers for zero-sized allocas.
1337 if (AllocaArraySize == 0)
1340 // Check that array size doesn't saturate uint64_t and doesn't
1341 // overflow when it's multiplied by type size.
1342 if (AllocaArraySize != ~0ULL &&
1343 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
1344 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
1345 AllocaArraySize * AllocaTypeSize);
1349 builder.CreateLifetimeStart(AI, AllocaSize);
1350 for (ReturnInst *RI : Returns) {
1351 // Don't insert llvm.lifetime.end calls between a musttail call and a
1352 // return. The return kills all local allocas.
1353 if (InlinedMustTailCalls &&
1354 RI->getParent()->getTerminatingMustTailCall())
1356 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
1361 // If the inlined code contained dynamic alloca instructions, wrap the inlined
1362 // code with llvm.stacksave/llvm.stackrestore intrinsics.
1363 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
1364 Module *M = Caller->getParent();
1365 // Get the two intrinsics we care about.
1366 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
1367 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
1369 // Insert the llvm.stacksave.
1370 CallInst *SavedPtr = IRBuilder<>(&*FirstNewBlock, FirstNewBlock->begin())
1371 .CreateCall(StackSave, {}, "savedstack");
1373 // Insert a call to llvm.stackrestore before any return instructions in the
1374 // inlined function.
1375 for (ReturnInst *RI : Returns) {
1376 // Don't insert llvm.stackrestore calls between a musttail call and a
1377 // return. The return will restore the stack pointer.
1378 if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
1380 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
1384 // If we are inlining for an invoke instruction, we must make sure to rewrite
1385 // any call instructions into invoke instructions.
1386 if (auto *II = dyn_cast<InvokeInst>(TheCall)) {
1387 BasicBlock *UnwindDest = II->getUnwindDest();
1388 Instruction *FirstNonPHI = UnwindDest->getFirstNonPHI();
1389 if (isa<LandingPadInst>(FirstNonPHI)) {
1390 HandleInlinedLandingPad(II, &*FirstNewBlock, InlinedFunctionInfo);
1392 HandleInlinedEHPad(II, &*FirstNewBlock, InlinedFunctionInfo);
1396 // Handle any inlined musttail call sites. In order for a new call site to be
1397 // musttail, the source of the clone and the inlined call site must have been
1398 // musttail. Therefore it's safe to return without merging control into the
1400 if (InlinedMustTailCalls) {
1401 // Check if we need to bitcast the result of any musttail calls.
1402 Type *NewRetTy = Caller->getReturnType();
1403 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
1405 // Handle the returns preceded by musttail calls separately.
1406 SmallVector<ReturnInst *, 8> NormalReturns;
1407 for (ReturnInst *RI : Returns) {
1408 CallInst *ReturnedMustTail =
1409 RI->getParent()->getTerminatingMustTailCall();
1410 if (!ReturnedMustTail) {
1411 NormalReturns.push_back(RI);
1417 // Delete the old return and any preceding bitcast.
1418 BasicBlock *CurBB = RI->getParent();
1419 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
1420 RI->eraseFromParent();
1422 OldCast->eraseFromParent();
1424 // Insert a new bitcast and return with the right type.
1425 IRBuilder<> Builder(CurBB);
1426 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
1429 // Leave behind the normal returns so we can merge control flow.
1430 std::swap(Returns, NormalReturns);
1433 // If we cloned in _exactly one_ basic block, and if that block ends in a
1434 // return instruction, we splice the body of the inlined callee directly into
1435 // the calling basic block.
1436 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
1437 // Move all of the instructions right before the call.
1438 OrigBB->getInstList().splice(TheCall->getIterator(),
1439 FirstNewBlock->getInstList(),
1440 FirstNewBlock->begin(), FirstNewBlock->end());
1441 // Remove the cloned basic block.
1442 Caller->getBasicBlockList().pop_back();
1444 // If the call site was an invoke instruction, add a branch to the normal
1446 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1447 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
1448 NewBr->setDebugLoc(Returns[0]->getDebugLoc());
1451 // If the return instruction returned a value, replace uses of the call with
1452 // uses of the returned value.
1453 if (!TheCall->use_empty()) {
1454 ReturnInst *R = Returns[0];
1455 if (TheCall == R->getReturnValue())
1456 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1458 TheCall->replaceAllUsesWith(R->getReturnValue());
1460 // Since we are now done with the Call/Invoke, we can delete it.
1461 TheCall->eraseFromParent();
1463 // Since we are now done with the return instruction, delete it also.
1464 Returns[0]->eraseFromParent();
1466 // We are now done with the inlining.
1470 // Otherwise, we have the normal case, of more than one block to inline or
1471 // multiple return sites.
1473 // We want to clone the entire callee function into the hole between the
1474 // "starter" and "ender" blocks. How we accomplish this depends on whether
1475 // this is an invoke instruction or a call instruction.
1476 BasicBlock *AfterCallBB;
1477 BranchInst *CreatedBranchToNormalDest = nullptr;
1478 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1480 // Add an unconditional branch to make this look like the CallInst case...
1481 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
1483 // Split the basic block. This guarantees that no PHI nodes will have to be
1484 // updated due to new incoming edges, and make the invoke case more
1485 // symmetric to the call case.
1487 OrigBB->splitBasicBlock(CreatedBranchToNormalDest->getIterator(),
1488 CalledFunc->getName() + ".exit");
1490 } else { // It's a call
1491 // If this is a call instruction, we need to split the basic block that
1492 // the call lives in.
1494 AfterCallBB = OrigBB->splitBasicBlock(TheCall->getIterator(),
1495 CalledFunc->getName() + ".exit");
1498 // Change the branch that used to go to AfterCallBB to branch to the first
1499 // basic block of the inlined function.
1501 TerminatorInst *Br = OrigBB->getTerminator();
1502 assert(Br && Br->getOpcode() == Instruction::Br &&
1503 "splitBasicBlock broken!");
1504 Br->setOperand(0, &*FirstNewBlock);
1506 // Now that the function is correct, make it a little bit nicer. In
1507 // particular, move the basic blocks inserted from the end of the function
1508 // into the space made by splitting the source basic block.
1509 Caller->getBasicBlockList().splice(AfterCallBB->getIterator(),
1510 Caller->getBasicBlockList(), FirstNewBlock,
1513 // Handle all of the return instructions that we just cloned in, and eliminate
1514 // any users of the original call/invoke instruction.
1515 Type *RTy = CalledFunc->getReturnType();
1517 PHINode *PHI = nullptr;
1518 if (Returns.size() > 1) {
1519 // The PHI node should go at the front of the new basic block to merge all
1520 // possible incoming values.
1521 if (!TheCall->use_empty()) {
1522 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
1523 &AfterCallBB->front());
1524 // Anything that used the result of the function call should now use the
1525 // PHI node as their operand.
1526 TheCall->replaceAllUsesWith(PHI);
1529 // Loop over all of the return instructions adding entries to the PHI node
1532 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1533 ReturnInst *RI = Returns[i];
1534 assert(RI->getReturnValue()->getType() == PHI->getType() &&
1535 "Ret value not consistent in function!");
1536 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
1540 // Add a branch to the merge points and remove return instructions.
1542 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1543 ReturnInst *RI = Returns[i];
1544 BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
1545 Loc = RI->getDebugLoc();
1546 BI->setDebugLoc(Loc);
1547 RI->eraseFromParent();
1549 // We need to set the debug location to *somewhere* inside the
1550 // inlined function. The line number may be nonsensical, but the
1551 // instruction will at least be associated with the right
1553 if (CreatedBranchToNormalDest)
1554 CreatedBranchToNormalDest->setDebugLoc(Loc);
1555 } else if (!Returns.empty()) {
1556 // Otherwise, if there is exactly one return value, just replace anything
1557 // using the return value of the call with the computed value.
1558 if (!TheCall->use_empty()) {
1559 if (TheCall == Returns[0]->getReturnValue())
1560 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1562 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
1565 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
1566 BasicBlock *ReturnBB = Returns[0]->getParent();
1567 ReturnBB->replaceAllUsesWith(AfterCallBB);
1569 // Splice the code from the return block into the block that it will return
1570 // to, which contains the code that was after the call.
1571 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
1572 ReturnBB->getInstList());
1574 if (CreatedBranchToNormalDest)
1575 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
1577 // Delete the return instruction now and empty ReturnBB now.
1578 Returns[0]->eraseFromParent();
1579 ReturnBB->eraseFromParent();
1580 } else if (!TheCall->use_empty()) {
1581 // No returns, but something is using the return value of the call. Just
1583 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1586 // Since we are now done with the Call/Invoke, we can delete it.
1587 TheCall->eraseFromParent();
1589 // If we inlined any musttail calls and the original return is now
1590 // unreachable, delete it. It can only contain a bitcast and ret.
1591 if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB))
1592 AfterCallBB->eraseFromParent();
1594 // We should always be able to fold the entry block of the function into the
1595 // single predecessor of the block...
1596 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
1597 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
1599 // Splice the code entry block into calling block, right before the
1600 // unconditional branch.
1601 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
1602 OrigBB->getInstList().splice(Br->getIterator(), CalleeEntry->getInstList());
1604 // Remove the unconditional branch.
1605 OrigBB->getInstList().erase(Br);
1607 // Now we can remove the CalleeEntry block, which is now empty.
1608 Caller->getBasicBlockList().erase(CalleeEntry);
1610 // If we inserted a phi node, check to see if it has a single value (e.g. all
1611 // the entries are the same or undef). If so, remove the PHI so it doesn't
1612 // block other optimizations.
1614 auto &DL = Caller->getParent()->getDataLayout();
1615 if (Value *V = SimplifyInstruction(PHI, DL, nullptr, nullptr,
1616 &IFI.ACT->getAssumptionCache(*Caller))) {
1617 PHI->replaceAllUsesWith(V);
1618 PHI->eraseFromParent();