1 //===- InlineFunction.cpp - Code to perform function inlining -------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements inlining of a function into a call site, resolving
11 // parameters and the return value as appropriate.
12 //
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Utils/Cloning.h"
16 #include "llvm/ADT/SmallSet.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/StringExtras.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/CallGraph.h"
22 #include "llvm/Analysis/CaptureTracking.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/Attributes.h"
26 #include "llvm/IR/CallSite.h"
27 #include "llvm/IR/CFG.h"
28 #include "llvm/IR/Constants.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/DebugInfo.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/Dominators.h"
33 #include "llvm/IR/IRBuilder.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/IntrinsicInst.h"
36 #include "llvm/IR/Intrinsics.h"
37 #include "llvm/IR/MDBuilder.h"
38 #include "llvm/IR/Module.h"
39 #include "llvm/Transforms/Utils/Local.h"
40 #include "llvm/Support/CommandLine.h"
41 #include <algorithm>
42 using namespace llvm;
44 static cl::opt<bool>
45 EnableNoAliasConversion("enable-noalias-to-md-conversion", cl::init(false),
46 cl::Hidden,
47 cl::desc("Convert noalias attributes to metadata during inlining."));
49 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI,
50 bool InsertLifetime) {
51 return InlineFunction(CallSite(CI), IFI, InsertLifetime);
52 }
53 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI,
54 bool InsertLifetime) {
55 return InlineFunction(CallSite(II), IFI, InsertLifetime);
56 }
58 namespace {
59 /// A class for recording information about inlining through an invoke.
60 class InvokeInliningInfo {
61 BasicBlock *OuterResumeDest; ///< Destination of the invoke's unwind.
62 BasicBlock *InnerResumeDest; ///< Destination for the callee's resume.
63 LandingPadInst *CallerLPad; ///< LandingPadInst associated with the invoke.
64 PHINode *InnerEHValuesPHI; ///< PHI for EH values from landingpad insts.
65 SmallVector<Value*, 8> UnwindDestPHIValues;
67 public:
68 InvokeInliningInfo(InvokeInst *II)
69 : OuterResumeDest(II->getUnwindDest()), InnerResumeDest(nullptr),
70 CallerLPad(nullptr), InnerEHValuesPHI(nullptr) {
71 // If there are PHI nodes in the unwind destination block, we need to keep
72 // track of which values came into them from the invoke before removing
73 // the edge from this block.
74 llvm::BasicBlock *InvokeBB = II->getParent();
75 BasicBlock::iterator I = OuterResumeDest->begin();
76 for (; isa<PHINode>(I); ++I) {
77 // Save the value to use for this edge.
78 PHINode *PHI = cast<PHINode>(I);
79 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
80 }
82 CallerLPad = cast<LandingPadInst>(I);
83 }
85 /// getOuterResumeDest - The outer unwind destination is the target of
86 /// unwind edges introduced for calls within the inlined function.
87 BasicBlock *getOuterResumeDest() const {
88 return OuterResumeDest;
89 }
91 BasicBlock *getInnerResumeDest();
93 LandingPadInst *getLandingPadInst() const { return CallerLPad; }
95 /// forwardResume - Forward the 'resume' instruction to the caller's landing
96 /// pad block. When the landing pad block has only one predecessor, this is
97 /// a simple branch. When there is more than one predecessor, we need to
98 /// split the landing pad block after the landingpad instruction and jump
99 /// to there.
100 void forwardResume(ResumeInst *RI,
101 SmallPtrSet<LandingPadInst*, 16> &InlinedLPads);
103 /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind
104 /// destination block for the given basic block, using the values for the
105 /// original invoke's source block.
106 void addIncomingPHIValuesFor(BasicBlock *BB) const {
107 addIncomingPHIValuesForInto(BB, OuterResumeDest);
108 }
110 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const {
111 BasicBlock::iterator I = dest->begin();
112 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
113 PHINode *phi = cast<PHINode>(I);
114 phi->addIncoming(UnwindDestPHIValues[i], src);
115 }
116 }
117 };
118 }
120 /// getInnerResumeDest - Get or create a target for the branch from ResumeInsts.
121 BasicBlock *InvokeInliningInfo::getInnerResumeDest() {
122 if (InnerResumeDest) return InnerResumeDest;
124 // Split the landing pad.
125 BasicBlock::iterator SplitPoint = CallerLPad; ++SplitPoint;
126 InnerResumeDest =
127 OuterResumeDest->splitBasicBlock(SplitPoint,
128 OuterResumeDest->getName() + ".body");
130 // The number of incoming edges we expect to the inner landing pad.
131 const unsigned PHICapacity = 2;
133 // Create corresponding new PHIs for all the PHIs in the outer landing pad.
134 BasicBlock::iterator InsertPoint = InnerResumeDest->begin();
135 BasicBlock::iterator I = OuterResumeDest->begin();
136 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
137 PHINode *OuterPHI = cast<PHINode>(I);
138 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity,
139 OuterPHI->getName() + ".lpad-body",
140 InsertPoint);
141 OuterPHI->replaceAllUsesWith(InnerPHI);
142 InnerPHI->addIncoming(OuterPHI, OuterResumeDest);
143 }
145 // Create a PHI for the exception values.
146 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity,
147 "eh.lpad-body", InsertPoint);
148 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI);
149 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest);
151 // All done.
152 return InnerResumeDest;
153 }
155 /// forwardResume - Forward the 'resume' instruction to the caller's landing pad
156 /// block. When the landing pad block has only one predecessor, this is a simple
157 /// branch. When there is more than one predecessor, we need to split the
158 /// landing pad block after the landingpad instruction and jump to there.
159 void InvokeInliningInfo::forwardResume(ResumeInst *RI,
160 SmallPtrSet<LandingPadInst*, 16> &InlinedLPads) {
161 BasicBlock *Dest = getInnerResumeDest();
162 BasicBlock *Src = RI->getParent();
164 BranchInst::Create(Dest, Src);
166 // Update the PHIs in the destination. They were inserted in an order which
167 // makes this work.
168 addIncomingPHIValuesForInto(Src, Dest);
170 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src);
171 RI->eraseFromParent();
172 }
174 /// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
175 /// an invoke, we have to turn all of the calls that can throw into
176 /// invokes. This function analyze BB to see if there are any calls, and if so,
177 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
178 /// nodes in that block with the values specified in InvokeDestPHIValues.
179 static void HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
180 InvokeInliningInfo &Invoke) {
181 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
182 Instruction *I = BBI++;
184 // We only need to check for function calls: inlined invoke
185 // instructions require no special handling.
186 CallInst *CI = dyn_cast<CallInst>(I);
188 // If this call cannot unwind, don't convert it to an invoke.
189 // Inline asm calls cannot throw.
190 if (!CI || CI->doesNotThrow() || isa<InlineAsm>(CI->getCalledValue()))
191 continue;
193 // Convert this function call into an invoke instruction. First, split the
194 // basic block.
195 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
197 // Delete the unconditional branch inserted by splitBasicBlock
198 BB->getInstList().pop_back();
200 // Create the new invoke instruction.
201 ImmutableCallSite CS(CI);
202 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
203 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split,
204 Invoke.getOuterResumeDest(),
205 InvokeArgs, CI->getName(), BB);
206 II->setDebugLoc(CI->getDebugLoc());
207 II->setCallingConv(CI->getCallingConv());
208 II->setAttributes(CI->getAttributes());
210 // Make sure that anything using the call now uses the invoke! This also
211 // updates the CallGraph if present, because it uses a WeakVH.
212 CI->replaceAllUsesWith(II);
214 // Delete the original call
215 Split->getInstList().pop_front();
217 // Update any PHI nodes in the exceptional block to indicate that there is
218 // now a new entry in them.
219 Invoke.addIncomingPHIValuesFor(BB);
220 return;
221 }
222 }
224 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
225 /// in the body of the inlined function into invokes.
226 ///
227 /// II is the invoke instruction being inlined. FirstNewBlock is the first
228 /// block of the inlined code (the last block is the end of the function),
229 /// and InlineCodeInfo is information about the code that got inlined.
230 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
231 ClonedCodeInfo &InlinedCodeInfo) {
232 BasicBlock *InvokeDest = II->getUnwindDest();
234 Function *Caller = FirstNewBlock->getParent();
236 // The inlined code is currently at the end of the function, scan from the
237 // start of the inlined code to its end, checking for stuff we need to
238 // rewrite.
239 InvokeInliningInfo Invoke(II);
241 // Get all of the inlined landing pad instructions.
242 SmallPtrSet<LandingPadInst*, 16> InlinedLPads;
243 for (Function::iterator I = FirstNewBlock, E = Caller->end(); I != E; ++I)
244 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator()))
245 InlinedLPads.insert(II->getLandingPadInst());
247 // Append the clauses from the outer landing pad instruction into the inlined
248 // landing pad instructions.
249 LandingPadInst *OuterLPad = Invoke.getLandingPadInst();
250 for (SmallPtrSet<LandingPadInst*, 16>::iterator I = InlinedLPads.begin(),
251 E = InlinedLPads.end(); I != E; ++I) {
252 LandingPadInst *InlinedLPad = *I;
253 unsigned OuterNum = OuterLPad->getNumClauses();
254 InlinedLPad->reserveClauses(OuterNum);
255 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx)
256 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx));
257 if (OuterLPad->isCleanup())
258 InlinedLPad->setCleanup(true);
259 }
261 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
262 if (InlinedCodeInfo.ContainsCalls)
263 HandleCallsInBlockInlinedThroughInvoke(BB, Invoke);
265 // Forward any resumes that are remaining here.
266 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator()))
267 Invoke.forwardResume(RI, InlinedLPads);
268 }
270 // Now that everything is happy, we have one final detail. The PHI nodes in
271 // the exception destination block still have entries due to the original
272 // invoke instruction. Eliminate these entries (which might even delete the
273 // PHI node) now.
274 InvokeDest->removePredecessor(II->getParent());
275 }
277 /// CloneAliasScopeMetadata - When inlining a function that contains noalias
278 /// scope metadata, this metadata needs to be cloned so that the inlined blocks
279 /// have different "unqiue scopes" at every call site. Were this not done, then
280 /// aliasing scopes from a function inlined into a caller multiple times could
281 /// not be differentiated (and this would lead to miscompiles because the
282 /// non-aliasing property communicated by the metadata could have
283 /// call-site-specific control dependencies).
284 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) {
285 const Function *CalledFunc = CS.getCalledFunction();
286 SetVector<const MDNode *> MD;
288 // Note: We could only clone the metadata if it is already used in the
289 // caller. I'm omitting that check here because it might confuse
290 // inter-procedural alias analysis passes. We can revisit this if it becomes
291 // an efficiency or overhead problem.
293 for (Function::const_iterator I = CalledFunc->begin(), IE = CalledFunc->end();
294 I != IE; ++I)
295 for (BasicBlock::const_iterator J = I->begin(), JE = I->end(); J != JE; ++J) {
296 if (const MDNode *M = J->getMetadata(LLVMContext::MD_alias_scope))
297 MD.insert(M);
298 if (const MDNode *M = J->getMetadata(LLVMContext::MD_noalias))
299 MD.insert(M);
300 }
302 if (MD.empty())
303 return;
305 // Walk the existing metadata, adding the complete (perhaps cyclic) chain to
306 // the set.
307 SmallVector<const Value *, 16> Queue(MD.begin(), MD.end());
308 while (!Queue.empty()) {
309 const MDNode *M = cast<MDNode>(Queue.pop_back_val());
310 for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i)
311 if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i)))
312 if (MD.insert(M1))
313 Queue.push_back(M1);
314 }
316 // Now we have a complete set of all metadata in the chains used to specify
317 // the noalias scopes and the lists of those scopes.
318 SmallVector<MDNode *, 16> DummyNodes;
319 DenseMap<const MDNode *, TrackingVH<MDNode> > MDMap;
320 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
321 I != IE; ++I) {
322 MDNode *Dummy = MDNode::getTemporary(CalledFunc->getContext(),
323 ArrayRef<Value*>());
324 DummyNodes.push_back(Dummy);
325 MDMap[*I] = Dummy;
326 }
328 // Create new metadata nodes to replace the dummy nodes, replacing old
329 // metadata references with either a dummy node or an already-created new
330 // node.
331 for (SetVector<const MDNode *>::iterator I = MD.begin(), IE = MD.end();
332 I != IE; ++I) {
333 SmallVector<Value *, 4> NewOps;
334 for (unsigned i = 0, ie = (*I)->getNumOperands(); i != ie; ++i) {
335 const Value *V = (*I)->getOperand(i);
336 if (const MDNode *M = dyn_cast<MDNode>(V))
337 NewOps.push_back(MDMap[M]);
338 else
339 NewOps.push_back(const_cast<Value *>(V));
340 }
342 MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps),
343 *TempM = MDMap[*I];
345 TempM->replaceAllUsesWith(NewM);
346 }
348 // Now replace the metadata in the new inlined instructions with the
349 // repacements from the map.
350 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
351 VMI != VMIE; ++VMI) {
352 if (!VMI->second)
353 continue;
355 Instruction *NI = dyn_cast<Instruction>(VMI->second);
356 if (!NI)
357 continue;
359 if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope))
360 NI->setMetadata(LLVMContext::MD_alias_scope, MDMap[M]);
362 if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias))
363 NI->setMetadata(LLVMContext::MD_noalias, MDMap[M]);
364 }
366 // Now that everything has been replaced, delete the dummy nodes.
367 for (unsigned i = 0, ie = DummyNodes.size(); i != ie; ++i)
368 MDNode::deleteTemporary(DummyNodes[i]);
369 }
371 /// AddAliasScopeMetadata - If the inlined function has noalias arguments, then
372 /// add new alias scopes for each noalias argument, tag the mapped noalias
373 /// parameters with noalias metadata specifying the new scope, and tag all
374 /// non-derived loads, stores and memory intrinsics with the new alias scopes.
375 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
376 const DataLayout *DL) {
377 if (!EnableNoAliasConversion)
378 return;
380 const Function *CalledFunc = CS.getCalledFunction();
381 SmallVector<const Argument *, 4> NoAliasArgs;
383 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
384 E = CalledFunc->arg_end(); I != E; ++I) {
385 if (I->hasNoAliasAttr() && !I->hasNUses(0))
386 NoAliasArgs.push_back(I);
387 }
389 if (NoAliasArgs.empty())
390 return;
392 // To do a good job, if a noalias variable is captured, we need to know if
393 // the capture point dominates the particular use we're considering.
394 DominatorTree DT;
395 DT.recalculate(const_cast<Function&>(*CalledFunc));
397 // noalias indicates that pointer values based on the argument do not alias
398 // pointer values which are not based on it. So we add a new "scope" for each
399 // noalias function argument. Accesses using pointers based on that argument
400 // become part of that alias scope, accesses using pointers not based on that
401 // argument are tagged as noalias with that scope.
403 DenseMap<const Argument *, MDNode *> NewScopes;
404 MDBuilder MDB(CalledFunc->getContext());
406 // Create a new scope domain for this function.
407 MDNode *NewDomain =
408 MDB.createAnonymousAliasScopeDomain(CalledFunc->getName());
409 for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) {
410 const Argument *A = NoAliasArgs[i];
412 std::string Name = CalledFunc->getName();
413 if (A->hasName()) {
414 Name += ": %";
415 Name += A->getName();
416 } else {
417 Name += ": argument ";
418 Name += utostr(i);
419 }
421 // Note: We always create a new anonymous root here. This is true regardless
422 // of the linkage of the callee because the aliasing "scope" is not just a
423 // property of the callee, but also all control dependencies in the caller.
424 MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
425 NewScopes.insert(std::make_pair(A, NewScope));
426 }
428 // Iterate over all new instructions in the map; for all memory-access
429 // instructions, add the alias scope metadata.
430 for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
431 VMI != VMIE; ++VMI) {
432 if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) {
433 if (!VMI->second)
434 continue;
436 Instruction *NI = dyn_cast<Instruction>(VMI->second);
437 if (!NI)
438 continue;
440 SmallVector<const Value *, 2> PtrArgs;
442 if (const LoadInst *LI = dyn_cast<LoadInst>(I))
443 PtrArgs.push_back(LI->getPointerOperand());
444 else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
445 PtrArgs.push_back(SI->getPointerOperand());
446 else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
447 PtrArgs.push_back(VAAI->getPointerOperand());
448 else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
449 PtrArgs.push_back(CXI->getPointerOperand());
450 else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
451 PtrArgs.push_back(RMWI->getPointerOperand());
452 else if (const MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) {
453 PtrArgs.push_back(MI->getRawDest());
454 if (const MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
455 PtrArgs.push_back(MTI->getRawSource());
456 }
458 // If we found no pointers, then this instruction is not suitable for
459 // pairing with an instruction to receive aliasing metadata.
460 // Simplification during cloning could make this happen, and skip these
461 // cases for now.
462 if (PtrArgs.empty())
463 continue;
465 // It is possible that there is only one underlying object, but you
466 // need to go through several PHIs to see it, and thus could be
467 // repeated in the Objects list.
468 SmallPtrSet<const Value *, 4> ObjSet;
469 SmallVector<Value *, 4> Scopes, NoAliases;
471 SmallSetVector<const Argument *, 4> NAPtrArgs;
472 for (unsigned i = 0, ie = PtrArgs.size(); i != ie; ++i) {
473 SmallVector<Value *, 4> Objects;
474 GetUnderlyingObjects(const_cast<Value*>(PtrArgs[i]),
475 Objects, DL, /* MaxLookup = */ 0);
477 for (Value *O : Objects)
478 ObjSet.insert(O);
479 }
481 // Figure out if we're derived from anyhing that is not a noalias
482 // argument.
483 bool CanDeriveViaCapture = false;
484 for (const Value *V : ObjSet)
485 if (!isIdentifiedFunctionLocal(const_cast<Value*>(V))) {
486 CanDeriveViaCapture = true;
487 break;
488 }
490 // First, we want to figure out all of the sets with which we definitely
491 // don't alias. Iterate over all noalias set, and add those for which:
492 // 1. The noalias argument is not in the set of objects from which we
493 // definitely derive.
494 // 2. The noalias argument has not yet been captured.
495 for (const Argument *A : NoAliasArgs) {
496 if (!ObjSet.count(A) && (!CanDeriveViaCapture ||
497 A->hasNoCaptureAttr() ||
498 !PointerMayBeCapturedBefore(A,
499 /* ReturnCaptures */ false,
500 /* StoreCaptures */ false, I, &DT)))
501 NoAliases.push_back(NewScopes[A]);
502 }
504 if (!NoAliases.empty())
505 NI->setMetadata(LLVMContext::MD_noalias, MDNode::concatenate(
506 NI->getMetadata(LLVMContext::MD_noalias),
507 MDNode::get(CalledFunc->getContext(), NoAliases)));
508 // Next, we want to figure out all of the sets to which we might belong.
509 // We might below to a set if:
510 // 1. The noalias argument is in the set of underlying objects
511 // or
512 // 2. There is some non-noalias argument in our list and the no-alias
513 // argument has been captured.
515 for (const Argument *A : NoAliasArgs) {
516 if (ObjSet.count(A) || (CanDeriveViaCapture &&
517 PointerMayBeCapturedBefore(A,
518 /* ReturnCaptures */ false,
519 /* StoreCaptures */ false,
520 I, &DT)))
521 Scopes.push_back(NewScopes[A]);
522 }
524 if (!Scopes.empty())
525 NI->setMetadata(LLVMContext::MD_alias_scope, MDNode::concatenate(
526 NI->getMetadata(LLVMContext::MD_alias_scope),
527 MDNode::get(CalledFunc->getContext(), Scopes)));
528 }
529 }
530 }
532 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
533 /// into the caller, update the specified callgraph to reflect the changes we
534 /// made. Note that it's possible that not all code was copied over, so only
535 /// some edges of the callgraph may remain.
536 static void UpdateCallGraphAfterInlining(CallSite CS,
537 Function::iterator FirstNewBlock,
538 ValueToValueMapTy &VMap,
539 InlineFunctionInfo &IFI) {
540 CallGraph &CG = *IFI.CG;
541 const Function *Caller = CS.getInstruction()->getParent()->getParent();
542 const Function *Callee = CS.getCalledFunction();
543 CallGraphNode *CalleeNode = CG[Callee];
544 CallGraphNode *CallerNode = CG[Caller];
546 // Since we inlined some uninlined call sites in the callee into the caller,
547 // add edges from the caller to all of the callees of the callee.
548 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
550 // Consider the case where CalleeNode == CallerNode.
551 CallGraphNode::CalledFunctionsVector CallCache;
552 if (CalleeNode == CallerNode) {
553 CallCache.assign(I, E);
554 I = CallCache.begin();
555 E = CallCache.end();
556 }
558 for (; I != E; ++I) {
559 const Value *OrigCall = I->first;
561 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
562 // Only copy the edge if the call was inlined!
563 if (VMI == VMap.end() || VMI->second == nullptr)
564 continue;
566 // If the call was inlined, but then constant folded, there is no edge to
567 // add. Check for this case.
568 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
569 if (!NewCall) continue;
571 // Remember that this call site got inlined for the client of
572 // InlineFunction.
573 IFI.InlinedCalls.push_back(NewCall);
575 // It's possible that inlining the callsite will cause it to go from an
576 // indirect to a direct call by resolving a function pointer. If this
577 // happens, set the callee of the new call site to a more precise
578 // destination. This can also happen if the call graph node of the caller
579 // was just unnecessarily imprecise.
580 if (!I->second->getFunction())
581 if (Function *F = CallSite(NewCall).getCalledFunction()) {
582 // Indirect call site resolved to direct call.
583 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
585 continue;
586 }
588 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
589 }
591 // Update the call graph by deleting the edge from Callee to Caller. We must
592 // do this after the loop above in case Caller and Callee are the same.
593 CallerNode->removeCallEdgeFor(CS);
594 }
596 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
597 BasicBlock *InsertBlock,
598 InlineFunctionInfo &IFI) {
599 LLVMContext &Context = Src->getContext();
600 Type *VoidPtrTy = Type::getInt8PtrTy(Context);
601 Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
602 Type *Tys[3] = { VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context) };
603 Function *MemCpyFn = Intrinsic::getDeclaration(M, Intrinsic::memcpy, Tys);
604 IRBuilder<> builder(InsertBlock->begin());
605 Value *DstCast = builder.CreateBitCast(Dst, VoidPtrTy, "tmp");
606 Value *SrcCast = builder.CreateBitCast(Src, VoidPtrTy, "tmp");
608 Value *Size;
609 if (IFI.DL == nullptr)
610 Size = ConstantExpr::getSizeOf(AggTy);
611 else
612 Size = ConstantInt::get(Type::getInt64Ty(Context),
613 IFI.DL->getTypeStoreSize(AggTy));
615 // Always generate a memcpy of alignment 1 here because we don't know
616 // the alignment of the src pointer. Other optimizations can infer
617 // better alignment.
618 Value *CallArgs[] = {
619 DstCast, SrcCast, Size,
620 ConstantInt::get(Type::getInt32Ty(Context), 1),
621 ConstantInt::getFalse(Context) // isVolatile
622 };
623 builder.CreateCall(MemCpyFn, CallArgs);
624 }
626 /// HandleByValArgument - When inlining a call site that has a byval argument,
627 /// we have to make the implicit memcpy explicit by adding it.
628 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
629 const Function *CalledFunc,
630 InlineFunctionInfo &IFI,
631 unsigned ByValAlignment) {
632 PointerType *ArgTy = cast<PointerType>(Arg->getType());
633 Type *AggTy = ArgTy->getElementType();
635 // If the called function is readonly, then it could not mutate the caller's
636 // copy of the byval'd memory. In this case, it is safe to elide the copy and
637 // temporary.
638 if (CalledFunc->onlyReadsMemory()) {
639 // If the byval argument has a specified alignment that is greater than the
640 // passed in pointer, then we either have to round up the input pointer or
641 // give up on this transformation.
642 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
643 return Arg;
645 // If the pointer is already known to be sufficiently aligned, or if we can
646 // round it up to a larger alignment, then we don't need a temporary.
647 if (getOrEnforceKnownAlignment(Arg, ByValAlignment,
648 IFI.DL) >= ByValAlignment)
649 return Arg;
651 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
652 // for code quality, but rarely happens and is required for correctness.
653 }
655 // Create the alloca. If we have DataLayout, use nice alignment.
656 unsigned Align = 1;
657 if (IFI.DL)
658 Align = IFI.DL->getPrefTypeAlignment(AggTy);
660 // If the byval had an alignment specified, we *must* use at least that
661 // alignment, as it is required by the byval argument (and uses of the
662 // pointer inside the callee).
663 Align = std::max(Align, ByValAlignment);
665 Function *Caller = TheCall->getParent()->getParent();
667 Value *NewAlloca = new AllocaInst(AggTy, nullptr, Align, Arg->getName(),
668 &*Caller->begin()->begin());
669 IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
671 // Uses of the argument in the function should use our new alloca
672 // instead.
673 return NewAlloca;
674 }
676 // isUsedByLifetimeMarker - Check whether this Value is used by a lifetime
677 // intrinsic.
678 static bool isUsedByLifetimeMarker(Value *V) {
679 for (User *U : V->users()) {
680 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
681 switch (II->getIntrinsicID()) {
682 default: break;
683 case Intrinsic::lifetime_start:
684 case Intrinsic::lifetime_end:
685 return true;
686 }
687 }
688 }
689 return false;
690 }
692 // hasLifetimeMarkers - Check whether the given alloca already has
693 // lifetime.start or lifetime.end intrinsics.
694 static bool hasLifetimeMarkers(AllocaInst *AI) {
695 Type *Ty = AI->getType();
696 Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
697 Ty->getPointerAddressSpace());
698 if (Ty == Int8PtrTy)
699 return isUsedByLifetimeMarker(AI);
701 // Do a scan to find all the casts to i8*.
702 for (User *U : AI->users()) {
703 if (U->getType() != Int8PtrTy) continue;
704 if (U->stripPointerCasts() != AI) continue;
705 if (isUsedByLifetimeMarker(U))
706 return true;
707 }
708 return false;
709 }
711 /// updateInlinedAtInfo - Helper function used by fixupLineNumbers to
712 /// recursively update InlinedAtEntry of a DebugLoc.
713 static DebugLoc updateInlinedAtInfo(const DebugLoc &DL,
714 const DebugLoc &InlinedAtDL,
715 LLVMContext &Ctx) {
716 if (MDNode *IA = DL.getInlinedAt(Ctx)) {
717 DebugLoc NewInlinedAtDL
718 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx);
719 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
720 NewInlinedAtDL.getAsMDNode(Ctx));
721 }
723 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
724 InlinedAtDL.getAsMDNode(Ctx));
725 }
727 /// fixupLineNumbers - Update inlined instructions' line numbers to
728 /// to encode location where these instructions are inlined.
729 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
730 Instruction *TheCall) {
731 DebugLoc TheCallDL = TheCall->getDebugLoc();
732 if (TheCallDL.isUnknown())
733 return;
735 for (; FI != Fn->end(); ++FI) {
736 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
737 BI != BE; ++BI) {
738 DebugLoc DL = BI->getDebugLoc();
739 if (DL.isUnknown()) {
740 // If the inlined instruction has no line number, make it look as if it
741 // originates from the call location. This is important for
742 // ((__always_inline__, __nodebug__)) functions which must use caller
743 // location for all instructions in their function body.
744 BI->setDebugLoc(TheCallDL);
745 } else {
746 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext()));
747 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) {
748 LLVMContext &Ctx = BI->getContext();
749 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
750 DVI->setOperand(2, createInlinedVariable(DVI->getVariable(),
751 InlinedAt, Ctx));
752 }
753 }
754 }
755 }
756 }
758 /// InlineFunction - This function inlines the called function into the basic
759 /// block of the caller. This returns false if it is not possible to inline
760 /// this call. The program is still in a well defined state if this occurs
761 /// though.
762 ///
763 /// Note that this only does one level of inlining. For example, if the
764 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
765 /// exists in the instruction stream. Similarly this will inline a recursive
766 /// function by one level.
767 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
768 bool InsertLifetime) {
769 Instruction *TheCall = CS.getInstruction();
770 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
771 "Instruction not in function!");
773 // If IFI has any state in it, zap it before we fill it in.
774 IFI.reset();
776 const Function *CalledFunc = CS.getCalledFunction();
777 if (!CalledFunc || // Can't inline external function or indirect
778 CalledFunc->isDeclaration() || // call, or call to a vararg function!
779 CalledFunc->getFunctionType()->isVarArg()) return false;
781 // If the call to the callee cannot throw, set the 'nounwind' flag on any
782 // calls that we inline.
783 bool MarkNoUnwind = CS.doesNotThrow();
785 BasicBlock *OrigBB = TheCall->getParent();
786 Function *Caller = OrigBB->getParent();
788 // GC poses two hazards to inlining, which only occur when the callee has GC:
789 // 1. If the caller has no GC, then the callee's GC must be propagated to the
790 // caller.
791 // 2. If the caller has a differing GC, it is invalid to inline.
792 if (CalledFunc->hasGC()) {
793 if (!Caller->hasGC())
794 Caller->setGC(CalledFunc->getGC());
795 else if (CalledFunc->getGC() != Caller->getGC())
796 return false;
797 }
799 // Get the personality function from the callee if it contains a landing pad.
800 Value *CalleePersonality = nullptr;
801 for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end();
802 I != E; ++I)
803 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
804 const BasicBlock *BB = II->getUnwindDest();
805 const LandingPadInst *LP = BB->getLandingPadInst();
806 CalleePersonality = LP->getPersonalityFn();
807 break;
808 }
810 // Find the personality function used by the landing pads of the caller. If it
811 // exists, then check to see that it matches the personality function used in
812 // the callee.
813 if (CalleePersonality) {
814 for (Function::const_iterator I = Caller->begin(), E = Caller->end();
815 I != E; ++I)
816 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
817 const BasicBlock *BB = II->getUnwindDest();
818 const LandingPadInst *LP = BB->getLandingPadInst();
820 // If the personality functions match, then we can perform the
821 // inlining. Otherwise, we can't inline.
822 // TODO: This isn't 100% true. Some personality functions are proper
823 // supersets of others and can be used in place of the other.
824 if (LP->getPersonalityFn() != CalleePersonality)
825 return false;
827 break;
828 }
829 }
831 // Get an iterator to the last basic block in the function, which will have
832 // the new function inlined after it.
833 Function::iterator LastBlock = &Caller->back();
835 // Make sure to capture all of the return instructions from the cloned
836 // function.
837 SmallVector<ReturnInst*, 8> Returns;
838 ClonedCodeInfo InlinedFunctionInfo;
839 Function::iterator FirstNewBlock;
841 { // Scope to destroy VMap after cloning.
842 ValueToValueMapTy VMap;
843 // Keep a list of pair (dst, src) to emit byval initializations.
844 SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
846 assert(CalledFunc->arg_size() == CS.arg_size() &&
847 "No varargs calls can be inlined!");
849 // Calculate the vector of arguments to pass into the function cloner, which
850 // matches up the formal to the actual argument values.
851 CallSite::arg_iterator AI = CS.arg_begin();
852 unsigned ArgNo = 0;
853 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
854 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
855 Value *ActualArg = *AI;
857 // When byval arguments actually inlined, we need to make the copy implied
858 // by them explicit. However, we don't do this if the callee is readonly
859 // or readnone, because the copy would be unneeded: the callee doesn't
860 // modify the struct.
861 if (CS.isByValArgument(ArgNo)) {
862 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
863 CalledFunc->getParamAlignment(ArgNo+1));
864 if (ActualArg != *AI)
865 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
866 }
868 VMap[I] = ActualArg;
869 }
871 // We want the inliner to prune the code as it copies. We would LOVE to
872 // have no dead or constant instructions leftover after inlining occurs
873 // (which can happen, e.g., because an argument was constant), but we'll be
874 // happy with whatever the cloner can do.
875 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
876 /*ModuleLevelChanges=*/false, Returns, ".i",
877 &InlinedFunctionInfo, IFI.DL, TheCall);
879 // Remember the first block that is newly cloned over.
880 FirstNewBlock = LastBlock; ++FirstNewBlock;
882 // Inject byval arguments initialization.
883 for (std::pair<Value*, Value*> &Init : ByValInit)
884 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
885 FirstNewBlock, IFI);
887 // Update the callgraph if requested.
888 if (IFI.CG)
889 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
891 // Update inlined instructions' line number information.
892 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
894 // Clone existing noalias metadata if necessary.
895 CloneAliasScopeMetadata(CS, VMap);
897 // Add noalias metadata if necessary.
898 AddAliasScopeMetadata(CS, VMap, IFI.DL);
899 }
901 // If there are any alloca instructions in the block that used to be the entry
902 // block for the callee, move them to the entry block of the caller. First
903 // calculate which instruction they should be inserted before. We insert the
904 // instructions at the end of the current alloca list.
905 {
906 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
907 for (BasicBlock::iterator I = FirstNewBlock->begin(),
908 E = FirstNewBlock->end(); I != E; ) {
909 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
910 if (!AI) continue;
912 // If the alloca is now dead, remove it. This often occurs due to code
913 // specialization.
914 if (AI->use_empty()) {
915 AI->eraseFromParent();
916 continue;
917 }
919 if (!isa<Constant>(AI->getArraySize()))
920 continue;
922 // Keep track of the static allocas that we inline into the caller.
923 IFI.StaticAllocas.push_back(AI);
925 // Scan for the block of allocas that we can move over, and move them
926 // all at once.
927 while (isa<AllocaInst>(I) &&
928 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
929 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
930 ++I;
931 }
933 // Transfer all of the allocas over in a block. Using splice means
934 // that the instructions aren't removed from the symbol table, then
935 // reinserted.
936 Caller->getEntryBlock().getInstList().splice(InsertPoint,
937 FirstNewBlock->getInstList(),
938 AI, I);
939 }
940 }
942 bool InlinedMustTailCalls = false;
943 if (InlinedFunctionInfo.ContainsCalls) {
944 CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
945 if (CallInst *CI = dyn_cast<CallInst>(TheCall))
946 CallSiteTailKind = CI->getTailCallKind();
948 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
949 ++BB) {
950 for (Instruction &I : *BB) {
951 CallInst *CI = dyn_cast<CallInst>(&I);
952 if (!CI)
953 continue;
955 // We need to reduce the strength of any inlined tail calls. For
956 // musttail, we have to avoid introducing potential unbounded stack
957 // growth. For example, if functions 'f' and 'g' are mutually recursive
958 // with musttail, we can inline 'g' into 'f' so long as we preserve
959 // musttail on the cloned call to 'f'. If either the inlined call site
960 // or the cloned call site is *not* musttail, the program already has
961 // one frame of stack growth, so it's safe to remove musttail. Here is
962 // a table of example transformations:
963 //
964 // f -> musttail g -> musttail f ==> f -> musttail f
965 // f -> musttail g -> tail f ==> f -> tail f
966 // f -> g -> musttail f ==> f -> f
967 // f -> g -> tail f ==> f -> f
968 CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
969 ChildTCK = std::min(CallSiteTailKind, ChildTCK);
970 CI->setTailCallKind(ChildTCK);
971 InlinedMustTailCalls |= CI->isMustTailCall();
973 // Calls inlined through a 'nounwind' call site should be marked
974 // 'nounwind'.
975 if (MarkNoUnwind)
976 CI->setDoesNotThrow();
977 }
978 }
979 }
981 // Leave lifetime markers for the static alloca's, scoping them to the
982 // function we just inlined.
983 if (InsertLifetime && !IFI.StaticAllocas.empty()) {
984 IRBuilder<> builder(FirstNewBlock->begin());
985 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
986 AllocaInst *AI = IFI.StaticAllocas[ai];
988 // If the alloca is already scoped to something smaller than the whole
989 // function then there's no need to add redundant, less accurate markers.
990 if (hasLifetimeMarkers(AI))
991 continue;
993 // Try to determine the size of the allocation.
994 ConstantInt *AllocaSize = nullptr;
995 if (ConstantInt *AIArraySize =
996 dyn_cast<ConstantInt>(AI->getArraySize())) {
997 if (IFI.DL) {
998 Type *AllocaType = AI->getAllocatedType();
999 uint64_t AllocaTypeSize = IFI.DL->getTypeAllocSize(AllocaType);
1000 uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
1001 assert(AllocaArraySize > 0 && "array size of AllocaInst is zero");
1002 // Check that array size doesn't saturate uint64_t and doesn't
1003 // overflow when it's multiplied by type size.
1004 if (AllocaArraySize != ~0ULL &&
1005 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
1006 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
1007 AllocaArraySize * AllocaTypeSize);
1008 }
1009 }
1010 }
1012 builder.CreateLifetimeStart(AI, AllocaSize);
1013 for (ReturnInst *RI : Returns) {
1014 // Don't insert llvm.lifetime.end calls between a musttail call and a
1015 // return. The return kills all local allocas.
1016 if (InlinedMustTailCalls &&
1017 RI->getParent()->getTerminatingMustTailCall())
1018 continue;
1019 IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
1020 }
1021 }
1022 }
1024 // If the inlined code contained dynamic alloca instructions, wrap the inlined
1025 // code with llvm.stacksave/llvm.stackrestore intrinsics.
1026 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
1027 Module *M = Caller->getParent();
1028 // Get the two intrinsics we care about.
1029 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
1030 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
1032 // Insert the llvm.stacksave.
1033 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
1034 .CreateCall(StackSave, "savedstack");
1036 // Insert a call to llvm.stackrestore before any return instructions in the
1037 // inlined function.
1038 for (ReturnInst *RI : Returns) {
1039 // Don't insert llvm.stackrestore calls between a musttail call and a
1040 // return. The return will restore the stack pointer.
1041 if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
1042 continue;
1043 IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
1044 }
1045 }
1047 // If we are inlining for an invoke instruction, we must make sure to rewrite
1048 // any call instructions into invoke instructions.
1049 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
1050 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
1052 // Handle any inlined musttail call sites. In order for a new call site to be
1053 // musttail, the source of the clone and the inlined call site must have been
1054 // musttail. Therefore it's safe to return without merging control into the
1055 // phi below.
1056 if (InlinedMustTailCalls) {
1057 // Check if we need to bitcast the result of any musttail calls.
1058 Type *NewRetTy = Caller->getReturnType();
1059 bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
1061 // Handle the returns preceded by musttail calls separately.
1062 SmallVector<ReturnInst *, 8> NormalReturns;
1063 for (ReturnInst *RI : Returns) {
1064 CallInst *ReturnedMustTail =
1065 RI->getParent()->getTerminatingMustTailCall();
1066 if (!ReturnedMustTail) {
1067 NormalReturns.push_back(RI);
1068 continue;
1069 }
1070 if (!NeedBitCast)
1071 continue;
1073 // Delete the old return and any preceding bitcast.
1074 BasicBlock *CurBB = RI->getParent();
1075 auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
1076 RI->eraseFromParent();
1077 if (OldCast)
1078 OldCast->eraseFromParent();
1080 // Insert a new bitcast and return with the right type.
1081 IRBuilder<> Builder(CurBB);
1082 Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
1083 }
1085 // Leave behind the normal returns so we can merge control flow.
1086 std::swap(Returns, NormalReturns);
1087 }
1089 // If we cloned in _exactly one_ basic block, and if that block ends in a
1090 // return instruction, we splice the body of the inlined callee directly into
1091 // the calling basic block.
1092 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
1093 // Move all of the instructions right before the call.
1094 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
1095 FirstNewBlock->begin(), FirstNewBlock->end());
1096 // Remove the cloned basic block.
1097 Caller->getBasicBlockList().pop_back();
1099 // If the call site was an invoke instruction, add a branch to the normal
1100 // destination.
1101 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1102 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
1103 NewBr->setDebugLoc(Returns[0]->getDebugLoc());
1104 }
1106 // If the return instruction returned a value, replace uses of the call with
1107 // uses of the returned value.
1108 if (!TheCall->use_empty()) {
1109 ReturnInst *R = Returns[0];
1110 if (TheCall == R->getReturnValue())
1111 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1112 else
1113 TheCall->replaceAllUsesWith(R->getReturnValue());
1114 }
1115 // Since we are now done with the Call/Invoke, we can delete it.
1116 TheCall->eraseFromParent();
1118 // Since we are now done with the return instruction, delete it also.
1119 Returns[0]->eraseFromParent();
1121 // We are now done with the inlining.
1122 return true;
1123 }
1125 // Otherwise, we have the normal case, of more than one block to inline or
1126 // multiple return sites.
1128 // We want to clone the entire callee function into the hole between the
1129 // "starter" and "ender" blocks. How we accomplish this depends on whether
1130 // this is an invoke instruction or a call instruction.
1131 BasicBlock *AfterCallBB;
1132 BranchInst *CreatedBranchToNormalDest = nullptr;
1133 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1135 // Add an unconditional branch to make this look like the CallInst case...
1136 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
1138 // Split the basic block. This guarantees that no PHI nodes will have to be
1139 // updated due to new incoming edges, and make the invoke case more
1140 // symmetric to the call case.
1141 AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest,
1142 CalledFunc->getName()+".exit");
1144 } else { // It's a call
1145 // If this is a call instruction, we need to split the basic block that
1146 // the call lives in.
1147 //
1148 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
1149 CalledFunc->getName()+".exit");
1150 }
1152 // Change the branch that used to go to AfterCallBB to branch to the first
1153 // basic block of the inlined function.
1154 //
1155 TerminatorInst *Br = OrigBB->getTerminator();
1156 assert(Br && Br->getOpcode() == Instruction::Br &&
1157 "splitBasicBlock broken!");
1158 Br->setOperand(0, FirstNewBlock);
1161 // Now that the function is correct, make it a little bit nicer. In
1162 // particular, move the basic blocks inserted from the end of the function
1163 // into the space made by splitting the source basic block.
1164 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
1165 FirstNewBlock, Caller->end());
1167 // Handle all of the return instructions that we just cloned in, and eliminate
1168 // any users of the original call/invoke instruction.
1169 Type *RTy = CalledFunc->getReturnType();
1171 PHINode *PHI = nullptr;
1172 if (Returns.size() > 1) {
1173 // The PHI node should go at the front of the new basic block to merge all
1174 // possible incoming values.
1175 if (!TheCall->use_empty()) {
1176 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
1177 AfterCallBB->begin());
1178 // Anything that used the result of the function call should now use the
1179 // PHI node as their operand.
1180 TheCall->replaceAllUsesWith(PHI);
1181 }
1183 // Loop over all of the return instructions adding entries to the PHI node
1184 // as appropriate.
1185 if (PHI) {
1186 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1187 ReturnInst *RI = Returns[i];
1188 assert(RI->getReturnValue()->getType() == PHI->getType() &&
1189 "Ret value not consistent in function!");
1190 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
1191 }
1192 }
1195 // Add a branch to the merge points and remove return instructions.
1196 DebugLoc Loc;
1197 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1198 ReturnInst *RI = Returns[i];
1199 BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
1200 Loc = RI->getDebugLoc();
1201 BI->setDebugLoc(Loc);
1202 RI->eraseFromParent();
1203 }
1204 // We need to set the debug location to *somewhere* inside the
1205 // inlined function. The line number may be nonsensical, but the
1206 // instruction will at least be associated with the right
1207 // function.
1208 if (CreatedBranchToNormalDest)
1209 CreatedBranchToNormalDest->setDebugLoc(Loc);
1210 } else if (!Returns.empty()) {
1211 // Otherwise, if there is exactly one return value, just replace anything
1212 // using the return value of the call with the computed value.
1213 if (!TheCall->use_empty()) {
1214 if (TheCall == Returns[0]->getReturnValue())
1215 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1216 else
1217 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
1218 }
1220 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
1221 BasicBlock *ReturnBB = Returns[0]->getParent();
1222 ReturnBB->replaceAllUsesWith(AfterCallBB);
1224 // Splice the code from the return block into the block that it will return
1225 // to, which contains the code that was after the call.
1226 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
1227 ReturnBB->getInstList());
1229 if (CreatedBranchToNormalDest)
1230 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
1232 // Delete the return instruction now and empty ReturnBB now.
1233 Returns[0]->eraseFromParent();
1234 ReturnBB->eraseFromParent();
1235 } else if (!TheCall->use_empty()) {
1236 // No returns, but something is using the return value of the call. Just
1237 // nuke the result.
1238 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1239 }
1241 // Since we are now done with the Call/Invoke, we can delete it.
1242 TheCall->eraseFromParent();
1244 // If we inlined any musttail calls and the original return is now
1245 // unreachable, delete it. It can only contain a bitcast and ret.
1246 if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB))
1247 AfterCallBB->eraseFromParent();
1249 // We should always be able to fold the entry block of the function into the
1250 // single predecessor of the block...
1251 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
1252 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
1254 // Splice the code entry block into calling block, right before the
1255 // unconditional branch.
1256 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
1257 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
1259 // Remove the unconditional branch.
1260 OrigBB->getInstList().erase(Br);
1262 // Now we can remove the CalleeEntry block, which is now empty.
1263 Caller->getBasicBlockList().erase(CalleeEntry);
1265 // If we inserted a phi node, check to see if it has a single value (e.g. all
1266 // the entries are the same or undef). If so, remove the PHI so it doesn't
1267 // block other optimizations.
1268 if (PHI) {
1269 if (Value *V = SimplifyInstruction(PHI, IFI.DL)) {
1270 PHI->replaceAllUsesWith(V);
1271 PHI->eraseFromParent();
1272 }
1273 }
1275 return true;
1276 }