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/Constants.h"
17 #include "llvm/DerivedTypes.h"
18 #include "llvm/Module.h"
19 #include "llvm/Instructions.h"
20 #include "llvm/IntrinsicInst.h"
21 #include "llvm/Intrinsics.h"
22 #include "llvm/Attributes.h"
23 #include "llvm/Analysis/CallGraph.h"
24 #include "llvm/Analysis/DebugInfo.h"
25 #include "llvm/Analysis/InstructionSimplify.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Transforms/Utils/Local.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/ADT/StringExtras.h"
30 #include "llvm/Support/CallSite.h"
31 #include "llvm/Support/IRBuilder.h"
34 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI) {
35 return InlineFunction(CallSite(CI), IFI);
37 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI) {
38 return InlineFunction(CallSite(II), IFI);
42 /// A class for recording information about inlining through an invoke.
43 class InvokeInliningInfo {
44 BasicBlock *OuterResumeDest; //< Destination of the invoke's unwind.
45 BasicBlock *InnerResumeDest; //< Destination for the callee's resume.
46 LandingPadInst *CallerLPad; //< LandingPadInst associated with the invoke.
47 PHINode *InnerEHValuesPHI; //< PHI for EH values from landingpad insts.
48 SmallVector<Value*, 8> UnwindDestPHIValues;
51 InvokeInliningInfo(InvokeInst *II)
52 : OuterResumeDest(II->getUnwindDest()), InnerResumeDest(0),
53 CallerLPad(0), InnerEHValuesPHI(0) {
54 // If there are PHI nodes in the unwind destination block, we need to keep
55 // track of which values came into them from the invoke before removing
56 // the edge from this block.
57 llvm::BasicBlock *InvokeBB = II->getParent();
58 BasicBlock::iterator I = OuterResumeDest->begin();
59 for (; isa<PHINode>(I); ++I) {
60 // Save the value to use for this edge.
61 PHINode *PHI = cast<PHINode>(I);
62 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
65 CallerLPad = cast<LandingPadInst>(I);
68 /// getOuterResumeDest - The outer unwind destination is the target of
69 /// unwind edges introduced for calls within the inlined function.
70 BasicBlock *getOuterResumeDest() const {
71 return OuterResumeDest;
74 BasicBlock *getInnerResumeDest();
76 LandingPadInst *getLandingPadInst() const { return CallerLPad; }
78 /// forwardResume - Forward the 'resume' instruction to the caller's landing
79 /// pad block. When the landing pad block has only one predecessor, this is
80 /// a simple branch. When there is more than one predecessor, we need to
81 /// split the landing pad block after the landingpad instruction and jump
83 void forwardResume(ResumeInst *RI);
85 /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind
86 /// destination block for the given basic block, using the values for the
87 /// original invoke's source block.
88 void addIncomingPHIValuesFor(BasicBlock *BB) const {
89 addIncomingPHIValuesForInto(BB, OuterResumeDest);
92 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const {
93 BasicBlock::iterator I = dest->begin();
94 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
95 PHINode *phi = cast<PHINode>(I);
96 phi->addIncoming(UnwindDestPHIValues[i], src);
102 /// getInnerResumeDest - Get or create a target for the branch from ResumeInsts.
103 BasicBlock *InvokeInliningInfo::getInnerResumeDest() {
104 if (InnerResumeDest) return InnerResumeDest;
106 // Split the landing pad.
107 BasicBlock::iterator SplitPoint = CallerLPad; ++SplitPoint;
109 OuterResumeDest->splitBasicBlock(SplitPoint,
110 OuterResumeDest->getName() + ".body");
112 // The number of incoming edges we expect to the inner landing pad.
113 const unsigned PHICapacity = 2;
115 // Create corresponding new PHIs for all the PHIs in the outer landing pad.
116 BasicBlock::iterator InsertPoint = InnerResumeDest->begin();
117 BasicBlock::iterator I = OuterResumeDest->begin();
118 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
119 PHINode *OuterPHI = cast<PHINode>(I);
120 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity,
121 OuterPHI->getName() + ".lpad-body",
123 OuterPHI->replaceAllUsesWith(InnerPHI);
124 InnerPHI->addIncoming(OuterPHI, OuterResumeDest);
127 // Create a PHI for the exception values.
128 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity,
129 "eh.lpad-body", InsertPoint);
130 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI);
131 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest);
134 return InnerResumeDest;
137 /// forwardResume - Forward the 'resume' instruction to the caller's landing pad
138 /// block. When the landing pad block has only one predecessor, this is a simple
139 /// branch. When there is more than one predecessor, we need to split the
140 /// landing pad block after the landingpad instruction and jump to there.
141 void InvokeInliningInfo::forwardResume(ResumeInst *RI) {
142 BasicBlock *Dest = getInnerResumeDest();
143 BasicBlock *Src = RI->getParent();
145 BranchInst::Create(Dest, Src);
147 // Update the PHIs in the destination. They were inserted in an order which
149 addIncomingPHIValuesForInto(Src, Dest);
151 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src);
152 RI->eraseFromParent();
155 /// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
156 /// an invoke, we have to turn all of the calls that can throw into
157 /// invokes. This function analyze BB to see if there are any calls, and if so,
158 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
159 /// nodes in that block with the values specified in InvokeDestPHIValues.
161 /// Returns true to indicate that the next block should be skipped.
162 static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
163 InvokeInliningInfo &Invoke) {
164 LandingPadInst *LPI = Invoke.getLandingPadInst();
166 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
167 Instruction *I = BBI++;
169 if (LandingPadInst *L = dyn_cast<LandingPadInst>(I)) {
170 unsigned NumClauses = LPI->getNumClauses();
171 L->reserveClauses(NumClauses);
172 for (unsigned i = 0; i != NumClauses; ++i)
173 L->addClause(LPI->getClause(i));
176 // We only need to check for function calls: inlined invoke
177 // instructions require no special handling.
178 CallInst *CI = dyn_cast<CallInst>(I);
180 // If this call cannot unwind, don't convert it to an invoke.
181 if (!CI || CI->doesNotThrow())
184 // Convert this function call into an invoke instruction. First, split the
186 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
188 // Delete the unconditional branch inserted by splitBasicBlock
189 BB->getInstList().pop_back();
191 // Create the new invoke instruction.
192 ImmutableCallSite CS(CI);
193 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
194 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split,
195 Invoke.getOuterResumeDest(),
196 InvokeArgs, CI->getName(), BB);
197 II->setCallingConv(CI->getCallingConv());
198 II->setAttributes(CI->getAttributes());
200 // Make sure that anything using the call now uses the invoke! This also
201 // updates the CallGraph if present, because it uses a WeakVH.
202 CI->replaceAllUsesWith(II);
204 // Delete the original call
205 Split->getInstList().pop_front();
207 // Update any PHI nodes in the exceptional block to indicate that there is
208 // now a new entry in them.
209 Invoke.addIncomingPHIValuesFor(BB);
216 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
217 /// in the body of the inlined function into invokes.
219 /// II is the invoke instruction being inlined. FirstNewBlock is the first
220 /// block of the inlined code (the last block is the end of the function),
221 /// and InlineCodeInfo is information about the code that got inlined.
222 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
223 ClonedCodeInfo &InlinedCodeInfo) {
224 BasicBlock *InvokeDest = II->getUnwindDest();
226 Function *Caller = FirstNewBlock->getParent();
228 // The inlined code is currently at the end of the function, scan from the
229 // start of the inlined code to its end, checking for stuff we need to
230 // rewrite. If the code doesn't have calls or unwinds, we know there is
231 // nothing to rewrite.
232 if (!InlinedCodeInfo.ContainsCalls) {
233 // Now that everything is happy, we have one final detail. The PHI nodes in
234 // the exception destination block still have entries due to the original
235 // invoke instruction. Eliminate these entries (which might even delete the
237 InvokeDest->removePredecessor(II->getParent());
241 InvokeInliningInfo Invoke(II);
243 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
244 if (InlinedCodeInfo.ContainsCalls)
245 if (HandleCallsInBlockInlinedThroughInvoke(BB, Invoke)) {
246 // Honor a request to skip the next block.
251 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator()))
252 Invoke.forwardResume(RI);
255 // Now that everything is happy, we have one final detail. The PHI nodes in
256 // the exception destination block still have entries due to the original
257 // invoke instruction. Eliminate these entries (which might even delete the
259 InvokeDest->removePredecessor(II->getParent());
262 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
263 /// into the caller, update the specified callgraph to reflect the changes we
264 /// made. Note that it's possible that not all code was copied over, so only
265 /// some edges of the callgraph may remain.
266 static void UpdateCallGraphAfterInlining(CallSite CS,
267 Function::iterator FirstNewBlock,
268 ValueToValueMapTy &VMap,
269 InlineFunctionInfo &IFI) {
270 CallGraph &CG = *IFI.CG;
271 const Function *Caller = CS.getInstruction()->getParent()->getParent();
272 const Function *Callee = CS.getCalledFunction();
273 CallGraphNode *CalleeNode = CG[Callee];
274 CallGraphNode *CallerNode = CG[Caller];
276 // Since we inlined some uninlined call sites in the callee into the caller,
277 // add edges from the caller to all of the callees of the callee.
278 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
280 // Consider the case where CalleeNode == CallerNode.
281 CallGraphNode::CalledFunctionsVector CallCache;
282 if (CalleeNode == CallerNode) {
283 CallCache.assign(I, E);
284 I = CallCache.begin();
288 for (; I != E; ++I) {
289 const Value *OrigCall = I->first;
291 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
292 // Only copy the edge if the call was inlined!
293 if (VMI == VMap.end() || VMI->second == 0)
296 // If the call was inlined, but then constant folded, there is no edge to
297 // add. Check for this case.
298 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
299 if (NewCall == 0) continue;
301 // Remember that this call site got inlined for the client of
303 IFI.InlinedCalls.push_back(NewCall);
305 // It's possible that inlining the callsite will cause it to go from an
306 // indirect to a direct call by resolving a function pointer. If this
307 // happens, set the callee of the new call site to a more precise
308 // destination. This can also happen if the call graph node of the caller
309 // was just unnecessarily imprecise.
310 if (I->second->getFunction() == 0)
311 if (Function *F = CallSite(NewCall).getCalledFunction()) {
312 // Indirect call site resolved to direct call.
313 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
318 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
321 // Update the call graph by deleting the edge from Callee to Caller. We must
322 // do this after the loop above in case Caller and Callee are the same.
323 CallerNode->removeCallEdgeFor(CS);
326 /// HandleByValArgument - When inlining a call site that has a byval argument,
327 /// we have to make the implicit memcpy explicit by adding it.
328 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
329 const Function *CalledFunc,
330 InlineFunctionInfo &IFI,
331 unsigned ByValAlignment) {
332 Type *AggTy = cast<PointerType>(Arg->getType())->getElementType();
334 // If the called function is readonly, then it could not mutate the caller's
335 // copy of the byval'd memory. In this case, it is safe to elide the copy and
337 if (CalledFunc->onlyReadsMemory()) {
338 // If the byval argument has a specified alignment that is greater than the
339 // passed in pointer, then we either have to round up the input pointer or
340 // give up on this transformation.
341 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
344 // If the pointer is already known to be sufficiently aligned, or if we can
345 // round it up to a larger alignment, then we don't need a temporary.
346 if (getOrEnforceKnownAlignment(Arg, ByValAlignment,
347 IFI.TD) >= ByValAlignment)
350 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
351 // for code quality, but rarely happens and is required for correctness.
354 LLVMContext &Context = Arg->getContext();
356 Type *VoidPtrTy = Type::getInt8PtrTy(Context);
358 // Create the alloca. If we have TargetData, use nice alignment.
361 Align = IFI.TD->getPrefTypeAlignment(AggTy);
363 // If the byval had an alignment specified, we *must* use at least that
364 // alignment, as it is required by the byval argument (and uses of the
365 // pointer inside the callee).
366 Align = std::max(Align, ByValAlignment);
368 Function *Caller = TheCall->getParent()->getParent();
370 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(),
371 &*Caller->begin()->begin());
373 Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)};
374 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(),
377 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall);
378 Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall);
382 Size = ConstantExpr::getSizeOf(AggTy);
384 Size = ConstantInt::get(Type::getInt64Ty(Context),
385 IFI.TD->getTypeStoreSize(AggTy));
387 // Always generate a memcpy of alignment 1 here because we don't know
388 // the alignment of the src pointer. Other optimizations can infer
390 Value *CallArgs[] = {
391 DestCast, SrcCast, Size,
392 ConstantInt::get(Type::getInt32Ty(Context), 1),
393 ConstantInt::getFalse(Context) // isVolatile
395 IRBuilder<>(TheCall).CreateCall(MemCpyFn, CallArgs);
397 // Uses of the argument in the function should use our new alloca
402 // isUsedByLifetimeMarker - Check whether this Value is used by a lifetime
404 static bool isUsedByLifetimeMarker(Value *V) {
405 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE;
407 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI)) {
408 switch (II->getIntrinsicID()) {
410 case Intrinsic::lifetime_start:
411 case Intrinsic::lifetime_end:
419 // hasLifetimeMarkers - Check whether the given alloca already has
420 // lifetime.start or lifetime.end intrinsics.
421 static bool hasLifetimeMarkers(AllocaInst *AI) {
422 Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext());
423 if (AI->getType() == Int8PtrTy)
424 return isUsedByLifetimeMarker(AI);
426 // Do a scan to find all the casts to i8*.
427 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); I != E;
429 if (I->getType() != Int8PtrTy) continue;
430 if (I->stripPointerCasts() != AI) continue;
431 if (isUsedByLifetimeMarker(*I))
437 /// updateInlinedAtInfo - Helper function used by fixupLineNumbers to recursively
438 /// update InlinedAtEntry of a DebugLoc.
439 static DebugLoc updateInlinedAtInfo(const DebugLoc &DL,
440 const DebugLoc &InlinedAtDL,
442 if (MDNode *IA = DL.getInlinedAt(Ctx)) {
443 DebugLoc NewInlinedAtDL
444 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx);
445 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
446 NewInlinedAtDL.getAsMDNode(Ctx));
449 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
450 InlinedAtDL.getAsMDNode(Ctx));
453 /// fixupLineNumbers - Update inlined instructions' line numbers to
454 /// to encode location where these instructions are inlined.
455 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
456 Instruction *TheCall) {
457 DebugLoc TheCallDL = TheCall->getDebugLoc();
458 if (TheCallDL.isUnknown())
461 for (; FI != Fn->end(); ++FI) {
462 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
464 DebugLoc DL = BI->getDebugLoc();
465 if (!DL.isUnknown()) {
466 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext()));
467 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) {
468 LLVMContext &Ctx = BI->getContext();
469 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
470 DVI->setOperand(2, createInlinedVariable(DVI->getVariable(),
478 /// InlineFunction - This function inlines the called function into the basic
479 /// block of the caller. This returns false if it is not possible to inline
480 /// this call. The program is still in a well defined state if this occurs
483 /// Note that this only does one level of inlining. For example, if the
484 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
485 /// exists in the instruction stream. Similarly this will inline a recursive
486 /// function by one level.
487 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI) {
488 Instruction *TheCall = CS.getInstruction();
489 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
490 "Instruction not in function!");
492 // If IFI has any state in it, zap it before we fill it in.
495 const Function *CalledFunc = CS.getCalledFunction();
496 if (CalledFunc == 0 || // Can't inline external function or indirect
497 CalledFunc->isDeclaration() || // call, or call to a vararg function!
498 CalledFunc->getFunctionType()->isVarArg()) return false;
500 // If the call to the callee is not a tail call, we must clear the 'tail'
501 // flags on any calls that we inline.
502 bool MustClearTailCallFlags =
503 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall());
505 // If the call to the callee cannot throw, set the 'nounwind' flag on any
506 // calls that we inline.
507 bool MarkNoUnwind = CS.doesNotThrow();
509 BasicBlock *OrigBB = TheCall->getParent();
510 Function *Caller = OrigBB->getParent();
512 // GC poses two hazards to inlining, which only occur when the callee has GC:
513 // 1. If the caller has no GC, then the callee's GC must be propagated to the
515 // 2. If the caller has a differing GC, it is invalid to inline.
516 if (CalledFunc->hasGC()) {
517 if (!Caller->hasGC())
518 Caller->setGC(CalledFunc->getGC());
519 else if (CalledFunc->getGC() != Caller->getGC())
523 // Get the personality function from the callee if it contains a landing pad.
524 Value *CalleePersonality = 0;
525 for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end();
527 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
528 const BasicBlock *BB = II->getUnwindDest();
529 const LandingPadInst *LP = BB->getLandingPadInst();
530 CalleePersonality = LP->getPersonalityFn();
534 // Find the personality function used by the landing pads of the caller. If it
535 // exists, then check to see that it matches the personality function used in
537 if (CalleePersonality) {
538 for (Function::const_iterator I = Caller->begin(), E = Caller->end();
540 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
541 const BasicBlock *BB = II->getUnwindDest();
542 const LandingPadInst *LP = BB->getLandingPadInst();
544 // If the personality functions match, then we can perform the
545 // inlining. Otherwise, we can't inline.
546 // TODO: This isn't 100% true. Some personality functions are proper
547 // supersets of others and can be used in place of the other.
548 if (LP->getPersonalityFn() != CalleePersonality)
555 // Get an iterator to the last basic block in the function, which will have
556 // the new function inlined after it.
557 Function::iterator LastBlock = &Caller->back();
559 // Make sure to capture all of the return instructions from the cloned
561 SmallVector<ReturnInst*, 8> Returns;
562 ClonedCodeInfo InlinedFunctionInfo;
563 Function::iterator FirstNewBlock;
565 { // Scope to destroy VMap after cloning.
566 ValueToValueMapTy VMap;
568 assert(CalledFunc->arg_size() == CS.arg_size() &&
569 "No varargs calls can be inlined!");
571 // Calculate the vector of arguments to pass into the function cloner, which
572 // matches up the formal to the actual argument values.
573 CallSite::arg_iterator AI = CS.arg_begin();
575 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
576 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
577 Value *ActualArg = *AI;
579 // When byval arguments actually inlined, we need to make the copy implied
580 // by them explicit. However, we don't do this if the callee is readonly
581 // or readnone, because the copy would be unneeded: the callee doesn't
582 // modify the struct.
583 if (CS.isByValArgument(ArgNo)) {
584 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
585 CalledFunc->getParamAlignment(ArgNo+1));
587 // Calls that we inline may use the new alloca, so we need to clear
588 // their 'tail' flags if HandleByValArgument introduced a new alloca and
589 // the callee has calls.
590 MustClearTailCallFlags |= ActualArg != *AI;
596 // We want the inliner to prune the code as it copies. We would LOVE to
597 // have no dead or constant instructions leftover after inlining occurs
598 // (which can happen, e.g., because an argument was constant), but we'll be
599 // happy with whatever the cloner can do.
600 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
601 /*ModuleLevelChanges=*/false, Returns, ".i",
602 &InlinedFunctionInfo, IFI.TD, TheCall);
604 // Remember the first block that is newly cloned over.
605 FirstNewBlock = LastBlock; ++FirstNewBlock;
607 // Update the callgraph if requested.
609 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
611 // Update inlined instructions' line number information.
612 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
615 // If there are any alloca instructions in the block that used to be the entry
616 // block for the callee, move them to the entry block of the caller. First
617 // calculate which instruction they should be inserted before. We insert the
618 // instructions at the end of the current alloca list.
620 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
621 for (BasicBlock::iterator I = FirstNewBlock->begin(),
622 E = FirstNewBlock->end(); I != E; ) {
623 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
624 if (AI == 0) continue;
626 // If the alloca is now dead, remove it. This often occurs due to code
628 if (AI->use_empty()) {
629 AI->eraseFromParent();
633 if (!isa<Constant>(AI->getArraySize()))
636 // Keep track of the static allocas that we inline into the caller.
637 IFI.StaticAllocas.push_back(AI);
639 // Scan for the block of allocas that we can move over, and move them
641 while (isa<AllocaInst>(I) &&
642 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
643 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
647 // Transfer all of the allocas over in a block. Using splice means
648 // that the instructions aren't removed from the symbol table, then
650 Caller->getEntryBlock().getInstList().splice(InsertPoint,
651 FirstNewBlock->getInstList(),
656 // Leave lifetime markers for the static alloca's, scoping them to the
657 // function we just inlined.
658 if (!IFI.StaticAllocas.empty()) {
659 IRBuilder<> builder(FirstNewBlock->begin());
660 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
661 AllocaInst *AI = IFI.StaticAllocas[ai];
663 // If the alloca is already scoped to something smaller than the whole
664 // function then there's no need to add redundant, less accurate markers.
665 if (hasLifetimeMarkers(AI))
668 builder.CreateLifetimeStart(AI);
669 for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) {
670 IRBuilder<> builder(Returns[ri]);
671 builder.CreateLifetimeEnd(AI);
676 // If the inlined code contained dynamic alloca instructions, wrap the inlined
677 // code with llvm.stacksave/llvm.stackrestore intrinsics.
678 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
679 Module *M = Caller->getParent();
680 // Get the two intrinsics we care about.
681 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
682 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
684 // Insert the llvm.stacksave.
685 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
686 .CreateCall(StackSave, "savedstack");
688 // Insert a call to llvm.stackrestore before any return instructions in the
690 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
691 IRBuilder<>(Returns[i]).CreateCall(StackRestore, SavedPtr);
695 // If we are inlining tail call instruction through a call site that isn't
696 // marked 'tail', we must remove the tail marker for any calls in the inlined
697 // code. Also, calls inlined through a 'nounwind' call site should be marked
699 if (InlinedFunctionInfo.ContainsCalls &&
700 (MustClearTailCallFlags || MarkNoUnwind)) {
701 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
703 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
704 if (CallInst *CI = dyn_cast<CallInst>(I)) {
705 if (MustClearTailCallFlags)
706 CI->setTailCall(false);
708 CI->setDoesNotThrow();
712 // If we are inlining for an invoke instruction, we must make sure to rewrite
713 // any call instructions into invoke instructions.
714 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
715 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
717 // If we cloned in _exactly one_ basic block, and if that block ends in a
718 // return instruction, we splice the body of the inlined callee directly into
719 // the calling basic block.
720 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
721 // Move all of the instructions right before the call.
722 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
723 FirstNewBlock->begin(), FirstNewBlock->end());
724 // Remove the cloned basic block.
725 Caller->getBasicBlockList().pop_back();
727 // If the call site was an invoke instruction, add a branch to the normal
729 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
730 BranchInst::Create(II->getNormalDest(), TheCall);
732 // If the return instruction returned a value, replace uses of the call with
733 // uses of the returned value.
734 if (!TheCall->use_empty()) {
735 ReturnInst *R = Returns[0];
736 if (TheCall == R->getReturnValue())
737 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
739 TheCall->replaceAllUsesWith(R->getReturnValue());
741 // Since we are now done with the Call/Invoke, we can delete it.
742 TheCall->eraseFromParent();
744 // Since we are now done with the return instruction, delete it also.
745 Returns[0]->eraseFromParent();
747 // We are now done with the inlining.
751 // Otherwise, we have the normal case, of more than one block to inline or
752 // multiple return sites.
754 // We want to clone the entire callee function into the hole between the
755 // "starter" and "ender" blocks. How we accomplish this depends on whether
756 // this is an invoke instruction or a call instruction.
757 BasicBlock *AfterCallBB;
758 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
760 // Add an unconditional branch to make this look like the CallInst case...
761 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
763 // Split the basic block. This guarantees that no PHI nodes will have to be
764 // updated due to new incoming edges, and make the invoke case more
765 // symmetric to the call case.
766 AfterCallBB = OrigBB->splitBasicBlock(NewBr,
767 CalledFunc->getName()+".exit");
769 } else { // It's a call
770 // If this is a call instruction, we need to split the basic block that
771 // the call lives in.
773 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
774 CalledFunc->getName()+".exit");
777 // Change the branch that used to go to AfterCallBB to branch to the first
778 // basic block of the inlined function.
780 TerminatorInst *Br = OrigBB->getTerminator();
781 assert(Br && Br->getOpcode() == Instruction::Br &&
782 "splitBasicBlock broken!");
783 Br->setOperand(0, FirstNewBlock);
786 // Now that the function is correct, make it a little bit nicer. In
787 // particular, move the basic blocks inserted from the end of the function
788 // into the space made by splitting the source basic block.
789 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
790 FirstNewBlock, Caller->end());
792 // Handle all of the return instructions that we just cloned in, and eliminate
793 // any users of the original call/invoke instruction.
794 Type *RTy = CalledFunc->getReturnType();
797 if (Returns.size() > 1) {
798 // The PHI node should go at the front of the new basic block to merge all
799 // possible incoming values.
800 if (!TheCall->use_empty()) {
801 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
802 AfterCallBB->begin());
803 // Anything that used the result of the function call should now use the
804 // PHI node as their operand.
805 TheCall->replaceAllUsesWith(PHI);
808 // Loop over all of the return instructions adding entries to the PHI node
811 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
812 ReturnInst *RI = Returns[i];
813 assert(RI->getReturnValue()->getType() == PHI->getType() &&
814 "Ret value not consistent in function!");
815 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
820 // Add a branch to the merge points and remove return instructions.
821 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
822 ReturnInst *RI = Returns[i];
823 BranchInst::Create(AfterCallBB, RI);
824 RI->eraseFromParent();
826 } else if (!Returns.empty()) {
827 // Otherwise, if there is exactly one return value, just replace anything
828 // using the return value of the call with the computed value.
829 if (!TheCall->use_empty()) {
830 if (TheCall == Returns[0]->getReturnValue())
831 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
833 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
836 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
837 BasicBlock *ReturnBB = Returns[0]->getParent();
838 ReturnBB->replaceAllUsesWith(AfterCallBB);
840 // Splice the code from the return block into the block that it will return
841 // to, which contains the code that was after the call.
842 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
843 ReturnBB->getInstList());
845 // Delete the return instruction now and empty ReturnBB now.
846 Returns[0]->eraseFromParent();
847 ReturnBB->eraseFromParent();
848 } else if (!TheCall->use_empty()) {
849 // No returns, but something is using the return value of the call. Just
851 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
854 // Since we are now done with the Call/Invoke, we can delete it.
855 TheCall->eraseFromParent();
857 // We should always be able to fold the entry block of the function into the
858 // single predecessor of the block...
859 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
860 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
862 // Splice the code entry block into calling block, right before the
863 // unconditional branch.
864 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
865 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
867 // Remove the unconditional branch.
868 OrigBB->getInstList().erase(Br);
870 // Now we can remove the CalleeEntry block, which is now empty.
871 Caller->getBasicBlockList().erase(CalleeEntry);
873 // If we inserted a phi node, check to see if it has a single value (e.g. all
874 // the entries are the same or undef). If so, remove the PHI so it doesn't
875 // block other optimizations.
877 if (Value *V = SimplifyInstruction(PHI, IFI.TD)) {
878 PHI->replaceAllUsesWith(V);
879 PHI->eraseFromParent();