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 // The code in this file for handling inlines through invoke
14 // instructions preserves semantics only under some assumptions about
15 // the behavior of unwinders which correspond to gcc-style libUnwind
16 // exception personality functions. Eventually the IR will be
17 // improved to make this unnecessary, but until then, this code is
18 // marked [LIBUNWIND].
20 //===----------------------------------------------------------------------===//
22 #include "llvm/Transforms/Utils/Cloning.h"
23 #include "llvm/Constants.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/Module.h"
26 #include "llvm/Instructions.h"
27 #include "llvm/IntrinsicInst.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/Attributes.h"
30 #include "llvm/Analysis/CallGraph.h"
31 #include "llvm/Analysis/DebugInfo.h"
32 #include "llvm/Analysis/InstructionSimplify.h"
33 #include "llvm/Target/TargetData.h"
34 #include "llvm/Transforms/Utils/Local.h"
35 #include "llvm/ADT/SmallVector.h"
36 #include "llvm/ADT/StringExtras.h"
37 #include "llvm/Support/CallSite.h"
38 #include "llvm/Support/IRBuilder.h"
41 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI) {
42 return InlineFunction(CallSite(CI), IFI);
44 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI) {
45 return InlineFunction(CallSite(II), IFI);
48 /// [LIBUNWIND] Find the (possibly absent) call to @llvm.eh.selector in
49 /// the given landing pad.
50 static EHSelectorInst *findSelectorForLandingPad(BasicBlock *lpad) {
51 // The llvm.eh.exception call is required to be in the landing pad.
52 for (BasicBlock::iterator i = lpad->begin(), e = lpad->end(); i != e; i++) {
53 EHExceptionInst *exn = dyn_cast<EHExceptionInst>(i);
56 EHSelectorInst *selector = 0;
57 for (Instruction::use_iterator
58 ui = exn->use_begin(), ue = exn->use_end(); ui != ue; ++ui) {
59 EHSelectorInst *sel = dyn_cast<EHSelectorInst>(*ui);
62 // Immediately accept an eh.selector in the landing pad.
63 if (sel->getParent() == lpad) return sel;
65 // Otherwise, use the first selector we see.
66 if (!selector) selector = sel;
76 /// A class for recording information about inlining through an invoke.
77 class InvokeInliningInfo {
78 BasicBlock *OuterUnwindDest;
79 EHSelectorInst *OuterSelector;
80 BasicBlock *InnerUnwindDest;
81 PHINode *InnerExceptionPHI;
82 PHINode *InnerSelectorPHI;
83 SmallVector<Value*, 8> UnwindDestPHIValues;
86 InvokeInliningInfo(InvokeInst *II) :
87 OuterUnwindDest(II->getUnwindDest()), OuterSelector(0),
88 InnerUnwindDest(0), InnerExceptionPHI(0), InnerSelectorPHI(0) {
90 // If there are PHI nodes in the unwind destination block, we
91 // need to keep track of which values came into them from the
92 // invoke before removing the edge from this block.
93 llvm::BasicBlock *invokeBB = II->getParent();
94 for (BasicBlock::iterator I = OuterUnwindDest->begin();
95 isa<PHINode>(I); ++I) {
96 // Save the value to use for this edge.
97 PHINode *phi = cast<PHINode>(I);
98 UnwindDestPHIValues.push_back(phi->getIncomingValueForBlock(invokeBB));
102 /// The outer unwind destination is the target of unwind edges
103 /// introduced for calls within the inlined function.
104 BasicBlock *getOuterUnwindDest() const {
105 return OuterUnwindDest;
108 EHSelectorInst *getOuterSelector() {
110 OuterSelector = findSelectorForLandingPad(OuterUnwindDest);
111 return OuterSelector;
114 BasicBlock *getInnerUnwindDest();
116 bool forwardEHResume(CallInst *call, BasicBlock *src);
118 /// Add incoming-PHI values to the unwind destination block for
119 /// the given basic block, using the values for the original
120 /// invoke's source block.
121 void addIncomingPHIValuesFor(BasicBlock *BB) const {
122 addIncomingPHIValuesForInto(BB, OuterUnwindDest);
125 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const {
126 BasicBlock::iterator I = dest->begin();
127 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
128 PHINode *phi = cast<PHINode>(I);
129 phi->addIncoming(UnwindDestPHIValues[i], src);
135 /// Replace all the instruction uses of a value with a different value.
136 /// This has the advantage of not screwing up the CallGraph.
137 static void replaceAllInsnUsesWith(Instruction *insn, Value *replacement) {
138 for (Value::use_iterator i = insn->use_begin(), e = insn->use_end();
140 Use &use = i.getUse();
142 if (isa<Instruction>(use.getUser()))
143 use.set(replacement);
147 /// Get or create a target for the branch out of rewritten calls to
149 BasicBlock *InvokeInliningInfo::getInnerUnwindDest() {
150 if (InnerUnwindDest) return InnerUnwindDest;
152 // Find and hoist the llvm.eh.exception and llvm.eh.selector calls
153 // in the outer landing pad to immediately following the phis.
154 EHSelectorInst *selector = getOuterSelector();
155 if (!selector) return 0;
157 // The call to llvm.eh.exception *must* be in the landing pad.
158 Instruction *exn = cast<Instruction>(selector->getArgOperand(0));
159 assert(exn->getParent() == OuterUnwindDest);
161 // TODO: recognize when we've already done this, so that we don't
162 // get a linear number of these when inlining calls into lots of
163 // invokes with the same landing pad.
166 Instruction *splitPoint = exn->getParent()->getFirstNonPHI();
167 assert(splitPoint != selector && "selector-on-exception dominance broken!");
168 if (splitPoint == exn) {
169 selector->removeFromParent();
170 selector->insertAfter(exn);
171 splitPoint = selector->getNextNode();
173 exn->moveBefore(splitPoint);
174 selector->moveBefore(splitPoint);
177 // Split the landing pad.
178 InnerUnwindDest = OuterUnwindDest->splitBasicBlock(splitPoint,
179 OuterUnwindDest->getName() + ".body");
181 // The number of incoming edges we expect to the inner landing pad.
182 const unsigned phiCapacity = 2;
184 // Create corresponding new phis for all the phis in the outer landing pad.
185 BasicBlock::iterator insertPoint = InnerUnwindDest->begin();
186 BasicBlock::iterator I = OuterUnwindDest->begin();
187 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
188 PHINode *outerPhi = cast<PHINode>(I);
189 PHINode *innerPhi = PHINode::Create(outerPhi->getType(), phiCapacity,
190 outerPhi->getName() + ".lpad-body",
192 outerPhi->replaceAllUsesWith(innerPhi);
193 innerPhi->addIncoming(outerPhi, OuterUnwindDest);
196 // Create a phi for the exception value...
197 InnerExceptionPHI = PHINode::Create(exn->getType(), phiCapacity,
198 "exn.lpad-body", insertPoint);
199 replaceAllInsnUsesWith(exn, InnerExceptionPHI);
200 selector->setArgOperand(0, exn); // restore this use
201 InnerExceptionPHI->addIncoming(exn, OuterUnwindDest);
203 // ...and the selector.
204 InnerSelectorPHI = PHINode::Create(selector->getType(), phiCapacity,
205 "selector.lpad-body", insertPoint);
206 replaceAllInsnUsesWith(selector, InnerSelectorPHI);
207 InnerSelectorPHI->addIncoming(selector, OuterUnwindDest);
210 return InnerUnwindDest;
213 /// [LIBUNWIND] Try to forward the given call, which logically occurs
214 /// at the end of the given block, as a branch to the inner unwind
215 /// block. Returns true if the call was forwarded.
216 bool InvokeInliningInfo::forwardEHResume(CallInst *call, BasicBlock *src) {
217 Function *fn = dyn_cast<Function>(call->getCalledValue());
218 if (!fn || fn->getName() != "llvm.eh.resume")
221 // If this fails, maybe it should be a fatal error.
222 BasicBlock *dest = getInnerUnwindDest();
223 if (!dest) return false;
226 BranchInst::Create(dest, src);
228 // Update the phis in the destination. They were inserted in an
229 // order which makes this work.
230 addIncomingPHIValuesForInto(src, dest);
231 InnerExceptionPHI->addIncoming(call->getArgOperand(0), src);
232 InnerSelectorPHI->addIncoming(call->getArgOperand(1), src);
237 /// [LIBUNWIND] Check whether this selector is "only cleanups":
238 /// call i32 @llvm.eh.selector(blah, blah, i32 0)
239 static bool isCleanupOnlySelector(EHSelectorInst *selector) {
240 if (selector->getNumArgOperands() != 3) return false;
241 ConstantInt *val = dyn_cast<ConstantInt>(selector->getArgOperand(2));
242 return (val && val->isZero());
245 /// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
246 /// an invoke, we have to turn all of the calls that can throw into
247 /// invokes. This function analyze BB to see if there are any calls, and if so,
248 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
249 /// nodes in that block with the values specified in InvokeDestPHIValues.
251 /// Returns true to indicate that the next block should be skipped.
252 static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
253 InvokeInliningInfo &Invoke) {
254 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
255 Instruction *I = BBI++;
257 // We only need to check for function calls: inlined invoke
258 // instructions require no special handling.
259 CallInst *CI = dyn_cast<CallInst>(I);
260 if (CI == 0) continue;
262 // LIBUNWIND: merge selector instructions.
263 if (EHSelectorInst *Inner = dyn_cast<EHSelectorInst>(CI)) {
264 EHSelectorInst *Outer = Invoke.getOuterSelector();
265 if (!Outer) continue;
267 bool innerIsOnlyCleanup = isCleanupOnlySelector(Inner);
268 bool outerIsOnlyCleanup = isCleanupOnlySelector(Outer);
270 // If both selectors contain only cleanups, we don't need to do
271 // anything. TODO: this is really just a very specific instance
272 // of a much more general optimization.
273 if (innerIsOnlyCleanup && outerIsOnlyCleanup) continue;
275 // Otherwise, we just append the outer selector to the inner selector.
276 SmallVector<Value*, 16> NewSelector;
277 for (unsigned i = 0, e = Inner->getNumArgOperands(); i != e; ++i)
278 NewSelector.push_back(Inner->getArgOperand(i));
279 for (unsigned i = 2, e = Outer->getNumArgOperands(); i != e; ++i)
280 NewSelector.push_back(Outer->getArgOperand(i));
282 CallInst *NewInner = CallInst::Create(Inner->getCalledValue(),
287 // No need to copy attributes, calling convention, etc.
288 NewInner->takeName(Inner);
289 Inner->replaceAllUsesWith(NewInner);
290 Inner->eraseFromParent();
294 // If this call cannot unwind, don't convert it to an invoke.
295 if (CI->doesNotThrow())
298 // Convert this function call into an invoke instruction.
299 // First, split the basic block.
300 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
302 // Delete the unconditional branch inserted by splitBasicBlock
303 BB->getInstList().pop_back();
305 // LIBUNWIND: If this is a call to @llvm.eh.resume, just branch
306 // directly to the new landing pad.
307 if (Invoke.forwardEHResume(CI, BB)) {
308 // TODO: 'Split' is now unreachable; clean it up.
310 // We want to leave the original call intact so that the call
311 // graph and other structures won't get misled. We also have to
312 // avoid processing the next block, or we'll iterate here forever.
316 // Otherwise, create the new invoke instruction.
317 ImmutableCallSite CS(CI);
318 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
320 InvokeInst::Create(CI->getCalledValue(), Split,
321 Invoke.getOuterUnwindDest(),
322 InvokeArgs.begin(), InvokeArgs.end(),
324 II->setCallingConv(CI->getCallingConv());
325 II->setAttributes(CI->getAttributes());
327 // Make sure that anything using the call now uses the invoke! This also
328 // updates the CallGraph if present, because it uses a WeakVH.
329 CI->replaceAllUsesWith(II);
331 Split->getInstList().pop_front(); // Delete the original call
333 // Update any PHI nodes in the exceptional block to indicate that
334 // there is now a new entry in them.
335 Invoke.addIncomingPHIValuesFor(BB);
343 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
344 /// in the body of the inlined function into invokes and turn unwind
345 /// instructions into branches to the invoke unwind dest.
347 /// II is the invoke instruction being inlined. FirstNewBlock is the first
348 /// block of the inlined code (the last block is the end of the function),
349 /// and InlineCodeInfo is information about the code that got inlined.
350 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
351 ClonedCodeInfo &InlinedCodeInfo) {
352 BasicBlock *InvokeDest = II->getUnwindDest();
354 Function *Caller = FirstNewBlock->getParent();
356 // The inlined code is currently at the end of the function, scan from the
357 // start of the inlined code to its end, checking for stuff we need to
358 // rewrite. If the code doesn't have calls or unwinds, we know there is
359 // nothing to rewrite.
360 if (!InlinedCodeInfo.ContainsCalls && !InlinedCodeInfo.ContainsUnwinds) {
361 // Now that everything is happy, we have one final detail. The PHI nodes in
362 // the exception destination block still have entries due to the original
363 // invoke instruction. Eliminate these entries (which might even delete the
365 InvokeDest->removePredecessor(II->getParent());
369 InvokeInliningInfo Invoke(II);
371 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
372 if (InlinedCodeInfo.ContainsCalls)
373 if (HandleCallsInBlockInlinedThroughInvoke(BB, Invoke)) {
374 // Honor a request to skip the next block. We don't need to
375 // consider UnwindInsts in this case either.
380 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
381 // An UnwindInst requires special handling when it gets inlined into an
382 // invoke site. Once this happens, we know that the unwind would cause
383 // a control transfer to the invoke exception destination, so we can
384 // transform it into a direct branch to the exception destination.
385 BranchInst::Create(InvokeDest, UI);
387 // Delete the unwind instruction!
388 UI->eraseFromParent();
390 // Update any PHI nodes in the exceptional block to indicate that
391 // there is now a new entry in them.
392 Invoke.addIncomingPHIValuesFor(BB);
396 // Now that everything is happy, we have one final detail. The PHI nodes in
397 // the exception destination block still have entries due to the original
398 // invoke instruction. Eliminate these entries (which might even delete the
400 InvokeDest->removePredecessor(II->getParent());
403 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
404 /// into the caller, update the specified callgraph to reflect the changes we
405 /// made. Note that it's possible that not all code was copied over, so only
406 /// some edges of the callgraph may remain.
407 static void UpdateCallGraphAfterInlining(CallSite CS,
408 Function::iterator FirstNewBlock,
409 ValueToValueMapTy &VMap,
410 InlineFunctionInfo &IFI) {
411 CallGraph &CG = *IFI.CG;
412 const Function *Caller = CS.getInstruction()->getParent()->getParent();
413 const Function *Callee = CS.getCalledFunction();
414 CallGraphNode *CalleeNode = CG[Callee];
415 CallGraphNode *CallerNode = CG[Caller];
417 // Since we inlined some uninlined call sites in the callee into the caller,
418 // add edges from the caller to all of the callees of the callee.
419 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
421 // Consider the case where CalleeNode == CallerNode.
422 CallGraphNode::CalledFunctionsVector CallCache;
423 if (CalleeNode == CallerNode) {
424 CallCache.assign(I, E);
425 I = CallCache.begin();
429 for (; I != E; ++I) {
430 const Value *OrigCall = I->first;
432 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
433 // Only copy the edge if the call was inlined!
434 if (VMI == VMap.end() || VMI->second == 0)
437 // If the call was inlined, but then constant folded, there is no edge to
438 // add. Check for this case.
439 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
440 if (NewCall == 0) continue;
442 // Remember that this call site got inlined for the client of
444 IFI.InlinedCalls.push_back(NewCall);
446 // It's possible that inlining the callsite will cause it to go from an
447 // indirect to a direct call by resolving a function pointer. If this
448 // happens, set the callee of the new call site to a more precise
449 // destination. This can also happen if the call graph node of the caller
450 // was just unnecessarily imprecise.
451 if (I->second->getFunction() == 0)
452 if (Function *F = CallSite(NewCall).getCalledFunction()) {
453 // Indirect call site resolved to direct call.
454 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
459 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
462 // Update the call graph by deleting the edge from Callee to Caller. We must
463 // do this after the loop above in case Caller and Callee are the same.
464 CallerNode->removeCallEdgeFor(CS);
467 /// HandleByValArgument - When inlining a call site that has a byval argument,
468 /// we have to make the implicit memcpy explicit by adding it.
469 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
470 const Function *CalledFunc,
471 InlineFunctionInfo &IFI,
472 unsigned ByValAlignment) {
473 const Type *AggTy = cast<PointerType>(Arg->getType())->getElementType();
475 // If the called function is readonly, then it could not mutate the caller's
476 // copy of the byval'd memory. In this case, it is safe to elide the copy and
478 if (CalledFunc->onlyReadsMemory()) {
479 // If the byval argument has a specified alignment that is greater than the
480 // passed in pointer, then we either have to round up the input pointer or
481 // give up on this transformation.
482 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
485 // If the pointer is already known to be sufficiently aligned, or if we can
486 // round it up to a larger alignment, then we don't need a temporary.
487 if (getOrEnforceKnownAlignment(Arg, ByValAlignment,
488 IFI.TD) >= ByValAlignment)
491 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
492 // for code quality, but rarely happens and is required for correctness.
495 LLVMContext &Context = Arg->getContext();
497 const Type *VoidPtrTy = Type::getInt8PtrTy(Context);
499 // Create the alloca. If we have TargetData, use nice alignment.
502 Align = IFI.TD->getPrefTypeAlignment(AggTy);
504 // If the byval had an alignment specified, we *must* use at least that
505 // alignment, as it is required by the byval argument (and uses of the
506 // pointer inside the callee).
507 Align = std::max(Align, ByValAlignment);
509 Function *Caller = TheCall->getParent()->getParent();
511 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(),
512 &*Caller->begin()->begin());
514 const Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)};
515 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(),
518 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall);
519 Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall);
523 Size = ConstantExpr::getSizeOf(AggTy);
525 Size = ConstantInt::get(Type::getInt64Ty(Context),
526 IFI.TD->getTypeStoreSize(AggTy));
528 // Always generate a memcpy of alignment 1 here because we don't know
529 // the alignment of the src pointer. Other optimizations can infer
531 Value *CallArgs[] = {
532 DestCast, SrcCast, Size,
533 ConstantInt::get(Type::getInt32Ty(Context), 1),
534 ConstantInt::getFalse(Context) // isVolatile
536 CallInst *TheMemCpy =
537 CallInst::Create(MemCpyFn, CallArgs, CallArgs+5, "", TheCall);
539 // If we have a call graph, update it.
540 if (CallGraph *CG = IFI.CG) {
541 CallGraphNode *MemCpyCGN = CG->getOrInsertFunction(MemCpyFn);
542 CallGraphNode *CallerNode = (*CG)[Caller];
543 CallerNode->addCalledFunction(TheMemCpy, MemCpyCGN);
546 // Uses of the argument in the function should use our new alloca
551 // isUsedByLifetimeMarker - Check whether this Value is used by a lifetime
553 static bool isUsedByLifetimeMarker(Value *V) {
554 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE;
556 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI)) {
557 switch (II->getIntrinsicID()) {
559 case Intrinsic::lifetime_start:
560 case Intrinsic::lifetime_end:
568 // hasLifetimeMarkers - Check whether the given alloca already has
569 // lifetime.start or lifetime.end intrinsics.
570 static bool hasLifetimeMarkers(AllocaInst *AI) {
571 const Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext());
572 if (AI->getType() == Int8PtrTy)
573 return isUsedByLifetimeMarker(AI);
575 // Do a scan to find all the bitcasts to i8*.
576 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); I != E;
578 if (I->getType() != Int8PtrTy) continue;
579 if (!isa<BitCastInst>(*I)) continue;
580 if (isUsedByLifetimeMarker(*I))
586 // InlineFunction - This function inlines the called function into the basic
587 // block of the caller. This returns false if it is not possible to inline this
588 // call. The program is still in a well defined state if this occurs though.
590 // Note that this only does one level of inlining. For example, if the
591 // instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
592 // exists in the instruction stream. Similarly this will inline a recursive
593 // function by one level.
595 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI) {
596 Instruction *TheCall = CS.getInstruction();
597 LLVMContext &Context = TheCall->getContext();
598 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
599 "Instruction not in function!");
601 // If IFI has any state in it, zap it before we fill it in.
604 const Function *CalledFunc = CS.getCalledFunction();
605 if (CalledFunc == 0 || // Can't inline external function or indirect
606 CalledFunc->isDeclaration() || // call, or call to a vararg function!
607 CalledFunc->getFunctionType()->isVarArg()) return false;
609 // If the call to the callee is not a tail call, we must clear the 'tail'
610 // flags on any calls that we inline.
611 bool MustClearTailCallFlags =
612 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall());
614 // If the call to the callee cannot throw, set the 'nounwind' flag on any
615 // calls that we inline.
616 bool MarkNoUnwind = CS.doesNotThrow();
618 BasicBlock *OrigBB = TheCall->getParent();
619 Function *Caller = OrigBB->getParent();
621 // GC poses two hazards to inlining, which only occur when the callee has GC:
622 // 1. If the caller has no GC, then the callee's GC must be propagated to the
624 // 2. If the caller has a differing GC, it is invalid to inline.
625 if (CalledFunc->hasGC()) {
626 if (!Caller->hasGC())
627 Caller->setGC(CalledFunc->getGC());
628 else if (CalledFunc->getGC() != Caller->getGC())
632 // Get an iterator to the last basic block in the function, which will have
633 // the new function inlined after it.
635 Function::iterator LastBlock = &Caller->back();
637 // Make sure to capture all of the return instructions from the cloned
639 SmallVector<ReturnInst*, 8> Returns;
640 ClonedCodeInfo InlinedFunctionInfo;
641 Function::iterator FirstNewBlock;
643 { // Scope to destroy VMap after cloning.
644 ValueToValueMapTy VMap;
646 assert(CalledFunc->arg_size() == CS.arg_size() &&
647 "No varargs calls can be inlined!");
649 // Calculate the vector of arguments to pass into the function cloner, which
650 // matches up the formal to the actual argument values.
651 CallSite::arg_iterator AI = CS.arg_begin();
653 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
654 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
655 Value *ActualArg = *AI;
657 // When byval arguments actually inlined, we need to make the copy implied
658 // by them explicit. However, we don't do this if the callee is readonly
659 // or readnone, because the copy would be unneeded: the callee doesn't
660 // modify the struct.
661 if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal)) {
662 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
663 CalledFunc->getParamAlignment(ArgNo+1));
665 // Calls that we inline may use the new alloca, so we need to clear
666 // their 'tail' flags if HandleByValArgument introduced a new alloca and
667 // the callee has calls.
668 MustClearTailCallFlags |= ActualArg != *AI;
674 // We want the inliner to prune the code as it copies. We would LOVE to
675 // have no dead or constant instructions leftover after inlining occurs
676 // (which can happen, e.g., because an argument was constant), but we'll be
677 // happy with whatever the cloner can do.
678 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
679 /*ModuleLevelChanges=*/false, Returns, ".i",
680 &InlinedFunctionInfo, IFI.TD, TheCall);
682 // Remember the first block that is newly cloned over.
683 FirstNewBlock = LastBlock; ++FirstNewBlock;
685 // Update the callgraph if requested.
687 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
690 // If there are any alloca instructions in the block that used to be the entry
691 // block for the callee, move them to the entry block of the caller. First
692 // calculate which instruction they should be inserted before. We insert the
693 // instructions at the end of the current alloca list.
696 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
697 for (BasicBlock::iterator I = FirstNewBlock->begin(),
698 E = FirstNewBlock->end(); I != E; ) {
699 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
700 if (AI == 0) continue;
702 // If the alloca is now dead, remove it. This often occurs due to code
704 if (AI->use_empty()) {
705 AI->eraseFromParent();
709 if (!isa<Constant>(AI->getArraySize()))
712 // Keep track of the static allocas that we inline into the caller.
713 IFI.StaticAllocas.push_back(AI);
715 // Scan for the block of allocas that we can move over, and move them
717 while (isa<AllocaInst>(I) &&
718 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
719 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
723 // Transfer all of the allocas over in a block. Using splice means
724 // that the instructions aren't removed from the symbol table, then
726 Caller->getEntryBlock().getInstList().splice(InsertPoint,
727 FirstNewBlock->getInstList(),
732 // Leave lifetime markers for the static alloca's, scoping them to the
733 // function we just inlined.
734 if (!IFI.StaticAllocas.empty()) {
735 // Also preserve the call graph, if applicable.
736 CallGraphNode *StartCGN = 0, *EndCGN = 0, *CallerNode = 0;
737 if (CallGraph *CG = IFI.CG) {
738 Function *Start = Intrinsic::getDeclaration(Caller->getParent(),
739 Intrinsic::lifetime_start);
740 Function *End = Intrinsic::getDeclaration(Caller->getParent(),
741 Intrinsic::lifetime_end);
742 StartCGN = CG->getOrInsertFunction(Start);
743 EndCGN = CG->getOrInsertFunction(End);
744 CallerNode = (*CG)[Caller];
747 IRBuilder<> builder(FirstNewBlock->begin());
748 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
749 AllocaInst *AI = IFI.StaticAllocas[ai];
751 // If the alloca is already scoped to something smaller than the whole
752 // function then there's no need to add redundant, less accurate markers.
753 if (hasLifetimeMarkers(AI))
756 CallInst *StartCall = builder.CreateLifetimeStart(AI);
757 if (IFI.CG) CallerNode->addCalledFunction(StartCall, StartCGN);
758 for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) {
759 IRBuilder<> builder(Returns[ri]);
760 CallInst *EndCall = builder.CreateLifetimeEnd(AI);
761 if (IFI.CG) CallerNode->addCalledFunction(EndCall, EndCGN);
766 // If the inlined code contained dynamic alloca instructions, wrap the inlined
767 // code with llvm.stacksave/llvm.stackrestore intrinsics.
768 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
769 Module *M = Caller->getParent();
770 // Get the two intrinsics we care about.
771 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
772 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
774 // If we are preserving the callgraph, add edges to the stacksave/restore
775 // functions for the calls we insert.
776 CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0;
777 if (CallGraph *CG = IFI.CG) {
778 StackSaveCGN = CG->getOrInsertFunction(StackSave);
779 StackRestoreCGN = CG->getOrInsertFunction(StackRestore);
780 CallerNode = (*CG)[Caller];
783 // Insert the llvm.stacksave.
784 CallInst *SavedPtr = CallInst::Create(StackSave, "savedstack",
785 FirstNewBlock->begin());
786 if (IFI.CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN);
788 // Insert a call to llvm.stackrestore before any return instructions in the
790 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
791 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", Returns[i]);
792 if (IFI.CG) CallerNode->addCalledFunction(CI, StackRestoreCGN);
795 // Count the number of StackRestore calls we insert.
796 unsigned NumStackRestores = Returns.size();
798 // If we are inlining an invoke instruction, insert restores before each
799 // unwind. These unwinds will be rewritten into branches later.
800 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) {
801 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
803 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
804 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", UI);
805 if (IFI.CG) CallerNode->addCalledFunction(CI, StackRestoreCGN);
811 // If we are inlining tail call instruction through a call site that isn't
812 // marked 'tail', we must remove the tail marker for any calls in the inlined
813 // code. Also, calls inlined through a 'nounwind' call site should be marked
815 if (InlinedFunctionInfo.ContainsCalls &&
816 (MustClearTailCallFlags || MarkNoUnwind)) {
817 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
819 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
820 if (CallInst *CI = dyn_cast<CallInst>(I)) {
821 if (MustClearTailCallFlags)
822 CI->setTailCall(false);
824 CI->setDoesNotThrow();
828 // If we are inlining through a 'nounwind' call site then any inlined 'unwind'
829 // instructions are unreachable.
830 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind)
831 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
833 TerminatorInst *Term = BB->getTerminator();
834 if (isa<UnwindInst>(Term)) {
835 new UnreachableInst(Context, Term);
836 BB->getInstList().erase(Term);
840 // If we are inlining for an invoke instruction, we must make sure to rewrite
841 // any inlined 'unwind' instructions into branches to the invoke exception
842 // destination, and call instructions into invoke instructions.
843 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
844 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
846 // If we cloned in _exactly one_ basic block, and if that block ends in a
847 // return instruction, we splice the body of the inlined callee directly into
848 // the calling basic block.
849 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
850 // Move all of the instructions right before the call.
851 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
852 FirstNewBlock->begin(), FirstNewBlock->end());
853 // Remove the cloned basic block.
854 Caller->getBasicBlockList().pop_back();
856 // If the call site was an invoke instruction, add a branch to the normal
858 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
859 BranchInst::Create(II->getNormalDest(), TheCall);
861 // If the return instruction returned a value, replace uses of the call with
862 // uses of the returned value.
863 if (!TheCall->use_empty()) {
864 ReturnInst *R = Returns[0];
865 if (TheCall == R->getReturnValue())
866 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
868 TheCall->replaceAllUsesWith(R->getReturnValue());
870 // Since we are now done with the Call/Invoke, we can delete it.
871 TheCall->eraseFromParent();
873 // Since we are now done with the return instruction, delete it also.
874 Returns[0]->eraseFromParent();
876 // We are now done with the inlining.
880 // Otherwise, we have the normal case, of more than one block to inline or
881 // multiple return sites.
883 // We want to clone the entire callee function into the hole between the
884 // "starter" and "ender" blocks. How we accomplish this depends on whether
885 // this is an invoke instruction or a call instruction.
886 BasicBlock *AfterCallBB;
887 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
889 // Add an unconditional branch to make this look like the CallInst case...
890 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
892 // Split the basic block. This guarantees that no PHI nodes will have to be
893 // updated due to new incoming edges, and make the invoke case more
894 // symmetric to the call case.
895 AfterCallBB = OrigBB->splitBasicBlock(NewBr,
896 CalledFunc->getName()+".exit");
898 } else { // It's a call
899 // If this is a call instruction, we need to split the basic block that
900 // the call lives in.
902 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
903 CalledFunc->getName()+".exit");
906 // Change the branch that used to go to AfterCallBB to branch to the first
907 // basic block of the inlined function.
909 TerminatorInst *Br = OrigBB->getTerminator();
910 assert(Br && Br->getOpcode() == Instruction::Br &&
911 "splitBasicBlock broken!");
912 Br->setOperand(0, FirstNewBlock);
915 // Now that the function is correct, make it a little bit nicer. In
916 // particular, move the basic blocks inserted from the end of the function
917 // into the space made by splitting the source basic block.
918 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
919 FirstNewBlock, Caller->end());
921 // Handle all of the return instructions that we just cloned in, and eliminate
922 // any users of the original call/invoke instruction.
923 const Type *RTy = CalledFunc->getReturnType();
926 if (Returns.size() > 1) {
927 // The PHI node should go at the front of the new basic block to merge all
928 // possible incoming values.
929 if (!TheCall->use_empty()) {
930 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
931 AfterCallBB->begin());
932 // Anything that used the result of the function call should now use the
933 // PHI node as their operand.
934 TheCall->replaceAllUsesWith(PHI);
937 // Loop over all of the return instructions adding entries to the PHI node
940 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
941 ReturnInst *RI = Returns[i];
942 assert(RI->getReturnValue()->getType() == PHI->getType() &&
943 "Ret value not consistent in function!");
944 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
949 // Add a branch to the merge points and remove return instructions.
950 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
951 ReturnInst *RI = Returns[i];
952 BranchInst::Create(AfterCallBB, RI);
953 RI->eraseFromParent();
955 } else if (!Returns.empty()) {
956 // Otherwise, if there is exactly one return value, just replace anything
957 // using the return value of the call with the computed value.
958 if (!TheCall->use_empty()) {
959 if (TheCall == Returns[0]->getReturnValue())
960 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
962 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
965 // Splice the code from the return block into the block that it will return
966 // to, which contains the code that was after the call.
967 BasicBlock *ReturnBB = Returns[0]->getParent();
968 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
969 ReturnBB->getInstList());
971 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
972 ReturnBB->replaceAllUsesWith(AfterCallBB);
974 // Delete the return instruction now and empty ReturnBB now.
975 Returns[0]->eraseFromParent();
976 ReturnBB->eraseFromParent();
977 } else if (!TheCall->use_empty()) {
978 // No returns, but something is using the return value of the call. Just
980 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
983 // Since we are now done with the Call/Invoke, we can delete it.
984 TheCall->eraseFromParent();
986 // We should always be able to fold the entry block of the function into the
987 // single predecessor of the block...
988 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
989 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
991 // Splice the code entry block into calling block, right before the
992 // unconditional branch.
993 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
994 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
996 // Remove the unconditional branch.
997 OrigBB->getInstList().erase(Br);
999 // Now we can remove the CalleeEntry block, which is now empty.
1000 Caller->getBasicBlockList().erase(CalleeEntry);
1002 // If we inserted a phi node, check to see if it has a single value (e.g. all
1003 // the entries are the same or undef). If so, remove the PHI so it doesn't
1004 // block other optimizations.
1006 if (Value *V = SimplifyInstruction(PHI, IFI.TD)) {
1007 PHI->replaceAllUsesWith(V);
1008 PHI->eraseFromParent();