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] Look for an llvm.eh.exception call in the given block.
49 static EHExceptionInst *findExceptionInBlock(BasicBlock *bb) {
50 for (BasicBlock::iterator i = bb->begin(), e = bb->end(); i != e; i++) {
51 EHExceptionInst *exn = dyn_cast<EHExceptionInst>(i);
58 /// [LIBUNWIND] Look for the 'best' llvm.eh.selector instruction for
59 /// the given llvm.eh.exception call.
60 static EHSelectorInst *findSelectorForException(EHExceptionInst *exn) {
61 BasicBlock *exnBlock = exn->getParent();
63 EHSelectorInst *outOfBlockSelector = 0;
64 for (Instruction::use_iterator
65 ui = exn->use_begin(), ue = exn->use_end(); ui != ue; ++ui) {
66 EHSelectorInst *sel = dyn_cast<EHSelectorInst>(*ui);
69 // Immediately accept an eh.selector in the same block as the
71 if (sel->getParent() == exnBlock) return sel;
73 // Otherwise, use the first selector we see.
74 if (!outOfBlockSelector) outOfBlockSelector = sel;
77 return outOfBlockSelector;
80 /// [LIBUNWIND] Find the (possibly absent) call to @llvm.eh.selector
81 /// in the given landing pad. In principle, llvm.eh.exception is
82 /// required to be in the landing pad; in practice, SplitCriticalEdge
83 /// can break that invariant, and then inlining can break it further.
84 /// There's a real need for a reliable solution here, but until that
85 /// happens, we have some fragile workarounds here.
86 static EHSelectorInst *findSelectorForLandingPad(BasicBlock *lpad) {
87 // Look for an exception call in the actual landing pad.
88 EHExceptionInst *exn = findExceptionInBlock(lpad);
89 if (exn) return findSelectorForException(exn);
91 // Okay, if that failed, look for one in an obvious successor. If
92 // we find one, we'll fix the IR by moving things back to the
95 bool dominates = true; // does the lpad dominate the exn call
96 BasicBlock *nonDominated = 0; // if not, the first non-dominated block
97 BasicBlock *lastDominated = 0; // and the block which branched to it
99 BasicBlock *exnBlock = lpad;
101 // We need to protect against lpads that lead into infinite loops.
102 SmallPtrSet<BasicBlock*,4> visited;
103 visited.insert(exnBlock);
106 // We're not going to apply this hack to anything more complicated
107 // than a series of unconditional branches, so if the block
108 // doesn't terminate in an unconditional branch, just fail. More
109 // complicated cases can arise when, say, sinking a call into a
110 // split unwind edge and then inlining it; but that can do almost
111 // *anything* to the CFG, including leaving the selector
112 // completely unreachable. The only way to fix that properly is
113 // to (1) prohibit transforms which move the exception or selector
114 // values away from the landing pad, e.g. by producing them with
115 // instructions that are pinned to an edge like a phi, or
116 // producing them with not-really-instructions, and (2) making
117 // transforms which split edges deal with that.
118 BranchInst *branch = dyn_cast<BranchInst>(&exnBlock->back());
119 if (!branch || branch->isConditional()) return 0;
121 BasicBlock *successor = branch->getSuccessor(0);
123 // Fail if we found an infinite loop.
124 if (!visited.insert(successor)) return 0;
126 // If the successor isn't dominated by exnBlock:
127 if (!successor->getSinglePredecessor()) {
128 // We don't want to have to deal with threading the exception
129 // through multiple levels of phi, so give up if we've already
130 // followed a non-dominating edge.
131 if (!dominates) return 0;
133 // Otherwise, remember this as a non-dominating edge.
135 nonDominated = successor;
136 lastDominated = exnBlock;
139 exnBlock = successor;
142 exn = findExceptionInBlock(exnBlock);
145 // Look for a selector call for the exception we found.
146 EHSelectorInst *selector = findSelectorForException(exn);
147 if (!selector) return 0;
149 // The easy case is when the landing pad still dominates the
150 // exception call, in which case we can just move both calls back to
153 selector->moveBefore(lpad->getFirstNonPHI());
154 exn->moveBefore(selector);
158 // Otherwise, we have to split at the first non-dominating block.
159 // The CFG looks basically like this:
162 // insnsAndBranches_1
163 // br label %nonDominated
167 // %exn = call i8* @llvm.eh.exception()
168 // insnsAndBranches_4
169 // %selector = call @llvm.eh.selector(i8* %exn, ...
170 // We need to turn this into:
173 // %exn0 = call i8* @llvm.eh.exception()
174 // %selector0 = call @llvm.eh.selector(i8* %exn0, ...
175 // insnsAndBranches_1
176 // br label %split // from lastDominated
178 // phis_2 (without edge from lastDominated)
179 // %exn1 = call i8* @llvm.eh.exception()
180 // %selector1 = call i8* @llvm.eh.selector(i8* %exn1, ...
183 // phis_2 (edge from lastDominated, edge from split)
185 // %selector = phi ...
187 // insnsAndBranches_4
189 assert(nonDominated);
190 assert(lastDominated);
192 // First, make clones of the intrinsics to go in lpad.
193 EHExceptionInst *lpadExn = cast<EHExceptionInst>(exn->clone());
194 EHSelectorInst *lpadSelector = cast<EHSelectorInst>(selector->clone());
195 lpadSelector->setArgOperand(0, lpadExn);
196 lpadSelector->insertBefore(lpad->getFirstNonPHI());
197 lpadExn->insertBefore(lpadSelector);
199 // Split the non-dominated block.
201 nonDominated->splitBasicBlock(nonDominated->getFirstNonPHI(),
202 nonDominated->getName() + ".lpad-fix");
204 // Redirect the last dominated branch there.
205 cast<BranchInst>(lastDominated->back()).setSuccessor(0, split);
207 // Move the existing intrinsics to the end of the old block.
208 selector->moveBefore(&nonDominated->back());
209 exn->moveBefore(selector);
211 Instruction *splitIP = &split->front();
213 // For all the phis in nonDominated, make a new phi in split to join
214 // that phi with the edge from lastDominated.
215 for (BasicBlock::iterator
216 i = nonDominated->begin(), e = nonDominated->end(); i != e; ++i) {
217 PHINode *phi = dyn_cast<PHINode>(i);
220 PHINode *splitPhi = PHINode::Create(phi->getType(), 2, phi->getName(),
222 phi->replaceAllUsesWith(splitPhi);
223 splitPhi->addIncoming(phi, nonDominated);
224 splitPhi->addIncoming(phi->removeIncomingValue(lastDominated),
228 // Make new phis for the exception and selector.
229 PHINode *exnPhi = PHINode::Create(exn->getType(), 2, "", splitIP);
230 exn->replaceAllUsesWith(exnPhi);
231 selector->setArgOperand(0, exn); // except for this use
232 exnPhi->addIncoming(exn, nonDominated);
233 exnPhi->addIncoming(lpadExn, lastDominated);
235 PHINode *selectorPhi = PHINode::Create(selector->getType(), 2, "", splitIP);
236 selector->replaceAllUsesWith(selectorPhi);
237 selectorPhi->addIncoming(selector, nonDominated);
238 selectorPhi->addIncoming(lpadSelector, lastDominated);
244 /// A class for recording information about inlining through an invoke.
245 class InvokeInliningInfo {
246 BasicBlock *OuterUnwindDest;
247 EHSelectorInst *OuterSelector;
248 BasicBlock *InnerUnwindDest;
249 PHINode *InnerExceptionPHI;
250 PHINode *InnerSelectorPHI;
251 SmallVector<Value*, 8> UnwindDestPHIValues;
254 BasicBlock *OuterResumeDest;
255 BasicBlock *InnerResumeDest;
256 LandingPadInst *CallerLPad;
257 PHINode *InnerEHValuesPHI;
258 BasicBlock *SplitLPad;
261 InvokeInliningInfo(InvokeInst *II)
262 : OuterUnwindDest(II->getUnwindDest()), OuterSelector(0),
263 InnerUnwindDest(0), InnerExceptionPHI(0), InnerSelectorPHI(0),
265 OuterResumeDest(II->getUnwindDest()), InnerResumeDest(0),
266 CallerLPad(0), InnerEHValuesPHI(0), SplitLPad(0) {
267 // If there are PHI nodes in the unwind destination block, we
268 // need to keep track of which values came into them from the
269 // invoke before removing the edge from this block.
270 llvm::BasicBlock *InvokeBB = II->getParent();
271 BasicBlock::iterator I = OuterUnwindDest->begin();
272 for (; isa<PHINode>(I); ++I) {
273 // Save the value to use for this edge.
274 PHINode *PHI = cast<PHINode>(I);
275 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
278 // FIXME: With the new EH, this if/dyn_cast should be a 'cast'.
279 if (LandingPadInst *LPI = dyn_cast<LandingPadInst>(I))
283 /// The outer unwind destination is the target of unwind edges
284 /// introduced for calls within the inlined function.
285 BasicBlock *getOuterUnwindDest() const {
286 return OuterUnwindDest;
289 EHSelectorInst *getOuterSelector() {
291 OuterSelector = findSelectorForLandingPad(OuterUnwindDest);
292 return OuterSelector;
295 BasicBlock *getInnerUnwindDest();
296 BasicBlock *getInnerUnwindDest_new();
298 LandingPadInst *getLandingPadInst() const { return CallerLPad; }
299 BasicBlock *getSplitLandingPad() {
300 if (SplitLPad) return SplitLPad;
301 assert(CallerLPad && "Trying to split a block that isn't a landing pad!");
302 BasicBlock::iterator I = CallerLPad; ++I;
303 SplitLPad = CallerLPad->getParent()->splitBasicBlock(I, "split.lpad");
307 bool forwardEHResume(CallInst *call, BasicBlock *src);
309 /// forwardResume - Forward the 'resume' instruction to the caller's landing
310 /// pad block. When the landing pad block has only one predecessor, this is
311 /// a simple branch. When there is more than one predecessor, we need to
312 /// split the landing pad block after the landingpad instruction and jump
314 void forwardResume(ResumeInst *RI);
316 /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind
317 /// destination block for the given basic block, using the values for the
318 /// original invoke's source block.
319 void addIncomingPHIValuesFor(BasicBlock *BB) const {
320 addIncomingPHIValuesForInto(BB, OuterUnwindDest);
322 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const {
323 BasicBlock::iterator I = dest->begin();
324 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
325 PHINode *PHI = cast<PHINode>(I);
326 PHI->addIncoming(UnwindDestPHIValues[i], src);
332 /// Get or create a target for the branch out of rewritten calls to
334 BasicBlock *InvokeInliningInfo::getInnerUnwindDest() {
335 if (InnerUnwindDest) return InnerUnwindDest;
337 // Find and hoist the llvm.eh.exception and llvm.eh.selector calls
338 // in the outer landing pad to immediately following the phis.
339 EHSelectorInst *selector = getOuterSelector();
340 if (!selector) return 0;
342 // The call to llvm.eh.exception *must* be in the landing pad.
343 Instruction *exn = cast<Instruction>(selector->getArgOperand(0));
344 assert(exn->getParent() == OuterUnwindDest);
346 // TODO: recognize when we've already done this, so that we don't
347 // get a linear number of these when inlining calls into lots of
348 // invokes with the same landing pad.
351 Instruction *splitPoint = exn->getParent()->getFirstNonPHI();
352 assert(splitPoint != selector && "selector-on-exception dominance broken!");
353 if (splitPoint == exn) {
354 selector->removeFromParent();
355 selector->insertAfter(exn);
356 splitPoint = selector->getNextNode();
358 exn->moveBefore(splitPoint);
359 selector->moveBefore(splitPoint);
362 // Split the landing pad.
363 InnerUnwindDest = OuterUnwindDest->splitBasicBlock(splitPoint,
364 OuterUnwindDest->getName() + ".body");
366 // The number of incoming edges we expect to the inner landing pad.
367 const unsigned phiCapacity = 2;
369 // Create corresponding new phis for all the phis in the outer landing pad.
370 BasicBlock::iterator insertPoint = InnerUnwindDest->begin();
371 BasicBlock::iterator I = OuterUnwindDest->begin();
372 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
373 PHINode *outerPhi = cast<PHINode>(I);
374 PHINode *innerPhi = PHINode::Create(outerPhi->getType(), phiCapacity,
375 outerPhi->getName() + ".lpad-body",
377 outerPhi->replaceAllUsesWith(innerPhi);
378 innerPhi->addIncoming(outerPhi, OuterUnwindDest);
381 // Create a phi for the exception value...
382 InnerExceptionPHI = PHINode::Create(exn->getType(), phiCapacity,
383 "exn.lpad-body", insertPoint);
384 exn->replaceAllUsesWith(InnerExceptionPHI);
385 selector->setArgOperand(0, exn); // restore this use
386 InnerExceptionPHI->addIncoming(exn, OuterUnwindDest);
388 // ...and the selector.
389 InnerSelectorPHI = PHINode::Create(selector->getType(), phiCapacity,
390 "selector.lpad-body", insertPoint);
391 selector->replaceAllUsesWith(InnerSelectorPHI);
392 InnerSelectorPHI->addIncoming(selector, OuterUnwindDest);
395 return InnerUnwindDest;
398 /// [LIBUNWIND] Try to forward the given call, which logically occurs
399 /// at the end of the given block, as a branch to the inner unwind
400 /// block. Returns true if the call was forwarded.
401 bool InvokeInliningInfo::forwardEHResume(CallInst *call, BasicBlock *src) {
402 // First, check whether this is a call to the intrinsic.
403 Function *fn = dyn_cast<Function>(call->getCalledValue());
404 if (!fn || fn->getName() != "llvm.eh.resume")
407 // At this point, we need to return true on all paths, because
408 // otherwise we'll construct an invoke of the intrinsic, which is
411 // Try to find or make an inner unwind dest, which will fail if we
412 // can't find a selector call for the outer unwind dest.
413 BasicBlock *dest = getInnerUnwindDest();
414 bool hasSelector = (dest != 0);
416 // If we failed, just use the outer unwind dest, dropping the
417 // exception and selector on the floor.
419 dest = OuterUnwindDest;
422 BranchInst::Create(dest, src);
424 // Update the phis in the destination. They were inserted in an
425 // order which makes this work.
426 addIncomingPHIValuesForInto(src, dest);
429 InnerExceptionPHI->addIncoming(call->getArgOperand(0), src);
430 InnerSelectorPHI->addIncoming(call->getArgOperand(1), src);
436 /// Get or create a target for the branch from ResumeInsts.
437 BasicBlock *InvokeInliningInfo::getInnerUnwindDest_new() {
438 if (InnerResumeDest) return InnerResumeDest;
440 // Split the landing pad.
441 BasicBlock::iterator SplitPoint = CallerLPad; ++SplitPoint;
443 OuterResumeDest->splitBasicBlock(SplitPoint,
444 OuterResumeDest->getName() + ".body");
446 // The number of incoming edges we expect to the inner landing pad.
447 const unsigned PHICapacity = 2;
449 // Create corresponding new PHIs for all the PHIs in the outer landing pad.
450 BasicBlock::iterator InsertPoint = InnerResumeDest->begin();
451 BasicBlock::iterator I = OuterResumeDest->begin();
452 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
453 PHINode *OuterPHI = cast<PHINode>(I);
454 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity,
455 OuterPHI->getName() + ".lpad-body",
457 OuterPHI->replaceAllUsesWith(InnerPHI);
458 InnerPHI->addIncoming(OuterPHI, OuterResumeDest);
461 // Create a PHI for the exception values.
462 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity,
463 "eh.lpad-body", InsertPoint);
464 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI);
465 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest);
468 return InnerResumeDest;
471 /// forwardResume - Forward the 'resume' instruction to the caller's landing pad
472 /// block. When the landing pad block has only one predecessor, this is a simple
473 /// branch. When there is more than one predecessor, we need to split the
474 /// landing pad block after the landingpad instruction and jump to there.
475 void InvokeInliningInfo::forwardResume(ResumeInst *RI) {
476 BasicBlock *Dest = getInnerUnwindDest_new();
477 BasicBlock *Src = RI->getParent();
479 BranchInst::Create(Dest, Src);
481 // Update the PHIs in the destination. They were inserted in an order which
483 addIncomingPHIValuesForInto(Src, Dest);
485 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src);
486 RI->eraseFromParent();
489 /// [LIBUNWIND] Check whether this selector is "only cleanups":
490 /// call i32 @llvm.eh.selector(blah, blah, i32 0)
491 static bool isCleanupOnlySelector(EHSelectorInst *selector) {
492 if (selector->getNumArgOperands() != 3) return false;
493 ConstantInt *val = dyn_cast<ConstantInt>(selector->getArgOperand(2));
494 return (val && val->isZero());
497 /// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
498 /// an invoke, we have to turn all of the calls that can throw into
499 /// invokes. This function analyze BB to see if there are any calls, and if so,
500 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
501 /// nodes in that block with the values specified in InvokeDestPHIValues.
503 /// Returns true to indicate that the next block should be skipped.
504 static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
505 InvokeInliningInfo &Invoke) {
506 LandingPadInst *LPI = Invoke.getLandingPadInst();
508 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
509 Instruction *I = BBI++;
511 if (LPI) // FIXME: This won't be NULL in the new EH.
512 if (LandingPadInst *L = dyn_cast<LandingPadInst>(I)) {
513 unsigned NumClauses = LPI->getNumClauses();
514 L->reserveClauses(NumClauses);
515 for (unsigned i = 0; i != NumClauses; ++i)
516 L->addClause(LPI->getClauseType(i), LPI->getClauseValue(i));
519 // We only need to check for function calls: inlined invoke
520 // instructions require no special handling.
521 CallInst *CI = dyn_cast<CallInst>(I);
522 if (CI == 0) continue;
524 // LIBUNWIND: merge selector instructions.
525 if (EHSelectorInst *Inner = dyn_cast<EHSelectorInst>(CI)) {
526 EHSelectorInst *Outer = Invoke.getOuterSelector();
527 if (!Outer) continue;
529 bool innerIsOnlyCleanup = isCleanupOnlySelector(Inner);
530 bool outerIsOnlyCleanup = isCleanupOnlySelector(Outer);
532 // If both selectors contain only cleanups, we don't need to do
533 // anything. TODO: this is really just a very specific instance
534 // of a much more general optimization.
535 if (innerIsOnlyCleanup && outerIsOnlyCleanup) continue;
537 // Otherwise, we just append the outer selector to the inner selector.
538 SmallVector<Value*, 16> NewSelector;
539 for (unsigned i = 0, e = Inner->getNumArgOperands(); i != e; ++i)
540 NewSelector.push_back(Inner->getArgOperand(i));
541 for (unsigned i = 2, e = Outer->getNumArgOperands(); i != e; ++i)
542 NewSelector.push_back(Outer->getArgOperand(i));
545 IRBuilder<>(Inner).CreateCall(Inner->getCalledValue(), NewSelector);
546 // No need to copy attributes, calling convention, etc.
547 NewInner->takeName(Inner);
548 Inner->replaceAllUsesWith(NewInner);
549 Inner->eraseFromParent();
553 // If this call cannot unwind, don't convert it to an invoke.
554 if (CI->doesNotThrow())
557 // Convert this function call into an invoke instruction.
558 // First, split the basic block.
559 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
561 // Delete the unconditional branch inserted by splitBasicBlock
562 BB->getInstList().pop_back();
564 // LIBUNWIND: If this is a call to @llvm.eh.resume, just branch
565 // directly to the new landing pad.
566 if (Invoke.forwardEHResume(CI, BB)) {
567 // TODO: 'Split' is now unreachable; clean it up.
569 // We want to leave the original call intact so that the call
570 // graph and other structures won't get misled. We also have to
571 // avoid processing the next block, or we'll iterate here forever.
575 // Otherwise, create the new invoke instruction.
576 ImmutableCallSite CS(CI);
577 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
579 InvokeInst::Create(CI->getCalledValue(), Split,
580 Invoke.getOuterUnwindDest(),
581 InvokeArgs, CI->getName(), BB);
582 II->setCallingConv(CI->getCallingConv());
583 II->setAttributes(CI->getAttributes());
585 // Make sure that anything using the call now uses the invoke! This also
586 // updates the CallGraph if present, because it uses a WeakVH.
587 CI->replaceAllUsesWith(II);
589 Split->getInstList().pop_front(); // Delete the original call
591 // Update any PHI nodes in the exceptional block to indicate that
592 // there is now a new entry in them.
593 Invoke.addIncomingPHIValuesFor(BB);
601 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
602 /// in the body of the inlined function into invokes and turn unwind
603 /// instructions into branches to the invoke unwind dest.
605 /// II is the invoke instruction being inlined. FirstNewBlock is the first
606 /// block of the inlined code (the last block is the end of the function),
607 /// and InlineCodeInfo is information about the code that got inlined.
608 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
609 ClonedCodeInfo &InlinedCodeInfo) {
610 BasicBlock *InvokeDest = II->getUnwindDest();
612 Function *Caller = FirstNewBlock->getParent();
614 // The inlined code is currently at the end of the function, scan from the
615 // start of the inlined code to its end, checking for stuff we need to
616 // rewrite. If the code doesn't have calls or unwinds, we know there is
617 // nothing to rewrite.
618 if (!InlinedCodeInfo.ContainsCalls && !InlinedCodeInfo.ContainsUnwinds) {
619 // Now that everything is happy, we have one final detail. The PHI nodes in
620 // the exception destination block still have entries due to the original
621 // invoke instruction. Eliminate these entries (which might even delete the
623 InvokeDest->removePredecessor(II->getParent());
627 InvokeInliningInfo Invoke(II);
629 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
630 if (InlinedCodeInfo.ContainsCalls)
631 if (HandleCallsInBlockInlinedThroughInvoke(BB, Invoke)) {
632 // Honor a request to skip the next block. We don't need to
633 // consider UnwindInsts in this case either.
638 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
639 // An UnwindInst requires special handling when it gets inlined into an
640 // invoke site. Once this happens, we know that the unwind would cause
641 // a control transfer to the invoke exception destination, so we can
642 // transform it into a direct branch to the exception destination.
643 BranchInst::Create(InvokeDest, UI);
645 // Delete the unwind instruction!
646 UI->eraseFromParent();
648 // Update any PHI nodes in the exceptional block to indicate that
649 // there is now a new entry in them.
650 Invoke.addIncomingPHIValuesFor(BB);
653 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
654 Invoke.forwardResume(RI);
658 // Now that everything is happy, we have one final detail. The PHI nodes in
659 // the exception destination block still have entries due to the original
660 // invoke instruction. Eliminate these entries (which might even delete the
662 InvokeDest->removePredecessor(II->getParent());
665 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
666 /// into the caller, update the specified callgraph to reflect the changes we
667 /// made. Note that it's possible that not all code was copied over, so only
668 /// some edges of the callgraph may remain.
669 static void UpdateCallGraphAfterInlining(CallSite CS,
670 Function::iterator FirstNewBlock,
671 ValueToValueMapTy &VMap,
672 InlineFunctionInfo &IFI) {
673 CallGraph &CG = *IFI.CG;
674 const Function *Caller = CS.getInstruction()->getParent()->getParent();
675 const Function *Callee = CS.getCalledFunction();
676 CallGraphNode *CalleeNode = CG[Callee];
677 CallGraphNode *CallerNode = CG[Caller];
679 // Since we inlined some uninlined call sites in the callee into the caller,
680 // add edges from the caller to all of the callees of the callee.
681 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
683 // Consider the case where CalleeNode == CallerNode.
684 CallGraphNode::CalledFunctionsVector CallCache;
685 if (CalleeNode == CallerNode) {
686 CallCache.assign(I, E);
687 I = CallCache.begin();
691 for (; I != E; ++I) {
692 const Value *OrigCall = I->first;
694 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
695 // Only copy the edge if the call was inlined!
696 if (VMI == VMap.end() || VMI->second == 0)
699 // If the call was inlined, but then constant folded, there is no edge to
700 // add. Check for this case.
701 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
702 if (NewCall == 0) continue;
704 // Remember that this call site got inlined for the client of
706 IFI.InlinedCalls.push_back(NewCall);
708 // It's possible that inlining the callsite will cause it to go from an
709 // indirect to a direct call by resolving a function pointer. If this
710 // happens, set the callee of the new call site to a more precise
711 // destination. This can also happen if the call graph node of the caller
712 // was just unnecessarily imprecise.
713 if (I->second->getFunction() == 0)
714 if (Function *F = CallSite(NewCall).getCalledFunction()) {
715 // Indirect call site resolved to direct call.
716 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
721 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
724 // Update the call graph by deleting the edge from Callee to Caller. We must
725 // do this after the loop above in case Caller and Callee are the same.
726 CallerNode->removeCallEdgeFor(CS);
729 /// HandleByValArgument - When inlining a call site that has a byval argument,
730 /// we have to make the implicit memcpy explicit by adding it.
731 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
732 const Function *CalledFunc,
733 InlineFunctionInfo &IFI,
734 unsigned ByValAlignment) {
735 Type *AggTy = cast<PointerType>(Arg->getType())->getElementType();
737 // If the called function is readonly, then it could not mutate the caller's
738 // copy of the byval'd memory. In this case, it is safe to elide the copy and
740 if (CalledFunc->onlyReadsMemory()) {
741 // If the byval argument has a specified alignment that is greater than the
742 // passed in pointer, then we either have to round up the input pointer or
743 // give up on this transformation.
744 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
747 // If the pointer is already known to be sufficiently aligned, or if we can
748 // round it up to a larger alignment, then we don't need a temporary.
749 if (getOrEnforceKnownAlignment(Arg, ByValAlignment,
750 IFI.TD) >= ByValAlignment)
753 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
754 // for code quality, but rarely happens and is required for correctness.
757 LLVMContext &Context = Arg->getContext();
759 Type *VoidPtrTy = Type::getInt8PtrTy(Context);
761 // Create the alloca. If we have TargetData, use nice alignment.
764 Align = IFI.TD->getPrefTypeAlignment(AggTy);
766 // If the byval had an alignment specified, we *must* use at least that
767 // alignment, as it is required by the byval argument (and uses of the
768 // pointer inside the callee).
769 Align = std::max(Align, ByValAlignment);
771 Function *Caller = TheCall->getParent()->getParent();
773 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(),
774 &*Caller->begin()->begin());
776 Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)};
777 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(),
780 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall);
781 Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall);
785 Size = ConstantExpr::getSizeOf(AggTy);
787 Size = ConstantInt::get(Type::getInt64Ty(Context),
788 IFI.TD->getTypeStoreSize(AggTy));
790 // Always generate a memcpy of alignment 1 here because we don't know
791 // the alignment of the src pointer. Other optimizations can infer
793 Value *CallArgs[] = {
794 DestCast, SrcCast, Size,
795 ConstantInt::get(Type::getInt32Ty(Context), 1),
796 ConstantInt::getFalse(Context) // isVolatile
798 IRBuilder<>(TheCall).CreateCall(MemCpyFn, CallArgs);
800 // Uses of the argument in the function should use our new alloca
805 // isUsedByLifetimeMarker - Check whether this Value is used by a lifetime
807 static bool isUsedByLifetimeMarker(Value *V) {
808 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE;
810 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI)) {
811 switch (II->getIntrinsicID()) {
813 case Intrinsic::lifetime_start:
814 case Intrinsic::lifetime_end:
822 // hasLifetimeMarkers - Check whether the given alloca already has
823 // lifetime.start or lifetime.end intrinsics.
824 static bool hasLifetimeMarkers(AllocaInst *AI) {
825 Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext());
826 if (AI->getType() == Int8PtrTy)
827 return isUsedByLifetimeMarker(AI);
829 // Do a scan to find all the casts to i8*.
830 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); I != E;
832 if (I->getType() != Int8PtrTy) continue;
833 if (I->stripPointerCasts() != AI) continue;
834 if (isUsedByLifetimeMarker(*I))
840 /// updateInlinedAtInfo - Helper function used by fixupLineNumbers to recursively
841 /// update InlinedAtEntry of a DebugLoc.
842 static DebugLoc updateInlinedAtInfo(const DebugLoc &DL,
843 const DebugLoc &InlinedAtDL,
845 if (MDNode *IA = DL.getInlinedAt(Ctx)) {
846 DebugLoc NewInlinedAtDL
847 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx);
848 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
849 NewInlinedAtDL.getAsMDNode(Ctx));
852 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
853 InlinedAtDL.getAsMDNode(Ctx));
857 /// fixupLineNumbers - Update inlined instructions' line numbers to
858 /// to encode location where these instructions are inlined.
859 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
860 Instruction *TheCall) {
861 DebugLoc TheCallDL = TheCall->getDebugLoc();
862 if (TheCallDL.isUnknown())
865 for (; FI != Fn->end(); ++FI) {
866 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
868 DebugLoc DL = BI->getDebugLoc();
869 if (!DL.isUnknown()) {
870 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext()));
871 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) {
872 LLVMContext &Ctx = BI->getContext();
873 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
874 DVI->setOperand(2, createInlinedVariable(DVI->getVariable(),
882 // InlineFunction - This function inlines the called function into the basic
883 // block of the caller. This returns false if it is not possible to inline this
884 // call. The program is still in a well defined state if this occurs though.
886 // Note that this only does one level of inlining. For example, if the
887 // instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
888 // exists in the instruction stream. Similarly this will inline a recursive
889 // function by one level.
891 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI) {
892 Instruction *TheCall = CS.getInstruction();
893 LLVMContext &Context = TheCall->getContext();
894 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
895 "Instruction not in function!");
897 // If IFI has any state in it, zap it before we fill it in.
900 const Function *CalledFunc = CS.getCalledFunction();
901 if (CalledFunc == 0 || // Can't inline external function or indirect
902 CalledFunc->isDeclaration() || // call, or call to a vararg function!
903 CalledFunc->getFunctionType()->isVarArg()) return false;
905 // If the call to the callee is not a tail call, we must clear the 'tail'
906 // flags on any calls that we inline.
907 bool MustClearTailCallFlags =
908 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall());
910 // If the call to the callee cannot throw, set the 'nounwind' flag on any
911 // calls that we inline.
912 bool MarkNoUnwind = CS.doesNotThrow();
914 BasicBlock *OrigBB = TheCall->getParent();
915 Function *Caller = OrigBB->getParent();
917 // GC poses two hazards to inlining, which only occur when the callee has GC:
918 // 1. If the caller has no GC, then the callee's GC must be propagated to the
920 // 2. If the caller has a differing GC, it is invalid to inline.
921 if (CalledFunc->hasGC()) {
922 if (!Caller->hasGC())
923 Caller->setGC(CalledFunc->getGC());
924 else if (CalledFunc->getGC() != Caller->getGC())
928 // Find the personality function used by the landing pads of the caller. If it
929 // exists, then check to see that it matches the personality function used in
931 for (Function::const_iterator
932 I = Caller->begin(), E = Caller->end(); I != E; ++I)
933 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
934 const BasicBlock *BB = II->getUnwindDest();
935 // FIXME: This 'isa' here should become go away once the new EH system is
937 if (!isa<LandingPadInst>(BB->getFirstNonPHI()))
939 const LandingPadInst *LP = cast<LandingPadInst>(BB->getFirstNonPHI());
940 const Value *CallerPersFn = LP->getPersonalityFn();
942 // If the personality functions match, then we can perform the
943 // inlining. Otherwise, we can't inline.
944 // TODO: This isn't 100% true. Some personality functions are proper
945 // supersets of others and can be used in place of the other.
946 for (Function::const_iterator
947 I = CalledFunc->begin(), E = CalledFunc->end(); I != E; ++I)
948 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
949 const BasicBlock *BB = II->getUnwindDest();
950 // FIXME: This 'if/dyn_cast' here should become a normal 'cast' once
951 // the new EH system is in place.
952 if (const LandingPadInst *LP =
953 dyn_cast<LandingPadInst>(BB->getFirstNonPHI()))
954 if (CallerPersFn != LP->getPersonalityFn())
962 // Get an iterator to the last basic block in the function, which will have
963 // the new function inlined after it.
965 Function::iterator LastBlock = &Caller->back();
967 // Make sure to capture all of the return instructions from the cloned
969 SmallVector<ReturnInst*, 8> Returns;
970 ClonedCodeInfo InlinedFunctionInfo;
971 Function::iterator FirstNewBlock;
973 { // Scope to destroy VMap after cloning.
974 ValueToValueMapTy VMap;
976 assert(CalledFunc->arg_size() == CS.arg_size() &&
977 "No varargs calls can be inlined!");
979 // Calculate the vector of arguments to pass into the function cloner, which
980 // matches up the formal to the actual argument values.
981 CallSite::arg_iterator AI = CS.arg_begin();
983 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
984 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
985 Value *ActualArg = *AI;
987 // When byval arguments actually inlined, we need to make the copy implied
988 // by them explicit. However, we don't do this if the callee is readonly
989 // or readnone, because the copy would be unneeded: the callee doesn't
990 // modify the struct.
991 if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal)) {
992 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
993 CalledFunc->getParamAlignment(ArgNo+1));
995 // Calls that we inline may use the new alloca, so we need to clear
996 // their 'tail' flags if HandleByValArgument introduced a new alloca and
997 // the callee has calls.
998 MustClearTailCallFlags |= ActualArg != *AI;
1001 VMap[I] = ActualArg;
1004 // We want the inliner to prune the code as it copies. We would LOVE to
1005 // have no dead or constant instructions leftover after inlining occurs
1006 // (which can happen, e.g., because an argument was constant), but we'll be
1007 // happy with whatever the cloner can do.
1008 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
1009 /*ModuleLevelChanges=*/false, Returns, ".i",
1010 &InlinedFunctionInfo, IFI.TD, TheCall);
1012 // Remember the first block that is newly cloned over.
1013 FirstNewBlock = LastBlock; ++FirstNewBlock;
1015 // Update the callgraph if requested.
1017 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
1019 // Update inlined instructions' line number information.
1020 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
1023 // If there are any alloca instructions in the block that used to be the entry
1024 // block for the callee, move them to the entry block of the caller. First
1025 // calculate which instruction they should be inserted before. We insert the
1026 // instructions at the end of the current alloca list.
1029 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
1030 for (BasicBlock::iterator I = FirstNewBlock->begin(),
1031 E = FirstNewBlock->end(); I != E; ) {
1032 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
1033 if (AI == 0) continue;
1035 // If the alloca is now dead, remove it. This often occurs due to code
1037 if (AI->use_empty()) {
1038 AI->eraseFromParent();
1042 if (!isa<Constant>(AI->getArraySize()))
1045 // Keep track of the static allocas that we inline into the caller.
1046 IFI.StaticAllocas.push_back(AI);
1048 // Scan for the block of allocas that we can move over, and move them
1050 while (isa<AllocaInst>(I) &&
1051 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
1052 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
1056 // Transfer all of the allocas over in a block. Using splice means
1057 // that the instructions aren't removed from the symbol table, then
1059 Caller->getEntryBlock().getInstList().splice(InsertPoint,
1060 FirstNewBlock->getInstList(),
1065 // Leave lifetime markers for the static alloca's, scoping them to the
1066 // function we just inlined.
1067 if (!IFI.StaticAllocas.empty()) {
1068 IRBuilder<> builder(FirstNewBlock->begin());
1069 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
1070 AllocaInst *AI = IFI.StaticAllocas[ai];
1072 // If the alloca is already scoped to something smaller than the whole
1073 // function then there's no need to add redundant, less accurate markers.
1074 if (hasLifetimeMarkers(AI))
1077 builder.CreateLifetimeStart(AI);
1078 for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) {
1079 IRBuilder<> builder(Returns[ri]);
1080 builder.CreateLifetimeEnd(AI);
1085 // If the inlined code contained dynamic alloca instructions, wrap the inlined
1086 // code with llvm.stacksave/llvm.stackrestore intrinsics.
1087 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
1088 Module *M = Caller->getParent();
1089 // Get the two intrinsics we care about.
1090 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
1091 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
1093 // Insert the llvm.stacksave.
1094 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
1095 .CreateCall(StackSave, "savedstack");
1097 // Insert a call to llvm.stackrestore before any return instructions in the
1098 // inlined function.
1099 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1100 IRBuilder<>(Returns[i]).CreateCall(StackRestore, SavedPtr);
1103 // Count the number of StackRestore calls we insert.
1104 unsigned NumStackRestores = Returns.size();
1106 // If we are inlining an invoke instruction, insert restores before each
1107 // unwind. These unwinds will be rewritten into branches later.
1108 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) {
1109 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
1111 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
1112 IRBuilder<>(UI).CreateCall(StackRestore, SavedPtr);
1118 // If we are inlining tail call instruction through a call site that isn't
1119 // marked 'tail', we must remove the tail marker for any calls in the inlined
1120 // code. Also, calls inlined through a 'nounwind' call site should be marked
1122 if (InlinedFunctionInfo.ContainsCalls &&
1123 (MustClearTailCallFlags || MarkNoUnwind)) {
1124 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
1126 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
1127 if (CallInst *CI = dyn_cast<CallInst>(I)) {
1128 if (MustClearTailCallFlags)
1129 CI->setTailCall(false);
1131 CI->setDoesNotThrow();
1135 // If we are inlining through a 'nounwind' call site then any inlined 'unwind'
1136 // instructions are unreachable.
1137 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind)
1138 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
1140 TerminatorInst *Term = BB->getTerminator();
1141 if (isa<UnwindInst>(Term)) {
1142 new UnreachableInst(Context, Term);
1143 BB->getInstList().erase(Term);
1147 // If we are inlining for an invoke instruction, we must make sure to rewrite
1148 // any inlined 'unwind' instructions into branches to the invoke exception
1149 // destination, and call instructions into invoke instructions.
1150 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
1151 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
1153 // If we cloned in _exactly one_ basic block, and if that block ends in a
1154 // return instruction, we splice the body of the inlined callee directly into
1155 // the calling basic block.
1156 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
1157 // Move all of the instructions right before the call.
1158 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
1159 FirstNewBlock->begin(), FirstNewBlock->end());
1160 // Remove the cloned basic block.
1161 Caller->getBasicBlockList().pop_back();
1163 // If the call site was an invoke instruction, add a branch to the normal
1165 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
1166 BranchInst::Create(II->getNormalDest(), TheCall);
1168 // If the return instruction returned a value, replace uses of the call with
1169 // uses of the returned value.
1170 if (!TheCall->use_empty()) {
1171 ReturnInst *R = Returns[0];
1172 if (TheCall == R->getReturnValue())
1173 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1175 TheCall->replaceAllUsesWith(R->getReturnValue());
1177 // Since we are now done with the Call/Invoke, we can delete it.
1178 TheCall->eraseFromParent();
1180 // Since we are now done with the return instruction, delete it also.
1181 Returns[0]->eraseFromParent();
1183 // We are now done with the inlining.
1187 // Otherwise, we have the normal case, of more than one block to inline or
1188 // multiple return sites.
1190 // We want to clone the entire callee function into the hole between the
1191 // "starter" and "ender" blocks. How we accomplish this depends on whether
1192 // this is an invoke instruction or a call instruction.
1193 BasicBlock *AfterCallBB;
1194 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1196 // Add an unconditional branch to make this look like the CallInst case...
1197 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
1199 // Split the basic block. This guarantees that no PHI nodes will have to be
1200 // updated due to new incoming edges, and make the invoke case more
1201 // symmetric to the call case.
1202 AfterCallBB = OrigBB->splitBasicBlock(NewBr,
1203 CalledFunc->getName()+".exit");
1205 } else { // It's a call
1206 // If this is a call instruction, we need to split the basic block that
1207 // the call lives in.
1209 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
1210 CalledFunc->getName()+".exit");
1213 // Change the branch that used to go to AfterCallBB to branch to the first
1214 // basic block of the inlined function.
1216 TerminatorInst *Br = OrigBB->getTerminator();
1217 assert(Br && Br->getOpcode() == Instruction::Br &&
1218 "splitBasicBlock broken!");
1219 Br->setOperand(0, FirstNewBlock);
1222 // Now that the function is correct, make it a little bit nicer. In
1223 // particular, move the basic blocks inserted from the end of the function
1224 // into the space made by splitting the source basic block.
1225 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
1226 FirstNewBlock, Caller->end());
1228 // Handle all of the return instructions that we just cloned in, and eliminate
1229 // any users of the original call/invoke instruction.
1230 Type *RTy = CalledFunc->getReturnType();
1233 if (Returns.size() > 1) {
1234 // The PHI node should go at the front of the new basic block to merge all
1235 // possible incoming values.
1236 if (!TheCall->use_empty()) {
1237 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
1238 AfterCallBB->begin());
1239 // Anything that used the result of the function call should now use the
1240 // PHI node as their operand.
1241 TheCall->replaceAllUsesWith(PHI);
1244 // Loop over all of the return instructions adding entries to the PHI node
1247 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1248 ReturnInst *RI = Returns[i];
1249 assert(RI->getReturnValue()->getType() == PHI->getType() &&
1250 "Ret value not consistent in function!");
1251 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
1256 // Add a branch to the merge points and remove return instructions.
1257 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1258 ReturnInst *RI = Returns[i];
1259 BranchInst::Create(AfterCallBB, RI);
1260 RI->eraseFromParent();
1262 } else if (!Returns.empty()) {
1263 // Otherwise, if there is exactly one return value, just replace anything
1264 // using the return value of the call with the computed value.
1265 if (!TheCall->use_empty()) {
1266 if (TheCall == Returns[0]->getReturnValue())
1267 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1269 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
1272 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
1273 BasicBlock *ReturnBB = Returns[0]->getParent();
1274 ReturnBB->replaceAllUsesWith(AfterCallBB);
1276 // Splice the code from the return block into the block that it will return
1277 // to, which contains the code that was after the call.
1278 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
1279 ReturnBB->getInstList());
1281 // Delete the return instruction now and empty ReturnBB now.
1282 Returns[0]->eraseFromParent();
1283 ReturnBB->eraseFromParent();
1284 } else if (!TheCall->use_empty()) {
1285 // No returns, but something is using the return value of the call. Just
1287 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1290 // Since we are now done with the Call/Invoke, we can delete it.
1291 TheCall->eraseFromParent();
1293 // We should always be able to fold the entry block of the function into the
1294 // single predecessor of the block...
1295 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
1296 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
1298 // Splice the code entry block into calling block, right before the
1299 // unconditional branch.
1300 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
1301 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
1303 // Remove the unconditional branch.
1304 OrigBB->getInstList().erase(Br);
1306 // Now we can remove the CalleeEntry block, which is now empty.
1307 Caller->getBasicBlockList().erase(CalleeEntry);
1309 // If we inserted a phi node, check to see if it has a single value (e.g. all
1310 // the entries are the same or undef). If so, remove the PHI so it doesn't
1311 // block other optimizations.
1313 if (Value *V = SimplifyInstruction(PHI, IFI.TD)) {
1314 PHI->replaceAllUsesWith(V);
1315 PHI->eraseFromParent();