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 // FIXME: New EH - Remove the functions marked [LIBUNWIND] when new EH is
51 /// [LIBUNWIND] Look for an llvm.eh.exception call in the given block.
52 static EHExceptionInst *findExceptionInBlock(BasicBlock *bb) {
53 for (BasicBlock::iterator i = bb->begin(), e = bb->end(); i != e; i++) {
54 EHExceptionInst *exn = dyn_cast<EHExceptionInst>(i);
61 /// [LIBUNWIND] Look for the 'best' llvm.eh.selector instruction for
62 /// the given llvm.eh.exception call.
63 static EHSelectorInst *findSelectorForException(EHExceptionInst *exn) {
64 BasicBlock *exnBlock = exn->getParent();
66 EHSelectorInst *outOfBlockSelector = 0;
67 for (Instruction::use_iterator
68 ui = exn->use_begin(), ue = exn->use_end(); ui != ue; ++ui) {
69 EHSelectorInst *sel = dyn_cast<EHSelectorInst>(*ui);
72 // Immediately accept an eh.selector in the same block as the
74 if (sel->getParent() == exnBlock) return sel;
76 // Otherwise, use the first selector we see.
77 if (!outOfBlockSelector) outOfBlockSelector = sel;
80 return outOfBlockSelector;
83 /// [LIBUNWIND] Find the (possibly absent) call to @llvm.eh.selector
84 /// in the given landing pad. In principle, llvm.eh.exception is
85 /// required to be in the landing pad; in practice, SplitCriticalEdge
86 /// can break that invariant, and then inlining can break it further.
87 /// There's a real need for a reliable solution here, but until that
88 /// happens, we have some fragile workarounds here.
89 static EHSelectorInst *findSelectorForLandingPad(BasicBlock *lpad) {
90 // Look for an exception call in the actual landing pad.
91 EHExceptionInst *exn = findExceptionInBlock(lpad);
92 if (exn) return findSelectorForException(exn);
94 // Okay, if that failed, look for one in an obvious successor. If
95 // we find one, we'll fix the IR by moving things back to the
98 bool dominates = true; // does the lpad dominate the exn call
99 BasicBlock *nonDominated = 0; // if not, the first non-dominated block
100 BasicBlock *lastDominated = 0; // and the block which branched to it
102 BasicBlock *exnBlock = lpad;
104 // We need to protect against lpads that lead into infinite loops.
105 SmallPtrSet<BasicBlock*,4> visited;
106 visited.insert(exnBlock);
109 // We're not going to apply this hack to anything more complicated
110 // than a series of unconditional branches, so if the block
111 // doesn't terminate in an unconditional branch, just fail. More
112 // complicated cases can arise when, say, sinking a call into a
113 // split unwind edge and then inlining it; but that can do almost
114 // *anything* to the CFG, including leaving the selector
115 // completely unreachable. The only way to fix that properly is
116 // to (1) prohibit transforms which move the exception or selector
117 // values away from the landing pad, e.g. by producing them with
118 // instructions that are pinned to an edge like a phi, or
119 // producing them with not-really-instructions, and (2) making
120 // transforms which split edges deal with that.
121 BranchInst *branch = dyn_cast<BranchInst>(&exnBlock->back());
122 if (!branch || branch->isConditional()) return 0;
124 BasicBlock *successor = branch->getSuccessor(0);
126 // Fail if we found an infinite loop.
127 if (!visited.insert(successor)) return 0;
129 // If the successor isn't dominated by exnBlock:
130 if (!successor->getSinglePredecessor()) {
131 // We don't want to have to deal with threading the exception
132 // through multiple levels of phi, so give up if we've already
133 // followed a non-dominating edge.
134 if (!dominates) return 0;
136 // Otherwise, remember this as a non-dominating edge.
138 nonDominated = successor;
139 lastDominated = exnBlock;
142 exnBlock = successor;
145 exn = findExceptionInBlock(exnBlock);
148 // Look for a selector call for the exception we found.
149 EHSelectorInst *selector = findSelectorForException(exn);
150 if (!selector) return 0;
152 // The easy case is when the landing pad still dominates the
153 // exception call, in which case we can just move both calls back to
156 selector->moveBefore(lpad->getFirstNonPHI());
157 exn->moveBefore(selector);
161 // Otherwise, we have to split at the first non-dominating block.
162 // The CFG looks basically like this:
165 // insnsAndBranches_1
166 // br label %nonDominated
170 // %exn = call i8* @llvm.eh.exception()
171 // insnsAndBranches_4
172 // %selector = call @llvm.eh.selector(i8* %exn, ...
173 // We need to turn this into:
176 // %exn0 = call i8* @llvm.eh.exception()
177 // %selector0 = call @llvm.eh.selector(i8* %exn0, ...
178 // insnsAndBranches_1
179 // br label %split // from lastDominated
181 // phis_2 (without edge from lastDominated)
182 // %exn1 = call i8* @llvm.eh.exception()
183 // %selector1 = call i8* @llvm.eh.selector(i8* %exn1, ...
186 // phis_2 (edge from lastDominated, edge from split)
188 // %selector = phi ...
190 // insnsAndBranches_4
192 assert(nonDominated);
193 assert(lastDominated);
195 // First, make clones of the intrinsics to go in lpad.
196 EHExceptionInst *lpadExn = cast<EHExceptionInst>(exn->clone());
197 EHSelectorInst *lpadSelector = cast<EHSelectorInst>(selector->clone());
198 lpadSelector->setArgOperand(0, lpadExn);
199 lpadSelector->insertBefore(lpad->getFirstNonPHI());
200 lpadExn->insertBefore(lpadSelector);
202 // Split the non-dominated block.
204 nonDominated->splitBasicBlock(nonDominated->getFirstNonPHI(),
205 nonDominated->getName() + ".lpad-fix");
207 // Redirect the last dominated branch there.
208 cast<BranchInst>(lastDominated->back()).setSuccessor(0, split);
210 // Move the existing intrinsics to the end of the old block.
211 selector->moveBefore(&nonDominated->back());
212 exn->moveBefore(selector);
214 Instruction *splitIP = &split->front();
216 // For all the phis in nonDominated, make a new phi in split to join
217 // that phi with the edge from lastDominated.
218 for (BasicBlock::iterator
219 i = nonDominated->begin(), e = nonDominated->end(); i != e; ++i) {
220 PHINode *phi = dyn_cast<PHINode>(i);
223 PHINode *splitPhi = PHINode::Create(phi->getType(), 2, phi->getName(),
225 phi->replaceAllUsesWith(splitPhi);
226 splitPhi->addIncoming(phi, nonDominated);
227 splitPhi->addIncoming(phi->removeIncomingValue(lastDominated),
231 // Make new phis for the exception and selector.
232 PHINode *exnPhi = PHINode::Create(exn->getType(), 2, "", splitIP);
233 exn->replaceAllUsesWith(exnPhi);
234 selector->setArgOperand(0, exn); // except for this use
235 exnPhi->addIncoming(exn, nonDominated);
236 exnPhi->addIncoming(lpadExn, lastDominated);
238 PHINode *selectorPhi = PHINode::Create(selector->getType(), 2, "", splitIP);
239 selector->replaceAllUsesWith(selectorPhi);
240 selectorPhi->addIncoming(selector, nonDominated);
241 selectorPhi->addIncoming(lpadSelector, lastDominated);
247 /// A class for recording information about inlining through an invoke.
248 class InvokeInliningInfo {
249 BasicBlock *OuterUnwindDest;
250 EHSelectorInst *OuterSelector;
251 BasicBlock *InnerUnwindDest;
252 PHINode *InnerExceptionPHI;
253 PHINode *InnerSelectorPHI;
254 SmallVector<Value*, 8> UnwindDestPHIValues;
256 // FIXME: New EH - These will replace the analogous ones above.
257 BasicBlock *OuterResumeDest; //< Destination of the invoke's unwind.
258 BasicBlock *InnerResumeDest; //< Destination for the callee's resume.
259 LandingPadInst *CallerLPad; //< LandingPadInst associated with the invoke.
260 PHINode *InnerEHValuesPHI; //< PHI for EH values from landingpad insts.
263 InvokeInliningInfo(InvokeInst *II)
264 : OuterUnwindDest(II->getUnwindDest()), OuterSelector(0),
265 InnerUnwindDest(0), InnerExceptionPHI(0), InnerSelectorPHI(0),
266 OuterResumeDest(II->getUnwindDest()), InnerResumeDest(0),
267 CallerLPad(0), InnerEHValuesPHI(0) {
268 // If there are PHI nodes in the unwind destination block, we need to keep
269 // track of which values came into them from the invoke before removing
270 // the edge from this block.
271 llvm::BasicBlock *InvokeBB = II->getParent();
272 BasicBlock::iterator I = OuterUnwindDest->begin();
273 for (; isa<PHINode>(I); ++I) {
274 // Save the value to use for this edge.
275 PHINode *PHI = cast<PHINode>(I);
276 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
279 CallerLPad = cast<LandingPadInst>(I);
282 /// The outer unwind destination is the target of unwind edges
283 /// introduced for calls within the inlined function.
284 BasicBlock *getOuterUnwindDest() const {
285 return OuterUnwindDest;
288 EHSelectorInst *getOuterSelector() {
290 OuterSelector = findSelectorForLandingPad(OuterUnwindDest);
291 return OuterSelector;
294 BasicBlock *getInnerUnwindDest();
296 // FIXME: New EH - Rename when new EH is turned on.
297 BasicBlock *getInnerUnwindDestNewEH();
299 LandingPadInst *getLandingPadInst() const { return CallerLPad; }
301 bool forwardEHResume(CallInst *call, BasicBlock *src);
303 /// forwardResume - Forward the 'resume' instruction to the caller's landing
304 /// pad block. When the landing pad block has only one predecessor, this is
305 /// a simple branch. When there is more than one predecessor, we need to
306 /// split the landing pad block after the landingpad instruction and jump
308 void forwardResume(ResumeInst *RI);
310 /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind
311 /// destination block for the given basic block, using the values for the
312 /// original invoke's source block.
313 void addIncomingPHIValuesFor(BasicBlock *BB) const {
314 addIncomingPHIValuesForInto(BB, OuterUnwindDest);
317 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const {
318 BasicBlock::iterator I = dest->begin();
319 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
320 PHINode *phi = cast<PHINode>(I);
321 phi->addIncoming(UnwindDestPHIValues[i], src);
327 /// [LIBUNWIND] Get or create a target for the branch out of rewritten calls to
329 BasicBlock *InvokeInliningInfo::getInnerUnwindDest() {
330 if (InnerUnwindDest) return InnerUnwindDest;
332 // Find and hoist the llvm.eh.exception and llvm.eh.selector calls
333 // in the outer landing pad to immediately following the phis.
334 EHSelectorInst *selector = getOuterSelector();
335 if (!selector) return 0;
337 // The call to llvm.eh.exception *must* be in the landing pad.
338 Instruction *exn = cast<Instruction>(selector->getArgOperand(0));
339 assert(exn->getParent() == OuterUnwindDest);
341 // TODO: recognize when we've already done this, so that we don't
342 // get a linear number of these when inlining calls into lots of
343 // invokes with the same landing pad.
346 Instruction *splitPoint = exn->getParent()->getFirstNonPHI();
347 assert(splitPoint != selector && "selector-on-exception dominance broken!");
348 if (splitPoint == exn) {
349 selector->removeFromParent();
350 selector->insertAfter(exn);
351 splitPoint = selector->getNextNode();
353 exn->moveBefore(splitPoint);
354 selector->moveBefore(splitPoint);
357 // Split the landing pad.
358 InnerUnwindDest = OuterUnwindDest->splitBasicBlock(splitPoint,
359 OuterUnwindDest->getName() + ".body");
361 // The number of incoming edges we expect to the inner landing pad.
362 const unsigned phiCapacity = 2;
364 // Create corresponding new phis for all the phis in the outer landing pad.
365 BasicBlock::iterator insertPoint = InnerUnwindDest->begin();
366 BasicBlock::iterator I = OuterUnwindDest->begin();
367 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
368 PHINode *outerPhi = cast<PHINode>(I);
369 PHINode *innerPhi = PHINode::Create(outerPhi->getType(), phiCapacity,
370 outerPhi->getName() + ".lpad-body",
372 outerPhi->replaceAllUsesWith(innerPhi);
373 innerPhi->addIncoming(outerPhi, OuterUnwindDest);
376 // Create a phi for the exception value...
377 InnerExceptionPHI = PHINode::Create(exn->getType(), phiCapacity,
378 "exn.lpad-body", insertPoint);
379 exn->replaceAllUsesWith(InnerExceptionPHI);
380 selector->setArgOperand(0, exn); // restore this use
381 InnerExceptionPHI->addIncoming(exn, OuterUnwindDest);
383 // ...and the selector.
384 InnerSelectorPHI = PHINode::Create(selector->getType(), phiCapacity,
385 "selector.lpad-body", insertPoint);
386 selector->replaceAllUsesWith(InnerSelectorPHI);
387 InnerSelectorPHI->addIncoming(selector, OuterUnwindDest);
390 return InnerUnwindDest;
393 /// [LIBUNWIND] Try to forward the given call, which logically occurs
394 /// at the end of the given block, as a branch to the inner unwind
395 /// block. Returns true if the call was forwarded.
396 bool InvokeInliningInfo::forwardEHResume(CallInst *call, BasicBlock *src) {
397 // First, check whether this is a call to the intrinsic.
398 Function *fn = dyn_cast<Function>(call->getCalledValue());
399 if (!fn || fn->getName() != "llvm.eh.resume")
402 // At this point, we need to return true on all paths, because
403 // otherwise we'll construct an invoke of the intrinsic, which is
406 // Try to find or make an inner unwind dest, which will fail if we
407 // can't find a selector call for the outer unwind dest.
408 BasicBlock *dest = getInnerUnwindDest();
409 bool hasSelector = (dest != 0);
411 // If we failed, just use the outer unwind dest, dropping the
412 // exception and selector on the floor.
414 dest = OuterUnwindDest;
417 BranchInst::Create(dest, src);
419 // Update the phis in the destination. They were inserted in an
420 // order which makes this work.
421 addIncomingPHIValuesForInto(src, dest);
424 InnerExceptionPHI->addIncoming(call->getArgOperand(0), src);
425 InnerSelectorPHI->addIncoming(call->getArgOperand(1), src);
431 /// Get or create a target for the branch from ResumeInsts.
432 BasicBlock *InvokeInliningInfo::getInnerUnwindDestNewEH() {
433 // FIXME: New EH - rename this function when new EH is turned on.
434 if (InnerResumeDest) return InnerResumeDest;
436 // Split the landing pad.
437 BasicBlock::iterator SplitPoint = CallerLPad; ++SplitPoint;
439 OuterResumeDest->splitBasicBlock(SplitPoint,
440 OuterResumeDest->getName() + ".body");
442 // The number of incoming edges we expect to the inner landing pad.
443 const unsigned PHICapacity = 2;
445 // Create corresponding new PHIs for all the PHIs in the outer landing pad.
446 BasicBlock::iterator InsertPoint = InnerResumeDest->begin();
447 BasicBlock::iterator I = OuterResumeDest->begin();
448 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
449 PHINode *OuterPHI = cast<PHINode>(I);
450 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity,
451 OuterPHI->getName() + ".lpad-body",
453 OuterPHI->replaceAllUsesWith(InnerPHI);
454 InnerPHI->addIncoming(OuterPHI, OuterResumeDest);
457 // Create a PHI for the exception values.
458 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity,
459 "eh.lpad-body", InsertPoint);
460 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI);
461 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest);
464 return InnerResumeDest;
467 /// forwardResume - Forward the 'resume' instruction to the caller's landing pad
468 /// block. When the landing pad block has only one predecessor, this is a simple
469 /// branch. When there is more than one predecessor, we need to split the
470 /// landing pad block after the landingpad instruction and jump to there.
471 void InvokeInliningInfo::forwardResume(ResumeInst *RI) {
472 BasicBlock *Dest = getInnerUnwindDestNewEH();
473 BasicBlock *Src = RI->getParent();
475 BranchInst::Create(Dest, Src);
477 // Update the PHIs in the destination. They were inserted in an order which
479 addIncomingPHIValuesForInto(Src, Dest);
481 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src);
482 RI->eraseFromParent();
485 /// [LIBUNWIND] Check whether this selector is "only cleanups":
486 /// call i32 @llvm.eh.selector(blah, blah, i32 0)
487 static bool isCleanupOnlySelector(EHSelectorInst *selector) {
488 if (selector->getNumArgOperands() != 3) return false;
489 ConstantInt *val = dyn_cast<ConstantInt>(selector->getArgOperand(2));
490 return (val && val->isZero());
493 /// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
494 /// an invoke, we have to turn all of the calls that can throw into
495 /// invokes. This function analyze BB to see if there are any calls, and if so,
496 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
497 /// nodes in that block with the values specified in InvokeDestPHIValues.
499 /// Returns true to indicate that the next block should be skipped.
500 static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
501 InvokeInliningInfo &Invoke) {
502 LandingPadInst *LPI = Invoke.getLandingPadInst();
504 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
505 Instruction *I = BBI++;
507 if (LandingPadInst *L = dyn_cast<LandingPadInst>(I)) {
508 unsigned NumClauses = LPI->getNumClauses();
509 L->reserveClauses(NumClauses);
510 for (unsigned i = 0; i != NumClauses; ++i)
511 L->addClause(LPI->getClause(i));
514 // We only need to check for function calls: inlined invoke
515 // instructions require no special handling.
516 CallInst *CI = dyn_cast<CallInst>(I);
517 if (CI == 0) continue;
519 // LIBUNWIND: merge selector instructions.
520 if (EHSelectorInst *Inner = dyn_cast<EHSelectorInst>(CI)) {
521 EHSelectorInst *Outer = Invoke.getOuterSelector();
522 if (!Outer) continue;
524 bool innerIsOnlyCleanup = isCleanupOnlySelector(Inner);
525 bool outerIsOnlyCleanup = isCleanupOnlySelector(Outer);
527 // If both selectors contain only cleanups, we don't need to do
528 // anything. TODO: this is really just a very specific instance
529 // of a much more general optimization.
530 if (innerIsOnlyCleanup && outerIsOnlyCleanup) continue;
532 // Otherwise, we just append the outer selector to the inner selector.
533 SmallVector<Value*, 16> NewSelector;
534 for (unsigned i = 0, e = Inner->getNumArgOperands(); i != e; ++i)
535 NewSelector.push_back(Inner->getArgOperand(i));
536 for (unsigned i = 2, e = Outer->getNumArgOperands(); i != e; ++i)
537 NewSelector.push_back(Outer->getArgOperand(i));
540 IRBuilder<>(Inner).CreateCall(Inner->getCalledValue(), NewSelector);
541 // No need to copy attributes, calling convention, etc.
542 NewInner->takeName(Inner);
543 Inner->replaceAllUsesWith(NewInner);
544 Inner->eraseFromParent();
548 // If this call cannot unwind, don't convert it to an invoke.
549 if (CI->doesNotThrow())
552 // Convert this function call into an invoke instruction.
553 // First, split the basic block.
554 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
556 // Delete the unconditional branch inserted by splitBasicBlock
557 BB->getInstList().pop_back();
559 // LIBUNWIND: If this is a call to @llvm.eh.resume, just branch
560 // directly to the new landing pad.
561 if (Invoke.forwardEHResume(CI, BB)) {
562 // TODO: 'Split' is now unreachable; clean it up.
564 // We want to leave the original call intact so that the call
565 // graph and other structures won't get misled. We also have to
566 // avoid processing the next block, or we'll iterate here forever.
570 // Otherwise, create the new invoke instruction.
571 ImmutableCallSite CS(CI);
572 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
574 InvokeInst::Create(CI->getCalledValue(), Split,
575 Invoke.getOuterUnwindDest(),
576 InvokeArgs, CI->getName(), BB);
577 II->setCallingConv(CI->getCallingConv());
578 II->setAttributes(CI->getAttributes());
580 // Make sure that anything using the call now uses the invoke! This also
581 // updates the CallGraph if present, because it uses a WeakVH.
582 CI->replaceAllUsesWith(II);
584 Split->getInstList().pop_front(); // Delete the original call
586 // Update any PHI nodes in the exceptional block to indicate that
587 // there is now a new entry in them.
588 Invoke.addIncomingPHIValuesFor(BB);
596 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
597 /// in the body of the inlined function into invokes and turn unwind
598 /// instructions into branches to the invoke unwind dest.
600 /// II is the invoke instruction being inlined. FirstNewBlock is the first
601 /// block of the inlined code (the last block is the end of the function),
602 /// and InlineCodeInfo is information about the code that got inlined.
603 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
604 ClonedCodeInfo &InlinedCodeInfo) {
605 BasicBlock *InvokeDest = II->getUnwindDest();
607 Function *Caller = FirstNewBlock->getParent();
609 // The inlined code is currently at the end of the function, scan from the
610 // start of the inlined code to its end, checking for stuff we need to
611 // rewrite. If the code doesn't have calls or unwinds, we know there is
612 // nothing to rewrite.
613 if (!InlinedCodeInfo.ContainsCalls && !InlinedCodeInfo.ContainsUnwinds) {
614 // Now that everything is happy, we have one final detail. The PHI nodes in
615 // the exception destination block still have entries due to the original
616 // invoke instruction. Eliminate these entries (which might even delete the
618 InvokeDest->removePredecessor(II->getParent());
622 InvokeInliningInfo Invoke(II);
624 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
625 if (InlinedCodeInfo.ContainsCalls)
626 if (HandleCallsInBlockInlinedThroughInvoke(BB, Invoke)) {
627 // Honor a request to skip the next block. We don't need to
628 // consider UnwindInsts in this case either.
633 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
634 // An UnwindInst requires special handling when it gets inlined into an
635 // invoke site. Once this happens, we know that the unwind would cause
636 // a control transfer to the invoke exception destination, so we can
637 // transform it into a direct branch to the exception destination.
638 BranchInst::Create(InvokeDest, UI);
640 // Delete the unwind instruction!
641 UI->eraseFromParent();
643 // Update any PHI nodes in the exceptional block to indicate that
644 // there is now a new entry in them.
645 Invoke.addIncomingPHIValuesFor(BB);
648 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
649 Invoke.forwardResume(RI);
653 // Now that everything is happy, we have one final detail. The PHI nodes in
654 // the exception destination block still have entries due to the original
655 // invoke instruction. Eliminate these entries (which might even delete the
657 InvokeDest->removePredecessor(II->getParent());
660 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
661 /// into the caller, update the specified callgraph to reflect the changes we
662 /// made. Note that it's possible that not all code was copied over, so only
663 /// some edges of the callgraph may remain.
664 static void UpdateCallGraphAfterInlining(CallSite CS,
665 Function::iterator FirstNewBlock,
666 ValueToValueMapTy &VMap,
667 InlineFunctionInfo &IFI) {
668 CallGraph &CG = *IFI.CG;
669 const Function *Caller = CS.getInstruction()->getParent()->getParent();
670 const Function *Callee = CS.getCalledFunction();
671 CallGraphNode *CalleeNode = CG[Callee];
672 CallGraphNode *CallerNode = CG[Caller];
674 // Since we inlined some uninlined call sites in the callee into the caller,
675 // add edges from the caller to all of the callees of the callee.
676 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
678 // Consider the case where CalleeNode == CallerNode.
679 CallGraphNode::CalledFunctionsVector CallCache;
680 if (CalleeNode == CallerNode) {
681 CallCache.assign(I, E);
682 I = CallCache.begin();
686 for (; I != E; ++I) {
687 const Value *OrigCall = I->first;
689 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
690 // Only copy the edge if the call was inlined!
691 if (VMI == VMap.end() || VMI->second == 0)
694 // If the call was inlined, but then constant folded, there is no edge to
695 // add. Check for this case.
696 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
697 if (NewCall == 0) continue;
699 // Remember that this call site got inlined for the client of
701 IFI.InlinedCalls.push_back(NewCall);
703 // It's possible that inlining the callsite will cause it to go from an
704 // indirect to a direct call by resolving a function pointer. If this
705 // happens, set the callee of the new call site to a more precise
706 // destination. This can also happen if the call graph node of the caller
707 // was just unnecessarily imprecise.
708 if (I->second->getFunction() == 0)
709 if (Function *F = CallSite(NewCall).getCalledFunction()) {
710 // Indirect call site resolved to direct call.
711 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
716 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
719 // Update the call graph by deleting the edge from Callee to Caller. We must
720 // do this after the loop above in case Caller and Callee are the same.
721 CallerNode->removeCallEdgeFor(CS);
724 /// HandleByValArgument - When inlining a call site that has a byval argument,
725 /// we have to make the implicit memcpy explicit by adding it.
726 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
727 const Function *CalledFunc,
728 InlineFunctionInfo &IFI,
729 unsigned ByValAlignment) {
730 Type *AggTy = cast<PointerType>(Arg->getType())->getElementType();
732 // If the called function is readonly, then it could not mutate the caller's
733 // copy of the byval'd memory. In this case, it is safe to elide the copy and
735 if (CalledFunc->onlyReadsMemory()) {
736 // If the byval argument has a specified alignment that is greater than the
737 // passed in pointer, then we either have to round up the input pointer or
738 // give up on this transformation.
739 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
742 // If the pointer is already known to be sufficiently aligned, or if we can
743 // round it up to a larger alignment, then we don't need a temporary.
744 if (getOrEnforceKnownAlignment(Arg, ByValAlignment,
745 IFI.TD) >= ByValAlignment)
748 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
749 // for code quality, but rarely happens and is required for correctness.
752 LLVMContext &Context = Arg->getContext();
754 Type *VoidPtrTy = Type::getInt8PtrTy(Context);
756 // Create the alloca. If we have TargetData, use nice alignment.
759 Align = IFI.TD->getPrefTypeAlignment(AggTy);
761 // If the byval had an alignment specified, we *must* use at least that
762 // alignment, as it is required by the byval argument (and uses of the
763 // pointer inside the callee).
764 Align = std::max(Align, ByValAlignment);
766 Function *Caller = TheCall->getParent()->getParent();
768 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(),
769 &*Caller->begin()->begin());
771 Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)};
772 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(),
775 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall);
776 Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall);
780 Size = ConstantExpr::getSizeOf(AggTy);
782 Size = ConstantInt::get(Type::getInt64Ty(Context),
783 IFI.TD->getTypeStoreSize(AggTy));
785 // Always generate a memcpy of alignment 1 here because we don't know
786 // the alignment of the src pointer. Other optimizations can infer
788 Value *CallArgs[] = {
789 DestCast, SrcCast, Size,
790 ConstantInt::get(Type::getInt32Ty(Context), 1),
791 ConstantInt::getFalse(Context) // isVolatile
793 IRBuilder<>(TheCall).CreateCall(MemCpyFn, CallArgs);
795 // Uses of the argument in the function should use our new alloca
800 // isUsedByLifetimeMarker - Check whether this Value is used by a lifetime
802 static bool isUsedByLifetimeMarker(Value *V) {
803 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE;
805 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI)) {
806 switch (II->getIntrinsicID()) {
808 case Intrinsic::lifetime_start:
809 case Intrinsic::lifetime_end:
817 // hasLifetimeMarkers - Check whether the given alloca already has
818 // lifetime.start or lifetime.end intrinsics.
819 static bool hasLifetimeMarkers(AllocaInst *AI) {
820 Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext());
821 if (AI->getType() == Int8PtrTy)
822 return isUsedByLifetimeMarker(AI);
824 // Do a scan to find all the casts to i8*.
825 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); I != E;
827 if (I->getType() != Int8PtrTy) continue;
828 if (I->stripPointerCasts() != AI) continue;
829 if (isUsedByLifetimeMarker(*I))
835 /// updateInlinedAtInfo - Helper function used by fixupLineNumbers to recursively
836 /// update InlinedAtEntry of a DebugLoc.
837 static DebugLoc updateInlinedAtInfo(const DebugLoc &DL,
838 const DebugLoc &InlinedAtDL,
840 if (MDNode *IA = DL.getInlinedAt(Ctx)) {
841 DebugLoc NewInlinedAtDL
842 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx);
843 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
844 NewInlinedAtDL.getAsMDNode(Ctx));
847 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
848 InlinedAtDL.getAsMDNode(Ctx));
852 /// fixupLineNumbers - Update inlined instructions' line numbers to
853 /// to encode location where these instructions are inlined.
854 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
855 Instruction *TheCall) {
856 DebugLoc TheCallDL = TheCall->getDebugLoc();
857 if (TheCallDL.isUnknown())
860 for (; FI != Fn->end(); ++FI) {
861 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
863 DebugLoc DL = BI->getDebugLoc();
864 if (!DL.isUnknown()) {
865 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext()));
866 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) {
867 LLVMContext &Ctx = BI->getContext();
868 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
869 DVI->setOperand(2, createInlinedVariable(DVI->getVariable(),
877 // InlineFunction - This function inlines the called function into the basic
878 // block of the caller. This returns false if it is not possible to inline this
879 // call. The program is still in a well defined state if this occurs though.
881 // Note that this only does one level of inlining. For example, if the
882 // instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
883 // exists in the instruction stream. Similarly this will inline a recursive
884 // function by one level.
886 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI) {
887 Instruction *TheCall = CS.getInstruction();
888 LLVMContext &Context = TheCall->getContext();
889 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
890 "Instruction not in function!");
892 // If IFI has any state in it, zap it before we fill it in.
895 const Function *CalledFunc = CS.getCalledFunction();
896 if (CalledFunc == 0 || // Can't inline external function or indirect
897 CalledFunc->isDeclaration() || // call, or call to a vararg function!
898 CalledFunc->getFunctionType()->isVarArg()) return false;
900 // If the call to the callee is not a tail call, we must clear the 'tail'
901 // flags on any calls that we inline.
902 bool MustClearTailCallFlags =
903 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall());
905 // If the call to the callee cannot throw, set the 'nounwind' flag on any
906 // calls that we inline.
907 bool MarkNoUnwind = CS.doesNotThrow();
909 BasicBlock *OrigBB = TheCall->getParent();
910 Function *Caller = OrigBB->getParent();
912 // GC poses two hazards to inlining, which only occur when the callee has GC:
913 // 1. If the caller has no GC, then the callee's GC must be propagated to the
915 // 2. If the caller has a differing GC, it is invalid to inline.
916 if (CalledFunc->hasGC()) {
917 if (!Caller->hasGC())
918 Caller->setGC(CalledFunc->getGC());
919 else if (CalledFunc->getGC() != Caller->getGC())
923 // Get the personality function from the callee if it contains a landing pad.
924 Value *CalleePersonality = 0;
925 for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end();
927 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
928 const BasicBlock *BB = II->getUnwindDest();
929 const LandingPadInst *LP = BB->getLandingPadInst();
930 CalleePersonality = LP->getPersonalityFn();
934 // Find the personality function used by the landing pads of the caller. If it
935 // exists, then check to see that it matches the personality function used in
937 if (CalleePersonality)
938 for (Function::const_iterator I = Caller->begin(), E = Caller->end();
940 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
941 const BasicBlock *BB = II->getUnwindDest();
942 const LandingPadInst *LP = BB->getLandingPadInst();
944 // If the personality functions match, then we can perform the
945 // inlining. Otherwise, we can't inline.
946 // TODO: This isn't 100% true. Some personality functions are proper
947 // supersets of others and can be used in place of the other.
948 if (LP->getPersonalityFn() != CalleePersonality)
954 // Get an iterator to the last basic block in the function, which will have
955 // the new function inlined after it.
957 Function::iterator LastBlock = &Caller->back();
959 // Make sure to capture all of the return instructions from the cloned
961 SmallVector<ReturnInst*, 8> Returns;
962 ClonedCodeInfo InlinedFunctionInfo;
963 Function::iterator FirstNewBlock;
965 { // Scope to destroy VMap after cloning.
966 ValueToValueMapTy VMap;
968 assert(CalledFunc->arg_size() == CS.arg_size() &&
969 "No varargs calls can be inlined!");
971 // Calculate the vector of arguments to pass into the function cloner, which
972 // matches up the formal to the actual argument values.
973 CallSite::arg_iterator AI = CS.arg_begin();
975 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
976 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
977 Value *ActualArg = *AI;
979 // When byval arguments actually inlined, we need to make the copy implied
980 // by them explicit. However, we don't do this if the callee is readonly
981 // or readnone, because the copy would be unneeded: the callee doesn't
982 // modify the struct.
983 if (CS.isByValArgument(ArgNo)) {
984 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
985 CalledFunc->getParamAlignment(ArgNo+1));
987 // Calls that we inline may use the new alloca, so we need to clear
988 // their 'tail' flags if HandleByValArgument introduced a new alloca and
989 // the callee has calls.
990 MustClearTailCallFlags |= ActualArg != *AI;
996 // We want the inliner to prune the code as it copies. We would LOVE to
997 // have no dead or constant instructions leftover after inlining occurs
998 // (which can happen, e.g., because an argument was constant), but we'll be
999 // happy with whatever the cloner can do.
1000 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
1001 /*ModuleLevelChanges=*/false, Returns, ".i",
1002 &InlinedFunctionInfo, IFI.TD, TheCall);
1004 // Remember the first block that is newly cloned over.
1005 FirstNewBlock = LastBlock; ++FirstNewBlock;
1007 // Update the callgraph if requested.
1009 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
1011 // Update inlined instructions' line number information.
1012 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
1015 // If there are any alloca instructions in the block that used to be the entry
1016 // block for the callee, move them to the entry block of the caller. First
1017 // calculate which instruction they should be inserted before. We insert the
1018 // instructions at the end of the current alloca list.
1021 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
1022 for (BasicBlock::iterator I = FirstNewBlock->begin(),
1023 E = FirstNewBlock->end(); I != E; ) {
1024 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
1025 if (AI == 0) continue;
1027 // If the alloca is now dead, remove it. This often occurs due to code
1029 if (AI->use_empty()) {
1030 AI->eraseFromParent();
1034 if (!isa<Constant>(AI->getArraySize()))
1037 // Keep track of the static allocas that we inline into the caller.
1038 IFI.StaticAllocas.push_back(AI);
1040 // Scan for the block of allocas that we can move over, and move them
1042 while (isa<AllocaInst>(I) &&
1043 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
1044 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
1048 // Transfer all of the allocas over in a block. Using splice means
1049 // that the instructions aren't removed from the symbol table, then
1051 Caller->getEntryBlock().getInstList().splice(InsertPoint,
1052 FirstNewBlock->getInstList(),
1057 // Leave lifetime markers for the static alloca's, scoping them to the
1058 // function we just inlined.
1059 if (!IFI.StaticAllocas.empty()) {
1060 IRBuilder<> builder(FirstNewBlock->begin());
1061 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
1062 AllocaInst *AI = IFI.StaticAllocas[ai];
1064 // If the alloca is already scoped to something smaller than the whole
1065 // function then there's no need to add redundant, less accurate markers.
1066 if (hasLifetimeMarkers(AI))
1069 builder.CreateLifetimeStart(AI);
1070 for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) {
1071 IRBuilder<> builder(Returns[ri]);
1072 builder.CreateLifetimeEnd(AI);
1077 // If the inlined code contained dynamic alloca instructions, wrap the inlined
1078 // code with llvm.stacksave/llvm.stackrestore intrinsics.
1079 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
1080 Module *M = Caller->getParent();
1081 // Get the two intrinsics we care about.
1082 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
1083 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
1085 // Insert the llvm.stacksave.
1086 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
1087 .CreateCall(StackSave, "savedstack");
1089 // Insert a call to llvm.stackrestore before any return instructions in the
1090 // inlined function.
1091 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1092 IRBuilder<>(Returns[i]).CreateCall(StackRestore, SavedPtr);
1095 // Count the number of StackRestore calls we insert.
1096 unsigned NumStackRestores = Returns.size();
1098 // If we are inlining an invoke instruction, insert restores before each
1099 // unwind. These unwinds will be rewritten into branches later.
1100 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) {
1101 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
1103 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
1104 IRBuilder<>(UI).CreateCall(StackRestore, SavedPtr);
1110 // If we are inlining tail call instruction through a call site that isn't
1111 // marked 'tail', we must remove the tail marker for any calls in the inlined
1112 // code. Also, calls inlined through a 'nounwind' call site should be marked
1114 if (InlinedFunctionInfo.ContainsCalls &&
1115 (MustClearTailCallFlags || MarkNoUnwind)) {
1116 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
1118 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
1119 if (CallInst *CI = dyn_cast<CallInst>(I)) {
1120 if (MustClearTailCallFlags)
1121 CI->setTailCall(false);
1123 CI->setDoesNotThrow();
1127 // If we are inlining through a 'nounwind' call site then any inlined 'unwind'
1128 // instructions are unreachable.
1129 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind)
1130 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
1132 TerminatorInst *Term = BB->getTerminator();
1133 if (isa<UnwindInst>(Term)) {
1134 new UnreachableInst(Context, Term);
1135 BB->getInstList().erase(Term);
1139 // If we are inlining for an invoke instruction, we must make sure to rewrite
1140 // any inlined 'unwind' instructions into branches to the invoke exception
1141 // destination, and call instructions into invoke instructions.
1142 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
1143 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
1145 // If we cloned in _exactly one_ basic block, and if that block ends in a
1146 // return instruction, we splice the body of the inlined callee directly into
1147 // the calling basic block.
1148 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
1149 // Move all of the instructions right before the call.
1150 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
1151 FirstNewBlock->begin(), FirstNewBlock->end());
1152 // Remove the cloned basic block.
1153 Caller->getBasicBlockList().pop_back();
1155 // If the call site was an invoke instruction, add a branch to the normal
1157 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
1158 BranchInst::Create(II->getNormalDest(), TheCall);
1160 // If the return instruction returned a value, replace uses of the call with
1161 // uses of the returned value.
1162 if (!TheCall->use_empty()) {
1163 ReturnInst *R = Returns[0];
1164 if (TheCall == R->getReturnValue())
1165 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1167 TheCall->replaceAllUsesWith(R->getReturnValue());
1169 // Since we are now done with the Call/Invoke, we can delete it.
1170 TheCall->eraseFromParent();
1172 // Since we are now done with the return instruction, delete it also.
1173 Returns[0]->eraseFromParent();
1175 // We are now done with the inlining.
1179 // Otherwise, we have the normal case, of more than one block to inline or
1180 // multiple return sites.
1182 // We want to clone the entire callee function into the hole between the
1183 // "starter" and "ender" blocks. How we accomplish this depends on whether
1184 // this is an invoke instruction or a call instruction.
1185 BasicBlock *AfterCallBB;
1186 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1188 // Add an unconditional branch to make this look like the CallInst case...
1189 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
1191 // Split the basic block. This guarantees that no PHI nodes will have to be
1192 // updated due to new incoming edges, and make the invoke case more
1193 // symmetric to the call case.
1194 AfterCallBB = OrigBB->splitBasicBlock(NewBr,
1195 CalledFunc->getName()+".exit");
1197 } else { // It's a call
1198 // If this is a call instruction, we need to split the basic block that
1199 // the call lives in.
1201 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
1202 CalledFunc->getName()+".exit");
1205 // Change the branch that used to go to AfterCallBB to branch to the first
1206 // basic block of the inlined function.
1208 TerminatorInst *Br = OrigBB->getTerminator();
1209 assert(Br && Br->getOpcode() == Instruction::Br &&
1210 "splitBasicBlock broken!");
1211 Br->setOperand(0, FirstNewBlock);
1214 // Now that the function is correct, make it a little bit nicer. In
1215 // particular, move the basic blocks inserted from the end of the function
1216 // into the space made by splitting the source basic block.
1217 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
1218 FirstNewBlock, Caller->end());
1220 // Handle all of the return instructions that we just cloned in, and eliminate
1221 // any users of the original call/invoke instruction.
1222 Type *RTy = CalledFunc->getReturnType();
1225 if (Returns.size() > 1) {
1226 // The PHI node should go at the front of the new basic block to merge all
1227 // possible incoming values.
1228 if (!TheCall->use_empty()) {
1229 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
1230 AfterCallBB->begin());
1231 // Anything that used the result of the function call should now use the
1232 // PHI node as their operand.
1233 TheCall->replaceAllUsesWith(PHI);
1236 // Loop over all of the return instructions adding entries to the PHI node
1239 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1240 ReturnInst *RI = Returns[i];
1241 assert(RI->getReturnValue()->getType() == PHI->getType() &&
1242 "Ret value not consistent in function!");
1243 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
1248 // Add a branch to the merge points and remove return instructions.
1249 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1250 ReturnInst *RI = Returns[i];
1251 BranchInst::Create(AfterCallBB, RI);
1252 RI->eraseFromParent();
1254 } else if (!Returns.empty()) {
1255 // Otherwise, if there is exactly one return value, just replace anything
1256 // using the return value of the call with the computed value.
1257 if (!TheCall->use_empty()) {
1258 if (TheCall == Returns[0]->getReturnValue())
1259 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1261 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
1264 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
1265 BasicBlock *ReturnBB = Returns[0]->getParent();
1266 ReturnBB->replaceAllUsesWith(AfterCallBB);
1268 // Splice the code from the return block into the block that it will return
1269 // to, which contains the code that was after the call.
1270 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
1271 ReturnBB->getInstList());
1273 // Delete the return instruction now and empty ReturnBB now.
1274 Returns[0]->eraseFromParent();
1275 ReturnBB->eraseFromParent();
1276 } else if (!TheCall->use_empty()) {
1277 // No returns, but something is using the return value of the call. Just
1279 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1282 // Since we are now done with the Call/Invoke, we can delete it.
1283 TheCall->eraseFromParent();
1285 // We should always be able to fold the entry block of the function into the
1286 // single predecessor of the block...
1287 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
1288 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
1290 // Splice the code entry block into calling block, right before the
1291 // unconditional branch.
1292 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
1293 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
1295 // Remove the unconditional branch.
1296 OrigBB->getInstList().erase(Br);
1298 // Now we can remove the CalleeEntry block, which is now empty.
1299 Caller->getBasicBlockList().erase(CalleeEntry);
1301 // If we inserted a phi node, check to see if it has a single value (e.g. all
1302 // the entries are the same or undef). If so, remove the PHI so it doesn't
1303 // block other optimizations.
1305 if (Value *V = SimplifyInstruction(PHI, IFI.TD)) {
1306 PHI->replaceAllUsesWith(V);
1307 PHI->eraseFromParent();