1 //===- InlineFunction.cpp - Code to perform function inlining -------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements inlining of a function into a call site, resolving
11 // parameters and the return value as appropriate.
12 //
13 // 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].
19 //
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"
39 using namespace llvm;
41 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI) {
42 return InlineFunction(CallSite(CI), IFI);
43 }
44 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI) {
45 return InlineFunction(CallSite(II), IFI);
46 }
48 // FIXME: New EH - Remove the functions marked [LIBUNWIND] when new EH is
49 // turned on.
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);
55 if (exn) return exn;
56 }
58 return 0;
59 }
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);
70 if (!sel) continue;
72 // Immediately accept an eh.selector in the same block as the
73 // excepton call.
74 if (sel->getParent() == exnBlock) return sel;
76 // Otherwise, use the first selector we see.
77 if (!outOfBlockSelector) outOfBlockSelector = sel;
78 }
80 return outOfBlockSelector;
81 }
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
96 // landing pad.
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);
108 do {
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.
137 dominates = false;
138 nonDominated = successor;
139 lastDominated = exnBlock;
140 }
142 exnBlock = successor;
144 // Can we stop here?
145 exn = findExceptionInBlock(exnBlock);
146 } while (!exn);
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
154 // the landing pad.
155 if (dominates) {
156 selector->moveBefore(lpad->getFirstNonPHI());
157 exn->moveBefore(selector);
158 return selector;
159 }
161 // Otherwise, we have to split at the first non-dominating block.
162 // The CFG looks basically like this:
163 // lpad:
164 // phis_0
165 // insnsAndBranches_1
166 // br label %nonDominated
167 // nonDominated:
168 // phis_2
169 // insns_3
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:
174 // lpad:
175 // phis_0
176 // %exn0 = call i8* @llvm.eh.exception()
177 // %selector0 = call @llvm.eh.selector(i8* %exn0, ...
178 // insnsAndBranches_1
179 // br label %split // from lastDominated
180 // nonDominated:
181 // phis_2 (without edge from lastDominated)
182 // %exn1 = call i8* @llvm.eh.exception()
183 // %selector1 = call i8* @llvm.eh.selector(i8* %exn1, ...
184 // br label %split
185 // split:
186 // phis_2 (edge from lastDominated, edge from split)
187 // %exn = phi ...
188 // %selector = phi ...
189 // insns_3
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.
203 BasicBlock *split =
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);
221 if (!phi) break;
223 PHINode *splitPhi = PHINode::Create(phi->getType(), 2, phi->getName(),
224 splitIP);
225 phi->replaceAllUsesWith(splitPhi);
226 splitPhi->addIncoming(phi, nonDominated);
227 splitPhi->addIncoming(phi->removeIncomingValue(lastDominated),
228 lastDominated);
229 }
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);
243 return lpadSelector;
244 }
246 namespace {
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.
262 public:
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));
277 }
279 CallerLPad = cast<LandingPadInst>(I);
280 }
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;
286 }
288 EHSelectorInst *getOuterSelector() {
289 if (!OuterSelector)
290 OuterSelector = findSelectorForLandingPad(OuterUnwindDest);
291 return OuterSelector;
292 }
294 BasicBlock *getInnerUnwindDest();
296 LandingPadInst *getLandingPadInst() const { return CallerLPad; }
298 /// forwardResume - Forward the 'resume' instruction to the caller's landing
299 /// pad block. When the landing pad block has only one predecessor, this is
300 /// a simple branch. When there is more than one predecessor, we need to
301 /// split the landing pad block after the landingpad instruction and jump
302 /// to there.
303 void forwardResume(ResumeInst *RI);
305 /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind
306 /// destination block for the given basic block, using the values for the
307 /// original invoke's source block.
308 void addIncomingPHIValuesFor(BasicBlock *BB) const {
309 addIncomingPHIValuesForInto(BB, OuterUnwindDest);
310 }
312 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const {
313 BasicBlock::iterator I = dest->begin();
314 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
315 PHINode *phi = cast<PHINode>(I);
316 phi->addIncoming(UnwindDestPHIValues[i], src);
317 }
318 }
319 };
320 }
322 /// getInnerUnwindDest - Get or create a target for the branch from ResumeInsts.
323 BasicBlock *InvokeInliningInfo::getInnerUnwindDest() {
324 if (InnerResumeDest) return InnerResumeDest;
326 // Split the landing pad.
327 BasicBlock::iterator SplitPoint = CallerLPad; ++SplitPoint;
328 InnerResumeDest =
329 OuterResumeDest->splitBasicBlock(SplitPoint,
330 OuterResumeDest->getName() + ".body");
332 // The number of incoming edges we expect to the inner landing pad.
333 const unsigned PHICapacity = 2;
335 // Create corresponding new PHIs for all the PHIs in the outer landing pad.
336 BasicBlock::iterator InsertPoint = InnerResumeDest->begin();
337 BasicBlock::iterator I = OuterResumeDest->begin();
338 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
339 PHINode *OuterPHI = cast<PHINode>(I);
340 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity,
341 OuterPHI->getName() + ".lpad-body",
342 InsertPoint);
343 OuterPHI->replaceAllUsesWith(InnerPHI);
344 InnerPHI->addIncoming(OuterPHI, OuterResumeDest);
345 }
347 // Create a PHI for the exception values.
348 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity,
349 "eh.lpad-body", InsertPoint);
350 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI);
351 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest);
353 // All done.
354 return InnerResumeDest;
355 }
357 /// forwardResume - Forward the 'resume' instruction to the caller's landing pad
358 /// block. When the landing pad block has only one predecessor, this is a simple
359 /// branch. When there is more than one predecessor, we need to split the
360 /// landing pad block after the landingpad instruction and jump to there.
361 void InvokeInliningInfo::forwardResume(ResumeInst *RI) {
362 BasicBlock *Dest = getInnerUnwindDest();
363 BasicBlock *Src = RI->getParent();
365 BranchInst::Create(Dest, Src);
367 // Update the PHIs in the destination. They were inserted in an order which
368 // makes this work.
369 addIncomingPHIValuesForInto(Src, Dest);
371 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src);
372 RI->eraseFromParent();
373 }
375 /// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
376 /// an invoke, we have to turn all of the calls that can throw into
377 /// invokes. This function analyze BB to see if there are any calls, and if so,
378 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
379 /// nodes in that block with the values specified in InvokeDestPHIValues.
380 ///
381 /// Returns true to indicate that the next block should be skipped.
382 static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
383 InvokeInliningInfo &Invoke) {
384 LandingPadInst *LPI = Invoke.getLandingPadInst();
386 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
387 Instruction *I = BBI++;
389 if (LandingPadInst *L = dyn_cast<LandingPadInst>(I)) {
390 unsigned NumClauses = LPI->getNumClauses();
391 L->reserveClauses(NumClauses);
392 for (unsigned i = 0; i != NumClauses; ++i)
393 L->addClause(LPI->getClause(i));
394 }
396 // We only need to check for function calls: inlined invoke
397 // instructions require no special handling.
398 CallInst *CI = dyn_cast<CallInst>(I);
400 // If this call cannot unwind, don't convert it to an invoke.
401 if (!CI || CI->doesNotThrow())
402 continue;
404 // Convert this function call into an invoke instruction. First, split the
405 // basic block.
406 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
408 // Delete the unconditional branch inserted by splitBasicBlock
409 BB->getInstList().pop_back();
411 // Create the new invoke instruction.
412 ImmutableCallSite CS(CI);
413 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
414 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split,
415 Invoke.getOuterUnwindDest(),
416 InvokeArgs, CI->getName(), BB);
417 II->setCallingConv(CI->getCallingConv());
418 II->setAttributes(CI->getAttributes());
420 // Make sure that anything using the call now uses the invoke! This also
421 // updates the CallGraph if present, because it uses a WeakVH.
422 CI->replaceAllUsesWith(II);
424 // Delete the original call
425 Split->getInstList().pop_front();
427 // Update any PHI nodes in the exceptional block to indicate that there is
428 // now a new entry in them.
429 Invoke.addIncomingPHIValuesFor(BB);
430 return false;
431 }
433 return false;
434 }
436 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
437 /// in the body of the inlined function into invokes and turn unwind
438 /// instructions into branches to the invoke unwind dest.
439 ///
440 /// II is the invoke instruction being inlined. FirstNewBlock is the first
441 /// block of the inlined code (the last block is the end of the function),
442 /// and InlineCodeInfo is information about the code that got inlined.
443 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
444 ClonedCodeInfo &InlinedCodeInfo) {
445 BasicBlock *InvokeDest = II->getUnwindDest();
447 Function *Caller = FirstNewBlock->getParent();
449 // The inlined code is currently at the end of the function, scan from the
450 // start of the inlined code to its end, checking for stuff we need to
451 // rewrite. If the code doesn't have calls or unwinds, we know there is
452 // nothing to rewrite.
453 if (!InlinedCodeInfo.ContainsCalls && !InlinedCodeInfo.ContainsUnwinds) {
454 // Now that everything is happy, we have one final detail. The PHI nodes in
455 // the exception destination block still have entries due to the original
456 // invoke instruction. Eliminate these entries (which might even delete the
457 // PHI node) now.
458 InvokeDest->removePredecessor(II->getParent());
459 return;
460 }
462 InvokeInliningInfo Invoke(II);
464 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
465 if (InlinedCodeInfo.ContainsCalls)
466 if (HandleCallsInBlockInlinedThroughInvoke(BB, Invoke)) {
467 // Honor a request to skip the next block. We don't need to
468 // consider UnwindInsts in this case either.
469 ++BB;
470 continue;
471 }
473 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
474 // An UnwindInst requires special handling when it gets inlined into an
475 // invoke site. Once this happens, we know that the unwind would cause
476 // a control transfer to the invoke exception destination, so we can
477 // transform it into a direct branch to the exception destination.
478 BranchInst::Create(InvokeDest, UI);
480 // Delete the unwind instruction!
481 UI->eraseFromParent();
483 // Update any PHI nodes in the exceptional block to indicate that
484 // there is now a new entry in them.
485 Invoke.addIncomingPHIValuesFor(BB);
486 }
488 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator()))
489 Invoke.forwardResume(RI);
490 }
492 // Now that everything is happy, we have one final detail. The PHI nodes in
493 // the exception destination block still have entries due to the original
494 // invoke instruction. Eliminate these entries (which might even delete the
495 // PHI node) now.
496 InvokeDest->removePredecessor(II->getParent());
497 }
499 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
500 /// into the caller, update the specified callgraph to reflect the changes we
501 /// made. Note that it's possible that not all code was copied over, so only
502 /// some edges of the callgraph may remain.
503 static void UpdateCallGraphAfterInlining(CallSite CS,
504 Function::iterator FirstNewBlock,
505 ValueToValueMapTy &VMap,
506 InlineFunctionInfo &IFI) {
507 CallGraph &CG = *IFI.CG;
508 const Function *Caller = CS.getInstruction()->getParent()->getParent();
509 const Function *Callee = CS.getCalledFunction();
510 CallGraphNode *CalleeNode = CG[Callee];
511 CallGraphNode *CallerNode = CG[Caller];
513 // Since we inlined some uninlined call sites in the callee into the caller,
514 // add edges from the caller to all of the callees of the callee.
515 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
517 // Consider the case where CalleeNode == CallerNode.
518 CallGraphNode::CalledFunctionsVector CallCache;
519 if (CalleeNode == CallerNode) {
520 CallCache.assign(I, E);
521 I = CallCache.begin();
522 E = CallCache.end();
523 }
525 for (; I != E; ++I) {
526 const Value *OrigCall = I->first;
528 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
529 // Only copy the edge if the call was inlined!
530 if (VMI == VMap.end() || VMI->second == 0)
531 continue;
533 // If the call was inlined, but then constant folded, there is no edge to
534 // add. Check for this case.
535 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
536 if (NewCall == 0) continue;
538 // Remember that this call site got inlined for the client of
539 // InlineFunction.
540 IFI.InlinedCalls.push_back(NewCall);
542 // It's possible that inlining the callsite will cause it to go from an
543 // indirect to a direct call by resolving a function pointer. If this
544 // happens, set the callee of the new call site to a more precise
545 // destination. This can also happen if the call graph node of the caller
546 // was just unnecessarily imprecise.
547 if (I->second->getFunction() == 0)
548 if (Function *F = CallSite(NewCall).getCalledFunction()) {
549 // Indirect call site resolved to direct call.
550 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
552 continue;
553 }
555 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
556 }
558 // Update the call graph by deleting the edge from Callee to Caller. We must
559 // do this after the loop above in case Caller and Callee are the same.
560 CallerNode->removeCallEdgeFor(CS);
561 }
563 /// HandleByValArgument - When inlining a call site that has a byval argument,
564 /// we have to make the implicit memcpy explicit by adding it.
565 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
566 const Function *CalledFunc,
567 InlineFunctionInfo &IFI,
568 unsigned ByValAlignment) {
569 Type *AggTy = cast<PointerType>(Arg->getType())->getElementType();
571 // If the called function is readonly, then it could not mutate the caller's
572 // copy of the byval'd memory. In this case, it is safe to elide the copy and
573 // temporary.
574 if (CalledFunc->onlyReadsMemory()) {
575 // If the byval argument has a specified alignment that is greater than the
576 // passed in pointer, then we either have to round up the input pointer or
577 // give up on this transformation.
578 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
579 return Arg;
581 // If the pointer is already known to be sufficiently aligned, or if we can
582 // round it up to a larger alignment, then we don't need a temporary.
583 if (getOrEnforceKnownAlignment(Arg, ByValAlignment,
584 IFI.TD) >= ByValAlignment)
585 return Arg;
587 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
588 // for code quality, but rarely happens and is required for correctness.
589 }
591 LLVMContext &Context = Arg->getContext();
593 Type *VoidPtrTy = Type::getInt8PtrTy(Context);
595 // Create the alloca. If we have TargetData, use nice alignment.
596 unsigned Align = 1;
597 if (IFI.TD)
598 Align = IFI.TD->getPrefTypeAlignment(AggTy);
600 // If the byval had an alignment specified, we *must* use at least that
601 // alignment, as it is required by the byval argument (and uses of the
602 // pointer inside the callee).
603 Align = std::max(Align, ByValAlignment);
605 Function *Caller = TheCall->getParent()->getParent();
607 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(),
608 &*Caller->begin()->begin());
609 // Emit a memcpy.
610 Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)};
611 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(),
612 Intrinsic::memcpy,
613 Tys);
614 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall);
615 Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall);
617 Value *Size;
618 if (IFI.TD == 0)
619 Size = ConstantExpr::getSizeOf(AggTy);
620 else
621 Size = ConstantInt::get(Type::getInt64Ty(Context),
622 IFI.TD->getTypeStoreSize(AggTy));
624 // Always generate a memcpy of alignment 1 here because we don't know
625 // the alignment of the src pointer. Other optimizations can infer
626 // better alignment.
627 Value *CallArgs[] = {
628 DestCast, SrcCast, Size,
629 ConstantInt::get(Type::getInt32Ty(Context), 1),
630 ConstantInt::getFalse(Context) // isVolatile
631 };
632 IRBuilder<>(TheCall).CreateCall(MemCpyFn, CallArgs);
634 // Uses of the argument in the function should use our new alloca
635 // instead.
636 return NewAlloca;
637 }
639 // isUsedByLifetimeMarker - Check whether this Value is used by a lifetime
640 // intrinsic.
641 static bool isUsedByLifetimeMarker(Value *V) {
642 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE;
643 ++UI) {
644 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI)) {
645 switch (II->getIntrinsicID()) {
646 default: break;
647 case Intrinsic::lifetime_start:
648 case Intrinsic::lifetime_end:
649 return true;
650 }
651 }
652 }
653 return false;
654 }
656 // hasLifetimeMarkers - Check whether the given alloca already has
657 // lifetime.start or lifetime.end intrinsics.
658 static bool hasLifetimeMarkers(AllocaInst *AI) {
659 Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext());
660 if (AI->getType() == Int8PtrTy)
661 return isUsedByLifetimeMarker(AI);
663 // Do a scan to find all the casts to i8*.
664 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); I != E;
665 ++I) {
666 if (I->getType() != Int8PtrTy) continue;
667 if (I->stripPointerCasts() != AI) continue;
668 if (isUsedByLifetimeMarker(*I))
669 return true;
670 }
671 return false;
672 }
674 /// updateInlinedAtInfo - Helper function used by fixupLineNumbers to recursively
675 /// update InlinedAtEntry of a DebugLoc.
676 static DebugLoc updateInlinedAtInfo(const DebugLoc &DL,
677 const DebugLoc &InlinedAtDL,
678 LLVMContext &Ctx) {
679 if (MDNode *IA = DL.getInlinedAt(Ctx)) {
680 DebugLoc NewInlinedAtDL
681 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx);
682 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
683 NewInlinedAtDL.getAsMDNode(Ctx));
684 }
686 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
687 InlinedAtDL.getAsMDNode(Ctx));
688 }
690 /// fixupLineNumbers - Update inlined instructions' line numbers to
691 /// to encode location where these instructions are inlined.
692 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
693 Instruction *TheCall) {
694 DebugLoc TheCallDL = TheCall->getDebugLoc();
695 if (TheCallDL.isUnknown())
696 return;
698 for (; FI != Fn->end(); ++FI) {
699 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
700 BI != BE; ++BI) {
701 DebugLoc DL = BI->getDebugLoc();
702 if (!DL.isUnknown()) {
703 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext()));
704 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) {
705 LLVMContext &Ctx = BI->getContext();
706 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
707 DVI->setOperand(2, createInlinedVariable(DVI->getVariable(),
708 InlinedAt, Ctx));
709 }
710 }
711 }
712 }
713 }
715 /// InlineFunction - This function inlines the called function into the basic
716 /// block of the caller. This returns false if it is not possible to inline
717 /// this call. The program is still in a well defined state if this occurs
718 /// though.
719 ///
720 /// Note that this only does one level of inlining. For example, if the
721 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
722 /// exists in the instruction stream. Similarly this will inline a recursive
723 /// function by one level.
724 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI) {
725 Instruction *TheCall = CS.getInstruction();
726 LLVMContext &Context = TheCall->getContext();
727 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
728 "Instruction not in function!");
730 // If IFI has any state in it, zap it before we fill it in.
731 IFI.reset();
733 const Function *CalledFunc = CS.getCalledFunction();
734 if (CalledFunc == 0 || // Can't inline external function or indirect
735 CalledFunc->isDeclaration() || // call, or call to a vararg function!
736 CalledFunc->getFunctionType()->isVarArg()) return false;
738 // If the call to the callee is not a tail call, we must clear the 'tail'
739 // flags on any calls that we inline.
740 bool MustClearTailCallFlags =
741 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall());
743 // If the call to the callee cannot throw, set the 'nounwind' flag on any
744 // calls that we inline.
745 bool MarkNoUnwind = CS.doesNotThrow();
747 BasicBlock *OrigBB = TheCall->getParent();
748 Function *Caller = OrigBB->getParent();
750 // GC poses two hazards to inlining, which only occur when the callee has GC:
751 // 1. If the caller has no GC, then the callee's GC must be propagated to the
752 // caller.
753 // 2. If the caller has a differing GC, it is invalid to inline.
754 if (CalledFunc->hasGC()) {
755 if (!Caller->hasGC())
756 Caller->setGC(CalledFunc->getGC());
757 else if (CalledFunc->getGC() != Caller->getGC())
758 return false;
759 }
761 // Get the personality function from the callee if it contains a landing pad.
762 Value *CalleePersonality = 0;
763 for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end();
764 I != E; ++I)
765 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
766 const BasicBlock *BB = II->getUnwindDest();
767 const LandingPadInst *LP = BB->getLandingPadInst();
768 CalleePersonality = LP->getPersonalityFn();
769 break;
770 }
772 // Find the personality function used by the landing pads of the caller. If it
773 // exists, then check to see that it matches the personality function used in
774 // the callee.
775 if (CalleePersonality) {
776 for (Function::const_iterator I = Caller->begin(), E = Caller->end();
777 I != E; ++I)
778 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
779 const BasicBlock *BB = II->getUnwindDest();
780 const LandingPadInst *LP = BB->getLandingPadInst();
782 // If the personality functions match, then we can perform the
783 // inlining. Otherwise, we can't inline.
784 // TODO: This isn't 100% true. Some personality functions are proper
785 // supersets of others and can be used in place of the other.
786 if (LP->getPersonalityFn() != CalleePersonality)
787 return false;
789 break;
790 }
791 }
793 // Get an iterator to the last basic block in the function, which will have
794 // the new function inlined after it.
795 Function::iterator LastBlock = &Caller->back();
797 // Make sure to capture all of the return instructions from the cloned
798 // function.
799 SmallVector<ReturnInst*, 8> Returns;
800 ClonedCodeInfo InlinedFunctionInfo;
801 Function::iterator FirstNewBlock;
803 { // Scope to destroy VMap after cloning.
804 ValueToValueMapTy VMap;
806 assert(CalledFunc->arg_size() == CS.arg_size() &&
807 "No varargs calls can be inlined!");
809 // Calculate the vector of arguments to pass into the function cloner, which
810 // matches up the formal to the actual argument values.
811 CallSite::arg_iterator AI = CS.arg_begin();
812 unsigned ArgNo = 0;
813 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
814 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
815 Value *ActualArg = *AI;
817 // When byval arguments actually inlined, we need to make the copy implied
818 // by them explicit. However, we don't do this if the callee is readonly
819 // or readnone, because the copy would be unneeded: the callee doesn't
820 // modify the struct.
821 if (CS.isByValArgument(ArgNo)) {
822 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
823 CalledFunc->getParamAlignment(ArgNo+1));
825 // Calls that we inline may use the new alloca, so we need to clear
826 // their 'tail' flags if HandleByValArgument introduced a new alloca and
827 // the callee has calls.
828 MustClearTailCallFlags |= ActualArg != *AI;
829 }
831 VMap[I] = ActualArg;
832 }
834 // We want the inliner to prune the code as it copies. We would LOVE to
835 // have no dead or constant instructions leftover after inlining occurs
836 // (which can happen, e.g., because an argument was constant), but we'll be
837 // happy with whatever the cloner can do.
838 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
839 /*ModuleLevelChanges=*/false, Returns, ".i",
840 &InlinedFunctionInfo, IFI.TD, TheCall);
842 // Remember the first block that is newly cloned over.
843 FirstNewBlock = LastBlock; ++FirstNewBlock;
845 // Update the callgraph if requested.
846 if (IFI.CG)
847 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
849 // Update inlined instructions' line number information.
850 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
851 }
853 // If there are any alloca instructions in the block that used to be the entry
854 // block for the callee, move them to the entry block of the caller. First
855 // calculate which instruction they should be inserted before. We insert the
856 // instructions at the end of the current alloca list.
857 {
858 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
859 for (BasicBlock::iterator I = FirstNewBlock->begin(),
860 E = FirstNewBlock->end(); I != E; ) {
861 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
862 if (AI == 0) continue;
864 // If the alloca is now dead, remove it. This often occurs due to code
865 // specialization.
866 if (AI->use_empty()) {
867 AI->eraseFromParent();
868 continue;
869 }
871 if (!isa<Constant>(AI->getArraySize()))
872 continue;
874 // Keep track of the static allocas that we inline into the caller.
875 IFI.StaticAllocas.push_back(AI);
877 // Scan for the block of allocas that we can move over, and move them
878 // all at once.
879 while (isa<AllocaInst>(I) &&
880 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
881 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
882 ++I;
883 }
885 // Transfer all of the allocas over in a block. Using splice means
886 // that the instructions aren't removed from the symbol table, then
887 // reinserted.
888 Caller->getEntryBlock().getInstList().splice(InsertPoint,
889 FirstNewBlock->getInstList(),
890 AI, I);
891 }
892 }
894 // Leave lifetime markers for the static alloca's, scoping them to the
895 // function we just inlined.
896 if (!IFI.StaticAllocas.empty()) {
897 IRBuilder<> builder(FirstNewBlock->begin());
898 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
899 AllocaInst *AI = IFI.StaticAllocas[ai];
901 // If the alloca is already scoped to something smaller than the whole
902 // function then there's no need to add redundant, less accurate markers.
903 if (hasLifetimeMarkers(AI))
904 continue;
906 builder.CreateLifetimeStart(AI);
907 for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) {
908 IRBuilder<> builder(Returns[ri]);
909 builder.CreateLifetimeEnd(AI);
910 }
911 }
912 }
914 // If the inlined code contained dynamic alloca instructions, wrap the inlined
915 // code with llvm.stacksave/llvm.stackrestore intrinsics.
916 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
917 Module *M = Caller->getParent();
918 // Get the two intrinsics we care about.
919 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
920 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
922 // Insert the llvm.stacksave.
923 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
924 .CreateCall(StackSave, "savedstack");
926 // Insert a call to llvm.stackrestore before any return instructions in the
927 // inlined function.
928 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
929 IRBuilder<>(Returns[i]).CreateCall(StackRestore, SavedPtr);
930 }
932 // Count the number of StackRestore calls we insert.
933 unsigned NumStackRestores = Returns.size();
935 // If we are inlining an invoke instruction, insert restores before each
936 // unwind. These unwinds will be rewritten into branches later.
937 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) {
938 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
939 BB != E; ++BB)
940 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
941 IRBuilder<>(UI).CreateCall(StackRestore, SavedPtr);
942 ++NumStackRestores;
943 }
944 }
945 }
947 // If we are inlining tail call instruction through a call site that isn't
948 // marked 'tail', we must remove the tail marker for any calls in the inlined
949 // code. Also, calls inlined through a 'nounwind' call site should be marked
950 // 'nounwind'.
951 if (InlinedFunctionInfo.ContainsCalls &&
952 (MustClearTailCallFlags || MarkNoUnwind)) {
953 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
954 BB != E; ++BB)
955 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
956 if (CallInst *CI = dyn_cast<CallInst>(I)) {
957 if (MustClearTailCallFlags)
958 CI->setTailCall(false);
959 if (MarkNoUnwind)
960 CI->setDoesNotThrow();
961 }
962 }
964 // If we are inlining through a 'nounwind' call site then any inlined 'unwind'
965 // instructions are unreachable.
966 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind)
967 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
968 BB != E; ++BB) {
969 TerminatorInst *Term = BB->getTerminator();
970 if (isa<UnwindInst>(Term)) {
971 new UnreachableInst(Context, Term);
972 BB->getInstList().erase(Term);
973 }
974 }
976 // If we are inlining for an invoke instruction, we must make sure to rewrite
977 // any inlined 'unwind' instructions into branches to the invoke exception
978 // destination, and call instructions into invoke instructions.
979 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
980 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
982 // If we cloned in _exactly one_ basic block, and if that block ends in a
983 // return instruction, we splice the body of the inlined callee directly into
984 // the calling basic block.
985 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
986 // Move all of the instructions right before the call.
987 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
988 FirstNewBlock->begin(), FirstNewBlock->end());
989 // Remove the cloned basic block.
990 Caller->getBasicBlockList().pop_back();
992 // If the call site was an invoke instruction, add a branch to the normal
993 // destination.
994 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
995 BranchInst::Create(II->getNormalDest(), TheCall);
997 // If the return instruction returned a value, replace uses of the call with
998 // uses of the returned value.
999 if (!TheCall->use_empty()) {
1000 ReturnInst *R = Returns[0];
1001 if (TheCall == R->getReturnValue())
1002 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1003 else
1004 TheCall->replaceAllUsesWith(R->getReturnValue());
1005 }
1006 // Since we are now done with the Call/Invoke, we can delete it.
1007 TheCall->eraseFromParent();
1009 // Since we are now done with the return instruction, delete it also.
1010 Returns[0]->eraseFromParent();
1012 // We are now done with the inlining.
1013 return true;
1014 }
1016 // Otherwise, we have the normal case, of more than one block to inline or
1017 // multiple return sites.
1019 // We want to clone the entire callee function into the hole between the
1020 // "starter" and "ender" blocks. How we accomplish this depends on whether
1021 // this is an invoke instruction or a call instruction.
1022 BasicBlock *AfterCallBB;
1023 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
1025 // Add an unconditional branch to make this look like the CallInst case...
1026 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
1028 // Split the basic block. This guarantees that no PHI nodes will have to be
1029 // updated due to new incoming edges, and make the invoke case more
1030 // symmetric to the call case.
1031 AfterCallBB = OrigBB->splitBasicBlock(NewBr,
1032 CalledFunc->getName()+".exit");
1034 } else { // It's a call
1035 // If this is a call instruction, we need to split the basic block that
1036 // the call lives in.
1037 //
1038 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
1039 CalledFunc->getName()+".exit");
1040 }
1042 // Change the branch that used to go to AfterCallBB to branch to the first
1043 // basic block of the inlined function.
1044 //
1045 TerminatorInst *Br = OrigBB->getTerminator();
1046 assert(Br && Br->getOpcode() == Instruction::Br &&
1047 "splitBasicBlock broken!");
1048 Br->setOperand(0, FirstNewBlock);
1051 // Now that the function is correct, make it a little bit nicer. In
1052 // particular, move the basic blocks inserted from the end of the function
1053 // into the space made by splitting the source basic block.
1054 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
1055 FirstNewBlock, Caller->end());
1057 // Handle all of the return instructions that we just cloned in, and eliminate
1058 // any users of the original call/invoke instruction.
1059 Type *RTy = CalledFunc->getReturnType();
1061 PHINode *PHI = 0;
1062 if (Returns.size() > 1) {
1063 // The PHI node should go at the front of the new basic block to merge all
1064 // possible incoming values.
1065 if (!TheCall->use_empty()) {
1066 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
1067 AfterCallBB->begin());
1068 // Anything that used the result of the function call should now use the
1069 // PHI node as their operand.
1070 TheCall->replaceAllUsesWith(PHI);
1071 }
1073 // Loop over all of the return instructions adding entries to the PHI node
1074 // as appropriate.
1075 if (PHI) {
1076 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1077 ReturnInst *RI = Returns[i];
1078 assert(RI->getReturnValue()->getType() == PHI->getType() &&
1079 "Ret value not consistent in function!");
1080 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
1081 }
1082 }
1085 // Add a branch to the merge points and remove return instructions.
1086 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
1087 ReturnInst *RI = Returns[i];
1088 BranchInst::Create(AfterCallBB, RI);
1089 RI->eraseFromParent();
1090 }
1091 } else if (!Returns.empty()) {
1092 // Otherwise, if there is exactly one return value, just replace anything
1093 // using the return value of the call with the computed value.
1094 if (!TheCall->use_empty()) {
1095 if (TheCall == Returns[0]->getReturnValue())
1096 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1097 else
1098 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
1099 }
1101 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
1102 BasicBlock *ReturnBB = Returns[0]->getParent();
1103 ReturnBB->replaceAllUsesWith(AfterCallBB);
1105 // Splice the code from the return block into the block that it will return
1106 // to, which contains the code that was after the call.
1107 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
1108 ReturnBB->getInstList());
1110 // Delete the return instruction now and empty ReturnBB now.
1111 Returns[0]->eraseFromParent();
1112 ReturnBB->eraseFromParent();
1113 } else if (!TheCall->use_empty()) {
1114 // No returns, but something is using the return value of the call. Just
1115 // nuke the result.
1116 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
1117 }
1119 // Since we are now done with the Call/Invoke, we can delete it.
1120 TheCall->eraseFromParent();
1122 // We should always be able to fold the entry block of the function into the
1123 // single predecessor of the block...
1124 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
1125 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
1127 // Splice the code entry block into calling block, right before the
1128 // unconditional branch.
1129 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
1130 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
1132 // Remove the unconditional branch.
1133 OrigBB->getInstList().erase(Br);
1135 // Now we can remove the CalleeEntry block, which is now empty.
1136 Caller->getBasicBlockList().erase(CalleeEntry);
1138 // If we inserted a phi node, check to see if it has a single value (e.g. all
1139 // the entries are the same or undef). If so, remove the PHI so it doesn't
1140 // block other optimizations.
1141 if (PHI) {
1142 if (Value *V = SimplifyInstruction(PHI, IFI.TD)) {
1143 PHI->replaceAllUsesWith(V);
1144 PHI->eraseFromParent();
1145 }
1146 }
1148 return true;
1149 }