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 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Utils/Cloning.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/ADT/StringExtras.h"
18 #include "llvm/Analysis/CallGraph.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/IR/Attributes.h"
21 #include "llvm/IR/CallSite.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/DebugInfo.h"
25 #include "llvm/IR/DerivedTypes.h"
26 #include "llvm/IR/IRBuilder.h"
27 #include "llvm/IR/Instructions.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/IR/Intrinsics.h"
30 #include "llvm/IR/Module.h"
31 #include "llvm/Transforms/Utils/Local.h"
32 using namespace llvm;
34 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI,
35 bool InsertLifetime) {
36 return InlineFunction(CallSite(CI), IFI, InsertLifetime);
37 }
38 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI,
39 bool InsertLifetime) {
40 return InlineFunction(CallSite(II), IFI, InsertLifetime);
41 }
43 namespace {
44 /// A class for recording information about inlining through an invoke.
45 class InvokeInliningInfo {
46 BasicBlock *OuterResumeDest; ///< Destination of the invoke's unwind.
47 BasicBlock *InnerResumeDest; ///< Destination for the callee's resume.
48 LandingPadInst *CallerLPad; ///< LandingPadInst associated with the invoke.
49 PHINode *InnerEHValuesPHI; ///< PHI for EH values from landingpad insts.
50 SmallVector<Value*, 8> UnwindDestPHIValues;
52 public:
53 InvokeInliningInfo(InvokeInst *II)
54 : OuterResumeDest(II->getUnwindDest()), InnerResumeDest(0),
55 CallerLPad(0), InnerEHValuesPHI(0) {
56 // If there are PHI nodes in the unwind destination block, we need to keep
57 // track of which values came into them from the invoke before removing
58 // the edge from this block.
59 llvm::BasicBlock *InvokeBB = II->getParent();
60 BasicBlock::iterator I = OuterResumeDest->begin();
61 for (; isa<PHINode>(I); ++I) {
62 // Save the value to use for this edge.
63 PHINode *PHI = cast<PHINode>(I);
64 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
65 }
67 CallerLPad = cast<LandingPadInst>(I);
68 }
70 /// getOuterResumeDest - The outer unwind destination is the target of
71 /// unwind edges introduced for calls within the inlined function.
72 BasicBlock *getOuterResumeDest() const {
73 return OuterResumeDest;
74 }
76 BasicBlock *getInnerResumeDest();
78 LandingPadInst *getLandingPadInst() const { return CallerLPad; }
80 /// forwardResume - Forward the 'resume' instruction to the caller's landing
81 /// pad block. When the landing pad block has only one predecessor, this is
82 /// a simple branch. When there is more than one predecessor, we need to
83 /// split the landing pad block after the landingpad instruction and jump
84 /// to there.
85 void forwardResume(ResumeInst *RI,
86 SmallPtrSet<LandingPadInst*, 16> &InlinedLPads);
88 /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind
89 /// destination block for the given basic block, using the values for the
90 /// original invoke's source block.
91 void addIncomingPHIValuesFor(BasicBlock *BB) const {
92 addIncomingPHIValuesForInto(BB, OuterResumeDest);
93 }
95 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const {
96 BasicBlock::iterator I = dest->begin();
97 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
98 PHINode *phi = cast<PHINode>(I);
99 phi->addIncoming(UnwindDestPHIValues[i], src);
100 }
101 }
102 };
103 }
105 /// getInnerResumeDest - Get or create a target for the branch from ResumeInsts.
106 BasicBlock *InvokeInliningInfo::getInnerResumeDest() {
107 if (InnerResumeDest) return InnerResumeDest;
109 // Split the landing pad.
110 BasicBlock::iterator SplitPoint = CallerLPad; ++SplitPoint;
111 InnerResumeDest =
112 OuterResumeDest->splitBasicBlock(SplitPoint,
113 OuterResumeDest->getName() + ".body");
115 // The number of incoming edges we expect to the inner landing pad.
116 const unsigned PHICapacity = 2;
118 // Create corresponding new PHIs for all the PHIs in the outer landing pad.
119 BasicBlock::iterator InsertPoint = InnerResumeDest->begin();
120 BasicBlock::iterator I = OuterResumeDest->begin();
121 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
122 PHINode *OuterPHI = cast<PHINode>(I);
123 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity,
124 OuterPHI->getName() + ".lpad-body",
125 InsertPoint);
126 OuterPHI->replaceAllUsesWith(InnerPHI);
127 InnerPHI->addIncoming(OuterPHI, OuterResumeDest);
128 }
130 // Create a PHI for the exception values.
131 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity,
132 "eh.lpad-body", InsertPoint);
133 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI);
134 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest);
136 // All done.
137 return InnerResumeDest;
138 }
140 /// forwardResume - Forward the 'resume' instruction to the caller's landing pad
141 /// block. When the landing pad block has only one predecessor, this is a simple
142 /// branch. When there is more than one predecessor, we need to split the
143 /// landing pad block after the landingpad instruction and jump to there.
144 void InvokeInliningInfo::forwardResume(ResumeInst *RI,
145 SmallPtrSet<LandingPadInst*, 16> &InlinedLPads) {
146 BasicBlock *Dest = getInnerResumeDest();
147 BasicBlock *Src = RI->getParent();
149 BranchInst::Create(Dest, Src);
151 // Update the PHIs in the destination. They were inserted in an order which
152 // makes this work.
153 addIncomingPHIValuesForInto(Src, Dest);
155 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src);
156 RI->eraseFromParent();
157 }
159 /// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
160 /// an invoke, we have to turn all of the calls that can throw into
161 /// invokes. This function analyze BB to see if there are any calls, and if so,
162 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
163 /// nodes in that block with the values specified in InvokeDestPHIValues.
164 static void HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
165 InvokeInliningInfo &Invoke) {
166 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
167 Instruction *I = BBI++;
169 // We only need to check for function calls: inlined invoke
170 // instructions require no special handling.
171 CallInst *CI = dyn_cast<CallInst>(I);
173 // If this call cannot unwind, don't convert it to an invoke.
174 // Inline asm calls cannot throw.
175 if (!CI || CI->doesNotThrow() || isa<InlineAsm>(CI->getCalledValue()))
176 continue;
178 // Convert this function call into an invoke instruction. First, split the
179 // basic block.
180 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
182 // Delete the unconditional branch inserted by splitBasicBlock
183 BB->getInstList().pop_back();
185 // Create the new invoke instruction.
186 ImmutableCallSite CS(CI);
187 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
188 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split,
189 Invoke.getOuterResumeDest(),
190 InvokeArgs, CI->getName(), BB);
191 II->setCallingConv(CI->getCallingConv());
192 II->setAttributes(CI->getAttributes());
194 // Make sure that anything using the call now uses the invoke! This also
195 // updates the CallGraph if present, because it uses a WeakVH.
196 CI->replaceAllUsesWith(II);
198 // Delete the original call
199 Split->getInstList().pop_front();
201 // Update any PHI nodes in the exceptional block to indicate that there is
202 // now a new entry in them.
203 Invoke.addIncomingPHIValuesFor(BB);
204 return;
205 }
206 }
208 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
209 /// in the body of the inlined function into invokes.
210 ///
211 /// II is the invoke instruction being inlined. FirstNewBlock is the first
212 /// block of the inlined code (the last block is the end of the function),
213 /// and InlineCodeInfo is information about the code that got inlined.
214 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
215 ClonedCodeInfo &InlinedCodeInfo) {
216 BasicBlock *InvokeDest = II->getUnwindDest();
218 Function *Caller = FirstNewBlock->getParent();
220 // The inlined code is currently at the end of the function, scan from the
221 // start of the inlined code to its end, checking for stuff we need to
222 // rewrite.
223 InvokeInliningInfo Invoke(II);
225 // Get all of the inlined landing pad instructions.
226 SmallPtrSet<LandingPadInst*, 16> InlinedLPads;
227 for (Function::iterator I = FirstNewBlock, E = Caller->end(); I != E; ++I)
228 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator()))
229 InlinedLPads.insert(II->getLandingPadInst());
231 // Append the clauses from the outer landing pad instruction into the inlined
232 // landing pad instructions.
233 LandingPadInst *OuterLPad = Invoke.getLandingPadInst();
234 for (SmallPtrSet<LandingPadInst*, 16>::iterator I = InlinedLPads.begin(),
235 E = InlinedLPads.end(); I != E; ++I) {
236 LandingPadInst *InlinedLPad = *I;
237 unsigned OuterNum = OuterLPad->getNumClauses();
238 InlinedLPad->reserveClauses(OuterNum);
239 for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx)
240 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx));
241 if (OuterLPad->isCleanup())
242 InlinedLPad->setCleanup(true);
243 }
245 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
246 if (InlinedCodeInfo.ContainsCalls)
247 HandleCallsInBlockInlinedThroughInvoke(BB, Invoke);
249 // Forward any resumes that are remaining here.
250 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator()))
251 Invoke.forwardResume(RI, InlinedLPads);
252 }
254 // Now that everything is happy, we have one final detail. The PHI nodes in
255 // the exception destination block still have entries due to the original
256 // invoke instruction. Eliminate these entries (which might even delete the
257 // PHI node) now.
258 InvokeDest->removePredecessor(II->getParent());
259 }
261 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
262 /// into the caller, update the specified callgraph to reflect the changes we
263 /// made. Note that it's possible that not all code was copied over, so only
264 /// some edges of the callgraph may remain.
265 static void UpdateCallGraphAfterInlining(CallSite CS,
266 Function::iterator FirstNewBlock,
267 ValueToValueMapTy &VMap,
268 InlineFunctionInfo &IFI) {
269 CallGraph &CG = *IFI.CG;
270 const Function *Caller = CS.getInstruction()->getParent()->getParent();
271 const Function *Callee = CS.getCalledFunction();
272 CallGraphNode *CalleeNode = CG[Callee];
273 CallGraphNode *CallerNode = CG[Caller];
275 // Since we inlined some uninlined call sites in the callee into the caller,
276 // add edges from the caller to all of the callees of the callee.
277 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
279 // Consider the case where CalleeNode == CallerNode.
280 CallGraphNode::CalledFunctionsVector CallCache;
281 if (CalleeNode == CallerNode) {
282 CallCache.assign(I, E);
283 I = CallCache.begin();
284 E = CallCache.end();
285 }
287 for (; I != E; ++I) {
288 const Value *OrigCall = I->first;
290 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
291 // Only copy the edge if the call was inlined!
292 if (VMI == VMap.end() || VMI->second == 0)
293 continue;
295 // If the call was inlined, but then constant folded, there is no edge to
296 // add. Check for this case.
297 Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
298 if (NewCall == 0) continue;
300 // Remember that this call site got inlined for the client of
301 // InlineFunction.
302 IFI.InlinedCalls.push_back(NewCall);
304 // It's possible that inlining the callsite will cause it to go from an
305 // indirect to a direct call by resolving a function pointer. If this
306 // happens, set the callee of the new call site to a more precise
307 // destination. This can also happen if the call graph node of the caller
308 // was just unnecessarily imprecise.
309 if (I->second->getFunction() == 0)
310 if (Function *F = CallSite(NewCall).getCalledFunction()) {
311 // Indirect call site resolved to direct call.
312 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
314 continue;
315 }
317 CallerNode->addCalledFunction(CallSite(NewCall), I->second);
318 }
320 // Update the call graph by deleting the edge from Callee to Caller. We must
321 // do this after the loop above in case Caller and Callee are the same.
322 CallerNode->removeCallEdgeFor(CS);
323 }
325 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
326 BasicBlock *InsertBlock,
327 InlineFunctionInfo &IFI) {
328 LLVMContext &Context = Src->getContext();
329 Type *VoidPtrTy = Type::getInt8PtrTy(Context);
330 Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
331 Type *Tys[3] = { VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context) };
332 Function *MemCpyFn = Intrinsic::getDeclaration(M, Intrinsic::memcpy, Tys);
333 IRBuilder<> builder(InsertBlock->begin());
334 Value *DstCast = builder.CreateBitCast(Dst, VoidPtrTy, "tmp");
335 Value *SrcCast = builder.CreateBitCast(Src, VoidPtrTy, "tmp");
337 Value *Size;
338 if (IFI.DL == 0)
339 Size = ConstantExpr::getSizeOf(AggTy);
340 else
341 Size = ConstantInt::get(Type::getInt64Ty(Context),
342 IFI.DL->getTypeStoreSize(AggTy));
344 // Always generate a memcpy of alignment 1 here because we don't know
345 // the alignment of the src pointer. Other optimizations can infer
346 // better alignment.
347 Value *CallArgs[] = {
348 DstCast, SrcCast, Size,
349 ConstantInt::get(Type::getInt32Ty(Context), 1),
350 ConstantInt::getFalse(Context) // isVolatile
351 };
352 builder.CreateCall(MemCpyFn, CallArgs);
353 }
355 /// HandleByValArgument - When inlining a call site that has a byval argument,
356 /// we have to make the implicit memcpy explicit by adding it.
357 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
358 const Function *CalledFunc,
359 InlineFunctionInfo &IFI,
360 unsigned ByValAlignment) {
361 Type *AggTy = cast<PointerType>(Arg->getType())->getElementType();
363 // If the called function is readonly, then it could not mutate the caller's
364 // copy of the byval'd memory. In this case, it is safe to elide the copy and
365 // temporary.
366 if (CalledFunc->onlyReadsMemory()) {
367 // If the byval argument has a specified alignment that is greater than the
368 // passed in pointer, then we either have to round up the input pointer or
369 // give up on this transformation.
370 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment.
371 return Arg;
373 // If the pointer is already known to be sufficiently aligned, or if we can
374 // round it up to a larger alignment, then we don't need a temporary.
375 if (getOrEnforceKnownAlignment(Arg, ByValAlignment,
376 IFI.DL) >= ByValAlignment)
377 return Arg;
379 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad
380 // for code quality, but rarely happens and is required for correctness.
381 }
383 // Create the alloca. If we have DataLayout, use nice alignment.
384 unsigned Align = 1;
385 if (IFI.DL)
386 Align = IFI.DL->getPrefTypeAlignment(AggTy);
388 // If the byval had an alignment specified, we *must* use at least that
389 // alignment, as it is required by the byval argument (and uses of the
390 // pointer inside the callee).
391 Align = std::max(Align, ByValAlignment);
393 Function *Caller = TheCall->getParent()->getParent();
395 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(),
396 &*Caller->begin()->begin());
398 // Uses of the argument in the function should use our new alloca
399 // instead.
400 return NewAlloca;
401 }
403 // isUsedByLifetimeMarker - Check whether this Value is used by a lifetime
404 // intrinsic.
405 static bool isUsedByLifetimeMarker(Value *V) {
406 for (User *U : V->users()) {
407 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
408 switch (II->getIntrinsicID()) {
409 default: break;
410 case Intrinsic::lifetime_start:
411 case Intrinsic::lifetime_end:
412 return true;
413 }
414 }
415 }
416 return false;
417 }
419 // hasLifetimeMarkers - Check whether the given alloca already has
420 // lifetime.start or lifetime.end intrinsics.
421 static bool hasLifetimeMarkers(AllocaInst *AI) {
422 Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext());
423 if (AI->getType() == Int8PtrTy)
424 return isUsedByLifetimeMarker(AI);
426 // Do a scan to find all the casts to i8*.
427 for (User *U : AI->users()) {
428 if (U->getType() != Int8PtrTy) continue;
429 if (U->stripPointerCasts() != AI) continue;
430 if (isUsedByLifetimeMarker(U))
431 return true;
432 }
433 return false;
434 }
436 /// updateInlinedAtInfo - Helper function used by fixupLineNumbers to
437 /// recursively update InlinedAtEntry of a DebugLoc.
438 static DebugLoc updateInlinedAtInfo(const DebugLoc &DL,
439 const DebugLoc &InlinedAtDL,
440 LLVMContext &Ctx) {
441 if (MDNode *IA = DL.getInlinedAt(Ctx)) {
442 DebugLoc NewInlinedAtDL
443 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx);
444 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
445 NewInlinedAtDL.getAsMDNode(Ctx));
446 }
448 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx),
449 InlinedAtDL.getAsMDNode(Ctx));
450 }
452 /// fixupLineNumbers - Update inlined instructions' line numbers to
453 /// to encode location where these instructions are inlined.
454 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
455 Instruction *TheCall) {
456 DebugLoc TheCallDL = TheCall->getDebugLoc();
457 if (TheCallDL.isUnknown())
458 return;
460 for (; FI != Fn->end(); ++FI) {
461 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
462 BI != BE; ++BI) {
463 DebugLoc DL = BI->getDebugLoc();
464 if (!DL.isUnknown()) {
465 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext()));
466 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) {
467 LLVMContext &Ctx = BI->getContext();
468 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx);
469 DVI->setOperand(2, createInlinedVariable(DVI->getVariable(),
470 InlinedAt, Ctx));
471 }
472 }
473 }
474 }
475 }
477 /// InlineFunction - This function inlines the called function into the basic
478 /// block of the caller. This returns false if it is not possible to inline
479 /// this call. The program is still in a well defined state if this occurs
480 /// though.
481 ///
482 /// Note that this only does one level of inlining. For example, if the
483 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
484 /// exists in the instruction stream. Similarly this will inline a recursive
485 /// function by one level.
486 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
487 bool InsertLifetime) {
488 Instruction *TheCall = CS.getInstruction();
489 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
490 "Instruction not in function!");
492 // If IFI has any state in it, zap it before we fill it in.
493 IFI.reset();
495 const Function *CalledFunc = CS.getCalledFunction();
496 if (CalledFunc == 0 || // Can't inline external function or indirect
497 CalledFunc->isDeclaration() || // call, or call to a vararg function!
498 CalledFunc->getFunctionType()->isVarArg()) return false;
500 // If the call to the callee is not a tail call, we must clear the 'tail'
501 // flags on any calls that we inline.
502 bool MustClearTailCallFlags =
503 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall());
505 // If the call to the callee cannot throw, set the 'nounwind' flag on any
506 // calls that we inline.
507 bool MarkNoUnwind = CS.doesNotThrow();
509 BasicBlock *OrigBB = TheCall->getParent();
510 Function *Caller = OrigBB->getParent();
512 // GC poses two hazards to inlining, which only occur when the callee has GC:
513 // 1. If the caller has no GC, then the callee's GC must be propagated to the
514 // caller.
515 // 2. If the caller has a differing GC, it is invalid to inline.
516 if (CalledFunc->hasGC()) {
517 if (!Caller->hasGC())
518 Caller->setGC(CalledFunc->getGC());
519 else if (CalledFunc->getGC() != Caller->getGC())
520 return false;
521 }
523 // Get the personality function from the callee if it contains a landing pad.
524 Value *CalleePersonality = 0;
525 for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end();
526 I != E; ++I)
527 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
528 const BasicBlock *BB = II->getUnwindDest();
529 const LandingPadInst *LP = BB->getLandingPadInst();
530 CalleePersonality = LP->getPersonalityFn();
531 break;
532 }
534 // Find the personality function used by the landing pads of the caller. If it
535 // exists, then check to see that it matches the personality function used in
536 // the callee.
537 if (CalleePersonality) {
538 for (Function::const_iterator I = Caller->begin(), E = Caller->end();
539 I != E; ++I)
540 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
541 const BasicBlock *BB = II->getUnwindDest();
542 const LandingPadInst *LP = BB->getLandingPadInst();
544 // If the personality functions match, then we can perform the
545 // inlining. Otherwise, we can't inline.
546 // TODO: This isn't 100% true. Some personality functions are proper
547 // supersets of others and can be used in place of the other.
548 if (LP->getPersonalityFn() != CalleePersonality)
549 return false;
551 break;
552 }
553 }
555 // Get an iterator to the last basic block in the function, which will have
556 // the new function inlined after it.
557 Function::iterator LastBlock = &Caller->back();
559 // Make sure to capture all of the return instructions from the cloned
560 // function.
561 SmallVector<ReturnInst*, 8> Returns;
562 ClonedCodeInfo InlinedFunctionInfo;
563 Function::iterator FirstNewBlock;
565 { // Scope to destroy VMap after cloning.
566 ValueToValueMapTy VMap;
567 // Keep a list of pair (dst, src) to emit byval initializations.
568 SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
570 assert(CalledFunc->arg_size() == CS.arg_size() &&
571 "No varargs calls can be inlined!");
573 // Calculate the vector of arguments to pass into the function cloner, which
574 // matches up the formal to the actual argument values.
575 CallSite::arg_iterator AI = CS.arg_begin();
576 unsigned ArgNo = 0;
577 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
578 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
579 Value *ActualArg = *AI;
581 // When byval arguments actually inlined, we need to make the copy implied
582 // by them explicit. However, we don't do this if the callee is readonly
583 // or readnone, because the copy would be unneeded: the callee doesn't
584 // modify the struct.
585 if (CS.isByValArgument(ArgNo)) {
586 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
587 CalledFunc->getParamAlignment(ArgNo+1));
589 // Calls that we inline may use the new alloca, so we need to clear
590 // their 'tail' flags if HandleByValArgument introduced a new alloca and
591 // the callee has calls.
592 if (ActualArg != *AI) {
593 MustClearTailCallFlags = true;
594 ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
595 }
597 }
599 VMap[I] = ActualArg;
600 }
602 // We want the inliner to prune the code as it copies. We would LOVE to
603 // have no dead or constant instructions leftover after inlining occurs
604 // (which can happen, e.g., because an argument was constant), but we'll be
605 // happy with whatever the cloner can do.
606 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
607 /*ModuleLevelChanges=*/false, Returns, ".i",
608 &InlinedFunctionInfo, IFI.DL, TheCall);
610 // Remember the first block that is newly cloned over.
611 FirstNewBlock = LastBlock; ++FirstNewBlock;
613 // Inject byval arguments initialization.
614 for (std::pair<Value*, Value*> &Init : ByValInit)
615 HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
616 FirstNewBlock, IFI);
618 // Update the callgraph if requested.
619 if (IFI.CG)
620 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
622 // Update inlined instructions' line number information.
623 fixupLineNumbers(Caller, FirstNewBlock, TheCall);
624 }
626 // If there are any alloca instructions in the block that used to be the entry
627 // block for the callee, move them to the entry block of the caller. First
628 // calculate which instruction they should be inserted before. We insert the
629 // instructions at the end of the current alloca list.
630 {
631 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
632 for (BasicBlock::iterator I = FirstNewBlock->begin(),
633 E = FirstNewBlock->end(); I != E; ) {
634 AllocaInst *AI = dyn_cast<AllocaInst>(I++);
635 if (AI == 0) continue;
637 // If the alloca is now dead, remove it. This often occurs due to code
638 // specialization.
639 if (AI->use_empty()) {
640 AI->eraseFromParent();
641 continue;
642 }
644 if (!isa<Constant>(AI->getArraySize()))
645 continue;
647 // Keep track of the static allocas that we inline into the caller.
648 IFI.StaticAllocas.push_back(AI);
650 // Scan for the block of allocas that we can move over, and move them
651 // all at once.
652 while (isa<AllocaInst>(I) &&
653 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
654 IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
655 ++I;
656 }
658 // Transfer all of the allocas over in a block. Using splice means
659 // that the instructions aren't removed from the symbol table, then
660 // reinserted.
661 Caller->getEntryBlock().getInstList().splice(InsertPoint,
662 FirstNewBlock->getInstList(),
663 AI, I);
664 }
665 }
667 // Leave lifetime markers for the static alloca's, scoping them to the
668 // function we just inlined.
669 if (InsertLifetime && !IFI.StaticAllocas.empty()) {
670 IRBuilder<> builder(FirstNewBlock->begin());
671 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
672 AllocaInst *AI = IFI.StaticAllocas[ai];
674 // If the alloca is already scoped to something smaller than the whole
675 // function then there's no need to add redundant, less accurate markers.
676 if (hasLifetimeMarkers(AI))
677 continue;
679 // Try to determine the size of the allocation.
680 ConstantInt *AllocaSize = 0;
681 if (ConstantInt *AIArraySize =
682 dyn_cast<ConstantInt>(AI->getArraySize())) {
683 if (IFI.DL) {
684 Type *AllocaType = AI->getAllocatedType();
685 uint64_t AllocaTypeSize = IFI.DL->getTypeAllocSize(AllocaType);
686 uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
687 assert(AllocaArraySize > 0 && "array size of AllocaInst is zero");
688 // Check that array size doesn't saturate uint64_t and doesn't
689 // overflow when it's multiplied by type size.
690 if (AllocaArraySize != ~0ULL &&
691 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
692 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
693 AllocaArraySize * AllocaTypeSize);
694 }
695 }
696 }
698 builder.CreateLifetimeStart(AI, AllocaSize);
699 for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) {
700 IRBuilder<> builder(Returns[ri]);
701 builder.CreateLifetimeEnd(AI, AllocaSize);
702 }
703 }
704 }
706 // If the inlined code contained dynamic alloca instructions, wrap the inlined
707 // code with llvm.stacksave/llvm.stackrestore intrinsics.
708 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
709 Module *M = Caller->getParent();
710 // Get the two intrinsics we care about.
711 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
712 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
714 // Insert the llvm.stacksave.
715 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
716 .CreateCall(StackSave, "savedstack");
718 // Insert a call to llvm.stackrestore before any return instructions in the
719 // inlined function.
720 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
721 IRBuilder<>(Returns[i]).CreateCall(StackRestore, SavedPtr);
722 }
723 }
725 // If we are inlining tail call instruction through a call site that isn't
726 // marked 'tail', we must remove the tail marker for any calls in the inlined
727 // code. Also, calls inlined through a 'nounwind' call site should be marked
728 // 'nounwind'.
729 if (InlinedFunctionInfo.ContainsCalls &&
730 (MustClearTailCallFlags || MarkNoUnwind)) {
731 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
732 BB != E; ++BB)
733 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
734 if (CallInst *CI = dyn_cast<CallInst>(I)) {
735 if (MustClearTailCallFlags)
736 CI->setTailCall(false);
737 if (MarkNoUnwind)
738 CI->setDoesNotThrow();
739 }
740 }
742 // If we are inlining for an invoke instruction, we must make sure to rewrite
743 // any call instructions into invoke instructions.
744 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
745 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
747 // If we cloned in _exactly one_ basic block, and if that block ends in a
748 // return instruction, we splice the body of the inlined callee directly into
749 // the calling basic block.
750 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
751 // Move all of the instructions right before the call.
752 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
753 FirstNewBlock->begin(), FirstNewBlock->end());
754 // Remove the cloned basic block.
755 Caller->getBasicBlockList().pop_back();
757 // If the call site was an invoke instruction, add a branch to the normal
758 // destination.
759 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
760 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
761 NewBr->setDebugLoc(Returns[0]->getDebugLoc());
762 }
764 // If the return instruction returned a value, replace uses of the call with
765 // uses of the returned value.
766 if (!TheCall->use_empty()) {
767 ReturnInst *R = Returns[0];
768 if (TheCall == R->getReturnValue())
769 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
770 else
771 TheCall->replaceAllUsesWith(R->getReturnValue());
772 }
773 // Since we are now done with the Call/Invoke, we can delete it.
774 TheCall->eraseFromParent();
776 // Since we are now done with the return instruction, delete it also.
777 Returns[0]->eraseFromParent();
779 // We are now done with the inlining.
780 return true;
781 }
783 // Otherwise, we have the normal case, of more than one block to inline or
784 // multiple return sites.
786 // We want to clone the entire callee function into the hole between the
787 // "starter" and "ender" blocks. How we accomplish this depends on whether
788 // this is an invoke instruction or a call instruction.
789 BasicBlock *AfterCallBB;
790 BranchInst *CreatedBranchToNormalDest = NULL;
791 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
793 // Add an unconditional branch to make this look like the CallInst case...
794 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
796 // Split the basic block. This guarantees that no PHI nodes will have to be
797 // updated due to new incoming edges, and make the invoke case more
798 // symmetric to the call case.
799 AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest,
800 CalledFunc->getName()+".exit");
802 } else { // It's a call
803 // If this is a call instruction, we need to split the basic block that
804 // the call lives in.
805 //
806 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
807 CalledFunc->getName()+".exit");
808 }
810 // Change the branch that used to go to AfterCallBB to branch to the first
811 // basic block of the inlined function.
812 //
813 TerminatorInst *Br = OrigBB->getTerminator();
814 assert(Br && Br->getOpcode() == Instruction::Br &&
815 "splitBasicBlock broken!");
816 Br->setOperand(0, FirstNewBlock);
819 // Now that the function is correct, make it a little bit nicer. In
820 // particular, move the basic blocks inserted from the end of the function
821 // into the space made by splitting the source basic block.
822 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
823 FirstNewBlock, Caller->end());
825 // Handle all of the return instructions that we just cloned in, and eliminate
826 // any users of the original call/invoke instruction.
827 Type *RTy = CalledFunc->getReturnType();
829 PHINode *PHI = 0;
830 if (Returns.size() > 1) {
831 // The PHI node should go at the front of the new basic block to merge all
832 // possible incoming values.
833 if (!TheCall->use_empty()) {
834 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
835 AfterCallBB->begin());
836 // Anything that used the result of the function call should now use the
837 // PHI node as their operand.
838 TheCall->replaceAllUsesWith(PHI);
839 }
841 // Loop over all of the return instructions adding entries to the PHI node
842 // as appropriate.
843 if (PHI) {
844 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
845 ReturnInst *RI = Returns[i];
846 assert(RI->getReturnValue()->getType() == PHI->getType() &&
847 "Ret value not consistent in function!");
848 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
849 }
850 }
853 // Add a branch to the merge points and remove return instructions.
854 DebugLoc Loc;
855 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
856 ReturnInst *RI = Returns[i];
857 BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
858 Loc = RI->getDebugLoc();
859 BI->setDebugLoc(Loc);
860 RI->eraseFromParent();
861 }
862 // We need to set the debug location to *somewhere* inside the
863 // inlined function. The line number may be nonsensical, but the
864 // instruction will at least be associated with the right
865 // function.
866 if (CreatedBranchToNormalDest)
867 CreatedBranchToNormalDest->setDebugLoc(Loc);
868 } else if (!Returns.empty()) {
869 // Otherwise, if there is exactly one return value, just replace anything
870 // using the return value of the call with the computed value.
871 if (!TheCall->use_empty()) {
872 if (TheCall == Returns[0]->getReturnValue())
873 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
874 else
875 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
876 }
878 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
879 BasicBlock *ReturnBB = Returns[0]->getParent();
880 ReturnBB->replaceAllUsesWith(AfterCallBB);
882 // Splice the code from the return block into the block that it will return
883 // to, which contains the code that was after the call.
884 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
885 ReturnBB->getInstList());
887 if (CreatedBranchToNormalDest)
888 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
890 // Delete the return instruction now and empty ReturnBB now.
891 Returns[0]->eraseFromParent();
892 ReturnBB->eraseFromParent();
893 } else if (!TheCall->use_empty()) {
894 // No returns, but something is using the return value of the call. Just
895 // nuke the result.
896 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
897 }
899 // Since we are now done with the Call/Invoke, we can delete it.
900 TheCall->eraseFromParent();
902 // We should always be able to fold the entry block of the function into the
903 // single predecessor of the block...
904 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
905 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
907 // Splice the code entry block into calling block, right before the
908 // unconditional branch.
909 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
910 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
912 // Remove the unconditional branch.
913 OrigBB->getInstList().erase(Br);
915 // Now we can remove the CalleeEntry block, which is now empty.
916 Caller->getBasicBlockList().erase(CalleeEntry);
918 // If we inserted a phi node, check to see if it has a single value (e.g. all
919 // the entries are the same or undef). If so, remove the PHI so it doesn't
920 // block other optimizations.
921 if (PHI) {
922 if (Value *V = SimplifyInstruction(PHI, IFI.DL)) {
923 PHI->replaceAllUsesWith(V);
924 PHI->eraseFromParent();
925 }
926 }
928 return true;
929 }