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 //===----------------------------------------------------------------------===//
14 
15 #include "llvm/Transforms/Utils/Cloning.h"
16 #include "llvm/Constants.h"
17 #include "llvm/DerivedTypes.h"
18 #include "llvm/Module.h"
19 #include "llvm/Instructions.h"
20 #include "llvm/IntrinsicInst.h"
21 #include "llvm/Intrinsics.h"
22 #include "llvm/Attributes.h"
23 #include "llvm/Analysis/CallGraph.h"
24 #include "llvm/Analysis/DebugInfo.h"
25 #include "llvm/Target/TargetData.h"
26 #include "llvm/ADT/SmallVector.h"
27 #include "llvm/ADT/StringExtras.h"
28 #include "llvm/Support/CallSite.h"
29 using namespace llvm;
30 
InlineFunction(CallInst * CI,InlineFunctionInfo & IFI)31 bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI) {
32   return InlineFunction(CallSite(CI), IFI);
33 }
InlineFunction(InvokeInst * II,InlineFunctionInfo & IFI)34 bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI) {
35   return InlineFunction(CallSite(II), IFI);
36 }
37 
38 
39 /// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into
40 /// an invoke, we have to turn all of the calls that can throw into
41 /// invokes.  This function analyze BB to see if there are any calls, and if so,
42 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
43 /// nodes in that block with the values specified in InvokeDestPHIValues.
44 ///
HandleCallsInBlockInlinedThroughInvoke(BasicBlock * BB,BasicBlock * InvokeDest,const SmallVectorImpl<Value * > & InvokeDestPHIValues)45 static void HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB,
46                                                    BasicBlock *InvokeDest,
47                            const SmallVectorImpl<Value*> &InvokeDestPHIValues) {
48   for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
49     Instruction *I = BBI++;
50 
51     // We only need to check for function calls: inlined invoke
52     // instructions require no special handling.
53     CallInst *CI = dyn_cast<CallInst>(I);
54     if (CI == 0) continue;
55 
56     // If this call cannot unwind, don't convert it to an invoke.
57     if (CI->doesNotThrow())
58       continue;
59 
60     // Convert this function call into an invoke instruction.
61     // First, split the basic block.
62     BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
63 
64     // Next, create the new invoke instruction, inserting it at the end
65     // of the old basic block.
66     ImmutableCallSite CS(CI);
67     SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end());
68     InvokeInst *II =
69       InvokeInst::Create(CI->getCalledValue(), Split, InvokeDest,
70                          InvokeArgs.begin(), InvokeArgs.end(),
71                          CI->getName(), BB->getTerminator());
72     II->setCallingConv(CI->getCallingConv());
73     II->setAttributes(CI->getAttributes());
74 
75     // Make sure that anything using the call now uses the invoke!  This also
76     // updates the CallGraph if present, because it uses a WeakVH.
77     CI->replaceAllUsesWith(II);
78 
79     // Delete the unconditional branch inserted by splitBasicBlock
80     BB->getInstList().pop_back();
81     Split->getInstList().pop_front();  // Delete the original call
82 
83     // Update any PHI nodes in the exceptional block to indicate that
84     // there is now a new entry in them.
85     unsigned i = 0;
86     for (BasicBlock::iterator I = InvokeDest->begin();
87          isa<PHINode>(I); ++I, ++i)
88       cast<PHINode>(I)->addIncoming(InvokeDestPHIValues[i], BB);
89 
90     // This basic block is now complete, the caller will continue scanning the
91     // next one.
92     return;
93   }
94 }
95 
96 
97 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
98 /// in the body of the inlined function into invokes and turn unwind
99 /// instructions into branches to the invoke unwind dest.
100 ///
101 /// II is the invoke instruction being inlined.  FirstNewBlock is the first
102 /// block of the inlined code (the last block is the end of the function),
103 /// and InlineCodeInfo is information about the code that got inlined.
HandleInlinedInvoke(InvokeInst * II,BasicBlock * FirstNewBlock,ClonedCodeInfo & InlinedCodeInfo)104 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
105                                 ClonedCodeInfo &InlinedCodeInfo) {
106   BasicBlock *InvokeDest = II->getUnwindDest();
107   SmallVector<Value*, 8> InvokeDestPHIValues;
108 
109   // If there are PHI nodes in the unwind destination block, we need to
110   // keep track of which values came into them from this invoke, then remove
111   // the entry for this block.
112   BasicBlock *InvokeBlock = II->getParent();
113   for (BasicBlock::iterator I = InvokeDest->begin(); isa<PHINode>(I); ++I) {
114     PHINode *PN = cast<PHINode>(I);
115     // Save the value to use for this edge.
116     InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(InvokeBlock));
117   }
118 
119   Function *Caller = FirstNewBlock->getParent();
120 
121   // The inlined code is currently at the end of the function, scan from the
122   // start of the inlined code to its end, checking for stuff we need to
123   // rewrite.  If the code doesn't have calls or unwinds, we know there is
124   // nothing to rewrite.
125   if (!InlinedCodeInfo.ContainsCalls && !InlinedCodeInfo.ContainsUnwinds) {
126     // Now that everything is happy, we have one final detail.  The PHI nodes in
127     // the exception destination block still have entries due to the original
128     // invoke instruction.  Eliminate these entries (which might even delete the
129     // PHI node) now.
130     InvokeDest->removePredecessor(II->getParent());
131     return;
132   }
133 
134   for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){
135     if (InlinedCodeInfo.ContainsCalls)
136       HandleCallsInBlockInlinedThroughInvoke(BB, InvokeDest,
137                                              InvokeDestPHIValues);
138 
139     if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
140       // An UnwindInst requires special handling when it gets inlined into an
141       // invoke site.  Once this happens, we know that the unwind would cause
142       // a control transfer to the invoke exception destination, so we can
143       // transform it into a direct branch to the exception destination.
144       BranchInst::Create(InvokeDest, UI);
145 
146       // Delete the unwind instruction!
147       UI->eraseFromParent();
148 
149       // Update any PHI nodes in the exceptional block to indicate that
150       // there is now a new entry in them.
151       unsigned i = 0;
152       for (BasicBlock::iterator I = InvokeDest->begin();
153            isa<PHINode>(I); ++I, ++i) {
154         PHINode *PN = cast<PHINode>(I);
155         PN->addIncoming(InvokeDestPHIValues[i], BB);
156       }
157     }
158   }
159 
160   // Now that everything is happy, we have one final detail.  The PHI nodes in
161   // the exception destination block still have entries due to the original
162   // invoke instruction.  Eliminate these entries (which might even delete the
163   // PHI node) now.
164   InvokeDest->removePredecessor(II->getParent());
165 }
166 
167 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
168 /// into the caller, update the specified callgraph to reflect the changes we
169 /// made.  Note that it's possible that not all code was copied over, so only
170 /// some edges of the callgraph may remain.
UpdateCallGraphAfterInlining(CallSite CS,Function::iterator FirstNewBlock,ValueMap<const Value *,Value * > & VMap,InlineFunctionInfo & IFI)171 static void UpdateCallGraphAfterInlining(CallSite CS,
172                                          Function::iterator FirstNewBlock,
173                                          ValueMap<const Value*, Value*> &VMap,
174                                          InlineFunctionInfo &IFI) {
175   CallGraph &CG = *IFI.CG;
176   const Function *Caller = CS.getInstruction()->getParent()->getParent();
177   const Function *Callee = CS.getCalledFunction();
178   CallGraphNode *CalleeNode = CG[Callee];
179   CallGraphNode *CallerNode = CG[Caller];
180 
181   // Since we inlined some uninlined call sites in the callee into the caller,
182   // add edges from the caller to all of the callees of the callee.
183   CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
184 
185   // Consider the case where CalleeNode == CallerNode.
186   CallGraphNode::CalledFunctionsVector CallCache;
187   if (CalleeNode == CallerNode) {
188     CallCache.assign(I, E);
189     I = CallCache.begin();
190     E = CallCache.end();
191   }
192 
193   for (; I != E; ++I) {
194     const Value *OrigCall = I->first;
195 
196     ValueMap<const Value*, Value*>::iterator VMI = VMap.find(OrigCall);
197     // Only copy the edge if the call was inlined!
198     if (VMI == VMap.end() || VMI->second == 0)
199       continue;
200 
201     // If the call was inlined, but then constant folded, there is no edge to
202     // add.  Check for this case.
203     Instruction *NewCall = dyn_cast<Instruction>(VMI->second);
204     if (NewCall == 0) continue;
205 
206     // Remember that this call site got inlined for the client of
207     // InlineFunction.
208     IFI.InlinedCalls.push_back(NewCall);
209 
210     // It's possible that inlining the callsite will cause it to go from an
211     // indirect to a direct call by resolving a function pointer.  If this
212     // happens, set the callee of the new call site to a more precise
213     // destination.  This can also happen if the call graph node of the caller
214     // was just unnecessarily imprecise.
215     if (I->second->getFunction() == 0)
216       if (Function *F = CallSite(NewCall).getCalledFunction()) {
217         // Indirect call site resolved to direct call.
218         CallerNode->addCalledFunction(CallSite(NewCall), CG[F]);
219 
220         continue;
221       }
222 
223     CallerNode->addCalledFunction(CallSite(NewCall), I->second);
224   }
225 
226   // Update the call graph by deleting the edge from Callee to Caller.  We must
227   // do this after the loop above in case Caller and Callee are the same.
228   CallerNode->removeCallEdgeFor(CS);
229 }
230 
231 // InlineFunction - This function inlines the called function into the basic
232 // block of the caller.  This returns false if it is not possible to inline this
233 // call.  The program is still in a well defined state if this occurs though.
234 //
235 // Note that this only does one level of inlining.  For example, if the
236 // instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
237 // exists in the instruction stream.  Similiarly this will inline a recursive
238 // function by one level.
239 //
InlineFunction(CallSite CS,InlineFunctionInfo & IFI)240 bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI) {
241   Instruction *TheCall = CS.getInstruction();
242   LLVMContext &Context = TheCall->getContext();
243   assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
244          "Instruction not in function!");
245 
246   // If IFI has any state in it, zap it before we fill it in.
247   IFI.reset();
248 
249   const Function *CalledFunc = CS.getCalledFunction();
250   if (CalledFunc == 0 ||          // Can't inline external function or indirect
251       CalledFunc->isDeclaration() || // call, or call to a vararg function!
252       CalledFunc->getFunctionType()->isVarArg()) return false;
253 
254 
255   // If the call to the callee is not a tail call, we must clear the 'tail'
256   // flags on any calls that we inline.
257   bool MustClearTailCallFlags =
258     !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall());
259 
260   // If the call to the callee cannot throw, set the 'nounwind' flag on any
261   // calls that we inline.
262   bool MarkNoUnwind = CS.doesNotThrow();
263 
264   BasicBlock *OrigBB = TheCall->getParent();
265   Function *Caller = OrigBB->getParent();
266 
267   // GC poses two hazards to inlining, which only occur when the callee has GC:
268   //  1. If the caller has no GC, then the callee's GC must be propagated to the
269   //     caller.
270   //  2. If the caller has a differing GC, it is invalid to inline.
271   if (CalledFunc->hasGC()) {
272     if (!Caller->hasGC())
273       Caller->setGC(CalledFunc->getGC());
274     else if (CalledFunc->getGC() != Caller->getGC())
275       return false;
276   }
277 
278   // Get an iterator to the last basic block in the function, which will have
279   // the new function inlined after it.
280   //
281   Function::iterator LastBlock = &Caller->back();
282 
283   // Make sure to capture all of the return instructions from the cloned
284   // function.
285   SmallVector<ReturnInst*, 8> Returns;
286   ClonedCodeInfo InlinedFunctionInfo;
287   Function::iterator FirstNewBlock;
288 
289   { // Scope to destroy VMap after cloning.
290     ValueMap<const Value*, Value*> VMap;
291 
292     assert(CalledFunc->arg_size() == CS.arg_size() &&
293            "No varargs calls can be inlined!");
294 
295     // Calculate the vector of arguments to pass into the function cloner, which
296     // matches up the formal to the actual argument values.
297     CallSite::arg_iterator AI = CS.arg_begin();
298     unsigned ArgNo = 0;
299     for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
300          E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
301       Value *ActualArg = *AI;
302 
303       // When byval arguments actually inlined, we need to make the copy implied
304       // by them explicit.  However, we don't do this if the callee is readonly
305       // or readnone, because the copy would be unneeded: the callee doesn't
306       // modify the struct.
307       if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal) &&
308           !CalledFunc->onlyReadsMemory()) {
309         const Type *AggTy = cast<PointerType>(I->getType())->getElementType();
310         const Type *VoidPtrTy =
311             Type::getInt8PtrTy(Context);
312 
313         // Create the alloca.  If we have TargetData, use nice alignment.
314         unsigned Align = 1;
315         if (IFI.TD) Align = IFI.TD->getPrefTypeAlignment(AggTy);
316         Value *NewAlloca = new AllocaInst(AggTy, 0, Align,
317                                           I->getName(),
318                                           &*Caller->begin()->begin());
319         // Emit a memcpy.
320         const Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)};
321         Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(),
322                                                        Intrinsic::memcpy,
323                                                        Tys, 3);
324         Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall);
325         Value *SrcCast = new BitCastInst(*AI, VoidPtrTy, "tmp", TheCall);
326 
327         Value *Size;
328         if (IFI.TD == 0)
329           Size = ConstantExpr::getSizeOf(AggTy);
330         else
331           Size = ConstantInt::get(Type::getInt64Ty(Context),
332                                   IFI.TD->getTypeStoreSize(AggTy));
333 
334         // Always generate a memcpy of alignment 1 here because we don't know
335         // the alignment of the src pointer.  Other optimizations can infer
336         // better alignment.
337         Value *CallArgs[] = {
338           DestCast, SrcCast, Size,
339           ConstantInt::get(Type::getInt32Ty(Context), 1),
340           ConstantInt::get(Type::getInt1Ty(Context), 0)
341         };
342         CallInst *TheMemCpy =
343           CallInst::Create(MemCpyFn, CallArgs, CallArgs+5, "", TheCall);
344 
345         // If we have a call graph, update it.
346         if (CallGraph *CG = IFI.CG) {
347           CallGraphNode *MemCpyCGN = CG->getOrInsertFunction(MemCpyFn);
348           CallGraphNode *CallerNode = (*CG)[Caller];
349           CallerNode->addCalledFunction(TheMemCpy, MemCpyCGN);
350         }
351 
352         // Uses of the argument in the function should use our new alloca
353         // instead.
354         ActualArg = NewAlloca;
355 
356         // Calls that we inline may use the new alloca, so we need to clear
357         // their 'tail' flags.
358         MustClearTailCallFlags = true;
359       }
360 
361       VMap[I] = ActualArg;
362     }
363 
364     // We want the inliner to prune the code as it copies.  We would LOVE to
365     // have no dead or constant instructions leftover after inlining occurs
366     // (which can happen, e.g., because an argument was constant), but we'll be
367     // happy with whatever the cloner can do.
368     CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
369                               /*ModuleLevelChanges=*/false, Returns, ".i",
370                               &InlinedFunctionInfo, IFI.TD, TheCall);
371 
372     // Remember the first block that is newly cloned over.
373     FirstNewBlock = LastBlock; ++FirstNewBlock;
374 
375     // Update the callgraph if requested.
376     if (IFI.CG)
377       UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
378   }
379 
380   // If there are any alloca instructions in the block that used to be the entry
381   // block for the callee, move them to the entry block of the caller.  First
382   // calculate which instruction they should be inserted before.  We insert the
383   // instructions at the end of the current alloca list.
384   //
385   {
386     BasicBlock::iterator InsertPoint = Caller->begin()->begin();
387     for (BasicBlock::iterator I = FirstNewBlock->begin(),
388          E = FirstNewBlock->end(); I != E; ) {
389       AllocaInst *AI = dyn_cast<AllocaInst>(I++);
390       if (AI == 0) continue;
391 
392       // If the alloca is now dead, remove it.  This often occurs due to code
393       // specialization.
394       if (AI->use_empty()) {
395         AI->eraseFromParent();
396         continue;
397       }
398 
399       if (!isa<Constant>(AI->getArraySize()))
400         continue;
401 
402       // Keep track of the static allocas that we inline into the caller if the
403       // StaticAllocas pointer is non-null.
404       IFI.StaticAllocas.push_back(AI);
405 
406       // Scan for the block of allocas that we can move over, and move them
407       // all at once.
408       while (isa<AllocaInst>(I) &&
409              isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
410         IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
411         ++I;
412       }
413 
414       // Transfer all of the allocas over in a block.  Using splice means
415       // that the instructions aren't removed from the symbol table, then
416       // reinserted.
417       Caller->getEntryBlock().getInstList().splice(InsertPoint,
418                                                    FirstNewBlock->getInstList(),
419                                                    AI, I);
420     }
421   }
422 
423   // If the inlined code contained dynamic alloca instructions, wrap the inlined
424   // code with llvm.stacksave/llvm.stackrestore intrinsics.
425   if (InlinedFunctionInfo.ContainsDynamicAllocas) {
426     Module *M = Caller->getParent();
427     // Get the two intrinsics we care about.
428     Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
429     Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
430 
431     // If we are preserving the callgraph, add edges to the stacksave/restore
432     // functions for the calls we insert.
433     CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0;
434     if (CallGraph *CG = IFI.CG) {
435       StackSaveCGN    = CG->getOrInsertFunction(StackSave);
436       StackRestoreCGN = CG->getOrInsertFunction(StackRestore);
437       CallerNode = (*CG)[Caller];
438     }
439 
440     // Insert the llvm.stacksave.
441     CallInst *SavedPtr = CallInst::Create(StackSave, "savedstack",
442                                           FirstNewBlock->begin());
443     if (IFI.CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN);
444 
445     // Insert a call to llvm.stackrestore before any return instructions in the
446     // inlined function.
447     for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
448       CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", Returns[i]);
449       if (IFI.CG) CallerNode->addCalledFunction(CI, StackRestoreCGN);
450     }
451 
452     // Count the number of StackRestore calls we insert.
453     unsigned NumStackRestores = Returns.size();
454 
455     // If we are inlining an invoke instruction, insert restores before each
456     // unwind.  These unwinds will be rewritten into branches later.
457     if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) {
458       for (Function::iterator BB = FirstNewBlock, E = Caller->end();
459            BB != E; ++BB)
460         if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
461           CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", UI);
462           if (IFI.CG) CallerNode->addCalledFunction(CI, StackRestoreCGN);
463           ++NumStackRestores;
464         }
465     }
466   }
467 
468   // If we are inlining tail call instruction through a call site that isn't
469   // marked 'tail', we must remove the tail marker for any calls in the inlined
470   // code.  Also, calls inlined through a 'nounwind' call site should be marked
471   // 'nounwind'.
472   if (InlinedFunctionInfo.ContainsCalls &&
473       (MustClearTailCallFlags || MarkNoUnwind)) {
474     for (Function::iterator BB = FirstNewBlock, E = Caller->end();
475          BB != E; ++BB)
476       for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
477         if (CallInst *CI = dyn_cast<CallInst>(I)) {
478           if (MustClearTailCallFlags)
479             CI->setTailCall(false);
480           if (MarkNoUnwind)
481             CI->setDoesNotThrow();
482         }
483   }
484 
485   // If we are inlining through a 'nounwind' call site then any inlined 'unwind'
486   // instructions are unreachable.
487   if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind)
488     for (Function::iterator BB = FirstNewBlock, E = Caller->end();
489          BB != E; ++BB) {
490       TerminatorInst *Term = BB->getTerminator();
491       if (isa<UnwindInst>(Term)) {
492         new UnreachableInst(Context, Term);
493         BB->getInstList().erase(Term);
494       }
495     }
496 
497   // If we are inlining for an invoke instruction, we must make sure to rewrite
498   // any inlined 'unwind' instructions into branches to the invoke exception
499   // destination, and call instructions into invoke instructions.
500   if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
501     HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
502 
503   // If we cloned in _exactly one_ basic block, and if that block ends in a
504   // return instruction, we splice the body of the inlined callee directly into
505   // the calling basic block.
506   if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
507     // Move all of the instructions right before the call.
508     OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
509                                  FirstNewBlock->begin(), FirstNewBlock->end());
510     // Remove the cloned basic block.
511     Caller->getBasicBlockList().pop_back();
512 
513     // If the call site was an invoke instruction, add a branch to the normal
514     // destination.
515     if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
516       BranchInst::Create(II->getNormalDest(), TheCall);
517 
518     // If the return instruction returned a value, replace uses of the call with
519     // uses of the returned value.
520     if (!TheCall->use_empty()) {
521       ReturnInst *R = Returns[0];
522       if (TheCall == R->getReturnValue())
523         TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
524       else
525         TheCall->replaceAllUsesWith(R->getReturnValue());
526     }
527     // Since we are now done with the Call/Invoke, we can delete it.
528     TheCall->eraseFromParent();
529 
530     // Since we are now done with the return instruction, delete it also.
531     Returns[0]->eraseFromParent();
532 
533     // We are now done with the inlining.
534     return true;
535   }
536 
537   // Otherwise, we have the normal case, of more than one block to inline or
538   // multiple return sites.
539 
540   // We want to clone the entire callee function into the hole between the
541   // "starter" and "ender" blocks.  How we accomplish this depends on whether
542   // this is an invoke instruction or a call instruction.
543   BasicBlock *AfterCallBB;
544   if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
545 
546     // Add an unconditional branch to make this look like the CallInst case...
547     BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
548 
549     // Split the basic block.  This guarantees that no PHI nodes will have to be
550     // updated due to new incoming edges, and make the invoke case more
551     // symmetric to the call case.
552     AfterCallBB = OrigBB->splitBasicBlock(NewBr,
553                                           CalledFunc->getName()+".exit");
554 
555   } else {  // It's a call
556     // If this is a call instruction, we need to split the basic block that
557     // the call lives in.
558     //
559     AfterCallBB = OrigBB->splitBasicBlock(TheCall,
560                                           CalledFunc->getName()+".exit");
561   }
562 
563   // Change the branch that used to go to AfterCallBB to branch to the first
564   // basic block of the inlined function.
565   //
566   TerminatorInst *Br = OrigBB->getTerminator();
567   assert(Br && Br->getOpcode() == Instruction::Br &&
568          "splitBasicBlock broken!");
569   Br->setOperand(0, FirstNewBlock);
570 
571 
572   // Now that the function is correct, make it a little bit nicer.  In
573   // particular, move the basic blocks inserted from the end of the function
574   // into the space made by splitting the source basic block.
575   Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
576                                      FirstNewBlock, Caller->end());
577 
578   // Handle all of the return instructions that we just cloned in, and eliminate
579   // any users of the original call/invoke instruction.
580   const Type *RTy = CalledFunc->getReturnType();
581 
582   if (Returns.size() > 1) {
583     // The PHI node should go at the front of the new basic block to merge all
584     // possible incoming values.
585     PHINode *PHI = 0;
586     if (!TheCall->use_empty()) {
587       PHI = PHINode::Create(RTy, TheCall->getName(),
588                             AfterCallBB->begin());
589       // Anything that used the result of the function call should now use the
590       // PHI node as their operand.
591       TheCall->replaceAllUsesWith(PHI);
592     }
593 
594     // Loop over all of the return instructions adding entries to the PHI node
595     // as appropriate.
596     if (PHI) {
597       for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
598         ReturnInst *RI = Returns[i];
599         assert(RI->getReturnValue()->getType() == PHI->getType() &&
600                "Ret value not consistent in function!");
601         PHI->addIncoming(RI->getReturnValue(), RI->getParent());
602       }
603 
604       // Now that we inserted the PHI, check to see if it has a single value
605       // (e.g. all the entries are the same or undef).  If so, remove the PHI so
606       // it doesn't block other optimizations.
607       if (Value *V = PHI->hasConstantValue()) {
608         PHI->replaceAllUsesWith(V);
609         PHI->eraseFromParent();
610       }
611     }
612 
613 
614     // Add a branch to the merge points and remove return instructions.
615     for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
616       ReturnInst *RI = Returns[i];
617       BranchInst::Create(AfterCallBB, RI);
618       RI->eraseFromParent();
619     }
620   } else if (!Returns.empty()) {
621     // Otherwise, if there is exactly one return value, just replace anything
622     // using the return value of the call with the computed value.
623     if (!TheCall->use_empty()) {
624       if (TheCall == Returns[0]->getReturnValue())
625         TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
626       else
627         TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
628     }
629 
630     // Splice the code from the return block into the block that it will return
631     // to, which contains the code that was after the call.
632     BasicBlock *ReturnBB = Returns[0]->getParent();
633     AfterCallBB->getInstList().splice(AfterCallBB->begin(),
634                                       ReturnBB->getInstList());
635 
636     // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
637     ReturnBB->replaceAllUsesWith(AfterCallBB);
638 
639     // Delete the return instruction now and empty ReturnBB now.
640     Returns[0]->eraseFromParent();
641     ReturnBB->eraseFromParent();
642   } else if (!TheCall->use_empty()) {
643     // No returns, but something is using the return value of the call.  Just
644     // nuke the result.
645     TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
646   }
647 
648   // Since we are now done with the Call/Invoke, we can delete it.
649   TheCall->eraseFromParent();
650 
651   // We should always be able to fold the entry block of the function into the
652   // single predecessor of the block...
653   assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
654   BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
655 
656   // Splice the code entry block into calling block, right before the
657   // unconditional branch.
658   OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
659   CalleeEntry->replaceAllUsesWith(OrigBB);  // Update PHI nodes
660 
661   // Remove the unconditional branch.
662   OrigBB->getInstList().erase(Br);
663 
664   // Now we can remove the CalleeEntry block, which is now empty.
665   Caller->getBasicBlockList().erase(CalleeEntry);
666 
667   return true;
668 }
669