• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- CodeExtractor.cpp - Pull code region into a new function -----------===//
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 the interface to tear out a code region, such as an
11 // individual loop or a parallel section, into a new function, replacing it with
12 // a call to the new function.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "llvm/Transforms/Utils/FunctionUtils.h"
17 #include "llvm/Constants.h"
18 #include "llvm/DerivedTypes.h"
19 #include "llvm/Instructions.h"
20 #include "llvm/Intrinsics.h"
21 #include "llvm/LLVMContext.h"
22 #include "llvm/Module.h"
23 #include "llvm/Pass.h"
24 #include "llvm/Analysis/Dominators.h"
25 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Analysis/Verifier.h"
27 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/raw_ostream.h"
32 #include "llvm/ADT/SetVector.h"
33 #include "llvm/ADT/StringExtras.h"
34 #include <algorithm>
35 #include <set>
36 using namespace llvm;
37 
38 // Provide a command-line option to aggregate function arguments into a struct
39 // for functions produced by the code extractor. This is useful when converting
40 // extracted functions to pthread-based code, as only one argument (void*) can
41 // be passed in to pthread_create().
42 static cl::opt<bool>
43 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden,
44                  cl::desc("Aggregate arguments to code-extracted functions"));
45 
46 namespace {
47   class CodeExtractor {
48     typedef SetVector<Value*> Values;
49     SetVector<BasicBlock*> BlocksToExtract;
50     DominatorTree* DT;
51     bool AggregateArgs;
52     unsigned NumExitBlocks;
53     Type *RetTy;
54   public:
CodeExtractor(DominatorTree * dt=0,bool AggArgs=false)55     CodeExtractor(DominatorTree* dt = 0, bool AggArgs = false)
56       : DT(dt), AggregateArgs(AggArgs||AggregateArgsOpt), NumExitBlocks(~0U) {}
57 
58     Function *ExtractCodeRegion(ArrayRef<BasicBlock*> code);
59 
60     bool isEligible(ArrayRef<BasicBlock*> code);
61 
62   private:
63     /// definedInRegion - Return true if the specified value is defined in the
64     /// extracted region.
definedInRegion(Value * V) const65     bool definedInRegion(Value *V) const {
66       if (Instruction *I = dyn_cast<Instruction>(V))
67         if (BlocksToExtract.count(I->getParent()))
68           return true;
69       return false;
70     }
71 
72     /// definedInCaller - Return true if the specified value is defined in the
73     /// function being code extracted, but not in the region being extracted.
74     /// These values must be passed in as live-ins to the function.
definedInCaller(Value * V) const75     bool definedInCaller(Value *V) const {
76       if (isa<Argument>(V)) return true;
77       if (Instruction *I = dyn_cast<Instruction>(V))
78         if (!BlocksToExtract.count(I->getParent()))
79           return true;
80       return false;
81     }
82 
83     void severSplitPHINodes(BasicBlock *&Header);
84     void splitReturnBlocks();
85     void findInputsOutputs(Values &inputs, Values &outputs);
86 
87     Function *constructFunction(const Values &inputs,
88                                 const Values &outputs,
89                                 BasicBlock *header,
90                                 BasicBlock *newRootNode, BasicBlock *newHeader,
91                                 Function *oldFunction, Module *M);
92 
93     void moveCodeToFunction(Function *newFunction);
94 
95     void emitCallAndSwitchStatement(Function *newFunction,
96                                     BasicBlock *newHeader,
97                                     Values &inputs,
98                                     Values &outputs);
99 
100   };
101 }
102 
103 /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the
104 /// region, we need to split the entry block of the region so that the PHI node
105 /// is easier to deal with.
severSplitPHINodes(BasicBlock * & Header)106 void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) {
107   unsigned NumPredsFromRegion = 0;
108   unsigned NumPredsOutsideRegion = 0;
109 
110   if (Header != &Header->getParent()->getEntryBlock()) {
111     PHINode *PN = dyn_cast<PHINode>(Header->begin());
112     if (!PN) return;  // No PHI nodes.
113 
114     // If the header node contains any PHI nodes, check to see if there is more
115     // than one entry from outside the region.  If so, we need to sever the
116     // header block into two.
117     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
118       if (BlocksToExtract.count(PN->getIncomingBlock(i)))
119         ++NumPredsFromRegion;
120       else
121         ++NumPredsOutsideRegion;
122 
123     // If there is one (or fewer) predecessor from outside the region, we don't
124     // need to do anything special.
125     if (NumPredsOutsideRegion <= 1) return;
126   }
127 
128   // Otherwise, we need to split the header block into two pieces: one
129   // containing PHI nodes merging values from outside of the region, and a
130   // second that contains all of the code for the block and merges back any
131   // incoming values from inside of the region.
132   BasicBlock::iterator AfterPHIs = Header->getFirstNonPHI();
133   BasicBlock *NewBB = Header->splitBasicBlock(AfterPHIs,
134                                               Header->getName()+".ce");
135 
136   // We only want to code extract the second block now, and it becomes the new
137   // header of the region.
138   BasicBlock *OldPred = Header;
139   BlocksToExtract.remove(OldPred);
140   BlocksToExtract.insert(NewBB);
141   Header = NewBB;
142 
143   // Okay, update dominator sets. The blocks that dominate the new one are the
144   // blocks that dominate TIBB plus the new block itself.
145   if (DT)
146     DT->splitBlock(NewBB);
147 
148   // Okay, now we need to adjust the PHI nodes and any branches from within the
149   // region to go to the new header block instead of the old header block.
150   if (NumPredsFromRegion) {
151     PHINode *PN = cast<PHINode>(OldPred->begin());
152     // Loop over all of the predecessors of OldPred that are in the region,
153     // changing them to branch to NewBB instead.
154     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
155       if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
156         TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator();
157         TI->replaceUsesOfWith(OldPred, NewBB);
158       }
159 
160     // Okay, everything within the region is now branching to the right block, we
161     // just have to update the PHI nodes now, inserting PHI nodes into NewBB.
162     for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) {
163       PHINode *PN = cast<PHINode>(AfterPHIs);
164       // Create a new PHI node in the new region, which has an incoming value
165       // from OldPred of PN.
166       PHINode *NewPN = PHINode::Create(PN->getType(), 1 + NumPredsFromRegion,
167                                        PN->getName()+".ce", NewBB->begin());
168       NewPN->addIncoming(PN, OldPred);
169 
170       // Loop over all of the incoming value in PN, moving them to NewPN if they
171       // are from the extracted region.
172       for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
173         if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
174           NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i));
175           PN->removeIncomingValue(i);
176           --i;
177         }
178       }
179     }
180   }
181 }
182 
splitReturnBlocks()183 void CodeExtractor::splitReturnBlocks() {
184   for (SetVector<BasicBlock*>::iterator I = BlocksToExtract.begin(),
185          E = BlocksToExtract.end(); I != E; ++I)
186     if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator())) {
187       BasicBlock *New = (*I)->splitBasicBlock(RI, (*I)->getName()+".ret");
188       if (DT) {
189         // Old dominates New. New node dominates all other nodes dominated
190         // by Old.
191         DomTreeNode *OldNode = DT->getNode(*I);
192         SmallVector<DomTreeNode*, 8> Children;
193         for (DomTreeNode::iterator DI = OldNode->begin(), DE = OldNode->end();
194              DI != DE; ++DI)
195           Children.push_back(*DI);
196 
197         DomTreeNode *NewNode = DT->addNewBlock(New, *I);
198 
199         for (SmallVector<DomTreeNode*, 8>::iterator I = Children.begin(),
200                E = Children.end(); I != E; ++I)
201           DT->changeImmediateDominator(*I, NewNode);
202       }
203     }
204 }
205 
206 // findInputsOutputs - Find inputs to, outputs from the code region.
207 //
findInputsOutputs(Values & inputs,Values & outputs)208 void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs) {
209   std::set<BasicBlock*> ExitBlocks;
210   for (SetVector<BasicBlock*>::const_iterator ci = BlocksToExtract.begin(),
211        ce = BlocksToExtract.end(); ci != ce; ++ci) {
212     BasicBlock *BB = *ci;
213 
214     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
215       // If a used value is defined outside the region, it's an input.  If an
216       // instruction is used outside the region, it's an output.
217       for (User::op_iterator O = I->op_begin(), E = I->op_end(); O != E; ++O)
218         if (definedInCaller(*O))
219           inputs.insert(*O);
220 
221       // Consider uses of this instruction (outputs).
222       for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
223            UI != E; ++UI)
224         if (!definedInRegion(*UI)) {
225           outputs.insert(I);
226           break;
227         }
228     } // for: insts
229 
230     // Keep track of the exit blocks from the region.
231     TerminatorInst *TI = BB->getTerminator();
232     for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
233       if (!BlocksToExtract.count(TI->getSuccessor(i)))
234         ExitBlocks.insert(TI->getSuccessor(i));
235   } // for: basic blocks
236 
237   NumExitBlocks = ExitBlocks.size();
238 }
239 
240 /// constructFunction - make a function based on inputs and outputs, as follows:
241 /// f(in0, ..., inN, out0, ..., outN)
242 ///
constructFunction(const Values & inputs,const Values & outputs,BasicBlock * header,BasicBlock * newRootNode,BasicBlock * newHeader,Function * oldFunction,Module * M)243 Function *CodeExtractor::constructFunction(const Values &inputs,
244                                            const Values &outputs,
245                                            BasicBlock *header,
246                                            BasicBlock *newRootNode,
247                                            BasicBlock *newHeader,
248                                            Function *oldFunction,
249                                            Module *M) {
250   DEBUG(dbgs() << "inputs: " << inputs.size() << "\n");
251   DEBUG(dbgs() << "outputs: " << outputs.size() << "\n");
252 
253   // This function returns unsigned, outputs will go back by reference.
254   switch (NumExitBlocks) {
255   case 0:
256   case 1: RetTy = Type::getVoidTy(header->getContext()); break;
257   case 2: RetTy = Type::getInt1Ty(header->getContext()); break;
258   default: RetTy = Type::getInt16Ty(header->getContext()); break;
259   }
260 
261   std::vector<Type*> paramTy;
262 
263   // Add the types of the input values to the function's argument list
264   for (Values::const_iterator i = inputs.begin(),
265          e = inputs.end(); i != e; ++i) {
266     const Value *value = *i;
267     DEBUG(dbgs() << "value used in func: " << *value << "\n");
268     paramTy.push_back(value->getType());
269   }
270 
271   // Add the types of the output values to the function's argument list.
272   for (Values::const_iterator I = outputs.begin(), E = outputs.end();
273        I != E; ++I) {
274     DEBUG(dbgs() << "instr used in func: " << **I << "\n");
275     if (AggregateArgs)
276       paramTy.push_back((*I)->getType());
277     else
278       paramTy.push_back(PointerType::getUnqual((*I)->getType()));
279   }
280 
281   DEBUG(dbgs() << "Function type: " << *RetTy << " f(");
282   for (std::vector<Type*>::iterator i = paramTy.begin(),
283          e = paramTy.end(); i != e; ++i)
284     DEBUG(dbgs() << **i << ", ");
285   DEBUG(dbgs() << ")\n");
286 
287   if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
288     PointerType *StructPtr =
289            PointerType::getUnqual(StructType::get(M->getContext(), paramTy));
290     paramTy.clear();
291     paramTy.push_back(StructPtr);
292   }
293   FunctionType *funcType =
294                   FunctionType::get(RetTy, paramTy, false);
295 
296   // Create the new function
297   Function *newFunction = Function::Create(funcType,
298                                            GlobalValue::InternalLinkage,
299                                            oldFunction->getName() + "_" +
300                                            header->getName(), M);
301   // If the old function is no-throw, so is the new one.
302   if (oldFunction->doesNotThrow())
303     newFunction->setDoesNotThrow(true);
304 
305   newFunction->getBasicBlockList().push_back(newRootNode);
306 
307   // Create an iterator to name all of the arguments we inserted.
308   Function::arg_iterator AI = newFunction->arg_begin();
309 
310   // Rewrite all users of the inputs in the extracted region to use the
311   // arguments (or appropriate addressing into struct) instead.
312   for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
313     Value *RewriteVal;
314     if (AggregateArgs) {
315       Value *Idx[2];
316       Idx[0] = Constant::getNullValue(Type::getInt32Ty(header->getContext()));
317       Idx[1] = ConstantInt::get(Type::getInt32Ty(header->getContext()), i);
318       TerminatorInst *TI = newFunction->begin()->getTerminator();
319       GetElementPtrInst *GEP =
320         GetElementPtrInst::Create(AI, Idx, "gep_" + inputs[i]->getName(), TI);
321       RewriteVal = new LoadInst(GEP, "loadgep_" + inputs[i]->getName(), TI);
322     } else
323       RewriteVal = AI++;
324 
325     std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
326     for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
327          use != useE; ++use)
328       if (Instruction* inst = dyn_cast<Instruction>(*use))
329         if (BlocksToExtract.count(inst->getParent()))
330           inst->replaceUsesOfWith(inputs[i], RewriteVal);
331   }
332 
333   // Set names for input and output arguments.
334   if (!AggregateArgs) {
335     AI = newFunction->arg_begin();
336     for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
337       AI->setName(inputs[i]->getName());
338     for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
339       AI->setName(outputs[i]->getName()+".out");
340   }
341 
342   // Rewrite branches to basic blocks outside of the loop to new dummy blocks
343   // within the new function. This must be done before we lose track of which
344   // blocks were originally in the code region.
345   std::vector<User*> Users(header->use_begin(), header->use_end());
346   for (unsigned i = 0, e = Users.size(); i != e; ++i)
347     // The BasicBlock which contains the branch is not in the region
348     // modify the branch target to a new block
349     if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
350       if (!BlocksToExtract.count(TI->getParent()) &&
351           TI->getParent()->getParent() == oldFunction)
352         TI->replaceUsesOfWith(header, newHeader);
353 
354   return newFunction;
355 }
356 
357 /// FindPhiPredForUseInBlock - Given a value and a basic block, find a PHI
358 /// that uses the value within the basic block, and return the predecessor
359 /// block associated with that use, or return 0 if none is found.
FindPhiPredForUseInBlock(Value * Used,BasicBlock * BB)360 static BasicBlock* FindPhiPredForUseInBlock(Value* Used, BasicBlock* BB) {
361   for (Value::use_iterator UI = Used->use_begin(),
362        UE = Used->use_end(); UI != UE; ++UI) {
363      PHINode *P = dyn_cast<PHINode>(*UI);
364      if (P && P->getParent() == BB)
365        return P->getIncomingBlock(UI);
366   }
367 
368   return 0;
369 }
370 
371 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
372 /// the call instruction, splitting any PHI nodes in the header block as
373 /// necessary.
374 void CodeExtractor::
emitCallAndSwitchStatement(Function * newFunction,BasicBlock * codeReplacer,Values & inputs,Values & outputs)375 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
376                            Values &inputs, Values &outputs) {
377   // Emit a call to the new function, passing in: *pointer to struct (if
378   // aggregating parameters), or plan inputs and allocated memory for outputs
379   std::vector<Value*> params, StructValues, ReloadOutputs, Reloads;
380 
381   LLVMContext &Context = newFunction->getContext();
382 
383   // Add inputs as params, or to be filled into the struct
384   for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
385     if (AggregateArgs)
386       StructValues.push_back(*i);
387     else
388       params.push_back(*i);
389 
390   // Create allocas for the outputs
391   for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
392     if (AggregateArgs) {
393       StructValues.push_back(*i);
394     } else {
395       AllocaInst *alloca =
396         new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
397                        codeReplacer->getParent()->begin()->begin());
398       ReloadOutputs.push_back(alloca);
399       params.push_back(alloca);
400     }
401   }
402 
403   AllocaInst *Struct = 0;
404   if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
405     std::vector<Type*> ArgTypes;
406     for (Values::iterator v = StructValues.begin(),
407            ve = StructValues.end(); v != ve; ++v)
408       ArgTypes.push_back((*v)->getType());
409 
410     // Allocate a struct at the beginning of this function
411     Type *StructArgTy = StructType::get(newFunction->getContext(), ArgTypes);
412     Struct =
413       new AllocaInst(StructArgTy, 0, "structArg",
414                      codeReplacer->getParent()->begin()->begin());
415     params.push_back(Struct);
416 
417     for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
418       Value *Idx[2];
419       Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
420       Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), i);
421       GetElementPtrInst *GEP =
422         GetElementPtrInst::Create(Struct, Idx,
423                                   "gep_" + StructValues[i]->getName());
424       codeReplacer->getInstList().push_back(GEP);
425       StoreInst *SI = new StoreInst(StructValues[i], GEP);
426       codeReplacer->getInstList().push_back(SI);
427     }
428   }
429 
430   // Emit the call to the function
431   CallInst *call = CallInst::Create(newFunction, params,
432                                     NumExitBlocks > 1 ? "targetBlock" : "");
433   codeReplacer->getInstList().push_back(call);
434 
435   Function::arg_iterator OutputArgBegin = newFunction->arg_begin();
436   unsigned FirstOut = inputs.size();
437   if (!AggregateArgs)
438     std::advance(OutputArgBegin, inputs.size());
439 
440   // Reload the outputs passed in by reference
441   for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
442     Value *Output = 0;
443     if (AggregateArgs) {
444       Value *Idx[2];
445       Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
446       Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), FirstOut + i);
447       GetElementPtrInst *GEP
448         = GetElementPtrInst::Create(Struct, Idx,
449                                     "gep_reload_" + outputs[i]->getName());
450       codeReplacer->getInstList().push_back(GEP);
451       Output = GEP;
452     } else {
453       Output = ReloadOutputs[i];
454     }
455     LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
456     Reloads.push_back(load);
457     codeReplacer->getInstList().push_back(load);
458     std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
459     for (unsigned u = 0, e = Users.size(); u != e; ++u) {
460       Instruction *inst = cast<Instruction>(Users[u]);
461       if (!BlocksToExtract.count(inst->getParent()))
462         inst->replaceUsesOfWith(outputs[i], load);
463     }
464   }
465 
466   // Now we can emit a switch statement using the call as a value.
467   SwitchInst *TheSwitch =
468       SwitchInst::Create(Constant::getNullValue(Type::getInt16Ty(Context)),
469                          codeReplacer, 0, codeReplacer);
470 
471   // Since there may be multiple exits from the original region, make the new
472   // function return an unsigned, switch on that number.  This loop iterates
473   // over all of the blocks in the extracted region, updating any terminator
474   // instructions in the to-be-extracted region that branch to blocks that are
475   // not in the region to be extracted.
476   std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
477 
478   unsigned switchVal = 0;
479   for (SetVector<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
480          e = BlocksToExtract.end(); i != e; ++i) {
481     TerminatorInst *TI = (*i)->getTerminator();
482     for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
483       if (!BlocksToExtract.count(TI->getSuccessor(i))) {
484         BasicBlock *OldTarget = TI->getSuccessor(i);
485         // add a new basic block which returns the appropriate value
486         BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
487         if (!NewTarget) {
488           // If we don't already have an exit stub for this non-extracted
489           // destination, create one now!
490           NewTarget = BasicBlock::Create(Context,
491                                          OldTarget->getName() + ".exitStub",
492                                          newFunction);
493           unsigned SuccNum = switchVal++;
494 
495           Value *brVal = 0;
496           switch (NumExitBlocks) {
497           case 0:
498           case 1: break;  // No value needed.
499           case 2:         // Conditional branch, return a bool
500             brVal = ConstantInt::get(Type::getInt1Ty(Context), !SuccNum);
501             break;
502           default:
503             brVal = ConstantInt::get(Type::getInt16Ty(Context), SuccNum);
504             break;
505           }
506 
507           ReturnInst *NTRet = ReturnInst::Create(Context, brVal, NewTarget);
508 
509           // Update the switch instruction.
510           TheSwitch->addCase(ConstantInt::get(Type::getInt16Ty(Context),
511                                               SuccNum),
512                              OldTarget);
513 
514           // Restore values just before we exit
515           Function::arg_iterator OAI = OutputArgBegin;
516           for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
517             // For an invoke, the normal destination is the only one that is
518             // dominated by the result of the invocation
519             BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
520 
521             bool DominatesDef = true;
522 
523             if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) {
524               DefBlock = Invoke->getNormalDest();
525 
526               // Make sure we are looking at the original successor block, not
527               // at a newly inserted exit block, which won't be in the dominator
528               // info.
529               for (std::map<BasicBlock*, BasicBlock*>::iterator I =
530                      ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I)
531                 if (DefBlock == I->second) {
532                   DefBlock = I->first;
533                   break;
534                 }
535 
536               // In the extract block case, if the block we are extracting ends
537               // with an invoke instruction, make sure that we don't emit a
538               // store of the invoke value for the unwind block.
539               if (!DT && DefBlock != OldTarget)
540                 DominatesDef = false;
541             }
542 
543             if (DT) {
544               DominatesDef = DT->dominates(DefBlock, OldTarget);
545 
546               // If the output value is used by a phi in the target block,
547               // then we need to test for dominance of the phi's predecessor
548               // instead.  Unfortunately, this a little complicated since we
549               // have already rewritten uses of the value to uses of the reload.
550               BasicBlock* pred = FindPhiPredForUseInBlock(Reloads[out],
551                                                           OldTarget);
552               if (pred && DT && DT->dominates(DefBlock, pred))
553                 DominatesDef = true;
554             }
555 
556             if (DominatesDef) {
557               if (AggregateArgs) {
558                 Value *Idx[2];
559                 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context));
560                 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context),
561                                           FirstOut+out);
562                 GetElementPtrInst *GEP =
563                   GetElementPtrInst::Create(OAI, Idx,
564                                             "gep_" + outputs[out]->getName(),
565                                             NTRet);
566                 new StoreInst(outputs[out], GEP, NTRet);
567               } else {
568                 new StoreInst(outputs[out], OAI, NTRet);
569               }
570             }
571             // Advance output iterator even if we don't emit a store
572             if (!AggregateArgs) ++OAI;
573           }
574         }
575 
576         // rewrite the original branch instruction with this new target
577         TI->setSuccessor(i, NewTarget);
578       }
579   }
580 
581   // Now that we've done the deed, simplify the switch instruction.
582   Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
583   switch (NumExitBlocks) {
584   case 0:
585     // There are no successors (the block containing the switch itself), which
586     // means that previously this was the last part of the function, and hence
587     // this should be rewritten as a `ret'
588 
589     // Check if the function should return a value
590     if (OldFnRetTy->isVoidTy()) {
591       ReturnInst::Create(Context, 0, TheSwitch);  // Return void
592     } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
593       // return what we have
594       ReturnInst::Create(Context, TheSwitch->getCondition(), TheSwitch);
595     } else {
596       // Otherwise we must have code extracted an unwind or something, just
597       // return whatever we want.
598       ReturnInst::Create(Context,
599                          Constant::getNullValue(OldFnRetTy), TheSwitch);
600     }
601 
602     TheSwitch->eraseFromParent();
603     break;
604   case 1:
605     // Only a single destination, change the switch into an unconditional
606     // branch.
607     BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch);
608     TheSwitch->eraseFromParent();
609     break;
610   case 2:
611     BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
612                        call, TheSwitch);
613     TheSwitch->eraseFromParent();
614     break;
615   default:
616     // Otherwise, make the default destination of the switch instruction be one
617     // of the other successors.
618     TheSwitch->setCondition(call);
619     TheSwitch->setDefaultDest(TheSwitch->getSuccessor(NumExitBlocks));
620     // Remove redundant case
621     TheSwitch->removeCase(SwitchInst::CaseIt(TheSwitch, NumExitBlocks-1));
622     break;
623   }
624 }
625 
moveCodeToFunction(Function * newFunction)626 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
627   Function *oldFunc = (*BlocksToExtract.begin())->getParent();
628   Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
629   Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
630 
631   for (SetVector<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
632          e = BlocksToExtract.end(); i != e; ++i) {
633     // Delete the basic block from the old function, and the list of blocks
634     oldBlocks.remove(*i);
635 
636     // Insert this basic block into the new function
637     newBlocks.push_back(*i);
638   }
639 }
640 
641 /// ExtractRegion - Removes a loop from a function, replaces it with a call to
642 /// new function. Returns pointer to the new function.
643 ///
644 /// algorithm:
645 ///
646 /// find inputs and outputs for the region
647 ///
648 /// for inputs: add to function as args, map input instr* to arg#
649 /// for outputs: add allocas for scalars,
650 ///             add to func as args, map output instr* to arg#
651 ///
652 /// rewrite func to use argument #s instead of instr*
653 ///
654 /// for each scalar output in the function: at every exit, store intermediate
655 /// computed result back into memory.
656 ///
657 Function *CodeExtractor::
ExtractCodeRegion(ArrayRef<BasicBlock * > code)658 ExtractCodeRegion(ArrayRef<BasicBlock*> code) {
659   if (!isEligible(code))
660     return 0;
661 
662   // 1) Find inputs, outputs
663   // 2) Construct new function
664   //  * Add allocas for defs, pass as args by reference
665   //  * Pass in uses as args
666   // 3) Move code region, add call instr to func
667   //
668   BlocksToExtract.insert(code.begin(), code.end());
669 
670   Values inputs, outputs;
671 
672   // Assumption: this is a single-entry code region, and the header is the first
673   // block in the region.
674   BasicBlock *header = code[0];
675 
676   for (unsigned i = 1, e = code.size(); i != e; ++i)
677     for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]);
678          PI != E; ++PI)
679       assert(BlocksToExtract.count(*PI) &&
680              "No blocks in this region may have entries from outside the region"
681              " except for the first block!");
682 
683   // If we have to split PHI nodes or the entry block, do so now.
684   severSplitPHINodes(header);
685 
686   // If we have any return instructions in the region, split those blocks so
687   // that the return is not in the region.
688   splitReturnBlocks();
689 
690   Function *oldFunction = header->getParent();
691 
692   // This takes place of the original loop
693   BasicBlock *codeReplacer = BasicBlock::Create(header->getContext(),
694                                                 "codeRepl", oldFunction,
695                                                 header);
696 
697   // The new function needs a root node because other nodes can branch to the
698   // head of the region, but the entry node of a function cannot have preds.
699   BasicBlock *newFuncRoot = BasicBlock::Create(header->getContext(),
700                                                "newFuncRoot");
701   newFuncRoot->getInstList().push_back(BranchInst::Create(header));
702 
703   // Find inputs to, outputs from the code region.
704   findInputsOutputs(inputs, outputs);
705 
706   // Construct new function based on inputs/outputs & add allocas for all defs.
707   Function *newFunction = constructFunction(inputs, outputs, header,
708                                             newFuncRoot,
709                                             codeReplacer, oldFunction,
710                                             oldFunction->getParent());
711 
712   emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
713 
714   moveCodeToFunction(newFunction);
715 
716   // Loop over all of the PHI nodes in the header block, and change any
717   // references to the old incoming edge to be the new incoming edge.
718   for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
719     PHINode *PN = cast<PHINode>(I);
720     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
721       if (!BlocksToExtract.count(PN->getIncomingBlock(i)))
722         PN->setIncomingBlock(i, newFuncRoot);
723   }
724 
725   // Look at all successors of the codeReplacer block.  If any of these blocks
726   // had PHI nodes in them, we need to update the "from" block to be the code
727   // replacer, not the original block in the extracted region.
728   std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
729                                  succ_end(codeReplacer));
730   for (unsigned i = 0, e = Succs.size(); i != e; ++i)
731     for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) {
732       PHINode *PN = cast<PHINode>(I);
733       std::set<BasicBlock*> ProcessedPreds;
734       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
735         if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
736           if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
737             PN->setIncomingBlock(i, codeReplacer);
738           else {
739             // There were multiple entries in the PHI for this block, now there
740             // is only one, so remove the duplicated entries.
741             PN->removeIncomingValue(i, false);
742             --i; --e;
743           }
744         }
745     }
746 
747   //cerr << "NEW FUNCTION: " << *newFunction;
748   //  verifyFunction(*newFunction);
749 
750   //  cerr << "OLD FUNCTION: " << *oldFunction;
751   //  verifyFunction(*oldFunction);
752 
753   DEBUG(if (verifyFunction(*newFunction))
754         report_fatal_error("verifyFunction failed!"));
755   return newFunction;
756 }
757 
isEligible(ArrayRef<BasicBlock * > code)758 bool CodeExtractor::isEligible(ArrayRef<BasicBlock*> code) {
759   // Deny a single basic block that's a landing pad block.
760   if (code.size() == 1 && code[0]->isLandingPad())
761     return false;
762 
763   // Deny code region if it contains allocas or vastarts.
764   for (ArrayRef<BasicBlock*>::iterator BB = code.begin(), e=code.end();
765        BB != e; ++BB)
766     for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end();
767          I != Ie; ++I)
768       if (isa<AllocaInst>(*I))
769         return false;
770       else if (const CallInst *CI = dyn_cast<CallInst>(I))
771         if (const Function *F = CI->getCalledFunction())
772           if (F->getIntrinsicID() == Intrinsic::vastart)
773             return false;
774   return true;
775 }
776 
777 
778 /// ExtractCodeRegion - Slurp a sequence of basic blocks into a brand new
779 /// function.
780 ///
ExtractCodeRegion(DominatorTree & DT,ArrayRef<BasicBlock * > code,bool AggregateArgs)781 Function* llvm::ExtractCodeRegion(DominatorTree &DT,
782                                   ArrayRef<BasicBlock*> code,
783                                   bool AggregateArgs) {
784   return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(code);
785 }
786 
787 /// ExtractLoop - Slurp a natural loop into a brand new function.
788 ///
ExtractLoop(DominatorTree & DT,Loop * L,bool AggregateArgs)789 Function* llvm::ExtractLoop(DominatorTree &DT, Loop *L, bool AggregateArgs) {
790   return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(L->getBlocks());
791 }
792 
793 /// ExtractBasicBlock - Slurp a basic block into a brand new function.
794 ///
ExtractBasicBlock(ArrayRef<BasicBlock * > BBs,bool AggregateArgs)795 Function* llvm::ExtractBasicBlock(ArrayRef<BasicBlock*> BBs, bool AggregateArgs){
796   return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(BBs);
797 }
798