LLVM API Documentation

Main Page | Namespace List | Class Hierarchy | Alphabetical List | Class List | Directories | File List | Namespace Members | Class Members | File Members | Related Pages

CodeExtractor.cpp

Go to the documentation of this file.
00001 //===- CodeExtractor.cpp - Pull code region into a new function -----------===//
00002 // 
00003 //                     The LLVM Compiler Infrastructure
00004 //
00005 // This file was developed by the LLVM research group and is distributed under
00006 // the University of Illinois Open Source License. See LICENSE.TXT for details.
00007 // 
00008 //===----------------------------------------------------------------------===//
00009 //
00010 // This file implements the interface to tear out a code region, such as an
00011 // individual loop or a parallel section, into a new function, replacing it with
00012 // a call to the new function.
00013 //
00014 //===----------------------------------------------------------------------===//
00015 
00016 #include "llvm/Transforms/Utils/FunctionUtils.h"
00017 #include "llvm/Constants.h"
00018 #include "llvm/DerivedTypes.h"
00019 #include "llvm/Instructions.h"
00020 #include "llvm/Intrinsics.h"
00021 #include "llvm/Module.h"
00022 #include "llvm/Pass.h"
00023 #include "llvm/Analysis/Dominators.h"
00024 #include "llvm/Analysis/LoopInfo.h"
00025 #include "llvm/Analysis/Verifier.h"
00026 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
00027 #include "llvm/Support/CommandLine.h"
00028 #include "llvm/Support/Debug.h"
00029 #include "llvm/ADT/StringExtras.h"
00030 #include <algorithm>
00031 #include <set>
00032 using namespace llvm;
00033 
00034 // Provide a command-line option to aggregate function arguments into a struct
00035 // for functions produced by the code extrator. This is useful when converting
00036 // extracted functions to pthread-based code, as only one argument (void*) can
00037 // be passed in to pthread_create().
00038 static cl::opt<bool>
00039 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden,
00040                  cl::desc("Aggregate arguments to code-extracted functions"));
00041 
00042 namespace {
00043   class CodeExtractor {
00044     typedef std::vector<Value*> Values;
00045     std::set<BasicBlock*> BlocksToExtract;
00046     DominatorSet *DS;
00047     bool AggregateArgs;
00048     unsigned NumExitBlocks;
00049     const Type *RetTy;
00050   public:
00051     CodeExtractor(DominatorSet *ds = 0, bool AggArgs = false)
00052       : DS(ds), AggregateArgs(AggArgs||AggregateArgsOpt), NumExitBlocks(~0U) {}
00053 
00054     Function *ExtractCodeRegion(const std::vector<BasicBlock*> &code);
00055 
00056     bool isEligible(const std::vector<BasicBlock*> &code);
00057 
00058   private:
00059     /// definedInRegion - Return true if the specified value is defined in the
00060     /// extracted region.
00061     bool definedInRegion(Value *V) const {
00062       if (Instruction *I = dyn_cast<Instruction>(V))
00063         if (BlocksToExtract.count(I->getParent()))
00064           return true;
00065       return false;
00066     }
00067     
00068     /// definedInCaller - Return true if the specified value is defined in the
00069     /// function being code extracted, but not in the region being extracted.
00070     /// These values must be passed in as live-ins to the function.
00071     bool definedInCaller(Value *V) const {
00072       if (isa<Argument>(V)) return true;
00073       if (Instruction *I = dyn_cast<Instruction>(V))
00074         if (!BlocksToExtract.count(I->getParent()))
00075           return true;
00076       return false;
00077     }
00078 
00079     void severSplitPHINodes(BasicBlock *&Header);
00080     void splitReturnBlocks();
00081     void findInputsOutputs(Values &inputs, Values &outputs);
00082 
00083     Function *constructFunction(const Values &inputs,
00084                                 const Values &outputs,
00085                                 BasicBlock *header,
00086                                 BasicBlock *newRootNode, BasicBlock *newHeader,
00087                                 Function *oldFunction, Module *M);
00088 
00089     void moveCodeToFunction(Function *newFunction);
00090 
00091     void emitCallAndSwitchStatement(Function *newFunction,
00092                                     BasicBlock *newHeader,
00093                                     Values &inputs,
00094                                     Values &outputs);
00095 
00096   };
00097 }
00098 
00099 /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the
00100 /// region, we need to split the entry block of the region so that the PHI node
00101 /// is easier to deal with.
00102 void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) {
00103   bool HasPredsFromRegion = false;
00104   unsigned NumPredsOutsideRegion = 0;
00105 
00106   if (Header != &Header->getParent()->front()) {
00107     PHINode *PN = dyn_cast<PHINode>(Header->begin());
00108     if (!PN) return;  // No PHI nodes.
00109 
00110     // If the header node contains any PHI nodes, check to see if there is more
00111     // than one entry from outside the region.  If so, we need to sever the
00112     // header block into two.
00113     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
00114       if (BlocksToExtract.count(PN->getIncomingBlock(i)))
00115         HasPredsFromRegion = true;
00116       else
00117         ++NumPredsOutsideRegion;
00118 
00119     // If there is one (or fewer) predecessor from outside the region, we don't
00120     // need to do anything special.
00121     if (NumPredsOutsideRegion <= 1) return;
00122   }
00123 
00124   // Otherwise, we need to split the header block into two pieces: one
00125   // containing PHI nodes merging values from outside of the region, and a
00126   // second that contains all of the code for the block and merges back any
00127   // incoming values from inside of the region.
00128   BasicBlock::iterator AfterPHIs = Header->begin();
00129   while (isa<PHINode>(AfterPHIs)) ++AfterPHIs;
00130   BasicBlock *NewBB = Header->splitBasicBlock(AfterPHIs,
00131                                               Header->getName()+".ce");
00132 
00133   // We only want to code extract the second block now, and it becomes the new
00134   // header of the region.
00135   BasicBlock *OldPred = Header;
00136   BlocksToExtract.erase(OldPred);
00137   BlocksToExtract.insert(NewBB);
00138   Header = NewBB;
00139 
00140   // Okay, update dominator sets. The blocks that dominate the new one are the
00141   // blocks that dominate TIBB plus the new block itself.
00142   if (DS) {
00143     DominatorSet::DomSetType DomSet = DS->getDominators(OldPred);
00144     DomSet.insert(NewBB);  // A block always dominates itself.
00145     DS->addBasicBlock(NewBB, DomSet);
00146 
00147     // Additionally, NewBB dominates all blocks in the function that are
00148     // dominated by OldPred.
00149     Function *F = Header->getParent();
00150     for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
00151       if (DS->properlyDominates(OldPred, I))
00152         DS->addDominator(I, NewBB);
00153   }
00154 
00155   // Okay, now we need to adjust the PHI nodes and any branches from within the
00156   // region to go to the new header block instead of the old header block.
00157   if (HasPredsFromRegion) {
00158     PHINode *PN = cast<PHINode>(OldPred->begin());
00159     // Loop over all of the predecessors of OldPred that are in the region,
00160     // changing them to branch to NewBB instead.
00161     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
00162       if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
00163         TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator();
00164         TI->replaceUsesOfWith(OldPred, NewBB);
00165       }
00166 
00167     // Okay, everthing within the region is now branching to the right block, we
00168     // just have to update the PHI nodes now, inserting PHI nodes into NewBB.
00169     for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) {
00170       PHINode *PN = cast<PHINode>(AfterPHIs);
00171       // Create a new PHI node in the new region, which has an incoming value
00172       // from OldPred of PN.
00173       PHINode *NewPN = new PHINode(PN->getType(), PN->getName()+".ce",
00174                                    NewBB->begin());
00175       NewPN->addIncoming(PN, OldPred);
00176 
00177       // Loop over all of the incoming value in PN, moving them to NewPN if they
00178       // are from the extracted region.
00179       for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
00180         if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
00181           NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i));
00182           PN->removeIncomingValue(i);
00183           --i;
00184         }
00185       }
00186     }
00187   }
00188 }
00189 
00190 void CodeExtractor::splitReturnBlocks() {
00191   for (std::set<BasicBlock*>::iterator I = BlocksToExtract.begin(),
00192          E = BlocksToExtract.end(); I != E; ++I)
00193     if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator()))
00194       (*I)->splitBasicBlock(RI, (*I)->getName()+".ret");
00195 }
00196 
00197 // findInputsOutputs - Find inputs to, outputs from the code region.
00198 //
00199 void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs) {
00200   std::set<BasicBlock*> ExitBlocks;
00201   for (std::set<BasicBlock*>::const_iterator ci = BlocksToExtract.begin(), 
00202        ce = BlocksToExtract.end(); ci != ce; ++ci) {
00203     BasicBlock *BB = *ci;
00204 
00205     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
00206       // If a used value is defined outside the region, it's an input.  If an
00207       // instruction is used outside the region, it's an output.
00208       for (User::op_iterator O = I->op_begin(), E = I->op_end(); O != E; ++O)
00209         if (definedInCaller(*O))
00210           inputs.push_back(*O);
00211       
00212       // Consider uses of this instruction (outputs).
00213       for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
00214            UI != E; ++UI)
00215         if (!definedInRegion(*UI)) {
00216           outputs.push_back(I);
00217           break;
00218         }
00219     } // for: insts
00220 
00221     // Keep track of the exit blocks from the region.
00222     TerminatorInst *TI = BB->getTerminator();
00223     for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
00224       if (!BlocksToExtract.count(TI->getSuccessor(i)))
00225         ExitBlocks.insert(TI->getSuccessor(i));
00226   } // for: basic blocks
00227 
00228   NumExitBlocks = ExitBlocks.size();
00229 
00230   // Eliminate duplicates.
00231   std::sort(inputs.begin(), inputs.end());
00232   inputs.erase(std::unique(inputs.begin(), inputs.end()), inputs.end());
00233   std::sort(outputs.begin(), outputs.end());
00234   outputs.erase(std::unique(outputs.begin(), outputs.end()), outputs.end());
00235 }
00236 
00237 /// constructFunction - make a function based on inputs and outputs, as follows:
00238 /// f(in0, ..., inN, out0, ..., outN)
00239 ///
00240 Function *CodeExtractor::constructFunction(const Values &inputs,
00241                                            const Values &outputs,
00242                                            BasicBlock *header,
00243                                            BasicBlock *newRootNode,
00244                                            BasicBlock *newHeader,
00245                                            Function *oldFunction,
00246                                            Module *M) {
00247   DEBUG(std::cerr << "inputs: " << inputs.size() << "\n");
00248   DEBUG(std::cerr << "outputs: " << outputs.size() << "\n");
00249 
00250   // This function returns unsigned, outputs will go back by reference.
00251   switch (NumExitBlocks) {
00252   case 0:
00253   case 1: RetTy = Type::VoidTy; break;
00254   case 2: RetTy = Type::BoolTy; break;
00255   default: RetTy = Type::UShortTy; break;
00256   }
00257 
00258   std::vector<const Type*> paramTy;
00259 
00260   // Add the types of the input values to the function's argument list
00261   for (Values::const_iterator i = inputs.begin(),
00262          e = inputs.end(); i != e; ++i) {
00263     const Value *value = *i;
00264     DEBUG(std::cerr << "value used in func: " << *value << "\n");
00265     paramTy.push_back(value->getType());
00266   }
00267 
00268   // Add the types of the output values to the function's argument list.
00269   for (Values::const_iterator I = outputs.begin(), E = outputs.end();
00270        I != E; ++I) {
00271     DEBUG(std::cerr << "instr used in func: " << **I << "\n");
00272     if (AggregateArgs)
00273       paramTy.push_back((*I)->getType());
00274     else
00275       paramTy.push_back(PointerType::get((*I)->getType()));
00276   }
00277 
00278   DEBUG(std::cerr << "Function type: " << *RetTy << " f(");
00279   DEBUG(for (std::vector<const Type*>::iterator i = paramTy.begin(),
00280                e = paramTy.end(); i != e; ++i) std::cerr << **i << ", ");
00281   DEBUG(std::cerr << ")\n");
00282 
00283   if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
00284     PointerType *StructPtr = PointerType::get(StructType::get(paramTy));
00285     paramTy.clear();
00286     paramTy.push_back(StructPtr);
00287   }
00288   const FunctionType *funcType = FunctionType::get(RetTy, paramTy, false);
00289 
00290   // Create the new function
00291   Function *newFunction = new Function(funcType,
00292                                        GlobalValue::InternalLinkage,
00293                                        oldFunction->getName() + "_" +
00294                                        header->getName(), M);
00295   newFunction->getBasicBlockList().push_back(newRootNode);
00296 
00297   // Create an iterator to name all of the arguments we inserted.
00298   Function::aiterator AI = newFunction->abegin();
00299 
00300   // Rewrite all users of the inputs in the extracted region to use the
00301   // arguments (or appropriate addressing into struct) instead.
00302   for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
00303     Value *RewriteVal;
00304     if (AggregateArgs) {
00305       std::vector<Value*> Indices;
00306       Indices.push_back(Constant::getNullValue(Type::UIntTy));
00307       Indices.push_back(ConstantUInt::get(Type::UIntTy, i));
00308       std::string GEPname = "gep_" + inputs[i]->getName();
00309       TerminatorInst *TI = newFunction->begin()->getTerminator();
00310       GetElementPtrInst *GEP = new GetElementPtrInst(AI, Indices, GEPname, TI);
00311       RewriteVal = new LoadInst(GEP, "load" + GEPname, TI);
00312     } else
00313       RewriteVal = AI++;
00314 
00315     std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
00316     for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
00317          use != useE; ++use)
00318       if (Instruction* inst = dyn_cast<Instruction>(*use))
00319         if (BlocksToExtract.count(inst->getParent()))
00320           inst->replaceUsesOfWith(inputs[i], RewriteVal);
00321   }
00322 
00323   // Set names for input and output arguments.
00324   if (!AggregateArgs) {
00325     AI = newFunction->abegin();
00326     for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
00327       AI->setName(inputs[i]->getName());
00328     for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
00329       AI->setName(outputs[i]->getName()+".out");  
00330   }
00331 
00332   // Rewrite branches to basic blocks outside of the loop to new dummy blocks
00333   // within the new function. This must be done before we lose track of which
00334   // blocks were originally in the code region.
00335   std::vector<User*> Users(header->use_begin(), header->use_end());
00336   for (unsigned i = 0, e = Users.size(); i != e; ++i)
00337     // The BasicBlock which contains the branch is not in the region
00338     // modify the branch target to a new block
00339     if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
00340       if (!BlocksToExtract.count(TI->getParent()) &&
00341           TI->getParent()->getParent() == oldFunction)
00342         TI->replaceUsesOfWith(header, newHeader);
00343 
00344   return newFunction;
00345 }
00346 
00347 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
00348 /// the call instruction, splitting any PHI nodes in the header block as
00349 /// necessary.
00350 void CodeExtractor::
00351 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
00352                            Values &inputs, Values &outputs) {
00353   // Emit a call to the new function, passing in: *pointer to struct (if
00354   // aggregating parameters), or plan inputs and allocated memory for outputs
00355   std::vector<Value*> params, StructValues, ReloadOutputs;
00356 
00357   // Add inputs as params, or to be filled into the struct
00358   for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
00359     if (AggregateArgs)
00360       StructValues.push_back(*i);
00361     else
00362       params.push_back(*i);
00363 
00364   // Create allocas for the outputs
00365   for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
00366     if (AggregateArgs) {
00367       StructValues.push_back(*i);
00368     } else {
00369       AllocaInst *alloca =
00370         new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
00371                        codeReplacer->getParent()->begin()->begin());
00372       ReloadOutputs.push_back(alloca);
00373       params.push_back(alloca);
00374     }
00375   }
00376 
00377   AllocaInst *Struct = 0;
00378   if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
00379     std::vector<const Type*> ArgTypes;
00380     for (Values::iterator v = StructValues.begin(),
00381            ve = StructValues.end(); v != ve; ++v)
00382       ArgTypes.push_back((*v)->getType());
00383 
00384     // Allocate a struct at the beginning of this function
00385     Type *StructArgTy = StructType::get(ArgTypes);
00386     Struct = 
00387       new AllocaInst(StructArgTy, 0, "structArg", 
00388                      codeReplacer->getParent()->begin()->begin());
00389     params.push_back(Struct);
00390 
00391     for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
00392       std::vector<Value*> Indices;
00393       Indices.push_back(Constant::getNullValue(Type::UIntTy));
00394       Indices.push_back(ConstantUInt::get(Type::UIntTy, i));
00395       GetElementPtrInst *GEP =
00396         new GetElementPtrInst(Struct, Indices,
00397                               "gep_" + StructValues[i]->getName());
00398       codeReplacer->getInstList().push_back(GEP);
00399       StoreInst *SI = new StoreInst(StructValues[i], GEP);
00400       codeReplacer->getInstList().push_back(SI);
00401     }
00402   } 
00403 
00404   // Emit the call to the function
00405   CallInst *call = new CallInst(newFunction, params,
00406                                 NumExitBlocks > 1 ? "targetBlock" : "");
00407   codeReplacer->getInstList().push_back(call);
00408 
00409   Function::aiterator OutputArgBegin = newFunction->abegin();
00410   unsigned FirstOut = inputs.size();
00411   if (!AggregateArgs)
00412     std::advance(OutputArgBegin, inputs.size());
00413 
00414   // Reload the outputs passed in by reference
00415   for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
00416     Value *Output = 0;
00417     if (AggregateArgs) {
00418       std::vector<Value*> Indices;
00419       Indices.push_back(Constant::getNullValue(Type::UIntTy));
00420       Indices.push_back(ConstantUInt::get(Type::UIntTy, FirstOut + i));
00421       GetElementPtrInst *GEP 
00422         = new GetElementPtrInst(Struct, Indices,
00423                                 "gep_reload_" + outputs[i]->getName());
00424       codeReplacer->getInstList().push_back(GEP);
00425       Output = GEP;
00426     } else {
00427       Output = ReloadOutputs[i];
00428     }
00429     LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
00430     codeReplacer->getInstList().push_back(load);
00431     std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
00432     for (unsigned u = 0, e = Users.size(); u != e; ++u) {
00433       Instruction *inst = cast<Instruction>(Users[u]);
00434       if (!BlocksToExtract.count(inst->getParent()))
00435         inst->replaceUsesOfWith(outputs[i], load);
00436     }
00437   }
00438 
00439   // Now we can emit a switch statement using the call as a value.
00440   SwitchInst *TheSwitch =
00441     new SwitchInst(ConstantUInt::getNullValue(Type::UShortTy),
00442                    codeReplacer, codeReplacer);
00443 
00444   // Since there may be multiple exits from the original region, make the new
00445   // function return an unsigned, switch on that number.  This loop iterates
00446   // over all of the blocks in the extracted region, updating any terminator
00447   // instructions in the to-be-extracted region that branch to blocks that are
00448   // not in the region to be extracted.
00449   std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
00450 
00451   unsigned switchVal = 0;
00452   for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
00453          e = BlocksToExtract.end(); i != e; ++i) {
00454     TerminatorInst *TI = (*i)->getTerminator();
00455     for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
00456       if (!BlocksToExtract.count(TI->getSuccessor(i))) {
00457         BasicBlock *OldTarget = TI->getSuccessor(i);
00458         // add a new basic block which returns the appropriate value
00459         BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
00460         if (!NewTarget) {
00461           // If we don't already have an exit stub for this non-extracted
00462           // destination, create one now!
00463           NewTarget = new BasicBlock(OldTarget->getName() + ".exitStub",
00464                                      newFunction);
00465           unsigned SuccNum = switchVal++;
00466 
00467           Value *brVal = 0;
00468           switch (NumExitBlocks) {
00469           case 0:
00470           case 1: break;  // No value needed.
00471           case 2:         // Conditional branch, return a bool
00472             brVal = SuccNum ? ConstantBool::False : ConstantBool::True;
00473             break;
00474           default:
00475             brVal = ConstantUInt::get(Type::UShortTy, SuccNum);
00476             break;
00477           }
00478 
00479           ReturnInst *NTRet = new ReturnInst(brVal, NewTarget);
00480 
00481           // Update the switch instruction.
00482           TheSwitch->addCase(ConstantUInt::get(Type::UShortTy, SuccNum),
00483                              OldTarget);
00484 
00485           // Restore values just before we exit
00486           Function::aiterator OAI = OutputArgBegin;
00487           for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
00488             // For an invoke, the normal destination is the only one that is
00489             // dominated by the result of the invocation
00490             BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
00491 
00492             bool DominatesDef = true;
00493 
00494             if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) {
00495               DefBlock = Invoke->getNormalDest();
00496 
00497               // Make sure we are looking at the original successor block, not
00498               // at a newly inserted exit block, which won't be in the dominator
00499               // info.
00500               for (std::map<BasicBlock*, BasicBlock*>::iterator I =
00501                      ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I)
00502                 if (DefBlock == I->second) {
00503                   DefBlock = I->first;
00504                   break;
00505                 }
00506 
00507               // In the extract block case, if the block we are extracting ends
00508               // with an invoke instruction, make sure that we don't emit a
00509               // store of the invoke value for the unwind block.
00510               if (!DS && DefBlock != OldTarget)
00511                 DominatesDef = false;
00512             }
00513 
00514             if (DS)
00515               DominatesDef = DS->dominates(DefBlock, OldTarget);
00516 
00517             if (DominatesDef) {
00518               if (AggregateArgs) {
00519                 std::vector<Value*> Indices;
00520                 Indices.push_back(Constant::getNullValue(Type::UIntTy));
00521                 Indices.push_back(ConstantUInt::get(Type::UIntTy,FirstOut+out));
00522                 GetElementPtrInst *GEP =
00523                   new GetElementPtrInst(OAI, Indices,
00524                                         "gep_" + outputs[out]->getName(), 
00525                                         NTRet);
00526                 new StoreInst(outputs[out], GEP, NTRet);
00527               } else {
00528                 new StoreInst(outputs[out], OAI, NTRet);
00529               }
00530             }
00531             // Advance output iterator even if we don't emit a store
00532             if (!AggregateArgs) ++OAI;
00533           }
00534         }
00535 
00536         // rewrite the original branch instruction with this new target
00537         TI->setSuccessor(i, NewTarget);
00538       }
00539   }
00540 
00541   // Now that we've done the deed, simplify the switch instruction.
00542   const Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
00543   switch (NumExitBlocks) {
00544   case 0:
00545     // There are no successors (the block containing the switch itself), which
00546     // means that previously this was the last part of the function, and hence
00547     // this should be rewritten as a `ret'
00548     
00549     // Check if the function should return a value
00550     if (OldFnRetTy == Type::VoidTy) {
00551       new ReturnInst(0, TheSwitch);  // Return void
00552     } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
00553       // return what we have
00554       new ReturnInst(TheSwitch->getCondition(), TheSwitch);
00555     } else {
00556       // Otherwise we must have code extracted an unwind or something, just
00557       // return whatever we want.
00558       new ReturnInst(Constant::getNullValue(OldFnRetTy), TheSwitch);
00559     }
00560 
00561     TheSwitch->getParent()->getInstList().erase(TheSwitch);
00562     break;
00563   case 1:
00564     // Only a single destination, change the switch into an unconditional
00565     // branch.
00566     new BranchInst(TheSwitch->getSuccessor(1), TheSwitch);
00567     TheSwitch->getParent()->getInstList().erase(TheSwitch);
00568     break;
00569   case 2:
00570     new BranchInst(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
00571                    call, TheSwitch);
00572     TheSwitch->getParent()->getInstList().erase(TheSwitch);
00573     break;
00574   default:
00575     // Otherwise, make the default destination of the switch instruction be one
00576     // of the other successors.
00577     TheSwitch->setOperand(0, call);
00578     TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks));
00579     TheSwitch->removeCase(NumExitBlocks);  // Remove redundant case
00580     break;
00581   }
00582 }
00583 
00584 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
00585   Function *oldFunc = (*BlocksToExtract.begin())->getParent();
00586   Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
00587   Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
00588 
00589   for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
00590          e = BlocksToExtract.end(); i != e; ++i) {
00591     // Delete the basic block from the old function, and the list of blocks
00592     oldBlocks.remove(*i);
00593 
00594     // Insert this basic block into the new function
00595     newBlocks.push_back(*i);
00596   }
00597 }
00598 
00599 /// ExtractRegion - Removes a loop from a function, replaces it with a call to
00600 /// new function. Returns pointer to the new function.
00601 ///
00602 /// algorithm:
00603 ///
00604 /// find inputs and outputs for the region
00605 ///
00606 /// for inputs: add to function as args, map input instr* to arg# 
00607 /// for outputs: add allocas for scalars, 
00608 ///             add to func as args, map output instr* to arg#
00609 ///
00610 /// rewrite func to use argument #s instead of instr*
00611 ///
00612 /// for each scalar output in the function: at every exit, store intermediate 
00613 /// computed result back into memory.
00614 ///
00615 Function *CodeExtractor::
00616 ExtractCodeRegion(const std::vector<BasicBlock*> &code) {
00617   if (!isEligible(code))
00618     return 0;
00619 
00620   // 1) Find inputs, outputs
00621   // 2) Construct new function
00622   //  * Add allocas for defs, pass as args by reference
00623   //  * Pass in uses as args
00624   // 3) Move code region, add call instr to func
00625   //
00626   BlocksToExtract.insert(code.begin(), code.end());
00627 
00628   Values inputs, outputs;
00629 
00630   // Assumption: this is a single-entry code region, and the header is the first
00631   // block in the region.
00632   BasicBlock *header = code[0];
00633 
00634   for (unsigned i = 1, e = code.size(); i != e; ++i)
00635     for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]);
00636          PI != E; ++PI)
00637       assert(BlocksToExtract.count(*PI) &&
00638              "No blocks in this region may have entries from outside the region"
00639              " except for the first block!");
00640   
00641   // If we have to split PHI nodes or the entry block, do so now.
00642   severSplitPHINodes(header);
00643 
00644   // If we have any return instructions in the region, split those blocks so
00645   // that the return is not in the region.
00646   splitReturnBlocks();
00647 
00648   Function *oldFunction = header->getParent();
00649 
00650   // This takes place of the original loop
00651   BasicBlock *codeReplacer = new BasicBlock("codeRepl", oldFunction, header);
00652 
00653   // The new function needs a root node because other nodes can branch to the
00654   // head of the region, but the entry node of a function cannot have preds.
00655   BasicBlock *newFuncRoot = new BasicBlock("newFuncRoot");
00656   newFuncRoot->getInstList().push_back(new BranchInst(header));
00657 
00658   // Find inputs to, outputs from the code region.
00659   findInputsOutputs(inputs, outputs);
00660 
00661   // Construct new function based on inputs/outputs & add allocas for all defs.
00662   Function *newFunction = constructFunction(inputs, outputs, header,
00663                                             newFuncRoot, 
00664                                             codeReplacer, oldFunction,
00665                                             oldFunction->getParent());
00666 
00667   emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
00668 
00669   moveCodeToFunction(newFunction);
00670 
00671   // Loop over all of the PHI nodes in the header block, and change any
00672   // references to the old incoming edge to be the new incoming edge.
00673   for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
00674     PHINode *PN = cast<PHINode>(I);
00675     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
00676       if (!BlocksToExtract.count(PN->getIncomingBlock(i)))
00677         PN->setIncomingBlock(i, newFuncRoot);
00678   }
00679   
00680   // Look at all successors of the codeReplacer block.  If any of these blocks
00681   // had PHI nodes in them, we need to update the "from" block to be the code
00682   // replacer, not the original block in the extracted region.
00683   std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
00684                                  succ_end(codeReplacer));
00685   for (unsigned i = 0, e = Succs.size(); i != e; ++i)
00686     for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) {
00687       PHINode *PN = cast<PHINode>(I);
00688       std::set<BasicBlock*> ProcessedPreds;
00689       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
00690         if (BlocksToExtract.count(PN->getIncomingBlock(i)))
00691           if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
00692             PN->setIncomingBlock(i, codeReplacer);
00693           else {
00694             // There were multiple entries in the PHI for this block, now there
00695             // is only one, so remove the duplicated entries.
00696             PN->removeIncomingValue(i, false);
00697             --i; --e;
00698           }
00699     }
00700   
00701   //std::cerr << "NEW FUNCTION: " << *newFunction;
00702   //  verifyFunction(*newFunction);
00703 
00704   //  std::cerr << "OLD FUNCTION: " << *oldFunction;
00705   //  verifyFunction(*oldFunction);
00706 
00707   DEBUG(if (verifyFunction(*newFunction)) abort());
00708   return newFunction;
00709 }
00710 
00711 bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) {
00712   // Deny code region if it contains allocas or vastarts.
00713   for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end();
00714        BB != e; ++BB)
00715     for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end();
00716          I != Ie; ++I)
00717       if (isa<AllocaInst>(*I))
00718         return false;
00719       else if (const CallInst *CI = dyn_cast<CallInst>(I))
00720         if (const Function *F = CI->getCalledFunction())
00721           if (F->getIntrinsicID() == Intrinsic::vastart)
00722             return false;
00723   return true;
00724 }
00725 
00726 
00727 /// ExtractCodeRegion - slurp a sequence of basic blocks into a brand new
00728 /// function
00729 ///
00730 Function* llvm::ExtractCodeRegion(DominatorSet &DS,
00731                                   const std::vector<BasicBlock*> &code,
00732                                   bool AggregateArgs) {
00733   return CodeExtractor(&DS, AggregateArgs).ExtractCodeRegion(code);
00734 }
00735 
00736 /// ExtractBasicBlock - slurp a natural loop into a brand new function
00737 ///
00738 Function* llvm::ExtractLoop(DominatorSet &DS, Loop *L, bool AggregateArgs) {
00739   return CodeExtractor(&DS, AggregateArgs).ExtractCodeRegion(L->getBlocks());
00740 }
00741 
00742 /// ExtractBasicBlock - slurp a basic block into a brand new function
00743 ///
00744 Function* llvm::ExtractBasicBlock(BasicBlock *BB, bool AggregateArgs) {
00745   std::vector<BasicBlock*> Blocks;
00746   Blocks.push_back(BB);
00747   return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(Blocks);  
00748 }