LLVM API Documentation

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 #include <iostream>
00033 using namespace llvm;
00034 
00035 // Provide a command-line option to aggregate function arguments into a struct
00036 // for functions produced by the code extrator. This is useful when converting
00037 // extracted functions to pthread-based code, as only one argument (void*) can
00038 // be passed in to pthread_create().
00039 static cl::opt<bool>
00040 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden,
00041                  cl::desc("Aggregate arguments to code-extracted functions"));
00042 
00043 namespace {
00044   class CodeExtractor {
00045     typedef std::vector<Value*> Values;
00046     std::set<BasicBlock*> BlocksToExtract;
00047     DominatorSet *DS;
00048     bool AggregateArgs;
00049     unsigned NumExitBlocks;
00050     const Type *RetTy;
00051   public:
00052     CodeExtractor(DominatorSet *ds = 0, bool AggArgs = false)
00053       : DS(ds), AggregateArgs(AggArgs||AggregateArgsOpt), NumExitBlocks(~0U) {}
00054 
00055     Function *ExtractCodeRegion(const std::vector<BasicBlock*> &code);
00056 
00057     bool isEligible(const std::vector<BasicBlock*> &code);
00058 
00059   private:
00060     /// definedInRegion - Return true if the specified value is defined in the
00061     /// extracted region.
00062     bool definedInRegion(Value *V) const {
00063       if (Instruction *I = dyn_cast<Instruction>(V))
00064         if (BlocksToExtract.count(I->getParent()))
00065           return true;
00066       return false;
00067     }
00068 
00069     /// definedInCaller - Return true if the specified value is defined in the
00070     /// function being code extracted, but not in the region being extracted.
00071     /// These values must be passed in as live-ins to the function.
00072     bool definedInCaller(Value *V) const {
00073       if (isa<Argument>(V)) return true;
00074       if (Instruction *I = dyn_cast<Instruction>(V))
00075         if (!BlocksToExtract.count(I->getParent()))
00076           return true;
00077       return false;
00078     }
00079 
00080     void severSplitPHINodes(BasicBlock *&Header);
00081     void splitReturnBlocks();
00082     void findInputsOutputs(Values &inputs, Values &outputs);
00083 
00084     Function *constructFunction(const Values &inputs,
00085                                 const Values &outputs,
00086                                 BasicBlock *header,
00087                                 BasicBlock *newRootNode, BasicBlock *newHeader,
00088                                 Function *oldFunction, Module *M);
00089 
00090     void moveCodeToFunction(Function *newFunction);
00091 
00092     void emitCallAndSwitchStatement(Function *newFunction,
00093                                     BasicBlock *newHeader,
00094                                     Values &inputs,
00095                                     Values &outputs);
00096 
00097   };
00098 }
00099 
00100 /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the
00101 /// region, we need to split the entry block of the region so that the PHI node
00102 /// is easier to deal with.
00103 void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) {
00104   bool HasPredsFromRegion = false;
00105   unsigned NumPredsOutsideRegion = 0;
00106 
00107   if (Header != &Header->getParent()->front()) {
00108     PHINode *PN = dyn_cast<PHINode>(Header->begin());
00109     if (!PN) return;  // No PHI nodes.
00110 
00111     // If the header node contains any PHI nodes, check to see if there is more
00112     // than one entry from outside the region.  If so, we need to sever the
00113     // header block into two.
00114     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
00115       if (BlocksToExtract.count(PN->getIncomingBlock(i)))
00116         HasPredsFromRegion = true;
00117       else
00118         ++NumPredsOutsideRegion;
00119 
00120     // If there is one (or fewer) predecessor from outside the region, we don't
00121     // need to do anything special.
00122     if (NumPredsOutsideRegion <= 1) return;
00123   }
00124 
00125   // Otherwise, we need to split the header block into two pieces: one
00126   // containing PHI nodes merging values from outside of the region, and a
00127   // second that contains all of the code for the block and merges back any
00128   // incoming values from inside of the region.
00129   BasicBlock::iterator AfterPHIs = Header->begin();
00130   while (isa<PHINode>(AfterPHIs)) ++AfterPHIs;
00131   BasicBlock *NewBB = Header->splitBasicBlock(AfterPHIs,
00132                                               Header->getName()+".ce");
00133 
00134   // We only want to code extract the second block now, and it becomes the new
00135   // header of the region.
00136   BasicBlock *OldPred = Header;
00137   BlocksToExtract.erase(OldPred);
00138   BlocksToExtract.insert(NewBB);
00139   Header = NewBB;
00140 
00141   // Okay, update dominator sets. The blocks that dominate the new one are the
00142   // blocks that dominate TIBB plus the new block itself.
00143   if (DS) {
00144     DominatorSet::DomSetType DomSet = DS->getDominators(OldPred);
00145     DomSet.insert(NewBB);  // A block always dominates itself.
00146     DS->addBasicBlock(NewBB, DomSet);
00147 
00148     // Additionally, NewBB dominates all blocks in the function that are
00149     // dominated by OldPred.
00150     Function *F = Header->getParent();
00151     for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
00152       if (DS->properlyDominates(OldPred, I))
00153         DS->addDominator(I, NewBB);
00154   }
00155 
00156   // Okay, now we need to adjust the PHI nodes and any branches from within the
00157   // region to go to the new header block instead of the old header block.
00158   if (HasPredsFromRegion) {
00159     PHINode *PN = cast<PHINode>(OldPred->begin());
00160     // Loop over all of the predecessors of OldPred that are in the region,
00161     // changing them to branch to NewBB instead.
00162     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
00163       if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
00164         TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator();
00165         TI->replaceUsesOfWith(OldPred, NewBB);
00166       }
00167 
00168     // Okay, everthing within the region is now branching to the right block, we
00169     // just have to update the PHI nodes now, inserting PHI nodes into NewBB.
00170     for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) {
00171       PHINode *PN = cast<PHINode>(AfterPHIs);
00172       // Create a new PHI node in the new region, which has an incoming value
00173       // from OldPred of PN.
00174       PHINode *NewPN = new PHINode(PN->getType(), PN->getName()+".ce",
00175                                    NewBB->begin());
00176       NewPN->addIncoming(PN, OldPred);
00177 
00178       // Loop over all of the incoming value in PN, moving them to NewPN if they
00179       // are from the extracted region.
00180       for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
00181         if (BlocksToExtract.count(PN->getIncomingBlock(i))) {
00182           NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i));
00183           PN->removeIncomingValue(i);
00184           --i;
00185         }
00186       }
00187     }
00188   }
00189 }
00190 
00191 void CodeExtractor::splitReturnBlocks() {
00192   for (std::set<BasicBlock*>::iterator I = BlocksToExtract.begin(),
00193          E = BlocksToExtract.end(); I != E; ++I)
00194     if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator()))
00195       (*I)->splitBasicBlock(RI, (*I)->getName()+".ret");
00196 }
00197 
00198 // findInputsOutputs - Find inputs to, outputs from the code region.
00199 //
00200 void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs) {
00201   std::set<BasicBlock*> ExitBlocks;
00202   for (std::set<BasicBlock*>::const_iterator ci = BlocksToExtract.begin(),
00203        ce = BlocksToExtract.end(); ci != ce; ++ci) {
00204     BasicBlock *BB = *ci;
00205 
00206     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
00207       // If a used value is defined outside the region, it's an input.  If an
00208       // instruction is used outside the region, it's an output.
00209       for (User::op_iterator O = I->op_begin(), E = I->op_end(); O != E; ++O)
00210         if (definedInCaller(*O))
00211           inputs.push_back(*O);
00212 
00213       // Consider uses of this instruction (outputs).
00214       for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
00215            UI != E; ++UI)
00216         if (!definedInRegion(*UI)) {
00217           outputs.push_back(I);
00218           break;
00219         }
00220     } // for: insts
00221 
00222     // Keep track of the exit blocks from the region.
00223     TerminatorInst *TI = BB->getTerminator();
00224     for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
00225       if (!BlocksToExtract.count(TI->getSuccessor(i)))
00226         ExitBlocks.insert(TI->getSuccessor(i));
00227   } // for: basic blocks
00228 
00229   NumExitBlocks = ExitBlocks.size();
00230 
00231   // Eliminate duplicates.
00232   std::sort(inputs.begin(), inputs.end());
00233   inputs.erase(std::unique(inputs.begin(), inputs.end()), inputs.end());
00234   std::sort(outputs.begin(), outputs.end());
00235   outputs.erase(std::unique(outputs.begin(), outputs.end()), outputs.end());
00236 }
00237 
00238 /// constructFunction - make a function based on inputs and outputs, as follows:
00239 /// f(in0, ..., inN, out0, ..., outN)
00240 ///
00241 Function *CodeExtractor::constructFunction(const Values &inputs,
00242                                            const Values &outputs,
00243                                            BasicBlock *header,
00244                                            BasicBlock *newRootNode,
00245                                            BasicBlock *newHeader,
00246                                            Function *oldFunction,
00247                                            Module *M) {
00248   DEBUG(std::cerr << "inputs: " << inputs.size() << "\n");
00249   DEBUG(std::cerr << "outputs: " << outputs.size() << "\n");
00250 
00251   // This function returns unsigned, outputs will go back by reference.
00252   switch (NumExitBlocks) {
00253   case 0:
00254   case 1: RetTy = Type::VoidTy; break;
00255   case 2: RetTy = Type::BoolTy; break;
00256   default: RetTy = Type::UShortTy; break;
00257   }
00258 
00259   std::vector<const Type*> paramTy;
00260 
00261   // Add the types of the input values to the function's argument list
00262   for (Values::const_iterator i = inputs.begin(),
00263          e = inputs.end(); i != e; ++i) {
00264     const Value *value = *i;
00265     DEBUG(std::cerr << "value used in func: " << *value << "\n");
00266     paramTy.push_back(value->getType());
00267   }
00268 
00269   // Add the types of the output values to the function's argument list.
00270   for (Values::const_iterator I = outputs.begin(), E = outputs.end();
00271        I != E; ++I) {
00272     DEBUG(std::cerr << "instr used in func: " << **I << "\n");
00273     if (AggregateArgs)
00274       paramTy.push_back((*I)->getType());
00275     else
00276       paramTy.push_back(PointerType::get((*I)->getType()));
00277   }
00278 
00279   DEBUG(std::cerr << "Function type: " << *RetTy << " f(");
00280   DEBUG(for (std::vector<const Type*>::iterator i = paramTy.begin(),
00281                e = paramTy.end(); i != e; ++i) std::cerr << **i << ", ");
00282   DEBUG(std::cerr << ")\n");
00283 
00284   if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
00285     PointerType *StructPtr = PointerType::get(StructType::get(paramTy));
00286     paramTy.clear();
00287     paramTy.push_back(StructPtr);
00288   }
00289   const FunctionType *funcType = FunctionType::get(RetTy, paramTy, false);
00290 
00291   // Create the new function
00292   Function *newFunction = new Function(funcType,
00293                                        GlobalValue::InternalLinkage,
00294                                        oldFunction->getName() + "_" +
00295                                        header->getName(), M);
00296   newFunction->getBasicBlockList().push_back(newRootNode);
00297 
00298   // Create an iterator to name all of the arguments we inserted.
00299   Function::arg_iterator AI = newFunction->arg_begin();
00300 
00301   // Rewrite all users of the inputs in the extracted region to use the
00302   // arguments (or appropriate addressing into struct) instead.
00303   for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
00304     Value *RewriteVal;
00305     if (AggregateArgs) {
00306       std::vector<Value*> Indices;
00307       Indices.push_back(Constant::getNullValue(Type::UIntTy));
00308       Indices.push_back(ConstantUInt::get(Type::UIntTy, i));
00309       std::string GEPname = "gep_" + inputs[i]->getName();
00310       TerminatorInst *TI = newFunction->begin()->getTerminator();
00311       GetElementPtrInst *GEP = new GetElementPtrInst(AI, Indices, GEPname, TI);
00312       RewriteVal = new LoadInst(GEP, "load" + GEPname, TI);
00313     } else
00314       RewriteVal = AI++;
00315 
00316     std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end());
00317     for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end();
00318          use != useE; ++use)
00319       if (Instruction* inst = dyn_cast<Instruction>(*use))
00320         if (BlocksToExtract.count(inst->getParent()))
00321           inst->replaceUsesOfWith(inputs[i], RewriteVal);
00322   }
00323 
00324   // Set names for input and output arguments.
00325   if (!AggregateArgs) {
00326     AI = newFunction->arg_begin();
00327     for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI)
00328       AI->setName(inputs[i]->getName());
00329     for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI)
00330       AI->setName(outputs[i]->getName()+".out");
00331   }
00332 
00333   // Rewrite branches to basic blocks outside of the loop to new dummy blocks
00334   // within the new function. This must be done before we lose track of which
00335   // blocks were originally in the code region.
00336   std::vector<User*> Users(header->use_begin(), header->use_end());
00337   for (unsigned i = 0, e = Users.size(); i != e; ++i)
00338     // The BasicBlock which contains the branch is not in the region
00339     // modify the branch target to a new block
00340     if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i]))
00341       if (!BlocksToExtract.count(TI->getParent()) &&
00342           TI->getParent()->getParent() == oldFunction)
00343         TI->replaceUsesOfWith(header, newHeader);
00344 
00345   return newFunction;
00346 }
00347 
00348 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
00349 /// the call instruction, splitting any PHI nodes in the header block as
00350 /// necessary.
00351 void CodeExtractor::
00352 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer,
00353                            Values &inputs, Values &outputs) {
00354   // Emit a call to the new function, passing in: *pointer to struct (if
00355   // aggregating parameters), or plan inputs and allocated memory for outputs
00356   std::vector<Value*> params, StructValues, ReloadOutputs;
00357 
00358   // Add inputs as params, or to be filled into the struct
00359   for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i)
00360     if (AggregateArgs)
00361       StructValues.push_back(*i);
00362     else
00363       params.push_back(*i);
00364 
00365   // Create allocas for the outputs
00366   for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) {
00367     if (AggregateArgs) {
00368       StructValues.push_back(*i);
00369     } else {
00370       AllocaInst *alloca =
00371         new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc",
00372                        codeReplacer->getParent()->begin()->begin());
00373       ReloadOutputs.push_back(alloca);
00374       params.push_back(alloca);
00375     }
00376   }
00377 
00378   AllocaInst *Struct = 0;
00379   if (AggregateArgs && (inputs.size() + outputs.size() > 0)) {
00380     std::vector<const Type*> ArgTypes;
00381     for (Values::iterator v = StructValues.begin(),
00382            ve = StructValues.end(); v != ve; ++v)
00383       ArgTypes.push_back((*v)->getType());
00384 
00385     // Allocate a struct at the beginning of this function
00386     Type *StructArgTy = StructType::get(ArgTypes);
00387     Struct =
00388       new AllocaInst(StructArgTy, 0, "structArg",
00389                      codeReplacer->getParent()->begin()->begin());
00390     params.push_back(Struct);
00391 
00392     for (unsigned i = 0, e = inputs.size(); i != e; ++i) {
00393       std::vector<Value*> Indices;
00394       Indices.push_back(Constant::getNullValue(Type::UIntTy));
00395       Indices.push_back(ConstantUInt::get(Type::UIntTy, i));
00396       GetElementPtrInst *GEP =
00397         new GetElementPtrInst(Struct, Indices,
00398                               "gep_" + StructValues[i]->getName());
00399       codeReplacer->getInstList().push_back(GEP);
00400       StoreInst *SI = new StoreInst(StructValues[i], GEP);
00401       codeReplacer->getInstList().push_back(SI);
00402     }
00403   }
00404 
00405   // Emit the call to the function
00406   CallInst *call = new CallInst(newFunction, params,
00407                                 NumExitBlocks > 1 ? "targetBlock" : "");
00408   codeReplacer->getInstList().push_back(call);
00409 
00410   Function::arg_iterator OutputArgBegin = newFunction->arg_begin();
00411   unsigned FirstOut = inputs.size();
00412   if (!AggregateArgs)
00413     std::advance(OutputArgBegin, inputs.size());
00414 
00415   // Reload the outputs passed in by reference
00416   for (unsigned i = 0, e = outputs.size(); i != e; ++i) {
00417     Value *Output = 0;
00418     if (AggregateArgs) {
00419       std::vector<Value*> Indices;
00420       Indices.push_back(Constant::getNullValue(Type::UIntTy));
00421       Indices.push_back(ConstantUInt::get(Type::UIntTy, FirstOut + i));
00422       GetElementPtrInst *GEP
00423         = new GetElementPtrInst(Struct, Indices,
00424                                 "gep_reload_" + outputs[i]->getName());
00425       codeReplacer->getInstList().push_back(GEP);
00426       Output = GEP;
00427     } else {
00428       Output = ReloadOutputs[i];
00429     }
00430     LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload");
00431     codeReplacer->getInstList().push_back(load);
00432     std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end());
00433     for (unsigned u = 0, e = Users.size(); u != e; ++u) {
00434       Instruction *inst = cast<Instruction>(Users[u]);
00435       if (!BlocksToExtract.count(inst->getParent()))
00436         inst->replaceUsesOfWith(outputs[i], load);
00437     }
00438   }
00439 
00440   // Now we can emit a switch statement using the call as a value.
00441   SwitchInst *TheSwitch =
00442     new SwitchInst(ConstantUInt::getNullValue(Type::UShortTy),
00443                    codeReplacer, 0, codeReplacer);
00444 
00445   // Since there may be multiple exits from the original region, make the new
00446   // function return an unsigned, switch on that number.  This loop iterates
00447   // over all of the blocks in the extracted region, updating any terminator
00448   // instructions in the to-be-extracted region that branch to blocks that are
00449   // not in the region to be extracted.
00450   std::map<BasicBlock*, BasicBlock*> ExitBlockMap;
00451 
00452   unsigned switchVal = 0;
00453   for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
00454          e = BlocksToExtract.end(); i != e; ++i) {
00455     TerminatorInst *TI = (*i)->getTerminator();
00456     for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
00457       if (!BlocksToExtract.count(TI->getSuccessor(i))) {
00458         BasicBlock *OldTarget = TI->getSuccessor(i);
00459         // add a new basic block which returns the appropriate value
00460         BasicBlock *&NewTarget = ExitBlockMap[OldTarget];
00461         if (!NewTarget) {
00462           // If we don't already have an exit stub for this non-extracted
00463           // destination, create one now!
00464           NewTarget = new BasicBlock(OldTarget->getName() + ".exitStub",
00465                                      newFunction);
00466           unsigned SuccNum = switchVal++;
00467 
00468           Value *brVal = 0;
00469           switch (NumExitBlocks) {
00470           case 0:
00471           case 1: break;  // No value needed.
00472           case 2:         // Conditional branch, return a bool
00473             brVal = SuccNum ? ConstantBool::False : ConstantBool::True;
00474             break;
00475           default:
00476             brVal = ConstantUInt::get(Type::UShortTy, SuccNum);
00477             break;
00478           }
00479 
00480           ReturnInst *NTRet = new ReturnInst(brVal, NewTarget);
00481 
00482           // Update the switch instruction.
00483           TheSwitch->addCase(ConstantUInt::get(Type::UShortTy, SuccNum),
00484                              OldTarget);
00485 
00486           // Restore values just before we exit
00487           Function::arg_iterator OAI = OutputArgBegin;
00488           for (unsigned out = 0, e = outputs.size(); out != e; ++out) {
00489             // For an invoke, the normal destination is the only one that is
00490             // dominated by the result of the invocation
00491             BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent();
00492 
00493             bool DominatesDef = true;
00494 
00495             if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) {
00496               DefBlock = Invoke->getNormalDest();
00497 
00498               // Make sure we are looking at the original successor block, not
00499               // at a newly inserted exit block, which won't be in the dominator
00500               // info.
00501               for (std::map<BasicBlock*, BasicBlock*>::iterator I =
00502                      ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I)
00503                 if (DefBlock == I->second) {
00504                   DefBlock = I->first;
00505                   break;
00506                 }
00507 
00508               // In the extract block case, if the block we are extracting ends
00509               // with an invoke instruction, make sure that we don't emit a
00510               // store of the invoke value for the unwind block.
00511               if (!DS && DefBlock != OldTarget)
00512                 DominatesDef = false;
00513             }
00514 
00515             if (DS)
00516               DominatesDef = DS->dominates(DefBlock, OldTarget);
00517 
00518             if (DominatesDef) {
00519               if (AggregateArgs) {
00520                 std::vector<Value*> Indices;
00521                 Indices.push_back(Constant::getNullValue(Type::UIntTy));
00522                 Indices.push_back(ConstantUInt::get(Type::UIntTy,FirstOut+out));
00523                 GetElementPtrInst *GEP =
00524                   new GetElementPtrInst(OAI, Indices,
00525                                         "gep_" + outputs[out]->getName(),
00526                                         NTRet);
00527                 new StoreInst(outputs[out], GEP, NTRet);
00528               } else {
00529                 new StoreInst(outputs[out], OAI, NTRet);
00530               }
00531             }
00532             // Advance output iterator even if we don't emit a store
00533             if (!AggregateArgs) ++OAI;
00534           }
00535         }
00536 
00537         // rewrite the original branch instruction with this new target
00538         TI->setSuccessor(i, NewTarget);
00539       }
00540   }
00541 
00542   // Now that we've done the deed, simplify the switch instruction.
00543   const Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType();
00544   switch (NumExitBlocks) {
00545   case 0:
00546     // There are no successors (the block containing the switch itself), which
00547     // means that previously this was the last part of the function, and hence
00548     // this should be rewritten as a `ret'
00549 
00550     // Check if the function should return a value
00551     if (OldFnRetTy == Type::VoidTy) {
00552       new ReturnInst(0, TheSwitch);  // Return void
00553     } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) {
00554       // return what we have
00555       new ReturnInst(TheSwitch->getCondition(), TheSwitch);
00556     } else {
00557       // Otherwise we must have code extracted an unwind or something, just
00558       // return whatever we want.
00559       new ReturnInst(Constant::getNullValue(OldFnRetTy), TheSwitch);
00560     }
00561 
00562     TheSwitch->getParent()->getInstList().erase(TheSwitch);
00563     break;
00564   case 1:
00565     // Only a single destination, change the switch into an unconditional
00566     // branch.
00567     new BranchInst(TheSwitch->getSuccessor(1), TheSwitch);
00568     TheSwitch->getParent()->getInstList().erase(TheSwitch);
00569     break;
00570   case 2:
00571     new BranchInst(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2),
00572                    call, TheSwitch);
00573     TheSwitch->getParent()->getInstList().erase(TheSwitch);
00574     break;
00575   default:
00576     // Otherwise, make the default destination of the switch instruction be one
00577     // of the other successors.
00578     TheSwitch->setOperand(0, call);
00579     TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks));
00580     TheSwitch->removeCase(NumExitBlocks);  // Remove redundant case
00581     break;
00582   }
00583 }
00584 
00585 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
00586   Function *oldFunc = (*BlocksToExtract.begin())->getParent();
00587   Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
00588   Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
00589 
00590   for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(),
00591          e = BlocksToExtract.end(); i != e; ++i) {
00592     // Delete the basic block from the old function, and the list of blocks
00593     oldBlocks.remove(*i);
00594 
00595     // Insert this basic block into the new function
00596     newBlocks.push_back(*i);
00597   }
00598 }
00599 
00600 /// ExtractRegion - Removes a loop from a function, replaces it with a call to
00601 /// new function. Returns pointer to the new function.
00602 ///
00603 /// algorithm:
00604 ///
00605 /// find inputs and outputs for the region
00606 ///
00607 /// for inputs: add to function as args, map input instr* to arg#
00608 /// for outputs: add allocas for scalars,
00609 ///             add to func as args, map output instr* to arg#
00610 ///
00611 /// rewrite func to use argument #s instead of instr*
00612 ///
00613 /// for each scalar output in the function: at every exit, store intermediate
00614 /// computed result back into memory.
00615 ///
00616 Function *CodeExtractor::
00617 ExtractCodeRegion(const std::vector<BasicBlock*> &code) {
00618   if (!isEligible(code))
00619     return 0;
00620 
00621   // 1) Find inputs, outputs
00622   // 2) Construct new function
00623   //  * Add allocas for defs, pass as args by reference
00624   //  * Pass in uses as args
00625   // 3) Move code region, add call instr to func
00626   //
00627   BlocksToExtract.insert(code.begin(), code.end());
00628 
00629   Values inputs, outputs;
00630 
00631   // Assumption: this is a single-entry code region, and the header is the first
00632   // block in the region.
00633   BasicBlock *header = code[0];
00634 
00635   for (unsigned i = 1, e = code.size(); i != e; ++i)
00636     for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]);
00637          PI != E; ++PI)
00638       assert(BlocksToExtract.count(*PI) &&
00639              "No blocks in this region may have entries from outside the region"
00640              " except for the first block!");
00641 
00642   // If we have to split PHI nodes or the entry block, do so now.
00643   severSplitPHINodes(header);
00644 
00645   // If we have any return instructions in the region, split those blocks so
00646   // that the return is not in the region.
00647   splitReturnBlocks();
00648 
00649   Function *oldFunction = header->getParent();
00650 
00651   // This takes place of the original loop
00652   BasicBlock *codeReplacer = new BasicBlock("codeRepl", oldFunction, header);
00653 
00654   // The new function needs a root node because other nodes can branch to the
00655   // head of the region, but the entry node of a function cannot have preds.
00656   BasicBlock *newFuncRoot = new BasicBlock("newFuncRoot");
00657   newFuncRoot->getInstList().push_back(new BranchInst(header));
00658 
00659   // Find inputs to, outputs from the code region.
00660   findInputsOutputs(inputs, outputs);
00661 
00662   // Construct new function based on inputs/outputs & add allocas for all defs.
00663   Function *newFunction = constructFunction(inputs, outputs, header,
00664                                             newFuncRoot,
00665                                             codeReplacer, oldFunction,
00666                                             oldFunction->getParent());
00667 
00668   emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
00669 
00670   moveCodeToFunction(newFunction);
00671 
00672   // Loop over all of the PHI nodes in the header block, and change any
00673   // references to the old incoming edge to be the new incoming edge.
00674   for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) {
00675     PHINode *PN = cast<PHINode>(I);
00676     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
00677       if (!BlocksToExtract.count(PN->getIncomingBlock(i)))
00678         PN->setIncomingBlock(i, newFuncRoot);
00679   }
00680 
00681   // Look at all successors of the codeReplacer block.  If any of these blocks
00682   // had PHI nodes in them, we need to update the "from" block to be the code
00683   // replacer, not the original block in the extracted region.
00684   std::vector<BasicBlock*> Succs(succ_begin(codeReplacer),
00685                                  succ_end(codeReplacer));
00686   for (unsigned i = 0, e = Succs.size(); i != e; ++i)
00687     for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) {
00688       PHINode *PN = cast<PHINode>(I);
00689       std::set<BasicBlock*> ProcessedPreds;
00690       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
00691         if (BlocksToExtract.count(PN->getIncomingBlock(i)))
00692           if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second)
00693             PN->setIncomingBlock(i, codeReplacer);
00694           else {
00695             // There were multiple entries in the PHI for this block, now there
00696             // is only one, so remove the duplicated entries.
00697             PN->removeIncomingValue(i, false);
00698             --i; --e;
00699           }
00700     }
00701 
00702   //std::cerr << "NEW FUNCTION: " << *newFunction;
00703   //  verifyFunction(*newFunction);
00704 
00705   //  std::cerr << "OLD FUNCTION: " << *oldFunction;
00706   //  verifyFunction(*oldFunction);
00707 
00708   DEBUG(if (verifyFunction(*newFunction)) abort());
00709   return newFunction;
00710 }
00711 
00712 bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) {
00713   // Deny code region if it contains allocas or vastarts.
00714   for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end();
00715        BB != e; ++BB)
00716     for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end();
00717          I != Ie; ++I)
00718       if (isa<AllocaInst>(*I))
00719         return false;
00720       else if (const CallInst *CI = dyn_cast<CallInst>(I))
00721         if (const Function *F = CI->getCalledFunction())
00722           if (F->getIntrinsicID() == Intrinsic::vastart)
00723             return false;
00724   return true;
00725 }
00726 
00727 
00728 /// ExtractCodeRegion - slurp a sequence of basic blocks into a brand new
00729 /// function
00730 ///
00731 Function* llvm::ExtractCodeRegion(DominatorSet &DS,
00732                                   const std::vector<BasicBlock*> &code,
00733                                   bool AggregateArgs) {
00734   return CodeExtractor(&DS, AggregateArgs).ExtractCodeRegion(code);
00735 }
00736 
00737 /// ExtractBasicBlock - slurp a natural loop into a brand new function
00738 ///
00739 Function* llvm::ExtractLoop(DominatorSet &DS, Loop *L, bool AggregateArgs) {
00740   return CodeExtractor(&DS, AggregateArgs).ExtractCodeRegion(L->getBlocks());
00741 }
00742 
00743 /// ExtractBasicBlock - slurp a basic block into a brand new function
00744 ///
00745 Function* llvm::ExtractBasicBlock(BasicBlock *BB, bool AggregateArgs) {
00746   std::vector<BasicBlock*> Blocks;
00747   Blocks.push_back(BB);
00748   return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(Blocks);
00749 }