CodeExtractor.cpp revision 613bf1ef018514e88f64f5e60f126096963248f3
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/Compiler.h" 30#include "llvm/Support/Debug.h" 31#include "llvm/Support/ErrorHandling.h" 32#include "llvm/Support/raw_ostream.h" 33#include "llvm/ADT/StringExtras.h" 34#include <algorithm> 35#include <set> 36using 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(). 42static cl::opt<bool> 43AggregateArgsOpt("aggregate-extracted-args", cl::Hidden, 44 cl::desc("Aggregate arguments to code-extracted functions")); 45 46namespace { 47 class VISIBILITY_HIDDEN CodeExtractor { 48 typedef std::vector<Value*> Values; 49 std::set<BasicBlock*> BlocksToExtract; 50 DominatorTree* DT; 51 bool AggregateArgs; 52 unsigned NumExitBlocks; 53 const Type *RetTy; 54 public: 55 CodeExtractor(DominatorTree* dt = 0, bool AggArgs = false) 56 : DT(dt), AggregateArgs(AggArgs||AggregateArgsOpt), NumExitBlocks(~0U) {} 57 58 Function *ExtractCodeRegion(const std::vector<BasicBlock*> &code); 59 60 bool isEligible(const std::vector<BasicBlock*> &code); 61 62 private: 63 /// definedInRegion - Return true if the specified value is defined in the 64 /// extracted region. 65 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. 75 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. 106void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) { 107 bool HasPredsFromRegion = false; 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 HasPredsFromRegion = true; 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.erase(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 (HasPredsFromRegion) { 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, everthing 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(), PN->getName()+".ce", 167 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 183void CodeExtractor::splitReturnBlocks() { 184 for (std::set<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 domiantes 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// 208void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs) { 209 std::set<BasicBlock*> ExitBlocks; 210 for (std::set<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.push_back(*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.push_back(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 // Eliminate duplicates. 240 std::sort(inputs.begin(), inputs.end()); 241 inputs.erase(std::unique(inputs.begin(), inputs.end()), inputs.end()); 242 std::sort(outputs.begin(), outputs.end()); 243 outputs.erase(std::unique(outputs.begin(), outputs.end()), outputs.end()); 244} 245 246/// constructFunction - make a function based on inputs and outputs, as follows: 247/// f(in0, ..., inN, out0, ..., outN) 248/// 249Function *CodeExtractor::constructFunction(const Values &inputs, 250 const Values &outputs, 251 BasicBlock *header, 252 BasicBlock *newRootNode, 253 BasicBlock *newHeader, 254 Function *oldFunction, 255 Module *M) { 256 DEBUG(errs() << "inputs: " << inputs.size() << "\n"); 257 DEBUG(errs() << "outputs: " << outputs.size() << "\n"); 258 259 // This function returns unsigned, outputs will go back by reference. 260 switch (NumExitBlocks) { 261 case 0: 262 case 1: RetTy = Type::getVoidTy(header->getContext()); break; 263 case 2: RetTy = Type::getInt1Ty(header->getContext()); break; 264 default: RetTy = Type::getInt16Ty(header->getContext()); break; 265 } 266 267 std::vector<const Type*> paramTy; 268 269 // Add the types of the input values to the function's argument list 270 for (Values::const_iterator i = inputs.begin(), 271 e = inputs.end(); i != e; ++i) { 272 const Value *value = *i; 273 DEBUG(errs() << "value used in func: " << *value << "\n"); 274 paramTy.push_back(value->getType()); 275 } 276 277 // Add the types of the output values to the function's argument list. 278 for (Values::const_iterator I = outputs.begin(), E = outputs.end(); 279 I != E; ++I) { 280 DEBUG(errs() << "instr used in func: " << **I << "\n"); 281 if (AggregateArgs) 282 paramTy.push_back((*I)->getType()); 283 else 284 paramTy.push_back(PointerType::getUnqual((*I)->getType())); 285 } 286 287 DEBUG(errs() << "Function type: " << *RetTy << " f("); 288 for (std::vector<const Type*>::iterator i = paramTy.begin(), 289 e = paramTy.end(); i != e; ++i) 290 DEBUG(errs() << **i << ", "); 291 DEBUG(errs() << ")\n"); 292 293 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) { 294 PointerType *StructPtr = 295 PointerType::getUnqual(StructType::get(M->getContext(), paramTy)); 296 paramTy.clear(); 297 paramTy.push_back(StructPtr); 298 } 299 const FunctionType *funcType = 300 FunctionType::get(RetTy, paramTy, false); 301 302 // Create the new function 303 Function *newFunction = Function::Create(funcType, 304 GlobalValue::InternalLinkage, 305 oldFunction->getName() + "_" + 306 header->getName(), M); 307 // If the old function is no-throw, so is the new one. 308 if (oldFunction->doesNotThrow()) 309 newFunction->setDoesNotThrow(true); 310 311 newFunction->getBasicBlockList().push_back(newRootNode); 312 313 // Create an iterator to name all of the arguments we inserted. 314 Function::arg_iterator AI = newFunction->arg_begin(); 315 316 // Rewrite all users of the inputs in the extracted region to use the 317 // arguments (or appropriate addressing into struct) instead. 318 for (unsigned i = 0, e = inputs.size(); i != e; ++i) { 319 Value *RewriteVal; 320 if (AggregateArgs) { 321 Value *Idx[2]; 322 Idx[0] = Constant::getNullValue(Type::getInt32Ty(header->getContext())); 323 Idx[1] = ConstantInt::get(Type::getInt32Ty(header->getContext()), i); 324 TerminatorInst *TI = newFunction->begin()->getTerminator(); 325 GetElementPtrInst *GEP = 326 GetElementPtrInst::Create(AI, Idx, Idx+2, 327 "gep_" + inputs[i]->getName(), TI); 328 RewriteVal = new LoadInst(GEP, "loadgep_" + inputs[i]->getName(), TI); 329 } else 330 RewriteVal = AI++; 331 332 std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end()); 333 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end(); 334 use != useE; ++use) 335 if (Instruction* inst = dyn_cast<Instruction>(*use)) 336 if (BlocksToExtract.count(inst->getParent())) 337 inst->replaceUsesOfWith(inputs[i], RewriteVal); 338 } 339 340 // Set names for input and output arguments. 341 if (!AggregateArgs) { 342 AI = newFunction->arg_begin(); 343 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI) 344 AI->setName(inputs[i]->getName()); 345 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI) 346 AI->setName(outputs[i]->getName()+".out"); 347 } 348 349 // Rewrite branches to basic blocks outside of the loop to new dummy blocks 350 // within the new function. This must be done before we lose track of which 351 // blocks were originally in the code region. 352 std::vector<User*> Users(header->use_begin(), header->use_end()); 353 for (unsigned i = 0, e = Users.size(); i != e; ++i) 354 // The BasicBlock which contains the branch is not in the region 355 // modify the branch target to a new block 356 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i])) 357 if (!BlocksToExtract.count(TI->getParent()) && 358 TI->getParent()->getParent() == oldFunction) 359 TI->replaceUsesOfWith(header, newHeader); 360 361 return newFunction; 362} 363 364/// FindPhiPredForUseInBlock - Given a value and a basic block, find a PHI 365/// that uses the value within the basic block, and return the predecessor 366/// block associated with that use, or return 0 if none is found. 367static BasicBlock* FindPhiPredForUseInBlock(Value* Used, BasicBlock* BB) { 368 for (Value::use_iterator UI = Used->use_begin(), 369 UE = Used->use_end(); UI != UE; ++UI) { 370 PHINode *P = dyn_cast<PHINode>(*UI); 371 if (P && P->getParent() == BB) 372 return P->getIncomingBlock(UI); 373 } 374 375 return 0; 376} 377 378/// emitCallAndSwitchStatement - This method sets up the caller side by adding 379/// the call instruction, splitting any PHI nodes in the header block as 380/// necessary. 381void CodeExtractor:: 382emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer, 383 Values &inputs, Values &outputs) { 384 // Emit a call to the new function, passing in: *pointer to struct (if 385 // aggregating parameters), or plan inputs and allocated memory for outputs 386 std::vector<Value*> params, StructValues, ReloadOutputs, Reloads; 387 388 LLVMContext &Context = newFunction->getContext(); 389 390 // Add inputs as params, or to be filled into the struct 391 for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i) 392 if (AggregateArgs) 393 StructValues.push_back(*i); 394 else 395 params.push_back(*i); 396 397 // Create allocas for the outputs 398 for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) { 399 if (AggregateArgs) { 400 StructValues.push_back(*i); 401 } else { 402 AllocaInst *alloca = 403 new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc", 404 codeReplacer->getParent()->begin()->begin()); 405 ReloadOutputs.push_back(alloca); 406 params.push_back(alloca); 407 } 408 } 409 410 AllocaInst *Struct = 0; 411 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) { 412 std::vector<const Type*> ArgTypes; 413 for (Values::iterator v = StructValues.begin(), 414 ve = StructValues.end(); v != ve; ++v) 415 ArgTypes.push_back((*v)->getType()); 416 417 // Allocate a struct at the beginning of this function 418 Type *StructArgTy = StructType::get(newFunction->getContext(), ArgTypes); 419 Struct = 420 new AllocaInst(StructArgTy, 0, "structArg", 421 codeReplacer->getParent()->begin()->begin()); 422 params.push_back(Struct); 423 424 for (unsigned i = 0, e = inputs.size(); i != e; ++i) { 425 Value *Idx[2]; 426 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context)); 427 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), i); 428 GetElementPtrInst *GEP = 429 GetElementPtrInst::Create(Struct, Idx, Idx + 2, 430 "gep_" + StructValues[i]->getName()); 431 codeReplacer->getInstList().push_back(GEP); 432 StoreInst *SI = new StoreInst(StructValues[i], GEP); 433 codeReplacer->getInstList().push_back(SI); 434 } 435 } 436 437 // Emit the call to the function 438 CallInst *call = CallInst::Create(newFunction, params.begin(), params.end(), 439 NumExitBlocks > 1 ? "targetBlock" : ""); 440 codeReplacer->getInstList().push_back(call); 441 442 Function::arg_iterator OutputArgBegin = newFunction->arg_begin(); 443 unsigned FirstOut = inputs.size(); 444 if (!AggregateArgs) 445 std::advance(OutputArgBegin, inputs.size()); 446 447 // Reload the outputs passed in by reference 448 for (unsigned i = 0, e = outputs.size(); i != e; ++i) { 449 Value *Output = 0; 450 if (AggregateArgs) { 451 Value *Idx[2]; 452 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context)); 453 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), FirstOut + i); 454 GetElementPtrInst *GEP 455 = GetElementPtrInst::Create(Struct, Idx, Idx + 2, 456 "gep_reload_" + outputs[i]->getName()); 457 codeReplacer->getInstList().push_back(GEP); 458 Output = GEP; 459 } else { 460 Output = ReloadOutputs[i]; 461 } 462 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload"); 463 Reloads.push_back(load); 464 codeReplacer->getInstList().push_back(load); 465 std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end()); 466 for (unsigned u = 0, e = Users.size(); u != e; ++u) { 467 Instruction *inst = cast<Instruction>(Users[u]); 468 if (!BlocksToExtract.count(inst->getParent())) 469 inst->replaceUsesOfWith(outputs[i], load); 470 } 471 } 472 473 // Now we can emit a switch statement using the call as a value. 474 SwitchInst *TheSwitch = 475 SwitchInst::Create(Constant::getNullValue(Type::getInt16Ty(Context)), 476 codeReplacer, 0, codeReplacer); 477 478 // Since there may be multiple exits from the original region, make the new 479 // function return an unsigned, switch on that number. This loop iterates 480 // over all of the blocks in the extracted region, updating any terminator 481 // instructions in the to-be-extracted region that branch to blocks that are 482 // not in the region to be extracted. 483 std::map<BasicBlock*, BasicBlock*> ExitBlockMap; 484 485 unsigned switchVal = 0; 486 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(), 487 e = BlocksToExtract.end(); i != e; ++i) { 488 TerminatorInst *TI = (*i)->getTerminator(); 489 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 490 if (!BlocksToExtract.count(TI->getSuccessor(i))) { 491 BasicBlock *OldTarget = TI->getSuccessor(i); 492 // add a new basic block which returns the appropriate value 493 BasicBlock *&NewTarget = ExitBlockMap[OldTarget]; 494 if (!NewTarget) { 495 // If we don't already have an exit stub for this non-extracted 496 // destination, create one now! 497 NewTarget = BasicBlock::Create(Context, 498 OldTarget->getName() + ".exitStub", 499 newFunction); 500 unsigned SuccNum = switchVal++; 501 502 Value *brVal = 0; 503 switch (NumExitBlocks) { 504 case 0: 505 case 1: break; // No value needed. 506 case 2: // Conditional branch, return a bool 507 brVal = ConstantInt::get(Type::getInt1Ty(Context), !SuccNum); 508 break; 509 default: 510 brVal = ConstantInt::get(Type::getInt16Ty(Context), SuccNum); 511 break; 512 } 513 514 ReturnInst *NTRet = ReturnInst::Create(Context, brVal, NewTarget); 515 516 // Update the switch instruction. 517 TheSwitch->addCase(ConstantInt::get(Type::getInt16Ty(Context), 518 SuccNum), 519 OldTarget); 520 521 // Restore values just before we exit 522 Function::arg_iterator OAI = OutputArgBegin; 523 for (unsigned out = 0, e = outputs.size(); out != e; ++out) { 524 // For an invoke, the normal destination is the only one that is 525 // dominated by the result of the invocation 526 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent(); 527 528 bool DominatesDef = true; 529 530 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) { 531 DefBlock = Invoke->getNormalDest(); 532 533 // Make sure we are looking at the original successor block, not 534 // at a newly inserted exit block, which won't be in the dominator 535 // info. 536 for (std::map<BasicBlock*, BasicBlock*>::iterator I = 537 ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I) 538 if (DefBlock == I->second) { 539 DefBlock = I->first; 540 break; 541 } 542 543 // In the extract block case, if the block we are extracting ends 544 // with an invoke instruction, make sure that we don't emit a 545 // store of the invoke value for the unwind block. 546 if (!DT && DefBlock != OldTarget) 547 DominatesDef = false; 548 } 549 550 if (DT) { 551 DominatesDef = DT->dominates(DefBlock, OldTarget); 552 553 // If the output value is used by a phi in the target block, 554 // then we need to test for dominance of the phi's predecessor 555 // instead. Unfortunately, this a little complicated since we 556 // have already rewritten uses of the value to uses of the reload. 557 BasicBlock* pred = FindPhiPredForUseInBlock(Reloads[out], 558 OldTarget); 559 if (pred && DT && DT->dominates(DefBlock, pred)) 560 DominatesDef = true; 561 } 562 563 if (DominatesDef) { 564 if (AggregateArgs) { 565 Value *Idx[2]; 566 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context)); 567 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), 568 FirstOut+out); 569 GetElementPtrInst *GEP = 570 GetElementPtrInst::Create(OAI, Idx, Idx + 2, 571 "gep_" + outputs[out]->getName(), 572 NTRet); 573 new StoreInst(outputs[out], GEP, NTRet); 574 } else { 575 new StoreInst(outputs[out], OAI, NTRet); 576 } 577 } 578 // Advance output iterator even if we don't emit a store 579 if (!AggregateArgs) ++OAI; 580 } 581 } 582 583 // rewrite the original branch instruction with this new target 584 TI->setSuccessor(i, NewTarget); 585 } 586 } 587 588 // Now that we've done the deed, simplify the switch instruction. 589 const Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType(); 590 switch (NumExitBlocks) { 591 case 0: 592 // There are no successors (the block containing the switch itself), which 593 // means that previously this was the last part of the function, and hence 594 // this should be rewritten as a `ret' 595 596 // Check if the function should return a value 597 if (OldFnRetTy == Type::getVoidTy(Context)) { 598 ReturnInst::Create(Context, 0, TheSwitch); // Return void 599 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) { 600 // return what we have 601 ReturnInst::Create(Context, TheSwitch->getCondition(), TheSwitch); 602 } else { 603 // Otherwise we must have code extracted an unwind or something, just 604 // return whatever we want. 605 ReturnInst::Create(Context, 606 Constant::getNullValue(OldFnRetTy), TheSwitch); 607 } 608 609 TheSwitch->eraseFromParent(); 610 break; 611 case 1: 612 // Only a single destination, change the switch into an unconditional 613 // branch. 614 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch); 615 TheSwitch->eraseFromParent(); 616 break; 617 case 2: 618 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2), 619 call, TheSwitch); 620 TheSwitch->eraseFromParent(); 621 break; 622 default: 623 // Otherwise, make the default destination of the switch instruction be one 624 // of the other successors. 625 TheSwitch->setOperand(0, call); 626 TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks)); 627 TheSwitch->removeCase(NumExitBlocks); // Remove redundant case 628 break; 629 } 630} 631 632void CodeExtractor::moveCodeToFunction(Function *newFunction) { 633 Function *oldFunc = (*BlocksToExtract.begin())->getParent(); 634 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList(); 635 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList(); 636 637 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(), 638 e = BlocksToExtract.end(); i != e; ++i) { 639 // Delete the basic block from the old function, and the list of blocks 640 oldBlocks.remove(*i); 641 642 // Insert this basic block into the new function 643 newBlocks.push_back(*i); 644 } 645} 646 647/// ExtractRegion - Removes a loop from a function, replaces it with a call to 648/// new function. Returns pointer to the new function. 649/// 650/// algorithm: 651/// 652/// find inputs and outputs for the region 653/// 654/// for inputs: add to function as args, map input instr* to arg# 655/// for outputs: add allocas for scalars, 656/// add to func as args, map output instr* to arg# 657/// 658/// rewrite func to use argument #s instead of instr* 659/// 660/// for each scalar output in the function: at every exit, store intermediate 661/// computed result back into memory. 662/// 663Function *CodeExtractor:: 664ExtractCodeRegion(const std::vector<BasicBlock*> &code) { 665 if (!isEligible(code)) 666 return 0; 667 668 // 1) Find inputs, outputs 669 // 2) Construct new function 670 // * Add allocas for defs, pass as args by reference 671 // * Pass in uses as args 672 // 3) Move code region, add call instr to func 673 // 674 BlocksToExtract.insert(code.begin(), code.end()); 675 676 Values inputs, outputs; 677 678 // Assumption: this is a single-entry code region, and the header is the first 679 // block in the region. 680 BasicBlock *header = code[0]; 681 682 for (unsigned i = 1, e = code.size(); i != e; ++i) 683 for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]); 684 PI != E; ++PI) 685 assert(BlocksToExtract.count(*PI) && 686 "No blocks in this region may have entries from outside the region" 687 " except for the first block!"); 688 689 // If we have to split PHI nodes or the entry block, do so now. 690 severSplitPHINodes(header); 691 692 // If we have any return instructions in the region, split those blocks so 693 // that the return is not in the region. 694 splitReturnBlocks(); 695 696 Function *oldFunction = header->getParent(); 697 698 // This takes place of the original loop 699 BasicBlock *codeReplacer = BasicBlock::Create(header->getContext(), 700 "codeRepl", oldFunction, 701 header); 702 703 // The new function needs a root node because other nodes can branch to the 704 // head of the region, but the entry node of a function cannot have preds. 705 BasicBlock *newFuncRoot = BasicBlock::Create(header->getContext(), 706 "newFuncRoot"); 707 newFuncRoot->getInstList().push_back(BranchInst::Create(header)); 708 709 // Find inputs to, outputs from the code region. 710 findInputsOutputs(inputs, outputs); 711 712 // Construct new function based on inputs/outputs & add allocas for all defs. 713 Function *newFunction = constructFunction(inputs, outputs, header, 714 newFuncRoot, 715 codeReplacer, oldFunction, 716 oldFunction->getParent()); 717 718 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs); 719 720 moveCodeToFunction(newFunction); 721 722 // Loop over all of the PHI nodes in the header block, and change any 723 // references to the old incoming edge to be the new incoming edge. 724 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) { 725 PHINode *PN = cast<PHINode>(I); 726 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 727 if (!BlocksToExtract.count(PN->getIncomingBlock(i))) 728 PN->setIncomingBlock(i, newFuncRoot); 729 } 730 731 // Look at all successors of the codeReplacer block. If any of these blocks 732 // had PHI nodes in them, we need to update the "from" block to be the code 733 // replacer, not the original block in the extracted region. 734 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer), 735 succ_end(codeReplacer)); 736 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 737 for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) { 738 PHINode *PN = cast<PHINode>(I); 739 std::set<BasicBlock*> ProcessedPreds; 740 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 741 if (BlocksToExtract.count(PN->getIncomingBlock(i))) { 742 if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second) 743 PN->setIncomingBlock(i, codeReplacer); 744 else { 745 // There were multiple entries in the PHI for this block, now there 746 // is only one, so remove the duplicated entries. 747 PN->removeIncomingValue(i, false); 748 --i; --e; 749 } 750 } 751 } 752 753 //cerr << "NEW FUNCTION: " << *newFunction; 754 // verifyFunction(*newFunction); 755 756 // cerr << "OLD FUNCTION: " << *oldFunction; 757 // verifyFunction(*oldFunction); 758 759 DEBUG(if (verifyFunction(*newFunction)) 760 llvm_report_error("verifyFunction failed!")); 761 return newFunction; 762} 763 764bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) { 765 // Deny code region if it contains allocas or vastarts. 766 for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end(); 767 BB != e; ++BB) 768 for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end(); 769 I != Ie; ++I) 770 if (isa<AllocaInst>(*I)) 771 return false; 772 else if (const CallInst *CI = dyn_cast<CallInst>(I)) 773 if (const Function *F = CI->getCalledFunction()) 774 if (F->getIntrinsicID() == Intrinsic::vastart) 775 return false; 776 return true; 777} 778 779 780/// ExtractCodeRegion - slurp a sequence of basic blocks into a brand new 781/// function 782/// 783Function* llvm::ExtractCodeRegion(DominatorTree &DT, 784 const std::vector<BasicBlock*> &code, 785 bool AggregateArgs) { 786 return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(code); 787} 788 789/// ExtractBasicBlock - slurp a natural loop into a brand new function 790/// 791Function* llvm::ExtractLoop(DominatorTree &DT, Loop *L, bool AggregateArgs) { 792 return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(L->getBlocks()); 793} 794 795/// ExtractBasicBlock - slurp a basic block into a brand new function 796/// 797Function* llvm::ExtractBasicBlock(BasicBlock *BB, bool AggregateArgs) { 798 std::vector<BasicBlock*> Blocks; 799 Blocks.push_back(BB); 800 return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(Blocks); 801} 802