CodeExtractor.cpp revision 193323
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/Module.h" 22#include "llvm/Pass.h" 23#include "llvm/Analysis/Dominators.h" 24#include "llvm/Analysis/LoopInfo.h" 25#include "llvm/Analysis/Verifier.h" 26#include "llvm/Transforms/Utils/BasicBlockUtils.h" 27#include "llvm/Support/CommandLine.h" 28#include "llvm/Support/Compiler.h" 29#include "llvm/Support/Debug.h" 30#include "llvm/ADT/StringExtras.h" 31#include <algorithm> 32#include <set> 33using namespace llvm; 34 35// Provide a command-line option to aggregate function arguments into a struct 36// for functions produced by the code extractor. This is useful when converting 37// extracted functions to pthread-based code, as only one argument (void*) can 38// be passed in to pthread_create(). 39static cl::opt<bool> 40AggregateArgsOpt("aggregate-extracted-args", cl::Hidden, 41 cl::desc("Aggregate arguments to code-extracted functions")); 42 43namespace { 44 class VISIBILITY_HIDDEN CodeExtractor { 45 typedef std::vector<Value*> Values; 46 std::set<BasicBlock*> BlocksToExtract; 47 DominatorTree* DT; 48 bool AggregateArgs; 49 unsigned NumExitBlocks; 50 const Type *RetTy; 51 public: 52 CodeExtractor(DominatorTree* dt = 0, bool AggArgs = false) 53 : DT(dt), AggregateArgs(AggArgs||AggregateArgsOpt), NumExitBlocks(~0U) {} 54 55 Function *ExtractCodeRegion(const std::vector<BasicBlock*> &code); 56 57 bool isEligible(const std::vector<BasicBlock*> &code); 58 59 private: 60 /// definedInRegion - Return true if the specified value is defined in the 61 /// extracted region. 62 bool definedInRegion(Value *V) const { 63 if (Instruction *I = dyn_cast<Instruction>(V)) 64 if (BlocksToExtract.count(I->getParent())) 65 return true; 66 return false; 67 } 68 69 /// definedInCaller - Return true if the specified value is defined in the 70 /// function being code extracted, but not in the region being extracted. 71 /// These values must be passed in as live-ins to the function. 72 bool definedInCaller(Value *V) const { 73 if (isa<Argument>(V)) return true; 74 if (Instruction *I = dyn_cast<Instruction>(V)) 75 if (!BlocksToExtract.count(I->getParent())) 76 return true; 77 return false; 78 } 79 80 void severSplitPHINodes(BasicBlock *&Header); 81 void splitReturnBlocks(); 82 void findInputsOutputs(Values &inputs, Values &outputs); 83 84 Function *constructFunction(const Values &inputs, 85 const Values &outputs, 86 BasicBlock *header, 87 BasicBlock *newRootNode, BasicBlock *newHeader, 88 Function *oldFunction, Module *M); 89 90 void moveCodeToFunction(Function *newFunction); 91 92 void emitCallAndSwitchStatement(Function *newFunction, 93 BasicBlock *newHeader, 94 Values &inputs, 95 Values &outputs); 96 97 }; 98} 99 100/// severSplitPHINodes - If a PHI node has multiple inputs from outside of the 101/// region, we need to split the entry block of the region so that the PHI node 102/// is easier to deal with. 103void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) { 104 bool HasPredsFromRegion = false; 105 unsigned NumPredsOutsideRegion = 0; 106 107 if (Header != &Header->getParent()->getEntryBlock()) { 108 PHINode *PN = dyn_cast<PHINode>(Header->begin()); 109 if (!PN) return; // No PHI nodes. 110 111 // If the header node contains any PHI nodes, check to see if there is more 112 // than one entry from outside the region. If so, we need to sever the 113 // header block into two. 114 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 115 if (BlocksToExtract.count(PN->getIncomingBlock(i))) 116 HasPredsFromRegion = true; 117 else 118 ++NumPredsOutsideRegion; 119 120 // If there is one (or fewer) predecessor from outside the region, we don't 121 // need to do anything special. 122 if (NumPredsOutsideRegion <= 1) return; 123 } 124 125 // Otherwise, we need to split the header block into two pieces: one 126 // containing PHI nodes merging values from outside of the region, and a 127 // second that contains all of the code for the block and merges back any 128 // incoming values from inside of the region. 129 BasicBlock::iterator AfterPHIs = Header->getFirstNonPHI(); 130 BasicBlock *NewBB = Header->splitBasicBlock(AfterPHIs, 131 Header->getName()+".ce"); 132 133 // We only want to code extract the second block now, and it becomes the new 134 // header of the region. 135 BasicBlock *OldPred = Header; 136 BlocksToExtract.erase(OldPred); 137 BlocksToExtract.insert(NewBB); 138 Header = NewBB; 139 140 // Okay, update dominator sets. The blocks that dominate the new one are the 141 // blocks that dominate TIBB plus the new block itself. 142 if (DT) 143 DT->splitBlock(NewBB); 144 145 // Okay, now we need to adjust the PHI nodes and any branches from within the 146 // region to go to the new header block instead of the old header block. 147 if (HasPredsFromRegion) { 148 PHINode *PN = cast<PHINode>(OldPred->begin()); 149 // Loop over all of the predecessors of OldPred that are in the region, 150 // changing them to branch to NewBB instead. 151 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 152 if (BlocksToExtract.count(PN->getIncomingBlock(i))) { 153 TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator(); 154 TI->replaceUsesOfWith(OldPred, NewBB); 155 } 156 157 // Okay, everthing within the region is now branching to the right block, we 158 // just have to update the PHI nodes now, inserting PHI nodes into NewBB. 159 for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) { 160 PHINode *PN = cast<PHINode>(AfterPHIs); 161 // Create a new PHI node in the new region, which has an incoming value 162 // from OldPred of PN. 163 PHINode *NewPN = PHINode::Create(PN->getType(), PN->getName()+".ce", 164 NewBB->begin()); 165 NewPN->addIncoming(PN, OldPred); 166 167 // Loop over all of the incoming value in PN, moving them to NewPN if they 168 // are from the extracted region. 169 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) { 170 if (BlocksToExtract.count(PN->getIncomingBlock(i))) { 171 NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i)); 172 PN->removeIncomingValue(i); 173 --i; 174 } 175 } 176 } 177 } 178} 179 180void CodeExtractor::splitReturnBlocks() { 181 for (std::set<BasicBlock*>::iterator I = BlocksToExtract.begin(), 182 E = BlocksToExtract.end(); I != E; ++I) 183 if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator())) 184 (*I)->splitBasicBlock(RI, (*I)->getName()+".ret"); 185} 186 187// findInputsOutputs - Find inputs to, outputs from the code region. 188// 189void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs) { 190 std::set<BasicBlock*> ExitBlocks; 191 for (std::set<BasicBlock*>::const_iterator ci = BlocksToExtract.begin(), 192 ce = BlocksToExtract.end(); ci != ce; ++ci) { 193 BasicBlock *BB = *ci; 194 195 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 196 // If a used value is defined outside the region, it's an input. If an 197 // instruction is used outside the region, it's an output. 198 for (User::op_iterator O = I->op_begin(), E = I->op_end(); O != E; ++O) 199 if (definedInCaller(*O)) 200 inputs.push_back(*O); 201 202 // Consider uses of this instruction (outputs). 203 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 204 UI != E; ++UI) 205 if (!definedInRegion(*UI)) { 206 outputs.push_back(I); 207 break; 208 } 209 } // for: insts 210 211 // Keep track of the exit blocks from the region. 212 TerminatorInst *TI = BB->getTerminator(); 213 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 214 if (!BlocksToExtract.count(TI->getSuccessor(i))) 215 ExitBlocks.insert(TI->getSuccessor(i)); 216 } // for: basic blocks 217 218 NumExitBlocks = ExitBlocks.size(); 219 220 // Eliminate duplicates. 221 std::sort(inputs.begin(), inputs.end()); 222 inputs.erase(std::unique(inputs.begin(), inputs.end()), inputs.end()); 223 std::sort(outputs.begin(), outputs.end()); 224 outputs.erase(std::unique(outputs.begin(), outputs.end()), outputs.end()); 225} 226 227/// constructFunction - make a function based on inputs and outputs, as follows: 228/// f(in0, ..., inN, out0, ..., outN) 229/// 230Function *CodeExtractor::constructFunction(const Values &inputs, 231 const Values &outputs, 232 BasicBlock *header, 233 BasicBlock *newRootNode, 234 BasicBlock *newHeader, 235 Function *oldFunction, 236 Module *M) { 237 DOUT << "inputs: " << inputs.size() << "\n"; 238 DOUT << "outputs: " << outputs.size() << "\n"; 239 240 // This function returns unsigned, outputs will go back by reference. 241 switch (NumExitBlocks) { 242 case 0: 243 case 1: RetTy = Type::VoidTy; break; 244 case 2: RetTy = Type::Int1Ty; break; 245 default: RetTy = Type::Int16Ty; break; 246 } 247 248 std::vector<const Type*> paramTy; 249 250 // Add the types of the input values to the function's argument list 251 for (Values::const_iterator i = inputs.begin(), 252 e = inputs.end(); i != e; ++i) { 253 const Value *value = *i; 254 DOUT << "value used in func: " << *value << "\n"; 255 paramTy.push_back(value->getType()); 256 } 257 258 // Add the types of the output values to the function's argument list. 259 for (Values::const_iterator I = outputs.begin(), E = outputs.end(); 260 I != E; ++I) { 261 DOUT << "instr used in func: " << **I << "\n"; 262 if (AggregateArgs) 263 paramTy.push_back((*I)->getType()); 264 else 265 paramTy.push_back(PointerType::getUnqual((*I)->getType())); 266 } 267 268 DOUT << "Function type: " << *RetTy << " f("; 269 for (std::vector<const Type*>::iterator i = paramTy.begin(), 270 e = paramTy.end(); i != e; ++i) 271 DOUT << **i << ", "; 272 DOUT << ")\n"; 273 274 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) { 275 PointerType *StructPtr = PointerType::getUnqual(StructType::get(paramTy)); 276 paramTy.clear(); 277 paramTy.push_back(StructPtr); 278 } 279 const FunctionType *funcType = FunctionType::get(RetTy, paramTy, false); 280 281 // Create the new function 282 Function *newFunction = Function::Create(funcType, 283 GlobalValue::InternalLinkage, 284 oldFunction->getName() + "_" + 285 header->getName(), M); 286 // If the old function is no-throw, so is the new one. 287 if (oldFunction->doesNotThrow()) 288 newFunction->setDoesNotThrow(true); 289 290 newFunction->getBasicBlockList().push_back(newRootNode); 291 292 // Create an iterator to name all of the arguments we inserted. 293 Function::arg_iterator AI = newFunction->arg_begin(); 294 295 // Rewrite all users of the inputs in the extracted region to use the 296 // arguments (or appropriate addressing into struct) instead. 297 for (unsigned i = 0, e = inputs.size(); i != e; ++i) { 298 Value *RewriteVal; 299 if (AggregateArgs) { 300 Value *Idx[2]; 301 Idx[0] = Constant::getNullValue(Type::Int32Ty); 302 Idx[1] = ConstantInt::get(Type::Int32Ty, i); 303 std::string GEPname = "gep_" + inputs[i]->getName(); 304 TerminatorInst *TI = newFunction->begin()->getTerminator(); 305 GetElementPtrInst *GEP = GetElementPtrInst::Create(AI, Idx, Idx+2, 306 GEPname, TI); 307 RewriteVal = new LoadInst(GEP, "load" + GEPname, TI); 308 } else 309 RewriteVal = AI++; 310 311 std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end()); 312 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end(); 313 use != useE; ++use) 314 if (Instruction* inst = dyn_cast<Instruction>(*use)) 315 if (BlocksToExtract.count(inst->getParent())) 316 inst->replaceUsesOfWith(inputs[i], RewriteVal); 317 } 318 319 // Set names for input and output arguments. 320 if (!AggregateArgs) { 321 AI = newFunction->arg_begin(); 322 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI) 323 AI->setName(inputs[i]->getName()); 324 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI) 325 AI->setName(outputs[i]->getName()+".out"); 326 } 327 328 // Rewrite branches to basic blocks outside of the loop to new dummy blocks 329 // within the new function. This must be done before we lose track of which 330 // blocks were originally in the code region. 331 std::vector<User*> Users(header->use_begin(), header->use_end()); 332 for (unsigned i = 0, e = Users.size(); i != e; ++i) 333 // The BasicBlock which contains the branch is not in the region 334 // modify the branch target to a new block 335 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i])) 336 if (!BlocksToExtract.count(TI->getParent()) && 337 TI->getParent()->getParent() == oldFunction) 338 TI->replaceUsesOfWith(header, newHeader); 339 340 return newFunction; 341} 342 343/// emitCallAndSwitchStatement - This method sets up the caller side by adding 344/// the call instruction, splitting any PHI nodes in the header block as 345/// necessary. 346void CodeExtractor:: 347emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer, 348 Values &inputs, Values &outputs) { 349 // Emit a call to the new function, passing in: *pointer to struct (if 350 // aggregating parameters), or plan inputs and allocated memory for outputs 351 std::vector<Value*> params, StructValues, ReloadOutputs; 352 353 // Add inputs as params, or to be filled into the struct 354 for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i) 355 if (AggregateArgs) 356 StructValues.push_back(*i); 357 else 358 params.push_back(*i); 359 360 // Create allocas for the outputs 361 for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) { 362 if (AggregateArgs) { 363 StructValues.push_back(*i); 364 } else { 365 AllocaInst *alloca = 366 new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc", 367 codeReplacer->getParent()->begin()->begin()); 368 ReloadOutputs.push_back(alloca); 369 params.push_back(alloca); 370 } 371 } 372 373 AllocaInst *Struct = 0; 374 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) { 375 std::vector<const Type*> ArgTypes; 376 for (Values::iterator v = StructValues.begin(), 377 ve = StructValues.end(); v != ve; ++v) 378 ArgTypes.push_back((*v)->getType()); 379 380 // Allocate a struct at the beginning of this function 381 Type *StructArgTy = StructType::get(ArgTypes); 382 Struct = 383 new AllocaInst(StructArgTy, 0, "structArg", 384 codeReplacer->getParent()->begin()->begin()); 385 params.push_back(Struct); 386 387 for (unsigned i = 0, e = inputs.size(); i != e; ++i) { 388 Value *Idx[2]; 389 Idx[0] = Constant::getNullValue(Type::Int32Ty); 390 Idx[1] = ConstantInt::get(Type::Int32Ty, i); 391 GetElementPtrInst *GEP = 392 GetElementPtrInst::Create(Struct, Idx, Idx + 2, 393 "gep_" + StructValues[i]->getName()); 394 codeReplacer->getInstList().push_back(GEP); 395 StoreInst *SI = new StoreInst(StructValues[i], GEP); 396 codeReplacer->getInstList().push_back(SI); 397 } 398 } 399 400 // Emit the call to the function 401 CallInst *call = CallInst::Create(newFunction, params.begin(), params.end(), 402 NumExitBlocks > 1 ? "targetBlock" : ""); 403 codeReplacer->getInstList().push_back(call); 404 405 Function::arg_iterator OutputArgBegin = newFunction->arg_begin(); 406 unsigned FirstOut = inputs.size(); 407 if (!AggregateArgs) 408 std::advance(OutputArgBegin, inputs.size()); 409 410 // Reload the outputs passed in by reference 411 for (unsigned i = 0, e = outputs.size(); i != e; ++i) { 412 Value *Output = 0; 413 if (AggregateArgs) { 414 Value *Idx[2]; 415 Idx[0] = Constant::getNullValue(Type::Int32Ty); 416 Idx[1] = ConstantInt::get(Type::Int32Ty, FirstOut + i); 417 GetElementPtrInst *GEP 418 = GetElementPtrInst::Create(Struct, Idx, Idx + 2, 419 "gep_reload_" + outputs[i]->getName()); 420 codeReplacer->getInstList().push_back(GEP); 421 Output = GEP; 422 } else { 423 Output = ReloadOutputs[i]; 424 } 425 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload"); 426 codeReplacer->getInstList().push_back(load); 427 std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end()); 428 for (unsigned u = 0, e = Users.size(); u != e; ++u) { 429 Instruction *inst = cast<Instruction>(Users[u]); 430 if (!BlocksToExtract.count(inst->getParent())) 431 inst->replaceUsesOfWith(outputs[i], load); 432 } 433 } 434 435 // Now we can emit a switch statement using the call as a value. 436 SwitchInst *TheSwitch = 437 SwitchInst::Create(ConstantInt::getNullValue(Type::Int16Ty), 438 codeReplacer, 0, codeReplacer); 439 440 // Since there may be multiple exits from the original region, make the new 441 // function return an unsigned, switch on that number. This loop iterates 442 // over all of the blocks in the extracted region, updating any terminator 443 // instructions in the to-be-extracted region that branch to blocks that are 444 // not in the region to be extracted. 445 std::map<BasicBlock*, BasicBlock*> ExitBlockMap; 446 447 unsigned switchVal = 0; 448 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(), 449 e = BlocksToExtract.end(); i != e; ++i) { 450 TerminatorInst *TI = (*i)->getTerminator(); 451 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 452 if (!BlocksToExtract.count(TI->getSuccessor(i))) { 453 BasicBlock *OldTarget = TI->getSuccessor(i); 454 // add a new basic block which returns the appropriate value 455 BasicBlock *&NewTarget = ExitBlockMap[OldTarget]; 456 if (!NewTarget) { 457 // If we don't already have an exit stub for this non-extracted 458 // destination, create one now! 459 NewTarget = BasicBlock::Create(OldTarget->getName() + ".exitStub", 460 newFunction); 461 unsigned SuccNum = switchVal++; 462 463 Value *brVal = 0; 464 switch (NumExitBlocks) { 465 case 0: 466 case 1: break; // No value needed. 467 case 2: // Conditional branch, return a bool 468 brVal = ConstantInt::get(Type::Int1Ty, !SuccNum); 469 break; 470 default: 471 brVal = ConstantInt::get(Type::Int16Ty, SuccNum); 472 break; 473 } 474 475 ReturnInst *NTRet = ReturnInst::Create(brVal, NewTarget); 476 477 // Update the switch instruction. 478 TheSwitch->addCase(ConstantInt::get(Type::Int16Ty, SuccNum), 479 OldTarget); 480 481 // Restore values just before we exit 482 Function::arg_iterator OAI = OutputArgBegin; 483 for (unsigned out = 0, e = outputs.size(); out != e; ++out) { 484 // For an invoke, the normal destination is the only one that is 485 // dominated by the result of the invocation 486 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent(); 487 488 bool DominatesDef = true; 489 490 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) { 491 DefBlock = Invoke->getNormalDest(); 492 493 // Make sure we are looking at the original successor block, not 494 // at a newly inserted exit block, which won't be in the dominator 495 // info. 496 for (std::map<BasicBlock*, BasicBlock*>::iterator I = 497 ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I) 498 if (DefBlock == I->second) { 499 DefBlock = I->first; 500 break; 501 } 502 503 // In the extract block case, if the block we are extracting ends 504 // with an invoke instruction, make sure that we don't emit a 505 // store of the invoke value for the unwind block. 506 if (!DT && DefBlock != OldTarget) 507 DominatesDef = false; 508 } 509 510 if (DT) 511 DominatesDef = DT->dominates(DefBlock, OldTarget); 512 513 if (DominatesDef) { 514 if (AggregateArgs) { 515 Value *Idx[2]; 516 Idx[0] = Constant::getNullValue(Type::Int32Ty); 517 Idx[1] = ConstantInt::get(Type::Int32Ty,FirstOut+out); 518 GetElementPtrInst *GEP = 519 GetElementPtrInst::Create(OAI, Idx, Idx + 2, 520 "gep_" + outputs[out]->getName(), 521 NTRet); 522 new StoreInst(outputs[out], GEP, NTRet); 523 } else { 524 new StoreInst(outputs[out], OAI, NTRet); 525 } 526 } 527 // Advance output iterator even if we don't emit a store 528 if (!AggregateArgs) ++OAI; 529 } 530 } 531 532 // rewrite the original branch instruction with this new target 533 TI->setSuccessor(i, NewTarget); 534 } 535 } 536 537 // Now that we've done the deed, simplify the switch instruction. 538 const Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType(); 539 switch (NumExitBlocks) { 540 case 0: 541 // There are no successors (the block containing the switch itself), which 542 // means that previously this was the last part of the function, and hence 543 // this should be rewritten as a `ret' 544 545 // Check if the function should return a value 546 if (OldFnRetTy == Type::VoidTy) { 547 ReturnInst::Create(0, TheSwitch); // Return void 548 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) { 549 // return what we have 550 ReturnInst::Create(TheSwitch->getCondition(), TheSwitch); 551 } else { 552 // Otherwise we must have code extracted an unwind or something, just 553 // return whatever we want. 554 ReturnInst::Create(Constant::getNullValue(OldFnRetTy), TheSwitch); 555 } 556 557 TheSwitch->eraseFromParent(); 558 break; 559 case 1: 560 // Only a single destination, change the switch into an unconditional 561 // branch. 562 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch); 563 TheSwitch->eraseFromParent(); 564 break; 565 case 2: 566 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2), 567 call, TheSwitch); 568 TheSwitch->eraseFromParent(); 569 break; 570 default: 571 // Otherwise, make the default destination of the switch instruction be one 572 // of the other successors. 573 TheSwitch->setOperand(0, call); 574 TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks)); 575 TheSwitch->removeCase(NumExitBlocks); // Remove redundant case 576 break; 577 } 578} 579 580void CodeExtractor::moveCodeToFunction(Function *newFunction) { 581 Function *oldFunc = (*BlocksToExtract.begin())->getParent(); 582 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList(); 583 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList(); 584 585 for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(), 586 e = BlocksToExtract.end(); i != e; ++i) { 587 // Delete the basic block from the old function, and the list of blocks 588 oldBlocks.remove(*i); 589 590 // Insert this basic block into the new function 591 newBlocks.push_back(*i); 592 } 593} 594 595/// ExtractRegion - Removes a loop from a function, replaces it with a call to 596/// new function. Returns pointer to the new function. 597/// 598/// algorithm: 599/// 600/// find inputs and outputs for the region 601/// 602/// for inputs: add to function as args, map input instr* to arg# 603/// for outputs: add allocas for scalars, 604/// add to func as args, map output instr* to arg# 605/// 606/// rewrite func to use argument #s instead of instr* 607/// 608/// for each scalar output in the function: at every exit, store intermediate 609/// computed result back into memory. 610/// 611Function *CodeExtractor:: 612ExtractCodeRegion(const std::vector<BasicBlock*> &code) { 613 if (!isEligible(code)) 614 return 0; 615 616 // 1) Find inputs, outputs 617 // 2) Construct new function 618 // * Add allocas for defs, pass as args by reference 619 // * Pass in uses as args 620 // 3) Move code region, add call instr to func 621 // 622 BlocksToExtract.insert(code.begin(), code.end()); 623 624 Values inputs, outputs; 625 626 // Assumption: this is a single-entry code region, and the header is the first 627 // block in the region. 628 BasicBlock *header = code[0]; 629 630 for (unsigned i = 1, e = code.size(); i != e; ++i) 631 for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]); 632 PI != E; ++PI) 633 assert(BlocksToExtract.count(*PI) && 634 "No blocks in this region may have entries from outside the region" 635 " except for the first block!"); 636 637 // If we have to split PHI nodes or the entry block, do so now. 638 severSplitPHINodes(header); 639 640 // If we have any return instructions in the region, split those blocks so 641 // that the return is not in the region. 642 splitReturnBlocks(); 643 644 Function *oldFunction = header->getParent(); 645 646 // This takes place of the original loop 647 BasicBlock *codeReplacer = BasicBlock::Create("codeRepl", oldFunction, 648 header); 649 650 // The new function needs a root node because other nodes can branch to the 651 // head of the region, but the entry node of a function cannot have preds. 652 BasicBlock *newFuncRoot = BasicBlock::Create("newFuncRoot"); 653 newFuncRoot->getInstList().push_back(BranchInst::Create(header)); 654 655 // Find inputs to, outputs from the code region. 656 findInputsOutputs(inputs, outputs); 657 658 // Construct new function based on inputs/outputs & add allocas for all defs. 659 Function *newFunction = constructFunction(inputs, outputs, header, 660 newFuncRoot, 661 codeReplacer, oldFunction, 662 oldFunction->getParent()); 663 664 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs); 665 666 moveCodeToFunction(newFunction); 667 668 // Loop over all of the PHI nodes in the header block, and change any 669 // references to the old incoming edge to be the new incoming edge. 670 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) { 671 PHINode *PN = cast<PHINode>(I); 672 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 673 if (!BlocksToExtract.count(PN->getIncomingBlock(i))) 674 PN->setIncomingBlock(i, newFuncRoot); 675 } 676 677 // Look at all successors of the codeReplacer block. If any of these blocks 678 // had PHI nodes in them, we need to update the "from" block to be the code 679 // replacer, not the original block in the extracted region. 680 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer), 681 succ_end(codeReplacer)); 682 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 683 for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) { 684 PHINode *PN = cast<PHINode>(I); 685 std::set<BasicBlock*> ProcessedPreds; 686 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 687 if (BlocksToExtract.count(PN->getIncomingBlock(i))) { 688 if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second) 689 PN->setIncomingBlock(i, codeReplacer); 690 else { 691 // There were multiple entries in the PHI for this block, now there 692 // is only one, so remove the duplicated entries. 693 PN->removeIncomingValue(i, false); 694 --i; --e; 695 } 696 } 697 } 698 699 //cerr << "NEW FUNCTION: " << *newFunction; 700 // verifyFunction(*newFunction); 701 702 // cerr << "OLD FUNCTION: " << *oldFunction; 703 // verifyFunction(*oldFunction); 704 705 DEBUG(if (verifyFunction(*newFunction)) abort()); 706 return newFunction; 707} 708 709bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) { 710 // Deny code region if it contains allocas or vastarts. 711 for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end(); 712 BB != e; ++BB) 713 for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end(); 714 I != Ie; ++I) 715 if (isa<AllocaInst>(*I)) 716 return false; 717 else if (const CallInst *CI = dyn_cast<CallInst>(I)) 718 if (const Function *F = CI->getCalledFunction()) 719 if (F->getIntrinsicID() == Intrinsic::vastart) 720 return false; 721 return true; 722} 723 724 725/// ExtractCodeRegion - slurp a sequence of basic blocks into a brand new 726/// function 727/// 728Function* llvm::ExtractCodeRegion(DominatorTree &DT, 729 const std::vector<BasicBlock*> &code, 730 bool AggregateArgs) { 731 return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(code); 732} 733 734/// ExtractBasicBlock - slurp a natural loop into a brand new function 735/// 736Function* llvm::ExtractLoop(DominatorTree &DT, Loop *L, bool AggregateArgs) { 737 return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(L->getBlocks()); 738} 739 740/// ExtractBasicBlock - slurp a basic block into a brand new function 741/// 742Function* llvm::ExtractBasicBlock(BasicBlock *BB, bool AggregateArgs) { 743 std::vector<BasicBlock*> Blocks; 744 Blocks.push_back(BB); 745 return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(Blocks); 746} 747