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/CodeExtractor.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/ADT/SetVector.h" 19 #include "llvm/ADT/StringExtras.h" 20 #include "llvm/Analysis/LoopInfo.h" 21 #include "llvm/Analysis/RegionInfo.h" 22 #include "llvm/Analysis/RegionIterator.h" 23 #include "llvm/IR/Constants.h" 24 #include "llvm/IR/DerivedTypes.h" 25 #include "llvm/IR/Dominators.h" 26 #include "llvm/IR/Instructions.h" 27 #include "llvm/IR/Intrinsics.h" 28 #include "llvm/IR/LLVMContext.h" 29 #include "llvm/IR/Module.h" 30 #include "llvm/IR/Verifier.h" 31 #include "llvm/Pass.h" 32 #include "llvm/Support/CommandLine.h" 33 #include "llvm/Support/Debug.h" 34 #include "llvm/Support/ErrorHandling.h" 35 #include "llvm/Support/raw_ostream.h" 36 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 37 #include <algorithm> 38 #include <set> 39 using namespace llvm; 40 41 #define DEBUG_TYPE "code-extractor" 42 43 // Provide a command-line option to aggregate function arguments into a struct 44 // for functions produced by the code extractor. This is useful when converting 45 // extracted functions to pthread-based code, as only one argument (void*) can 46 // be passed in to pthread_create(). 47 static cl::opt<bool> 48 AggregateArgsOpt("aggregate-extracted-args", cl::Hidden, 49 cl::desc("Aggregate arguments to code-extracted functions")); 50 51 /// \brief Test whether a block is valid for extraction. 52 static bool isBlockValidForExtraction(const BasicBlock &BB) { 53 // Landing pads must be in the function where they were inserted for cleanup. 54 if (BB.isLandingPad()) 55 return false; 56 57 // Don't hoist code containing allocas, invokes, or vastarts. 58 for (BasicBlock::const_iterator I = BB.begin(), E = BB.end(); I != E; ++I) { 59 if (isa<AllocaInst>(I) || isa<InvokeInst>(I)) 60 return false; 61 if (const CallInst *CI = dyn_cast<CallInst>(I)) 62 if (const Function *F = CI->getCalledFunction()) 63 if (F->getIntrinsicID() == Intrinsic::vastart) 64 return false; 65 } 66 67 return true; 68 } 69 70 /// \brief Build a set of blocks to extract if the input blocks are viable. 71 template <typename IteratorT> 72 static SetVector<BasicBlock *> buildExtractionBlockSet(IteratorT BBBegin, 73 IteratorT BBEnd) { 74 SetVector<BasicBlock *> Result; 75 76 assert(BBBegin != BBEnd); 77 78 // Loop over the blocks, adding them to our set-vector, and aborting with an 79 // empty set if we encounter invalid blocks. 80 for (IteratorT I = BBBegin, E = BBEnd; I != E; ++I) { 81 if (!Result.insert(*I)) 82 llvm_unreachable("Repeated basic blocks in extraction input"); 83 84 if (!isBlockValidForExtraction(**I)) { 85 Result.clear(); 86 return Result; 87 } 88 } 89 90 #ifndef NDEBUG 91 for (SetVector<BasicBlock *>::iterator I = std::next(Result.begin()), 92 E = Result.end(); 93 I != E; ++I) 94 for (pred_iterator PI = pred_begin(*I), PE = pred_end(*I); 95 PI != PE; ++PI) 96 assert(Result.count(*PI) && 97 "No blocks in this region may have entries from outside the region" 98 " except for the first block!"); 99 #endif 100 101 return Result; 102 } 103 104 /// \brief Helper to call buildExtractionBlockSet with an ArrayRef. 105 static SetVector<BasicBlock *> 106 buildExtractionBlockSet(ArrayRef<BasicBlock *> BBs) { 107 return buildExtractionBlockSet(BBs.begin(), BBs.end()); 108 } 109 110 /// \brief Helper to call buildExtractionBlockSet with a RegionNode. 111 static SetVector<BasicBlock *> 112 buildExtractionBlockSet(const RegionNode &RN) { 113 if (!RN.isSubRegion()) 114 // Just a single BasicBlock. 115 return buildExtractionBlockSet(RN.getNodeAs<BasicBlock>()); 116 117 const Region &R = *RN.getNodeAs<Region>(); 118 119 return buildExtractionBlockSet(R.block_begin(), R.block_end()); 120 } 121 122 CodeExtractor::CodeExtractor(BasicBlock *BB, bool AggregateArgs) 123 : DT(nullptr), AggregateArgs(AggregateArgs||AggregateArgsOpt), 124 Blocks(buildExtractionBlockSet(BB)), NumExitBlocks(~0U) {} 125 126 CodeExtractor::CodeExtractor(ArrayRef<BasicBlock *> BBs, DominatorTree *DT, 127 bool AggregateArgs) 128 : DT(DT), AggregateArgs(AggregateArgs||AggregateArgsOpt), 129 Blocks(buildExtractionBlockSet(BBs)), NumExitBlocks(~0U) {} 130 131 CodeExtractor::CodeExtractor(DominatorTree &DT, Loop &L, bool AggregateArgs) 132 : DT(&DT), AggregateArgs(AggregateArgs||AggregateArgsOpt), 133 Blocks(buildExtractionBlockSet(L.getBlocks())), NumExitBlocks(~0U) {} 134 135 CodeExtractor::CodeExtractor(DominatorTree &DT, const RegionNode &RN, 136 bool AggregateArgs) 137 : DT(&DT), AggregateArgs(AggregateArgs||AggregateArgsOpt), 138 Blocks(buildExtractionBlockSet(RN)), NumExitBlocks(~0U) {} 139 140 /// definedInRegion - Return true if the specified value is defined in the 141 /// extracted region. 142 static bool definedInRegion(const SetVector<BasicBlock *> &Blocks, Value *V) { 143 if (Instruction *I = dyn_cast<Instruction>(V)) 144 if (Blocks.count(I->getParent())) 145 return true; 146 return false; 147 } 148 149 /// definedInCaller - Return true if the specified value is defined in the 150 /// function being code extracted, but not in the region being extracted. 151 /// These values must be passed in as live-ins to the function. 152 static bool definedInCaller(const SetVector<BasicBlock *> &Blocks, Value *V) { 153 if (isa<Argument>(V)) return true; 154 if (Instruction *I = dyn_cast<Instruction>(V)) 155 if (!Blocks.count(I->getParent())) 156 return true; 157 return false; 158 } 159 160 void CodeExtractor::findInputsOutputs(ValueSet &Inputs, 161 ValueSet &Outputs) const { 162 for (SetVector<BasicBlock *>::const_iterator I = Blocks.begin(), 163 E = Blocks.end(); 164 I != E; ++I) { 165 BasicBlock *BB = *I; 166 167 // If a used value is defined outside the region, it's an input. If an 168 // instruction is used outside the region, it's an output. 169 for (BasicBlock::iterator II = BB->begin(), IE = BB->end(); 170 II != IE; ++II) { 171 for (User::op_iterator OI = II->op_begin(), OE = II->op_end(); 172 OI != OE; ++OI) 173 if (definedInCaller(Blocks, *OI)) 174 Inputs.insert(*OI); 175 176 for (User *U : II->users()) 177 if (!definedInRegion(Blocks, U)) { 178 Outputs.insert(II); 179 break; 180 } 181 } 182 } 183 } 184 185 /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the 186 /// region, we need to split the entry block of the region so that the PHI node 187 /// is easier to deal with. 188 void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) { 189 unsigned NumPredsFromRegion = 0; 190 unsigned NumPredsOutsideRegion = 0; 191 192 if (Header != &Header->getParent()->getEntryBlock()) { 193 PHINode *PN = dyn_cast<PHINode>(Header->begin()); 194 if (!PN) return; // No PHI nodes. 195 196 // If the header node contains any PHI nodes, check to see if there is more 197 // than one entry from outside the region. If so, we need to sever the 198 // header block into two. 199 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 200 if (Blocks.count(PN->getIncomingBlock(i))) 201 ++NumPredsFromRegion; 202 else 203 ++NumPredsOutsideRegion; 204 205 // If there is one (or fewer) predecessor from outside the region, we don't 206 // need to do anything special. 207 if (NumPredsOutsideRegion <= 1) return; 208 } 209 210 // Otherwise, we need to split the header block into two pieces: one 211 // containing PHI nodes merging values from outside of the region, and a 212 // second that contains all of the code for the block and merges back any 213 // incoming values from inside of the region. 214 BasicBlock::iterator AfterPHIs = Header->getFirstNonPHI(); 215 BasicBlock *NewBB = Header->splitBasicBlock(AfterPHIs, 216 Header->getName()+".ce"); 217 218 // We only want to code extract the second block now, and it becomes the new 219 // header of the region. 220 BasicBlock *OldPred = Header; 221 Blocks.remove(OldPred); 222 Blocks.insert(NewBB); 223 Header = NewBB; 224 225 // Okay, update dominator sets. The blocks that dominate the new one are the 226 // blocks that dominate TIBB plus the new block itself. 227 if (DT) 228 DT->splitBlock(NewBB); 229 230 // Okay, now we need to adjust the PHI nodes and any branches from within the 231 // region to go to the new header block instead of the old header block. 232 if (NumPredsFromRegion) { 233 PHINode *PN = cast<PHINode>(OldPred->begin()); 234 // Loop over all of the predecessors of OldPred that are in the region, 235 // changing them to branch to NewBB instead. 236 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 237 if (Blocks.count(PN->getIncomingBlock(i))) { 238 TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator(); 239 TI->replaceUsesOfWith(OldPred, NewBB); 240 } 241 242 // Okay, everything within the region is now branching to the right block, we 243 // just have to update the PHI nodes now, inserting PHI nodes into NewBB. 244 for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) { 245 PHINode *PN = cast<PHINode>(AfterPHIs); 246 // Create a new PHI node in the new region, which has an incoming value 247 // from OldPred of PN. 248 PHINode *NewPN = PHINode::Create(PN->getType(), 1 + NumPredsFromRegion, 249 PN->getName()+".ce", NewBB->begin()); 250 NewPN->addIncoming(PN, OldPred); 251 252 // Loop over all of the incoming value in PN, moving them to NewPN if they 253 // are from the extracted region. 254 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) { 255 if (Blocks.count(PN->getIncomingBlock(i))) { 256 NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i)); 257 PN->removeIncomingValue(i); 258 --i; 259 } 260 } 261 } 262 } 263 } 264 265 void CodeExtractor::splitReturnBlocks() { 266 for (SetVector<BasicBlock *>::iterator I = Blocks.begin(), E = Blocks.end(); 267 I != E; ++I) 268 if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator())) { 269 BasicBlock *New = (*I)->splitBasicBlock(RI, (*I)->getName()+".ret"); 270 if (DT) { 271 // Old dominates New. New node dominates all other nodes dominated 272 // by Old. 273 DomTreeNode *OldNode = DT->getNode(*I); 274 SmallVector<DomTreeNode*, 8> Children; 275 for (DomTreeNode::iterator DI = OldNode->begin(), DE = OldNode->end(); 276 DI != DE; ++DI) 277 Children.push_back(*DI); 278 279 DomTreeNode *NewNode = DT->addNewBlock(New, *I); 280 281 for (SmallVectorImpl<DomTreeNode *>::iterator I = Children.begin(), 282 E = Children.end(); I != E; ++I) 283 DT->changeImmediateDominator(*I, NewNode); 284 } 285 } 286 } 287 288 /// constructFunction - make a function based on inputs and outputs, as follows: 289 /// f(in0, ..., inN, out0, ..., outN) 290 /// 291 Function *CodeExtractor::constructFunction(const ValueSet &inputs, 292 const ValueSet &outputs, 293 BasicBlock *header, 294 BasicBlock *newRootNode, 295 BasicBlock *newHeader, 296 Function *oldFunction, 297 Module *M) { 298 DEBUG(dbgs() << "inputs: " << inputs.size() << "\n"); 299 DEBUG(dbgs() << "outputs: " << outputs.size() << "\n"); 300 301 // This function returns unsigned, outputs will go back by reference. 302 switch (NumExitBlocks) { 303 case 0: 304 case 1: RetTy = Type::getVoidTy(header->getContext()); break; 305 case 2: RetTy = Type::getInt1Ty(header->getContext()); break; 306 default: RetTy = Type::getInt16Ty(header->getContext()); break; 307 } 308 309 std::vector<Type*> paramTy; 310 311 // Add the types of the input values to the function's argument list 312 for (ValueSet::const_iterator i = inputs.begin(), e = inputs.end(); 313 i != e; ++i) { 314 const Value *value = *i; 315 DEBUG(dbgs() << "value used in func: " << *value << "\n"); 316 paramTy.push_back(value->getType()); 317 } 318 319 // Add the types of the output values to the function's argument list. 320 for (ValueSet::const_iterator I = outputs.begin(), E = outputs.end(); 321 I != E; ++I) { 322 DEBUG(dbgs() << "instr used in func: " << **I << "\n"); 323 if (AggregateArgs) 324 paramTy.push_back((*I)->getType()); 325 else 326 paramTy.push_back(PointerType::getUnqual((*I)->getType())); 327 } 328 329 DEBUG(dbgs() << "Function type: " << *RetTy << " f("); 330 for (std::vector<Type*>::iterator i = paramTy.begin(), 331 e = paramTy.end(); i != e; ++i) 332 DEBUG(dbgs() << **i << ", "); 333 DEBUG(dbgs() << ")\n"); 334 335 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) { 336 PointerType *StructPtr = 337 PointerType::getUnqual(StructType::get(M->getContext(), paramTy)); 338 paramTy.clear(); 339 paramTy.push_back(StructPtr); 340 } 341 FunctionType *funcType = 342 FunctionType::get(RetTy, paramTy, false); 343 344 // Create the new function 345 Function *newFunction = Function::Create(funcType, 346 GlobalValue::InternalLinkage, 347 oldFunction->getName() + "_" + 348 header->getName(), M); 349 // If the old function is no-throw, so is the new one. 350 if (oldFunction->doesNotThrow()) 351 newFunction->setDoesNotThrow(); 352 353 newFunction->getBasicBlockList().push_back(newRootNode); 354 355 // Create an iterator to name all of the arguments we inserted. 356 Function::arg_iterator AI = newFunction->arg_begin(); 357 358 // Rewrite all users of the inputs in the extracted region to use the 359 // arguments (or appropriate addressing into struct) instead. 360 for (unsigned i = 0, e = inputs.size(); i != e; ++i) { 361 Value *RewriteVal; 362 if (AggregateArgs) { 363 Value *Idx[2]; 364 Idx[0] = Constant::getNullValue(Type::getInt32Ty(header->getContext())); 365 Idx[1] = ConstantInt::get(Type::getInt32Ty(header->getContext()), i); 366 TerminatorInst *TI = newFunction->begin()->getTerminator(); 367 GetElementPtrInst *GEP = 368 GetElementPtrInst::Create(AI, Idx, "gep_" + inputs[i]->getName(), TI); 369 RewriteVal = new LoadInst(GEP, "loadgep_" + inputs[i]->getName(), TI); 370 } else 371 RewriteVal = AI++; 372 373 std::vector<User*> Users(inputs[i]->user_begin(), inputs[i]->user_end()); 374 for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end(); 375 use != useE; ++use) 376 if (Instruction* inst = dyn_cast<Instruction>(*use)) 377 if (Blocks.count(inst->getParent())) 378 inst->replaceUsesOfWith(inputs[i], RewriteVal); 379 } 380 381 // Set names for input and output arguments. 382 if (!AggregateArgs) { 383 AI = newFunction->arg_begin(); 384 for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI) 385 AI->setName(inputs[i]->getName()); 386 for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI) 387 AI->setName(outputs[i]->getName()+".out"); 388 } 389 390 // Rewrite branches to basic blocks outside of the loop to new dummy blocks 391 // within the new function. This must be done before we lose track of which 392 // blocks were originally in the code region. 393 std::vector<User*> Users(header->user_begin(), header->user_end()); 394 for (unsigned i = 0, e = Users.size(); i != e; ++i) 395 // The BasicBlock which contains the branch is not in the region 396 // modify the branch target to a new block 397 if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i])) 398 if (!Blocks.count(TI->getParent()) && 399 TI->getParent()->getParent() == oldFunction) 400 TI->replaceUsesOfWith(header, newHeader); 401 402 return newFunction; 403 } 404 405 /// FindPhiPredForUseInBlock - Given a value and a basic block, find a PHI 406 /// that uses the value within the basic block, and return the predecessor 407 /// block associated with that use, or return 0 if none is found. 408 static BasicBlock* FindPhiPredForUseInBlock(Value* Used, BasicBlock* BB) { 409 for (Use &U : Used->uses()) { 410 PHINode *P = dyn_cast<PHINode>(U.getUser()); 411 if (P && P->getParent() == BB) 412 return P->getIncomingBlock(U); 413 } 414 415 return nullptr; 416 } 417 418 /// emitCallAndSwitchStatement - This method sets up the caller side by adding 419 /// the call instruction, splitting any PHI nodes in the header block as 420 /// necessary. 421 void CodeExtractor:: 422 emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer, 423 ValueSet &inputs, ValueSet &outputs) { 424 // Emit a call to the new function, passing in: *pointer to struct (if 425 // aggregating parameters), or plan inputs and allocated memory for outputs 426 std::vector<Value*> params, StructValues, ReloadOutputs, Reloads; 427 428 LLVMContext &Context = newFunction->getContext(); 429 430 // Add inputs as params, or to be filled into the struct 431 for (ValueSet::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i) 432 if (AggregateArgs) 433 StructValues.push_back(*i); 434 else 435 params.push_back(*i); 436 437 // Create allocas for the outputs 438 for (ValueSet::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) { 439 if (AggregateArgs) { 440 StructValues.push_back(*i); 441 } else { 442 AllocaInst *alloca = 443 new AllocaInst((*i)->getType(), nullptr, (*i)->getName()+".loc", 444 codeReplacer->getParent()->begin()->begin()); 445 ReloadOutputs.push_back(alloca); 446 params.push_back(alloca); 447 } 448 } 449 450 AllocaInst *Struct = nullptr; 451 if (AggregateArgs && (inputs.size() + outputs.size() > 0)) { 452 std::vector<Type*> ArgTypes; 453 for (ValueSet::iterator v = StructValues.begin(), 454 ve = StructValues.end(); v != ve; ++v) 455 ArgTypes.push_back((*v)->getType()); 456 457 // Allocate a struct at the beginning of this function 458 Type *StructArgTy = StructType::get(newFunction->getContext(), ArgTypes); 459 Struct = 460 new AllocaInst(StructArgTy, nullptr, "structArg", 461 codeReplacer->getParent()->begin()->begin()); 462 params.push_back(Struct); 463 464 for (unsigned i = 0, e = inputs.size(); i != e; ++i) { 465 Value *Idx[2]; 466 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context)); 467 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), i); 468 GetElementPtrInst *GEP = 469 GetElementPtrInst::Create(Struct, Idx, 470 "gep_" + StructValues[i]->getName()); 471 codeReplacer->getInstList().push_back(GEP); 472 StoreInst *SI = new StoreInst(StructValues[i], GEP); 473 codeReplacer->getInstList().push_back(SI); 474 } 475 } 476 477 // Emit the call to the function 478 CallInst *call = CallInst::Create(newFunction, params, 479 NumExitBlocks > 1 ? "targetBlock" : ""); 480 codeReplacer->getInstList().push_back(call); 481 482 Function::arg_iterator OutputArgBegin = newFunction->arg_begin(); 483 unsigned FirstOut = inputs.size(); 484 if (!AggregateArgs) 485 std::advance(OutputArgBegin, inputs.size()); 486 487 // Reload the outputs passed in by reference 488 for (unsigned i = 0, e = outputs.size(); i != e; ++i) { 489 Value *Output = nullptr; 490 if (AggregateArgs) { 491 Value *Idx[2]; 492 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context)); 493 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), FirstOut + i); 494 GetElementPtrInst *GEP 495 = GetElementPtrInst::Create(Struct, Idx, 496 "gep_reload_" + outputs[i]->getName()); 497 codeReplacer->getInstList().push_back(GEP); 498 Output = GEP; 499 } else { 500 Output = ReloadOutputs[i]; 501 } 502 LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload"); 503 Reloads.push_back(load); 504 codeReplacer->getInstList().push_back(load); 505 std::vector<User*> Users(outputs[i]->user_begin(), outputs[i]->user_end()); 506 for (unsigned u = 0, e = Users.size(); u != e; ++u) { 507 Instruction *inst = cast<Instruction>(Users[u]); 508 if (!Blocks.count(inst->getParent())) 509 inst->replaceUsesOfWith(outputs[i], load); 510 } 511 } 512 513 // Now we can emit a switch statement using the call as a value. 514 SwitchInst *TheSwitch = 515 SwitchInst::Create(Constant::getNullValue(Type::getInt16Ty(Context)), 516 codeReplacer, 0, codeReplacer); 517 518 // Since there may be multiple exits from the original region, make the new 519 // function return an unsigned, switch on that number. This loop iterates 520 // over all of the blocks in the extracted region, updating any terminator 521 // instructions in the to-be-extracted region that branch to blocks that are 522 // not in the region to be extracted. 523 std::map<BasicBlock*, BasicBlock*> ExitBlockMap; 524 525 unsigned switchVal = 0; 526 for (SetVector<BasicBlock*>::const_iterator i = Blocks.begin(), 527 e = Blocks.end(); i != e; ++i) { 528 TerminatorInst *TI = (*i)->getTerminator(); 529 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 530 if (!Blocks.count(TI->getSuccessor(i))) { 531 BasicBlock *OldTarget = TI->getSuccessor(i); 532 // add a new basic block which returns the appropriate value 533 BasicBlock *&NewTarget = ExitBlockMap[OldTarget]; 534 if (!NewTarget) { 535 // If we don't already have an exit stub for this non-extracted 536 // destination, create one now! 537 NewTarget = BasicBlock::Create(Context, 538 OldTarget->getName() + ".exitStub", 539 newFunction); 540 unsigned SuccNum = switchVal++; 541 542 Value *brVal = nullptr; 543 switch (NumExitBlocks) { 544 case 0: 545 case 1: break; // No value needed. 546 case 2: // Conditional branch, return a bool 547 brVal = ConstantInt::get(Type::getInt1Ty(Context), !SuccNum); 548 break; 549 default: 550 brVal = ConstantInt::get(Type::getInt16Ty(Context), SuccNum); 551 break; 552 } 553 554 ReturnInst *NTRet = ReturnInst::Create(Context, brVal, NewTarget); 555 556 // Update the switch instruction. 557 TheSwitch->addCase(ConstantInt::get(Type::getInt16Ty(Context), 558 SuccNum), 559 OldTarget); 560 561 // Restore values just before we exit 562 Function::arg_iterator OAI = OutputArgBegin; 563 for (unsigned out = 0, e = outputs.size(); out != e; ++out) { 564 // For an invoke, the normal destination is the only one that is 565 // dominated by the result of the invocation 566 BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent(); 567 568 bool DominatesDef = true; 569 570 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) { 571 DefBlock = Invoke->getNormalDest(); 572 573 // Make sure we are looking at the original successor block, not 574 // at a newly inserted exit block, which won't be in the dominator 575 // info. 576 for (std::map<BasicBlock*, BasicBlock*>::iterator I = 577 ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I) 578 if (DefBlock == I->second) { 579 DefBlock = I->first; 580 break; 581 } 582 583 // In the extract block case, if the block we are extracting ends 584 // with an invoke instruction, make sure that we don't emit a 585 // store of the invoke value for the unwind block. 586 if (!DT && DefBlock != OldTarget) 587 DominatesDef = false; 588 } 589 590 if (DT) { 591 DominatesDef = DT->dominates(DefBlock, OldTarget); 592 593 // If the output value is used by a phi in the target block, 594 // then we need to test for dominance of the phi's predecessor 595 // instead. Unfortunately, this a little complicated since we 596 // have already rewritten uses of the value to uses of the reload. 597 BasicBlock* pred = FindPhiPredForUseInBlock(Reloads[out], 598 OldTarget); 599 if (pred && DT && DT->dominates(DefBlock, pred)) 600 DominatesDef = true; 601 } 602 603 if (DominatesDef) { 604 if (AggregateArgs) { 605 Value *Idx[2]; 606 Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context)); 607 Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), 608 FirstOut+out); 609 GetElementPtrInst *GEP = 610 GetElementPtrInst::Create(OAI, Idx, 611 "gep_" + outputs[out]->getName(), 612 NTRet); 613 new StoreInst(outputs[out], GEP, NTRet); 614 } else { 615 new StoreInst(outputs[out], OAI, NTRet); 616 } 617 } 618 // Advance output iterator even if we don't emit a store 619 if (!AggregateArgs) ++OAI; 620 } 621 } 622 623 // rewrite the original branch instruction with this new target 624 TI->setSuccessor(i, NewTarget); 625 } 626 } 627 628 // Now that we've done the deed, simplify the switch instruction. 629 Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType(); 630 switch (NumExitBlocks) { 631 case 0: 632 // There are no successors (the block containing the switch itself), which 633 // means that previously this was the last part of the function, and hence 634 // this should be rewritten as a `ret' 635 636 // Check if the function should return a value 637 if (OldFnRetTy->isVoidTy()) { 638 ReturnInst::Create(Context, nullptr, TheSwitch); // Return void 639 } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) { 640 // return what we have 641 ReturnInst::Create(Context, TheSwitch->getCondition(), TheSwitch); 642 } else { 643 // Otherwise we must have code extracted an unwind or something, just 644 // return whatever we want. 645 ReturnInst::Create(Context, 646 Constant::getNullValue(OldFnRetTy), TheSwitch); 647 } 648 649 TheSwitch->eraseFromParent(); 650 break; 651 case 1: 652 // Only a single destination, change the switch into an unconditional 653 // branch. 654 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch); 655 TheSwitch->eraseFromParent(); 656 break; 657 case 2: 658 BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2), 659 call, TheSwitch); 660 TheSwitch->eraseFromParent(); 661 break; 662 default: 663 // Otherwise, make the default destination of the switch instruction be one 664 // of the other successors. 665 TheSwitch->setCondition(call); 666 TheSwitch->setDefaultDest(TheSwitch->getSuccessor(NumExitBlocks)); 667 // Remove redundant case 668 TheSwitch->removeCase(SwitchInst::CaseIt(TheSwitch, NumExitBlocks-1)); 669 break; 670 } 671 } 672 673 void CodeExtractor::moveCodeToFunction(Function *newFunction) { 674 Function *oldFunc = (*Blocks.begin())->getParent(); 675 Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList(); 676 Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList(); 677 678 for (SetVector<BasicBlock*>::const_iterator i = Blocks.begin(), 679 e = Blocks.end(); i != e; ++i) { 680 // Delete the basic block from the old function, and the list of blocks 681 oldBlocks.remove(*i); 682 683 // Insert this basic block into the new function 684 newBlocks.push_back(*i); 685 } 686 } 687 688 Function *CodeExtractor::extractCodeRegion() { 689 if (!isEligible()) 690 return nullptr; 691 692 ValueSet inputs, outputs; 693 694 // Assumption: this is a single-entry code region, and the header is the first 695 // block in the region. 696 BasicBlock *header = *Blocks.begin(); 697 698 // If we have to split PHI nodes or the entry block, do so now. 699 severSplitPHINodes(header); 700 701 // If we have any return instructions in the region, split those blocks so 702 // that the return is not in the region. 703 splitReturnBlocks(); 704 705 Function *oldFunction = header->getParent(); 706 707 // This takes place of the original loop 708 BasicBlock *codeReplacer = BasicBlock::Create(header->getContext(), 709 "codeRepl", oldFunction, 710 header); 711 712 // The new function needs a root node because other nodes can branch to the 713 // head of the region, but the entry node of a function cannot have preds. 714 BasicBlock *newFuncRoot = BasicBlock::Create(header->getContext(), 715 "newFuncRoot"); 716 newFuncRoot->getInstList().push_back(BranchInst::Create(header)); 717 718 // Find inputs to, outputs from the code region. 719 findInputsOutputs(inputs, outputs); 720 721 SmallPtrSet<BasicBlock *, 1> ExitBlocks; 722 for (SetVector<BasicBlock *>::iterator I = Blocks.begin(), E = Blocks.end(); 723 I != E; ++I) 724 for (succ_iterator SI = succ_begin(*I), SE = succ_end(*I); SI != SE; ++SI) 725 if (!Blocks.count(*SI)) 726 ExitBlocks.insert(*SI); 727 NumExitBlocks = ExitBlocks.size(); 728 729 // Construct new function based on inputs/outputs & add allocas for all defs. 730 Function *newFunction = constructFunction(inputs, outputs, header, 731 newFuncRoot, 732 codeReplacer, oldFunction, 733 oldFunction->getParent()); 734 735 emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs); 736 737 moveCodeToFunction(newFunction); 738 739 // Loop over all of the PHI nodes in the header block, and change any 740 // references to the old incoming edge to be the new incoming edge. 741 for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) { 742 PHINode *PN = cast<PHINode>(I); 743 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 744 if (!Blocks.count(PN->getIncomingBlock(i))) 745 PN->setIncomingBlock(i, newFuncRoot); 746 } 747 748 // Look at all successors of the codeReplacer block. If any of these blocks 749 // had PHI nodes in them, we need to update the "from" block to be the code 750 // replacer, not the original block in the extracted region. 751 std::vector<BasicBlock*> Succs(succ_begin(codeReplacer), 752 succ_end(codeReplacer)); 753 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 754 for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) { 755 PHINode *PN = cast<PHINode>(I); 756 std::set<BasicBlock*> ProcessedPreds; 757 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 758 if (Blocks.count(PN->getIncomingBlock(i))) { 759 if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second) 760 PN->setIncomingBlock(i, codeReplacer); 761 else { 762 // There were multiple entries in the PHI for this block, now there 763 // is only one, so remove the duplicated entries. 764 PN->removeIncomingValue(i, false); 765 --i; --e; 766 } 767 } 768 } 769 770 //cerr << "NEW FUNCTION: " << *newFunction; 771 // verifyFunction(*newFunction); 772 773 // cerr << "OLD FUNCTION: " << *oldFunction; 774 // verifyFunction(*oldFunction); 775 776 DEBUG(if (verifyFunction(*newFunction)) 777 report_fatal_error("verifyFunction failed!")); 778 return newFunction; 779 } 780