JumpThreading.cpp revision e34537856a544c83513e390ac9552a8bc3823346
1//===- JumpThreading.cpp - Thread control through conditional blocks ------===// 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 Jump Threading pass. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "jump-threading" 15#include "llvm/Transforms/Scalar.h" 16#include "llvm/IntrinsicInst.h" 17#include "llvm/LLVMContext.h" 18#include "llvm/Pass.h" 19#include "llvm/Analysis/InstructionSimplify.h" 20#include "llvm/Transforms/Utils/BasicBlockUtils.h" 21#include "llvm/Transforms/Utils/Local.h" 22#include "llvm/Transforms/Utils/SSAUpdater.h" 23#include "llvm/Target/TargetData.h" 24#include "llvm/ADT/DenseMap.h" 25#include "llvm/ADT/Statistic.h" 26#include "llvm/ADT/STLExtras.h" 27#include "llvm/ADT/SmallPtrSet.h" 28#include "llvm/ADT/SmallSet.h" 29#include "llvm/Support/CommandLine.h" 30#include "llvm/Support/Debug.h" 31#include "llvm/Support/raw_ostream.h" 32using namespace llvm; 33 34STATISTIC(NumThreads, "Number of jumps threaded"); 35STATISTIC(NumFolds, "Number of terminators folded"); 36STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 37 38static cl::opt<unsigned> 39Threshold("jump-threading-threshold", 40 cl::desc("Max block size to duplicate for jump threading"), 41 cl::init(6), cl::Hidden); 42 43namespace { 44 /// This pass performs 'jump threading', which looks at blocks that have 45 /// multiple predecessors and multiple successors. If one or more of the 46 /// predecessors of the block can be proven to always jump to one of the 47 /// successors, we forward the edge from the predecessor to the successor by 48 /// duplicating the contents of this block. 49 /// 50 /// An example of when this can occur is code like this: 51 /// 52 /// if () { ... 53 /// X = 4; 54 /// } 55 /// if (X < 3) { 56 /// 57 /// In this case, the unconditional branch at the end of the first if can be 58 /// revectored to the false side of the second if. 59 /// 60 class JumpThreading : public FunctionPass { 61 TargetData *TD; 62#ifdef NDEBUG 63 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 64#else 65 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 66#endif 67 public: 68 static char ID; // Pass identification 69 JumpThreading() : FunctionPass(&ID) {} 70 71 bool runOnFunction(Function &F); 72 void FindLoopHeaders(Function &F); 73 74 bool ProcessBlock(BasicBlock *BB); 75 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 76 BasicBlock *SuccBB); 77 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 78 BasicBlock *PredBB); 79 80 typedef SmallVectorImpl<std::pair<ConstantInt*, 81 BasicBlock*> > PredValueInfo; 82 83 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 84 PredValueInfo &Result); 85 bool ProcessThreadableEdges(Instruction *CondInst, BasicBlock *BB); 86 87 88 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 89 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 90 91 bool ProcessJumpOnPHI(PHINode *PN); 92 93 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 94 }; 95} 96 97char JumpThreading::ID = 0; 98static RegisterPass<JumpThreading> 99X("jump-threading", "Jump Threading"); 100 101// Public interface to the Jump Threading pass 102FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 103 104/// runOnFunction - Top level algorithm. 105/// 106bool JumpThreading::runOnFunction(Function &F) { 107 DEBUG(errs() << "Jump threading on function '" << F.getName() << "'\n"); 108 TD = getAnalysisIfAvailable<TargetData>(); 109 110 FindLoopHeaders(F); 111 112 bool AnotherIteration = true, EverChanged = false; 113 while (AnotherIteration) { 114 AnotherIteration = false; 115 bool Changed = false; 116 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 117 BasicBlock *BB = I; 118 while (ProcessBlock(BB)) 119 Changed = true; 120 121 ++I; 122 123 // If the block is trivially dead, zap it. This eliminates the successor 124 // edges which simplifies the CFG. 125 if (pred_begin(BB) == pred_end(BB) && 126 BB != &BB->getParent()->getEntryBlock()) { 127 DEBUG(errs() << " JT: Deleting dead block '" << BB->getName() 128 << "' with terminator: " << *BB->getTerminator() << '\n'); 129 LoopHeaders.erase(BB); 130 DeleteDeadBlock(BB); 131 Changed = true; 132 } 133 } 134 AnotherIteration = Changed; 135 EverChanged |= Changed; 136 } 137 138 LoopHeaders.clear(); 139 return EverChanged; 140} 141 142/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 143/// thread across it. 144static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 145 /// Ignore PHI nodes, these will be flattened when duplication happens. 146 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 147 148 // Sum up the cost of each instruction until we get to the terminator. Don't 149 // include the terminator because the copy won't include it. 150 unsigned Size = 0; 151 for (; !isa<TerminatorInst>(I); ++I) { 152 // Debugger intrinsics don't incur code size. 153 if (isa<DbgInfoIntrinsic>(I)) continue; 154 155 // If this is a pointer->pointer bitcast, it is free. 156 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType())) 157 continue; 158 159 // All other instructions count for at least one unit. 160 ++Size; 161 162 // Calls are more expensive. If they are non-intrinsic calls, we model them 163 // as having cost of 4. If they are a non-vector intrinsic, we model them 164 // as having cost of 2 total, and if they are a vector intrinsic, we model 165 // them as having cost 1. 166 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 167 if (!isa<IntrinsicInst>(CI)) 168 Size += 3; 169 else if (!isa<VectorType>(CI->getType())) 170 Size += 1; 171 } 172 } 173 174 // Threading through a switch statement is particularly profitable. If this 175 // block ends in a switch, decrease its cost to make it more likely to happen. 176 if (isa<SwitchInst>(I)) 177 Size = Size > 6 ? Size-6 : 0; 178 179 return Size; 180} 181 182 183//===----------------------------------------------------------------------===// 184 185 186/// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this 187/// method is called when we're about to delete Pred as a predecessor of BB. If 188/// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred. 189/// 190/// Unlike the removePredecessor method, this attempts to simplify uses of PHI 191/// nodes that collapse into identity values. For example, if we have: 192/// x = phi(1, 0, 0, 0) 193/// y = and x, z 194/// 195/// .. and delete the predecessor corresponding to the '1', this will attempt to 196/// recursively fold the and to 0. 197static void RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred, 198 TargetData *TD) { 199 // This only adjusts blocks with PHI nodes. 200 if (!isa<PHINode>(BB->begin())) 201 return; 202 203 // Remove the entries for Pred from the PHI nodes in BB, but do not simplify 204 // them down. This will leave us with single entry phi nodes and other phis 205 // that can be removed. 206 BB->removePredecessor(Pred, true); 207 208 WeakVH PhiIt = &BB->front(); 209 while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) { 210 PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt)); 211 212 Value *PNV = PN->hasConstantValue(); 213 if (PNV == 0) continue; 214 215 assert(PNV != PN && "hasConstantValue broken"); 216 217 // If we're able to simplify the phi to a constant, simplify it into its 218 // uses. 219 while (!PN->use_empty()) { 220 // Update the instruction to use the new value. 221 Use &U = PN->use_begin().getUse(); 222 Instruction *User = cast<Instruction>(U.getUser()); 223 U = PNV; 224 225 // See if we can simplify it. 226 if (Value *V = SimplifyInstruction(User, TD)) { 227 User->replaceAllUsesWith(V); 228 User->eraseFromParent(); 229 } 230 } 231 232 PN->replaceAllUsesWith(PNV); 233 PN->eraseFromParent(); 234 235 // If recursive simplification ended up deleting the next PHI node we would 236 // iterate to, then our iterator is invalid, restart scanning from the top 237 // of the block. 238 if (PhiIt == 0) PhiIt = &BB->front(); 239 } 240} 241 242//===----------------------------------------------------------------------===// 243 244 245/// FindLoopHeaders - We do not want jump threading to turn proper loop 246/// structures into irreducible loops. Doing this breaks up the loop nesting 247/// hierarchy and pessimizes later transformations. To prevent this from 248/// happening, we first have to find the loop headers. Here we approximate this 249/// by finding targets of backedges in the CFG. 250/// 251/// Note that there definitely are cases when we want to allow threading of 252/// edges across a loop header. For example, threading a jump from outside the 253/// loop (the preheader) to an exit block of the loop is definitely profitable. 254/// It is also almost always profitable to thread backedges from within the loop 255/// to exit blocks, and is often profitable to thread backedges to other blocks 256/// within the loop (forming a nested loop). This simple analysis is not rich 257/// enough to track all of these properties and keep it up-to-date as the CFG 258/// mutates, so we don't allow any of these transformations. 259/// 260void JumpThreading::FindLoopHeaders(Function &F) { 261 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 262 FindFunctionBackedges(F, Edges); 263 264 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 265 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 266} 267 268/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 269/// if we can infer that the value is a known ConstantInt in any of our 270/// predecessors. If so, return the known list of value and pred BB in the 271/// result vector. If a value is known to be undef, it is returned as null. 272/// 273/// The BB basic block is known to start with a PHI node. 274/// 275/// This returns true if there were any known values. 276/// 277/// 278/// TODO: Per PR2563, we could infer value range information about a predecessor 279/// based on its terminator. 280bool JumpThreading:: 281ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 282 PHINode *TheFirstPHI = cast<PHINode>(BB->begin()); 283 284 // If V is a constantint, then it is known in all predecessors. 285 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 286 ConstantInt *CI = dyn_cast<ConstantInt>(V); 287 Result.resize(TheFirstPHI->getNumIncomingValues()); 288 for (unsigned i = 0, e = Result.size(); i != e; ++i) 289 Result[i] = std::make_pair(CI, TheFirstPHI->getIncomingBlock(i)); 290 return true; 291 } 292 293 // If V is a non-instruction value, or an instruction in a different block, 294 // then it can't be derived from a PHI. 295 Instruction *I = dyn_cast<Instruction>(V); 296 if (I == 0 || I->getParent() != BB) 297 return false; 298 299 /// If I is a PHI node, then we know the incoming values for any constants. 300 if (PHINode *PN = dyn_cast<PHINode>(I)) { 301 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 302 Value *InVal = PN->getIncomingValue(i); 303 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 304 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 305 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 306 } 307 } 308 return !Result.empty(); 309 } 310 311 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 312 313 // Handle some boolean conditions. 314 if (I->getType()->getPrimitiveSizeInBits() == 1) { 315 // X | true -> true 316 // X & false -> false 317 if (I->getOpcode() == Instruction::Or || 318 I->getOpcode() == Instruction::And) { 319 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 320 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 321 322 if (LHSVals.empty() && RHSVals.empty()) 323 return false; 324 325 ConstantInt *InterestingVal; 326 if (I->getOpcode() == Instruction::Or) 327 InterestingVal = ConstantInt::getTrue(I->getContext()); 328 else 329 InterestingVal = ConstantInt::getFalse(I->getContext()); 330 331 // Scan for the sentinel. 332 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 333 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) 334 Result.push_back(LHSVals[i]); 335 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 336 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) 337 Result.push_back(RHSVals[i]); 338 return !Result.empty(); 339 } 340 341 // TODO: Should handle the NOT form of XOR. 342 343 } 344 345 // Handle compare with phi operand, where the PHI is defined in this block. 346 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 347 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 348 if (PN && PN->getParent() == BB) { 349 // We can do this simplification if any comparisons fold to true or false. 350 // See if any do. 351 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 352 BasicBlock *PredBB = PN->getIncomingBlock(i); 353 Value *LHS = PN->getIncomingValue(i); 354 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 355 356 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS); 357 if (Res == 0) continue; 358 359 if (isa<UndefValue>(Res)) 360 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 361 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 362 Result.push_back(std::make_pair(CI, PredBB)); 363 } 364 365 return !Result.empty(); 366 } 367 368 // TODO: We could also recurse to see if we can determine constants another 369 // way. 370 } 371 return false; 372} 373 374 375 376/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 377/// in an undefined jump, decide which block is best to revector to. 378/// 379/// Since we can pick an arbitrary destination, we pick the successor with the 380/// fewest predecessors. This should reduce the in-degree of the others. 381/// 382static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 383 TerminatorInst *BBTerm = BB->getTerminator(); 384 unsigned MinSucc = 0; 385 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 386 // Compute the successor with the minimum number of predecessors. 387 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 388 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 389 TestBB = BBTerm->getSuccessor(i); 390 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 391 if (NumPreds < MinNumPreds) 392 MinSucc = i; 393 } 394 395 return MinSucc; 396} 397 398/// ProcessBlock - If there are any predecessors whose control can be threaded 399/// through to a successor, transform them now. 400bool JumpThreading::ProcessBlock(BasicBlock *BB) { 401 // If this block has a single predecessor, and if that pred has a single 402 // successor, merge the blocks. This encourages recursive jump threading 403 // because now the condition in this block can be threaded through 404 // predecessors of our predecessor block. 405 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 406 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 407 SinglePred != BB) { 408 // If SinglePred was a loop header, BB becomes one. 409 if (LoopHeaders.erase(SinglePred)) 410 LoopHeaders.insert(BB); 411 412 // Remember if SinglePred was the entry block of the function. If so, we 413 // will need to move BB back to the entry position. 414 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 415 MergeBasicBlockIntoOnlyPred(BB); 416 417 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 418 BB->moveBefore(&BB->getParent()->getEntryBlock()); 419 return true; 420 } 421 } 422 423 // Look to see if the terminator is a branch of switch, if not we can't thread 424 // it. 425 Value *Condition; 426 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 427 // Can't thread an unconditional jump. 428 if (BI->isUnconditional()) return false; 429 Condition = BI->getCondition(); 430 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 431 Condition = SI->getCondition(); 432 else 433 return false; // Must be an invoke. 434 435 // If the terminator of this block is branching on a constant, simplify the 436 // terminator to an unconditional branch. This can occur due to threading in 437 // other blocks. 438 if (isa<ConstantInt>(Condition)) { 439 DEBUG(errs() << " In block '" << BB->getName() 440 << "' folding terminator: " << *BB->getTerminator() << '\n'); 441 ++NumFolds; 442 ConstantFoldTerminator(BB); 443 return true; 444 } 445 446 // If the terminator is branching on an undef, we can pick any of the 447 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 448 if (isa<UndefValue>(Condition)) { 449 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 450 451 // Fold the branch/switch. 452 TerminatorInst *BBTerm = BB->getTerminator(); 453 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 454 if (i == BestSucc) continue; 455 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD); 456 } 457 458 DEBUG(errs() << " In block '" << BB->getName() 459 << "' folding undef terminator: " << *BBTerm << '\n'); 460 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 461 BBTerm->eraseFromParent(); 462 return true; 463 } 464 465 Instruction *CondInst = dyn_cast<Instruction>(Condition); 466 467 // If the condition is an instruction defined in another block, see if a 468 // predecessor has the same condition: 469 // br COND, BBX, BBY 470 // BBX: 471 // br COND, BBZ, BBW 472 if (!Condition->hasOneUse() && // Multiple uses. 473 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 474 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 475 if (isa<BranchInst>(BB->getTerminator())) { 476 for (; PI != E; ++PI) 477 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 478 if (PBI->isConditional() && PBI->getCondition() == Condition && 479 ProcessBranchOnDuplicateCond(*PI, BB)) 480 return true; 481 } else { 482 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 483 for (; PI != E; ++PI) 484 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator())) 485 if (PSI->getCondition() == Condition && 486 ProcessSwitchOnDuplicateCond(*PI, BB)) 487 return true; 488 } 489 } 490 491 // All the rest of our checks depend on the condition being an instruction. 492 if (CondInst == 0) 493 return false; 494 495 // See if this is a phi node in the current block. 496 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 497 if (PN->getParent() == BB) 498 return ProcessJumpOnPHI(PN); 499 500 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 501 if (!isa<PHINode>(CondCmp->getOperand(0)) || 502 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB) { 503 // If we have a comparison, loop over the predecessors to see if there is 504 // a condition with a lexically identical value. 505 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 506 for (; PI != E; ++PI) 507 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 508 if (PBI->isConditional() && *PI != BB) { 509 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 510 if (CI->getOperand(0) == CondCmp->getOperand(0) && 511 CI->getOperand(1) == CondCmp->getOperand(1) && 512 CI->getPredicate() == CondCmp->getPredicate()) { 513 // TODO: Could handle things like (x != 4) --> (x == 17) 514 if (ProcessBranchOnDuplicateCond(*PI, BB)) 515 return true; 516 } 517 } 518 } 519 } 520 } 521 522 // Check for some cases that are worth simplifying. Right now we want to look 523 // for loads that are used by a switch or by the condition for the branch. If 524 // we see one, check to see if it's partially redundant. If so, insert a PHI 525 // which can then be used to thread the values. 526 // 527 // This is particularly important because reg2mem inserts loads and stores all 528 // over the place, and this blocks jump threading if we don't zap them. 529 Value *SimplifyValue = CondInst; 530 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 531 if (isa<Constant>(CondCmp->getOperand(1))) 532 SimplifyValue = CondCmp->getOperand(0); 533 534 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 535 if (SimplifyPartiallyRedundantLoad(LI)) 536 return true; 537 538 539 // Handle a variety of cases where we are branching on something derived from 540 // a PHI node in the current block. If we can prove that any predecessors 541 // compute a predictable value based on a PHI node, thread those predecessors. 542 // 543 // We only bother doing this if the current block has a PHI node and if the 544 // conditional instruction lives in the current block. If either condition 545 // fails, this won't be a computable value anyway. 546 if (CondInst->getParent() == BB && isa<PHINode>(BB->front())) 547 if (ProcessThreadableEdges(CondInst, BB)) 548 return true; 549 550 551 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 552 // "(X == 4)" thread through this block. 553 554 return false; 555} 556 557/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 558/// block that jump on exactly the same condition. This means that we almost 559/// always know the direction of the edge in the DESTBB: 560/// PREDBB: 561/// br COND, DESTBB, BBY 562/// DESTBB: 563/// br COND, BBZ, BBW 564/// 565/// If DESTBB has multiple predecessors, we can't just constant fold the branch 566/// in DESTBB, we have to thread over it. 567bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 568 BasicBlock *BB) { 569 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 570 571 // If both successors of PredBB go to DESTBB, we don't know anything. We can 572 // fold the branch to an unconditional one, which allows other recursive 573 // simplifications. 574 bool BranchDir; 575 if (PredBI->getSuccessor(1) != BB) 576 BranchDir = true; 577 else if (PredBI->getSuccessor(0) != BB) 578 BranchDir = false; 579 else { 580 DEBUG(errs() << " In block '" << PredBB->getName() 581 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 582 ++NumFolds; 583 ConstantFoldTerminator(PredBB); 584 return true; 585 } 586 587 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 588 589 // If the dest block has one predecessor, just fix the branch condition to a 590 // constant and fold it. 591 if (BB->getSinglePredecessor()) { 592 DEBUG(errs() << " In block '" << BB->getName() 593 << "' folding condition to '" << BranchDir << "': " 594 << *BB->getTerminator() << '\n'); 595 ++NumFolds; 596 Value *OldCond = DestBI->getCondition(); 597 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 598 BranchDir)); 599 ConstantFoldTerminator(BB); 600 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 601 return true; 602 } 603 604 605 // Next, figure out which successor we are threading to. 606 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 607 608 SmallVector<BasicBlock*, 2> Preds; 609 Preds.push_back(PredBB); 610 611 // Ok, try to thread it! 612 return ThreadEdge(BB, Preds, SuccBB); 613} 614 615/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 616/// block that switch on exactly the same condition. This means that we almost 617/// always know the direction of the edge in the DESTBB: 618/// PREDBB: 619/// switch COND [... DESTBB, BBY ... ] 620/// DESTBB: 621/// switch COND [... BBZ, BBW ] 622/// 623/// Optimizing switches like this is very important, because simplifycfg builds 624/// switches out of repeated 'if' conditions. 625bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 626 BasicBlock *DestBB) { 627 // Can't thread edge to self. 628 if (PredBB == DestBB) 629 return false; 630 631 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 632 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 633 634 // There are a variety of optimizations that we can potentially do on these 635 // blocks: we order them from most to least preferable. 636 637 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 638 // directly to their destination. This does not introduce *any* code size 639 // growth. Skip debug info first. 640 BasicBlock::iterator BBI = DestBB->begin(); 641 while (isa<DbgInfoIntrinsic>(BBI)) 642 BBI++; 643 644 // FIXME: Thread if it just contains a PHI. 645 if (isa<SwitchInst>(BBI)) { 646 bool MadeChange = false; 647 // Ignore the default edge for now. 648 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 649 ConstantInt *DestVal = DestSI->getCaseValue(i); 650 BasicBlock *DestSucc = DestSI->getSuccessor(i); 651 652 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 653 // PredSI has an explicit case for it. If so, forward. If it is covered 654 // by the default case, we can't update PredSI. 655 unsigned PredCase = PredSI->findCaseValue(DestVal); 656 if (PredCase == 0) continue; 657 658 // If PredSI doesn't go to DestBB on this value, then it won't reach the 659 // case on this condition. 660 if (PredSI->getSuccessor(PredCase) != DestBB && 661 DestSI->getSuccessor(i) != DestBB) 662 continue; 663 664 // Otherwise, we're safe to make the change. Make sure that the edge from 665 // DestSI to DestSucc is not critical and has no PHI nodes. 666 DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 667 DEBUG(errs() << "THROUGH: " << *DestSI); 668 669 // If the destination has PHI nodes, just split the edge for updating 670 // simplicity. 671 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 672 SplitCriticalEdge(DestSI, i, this); 673 DestSucc = DestSI->getSuccessor(i); 674 } 675 FoldSingleEntryPHINodes(DestSucc); 676 PredSI->setSuccessor(PredCase, DestSucc); 677 MadeChange = true; 678 } 679 680 if (MadeChange) 681 return true; 682 } 683 684 return false; 685} 686 687 688/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 689/// load instruction, eliminate it by replacing it with a PHI node. This is an 690/// important optimization that encourages jump threading, and needs to be run 691/// interlaced with other jump threading tasks. 692bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 693 // Don't hack volatile loads. 694 if (LI->isVolatile()) return false; 695 696 // If the load is defined in a block with exactly one predecessor, it can't be 697 // partially redundant. 698 BasicBlock *LoadBB = LI->getParent(); 699 if (LoadBB->getSinglePredecessor()) 700 return false; 701 702 Value *LoadedPtr = LI->getOperand(0); 703 704 // If the loaded operand is defined in the LoadBB, it can't be available. 705 // FIXME: Could do PHI translation, that would be fun :) 706 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 707 if (PtrOp->getParent() == LoadBB) 708 return false; 709 710 // Scan a few instructions up from the load, to see if it is obviously live at 711 // the entry to its block. 712 BasicBlock::iterator BBIt = LI; 713 714 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB, 715 BBIt, 6)) { 716 // If the value if the load is locally available within the block, just use 717 // it. This frequently occurs for reg2mem'd allocas. 718 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 719 720 // If the returned value is the load itself, replace with an undef. This can 721 // only happen in dead loops. 722 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 723 LI->replaceAllUsesWith(AvailableVal); 724 LI->eraseFromParent(); 725 return true; 726 } 727 728 // Otherwise, if we scanned the whole block and got to the top of the block, 729 // we know the block is locally transparent to the load. If not, something 730 // might clobber its value. 731 if (BBIt != LoadBB->begin()) 732 return false; 733 734 735 SmallPtrSet<BasicBlock*, 8> PredsScanned; 736 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 737 AvailablePredsTy AvailablePreds; 738 BasicBlock *OneUnavailablePred = 0; 739 740 // If we got here, the loaded value is transparent through to the start of the 741 // block. Check to see if it is available in any of the predecessor blocks. 742 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 743 PI != PE; ++PI) { 744 BasicBlock *PredBB = *PI; 745 746 // If we already scanned this predecessor, skip it. 747 if (!PredsScanned.insert(PredBB)) 748 continue; 749 750 // Scan the predecessor to see if the value is available in the pred. 751 BBIt = PredBB->end(); 752 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 753 if (!PredAvailable) { 754 OneUnavailablePred = PredBB; 755 continue; 756 } 757 758 // If so, this load is partially redundant. Remember this info so that we 759 // can create a PHI node. 760 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 761 } 762 763 // If the loaded value isn't available in any predecessor, it isn't partially 764 // redundant. 765 if (AvailablePreds.empty()) return false; 766 767 // Okay, the loaded value is available in at least one (and maybe all!) 768 // predecessors. If the value is unavailable in more than one unique 769 // predecessor, we want to insert a merge block for those common predecessors. 770 // This ensures that we only have to insert one reload, thus not increasing 771 // code size. 772 BasicBlock *UnavailablePred = 0; 773 774 // If there is exactly one predecessor where the value is unavailable, the 775 // already computed 'OneUnavailablePred' block is it. If it ends in an 776 // unconditional branch, we know that it isn't a critical edge. 777 if (PredsScanned.size() == AvailablePreds.size()+1 && 778 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 779 UnavailablePred = OneUnavailablePred; 780 } else if (PredsScanned.size() != AvailablePreds.size()) { 781 // Otherwise, we had multiple unavailable predecessors or we had a critical 782 // edge from the one. 783 SmallVector<BasicBlock*, 8> PredsToSplit; 784 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 785 786 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 787 AvailablePredSet.insert(AvailablePreds[i].first); 788 789 // Add all the unavailable predecessors to the PredsToSplit list. 790 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 791 PI != PE; ++PI) 792 if (!AvailablePredSet.count(*PI)) 793 PredsToSplit.push_back(*PI); 794 795 // Split them out to their own block. 796 UnavailablePred = 797 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 798 "thread-split", this); 799 } 800 801 // If the value isn't available in all predecessors, then there will be 802 // exactly one where it isn't available. Insert a load on that edge and add 803 // it to the AvailablePreds list. 804 if (UnavailablePred) { 805 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 806 "Can't handle critical edge here!"); 807 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", 808 UnavailablePred->getTerminator()); 809 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 810 } 811 812 // Now we know that each predecessor of this block has a value in 813 // AvailablePreds, sort them for efficient access as we're walking the preds. 814 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 815 816 // Create a PHI node at the start of the block for the PRE'd load value. 817 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 818 PN->takeName(LI); 819 820 // Insert new entries into the PHI for each predecessor. A single block may 821 // have multiple entries here. 822 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 823 ++PI) { 824 AvailablePredsTy::iterator I = 825 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 826 std::make_pair(*PI, (Value*)0)); 827 828 assert(I != AvailablePreds.end() && I->first == *PI && 829 "Didn't find entry for predecessor!"); 830 831 PN->addIncoming(I->second, I->first); 832 } 833 834 //cerr << "PRE: " << *LI << *PN << "\n"; 835 836 LI->replaceAllUsesWith(PN); 837 LI->eraseFromParent(); 838 839 return true; 840} 841 842/// FindMostPopularDest - The specified list contains multiple possible 843/// threadable destinations. Pick the one that occurs the most frequently in 844/// the list. 845static BasicBlock * 846FindMostPopularDest(BasicBlock *BB, 847 const SmallVectorImpl<std::pair<BasicBlock*, 848 BasicBlock*> > &PredToDestList) { 849 assert(!PredToDestList.empty()); 850 851 // Determine popularity. If there are multiple possible destinations, we 852 // explicitly choose to ignore 'undef' destinations. We prefer to thread 853 // blocks with known and real destinations to threading undef. We'll handle 854 // them later if interesting. 855 DenseMap<BasicBlock*, unsigned> DestPopularity; 856 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 857 if (PredToDestList[i].second) 858 DestPopularity[PredToDestList[i].second]++; 859 860 // Find the most popular dest. 861 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 862 BasicBlock *MostPopularDest = DPI->first; 863 unsigned Popularity = DPI->second; 864 SmallVector<BasicBlock*, 4> SamePopularity; 865 866 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 867 // If the popularity of this entry isn't higher than the popularity we've 868 // seen so far, ignore it. 869 if (DPI->second < Popularity) 870 ; // ignore. 871 else if (DPI->second == Popularity) { 872 // If it is the same as what we've seen so far, keep track of it. 873 SamePopularity.push_back(DPI->first); 874 } else { 875 // If it is more popular, remember it. 876 SamePopularity.clear(); 877 MostPopularDest = DPI->first; 878 Popularity = DPI->second; 879 } 880 } 881 882 // Okay, now we know the most popular destination. If there is more than 883 // destination, we need to determine one. This is arbitrary, but we need 884 // to make a deterministic decision. Pick the first one that appears in the 885 // successor list. 886 if (!SamePopularity.empty()) { 887 SamePopularity.push_back(MostPopularDest); 888 TerminatorInst *TI = BB->getTerminator(); 889 for (unsigned i = 0; ; ++i) { 890 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 891 892 if (std::find(SamePopularity.begin(), SamePopularity.end(), 893 TI->getSuccessor(i)) == SamePopularity.end()) 894 continue; 895 896 MostPopularDest = TI->getSuccessor(i); 897 break; 898 } 899 } 900 901 // Okay, we have finally picked the most popular destination. 902 return MostPopularDest; 903} 904 905bool JumpThreading::ProcessThreadableEdges(Instruction *CondInst, 906 BasicBlock *BB) { 907 // If threading this would thread across a loop header, don't even try to 908 // thread the edge. 909 if (LoopHeaders.count(BB)) 910 return false; 911 912 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 913 if (!ComputeValueKnownInPredecessors(CondInst, BB, PredValues)) 914 return false; 915 assert(!PredValues.empty() && 916 "ComputeValueKnownInPredecessors returned true with no values"); 917 918 DEBUG(errs() << "IN BB: " << *BB; 919 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 920 errs() << " BB '" << BB->getName() << "': FOUND condition = "; 921 if (PredValues[i].first) 922 errs() << *PredValues[i].first; 923 else 924 errs() << "UNDEF"; 925 errs() << " for pred '" << PredValues[i].second->getName() 926 << "'.\n"; 927 }); 928 929 // Decide what we want to thread through. Convert our list of known values to 930 // a list of known destinations for each pred. This also discards duplicate 931 // predecessors and keeps track of the undefined inputs (which are represented 932 // as a null dest in the PredToDestList). 933 SmallPtrSet<BasicBlock*, 16> SeenPreds; 934 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 935 936 BasicBlock *OnlyDest = 0; 937 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 938 939 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 940 BasicBlock *Pred = PredValues[i].second; 941 if (!SeenPreds.insert(Pred)) 942 continue; // Duplicate predecessor entry. 943 944 // If the predecessor ends with an indirect goto, we can't change its 945 // destination. 946 if (isa<IndirectBrInst>(Pred->getTerminator())) 947 continue; 948 949 ConstantInt *Val = PredValues[i].first; 950 951 BasicBlock *DestBB; 952 if (Val == 0) // Undef. 953 DestBB = 0; 954 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 955 DestBB = BI->getSuccessor(Val->isZero()); 956 else { 957 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 958 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 959 } 960 961 // If we have exactly one destination, remember it for efficiency below. 962 if (i == 0) 963 OnlyDest = DestBB; 964 else if (OnlyDest != DestBB) 965 OnlyDest = MultipleDestSentinel; 966 967 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 968 } 969 970 // If all edges were unthreadable, we fail. 971 if (PredToDestList.empty()) 972 return false; 973 974 // Determine which is the most common successor. If we have many inputs and 975 // this block is a switch, we want to start by threading the batch that goes 976 // to the most popular destination first. If we only know about one 977 // threadable destination (the common case) we can avoid this. 978 BasicBlock *MostPopularDest = OnlyDest; 979 980 if (MostPopularDest == MultipleDestSentinel) 981 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 982 983 // Now that we know what the most popular destination is, factor all 984 // predecessors that will jump to it into a single predecessor. 985 SmallVector<BasicBlock*, 16> PredsToFactor; 986 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 987 if (PredToDestList[i].second == MostPopularDest) { 988 BasicBlock *Pred = PredToDestList[i].first; 989 990 // This predecessor may be a switch or something else that has multiple 991 // edges to the block. Factor each of these edges by listing them 992 // according to # occurrences in PredsToFactor. 993 TerminatorInst *PredTI = Pred->getTerminator(); 994 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 995 if (PredTI->getSuccessor(i) == BB) 996 PredsToFactor.push_back(Pred); 997 } 998 999 // If the threadable edges are branching on an undefined value, we get to pick 1000 // the destination that these predecessors should get to. 1001 if (MostPopularDest == 0) 1002 MostPopularDest = BB->getTerminator()-> 1003 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1004 1005 // Ok, try to thread it! 1006 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1007} 1008 1009/// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in 1010/// the current block. See if there are any simplifications we can do based on 1011/// inputs to the phi node. 1012/// 1013bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) { 1014 BasicBlock *BB = PN->getParent(); 1015 1016 // If any of the predecessor blocks end in an unconditional branch, we can 1017 // *duplicate* the jump into that block in order to further encourage jump 1018 // threading and to eliminate cases where we have branch on a phi of an icmp 1019 // (branch on icmp is much better). 1020 1021 // We don't want to do this tranformation for switches, because we don't 1022 // really want to duplicate a switch. 1023 if (isa<SwitchInst>(BB->getTerminator())) 1024 return false; 1025 1026 // Look for unconditional branch predecessors. 1027 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1028 BasicBlock *PredBB = PN->getIncomingBlock(i); 1029 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1030 if (PredBr->isUnconditional() && 1031 // Try to duplicate BB into PredBB. 1032 DuplicateCondBranchOnPHIIntoPred(BB, PredBB)) 1033 return true; 1034 } 1035 1036 return false; 1037} 1038 1039 1040/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1041/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1042/// NewPred using the entries from OldPred (suitably mapped). 1043static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1044 BasicBlock *OldPred, 1045 BasicBlock *NewPred, 1046 DenseMap<Instruction*, Value*> &ValueMap) { 1047 for (BasicBlock::iterator PNI = PHIBB->begin(); 1048 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1049 // Ok, we have a PHI node. Figure out what the incoming value was for the 1050 // DestBlock. 1051 Value *IV = PN->getIncomingValueForBlock(OldPred); 1052 1053 // Remap the value if necessary. 1054 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1055 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1056 if (I != ValueMap.end()) 1057 IV = I->second; 1058 } 1059 1060 PN->addIncoming(IV, NewPred); 1061 } 1062} 1063 1064/// ThreadEdge - We have decided that it is safe and profitable to factor the 1065/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1066/// across BB. Transform the IR to reflect this change. 1067bool JumpThreading::ThreadEdge(BasicBlock *BB, 1068 const SmallVectorImpl<BasicBlock*> &PredBBs, 1069 BasicBlock *SuccBB) { 1070 // If threading to the same block as we come from, we would infinite loop. 1071 if (SuccBB == BB) { 1072 DEBUG(errs() << " Not threading across BB '" << BB->getName() 1073 << "' - would thread to self!\n"); 1074 return false; 1075 } 1076 1077 // If threading this would thread across a loop header, don't thread the edge. 1078 // See the comments above FindLoopHeaders for justifications and caveats. 1079 if (LoopHeaders.count(BB)) { 1080 DEBUG(errs() << " Not threading across loop header BB '" << BB->getName() 1081 << "' to dest BB '" << SuccBB->getName() 1082 << "' - it might create an irreducible loop!\n"); 1083 return false; 1084 } 1085 1086 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1087 if (JumpThreadCost > Threshold) { 1088 DEBUG(errs() << " Not threading BB '" << BB->getName() 1089 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1090 return false; 1091 } 1092 1093 // And finally, do it! Start by factoring the predecessors is needed. 1094 BasicBlock *PredBB; 1095 if (PredBBs.size() == 1) 1096 PredBB = PredBBs[0]; 1097 else { 1098 DEBUG(errs() << " Factoring out " << PredBBs.size() 1099 << " common predecessors.\n"); 1100 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1101 ".thr_comm", this); 1102 } 1103 1104 // And finally, do it! 1105 DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '" 1106 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1107 << ", across block:\n " 1108 << *BB << "\n"); 1109 1110 // We are going to have to map operands from the original BB block to the new 1111 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1112 // account for entry from PredBB. 1113 DenseMap<Instruction*, Value*> ValueMapping; 1114 1115 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1116 BB->getName()+".thread", 1117 BB->getParent(), BB); 1118 NewBB->moveAfter(PredBB); 1119 1120 BasicBlock::iterator BI = BB->begin(); 1121 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1122 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1123 1124 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1125 // mapping and using it to remap operands in the cloned instructions. 1126 for (; !isa<TerminatorInst>(BI); ++BI) { 1127 Instruction *New = BI->clone(); 1128 New->setName(BI->getName()); 1129 NewBB->getInstList().push_back(New); 1130 ValueMapping[BI] = New; 1131 1132 // Remap operands to patch up intra-block references. 1133 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1134 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1135 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1136 if (I != ValueMapping.end()) 1137 New->setOperand(i, I->second); 1138 } 1139 } 1140 1141 // We didn't copy the terminator from BB over to NewBB, because there is now 1142 // an unconditional jump to SuccBB. Insert the unconditional jump. 1143 BranchInst::Create(SuccBB, NewBB); 1144 1145 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1146 // PHI nodes for NewBB now. 1147 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1148 1149 // If there were values defined in BB that are used outside the block, then we 1150 // now have to update all uses of the value to use either the original value, 1151 // the cloned value, or some PHI derived value. This can require arbitrary 1152 // PHI insertion, of which we are prepared to do, clean these up now. 1153 SSAUpdater SSAUpdate; 1154 SmallVector<Use*, 16> UsesToRename; 1155 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1156 // Scan all uses of this instruction to see if it is used outside of its 1157 // block, and if so, record them in UsesToRename. 1158 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1159 ++UI) { 1160 Instruction *User = cast<Instruction>(*UI); 1161 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1162 if (UserPN->getIncomingBlock(UI) == BB) 1163 continue; 1164 } else if (User->getParent() == BB) 1165 continue; 1166 1167 UsesToRename.push_back(&UI.getUse()); 1168 } 1169 1170 // If there are no uses outside the block, we're done with this instruction. 1171 if (UsesToRename.empty()) 1172 continue; 1173 1174 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1175 1176 // We found a use of I outside of BB. Rename all uses of I that are outside 1177 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1178 // with the two values we know. 1179 SSAUpdate.Initialize(I); 1180 SSAUpdate.AddAvailableValue(BB, I); 1181 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1182 1183 while (!UsesToRename.empty()) 1184 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1185 DEBUG(errs() << "\n"); 1186 } 1187 1188 1189 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1190 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1191 // us to simplify any PHI nodes in BB. 1192 TerminatorInst *PredTerm = PredBB->getTerminator(); 1193 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1194 if (PredTerm->getSuccessor(i) == BB) { 1195 RemovePredecessorAndSimplify(BB, PredBB, TD); 1196 PredTerm->setSuccessor(i, NewBB); 1197 } 1198 1199 // At this point, the IR is fully up to date and consistent. Do a quick scan 1200 // over the new instructions and zap any that are constants or dead. This 1201 // frequently happens because of phi translation. 1202 BI = NewBB->begin(); 1203 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) { 1204 Instruction *Inst = BI++; 1205 if (Value *V = SimplifyInstruction(Inst, TD)) { 1206 Inst->replaceAllUsesWith(V); 1207 Inst->eraseFromParent(); 1208 continue; 1209 } 1210 1211 RecursivelyDeleteTriviallyDeadInstructions(Inst); 1212 } 1213 1214 // Threaded an edge! 1215 ++NumThreads; 1216 return true; 1217} 1218 1219/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1220/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1221/// If we can duplicate the contents of BB up into PredBB do so now, this 1222/// improves the odds that the branch will be on an analyzable instruction like 1223/// a compare. 1224bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1225 BasicBlock *PredBB) { 1226 // If BB is a loop header, then duplicating this block outside the loop would 1227 // cause us to transform this into an irreducible loop, don't do this. 1228 // See the comments above FindLoopHeaders for justifications and caveats. 1229 if (LoopHeaders.count(BB)) { 1230 DEBUG(errs() << " Not duplicating loop header '" << BB->getName() 1231 << "' into predecessor block '" << PredBB->getName() 1232 << "' - it might create an irreducible loop!\n"); 1233 return false; 1234 } 1235 1236 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1237 if (DuplicationCost > Threshold) { 1238 DEBUG(errs() << " Not duplicating BB '" << BB->getName() 1239 << "' - Cost is too high: " << DuplicationCost << "\n"); 1240 return false; 1241 } 1242 1243 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1244 // of PredBB. 1245 DEBUG(errs() << " Duplicating block '" << BB->getName() << "' into end of '" 1246 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1247 << DuplicationCost << " block is:" << *BB << "\n"); 1248 1249 // We are going to have to map operands from the original BB block into the 1250 // PredBB block. Evaluate PHI nodes in BB. 1251 DenseMap<Instruction*, Value*> ValueMapping; 1252 1253 BasicBlock::iterator BI = BB->begin(); 1254 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1255 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1256 1257 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1258 1259 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1260 // mapping and using it to remap operands in the cloned instructions. 1261 for (; BI != BB->end(); ++BI) { 1262 Instruction *New = BI->clone(); 1263 New->setName(BI->getName()); 1264 PredBB->getInstList().insert(OldPredBranch, New); 1265 ValueMapping[BI] = New; 1266 1267 // Remap operands to patch up intra-block references. 1268 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1269 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1270 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1271 if (I != ValueMapping.end()) 1272 New->setOperand(i, I->second); 1273 } 1274 } 1275 1276 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1277 // add entries to the PHI nodes for branch from PredBB now. 1278 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1279 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1280 ValueMapping); 1281 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1282 ValueMapping); 1283 1284 // If there were values defined in BB that are used outside the block, then we 1285 // now have to update all uses of the value to use either the original value, 1286 // the cloned value, or some PHI derived value. This can require arbitrary 1287 // PHI insertion, of which we are prepared to do, clean these up now. 1288 SSAUpdater SSAUpdate; 1289 SmallVector<Use*, 16> UsesToRename; 1290 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1291 // Scan all uses of this instruction to see if it is used outside of its 1292 // block, and if so, record them in UsesToRename. 1293 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1294 ++UI) { 1295 Instruction *User = cast<Instruction>(*UI); 1296 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1297 if (UserPN->getIncomingBlock(UI) == BB) 1298 continue; 1299 } else if (User->getParent() == BB) 1300 continue; 1301 1302 UsesToRename.push_back(&UI.getUse()); 1303 } 1304 1305 // If there are no uses outside the block, we're done with this instruction. 1306 if (UsesToRename.empty()) 1307 continue; 1308 1309 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1310 1311 // We found a use of I outside of BB. Rename all uses of I that are outside 1312 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1313 // with the two values we know. 1314 SSAUpdate.Initialize(I); 1315 SSAUpdate.AddAvailableValue(BB, I); 1316 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1317 1318 while (!UsesToRename.empty()) 1319 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1320 DEBUG(errs() << "\n"); 1321 } 1322 1323 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1324 // that we nuked. 1325 RemovePredecessorAndSimplify(BB, PredBB, TD); 1326 1327 // Remove the unconditional branch at the end of the PredBB block. 1328 OldPredBranch->eraseFromParent(); 1329 1330 ++NumDupes; 1331 return true; 1332} 1333 1334 1335