JumpThreading.cpp revision e7e63fe9650fc01043c96e2bf8a1ecc19e49c5b7
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/ConstantFolding.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/// FindLoopHeaders - We do not want jump threading to turn proper loop 185/// structures into irreducible loops. Doing this breaks up the loop nesting 186/// hierarchy and pessimizes later transformations. To prevent this from 187/// happening, we first have to find the loop headers. Here we approximate this 188/// by finding targets of backedges in the CFG. 189/// 190/// Note that there definitely are cases when we want to allow threading of 191/// edges across a loop header. For example, threading a jump from outside the 192/// loop (the preheader) to an exit block of the loop is definitely profitable. 193/// It is also almost always profitable to thread backedges from within the loop 194/// to exit blocks, and is often profitable to thread backedges to other blocks 195/// within the loop (forming a nested loop). This simple analysis is not rich 196/// enough to track all of these properties and keep it up-to-date as the CFG 197/// mutates, so we don't allow any of these transformations. 198/// 199void JumpThreading::FindLoopHeaders(Function &F) { 200 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 201 FindFunctionBackedges(F, Edges); 202 203 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 204 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 205} 206 207/// GetResultOfComparison - Given an icmp/fcmp predicate and the left and right 208/// hand sides of the compare instruction, try to determine the result. If the 209/// result can not be determined, a null pointer is returned. 210static Constant *GetResultOfComparison(CmpInst::Predicate pred, 211 Value *LHS, Value *RHS) { 212 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 213 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 214 return ConstantExpr::getCompare(pred, CLHS, CRHS); 215 216 if (LHS == RHS) 217 if (isa<IntegerType>(LHS->getType()) || isa<PointerType>(LHS->getType())) { 218 if (ICmpInst::isTrueWhenEqual(pred)) 219 return ConstantInt::getTrue(LHS->getContext()); 220 else 221 return ConstantInt::getFalse(LHS->getContext()); 222 } 223 return 0; 224} 225 226 227/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 228/// if we can infer that the value is a known ConstantInt in any of our 229/// predecessors. If so, return the known list of value and pred BB in the 230/// result vector. If a value is known to be undef, it is returned as null. 231/// 232/// The BB basic block is known to start with a PHI node. 233/// 234/// This returns true if there were any known values. 235/// 236/// 237/// TODO: Per PR2563, we could infer value range information about a predecessor 238/// based on its terminator. 239bool JumpThreading:: 240ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 241 PHINode *TheFirstPHI = cast<PHINode>(BB->begin()); 242 243 // If V is a constantint, then it is known in all predecessors. 244 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 245 ConstantInt *CI = dyn_cast<ConstantInt>(V); 246 Result.resize(TheFirstPHI->getNumIncomingValues()); 247 for (unsigned i = 0, e = Result.size(); i != e; ++i) 248 Result[i] = std::make_pair(CI, TheFirstPHI->getIncomingBlock(i)); 249 return true; 250 } 251 252 // If V is a non-instruction value, or an instruction in a different block, 253 // then it can't be derived from a PHI. 254 Instruction *I = dyn_cast<Instruction>(V); 255 if (I == 0 || I->getParent() != BB) 256 return false; 257 258 /// If I is a PHI node, then we know the incoming values for any constants. 259 if (PHINode *PN = dyn_cast<PHINode>(I)) { 260 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 261 Value *InVal = PN->getIncomingValue(i); 262 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 263 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 264 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 265 } 266 } 267 return !Result.empty(); 268 } 269 270 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 271 272 // Handle some boolean conditions. 273 if (I->getType()->getPrimitiveSizeInBits() == 1) { 274 // X | true -> true 275 // X & false -> false 276 if (I->getOpcode() == Instruction::Or || 277 I->getOpcode() == Instruction::And) { 278 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 279 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 280 281 if (LHSVals.empty() && RHSVals.empty()) 282 return false; 283 284 ConstantInt *InterestingVal; 285 if (I->getOpcode() == Instruction::Or) 286 InterestingVal = ConstantInt::getTrue(I->getContext()); 287 else 288 InterestingVal = ConstantInt::getFalse(I->getContext()); 289 290 // Scan for the sentinel. 291 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 292 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) 293 Result.push_back(LHSVals[i]); 294 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 295 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) 296 Result.push_back(RHSVals[i]); 297 return !Result.empty(); 298 } 299 300 // TODO: Should handle the NOT form of XOR. 301 302 } 303 304 // Handle compare with phi operand, where the PHI is defined in this block. 305 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 306 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 307 if (PN && PN->getParent() == BB) { 308 // We can do this simplification if any comparisons fold to true or false. 309 // See if any do. 310 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 311 BasicBlock *PredBB = PN->getIncomingBlock(i); 312 Value *LHS = PN->getIncomingValue(i); 313 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 314 315 Constant *Res = GetResultOfComparison(Cmp->getPredicate(), LHS, RHS); 316 if (Res == 0) continue; 317 318 if (isa<UndefValue>(Res)) 319 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 320 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 321 Result.push_back(std::make_pair(CI, PredBB)); 322 } 323 324 return !Result.empty(); 325 } 326 327 // TODO: We could also recurse to see if we can determine constants another 328 // way. 329 } 330 return false; 331} 332 333 334 335/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 336/// in an undefined jump, decide which block is best to revector to. 337/// 338/// Since we can pick an arbitrary destination, we pick the successor with the 339/// fewest predecessors. This should reduce the in-degree of the others. 340/// 341static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 342 TerminatorInst *BBTerm = BB->getTerminator(); 343 unsigned MinSucc = 0; 344 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 345 // Compute the successor with the minimum number of predecessors. 346 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 347 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 348 TestBB = BBTerm->getSuccessor(i); 349 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 350 if (NumPreds < MinNumPreds) 351 MinSucc = i; 352 } 353 354 return MinSucc; 355} 356 357/// ProcessBlock - If there are any predecessors whose control can be threaded 358/// through to a successor, transform them now. 359bool JumpThreading::ProcessBlock(BasicBlock *BB) { 360 // If this block has a single predecessor, and if that pred has a single 361 // successor, merge the blocks. This encourages recursive jump threading 362 // because now the condition in this block can be threaded through 363 // predecessors of our predecessor block. 364 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 365 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 366 SinglePred != BB) { 367 // If SinglePred was a loop header, BB becomes one. 368 if (LoopHeaders.erase(SinglePred)) 369 LoopHeaders.insert(BB); 370 371 // Remember if SinglePred was the entry block of the function. If so, we 372 // will need to move BB back to the entry position. 373 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 374 MergeBasicBlockIntoOnlyPred(BB); 375 376 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 377 BB->moveBefore(&BB->getParent()->getEntryBlock()); 378 return true; 379 } 380 } 381 382 // Look to see if the terminator is a branch of switch, if not we can't thread 383 // it. 384 Value *Condition; 385 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 386 // Can't thread an unconditional jump. 387 if (BI->isUnconditional()) return false; 388 Condition = BI->getCondition(); 389 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 390 Condition = SI->getCondition(); 391 else 392 return false; // Must be an invoke. 393 394 // If the terminator of this block is branching on a constant, simplify the 395 // terminator to an unconditional branch. This can occur due to threading in 396 // other blocks. 397 if (isa<ConstantInt>(Condition)) { 398 DEBUG(errs() << " In block '" << BB->getName() 399 << "' folding terminator: " << *BB->getTerminator() << '\n'); 400 ++NumFolds; 401 ConstantFoldTerminator(BB); 402 return true; 403 } 404 405 // If the terminator is branching on an undef, we can pick any of the 406 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 407 if (isa<UndefValue>(Condition)) { 408 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 409 410 // Fold the branch/switch. 411 TerminatorInst *BBTerm = BB->getTerminator(); 412 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 413 if (i == BestSucc) continue; 414 BBTerm->getSuccessor(i)->removePredecessor(BB); 415 } 416 417 DEBUG(errs() << " In block '" << BB->getName() 418 << "' folding undef terminator: " << *BBTerm << '\n'); 419 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 420 BBTerm->eraseFromParent(); 421 return true; 422 } 423 424 Instruction *CondInst = dyn_cast<Instruction>(Condition); 425 426 // If the condition is an instruction defined in another block, see if a 427 // predecessor has the same condition: 428 // br COND, BBX, BBY 429 // BBX: 430 // br COND, BBZ, BBW 431 if (!Condition->hasOneUse() && // Multiple uses. 432 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 433 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 434 if (isa<BranchInst>(BB->getTerminator())) { 435 for (; PI != E; ++PI) 436 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 437 if (PBI->isConditional() && PBI->getCondition() == Condition && 438 ProcessBranchOnDuplicateCond(*PI, BB)) 439 return true; 440 } else { 441 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 442 for (; PI != E; ++PI) 443 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator())) 444 if (PSI->getCondition() == Condition && 445 ProcessSwitchOnDuplicateCond(*PI, BB)) 446 return true; 447 } 448 } 449 450 // All the rest of our checks depend on the condition being an instruction. 451 if (CondInst == 0) 452 return false; 453 454 // See if this is a phi node in the current block. 455 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 456 if (PN->getParent() == BB) 457 return ProcessJumpOnPHI(PN); 458 459 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 460 if (!isa<PHINode>(CondCmp->getOperand(0)) || 461 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB) { 462 // If we have a comparison, loop over the predecessors to see if there is 463 // a condition with a lexically identical value. 464 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 465 for (; PI != E; ++PI) 466 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 467 if (PBI->isConditional() && *PI != BB) { 468 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 469 if (CI->getOperand(0) == CondCmp->getOperand(0) && 470 CI->getOperand(1) == CondCmp->getOperand(1) && 471 CI->getPredicate() == CondCmp->getPredicate()) { 472 // TODO: Could handle things like (x != 4) --> (x == 17) 473 if (ProcessBranchOnDuplicateCond(*PI, BB)) 474 return true; 475 } 476 } 477 } 478 } 479 } 480 481 // Check for some cases that are worth simplifying. Right now we want to look 482 // for loads that are used by a switch or by the condition for the branch. If 483 // we see one, check to see if it's partially redundant. If so, insert a PHI 484 // which can then be used to thread the values. 485 // 486 // This is particularly important because reg2mem inserts loads and stores all 487 // over the place, and this blocks jump threading if we don't zap them. 488 Value *SimplifyValue = CondInst; 489 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 490 if (isa<Constant>(CondCmp->getOperand(1))) 491 SimplifyValue = CondCmp->getOperand(0); 492 493 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 494 if (SimplifyPartiallyRedundantLoad(LI)) 495 return true; 496 497 498 // Handle a variety of cases where we are branching on something derived from 499 // a PHI node in the current block. If we can prove that any predecessors 500 // compute a predictable value based on a PHI node, thread those predecessors. 501 // 502 // We only bother doing this if the current block has a PHI node and if the 503 // conditional instruction lives in the current block. If either condition 504 // fails, this won't be a computable value anyway. 505 if (CondInst->getParent() == BB && isa<PHINode>(BB->front())) 506 if (ProcessThreadableEdges(CondInst, BB)) 507 return true; 508 509 510 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 511 // "(X == 4)" thread through this block. 512 513 return false; 514} 515 516/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 517/// block that jump on exactly the same condition. This means that we almost 518/// always know the direction of the edge in the DESTBB: 519/// PREDBB: 520/// br COND, DESTBB, BBY 521/// DESTBB: 522/// br COND, BBZ, BBW 523/// 524/// If DESTBB has multiple predecessors, we can't just constant fold the branch 525/// in DESTBB, we have to thread over it. 526bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 527 BasicBlock *BB) { 528 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 529 530 // If both successors of PredBB go to DESTBB, we don't know anything. We can 531 // fold the branch to an unconditional one, which allows other recursive 532 // simplifications. 533 bool BranchDir; 534 if (PredBI->getSuccessor(1) != BB) 535 BranchDir = true; 536 else if (PredBI->getSuccessor(0) != BB) 537 BranchDir = false; 538 else { 539 DEBUG(errs() << " In block '" << PredBB->getName() 540 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 541 ++NumFolds; 542 ConstantFoldTerminator(PredBB); 543 return true; 544 } 545 546 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 547 548 // If the dest block has one predecessor, just fix the branch condition to a 549 // constant and fold it. 550 if (BB->getSinglePredecessor()) { 551 DEBUG(errs() << " In block '" << BB->getName() 552 << "' folding condition to '" << BranchDir << "': " 553 << *BB->getTerminator() << '\n'); 554 ++NumFolds; 555 Value *OldCond = DestBI->getCondition(); 556 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 557 BranchDir)); 558 ConstantFoldTerminator(BB); 559 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 560 return true; 561 } 562 563 564 // Next, figure out which successor we are threading to. 565 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 566 567 SmallVector<BasicBlock*, 2> Preds; 568 Preds.push_back(PredBB); 569 570 // Ok, try to thread it! 571 return ThreadEdge(BB, Preds, SuccBB); 572} 573 574/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 575/// block that switch on exactly the same condition. This means that we almost 576/// always know the direction of the edge in the DESTBB: 577/// PREDBB: 578/// switch COND [... DESTBB, BBY ... ] 579/// DESTBB: 580/// switch COND [... BBZ, BBW ] 581/// 582/// Optimizing switches like this is very important, because simplifycfg builds 583/// switches out of repeated 'if' conditions. 584bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 585 BasicBlock *DestBB) { 586 // Can't thread edge to self. 587 if (PredBB == DestBB) 588 return false; 589 590 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 591 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 592 593 // There are a variety of optimizations that we can potentially do on these 594 // blocks: we order them from most to least preferable. 595 596 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 597 // directly to their destination. This does not introduce *any* code size 598 // growth. Skip debug info first. 599 BasicBlock::iterator BBI = DestBB->begin(); 600 while (isa<DbgInfoIntrinsic>(BBI)) 601 BBI++; 602 603 // FIXME: Thread if it just contains a PHI. 604 if (isa<SwitchInst>(BBI)) { 605 bool MadeChange = false; 606 // Ignore the default edge for now. 607 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 608 ConstantInt *DestVal = DestSI->getCaseValue(i); 609 BasicBlock *DestSucc = DestSI->getSuccessor(i); 610 611 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 612 // PredSI has an explicit case for it. If so, forward. If it is covered 613 // by the default case, we can't update PredSI. 614 unsigned PredCase = PredSI->findCaseValue(DestVal); 615 if (PredCase == 0) continue; 616 617 // If PredSI doesn't go to DestBB on this value, then it won't reach the 618 // case on this condition. 619 if (PredSI->getSuccessor(PredCase) != DestBB && 620 DestSI->getSuccessor(i) != DestBB) 621 continue; 622 623 // Otherwise, we're safe to make the change. Make sure that the edge from 624 // DestSI to DestSucc is not critical and has no PHI nodes. 625 DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 626 DEBUG(errs() << "THROUGH: " << *DestSI); 627 628 // If the destination has PHI nodes, just split the edge for updating 629 // simplicity. 630 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 631 SplitCriticalEdge(DestSI, i, this); 632 DestSucc = DestSI->getSuccessor(i); 633 } 634 FoldSingleEntryPHINodes(DestSucc); 635 PredSI->setSuccessor(PredCase, DestSucc); 636 MadeChange = true; 637 } 638 639 if (MadeChange) 640 return true; 641 } 642 643 return false; 644} 645 646 647/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 648/// load instruction, eliminate it by replacing it with a PHI node. This is an 649/// important optimization that encourages jump threading, and needs to be run 650/// interlaced with other jump threading tasks. 651bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 652 // Don't hack volatile loads. 653 if (LI->isVolatile()) return false; 654 655 // If the load is defined in a block with exactly one predecessor, it can't be 656 // partially redundant. 657 BasicBlock *LoadBB = LI->getParent(); 658 if (LoadBB->getSinglePredecessor()) 659 return false; 660 661 Value *LoadedPtr = LI->getOperand(0); 662 663 // If the loaded operand is defined in the LoadBB, it can't be available. 664 // FIXME: Could do PHI translation, that would be fun :) 665 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 666 if (PtrOp->getParent() == LoadBB) 667 return false; 668 669 // Scan a few instructions up from the load, to see if it is obviously live at 670 // the entry to its block. 671 BasicBlock::iterator BBIt = LI; 672 673 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB, 674 BBIt, 6)) { 675 // If the value if the load is locally available within the block, just use 676 // it. This frequently occurs for reg2mem'd allocas. 677 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 678 679 // If the returned value is the load itself, replace with an undef. This can 680 // only happen in dead loops. 681 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 682 LI->replaceAllUsesWith(AvailableVal); 683 LI->eraseFromParent(); 684 return true; 685 } 686 687 // Otherwise, if we scanned the whole block and got to the top of the block, 688 // we know the block is locally transparent to the load. If not, something 689 // might clobber its value. 690 if (BBIt != LoadBB->begin()) 691 return false; 692 693 694 SmallPtrSet<BasicBlock*, 8> PredsScanned; 695 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 696 AvailablePredsTy AvailablePreds; 697 BasicBlock *OneUnavailablePred = 0; 698 699 // If we got here, the loaded value is transparent through to the start of the 700 // block. Check to see if it is available in any of the predecessor blocks. 701 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 702 PI != PE; ++PI) { 703 BasicBlock *PredBB = *PI; 704 705 // If we already scanned this predecessor, skip it. 706 if (!PredsScanned.insert(PredBB)) 707 continue; 708 709 // Scan the predecessor to see if the value is available in the pred. 710 BBIt = PredBB->end(); 711 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 712 if (!PredAvailable) { 713 OneUnavailablePred = PredBB; 714 continue; 715 } 716 717 // If so, this load is partially redundant. Remember this info so that we 718 // can create a PHI node. 719 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 720 } 721 722 // If the loaded value isn't available in any predecessor, it isn't partially 723 // redundant. 724 if (AvailablePreds.empty()) return false; 725 726 // Okay, the loaded value is available in at least one (and maybe all!) 727 // predecessors. If the value is unavailable in more than one unique 728 // predecessor, we want to insert a merge block for those common predecessors. 729 // This ensures that we only have to insert one reload, thus not increasing 730 // code size. 731 BasicBlock *UnavailablePred = 0; 732 733 // If there is exactly one predecessor where the value is unavailable, the 734 // already computed 'OneUnavailablePred' block is it. If it ends in an 735 // unconditional branch, we know that it isn't a critical edge. 736 if (PredsScanned.size() == AvailablePreds.size()+1 && 737 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 738 UnavailablePred = OneUnavailablePred; 739 } else if (PredsScanned.size() != AvailablePreds.size()) { 740 // Otherwise, we had multiple unavailable predecessors or we had a critical 741 // edge from the one. 742 SmallVector<BasicBlock*, 8> PredsToSplit; 743 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 744 745 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 746 AvailablePredSet.insert(AvailablePreds[i].first); 747 748 // Add all the unavailable predecessors to the PredsToSplit list. 749 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 750 PI != PE; ++PI) 751 if (!AvailablePredSet.count(*PI)) 752 PredsToSplit.push_back(*PI); 753 754 // Split them out to their own block. 755 UnavailablePred = 756 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 757 "thread-split", this); 758 } 759 760 // If the value isn't available in all predecessors, then there will be 761 // exactly one where it isn't available. Insert a load on that edge and add 762 // it to the AvailablePreds list. 763 if (UnavailablePred) { 764 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 765 "Can't handle critical edge here!"); 766 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", 767 UnavailablePred->getTerminator()); 768 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 769 } 770 771 // Now we know that each predecessor of this block has a value in 772 // AvailablePreds, sort them for efficient access as we're walking the preds. 773 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 774 775 // Create a PHI node at the start of the block for the PRE'd load value. 776 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 777 PN->takeName(LI); 778 779 // Insert new entries into the PHI for each predecessor. A single block may 780 // have multiple entries here. 781 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 782 ++PI) { 783 AvailablePredsTy::iterator I = 784 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 785 std::make_pair(*PI, (Value*)0)); 786 787 assert(I != AvailablePreds.end() && I->first == *PI && 788 "Didn't find entry for predecessor!"); 789 790 PN->addIncoming(I->second, I->first); 791 } 792 793 //cerr << "PRE: " << *LI << *PN << "\n"; 794 795 LI->replaceAllUsesWith(PN); 796 LI->eraseFromParent(); 797 798 return true; 799} 800 801/// FindMostPopularDest - The specified list contains multiple possible 802/// threadable destinations. Pick the one that occurs the most frequently in 803/// the list. 804static BasicBlock * 805FindMostPopularDest(BasicBlock *BB, 806 const SmallVectorImpl<std::pair<BasicBlock*, 807 BasicBlock*> > &PredToDestList) { 808 assert(!PredToDestList.empty()); 809 810 // Determine popularity. If there are multiple possible destinations, we 811 // explicitly choose to ignore 'undef' destinations. We prefer to thread 812 // blocks with known and real destinations to threading undef. We'll handle 813 // them later if interesting. 814 DenseMap<BasicBlock*, unsigned> DestPopularity; 815 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 816 if (PredToDestList[i].second) 817 DestPopularity[PredToDestList[i].second]++; 818 819 // Find the most popular dest. 820 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 821 BasicBlock *MostPopularDest = DPI->first; 822 unsigned Popularity = DPI->second; 823 SmallVector<BasicBlock*, 4> SamePopularity; 824 825 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 826 // If the popularity of this entry isn't higher than the popularity we've 827 // seen so far, ignore it. 828 if (DPI->second < Popularity) 829 ; // ignore. 830 else if (DPI->second == Popularity) { 831 // If it is the same as what we've seen so far, keep track of it. 832 SamePopularity.push_back(DPI->first); 833 } else { 834 // If it is more popular, remember it. 835 SamePopularity.clear(); 836 MostPopularDest = DPI->first; 837 Popularity = DPI->second; 838 } 839 } 840 841 // Okay, now we know the most popular destination. If there is more than 842 // destination, we need to determine one. This is arbitrary, but we need 843 // to make a deterministic decision. Pick the first one that appears in the 844 // successor list. 845 if (!SamePopularity.empty()) { 846 SamePopularity.push_back(MostPopularDest); 847 TerminatorInst *TI = BB->getTerminator(); 848 for (unsigned i = 0; ; ++i) { 849 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 850 851 if (std::find(SamePopularity.begin(), SamePopularity.end(), 852 TI->getSuccessor(i)) == SamePopularity.end()) 853 continue; 854 855 MostPopularDest = TI->getSuccessor(i); 856 break; 857 } 858 } 859 860 // Okay, we have finally picked the most popular destination. 861 return MostPopularDest; 862} 863 864bool JumpThreading::ProcessThreadableEdges(Instruction *CondInst, 865 BasicBlock *BB) { 866 // If threading this would thread across a loop header, don't even try to 867 // thread the edge. 868 if (LoopHeaders.count(BB)) 869 return false; 870 871 872 873 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 874 if (!ComputeValueKnownInPredecessors(CondInst, BB, PredValues)) 875 return false; 876 assert(!PredValues.empty() && 877 "ComputeValueKnownInPredecessors returned true with no values"); 878 879 DEBUG(errs() << "IN BB: " << *BB; 880 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 881 errs() << " BB '" << BB->getName() << "': FOUND condition = "; 882 if (PredValues[i].first) 883 errs() << *PredValues[i].first; 884 else 885 errs() << "UNDEF"; 886 errs() << " for pred '" << PredValues[i].second->getName() 887 << "'.\n"; 888 }); 889 890 // Decide what we want to thread through. Convert our list of known values to 891 // a list of known destinations for each pred. This also discards duplicate 892 // predecessors and keeps track of the undefined inputs (which are represented 893 // as a null dest in the PredToDestList). 894 SmallPtrSet<BasicBlock*, 16> SeenPreds; 895 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 896 897 BasicBlock *OnlyDest = 0; 898 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 899 900 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 901 BasicBlock *Pred = PredValues[i].second; 902 if (!SeenPreds.insert(Pred)) 903 continue; // Duplicate predecessor entry. 904 905 // If the predecessor ends with an indirect goto, we can't change its 906 // destination. 907 if (isa<IndirectBrInst>(Pred->getTerminator())) 908 continue; 909 910 ConstantInt *Val = PredValues[i].first; 911 912 BasicBlock *DestBB; 913 if (Val == 0) // Undef. 914 DestBB = 0; 915 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 916 DestBB = BI->getSuccessor(Val->isZero()); 917 else { 918 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 919 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 920 } 921 922 // If we have exactly one destination, remember it for efficiency below. 923 if (i == 0) 924 OnlyDest = DestBB; 925 else if (OnlyDest != DestBB) 926 OnlyDest = MultipleDestSentinel; 927 928 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 929 } 930 931 // If all edges were unthreadable, we fail. 932 if (PredToDestList.empty()) 933 return false; 934 935 // Determine which is the most common successor. If we have many inputs and 936 // this block is a switch, we want to start by threading the batch that goes 937 // to the most popular destination first. If we only know about one 938 // threadable destination (the common case) we can avoid this. 939 BasicBlock *MostPopularDest = OnlyDest; 940 941 if (MostPopularDest == MultipleDestSentinel) 942 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 943 944 // Now that we know what the most popular destination is, factor all 945 // predecessors that will jump to it into a single predecessor. 946 SmallVector<BasicBlock*, 16> PredsToFactor; 947 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 948 if (PredToDestList[i].second == MostPopularDest) { 949 BasicBlock *Pred = PredToDestList[i].first; 950 951 // This predecessor may be a switch or something else that has multiple 952 // edges to the block. Factor each of these edges by listing them 953 // according to # occurrences in PredsToFactor. 954 TerminatorInst *PredTI = Pred->getTerminator(); 955 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 956 if (PredTI->getSuccessor(i) == BB) 957 PredsToFactor.push_back(Pred); 958 } 959 960 // If the threadable edges are branching on an undefined value, we get to pick 961 // the destination that these predecessors should get to. 962 if (MostPopularDest == 0) 963 MostPopularDest = BB->getTerminator()-> 964 getSuccessor(GetBestDestForJumpOnUndef(BB)); 965 966 // Ok, try to thread it! 967 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 968} 969 970/// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in 971/// the current block. See if there are any simplifications we can do based on 972/// inputs to the phi node. 973/// 974bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) { 975 BasicBlock *BB = PN->getParent(); 976 977 // If any of the predecessor blocks end in an unconditional branch, we can 978 // *duplicate* the jump into that block in order to further encourage jump 979 // threading and to eliminate cases where we have branch on a phi of an icmp 980 // (branch on icmp is much better). 981 982 // We don't want to do this tranformation for switches, because we don't 983 // really want to duplicate a switch. 984 if (isa<SwitchInst>(BB->getTerminator())) 985 return false; 986 987 // Look for unconditional branch predecessors. 988 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 989 BasicBlock *PredBB = PN->getIncomingBlock(i); 990 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 991 if (PredBr->isUnconditional() && 992 // Try to duplicate BB into PredBB. 993 DuplicateCondBranchOnPHIIntoPred(BB, PredBB)) 994 return true; 995 } 996 997 return false; 998} 999 1000 1001/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1002/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1003/// NewPred using the entries from OldPred (suitably mapped). 1004static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1005 BasicBlock *OldPred, 1006 BasicBlock *NewPred, 1007 DenseMap<Instruction*, Value*> &ValueMap) { 1008 for (BasicBlock::iterator PNI = PHIBB->begin(); 1009 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1010 // Ok, we have a PHI node. Figure out what the incoming value was for the 1011 // DestBlock. 1012 Value *IV = PN->getIncomingValueForBlock(OldPred); 1013 1014 // Remap the value if necessary. 1015 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1016 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1017 if (I != ValueMap.end()) 1018 IV = I->second; 1019 } 1020 1021 PN->addIncoming(IV, NewPred); 1022 } 1023} 1024 1025/// ThreadEdge - We have decided that it is safe and profitable to factor the 1026/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1027/// across BB. Transform the IR to reflect this change. 1028bool JumpThreading::ThreadEdge(BasicBlock *BB, 1029 const SmallVectorImpl<BasicBlock*> &PredBBs, 1030 BasicBlock *SuccBB) { 1031 // If threading to the same block as we come from, we would infinite loop. 1032 if (SuccBB == BB) { 1033 DEBUG(errs() << " Not threading across BB '" << BB->getName() 1034 << "' - would thread to self!\n"); 1035 return false; 1036 } 1037 1038 // If threading this would thread across a loop header, don't thread the edge. 1039 // See the comments above FindLoopHeaders for justifications and caveats. 1040 if (LoopHeaders.count(BB)) { 1041 DEBUG(errs() << " Not threading across loop header BB '" << BB->getName() 1042 << "' to dest BB '" << SuccBB->getName() 1043 << "' - it might create an irreducible loop!\n"); 1044 return false; 1045 } 1046 1047 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1048 if (JumpThreadCost > Threshold) { 1049 DEBUG(errs() << " Not threading BB '" << BB->getName() 1050 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1051 return false; 1052 } 1053 1054 // And finally, do it! Start by factoring the predecessors is needed. 1055 BasicBlock *PredBB; 1056 if (PredBBs.size() == 1) 1057 PredBB = PredBBs[0]; 1058 else { 1059 DEBUG(errs() << " Factoring out " << PredBBs.size() 1060 << " common predecessors.\n"); 1061 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1062 ".thr_comm", this); 1063 } 1064 1065 // And finally, do it! 1066 DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '" 1067 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1068 << ", across block:\n " 1069 << *BB << "\n"); 1070 1071 // We are going to have to map operands from the original BB block to the new 1072 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1073 // account for entry from PredBB. 1074 DenseMap<Instruction*, Value*> ValueMapping; 1075 1076 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1077 BB->getName()+".thread", 1078 BB->getParent(), BB); 1079 NewBB->moveAfter(PredBB); 1080 1081 BasicBlock::iterator BI = BB->begin(); 1082 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1083 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1084 1085 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1086 // mapping and using it to remap operands in the cloned instructions. 1087 for (; !isa<TerminatorInst>(BI); ++BI) { 1088 Instruction *New = BI->clone(); 1089 New->setName(BI->getName()); 1090 NewBB->getInstList().push_back(New); 1091 ValueMapping[BI] = New; 1092 1093 // Remap operands to patch up intra-block references. 1094 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1095 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1096 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1097 if (I != ValueMapping.end()) 1098 New->setOperand(i, I->second); 1099 } 1100 } 1101 1102 // We didn't copy the terminator from BB over to NewBB, because there is now 1103 // an unconditional jump to SuccBB. Insert the unconditional jump. 1104 BranchInst::Create(SuccBB, NewBB); 1105 1106 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1107 // PHI nodes for NewBB now. 1108 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1109 1110 // If there were values defined in BB that are used outside the block, then we 1111 // now have to update all uses of the value to use either the original value, 1112 // the cloned value, or some PHI derived value. This can require arbitrary 1113 // PHI insertion, of which we are prepared to do, clean these up now. 1114 SSAUpdater SSAUpdate; 1115 SmallVector<Use*, 16> UsesToRename; 1116 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1117 // Scan all uses of this instruction to see if it is used outside of its 1118 // block, and if so, record them in UsesToRename. 1119 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1120 ++UI) { 1121 Instruction *User = cast<Instruction>(*UI); 1122 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1123 if (UserPN->getIncomingBlock(UI) == BB) 1124 continue; 1125 } else if (User->getParent() == BB) 1126 continue; 1127 1128 UsesToRename.push_back(&UI.getUse()); 1129 } 1130 1131 // If there are no uses outside the block, we're done with this instruction. 1132 if (UsesToRename.empty()) 1133 continue; 1134 1135 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1136 1137 // We found a use of I outside of BB. Rename all uses of I that are outside 1138 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1139 // with the two values we know. 1140 SSAUpdate.Initialize(I); 1141 SSAUpdate.AddAvailableValue(BB, I); 1142 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1143 1144 while (!UsesToRename.empty()) 1145 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1146 DEBUG(errs() << "\n"); 1147 } 1148 1149 1150 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1151 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1152 // us to simplify any PHI nodes in BB. 1153 TerminatorInst *PredTerm = PredBB->getTerminator(); 1154 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1155 if (PredTerm->getSuccessor(i) == BB) { 1156 BB->removePredecessor(PredBB); 1157 PredTerm->setSuccessor(i, NewBB); 1158 } 1159 1160 // At this point, the IR is fully up to date and consistent. Do a quick scan 1161 // over the new instructions and zap any that are constants or dead. This 1162 // frequently happens because of phi translation. 1163 BI = NewBB->begin(); 1164 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) { 1165 Instruction *Inst = BI++; 1166 if (Constant *C = ConstantFoldInstruction(Inst, TD)) { 1167 Inst->replaceAllUsesWith(C); 1168 Inst->eraseFromParent(); 1169 continue; 1170 } 1171 1172 RecursivelyDeleteTriviallyDeadInstructions(Inst); 1173 } 1174 1175 // Threaded an edge! 1176 ++NumThreads; 1177 return true; 1178} 1179 1180/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1181/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1182/// If we can duplicate the contents of BB up into PredBB do so now, this 1183/// improves the odds that the branch will be on an analyzable instruction like 1184/// a compare. 1185bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1186 BasicBlock *PredBB) { 1187 // If BB is a loop header, then duplicating this block outside the loop would 1188 // cause us to transform this into an irreducible loop, don't do this. 1189 // See the comments above FindLoopHeaders for justifications and caveats. 1190 if (LoopHeaders.count(BB)) { 1191 DEBUG(errs() << " Not duplicating loop header '" << BB->getName() 1192 << "' into predecessor block '" << PredBB->getName() 1193 << "' - it might create an irreducible loop!\n"); 1194 return false; 1195 } 1196 1197 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1198 if (DuplicationCost > Threshold) { 1199 DEBUG(errs() << " Not duplicating BB '" << BB->getName() 1200 << "' - Cost is too high: " << DuplicationCost << "\n"); 1201 return false; 1202 } 1203 1204 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1205 // of PredBB. 1206 DEBUG(errs() << " Duplicating block '" << BB->getName() << "' into end of '" 1207 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1208 << DuplicationCost << " block is:" << *BB << "\n"); 1209 1210 // We are going to have to map operands from the original BB block into the 1211 // PredBB block. Evaluate PHI nodes in BB. 1212 DenseMap<Instruction*, Value*> ValueMapping; 1213 1214 BasicBlock::iterator BI = BB->begin(); 1215 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1216 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1217 1218 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1219 1220 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1221 // mapping and using it to remap operands in the cloned instructions. 1222 for (; BI != BB->end(); ++BI) { 1223 Instruction *New = BI->clone(); 1224 New->setName(BI->getName()); 1225 PredBB->getInstList().insert(OldPredBranch, New); 1226 ValueMapping[BI] = New; 1227 1228 // Remap operands to patch up intra-block references. 1229 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1230 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1231 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1232 if (I != ValueMapping.end()) 1233 New->setOperand(i, I->second); 1234 } 1235 } 1236 1237 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1238 // add entries to the PHI nodes for branch from PredBB now. 1239 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1240 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1241 ValueMapping); 1242 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1243 ValueMapping); 1244 1245 // If there were values defined in BB that are used outside the block, then we 1246 // now have to update all uses of the value to use either the original value, 1247 // the cloned value, or some PHI derived value. This can require arbitrary 1248 // PHI insertion, of which we are prepared to do, clean these up now. 1249 SSAUpdater SSAUpdate; 1250 SmallVector<Use*, 16> UsesToRename; 1251 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1252 // Scan all uses of this instruction to see if it is used outside of its 1253 // block, and if so, record them in UsesToRename. 1254 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1255 ++UI) { 1256 Instruction *User = cast<Instruction>(*UI); 1257 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1258 if (UserPN->getIncomingBlock(UI) == BB) 1259 continue; 1260 } else if (User->getParent() == BB) 1261 continue; 1262 1263 UsesToRename.push_back(&UI.getUse()); 1264 } 1265 1266 // If there are no uses outside the block, we're done with this instruction. 1267 if (UsesToRename.empty()) 1268 continue; 1269 1270 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1271 1272 // We found a use of I outside of BB. Rename all uses of I that are outside 1273 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1274 // with the two values we know. 1275 SSAUpdate.Initialize(I); 1276 SSAUpdate.AddAvailableValue(BB, I); 1277 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1278 1279 while (!UsesToRename.empty()) 1280 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1281 DEBUG(errs() << "\n"); 1282 } 1283 1284 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1285 // that we nuked. 1286 BB->removePredecessor(PredBB); 1287 1288 // Remove the unconditional branch at the end of the PredBB block. 1289 OldPredBranch->eraseFromParent(); 1290 1291 ++NumDupes; 1292 return true; 1293} 1294 1295 1296