JumpThreading.cpp revision 1e35610f20ba04daf3aa3048219914421b9931af
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/Analysis/InstructionSimplify.h" 21#include "llvm/Analysis/LazyValueInfo.h" 22#include "llvm/Analysis/Loads.h" 23#include "llvm/Transforms/Utils/BasicBlockUtils.h" 24#include "llvm/Transforms/Utils/Local.h" 25#include "llvm/Transforms/Utils/SSAUpdater.h" 26#include "llvm/Target/TargetData.h" 27#include "llvm/ADT/DenseMap.h" 28#include "llvm/ADT/Statistic.h" 29#include "llvm/ADT/STLExtras.h" 30#include "llvm/ADT/SmallPtrSet.h" 31#include "llvm/ADT/SmallSet.h" 32#include "llvm/Support/CommandLine.h" 33#include "llvm/Support/Debug.h" 34#include "llvm/Support/ValueHandle.h" 35#include "llvm/Support/raw_ostream.h" 36using namespace llvm; 37 38STATISTIC(NumThreads, "Number of jumps threaded"); 39STATISTIC(NumFolds, "Number of terminators folded"); 40STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 41 42static cl::opt<unsigned> 43Threshold("jump-threading-threshold", 44 cl::desc("Max block size to duplicate for jump threading"), 45 cl::init(6), cl::Hidden); 46 47// Turn on use of LazyValueInfo. 48static cl::opt<bool> 49EnableLVI("enable-jump-threading-lvi", 50 cl::desc("Use LVI for jump threading"), 51 cl::init(true), 52 cl::ReallyHidden); 53 54 55 56namespace { 57 /// This pass performs 'jump threading', which looks at blocks that have 58 /// multiple predecessors and multiple successors. If one or more of the 59 /// predecessors of the block can be proven to always jump to one of the 60 /// successors, we forward the edge from the predecessor to the successor by 61 /// duplicating the contents of this block. 62 /// 63 /// An example of when this can occur is code like this: 64 /// 65 /// if () { ... 66 /// X = 4; 67 /// } 68 /// if (X < 3) { 69 /// 70 /// In this case, the unconditional branch at the end of the first if can be 71 /// revectored to the false side of the second if. 72 /// 73 class JumpThreading : public FunctionPass { 74 TargetData *TD; 75 LazyValueInfo *LVI; 76#ifdef NDEBUG 77 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 78#else 79 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 80#endif 81 public: 82 static char ID; // Pass identification 83 JumpThreading() : FunctionPass(ID) {} 84 85 bool runOnFunction(Function &F); 86 87 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 88 if (EnableLVI) 89 AU.addRequired<LazyValueInfo>(); 90 AU.addPreserved<LazyValueInfo>(); 91 } 92 93 void FindLoopHeaders(Function &F); 94 bool ProcessBlock(BasicBlock *BB); 95 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 96 BasicBlock *SuccBB); 97 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 98 const SmallVectorImpl<BasicBlock *> &PredBBs); 99 100 typedef SmallVectorImpl<std::pair<ConstantInt*, 101 BasicBlock*> > PredValueInfo; 102 103 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 104 PredValueInfo &Result); 105 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB); 106 107 108 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 109 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 110 111 bool ProcessBranchOnPHI(PHINode *PN); 112 bool ProcessBranchOnXOR(BinaryOperator *BO); 113 114 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 115 }; 116} 117 118char JumpThreading::ID = 0; 119INITIALIZE_PASS(JumpThreading, "jump-threading", 120 "Jump Threading", false, false); 121 122// Public interface to the Jump Threading pass 123FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 124 125/// runOnFunction - Top level algorithm. 126/// 127bool JumpThreading::runOnFunction(Function &F) { 128 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 129 TD = getAnalysisIfAvailable<TargetData>(); 130 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0; 131 132 FindLoopHeaders(F); 133 134 bool Changed, EverChanged = false; 135 do { 136 Changed = false; 137 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 138 BasicBlock *BB = I; 139 // Thread all of the branches we can over this block. 140 while (ProcessBlock(BB)) 141 Changed = true; 142 143 ++I; 144 145 // If the block is trivially dead, zap it. This eliminates the successor 146 // edges which simplifies the CFG. 147 if (pred_begin(BB) == pred_end(BB) && 148 BB != &BB->getParent()->getEntryBlock()) { 149 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 150 << "' with terminator: " << *BB->getTerminator() << '\n'); 151 LoopHeaders.erase(BB); 152 if (LVI) LVI->eraseBlock(BB); 153 DeleteDeadBlock(BB); 154 Changed = true; 155 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 156 // Can't thread an unconditional jump, but if the block is "almost 157 // empty", we can replace uses of it with uses of the successor and make 158 // this dead. 159 if (BI->isUnconditional() && 160 BB != &BB->getParent()->getEntryBlock()) { 161 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 162 // Ignore dbg intrinsics. 163 while (isa<DbgInfoIntrinsic>(BBI)) 164 ++BBI; 165 // If the terminator is the only non-phi instruction, try to nuke it. 166 if (BBI->isTerminator()) { 167 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 168 // block, we have to make sure it isn't in the LoopHeaders set. We 169 // reinsert afterward if needed. 170 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 171 BasicBlock *Succ = BI->getSuccessor(0); 172 173 // FIXME: It is always conservatively correct to drop the info 174 // for a block even if it doesn't get erased. This isn't totally 175 // awesome, but it allows us to use AssertingVH to prevent nasty 176 // dangling pointer issues within LazyValueInfo. 177 if (LVI) LVI->eraseBlock(BB); 178 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 179 Changed = true; 180 // If we deleted BB and BB was the header of a loop, then the 181 // successor is now the header of the loop. 182 BB = Succ; 183 } 184 185 if (ErasedFromLoopHeaders) 186 LoopHeaders.insert(BB); 187 } 188 } 189 } 190 } 191 EverChanged |= Changed; 192 } while (Changed); 193 194 LoopHeaders.clear(); 195 return EverChanged; 196} 197 198/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 199/// thread across it. 200static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 201 /// Ignore PHI nodes, these will be flattened when duplication happens. 202 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 203 204 // FIXME: THREADING will delete values that are just used to compute the 205 // branch, so they shouldn't count against the duplication cost. 206 207 208 // Sum up the cost of each instruction until we get to the terminator. Don't 209 // include the terminator because the copy won't include it. 210 unsigned Size = 0; 211 for (; !isa<TerminatorInst>(I); ++I) { 212 // Debugger intrinsics don't incur code size. 213 if (isa<DbgInfoIntrinsic>(I)) continue; 214 215 // If this is a pointer->pointer bitcast, it is free. 216 if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) 217 continue; 218 219 // All other instructions count for at least one unit. 220 ++Size; 221 222 // Calls are more expensive. If they are non-intrinsic calls, we model them 223 // as having cost of 4. If they are a non-vector intrinsic, we model them 224 // as having cost of 2 total, and if they are a vector intrinsic, we model 225 // them as having cost 1. 226 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 227 if (!isa<IntrinsicInst>(CI)) 228 Size += 3; 229 else if (!CI->getType()->isVectorTy()) 230 Size += 1; 231 } 232 } 233 234 // Threading through a switch statement is particularly profitable. If this 235 // block ends in a switch, decrease its cost to make it more likely to happen. 236 if (isa<SwitchInst>(I)) 237 Size = Size > 6 ? Size-6 : 0; 238 239 return Size; 240} 241 242/// FindLoopHeaders - We do not want jump threading to turn proper loop 243/// structures into irreducible loops. Doing this breaks up the loop nesting 244/// hierarchy and pessimizes later transformations. To prevent this from 245/// happening, we first have to find the loop headers. Here we approximate this 246/// by finding targets of backedges in the CFG. 247/// 248/// Note that there definitely are cases when we want to allow threading of 249/// edges across a loop header. For example, threading a jump from outside the 250/// loop (the preheader) to an exit block of the loop is definitely profitable. 251/// It is also almost always profitable to thread backedges from within the loop 252/// to exit blocks, and is often profitable to thread backedges to other blocks 253/// within the loop (forming a nested loop). This simple analysis is not rich 254/// enough to track all of these properties and keep it up-to-date as the CFG 255/// mutates, so we don't allow any of these transformations. 256/// 257void JumpThreading::FindLoopHeaders(Function &F) { 258 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 259 FindFunctionBackedges(F, Edges); 260 261 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 262 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 263} 264 265/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 266/// if we can infer that the value is a known ConstantInt in any of our 267/// predecessors. If so, return the known list of value and pred BB in the 268/// result vector. If a value is known to be undef, it is returned as null. 269/// 270/// This returns true if there were any known values. 271/// 272bool JumpThreading:: 273ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 274 // If V is a constantint, then it is known in all predecessors. 275 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 276 ConstantInt *CI = dyn_cast<ConstantInt>(V); 277 278 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 279 Result.push_back(std::make_pair(CI, *PI)); 280 return true; 281 } 282 283 // If V is a non-instruction value, or an instruction in a different block, 284 // then it can't be derived from a PHI. 285 Instruction *I = dyn_cast<Instruction>(V); 286 if (I == 0 || I->getParent() != BB) { 287 288 // Okay, if this is a live-in value, see if it has a known value at the end 289 // of any of our predecessors. 290 // 291 // FIXME: This should be an edge property, not a block end property. 292 /// TODO: Per PR2563, we could infer value range information about a 293 /// predecessor based on its terminator. 294 // 295 if (LVI) { 296 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 297 // "I" is a non-local compare-with-a-constant instruction. This would be 298 // able to handle value inequalities better, for example if the compare is 299 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 300 // Perhaps getConstantOnEdge should be smart enough to do this? 301 302 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 303 BasicBlock *P = *PI; 304 // If the value is known by LazyValueInfo to be a constant in a 305 // predecessor, use that information to try to thread this block. 306 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB); 307 if (PredCst == 0 || 308 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst))) 309 continue; 310 311 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), P)); 312 } 313 314 return !Result.empty(); 315 } 316 317 return false; 318 } 319 320 /// If I is a PHI node, then we know the incoming values for any constants. 321 if (PHINode *PN = dyn_cast<PHINode>(I)) { 322 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 323 Value *InVal = PN->getIncomingValue(i); 324 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 325 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 326 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 327 } else if (LVI) { 328 Constant *CI = LVI->getConstantOnEdge(InVal, 329 PN->getIncomingBlock(i), BB); 330 // LVI returns null is no value could be determined. 331 if (!CI) continue; 332 ConstantInt *CInt = dyn_cast<ConstantInt>(CI); 333 Result.push_back(std::make_pair(CInt, PN->getIncomingBlock(i))); 334 } 335 } 336 return !Result.empty(); 337 } 338 339 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 340 341 // Handle some boolean conditions. 342 if (I->getType()->getPrimitiveSizeInBits() == 1) { 343 // X | true -> true 344 // X & false -> false 345 if (I->getOpcode() == Instruction::Or || 346 I->getOpcode() == Instruction::And) { 347 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 348 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 349 350 if (LHSVals.empty() && RHSVals.empty()) 351 return false; 352 353 ConstantInt *InterestingVal; 354 if (I->getOpcode() == Instruction::Or) 355 InterestingVal = ConstantInt::getTrue(I->getContext()); 356 else 357 InterestingVal = ConstantInt::getFalse(I->getContext()); 358 359 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; 360 361 // Scan for the sentinel. If we find an undef, force it to the 362 // interesting value: x|undef -> true and x&undef -> false. 363 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 364 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) { 365 Result.push_back(LHSVals[i]); 366 Result.back().first = InterestingVal; 367 LHSKnownBBs.insert(LHSVals[i].second); 368 } 369 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 370 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) { 371 // If we already inferred a value for this block on the LHS, don't 372 // re-add it. 373 if (!LHSKnownBBs.count(RHSVals[i].second)) { 374 Result.push_back(RHSVals[i]); 375 Result.back().first = InterestingVal; 376 } 377 } 378 return !Result.empty(); 379 } 380 381 // Handle the NOT form of XOR. 382 if (I->getOpcode() == Instruction::Xor && 383 isa<ConstantInt>(I->getOperand(1)) && 384 cast<ConstantInt>(I->getOperand(1))->isOne()) { 385 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result); 386 if (Result.empty()) 387 return false; 388 389 // Invert the known values. 390 for (unsigned i = 0, e = Result.size(); i != e; ++i) 391 if (Result[i].first) 392 Result[i].first = 393 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first)); 394 return true; 395 } 396 397 // Try to simplify some other binary operator values. 398 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 399 // AND or OR of a value with itself is that value. 400 ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1)); 401 if (CI && (BO->getOpcode() == Instruction::And || 402 BO->getOpcode() == Instruction::Or)) { 403 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals; 404 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals); 405 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 406 if (LHSVals[i].first == 0) 407 Result.push_back(std::make_pair((ConstantInt*)0, LHSVals[i].second)); 408 else if (LHSVals[i].first == CI) 409 Result.push_back(std::make_pair(CI, LHSVals[i].second)); 410 411 return !Result.empty(); 412 } 413 } 414 415 // Handle compare with phi operand, where the PHI is defined in this block. 416 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 417 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 418 if (PN && PN->getParent() == BB) { 419 // We can do this simplification if any comparisons fold to true or false. 420 // See if any do. 421 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 422 BasicBlock *PredBB = PN->getIncomingBlock(i); 423 Value *LHS = PN->getIncomingValue(i); 424 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 425 426 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD); 427 if (Res == 0) { 428 if (!LVI || !isa<Constant>(RHS)) 429 continue; 430 431 LazyValueInfo::Tristate 432 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 433 cast<Constant>(RHS), PredBB, BB); 434 if (ResT == LazyValueInfo::Unknown) 435 continue; 436 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 437 } 438 439 if (isa<UndefValue>(Res)) 440 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 441 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 442 Result.push_back(std::make_pair(CI, PredBB)); 443 } 444 445 return !Result.empty(); 446 } 447 448 449 // If comparing a live-in value against a constant, see if we know the 450 // live-in value on any predecessors. 451 if (LVI && isa<Constant>(Cmp->getOperand(1)) && 452 Cmp->getType()->isIntegerTy()) { 453 if (!isa<Instruction>(Cmp->getOperand(0)) || 454 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) { 455 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 456 457 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){ 458 BasicBlock *P = *PI; 459 // If the value is known by LazyValueInfo to be a constant in a 460 // predecessor, use that information to try to thread this block. 461 LazyValueInfo::Tristate Res = 462 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 463 RHSCst, P, BB); 464 if (Res == LazyValueInfo::Unknown) 465 continue; 466 467 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 468 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), P)); 469 } 470 471 return !Result.empty(); 472 } 473 474 // Try to find a constant value for the LHS of an equality comparison, 475 // and evaluate it statically if we can. 476 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) { 477 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals; 478 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 479 480 ConstantInt *True = ConstantInt::getTrue(I->getContext()); 481 ConstantInt *False = ConstantInt::getFalse(I->getContext()); 482 483 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 484 if (LHSVals[i].first == 0) 485 Result.push_back(std::make_pair((ConstantInt*)0, 486 LHSVals[i].second)); 487 else if (ConstantFoldCompareInstOperands(Cmp->getPredicate(), 488 LHSVals[i].first, 489 CmpConst)) 490 Result.push_back(std::make_pair(True, LHSVals[i].second)); 491 else 492 Result.push_back(std::make_pair(False, LHSVals[i].second)); 493 } 494 495 return !Result.empty(); 496 } 497 } 498 } 499 500 if (LVI) { 501 // If all else fails, see if LVI can figure out a constant value for us. 502 Constant *CI = LVI->getConstant(V, BB); 503 ConstantInt *CInt = dyn_cast_or_null<ConstantInt>(CI); 504 if (CInt) { 505 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 506 Result.push_back(std::make_pair(CInt, *PI)); 507 } 508 509 return !Result.empty(); 510 } 511 512 return false; 513} 514 515 516 517/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 518/// in an undefined jump, decide which block is best to revector to. 519/// 520/// Since we can pick an arbitrary destination, we pick the successor with the 521/// fewest predecessors. This should reduce the in-degree of the others. 522/// 523static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 524 TerminatorInst *BBTerm = BB->getTerminator(); 525 unsigned MinSucc = 0; 526 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 527 // Compute the successor with the minimum number of predecessors. 528 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 529 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 530 TestBB = BBTerm->getSuccessor(i); 531 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 532 if (NumPreds < MinNumPreds) 533 MinSucc = i; 534 } 535 536 return MinSucc; 537} 538 539/// ProcessBlock - If there are any predecessors whose control can be threaded 540/// through to a successor, transform them now. 541bool JumpThreading::ProcessBlock(BasicBlock *BB) { 542 // If the block is trivially dead, just return and let the caller nuke it. 543 // This simplifies other transformations. 544 if (pred_begin(BB) == pred_end(BB) && 545 BB != &BB->getParent()->getEntryBlock()) 546 return false; 547 548 // If this block has a single predecessor, and if that pred has a single 549 // successor, merge the blocks. This encourages recursive jump threading 550 // because now the condition in this block can be threaded through 551 // predecessors of our predecessor block. 552 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 553 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 554 SinglePred != BB) { 555 // If SinglePred was a loop header, BB becomes one. 556 if (LoopHeaders.erase(SinglePred)) 557 LoopHeaders.insert(BB); 558 559 // Remember if SinglePred was the entry block of the function. If so, we 560 // will need to move BB back to the entry position. 561 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 562 if (LVI) LVI->eraseBlock(SinglePred); 563 MergeBasicBlockIntoOnlyPred(BB); 564 565 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 566 BB->moveBefore(&BB->getParent()->getEntryBlock()); 567 return true; 568 } 569 } 570 571 // Look to see if the terminator is a branch of switch, if not we can't thread 572 // it. 573 Value *Condition; 574 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 575 // Can't thread an unconditional jump. 576 if (BI->isUnconditional()) return false; 577 Condition = BI->getCondition(); 578 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 579 Condition = SI->getCondition(); 580 else 581 return false; // Must be an invoke. 582 583 // If the terminator of this block is branching on a constant, simplify the 584 // terminator to an unconditional branch. This can occur due to threading in 585 // other blocks. 586 if (isa<ConstantInt>(Condition)) { 587 DEBUG(dbgs() << " In block '" << BB->getName() 588 << "' folding terminator: " << *BB->getTerminator() << '\n'); 589 ++NumFolds; 590 ConstantFoldTerminator(BB); 591 return true; 592 } 593 594 // If the terminator is branching on an undef, we can pick any of the 595 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 596 if (isa<UndefValue>(Condition)) { 597 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 598 599 // Fold the branch/switch. 600 TerminatorInst *BBTerm = BB->getTerminator(); 601 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 602 if (i == BestSucc) continue; 603 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD); 604 } 605 606 DEBUG(dbgs() << " In block '" << BB->getName() 607 << "' folding undef terminator: " << *BBTerm << '\n'); 608 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 609 BBTerm->eraseFromParent(); 610 return true; 611 } 612 613 Instruction *CondInst = dyn_cast<Instruction>(Condition); 614 615 // If the condition is an instruction defined in another block, see if a 616 // predecessor has the same condition: 617 // br COND, BBX, BBY 618 // BBX: 619 // br COND, BBZ, BBW 620 if (!LVI && 621 !Condition->hasOneUse() && // Multiple uses. 622 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 623 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 624 if (isa<BranchInst>(BB->getTerminator())) { 625 for (; PI != E; ++PI) { 626 BasicBlock *P = *PI; 627 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator())) 628 if (PBI->isConditional() && PBI->getCondition() == Condition && 629 ProcessBranchOnDuplicateCond(P, BB)) 630 return true; 631 } 632 } else { 633 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 634 for (; PI != E; ++PI) { 635 BasicBlock *P = *PI; 636 if (SwitchInst *PSI = dyn_cast<SwitchInst>(P->getTerminator())) 637 if (PSI->getCondition() == Condition && 638 ProcessSwitchOnDuplicateCond(P, BB)) 639 return true; 640 } 641 } 642 } 643 644 // All the rest of our checks depend on the condition being an instruction. 645 if (CondInst == 0) { 646 // FIXME: Unify this with code below. 647 if (LVI && ProcessThreadableEdges(Condition, BB)) 648 return true; 649 return false; 650 } 651 652 653 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 654 if (!LVI && 655 (!isa<PHINode>(CondCmp->getOperand(0)) || 656 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) { 657 // If we have a comparison, loop over the predecessors to see if there is 658 // a condition with a lexically identical value. 659 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 660 for (; PI != E; ++PI) { 661 BasicBlock *P = *PI; 662 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator())) 663 if (PBI->isConditional() && P != BB) { 664 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 665 if (CI->getOperand(0) == CondCmp->getOperand(0) && 666 CI->getOperand(1) == CondCmp->getOperand(1) && 667 CI->getPredicate() == CondCmp->getPredicate()) { 668 // TODO: Could handle things like (x != 4) --> (x == 17) 669 if (ProcessBranchOnDuplicateCond(P, BB)) 670 return true; 671 } 672 } 673 } 674 } 675 } 676 677 // For a comparison where the LHS is outside this block, it's possible 678 // that we've branched on it before. Used LVI to see if we can simplify 679 // the branch based on that. 680 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 681 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1)); 682 if (LVI && CondBr && CondConst && CondBr->isConditional() && 683 (!isa<Instruction>(CondCmp->getOperand(0)) || 684 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) { 685 // For predecessor edge, determine if the comparison is true or false 686 // on that edge. If they're all true or all false, we can simplify the 687 // branch. 688 // FIXME: We could handle mixed true/false by duplicating code. 689 unsigned Trues = 0, Falses = 0, predcount = 0; 690 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB);PI != PE; ++PI){ 691 ++predcount; 692 LazyValueInfo::Tristate Ret = 693 LVI->getPredicateOnEdge(CondCmp->getPredicate(), 694 CondCmp->getOperand(0), CondConst, *PI, BB); 695 if (Ret == LazyValueInfo::True) 696 ++Trues; 697 else if (Ret == LazyValueInfo::False) 698 ++Falses; 699 } 700 701 // If we can determine the branch direction statically, convert 702 // the conditional branch to an unconditional one. 703 if (Trues && Trues == predcount) { 704 RemovePredecessorAndSimplify(CondBr->getSuccessor(1), BB, TD); 705 BranchInst::Create(CondBr->getSuccessor(0), CondBr); 706 CondBr->eraseFromParent(); 707 return true; 708 } else if (Falses && Falses == predcount) { 709 RemovePredecessorAndSimplify(CondBr->getSuccessor(0), BB, TD); 710 BranchInst::Create(CondBr->getSuccessor(1), CondBr); 711 CondBr->eraseFromParent(); 712 return true; 713 } 714 } 715 } 716 717 // Check for some cases that are worth simplifying. Right now we want to look 718 // for loads that are used by a switch or by the condition for the branch. If 719 // we see one, check to see if it's partially redundant. If so, insert a PHI 720 // which can then be used to thread the values. 721 // 722 Value *SimplifyValue = CondInst; 723 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 724 if (isa<Constant>(CondCmp->getOperand(1))) 725 SimplifyValue = CondCmp->getOperand(0); 726 727 // TODO: There are other places where load PRE would be profitable, such as 728 // more complex comparisons. 729 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 730 if (SimplifyPartiallyRedundantLoad(LI)) 731 return true; 732 733 734 // Handle a variety of cases where we are branching on something derived from 735 // a PHI node in the current block. If we can prove that any predecessors 736 // compute a predictable value based on a PHI node, thread those predecessors. 737 // 738 if (ProcessThreadableEdges(CondInst, BB)) 739 return true; 740 741 // If this is an otherwise-unfoldable branch on a phi node in the current 742 // block, see if we can simplify. 743 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 744 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 745 return ProcessBranchOnPHI(PN); 746 747 748 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 749 if (CondInst->getOpcode() == Instruction::Xor && 750 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 751 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 752 753 754 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 755 // "(X == 4)", thread through this block. 756 757 return false; 758} 759 760/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 761/// block that jump on exactly the same condition. This means that we almost 762/// always know the direction of the edge in the DESTBB: 763/// PREDBB: 764/// br COND, DESTBB, BBY 765/// DESTBB: 766/// br COND, BBZ, BBW 767/// 768/// If DESTBB has multiple predecessors, we can't just constant fold the branch 769/// in DESTBB, we have to thread over it. 770bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 771 BasicBlock *BB) { 772 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 773 774 // If both successors of PredBB go to DESTBB, we don't know anything. We can 775 // fold the branch to an unconditional one, which allows other recursive 776 // simplifications. 777 bool BranchDir; 778 if (PredBI->getSuccessor(1) != BB) 779 BranchDir = true; 780 else if (PredBI->getSuccessor(0) != BB) 781 BranchDir = false; 782 else { 783 DEBUG(dbgs() << " In block '" << PredBB->getName() 784 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 785 ++NumFolds; 786 ConstantFoldTerminator(PredBB); 787 return true; 788 } 789 790 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 791 792 // If the dest block has one predecessor, just fix the branch condition to a 793 // constant and fold it. 794 if (BB->getSinglePredecessor()) { 795 DEBUG(dbgs() << " In block '" << BB->getName() 796 << "' folding condition to '" << BranchDir << "': " 797 << *BB->getTerminator() << '\n'); 798 ++NumFolds; 799 Value *OldCond = DestBI->getCondition(); 800 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 801 BranchDir)); 802 // Delete dead instructions before we fold the branch. Folding the branch 803 // can eliminate edges from the CFG which can end up deleting OldCond. 804 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 805 ConstantFoldTerminator(BB); 806 return true; 807 } 808 809 810 // Next, figure out which successor we are threading to. 811 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 812 813 SmallVector<BasicBlock*, 2> Preds; 814 Preds.push_back(PredBB); 815 816 // Ok, try to thread it! 817 return ThreadEdge(BB, Preds, SuccBB); 818} 819 820/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 821/// block that switch on exactly the same condition. This means that we almost 822/// always know the direction of the edge in the DESTBB: 823/// PREDBB: 824/// switch COND [... DESTBB, BBY ... ] 825/// DESTBB: 826/// switch COND [... BBZ, BBW ] 827/// 828/// Optimizing switches like this is very important, because simplifycfg builds 829/// switches out of repeated 'if' conditions. 830bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 831 BasicBlock *DestBB) { 832 // Can't thread edge to self. 833 if (PredBB == DestBB) 834 return false; 835 836 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 837 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 838 839 // There are a variety of optimizations that we can potentially do on these 840 // blocks: we order them from most to least preferable. 841 842 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 843 // directly to their destination. This does not introduce *any* code size 844 // growth. Skip debug info first. 845 BasicBlock::iterator BBI = DestBB->begin(); 846 while (isa<DbgInfoIntrinsic>(BBI)) 847 BBI++; 848 849 // FIXME: Thread if it just contains a PHI. 850 if (isa<SwitchInst>(BBI)) { 851 bool MadeChange = false; 852 // Ignore the default edge for now. 853 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 854 ConstantInt *DestVal = DestSI->getCaseValue(i); 855 BasicBlock *DestSucc = DestSI->getSuccessor(i); 856 857 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 858 // PredSI has an explicit case for it. If so, forward. If it is covered 859 // by the default case, we can't update PredSI. 860 unsigned PredCase = PredSI->findCaseValue(DestVal); 861 if (PredCase == 0) continue; 862 863 // If PredSI doesn't go to DestBB on this value, then it won't reach the 864 // case on this condition. 865 if (PredSI->getSuccessor(PredCase) != DestBB && 866 DestSI->getSuccessor(i) != DestBB) 867 continue; 868 869 // Do not forward this if it already goes to this destination, this would 870 // be an infinite loop. 871 if (PredSI->getSuccessor(PredCase) == DestSucc) 872 continue; 873 874 // Otherwise, we're safe to make the change. Make sure that the edge from 875 // DestSI to DestSucc is not critical and has no PHI nodes. 876 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 877 DEBUG(dbgs() << "THROUGH: " << *DestSI); 878 879 // If the destination has PHI nodes, just split the edge for updating 880 // simplicity. 881 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 882 SplitCriticalEdge(DestSI, i, this); 883 DestSucc = DestSI->getSuccessor(i); 884 } 885 FoldSingleEntryPHINodes(DestSucc); 886 PredSI->setSuccessor(PredCase, DestSucc); 887 MadeChange = true; 888 } 889 890 if (MadeChange) 891 return true; 892 } 893 894 return false; 895} 896 897 898/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 899/// load instruction, eliminate it by replacing it with a PHI node. This is an 900/// important optimization that encourages jump threading, and needs to be run 901/// interlaced with other jump threading tasks. 902bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 903 // Don't hack volatile loads. 904 if (LI->isVolatile()) return false; 905 906 // If the load is defined in a block with exactly one predecessor, it can't be 907 // partially redundant. 908 BasicBlock *LoadBB = LI->getParent(); 909 if (LoadBB->getSinglePredecessor()) 910 return false; 911 912 Value *LoadedPtr = LI->getOperand(0); 913 914 // If the loaded operand is defined in the LoadBB, it can't be available. 915 // TODO: Could do simple PHI translation, that would be fun :) 916 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 917 if (PtrOp->getParent() == LoadBB) 918 return false; 919 920 // Scan a few instructions up from the load, to see if it is obviously live at 921 // the entry to its block. 922 BasicBlock::iterator BBIt = LI; 923 924 if (Value *AvailableVal = 925 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 926 // If the value if the load is locally available within the block, just use 927 // it. This frequently occurs for reg2mem'd allocas. 928 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 929 930 // If the returned value is the load itself, replace with an undef. This can 931 // only happen in dead loops. 932 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 933 LI->replaceAllUsesWith(AvailableVal); 934 LI->eraseFromParent(); 935 return true; 936 } 937 938 // Otherwise, if we scanned the whole block and got to the top of the block, 939 // we know the block is locally transparent to the load. If not, something 940 // might clobber its value. 941 if (BBIt != LoadBB->begin()) 942 return false; 943 944 945 SmallPtrSet<BasicBlock*, 8> PredsScanned; 946 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 947 AvailablePredsTy AvailablePreds; 948 BasicBlock *OneUnavailablePred = 0; 949 950 // If we got here, the loaded value is transparent through to the start of the 951 // block. Check to see if it is available in any of the predecessor blocks. 952 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 953 PI != PE; ++PI) { 954 BasicBlock *PredBB = *PI; 955 956 // If we already scanned this predecessor, skip it. 957 if (!PredsScanned.insert(PredBB)) 958 continue; 959 960 // Scan the predecessor to see if the value is available in the pred. 961 BBIt = PredBB->end(); 962 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 963 if (!PredAvailable) { 964 OneUnavailablePred = PredBB; 965 continue; 966 } 967 968 // If so, this load is partially redundant. Remember this info so that we 969 // can create a PHI node. 970 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 971 } 972 973 // If the loaded value isn't available in any predecessor, it isn't partially 974 // redundant. 975 if (AvailablePreds.empty()) return false; 976 977 // Okay, the loaded value is available in at least one (and maybe all!) 978 // predecessors. If the value is unavailable in more than one unique 979 // predecessor, we want to insert a merge block for those common predecessors. 980 // This ensures that we only have to insert one reload, thus not increasing 981 // code size. 982 BasicBlock *UnavailablePred = 0; 983 984 // If there is exactly one predecessor where the value is unavailable, the 985 // already computed 'OneUnavailablePred' block is it. If it ends in an 986 // unconditional branch, we know that it isn't a critical edge. 987 if (PredsScanned.size() == AvailablePreds.size()+1 && 988 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 989 UnavailablePred = OneUnavailablePred; 990 } else if (PredsScanned.size() != AvailablePreds.size()) { 991 // Otherwise, we had multiple unavailable predecessors or we had a critical 992 // edge from the one. 993 SmallVector<BasicBlock*, 8> PredsToSplit; 994 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 995 996 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 997 AvailablePredSet.insert(AvailablePreds[i].first); 998 999 // Add all the unavailable predecessors to the PredsToSplit list. 1000 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 1001 PI != PE; ++PI) { 1002 BasicBlock *P = *PI; 1003 // If the predecessor is an indirect goto, we can't split the edge. 1004 if (isa<IndirectBrInst>(P->getTerminator())) 1005 return false; 1006 1007 if (!AvailablePredSet.count(P)) 1008 PredsToSplit.push_back(P); 1009 } 1010 1011 // Split them out to their own block. 1012 UnavailablePred = 1013 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 1014 "thread-pre-split", this); 1015 } 1016 1017 // If the value isn't available in all predecessors, then there will be 1018 // exactly one where it isn't available. Insert a load on that edge and add 1019 // it to the AvailablePreds list. 1020 if (UnavailablePred) { 1021 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 1022 "Can't handle critical edge here!"); 1023 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 1024 LI->getAlignment(), 1025 UnavailablePred->getTerminator()); 1026 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 1027 } 1028 1029 // Now we know that each predecessor of this block has a value in 1030 // AvailablePreds, sort them for efficient access as we're walking the preds. 1031 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 1032 1033 // Create a PHI node at the start of the block for the PRE'd load value. 1034 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 1035 PN->takeName(LI); 1036 1037 // Insert new entries into the PHI for each predecessor. A single block may 1038 // have multiple entries here. 1039 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 1040 ++PI) { 1041 BasicBlock *P = *PI; 1042 AvailablePredsTy::iterator I = 1043 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 1044 std::make_pair(P, (Value*)0)); 1045 1046 assert(I != AvailablePreds.end() && I->first == P && 1047 "Didn't find entry for predecessor!"); 1048 1049 PN->addIncoming(I->second, I->first); 1050 } 1051 1052 //cerr << "PRE: " << *LI << *PN << "\n"; 1053 1054 LI->replaceAllUsesWith(PN); 1055 LI->eraseFromParent(); 1056 1057 return true; 1058} 1059 1060/// FindMostPopularDest - The specified list contains multiple possible 1061/// threadable destinations. Pick the one that occurs the most frequently in 1062/// the list. 1063static BasicBlock * 1064FindMostPopularDest(BasicBlock *BB, 1065 const SmallVectorImpl<std::pair<BasicBlock*, 1066 BasicBlock*> > &PredToDestList) { 1067 assert(!PredToDestList.empty()); 1068 1069 // Determine popularity. If there are multiple possible destinations, we 1070 // explicitly choose to ignore 'undef' destinations. We prefer to thread 1071 // blocks with known and real destinations to threading undef. We'll handle 1072 // them later if interesting. 1073 DenseMap<BasicBlock*, unsigned> DestPopularity; 1074 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1075 if (PredToDestList[i].second) 1076 DestPopularity[PredToDestList[i].second]++; 1077 1078 // Find the most popular dest. 1079 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 1080 BasicBlock *MostPopularDest = DPI->first; 1081 unsigned Popularity = DPI->second; 1082 SmallVector<BasicBlock*, 4> SamePopularity; 1083 1084 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 1085 // If the popularity of this entry isn't higher than the popularity we've 1086 // seen so far, ignore it. 1087 if (DPI->second < Popularity) 1088 ; // ignore. 1089 else if (DPI->second == Popularity) { 1090 // If it is the same as what we've seen so far, keep track of it. 1091 SamePopularity.push_back(DPI->first); 1092 } else { 1093 // If it is more popular, remember it. 1094 SamePopularity.clear(); 1095 MostPopularDest = DPI->first; 1096 Popularity = DPI->second; 1097 } 1098 } 1099 1100 // Okay, now we know the most popular destination. If there is more than 1101 // destination, we need to determine one. This is arbitrary, but we need 1102 // to make a deterministic decision. Pick the first one that appears in the 1103 // successor list. 1104 if (!SamePopularity.empty()) { 1105 SamePopularity.push_back(MostPopularDest); 1106 TerminatorInst *TI = BB->getTerminator(); 1107 for (unsigned i = 0; ; ++i) { 1108 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 1109 1110 if (std::find(SamePopularity.begin(), SamePopularity.end(), 1111 TI->getSuccessor(i)) == SamePopularity.end()) 1112 continue; 1113 1114 MostPopularDest = TI->getSuccessor(i); 1115 break; 1116 } 1117 } 1118 1119 // Okay, we have finally picked the most popular destination. 1120 return MostPopularDest; 1121} 1122 1123bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) { 1124 // If threading this would thread across a loop header, don't even try to 1125 // thread the edge. 1126 if (LoopHeaders.count(BB)) 1127 return false; 1128 1129 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 1130 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues)) 1131 return false; 1132 assert(!PredValues.empty() && 1133 "ComputeValueKnownInPredecessors returned true with no values"); 1134 1135 DEBUG(dbgs() << "IN BB: " << *BB; 1136 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1137 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "; 1138 if (PredValues[i].first) 1139 dbgs() << *PredValues[i].first; 1140 else 1141 dbgs() << "UNDEF"; 1142 dbgs() << " for pred '" << PredValues[i].second->getName() 1143 << "'.\n"; 1144 }); 1145 1146 // Decide what we want to thread through. Convert our list of known values to 1147 // a list of known destinations for each pred. This also discards duplicate 1148 // predecessors and keeps track of the undefined inputs (which are represented 1149 // as a null dest in the PredToDestList). 1150 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1151 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1152 1153 BasicBlock *OnlyDest = 0; 1154 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1155 1156 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1157 BasicBlock *Pred = PredValues[i].second; 1158 if (!SeenPreds.insert(Pred)) 1159 continue; // Duplicate predecessor entry. 1160 1161 // If the predecessor ends with an indirect goto, we can't change its 1162 // destination. 1163 if (isa<IndirectBrInst>(Pred->getTerminator())) 1164 continue; 1165 1166 ConstantInt *Val = PredValues[i].first; 1167 1168 BasicBlock *DestBB; 1169 if (Val == 0) // Undef. 1170 DestBB = 0; 1171 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1172 DestBB = BI->getSuccessor(Val->isZero()); 1173 else { 1174 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 1175 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 1176 } 1177 1178 // If we have exactly one destination, remember it for efficiency below. 1179 if (i == 0) 1180 OnlyDest = DestBB; 1181 else if (OnlyDest != DestBB) 1182 OnlyDest = MultipleDestSentinel; 1183 1184 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1185 } 1186 1187 // If all edges were unthreadable, we fail. 1188 if (PredToDestList.empty()) 1189 return false; 1190 1191 // Determine which is the most common successor. If we have many inputs and 1192 // this block is a switch, we want to start by threading the batch that goes 1193 // to the most popular destination first. If we only know about one 1194 // threadable destination (the common case) we can avoid this. 1195 BasicBlock *MostPopularDest = OnlyDest; 1196 1197 if (MostPopularDest == MultipleDestSentinel) 1198 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1199 1200 // Now that we know what the most popular destination is, factor all 1201 // predecessors that will jump to it into a single predecessor. 1202 SmallVector<BasicBlock*, 16> PredsToFactor; 1203 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1204 if (PredToDestList[i].second == MostPopularDest) { 1205 BasicBlock *Pred = PredToDestList[i].first; 1206 1207 // This predecessor may be a switch or something else that has multiple 1208 // edges to the block. Factor each of these edges by listing them 1209 // according to # occurrences in PredsToFactor. 1210 TerminatorInst *PredTI = Pred->getTerminator(); 1211 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1212 if (PredTI->getSuccessor(i) == BB) 1213 PredsToFactor.push_back(Pred); 1214 } 1215 1216 // If the threadable edges are branching on an undefined value, we get to pick 1217 // the destination that these predecessors should get to. 1218 if (MostPopularDest == 0) 1219 MostPopularDest = BB->getTerminator()-> 1220 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1221 1222 // Ok, try to thread it! 1223 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1224} 1225 1226/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1227/// a PHI node in the current block. See if there are any simplifications we 1228/// can do based on inputs to the phi node. 1229/// 1230bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1231 BasicBlock *BB = PN->getParent(); 1232 1233 // TODO: We could make use of this to do it once for blocks with common PHI 1234 // values. 1235 SmallVector<BasicBlock*, 1> PredBBs; 1236 PredBBs.resize(1); 1237 1238 // If any of the predecessor blocks end in an unconditional branch, we can 1239 // *duplicate* the conditional branch into that block in order to further 1240 // encourage jump threading and to eliminate cases where we have branch on a 1241 // phi of an icmp (branch on icmp is much better). 1242 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1243 BasicBlock *PredBB = PN->getIncomingBlock(i); 1244 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1245 if (PredBr->isUnconditional()) { 1246 PredBBs[0] = PredBB; 1247 // Try to duplicate BB into PredBB. 1248 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1249 return true; 1250 } 1251 } 1252 1253 return false; 1254} 1255 1256/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1257/// a xor instruction in the current block. See if there are any 1258/// simplifications we can do based on inputs to the xor. 1259/// 1260bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1261 BasicBlock *BB = BO->getParent(); 1262 1263 // If either the LHS or RHS of the xor is a constant, don't do this 1264 // optimization. 1265 if (isa<ConstantInt>(BO->getOperand(0)) || 1266 isa<ConstantInt>(BO->getOperand(1))) 1267 return false; 1268 1269 // If the first instruction in BB isn't a phi, we won't be able to infer 1270 // anything special about any particular predecessor. 1271 if (!isa<PHINode>(BB->front())) 1272 return false; 1273 1274 // If we have a xor as the branch input to this block, and we know that the 1275 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1276 // the condition into the predecessor and fix that value to true, saving some 1277 // logical ops on that path and encouraging other paths to simplify. 1278 // 1279 // This copies something like this: 1280 // 1281 // BB: 1282 // %X = phi i1 [1], [%X'] 1283 // %Y = icmp eq i32 %A, %B 1284 // %Z = xor i1 %X, %Y 1285 // br i1 %Z, ... 1286 // 1287 // Into: 1288 // BB': 1289 // %Y = icmp ne i32 %A, %B 1290 // br i1 %Z, ... 1291 1292 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues; 1293 bool isLHS = true; 1294 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) { 1295 assert(XorOpValues.empty()); 1296 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues)) 1297 return false; 1298 isLHS = false; 1299 } 1300 1301 assert(!XorOpValues.empty() && 1302 "ComputeValueKnownInPredecessors returned true with no values"); 1303 1304 // Scan the information to see which is most popular: true or false. The 1305 // predecessors can be of the set true, false, or undef. 1306 unsigned NumTrue = 0, NumFalse = 0; 1307 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1308 if (!XorOpValues[i].first) continue; // Ignore undefs for the count. 1309 if (XorOpValues[i].first->isZero()) 1310 ++NumFalse; 1311 else 1312 ++NumTrue; 1313 } 1314 1315 // Determine which value to split on, true, false, or undef if neither. 1316 ConstantInt *SplitVal = 0; 1317 if (NumTrue > NumFalse) 1318 SplitVal = ConstantInt::getTrue(BB->getContext()); 1319 else if (NumTrue != 0 || NumFalse != 0) 1320 SplitVal = ConstantInt::getFalse(BB->getContext()); 1321 1322 // Collect all of the blocks that this can be folded into so that we can 1323 // factor this once and clone it once. 1324 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1325 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1326 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue; 1327 1328 BlocksToFoldInto.push_back(XorOpValues[i].second); 1329 } 1330 1331 // If we inferred a value for all of the predecessors, then duplication won't 1332 // help us. However, we can just replace the LHS or RHS with the constant. 1333 if (BlocksToFoldInto.size() == 1334 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1335 if (SplitVal == 0) { 1336 // If all preds provide undef, just nuke the xor, because it is undef too. 1337 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1338 BO->eraseFromParent(); 1339 } else if (SplitVal->isZero()) { 1340 // If all preds provide 0, replace the xor with the other input. 1341 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1342 BO->eraseFromParent(); 1343 } else { 1344 // If all preds provide 1, set the computed value to 1. 1345 BO->setOperand(!isLHS, SplitVal); 1346 } 1347 1348 return true; 1349 } 1350 1351 // Try to duplicate BB into PredBB. 1352 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1353} 1354 1355 1356/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1357/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1358/// NewPred using the entries from OldPred (suitably mapped). 1359static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1360 BasicBlock *OldPred, 1361 BasicBlock *NewPred, 1362 DenseMap<Instruction*, Value*> &ValueMap) { 1363 for (BasicBlock::iterator PNI = PHIBB->begin(); 1364 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1365 // Ok, we have a PHI node. Figure out what the incoming value was for the 1366 // DestBlock. 1367 Value *IV = PN->getIncomingValueForBlock(OldPred); 1368 1369 // Remap the value if necessary. 1370 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1371 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1372 if (I != ValueMap.end()) 1373 IV = I->second; 1374 } 1375 1376 PN->addIncoming(IV, NewPred); 1377 } 1378} 1379 1380/// ThreadEdge - We have decided that it is safe and profitable to factor the 1381/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1382/// across BB. Transform the IR to reflect this change. 1383bool JumpThreading::ThreadEdge(BasicBlock *BB, 1384 const SmallVectorImpl<BasicBlock*> &PredBBs, 1385 BasicBlock *SuccBB) { 1386 // If threading to the same block as we come from, we would infinite loop. 1387 if (SuccBB == BB) { 1388 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1389 << "' - would thread to self!\n"); 1390 return false; 1391 } 1392 1393 // If threading this would thread across a loop header, don't thread the edge. 1394 // See the comments above FindLoopHeaders for justifications and caveats. 1395 if (LoopHeaders.count(BB)) { 1396 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1397 << "' to dest BB '" << SuccBB->getName() 1398 << "' - it might create an irreducible loop!\n"); 1399 return false; 1400 } 1401 1402 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1403 if (JumpThreadCost > Threshold) { 1404 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1405 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1406 return false; 1407 } 1408 1409 // And finally, do it! Start by factoring the predecessors is needed. 1410 BasicBlock *PredBB; 1411 if (PredBBs.size() == 1) 1412 PredBB = PredBBs[0]; 1413 else { 1414 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1415 << " common predecessors.\n"); 1416 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1417 ".thr_comm", this); 1418 } 1419 1420 // And finally, do it! 1421 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1422 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1423 << ", across block:\n " 1424 << *BB << "\n"); 1425 1426 if (LVI) 1427 LVI->threadEdge(PredBB, BB, SuccBB); 1428 1429 // We are going to have to map operands from the original BB block to the new 1430 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1431 // account for entry from PredBB. 1432 DenseMap<Instruction*, Value*> ValueMapping; 1433 1434 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1435 BB->getName()+".thread", 1436 BB->getParent(), BB); 1437 NewBB->moveAfter(PredBB); 1438 1439 BasicBlock::iterator BI = BB->begin(); 1440 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1441 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1442 1443 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1444 // mapping and using it to remap operands in the cloned instructions. 1445 for (; !isa<TerminatorInst>(BI); ++BI) { 1446 Instruction *New = BI->clone(); 1447 New->setName(BI->getName()); 1448 NewBB->getInstList().push_back(New); 1449 ValueMapping[BI] = New; 1450 1451 // Remap operands to patch up intra-block references. 1452 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1453 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1454 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1455 if (I != ValueMapping.end()) 1456 New->setOperand(i, I->second); 1457 } 1458 } 1459 1460 // We didn't copy the terminator from BB over to NewBB, because there is now 1461 // an unconditional jump to SuccBB. Insert the unconditional jump. 1462 BranchInst::Create(SuccBB, NewBB); 1463 1464 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1465 // PHI nodes for NewBB now. 1466 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1467 1468 // If there were values defined in BB that are used outside the block, then we 1469 // now have to update all uses of the value to use either the original value, 1470 // the cloned value, or some PHI derived value. This can require arbitrary 1471 // PHI insertion, of which we are prepared to do, clean these up now. 1472 SSAUpdater SSAUpdate; 1473 SmallVector<Use*, 16> UsesToRename; 1474 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1475 // Scan all uses of this instruction to see if it is used outside of its 1476 // block, and if so, record them in UsesToRename. 1477 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1478 ++UI) { 1479 Instruction *User = cast<Instruction>(*UI); 1480 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1481 if (UserPN->getIncomingBlock(UI) == BB) 1482 continue; 1483 } else if (User->getParent() == BB) 1484 continue; 1485 1486 UsesToRename.push_back(&UI.getUse()); 1487 } 1488 1489 // If there are no uses outside the block, we're done with this instruction. 1490 if (UsesToRename.empty()) 1491 continue; 1492 1493 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1494 1495 // We found a use of I outside of BB. Rename all uses of I that are outside 1496 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1497 // with the two values we know. 1498 SSAUpdate.Initialize(I); 1499 SSAUpdate.AddAvailableValue(BB, I); 1500 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1501 1502 while (!UsesToRename.empty()) 1503 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1504 DEBUG(dbgs() << "\n"); 1505 } 1506 1507 1508 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1509 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1510 // us to simplify any PHI nodes in BB. 1511 TerminatorInst *PredTerm = PredBB->getTerminator(); 1512 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1513 if (PredTerm->getSuccessor(i) == BB) { 1514 RemovePredecessorAndSimplify(BB, PredBB, TD); 1515 PredTerm->setSuccessor(i, NewBB); 1516 } 1517 1518 // At this point, the IR is fully up to date and consistent. Do a quick scan 1519 // over the new instructions and zap any that are constants or dead. This 1520 // frequently happens because of phi translation. 1521 SimplifyInstructionsInBlock(NewBB, TD); 1522 1523 // Threaded an edge! 1524 ++NumThreads; 1525 return true; 1526} 1527 1528/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1529/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1530/// If we can duplicate the contents of BB up into PredBB do so now, this 1531/// improves the odds that the branch will be on an analyzable instruction like 1532/// a compare. 1533bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1534 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1535 assert(!PredBBs.empty() && "Can't handle an empty set"); 1536 1537 // If BB is a loop header, then duplicating this block outside the loop would 1538 // cause us to transform this into an irreducible loop, don't do this. 1539 // See the comments above FindLoopHeaders for justifications and caveats. 1540 if (LoopHeaders.count(BB)) { 1541 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1542 << "' into predecessor block '" << PredBBs[0]->getName() 1543 << "' - it might create an irreducible loop!\n"); 1544 return false; 1545 } 1546 1547 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1548 if (DuplicationCost > Threshold) { 1549 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1550 << "' - Cost is too high: " << DuplicationCost << "\n"); 1551 return false; 1552 } 1553 1554 // And finally, do it! Start by factoring the predecessors is needed. 1555 BasicBlock *PredBB; 1556 if (PredBBs.size() == 1) 1557 PredBB = PredBBs[0]; 1558 else { 1559 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1560 << " common predecessors.\n"); 1561 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1562 ".thr_comm", this); 1563 } 1564 1565 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1566 // of PredBB. 1567 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1568 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1569 << DuplicationCost << " block is:" << *BB << "\n"); 1570 1571 // Unless PredBB ends with an unconditional branch, split the edge so that we 1572 // can just clone the bits from BB into the end of the new PredBB. 1573 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1574 1575 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { 1576 PredBB = SplitEdge(PredBB, BB, this); 1577 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1578 } 1579 1580 // We are going to have to map operands from the original BB block into the 1581 // PredBB block. Evaluate PHI nodes in BB. 1582 DenseMap<Instruction*, Value*> ValueMapping; 1583 1584 BasicBlock::iterator BI = BB->begin(); 1585 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1586 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1587 1588 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1589 // mapping and using it to remap operands in the cloned instructions. 1590 for (; BI != BB->end(); ++BI) { 1591 Instruction *New = BI->clone(); 1592 1593 // Remap operands to patch up intra-block references. 1594 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1595 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1596 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1597 if (I != ValueMapping.end()) 1598 New->setOperand(i, I->second); 1599 } 1600 1601 // If this instruction can be simplified after the operands are updated, 1602 // just use the simplified value instead. This frequently happens due to 1603 // phi translation. 1604 if (Value *IV = SimplifyInstruction(New, TD)) { 1605 delete New; 1606 ValueMapping[BI] = IV; 1607 } else { 1608 // Otherwise, insert the new instruction into the block. 1609 New->setName(BI->getName()); 1610 PredBB->getInstList().insert(OldPredBranch, New); 1611 ValueMapping[BI] = New; 1612 } 1613 } 1614 1615 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1616 // add entries to the PHI nodes for branch from PredBB now. 1617 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1618 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1619 ValueMapping); 1620 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1621 ValueMapping); 1622 1623 // If there were values defined in BB that are used outside the block, then we 1624 // now have to update all uses of the value to use either the original value, 1625 // the cloned value, or some PHI derived value. This can require arbitrary 1626 // PHI insertion, of which we are prepared to do, clean these up now. 1627 SSAUpdater SSAUpdate; 1628 SmallVector<Use*, 16> UsesToRename; 1629 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1630 // Scan all uses of this instruction to see if it is used outside of its 1631 // block, and if so, record them in UsesToRename. 1632 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1633 ++UI) { 1634 Instruction *User = cast<Instruction>(*UI); 1635 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1636 if (UserPN->getIncomingBlock(UI) == BB) 1637 continue; 1638 } else if (User->getParent() == BB) 1639 continue; 1640 1641 UsesToRename.push_back(&UI.getUse()); 1642 } 1643 1644 // If there are no uses outside the block, we're done with this instruction. 1645 if (UsesToRename.empty()) 1646 continue; 1647 1648 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1649 1650 // We found a use of I outside of BB. Rename all uses of I that are outside 1651 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1652 // with the two values we know. 1653 SSAUpdate.Initialize(I); 1654 SSAUpdate.AddAvailableValue(BB, I); 1655 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1656 1657 while (!UsesToRename.empty()) 1658 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1659 DEBUG(dbgs() << "\n"); 1660 } 1661 1662 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1663 // that we nuked. 1664 RemovePredecessorAndSimplify(BB, PredBB, TD); 1665 1666 // Remove the unconditional branch at the end of the PredBB block. 1667 OldPredBranch->eraseFromParent(); 1668 1669 ++NumDupes; 1670 return true; 1671} 1672 1673 1674