JumpThreading.cpp revision c2c23d03df415d064c7789349cd82e438f9d44bb
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//===----------------------------------------------------------------------===// 184 185 186/// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this 187/// method is called when we're about to delete Pred as a predecessor of BB. If 188/// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred. 189/// 190/// Unlike the removePredecessor method, this attempts to simplify uses of PHI 191/// nodes that collapse into identity values. For example, if we have: 192/// x = phi(1, 0, 0, 0) 193/// y = and x, z 194/// 195/// .. and delete the predecessor corresponding to the '1', this will attempt to 196/// recursively fold the and to 0. 197static void RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred, 198 TargetData *TD) { 199 // This only adjusts blocks with PHI nodes. 200 if (!isa<PHINode>(BB->begin())) 201 return; 202 203 // Remove the entries for Pred from the PHI nodes in BB, but do not simplify 204 // them down. This will leave us with single entry phi nodes and other phis 205 // that can be removed. 206 //BB->removePredecessor(Pred, true); 207 BB->removePredecessor(Pred, true); 208 209 WeakVH PhiIt = &BB->front(); 210 while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) { 211 PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt)); 212 213 Value *PNV = PN->hasConstantValue(); 214 if (PNV == 0) continue; 215 216 assert(PNV != PN && "hasConstantValue broken"); 217 218 // If we're able to simplify the phi to a constant, simplify it into its 219 // uses. 220 while (!PN->use_empty()) { 221 // Update the instruction to use the new value. 222 Use &U = PN->use_begin().getUse(); 223 Instruction *User = cast<Instruction>(U.getUser()); 224 U = PNV; 225 226 // See if we can simplify it (constant folding). 227 if (Constant *C = ConstantFoldInstruction(User, TD)) { 228 User->replaceAllUsesWith(C); 229 User->eraseFromParent(); 230 } 231 } 232 233 PN->replaceAllUsesWith(PNV); 234 PN->eraseFromParent(); 235 236 // If recursive simplification ended up deleting the next PHI node we would 237 // iterate to, then our iterator is invalid, restart scanning from the top 238 // of the block. 239 if (PhiIt == 0) PhiIt = &BB->front(); 240 } 241} 242 243//===----------------------------------------------------------------------===// 244 245 246/// FindLoopHeaders - We do not want jump threading to turn proper loop 247/// structures into irreducible loops. Doing this breaks up the loop nesting 248/// hierarchy and pessimizes later transformations. To prevent this from 249/// happening, we first have to find the loop headers. Here we approximate this 250/// by finding targets of backedges in the CFG. 251/// 252/// Note that there definitely are cases when we want to allow threading of 253/// edges across a loop header. For example, threading a jump from outside the 254/// loop (the preheader) to an exit block of the loop is definitely profitable. 255/// It is also almost always profitable to thread backedges from within the loop 256/// to exit blocks, and is often profitable to thread backedges to other blocks 257/// within the loop (forming a nested loop). This simple analysis is not rich 258/// enough to track all of these properties and keep it up-to-date as the CFG 259/// mutates, so we don't allow any of these transformations. 260/// 261void JumpThreading::FindLoopHeaders(Function &F) { 262 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 263 FindFunctionBackedges(F, Edges); 264 265 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 266 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 267} 268 269/// GetResultOfComparison - Given an icmp/fcmp predicate and the left and right 270/// hand sides of the compare instruction, try to determine the result. If the 271/// result can not be determined, a null pointer is returned. 272static Constant *GetResultOfComparison(CmpInst::Predicate pred, 273 Value *LHS, Value *RHS) { 274 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 275 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 276 return ConstantExpr::getCompare(pred, CLHS, CRHS); 277 278 if (LHS == RHS) 279 if (isa<IntegerType>(LHS->getType()) || isa<PointerType>(LHS->getType())) { 280 if (ICmpInst::isTrueWhenEqual(pred)) 281 return ConstantInt::getTrue(LHS->getContext()); 282 else 283 return ConstantInt::getFalse(LHS->getContext()); 284 } 285 return 0; 286} 287 288 289/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 290/// if we can infer that the value is a known ConstantInt in any of our 291/// predecessors. If so, return the known list of value and pred BB in the 292/// result vector. If a value is known to be undef, it is returned as null. 293/// 294/// The BB basic block is known to start with a PHI node. 295/// 296/// This returns true if there were any known values. 297/// 298/// 299/// TODO: Per PR2563, we could infer value range information about a predecessor 300/// based on its terminator. 301bool JumpThreading:: 302ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 303 PHINode *TheFirstPHI = cast<PHINode>(BB->begin()); 304 305 // If V is a constantint, then it is known in all predecessors. 306 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 307 ConstantInt *CI = dyn_cast<ConstantInt>(V); 308 Result.resize(TheFirstPHI->getNumIncomingValues()); 309 for (unsigned i = 0, e = Result.size(); i != e; ++i) 310 Result[i] = std::make_pair(CI, TheFirstPHI->getIncomingBlock(i)); 311 return true; 312 } 313 314 // If V is a non-instruction value, or an instruction in a different block, 315 // then it can't be derived from a PHI. 316 Instruction *I = dyn_cast<Instruction>(V); 317 if (I == 0 || I->getParent() != BB) 318 return false; 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 } 328 } 329 return !Result.empty(); 330 } 331 332 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 333 334 // Handle some boolean conditions. 335 if (I->getType()->getPrimitiveSizeInBits() == 1) { 336 // X | true -> true 337 // X & false -> false 338 if (I->getOpcode() == Instruction::Or || 339 I->getOpcode() == Instruction::And) { 340 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 341 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 342 343 if (LHSVals.empty() && RHSVals.empty()) 344 return false; 345 346 ConstantInt *InterestingVal; 347 if (I->getOpcode() == Instruction::Or) 348 InterestingVal = ConstantInt::getTrue(I->getContext()); 349 else 350 InterestingVal = ConstantInt::getFalse(I->getContext()); 351 352 // Scan for the sentinel. 353 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 354 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) 355 Result.push_back(LHSVals[i]); 356 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 357 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) 358 Result.push_back(RHSVals[i]); 359 return !Result.empty(); 360 } 361 362 // TODO: Should handle the NOT form of XOR. 363 364 } 365 366 // Handle compare with phi operand, where the PHI is defined in this block. 367 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 368 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 369 if (PN && PN->getParent() == BB) { 370 // We can do this simplification if any comparisons fold to true or false. 371 // See if any do. 372 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 373 BasicBlock *PredBB = PN->getIncomingBlock(i); 374 Value *LHS = PN->getIncomingValue(i); 375 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 376 377 Constant *Res = GetResultOfComparison(Cmp->getPredicate(), LHS, RHS); 378 if (Res == 0) continue; 379 380 if (isa<UndefValue>(Res)) 381 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 382 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 383 Result.push_back(std::make_pair(CI, PredBB)); 384 } 385 386 return !Result.empty(); 387 } 388 389 // TODO: We could also recurse to see if we can determine constants another 390 // way. 391 } 392 return false; 393} 394 395 396 397/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 398/// in an undefined jump, decide which block is best to revector to. 399/// 400/// Since we can pick an arbitrary destination, we pick the successor with the 401/// fewest predecessors. This should reduce the in-degree of the others. 402/// 403static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 404 TerminatorInst *BBTerm = BB->getTerminator(); 405 unsigned MinSucc = 0; 406 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 407 // Compute the successor with the minimum number of predecessors. 408 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 409 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 410 TestBB = BBTerm->getSuccessor(i); 411 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 412 if (NumPreds < MinNumPreds) 413 MinSucc = i; 414 } 415 416 return MinSucc; 417} 418 419/// ProcessBlock - If there are any predecessors whose control can be threaded 420/// through to a successor, transform them now. 421bool JumpThreading::ProcessBlock(BasicBlock *BB) { 422 // If this block has a single predecessor, and if that pred has a single 423 // successor, merge the blocks. This encourages recursive jump threading 424 // because now the condition in this block can be threaded through 425 // predecessors of our predecessor block. 426 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 427 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 428 SinglePred != BB) { 429 // If SinglePred was a loop header, BB becomes one. 430 if (LoopHeaders.erase(SinglePred)) 431 LoopHeaders.insert(BB); 432 433 // Remember if SinglePred was the entry block of the function. If so, we 434 // will need to move BB back to the entry position. 435 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 436 MergeBasicBlockIntoOnlyPred(BB); 437 438 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 439 BB->moveBefore(&BB->getParent()->getEntryBlock()); 440 return true; 441 } 442 } 443 444 // Look to see if the terminator is a branch of switch, if not we can't thread 445 // it. 446 Value *Condition; 447 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 448 // Can't thread an unconditional jump. 449 if (BI->isUnconditional()) return false; 450 Condition = BI->getCondition(); 451 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 452 Condition = SI->getCondition(); 453 else 454 return false; // Must be an invoke. 455 456 // If the terminator of this block is branching on a constant, simplify the 457 // terminator to an unconditional branch. This can occur due to threading in 458 // other blocks. 459 if (isa<ConstantInt>(Condition)) { 460 DEBUG(errs() << " In block '" << BB->getName() 461 << "' folding terminator: " << *BB->getTerminator() << '\n'); 462 ++NumFolds; 463 ConstantFoldTerminator(BB); 464 return true; 465 } 466 467 // If the terminator is branching on an undef, we can pick any of the 468 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 469 if (isa<UndefValue>(Condition)) { 470 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 471 472 // Fold the branch/switch. 473 TerminatorInst *BBTerm = BB->getTerminator(); 474 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 475 if (i == BestSucc) continue; 476 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD); 477 } 478 479 DEBUG(errs() << " In block '" << BB->getName() 480 << "' folding undef terminator: " << *BBTerm << '\n'); 481 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 482 BBTerm->eraseFromParent(); 483 return true; 484 } 485 486 Instruction *CondInst = dyn_cast<Instruction>(Condition); 487 488 // If the condition is an instruction defined in another block, see if a 489 // predecessor has the same condition: 490 // br COND, BBX, BBY 491 // BBX: 492 // br COND, BBZ, BBW 493 if (!Condition->hasOneUse() && // Multiple uses. 494 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 495 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 496 if (isa<BranchInst>(BB->getTerminator())) { 497 for (; PI != E; ++PI) 498 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 499 if (PBI->isConditional() && PBI->getCondition() == Condition && 500 ProcessBranchOnDuplicateCond(*PI, BB)) 501 return true; 502 } else { 503 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 504 for (; PI != E; ++PI) 505 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator())) 506 if (PSI->getCondition() == Condition && 507 ProcessSwitchOnDuplicateCond(*PI, BB)) 508 return true; 509 } 510 } 511 512 // All the rest of our checks depend on the condition being an instruction. 513 if (CondInst == 0) 514 return false; 515 516 // See if this is a phi node in the current block. 517 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 518 if (PN->getParent() == BB) 519 return ProcessJumpOnPHI(PN); 520 521 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 522 if (!isa<PHINode>(CondCmp->getOperand(0)) || 523 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB) { 524 // If we have a comparison, loop over the predecessors to see if there is 525 // a condition with a lexically identical value. 526 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 527 for (; PI != E; ++PI) 528 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 529 if (PBI->isConditional() && *PI != BB) { 530 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 531 if (CI->getOperand(0) == CondCmp->getOperand(0) && 532 CI->getOperand(1) == CondCmp->getOperand(1) && 533 CI->getPredicate() == CondCmp->getPredicate()) { 534 // TODO: Could handle things like (x != 4) --> (x == 17) 535 if (ProcessBranchOnDuplicateCond(*PI, BB)) 536 return true; 537 } 538 } 539 } 540 } 541 } 542 543 // Check for some cases that are worth simplifying. Right now we want to look 544 // for loads that are used by a switch or by the condition for the branch. If 545 // we see one, check to see if it's partially redundant. If so, insert a PHI 546 // which can then be used to thread the values. 547 // 548 // This is particularly important because reg2mem inserts loads and stores all 549 // over the place, and this blocks jump threading if we don't zap them. 550 Value *SimplifyValue = CondInst; 551 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 552 if (isa<Constant>(CondCmp->getOperand(1))) 553 SimplifyValue = CondCmp->getOperand(0); 554 555 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 556 if (SimplifyPartiallyRedundantLoad(LI)) 557 return true; 558 559 560 // Handle a variety of cases where we are branching on something derived from 561 // a PHI node in the current block. If we can prove that any predecessors 562 // compute a predictable value based on a PHI node, thread those predecessors. 563 // 564 // We only bother doing this if the current block has a PHI node and if the 565 // conditional instruction lives in the current block. If either condition 566 // fails, this won't be a computable value anyway. 567 if (CondInst->getParent() == BB && isa<PHINode>(BB->front())) 568 if (ProcessThreadableEdges(CondInst, BB)) 569 return true; 570 571 572 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 573 // "(X == 4)" thread through this block. 574 575 return false; 576} 577 578/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 579/// block that jump on exactly the same condition. This means that we almost 580/// always know the direction of the edge in the DESTBB: 581/// PREDBB: 582/// br COND, DESTBB, BBY 583/// DESTBB: 584/// br COND, BBZ, BBW 585/// 586/// If DESTBB has multiple predecessors, we can't just constant fold the branch 587/// in DESTBB, we have to thread over it. 588bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 589 BasicBlock *BB) { 590 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 591 592 // If both successors of PredBB go to DESTBB, we don't know anything. We can 593 // fold the branch to an unconditional one, which allows other recursive 594 // simplifications. 595 bool BranchDir; 596 if (PredBI->getSuccessor(1) != BB) 597 BranchDir = true; 598 else if (PredBI->getSuccessor(0) != BB) 599 BranchDir = false; 600 else { 601 DEBUG(errs() << " In block '" << PredBB->getName() 602 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 603 ++NumFolds; 604 ConstantFoldTerminator(PredBB); 605 return true; 606 } 607 608 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 609 610 // If the dest block has one predecessor, just fix the branch condition to a 611 // constant and fold it. 612 if (BB->getSinglePredecessor()) { 613 DEBUG(errs() << " In block '" << BB->getName() 614 << "' folding condition to '" << BranchDir << "': " 615 << *BB->getTerminator() << '\n'); 616 ++NumFolds; 617 Value *OldCond = DestBI->getCondition(); 618 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 619 BranchDir)); 620 ConstantFoldTerminator(BB); 621 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 622 return true; 623 } 624 625 626 // Next, figure out which successor we are threading to. 627 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 628 629 SmallVector<BasicBlock*, 2> Preds; 630 Preds.push_back(PredBB); 631 632 // Ok, try to thread it! 633 return ThreadEdge(BB, Preds, SuccBB); 634} 635 636/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 637/// block that switch on exactly the same condition. This means that we almost 638/// always know the direction of the edge in the DESTBB: 639/// PREDBB: 640/// switch COND [... DESTBB, BBY ... ] 641/// DESTBB: 642/// switch COND [... BBZ, BBW ] 643/// 644/// Optimizing switches like this is very important, because simplifycfg builds 645/// switches out of repeated 'if' conditions. 646bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 647 BasicBlock *DestBB) { 648 // Can't thread edge to self. 649 if (PredBB == DestBB) 650 return false; 651 652 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 653 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 654 655 // There are a variety of optimizations that we can potentially do on these 656 // blocks: we order them from most to least preferable. 657 658 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 659 // directly to their destination. This does not introduce *any* code size 660 // growth. Skip debug info first. 661 BasicBlock::iterator BBI = DestBB->begin(); 662 while (isa<DbgInfoIntrinsic>(BBI)) 663 BBI++; 664 665 // FIXME: Thread if it just contains a PHI. 666 if (isa<SwitchInst>(BBI)) { 667 bool MadeChange = false; 668 // Ignore the default edge for now. 669 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 670 ConstantInt *DestVal = DestSI->getCaseValue(i); 671 BasicBlock *DestSucc = DestSI->getSuccessor(i); 672 673 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 674 // PredSI has an explicit case for it. If so, forward. If it is covered 675 // by the default case, we can't update PredSI. 676 unsigned PredCase = PredSI->findCaseValue(DestVal); 677 if (PredCase == 0) continue; 678 679 // If PredSI doesn't go to DestBB on this value, then it won't reach the 680 // case on this condition. 681 if (PredSI->getSuccessor(PredCase) != DestBB && 682 DestSI->getSuccessor(i) != DestBB) 683 continue; 684 685 // Otherwise, we're safe to make the change. Make sure that the edge from 686 // DestSI to DestSucc is not critical and has no PHI nodes. 687 DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 688 DEBUG(errs() << "THROUGH: " << *DestSI); 689 690 // If the destination has PHI nodes, just split the edge for updating 691 // simplicity. 692 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 693 SplitCriticalEdge(DestSI, i, this); 694 DestSucc = DestSI->getSuccessor(i); 695 } 696 FoldSingleEntryPHINodes(DestSucc); 697 PredSI->setSuccessor(PredCase, DestSucc); 698 MadeChange = true; 699 } 700 701 if (MadeChange) 702 return true; 703 } 704 705 return false; 706} 707 708 709/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 710/// load instruction, eliminate it by replacing it with a PHI node. This is an 711/// important optimization that encourages jump threading, and needs to be run 712/// interlaced with other jump threading tasks. 713bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 714 // Don't hack volatile loads. 715 if (LI->isVolatile()) return false; 716 717 // If the load is defined in a block with exactly one predecessor, it can't be 718 // partially redundant. 719 BasicBlock *LoadBB = LI->getParent(); 720 if (LoadBB->getSinglePredecessor()) 721 return false; 722 723 Value *LoadedPtr = LI->getOperand(0); 724 725 // If the loaded operand is defined in the LoadBB, it can't be available. 726 // FIXME: Could do PHI translation, that would be fun :) 727 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 728 if (PtrOp->getParent() == LoadBB) 729 return false; 730 731 // Scan a few instructions up from the load, to see if it is obviously live at 732 // the entry to its block. 733 BasicBlock::iterator BBIt = LI; 734 735 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB, 736 BBIt, 6)) { 737 // If the value if the load is locally available within the block, just use 738 // it. This frequently occurs for reg2mem'd allocas. 739 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 740 741 // If the returned value is the load itself, replace with an undef. This can 742 // only happen in dead loops. 743 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 744 LI->replaceAllUsesWith(AvailableVal); 745 LI->eraseFromParent(); 746 return true; 747 } 748 749 // Otherwise, if we scanned the whole block and got to the top of the block, 750 // we know the block is locally transparent to the load. If not, something 751 // might clobber its value. 752 if (BBIt != LoadBB->begin()) 753 return false; 754 755 756 SmallPtrSet<BasicBlock*, 8> PredsScanned; 757 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 758 AvailablePredsTy AvailablePreds; 759 BasicBlock *OneUnavailablePred = 0; 760 761 // If we got here, the loaded value is transparent through to the start of the 762 // block. Check to see if it is available in any of the predecessor blocks. 763 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 764 PI != PE; ++PI) { 765 BasicBlock *PredBB = *PI; 766 767 // If we already scanned this predecessor, skip it. 768 if (!PredsScanned.insert(PredBB)) 769 continue; 770 771 // Scan the predecessor to see if the value is available in the pred. 772 BBIt = PredBB->end(); 773 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 774 if (!PredAvailable) { 775 OneUnavailablePred = PredBB; 776 continue; 777 } 778 779 // If so, this load is partially redundant. Remember this info so that we 780 // can create a PHI node. 781 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 782 } 783 784 // If the loaded value isn't available in any predecessor, it isn't partially 785 // redundant. 786 if (AvailablePreds.empty()) return false; 787 788 // Okay, the loaded value is available in at least one (and maybe all!) 789 // predecessors. If the value is unavailable in more than one unique 790 // predecessor, we want to insert a merge block for those common predecessors. 791 // This ensures that we only have to insert one reload, thus not increasing 792 // code size. 793 BasicBlock *UnavailablePred = 0; 794 795 // If there is exactly one predecessor where the value is unavailable, the 796 // already computed 'OneUnavailablePred' block is it. If it ends in an 797 // unconditional branch, we know that it isn't a critical edge. 798 if (PredsScanned.size() == AvailablePreds.size()+1 && 799 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 800 UnavailablePred = OneUnavailablePred; 801 } else if (PredsScanned.size() != AvailablePreds.size()) { 802 // Otherwise, we had multiple unavailable predecessors or we had a critical 803 // edge from the one. 804 SmallVector<BasicBlock*, 8> PredsToSplit; 805 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 806 807 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 808 AvailablePredSet.insert(AvailablePreds[i].first); 809 810 // Add all the unavailable predecessors to the PredsToSplit list. 811 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 812 PI != PE; ++PI) 813 if (!AvailablePredSet.count(*PI)) 814 PredsToSplit.push_back(*PI); 815 816 // Split them out to their own block. 817 UnavailablePred = 818 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 819 "thread-split", this); 820 } 821 822 // If the value isn't available in all predecessors, then there will be 823 // exactly one where it isn't available. Insert a load on that edge and add 824 // it to the AvailablePreds list. 825 if (UnavailablePred) { 826 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 827 "Can't handle critical edge here!"); 828 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", 829 UnavailablePred->getTerminator()); 830 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 831 } 832 833 // Now we know that each predecessor of this block has a value in 834 // AvailablePreds, sort them for efficient access as we're walking the preds. 835 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 836 837 // Create a PHI node at the start of the block for the PRE'd load value. 838 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 839 PN->takeName(LI); 840 841 // Insert new entries into the PHI for each predecessor. A single block may 842 // have multiple entries here. 843 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 844 ++PI) { 845 AvailablePredsTy::iterator I = 846 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 847 std::make_pair(*PI, (Value*)0)); 848 849 assert(I != AvailablePreds.end() && I->first == *PI && 850 "Didn't find entry for predecessor!"); 851 852 PN->addIncoming(I->second, I->first); 853 } 854 855 //cerr << "PRE: " << *LI << *PN << "\n"; 856 857 LI->replaceAllUsesWith(PN); 858 LI->eraseFromParent(); 859 860 return true; 861} 862 863/// FindMostPopularDest - The specified list contains multiple possible 864/// threadable destinations. Pick the one that occurs the most frequently in 865/// the list. 866static BasicBlock * 867FindMostPopularDest(BasicBlock *BB, 868 const SmallVectorImpl<std::pair<BasicBlock*, 869 BasicBlock*> > &PredToDestList) { 870 assert(!PredToDestList.empty()); 871 872 // Determine popularity. If there are multiple possible destinations, we 873 // explicitly choose to ignore 'undef' destinations. We prefer to thread 874 // blocks with known and real destinations to threading undef. We'll handle 875 // them later if interesting. 876 DenseMap<BasicBlock*, unsigned> DestPopularity; 877 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 878 if (PredToDestList[i].second) 879 DestPopularity[PredToDestList[i].second]++; 880 881 // Find the most popular dest. 882 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 883 BasicBlock *MostPopularDest = DPI->first; 884 unsigned Popularity = DPI->second; 885 SmallVector<BasicBlock*, 4> SamePopularity; 886 887 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 888 // If the popularity of this entry isn't higher than the popularity we've 889 // seen so far, ignore it. 890 if (DPI->second < Popularity) 891 ; // ignore. 892 else if (DPI->second == Popularity) { 893 // If it is the same as what we've seen so far, keep track of it. 894 SamePopularity.push_back(DPI->first); 895 } else { 896 // If it is more popular, remember it. 897 SamePopularity.clear(); 898 MostPopularDest = DPI->first; 899 Popularity = DPI->second; 900 } 901 } 902 903 // Okay, now we know the most popular destination. If there is more than 904 // destination, we need to determine one. This is arbitrary, but we need 905 // to make a deterministic decision. Pick the first one that appears in the 906 // successor list. 907 if (!SamePopularity.empty()) { 908 SamePopularity.push_back(MostPopularDest); 909 TerminatorInst *TI = BB->getTerminator(); 910 for (unsigned i = 0; ; ++i) { 911 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 912 913 if (std::find(SamePopularity.begin(), SamePopularity.end(), 914 TI->getSuccessor(i)) == SamePopularity.end()) 915 continue; 916 917 MostPopularDest = TI->getSuccessor(i); 918 break; 919 } 920 } 921 922 // Okay, we have finally picked the most popular destination. 923 return MostPopularDest; 924} 925 926bool JumpThreading::ProcessThreadableEdges(Instruction *CondInst, 927 BasicBlock *BB) { 928 // If threading this would thread across a loop header, don't even try to 929 // thread the edge. 930 if (LoopHeaders.count(BB)) 931 return false; 932 933 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 934 if (!ComputeValueKnownInPredecessors(CondInst, BB, PredValues)) 935 return false; 936 assert(!PredValues.empty() && 937 "ComputeValueKnownInPredecessors returned true with no values"); 938 939 DEBUG(errs() << "IN BB: " << *BB; 940 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 941 errs() << " BB '" << BB->getName() << "': FOUND condition = "; 942 if (PredValues[i].first) 943 errs() << *PredValues[i].first; 944 else 945 errs() << "UNDEF"; 946 errs() << " for pred '" << PredValues[i].second->getName() 947 << "'.\n"; 948 }); 949 950 // Decide what we want to thread through. Convert our list of known values to 951 // a list of known destinations for each pred. This also discards duplicate 952 // predecessors and keeps track of the undefined inputs (which are represented 953 // as a null dest in the PredToDestList). 954 SmallPtrSet<BasicBlock*, 16> SeenPreds; 955 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 956 957 BasicBlock *OnlyDest = 0; 958 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 959 960 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 961 BasicBlock *Pred = PredValues[i].second; 962 if (!SeenPreds.insert(Pred)) 963 continue; // Duplicate predecessor entry. 964 965 // If the predecessor ends with an indirect goto, we can't change its 966 // destination. 967 if (isa<IndirectBrInst>(Pred->getTerminator())) 968 continue; 969 970 ConstantInt *Val = PredValues[i].first; 971 972 BasicBlock *DestBB; 973 if (Val == 0) // Undef. 974 DestBB = 0; 975 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 976 DestBB = BI->getSuccessor(Val->isZero()); 977 else { 978 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 979 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 980 } 981 982 // If we have exactly one destination, remember it for efficiency below. 983 if (i == 0) 984 OnlyDest = DestBB; 985 else if (OnlyDest != DestBB) 986 OnlyDest = MultipleDestSentinel; 987 988 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 989 } 990 991 // If all edges were unthreadable, we fail. 992 if (PredToDestList.empty()) 993 return false; 994 995 // Determine which is the most common successor. If we have many inputs and 996 // this block is a switch, we want to start by threading the batch that goes 997 // to the most popular destination first. If we only know about one 998 // threadable destination (the common case) we can avoid this. 999 BasicBlock *MostPopularDest = OnlyDest; 1000 1001 if (MostPopularDest == MultipleDestSentinel) 1002 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1003 1004 // Now that we know what the most popular destination is, factor all 1005 // predecessors that will jump to it into a single predecessor. 1006 SmallVector<BasicBlock*, 16> PredsToFactor; 1007 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1008 if (PredToDestList[i].second == MostPopularDest) { 1009 BasicBlock *Pred = PredToDestList[i].first; 1010 1011 // This predecessor may be a switch or something else that has multiple 1012 // edges to the block. Factor each of these edges by listing them 1013 // according to # occurrences in PredsToFactor. 1014 TerminatorInst *PredTI = Pred->getTerminator(); 1015 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1016 if (PredTI->getSuccessor(i) == BB) 1017 PredsToFactor.push_back(Pred); 1018 } 1019 1020 // If the threadable edges are branching on an undefined value, we get to pick 1021 // the destination that these predecessors should get to. 1022 if (MostPopularDest == 0) 1023 MostPopularDest = BB->getTerminator()-> 1024 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1025 1026 // Ok, try to thread it! 1027 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1028} 1029 1030/// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in 1031/// the current block. See if there are any simplifications we can do based on 1032/// inputs to the phi node. 1033/// 1034bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) { 1035 BasicBlock *BB = PN->getParent(); 1036 1037 // If any of the predecessor blocks end in an unconditional branch, we can 1038 // *duplicate* the jump into that block in order to further encourage jump 1039 // threading and to eliminate cases where we have branch on a phi of an icmp 1040 // (branch on icmp is much better). 1041 1042 // We don't want to do this tranformation for switches, because we don't 1043 // really want to duplicate a switch. 1044 if (isa<SwitchInst>(BB->getTerminator())) 1045 return false; 1046 1047 // Look for unconditional branch predecessors. 1048 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1049 BasicBlock *PredBB = PN->getIncomingBlock(i); 1050 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1051 if (PredBr->isUnconditional() && 1052 // Try to duplicate BB into PredBB. 1053 DuplicateCondBranchOnPHIIntoPred(BB, PredBB)) 1054 return true; 1055 } 1056 1057 return false; 1058} 1059 1060 1061/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1062/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1063/// NewPred using the entries from OldPred (suitably mapped). 1064static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1065 BasicBlock *OldPred, 1066 BasicBlock *NewPred, 1067 DenseMap<Instruction*, Value*> &ValueMap) { 1068 for (BasicBlock::iterator PNI = PHIBB->begin(); 1069 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1070 // Ok, we have a PHI node. Figure out what the incoming value was for the 1071 // DestBlock. 1072 Value *IV = PN->getIncomingValueForBlock(OldPred); 1073 1074 // Remap the value if necessary. 1075 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1076 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1077 if (I != ValueMap.end()) 1078 IV = I->second; 1079 } 1080 1081 PN->addIncoming(IV, NewPred); 1082 } 1083} 1084 1085/// ThreadEdge - We have decided that it is safe and profitable to factor the 1086/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1087/// across BB. Transform the IR to reflect this change. 1088bool JumpThreading::ThreadEdge(BasicBlock *BB, 1089 const SmallVectorImpl<BasicBlock*> &PredBBs, 1090 BasicBlock *SuccBB) { 1091 // If threading to the same block as we come from, we would infinite loop. 1092 if (SuccBB == BB) { 1093 DEBUG(errs() << " Not threading across BB '" << BB->getName() 1094 << "' - would thread to self!\n"); 1095 return false; 1096 } 1097 1098 // If threading this would thread across a loop header, don't thread the edge. 1099 // See the comments above FindLoopHeaders for justifications and caveats. 1100 if (LoopHeaders.count(BB)) { 1101 DEBUG(errs() << " Not threading across loop header BB '" << BB->getName() 1102 << "' to dest BB '" << SuccBB->getName() 1103 << "' - it might create an irreducible loop!\n"); 1104 return false; 1105 } 1106 1107 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1108 if (JumpThreadCost > Threshold) { 1109 DEBUG(errs() << " Not threading BB '" << BB->getName() 1110 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1111 return false; 1112 } 1113 1114 // And finally, do it! Start by factoring the predecessors is needed. 1115 BasicBlock *PredBB; 1116 if (PredBBs.size() == 1) 1117 PredBB = PredBBs[0]; 1118 else { 1119 DEBUG(errs() << " Factoring out " << PredBBs.size() 1120 << " common predecessors.\n"); 1121 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1122 ".thr_comm", this); 1123 } 1124 1125 // And finally, do it! 1126 DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '" 1127 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1128 << ", across block:\n " 1129 << *BB << "\n"); 1130 1131 // We are going to have to map operands from the original BB block to the new 1132 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1133 // account for entry from PredBB. 1134 DenseMap<Instruction*, Value*> ValueMapping; 1135 1136 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1137 BB->getName()+".thread", 1138 BB->getParent(), BB); 1139 NewBB->moveAfter(PredBB); 1140 1141 BasicBlock::iterator BI = BB->begin(); 1142 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1143 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1144 1145 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1146 // mapping and using it to remap operands in the cloned instructions. 1147 for (; !isa<TerminatorInst>(BI); ++BI) { 1148 Instruction *New = BI->clone(); 1149 New->setName(BI->getName()); 1150 NewBB->getInstList().push_back(New); 1151 ValueMapping[BI] = New; 1152 1153 // Remap operands to patch up intra-block references. 1154 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1155 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1156 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1157 if (I != ValueMapping.end()) 1158 New->setOperand(i, I->second); 1159 } 1160 } 1161 1162 // We didn't copy the terminator from BB over to NewBB, because there is now 1163 // an unconditional jump to SuccBB. Insert the unconditional jump. 1164 BranchInst::Create(SuccBB, NewBB); 1165 1166 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1167 // PHI nodes for NewBB now. 1168 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1169 1170 // If there were values defined in BB that are used outside the block, then we 1171 // now have to update all uses of the value to use either the original value, 1172 // the cloned value, or some PHI derived value. This can require arbitrary 1173 // PHI insertion, of which we are prepared to do, clean these up now. 1174 SSAUpdater SSAUpdate; 1175 SmallVector<Use*, 16> UsesToRename; 1176 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1177 // Scan all uses of this instruction to see if it is used outside of its 1178 // block, and if so, record them in UsesToRename. 1179 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1180 ++UI) { 1181 Instruction *User = cast<Instruction>(*UI); 1182 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1183 if (UserPN->getIncomingBlock(UI) == BB) 1184 continue; 1185 } else if (User->getParent() == BB) 1186 continue; 1187 1188 UsesToRename.push_back(&UI.getUse()); 1189 } 1190 1191 // If there are no uses outside the block, we're done with this instruction. 1192 if (UsesToRename.empty()) 1193 continue; 1194 1195 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1196 1197 // We found a use of I outside of BB. Rename all uses of I that are outside 1198 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1199 // with the two values we know. 1200 SSAUpdate.Initialize(I); 1201 SSAUpdate.AddAvailableValue(BB, I); 1202 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1203 1204 while (!UsesToRename.empty()) 1205 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1206 DEBUG(errs() << "\n"); 1207 } 1208 1209 1210 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1211 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1212 // us to simplify any PHI nodes in BB. 1213 TerminatorInst *PredTerm = PredBB->getTerminator(); 1214 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1215 if (PredTerm->getSuccessor(i) == BB) { 1216 RemovePredecessorAndSimplify(BB, PredBB, TD); 1217 PredTerm->setSuccessor(i, NewBB); 1218 } 1219 1220 // At this point, the IR is fully up to date and consistent. Do a quick scan 1221 // over the new instructions and zap any that are constants or dead. This 1222 // frequently happens because of phi translation. 1223 BI = NewBB->begin(); 1224 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) { 1225 Instruction *Inst = BI++; 1226 if (Constant *C = ConstantFoldInstruction(Inst, TD)) { 1227 Inst->replaceAllUsesWith(C); 1228 Inst->eraseFromParent(); 1229 continue; 1230 } 1231 1232 RecursivelyDeleteTriviallyDeadInstructions(Inst); 1233 } 1234 1235 // Threaded an edge! 1236 ++NumThreads; 1237 return true; 1238} 1239 1240/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1241/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1242/// If we can duplicate the contents of BB up into PredBB do so now, this 1243/// improves the odds that the branch will be on an analyzable instruction like 1244/// a compare. 1245bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1246 BasicBlock *PredBB) { 1247 // If BB is a loop header, then duplicating this block outside the loop would 1248 // cause us to transform this into an irreducible loop, don't do this. 1249 // See the comments above FindLoopHeaders for justifications and caveats. 1250 if (LoopHeaders.count(BB)) { 1251 DEBUG(errs() << " Not duplicating loop header '" << BB->getName() 1252 << "' into predecessor block '" << PredBB->getName() 1253 << "' - it might create an irreducible loop!\n"); 1254 return false; 1255 } 1256 1257 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1258 if (DuplicationCost > Threshold) { 1259 DEBUG(errs() << " Not duplicating BB '" << BB->getName() 1260 << "' - Cost is too high: " << DuplicationCost << "\n"); 1261 return false; 1262 } 1263 1264 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1265 // of PredBB. 1266 DEBUG(errs() << " Duplicating block '" << BB->getName() << "' into end of '" 1267 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1268 << DuplicationCost << " block is:" << *BB << "\n"); 1269 1270 // We are going to have to map operands from the original BB block into the 1271 // PredBB block. Evaluate PHI nodes in BB. 1272 DenseMap<Instruction*, Value*> ValueMapping; 1273 1274 BasicBlock::iterator BI = BB->begin(); 1275 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1276 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1277 1278 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1279 1280 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1281 // mapping and using it to remap operands in the cloned instructions. 1282 for (; BI != BB->end(); ++BI) { 1283 Instruction *New = BI->clone(); 1284 New->setName(BI->getName()); 1285 PredBB->getInstList().insert(OldPredBranch, New); 1286 ValueMapping[BI] = New; 1287 1288 // Remap operands to patch up intra-block references. 1289 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1290 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1291 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1292 if (I != ValueMapping.end()) 1293 New->setOperand(i, I->second); 1294 } 1295 } 1296 1297 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1298 // add entries to the PHI nodes for branch from PredBB now. 1299 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1300 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1301 ValueMapping); 1302 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1303 ValueMapping); 1304 1305 // If there were values defined in BB that are used outside the block, then we 1306 // now have to update all uses of the value to use either the original value, 1307 // the cloned value, or some PHI derived value. This can require arbitrary 1308 // PHI insertion, of which we are prepared to do, clean these up now. 1309 SSAUpdater SSAUpdate; 1310 SmallVector<Use*, 16> UsesToRename; 1311 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1312 // Scan all uses of this instruction to see if it is used outside of its 1313 // block, and if so, record them in UsesToRename. 1314 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1315 ++UI) { 1316 Instruction *User = cast<Instruction>(*UI); 1317 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1318 if (UserPN->getIncomingBlock(UI) == BB) 1319 continue; 1320 } else if (User->getParent() == BB) 1321 continue; 1322 1323 UsesToRename.push_back(&UI.getUse()); 1324 } 1325 1326 // If there are no uses outside the block, we're done with this instruction. 1327 if (UsesToRename.empty()) 1328 continue; 1329 1330 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1331 1332 // We found a use of I outside of BB. Rename all uses of I that are outside 1333 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1334 // with the two values we know. 1335 SSAUpdate.Initialize(I); 1336 SSAUpdate.AddAvailableValue(BB, I); 1337 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1338 1339 while (!UsesToRename.empty()) 1340 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1341 DEBUG(errs() << "\n"); 1342 } 1343 1344 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1345 // that we nuked. 1346 RemovePredecessorAndSimplify(BB, PredBB, TD); 1347 1348 // Remove the unconditional branch at the end of the PredBB block. 1349 OldPredBranch->eraseFromParent(); 1350 1351 ++NumDupes; 1352 return true; 1353} 1354 1355 1356