JumpThreading.cpp revision 50591ab1ea2e8d483fca5c6984f47edd6e726f10
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, BasicBlock *PredBB, BasicBlock *SuccBB); 76 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 77 BasicBlock *PredBB); 78 79 typedef SmallVectorImpl<std::pair<ConstantInt*, 80 BasicBlock*> > PredValueInfo; 81 82 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 83 PredValueInfo &Result); 84 bool ProcessThreadableEdges(Instruction *CondInst, BasicBlock *BB); 85 86 87 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 88 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 89 90 bool ProcessJumpOnPHI(PHINode *PN); 91 92 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 93 }; 94} 95 96char JumpThreading::ID = 0; 97static RegisterPass<JumpThreading> 98X("jump-threading", "Jump Threading"); 99 100// Public interface to the Jump Threading pass 101FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 102 103/// runOnFunction - Top level algorithm. 104/// 105bool JumpThreading::runOnFunction(Function &F) { 106 DEBUG(errs() << "Jump threading on function '" << F.getName() << "'\n"); 107 TD = getAnalysisIfAvailable<TargetData>(); 108 109 FindLoopHeaders(F); 110 111 bool AnotherIteration = true, EverChanged = false; 112 while (AnotherIteration) { 113 AnotherIteration = false; 114 bool Changed = false; 115 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 116 BasicBlock *BB = I; 117 while (ProcessBlock(BB)) 118 Changed = true; 119 120 ++I; 121 122 // If the block is trivially dead, zap it. This eliminates the successor 123 // edges which simplifies the CFG. 124 if (pred_begin(BB) == pred_end(BB) && 125 BB != &BB->getParent()->getEntryBlock()) { 126 DEBUG(errs() << " JT: Deleting dead block '" << BB->getName() 127 << "' with terminator: " << *BB->getTerminator() << '\n'); 128 LoopHeaders.erase(BB); 129 DeleteDeadBlock(BB); 130 Changed = true; 131 } 132 } 133 AnotherIteration = Changed; 134 EverChanged |= Changed; 135 } 136 137 LoopHeaders.clear(); 138 return EverChanged; 139} 140 141/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 142/// thread across it. 143static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 144 /// Ignore PHI nodes, these will be flattened when duplication happens. 145 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 146 147 // Sum up the cost of each instruction until we get to the terminator. Don't 148 // include the terminator because the copy won't include it. 149 unsigned Size = 0; 150 for (; !isa<TerminatorInst>(I); ++I) { 151 // Debugger intrinsics don't incur code size. 152 if (isa<DbgInfoIntrinsic>(I)) continue; 153 154 // If this is a pointer->pointer bitcast, it is free. 155 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType())) 156 continue; 157 158 // All other instructions count for at least one unit. 159 ++Size; 160 161 // Calls are more expensive. If they are non-intrinsic calls, we model them 162 // as having cost of 4. If they are a non-vector intrinsic, we model them 163 // as having cost of 2 total, and if they are a vector intrinsic, we model 164 // them as having cost 1. 165 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 166 if (!isa<IntrinsicInst>(CI)) 167 Size += 3; 168 else if (!isa<VectorType>(CI->getType())) 169 Size += 1; 170 } 171 } 172 173 // Threading through a switch statement is particularly profitable. If this 174 // block ends in a switch, decrease its cost to make it more likely to happen. 175 if (isa<SwitchInst>(I)) 176 Size = Size > 6 ? Size-6 : 0; 177 178 return Size; 179} 180 181 182 183/// FindLoopHeaders - We do not want jump threading to turn proper loop 184/// structures into irreducible loops. Doing this breaks up the loop nesting 185/// hierarchy and pessimizes later transformations. To prevent this from 186/// happening, we first have to find the loop headers. Here we approximate this 187/// by finding targets of backedges in the CFG. 188/// 189/// Note that there definitely are cases when we want to allow threading of 190/// edges across a loop header. For example, threading a jump from outside the 191/// loop (the preheader) to an exit block of the loop is definitely profitable. 192/// It is also almost always profitable to thread backedges from within the loop 193/// to exit blocks, and is often profitable to thread backedges to other blocks 194/// within the loop (forming a nested loop). This simple analysis is not rich 195/// enough to track all of these properties and keep it up-to-date as the CFG 196/// mutates, so we don't allow any of these transformations. 197/// 198void JumpThreading::FindLoopHeaders(Function &F) { 199 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 200 FindFunctionBackedges(F, Edges); 201 202 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 203 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 204} 205 206/// GetResultOfComparison - Given an icmp/fcmp predicate and the left and right 207/// hand sides of the compare instruction, try to determine the result. If the 208/// result can not be determined, a null pointer is returned. 209static Constant *GetResultOfComparison(CmpInst::Predicate pred, 210 Value *LHS, Value *RHS) { 211 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 212 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 213 return ConstantExpr::getCompare(pred, CLHS, CRHS); 214 215 if (LHS == RHS) 216 if (isa<IntegerType>(LHS->getType()) || isa<PointerType>(LHS->getType())) 217 if (ICmpInst::isTrueWhenEqual(pred)) 218 return ConstantInt::getTrue(LHS->getContext()); 219 else 220 return ConstantInt::getFalse(LHS->getContext()); 221 return 0; 222} 223 224 225/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 226/// if we can infer that the value is a known ConstantInt in any of our 227/// predecessors. If so, return the known the list of value and pred BB in the 228/// result vector. If a value is known to be undef, it is returned as null. 229/// 230/// The BB basic block is known to start with a PHI node. 231/// 232/// This returns true if there were any known values. 233/// 234/// 235/// TODO: Per PR2563, we could infer value range information about a predecessor 236/// based on its terminator. 237bool JumpThreading:: 238ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 239 PHINode *TheFirstPHI = cast<PHINode>(BB->begin()); 240 241 // If V is a constantint, then it is known in all predecessors. 242 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 243 ConstantInt *CI = dyn_cast<ConstantInt>(V); 244 Result.resize(TheFirstPHI->getNumIncomingValues()); 245 for (unsigned i = 0, e = Result.size(); i != e; ++i) 246 Result[i] = std::make_pair(CI, TheFirstPHI->getIncomingBlock(i)); 247 return true; 248 } 249 250 // If V is a non-instruction value, or an instruction in a different block, 251 // then it can't be derived from a PHI. 252 Instruction *I = dyn_cast<Instruction>(V); 253 if (I == 0 || I->getParent() != BB) 254 return false; 255 256 /// If I is a PHI node, then we know the incoming values for any constants. 257 if (PHINode *PN = dyn_cast<PHINode>(I)) { 258 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 259 Value *InVal = PN->getIncomingValue(i); 260 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 261 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 262 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 263 } 264 } 265 return !Result.empty(); 266 } 267 268 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 269 270 // Handle some boolean conditions. 271 if (I->getType()->getPrimitiveSizeInBits() == 1) { 272 // X | true -> true 273 // X & false -> false 274 if (I->getOpcode() == Instruction::Or || 275 I->getOpcode() == Instruction::And) { 276 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 277 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 278 279 if (LHSVals.empty() && RHSVals.empty()) 280 return false; 281 282 ConstantInt *InterestingVal; 283 if (I->getOpcode() == Instruction::Or) 284 InterestingVal = ConstantInt::getTrue(I->getContext()); 285 else 286 InterestingVal = ConstantInt::getFalse(I->getContext()); 287 288 // Scan for the sentinel. 289 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 290 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) 291 Result.push_back(LHSVals[i]); 292 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 293 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) 294 Result.push_back(RHSVals[i]); 295 return !Result.empty(); 296 } 297 298 // TODO: Should handle the NOT form of XOR. 299 300 } 301 302 // Handle compare with phi operand, where the PHI is defined in this block. 303 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 304 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 305 if (PN && PN->getParent() == BB) { 306 // We can do this simplification if any comparisons fold to true or false. 307 // See if any do. 308 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 309 BasicBlock *PredBB = PN->getIncomingBlock(i); 310 Value *LHS = PN->getIncomingValue(i); 311 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 312 313 Constant *Res = GetResultOfComparison(Cmp->getPredicate(), LHS, RHS); 314 if (Res == 0) continue; 315 316 if (isa<UndefValue>(Res)) 317 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 318 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 319 Result.push_back(std::make_pair(CI, PredBB)); 320 } 321 322 return !Result.empty(); 323 } 324 325 // TODO: We could also recurse to see if we can determine constants another 326 // way. 327 } 328 return false; 329} 330 331 332 333/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 334/// in an undefined jump, decide which block is best to revector to. 335/// 336/// Since we can pick an arbitrary destination, we pick the successor with the 337/// fewest predecessors. This should reduce the in-degree of the others. 338/// 339static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 340 TerminatorInst *BBTerm = BB->getTerminator(); 341 unsigned MinSucc = 0; 342 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 343 // Compute the successor with the minimum number of predecessors. 344 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 345 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 346 TestBB = BBTerm->getSuccessor(i); 347 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 348 if (NumPreds < MinNumPreds) 349 MinSucc = i; 350 } 351 352 return MinSucc; 353} 354 355/// ProcessBlock - If there are any predecessors whose control can be threaded 356/// through to a successor, transform them now. 357bool JumpThreading::ProcessBlock(BasicBlock *BB) { 358 // If this block has a single predecessor, and if that pred has a single 359 // successor, merge the blocks. This encourages recursive jump threading 360 // because now the condition in this block can be threaded through 361 // predecessors of our predecessor block. 362 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 363 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 364 SinglePred != BB) { 365 // If SinglePred was a loop header, BB becomes one. 366 if (LoopHeaders.erase(SinglePred)) 367 LoopHeaders.insert(BB); 368 369 // Remember if SinglePred was the entry block of the function. If so, we 370 // will need to move BB back to the entry position. 371 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 372 MergeBasicBlockIntoOnlyPred(BB); 373 374 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 375 BB->moveBefore(&BB->getParent()->getEntryBlock()); 376 return true; 377 } 378 } 379 380 // Look to see if the terminator is a branch of switch, if not we can't thread 381 // it. 382 Value *Condition; 383 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 384 // Can't thread an unconditional jump. 385 if (BI->isUnconditional()) return false; 386 Condition = BI->getCondition(); 387 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 388 Condition = SI->getCondition(); 389 else 390 return false; // Must be an invoke. 391 392 // If the terminator of this block is branching on a constant, simplify the 393 // terminator to an unconditional branch. This can occur due to threading in 394 // other blocks. 395 if (isa<ConstantInt>(Condition)) { 396 DEBUG(errs() << " In block '" << BB->getName() 397 << "' folding terminator: " << *BB->getTerminator() << '\n'); 398 ++NumFolds; 399 ConstantFoldTerminator(BB); 400 return true; 401 } 402 403 // If the terminator is branching on an undef, we can pick any of the 404 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 405 if (isa<UndefValue>(Condition)) { 406 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 407 408 // Fold the branch/switch. 409 TerminatorInst *BBTerm = BB->getTerminator(); 410 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 411 if (i == BestSucc) continue; 412 BBTerm->getSuccessor(i)->removePredecessor(BB); 413 } 414 415 DEBUG(errs() << " In block '" << BB->getName() 416 << "' folding undef terminator: " << *BBTerm << '\n'); 417 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 418 BBTerm->eraseFromParent(); 419 return true; 420 } 421 422 Instruction *CondInst = dyn_cast<Instruction>(Condition); 423 424 // If the condition is an instruction defined in another block, see if a 425 // predecessor has the same condition: 426 // br COND, BBX, BBY 427 // BBX: 428 // br COND, BBZ, BBW 429 if (!Condition->hasOneUse() && // Multiple uses. 430 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 431 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 432 if (isa<BranchInst>(BB->getTerminator())) { 433 for (; PI != E; ++PI) 434 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 435 if (PBI->isConditional() && PBI->getCondition() == Condition && 436 ProcessBranchOnDuplicateCond(*PI, BB)) 437 return true; 438 } else { 439 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 440 for (; PI != E; ++PI) 441 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator())) 442 if (PSI->getCondition() == Condition && 443 ProcessSwitchOnDuplicateCond(*PI, BB)) 444 return true; 445 } 446 } 447 448 // All the rest of our checks depend on the condition being an instruction. 449 if (CondInst == 0) 450 return false; 451 452 // See if this is a phi node in the current block. 453 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 454 if (PN->getParent() == BB) 455 return ProcessJumpOnPHI(PN); 456 457 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 458 if (!isa<PHINode>(CondCmp->getOperand(0)) || 459 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB) { 460 // If we have a comparison, loop over the predecessors to see if there is 461 // a condition with a lexically identical value. 462 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 463 for (; PI != E; ++PI) 464 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 465 if (PBI->isConditional() && *PI != BB) { 466 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 467 if (CI->getOperand(0) == CondCmp->getOperand(0) && 468 CI->getOperand(1) == CondCmp->getOperand(1) && 469 CI->getPredicate() == CondCmp->getPredicate()) { 470 // TODO: Could handle things like (x != 4) --> (x == 17) 471 if (ProcessBranchOnDuplicateCond(*PI, BB)) 472 return true; 473 } 474 } 475 } 476 } 477 } 478 479 // Check for some cases that are worth simplifying. Right now we want to look 480 // for loads that are used by a switch or by the condition for the branch. If 481 // we see one, check to see if it's partially redundant. If so, insert a PHI 482 // which can then be used to thread the values. 483 // 484 // This is particularly important because reg2mem inserts loads and stores all 485 // over the place, and this blocks jump threading if we don't zap them. 486 Value *SimplifyValue = CondInst; 487 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 488 if (isa<Constant>(CondCmp->getOperand(1))) 489 SimplifyValue = CondCmp->getOperand(0); 490 491 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 492 if (SimplifyPartiallyRedundantLoad(LI)) 493 return true; 494 495 496 // Handle a variety of cases where we are branching on something derived from 497 // a PHI node in the current block. If we can prove that any predecessors 498 // compute a predictable value based on a PHI node, thread those predecessors. 499 // 500 // We only bother doing this if the current block has a PHI node and if the 501 // conditional instruction lives in the current block. If either condition 502 // fail, this won't be a computable value anyway. 503 if (CondInst->getParent() == BB && isa<PHINode>(BB->front())) 504 if (ProcessThreadableEdges(CondInst, BB)) 505 return true; 506 507 508 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 509 // "(X == 4)" thread through this block. 510 511 return false; 512} 513 514/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 515/// block that jump on exactly the same condition. This means that we almost 516/// always know the direction of the edge in the DESTBB: 517/// PREDBB: 518/// br COND, DESTBB, BBY 519/// DESTBB: 520/// br COND, BBZ, BBW 521/// 522/// If DESTBB has multiple predecessors, we can't just constant fold the branch 523/// in DESTBB, we have to thread over it. 524bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 525 BasicBlock *BB) { 526 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 527 528 // If both successors of PredBB go to DESTBB, we don't know anything. We can 529 // fold the branch to an unconditional one, which allows other recursive 530 // simplifications. 531 bool BranchDir; 532 if (PredBI->getSuccessor(1) != BB) 533 BranchDir = true; 534 else if (PredBI->getSuccessor(0) != BB) 535 BranchDir = false; 536 else { 537 DEBUG(errs() << " In block '" << PredBB->getName() 538 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 539 ++NumFolds; 540 ConstantFoldTerminator(PredBB); 541 return true; 542 } 543 544 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 545 546 // If the dest block has one predecessor, just fix the branch condition to a 547 // constant and fold it. 548 if (BB->getSinglePredecessor()) { 549 DEBUG(errs() << " In block '" << BB->getName() 550 << "' folding condition to '" << BranchDir << "': " 551 << *BB->getTerminator() << '\n'); 552 ++NumFolds; 553 Value *OldCond = DestBI->getCondition(); 554 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 555 BranchDir)); 556 ConstantFoldTerminator(BB); 557 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 558 return true; 559 } 560 561 562 // Next, figure out which successor we are threading to. 563 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 564 565 // Ok, try to thread it! 566 return ThreadEdge(BB, PredBB, SuccBB); 567} 568 569/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 570/// block that switch on exactly the same condition. This means that we almost 571/// always know the direction of the edge in the DESTBB: 572/// PREDBB: 573/// switch COND [... DESTBB, BBY ... ] 574/// DESTBB: 575/// switch COND [... BBZ, BBW ] 576/// 577/// Optimizing switches like this is very important, because simplifycfg builds 578/// switches out of repeated 'if' conditions. 579bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 580 BasicBlock *DestBB) { 581 // Can't thread edge to self. 582 if (PredBB == DestBB) 583 return false; 584 585 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 586 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 587 588 // There are a variety of optimizations that we can potentially do on these 589 // blocks: we order them from most to least preferable. 590 591 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 592 // directly to their destination. This does not introduce *any* code size 593 // growth. Skip debug info first. 594 BasicBlock::iterator BBI = DestBB->begin(); 595 while (isa<DbgInfoIntrinsic>(BBI)) 596 BBI++; 597 598 // FIXME: Thread if it just contains a PHI. 599 if (isa<SwitchInst>(BBI)) { 600 bool MadeChange = false; 601 // Ignore the default edge for now. 602 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 603 ConstantInt *DestVal = DestSI->getCaseValue(i); 604 BasicBlock *DestSucc = DestSI->getSuccessor(i); 605 606 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 607 // PredSI has an explicit case for it. If so, forward. If it is covered 608 // by the default case, we can't update PredSI. 609 unsigned PredCase = PredSI->findCaseValue(DestVal); 610 if (PredCase == 0) continue; 611 612 // If PredSI doesn't go to DestBB on this value, then it won't reach the 613 // case on this condition. 614 if (PredSI->getSuccessor(PredCase) != DestBB && 615 DestSI->getSuccessor(i) != DestBB) 616 continue; 617 618 // Otherwise, we're safe to make the change. Make sure that the edge from 619 // DestSI to DestSucc is not critical and has no PHI nodes. 620 DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 621 DEBUG(errs() << "THROUGH: " << *DestSI); 622 623 // If the destination has PHI nodes, just split the edge for updating 624 // simplicity. 625 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 626 SplitCriticalEdge(DestSI, i, this); 627 DestSucc = DestSI->getSuccessor(i); 628 } 629 FoldSingleEntryPHINodes(DestSucc); 630 PredSI->setSuccessor(PredCase, DestSucc); 631 MadeChange = true; 632 } 633 634 if (MadeChange) 635 return true; 636 } 637 638 return false; 639} 640 641 642/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 643/// load instruction, eliminate it by replacing it with a PHI node. This is an 644/// important optimization that encourages jump threading, and needs to be run 645/// interlaced with other jump threading tasks. 646bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 647 // Don't hack volatile loads. 648 if (LI->isVolatile()) return false; 649 650 // If the load is defined in a block with exactly one predecessor, it can't be 651 // partially redundant. 652 BasicBlock *LoadBB = LI->getParent(); 653 if (LoadBB->getSinglePredecessor()) 654 return false; 655 656 Value *LoadedPtr = LI->getOperand(0); 657 658 // If the loaded operand is defined in the LoadBB, it can't be available. 659 // FIXME: Could do PHI translation, that would be fun :) 660 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 661 if (PtrOp->getParent() == LoadBB) 662 return false; 663 664 // Scan a few instructions up from the load, to see if it is obviously live at 665 // the entry to its block. 666 BasicBlock::iterator BBIt = LI; 667 668 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB, 669 BBIt, 6)) { 670 // If the value if the load is locally available within the block, just use 671 // it. This frequently occurs for reg2mem'd allocas. 672 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 673 674 // If the returned value is the load itself, replace with an undef. This can 675 // only happen in dead loops. 676 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 677 LI->replaceAllUsesWith(AvailableVal); 678 LI->eraseFromParent(); 679 return true; 680 } 681 682 // Otherwise, if we scanned the whole block and got to the top of the block, 683 // we know the block is locally transparent to the load. If not, something 684 // might clobber its value. 685 if (BBIt != LoadBB->begin()) 686 return false; 687 688 689 SmallPtrSet<BasicBlock*, 8> PredsScanned; 690 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 691 AvailablePredsTy AvailablePreds; 692 BasicBlock *OneUnavailablePred = 0; 693 694 // If we got here, the loaded value is transparent through to the start of the 695 // block. Check to see if it is available in any of the predecessor blocks. 696 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 697 PI != PE; ++PI) { 698 BasicBlock *PredBB = *PI; 699 700 // If we already scanned this predecessor, skip it. 701 if (!PredsScanned.insert(PredBB)) 702 continue; 703 704 // Scan the predecessor to see if the value is available in the pred. 705 BBIt = PredBB->end(); 706 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 707 if (!PredAvailable) { 708 OneUnavailablePred = PredBB; 709 continue; 710 } 711 712 // If so, this load is partially redundant. Remember this info so that we 713 // can create a PHI node. 714 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 715 } 716 717 // If the loaded value isn't available in any predecessor, it isn't partially 718 // redundant. 719 if (AvailablePreds.empty()) return false; 720 721 // Okay, the loaded value is available in at least one (and maybe all!) 722 // predecessors. If the value is unavailable in more than one unique 723 // predecessor, we want to insert a merge block for those common predecessors. 724 // This ensures that we only have to insert one reload, thus not increasing 725 // code size. 726 BasicBlock *UnavailablePred = 0; 727 728 // If there is exactly one predecessor where the value is unavailable, the 729 // already computed 'OneUnavailablePred' block is it. If it ends in an 730 // unconditional branch, we know that it isn't a critical edge. 731 if (PredsScanned.size() == AvailablePreds.size()+1 && 732 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 733 UnavailablePred = OneUnavailablePred; 734 } else if (PredsScanned.size() != AvailablePreds.size()) { 735 // Otherwise, we had multiple unavailable predecessors or we had a critical 736 // edge from the one. 737 SmallVector<BasicBlock*, 8> PredsToSplit; 738 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 739 740 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 741 AvailablePredSet.insert(AvailablePreds[i].first); 742 743 // Add all the unavailable predecessors to the PredsToSplit list. 744 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 745 PI != PE; ++PI) 746 if (!AvailablePredSet.count(*PI)) 747 PredsToSplit.push_back(*PI); 748 749 // Split them out to their own block. 750 UnavailablePred = 751 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 752 "thread-split", this); 753 } 754 755 // If the value isn't available in all predecessors, then there will be 756 // exactly one where it isn't available. Insert a load on that edge and add 757 // it to the AvailablePreds list. 758 if (UnavailablePred) { 759 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 760 "Can't handle critical edge here!"); 761 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", 762 UnavailablePred->getTerminator()); 763 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 764 } 765 766 // Now we know that each predecessor of this block has a value in 767 // AvailablePreds, sort them for efficient access as we're walking the preds. 768 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 769 770 // Create a PHI node at the start of the block for the PRE'd load value. 771 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 772 PN->takeName(LI); 773 774 // Insert new entries into the PHI for each predecessor. A single block may 775 // have multiple entries here. 776 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 777 ++PI) { 778 AvailablePredsTy::iterator I = 779 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 780 std::make_pair(*PI, (Value*)0)); 781 782 assert(I != AvailablePreds.end() && I->first == *PI && 783 "Didn't find entry for predecessor!"); 784 785 PN->addIncoming(I->second, I->first); 786 } 787 788 //cerr << "PRE: " << *LI << *PN << "\n"; 789 790 LI->replaceAllUsesWith(PN); 791 LI->eraseFromParent(); 792 793 return true; 794} 795 796/// FindMostPopularDest - The specified list contains multiple possible 797/// threadable destinations. Pick the one that occurs the most frequently in 798/// the list. 799static BasicBlock * 800FindMostPopularDest(BasicBlock *BB, 801 const SmallVectorImpl<std::pair<BasicBlock*, 802 BasicBlock*> > &PredToDestList) { 803 assert(!PredToDestList.empty()); 804 805 // Determine popularity. If there are multiple possible destinations, we 806 // explicitly choose to ignore 'undef' destinations. We prefer to thread 807 // blocks with known and real destinations to threading undef. We'll handle 808 // them later if interesting. 809 DenseMap<BasicBlock*, unsigned> DestPopularity; 810 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 811 if (PredToDestList[i].second) 812 DestPopularity[PredToDestList[i].second]++; 813 814 // Find the most popular dest. 815 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 816 BasicBlock *MostPopularDest = DPI->first; 817 unsigned Popularity = DPI->second; 818 SmallVector<BasicBlock*, 4> SamePopularity; 819 820 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 821 // If the popularity of this entry isn't higher than the popularity we've 822 // seen so far, ignore it. 823 if (DPI->second < Popularity) 824 ; // ignore. 825 else if (DPI->second == Popularity) { 826 // If it is the same as what we've seen so far, keep track of it. 827 SamePopularity.push_back(DPI->first); 828 } else { 829 // If it is more popular, remember it. 830 SamePopularity.clear(); 831 MostPopularDest = DPI->first; 832 Popularity = DPI->second; 833 } 834 } 835 836 // Okay, now we know the most popular destination. If there is more than 837 // destination, we need to determine one. This is arbitrary, but we need 838 // to make a deterministic decision. Pick the first one that appears in the 839 // successor list. 840 if (!SamePopularity.empty()) { 841 SamePopularity.push_back(MostPopularDest); 842 TerminatorInst *TI = BB->getTerminator(); 843 for (unsigned i = 0; ; ++i) { 844 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 845 846 if (std::find(SamePopularity.begin(), SamePopularity.end(), 847 TI->getSuccessor(i)) == SamePopularity.end()) 848 continue; 849 850 MostPopularDest = TI->getSuccessor(i); 851 break; 852 } 853 } 854 855 // Okay, we have finally picked the most popular destination. 856 return MostPopularDest; 857} 858 859bool JumpThreading::ProcessThreadableEdges(Instruction *CondInst, 860 BasicBlock *BB) { 861 // If threading this would thread across a loop header, don't even try to 862 // thread the edge. 863 if (LoopHeaders.count(BB)) 864 return false; 865 866 867 868 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 869 if (!ComputeValueKnownInPredecessors(CondInst, BB, PredValues)) 870 return false; 871 assert(!PredValues.empty() && 872 "ComputeValueKnownInPredecessors returned true with no values"); 873 874 DEBUG(errs() << "IN BB: " << *BB; 875 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 876 errs() << " BB '" << BB->getName() << "': FOUND condition = "; 877 if (PredValues[i].first) 878 errs() << *PredValues[i].first; 879 else 880 errs() << "UNDEF"; 881 errs() << " for pred '" << PredValues[i].second->getName() 882 << "'.\n"; 883 }); 884 885 // Decide what we want to thread through. Convert our list of known values to 886 // a list of known destinations for each pred. This also discards duplicate 887 // predecessors and keeps track of the undefined inputs (which are represented 888 // as a null dest in the PredToDestList. 889 SmallPtrSet<BasicBlock*, 16> SeenPreds; 890 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 891 892 BasicBlock *OnlyDest = 0; 893 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 894 895 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 896 BasicBlock *Pred = PredValues[i].second; 897 if (!SeenPreds.insert(Pred)) 898 continue; // Duplicate predecessor entry. 899 900 // If the predecessor ends with an indirect goto, we can't change its 901 // destination. 902 if (isa<IndirectBrInst>(Pred->getTerminator())) 903 continue; 904 905 ConstantInt *Val = PredValues[i].first; 906 907 BasicBlock *DestBB; 908 if (Val == 0) // Undef. 909 DestBB = 0; 910 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 911 DestBB = BI->getSuccessor(Val->isZero()); 912 else { 913 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 914 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 915 } 916 917 // If we have exactly one destination, remember it for efficiency below. 918 if (i == 0) 919 OnlyDest = DestBB; 920 else if (OnlyDest != DestBB) 921 OnlyDest = MultipleDestSentinel; 922 923 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 924 } 925 926 // If all edges were unthreadable, we fail. 927 if (PredToDestList.empty()) 928 return false; 929 930 // Determine which is the most common successor. If we have many inputs and 931 // this block is a switch, we want to start by threading the batch that goes 932 // to the most popular destination first. If we only know about one 933 // threadable destination (the common case) we can avoid this. 934 BasicBlock *MostPopularDest = OnlyDest; 935 936 if (MostPopularDest == MultipleDestSentinel) 937 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 938 939 // Now that we know what the most popular destination is, factor all 940 // predecessors that will jump to it into a single predecessor. 941 SmallVector<BasicBlock*, 16> PredsToFactor; 942 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 943 if (PredToDestList[i].second == MostPopularDest) 944 PredsToFactor.push_back(PredToDestList[i].first); 945 946 BasicBlock *PredToThread; 947 if (PredsToFactor.size() == 1) 948 PredToThread = PredsToFactor[0]; 949 else { 950 DEBUG(errs() << " Factoring out " << PredsToFactor.size() 951 << " common predecessors.\n"); 952 PredToThread = SplitBlockPredecessors(BB, &PredsToFactor[0], 953 PredsToFactor.size(), 954 ".thr_comm", this); 955 } 956 957 // If the threadable edges are branching on an undefined value, we get to pick 958 // the destination that these predecessors should get to. 959 if (MostPopularDest == 0) 960 MostPopularDest = BB->getTerminator()-> 961 getSuccessor(GetBestDestForJumpOnUndef(BB)); 962 963 // Ok, try to thread it! 964 return ThreadEdge(BB, PredToThread, MostPopularDest); 965} 966 967/// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in 968/// the current block. See if there are any simplifications we can do based on 969/// inputs to the phi node. 970/// 971bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) { 972 BasicBlock *BB = PN->getParent(); 973 974 // If any of the predecessor blocks end in an unconditional branch, we can 975 // *duplicate* the jump into that block in order to further encourage jump 976 // threading and to eliminate cases where we have branch on a phi of an icmp 977 // (branch on icmp is much better). 978 979 // We don't want to do this tranformation for switches, because we don't 980 // really want to duplicate a switch. 981 if (isa<SwitchInst>(BB->getTerminator())) 982 return false; 983 984 // Look for unconditional branch predecessors. 985 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 986 BasicBlock *PredBB = PN->getIncomingBlock(i); 987 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 988 if (PredBr->isUnconditional() && 989 // Try to duplicate BB into PredBB. 990 DuplicateCondBranchOnPHIIntoPred(BB, PredBB)) 991 return true; 992 } 993 994 return false; 995} 996 997 998/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 999/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1000/// NewPred using the entries from OldPred (suitably mapped). 1001static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1002 BasicBlock *OldPred, 1003 BasicBlock *NewPred, 1004 DenseMap<Instruction*, Value*> &ValueMap) { 1005 for (BasicBlock::iterator PNI = PHIBB->begin(); 1006 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1007 // Ok, we have a PHI node. Figure out what the incoming value was for the 1008 // DestBlock. 1009 Value *IV = PN->getIncomingValueForBlock(OldPred); 1010 1011 // Remap the value if necessary. 1012 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1013 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1014 if (I != ValueMap.end()) 1015 IV = I->second; 1016 } 1017 1018 PN->addIncoming(IV, NewPred); 1019 } 1020} 1021 1022/// ThreadEdge - We have decided that it is safe and profitable to thread an 1023/// edge from PredBB to SuccBB across BB. Transform the IR to reflect this 1024/// change. 1025bool JumpThreading::ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, 1026 BasicBlock *SuccBB) { 1027 // If threading to the same block as we come from, we would infinite loop. 1028 if (SuccBB == BB) { 1029 DEBUG(errs() << " Not threading across BB '" << BB->getName() 1030 << "' - would thread to self!\n"); 1031 return false; 1032 } 1033 1034 // If threading this would thread across a loop header, don't thread the edge. 1035 // See the comments above FindLoopHeaders for justifications and caveats. 1036 if (LoopHeaders.count(BB)) { 1037 DEBUG(errs() << " Not threading from '" << PredBB->getName() 1038 << "' across loop header BB '" << BB->getName() 1039 << "' to dest BB '" << SuccBB->getName() 1040 << "' - it might create an irreducible loop!\n"); 1041 return false; 1042 } 1043 1044 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1045 if (JumpThreadCost > Threshold) { 1046 DEBUG(errs() << " Not threading BB '" << BB->getName() 1047 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1048 return false; 1049 } 1050 1051 // And finally, do it! 1052 DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '" 1053 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1054 << ", across block:\n " 1055 << *BB << "\n"); 1056 1057 // We are going to have to map operands from the original BB block to the new 1058 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1059 // account for entry from PredBB. 1060 DenseMap<Instruction*, Value*> ValueMapping; 1061 1062 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1063 BB->getName()+".thread", 1064 BB->getParent(), BB); 1065 NewBB->moveAfter(PredBB); 1066 1067 BasicBlock::iterator BI = BB->begin(); 1068 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1069 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1070 1071 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1072 // mapping and using it to remap operands in the cloned instructions. 1073 for (; !isa<TerminatorInst>(BI); ++BI) { 1074 Instruction *New = BI->clone(); 1075 New->setName(BI->getName()); 1076 NewBB->getInstList().push_back(New); 1077 ValueMapping[BI] = New; 1078 1079 // Remap operands to patch up intra-block references. 1080 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1081 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1082 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1083 if (I != ValueMapping.end()) 1084 New->setOperand(i, I->second); 1085 } 1086 } 1087 1088 // We didn't copy the terminator from BB over to NewBB, because there is now 1089 // an unconditional jump to SuccBB. Insert the unconditional jump. 1090 BranchInst::Create(SuccBB, NewBB); 1091 1092 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1093 // PHI nodes for NewBB now. 1094 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1095 1096 // If there were values defined in BB that are used outside the block, then we 1097 // now have to update all uses of the value to use either the original value, 1098 // the cloned value, or some PHI derived value. This can require arbitrary 1099 // PHI insertion, of which we are prepared to do, clean these up now. 1100 SSAUpdater SSAUpdate; 1101 SmallVector<Use*, 16> UsesToRename; 1102 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1103 // Scan all uses of this instruction to see if it is used outside of its 1104 // block, and if so, record them in UsesToRename. 1105 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1106 ++UI) { 1107 Instruction *User = cast<Instruction>(*UI); 1108 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1109 if (UserPN->getIncomingBlock(UI) == BB) 1110 continue; 1111 } else if (User->getParent() == BB) 1112 continue; 1113 1114 UsesToRename.push_back(&UI.getUse()); 1115 } 1116 1117 // If there are no uses outside the block, we're done with this instruction. 1118 if (UsesToRename.empty()) 1119 continue; 1120 1121 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1122 1123 // We found a use of I outside of BB. Rename all uses of I that are outside 1124 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1125 // with the two values we know. 1126 SSAUpdate.Initialize(I); 1127 SSAUpdate.AddAvailableValue(BB, I); 1128 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1129 1130 while (!UsesToRename.empty()) 1131 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1132 DEBUG(errs() << "\n"); 1133 } 1134 1135 1136 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1137 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1138 // us to simplify any PHI nodes in BB. 1139 TerminatorInst *PredTerm = PredBB->getTerminator(); 1140 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1141 if (PredTerm->getSuccessor(i) == BB) { 1142 BB->removePredecessor(PredBB); 1143 PredTerm->setSuccessor(i, NewBB); 1144 } 1145 1146 // At this point, the IR is fully up to date and consistent. Do a quick scan 1147 // over the new instructions and zap any that are constants or dead. This 1148 // frequently happens because of phi translation. 1149 BI = NewBB->begin(); 1150 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) { 1151 Instruction *Inst = BI++; 1152 if (Constant *C = ConstantFoldInstruction(Inst, TD)) { 1153 Inst->replaceAllUsesWith(C); 1154 Inst->eraseFromParent(); 1155 continue; 1156 } 1157 1158 RecursivelyDeleteTriviallyDeadInstructions(Inst); 1159 } 1160 1161 // Threaded an edge! 1162 ++NumThreads; 1163 return true; 1164} 1165 1166/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1167/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1168/// If we can duplicate the contents of BB up into PredBB do so now, this 1169/// improves the odds that the branch will be on an analyzable instruction like 1170/// a compare. 1171bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1172 BasicBlock *PredBB) { 1173 // If BB is a loop header, then duplicating this block outside the loop would 1174 // cause us to transform this into an irreducible loop, don't do this. 1175 // See the comments above FindLoopHeaders for justifications and caveats. 1176 if (LoopHeaders.count(BB)) { 1177 DEBUG(errs() << " Not duplicating loop header '" << BB->getName() 1178 << "' into predecessor block '" << PredBB->getName() 1179 << "' - it might create an irreducible loop!\n"); 1180 return false; 1181 } 1182 1183 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1184 if (DuplicationCost > Threshold) { 1185 DEBUG(errs() << " Not duplicating BB '" << BB->getName() 1186 << "' - Cost is too high: " << DuplicationCost << "\n"); 1187 return false; 1188 } 1189 1190 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1191 // of PredBB. 1192 DEBUG(errs() << " Duplicating block '" << BB->getName() << "' into end of '" 1193 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1194 << DuplicationCost << " block is:" << *BB << "\n"); 1195 1196 // We are going to have to map operands from the original BB block into the 1197 // PredBB block. Evaluate PHI nodes in BB. 1198 DenseMap<Instruction*, Value*> ValueMapping; 1199 1200 BasicBlock::iterator BI = BB->begin(); 1201 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1202 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1203 1204 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1205 1206 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1207 // mapping and using it to remap operands in the cloned instructions. 1208 for (; BI != BB->end(); ++BI) { 1209 Instruction *New = BI->clone(); 1210 New->setName(BI->getName()); 1211 PredBB->getInstList().insert(OldPredBranch, New); 1212 ValueMapping[BI] = New; 1213 1214 // Remap operands to patch up intra-block references. 1215 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1216 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1217 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1218 if (I != ValueMapping.end()) 1219 New->setOperand(i, I->second); 1220 } 1221 } 1222 1223 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1224 // add entries to the PHI nodes for branch from PredBB now. 1225 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1226 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1227 ValueMapping); 1228 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1229 ValueMapping); 1230 1231 // If there were values defined in BB that are used outside the block, then we 1232 // now have to update all uses of the value to use either the original value, 1233 // the cloned value, or some PHI derived value. This can require arbitrary 1234 // PHI insertion, of which we are prepared to do, clean these up now. 1235 SSAUpdater SSAUpdate; 1236 SmallVector<Use*, 16> UsesToRename; 1237 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1238 // Scan all uses of this instruction to see if it is used outside of its 1239 // block, and if so, record them in UsesToRename. 1240 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1241 ++UI) { 1242 Instruction *User = cast<Instruction>(*UI); 1243 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1244 if (UserPN->getIncomingBlock(UI) == BB) 1245 continue; 1246 } else if (User->getParent() == BB) 1247 continue; 1248 1249 UsesToRename.push_back(&UI.getUse()); 1250 } 1251 1252 // If there are no uses outside the block, we're done with this instruction. 1253 if (UsesToRename.empty()) 1254 continue; 1255 1256 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1257 1258 // We found a use of I outside of BB. Rename all uses of I that are outside 1259 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1260 // with the two values we know. 1261 SSAUpdate.Initialize(I); 1262 SSAUpdate.AddAvailableValue(BB, I); 1263 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1264 1265 while (!UsesToRename.empty()) 1266 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1267 DEBUG(errs() << "\n"); 1268 } 1269 1270 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1271 // that we nuked. 1272 BB->removePredecessor(PredBB); 1273 1274 // Remove the unconditional branch at the end of the PredBB block. 1275 OldPredBranch->eraseFromParent(); 1276 1277 ++NumDupes; 1278 return true; 1279} 1280 1281 1282