JumpThreading.cpp revision 82114b98c9da043d97a39a07255518dbc00ba522
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 BasicBlock *FactorCommonPHIPreds(PHINode *PN, Value *Val); 79 80 typedef SmallVectorImpl<std::pair<ConstantInt*, 81 BasicBlock*> > PredValueInfo; 82 83 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 84 PredValueInfo &Result); 85 bool ProcessThreadableEdges(Instruction *CondInst, BasicBlock *BB); 86 87 88 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 89 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 90 91 bool ProcessJumpOnPHI(PHINode *PN); 92 93 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 94 }; 95} 96 97char JumpThreading::ID = 0; 98static RegisterPass<JumpThreading> 99X("jump-threading", "Jump Threading"); 100 101// Public interface to the Jump Threading pass 102FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 103 104/// runOnFunction - Top level algorithm. 105/// 106bool JumpThreading::runOnFunction(Function &F) { 107 DEBUG(errs() << "Jump threading on function '" << F.getName() << "'\n"); 108 TD = getAnalysisIfAvailable<TargetData>(); 109 110 FindLoopHeaders(F); 111 112 bool AnotherIteration = true, EverChanged = false; 113 while (AnotherIteration) { 114 AnotherIteration = false; 115 bool Changed = false; 116 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 117 BasicBlock *BB = I; 118 while (ProcessBlock(BB)) 119 Changed = true; 120 121 ++I; 122 123 // If the block is trivially dead, zap it. This eliminates the successor 124 // edges which simplifies the CFG. 125 if (pred_begin(BB) == pred_end(BB) && 126 BB != &BB->getParent()->getEntryBlock()) { 127 DEBUG(errs() << " JT: Deleting dead block '" << BB->getName() 128 << "' with terminator: " << *BB->getTerminator() << '\n'); 129 LoopHeaders.erase(BB); 130 DeleteDeadBlock(BB); 131 Changed = true; 132 } 133 } 134 AnotherIteration = Changed; 135 EverChanged |= Changed; 136 } 137 138 LoopHeaders.clear(); 139 return EverChanged; 140} 141 142/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 143/// thread across it. 144static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 145 /// Ignore PHI nodes, these will be flattened when duplication happens. 146 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 147 148 // Sum up the cost of each instruction until we get to the terminator. Don't 149 // include the terminator because the copy won't include it. 150 unsigned Size = 0; 151 for (; !isa<TerminatorInst>(I); ++I) { 152 // Debugger intrinsics don't incur code size. 153 if (isa<DbgInfoIntrinsic>(I)) continue; 154 155 // If this is a pointer->pointer bitcast, it is free. 156 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType())) 157 continue; 158 159 // All other instructions count for at least one unit. 160 ++Size; 161 162 // Calls are more expensive. If they are non-intrinsic calls, we model them 163 // as having cost of 4. If they are a non-vector intrinsic, we model them 164 // as having cost of 2 total, and if they are a vector intrinsic, we model 165 // them as having cost 1. 166 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 167 if (!isa<IntrinsicInst>(CI)) 168 Size += 3; 169 else if (!isa<VectorType>(CI->getType())) 170 Size += 1; 171 } 172 } 173 174 // Threading through a switch statement is particularly profitable. If this 175 // block ends in a switch, decrease its cost to make it more likely to happen. 176 if (isa<SwitchInst>(I)) 177 Size = Size > 6 ? Size-6 : 0; 178 179 return Size; 180} 181 182 183 184/// FindLoopHeaders - We do not want jump threading to turn proper loop 185/// structures into irreducible loops. Doing this breaks up the loop nesting 186/// hierarchy and pessimizes later transformations. To prevent this from 187/// happening, we first have to find the loop headers. Here we approximate this 188/// by finding targets of backedges in the CFG. 189/// 190/// Note that there definitely are cases when we want to allow threading of 191/// edges across a loop header. For example, threading a jump from outside the 192/// loop (the preheader) to an exit block of the loop is definitely profitable. 193/// It is also almost always profitable to thread backedges from within the loop 194/// to exit blocks, and is often profitable to thread backedges to other blocks 195/// within the loop (forming a nested loop). This simple analysis is not rich 196/// enough to track all of these properties and keep it up-to-date as the CFG 197/// mutates, so we don't allow any of these transformations. 198/// 199void JumpThreading::FindLoopHeaders(Function &F) { 200 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 201 FindFunctionBackedges(F, Edges); 202 203 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 204 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 205} 206 207 208/// FactorCommonPHIPreds - If there are multiple preds with the same incoming 209/// value for the PHI, factor them together so we get one block to thread for 210/// the whole group. 211/// This is important for things like "phi i1 [true, true, false, true, x]" 212/// where we only need to clone the block for the true blocks once. 213/// 214BasicBlock *JumpThreading::FactorCommonPHIPreds(PHINode *PN, Value *Val) { 215 SmallVector<BasicBlock*, 16> CommonPreds; 216 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 217 if (PN->getIncomingValue(i) == Val) 218 CommonPreds.push_back(PN->getIncomingBlock(i)); 219 220 if (CommonPreds.size() == 1) 221 return CommonPreds[0]; 222 223 DEBUG(errs() << " Factoring out " << CommonPreds.size() 224 << " common predecessors.\n"); 225 return SplitBlockPredecessors(PN->getParent(), 226 &CommonPreds[0], CommonPreds.size(), 227 ".thr_comm", this); 228} 229 230/// GetResultOfComparison - Given an icmp/fcmp predicate and the left and right 231/// hand sides of the compare instruction, try to determine the result. If the 232/// result can not be determined, a null pointer is returned. 233static Constant *GetResultOfComparison(CmpInst::Predicate pred, 234 Value *LHS, Value *RHS) { 235 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 236 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 237 return ConstantExpr::getCompare(pred, CLHS, CRHS); 238 239 if (LHS == RHS) 240 if (isa<IntegerType>(LHS->getType()) || isa<PointerType>(LHS->getType())) 241 if (ICmpInst::isTrueWhenEqual(pred)) 242 return ConstantInt::getTrue(LHS->getContext()); 243 else 244 return ConstantInt::getFalse(LHS->getContext()); 245 return 0; 246} 247 248 249/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 250/// if we can infer that the value is a known ConstantInt in any of our 251/// predecessors. If so, return the known the list of value and pred BB in the 252/// result vector. If a value is known to be undef, it is returned as null. 253/// 254/// The BB basic block is known to start with a PHI node. 255/// 256/// This returns true if there were any known values. 257/// 258/// 259/// TODO: Per PR2563, we could infer value range information about a predecessor 260/// based on its terminator. 261bool JumpThreading:: 262ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 263 PHINode *TheFirstPHI = cast<PHINode>(BB->begin()); 264 265 // If V is a constantint, then it is known in all predecessors. 266 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 267 ConstantInt *CI = dyn_cast<ConstantInt>(V); 268 Result.resize(TheFirstPHI->getNumIncomingValues()); 269 for (unsigned i = 0, e = Result.size(); i != e; ++i) 270 Result[i] = std::make_pair(CI, TheFirstPHI->getIncomingBlock(i)); 271 return true; 272 } 273 274 // If V is a non-instruction value, or an instruction in a different block, 275 // then it can't be derived from a PHI. 276 Instruction *I = dyn_cast<Instruction>(V); 277 if (I == 0 || I->getParent() != BB) 278 return false; 279 280 /// If I is a PHI node, then we know the incoming values for any constants. 281 if (PHINode *PN = dyn_cast<PHINode>(I)) { 282 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 283 Value *InVal = PN->getIncomingValue(i); 284 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 285 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 286 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 287 } 288 } 289 return !Result.empty(); 290 } 291 292 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 293 294 // Handle some boolean conditions. 295 if (I->getType()->getPrimitiveSizeInBits() == 1) { 296 // X | true -> true 297 // X & false -> false 298 if (I->getOpcode() == Instruction::Or || 299 I->getOpcode() == Instruction::And) { 300 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 301 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 302 303 if (LHSVals.empty() && RHSVals.empty()) 304 return false; 305 306 ConstantInt *InterestingVal; 307 if (I->getOpcode() == Instruction::Or) 308 InterestingVal = ConstantInt::getTrue(I->getContext()); 309 else 310 InterestingVal = ConstantInt::getFalse(I->getContext()); 311 312 // Scan for the sentinel. 313 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 314 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) 315 Result.push_back(LHSVals[i]); 316 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 317 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) 318 Result.push_back(RHSVals[i]); 319 return !Result.empty(); 320 } 321 322 // TODO: Should handle the NOT form of XOR. 323 324 } 325 326 // Handle compare with phi operand, where the PHI is defined in this block. 327 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 328 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 329 if (PN && PN->getParent() == BB) { 330 // We can do this simplification if any comparisons fold to true or false. 331 // See if any do. 332 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 333 BasicBlock *PredBB = PN->getIncomingBlock(i); 334 Value *LHS = PN->getIncomingValue(i); 335 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 336 337 Constant *Res = GetResultOfComparison(Cmp->getPredicate(), LHS, RHS); 338 if (Res == 0) continue; 339 340 if (isa<UndefValue>(Res)) 341 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 342 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 343 Result.push_back(std::make_pair(CI, PredBB)); 344 } 345 346 return !Result.empty(); 347 } 348 349 // TODO: We could also recurse to see if we can determine constants another 350 // way. 351 } 352 return false; 353} 354 355 356 357/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 358/// in an undefined jump, decide which block is best to revector to. 359/// 360/// Since we can pick an arbitrary destination, we pick the successor with the 361/// fewest predecessors. This should reduce the in-degree of the others. 362/// 363static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 364 TerminatorInst *BBTerm = BB->getTerminator(); 365 unsigned MinSucc = 0; 366 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 367 // Compute the successor with the minimum number of predecessors. 368 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 369 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 370 TestBB = BBTerm->getSuccessor(i); 371 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 372 if (NumPreds < MinNumPreds) 373 MinSucc = i; 374 } 375 376 return MinSucc; 377} 378 379/// ProcessBlock - If there are any predecessors whose control can be threaded 380/// through to a successor, transform them now. 381bool JumpThreading::ProcessBlock(BasicBlock *BB) { 382 // If this block has a single predecessor, and if that pred has a single 383 // successor, merge the blocks. This encourages recursive jump threading 384 // because now the condition in this block can be threaded through 385 // predecessors of our predecessor block. 386 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 387 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 388 SinglePred != BB) { 389 // If SinglePred was a loop header, BB becomes one. 390 if (LoopHeaders.erase(SinglePred)) 391 LoopHeaders.insert(BB); 392 393 // Remember if SinglePred was the entry block of the function. If so, we 394 // will need to move BB back to the entry position. 395 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 396 MergeBasicBlockIntoOnlyPred(BB); 397 398 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 399 BB->moveBefore(&BB->getParent()->getEntryBlock()); 400 return true; 401 } 402 } 403 404 // Look to see if the terminator is a branch of switch, if not we can't thread 405 // it. 406 Value *Condition; 407 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 408 // Can't thread an unconditional jump. 409 if (BI->isUnconditional()) return false; 410 Condition = BI->getCondition(); 411 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 412 Condition = SI->getCondition(); 413 else 414 return false; // Must be an invoke. 415 416 // If the terminator of this block is branching on a constant, simplify the 417 // terminator to an unconditional branch. This can occur due to threading in 418 // other blocks. 419 if (isa<ConstantInt>(Condition)) { 420 DEBUG(errs() << " In block '" << BB->getName() 421 << "' folding terminator: " << *BB->getTerminator() << '\n'); 422 ++NumFolds; 423 ConstantFoldTerminator(BB); 424 return true; 425 } 426 427 // If the terminator is branching on an undef, we can pick any of the 428 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 429 if (isa<UndefValue>(Condition)) { 430 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 431 432 // Fold the branch/switch. 433 TerminatorInst *BBTerm = BB->getTerminator(); 434 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 435 if (i == BestSucc) continue; 436 BBTerm->getSuccessor(i)->removePredecessor(BB); 437 } 438 439 DEBUG(errs() << " In block '" << BB->getName() 440 << "' folding undef terminator: " << *BBTerm << '\n'); 441 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 442 BBTerm->eraseFromParent(); 443 return true; 444 } 445 446 Instruction *CondInst = dyn_cast<Instruction>(Condition); 447 448 // If the condition is an instruction defined in another block, see if a 449 // predecessor has the same condition: 450 // br COND, BBX, BBY 451 // BBX: 452 // br COND, BBZ, BBW 453 if (!Condition->hasOneUse() && // Multiple uses. 454 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 455 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 456 if (isa<BranchInst>(BB->getTerminator())) { 457 for (; PI != E; ++PI) 458 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 459 if (PBI->isConditional() && PBI->getCondition() == Condition && 460 ProcessBranchOnDuplicateCond(*PI, BB)) 461 return true; 462 } else { 463 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 464 for (; PI != E; ++PI) 465 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator())) 466 if (PSI->getCondition() == Condition && 467 ProcessSwitchOnDuplicateCond(*PI, BB)) 468 return true; 469 } 470 } 471 472 // All the rest of our checks depend on the condition being an instruction. 473 if (CondInst == 0) 474 return false; 475 476 // See if this is a phi node in the current block. 477 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 478 if (PN->getParent() == BB) 479 return ProcessJumpOnPHI(PN); 480 481 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 482 if (!isa<PHINode>(CondCmp->getOperand(0)) || 483 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB) { 484 // If we have a comparison, loop over the predecessors to see if there is 485 // a condition with a lexically identical value. 486 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 487 for (; PI != E; ++PI) 488 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 489 if (PBI->isConditional() && *PI != BB) { 490 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 491 if (CI->getOperand(0) == CondCmp->getOperand(0) && 492 CI->getOperand(1) == CondCmp->getOperand(1) && 493 CI->getPredicate() == CondCmp->getPredicate()) { 494 // TODO: Could handle things like (x != 4) --> (x == 17) 495 if (ProcessBranchOnDuplicateCond(*PI, BB)) 496 return true; 497 } 498 } 499 } 500 } 501 } 502 503 // Check for some cases that are worth simplifying. Right now we want to look 504 // for loads that are used by a switch or by the condition for the branch. If 505 // we see one, check to see if it's partially redundant. If so, insert a PHI 506 // which can then be used to thread the values. 507 // 508 // This is particularly important because reg2mem inserts loads and stores all 509 // over the place, and this blocks jump threading if we don't zap them. 510 Value *SimplifyValue = CondInst; 511 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 512 if (isa<Constant>(CondCmp->getOperand(1))) 513 SimplifyValue = CondCmp->getOperand(0); 514 515 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 516 if (SimplifyPartiallyRedundantLoad(LI)) 517 return true; 518 519 520 // Handle a variety of cases where we are branching on something derived from 521 // a PHI node in the current block. If we can prove that any predecessors 522 // compute a predictable value based on a PHI node, thread those predecessors. 523 // 524 // We only bother doing this if the current block has a PHI node and if the 525 // conditional instruction lives in the current block. If either condition 526 // fail, this won't be a computable value anyway. 527 if (CondInst->getParent() == BB && isa<PHINode>(BB->front())) 528 if (ProcessThreadableEdges(CondInst, BB)) 529 return true; 530 531 532 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 533 // "(X == 4)" thread through this block. 534 535 return false; 536} 537 538/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 539/// block that jump on exactly the same condition. This means that we almost 540/// always know the direction of the edge in the DESTBB: 541/// PREDBB: 542/// br COND, DESTBB, BBY 543/// DESTBB: 544/// br COND, BBZ, BBW 545/// 546/// If DESTBB has multiple predecessors, we can't just constant fold the branch 547/// in DESTBB, we have to thread over it. 548bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 549 BasicBlock *BB) { 550 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 551 552 // If both successors of PredBB go to DESTBB, we don't know anything. We can 553 // fold the branch to an unconditional one, which allows other recursive 554 // simplifications. 555 bool BranchDir; 556 if (PredBI->getSuccessor(1) != BB) 557 BranchDir = true; 558 else if (PredBI->getSuccessor(0) != BB) 559 BranchDir = false; 560 else { 561 DEBUG(errs() << " In block '" << PredBB->getName() 562 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 563 ++NumFolds; 564 ConstantFoldTerminator(PredBB); 565 return true; 566 } 567 568 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 569 570 // If the dest block has one predecessor, just fix the branch condition to a 571 // constant and fold it. 572 if (BB->getSinglePredecessor()) { 573 DEBUG(errs() << " In block '" << BB->getName() 574 << "' folding condition to '" << BranchDir << "': " 575 << *BB->getTerminator() << '\n'); 576 ++NumFolds; 577 Value *OldCond = DestBI->getCondition(); 578 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 579 BranchDir)); 580 ConstantFoldTerminator(BB); 581 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 582 return true; 583 } 584 585 586 // Next, figure out which successor we are threading to. 587 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 588 589 // Ok, try to thread it! 590 return ThreadEdge(BB, PredBB, SuccBB); 591} 592 593/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 594/// block that switch on exactly the same condition. This means that we almost 595/// always know the direction of the edge in the DESTBB: 596/// PREDBB: 597/// switch COND [... DESTBB, BBY ... ] 598/// DESTBB: 599/// switch COND [... BBZ, BBW ] 600/// 601/// Optimizing switches like this is very important, because simplifycfg builds 602/// switches out of repeated 'if' conditions. 603bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 604 BasicBlock *DestBB) { 605 // Can't thread edge to self. 606 if (PredBB == DestBB) 607 return false; 608 609 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 610 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 611 612 // There are a variety of optimizations that we can potentially do on these 613 // blocks: we order them from most to least preferable. 614 615 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 616 // directly to their destination. This does not introduce *any* code size 617 // growth. Skip debug info first. 618 BasicBlock::iterator BBI = DestBB->begin(); 619 while (isa<DbgInfoIntrinsic>(BBI)) 620 BBI++; 621 622 // FIXME: Thread if it just contains a PHI. 623 if (isa<SwitchInst>(BBI)) { 624 bool MadeChange = false; 625 // Ignore the default edge for now. 626 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 627 ConstantInt *DestVal = DestSI->getCaseValue(i); 628 BasicBlock *DestSucc = DestSI->getSuccessor(i); 629 630 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 631 // PredSI has an explicit case for it. If so, forward. If it is covered 632 // by the default case, we can't update PredSI. 633 unsigned PredCase = PredSI->findCaseValue(DestVal); 634 if (PredCase == 0) continue; 635 636 // If PredSI doesn't go to DestBB on this value, then it won't reach the 637 // case on this condition. 638 if (PredSI->getSuccessor(PredCase) != DestBB && 639 DestSI->getSuccessor(i) != DestBB) 640 continue; 641 642 // Otherwise, we're safe to make the change. Make sure that the edge from 643 // DestSI to DestSucc is not critical and has no PHI nodes. 644 DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 645 DEBUG(errs() << "THROUGH: " << *DestSI); 646 647 // If the destination has PHI nodes, just split the edge for updating 648 // simplicity. 649 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 650 SplitCriticalEdge(DestSI, i, this); 651 DestSucc = DestSI->getSuccessor(i); 652 } 653 FoldSingleEntryPHINodes(DestSucc); 654 PredSI->setSuccessor(PredCase, DestSucc); 655 MadeChange = true; 656 } 657 658 if (MadeChange) 659 return true; 660 } 661 662 return false; 663} 664 665 666/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 667/// load instruction, eliminate it by replacing it with a PHI node. This is an 668/// important optimization that encourages jump threading, and needs to be run 669/// interlaced with other jump threading tasks. 670bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 671 // Don't hack volatile loads. 672 if (LI->isVolatile()) return false; 673 674 // If the load is defined in a block with exactly one predecessor, it can't be 675 // partially redundant. 676 BasicBlock *LoadBB = LI->getParent(); 677 if (LoadBB->getSinglePredecessor()) 678 return false; 679 680 Value *LoadedPtr = LI->getOperand(0); 681 682 // If the loaded operand is defined in the LoadBB, it can't be available. 683 // FIXME: Could do PHI translation, that would be fun :) 684 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 685 if (PtrOp->getParent() == LoadBB) 686 return false; 687 688 // Scan a few instructions up from the load, to see if it is obviously live at 689 // the entry to its block. 690 BasicBlock::iterator BBIt = LI; 691 692 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB, 693 BBIt, 6)) { 694 // If the value if the load is locally available within the block, just use 695 // it. This frequently occurs for reg2mem'd allocas. 696 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 697 698 // If the returned value is the load itself, replace with an undef. This can 699 // only happen in dead loops. 700 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 701 LI->replaceAllUsesWith(AvailableVal); 702 LI->eraseFromParent(); 703 return true; 704 } 705 706 // Otherwise, if we scanned the whole block and got to the top of the block, 707 // we know the block is locally transparent to the load. If not, something 708 // might clobber its value. 709 if (BBIt != LoadBB->begin()) 710 return false; 711 712 713 SmallPtrSet<BasicBlock*, 8> PredsScanned; 714 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 715 AvailablePredsTy AvailablePreds; 716 BasicBlock *OneUnavailablePred = 0; 717 718 // If we got here, the loaded value is transparent through to the start of the 719 // block. Check to see if it is available in any of the predecessor blocks. 720 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 721 PI != PE; ++PI) { 722 BasicBlock *PredBB = *PI; 723 724 // If we already scanned this predecessor, skip it. 725 if (!PredsScanned.insert(PredBB)) 726 continue; 727 728 // Scan the predecessor to see if the value is available in the pred. 729 BBIt = PredBB->end(); 730 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 731 if (!PredAvailable) { 732 OneUnavailablePred = PredBB; 733 continue; 734 } 735 736 // If so, this load is partially redundant. Remember this info so that we 737 // can create a PHI node. 738 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 739 } 740 741 // If the loaded value isn't available in any predecessor, it isn't partially 742 // redundant. 743 if (AvailablePreds.empty()) return false; 744 745 // Okay, the loaded value is available in at least one (and maybe all!) 746 // predecessors. If the value is unavailable in more than one unique 747 // predecessor, we want to insert a merge block for those common predecessors. 748 // This ensures that we only have to insert one reload, thus not increasing 749 // code size. 750 BasicBlock *UnavailablePred = 0; 751 752 // If there is exactly one predecessor where the value is unavailable, the 753 // already computed 'OneUnavailablePred' block is it. If it ends in an 754 // unconditional branch, we know that it isn't a critical edge. 755 if (PredsScanned.size() == AvailablePreds.size()+1 && 756 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 757 UnavailablePred = OneUnavailablePred; 758 } else if (PredsScanned.size() != AvailablePreds.size()) { 759 // Otherwise, we had multiple unavailable predecessors or we had a critical 760 // edge from the one. 761 SmallVector<BasicBlock*, 8> PredsToSplit; 762 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 763 764 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 765 AvailablePredSet.insert(AvailablePreds[i].first); 766 767 // Add all the unavailable predecessors to the PredsToSplit list. 768 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 769 PI != PE; ++PI) 770 if (!AvailablePredSet.count(*PI)) 771 PredsToSplit.push_back(*PI); 772 773 // Split them out to their own block. 774 UnavailablePred = 775 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 776 "thread-split", this); 777 } 778 779 // If the value isn't available in all predecessors, then there will be 780 // exactly one where it isn't available. Insert a load on that edge and add 781 // it to the AvailablePreds list. 782 if (UnavailablePred) { 783 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 784 "Can't handle critical edge here!"); 785 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", 786 UnavailablePred->getTerminator()); 787 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 788 } 789 790 // Now we know that each predecessor of this block has a value in 791 // AvailablePreds, sort them for efficient access as we're walking the preds. 792 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 793 794 // Create a PHI node at the start of the block for the PRE'd load value. 795 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 796 PN->takeName(LI); 797 798 // Insert new entries into the PHI for each predecessor. A single block may 799 // have multiple entries here. 800 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 801 ++PI) { 802 AvailablePredsTy::iterator I = 803 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 804 std::make_pair(*PI, (Value*)0)); 805 806 assert(I != AvailablePreds.end() && I->first == *PI && 807 "Didn't find entry for predecessor!"); 808 809 PN->addIncoming(I->second, I->first); 810 } 811 812 //cerr << "PRE: " << *LI << *PN << "\n"; 813 814 LI->replaceAllUsesWith(PN); 815 LI->eraseFromParent(); 816 817 return true; 818} 819 820/// FindMostPopularDest - The specified list contains multiple possible 821/// threadable destinations. Pick the one that occurs the most frequently in 822/// the list. 823static BasicBlock * 824FindMostPopularDest(BasicBlock *BB, 825 const SmallVectorImpl<std::pair<BasicBlock*, 826 BasicBlock*> > &PredToDestList) { 827 assert(!PredToDestList.empty()); 828 829 // Determine popularity. If there are multiple possible destinations, we 830 // explicitly choose to ignore 'undef' destinations. We prefer to thread 831 // blocks with known and real destinations to threading undef. We'll handle 832 // them later if interesting. 833 DenseMap<BasicBlock*, unsigned> DestPopularity; 834 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 835 if (PredToDestList[i].second) 836 DestPopularity[PredToDestList[i].second]++; 837 838 // Find the most popular dest. 839 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 840 BasicBlock *MostPopularDest = DPI->first; 841 unsigned Popularity = DPI->second; 842 SmallVector<BasicBlock*, 4> SamePopularity; 843 844 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 845 // If the popularity of this entry isn't higher than the popularity we've 846 // seen so far, ignore it. 847 if (DPI->second < Popularity) 848 ; // ignore. 849 else if (DPI->second == Popularity) { 850 // If it is the same as what we've seen so far, keep track of it. 851 SamePopularity.push_back(DPI->first); 852 } else { 853 // If it is more popular, remember it. 854 SamePopularity.clear(); 855 MostPopularDest = DPI->first; 856 Popularity = DPI->second; 857 } 858 } 859 860 // Okay, now we know the most popular destination. If there is more than 861 // destination, we need to determine one. This is arbitrary, but we need 862 // to make a deterministic decision. Pick the first one that appears in the 863 // successor list. 864 if (!SamePopularity.empty()) { 865 SamePopularity.push_back(MostPopularDest); 866 TerminatorInst *TI = BB->getTerminator(); 867 for (unsigned i = 0; ; ++i) { 868 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 869 870 if (std::find(SamePopularity.begin(), SamePopularity.end(), 871 TI->getSuccessor(i)) == SamePopularity.end()) 872 continue; 873 874 MostPopularDest = TI->getSuccessor(i); 875 break; 876 } 877 } 878 879 // Okay, we have finally picked the most popular destination. 880 return MostPopularDest; 881} 882 883bool JumpThreading::ProcessThreadableEdges(Instruction *CondInst, 884 BasicBlock *BB) { 885 // If threading this would thread across a loop header, don't even try to 886 // thread the edge. 887 if (LoopHeaders.count(BB)) 888 return false; 889 890 891 892 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 893 if (!ComputeValueKnownInPredecessors(CondInst, BB, PredValues)) 894 return false; 895 assert(!PredValues.empty() && 896 "ComputeValueKnownInPredecessors returned true with no values"); 897 898 DEBUG(errs() << "IN BB: " << *BB; 899 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 900 errs() << " BB '" << BB->getName() << "': FOUND condition = "; 901 if (PredValues[i].first) 902 errs() << *PredValues[i].first; 903 else 904 errs() << "UNDEF"; 905 errs() << " for pred '" << PredValues[i].second->getName() 906 << "'.\n"; 907 }); 908 909 // Decide what we want to thread through. Convert our list of known values to 910 // a list of known destinations for each pred. This also discards duplicate 911 // predecessors and keeps track of the undefined inputs (which are represented 912 // as a null dest in the PredToDestList. 913 SmallPtrSet<BasicBlock*, 16> SeenPreds; 914 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 915 916 BasicBlock *OnlyDest = 0; 917 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 918 919 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 920 BasicBlock *Pred = PredValues[i].second; 921 if (!SeenPreds.insert(Pred)) 922 continue; // Duplicate predecessor entry. 923 924 // If the predecessor ends with an indirect goto, we can't change its 925 // destination. 926 if (isa<IndirectBrInst>(Pred->getTerminator())) 927 continue; 928 929 ConstantInt *Val = PredValues[i].first; 930 931 BasicBlock *DestBB; 932 if (Val == 0) // Undef. 933 DestBB = 0; 934 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 935 DestBB = BI->getSuccessor(Val->isZero()); 936 else { 937 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 938 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 939 } 940 941 // If we have exactly one destination, remember it for efficiency below. 942 if (i == 0) 943 OnlyDest = DestBB; 944 else if (OnlyDest != DestBB) 945 OnlyDest = MultipleDestSentinel; 946 947 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 948 } 949 950 // If all edges were unthreadable, we fail. 951 if (PredToDestList.empty()) 952 return false; 953 954 // Determine which is the most common successor. If we have many inputs and 955 // this block is a switch, we want to start by threading the batch that goes 956 // to the most popular destination first. If we only know about one 957 // threadable destination (the common case) we can avoid this. 958 BasicBlock *MostPopularDest = OnlyDest; 959 960 if (MostPopularDest == MultipleDestSentinel) 961 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 962 963 // Now that we know what the most popular destination is, factor all 964 // predecessors that will jump to it into a single predecessor. 965 SmallVector<BasicBlock*, 16> PredsToFactor; 966 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 967 if (PredToDestList[i].second == MostPopularDest) 968 PredsToFactor.push_back(PredToDestList[i].first); 969 970 BasicBlock *PredToThread; 971 if (PredsToFactor.size() == 1) 972 PredToThread = PredsToFactor[0]; 973 else { 974 DEBUG(errs() << " Factoring out " << PredsToFactor.size() 975 << " common predecessors.\n"); 976 PredToThread = SplitBlockPredecessors(BB, &PredsToFactor[0], 977 PredsToFactor.size(), 978 ".thr_comm", this); 979 } 980 981 // If the threadable edges are branching on an undefined value, we get to pick 982 // the destination that these predecessors should get to. 983 if (MostPopularDest == 0) 984 MostPopularDest = BB->getTerminator()-> 985 getSuccessor(GetBestDestForJumpOnUndef(BB)); 986 987 // Ok, try to thread it! 988 return ThreadEdge(BB, PredToThread, MostPopularDest); 989} 990 991/// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in 992/// the current block. See if there are any simplifications we can do based on 993/// inputs to the phi node. 994/// 995bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) { 996 BasicBlock *BB = PN->getParent(); 997 998 // If any of the predecessor blocks end in an unconditional branch, we can 999 // *duplicate* the jump into that block in order to further encourage jump 1000 // threading and to eliminate cases where we have branch on a phi of an icmp 1001 // (branch on icmp is much better). 1002 1003 // We don't want to do this tranformation for switches, because we don't 1004 // really want to duplicate a switch. 1005 if (isa<SwitchInst>(BB->getTerminator())) 1006 return false; 1007 1008 // Look for unconditional branch predecessors. 1009 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1010 BasicBlock *PredBB = PN->getIncomingBlock(i); 1011 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1012 if (PredBr->isUnconditional() && 1013 // Try to duplicate BB into PredBB. 1014 DuplicateCondBranchOnPHIIntoPred(BB, PredBB)) 1015 return true; 1016 } 1017 1018 return false; 1019} 1020 1021 1022/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1023/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1024/// NewPred using the entries from OldPred (suitably mapped). 1025static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1026 BasicBlock *OldPred, 1027 BasicBlock *NewPred, 1028 DenseMap<Instruction*, Value*> &ValueMap) { 1029 for (BasicBlock::iterator PNI = PHIBB->begin(); 1030 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1031 // Ok, we have a PHI node. Figure out what the incoming value was for the 1032 // DestBlock. 1033 Value *IV = PN->getIncomingValueForBlock(OldPred); 1034 1035 // Remap the value if necessary. 1036 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1037 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1038 if (I != ValueMap.end()) 1039 IV = I->second; 1040 } 1041 1042 PN->addIncoming(IV, NewPred); 1043 } 1044} 1045 1046/// ThreadEdge - We have decided that it is safe and profitable to thread an 1047/// edge from PredBB to SuccBB across BB. Transform the IR to reflect this 1048/// change. 1049bool JumpThreading::ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, 1050 BasicBlock *SuccBB) { 1051 // If threading to the same block as we come from, we would infinite loop. 1052 if (SuccBB == BB) { 1053 DEBUG(errs() << " Not threading across BB '" << BB->getName() 1054 << "' - would thread to self!\n"); 1055 return false; 1056 } 1057 1058 // If threading this would thread across a loop header, don't thread the edge. 1059 // See the comments above FindLoopHeaders for justifications and caveats. 1060 if (LoopHeaders.count(BB)) { 1061 DEBUG(errs() << " Not threading from '" << PredBB->getName() 1062 << "' across loop header BB '" << BB->getName() 1063 << "' to dest BB '" << SuccBB->getName() 1064 << "' - it might create an irreducible loop!\n"); 1065 return false; 1066 } 1067 1068 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1069 if (JumpThreadCost > Threshold) { 1070 DEBUG(errs() << " Not threading BB '" << BB->getName() 1071 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1072 return false; 1073 } 1074 1075 // And finally, do it! 1076 DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '" 1077 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1078 << ", across block:\n " 1079 << *BB << "\n"); 1080 1081 // We are going to have to map operands from the original BB block to the new 1082 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1083 // account for entry from PredBB. 1084 DenseMap<Instruction*, Value*> ValueMapping; 1085 1086 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1087 BB->getName()+".thread", 1088 BB->getParent(), BB); 1089 NewBB->moveAfter(PredBB); 1090 1091 BasicBlock::iterator BI = BB->begin(); 1092 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1093 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1094 1095 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1096 // mapping and using it to remap operands in the cloned instructions. 1097 for (; !isa<TerminatorInst>(BI); ++BI) { 1098 Instruction *New = BI->clone(); 1099 New->setName(BI->getName()); 1100 NewBB->getInstList().push_back(New); 1101 ValueMapping[BI] = New; 1102 1103 // Remap operands to patch up intra-block references. 1104 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1105 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1106 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1107 if (I != ValueMapping.end()) 1108 New->setOperand(i, I->second); 1109 } 1110 } 1111 1112 // We didn't copy the terminator from BB over to NewBB, because there is now 1113 // an unconditional jump to SuccBB. Insert the unconditional jump. 1114 BranchInst::Create(SuccBB, NewBB); 1115 1116 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1117 // PHI nodes for NewBB now. 1118 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1119 1120 // If there were values defined in BB that are used outside the block, then we 1121 // now have to update all uses of the value to use either the original value, 1122 // the cloned value, or some PHI derived value. This can require arbitrary 1123 // PHI insertion, of which we are prepared to do, clean these up now. 1124 SSAUpdater SSAUpdate; 1125 SmallVector<Use*, 16> UsesToRename; 1126 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1127 // Scan all uses of this instruction to see if it is used outside of its 1128 // block, and if so, record them in UsesToRename. 1129 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1130 ++UI) { 1131 Instruction *User = cast<Instruction>(*UI); 1132 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1133 if (UserPN->getIncomingBlock(UI) == BB) 1134 continue; 1135 } else if (User->getParent() == BB) 1136 continue; 1137 1138 UsesToRename.push_back(&UI.getUse()); 1139 } 1140 1141 // If there are no uses outside the block, we're done with this instruction. 1142 if (UsesToRename.empty()) 1143 continue; 1144 1145 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1146 1147 // We found a use of I outside of BB. Rename all uses of I that are outside 1148 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1149 // with the two values we know. 1150 SSAUpdate.Initialize(I); 1151 SSAUpdate.AddAvailableValue(BB, I); 1152 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1153 1154 while (!UsesToRename.empty()) 1155 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1156 DEBUG(errs() << "\n"); 1157 } 1158 1159 1160 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1161 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1162 // us to simplify any PHI nodes in BB. 1163 TerminatorInst *PredTerm = PredBB->getTerminator(); 1164 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1165 if (PredTerm->getSuccessor(i) == BB) { 1166 BB->removePredecessor(PredBB); 1167 PredTerm->setSuccessor(i, NewBB); 1168 } 1169 1170 // At this point, the IR is fully up to date and consistent. Do a quick scan 1171 // over the new instructions and zap any that are constants or dead. This 1172 // frequently happens because of phi translation. 1173 BI = NewBB->begin(); 1174 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) { 1175 Instruction *Inst = BI++; 1176 if (Constant *C = ConstantFoldInstruction(Inst, TD)) { 1177 Inst->replaceAllUsesWith(C); 1178 Inst->eraseFromParent(); 1179 continue; 1180 } 1181 1182 RecursivelyDeleteTriviallyDeadInstructions(Inst); 1183 } 1184 1185 // Threaded an edge! 1186 ++NumThreads; 1187 return true; 1188} 1189 1190/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1191/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1192/// If we can duplicate the contents of BB up into PredBB do so now, this 1193/// improves the odds that the branch will be on an analyzable instruction like 1194/// a compare. 1195bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1196 BasicBlock *PredBB) { 1197 // If BB is a loop header, then duplicating this block outside the loop would 1198 // cause us to transform this into an irreducible loop, don't do this. 1199 // See the comments above FindLoopHeaders for justifications and caveats. 1200 if (LoopHeaders.count(BB)) { 1201 DEBUG(errs() << " Not duplicating loop header '" << BB->getName() 1202 << "' into predecessor block '" << PredBB->getName() 1203 << "' - it might create an irreducible loop!\n"); 1204 return false; 1205 } 1206 1207 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1208 if (DuplicationCost > Threshold) { 1209 DEBUG(errs() << " Not duplicating BB '" << BB->getName() 1210 << "' - Cost is too high: " << DuplicationCost << "\n"); 1211 return false; 1212 } 1213 1214 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1215 // of PredBB. 1216 DEBUG(errs() << " Duplicating block '" << BB->getName() << "' into end of '" 1217 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1218 << DuplicationCost << " block is:" << *BB << "\n"); 1219 1220 // We are going to have to map operands from the original BB block into the 1221 // PredBB block. Evaluate PHI nodes in BB. 1222 DenseMap<Instruction*, Value*> ValueMapping; 1223 1224 BasicBlock::iterator BI = BB->begin(); 1225 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1226 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1227 1228 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1229 1230 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1231 // mapping and using it to remap operands in the cloned instructions. 1232 for (; BI != BB->end(); ++BI) { 1233 Instruction *New = BI->clone(); 1234 New->setName(BI->getName()); 1235 PredBB->getInstList().insert(OldPredBranch, New); 1236 ValueMapping[BI] = New; 1237 1238 // Remap operands to patch up intra-block references. 1239 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1240 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1241 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1242 if (I != ValueMapping.end()) 1243 New->setOperand(i, I->second); 1244 } 1245 } 1246 1247 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1248 // add entries to the PHI nodes for branch from PredBB now. 1249 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1250 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1251 ValueMapping); 1252 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1253 ValueMapping); 1254 1255 // If there were values defined in BB that are used outside the block, then we 1256 // now have to update all uses of the value to use either the original value, 1257 // the cloned value, or some PHI derived value. This can require arbitrary 1258 // PHI insertion, of which we are prepared to do, clean these up now. 1259 SSAUpdater SSAUpdate; 1260 SmallVector<Use*, 16> UsesToRename; 1261 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1262 // Scan all uses of this instruction to see if it is used outside of its 1263 // block, and if so, record them in UsesToRename. 1264 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1265 ++UI) { 1266 Instruction *User = cast<Instruction>(*UI); 1267 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1268 if (UserPN->getIncomingBlock(UI) == BB) 1269 continue; 1270 } else if (User->getParent() == BB) 1271 continue; 1272 1273 UsesToRename.push_back(&UI.getUse()); 1274 } 1275 1276 // If there are no uses outside the block, we're done with this instruction. 1277 if (UsesToRename.empty()) 1278 continue; 1279 1280 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1281 1282 // We found a use of I outside of BB. Rename all uses of I that are outside 1283 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1284 // with the two values we know. 1285 SSAUpdate.Initialize(I); 1286 SSAUpdate.AddAvailableValue(BB, I); 1287 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1288 1289 while (!UsesToRename.empty()) 1290 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1291 DEBUG(errs() << "\n"); 1292 } 1293 1294 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1295 // that we nuked. 1296 BB->removePredecessor(PredBB); 1297 1298 // Remove the unconditional branch at the end of the PredBB block. 1299 OldPredBranch->eraseFromParent(); 1300 1301 ++NumDupes; 1302 return true; 1303} 1304 1305 1306