JumpThreading.cpp revision 40d8c28b27377199b7465ba2c5a2c59c6fd12fa9
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/InstructionSimplify.h" 20#include "llvm/Transforms/Utils/BasicBlockUtils.h" 21#include "llvm/Transforms/Utils/Local.h" 22#include "llvm/Transforms/Utils/SSAUpdater.h" 23#include "llvm/Target/TargetData.h" 24#include "llvm/ADT/DenseMap.h" 25#include "llvm/ADT/Statistic.h" 26#include "llvm/ADT/STLExtras.h" 27#include "llvm/ADT/SmallPtrSet.h" 28#include "llvm/ADT/SmallSet.h" 29#include "llvm/Support/CommandLine.h" 30#include "llvm/Support/Debug.h" 31#include "llvm/Support/raw_ostream.h" 32using namespace llvm; 33 34STATISTIC(NumThreads, "Number of jumps threaded"); 35STATISTIC(NumFolds, "Number of terminators folded"); 36STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 37 38static cl::opt<unsigned> 39Threshold("jump-threading-threshold", 40 cl::desc("Max block size to duplicate for jump threading"), 41 cl::init(6), cl::Hidden); 42 43namespace { 44 /// This pass performs 'jump threading', which looks at blocks that have 45 /// multiple predecessors and multiple successors. If one or more of the 46 /// predecessors of the block can be proven to always jump to one of the 47 /// successors, we forward the edge from the predecessor to the successor by 48 /// duplicating the contents of this block. 49 /// 50 /// An example of when this can occur is code like this: 51 /// 52 /// if () { ... 53 /// X = 4; 54 /// } 55 /// if (X < 3) { 56 /// 57 /// In this case, the unconditional branch at the end of the first if can be 58 /// revectored to the false side of the second if. 59 /// 60 class JumpThreading : public FunctionPass { 61 TargetData *TD; 62#ifdef NDEBUG 63 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 64#else 65 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 66#endif 67 public: 68 static char ID; // Pass identification 69 JumpThreading() : FunctionPass(&ID) {} 70 71 bool runOnFunction(Function &F); 72 void FindLoopHeaders(Function &F); 73 74 bool ProcessBlock(BasicBlock *BB); 75 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 76 BasicBlock *SuccBB); 77 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 78 BasicBlock *PredBB); 79 80 typedef SmallVectorImpl<std::pair<ConstantInt*, 81 BasicBlock*> > PredValueInfo; 82 83 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 84 PredValueInfo &Result); 85 bool ProcessThreadableEdges(Instruction *CondInst, BasicBlock *BB); 86 87 88 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 89 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 90 91 bool ProcessJumpOnPHI(PHINode *PN); 92 93 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 94 }; 95} 96 97char JumpThreading::ID = 0; 98static RegisterPass<JumpThreading> 99X("jump-threading", "Jump Threading"); 100 101// Public interface to the Jump Threading pass 102FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 103 104/// runOnFunction - Top level algorithm. 105/// 106bool JumpThreading::runOnFunction(Function &F) { 107 DEBUG(errs() << "Jump threading on function '" << F.getName() << "'\n"); 108 TD = getAnalysisIfAvailable<TargetData>(); 109 110 FindLoopHeaders(F); 111 112 bool AnotherIteration = true, EverChanged = false; 113 while (AnotherIteration) { 114 AnotherIteration = false; 115 bool Changed = false; 116 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 117 BasicBlock *BB = I; 118 // Thread all of the branches we can over this block. 119 while (ProcessBlock(BB)) 120 Changed = true; 121 122 ++I; 123 124 // If the block is trivially dead, zap it. This eliminates the successor 125 // edges which simplifies the CFG. 126 if (pred_begin(BB) == pred_end(BB) && 127 BB != &BB->getParent()->getEntryBlock()) { 128 DEBUG(errs() << " JT: Deleting dead block '" << BB->getName() 129 << "' with terminator: " << *BB->getTerminator() << '\n'); 130 LoopHeaders.erase(BB); 131 DeleteDeadBlock(BB); 132 Changed = true; 133 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 134 // Can't thread an unconditional jump, but if the block is "almost 135 // empty", we can replace uses of it with uses of the successor and make 136 // this dead. 137 if (BI->isUnconditional() && 138 BB != &BB->getParent()->getEntryBlock()) { 139 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 140 // Ignore dbg intrinsics. 141 while (isa<DbgInfoIntrinsic>(BBI)) 142 ++BBI; 143 // If the terminator is the only non-phi instruction, try to nuke it. 144 if (BBI->isTerminator()) { 145 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 146 // block, we have to make sure it isn't in the LoopHeaders set. We 147 // reinsert afterward in the rare case when the block isn't deleted. 148 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 149 150 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) 151 Changed = true; 152 else if (ErasedFromLoopHeaders) 153 LoopHeaders.insert(BB); 154 } 155 } 156 } 157 } 158 AnotherIteration = Changed; 159 EverChanged |= Changed; 160 } 161 162 LoopHeaders.clear(); 163 return EverChanged; 164} 165 166/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 167/// thread across it. 168static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 169 /// Ignore PHI nodes, these will be flattened when duplication happens. 170 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 171 172 // Sum up the cost of each instruction until we get to the terminator. Don't 173 // include the terminator because the copy won't include it. 174 unsigned Size = 0; 175 for (; !isa<TerminatorInst>(I); ++I) { 176 // Debugger intrinsics don't incur code size. 177 if (isa<DbgInfoIntrinsic>(I)) continue; 178 179 // If this is a pointer->pointer bitcast, it is free. 180 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType())) 181 continue; 182 183 // All other instructions count for at least one unit. 184 ++Size; 185 186 // Calls are more expensive. If they are non-intrinsic calls, we model them 187 // as having cost of 4. If they are a non-vector intrinsic, we model them 188 // as having cost of 2 total, and if they are a vector intrinsic, we model 189 // them as having cost 1. 190 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 191 if (!isa<IntrinsicInst>(CI)) 192 Size += 3; 193 else if (!isa<VectorType>(CI->getType())) 194 Size += 1; 195 } 196 } 197 198 // Threading through a switch statement is particularly profitable. If this 199 // block ends in a switch, decrease its cost to make it more likely to happen. 200 if (isa<SwitchInst>(I)) 201 Size = Size > 6 ? Size-6 : 0; 202 203 return Size; 204} 205 206/// FindLoopHeaders - We do not want jump threading to turn proper loop 207/// structures into irreducible loops. Doing this breaks up the loop nesting 208/// hierarchy and pessimizes later transformations. To prevent this from 209/// happening, we first have to find the loop headers. Here we approximate this 210/// by finding targets of backedges in the CFG. 211/// 212/// Note that there definitely are cases when we want to allow threading of 213/// edges across a loop header. For example, threading a jump from outside the 214/// loop (the preheader) to an exit block of the loop is definitely profitable. 215/// It is also almost always profitable to thread backedges from within the loop 216/// to exit blocks, and is often profitable to thread backedges to other blocks 217/// within the loop (forming a nested loop). This simple analysis is not rich 218/// enough to track all of these properties and keep it up-to-date as the CFG 219/// mutates, so we don't allow any of these transformations. 220/// 221void JumpThreading::FindLoopHeaders(Function &F) { 222 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 223 FindFunctionBackedges(F, Edges); 224 225 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 226 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 227} 228 229/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 230/// if we can infer that the value is a known ConstantInt in any of our 231/// predecessors. If so, return the known list of value and pred BB in the 232/// result vector. If a value is known to be undef, it is returned as null. 233/// 234/// The BB basic block is known to start with a PHI node. 235/// 236/// This returns true if there were any known values. 237/// 238/// 239/// TODO: Per PR2563, we could infer value range information about a predecessor 240/// based on its terminator. 241bool JumpThreading:: 242ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 243 PHINode *TheFirstPHI = cast<PHINode>(BB->begin()); 244 245 // If V is a constantint, then it is known in all predecessors. 246 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 247 ConstantInt *CI = dyn_cast<ConstantInt>(V); 248 Result.resize(TheFirstPHI->getNumIncomingValues()); 249 for (unsigned i = 0, e = Result.size(); i != e; ++i) 250 Result[i] = std::make_pair(CI, TheFirstPHI->getIncomingBlock(i)); 251 return true; 252 } 253 254 // If V is a non-instruction value, or an instruction in a different block, 255 // then it can't be derived from a PHI. 256 Instruction *I = dyn_cast<Instruction>(V); 257 if (I == 0 || I->getParent() != BB) 258 return false; 259 260 /// If I is a PHI node, then we know the incoming values for any constants. 261 if (PHINode *PN = dyn_cast<PHINode>(I)) { 262 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 263 Value *InVal = PN->getIncomingValue(i); 264 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 265 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 266 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 267 } 268 } 269 return !Result.empty(); 270 } 271 272 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 273 274 // Handle some boolean conditions. 275 if (I->getType()->getPrimitiveSizeInBits() == 1) { 276 // X | true -> true 277 // X & false -> false 278 if (I->getOpcode() == Instruction::Or || 279 I->getOpcode() == Instruction::And) { 280 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 281 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 282 283 if (LHSVals.empty() && RHSVals.empty()) 284 return false; 285 286 ConstantInt *InterestingVal; 287 if (I->getOpcode() == Instruction::Or) 288 InterestingVal = ConstantInt::getTrue(I->getContext()); 289 else 290 InterestingVal = ConstantInt::getFalse(I->getContext()); 291 292 // Scan for the sentinel. 293 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 294 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) 295 Result.push_back(LHSVals[i]); 296 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 297 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) 298 Result.push_back(RHSVals[i]); 299 return !Result.empty(); 300 } 301 302 // TODO: Should handle the NOT form of XOR. 303 304 } 305 306 // Handle compare with phi operand, where the PHI is defined in this block. 307 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 308 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 309 if (PN && PN->getParent() == BB) { 310 // We can do this simplification if any comparisons fold to true or false. 311 // See if any do. 312 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 313 BasicBlock *PredBB = PN->getIncomingBlock(i); 314 Value *LHS = PN->getIncomingValue(i); 315 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 316 317 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS); 318 if (Res == 0) continue; 319 320 if (isa<UndefValue>(Res)) 321 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 322 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 323 Result.push_back(std::make_pair(CI, PredBB)); 324 } 325 326 return !Result.empty(); 327 } 328 329 // TODO: We could also recurse to see if we can determine constants another 330 // way. 331 } 332 return false; 333} 334 335 336 337/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 338/// in an undefined jump, decide which block is best to revector to. 339/// 340/// Since we can pick an arbitrary destination, we pick the successor with the 341/// fewest predecessors. This should reduce the in-degree of the others. 342/// 343static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 344 TerminatorInst *BBTerm = BB->getTerminator(); 345 unsigned MinSucc = 0; 346 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 347 // Compute the successor with the minimum number of predecessors. 348 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 349 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 350 TestBB = BBTerm->getSuccessor(i); 351 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 352 if (NumPreds < MinNumPreds) 353 MinSucc = i; 354 } 355 356 return MinSucc; 357} 358 359/// ProcessBlock - If there are any predecessors whose control can be threaded 360/// through to a successor, transform them now. 361bool JumpThreading::ProcessBlock(BasicBlock *BB) { 362 // If this block has a single predecessor, and if that pred has a single 363 // successor, merge the blocks. This encourages recursive jump threading 364 // because now the condition in this block can be threaded through 365 // predecessors of our predecessor block. 366 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 367 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 368 SinglePred != BB) { 369 // If SinglePred was a loop header, BB becomes one. 370 if (LoopHeaders.erase(SinglePred)) 371 LoopHeaders.insert(BB); 372 373 // Remember if SinglePred was the entry block of the function. If so, we 374 // will need to move BB back to the entry position. 375 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 376 MergeBasicBlockIntoOnlyPred(BB); 377 378 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 379 BB->moveBefore(&BB->getParent()->getEntryBlock()); 380 return true; 381 } 382 } 383 384 // Look to see if the terminator is a branch of switch, if not we can't thread 385 // it. 386 Value *Condition; 387 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 388 // Can't thread an unconditional jump. 389 if (BI->isUnconditional()) return false; 390 Condition = BI->getCondition(); 391 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 392 Condition = SI->getCondition(); 393 else 394 return false; // Must be an invoke. 395 396 // If the terminator of this block is branching on a constant, simplify the 397 // terminator to an unconditional branch. This can occur due to threading in 398 // other blocks. 399 if (isa<ConstantInt>(Condition)) { 400 DEBUG(errs() << " In block '" << BB->getName() 401 << "' folding terminator: " << *BB->getTerminator() << '\n'); 402 ++NumFolds; 403 ConstantFoldTerminator(BB); 404 return true; 405 } 406 407 // If the terminator is branching on an undef, we can pick any of the 408 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 409 if (isa<UndefValue>(Condition)) { 410 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 411 412 // Fold the branch/switch. 413 TerminatorInst *BBTerm = BB->getTerminator(); 414 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 415 if (i == BestSucc) continue; 416 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD); 417 } 418 419 DEBUG(errs() << " In block '" << BB->getName() 420 << "' folding undef terminator: " << *BBTerm << '\n'); 421 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 422 BBTerm->eraseFromParent(); 423 return true; 424 } 425 426 Instruction *CondInst = dyn_cast<Instruction>(Condition); 427 428 // If the condition is an instruction defined in another block, see if a 429 // predecessor has the same condition: 430 // br COND, BBX, BBY 431 // BBX: 432 // br COND, BBZ, BBW 433 if (!Condition->hasOneUse() && // Multiple uses. 434 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 435 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 436 if (isa<BranchInst>(BB->getTerminator())) { 437 for (; PI != E; ++PI) 438 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 439 if (PBI->isConditional() && PBI->getCondition() == Condition && 440 ProcessBranchOnDuplicateCond(*PI, BB)) 441 return true; 442 } else { 443 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 444 for (; PI != E; ++PI) 445 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator())) 446 if (PSI->getCondition() == Condition && 447 ProcessSwitchOnDuplicateCond(*PI, BB)) 448 return true; 449 } 450 } 451 452 // All the rest of our checks depend on the condition being an instruction. 453 if (CondInst == 0) 454 return false; 455 456 // See if this is a phi node in the current block. 457 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 458 if (PN->getParent() == BB) 459 return ProcessJumpOnPHI(PN); 460 461 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 462 if (!isa<PHINode>(CondCmp->getOperand(0)) || 463 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB) { 464 // If we have a comparison, loop over the predecessors to see if there is 465 // a condition with a lexically identical value. 466 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 467 for (; PI != E; ++PI) 468 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 469 if (PBI->isConditional() && *PI != BB) { 470 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 471 if (CI->getOperand(0) == CondCmp->getOperand(0) && 472 CI->getOperand(1) == CondCmp->getOperand(1) && 473 CI->getPredicate() == CondCmp->getPredicate()) { 474 // TODO: Could handle things like (x != 4) --> (x == 17) 475 if (ProcessBranchOnDuplicateCond(*PI, BB)) 476 return true; 477 } 478 } 479 } 480 } 481 } 482 483 // Check for some cases that are worth simplifying. Right now we want to look 484 // for loads that are used by a switch or by the condition for the branch. If 485 // we see one, check to see if it's partially redundant. If so, insert a PHI 486 // which can then be used to thread the values. 487 // 488 // This is particularly important because reg2mem inserts loads and stores all 489 // over the place, and this blocks jump threading if we don't zap them. 490 Value *SimplifyValue = CondInst; 491 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 492 if (isa<Constant>(CondCmp->getOperand(1))) 493 SimplifyValue = CondCmp->getOperand(0); 494 495 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 496 if (SimplifyPartiallyRedundantLoad(LI)) 497 return true; 498 499 500 // Handle a variety of cases where we are branching on something derived from 501 // a PHI node in the current block. If we can prove that any predecessors 502 // compute a predictable value based on a PHI node, thread those predecessors. 503 // 504 // We only bother doing this if the current block has a PHI node and if the 505 // conditional instruction lives in the current block. If either condition 506 // fails, this won't be a computable value anyway. 507 if (CondInst->getParent() == BB && isa<PHINode>(BB->front())) 508 if (ProcessThreadableEdges(CondInst, BB)) 509 return true; 510 511 512 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 513 // "(X == 4)" thread through this block. 514 515 return false; 516} 517 518/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 519/// block that jump on exactly the same condition. This means that we almost 520/// always know the direction of the edge in the DESTBB: 521/// PREDBB: 522/// br COND, DESTBB, BBY 523/// DESTBB: 524/// br COND, BBZ, BBW 525/// 526/// If DESTBB has multiple predecessors, we can't just constant fold the branch 527/// in DESTBB, we have to thread over it. 528bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 529 BasicBlock *BB) { 530 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 531 532 // If both successors of PredBB go to DESTBB, we don't know anything. We can 533 // fold the branch to an unconditional one, which allows other recursive 534 // simplifications. 535 bool BranchDir; 536 if (PredBI->getSuccessor(1) != BB) 537 BranchDir = true; 538 else if (PredBI->getSuccessor(0) != BB) 539 BranchDir = false; 540 else { 541 DEBUG(errs() << " In block '" << PredBB->getName() 542 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 543 ++NumFolds; 544 ConstantFoldTerminator(PredBB); 545 return true; 546 } 547 548 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 549 550 // If the dest block has one predecessor, just fix the branch condition to a 551 // constant and fold it. 552 if (BB->getSinglePredecessor()) { 553 DEBUG(errs() << " In block '" << BB->getName() 554 << "' folding condition to '" << BranchDir << "': " 555 << *BB->getTerminator() << '\n'); 556 ++NumFolds; 557 Value *OldCond = DestBI->getCondition(); 558 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 559 BranchDir)); 560 ConstantFoldTerminator(BB); 561 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 562 return true; 563 } 564 565 566 // Next, figure out which successor we are threading to. 567 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 568 569 SmallVector<BasicBlock*, 2> Preds; 570 Preds.push_back(PredBB); 571 572 // Ok, try to thread it! 573 return ThreadEdge(BB, Preds, SuccBB); 574} 575 576/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 577/// block that switch on exactly the same condition. This means that we almost 578/// always know the direction of the edge in the DESTBB: 579/// PREDBB: 580/// switch COND [... DESTBB, BBY ... ] 581/// DESTBB: 582/// switch COND [... BBZ, BBW ] 583/// 584/// Optimizing switches like this is very important, because simplifycfg builds 585/// switches out of repeated 'if' conditions. 586bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 587 BasicBlock *DestBB) { 588 // Can't thread edge to self. 589 if (PredBB == DestBB) 590 return false; 591 592 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 593 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 594 595 // There are a variety of optimizations that we can potentially do on these 596 // blocks: we order them from most to least preferable. 597 598 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 599 // directly to their destination. This does not introduce *any* code size 600 // growth. Skip debug info first. 601 BasicBlock::iterator BBI = DestBB->begin(); 602 while (isa<DbgInfoIntrinsic>(BBI)) 603 BBI++; 604 605 // FIXME: Thread if it just contains a PHI. 606 if (isa<SwitchInst>(BBI)) { 607 bool MadeChange = false; 608 // Ignore the default edge for now. 609 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 610 ConstantInt *DestVal = DestSI->getCaseValue(i); 611 BasicBlock *DestSucc = DestSI->getSuccessor(i); 612 613 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 614 // PredSI has an explicit case for it. If so, forward. If it is covered 615 // by the default case, we can't update PredSI. 616 unsigned PredCase = PredSI->findCaseValue(DestVal); 617 if (PredCase == 0) continue; 618 619 // If PredSI doesn't go to DestBB on this value, then it won't reach the 620 // case on this condition. 621 if (PredSI->getSuccessor(PredCase) != DestBB && 622 DestSI->getSuccessor(i) != DestBB) 623 continue; 624 625 // Otherwise, we're safe to make the change. Make sure that the edge from 626 // DestSI to DestSucc is not critical and has no PHI nodes. 627 DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 628 DEBUG(errs() << "THROUGH: " << *DestSI); 629 630 // If the destination has PHI nodes, just split the edge for updating 631 // simplicity. 632 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 633 SplitCriticalEdge(DestSI, i, this); 634 DestSucc = DestSI->getSuccessor(i); 635 } 636 FoldSingleEntryPHINodes(DestSucc); 637 PredSI->setSuccessor(PredCase, DestSucc); 638 MadeChange = true; 639 } 640 641 if (MadeChange) 642 return true; 643 } 644 645 return false; 646} 647 648 649/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 650/// load instruction, eliminate it by replacing it with a PHI node. This is an 651/// important optimization that encourages jump threading, and needs to be run 652/// interlaced with other jump threading tasks. 653bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 654 // Don't hack volatile loads. 655 if (LI->isVolatile()) return false; 656 657 // If the load is defined in a block with exactly one predecessor, it can't be 658 // partially redundant. 659 BasicBlock *LoadBB = LI->getParent(); 660 if (LoadBB->getSinglePredecessor()) 661 return false; 662 663 Value *LoadedPtr = LI->getOperand(0); 664 665 // If the loaded operand is defined in the LoadBB, it can't be available. 666 // FIXME: Could do PHI translation, that would be fun :) 667 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 668 if (PtrOp->getParent() == LoadBB) 669 return false; 670 671 // Scan a few instructions up from the load, to see if it is obviously live at 672 // the entry to its block. 673 BasicBlock::iterator BBIt = LI; 674 675 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB, 676 BBIt, 6)) { 677 // If the value if the load is locally available within the block, just use 678 // it. This frequently occurs for reg2mem'd allocas. 679 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 680 681 // If the returned value is the load itself, replace with an undef. This can 682 // only happen in dead loops. 683 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 684 LI->replaceAllUsesWith(AvailableVal); 685 LI->eraseFromParent(); 686 return true; 687 } 688 689 // Otherwise, if we scanned the whole block and got to the top of the block, 690 // we know the block is locally transparent to the load. If not, something 691 // might clobber its value. 692 if (BBIt != LoadBB->begin()) 693 return false; 694 695 696 SmallPtrSet<BasicBlock*, 8> PredsScanned; 697 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 698 AvailablePredsTy AvailablePreds; 699 BasicBlock *OneUnavailablePred = 0; 700 701 // If we got here, the loaded value is transparent through to the start of the 702 // block. Check to see if it is available in any of the predecessor blocks. 703 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 704 PI != PE; ++PI) { 705 BasicBlock *PredBB = *PI; 706 707 // If we already scanned this predecessor, skip it. 708 if (!PredsScanned.insert(PredBB)) 709 continue; 710 711 // Scan the predecessor to see if the value is available in the pred. 712 BBIt = PredBB->end(); 713 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 714 if (!PredAvailable) { 715 OneUnavailablePred = PredBB; 716 continue; 717 } 718 719 // If so, this load is partially redundant. Remember this info so that we 720 // can create a PHI node. 721 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 722 } 723 724 // If the loaded value isn't available in any predecessor, it isn't partially 725 // redundant. 726 if (AvailablePreds.empty()) return false; 727 728 // Okay, the loaded value is available in at least one (and maybe all!) 729 // predecessors. If the value is unavailable in more than one unique 730 // predecessor, we want to insert a merge block for those common predecessors. 731 // This ensures that we only have to insert one reload, thus not increasing 732 // code size. 733 BasicBlock *UnavailablePred = 0; 734 735 // If there is exactly one predecessor where the value is unavailable, the 736 // already computed 'OneUnavailablePred' block is it. If it ends in an 737 // unconditional branch, we know that it isn't a critical edge. 738 if (PredsScanned.size() == AvailablePreds.size()+1 && 739 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 740 UnavailablePred = OneUnavailablePred; 741 } else if (PredsScanned.size() != AvailablePreds.size()) { 742 // Otherwise, we had multiple unavailable predecessors or we had a critical 743 // edge from the one. 744 SmallVector<BasicBlock*, 8> PredsToSplit; 745 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 746 747 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 748 AvailablePredSet.insert(AvailablePreds[i].first); 749 750 // Add all the unavailable predecessors to the PredsToSplit list. 751 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 752 PI != PE; ++PI) 753 if (!AvailablePredSet.count(*PI)) 754 PredsToSplit.push_back(*PI); 755 756 // Split them out to their own block. 757 UnavailablePred = 758 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 759 "thread-split", this); 760 } 761 762 // If the value isn't available in all predecessors, then there will be 763 // exactly one where it isn't available. Insert a load on that edge and add 764 // it to the AvailablePreds list. 765 if (UnavailablePred) { 766 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 767 "Can't handle critical edge here!"); 768 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", 769 UnavailablePred->getTerminator()); 770 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 771 } 772 773 // Now we know that each predecessor of this block has a value in 774 // AvailablePreds, sort them for efficient access as we're walking the preds. 775 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 776 777 // Create a PHI node at the start of the block for the PRE'd load value. 778 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 779 PN->takeName(LI); 780 781 // Insert new entries into the PHI for each predecessor. A single block may 782 // have multiple entries here. 783 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 784 ++PI) { 785 AvailablePredsTy::iterator I = 786 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 787 std::make_pair(*PI, (Value*)0)); 788 789 assert(I != AvailablePreds.end() && I->first == *PI && 790 "Didn't find entry for predecessor!"); 791 792 PN->addIncoming(I->second, I->first); 793 } 794 795 //cerr << "PRE: " << *LI << *PN << "\n"; 796 797 LI->replaceAllUsesWith(PN); 798 LI->eraseFromParent(); 799 800 return true; 801} 802 803/// FindMostPopularDest - The specified list contains multiple possible 804/// threadable destinations. Pick the one that occurs the most frequently in 805/// the list. 806static BasicBlock * 807FindMostPopularDest(BasicBlock *BB, 808 const SmallVectorImpl<std::pair<BasicBlock*, 809 BasicBlock*> > &PredToDestList) { 810 assert(!PredToDestList.empty()); 811 812 // Determine popularity. If there are multiple possible destinations, we 813 // explicitly choose to ignore 'undef' destinations. We prefer to thread 814 // blocks with known and real destinations to threading undef. We'll handle 815 // them later if interesting. 816 DenseMap<BasicBlock*, unsigned> DestPopularity; 817 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 818 if (PredToDestList[i].second) 819 DestPopularity[PredToDestList[i].second]++; 820 821 // Find the most popular dest. 822 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 823 BasicBlock *MostPopularDest = DPI->first; 824 unsigned Popularity = DPI->second; 825 SmallVector<BasicBlock*, 4> SamePopularity; 826 827 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 828 // If the popularity of this entry isn't higher than the popularity we've 829 // seen so far, ignore it. 830 if (DPI->second < Popularity) 831 ; // ignore. 832 else if (DPI->second == Popularity) { 833 // If it is the same as what we've seen so far, keep track of it. 834 SamePopularity.push_back(DPI->first); 835 } else { 836 // If it is more popular, remember it. 837 SamePopularity.clear(); 838 MostPopularDest = DPI->first; 839 Popularity = DPI->second; 840 } 841 } 842 843 // Okay, now we know the most popular destination. If there is more than 844 // destination, we need to determine one. This is arbitrary, but we need 845 // to make a deterministic decision. Pick the first one that appears in the 846 // successor list. 847 if (!SamePopularity.empty()) { 848 SamePopularity.push_back(MostPopularDest); 849 TerminatorInst *TI = BB->getTerminator(); 850 for (unsigned i = 0; ; ++i) { 851 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 852 853 if (std::find(SamePopularity.begin(), SamePopularity.end(), 854 TI->getSuccessor(i)) == SamePopularity.end()) 855 continue; 856 857 MostPopularDest = TI->getSuccessor(i); 858 break; 859 } 860 } 861 862 // Okay, we have finally picked the most popular destination. 863 return MostPopularDest; 864} 865 866bool JumpThreading::ProcessThreadableEdges(Instruction *CondInst, 867 BasicBlock *BB) { 868 // If threading this would thread across a loop header, don't even try to 869 // thread the edge. 870 if (LoopHeaders.count(BB)) 871 return false; 872 873 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 874 if (!ComputeValueKnownInPredecessors(CondInst, BB, PredValues)) 875 return false; 876 assert(!PredValues.empty() && 877 "ComputeValueKnownInPredecessors returned true with no values"); 878 879 DEBUG(errs() << "IN BB: " << *BB; 880 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 881 errs() << " BB '" << BB->getName() << "': FOUND condition = "; 882 if (PredValues[i].first) 883 errs() << *PredValues[i].first; 884 else 885 errs() << "UNDEF"; 886 errs() << " for pred '" << PredValues[i].second->getName() 887 << "'.\n"; 888 }); 889 890 // Decide what we want to thread through. Convert our list of known values to 891 // a list of known destinations for each pred. This also discards duplicate 892 // predecessors and keeps track of the undefined inputs (which are represented 893 // as a null dest in the PredToDestList). 894 SmallPtrSet<BasicBlock*, 16> SeenPreds; 895 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 896 897 BasicBlock *OnlyDest = 0; 898 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 899 900 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 901 BasicBlock *Pred = PredValues[i].second; 902 if (!SeenPreds.insert(Pred)) 903 continue; // Duplicate predecessor entry. 904 905 // If the predecessor ends with an indirect goto, we can't change its 906 // destination. 907 if (isa<IndirectBrInst>(Pred->getTerminator())) 908 continue; 909 910 ConstantInt *Val = PredValues[i].first; 911 912 BasicBlock *DestBB; 913 if (Val == 0) // Undef. 914 DestBB = 0; 915 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 916 DestBB = BI->getSuccessor(Val->isZero()); 917 else { 918 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 919 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 920 } 921 922 // If we have exactly one destination, remember it for efficiency below. 923 if (i == 0) 924 OnlyDest = DestBB; 925 else if (OnlyDest != DestBB) 926 OnlyDest = MultipleDestSentinel; 927 928 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 929 } 930 931 // If all edges were unthreadable, we fail. 932 if (PredToDestList.empty()) 933 return false; 934 935 // Determine which is the most common successor. If we have many inputs and 936 // this block is a switch, we want to start by threading the batch that goes 937 // to the most popular destination first. If we only know about one 938 // threadable destination (the common case) we can avoid this. 939 BasicBlock *MostPopularDest = OnlyDest; 940 941 if (MostPopularDest == MultipleDestSentinel) 942 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 943 944 // Now that we know what the most popular destination is, factor all 945 // predecessors that will jump to it into a single predecessor. 946 SmallVector<BasicBlock*, 16> PredsToFactor; 947 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 948 if (PredToDestList[i].second == MostPopularDest) { 949 BasicBlock *Pred = PredToDestList[i].first; 950 951 // This predecessor may be a switch or something else that has multiple 952 // edges to the block. Factor each of these edges by listing them 953 // according to # occurrences in PredsToFactor. 954 TerminatorInst *PredTI = Pred->getTerminator(); 955 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 956 if (PredTI->getSuccessor(i) == BB) 957 PredsToFactor.push_back(Pred); 958 } 959 960 // If the threadable edges are branching on an undefined value, we get to pick 961 // the destination that these predecessors should get to. 962 if (MostPopularDest == 0) 963 MostPopularDest = BB->getTerminator()-> 964 getSuccessor(GetBestDestForJumpOnUndef(BB)); 965 966 // Ok, try to thread it! 967 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 968} 969 970/// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in 971/// the current block. See if there are any simplifications we can do based on 972/// inputs to the phi node. 973/// 974bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) { 975 BasicBlock *BB = PN->getParent(); 976 977 // If any of the predecessor blocks end in an unconditional branch, we can 978 // *duplicate* the jump into that block in order to further encourage jump 979 // threading and to eliminate cases where we have branch on a phi of an icmp 980 // (branch on icmp is much better). 981 982 // We don't want to do this tranformation for switches, because we don't 983 // really want to duplicate a switch. 984 if (isa<SwitchInst>(BB->getTerminator())) 985 return false; 986 987 // Look for unconditional branch predecessors. 988 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 989 BasicBlock *PredBB = PN->getIncomingBlock(i); 990 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 991 if (PredBr->isUnconditional() && 992 // Try to duplicate BB into PredBB. 993 DuplicateCondBranchOnPHIIntoPred(BB, PredBB)) 994 return true; 995 } 996 997 return false; 998} 999 1000 1001/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1002/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1003/// NewPred using the entries from OldPred (suitably mapped). 1004static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1005 BasicBlock *OldPred, 1006 BasicBlock *NewPred, 1007 DenseMap<Instruction*, Value*> &ValueMap) { 1008 for (BasicBlock::iterator PNI = PHIBB->begin(); 1009 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1010 // Ok, we have a PHI node. Figure out what the incoming value was for the 1011 // DestBlock. 1012 Value *IV = PN->getIncomingValueForBlock(OldPred); 1013 1014 // Remap the value if necessary. 1015 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1016 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1017 if (I != ValueMap.end()) 1018 IV = I->second; 1019 } 1020 1021 PN->addIncoming(IV, NewPred); 1022 } 1023} 1024 1025/// ThreadEdge - We have decided that it is safe and profitable to factor the 1026/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1027/// across BB. Transform the IR to reflect this change. 1028bool JumpThreading::ThreadEdge(BasicBlock *BB, 1029 const SmallVectorImpl<BasicBlock*> &PredBBs, 1030 BasicBlock *SuccBB) { 1031 // If threading to the same block as we come from, we would infinite loop. 1032 if (SuccBB == BB) { 1033 DEBUG(errs() << " Not threading across BB '" << BB->getName() 1034 << "' - would thread to self!\n"); 1035 return false; 1036 } 1037 1038 // If threading this would thread across a loop header, don't thread the edge. 1039 // See the comments above FindLoopHeaders for justifications and caveats. 1040 if (LoopHeaders.count(BB)) { 1041 DEBUG(errs() << " Not threading across loop header BB '" << BB->getName() 1042 << "' to dest BB '" << SuccBB->getName() 1043 << "' - it might create an irreducible loop!\n"); 1044 return false; 1045 } 1046 1047 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1048 if (JumpThreadCost > Threshold) { 1049 DEBUG(errs() << " Not threading BB '" << BB->getName() 1050 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1051 return false; 1052 } 1053 1054 // And finally, do it! Start by factoring the predecessors is needed. 1055 BasicBlock *PredBB; 1056 if (PredBBs.size() == 1) 1057 PredBB = PredBBs[0]; 1058 else { 1059 DEBUG(errs() << " Factoring out " << PredBBs.size() 1060 << " common predecessors.\n"); 1061 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1062 ".thr_comm", this); 1063 } 1064 1065 // And finally, do it! 1066 DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '" 1067 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1068 << ", across block:\n " 1069 << *BB << "\n"); 1070 1071 // We are going to have to map operands from the original BB block to the new 1072 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1073 // account for entry from PredBB. 1074 DenseMap<Instruction*, Value*> ValueMapping; 1075 1076 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1077 BB->getName()+".thread", 1078 BB->getParent(), BB); 1079 NewBB->moveAfter(PredBB); 1080 1081 BasicBlock::iterator BI = BB->begin(); 1082 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1083 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1084 1085 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1086 // mapping and using it to remap operands in the cloned instructions. 1087 for (; !isa<TerminatorInst>(BI); ++BI) { 1088 Instruction *New = BI->clone(); 1089 New->setName(BI->getName()); 1090 NewBB->getInstList().push_back(New); 1091 ValueMapping[BI] = New; 1092 1093 // Remap operands to patch up intra-block references. 1094 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1095 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1096 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1097 if (I != ValueMapping.end()) 1098 New->setOperand(i, I->second); 1099 } 1100 } 1101 1102 // We didn't copy the terminator from BB over to NewBB, because there is now 1103 // an unconditional jump to SuccBB. Insert the unconditional jump. 1104 BranchInst::Create(SuccBB, NewBB); 1105 1106 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1107 // PHI nodes for NewBB now. 1108 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1109 1110 // If there were values defined in BB that are used outside the block, then we 1111 // now have to update all uses of the value to use either the original value, 1112 // the cloned value, or some PHI derived value. This can require arbitrary 1113 // PHI insertion, of which we are prepared to do, clean these up now. 1114 SSAUpdater SSAUpdate; 1115 SmallVector<Use*, 16> UsesToRename; 1116 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1117 // Scan all uses of this instruction to see if it is used outside of its 1118 // block, and if so, record them in UsesToRename. 1119 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1120 ++UI) { 1121 Instruction *User = cast<Instruction>(*UI); 1122 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1123 if (UserPN->getIncomingBlock(UI) == BB) 1124 continue; 1125 } else if (User->getParent() == BB) 1126 continue; 1127 1128 UsesToRename.push_back(&UI.getUse()); 1129 } 1130 1131 // If there are no uses outside the block, we're done with this instruction. 1132 if (UsesToRename.empty()) 1133 continue; 1134 1135 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1136 1137 // We found a use of I outside of BB. Rename all uses of I that are outside 1138 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1139 // with the two values we know. 1140 SSAUpdate.Initialize(I); 1141 SSAUpdate.AddAvailableValue(BB, I); 1142 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1143 1144 while (!UsesToRename.empty()) 1145 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1146 DEBUG(errs() << "\n"); 1147 } 1148 1149 1150 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1151 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1152 // us to simplify any PHI nodes in BB. 1153 TerminatorInst *PredTerm = PredBB->getTerminator(); 1154 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1155 if (PredTerm->getSuccessor(i) == BB) { 1156 RemovePredecessorAndSimplify(BB, PredBB, TD); 1157 PredTerm->setSuccessor(i, NewBB); 1158 } 1159 1160 // At this point, the IR is fully up to date and consistent. Do a quick scan 1161 // over the new instructions and zap any that are constants or dead. This 1162 // frequently happens because of phi translation. 1163 BI = NewBB->begin(); 1164 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) { 1165 Instruction *Inst = BI++; 1166 1167 if (Value *V = SimplifyInstruction(Inst, TD)) { 1168 WeakVH BIHandle(BI); 1169 ReplaceAndSimplifyAllUses(Inst, V, TD); 1170 if (BIHandle == 0) 1171 BI = NewBB->begin(); 1172 continue; 1173 } 1174 1175 RecursivelyDeleteTriviallyDeadInstructions(Inst); 1176 } 1177 1178 // Threaded an edge! 1179 ++NumThreads; 1180 return true; 1181} 1182 1183/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1184/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1185/// If we can duplicate the contents of BB up into PredBB do so now, this 1186/// improves the odds that the branch will be on an analyzable instruction like 1187/// a compare. 1188bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1189 BasicBlock *PredBB) { 1190 // If BB is a loop header, then duplicating this block outside the loop would 1191 // cause us to transform this into an irreducible loop, don't do this. 1192 // See the comments above FindLoopHeaders for justifications and caveats. 1193 if (LoopHeaders.count(BB)) { 1194 DEBUG(errs() << " Not duplicating loop header '" << BB->getName() 1195 << "' into predecessor block '" << PredBB->getName() 1196 << "' - it might create an irreducible loop!\n"); 1197 return false; 1198 } 1199 1200 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1201 if (DuplicationCost > Threshold) { 1202 DEBUG(errs() << " Not duplicating BB '" << BB->getName() 1203 << "' - Cost is too high: " << DuplicationCost << "\n"); 1204 return false; 1205 } 1206 1207 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1208 // of PredBB. 1209 DEBUG(errs() << " Duplicating block '" << BB->getName() << "' into end of '" 1210 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1211 << DuplicationCost << " block is:" << *BB << "\n"); 1212 1213 // We are going to have to map operands from the original BB block into the 1214 // PredBB block. Evaluate PHI nodes in BB. 1215 DenseMap<Instruction*, Value*> ValueMapping; 1216 1217 BasicBlock::iterator BI = BB->begin(); 1218 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1219 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1220 1221 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1222 1223 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1224 // mapping and using it to remap operands in the cloned instructions. 1225 for (; BI != BB->end(); ++BI) { 1226 Instruction *New = BI->clone(); 1227 New->setName(BI->getName()); 1228 PredBB->getInstList().insert(OldPredBranch, New); 1229 ValueMapping[BI] = New; 1230 1231 // Remap operands to patch up intra-block references. 1232 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1233 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1234 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1235 if (I != ValueMapping.end()) 1236 New->setOperand(i, I->second); 1237 } 1238 } 1239 1240 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1241 // add entries to the PHI nodes for branch from PredBB now. 1242 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1243 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1244 ValueMapping); 1245 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1246 ValueMapping); 1247 1248 // If there were values defined in BB that are used outside the block, then we 1249 // now have to update all uses of the value to use either the original value, 1250 // the cloned value, or some PHI derived value. This can require arbitrary 1251 // PHI insertion, of which we are prepared to do, clean these up now. 1252 SSAUpdater SSAUpdate; 1253 SmallVector<Use*, 16> UsesToRename; 1254 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1255 // Scan all uses of this instruction to see if it is used outside of its 1256 // block, and if so, record them in UsesToRename. 1257 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1258 ++UI) { 1259 Instruction *User = cast<Instruction>(*UI); 1260 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1261 if (UserPN->getIncomingBlock(UI) == BB) 1262 continue; 1263 } else if (User->getParent() == BB) 1264 continue; 1265 1266 UsesToRename.push_back(&UI.getUse()); 1267 } 1268 1269 // If there are no uses outside the block, we're done with this instruction. 1270 if (UsesToRename.empty()) 1271 continue; 1272 1273 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1274 1275 // We found a use of I outside of BB. Rename all uses of I that are outside 1276 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1277 // with the two values we know. 1278 SSAUpdate.Initialize(I); 1279 SSAUpdate.AddAvailableValue(BB, I); 1280 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1281 1282 while (!UsesToRename.empty()) 1283 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1284 DEBUG(errs() << "\n"); 1285 } 1286 1287 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1288 // that we nuked. 1289 RemovePredecessorAndSimplify(BB, PredBB, TD); 1290 1291 // Remove the unconditional branch at the end of the PredBB block. 1292 OldPredBranch->eraseFromParent(); 1293 1294 ++NumDupes; 1295 return true; 1296} 1297 1298 1299