JumpThreading.cpp revision e33583b7c2b4d715a24e4d668bf446af9c836e54
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"); 36 37static cl::opt<unsigned> 38Threshold("jump-threading-threshold", 39 cl::desc("Max block size to duplicate for jump threading"), 40 cl::init(6), cl::Hidden); 41 42namespace { 43 /// This pass performs 'jump threading', which looks at blocks that have 44 /// multiple predecessors and multiple successors. If one or more of the 45 /// predecessors of the block can be proven to always jump to one of the 46 /// successors, we forward the edge from the predecessor to the successor by 47 /// duplicating the contents of this block. 48 /// 49 /// An example of when this can occur is code like this: 50 /// 51 /// if () { ... 52 /// X = 4; 53 /// } 54 /// if (X < 3) { 55 /// 56 /// In this case, the unconditional branch at the end of the first if can be 57 /// revectored to the false side of the second if. 58 /// 59 class JumpThreading : public FunctionPass { 60 TargetData *TD; 61#ifdef NDEBUG 62 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 63#else 64 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 65#endif 66 public: 67 static char ID; // Pass identification 68 JumpThreading() : FunctionPass(&ID) {} 69 70 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 71 } 72 73 bool runOnFunction(Function &F); 74 void FindLoopHeaders(Function &F); 75 76 bool ProcessBlock(BasicBlock *BB); 77 bool ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, BasicBlock *SuccBB, 78 unsigned JumpThreadCost); 79 BasicBlock *FactorCommonPHIPreds(PHINode *PN, Value *Val); 80 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 81 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 82 83 bool ProcessJumpOnPHI(PHINode *PN); 84 bool ProcessBranchOnLogical(Value *V, BasicBlock *BB, bool isAnd); 85 bool ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB); 86 87 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 88 }; 89} 90 91char JumpThreading::ID = 0; 92static RegisterPass<JumpThreading> 93X("jump-threading", "Jump Threading"); 94 95// Public interface to the Jump Threading pass 96FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 97 98/// runOnFunction - Top level algorithm. 99/// 100bool JumpThreading::runOnFunction(Function &F) { 101 DEBUG(errs() << "Jump threading on function '" << F.getName() << "'\n"); 102 TD = getAnalysisIfAvailable<TargetData>(); 103 104 FindLoopHeaders(F); 105 106 bool AnotherIteration = true, EverChanged = false; 107 while (AnotherIteration) { 108 AnotherIteration = false; 109 bool Changed = false; 110 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 111 BasicBlock *BB = I; 112 while (ProcessBlock(BB)) 113 Changed = true; 114 115 ++I; 116 117 // If the block is trivially dead, zap it. This eliminates the successor 118 // edges which simplifies the CFG. 119 if (pred_begin(BB) == pred_end(BB) && 120 BB != &BB->getParent()->getEntryBlock()) { 121 DEBUG(errs() << " JT: Deleting dead block '" << BB->getName() 122 << "' with terminator: " << *BB->getTerminator()); 123 LoopHeaders.erase(BB); 124 DeleteDeadBlock(BB); 125 Changed = true; 126 } 127 } 128 AnotherIteration = Changed; 129 EverChanged |= Changed; 130 } 131 132 LoopHeaders.clear(); 133 return EverChanged; 134} 135 136/// FindLoopHeaders - We do not want jump threading to turn proper loop 137/// structures into irreducible loops. Doing this breaks up the loop nesting 138/// hierarchy and pessimizes later transformations. To prevent this from 139/// happening, we first have to find the loop headers. Here we approximate this 140/// by finding targets of backedges in the CFG. 141/// 142/// Note that there definitely are cases when we want to allow threading of 143/// edges across a loop header. For example, threading a jump from outside the 144/// loop (the preheader) to an exit block of the loop is definitely profitable. 145/// It is also almost always profitable to thread backedges from within the loop 146/// to exit blocks, and is often profitable to thread backedges to other blocks 147/// within the loop (forming a nested loop). This simple analysis is not rich 148/// enough to track all of these properties and keep it up-to-date as the CFG 149/// mutates, so we don't allow any of these transformations. 150/// 151void JumpThreading::FindLoopHeaders(Function &F) { 152 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 153 FindFunctionBackedges(F, Edges); 154 155 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 156 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 157} 158 159 160/// FactorCommonPHIPreds - If there are multiple preds with the same incoming 161/// value for the PHI, factor them together so we get one block to thread for 162/// the whole group. 163/// This is important for things like "phi i1 [true, true, false, true, x]" 164/// where we only need to clone the block for the true blocks once. 165/// 166BasicBlock *JumpThreading::FactorCommonPHIPreds(PHINode *PN, Value *Val) { 167 SmallVector<BasicBlock*, 16> CommonPreds; 168 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 169 if (PN->getIncomingValue(i) == Val) 170 CommonPreds.push_back(PN->getIncomingBlock(i)); 171 172 if (CommonPreds.size() == 1) 173 return CommonPreds[0]; 174 175 DEBUG(errs() << " Factoring out " << CommonPreds.size() 176 << " common predecessors.\n"); 177 return SplitBlockPredecessors(PN->getParent(), 178 &CommonPreds[0], CommonPreds.size(), 179 ".thr_comm", this); 180} 181 182 183/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 184/// thread across it. 185static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 186 /// Ignore PHI nodes, these will be flattened when duplication happens. 187 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 188 189 // Sum up the cost of each instruction until we get to the terminator. Don't 190 // include the terminator because the copy won't include it. 191 unsigned Size = 0; 192 for (; !isa<TerminatorInst>(I); ++I) { 193 // Debugger intrinsics don't incur code size. 194 if (isa<DbgInfoIntrinsic>(I)) continue; 195 196 // If this is a pointer->pointer bitcast, it is free. 197 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType())) 198 continue; 199 200 // All other instructions count for at least one unit. 201 ++Size; 202 203 // Calls are more expensive. If they are non-intrinsic calls, we model them 204 // as having cost of 4. If they are a non-vector intrinsic, we model them 205 // as having cost of 2 total, and if they are a vector intrinsic, we model 206 // them as having cost 1. 207 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 208 if (!isa<IntrinsicInst>(CI)) 209 Size += 3; 210 else if (!isa<VectorType>(CI->getType())) 211 Size += 1; 212 } 213 } 214 215 // Threading through a switch statement is particularly profitable. If this 216 // block ends in a switch, decrease its cost to make it more likely to happen. 217 if (isa<SwitchInst>(I)) 218 Size = Size > 6 ? Size-6 : 0; 219 220 return Size; 221} 222 223/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 224/// in an undefined jump, decide which block is best to revector to. 225/// 226/// Since we can pick an arbitrary destination, we pick the successor with the 227/// fewest predecessors. This should reduce the in-degree of the others. 228/// 229static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 230 TerminatorInst *BBTerm = BB->getTerminator(); 231 unsigned MinSucc = 0; 232 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 233 // Compute the successor with the minimum number of predecessors. 234 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 235 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 236 TestBB = BBTerm->getSuccessor(i); 237 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 238 if (NumPreds < MinNumPreds) 239 MinSucc = i; 240 } 241 242 return MinSucc; 243} 244 245/// ProcessBlock - If there are any predecessors whose control can be threaded 246/// through to a successor, transform them now. 247bool JumpThreading::ProcessBlock(BasicBlock *BB) { 248 // If this block has a single predecessor, and if that pred has a single 249 // successor, merge the blocks. This encourages recursive jump threading 250 // because now the condition in this block can be threaded through 251 // predecessors of our predecessor block. 252 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) 253 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 254 SinglePred != BB) { 255 // If SinglePred was a loop header, BB becomes one. 256 if (LoopHeaders.erase(SinglePred)) 257 LoopHeaders.insert(BB); 258 259 // Remember if SinglePred was the entry block of the function. If so, we 260 // will need to move BB back to the entry position. 261 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 262 MergeBasicBlockIntoOnlyPred(BB); 263 264 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 265 BB->moveBefore(&BB->getParent()->getEntryBlock()); 266 return true; 267 } 268 269 // See if this block ends with a branch or switch. If so, see if the 270 // condition is a phi node. If so, and if an entry of the phi node is a 271 // constant, we can thread the block. 272 Value *Condition; 273 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 274 // Can't thread an unconditional jump. 275 if (BI->isUnconditional()) return false; 276 Condition = BI->getCondition(); 277 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 278 Condition = SI->getCondition(); 279 else 280 return false; // Must be an invoke. 281 282 // If the terminator of this block is branching on a constant, simplify the 283 // terminator to an unconditional branch. This can occur due to threading in 284 // other blocks. 285 if (isa<ConstantInt>(Condition)) { 286 DEBUG(errs() << " In block '" << BB->getName() 287 << "' folding terminator: " << *BB->getTerminator()); 288 ++NumFolds; 289 ConstantFoldTerminator(BB); 290 return true; 291 } 292 293 // If the terminator is branching on an undef, we can pick any of the 294 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 295 if (isa<UndefValue>(Condition)) { 296 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 297 298 // Fold the branch/switch. 299 TerminatorInst *BBTerm = BB->getTerminator(); 300 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 301 if (i == BestSucc) continue; 302 BBTerm->getSuccessor(i)->removePredecessor(BB); 303 } 304 305 DEBUG(errs() << " In block '" << BB->getName() 306 << "' folding undef terminator: " << *BBTerm); 307 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 308 BBTerm->eraseFromParent(); 309 return true; 310 } 311 312 Instruction *CondInst = dyn_cast<Instruction>(Condition); 313 314 // If the condition is an instruction defined in another block, see if a 315 // predecessor has the same condition: 316 // br COND, BBX, BBY 317 // BBX: 318 // br COND, BBZ, BBW 319 if (!Condition->hasOneUse() && // Multiple uses. 320 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 321 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 322 if (isa<BranchInst>(BB->getTerminator())) { 323 for (; PI != E; ++PI) 324 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 325 if (PBI->isConditional() && PBI->getCondition() == Condition && 326 ProcessBranchOnDuplicateCond(*PI, BB)) 327 return true; 328 } else { 329 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 330 for (; PI != E; ++PI) 331 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator())) 332 if (PSI->getCondition() == Condition && 333 ProcessSwitchOnDuplicateCond(*PI, BB)) 334 return true; 335 } 336 } 337 338 // All the rest of our checks depend on the condition being an instruction. 339 if (CondInst == 0) 340 return false; 341 342 // See if this is a phi node in the current block. 343 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 344 if (PN->getParent() == BB) 345 return ProcessJumpOnPHI(PN); 346 347 // If this is a conditional branch whose condition is and/or of a phi, try to 348 // simplify it. 349 if ((CondInst->getOpcode() == Instruction::And || 350 CondInst->getOpcode() == Instruction::Or) && 351 isa<BranchInst>(BB->getTerminator()) && 352 ProcessBranchOnLogical(CondInst, BB, 353 CondInst->getOpcode() == Instruction::And)) 354 return true; 355 356 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 357 if (isa<PHINode>(CondCmp->getOperand(0))) { 358 // If we have "br (phi != 42)" and the phi node has any constant values 359 // as operands, we can thread through this block. 360 // 361 // If we have "br (cmp phi, x)" and the phi node contains x such that the 362 // comparison uniquely identifies the branch target, we can thread 363 // through this block. 364 365 if (ProcessBranchOnCompare(CondCmp, BB)) 366 return true; 367 } 368 369 // If we have a comparison, loop over the predecessors to see if there is 370 // a condition with the same value. 371 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 372 for (; PI != E; ++PI) 373 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 374 if (PBI->isConditional() && *PI != BB) { 375 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 376 if (CI->getOperand(0) == CondCmp->getOperand(0) && 377 CI->getOperand(1) == CondCmp->getOperand(1) && 378 CI->getPredicate() == CondCmp->getPredicate()) { 379 // TODO: Could handle things like (x != 4) --> (x == 17) 380 if (ProcessBranchOnDuplicateCond(*PI, BB)) 381 return true; 382 } 383 } 384 } 385 } 386 387 // Check for some cases that are worth simplifying. Right now we want to look 388 // for loads that are used by a switch or by the condition for the branch. If 389 // we see one, check to see if it's partially redundant. If so, insert a PHI 390 // which can then be used to thread the values. 391 // 392 // This is particularly important because reg2mem inserts loads and stores all 393 // over the place, and this blocks jump threading if we don't zap them. 394 Value *SimplifyValue = CondInst; 395 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 396 if (isa<Constant>(CondCmp->getOperand(1))) 397 SimplifyValue = CondCmp->getOperand(0); 398 399 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 400 if (SimplifyPartiallyRedundantLoad(LI)) 401 return true; 402 403 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 404 // "(X == 4)" thread through this block. 405 406 return false; 407} 408 409/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 410/// block that jump on exactly the same condition. This means that we almost 411/// always know the direction of the edge in the DESTBB: 412/// PREDBB: 413/// br COND, DESTBB, BBY 414/// DESTBB: 415/// br COND, BBZ, BBW 416/// 417/// If DESTBB has multiple predecessors, we can't just constant fold the branch 418/// in DESTBB, we have to thread over it. 419bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 420 BasicBlock *BB) { 421 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 422 423 // If both successors of PredBB go to DESTBB, we don't know anything. We can 424 // fold the branch to an unconditional one, which allows other recursive 425 // simplifications. 426 bool BranchDir; 427 if (PredBI->getSuccessor(1) != BB) 428 BranchDir = true; 429 else if (PredBI->getSuccessor(0) != BB) 430 BranchDir = false; 431 else { 432 DEBUG(errs() << " In block '" << PredBB->getName() 433 << "' folding terminator: " << *PredBB->getTerminator()); 434 ++NumFolds; 435 ConstantFoldTerminator(PredBB); 436 return true; 437 } 438 439 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 440 441 // If the dest block has one predecessor, just fix the branch condition to a 442 // constant and fold it. 443 if (BB->getSinglePredecessor()) { 444 DEBUG(errs() << " In block '" << BB->getName() 445 << "' folding condition to '" << BranchDir << "': " 446 << *BB->getTerminator()); 447 ++NumFolds; 448 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 449 BranchDir)); 450 ConstantFoldTerminator(BB); 451 return true; 452 } 453 454 // Otherwise we need to thread from PredBB to DestBB's successor which 455 // involves code duplication. Check to see if it is worth it. 456 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 457 if (JumpThreadCost > Threshold) { 458 DEBUG(errs() << " Not threading BB '" << BB->getName() 459 << "' - Cost is too high: " << JumpThreadCost << "\n"); 460 return false; 461 } 462 463 // Next, figure out which successor we are threading to. 464 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 465 466 // Ok, try to thread it! 467 return ThreadEdge(BB, PredBB, SuccBB, JumpThreadCost); 468} 469 470/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 471/// block that switch on exactly the same condition. This means that we almost 472/// always know the direction of the edge in the DESTBB: 473/// PREDBB: 474/// switch COND [... DESTBB, BBY ... ] 475/// DESTBB: 476/// switch COND [... BBZ, BBW ] 477/// 478/// Optimizing switches like this is very important, because simplifycfg builds 479/// switches out of repeated 'if' conditions. 480bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 481 BasicBlock *DestBB) { 482 // Can't thread edge to self. 483 if (PredBB == DestBB) 484 return false; 485 486 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 487 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 488 489 // There are a variety of optimizations that we can potentially do on these 490 // blocks: we order them from most to least preferable. 491 492 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 493 // directly to their destination. This does not introduce *any* code size 494 // growth. Skip debug info first. 495 BasicBlock::iterator BBI = DestBB->begin(); 496 while (isa<DbgInfoIntrinsic>(BBI)) 497 BBI++; 498 499 // FIXME: Thread if it just contains a PHI. 500 if (isa<SwitchInst>(BBI)) { 501 bool MadeChange = false; 502 // Ignore the default edge for now. 503 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 504 ConstantInt *DestVal = DestSI->getCaseValue(i); 505 BasicBlock *DestSucc = DestSI->getSuccessor(i); 506 507 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 508 // PredSI has an explicit case for it. If so, forward. If it is covered 509 // by the default case, we can't update PredSI. 510 unsigned PredCase = PredSI->findCaseValue(DestVal); 511 if (PredCase == 0) continue; 512 513 // If PredSI doesn't go to DestBB on this value, then it won't reach the 514 // case on this condition. 515 if (PredSI->getSuccessor(PredCase) != DestBB && 516 DestSI->getSuccessor(i) != DestBB) 517 continue; 518 519 // Otherwise, we're safe to make the change. Make sure that the edge from 520 // DestSI to DestSucc is not critical and has no PHI nodes. 521 DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 522 DEBUG(errs() << "THROUGH: " << *DestSI); 523 524 // If the destination has PHI nodes, just split the edge for updating 525 // simplicity. 526 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 527 SplitCriticalEdge(DestSI, i, this); 528 DestSucc = DestSI->getSuccessor(i); 529 } 530 FoldSingleEntryPHINodes(DestSucc); 531 PredSI->setSuccessor(PredCase, DestSucc); 532 MadeChange = true; 533 } 534 535 if (MadeChange) 536 return true; 537 } 538 539 return false; 540} 541 542 543/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 544/// load instruction, eliminate it by replacing it with a PHI node. This is an 545/// important optimization that encourages jump threading, and needs to be run 546/// interlaced with other jump threading tasks. 547bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 548 // Don't hack volatile loads. 549 if (LI->isVolatile()) return false; 550 551 // If the load is defined in a block with exactly one predecessor, it can't be 552 // partially redundant. 553 BasicBlock *LoadBB = LI->getParent(); 554 if (LoadBB->getSinglePredecessor()) 555 return false; 556 557 Value *LoadedPtr = LI->getOperand(0); 558 559 // If the loaded operand is defined in the LoadBB, it can't be available. 560 // FIXME: Could do PHI translation, that would be fun :) 561 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 562 if (PtrOp->getParent() == LoadBB) 563 return false; 564 565 // Scan a few instructions up from the load, to see if it is obviously live at 566 // the entry to its block. 567 BasicBlock::iterator BBIt = LI; 568 569 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB, 570 BBIt, 6)) { 571 // If the value if the load is locally available within the block, just use 572 // it. This frequently occurs for reg2mem'd allocas. 573 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 574 575 // If the returned value is the load itself, replace with an undef. This can 576 // only happen in dead loops. 577 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 578 LI->replaceAllUsesWith(AvailableVal); 579 LI->eraseFromParent(); 580 return true; 581 } 582 583 // Otherwise, if we scanned the whole block and got to the top of the block, 584 // we know the block is locally transparent to the load. If not, something 585 // might clobber its value. 586 if (BBIt != LoadBB->begin()) 587 return false; 588 589 590 SmallPtrSet<BasicBlock*, 8> PredsScanned; 591 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 592 AvailablePredsTy AvailablePreds; 593 BasicBlock *OneUnavailablePred = 0; 594 595 // If we got here, the loaded value is transparent through to the start of the 596 // block. Check to see if it is available in any of the predecessor blocks. 597 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 598 PI != PE; ++PI) { 599 BasicBlock *PredBB = *PI; 600 601 // If we already scanned this predecessor, skip it. 602 if (!PredsScanned.insert(PredBB)) 603 continue; 604 605 // Scan the predecessor to see if the value is available in the pred. 606 BBIt = PredBB->end(); 607 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 608 if (!PredAvailable) { 609 OneUnavailablePred = PredBB; 610 continue; 611 } 612 613 // If so, this load is partially redundant. Remember this info so that we 614 // can create a PHI node. 615 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 616 } 617 618 // If the loaded value isn't available in any predecessor, it isn't partially 619 // redundant. 620 if (AvailablePreds.empty()) return false; 621 622 // Okay, the loaded value is available in at least one (and maybe all!) 623 // predecessors. If the value is unavailable in more than one unique 624 // predecessor, we want to insert a merge block for those common predecessors. 625 // This ensures that we only have to insert one reload, thus not increasing 626 // code size. 627 BasicBlock *UnavailablePred = 0; 628 629 // If there is exactly one predecessor where the value is unavailable, the 630 // already computed 'OneUnavailablePred' block is it. If it ends in an 631 // unconditional branch, we know that it isn't a critical edge. 632 if (PredsScanned.size() == AvailablePreds.size()+1 && 633 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 634 UnavailablePred = OneUnavailablePred; 635 } else if (PredsScanned.size() != AvailablePreds.size()) { 636 // Otherwise, we had multiple unavailable predecessors or we had a critical 637 // edge from the one. 638 SmallVector<BasicBlock*, 8> PredsToSplit; 639 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 640 641 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 642 AvailablePredSet.insert(AvailablePreds[i].first); 643 644 // Add all the unavailable predecessors to the PredsToSplit list. 645 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 646 PI != PE; ++PI) 647 if (!AvailablePredSet.count(*PI)) 648 PredsToSplit.push_back(*PI); 649 650 // Split them out to their own block. 651 UnavailablePred = 652 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 653 "thread-split", this); 654 } 655 656 // If the value isn't available in all predecessors, then there will be 657 // exactly one where it isn't available. Insert a load on that edge and add 658 // it to the AvailablePreds list. 659 if (UnavailablePred) { 660 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 661 "Can't handle critical edge here!"); 662 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", 663 UnavailablePred->getTerminator()); 664 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 665 } 666 667 // Now we know that each predecessor of this block has a value in 668 // AvailablePreds, sort them for efficient access as we're walking the preds. 669 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 670 671 // Create a PHI node at the start of the block for the PRE'd load value. 672 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 673 PN->takeName(LI); 674 675 // Insert new entries into the PHI for each predecessor. A single block may 676 // have multiple entries here. 677 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 678 ++PI) { 679 AvailablePredsTy::iterator I = 680 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 681 std::make_pair(*PI, (Value*)0)); 682 683 assert(I != AvailablePreds.end() && I->first == *PI && 684 "Didn't find entry for predecessor!"); 685 686 PN->addIncoming(I->second, I->first); 687 } 688 689 //cerr << "PRE: " << *LI << *PN << "\n"; 690 691 LI->replaceAllUsesWith(PN); 692 LI->eraseFromParent(); 693 694 return true; 695} 696 697 698/// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in 699/// the current block. See if there are any simplifications we can do based on 700/// inputs to the phi node. 701/// 702bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) { 703 // See if the phi node has any constant integer or undef values. If so, we 704 // can determine where the corresponding predecessor will branch. 705 Constant *PredCst = 0; 706 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 707 Value *PredVal = PN->getIncomingValue(i); 708 if (ConstantInt *CI = dyn_cast<ConstantInt>(PredVal)) { 709 PredCst = CI; 710 break; 711 } 712 713 if (UndefValue *UV = dyn_cast<UndefValue>(PredVal)) { 714 PredCst = UV; 715 break; 716 } 717 } 718 719 // If no incoming value has a constant, we don't know the destination of any 720 // predecessors. 721 if (PredCst == 0) { 722 return false; 723 } 724 725 // See if the cost of duplicating this block is low enough. 726 BasicBlock *BB = PN->getParent(); 727 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 728 if (JumpThreadCost > Threshold) { 729 DEBUG(errs() << " Not threading BB '" << BB->getName() 730 << "' - Cost is too high: " << JumpThreadCost << "\n"); 731 return false; 732 } 733 734 // If so, we can actually do this threading. Merge any common predecessors 735 // that will act the same. 736 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst); 737 738 739 TerminatorInst *BBTerm = BB->getTerminator(); 740 741 // Next, figure out which successor we are threading to. 742 BasicBlock *SuccBB; 743 if (isa<UndefValue>(PredCst)) { 744 // If the branch was going off an undef from PredBB, pick an arbitrary dest. 745 SuccBB = BBTerm->getSuccessor(GetBestDestForJumpOnUndef(BB)); 746 } else if (BranchInst *BI = dyn_cast<BranchInst>(BBTerm)) 747 SuccBB = BI->getSuccessor(cast<ConstantInt>(PredCst)->isZero()); 748 else { 749 SwitchInst *SI = cast<SwitchInst>(BBTerm); 750 SuccBB = SI->getSuccessor(SI->findCaseValue(cast<ConstantInt>(PredCst))); 751 } 752 753 // Ok, try to thread it! 754 return ThreadEdge(BB, PredBB, SuccBB, JumpThreadCost); 755} 756 757/// ProcessJumpOnLogicalPHI - PN's basic block contains a conditional branch 758/// whose condition is an AND/OR where one side is PN. If PN has constant 759/// operands that permit us to evaluate the condition for some operand, thread 760/// through the block. For example with: 761/// br (and X, phi(Y, Z, false)) 762/// the predecessor corresponding to the 'false' will always jump to the false 763/// destination of the branch. 764/// 765bool JumpThreading::ProcessBranchOnLogical(Value *V, BasicBlock *BB, 766 bool isAnd) { 767 // If this is a binary operator tree of the same AND/OR opcode, check the 768 // LHS/RHS. 769 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V)) 770 if ((isAnd && BO->getOpcode() == Instruction::And) || 771 (!isAnd && BO->getOpcode() == Instruction::Or)) { 772 if (ProcessBranchOnLogical(BO->getOperand(0), BB, isAnd)) 773 return true; 774 if (ProcessBranchOnLogical(BO->getOperand(1), BB, isAnd)) 775 return true; 776 } 777 778 // If this isn't a PHI node, we can't handle it. 779 PHINode *PN = dyn_cast<PHINode>(V); 780 if (!PN || PN->getParent() != BB) return false; 781 782 // We can only do the simplification for phi nodes of 'false' with AND or 783 // 'true' with OR. See if we have any entries in the phi for this. 784 unsigned PredNo = ~0U; 785 ConstantInt *PredCst = ConstantInt::get(Type::getInt1Ty(BB->getContext()), 786 !isAnd); 787 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 788 if (PN->getIncomingValue(i) == PredCst) { 789 PredNo = i; 790 break; 791 } 792 } 793 794 // If no match, bail out. 795 if (PredNo == ~0U) 796 return false; 797 798 // See if the cost of duplicating this block is low enough. 799 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 800 if (JumpThreadCost > Threshold) { 801 DEBUG(errs() << " Not threading BB '" << BB->getName() 802 << "' - Cost is too high: " << JumpThreadCost << "\n"); 803 return false; 804 } 805 806 // If so, we can actually do this threading. Merge any common predecessors 807 // that will act the same. 808 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst); 809 810 // Next, figure out which successor we are threading to. If this was an AND, 811 // the constant must be FALSE, and we must be targeting the 'false' block. 812 // If this is an OR, the constant must be TRUE, and we must be targeting the 813 // 'true' block. 814 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(isAnd); 815 816 // Ok, try to thread it! 817 return ThreadEdge(BB, PredBB, SuccBB, JumpThreadCost); 818} 819 820/// GetResultOfComparison - Given an icmp/fcmp predicate and the left and right 821/// hand sides of the compare instruction, try to determine the result. If the 822/// result can not be determined, a null pointer is returned. 823static Constant *GetResultOfComparison(CmpInst::Predicate pred, 824 Value *LHS, Value *RHS, 825 LLVMContext &Context) { 826 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 827 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 828 return ConstantExpr::getCompare(pred, CLHS, CRHS); 829 830 if (LHS == RHS) 831 if (isa<IntegerType>(LHS->getType()) || isa<PointerType>(LHS->getType())) 832 return ICmpInst::isTrueWhenEqual(pred) ? 833 ConstantInt::getTrue(Context) : ConstantInt::getFalse(Context); 834 835 return 0; 836} 837 838/// ProcessBranchOnCompare - We found a branch on a comparison between a phi 839/// node and a value. If we can identify when the comparison is true between 840/// the phi inputs and the value, we can fold the compare for that edge and 841/// thread through it. 842bool JumpThreading::ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB) { 843 PHINode *PN = cast<PHINode>(Cmp->getOperand(0)); 844 Value *RHS = Cmp->getOperand(1); 845 846 // If the phi isn't in the current block, an incoming edge to this block 847 // doesn't control the destination. 848 if (PN->getParent() != BB) 849 return false; 850 851 // We can do this simplification if any comparisons fold to true or false. 852 // See if any do. 853 Value *PredVal = 0; 854 bool TrueDirection = false; 855 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 856 PredVal = PN->getIncomingValue(i); 857 858 Constant *Res = GetResultOfComparison(Cmp->getPredicate(), PredVal, 859 RHS, Cmp->getContext()); 860 if (!Res) { 861 PredVal = 0; 862 continue; 863 } 864 865 // If this folded to a constant expr, we can't do anything. 866 if (ConstantInt *ResC = dyn_cast<ConstantInt>(Res)) { 867 TrueDirection = ResC->getZExtValue(); 868 break; 869 } 870 // If this folded to undef, just go the false way. 871 if (isa<UndefValue>(Res)) { 872 TrueDirection = false; 873 break; 874 } 875 876 // Otherwise, we can't fold this input. 877 PredVal = 0; 878 } 879 880 // If no match, bail out. 881 if (PredVal == 0) 882 return false; 883 884 // See if the cost of duplicating this block is low enough. 885 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 886 if (JumpThreadCost > Threshold) { 887 DEBUG(errs() << " Not threading BB '" << BB->getName() 888 << "' - Cost is too high: " << JumpThreadCost << "\n"); 889 return false; 890 } 891 892 // If so, we can actually do this threading. Merge any common predecessors 893 // that will act the same. 894 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredVal); 895 896 // Next, get our successor. 897 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(!TrueDirection); 898 899 // Ok, try to thread it! 900 return ThreadEdge(BB, PredBB, SuccBB, JumpThreadCost); 901} 902 903 904/// ThreadEdge - We have decided that it is safe and profitable to thread an 905/// edge from PredBB to SuccBB across BB. Transform the IR to reflect this 906/// change. 907bool JumpThreading::ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, 908 BasicBlock *SuccBB, unsigned JumpThreadCost) { 909 910 // If threading to the same block as we come from, we would infinite loop. 911 if (SuccBB == BB) { 912 DEBUG(errs() << " Not threading across BB '" << BB->getName() 913 << "' - would thread to self!\n"); 914 return false; 915 } 916 917 // If threading this would thread across a loop header, don't thread the edge. 918 // See the comments above FindLoopHeaders for justifications and caveats. 919 if (LoopHeaders.count(BB)) { 920 DEBUG(errs() << " Not threading from '" << PredBB->getName() 921 << "' across loop header BB '" << BB->getName() 922 << "' to dest BB '" << SuccBB->getName() 923 << "' - it might create an irreducible loop!\n"); 924 return false; 925 } 926 927 // And finally, do it! 928 DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '" 929 << SuccBB->getName() << "' with cost: " << JumpThreadCost 930 << ", across block:\n " 931 << *BB << "\n"); 932 933 // We are going to have to map operands from the original BB block to the new 934 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 935 // account for entry from PredBB. 936 DenseMap<Instruction*, Value*> ValueMapping; 937 938 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 939 BB->getName()+".thread", 940 BB->getParent(), BB); 941 NewBB->moveAfter(PredBB); 942 943 BasicBlock::iterator BI = BB->begin(); 944 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 945 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 946 947 // Clone the non-phi instructions of BB into NewBB, keeping track of the 948 // mapping and using it to remap operands in the cloned instructions. 949 for (; !isa<TerminatorInst>(BI); ++BI) { 950 Instruction *New = BI->clone(); 951 New->setName(BI->getName()); 952 NewBB->getInstList().push_back(New); 953 ValueMapping[BI] = New; 954 955 // Remap operands to patch up intra-block references. 956 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 957 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 958 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 959 if (I != ValueMapping.end()) 960 New->setOperand(i, I->second); 961 } 962 } 963 964 // We didn't copy the terminator from BB over to NewBB, because there is now 965 // an unconditional jump to SuccBB. Insert the unconditional jump. 966 BranchInst::Create(SuccBB, NewBB); 967 968 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 969 // PHI nodes for NewBB now. 970 for (BasicBlock::iterator PNI = SuccBB->begin(); 971 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 972 // Ok, we have a PHI node. Figure out what the incoming value was for the 973 // DestBlock. 974 Value *IV = PN->getIncomingValueForBlock(BB); 975 976 // Remap the value if necessary. 977 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 978 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 979 if (I != ValueMapping.end()) 980 IV = I->second; 981 } 982 PN->addIncoming(IV, NewBB); 983 } 984 985 // If there were values defined in BB that are used outside the block, then we 986 // now have to update all uses of the value to use either the original value, 987 // the cloned value, or some PHI derived value. This can require arbitrary 988 // PHI insertion, of which we are prepared to do, clean these up now. 989 SSAUpdater SSAUpdate; 990 SmallVector<Use*, 16> UsesToRename; 991 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 992 // Scan all uses of this instruction to see if it is used outside of its 993 // block, and if so, record them in UsesToRename. 994 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 995 ++UI) { 996 Instruction *User = cast<Instruction>(*UI); 997 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 998 if (UserPN->getIncomingBlock(UI) == BB) 999 continue; 1000 } else if (User->getParent() == BB) 1001 continue; 1002 1003 UsesToRename.push_back(&UI.getUse()); 1004 } 1005 1006 // If there are no uses outside the block, we're done with this instruction. 1007 if (UsesToRename.empty()) 1008 continue; 1009 1010 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1011 1012 // We found a use of I outside of BB. Rename all uses of I that are outside 1013 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1014 // with the two values we know. 1015 SSAUpdate.Initialize(I); 1016 SSAUpdate.AddAvailableValue(BB, I); 1017 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1018 1019 while (!UsesToRename.empty()) 1020 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1021 DEBUG(errs() << "\n"); 1022 } 1023 1024 1025 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1026 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1027 // us to simplify any PHI nodes in BB. 1028 TerminatorInst *PredTerm = PredBB->getTerminator(); 1029 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1030 if (PredTerm->getSuccessor(i) == BB) { 1031 BB->removePredecessor(PredBB); 1032 PredTerm->setSuccessor(i, NewBB); 1033 } 1034 1035 // At this point, the IR is fully up to date and consistent. Do a quick scan 1036 // over the new instructions and zap any that are constants or dead. This 1037 // frequently happens because of phi translation. 1038 BI = NewBB->begin(); 1039 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) { 1040 Instruction *Inst = BI++; 1041 if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) { 1042 Inst->replaceAllUsesWith(C); 1043 Inst->eraseFromParent(); 1044 continue; 1045 } 1046 1047 RecursivelyDeleteTriviallyDeadInstructions(Inst); 1048 } 1049 1050 // Threaded an edge! 1051 ++NumThreads; 1052 return true; 1053} 1054