JumpThreading.cpp revision 9683f181746c2312c8828c79b6e25a07fb441d57
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/Pass.h" 18#include "llvm/Analysis/ConstantFolding.h" 19#include "llvm/Transforms/Utils/BasicBlockUtils.h" 20#include "llvm/Transforms/Utils/Local.h" 21#include "llvm/Target/TargetData.h" 22#include "llvm/ADT/DenseMap.h" 23#include "llvm/ADT/Statistic.h" 24#include "llvm/ADT/STLExtras.h" 25#include "llvm/ADT/SmallPtrSet.h" 26#include "llvm/ADT/SmallSet.h" 27#include "llvm/Support/CommandLine.h" 28#include "llvm/Support/Compiler.h" 29#include "llvm/Support/Debug.h" 30#include "llvm/Support/ValueHandle.h" 31using namespace llvm; 32 33STATISTIC(NumThreads, "Number of jumps threaded"); 34STATISTIC(NumFolds, "Number of terminators folded"); 35 36static cl::opt<unsigned> 37Threshold("jump-threading-threshold", 38 cl::desc("Max block size to duplicate for jump threading"), 39 cl::init(6), cl::Hidden); 40 41namespace { 42 /// This pass performs 'jump threading', which looks at blocks that have 43 /// multiple predecessors and multiple successors. If one or more of the 44 /// predecessors of the block can be proven to always jump to one of the 45 /// successors, we forward the edge from the predecessor to the successor by 46 /// duplicating the contents of this block. 47 /// 48 /// An example of when this can occur is code like this: 49 /// 50 /// if () { ... 51 /// X = 4; 52 /// } 53 /// if (X < 3) { 54 /// 55 /// In this case, the unconditional branch at the end of the first if can be 56 /// revectored to the false side of the second if. 57 /// 58 class VISIBILITY_HIDDEN JumpThreading : public FunctionPass { 59 TargetData *TD; 60#ifdef NDEBUG 61 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 62#else 63 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 64#endif 65 public: 66 static char ID; // Pass identification 67 JumpThreading() : FunctionPass(&ID) {} 68 69 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 70 AU.addRequired<TargetData>(); 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 DOUT << "Jump threading on function '" << F.getNameStart() << "'\n"; 102 TD = &getAnalysis<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 DOUT << " JT: Deleting dead block '" << BB->getNameStart() 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 DOUT << " 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/// ProcessBlock - If there are any predecessors whose control can be threaded 224/// through to a successor, transform them now. 225bool JumpThreading::ProcessBlock(BasicBlock *BB) { 226 // If this block has a single predecessor, and if that pred has a single 227 // successor, merge the blocks. This encourages recursive jump threading 228 // because now the condition in this block can be threaded through 229 // predecessors of our predecessor block. 230 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) 231 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 232 SinglePred != BB) { 233 // If SinglePred was a loop header, BB becomes one. 234 if (LoopHeaders.erase(SinglePred)) 235 LoopHeaders.insert(BB); 236 237 // Remember if SinglePred was the entry block of the function. If so, we 238 // will need to move BB back to the entry position. 239 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 240 MergeBasicBlockIntoOnlyPred(BB); 241 242 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 243 BB->moveBefore(&BB->getParent()->getEntryBlock()); 244 return true; 245 } 246 247 // See if this block ends with a branch or switch. If so, see if the 248 // condition is a phi node. If so, and if an entry of the phi node is a 249 // constant, we can thread the block. 250 Value *Condition; 251 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 252 // Can't thread an unconditional jump. 253 if (BI->isUnconditional()) return false; 254 Condition = BI->getCondition(); 255 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 256 Condition = SI->getCondition(); 257 else 258 return false; // Must be an invoke. 259 260 // If the terminator of this block is branching on a constant, simplify the 261 // terminator to an unconditional branch. This can occur due to threading in 262 // other blocks. 263 if (isa<ConstantInt>(Condition)) { 264 DOUT << " In block '" << BB->getNameStart() 265 << "' folding terminator: " << *BB->getTerminator(); 266 ++NumFolds; 267 ConstantFoldTerminator(BB); 268 return true; 269 } 270 271 // If the terminator is branching on an undef, we can pick any of the 272 // successors to branch to. Since this is arbitrary, we pick the successor 273 // with the fewest predecessors. This should reduce the in-degree of the 274 // others. 275 if (isa<UndefValue>(Condition)) { 276 TerminatorInst *BBTerm = BB->getTerminator(); 277 unsigned MinSucc = 0; 278 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 279 // Compute the successor with the minimum number of predecessors. 280 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 281 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 282 TestBB = BBTerm->getSuccessor(i); 283 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 284 if (NumPreds < MinNumPreds) 285 MinSucc = i; 286 } 287 288 // Fold the branch/switch. 289 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 290 if (i == MinSucc) continue; 291 BBTerm->getSuccessor(i)->removePredecessor(BB); 292 } 293 294 DOUT << " In block '" << BB->getNameStart() 295 << "' folding undef terminator: " << *BBTerm; 296 BranchInst::Create(BBTerm->getSuccessor(MinSucc), BBTerm); 297 BBTerm->eraseFromParent(); 298 return true; 299 } 300 301 Instruction *CondInst = dyn_cast<Instruction>(Condition); 302 303 // If the condition is an instruction defined in another block, see if a 304 // predecessor has the same condition: 305 // br COND, BBX, BBY 306 // BBX: 307 // br COND, BBZ, BBW 308 if (!Condition->hasOneUse() && // Multiple uses. 309 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 310 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 311 if (isa<BranchInst>(BB->getTerminator())) { 312 for (; PI != E; ++PI) 313 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 314 if (PBI->isConditional() && PBI->getCondition() == Condition && 315 ProcessBranchOnDuplicateCond(*PI, BB)) 316 return true; 317 } else { 318 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 319 for (; PI != E; ++PI) 320 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator())) 321 if (PSI->getCondition() == Condition && 322 ProcessSwitchOnDuplicateCond(*PI, BB)) 323 return true; 324 } 325 } 326 327 // If there is only a single predecessor of this block, nothing to fold. 328 if (BB->getSinglePredecessor()) 329 return false; 330 331 // All the rest of our checks depend on the condition being an instruction. 332 if (CondInst == 0) 333 return false; 334 335 // See if this is a phi node in the current block. 336 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 337 if (PN->getParent() == BB) 338 return ProcessJumpOnPHI(PN); 339 340 // If this is a conditional branch whose condition is and/or of a phi, try to 341 // simplify it. 342 if ((CondInst->getOpcode() == Instruction::And || 343 CondInst->getOpcode() == Instruction::Or) && 344 isa<BranchInst>(BB->getTerminator()) && 345 ProcessBranchOnLogical(CondInst, BB, 346 CondInst->getOpcode() == Instruction::And)) 347 return true; 348 349 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 350 if (isa<PHINode>(CondCmp->getOperand(0))) { 351 // If we have "br (phi != 42)" and the phi node has any constant values 352 // as operands, we can thread through this block. 353 // 354 // If we have "br (cmp phi, x)" and the phi node contains x such that the 355 // comparison uniquely identifies the branch target, we can thread 356 // through this block. 357 358 if (ProcessBranchOnCompare(CondCmp, BB)) 359 return true; 360 } 361 } 362 363 // Check for some cases that are worth simplifying. Right now we want to look 364 // for loads that are used by a switch or by the condition for the branch. If 365 // we see one, check to see if it's partially redundant. If so, insert a PHI 366 // which can then be used to thread the values. 367 // 368 // This is particularly important because reg2mem inserts loads and stores all 369 // over the place, and this blocks jump threading if we don't zap them. 370 Value *SimplifyValue = CondInst; 371 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 372 if (isa<Constant>(CondCmp->getOperand(1))) 373 SimplifyValue = CondCmp->getOperand(0); 374 375 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 376 if (SimplifyPartiallyRedundantLoad(LI)) 377 return true; 378 379 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 380 // "(X == 4)" thread through this block. 381 382 return false; 383} 384 385/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 386/// block that jump on exactly the same condition. This means that we almost 387/// always know the direction of the edge in the DESTBB: 388/// PREDBB: 389/// br COND, DESTBB, BBY 390/// DESTBB: 391/// br COND, BBZ, BBW 392/// 393/// If DESTBB has multiple predecessors, we can't just constant fold the branch 394/// in DESTBB, we have to thread over it. 395bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 396 BasicBlock *BB) { 397 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 398 399 // If both successors of PredBB go to DESTBB, we don't know anything. We can 400 // fold the branch to an unconditional one, which allows other recursive 401 // simplifications. 402 bool BranchDir; 403 if (PredBI->getSuccessor(1) != BB) 404 BranchDir = true; 405 else if (PredBI->getSuccessor(0) != BB) 406 BranchDir = false; 407 else { 408 DOUT << " In block '" << PredBB->getNameStart() 409 << "' folding terminator: " << *PredBB->getTerminator(); 410 ++NumFolds; 411 ConstantFoldTerminator(PredBB); 412 return true; 413 } 414 415 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 416 417 // If the dest block has one predecessor, just fix the branch condition to a 418 // constant and fold it. 419 if (BB->getSinglePredecessor()) { 420 DOUT << " In block '" << BB->getNameStart() 421 << "' folding condition to '" << BranchDir << "': " 422 << *BB->getTerminator(); 423 ++NumFolds; 424 DestBI->setCondition(ConstantInt::get(Type::Int1Ty, BranchDir)); 425 ConstantFoldTerminator(BB); 426 return true; 427 } 428 429 // Otherwise we need to thread from PredBB to DestBB's successor which 430 // involves code duplication. Check to see if it is worth it. 431 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 432 if (JumpThreadCost > Threshold) { 433 DOUT << " Not threading BB '" << BB->getNameStart() 434 << "' - Cost is too high: " << JumpThreadCost << "\n"; 435 return false; 436 } 437 438 // Next, figure out which successor we are threading to. 439 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 440 441 // Ok, try to thread it! 442 return ThreadEdge(BB, PredBB, SuccBB, JumpThreadCost); 443} 444 445/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 446/// block that switch on exactly the same condition. This means that we almost 447/// always know the direction of the edge in the DESTBB: 448/// PREDBB: 449/// switch COND [... DESTBB, BBY ... ] 450/// DESTBB: 451/// switch COND [... BBZ, BBW ] 452/// 453/// Optimizing switches like this is very important, because simplifycfg builds 454/// switches out of repeated 'if' conditions. 455bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 456 BasicBlock *DestBB) { 457 // Can't thread edge to self. 458 if (PredBB == DestBB) 459 return false; 460 461 462 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 463 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 464 465 // There are a variety of optimizations that we can potentially do on these 466 // blocks: we order them from most to least preferable. 467 468 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 469 // directly to their destination. This does not introduce *any* code size 470 // growth. Skip debug info first. 471 BasicBlock::iterator BBI = DestBB->begin(); 472 while (isa<DbgInfoIntrinsic>(BBI)) 473 BBI++; 474 475 // FIXME: Thread if it just contains a PHI. 476 if (isa<SwitchInst>(BBI)) { 477 bool MadeChange = false; 478 // Ignore the default edge for now. 479 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 480 ConstantInt *DestVal = DestSI->getCaseValue(i); 481 BasicBlock *DestSucc = DestSI->getSuccessor(i); 482 483 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 484 // PredSI has an explicit case for it. If so, forward. If it is covered 485 // by the default case, we can't update PredSI. 486 unsigned PredCase = PredSI->findCaseValue(DestVal); 487 if (PredCase == 0) continue; 488 489 // If PredSI doesn't go to DestBB on this value, then it won't reach the 490 // case on this condition. 491 if (PredSI->getSuccessor(PredCase) != DestBB && 492 DestSI->getSuccessor(i) != DestBB) 493 continue; 494 495 // Otherwise, we're safe to make the change. Make sure that the edge from 496 // DestSI to DestSucc is not critical and has no PHI nodes. 497 DOUT << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI; 498 DOUT << "THROUGH: " << *DestSI; 499 500 // If the destination has PHI nodes, just split the edge for updating 501 // simplicity. 502 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 503 SplitCriticalEdge(DestSI, i, this); 504 DestSucc = DestSI->getSuccessor(i); 505 } 506 FoldSingleEntryPHINodes(DestSucc); 507 PredSI->setSuccessor(PredCase, DestSucc); 508 MadeChange = true; 509 } 510 511 if (MadeChange) 512 return true; 513 } 514 515 return false; 516} 517 518 519/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 520/// load instruction, eliminate it by replacing it with a PHI node. This is an 521/// important optimization that encourages jump threading, and needs to be run 522/// interlaced with other jump threading tasks. 523bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 524 // Don't hack volatile loads. 525 if (LI->isVolatile()) return false; 526 527 // If the load is defined in a block with exactly one predecessor, it can't be 528 // partially redundant. 529 BasicBlock *LoadBB = LI->getParent(); 530 if (LoadBB->getSinglePredecessor()) 531 return false; 532 533 Value *LoadedPtr = LI->getOperand(0); 534 535 // If the loaded operand is defined in the LoadBB, it can't be available. 536 // FIXME: Could do PHI translation, that would be fun :) 537 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 538 if (PtrOp->getParent() == LoadBB) 539 return false; 540 541 // Scan a few instructions up from the load, to see if it is obviously live at 542 // the entry to its block. 543 BasicBlock::iterator BBIt = LI; 544 545 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB, 546 BBIt, 6)) { 547 // If the value if the load is locally available within the block, just use 548 // it. This frequently occurs for reg2mem'd allocas. 549 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 550 551 // If the returned value is the load itself, replace with an undef. This can 552 // only happen in dead loops. 553 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 554 LI->replaceAllUsesWith(AvailableVal); 555 LI->eraseFromParent(); 556 return true; 557 } 558 559 // Otherwise, if we scanned the whole block and got to the top of the block, 560 // we know the block is locally transparent to the load. If not, something 561 // might clobber its value. 562 if (BBIt != LoadBB->begin()) 563 return false; 564 565 566 SmallPtrSet<BasicBlock*, 8> PredsScanned; 567 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 568 AvailablePredsTy AvailablePreds; 569 BasicBlock *OneUnavailablePred = 0; 570 571 // If we got here, the loaded value is transparent through to the start of the 572 // block. Check to see if it is available in any of the predecessor blocks. 573 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 574 PI != PE; ++PI) { 575 BasicBlock *PredBB = *PI; 576 577 // If we already scanned this predecessor, skip it. 578 if (!PredsScanned.insert(PredBB)) 579 continue; 580 581 // Scan the predecessor to see if the value is available in the pred. 582 BBIt = PredBB->end(); 583 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 584 if (!PredAvailable) { 585 OneUnavailablePred = PredBB; 586 continue; 587 } 588 589 // If so, this load is partially redundant. Remember this info so that we 590 // can create a PHI node. 591 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 592 } 593 594 // If the loaded value isn't available in any predecessor, it isn't partially 595 // redundant. 596 if (AvailablePreds.empty()) return false; 597 598 // Okay, the loaded value is available in at least one (and maybe all!) 599 // predecessors. If the value is unavailable in more than one unique 600 // predecessor, we want to insert a merge block for those common predecessors. 601 // This ensures that we only have to insert one reload, thus not increasing 602 // code size. 603 BasicBlock *UnavailablePred = 0; 604 605 // If there is exactly one predecessor where the value is unavailable, the 606 // already computed 'OneUnavailablePred' block is it. If it ends in an 607 // unconditional branch, we know that it isn't a critical edge. 608 if (PredsScanned.size() == AvailablePreds.size()+1 && 609 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 610 UnavailablePred = OneUnavailablePred; 611 } else if (PredsScanned.size() != AvailablePreds.size()) { 612 // Otherwise, we had multiple unavailable predecessors or we had a critical 613 // edge from the one. 614 SmallVector<BasicBlock*, 8> PredsToSplit; 615 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 616 617 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 618 AvailablePredSet.insert(AvailablePreds[i].first); 619 620 // Add all the unavailable predecessors to the PredsToSplit list. 621 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 622 PI != PE; ++PI) 623 if (!AvailablePredSet.count(*PI)) 624 PredsToSplit.push_back(*PI); 625 626 // Split them out to their own block. 627 UnavailablePred = 628 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 629 "thread-split", this); 630 } 631 632 // If the value isn't available in all predecessors, then there will be 633 // exactly one where it isn't available. Insert a load on that edge and add 634 // it to the AvailablePreds list. 635 if (UnavailablePred) { 636 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 637 "Can't handle critical edge here!"); 638 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", 639 UnavailablePred->getTerminator()); 640 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 641 } 642 643 // Now we know that each predecessor of this block has a value in 644 // AvailablePreds, sort them for efficient access as we're walking the preds. 645 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 646 647 // Create a PHI node at the start of the block for the PRE'd load value. 648 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 649 PN->takeName(LI); 650 651 // Insert new entries into the PHI for each predecessor. A single block may 652 // have multiple entries here. 653 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 654 ++PI) { 655 AvailablePredsTy::iterator I = 656 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 657 std::make_pair(*PI, (Value*)0)); 658 659 assert(I != AvailablePreds.end() && I->first == *PI && 660 "Didn't find entry for predecessor!"); 661 662 PN->addIncoming(I->second, I->first); 663 } 664 665 //cerr << "PRE: " << *LI << *PN << "\n"; 666 667 LI->replaceAllUsesWith(PN); 668 LI->eraseFromParent(); 669 670 return true; 671} 672 673 674/// ProcessJumpOnPHI - We have a conditional branch of switch on a PHI node in 675/// the current block. See if there are any simplifications we can do based on 676/// inputs to the phi node. 677/// 678bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) { 679 // See if the phi node has any constant values. If so, we can determine where 680 // the corresponding predecessor will branch. 681 ConstantInt *PredCst = 0; 682 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 683 if ((PredCst = dyn_cast<ConstantInt>(PN->getIncomingValue(i)))) 684 break; 685 686 // If no incoming value has a constant, we don't know the destination of any 687 // predecessors. 688 if (PredCst == 0) 689 return false; 690 691 // See if the cost of duplicating this block is low enough. 692 BasicBlock *BB = PN->getParent(); 693 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 694 if (JumpThreadCost > Threshold) { 695 DOUT << " Not threading BB '" << BB->getNameStart() 696 << "' - Cost is too high: " << JumpThreadCost << "\n"; 697 return false; 698 } 699 700 // If so, we can actually do this threading. Merge any common predecessors 701 // that will act the same. 702 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst); 703 704 // Next, figure out which successor we are threading to. 705 BasicBlock *SuccBB; 706 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 707 SuccBB = BI->getSuccessor(PredCst == ConstantInt::getFalse()); 708 else { 709 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 710 SuccBB = SI->getSuccessor(SI->findCaseValue(PredCst)); 711 } 712 713 // Ok, try to thread it! 714 return ThreadEdge(BB, PredBB, SuccBB, JumpThreadCost); 715} 716 717/// ProcessJumpOnLogicalPHI - PN's basic block contains a conditional branch 718/// whose condition is an AND/OR where one side is PN. If PN has constant 719/// operands that permit us to evaluate the condition for some operand, thread 720/// through the block. For example with: 721/// br (and X, phi(Y, Z, false)) 722/// the predecessor corresponding to the 'false' will always jump to the false 723/// destination of the branch. 724/// 725bool JumpThreading::ProcessBranchOnLogical(Value *V, BasicBlock *BB, 726 bool isAnd) { 727 // If this is a binary operator tree of the same AND/OR opcode, check the 728 // LHS/RHS. 729 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V)) 730 if ((isAnd && BO->getOpcode() == Instruction::And) || 731 (!isAnd && BO->getOpcode() == Instruction::Or)) { 732 if (ProcessBranchOnLogical(BO->getOperand(0), BB, isAnd)) 733 return true; 734 if (ProcessBranchOnLogical(BO->getOperand(1), BB, isAnd)) 735 return true; 736 } 737 738 // If this isn't a PHI node, we can't handle it. 739 PHINode *PN = dyn_cast<PHINode>(V); 740 if (!PN || PN->getParent() != BB) return false; 741 742 // We can only do the simplification for phi nodes of 'false' with AND or 743 // 'true' with OR. See if we have any entries in the phi for this. 744 unsigned PredNo = ~0U; 745 ConstantInt *PredCst = ConstantInt::get(Type::Int1Ty, !isAnd); 746 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 747 if (PN->getIncomingValue(i) == PredCst) { 748 PredNo = i; 749 break; 750 } 751 } 752 753 // If no match, bail out. 754 if (PredNo == ~0U) 755 return false; 756 757 // See if the cost of duplicating this block is low enough. 758 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 759 if (JumpThreadCost > Threshold) { 760 DOUT << " Not threading BB '" << BB->getNameStart() 761 << "' - Cost is too high: " << JumpThreadCost << "\n"; 762 return false; 763 } 764 765 // If so, we can actually do this threading. Merge any common predecessors 766 // that will act the same. 767 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst); 768 769 // Next, figure out which successor we are threading to. If this was an AND, 770 // the constant must be FALSE, and we must be targeting the 'false' block. 771 // If this is an OR, the constant must be TRUE, and we must be targeting the 772 // 'true' block. 773 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(isAnd); 774 775 // Ok, try to thread it! 776 return ThreadEdge(BB, PredBB, SuccBB, JumpThreadCost); 777} 778 779/// GetResultOfComparison - Given an icmp/fcmp predicate and the left and right 780/// hand sides of the compare instruction, try to determine the result. If the 781/// result can not be determined, a null pointer is returned. 782static Constant *GetResultOfComparison(CmpInst::Predicate pred, 783 Value *LHS, Value *RHS) { 784 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 785 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 786 return ConstantExpr::getCompare(pred, CLHS, CRHS); 787 788 if (LHS == RHS) 789 if (isa<IntegerType>(LHS->getType()) || isa<PointerType>(LHS->getType())) 790 return ICmpInst::isTrueWhenEqual(pred) ? 791 ConstantInt::getTrue() : ConstantInt::getFalse(); 792 793 return 0; 794} 795 796/// ProcessBranchOnCompare - We found a branch on a comparison between a phi 797/// node and a value. If we can identify when the comparison is true between 798/// the phi inputs and the value, we can fold the compare for that edge and 799/// thread through it. 800bool JumpThreading::ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB) { 801 PHINode *PN = cast<PHINode>(Cmp->getOperand(0)); 802 Value *RHS = Cmp->getOperand(1); 803 804 // If the phi isn't in the current block, an incoming edge to this block 805 // doesn't control the destination. 806 if (PN->getParent() != BB) 807 return false; 808 809 // We can do this simplification if any comparisons fold to true or false. 810 // See if any do. 811 Value *PredVal = 0; 812 bool TrueDirection = false; 813 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 814 PredVal = PN->getIncomingValue(i); 815 816 Constant *Res = GetResultOfComparison(Cmp->getPredicate(), PredVal, RHS); 817 if (!Res) { 818 PredVal = 0; 819 continue; 820 } 821 822 // If this folded to a constant expr, we can't do anything. 823 if (ConstantInt *ResC = dyn_cast<ConstantInt>(Res)) { 824 TrueDirection = ResC->getZExtValue(); 825 break; 826 } 827 // If this folded to undef, just go the false way. 828 if (isa<UndefValue>(Res)) { 829 TrueDirection = false; 830 break; 831 } 832 833 // Otherwise, we can't fold this input. 834 PredVal = 0; 835 } 836 837 // If no match, bail out. 838 if (PredVal == 0) 839 return false; 840 841 // See if the cost of duplicating this block is low enough. 842 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 843 if (JumpThreadCost > Threshold) { 844 DOUT << " Not threading BB '" << BB->getNameStart() 845 << "' - Cost is too high: " << JumpThreadCost << "\n"; 846 return false; 847 } 848 849 // If so, we can actually do this threading. Merge any common predecessors 850 // that will act the same. 851 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredVal); 852 853 // Next, get our successor. 854 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(!TrueDirection); 855 856 // Ok, try to thread it! 857 return ThreadEdge(BB, PredBB, SuccBB, JumpThreadCost); 858} 859 860 861/// ThreadEdge - We have decided that it is safe and profitable to thread an 862/// edge from PredBB to SuccBB across BB. Transform the IR to reflect this 863/// change. 864bool JumpThreading::ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, 865 BasicBlock *SuccBB, unsigned JumpThreadCost) { 866 867 // If threading to the same block as we come from, we would infinite loop. 868 if (SuccBB == BB) { 869 DOUT << " Not threading across BB '" << BB->getNameStart() 870 << "' - would thread to self!\n"; 871 return false; 872 } 873 874 // If threading this would thread across a loop header, don't thread the edge. 875 // See the comments above FindLoopHeaders for justifications and caveats. 876 if (LoopHeaders.count(BB)) { 877 DOUT << " Not threading from '" << PredBB->getNameStart() 878 << "' across loop header BB '" << BB->getNameStart() 879 << "' to dest BB '" << SuccBB->getNameStart() 880 << "' - it might create an irreducible loop!\n"; 881 return false; 882 } 883 884 // And finally, do it! 885 DOUT << " Threading edge from '" << PredBB->getNameStart() << "' to '" 886 << SuccBB->getNameStart() << "' with cost: " << JumpThreadCost 887 << ", across block:\n " 888 << *BB << "\n"; 889 890 // Jump Threading can not update SSA properties correctly if the values 891 // defined in the duplicated block are used outside of the block itself. For 892 // this reason, we spill all values that are used outside of BB to the stack. 893 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 894 if (!I->isUsedOutsideOfBlock(BB)) 895 continue; 896 897 // We found a use of I outside of BB. Create a new stack slot to 898 // break this inter-block usage pattern. 899 DemoteRegToStack(*I); 900 } 901 902 // We are going to have to map operands from the original BB block to the new 903 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 904 // account for entry from PredBB. 905 DenseMap<Instruction*, Value*> ValueMapping; 906 907 BasicBlock *NewBB = 908 BasicBlock::Create(BB->getName()+".thread", BB->getParent(), BB); 909 NewBB->moveAfter(PredBB); 910 911 BasicBlock::iterator BI = BB->begin(); 912 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 913 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 914 915 // Clone the non-phi instructions of BB into NewBB, keeping track of the 916 // mapping and using it to remap operands in the cloned instructions. 917 for (; !isa<TerminatorInst>(BI); ++BI) { 918 Instruction *New = BI->clone(); 919 New->setName(BI->getNameStart()); 920 NewBB->getInstList().push_back(New); 921 ValueMapping[BI] = New; 922 923 // Remap operands to patch up intra-block references. 924 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 925 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) 926 if (Value *Remapped = ValueMapping[Inst]) 927 New->setOperand(i, Remapped); 928 } 929 930 // We didn't copy the terminator from BB over to NewBB, because there is now 931 // an unconditional jump to SuccBB. Insert the unconditional jump. 932 BranchInst::Create(SuccBB, NewBB); 933 934 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 935 // PHI nodes for NewBB now. 936 for (BasicBlock::iterator PNI = SuccBB->begin(); isa<PHINode>(PNI); ++PNI) { 937 PHINode *PN = cast<PHINode>(PNI); 938 // Ok, we have a PHI node. Figure out what the incoming value was for the 939 // DestBlock. 940 Value *IV = PN->getIncomingValueForBlock(BB); 941 942 // Remap the value if necessary. 943 if (Instruction *Inst = dyn_cast<Instruction>(IV)) 944 if (Value *MappedIV = ValueMapping[Inst]) 945 IV = MappedIV; 946 PN->addIncoming(IV, NewBB); 947 } 948 949 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 950 // NewBB instead of BB. This eliminates predecessors from BB, which requires 951 // us to simplify any PHI nodes in BB. 952 TerminatorInst *PredTerm = PredBB->getTerminator(); 953 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 954 if (PredTerm->getSuccessor(i) == BB) { 955 BB->removePredecessor(PredBB); 956 PredTerm->setSuccessor(i, NewBB); 957 } 958 959 // At this point, the IR is fully up to date and consistent. Do a quick scan 960 // over the new instructions and zap any that are constants or dead. This 961 // frequently happens because of phi translation. 962 BI = NewBB->begin(); 963 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) { 964 Instruction *Inst = BI++; 965 if (Constant *C = ConstantFoldInstruction(Inst, TD)) { 966 Inst->replaceAllUsesWith(C); 967 Inst->eraseFromParent(); 968 continue; 969 } 970 971 RecursivelyDeleteTriviallyDeadInstructions(Inst); 972 } 973 974 // Threaded an edge! 975 ++NumThreads; 976 return true; 977} 978