JumpThreading.cpp revision d668839cb9b5db6865fd98e5e7dfccd64abf3e95
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/Analysis/LazyValueInfo.h" 21#include "llvm/Transforms/Utils/BasicBlockUtils.h" 22#include "llvm/Transforms/Utils/Local.h" 23#include "llvm/Transforms/Utils/SSAUpdater.h" 24#include "llvm/Target/TargetData.h" 25#include "llvm/ADT/DenseMap.h" 26#include "llvm/ADT/Statistic.h" 27#include "llvm/ADT/STLExtras.h" 28#include "llvm/ADT/SmallPtrSet.h" 29#include "llvm/ADT/SmallSet.h" 30#include "llvm/Support/CommandLine.h" 31#include "llvm/Support/Debug.h" 32#include "llvm/Support/ValueHandle.h" 33#include "llvm/Support/raw_ostream.h" 34using namespace llvm; 35 36STATISTIC(NumThreads, "Number of jumps threaded"); 37STATISTIC(NumFolds, "Number of terminators folded"); 38STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 39 40static cl::opt<unsigned> 41Threshold("jump-threading-threshold", 42 cl::desc("Max block size to duplicate for jump threading"), 43 cl::init(6), cl::Hidden); 44 45// Turn on use of LazyValueInfo. 46static cl::opt<bool> 47EnableLVI("enable-jump-threading-lvi", cl::ReallyHidden); 48 49 50 51namespace { 52 /// This pass performs 'jump threading', which looks at blocks that have 53 /// multiple predecessors and multiple successors. If one or more of the 54 /// predecessors of the block can be proven to always jump to one of the 55 /// successors, we forward the edge from the predecessor to the successor by 56 /// duplicating the contents of this block. 57 /// 58 /// An example of when this can occur is code like this: 59 /// 60 /// if () { ... 61 /// X = 4; 62 /// } 63 /// if (X < 3) { 64 /// 65 /// In this case, the unconditional branch at the end of the first if can be 66 /// revectored to the false side of the second if. 67 /// 68 class JumpThreading : public FunctionPass { 69 TargetData *TD; 70 LazyValueInfo *LVI; 71#ifdef NDEBUG 72 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 73#else 74 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 75#endif 76 public: 77 static char ID; // Pass identification 78 JumpThreading() : FunctionPass(&ID) {} 79 80 bool runOnFunction(Function &F); 81 82 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 83 if (EnableLVI) 84 AU.addRequired<LazyValueInfo>(); 85 } 86 87 void FindLoopHeaders(Function &F); 88 bool ProcessBlock(BasicBlock *BB); 89 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 90 BasicBlock *SuccBB); 91 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 92 const SmallVectorImpl<BasicBlock *> &PredBBs); 93 94 typedef SmallVectorImpl<std::pair<ConstantInt*, 95 BasicBlock*> > PredValueInfo; 96 97 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 98 PredValueInfo &Result); 99 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB); 100 101 102 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 103 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 104 105 bool ProcessBranchOnPHI(PHINode *PN); 106 bool ProcessBranchOnXOR(BinaryOperator *BO); 107 108 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 109 }; 110} 111 112char JumpThreading::ID = 0; 113static RegisterPass<JumpThreading> 114X("jump-threading", "Jump Threading"); 115 116// Public interface to the Jump Threading pass 117FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 118 119/// runOnFunction - Top level algorithm. 120/// 121bool JumpThreading::runOnFunction(Function &F) { 122 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 123 TD = getAnalysisIfAvailable<TargetData>(); 124 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0; 125 126 FindLoopHeaders(F); 127 128 bool Changed, EverChanged = false; 129 do { 130 Changed = false; 131 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 132 BasicBlock *BB = I; 133 // Thread all of the branches we can over this block. 134 while (ProcessBlock(BB)) 135 Changed = true; 136 137 ++I; 138 139 // If the block is trivially dead, zap it. This eliminates the successor 140 // edges which simplifies the CFG. 141 if (pred_begin(BB) == pred_end(BB) && 142 BB != &BB->getParent()->getEntryBlock()) { 143 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 144 << "' with terminator: " << *BB->getTerminator() << '\n'); 145 LoopHeaders.erase(BB); 146 DeleteDeadBlock(BB); 147 Changed = true; 148 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 149 // Can't thread an unconditional jump, but if the block is "almost 150 // empty", we can replace uses of it with uses of the successor and make 151 // this dead. 152 if (BI->isUnconditional() && 153 BB != &BB->getParent()->getEntryBlock()) { 154 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 155 // Ignore dbg intrinsics. 156 while (isa<DbgInfoIntrinsic>(BBI)) 157 ++BBI; 158 // If the terminator is the only non-phi instruction, try to nuke it. 159 if (BBI->isTerminator()) { 160 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 161 // block, we have to make sure it isn't in the LoopHeaders set. We 162 // reinsert afterward if needed. 163 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 164 BasicBlock *Succ = BI->getSuccessor(0); 165 166 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 167 Changed = true; 168 // If we deleted BB and BB was the header of a loop, then the 169 // successor is now the header of the loop. 170 BB = Succ; 171 } 172 173 if (ErasedFromLoopHeaders) 174 LoopHeaders.insert(BB); 175 } 176 } 177 } 178 } 179 EverChanged |= Changed; 180 } while (Changed); 181 182 LoopHeaders.clear(); 183 return EverChanged; 184} 185 186/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 187/// thread across it. 188static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 189 /// Ignore PHI nodes, these will be flattened when duplication happens. 190 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 191 192 // FIXME: THREADING will delete values that are just used to compute the 193 // branch, so they shouldn't count against the duplication cost. 194 195 196 // Sum up the cost of each instruction until we get to the terminator. Don't 197 // include the terminator because the copy won't include it. 198 unsigned Size = 0; 199 for (; !isa<TerminatorInst>(I); ++I) { 200 // Debugger intrinsics don't incur code size. 201 if (isa<DbgInfoIntrinsic>(I)) continue; 202 203 // If this is a pointer->pointer bitcast, it is free. 204 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType())) 205 continue; 206 207 // All other instructions count for at least one unit. 208 ++Size; 209 210 // Calls are more expensive. If they are non-intrinsic calls, we model them 211 // as having cost of 4. If they are a non-vector intrinsic, we model them 212 // as having cost of 2 total, and if they are a vector intrinsic, we model 213 // them as having cost 1. 214 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 215 if (!isa<IntrinsicInst>(CI)) 216 Size += 3; 217 else if (!isa<VectorType>(CI->getType())) 218 Size += 1; 219 } 220 } 221 222 // Threading through a switch statement is particularly profitable. If this 223 // block ends in a switch, decrease its cost to make it more likely to happen. 224 if (isa<SwitchInst>(I)) 225 Size = Size > 6 ? Size-6 : 0; 226 227 return Size; 228} 229 230/// FindLoopHeaders - We do not want jump threading to turn proper loop 231/// structures into irreducible loops. Doing this breaks up the loop nesting 232/// hierarchy and pessimizes later transformations. To prevent this from 233/// happening, we first have to find the loop headers. Here we approximate this 234/// by finding targets of backedges in the CFG. 235/// 236/// Note that there definitely are cases when we want to allow threading of 237/// edges across a loop header. For example, threading a jump from outside the 238/// loop (the preheader) to an exit block of the loop is definitely profitable. 239/// It is also almost always profitable to thread backedges from within the loop 240/// to exit blocks, and is often profitable to thread backedges to other blocks 241/// within the loop (forming a nested loop). This simple analysis is not rich 242/// enough to track all of these properties and keep it up-to-date as the CFG 243/// mutates, so we don't allow any of these transformations. 244/// 245void JumpThreading::FindLoopHeaders(Function &F) { 246 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 247 FindFunctionBackedges(F, Edges); 248 249 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 250 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 251} 252 253/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 254/// if we can infer that the value is a known ConstantInt in any of our 255/// predecessors. If so, return the known list of value and pred BB in the 256/// result vector. If a value is known to be undef, it is returned as null. 257/// 258/// This returns true if there were any known values. 259/// 260bool JumpThreading:: 261ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 262 // If V is a constantint, then it is known in all predecessors. 263 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 264 ConstantInt *CI = dyn_cast<ConstantInt>(V); 265 266 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 267 Result.push_back(std::make_pair(CI, *PI)); 268 return true; 269 } 270 271 // If V is a non-instruction value, or an instruction in a different block, 272 // then it can't be derived from a PHI. 273 Instruction *I = dyn_cast<Instruction>(V); 274 if (I == 0 || I->getParent() != BB) { 275 276 // Okay, if this is a live-in value, see if it has a known value at the end 277 // of any of our predecessors. 278 // 279 // FIXME: This should be an edge property, not a block end property. 280 /// TODO: Per PR2563, we could infer value range information about a 281 /// predecessor based on its terminator. 282 // 283 if (LVI) { 284 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 285 // "I" is a non-local compare-with-a-constant instruction. This would be 286 // able to handle value inequalities better, for example if the compare is 287 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 288 // Perhaps getConstantOnEdge should be smart enough to do this? 289 290 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 291 // If the value is known by LazyValueInfo to be a constant in a 292 // predecessor, use that information to try to thread this block. 293 Constant *PredCst = LVI->getConstantOnEdge(V, *PI, BB); 294 if (PredCst == 0 || 295 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst))) 296 continue; 297 298 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), *PI)); 299 } 300 301 return !Result.empty(); 302 } 303 304 return false; 305 } 306 307 /// If I is a PHI node, then we know the incoming values for any constants. 308 if (PHINode *PN = dyn_cast<PHINode>(I)) { 309 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 310 Value *InVal = PN->getIncomingValue(i); 311 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 312 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 313 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 314 } 315 } 316 return !Result.empty(); 317 } 318 319 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 320 321 // Handle some boolean conditions. 322 if (I->getType()->getPrimitiveSizeInBits() == 1) { 323 // X | true -> true 324 // X & false -> false 325 if (I->getOpcode() == Instruction::Or || 326 I->getOpcode() == Instruction::And) { 327 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 328 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 329 330 if (LHSVals.empty() && RHSVals.empty()) 331 return false; 332 333 ConstantInt *InterestingVal; 334 if (I->getOpcode() == Instruction::Or) 335 InterestingVal = ConstantInt::getTrue(I->getContext()); 336 else 337 InterestingVal = ConstantInt::getFalse(I->getContext()); 338 339 // Scan for the sentinel. 340 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 341 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) 342 Result.push_back(LHSVals[i]); 343 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 344 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) 345 Result.push_back(RHSVals[i]); 346 return !Result.empty(); 347 } 348 349 // Handle the NOT form of XOR. 350 if (I->getOpcode() == Instruction::Xor && 351 isa<ConstantInt>(I->getOperand(1)) && 352 cast<ConstantInt>(I->getOperand(1))->isOne()) { 353 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result); 354 if (Result.empty()) 355 return false; 356 357 // Invert the known values. 358 for (unsigned i = 0, e = Result.size(); i != e; ++i) 359 if (Result[i].first) 360 Result[i].first = 361 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first)); 362 return true; 363 } 364 } 365 366 // Handle compare with phi operand, where the PHI is defined in this block. 367 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 368 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 369 if (PN && PN->getParent() == BB) { 370 // We can do this simplification if any comparisons fold to true or false. 371 // See if any do. 372 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 373 BasicBlock *PredBB = PN->getIncomingBlock(i); 374 Value *LHS = PN->getIncomingValue(i); 375 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 376 377 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD); 378 if (Res == 0) { 379 if (!LVI || !isa<Constant>(RHS)) 380 continue; 381 382 LazyValueInfo::Tristate 383 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 384 cast<Constant>(RHS), PredBB, BB); 385 if (ResT == LazyValueInfo::Unknown) 386 continue; 387 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 388 } 389 390 if (isa<UndefValue>(Res)) 391 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 392 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 393 Result.push_back(std::make_pair(CI, PredBB)); 394 } 395 396 return !Result.empty(); 397 } 398 399 400 // If comparing a live-in value against a constant, see if we know the 401 // live-in value on any predecessors. 402 if (LVI && isa<Constant>(Cmp->getOperand(1)) && 403 Cmp->getType()->isInteger() && // Not vector compare. 404 (!isa<Instruction>(Cmp->getOperand(0)) || 405 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB)) { 406 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 407 408 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 409 // If the value is known by LazyValueInfo to be a constant in a 410 // predecessor, use that information to try to thread this block. 411 LazyValueInfo::Tristate 412 Res = LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 413 RHSCst, *PI, BB); 414 if (Res == LazyValueInfo::Unknown) 415 continue; 416 417 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 418 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), *PI)); 419 } 420 421 return !Result.empty(); 422 } 423 } 424 return false; 425} 426 427 428 429/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 430/// in an undefined jump, decide which block is best to revector to. 431/// 432/// Since we can pick an arbitrary destination, we pick the successor with the 433/// fewest predecessors. This should reduce the in-degree of the others. 434/// 435static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 436 TerminatorInst *BBTerm = BB->getTerminator(); 437 unsigned MinSucc = 0; 438 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 439 // Compute the successor with the minimum number of predecessors. 440 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 441 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 442 TestBB = BBTerm->getSuccessor(i); 443 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 444 if (NumPreds < MinNumPreds) 445 MinSucc = i; 446 } 447 448 return MinSucc; 449} 450 451/// ProcessBlock - If there are any predecessors whose control can be threaded 452/// through to a successor, transform them now. 453bool JumpThreading::ProcessBlock(BasicBlock *BB) { 454 // If the block is trivially dead, just return and let the caller nuke it. 455 // This simplifies other transformations. 456 if (pred_begin(BB) == pred_end(BB) && 457 BB != &BB->getParent()->getEntryBlock()) 458 return false; 459 460 // If this block has a single predecessor, and if that pred has a single 461 // successor, merge the blocks. This encourages recursive jump threading 462 // because now the condition in this block can be threaded through 463 // predecessors of our predecessor block. 464 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 465 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 466 SinglePred != BB) { 467 // If SinglePred was a loop header, BB becomes one. 468 if (LoopHeaders.erase(SinglePred)) 469 LoopHeaders.insert(BB); 470 471 // Remember if SinglePred was the entry block of the function. If so, we 472 // will need to move BB back to the entry position. 473 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 474 MergeBasicBlockIntoOnlyPred(BB); 475 476 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 477 BB->moveBefore(&BB->getParent()->getEntryBlock()); 478 return true; 479 } 480 } 481 482 // Look to see if the terminator is a branch of switch, if not we can't thread 483 // it. 484 Value *Condition; 485 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 486 // Can't thread an unconditional jump. 487 if (BI->isUnconditional()) return false; 488 Condition = BI->getCondition(); 489 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 490 Condition = SI->getCondition(); 491 else 492 return false; // Must be an invoke. 493 494 // If the terminator of this block is branching on a constant, simplify the 495 // terminator to an unconditional branch. This can occur due to threading in 496 // other blocks. 497 if (isa<ConstantInt>(Condition)) { 498 DEBUG(dbgs() << " In block '" << BB->getName() 499 << "' folding terminator: " << *BB->getTerminator() << '\n'); 500 ++NumFolds; 501 ConstantFoldTerminator(BB); 502 return true; 503 } 504 505 // If the terminator is branching on an undef, we can pick any of the 506 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 507 if (isa<UndefValue>(Condition)) { 508 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 509 510 // Fold the branch/switch. 511 TerminatorInst *BBTerm = BB->getTerminator(); 512 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 513 if (i == BestSucc) continue; 514 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD); 515 } 516 517 DEBUG(dbgs() << " In block '" << BB->getName() 518 << "' folding undef terminator: " << *BBTerm << '\n'); 519 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 520 BBTerm->eraseFromParent(); 521 return true; 522 } 523 524 Instruction *CondInst = dyn_cast<Instruction>(Condition); 525 526 // If the condition is an instruction defined in another block, see if a 527 // predecessor has the same condition: 528 // br COND, BBX, BBY 529 // BBX: 530 // br COND, BBZ, BBW 531 if (!LVI && 532 !Condition->hasOneUse() && // Multiple uses. 533 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 534 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 535 if (isa<BranchInst>(BB->getTerminator())) { 536 for (; PI != E; ++PI) 537 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 538 if (PBI->isConditional() && PBI->getCondition() == Condition && 539 ProcessBranchOnDuplicateCond(*PI, BB)) 540 return true; 541 } else { 542 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 543 for (; PI != E; ++PI) 544 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator())) 545 if (PSI->getCondition() == Condition && 546 ProcessSwitchOnDuplicateCond(*PI, BB)) 547 return true; 548 } 549 } 550 551 // All the rest of our checks depend on the condition being an instruction. 552 if (CondInst == 0) { 553 // FIXME: Unify this with code below. 554 if (LVI && ProcessThreadableEdges(Condition, BB)) 555 return true; 556 return false; 557 } 558 559 560 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 561 if (!LVI && 562 (!isa<PHINode>(CondCmp->getOperand(0)) || 563 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) { 564 // If we have a comparison, loop over the predecessors to see if there is 565 // a condition with a lexically identical value. 566 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 567 for (; PI != E; ++PI) 568 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 569 if (PBI->isConditional() && *PI != BB) { 570 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 571 if (CI->getOperand(0) == CondCmp->getOperand(0) && 572 CI->getOperand(1) == CondCmp->getOperand(1) && 573 CI->getPredicate() == CondCmp->getPredicate()) { 574 // TODO: Could handle things like (x != 4) --> (x == 17) 575 if (ProcessBranchOnDuplicateCond(*PI, BB)) 576 return true; 577 } 578 } 579 } 580 } 581 } 582 583 // Check for some cases that are worth simplifying. Right now we want to look 584 // for loads that are used by a switch or by the condition for the branch. If 585 // we see one, check to see if it's partially redundant. If so, insert a PHI 586 // which can then be used to thread the values. 587 // 588 Value *SimplifyValue = CondInst; 589 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 590 if (isa<Constant>(CondCmp->getOperand(1))) 591 SimplifyValue = CondCmp->getOperand(0); 592 593 // TODO: There are other places where load PRE would be profitable, such as 594 // more complex comparisons. 595 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 596 if (SimplifyPartiallyRedundantLoad(LI)) 597 return true; 598 599 600 // Handle a variety of cases where we are branching on something derived from 601 // a PHI node in the current block. If we can prove that any predecessors 602 // compute a predictable value based on a PHI node, thread those predecessors. 603 // 604 if (ProcessThreadableEdges(CondInst, BB)) 605 return true; 606 607 // If this is an otherwise-unfoldable branch on a phi node in the current 608 // block, see if we can simplify. 609 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 610 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 611 return ProcessBranchOnPHI(PN); 612 613 614 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 615 if (CondInst->getOpcode() == Instruction::Xor && 616 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 617 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 618 619 620 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 621 // "(X == 4)", thread through this block. 622 623 return false; 624} 625 626/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 627/// block that jump on exactly the same condition. This means that we almost 628/// always know the direction of the edge in the DESTBB: 629/// PREDBB: 630/// br COND, DESTBB, BBY 631/// DESTBB: 632/// br COND, BBZ, BBW 633/// 634/// If DESTBB has multiple predecessors, we can't just constant fold the branch 635/// in DESTBB, we have to thread over it. 636bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 637 BasicBlock *BB) { 638 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 639 640 // If both successors of PredBB go to DESTBB, we don't know anything. We can 641 // fold the branch to an unconditional one, which allows other recursive 642 // simplifications. 643 bool BranchDir; 644 if (PredBI->getSuccessor(1) != BB) 645 BranchDir = true; 646 else if (PredBI->getSuccessor(0) != BB) 647 BranchDir = false; 648 else { 649 DEBUG(dbgs() << " In block '" << PredBB->getName() 650 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 651 ++NumFolds; 652 ConstantFoldTerminator(PredBB); 653 return true; 654 } 655 656 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 657 658 // If the dest block has one predecessor, just fix the branch condition to a 659 // constant and fold it. 660 if (BB->getSinglePredecessor()) { 661 DEBUG(dbgs() << " In block '" << BB->getName() 662 << "' folding condition to '" << BranchDir << "': " 663 << *BB->getTerminator() << '\n'); 664 ++NumFolds; 665 Value *OldCond = DestBI->getCondition(); 666 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 667 BranchDir)); 668 ConstantFoldTerminator(BB); 669 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 670 return true; 671 } 672 673 674 // Next, figure out which successor we are threading to. 675 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 676 677 SmallVector<BasicBlock*, 2> Preds; 678 Preds.push_back(PredBB); 679 680 // Ok, try to thread it! 681 return ThreadEdge(BB, Preds, SuccBB); 682} 683 684/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 685/// block that switch on exactly the same condition. This means that we almost 686/// always know the direction of the edge in the DESTBB: 687/// PREDBB: 688/// switch COND [... DESTBB, BBY ... ] 689/// DESTBB: 690/// switch COND [... BBZ, BBW ] 691/// 692/// Optimizing switches like this is very important, because simplifycfg builds 693/// switches out of repeated 'if' conditions. 694bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 695 BasicBlock *DestBB) { 696 // Can't thread edge to self. 697 if (PredBB == DestBB) 698 return false; 699 700 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 701 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 702 703 // There are a variety of optimizations that we can potentially do on these 704 // blocks: we order them from most to least preferable. 705 706 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 707 // directly to their destination. This does not introduce *any* code size 708 // growth. Skip debug info first. 709 BasicBlock::iterator BBI = DestBB->begin(); 710 while (isa<DbgInfoIntrinsic>(BBI)) 711 BBI++; 712 713 // FIXME: Thread if it just contains a PHI. 714 if (isa<SwitchInst>(BBI)) { 715 bool MadeChange = false; 716 // Ignore the default edge for now. 717 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 718 ConstantInt *DestVal = DestSI->getCaseValue(i); 719 BasicBlock *DestSucc = DestSI->getSuccessor(i); 720 721 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 722 // PredSI has an explicit case for it. If so, forward. If it is covered 723 // by the default case, we can't update PredSI. 724 unsigned PredCase = PredSI->findCaseValue(DestVal); 725 if (PredCase == 0) continue; 726 727 // If PredSI doesn't go to DestBB on this value, then it won't reach the 728 // case on this condition. 729 if (PredSI->getSuccessor(PredCase) != DestBB && 730 DestSI->getSuccessor(i) != DestBB) 731 continue; 732 733 // Do not forward this if it already goes to this destination, this would 734 // be an infinite loop. 735 if (PredSI->getSuccessor(PredCase) == DestSucc) 736 continue; 737 738 // Otherwise, we're safe to make the change. Make sure that the edge from 739 // DestSI to DestSucc is not critical and has no PHI nodes. 740 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 741 DEBUG(dbgs() << "THROUGH: " << *DestSI); 742 743 // If the destination has PHI nodes, just split the edge for updating 744 // simplicity. 745 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 746 SplitCriticalEdge(DestSI, i, this); 747 DestSucc = DestSI->getSuccessor(i); 748 } 749 FoldSingleEntryPHINodes(DestSucc); 750 PredSI->setSuccessor(PredCase, DestSucc); 751 MadeChange = true; 752 } 753 754 if (MadeChange) 755 return true; 756 } 757 758 return false; 759} 760 761 762/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 763/// load instruction, eliminate it by replacing it with a PHI node. This is an 764/// important optimization that encourages jump threading, and needs to be run 765/// interlaced with other jump threading tasks. 766bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 767 // Don't hack volatile loads. 768 if (LI->isVolatile()) return false; 769 770 // If the load is defined in a block with exactly one predecessor, it can't be 771 // partially redundant. 772 BasicBlock *LoadBB = LI->getParent(); 773 if (LoadBB->getSinglePredecessor()) 774 return false; 775 776 Value *LoadedPtr = LI->getOperand(0); 777 778 // If the loaded operand is defined in the LoadBB, it can't be available. 779 // TODO: Could do simple PHI translation, that would be fun :) 780 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 781 if (PtrOp->getParent() == LoadBB) 782 return false; 783 784 // Scan a few instructions up from the load, to see if it is obviously live at 785 // the entry to its block. 786 BasicBlock::iterator BBIt = LI; 787 788 if (Value *AvailableVal = 789 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 790 // If the value if the load is locally available within the block, just use 791 // it. This frequently occurs for reg2mem'd allocas. 792 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 793 794 // If the returned value is the load itself, replace with an undef. This can 795 // only happen in dead loops. 796 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 797 LI->replaceAllUsesWith(AvailableVal); 798 LI->eraseFromParent(); 799 return true; 800 } 801 802 // Otherwise, if we scanned the whole block and got to the top of the block, 803 // we know the block is locally transparent to the load. If not, something 804 // might clobber its value. 805 if (BBIt != LoadBB->begin()) 806 return false; 807 808 809 SmallPtrSet<BasicBlock*, 8> PredsScanned; 810 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 811 AvailablePredsTy AvailablePreds; 812 BasicBlock *OneUnavailablePred = 0; 813 814 // If we got here, the loaded value is transparent through to the start of the 815 // block. Check to see if it is available in any of the predecessor blocks. 816 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 817 PI != PE; ++PI) { 818 BasicBlock *PredBB = *PI; 819 820 // If we already scanned this predecessor, skip it. 821 if (!PredsScanned.insert(PredBB)) 822 continue; 823 824 // Scan the predecessor to see if the value is available in the pred. 825 BBIt = PredBB->end(); 826 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 827 if (!PredAvailable) { 828 OneUnavailablePred = PredBB; 829 continue; 830 } 831 832 // If so, this load is partially redundant. Remember this info so that we 833 // can create a PHI node. 834 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 835 } 836 837 // If the loaded value isn't available in any predecessor, it isn't partially 838 // redundant. 839 if (AvailablePreds.empty()) return false; 840 841 // Okay, the loaded value is available in at least one (and maybe all!) 842 // predecessors. If the value is unavailable in more than one unique 843 // predecessor, we want to insert a merge block for those common predecessors. 844 // This ensures that we only have to insert one reload, thus not increasing 845 // code size. 846 BasicBlock *UnavailablePred = 0; 847 848 // If there is exactly one predecessor where the value is unavailable, the 849 // already computed 'OneUnavailablePred' block is it. If it ends in an 850 // unconditional branch, we know that it isn't a critical edge. 851 if (PredsScanned.size() == AvailablePreds.size()+1 && 852 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 853 UnavailablePred = OneUnavailablePred; 854 } else if (PredsScanned.size() != AvailablePreds.size()) { 855 // Otherwise, we had multiple unavailable predecessors or we had a critical 856 // edge from the one. 857 SmallVector<BasicBlock*, 8> PredsToSplit; 858 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 859 860 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 861 AvailablePredSet.insert(AvailablePreds[i].first); 862 863 // Add all the unavailable predecessors to the PredsToSplit list. 864 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 865 PI != PE; ++PI) 866 if (!AvailablePredSet.count(*PI)) 867 PredsToSplit.push_back(*PI); 868 869 // Split them out to their own block. 870 UnavailablePred = 871 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 872 "thread-pre-split", this); 873 } 874 875 // If the value isn't available in all predecessors, then there will be 876 // exactly one where it isn't available. Insert a load on that edge and add 877 // it to the AvailablePreds list. 878 if (UnavailablePred) { 879 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 880 "Can't handle critical edge here!"); 881 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 882 LI->getAlignment(), 883 UnavailablePred->getTerminator()); 884 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 885 } 886 887 // Now we know that each predecessor of this block has a value in 888 // AvailablePreds, sort them for efficient access as we're walking the preds. 889 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 890 891 // Create a PHI node at the start of the block for the PRE'd load value. 892 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 893 PN->takeName(LI); 894 895 // Insert new entries into the PHI for each predecessor. A single block may 896 // have multiple entries here. 897 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 898 ++PI) { 899 AvailablePredsTy::iterator I = 900 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 901 std::make_pair(*PI, (Value*)0)); 902 903 assert(I != AvailablePreds.end() && I->first == *PI && 904 "Didn't find entry for predecessor!"); 905 906 PN->addIncoming(I->second, I->first); 907 } 908 909 //cerr << "PRE: " << *LI << *PN << "\n"; 910 911 LI->replaceAllUsesWith(PN); 912 LI->eraseFromParent(); 913 914 return true; 915} 916 917/// FindMostPopularDest - The specified list contains multiple possible 918/// threadable destinations. Pick the one that occurs the most frequently in 919/// the list. 920static BasicBlock * 921FindMostPopularDest(BasicBlock *BB, 922 const SmallVectorImpl<std::pair<BasicBlock*, 923 BasicBlock*> > &PredToDestList) { 924 assert(!PredToDestList.empty()); 925 926 // Determine popularity. If there are multiple possible destinations, we 927 // explicitly choose to ignore 'undef' destinations. We prefer to thread 928 // blocks with known and real destinations to threading undef. We'll handle 929 // them later if interesting. 930 DenseMap<BasicBlock*, unsigned> DestPopularity; 931 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 932 if (PredToDestList[i].second) 933 DestPopularity[PredToDestList[i].second]++; 934 935 // Find the most popular dest. 936 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 937 BasicBlock *MostPopularDest = DPI->first; 938 unsigned Popularity = DPI->second; 939 SmallVector<BasicBlock*, 4> SamePopularity; 940 941 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 942 // If the popularity of this entry isn't higher than the popularity we've 943 // seen so far, ignore it. 944 if (DPI->second < Popularity) 945 ; // ignore. 946 else if (DPI->second == Popularity) { 947 // If it is the same as what we've seen so far, keep track of it. 948 SamePopularity.push_back(DPI->first); 949 } else { 950 // If it is more popular, remember it. 951 SamePopularity.clear(); 952 MostPopularDest = DPI->first; 953 Popularity = DPI->second; 954 } 955 } 956 957 // Okay, now we know the most popular destination. If there is more than 958 // destination, we need to determine one. This is arbitrary, but we need 959 // to make a deterministic decision. Pick the first one that appears in the 960 // successor list. 961 if (!SamePopularity.empty()) { 962 SamePopularity.push_back(MostPopularDest); 963 TerminatorInst *TI = BB->getTerminator(); 964 for (unsigned i = 0; ; ++i) { 965 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 966 967 if (std::find(SamePopularity.begin(), SamePopularity.end(), 968 TI->getSuccessor(i)) == SamePopularity.end()) 969 continue; 970 971 MostPopularDest = TI->getSuccessor(i); 972 break; 973 } 974 } 975 976 // Okay, we have finally picked the most popular destination. 977 return MostPopularDest; 978} 979 980bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) { 981 // If threading this would thread across a loop header, don't even try to 982 // thread the edge. 983 if (LoopHeaders.count(BB)) 984 return false; 985 986 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 987 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues)) 988 return false; 989 assert(!PredValues.empty() && 990 "ComputeValueKnownInPredecessors returned true with no values"); 991 992 DEBUG(dbgs() << "IN BB: " << *BB; 993 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 994 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "; 995 if (PredValues[i].first) 996 dbgs() << *PredValues[i].first; 997 else 998 dbgs() << "UNDEF"; 999 dbgs() << " for pred '" << PredValues[i].second->getName() 1000 << "'.\n"; 1001 }); 1002 1003 // Decide what we want to thread through. Convert our list of known values to 1004 // a list of known destinations for each pred. This also discards duplicate 1005 // predecessors and keeps track of the undefined inputs (which are represented 1006 // as a null dest in the PredToDestList). 1007 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1008 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1009 1010 BasicBlock *OnlyDest = 0; 1011 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1012 1013 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1014 BasicBlock *Pred = PredValues[i].second; 1015 if (!SeenPreds.insert(Pred)) 1016 continue; // Duplicate predecessor entry. 1017 1018 // If the predecessor ends with an indirect goto, we can't change its 1019 // destination. 1020 if (isa<IndirectBrInst>(Pred->getTerminator())) 1021 continue; 1022 1023 ConstantInt *Val = PredValues[i].first; 1024 1025 BasicBlock *DestBB; 1026 if (Val == 0) // Undef. 1027 DestBB = 0; 1028 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1029 DestBB = BI->getSuccessor(Val->isZero()); 1030 else { 1031 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 1032 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 1033 } 1034 1035 // If we have exactly one destination, remember it for efficiency below. 1036 if (i == 0) 1037 OnlyDest = DestBB; 1038 else if (OnlyDest != DestBB) 1039 OnlyDest = MultipleDestSentinel; 1040 1041 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1042 } 1043 1044 // If all edges were unthreadable, we fail. 1045 if (PredToDestList.empty()) 1046 return false; 1047 1048 // Determine which is the most common successor. If we have many inputs and 1049 // this block is a switch, we want to start by threading the batch that goes 1050 // to the most popular destination first. If we only know about one 1051 // threadable destination (the common case) we can avoid this. 1052 BasicBlock *MostPopularDest = OnlyDest; 1053 1054 if (MostPopularDest == MultipleDestSentinel) 1055 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1056 1057 // Now that we know what the most popular destination is, factor all 1058 // predecessors that will jump to it into a single predecessor. 1059 SmallVector<BasicBlock*, 16> PredsToFactor; 1060 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1061 if (PredToDestList[i].second == MostPopularDest) { 1062 BasicBlock *Pred = PredToDestList[i].first; 1063 1064 // This predecessor may be a switch or something else that has multiple 1065 // edges to the block. Factor each of these edges by listing them 1066 // according to # occurrences in PredsToFactor. 1067 TerminatorInst *PredTI = Pred->getTerminator(); 1068 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1069 if (PredTI->getSuccessor(i) == BB) 1070 PredsToFactor.push_back(Pred); 1071 } 1072 1073 // If the threadable edges are branching on an undefined value, we get to pick 1074 // the destination that these predecessors should get to. 1075 if (MostPopularDest == 0) 1076 MostPopularDest = BB->getTerminator()-> 1077 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1078 1079 // Ok, try to thread it! 1080 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1081} 1082 1083/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1084/// a PHI node in the current block. See if there are any simplifications we 1085/// can do based on inputs to the phi node. 1086/// 1087bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1088 BasicBlock *BB = PN->getParent(); 1089 1090 // TODO: We could make use of this to do it once for blocks with common PHI 1091 // values. 1092 SmallVector<BasicBlock*, 1> PredBBs; 1093 PredBBs.resize(1); 1094 1095 // If any of the predecessor blocks end in an unconditional branch, we can 1096 // *duplicate* the conditional branch into that block in order to further 1097 // encourage jump threading and to eliminate cases where we have branch on a 1098 // phi of an icmp (branch on icmp is much better). 1099 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1100 BasicBlock *PredBB = PN->getIncomingBlock(i); 1101 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1102 if (PredBr->isUnconditional()) { 1103 PredBBs[0] = PredBB; 1104 // Try to duplicate BB into PredBB. 1105 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1106 return true; 1107 } 1108 } 1109 1110 return false; 1111} 1112 1113/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1114/// a xor instruction in the current block. See if there are any 1115/// simplifications we can do based on inputs to the xor. 1116/// 1117bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1118 BasicBlock *BB = BO->getParent(); 1119 1120 // If either the LHS or RHS of the xor is a constant, don't do this 1121 // optimization. 1122 if (isa<ConstantInt>(BO->getOperand(0)) || 1123 isa<ConstantInt>(BO->getOperand(1))) 1124 return false; 1125 1126 // If the first instruction in BB isn't a phi, we won't be able to infer 1127 // anything special about any particular predecessor. 1128 if (!isa<PHINode>(BB->front())) 1129 return false; 1130 1131 // If we have a xor as the branch input to this block, and we know that the 1132 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1133 // the condition into the predecessor and fix that value to true, saving some 1134 // logical ops on that path and encouraging other paths to simplify. 1135 // 1136 // This copies something like this: 1137 // 1138 // BB: 1139 // %X = phi i1 [1], [%X'] 1140 // %Y = icmp eq i32 %A, %B 1141 // %Z = xor i1 %X, %Y 1142 // br i1 %Z, ... 1143 // 1144 // Into: 1145 // BB': 1146 // %Y = icmp ne i32 %A, %B 1147 // br i1 %Z, ... 1148 1149 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues; 1150 bool isLHS = true; 1151 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) { 1152 assert(XorOpValues.empty()); 1153 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues)) 1154 return false; 1155 isLHS = false; 1156 } 1157 1158 assert(!XorOpValues.empty() && 1159 "ComputeValueKnownInPredecessors returned true with no values"); 1160 1161 // Scan the information to see which is most popular: true or false. The 1162 // predecessors can be of the set true, false, or undef. 1163 unsigned NumTrue = 0, NumFalse = 0; 1164 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1165 if (!XorOpValues[i].first) continue; // Ignore undefs for the count. 1166 if (XorOpValues[i].first->isZero()) 1167 ++NumFalse; 1168 else 1169 ++NumTrue; 1170 } 1171 1172 // Determine which value to split on, true, false, or undef if neither. 1173 ConstantInt *SplitVal = 0; 1174 if (NumTrue > NumFalse) 1175 SplitVal = ConstantInt::getTrue(BB->getContext()); 1176 else if (NumTrue != 0 || NumFalse != 0) 1177 SplitVal = ConstantInt::getFalse(BB->getContext()); 1178 1179 // Collect all of the blocks that this can be folded into so that we can 1180 // factor this once and clone it once. 1181 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1182 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1183 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue; 1184 1185 BlocksToFoldInto.push_back(XorOpValues[i].second); 1186 } 1187 1188 // If we inferred a value for all of the predecessors, then duplication won't 1189 // help us. However, we can just replace the LHS or RHS with the constant. 1190 if (BlocksToFoldInto.size() == 1191 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1192 if (SplitVal == 0) { 1193 // If all preds provide undef, just nuke the xor, because it is undef too. 1194 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1195 BO->eraseFromParent(); 1196 } else if (SplitVal->isZero()) { 1197 // If all preds provide 0, replace the xor with the other input. 1198 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1199 BO->eraseFromParent(); 1200 } else { 1201 // If all preds provide 1, set the computed value to 1. 1202 BO->setOperand(!isLHS, SplitVal); 1203 } 1204 1205 return true; 1206 } 1207 1208 // Try to duplicate BB into PredBB. 1209 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1210} 1211 1212 1213/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1214/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1215/// NewPred using the entries from OldPred (suitably mapped). 1216static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1217 BasicBlock *OldPred, 1218 BasicBlock *NewPred, 1219 DenseMap<Instruction*, Value*> &ValueMap) { 1220 for (BasicBlock::iterator PNI = PHIBB->begin(); 1221 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1222 // Ok, we have a PHI node. Figure out what the incoming value was for the 1223 // DestBlock. 1224 Value *IV = PN->getIncomingValueForBlock(OldPred); 1225 1226 // Remap the value if necessary. 1227 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1228 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1229 if (I != ValueMap.end()) 1230 IV = I->second; 1231 } 1232 1233 PN->addIncoming(IV, NewPred); 1234 } 1235} 1236 1237/// ThreadEdge - We have decided that it is safe and profitable to factor the 1238/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1239/// across BB. Transform the IR to reflect this change. 1240bool JumpThreading::ThreadEdge(BasicBlock *BB, 1241 const SmallVectorImpl<BasicBlock*> &PredBBs, 1242 BasicBlock *SuccBB) { 1243 // If threading to the same block as we come from, we would infinite loop. 1244 if (SuccBB == BB) { 1245 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1246 << "' - would thread to self!\n"); 1247 return false; 1248 } 1249 1250 // If threading this would thread across a loop header, don't thread the edge. 1251 // See the comments above FindLoopHeaders for justifications and caveats. 1252 if (LoopHeaders.count(BB)) { 1253 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1254 << "' to dest BB '" << SuccBB->getName() 1255 << "' - it might create an irreducible loop!\n"); 1256 return false; 1257 } 1258 1259 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1260 if (JumpThreadCost > Threshold) { 1261 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1262 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1263 return false; 1264 } 1265 1266 // And finally, do it! Start by factoring the predecessors is needed. 1267 BasicBlock *PredBB; 1268 if (PredBBs.size() == 1) 1269 PredBB = PredBBs[0]; 1270 else { 1271 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1272 << " common predecessors.\n"); 1273 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1274 ".thr_comm", this); 1275 } 1276 1277 // And finally, do it! 1278 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1279 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1280 << ", across block:\n " 1281 << *BB << "\n"); 1282 1283 // We are going to have to map operands from the original BB block to the new 1284 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1285 // account for entry from PredBB. 1286 DenseMap<Instruction*, Value*> ValueMapping; 1287 1288 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1289 BB->getName()+".thread", 1290 BB->getParent(), BB); 1291 NewBB->moveAfter(PredBB); 1292 1293 BasicBlock::iterator BI = BB->begin(); 1294 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1295 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1296 1297 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1298 // mapping and using it to remap operands in the cloned instructions. 1299 for (; !isa<TerminatorInst>(BI); ++BI) { 1300 Instruction *New = BI->clone(); 1301 New->setName(BI->getName()); 1302 NewBB->getInstList().push_back(New); 1303 ValueMapping[BI] = New; 1304 1305 // Remap operands to patch up intra-block references. 1306 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1307 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1308 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1309 if (I != ValueMapping.end()) 1310 New->setOperand(i, I->second); 1311 } 1312 } 1313 1314 // We didn't copy the terminator from BB over to NewBB, because there is now 1315 // an unconditional jump to SuccBB. Insert the unconditional jump. 1316 BranchInst::Create(SuccBB, NewBB); 1317 1318 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1319 // PHI nodes for NewBB now. 1320 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1321 1322 // If there were values defined in BB that are used outside the block, then we 1323 // now have to update all uses of the value to use either the original value, 1324 // the cloned value, or some PHI derived value. This can require arbitrary 1325 // PHI insertion, of which we are prepared to do, clean these up now. 1326 SSAUpdater SSAUpdate; 1327 SmallVector<Use*, 16> UsesToRename; 1328 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1329 // Scan all uses of this instruction to see if it is used outside of its 1330 // block, and if so, record them in UsesToRename. 1331 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1332 ++UI) { 1333 Instruction *User = cast<Instruction>(*UI); 1334 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1335 if (UserPN->getIncomingBlock(UI) == BB) 1336 continue; 1337 } else if (User->getParent() == BB) 1338 continue; 1339 1340 UsesToRename.push_back(&UI.getUse()); 1341 } 1342 1343 // If there are no uses outside the block, we're done with this instruction. 1344 if (UsesToRename.empty()) 1345 continue; 1346 1347 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1348 1349 // We found a use of I outside of BB. Rename all uses of I that are outside 1350 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1351 // with the two values we know. 1352 SSAUpdate.Initialize(I); 1353 SSAUpdate.AddAvailableValue(BB, I); 1354 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1355 1356 while (!UsesToRename.empty()) 1357 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1358 DEBUG(dbgs() << "\n"); 1359 } 1360 1361 1362 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1363 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1364 // us to simplify any PHI nodes in BB. 1365 TerminatorInst *PredTerm = PredBB->getTerminator(); 1366 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1367 if (PredTerm->getSuccessor(i) == BB) { 1368 RemovePredecessorAndSimplify(BB, PredBB, TD); 1369 PredTerm->setSuccessor(i, NewBB); 1370 } 1371 1372 // At this point, the IR is fully up to date and consistent. Do a quick scan 1373 // over the new instructions and zap any that are constants or dead. This 1374 // frequently happens because of phi translation. 1375 SimplifyInstructionsInBlock(NewBB, TD); 1376 1377 // Threaded an edge! 1378 ++NumThreads; 1379 return true; 1380} 1381 1382/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1383/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1384/// If we can duplicate the contents of BB up into PredBB do so now, this 1385/// improves the odds that the branch will be on an analyzable instruction like 1386/// a compare. 1387bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1388 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1389 assert(!PredBBs.empty() && "Can't handle an empty set"); 1390 1391 // If BB is a loop header, then duplicating this block outside the loop would 1392 // cause us to transform this into an irreducible loop, don't do this. 1393 // See the comments above FindLoopHeaders for justifications and caveats. 1394 if (LoopHeaders.count(BB)) { 1395 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1396 << "' into predecessor block '" << PredBBs[0]->getName() 1397 << "' - it might create an irreducible loop!\n"); 1398 return false; 1399 } 1400 1401 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1402 if (DuplicationCost > Threshold) { 1403 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1404 << "' - Cost is too high: " << DuplicationCost << "\n"); 1405 return false; 1406 } 1407 1408 // And finally, do it! Start by factoring the predecessors is needed. 1409 BasicBlock *PredBB; 1410 if (PredBBs.size() == 1) 1411 PredBB = PredBBs[0]; 1412 else { 1413 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1414 << " common predecessors.\n"); 1415 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1416 ".thr_comm", this); 1417 } 1418 1419 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1420 // of PredBB. 1421 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1422 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1423 << DuplicationCost << " block is:" << *BB << "\n"); 1424 1425 // Unless PredBB ends with an unconditional branch, split the edge so that we 1426 // can just clone the bits from BB into the end of the new PredBB. 1427 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1428 1429 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { 1430 PredBB = SplitEdge(PredBB, BB, this); 1431 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1432 } 1433 1434 // We are going to have to map operands from the original BB block into the 1435 // PredBB block. Evaluate PHI nodes in BB. 1436 DenseMap<Instruction*, Value*> ValueMapping; 1437 1438 BasicBlock::iterator BI = BB->begin(); 1439 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1440 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1441 1442 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1443 // mapping and using it to remap operands in the cloned instructions. 1444 for (; BI != BB->end(); ++BI) { 1445 Instruction *New = BI->clone(); 1446 1447 // Remap operands to patch up intra-block references. 1448 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1449 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1450 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1451 if (I != ValueMapping.end()) 1452 New->setOperand(i, I->second); 1453 } 1454 1455 // If this instruction can be simplified after the operands are updated, 1456 // just use the simplified value instead. This frequently happens due to 1457 // phi translation. 1458 if (Value *IV = SimplifyInstruction(New, TD)) { 1459 delete New; 1460 ValueMapping[BI] = IV; 1461 } else { 1462 // Otherwise, insert the new instruction into the block. 1463 New->setName(BI->getName()); 1464 PredBB->getInstList().insert(OldPredBranch, New); 1465 ValueMapping[BI] = New; 1466 } 1467 } 1468 1469 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1470 // add entries to the PHI nodes for branch from PredBB now. 1471 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1472 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1473 ValueMapping); 1474 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1475 ValueMapping); 1476 1477 // If there were values defined in BB that are used outside the block, then we 1478 // now have to update all uses of the value to use either the original value, 1479 // the cloned value, or some PHI derived value. This can require arbitrary 1480 // PHI insertion, of which we are prepared to do, clean these up now. 1481 SSAUpdater SSAUpdate; 1482 SmallVector<Use*, 16> UsesToRename; 1483 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1484 // Scan all uses of this instruction to see if it is used outside of its 1485 // block, and if so, record them in UsesToRename. 1486 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1487 ++UI) { 1488 Instruction *User = cast<Instruction>(*UI); 1489 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1490 if (UserPN->getIncomingBlock(UI) == BB) 1491 continue; 1492 } else if (User->getParent() == BB) 1493 continue; 1494 1495 UsesToRename.push_back(&UI.getUse()); 1496 } 1497 1498 // If there are no uses outside the block, we're done with this instruction. 1499 if (UsesToRename.empty()) 1500 continue; 1501 1502 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1503 1504 // We found a use of I outside of BB. Rename all uses of I that are outside 1505 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1506 // with the two values we know. 1507 SSAUpdate.Initialize(I); 1508 SSAUpdate.AddAvailableValue(BB, I); 1509 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1510 1511 while (!UsesToRename.empty()) 1512 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1513 DEBUG(dbgs() << "\n"); 1514 } 1515 1516 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1517 // that we nuked. 1518 RemovePredecessorAndSimplify(BB, PredBB, TD); 1519 1520 // Remove the unconditional branch at the end of the PredBB block. 1521 OldPredBranch->eraseFromParent(); 1522 1523 ++NumDupes; 1524 return true; 1525} 1526 1527 1528