JumpThreading.cpp revision 684d8d5b5003402241235917bfcf67d60f70c2c7
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/Analysis/Loads.h" 22#include "llvm/Transforms/Utils/BasicBlockUtils.h" 23#include "llvm/Transforms/Utils/Local.h" 24#include "llvm/Transforms/Utils/SSAUpdater.h" 25#include "llvm/Target/TargetData.h" 26#include "llvm/ADT/DenseMap.h" 27#include "llvm/ADT/Statistic.h" 28#include "llvm/ADT/STLExtras.h" 29#include "llvm/ADT/SmallPtrSet.h" 30#include "llvm/ADT/SmallSet.h" 31#include "llvm/Support/CommandLine.h" 32#include "llvm/Support/Debug.h" 33#include "llvm/Support/ValueHandle.h" 34#include "llvm/Support/raw_ostream.h" 35using namespace llvm; 36 37STATISTIC(NumThreads, "Number of jumps threaded"); 38STATISTIC(NumFolds, "Number of terminators folded"); 39STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 40 41static cl::opt<unsigned> 42Threshold("jump-threading-threshold", 43 cl::desc("Max block size to duplicate for jump threading"), 44 cl::init(6), cl::Hidden); 45 46// Turn on use of LazyValueInfo. 47static cl::opt<bool> 48EnableLVI("enable-jump-threading-lvi", 49 cl::desc("Use LVI for jump threading"), 50 cl::init(false), 51 cl::ReallyHidden); 52 53 54 55namespace { 56 /// This pass performs 'jump threading', which looks at blocks that have 57 /// multiple predecessors and multiple successors. If one or more of the 58 /// predecessors of the block can be proven to always jump to one of the 59 /// successors, we forward the edge from the predecessor to the successor by 60 /// duplicating the contents of this block. 61 /// 62 /// An example of when this can occur is code like this: 63 /// 64 /// if () { ... 65 /// X = 4; 66 /// } 67 /// if (X < 3) { 68 /// 69 /// In this case, the unconditional branch at the end of the first if can be 70 /// revectored to the false side of the second if. 71 /// 72 class JumpThreading : public FunctionPass { 73 TargetData *TD; 74 LazyValueInfo *LVI; 75#ifdef NDEBUG 76 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 77#else 78 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 79#endif 80 public: 81 static char ID; // Pass identification 82 JumpThreading() : FunctionPass(ID) {} 83 84 bool runOnFunction(Function &F); 85 86 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 87 if (EnableLVI) 88 AU.addRequired<LazyValueInfo>(); 89 } 90 91 void FindLoopHeaders(Function &F); 92 bool ProcessBlock(BasicBlock *BB); 93 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 94 BasicBlock *SuccBB); 95 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 96 const SmallVectorImpl<BasicBlock *> &PredBBs); 97 98 typedef SmallVectorImpl<std::pair<ConstantInt*, 99 BasicBlock*> > PredValueInfo; 100 101 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 102 PredValueInfo &Result); 103 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB); 104 105 106 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 107 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 108 109 bool ProcessBranchOnPHI(PHINode *PN); 110 bool ProcessBranchOnXOR(BinaryOperator *BO); 111 112 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 113 }; 114} 115 116char JumpThreading::ID = 0; 117INITIALIZE_PASS(JumpThreading, "jump-threading", 118 "Jump Threading", false, false); 119 120// Public interface to the Jump Threading pass 121FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 122 123/// runOnFunction - Top level algorithm. 124/// 125bool JumpThreading::runOnFunction(Function &F) { 126 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 127 TD = getAnalysisIfAvailable<TargetData>(); 128 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0; 129 130 FindLoopHeaders(F); 131 132 bool Changed, EverChanged = false; 133 do { 134 Changed = false; 135 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 136 BasicBlock *BB = I; 137 // Thread all of the branches we can over this block. 138 while (ProcessBlock(BB)) 139 Changed = true; 140 141 ++I; 142 143 // If the block is trivially dead, zap it. This eliminates the successor 144 // edges which simplifies the CFG. 145 if (pred_begin(BB) == pred_end(BB) && 146 BB != &BB->getParent()->getEntryBlock()) { 147 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 148 << "' with terminator: " << *BB->getTerminator() << '\n'); 149 LoopHeaders.erase(BB); 150 if (LVI) LVI->eraseBlock(BB); 151 DeleteDeadBlock(BB); 152 Changed = true; 153 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 154 // Can't thread an unconditional jump, but if the block is "almost 155 // empty", we can replace uses of it with uses of the successor and make 156 // this dead. 157 if (BI->isUnconditional() && 158 BB != &BB->getParent()->getEntryBlock()) { 159 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 160 // Ignore dbg intrinsics. 161 while (isa<DbgInfoIntrinsic>(BBI)) 162 ++BBI; 163 // If the terminator is the only non-phi instruction, try to nuke it. 164 if (BBI->isTerminator()) { 165 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 166 // block, we have to make sure it isn't in the LoopHeaders set. We 167 // reinsert afterward if needed. 168 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 169 BasicBlock *Succ = BI->getSuccessor(0); 170 171 // FIXME: It is always conservatively correct to drop the info 172 // for a block even if it doesn't get erased. This isn't totally 173 // awesome, but it allows us to use AssertingVH to prevent nasty 174 // dangling pointer issues within LazyValueInfo. 175 if (LVI) LVI->eraseBlock(BB); 176 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 177 Changed = true; 178 // If we deleted BB and BB was the header of a loop, then the 179 // successor is now the header of the loop. 180 BB = Succ; 181 } 182 183 if (ErasedFromLoopHeaders) 184 LoopHeaders.insert(BB); 185 } 186 } 187 } 188 } 189 EverChanged |= Changed; 190 } while (Changed); 191 192 LoopHeaders.clear(); 193 return EverChanged; 194} 195 196/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 197/// thread across it. 198static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 199 /// Ignore PHI nodes, these will be flattened when duplication happens. 200 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 201 202 // FIXME: THREADING will delete values that are just used to compute the 203 // branch, so they shouldn't count against the duplication cost. 204 205 206 // Sum up the cost of each instruction until we get to the terminator. Don't 207 // include the terminator because the copy won't include it. 208 unsigned Size = 0; 209 for (; !isa<TerminatorInst>(I); ++I) { 210 // Debugger intrinsics don't incur code size. 211 if (isa<DbgInfoIntrinsic>(I)) continue; 212 213 // If this is a pointer->pointer bitcast, it is free. 214 if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) 215 continue; 216 217 // All other instructions count for at least one unit. 218 ++Size; 219 220 // Calls are more expensive. If they are non-intrinsic calls, we model them 221 // as having cost of 4. If they are a non-vector intrinsic, we model them 222 // as having cost of 2 total, and if they are a vector intrinsic, we model 223 // them as having cost 1. 224 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 225 if (!isa<IntrinsicInst>(CI)) 226 Size += 3; 227 else if (!CI->getType()->isVectorTy()) 228 Size += 1; 229 } 230 } 231 232 // Threading through a switch statement is particularly profitable. If this 233 // block ends in a switch, decrease its cost to make it more likely to happen. 234 if (isa<SwitchInst>(I)) 235 Size = Size > 6 ? Size-6 : 0; 236 237 return Size; 238} 239 240/// FindLoopHeaders - We do not want jump threading to turn proper loop 241/// structures into irreducible loops. Doing this breaks up the loop nesting 242/// hierarchy and pessimizes later transformations. To prevent this from 243/// happening, we first have to find the loop headers. Here we approximate this 244/// by finding targets of backedges in the CFG. 245/// 246/// Note that there definitely are cases when we want to allow threading of 247/// edges across a loop header. For example, threading a jump from outside the 248/// loop (the preheader) to an exit block of the loop is definitely profitable. 249/// It is also almost always profitable to thread backedges from within the loop 250/// to exit blocks, and is often profitable to thread backedges to other blocks 251/// within the loop (forming a nested loop). This simple analysis is not rich 252/// enough to track all of these properties and keep it up-to-date as the CFG 253/// mutates, so we don't allow any of these transformations. 254/// 255void JumpThreading::FindLoopHeaders(Function &F) { 256 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 257 FindFunctionBackedges(F, Edges); 258 259 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 260 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 261} 262 263/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 264/// if we can infer that the value is a known ConstantInt in any of our 265/// predecessors. If so, return the known list of value and pred BB in the 266/// result vector. If a value is known to be undef, it is returned as null. 267/// 268/// This returns true if there were any known values. 269/// 270bool JumpThreading:: 271ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 272 // If V is a constantint, then it is known in all predecessors. 273 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 274 ConstantInt *CI = dyn_cast<ConstantInt>(V); 275 276 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 277 Result.push_back(std::make_pair(CI, *PI)); 278 return true; 279 } 280 281 // If V is a non-instruction value, or an instruction in a different block, 282 // then it can't be derived from a PHI. 283 Instruction *I = dyn_cast<Instruction>(V); 284 if (I == 0 || I->getParent() != BB) { 285 286 // Okay, if this is a live-in value, see if it has a known value at the end 287 // of any of our predecessors. 288 // 289 // FIXME: This should be an edge property, not a block end property. 290 /// TODO: Per PR2563, we could infer value range information about a 291 /// predecessor based on its terminator. 292 // 293 if (LVI) { 294 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 295 // "I" is a non-local compare-with-a-constant instruction. This would be 296 // able to handle value inequalities better, for example if the compare is 297 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 298 // Perhaps getConstantOnEdge should be smart enough to do this? 299 300 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 301 BasicBlock *P = *PI; 302 // If the value is known by LazyValueInfo to be a constant in a 303 // predecessor, use that information to try to thread this block. 304 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB); 305 if (PredCst == 0 || 306 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst))) 307 continue; 308 309 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), P)); 310 } 311 312 return !Result.empty(); 313 } 314 315 return false; 316 } 317 318 /// If I is a PHI node, then we know the incoming values for any constants. 319 if (PHINode *PN = dyn_cast<PHINode>(I)) { 320 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 321 Value *InVal = PN->getIncomingValue(i); 322 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 323 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 324 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 325 } 326 } 327 return !Result.empty(); 328 } 329 330 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 331 332 // Handle some boolean conditions. 333 if (I->getType()->getPrimitiveSizeInBits() == 1) { 334 // X | true -> true 335 // X & false -> false 336 if (I->getOpcode() == Instruction::Or || 337 I->getOpcode() == Instruction::And) { 338 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 339 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 340 341 if (LHSVals.empty() && RHSVals.empty()) 342 return false; 343 344 ConstantInt *InterestingVal; 345 if (I->getOpcode() == Instruction::Or) 346 InterestingVal = ConstantInt::getTrue(I->getContext()); 347 else 348 InterestingVal = ConstantInt::getFalse(I->getContext()); 349 350 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; 351 352 // Scan for the sentinel. If we find an undef, force it to the 353 // interesting value: x|undef -> true and x&undef -> false. 354 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 355 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) { 356 Result.push_back(LHSVals[i]); 357 Result.back().first = InterestingVal; 358 LHSKnownBBs.insert(LHSVals[i].second); 359 } 360 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 361 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) { 362 // If we already inferred a value for this block on the LHS, don't 363 // re-add it. 364 if (!LHSKnownBBs.count(RHSVals[i].second)) { 365 Result.push_back(RHSVals[i]); 366 Result.back().first = InterestingVal; 367 } 368 } 369 return !Result.empty(); 370 } 371 372 // Handle the NOT form of XOR. 373 if (I->getOpcode() == Instruction::Xor && 374 isa<ConstantInt>(I->getOperand(1)) && 375 cast<ConstantInt>(I->getOperand(1))->isOne()) { 376 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result); 377 if (Result.empty()) 378 return false; 379 380 // Invert the known values. 381 for (unsigned i = 0, e = Result.size(); i != e; ++i) 382 if (Result[i].first) 383 Result[i].first = 384 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first)); 385 return true; 386 } 387 } 388 389 // Handle compare with phi operand, where the PHI is defined in this block. 390 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 391 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 392 if (PN && PN->getParent() == BB) { 393 // We can do this simplification if any comparisons fold to true or false. 394 // See if any do. 395 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 396 BasicBlock *PredBB = PN->getIncomingBlock(i); 397 Value *LHS = PN->getIncomingValue(i); 398 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 399 400 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD); 401 if (Res == 0) { 402 if (!LVI || !isa<Constant>(RHS)) 403 continue; 404 405 LazyValueInfo::Tristate 406 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 407 cast<Constant>(RHS), PredBB, BB); 408 if (ResT == LazyValueInfo::Unknown) 409 continue; 410 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 411 } 412 413 if (isa<UndefValue>(Res)) 414 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 415 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 416 Result.push_back(std::make_pair(CI, PredBB)); 417 } 418 419 return !Result.empty(); 420 } 421 422 423 // If comparing a live-in value against a constant, see if we know the 424 // live-in value on any predecessors. 425 if (LVI && isa<Constant>(Cmp->getOperand(1)) && 426 Cmp->getType()->isIntegerTy() && // Not vector compare. 427 (!isa<Instruction>(Cmp->getOperand(0)) || 428 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB)) { 429 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 430 431 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 432 BasicBlock *P = *PI; 433 // If the value is known by LazyValueInfo to be a constant in a 434 // predecessor, use that information to try to thread this block. 435 LazyValueInfo::Tristate 436 Res = LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 437 RHSCst, P, BB); 438 if (Res == LazyValueInfo::Unknown) 439 continue; 440 441 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 442 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), P)); 443 } 444 445 return !Result.empty(); 446 } 447 } 448 return false; 449} 450 451 452 453/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 454/// in an undefined jump, decide which block is best to revector to. 455/// 456/// Since we can pick an arbitrary destination, we pick the successor with the 457/// fewest predecessors. This should reduce the in-degree of the others. 458/// 459static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 460 TerminatorInst *BBTerm = BB->getTerminator(); 461 unsigned MinSucc = 0; 462 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 463 // Compute the successor with the minimum number of predecessors. 464 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 465 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 466 TestBB = BBTerm->getSuccessor(i); 467 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 468 if (NumPreds < MinNumPreds) 469 MinSucc = i; 470 } 471 472 return MinSucc; 473} 474 475/// ProcessBlock - If there are any predecessors whose control can be threaded 476/// through to a successor, transform them now. 477bool JumpThreading::ProcessBlock(BasicBlock *BB) { 478 // If the block is trivially dead, just return and let the caller nuke it. 479 // This simplifies other transformations. 480 if (pred_begin(BB) == pred_end(BB) && 481 BB != &BB->getParent()->getEntryBlock()) 482 return false; 483 484 // If this block has a single predecessor, and if that pred has a single 485 // successor, merge the blocks. This encourages recursive jump threading 486 // because now the condition in this block can be threaded through 487 // predecessors of our predecessor block. 488 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 489 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 490 SinglePred != BB) { 491 // If SinglePred was a loop header, BB becomes one. 492 if (LoopHeaders.erase(SinglePred)) 493 LoopHeaders.insert(BB); 494 495 // Remember if SinglePred was the entry block of the function. If so, we 496 // will need to move BB back to the entry position. 497 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 498 if (LVI) LVI->eraseBlock(SinglePred); 499 MergeBasicBlockIntoOnlyPred(BB); 500 501 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 502 BB->moveBefore(&BB->getParent()->getEntryBlock()); 503 return true; 504 } 505 } 506 507 // Look to see if the terminator is a branch of switch, if not we can't thread 508 // it. 509 Value *Condition; 510 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 511 // Can't thread an unconditional jump. 512 if (BI->isUnconditional()) return false; 513 Condition = BI->getCondition(); 514 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 515 Condition = SI->getCondition(); 516 else 517 return false; // Must be an invoke. 518 519 // If the terminator of this block is branching on a constant, simplify the 520 // terminator to an unconditional branch. This can occur due to threading in 521 // other blocks. 522 if (isa<ConstantInt>(Condition)) { 523 DEBUG(dbgs() << " In block '" << BB->getName() 524 << "' folding terminator: " << *BB->getTerminator() << '\n'); 525 ++NumFolds; 526 ConstantFoldTerminator(BB); 527 return true; 528 } 529 530 // If the terminator is branching on an undef, we can pick any of the 531 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 532 if (isa<UndefValue>(Condition)) { 533 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 534 535 // Fold the branch/switch. 536 TerminatorInst *BBTerm = BB->getTerminator(); 537 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 538 if (i == BestSucc) continue; 539 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD); 540 } 541 542 DEBUG(dbgs() << " In block '" << BB->getName() 543 << "' folding undef terminator: " << *BBTerm << '\n'); 544 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 545 BBTerm->eraseFromParent(); 546 return true; 547 } 548 549 Instruction *CondInst = dyn_cast<Instruction>(Condition); 550 551 // If the condition is an instruction defined in another block, see if a 552 // predecessor has the same condition: 553 // br COND, BBX, BBY 554 // BBX: 555 // br COND, BBZ, BBW 556 if (!LVI && 557 !Condition->hasOneUse() && // Multiple uses. 558 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 559 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 560 if (isa<BranchInst>(BB->getTerminator())) { 561 for (; PI != E; ++PI) { 562 BasicBlock *P = *PI; 563 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator())) 564 if (PBI->isConditional() && PBI->getCondition() == Condition && 565 ProcessBranchOnDuplicateCond(P, BB)) 566 return true; 567 } 568 } else { 569 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 570 for (; PI != E; ++PI) { 571 BasicBlock *P = *PI; 572 if (SwitchInst *PSI = dyn_cast<SwitchInst>(P->getTerminator())) 573 if (PSI->getCondition() == Condition && 574 ProcessSwitchOnDuplicateCond(P, BB)) 575 return true; 576 } 577 } 578 } 579 580 // All the rest of our checks depend on the condition being an instruction. 581 if (CondInst == 0) { 582 // FIXME: Unify this with code below. 583 if (LVI && ProcessThreadableEdges(Condition, BB)) 584 return true; 585 return false; 586 } 587 588 589 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 590 if (!LVI && 591 (!isa<PHINode>(CondCmp->getOperand(0)) || 592 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) { 593 // If we have a comparison, loop over the predecessors to see if there is 594 // a condition with a lexically identical value. 595 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 596 for (; PI != E; ++PI) { 597 BasicBlock *P = *PI; 598 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator())) 599 if (PBI->isConditional() && P != BB) { 600 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 601 if (CI->getOperand(0) == CondCmp->getOperand(0) && 602 CI->getOperand(1) == CondCmp->getOperand(1) && 603 CI->getPredicate() == CondCmp->getPredicate()) { 604 // TODO: Could handle things like (x != 4) --> (x == 17) 605 if (ProcessBranchOnDuplicateCond(P, BB)) 606 return true; 607 } 608 } 609 } 610 } 611 } 612 } 613 614 // Check for some cases that are worth simplifying. Right now we want to look 615 // for loads that are used by a switch or by the condition for the branch. If 616 // we see one, check to see if it's partially redundant. If so, insert a PHI 617 // which can then be used to thread the values. 618 // 619 Value *SimplifyValue = CondInst; 620 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 621 if (isa<Constant>(CondCmp->getOperand(1))) 622 SimplifyValue = CondCmp->getOperand(0); 623 624 // TODO: There are other places where load PRE would be profitable, such as 625 // more complex comparisons. 626 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 627 if (SimplifyPartiallyRedundantLoad(LI)) 628 return true; 629 630 631 // Handle a variety of cases where we are branching on something derived from 632 // a PHI node in the current block. If we can prove that any predecessors 633 // compute a predictable value based on a PHI node, thread those predecessors. 634 // 635 if (ProcessThreadableEdges(CondInst, BB)) 636 return true; 637 638 // If this is an otherwise-unfoldable branch on a phi node in the current 639 // block, see if we can simplify. 640 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 641 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 642 return ProcessBranchOnPHI(PN); 643 644 645 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 646 if (CondInst->getOpcode() == Instruction::Xor && 647 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 648 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 649 650 651 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 652 // "(X == 4)", thread through this block. 653 654 return false; 655} 656 657/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 658/// block that jump on exactly the same condition. This means that we almost 659/// always know the direction of the edge in the DESTBB: 660/// PREDBB: 661/// br COND, DESTBB, BBY 662/// DESTBB: 663/// br COND, BBZ, BBW 664/// 665/// If DESTBB has multiple predecessors, we can't just constant fold the branch 666/// in DESTBB, we have to thread over it. 667bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 668 BasicBlock *BB) { 669 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 670 671 // If both successors of PredBB go to DESTBB, we don't know anything. We can 672 // fold the branch to an unconditional one, which allows other recursive 673 // simplifications. 674 bool BranchDir; 675 if (PredBI->getSuccessor(1) != BB) 676 BranchDir = true; 677 else if (PredBI->getSuccessor(0) != BB) 678 BranchDir = false; 679 else { 680 DEBUG(dbgs() << " In block '" << PredBB->getName() 681 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 682 ++NumFolds; 683 ConstantFoldTerminator(PredBB); 684 return true; 685 } 686 687 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 688 689 // If the dest block has one predecessor, just fix the branch condition to a 690 // constant and fold it. 691 if (BB->getSinglePredecessor()) { 692 DEBUG(dbgs() << " In block '" << BB->getName() 693 << "' folding condition to '" << BranchDir << "': " 694 << *BB->getTerminator() << '\n'); 695 ++NumFolds; 696 Value *OldCond = DestBI->getCondition(); 697 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 698 BranchDir)); 699 // Delete dead instructions before we fold the branch. Folding the branch 700 // can eliminate edges from the CFG which can end up deleting OldCond. 701 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 702 ConstantFoldTerminator(BB); 703 return true; 704 } 705 706 707 // Next, figure out which successor we are threading to. 708 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 709 710 SmallVector<BasicBlock*, 2> Preds; 711 Preds.push_back(PredBB); 712 713 // Ok, try to thread it! 714 return ThreadEdge(BB, Preds, SuccBB); 715} 716 717/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 718/// block that switch on exactly the same condition. This means that we almost 719/// always know the direction of the edge in the DESTBB: 720/// PREDBB: 721/// switch COND [... DESTBB, BBY ... ] 722/// DESTBB: 723/// switch COND [... BBZ, BBW ] 724/// 725/// Optimizing switches like this is very important, because simplifycfg builds 726/// switches out of repeated 'if' conditions. 727bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 728 BasicBlock *DestBB) { 729 // Can't thread edge to self. 730 if (PredBB == DestBB) 731 return false; 732 733 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 734 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 735 736 // There are a variety of optimizations that we can potentially do on these 737 // blocks: we order them from most to least preferable. 738 739 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 740 // directly to their destination. This does not introduce *any* code size 741 // growth. Skip debug info first. 742 BasicBlock::iterator BBI = DestBB->begin(); 743 while (isa<DbgInfoIntrinsic>(BBI)) 744 BBI++; 745 746 // FIXME: Thread if it just contains a PHI. 747 if (isa<SwitchInst>(BBI)) { 748 bool MadeChange = false; 749 // Ignore the default edge for now. 750 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 751 ConstantInt *DestVal = DestSI->getCaseValue(i); 752 BasicBlock *DestSucc = DestSI->getSuccessor(i); 753 754 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 755 // PredSI has an explicit case for it. If so, forward. If it is covered 756 // by the default case, we can't update PredSI. 757 unsigned PredCase = PredSI->findCaseValue(DestVal); 758 if (PredCase == 0) continue; 759 760 // If PredSI doesn't go to DestBB on this value, then it won't reach the 761 // case on this condition. 762 if (PredSI->getSuccessor(PredCase) != DestBB && 763 DestSI->getSuccessor(i) != DestBB) 764 continue; 765 766 // Do not forward this if it already goes to this destination, this would 767 // be an infinite loop. 768 if (PredSI->getSuccessor(PredCase) == DestSucc) 769 continue; 770 771 // Otherwise, we're safe to make the change. Make sure that the edge from 772 // DestSI to DestSucc is not critical and has no PHI nodes. 773 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 774 DEBUG(dbgs() << "THROUGH: " << *DestSI); 775 776 // If the destination has PHI nodes, just split the edge for updating 777 // simplicity. 778 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 779 SplitCriticalEdge(DestSI, i, this); 780 DestSucc = DestSI->getSuccessor(i); 781 } 782 FoldSingleEntryPHINodes(DestSucc); 783 PredSI->setSuccessor(PredCase, DestSucc); 784 MadeChange = true; 785 } 786 787 if (MadeChange) 788 return true; 789 } 790 791 return false; 792} 793 794 795/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 796/// load instruction, eliminate it by replacing it with a PHI node. This is an 797/// important optimization that encourages jump threading, and needs to be run 798/// interlaced with other jump threading tasks. 799bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 800 // Don't hack volatile loads. 801 if (LI->isVolatile()) return false; 802 803 // If the load is defined in a block with exactly one predecessor, it can't be 804 // partially redundant. 805 BasicBlock *LoadBB = LI->getParent(); 806 if (LoadBB->getSinglePredecessor()) 807 return false; 808 809 Value *LoadedPtr = LI->getOperand(0); 810 811 // If the loaded operand is defined in the LoadBB, it can't be available. 812 // TODO: Could do simple PHI translation, that would be fun :) 813 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 814 if (PtrOp->getParent() == LoadBB) 815 return false; 816 817 // Scan a few instructions up from the load, to see if it is obviously live at 818 // the entry to its block. 819 BasicBlock::iterator BBIt = LI; 820 821 if (Value *AvailableVal = 822 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 823 // If the value if the load is locally available within the block, just use 824 // it. This frequently occurs for reg2mem'd allocas. 825 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 826 827 // If the returned value is the load itself, replace with an undef. This can 828 // only happen in dead loops. 829 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 830 LI->replaceAllUsesWith(AvailableVal); 831 LI->eraseFromParent(); 832 return true; 833 } 834 835 // Otherwise, if we scanned the whole block and got to the top of the block, 836 // we know the block is locally transparent to the load. If not, something 837 // might clobber its value. 838 if (BBIt != LoadBB->begin()) 839 return false; 840 841 842 SmallPtrSet<BasicBlock*, 8> PredsScanned; 843 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 844 AvailablePredsTy AvailablePreds; 845 BasicBlock *OneUnavailablePred = 0; 846 847 // If we got here, the loaded value is transparent through to the start of the 848 // block. Check to see if it is available in any of the predecessor blocks. 849 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 850 PI != PE; ++PI) { 851 BasicBlock *PredBB = *PI; 852 853 // If we already scanned this predecessor, skip it. 854 if (!PredsScanned.insert(PredBB)) 855 continue; 856 857 // Scan the predecessor to see if the value is available in the pred. 858 BBIt = PredBB->end(); 859 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 860 if (!PredAvailable) { 861 OneUnavailablePred = PredBB; 862 continue; 863 } 864 865 // If so, this load is partially redundant. Remember this info so that we 866 // can create a PHI node. 867 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 868 } 869 870 // If the loaded value isn't available in any predecessor, it isn't partially 871 // redundant. 872 if (AvailablePreds.empty()) return false; 873 874 // Okay, the loaded value is available in at least one (and maybe all!) 875 // predecessors. If the value is unavailable in more than one unique 876 // predecessor, we want to insert a merge block for those common predecessors. 877 // This ensures that we only have to insert one reload, thus not increasing 878 // code size. 879 BasicBlock *UnavailablePred = 0; 880 881 // If there is exactly one predecessor where the value is unavailable, the 882 // already computed 'OneUnavailablePred' block is it. If it ends in an 883 // unconditional branch, we know that it isn't a critical edge. 884 if (PredsScanned.size() == AvailablePreds.size()+1 && 885 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 886 UnavailablePred = OneUnavailablePred; 887 } else if (PredsScanned.size() != AvailablePreds.size()) { 888 // Otherwise, we had multiple unavailable predecessors or we had a critical 889 // edge from the one. 890 SmallVector<BasicBlock*, 8> PredsToSplit; 891 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 892 893 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 894 AvailablePredSet.insert(AvailablePreds[i].first); 895 896 // Add all the unavailable predecessors to the PredsToSplit list. 897 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 898 PI != PE; ++PI) { 899 BasicBlock *P = *PI; 900 // If the predecessor is an indirect goto, we can't split the edge. 901 if (isa<IndirectBrInst>(P->getTerminator())) 902 return false; 903 904 if (!AvailablePredSet.count(P)) 905 PredsToSplit.push_back(P); 906 } 907 908 // Split them out to their own block. 909 UnavailablePred = 910 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 911 "thread-pre-split", this); 912 } 913 914 // If the value isn't available in all predecessors, then there will be 915 // exactly one where it isn't available. Insert a load on that edge and add 916 // it to the AvailablePreds list. 917 if (UnavailablePred) { 918 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 919 "Can't handle critical edge here!"); 920 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 921 LI->getAlignment(), 922 UnavailablePred->getTerminator()); 923 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 924 } 925 926 // Now we know that each predecessor of this block has a value in 927 // AvailablePreds, sort them for efficient access as we're walking the preds. 928 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 929 930 // Create a PHI node at the start of the block for the PRE'd load value. 931 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 932 PN->takeName(LI); 933 934 // Insert new entries into the PHI for each predecessor. A single block may 935 // have multiple entries here. 936 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 937 ++PI) { 938 BasicBlock *P = *PI; 939 AvailablePredsTy::iterator I = 940 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 941 std::make_pair(P, (Value*)0)); 942 943 assert(I != AvailablePreds.end() && I->first == P && 944 "Didn't find entry for predecessor!"); 945 946 PN->addIncoming(I->second, I->first); 947 } 948 949 //cerr << "PRE: " << *LI << *PN << "\n"; 950 951 LI->replaceAllUsesWith(PN); 952 LI->eraseFromParent(); 953 954 return true; 955} 956 957/// FindMostPopularDest - The specified list contains multiple possible 958/// threadable destinations. Pick the one that occurs the most frequently in 959/// the list. 960static BasicBlock * 961FindMostPopularDest(BasicBlock *BB, 962 const SmallVectorImpl<std::pair<BasicBlock*, 963 BasicBlock*> > &PredToDestList) { 964 assert(!PredToDestList.empty()); 965 966 // Determine popularity. If there are multiple possible destinations, we 967 // explicitly choose to ignore 'undef' destinations. We prefer to thread 968 // blocks with known and real destinations to threading undef. We'll handle 969 // them later if interesting. 970 DenseMap<BasicBlock*, unsigned> DestPopularity; 971 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 972 if (PredToDestList[i].second) 973 DestPopularity[PredToDestList[i].second]++; 974 975 // Find the most popular dest. 976 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 977 BasicBlock *MostPopularDest = DPI->first; 978 unsigned Popularity = DPI->second; 979 SmallVector<BasicBlock*, 4> SamePopularity; 980 981 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 982 // If the popularity of this entry isn't higher than the popularity we've 983 // seen so far, ignore it. 984 if (DPI->second < Popularity) 985 ; // ignore. 986 else if (DPI->second == Popularity) { 987 // If it is the same as what we've seen so far, keep track of it. 988 SamePopularity.push_back(DPI->first); 989 } else { 990 // If it is more popular, remember it. 991 SamePopularity.clear(); 992 MostPopularDest = DPI->first; 993 Popularity = DPI->second; 994 } 995 } 996 997 // Okay, now we know the most popular destination. If there is more than 998 // destination, we need to determine one. This is arbitrary, but we need 999 // to make a deterministic decision. Pick the first one that appears in the 1000 // successor list. 1001 if (!SamePopularity.empty()) { 1002 SamePopularity.push_back(MostPopularDest); 1003 TerminatorInst *TI = BB->getTerminator(); 1004 for (unsigned i = 0; ; ++i) { 1005 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 1006 1007 if (std::find(SamePopularity.begin(), SamePopularity.end(), 1008 TI->getSuccessor(i)) == SamePopularity.end()) 1009 continue; 1010 1011 MostPopularDest = TI->getSuccessor(i); 1012 break; 1013 } 1014 } 1015 1016 // Okay, we have finally picked the most popular destination. 1017 return MostPopularDest; 1018} 1019 1020bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) { 1021 // If threading this would thread across a loop header, don't even try to 1022 // thread the edge. 1023 if (LoopHeaders.count(BB)) 1024 return false; 1025 1026 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 1027 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues)) 1028 return false; 1029 assert(!PredValues.empty() && 1030 "ComputeValueKnownInPredecessors returned true with no values"); 1031 1032 DEBUG(dbgs() << "IN BB: " << *BB; 1033 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1034 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "; 1035 if (PredValues[i].first) 1036 dbgs() << *PredValues[i].first; 1037 else 1038 dbgs() << "UNDEF"; 1039 dbgs() << " for pred '" << PredValues[i].second->getName() 1040 << "'.\n"; 1041 }); 1042 1043 // Decide what we want to thread through. Convert our list of known values to 1044 // a list of known destinations for each pred. This also discards duplicate 1045 // predecessors and keeps track of the undefined inputs (which are represented 1046 // as a null dest in the PredToDestList). 1047 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1048 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1049 1050 BasicBlock *OnlyDest = 0; 1051 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1052 1053 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1054 BasicBlock *Pred = PredValues[i].second; 1055 if (!SeenPreds.insert(Pred)) 1056 continue; // Duplicate predecessor entry. 1057 1058 // If the predecessor ends with an indirect goto, we can't change its 1059 // destination. 1060 if (isa<IndirectBrInst>(Pred->getTerminator())) 1061 continue; 1062 1063 ConstantInt *Val = PredValues[i].first; 1064 1065 BasicBlock *DestBB; 1066 if (Val == 0) // Undef. 1067 DestBB = 0; 1068 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1069 DestBB = BI->getSuccessor(Val->isZero()); 1070 else { 1071 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 1072 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 1073 } 1074 1075 // If we have exactly one destination, remember it for efficiency below. 1076 if (i == 0) 1077 OnlyDest = DestBB; 1078 else if (OnlyDest != DestBB) 1079 OnlyDest = MultipleDestSentinel; 1080 1081 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1082 } 1083 1084 // If all edges were unthreadable, we fail. 1085 if (PredToDestList.empty()) 1086 return false; 1087 1088 // Determine which is the most common successor. If we have many inputs and 1089 // this block is a switch, we want to start by threading the batch that goes 1090 // to the most popular destination first. If we only know about one 1091 // threadable destination (the common case) we can avoid this. 1092 BasicBlock *MostPopularDest = OnlyDest; 1093 1094 if (MostPopularDest == MultipleDestSentinel) 1095 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1096 1097 // Now that we know what the most popular destination is, factor all 1098 // predecessors that will jump to it into a single predecessor. 1099 SmallVector<BasicBlock*, 16> PredsToFactor; 1100 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1101 if (PredToDestList[i].second == MostPopularDest) { 1102 BasicBlock *Pred = PredToDestList[i].first; 1103 1104 // This predecessor may be a switch or something else that has multiple 1105 // edges to the block. Factor each of these edges by listing them 1106 // according to # occurrences in PredsToFactor. 1107 TerminatorInst *PredTI = Pred->getTerminator(); 1108 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1109 if (PredTI->getSuccessor(i) == BB) 1110 PredsToFactor.push_back(Pred); 1111 } 1112 1113 // If the threadable edges are branching on an undefined value, we get to pick 1114 // the destination that these predecessors should get to. 1115 if (MostPopularDest == 0) 1116 MostPopularDest = BB->getTerminator()-> 1117 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1118 1119 // Ok, try to thread it! 1120 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1121} 1122 1123/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1124/// a PHI node in the current block. See if there are any simplifications we 1125/// can do based on inputs to the phi node. 1126/// 1127bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1128 BasicBlock *BB = PN->getParent(); 1129 1130 // TODO: We could make use of this to do it once for blocks with common PHI 1131 // values. 1132 SmallVector<BasicBlock*, 1> PredBBs; 1133 PredBBs.resize(1); 1134 1135 // If any of the predecessor blocks end in an unconditional branch, we can 1136 // *duplicate* the conditional branch into that block in order to further 1137 // encourage jump threading and to eliminate cases where we have branch on a 1138 // phi of an icmp (branch on icmp is much better). 1139 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1140 BasicBlock *PredBB = PN->getIncomingBlock(i); 1141 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1142 if (PredBr->isUnconditional()) { 1143 PredBBs[0] = PredBB; 1144 // Try to duplicate BB into PredBB. 1145 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1146 return true; 1147 } 1148 } 1149 1150 return false; 1151} 1152 1153/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1154/// a xor instruction in the current block. See if there are any 1155/// simplifications we can do based on inputs to the xor. 1156/// 1157bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1158 BasicBlock *BB = BO->getParent(); 1159 1160 // If either the LHS or RHS of the xor is a constant, don't do this 1161 // optimization. 1162 if (isa<ConstantInt>(BO->getOperand(0)) || 1163 isa<ConstantInt>(BO->getOperand(1))) 1164 return false; 1165 1166 // If the first instruction in BB isn't a phi, we won't be able to infer 1167 // anything special about any particular predecessor. 1168 if (!isa<PHINode>(BB->front())) 1169 return false; 1170 1171 // If we have a xor as the branch input to this block, and we know that the 1172 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1173 // the condition into the predecessor and fix that value to true, saving some 1174 // logical ops on that path and encouraging other paths to simplify. 1175 // 1176 // This copies something like this: 1177 // 1178 // BB: 1179 // %X = phi i1 [1], [%X'] 1180 // %Y = icmp eq i32 %A, %B 1181 // %Z = xor i1 %X, %Y 1182 // br i1 %Z, ... 1183 // 1184 // Into: 1185 // BB': 1186 // %Y = icmp ne i32 %A, %B 1187 // br i1 %Z, ... 1188 1189 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues; 1190 bool isLHS = true; 1191 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) { 1192 assert(XorOpValues.empty()); 1193 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues)) 1194 return false; 1195 isLHS = false; 1196 } 1197 1198 assert(!XorOpValues.empty() && 1199 "ComputeValueKnownInPredecessors returned true with no values"); 1200 1201 // Scan the information to see which is most popular: true or false. The 1202 // predecessors can be of the set true, false, or undef. 1203 unsigned NumTrue = 0, NumFalse = 0; 1204 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1205 if (!XorOpValues[i].first) continue; // Ignore undefs for the count. 1206 if (XorOpValues[i].first->isZero()) 1207 ++NumFalse; 1208 else 1209 ++NumTrue; 1210 } 1211 1212 // Determine which value to split on, true, false, or undef if neither. 1213 ConstantInt *SplitVal = 0; 1214 if (NumTrue > NumFalse) 1215 SplitVal = ConstantInt::getTrue(BB->getContext()); 1216 else if (NumTrue != 0 || NumFalse != 0) 1217 SplitVal = ConstantInt::getFalse(BB->getContext()); 1218 1219 // Collect all of the blocks that this can be folded into so that we can 1220 // factor this once and clone it once. 1221 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1222 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1223 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue; 1224 1225 BlocksToFoldInto.push_back(XorOpValues[i].second); 1226 } 1227 1228 // If we inferred a value for all of the predecessors, then duplication won't 1229 // help us. However, we can just replace the LHS or RHS with the constant. 1230 if (BlocksToFoldInto.size() == 1231 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1232 if (SplitVal == 0) { 1233 // If all preds provide undef, just nuke the xor, because it is undef too. 1234 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1235 BO->eraseFromParent(); 1236 } else if (SplitVal->isZero()) { 1237 // If all preds provide 0, replace the xor with the other input. 1238 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1239 BO->eraseFromParent(); 1240 } else { 1241 // If all preds provide 1, set the computed value to 1. 1242 BO->setOperand(!isLHS, SplitVal); 1243 } 1244 1245 return true; 1246 } 1247 1248 // Try to duplicate BB into PredBB. 1249 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1250} 1251 1252 1253/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1254/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1255/// NewPred using the entries from OldPred (suitably mapped). 1256static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1257 BasicBlock *OldPred, 1258 BasicBlock *NewPred, 1259 DenseMap<Instruction*, Value*> &ValueMap) { 1260 for (BasicBlock::iterator PNI = PHIBB->begin(); 1261 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1262 // Ok, we have a PHI node. Figure out what the incoming value was for the 1263 // DestBlock. 1264 Value *IV = PN->getIncomingValueForBlock(OldPred); 1265 1266 // Remap the value if necessary. 1267 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1268 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1269 if (I != ValueMap.end()) 1270 IV = I->second; 1271 } 1272 1273 PN->addIncoming(IV, NewPred); 1274 } 1275} 1276 1277/// ThreadEdge - We have decided that it is safe and profitable to factor the 1278/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1279/// across BB. Transform the IR to reflect this change. 1280bool JumpThreading::ThreadEdge(BasicBlock *BB, 1281 const SmallVectorImpl<BasicBlock*> &PredBBs, 1282 BasicBlock *SuccBB) { 1283 // If threading to the same block as we come from, we would infinite loop. 1284 if (SuccBB == BB) { 1285 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1286 << "' - would thread to self!\n"); 1287 return false; 1288 } 1289 1290 // If threading this would thread across a loop header, don't thread the edge. 1291 // See the comments above FindLoopHeaders for justifications and caveats. 1292 if (LoopHeaders.count(BB)) { 1293 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1294 << "' to dest BB '" << SuccBB->getName() 1295 << "' - it might create an irreducible loop!\n"); 1296 return false; 1297 } 1298 1299 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1300 if (JumpThreadCost > Threshold) { 1301 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1302 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1303 return false; 1304 } 1305 1306 // And finally, do it! Start by factoring the predecessors is needed. 1307 BasicBlock *PredBB; 1308 if (PredBBs.size() == 1) 1309 PredBB = PredBBs[0]; 1310 else { 1311 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1312 << " common predecessors.\n"); 1313 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1314 ".thr_comm", this); 1315 } 1316 1317 // And finally, do it! 1318 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1319 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1320 << ", across block:\n " 1321 << *BB << "\n"); 1322 1323 if (LVI) 1324 LVI->threadEdge(PredBB, BB, SuccBB); 1325 1326 // We are going to have to map operands from the original BB block to the new 1327 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1328 // account for entry from PredBB. 1329 DenseMap<Instruction*, Value*> ValueMapping; 1330 1331 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1332 BB->getName()+".thread", 1333 BB->getParent(), BB); 1334 NewBB->moveAfter(PredBB); 1335 1336 BasicBlock::iterator BI = BB->begin(); 1337 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1338 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1339 1340 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1341 // mapping and using it to remap operands in the cloned instructions. 1342 for (; !isa<TerminatorInst>(BI); ++BI) { 1343 Instruction *New = BI->clone(); 1344 New->setName(BI->getName()); 1345 NewBB->getInstList().push_back(New); 1346 ValueMapping[BI] = New; 1347 1348 // Remap operands to patch up intra-block references. 1349 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1350 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1351 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1352 if (I != ValueMapping.end()) 1353 New->setOperand(i, I->second); 1354 } 1355 } 1356 1357 // We didn't copy the terminator from BB over to NewBB, because there is now 1358 // an unconditional jump to SuccBB. Insert the unconditional jump. 1359 BranchInst::Create(SuccBB, NewBB); 1360 1361 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1362 // PHI nodes for NewBB now. 1363 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1364 1365 // If there were values defined in BB that are used outside the block, then we 1366 // now have to update all uses of the value to use either the original value, 1367 // the cloned value, or some PHI derived value. This can require arbitrary 1368 // PHI insertion, of which we are prepared to do, clean these up now. 1369 SSAUpdater SSAUpdate; 1370 SmallVector<Use*, 16> UsesToRename; 1371 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1372 // Scan all uses of this instruction to see if it is used outside of its 1373 // block, and if so, record them in UsesToRename. 1374 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1375 ++UI) { 1376 Instruction *User = cast<Instruction>(*UI); 1377 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1378 if (UserPN->getIncomingBlock(UI) == BB) 1379 continue; 1380 } else if (User->getParent() == BB) 1381 continue; 1382 1383 UsesToRename.push_back(&UI.getUse()); 1384 } 1385 1386 // If there are no uses outside the block, we're done with this instruction. 1387 if (UsesToRename.empty()) 1388 continue; 1389 1390 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1391 1392 // We found a use of I outside of BB. Rename all uses of I that are outside 1393 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1394 // with the two values we know. 1395 SSAUpdate.Initialize(I); 1396 SSAUpdate.AddAvailableValue(BB, I); 1397 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1398 1399 while (!UsesToRename.empty()) 1400 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1401 DEBUG(dbgs() << "\n"); 1402 } 1403 1404 1405 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1406 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1407 // us to simplify any PHI nodes in BB. 1408 TerminatorInst *PredTerm = PredBB->getTerminator(); 1409 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1410 if (PredTerm->getSuccessor(i) == BB) { 1411 RemovePredecessorAndSimplify(BB, PredBB, TD); 1412 PredTerm->setSuccessor(i, NewBB); 1413 } 1414 1415 // At this point, the IR is fully up to date and consistent. Do a quick scan 1416 // over the new instructions and zap any that are constants or dead. This 1417 // frequently happens because of phi translation. 1418 SimplifyInstructionsInBlock(NewBB, TD); 1419 1420 // Threaded an edge! 1421 ++NumThreads; 1422 return true; 1423} 1424 1425/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1426/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1427/// If we can duplicate the contents of BB up into PredBB do so now, this 1428/// improves the odds that the branch will be on an analyzable instruction like 1429/// a compare. 1430bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1431 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1432 assert(!PredBBs.empty() && "Can't handle an empty set"); 1433 1434 // If BB is a loop header, then duplicating this block outside the loop would 1435 // cause us to transform this into an irreducible loop, don't do this. 1436 // See the comments above FindLoopHeaders for justifications and caveats. 1437 if (LoopHeaders.count(BB)) { 1438 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1439 << "' into predecessor block '" << PredBBs[0]->getName() 1440 << "' - it might create an irreducible loop!\n"); 1441 return false; 1442 } 1443 1444 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1445 if (DuplicationCost > Threshold) { 1446 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1447 << "' - Cost is too high: " << DuplicationCost << "\n"); 1448 return false; 1449 } 1450 1451 // And finally, do it! Start by factoring the predecessors is needed. 1452 BasicBlock *PredBB; 1453 if (PredBBs.size() == 1) 1454 PredBB = PredBBs[0]; 1455 else { 1456 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1457 << " common predecessors.\n"); 1458 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1459 ".thr_comm", this); 1460 } 1461 1462 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1463 // of PredBB. 1464 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1465 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1466 << DuplicationCost << " block is:" << *BB << "\n"); 1467 1468 // Unless PredBB ends with an unconditional branch, split the edge so that we 1469 // can just clone the bits from BB into the end of the new PredBB. 1470 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1471 1472 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { 1473 PredBB = SplitEdge(PredBB, BB, this); 1474 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1475 } 1476 1477 // We are going to have to map operands from the original BB block into the 1478 // PredBB block. Evaluate PHI nodes in BB. 1479 DenseMap<Instruction*, Value*> ValueMapping; 1480 1481 BasicBlock::iterator BI = BB->begin(); 1482 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1483 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1484 1485 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1486 // mapping and using it to remap operands in the cloned instructions. 1487 for (; BI != BB->end(); ++BI) { 1488 Instruction *New = BI->clone(); 1489 1490 // Remap operands to patch up intra-block references. 1491 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1492 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1493 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1494 if (I != ValueMapping.end()) 1495 New->setOperand(i, I->second); 1496 } 1497 1498 // If this instruction can be simplified after the operands are updated, 1499 // just use the simplified value instead. This frequently happens due to 1500 // phi translation. 1501 if (Value *IV = SimplifyInstruction(New, TD)) { 1502 delete New; 1503 ValueMapping[BI] = IV; 1504 } else { 1505 // Otherwise, insert the new instruction into the block. 1506 New->setName(BI->getName()); 1507 PredBB->getInstList().insert(OldPredBranch, New); 1508 ValueMapping[BI] = New; 1509 } 1510 } 1511 1512 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1513 // add entries to the PHI nodes for branch from PredBB now. 1514 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1515 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1516 ValueMapping); 1517 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1518 ValueMapping); 1519 1520 // If there were values defined in BB that are used outside the block, then we 1521 // now have to update all uses of the value to use either the original value, 1522 // the cloned value, or some PHI derived value. This can require arbitrary 1523 // PHI insertion, of which we are prepared to do, clean these up now. 1524 SSAUpdater SSAUpdate; 1525 SmallVector<Use*, 16> UsesToRename; 1526 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1527 // Scan all uses of this instruction to see if it is used outside of its 1528 // block, and if so, record them in UsesToRename. 1529 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1530 ++UI) { 1531 Instruction *User = cast<Instruction>(*UI); 1532 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1533 if (UserPN->getIncomingBlock(UI) == BB) 1534 continue; 1535 } else if (User->getParent() == BB) 1536 continue; 1537 1538 UsesToRename.push_back(&UI.getUse()); 1539 } 1540 1541 // If there are no uses outside the block, we're done with this instruction. 1542 if (UsesToRename.empty()) 1543 continue; 1544 1545 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1546 1547 // We found a use of I outside of BB. Rename all uses of I that are outside 1548 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1549 // with the two values we know. 1550 SSAUpdate.Initialize(I); 1551 SSAUpdate.AddAvailableValue(BB, I); 1552 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1553 1554 while (!UsesToRename.empty()) 1555 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1556 DEBUG(dbgs() << "\n"); 1557 } 1558 1559 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1560 // that we nuked. 1561 RemovePredecessorAndSimplify(BB, PredBB, TD); 1562 1563 // Remove the unconditional branch at the end of the PredBB block. 1564 OldPredBranch->eraseFromParent(); 1565 1566 ++NumDupes; 1567 return true; 1568} 1569 1570 1571