JumpThreading.cpp revision 9ccaf53ada99c63737547c0235baeb8454b04e80
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 DeleteDeadBlock(BB); 151 Changed = true; 152 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 153 // Can't thread an unconditional jump, but if the block is "almost 154 // empty", we can replace uses of it with uses of the successor and make 155 // this dead. 156 if (BI->isUnconditional() && 157 BB != &BB->getParent()->getEntryBlock()) { 158 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 159 // Ignore dbg intrinsics. 160 while (isa<DbgInfoIntrinsic>(BBI)) 161 ++BBI; 162 // If the terminator is the only non-phi instruction, try to nuke it. 163 if (BBI->isTerminator()) { 164 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 165 // block, we have to make sure it isn't in the LoopHeaders set. We 166 // reinsert afterward if needed. 167 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 168 BasicBlock *Succ = BI->getSuccessor(0); 169 170 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 171 Changed = true; 172 // If we deleted BB and BB was the header of a loop, then the 173 // successor is now the header of the loop. 174 BB = Succ; 175 } 176 177 if (ErasedFromLoopHeaders) 178 LoopHeaders.insert(BB); 179 } 180 } 181 } 182 } 183 EverChanged |= Changed; 184 } while (Changed); 185 186 LoopHeaders.clear(); 187 return EverChanged; 188} 189 190/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 191/// thread across it. 192static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 193 /// Ignore PHI nodes, these will be flattened when duplication happens. 194 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 195 196 // FIXME: THREADING will delete values that are just used to compute the 197 // branch, so they shouldn't count against the duplication cost. 198 199 200 // Sum up the cost of each instruction until we get to the terminator. Don't 201 // include the terminator because the copy won't include it. 202 unsigned Size = 0; 203 for (; !isa<TerminatorInst>(I); ++I) { 204 // Debugger intrinsics don't incur code size. 205 if (isa<DbgInfoIntrinsic>(I)) continue; 206 207 // If this is a pointer->pointer bitcast, it is free. 208 if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) 209 continue; 210 211 // All other instructions count for at least one unit. 212 ++Size; 213 214 // Calls are more expensive. If they are non-intrinsic calls, we model them 215 // as having cost of 4. If they are a non-vector intrinsic, we model them 216 // as having cost of 2 total, and if they are a vector intrinsic, we model 217 // them as having cost 1. 218 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 219 if (!isa<IntrinsicInst>(CI)) 220 Size += 3; 221 else if (!CI->getType()->isVectorTy()) 222 Size += 1; 223 } 224 } 225 226 // Threading through a switch statement is particularly profitable. If this 227 // block ends in a switch, decrease its cost to make it more likely to happen. 228 if (isa<SwitchInst>(I)) 229 Size = Size > 6 ? Size-6 : 0; 230 231 return Size; 232} 233 234/// FindLoopHeaders - We do not want jump threading to turn proper loop 235/// structures into irreducible loops. Doing this breaks up the loop nesting 236/// hierarchy and pessimizes later transformations. To prevent this from 237/// happening, we first have to find the loop headers. Here we approximate this 238/// by finding targets of backedges in the CFG. 239/// 240/// Note that there definitely are cases when we want to allow threading of 241/// edges across a loop header. For example, threading a jump from outside the 242/// loop (the preheader) to an exit block of the loop is definitely profitable. 243/// It is also almost always profitable to thread backedges from within the loop 244/// to exit blocks, and is often profitable to thread backedges to other blocks 245/// within the loop (forming a nested loop). This simple analysis is not rich 246/// enough to track all of these properties and keep it up-to-date as the CFG 247/// mutates, so we don't allow any of these transformations. 248/// 249void JumpThreading::FindLoopHeaders(Function &F) { 250 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 251 FindFunctionBackedges(F, Edges); 252 253 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 254 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 255} 256 257/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 258/// if we can infer that the value is a known ConstantInt in any of our 259/// predecessors. If so, return the known list of value and pred BB in the 260/// result vector. If a value is known to be undef, it is returned as null. 261/// 262/// This returns true if there were any known values. 263/// 264bool JumpThreading:: 265ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 266 // If V is a constantint, then it is known in all predecessors. 267 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 268 ConstantInt *CI = dyn_cast<ConstantInt>(V); 269 270 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 271 Result.push_back(std::make_pair(CI, *PI)); 272 return true; 273 } 274 275 // If V is a non-instruction value, or an instruction in a different block, 276 // then it can't be derived from a PHI. 277 Instruction *I = dyn_cast<Instruction>(V); 278 if (I == 0 || I->getParent() != BB) { 279 280 // Okay, if this is a live-in value, see if it has a known value at the end 281 // of any of our predecessors. 282 // 283 // FIXME: This should be an edge property, not a block end property. 284 /// TODO: Per PR2563, we could infer value range information about a 285 /// predecessor based on its terminator. 286 // 287 if (LVI) { 288 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 289 // "I" is a non-local compare-with-a-constant instruction. This would be 290 // able to handle value inequalities better, for example if the compare is 291 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 292 // Perhaps getConstantOnEdge should be smart enough to do this? 293 294 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 295 BasicBlock *P = *PI; 296 // If the value is known by LazyValueInfo to be a constant in a 297 // predecessor, use that information to try to thread this block. 298 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB); 299 if (PredCst == 0 || 300 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst))) 301 continue; 302 303 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), P)); 304 } 305 306 return !Result.empty(); 307 } 308 309 return false; 310 } 311 312 /// If I is a PHI node, then we know the incoming values for any constants. 313 if (PHINode *PN = dyn_cast<PHINode>(I)) { 314 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 315 Value *InVal = PN->getIncomingValue(i); 316 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 317 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 318 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 319 } 320 } 321 return !Result.empty(); 322 } 323 324 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 325 326 // Handle some boolean conditions. 327 if (I->getType()->getPrimitiveSizeInBits() == 1) { 328 // X | true -> true 329 // X & false -> false 330 if (I->getOpcode() == Instruction::Or || 331 I->getOpcode() == Instruction::And) { 332 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 333 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 334 335 if (LHSVals.empty() && RHSVals.empty()) 336 return false; 337 338 ConstantInt *InterestingVal; 339 if (I->getOpcode() == Instruction::Or) 340 InterestingVal = ConstantInt::getTrue(I->getContext()); 341 else 342 InterestingVal = ConstantInt::getFalse(I->getContext()); 343 344 // Scan for the sentinel. If we find an undef, force it to the 345 // interesting value: x|undef -> true and x&undef -> false. 346 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 347 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) { 348 Result.push_back(LHSVals[i]); 349 Result.back().first = InterestingVal; 350 } 351 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 352 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) { 353 // If we already inferred a value for this block on the LHS, don't 354 // re-add it. 355 bool HasValue = false; 356 for (unsigned r = 0, e = Result.size(); r != e; ++r) 357 if (Result[r].second == RHSVals[i].second) { 358 HasValue = true; 359 break; 360 } 361 362 if (!HasValue) { 363 Result.push_back(RHSVals[i]); 364 Result.back().first = InterestingVal; 365 } 366 } 367 return !Result.empty(); 368 } 369 370 // Handle the NOT form of XOR. 371 if (I->getOpcode() == Instruction::Xor && 372 isa<ConstantInt>(I->getOperand(1)) && 373 cast<ConstantInt>(I->getOperand(1))->isOne()) { 374 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result); 375 if (Result.empty()) 376 return false; 377 378 // Invert the known values. 379 for (unsigned i = 0, e = Result.size(); i != e; ++i) 380 if (Result[i].first) 381 Result[i].first = 382 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first)); 383 return true; 384 } 385 } 386 387 // Handle compare with phi operand, where the PHI is defined in this block. 388 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 389 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 390 if (PN && PN->getParent() == BB) { 391 // We can do this simplification if any comparisons fold to true or false. 392 // See if any do. 393 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 394 BasicBlock *PredBB = PN->getIncomingBlock(i); 395 Value *LHS = PN->getIncomingValue(i); 396 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 397 398 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD); 399 if (Res == 0) { 400 if (!LVI || !isa<Constant>(RHS)) 401 continue; 402 403 LazyValueInfo::Tristate 404 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 405 cast<Constant>(RHS), PredBB, BB); 406 if (ResT == LazyValueInfo::Unknown) 407 continue; 408 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 409 } 410 411 if (isa<UndefValue>(Res)) 412 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 413 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 414 Result.push_back(std::make_pair(CI, PredBB)); 415 } 416 417 return !Result.empty(); 418 } 419 420 421 // If comparing a live-in value against a constant, see if we know the 422 // live-in value on any predecessors. 423 if (LVI && isa<Constant>(Cmp->getOperand(1)) && 424 Cmp->getType()->isIntegerTy() && // Not vector compare. 425 (!isa<Instruction>(Cmp->getOperand(0)) || 426 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB)) { 427 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 428 429 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 430 BasicBlock *P = *PI; 431 // If the value is known by LazyValueInfo to be a constant in a 432 // predecessor, use that information to try to thread this block. 433 LazyValueInfo::Tristate 434 Res = LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 435 RHSCst, P, BB); 436 if (Res == LazyValueInfo::Unknown) 437 continue; 438 439 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 440 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), P)); 441 } 442 443 return !Result.empty(); 444 } 445 } 446 return false; 447} 448 449 450 451/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 452/// in an undefined jump, decide which block is best to revector to. 453/// 454/// Since we can pick an arbitrary destination, we pick the successor with the 455/// fewest predecessors. This should reduce the in-degree of the others. 456/// 457static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 458 TerminatorInst *BBTerm = BB->getTerminator(); 459 unsigned MinSucc = 0; 460 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 461 // Compute the successor with the minimum number of predecessors. 462 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 463 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 464 TestBB = BBTerm->getSuccessor(i); 465 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 466 if (NumPreds < MinNumPreds) 467 MinSucc = i; 468 } 469 470 return MinSucc; 471} 472 473/// ProcessBlock - If there are any predecessors whose control can be threaded 474/// through to a successor, transform them now. 475bool JumpThreading::ProcessBlock(BasicBlock *BB) { 476 // If the block is trivially dead, just return and let the caller nuke it. 477 // This simplifies other transformations. 478 if (pred_begin(BB) == pred_end(BB) && 479 BB != &BB->getParent()->getEntryBlock()) 480 return false; 481 482 // If this block has a single predecessor, and if that pred has a single 483 // successor, merge the blocks. This encourages recursive jump threading 484 // because now the condition in this block can be threaded through 485 // predecessors of our predecessor block. 486 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 487 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 488 SinglePred != BB) { 489 // If SinglePred was a loop header, BB becomes one. 490 if (LoopHeaders.erase(SinglePred)) 491 LoopHeaders.insert(BB); 492 493 // Remember if SinglePred was the entry block of the function. If so, we 494 // will need to move BB back to the entry position. 495 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 496 MergeBasicBlockIntoOnlyPred(BB); 497 498 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 499 BB->moveBefore(&BB->getParent()->getEntryBlock()); 500 return true; 501 } 502 } 503 504 // Look to see if the terminator is a branch of switch, if not we can't thread 505 // it. 506 Value *Condition; 507 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 508 // Can't thread an unconditional jump. 509 if (BI->isUnconditional()) return false; 510 Condition = BI->getCondition(); 511 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 512 Condition = SI->getCondition(); 513 else 514 return false; // Must be an invoke. 515 516 // If the terminator of this block is branching on a constant, simplify the 517 // terminator to an unconditional branch. This can occur due to threading in 518 // other blocks. 519 if (isa<ConstantInt>(Condition)) { 520 DEBUG(dbgs() << " In block '" << BB->getName() 521 << "' folding terminator: " << *BB->getTerminator() << '\n'); 522 ++NumFolds; 523 ConstantFoldTerminator(BB); 524 return true; 525 } 526 527 // If the terminator is branching on an undef, we can pick any of the 528 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 529 if (isa<UndefValue>(Condition)) { 530 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 531 532 // Fold the branch/switch. 533 TerminatorInst *BBTerm = BB->getTerminator(); 534 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 535 if (i == BestSucc) continue; 536 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD); 537 } 538 539 DEBUG(dbgs() << " In block '" << BB->getName() 540 << "' folding undef terminator: " << *BBTerm << '\n'); 541 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 542 BBTerm->eraseFromParent(); 543 return true; 544 } 545 546 Instruction *CondInst = dyn_cast<Instruction>(Condition); 547 548 // If the condition is an instruction defined in another block, see if a 549 // predecessor has the same condition: 550 // br COND, BBX, BBY 551 // BBX: 552 // br COND, BBZ, BBW 553 if (!LVI && 554 !Condition->hasOneUse() && // Multiple uses. 555 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 556 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 557 if (isa<BranchInst>(BB->getTerminator())) { 558 for (; PI != E; ++PI) { 559 BasicBlock *P = *PI; 560 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator())) 561 if (PBI->isConditional() && PBI->getCondition() == Condition && 562 ProcessBranchOnDuplicateCond(P, BB)) 563 return true; 564 } 565 } else { 566 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 567 for (; PI != E; ++PI) { 568 BasicBlock *P = *PI; 569 if (SwitchInst *PSI = dyn_cast<SwitchInst>(P->getTerminator())) 570 if (PSI->getCondition() == Condition && 571 ProcessSwitchOnDuplicateCond(P, BB)) 572 return true; 573 } 574 } 575 } 576 577 // All the rest of our checks depend on the condition being an instruction. 578 if (CondInst == 0) { 579 // FIXME: Unify this with code below. 580 if (LVI && ProcessThreadableEdges(Condition, BB)) 581 return true; 582 return false; 583 } 584 585 586 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 587 if (!LVI && 588 (!isa<PHINode>(CondCmp->getOperand(0)) || 589 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) { 590 // If we have a comparison, loop over the predecessors to see if there is 591 // a condition with a lexically identical value. 592 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 593 for (; PI != E; ++PI) { 594 BasicBlock *P = *PI; 595 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator())) 596 if (PBI->isConditional() && P != BB) { 597 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 598 if (CI->getOperand(0) == CondCmp->getOperand(0) && 599 CI->getOperand(1) == CondCmp->getOperand(1) && 600 CI->getPredicate() == CondCmp->getPredicate()) { 601 // TODO: Could handle things like (x != 4) --> (x == 17) 602 if (ProcessBranchOnDuplicateCond(P, BB)) 603 return true; 604 } 605 } 606 } 607 } 608 } 609 } 610 611 // Check for some cases that are worth simplifying. Right now we want to look 612 // for loads that are used by a switch or by the condition for the branch. If 613 // we see one, check to see if it's partially redundant. If so, insert a PHI 614 // which can then be used to thread the values. 615 // 616 Value *SimplifyValue = CondInst; 617 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 618 if (isa<Constant>(CondCmp->getOperand(1))) 619 SimplifyValue = CondCmp->getOperand(0); 620 621 // TODO: There are other places where load PRE would be profitable, such as 622 // more complex comparisons. 623 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 624 if (SimplifyPartiallyRedundantLoad(LI)) 625 return true; 626 627 628 // Handle a variety of cases where we are branching on something derived from 629 // a PHI node in the current block. If we can prove that any predecessors 630 // compute a predictable value based on a PHI node, thread those predecessors. 631 // 632 if (ProcessThreadableEdges(CondInst, BB)) 633 return true; 634 635 // If this is an otherwise-unfoldable branch on a phi node in the current 636 // block, see if we can simplify. 637 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 638 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 639 return ProcessBranchOnPHI(PN); 640 641 642 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 643 if (CondInst->getOpcode() == Instruction::Xor && 644 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 645 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 646 647 648 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 649 // "(X == 4)", thread through this block. 650 651 return false; 652} 653 654/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 655/// block that jump on exactly the same condition. This means that we almost 656/// always know the direction of the edge in the DESTBB: 657/// PREDBB: 658/// br COND, DESTBB, BBY 659/// DESTBB: 660/// br COND, BBZ, BBW 661/// 662/// If DESTBB has multiple predecessors, we can't just constant fold the branch 663/// in DESTBB, we have to thread over it. 664bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 665 BasicBlock *BB) { 666 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 667 668 // If both successors of PredBB go to DESTBB, we don't know anything. We can 669 // fold the branch to an unconditional one, which allows other recursive 670 // simplifications. 671 bool BranchDir; 672 if (PredBI->getSuccessor(1) != BB) 673 BranchDir = true; 674 else if (PredBI->getSuccessor(0) != BB) 675 BranchDir = false; 676 else { 677 DEBUG(dbgs() << " In block '" << PredBB->getName() 678 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 679 ++NumFolds; 680 ConstantFoldTerminator(PredBB); 681 return true; 682 } 683 684 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 685 686 // If the dest block has one predecessor, just fix the branch condition to a 687 // constant and fold it. 688 if (BB->getSinglePredecessor()) { 689 DEBUG(dbgs() << " In block '" << BB->getName() 690 << "' folding condition to '" << BranchDir << "': " 691 << *BB->getTerminator() << '\n'); 692 ++NumFolds; 693 Value *OldCond = DestBI->getCondition(); 694 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 695 BranchDir)); 696 // Delete dead instructions before we fold the branch. Folding the branch 697 // can eliminate edges from the CFG which can end up deleting OldCond. 698 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 699 ConstantFoldTerminator(BB); 700 return true; 701 } 702 703 704 // Next, figure out which successor we are threading to. 705 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 706 707 SmallVector<BasicBlock*, 2> Preds; 708 Preds.push_back(PredBB); 709 710 // Ok, try to thread it! 711 return ThreadEdge(BB, Preds, SuccBB); 712} 713 714/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 715/// block that switch on exactly the same condition. This means that we almost 716/// always know the direction of the edge in the DESTBB: 717/// PREDBB: 718/// switch COND [... DESTBB, BBY ... ] 719/// DESTBB: 720/// switch COND [... BBZ, BBW ] 721/// 722/// Optimizing switches like this is very important, because simplifycfg builds 723/// switches out of repeated 'if' conditions. 724bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 725 BasicBlock *DestBB) { 726 // Can't thread edge to self. 727 if (PredBB == DestBB) 728 return false; 729 730 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 731 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 732 733 // There are a variety of optimizations that we can potentially do on these 734 // blocks: we order them from most to least preferable. 735 736 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 737 // directly to their destination. This does not introduce *any* code size 738 // growth. Skip debug info first. 739 BasicBlock::iterator BBI = DestBB->begin(); 740 while (isa<DbgInfoIntrinsic>(BBI)) 741 BBI++; 742 743 // FIXME: Thread if it just contains a PHI. 744 if (isa<SwitchInst>(BBI)) { 745 bool MadeChange = false; 746 // Ignore the default edge for now. 747 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 748 ConstantInt *DestVal = DestSI->getCaseValue(i); 749 BasicBlock *DestSucc = DestSI->getSuccessor(i); 750 751 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 752 // PredSI has an explicit case for it. If so, forward. If it is covered 753 // by the default case, we can't update PredSI. 754 unsigned PredCase = PredSI->findCaseValue(DestVal); 755 if (PredCase == 0) continue; 756 757 // If PredSI doesn't go to DestBB on this value, then it won't reach the 758 // case on this condition. 759 if (PredSI->getSuccessor(PredCase) != DestBB && 760 DestSI->getSuccessor(i) != DestBB) 761 continue; 762 763 // Do not forward this if it already goes to this destination, this would 764 // be an infinite loop. 765 if (PredSI->getSuccessor(PredCase) == DestSucc) 766 continue; 767 768 // Otherwise, we're safe to make the change. Make sure that the edge from 769 // DestSI to DestSucc is not critical and has no PHI nodes. 770 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 771 DEBUG(dbgs() << "THROUGH: " << *DestSI); 772 773 // If the destination has PHI nodes, just split the edge for updating 774 // simplicity. 775 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 776 SplitCriticalEdge(DestSI, i, this); 777 DestSucc = DestSI->getSuccessor(i); 778 } 779 FoldSingleEntryPHINodes(DestSucc); 780 PredSI->setSuccessor(PredCase, DestSucc); 781 MadeChange = true; 782 } 783 784 if (MadeChange) 785 return true; 786 } 787 788 return false; 789} 790 791 792/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 793/// load instruction, eliminate it by replacing it with a PHI node. This is an 794/// important optimization that encourages jump threading, and needs to be run 795/// interlaced with other jump threading tasks. 796bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 797 // Don't hack volatile loads. 798 if (LI->isVolatile()) return false; 799 800 // If the load is defined in a block with exactly one predecessor, it can't be 801 // partially redundant. 802 BasicBlock *LoadBB = LI->getParent(); 803 if (LoadBB->getSinglePredecessor()) 804 return false; 805 806 Value *LoadedPtr = LI->getOperand(0); 807 808 // If the loaded operand is defined in the LoadBB, it can't be available. 809 // TODO: Could do simple PHI translation, that would be fun :) 810 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 811 if (PtrOp->getParent() == LoadBB) 812 return false; 813 814 // Scan a few instructions up from the load, to see if it is obviously live at 815 // the entry to its block. 816 BasicBlock::iterator BBIt = LI; 817 818 if (Value *AvailableVal = 819 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 820 // If the value if the load is locally available within the block, just use 821 // it. This frequently occurs for reg2mem'd allocas. 822 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 823 824 // If the returned value is the load itself, replace with an undef. This can 825 // only happen in dead loops. 826 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 827 LI->replaceAllUsesWith(AvailableVal); 828 LI->eraseFromParent(); 829 return true; 830 } 831 832 // Otherwise, if we scanned the whole block and got to the top of the block, 833 // we know the block is locally transparent to the load. If not, something 834 // might clobber its value. 835 if (BBIt != LoadBB->begin()) 836 return false; 837 838 839 SmallPtrSet<BasicBlock*, 8> PredsScanned; 840 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 841 AvailablePredsTy AvailablePreds; 842 BasicBlock *OneUnavailablePred = 0; 843 844 // If we got here, the loaded value is transparent through to the start of the 845 // block. Check to see if it is available in any of the predecessor blocks. 846 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 847 PI != PE; ++PI) { 848 BasicBlock *PredBB = *PI; 849 850 // If we already scanned this predecessor, skip it. 851 if (!PredsScanned.insert(PredBB)) 852 continue; 853 854 // Scan the predecessor to see if the value is available in the pred. 855 BBIt = PredBB->end(); 856 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 857 if (!PredAvailable) { 858 OneUnavailablePred = PredBB; 859 continue; 860 } 861 862 // If so, this load is partially redundant. Remember this info so that we 863 // can create a PHI node. 864 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 865 } 866 867 // If the loaded value isn't available in any predecessor, it isn't partially 868 // redundant. 869 if (AvailablePreds.empty()) return false; 870 871 // Okay, the loaded value is available in at least one (and maybe all!) 872 // predecessors. If the value is unavailable in more than one unique 873 // predecessor, we want to insert a merge block for those common predecessors. 874 // This ensures that we only have to insert one reload, thus not increasing 875 // code size. 876 BasicBlock *UnavailablePred = 0; 877 878 // If there is exactly one predecessor where the value is unavailable, the 879 // already computed 'OneUnavailablePred' block is it. If it ends in an 880 // unconditional branch, we know that it isn't a critical edge. 881 if (PredsScanned.size() == AvailablePreds.size()+1 && 882 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 883 UnavailablePred = OneUnavailablePred; 884 } else if (PredsScanned.size() != AvailablePreds.size()) { 885 // Otherwise, we had multiple unavailable predecessors or we had a critical 886 // edge from the one. 887 SmallVector<BasicBlock*, 8> PredsToSplit; 888 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 889 890 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 891 AvailablePredSet.insert(AvailablePreds[i].first); 892 893 // Add all the unavailable predecessors to the PredsToSplit list. 894 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 895 PI != PE; ++PI) { 896 BasicBlock *P = *PI; 897 // If the predecessor is an indirect goto, we can't split the edge. 898 if (isa<IndirectBrInst>(P->getTerminator())) 899 return false; 900 901 if (!AvailablePredSet.count(P)) 902 PredsToSplit.push_back(P); 903 } 904 905 // Split them out to their own block. 906 UnavailablePred = 907 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 908 "thread-pre-split", this); 909 } 910 911 // If the value isn't available in all predecessors, then there will be 912 // exactly one where it isn't available. Insert a load on that edge and add 913 // it to the AvailablePreds list. 914 if (UnavailablePred) { 915 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 916 "Can't handle critical edge here!"); 917 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 918 LI->getAlignment(), 919 UnavailablePred->getTerminator()); 920 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 921 } 922 923 // Now we know that each predecessor of this block has a value in 924 // AvailablePreds, sort them for efficient access as we're walking the preds. 925 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 926 927 // Create a PHI node at the start of the block for the PRE'd load value. 928 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 929 PN->takeName(LI); 930 931 // Insert new entries into the PHI for each predecessor. A single block may 932 // have multiple entries here. 933 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 934 ++PI) { 935 BasicBlock *P = *PI; 936 AvailablePredsTy::iterator I = 937 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 938 std::make_pair(P, (Value*)0)); 939 940 assert(I != AvailablePreds.end() && I->first == P && 941 "Didn't find entry for predecessor!"); 942 943 PN->addIncoming(I->second, I->first); 944 } 945 946 //cerr << "PRE: " << *LI << *PN << "\n"; 947 948 LI->replaceAllUsesWith(PN); 949 LI->eraseFromParent(); 950 951 return true; 952} 953 954/// FindMostPopularDest - The specified list contains multiple possible 955/// threadable destinations. Pick the one that occurs the most frequently in 956/// the list. 957static BasicBlock * 958FindMostPopularDest(BasicBlock *BB, 959 const SmallVectorImpl<std::pair<BasicBlock*, 960 BasicBlock*> > &PredToDestList) { 961 assert(!PredToDestList.empty()); 962 963 // Determine popularity. If there are multiple possible destinations, we 964 // explicitly choose to ignore 'undef' destinations. We prefer to thread 965 // blocks with known and real destinations to threading undef. We'll handle 966 // them later if interesting. 967 DenseMap<BasicBlock*, unsigned> DestPopularity; 968 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 969 if (PredToDestList[i].second) 970 DestPopularity[PredToDestList[i].second]++; 971 972 // Find the most popular dest. 973 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 974 BasicBlock *MostPopularDest = DPI->first; 975 unsigned Popularity = DPI->second; 976 SmallVector<BasicBlock*, 4> SamePopularity; 977 978 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 979 // If the popularity of this entry isn't higher than the popularity we've 980 // seen so far, ignore it. 981 if (DPI->second < Popularity) 982 ; // ignore. 983 else if (DPI->second == Popularity) { 984 // If it is the same as what we've seen so far, keep track of it. 985 SamePopularity.push_back(DPI->first); 986 } else { 987 // If it is more popular, remember it. 988 SamePopularity.clear(); 989 MostPopularDest = DPI->first; 990 Popularity = DPI->second; 991 } 992 } 993 994 // Okay, now we know the most popular destination. If there is more than 995 // destination, we need to determine one. This is arbitrary, but we need 996 // to make a deterministic decision. Pick the first one that appears in the 997 // successor list. 998 if (!SamePopularity.empty()) { 999 SamePopularity.push_back(MostPopularDest); 1000 TerminatorInst *TI = BB->getTerminator(); 1001 for (unsigned i = 0; ; ++i) { 1002 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 1003 1004 if (std::find(SamePopularity.begin(), SamePopularity.end(), 1005 TI->getSuccessor(i)) == SamePopularity.end()) 1006 continue; 1007 1008 MostPopularDest = TI->getSuccessor(i); 1009 break; 1010 } 1011 } 1012 1013 // Okay, we have finally picked the most popular destination. 1014 return MostPopularDest; 1015} 1016 1017bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) { 1018 // If threading this would thread across a loop header, don't even try to 1019 // thread the edge. 1020 if (LoopHeaders.count(BB)) 1021 return false; 1022 1023 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 1024 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues)) 1025 return false; 1026 assert(!PredValues.empty() && 1027 "ComputeValueKnownInPredecessors returned true with no values"); 1028 1029 DEBUG(dbgs() << "IN BB: " << *BB; 1030 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1031 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "; 1032 if (PredValues[i].first) 1033 dbgs() << *PredValues[i].first; 1034 else 1035 dbgs() << "UNDEF"; 1036 dbgs() << " for pred '" << PredValues[i].second->getName() 1037 << "'.\n"; 1038 }); 1039 1040 // Decide what we want to thread through. Convert our list of known values to 1041 // a list of known destinations for each pred. This also discards duplicate 1042 // predecessors and keeps track of the undefined inputs (which are represented 1043 // as a null dest in the PredToDestList). 1044 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1045 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1046 1047 BasicBlock *OnlyDest = 0; 1048 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1049 1050 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1051 BasicBlock *Pred = PredValues[i].second; 1052 if (!SeenPreds.insert(Pred)) 1053 continue; // Duplicate predecessor entry. 1054 1055 // If the predecessor ends with an indirect goto, we can't change its 1056 // destination. 1057 if (isa<IndirectBrInst>(Pred->getTerminator())) 1058 continue; 1059 1060 ConstantInt *Val = PredValues[i].first; 1061 1062 BasicBlock *DestBB; 1063 if (Val == 0) // Undef. 1064 DestBB = 0; 1065 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1066 DestBB = BI->getSuccessor(Val->isZero()); 1067 else { 1068 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 1069 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 1070 } 1071 1072 // If we have exactly one destination, remember it for efficiency below. 1073 if (i == 0) 1074 OnlyDest = DestBB; 1075 else if (OnlyDest != DestBB) 1076 OnlyDest = MultipleDestSentinel; 1077 1078 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1079 } 1080 1081 // If all edges were unthreadable, we fail. 1082 if (PredToDestList.empty()) 1083 return false; 1084 1085 // Determine which is the most common successor. If we have many inputs and 1086 // this block is a switch, we want to start by threading the batch that goes 1087 // to the most popular destination first. If we only know about one 1088 // threadable destination (the common case) we can avoid this. 1089 BasicBlock *MostPopularDest = OnlyDest; 1090 1091 if (MostPopularDest == MultipleDestSentinel) 1092 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1093 1094 // Now that we know what the most popular destination is, factor all 1095 // predecessors that will jump to it into a single predecessor. 1096 SmallVector<BasicBlock*, 16> PredsToFactor; 1097 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1098 if (PredToDestList[i].second == MostPopularDest) { 1099 BasicBlock *Pred = PredToDestList[i].first; 1100 1101 // This predecessor may be a switch or something else that has multiple 1102 // edges to the block. Factor each of these edges by listing them 1103 // according to # occurrences in PredsToFactor. 1104 TerminatorInst *PredTI = Pred->getTerminator(); 1105 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1106 if (PredTI->getSuccessor(i) == BB) 1107 PredsToFactor.push_back(Pred); 1108 } 1109 1110 // If the threadable edges are branching on an undefined value, we get to pick 1111 // the destination that these predecessors should get to. 1112 if (MostPopularDest == 0) 1113 MostPopularDest = BB->getTerminator()-> 1114 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1115 1116 // Ok, try to thread it! 1117 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1118} 1119 1120/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1121/// a PHI node in the current block. See if there are any simplifications we 1122/// can do based on inputs to the phi node. 1123/// 1124bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1125 BasicBlock *BB = PN->getParent(); 1126 1127 // TODO: We could make use of this to do it once for blocks with common PHI 1128 // values. 1129 SmallVector<BasicBlock*, 1> PredBBs; 1130 PredBBs.resize(1); 1131 1132 // If any of the predecessor blocks end in an unconditional branch, we can 1133 // *duplicate* the conditional branch into that block in order to further 1134 // encourage jump threading and to eliminate cases where we have branch on a 1135 // phi of an icmp (branch on icmp is much better). 1136 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1137 BasicBlock *PredBB = PN->getIncomingBlock(i); 1138 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1139 if (PredBr->isUnconditional()) { 1140 PredBBs[0] = PredBB; 1141 // Try to duplicate BB into PredBB. 1142 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1143 return true; 1144 } 1145 } 1146 1147 return false; 1148} 1149 1150/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1151/// a xor instruction in the current block. See if there are any 1152/// simplifications we can do based on inputs to the xor. 1153/// 1154bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1155 BasicBlock *BB = BO->getParent(); 1156 1157 // If either the LHS or RHS of the xor is a constant, don't do this 1158 // optimization. 1159 if (isa<ConstantInt>(BO->getOperand(0)) || 1160 isa<ConstantInt>(BO->getOperand(1))) 1161 return false; 1162 1163 // If the first instruction in BB isn't a phi, we won't be able to infer 1164 // anything special about any particular predecessor. 1165 if (!isa<PHINode>(BB->front())) 1166 return false; 1167 1168 // If we have a xor as the branch input to this block, and we know that the 1169 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1170 // the condition into the predecessor and fix that value to true, saving some 1171 // logical ops on that path and encouraging other paths to simplify. 1172 // 1173 // This copies something like this: 1174 // 1175 // BB: 1176 // %X = phi i1 [1], [%X'] 1177 // %Y = icmp eq i32 %A, %B 1178 // %Z = xor i1 %X, %Y 1179 // br i1 %Z, ... 1180 // 1181 // Into: 1182 // BB': 1183 // %Y = icmp ne i32 %A, %B 1184 // br i1 %Z, ... 1185 1186 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues; 1187 bool isLHS = true; 1188 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) { 1189 assert(XorOpValues.empty()); 1190 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues)) 1191 return false; 1192 isLHS = false; 1193 } 1194 1195 assert(!XorOpValues.empty() && 1196 "ComputeValueKnownInPredecessors returned true with no values"); 1197 1198 // Scan the information to see which is most popular: true or false. The 1199 // predecessors can be of the set true, false, or undef. 1200 unsigned NumTrue = 0, NumFalse = 0; 1201 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1202 if (!XorOpValues[i].first) continue; // Ignore undefs for the count. 1203 if (XorOpValues[i].first->isZero()) 1204 ++NumFalse; 1205 else 1206 ++NumTrue; 1207 } 1208 1209 // Determine which value to split on, true, false, or undef if neither. 1210 ConstantInt *SplitVal = 0; 1211 if (NumTrue > NumFalse) 1212 SplitVal = ConstantInt::getTrue(BB->getContext()); 1213 else if (NumTrue != 0 || NumFalse != 0) 1214 SplitVal = ConstantInt::getFalse(BB->getContext()); 1215 1216 // Collect all of the blocks that this can be folded into so that we can 1217 // factor this once and clone it once. 1218 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1219 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1220 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue; 1221 1222 BlocksToFoldInto.push_back(XorOpValues[i].second); 1223 } 1224 1225 // If we inferred a value for all of the predecessors, then duplication won't 1226 // help us. However, we can just replace the LHS or RHS with the constant. 1227 if (BlocksToFoldInto.size() == 1228 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1229 if (SplitVal == 0) { 1230 // If all preds provide undef, just nuke the xor, because it is undef too. 1231 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1232 BO->eraseFromParent(); 1233 } else if (SplitVal->isZero()) { 1234 // If all preds provide 0, replace the xor with the other input. 1235 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1236 BO->eraseFromParent(); 1237 } else { 1238 // If all preds provide 1, set the computed value to 1. 1239 BO->setOperand(!isLHS, SplitVal); 1240 } 1241 1242 return true; 1243 } 1244 1245 // Try to duplicate BB into PredBB. 1246 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1247} 1248 1249 1250/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1251/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1252/// NewPred using the entries from OldPred (suitably mapped). 1253static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1254 BasicBlock *OldPred, 1255 BasicBlock *NewPred, 1256 DenseMap<Instruction*, Value*> &ValueMap) { 1257 for (BasicBlock::iterator PNI = PHIBB->begin(); 1258 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1259 // Ok, we have a PHI node. Figure out what the incoming value was for the 1260 // DestBlock. 1261 Value *IV = PN->getIncomingValueForBlock(OldPred); 1262 1263 // Remap the value if necessary. 1264 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1265 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1266 if (I != ValueMap.end()) 1267 IV = I->second; 1268 } 1269 1270 PN->addIncoming(IV, NewPred); 1271 } 1272} 1273 1274/// ThreadEdge - We have decided that it is safe and profitable to factor the 1275/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1276/// across BB. Transform the IR to reflect this change. 1277bool JumpThreading::ThreadEdge(BasicBlock *BB, 1278 const SmallVectorImpl<BasicBlock*> &PredBBs, 1279 BasicBlock *SuccBB) { 1280 // If threading to the same block as we come from, we would infinite loop. 1281 if (SuccBB == BB) { 1282 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1283 << "' - would thread to self!\n"); 1284 return false; 1285 } 1286 1287 // If threading this would thread across a loop header, don't thread the edge. 1288 // See the comments above FindLoopHeaders for justifications and caveats. 1289 if (LoopHeaders.count(BB)) { 1290 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1291 << "' to dest BB '" << SuccBB->getName() 1292 << "' - it might create an irreducible loop!\n"); 1293 return false; 1294 } 1295 1296 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1297 if (JumpThreadCost > Threshold) { 1298 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1299 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1300 return false; 1301 } 1302 1303 // And finally, do it! Start by factoring the predecessors is needed. 1304 BasicBlock *PredBB; 1305 if (PredBBs.size() == 1) 1306 PredBB = PredBBs[0]; 1307 else { 1308 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1309 << " common predecessors.\n"); 1310 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1311 ".thr_comm", this); 1312 } 1313 1314 // And finally, do it! 1315 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1316 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1317 << ", across block:\n " 1318 << *BB << "\n"); 1319 1320 if (LVI) 1321 LVI->threadEdge(PredBB, BB, SuccBB); 1322 1323 // We are going to have to map operands from the original BB block to the new 1324 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1325 // account for entry from PredBB. 1326 DenseMap<Instruction*, Value*> ValueMapping; 1327 1328 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1329 BB->getName()+".thread", 1330 BB->getParent(), BB); 1331 NewBB->moveAfter(PredBB); 1332 1333 BasicBlock::iterator BI = BB->begin(); 1334 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1335 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1336 1337 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1338 // mapping and using it to remap operands in the cloned instructions. 1339 for (; !isa<TerminatorInst>(BI); ++BI) { 1340 Instruction *New = BI->clone(); 1341 New->setName(BI->getName()); 1342 NewBB->getInstList().push_back(New); 1343 ValueMapping[BI] = New; 1344 1345 // Remap operands to patch up intra-block references. 1346 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1347 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1348 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1349 if (I != ValueMapping.end()) 1350 New->setOperand(i, I->second); 1351 } 1352 } 1353 1354 // We didn't copy the terminator from BB over to NewBB, because there is now 1355 // an unconditional jump to SuccBB. Insert the unconditional jump. 1356 BranchInst::Create(SuccBB, NewBB); 1357 1358 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1359 // PHI nodes for NewBB now. 1360 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1361 1362 // If there were values defined in BB that are used outside the block, then we 1363 // now have to update all uses of the value to use either the original value, 1364 // the cloned value, or some PHI derived value. This can require arbitrary 1365 // PHI insertion, of which we are prepared to do, clean these up now. 1366 SSAUpdater SSAUpdate; 1367 SmallVector<Use*, 16> UsesToRename; 1368 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1369 // Scan all uses of this instruction to see if it is used outside of its 1370 // block, and if so, record them in UsesToRename. 1371 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1372 ++UI) { 1373 Instruction *User = cast<Instruction>(*UI); 1374 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1375 if (UserPN->getIncomingBlock(UI) == BB) 1376 continue; 1377 } else if (User->getParent() == BB) 1378 continue; 1379 1380 UsesToRename.push_back(&UI.getUse()); 1381 } 1382 1383 // If there are no uses outside the block, we're done with this instruction. 1384 if (UsesToRename.empty()) 1385 continue; 1386 1387 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1388 1389 // We found a use of I outside of BB. Rename all uses of I that are outside 1390 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1391 // with the two values we know. 1392 SSAUpdate.Initialize(I); 1393 SSAUpdate.AddAvailableValue(BB, I); 1394 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1395 1396 while (!UsesToRename.empty()) 1397 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1398 DEBUG(dbgs() << "\n"); 1399 } 1400 1401 1402 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1403 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1404 // us to simplify any PHI nodes in BB. 1405 TerminatorInst *PredTerm = PredBB->getTerminator(); 1406 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1407 if (PredTerm->getSuccessor(i) == BB) { 1408 RemovePredecessorAndSimplify(BB, PredBB, TD); 1409 PredTerm->setSuccessor(i, NewBB); 1410 } 1411 1412 // At this point, the IR is fully up to date and consistent. Do a quick scan 1413 // over the new instructions and zap any that are constants or dead. This 1414 // frequently happens because of phi translation. 1415 SimplifyInstructionsInBlock(NewBB, TD); 1416 1417 // Threaded an edge! 1418 ++NumThreads; 1419 return true; 1420} 1421 1422/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1423/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1424/// If we can duplicate the contents of BB up into PredBB do so now, this 1425/// improves the odds that the branch will be on an analyzable instruction like 1426/// a compare. 1427bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1428 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1429 assert(!PredBBs.empty() && "Can't handle an empty set"); 1430 1431 // If BB is a loop header, then duplicating this block outside the loop would 1432 // cause us to transform this into an irreducible loop, don't do this. 1433 // See the comments above FindLoopHeaders for justifications and caveats. 1434 if (LoopHeaders.count(BB)) { 1435 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1436 << "' into predecessor block '" << PredBBs[0]->getName() 1437 << "' - it might create an irreducible loop!\n"); 1438 return false; 1439 } 1440 1441 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1442 if (DuplicationCost > Threshold) { 1443 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1444 << "' - Cost is too high: " << DuplicationCost << "\n"); 1445 return false; 1446 } 1447 1448 // And finally, do it! Start by factoring the predecessors is needed. 1449 BasicBlock *PredBB; 1450 if (PredBBs.size() == 1) 1451 PredBB = PredBBs[0]; 1452 else { 1453 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1454 << " common predecessors.\n"); 1455 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1456 ".thr_comm", this); 1457 } 1458 1459 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1460 // of PredBB. 1461 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1462 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1463 << DuplicationCost << " block is:" << *BB << "\n"); 1464 1465 // Unless PredBB ends with an unconditional branch, split the edge so that we 1466 // can just clone the bits from BB into the end of the new PredBB. 1467 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1468 1469 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { 1470 PredBB = SplitEdge(PredBB, BB, this); 1471 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1472 } 1473 1474 // We are going to have to map operands from the original BB block into the 1475 // PredBB block. Evaluate PHI nodes in BB. 1476 DenseMap<Instruction*, Value*> ValueMapping; 1477 1478 BasicBlock::iterator BI = BB->begin(); 1479 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1480 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1481 1482 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1483 // mapping and using it to remap operands in the cloned instructions. 1484 for (; BI != BB->end(); ++BI) { 1485 Instruction *New = BI->clone(); 1486 1487 // Remap operands to patch up intra-block references. 1488 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1489 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1490 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1491 if (I != ValueMapping.end()) 1492 New->setOperand(i, I->second); 1493 } 1494 1495 // If this instruction can be simplified after the operands are updated, 1496 // just use the simplified value instead. This frequently happens due to 1497 // phi translation. 1498 if (Value *IV = SimplifyInstruction(New, TD)) { 1499 delete New; 1500 ValueMapping[BI] = IV; 1501 } else { 1502 // Otherwise, insert the new instruction into the block. 1503 New->setName(BI->getName()); 1504 PredBB->getInstList().insert(OldPredBranch, New); 1505 ValueMapping[BI] = New; 1506 } 1507 } 1508 1509 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1510 // add entries to the PHI nodes for branch from PredBB now. 1511 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1512 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1513 ValueMapping); 1514 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1515 ValueMapping); 1516 1517 // If there were values defined in BB that are used outside the block, then we 1518 // now have to update all uses of the value to use either the original value, 1519 // the cloned value, or some PHI derived value. This can require arbitrary 1520 // PHI insertion, of which we are prepared to do, clean these up now. 1521 SSAUpdater SSAUpdate; 1522 SmallVector<Use*, 16> UsesToRename; 1523 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1524 // Scan all uses of this instruction to see if it is used outside of its 1525 // block, and if so, record them in UsesToRename. 1526 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1527 ++UI) { 1528 Instruction *User = cast<Instruction>(*UI); 1529 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1530 if (UserPN->getIncomingBlock(UI) == BB) 1531 continue; 1532 } else if (User->getParent() == BB) 1533 continue; 1534 1535 UsesToRename.push_back(&UI.getUse()); 1536 } 1537 1538 // If there are no uses outside the block, we're done with this instruction. 1539 if (UsesToRename.empty()) 1540 continue; 1541 1542 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1543 1544 // We found a use of I outside of BB. Rename all uses of I that are outside 1545 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1546 // with the two values we know. 1547 SSAUpdate.Initialize(I); 1548 SSAUpdate.AddAvailableValue(BB, I); 1549 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1550 1551 while (!UsesToRename.empty()) 1552 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1553 DEBUG(dbgs() << "\n"); 1554 } 1555 1556 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1557 // that we nuked. 1558 RemovePredecessorAndSimplify(BB, PredBB, TD); 1559 1560 // Remove the unconditional branch at the end of the PredBB block. 1561 OldPredBranch->eraseFromParent(); 1562 1563 ++NumDupes; 1564 return true; 1565} 1566 1567 1568