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