JumpThreading.cpp revision d13db2c59cc94162d6cf0a04187d408bfef6d4a7
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", cl::ReallyHidden); 49 50 51 52namespace { 53 /// This pass performs 'jump threading', which looks at blocks that have 54 /// multiple predecessors and multiple successors. If one or more of the 55 /// predecessors of the block can be proven to always jump to one of the 56 /// successors, we forward the edge from the predecessor to the successor by 57 /// duplicating the contents of this block. 58 /// 59 /// An example of when this can occur is code like this: 60 /// 61 /// if () { ... 62 /// X = 4; 63 /// } 64 /// if (X < 3) { 65 /// 66 /// In this case, the unconditional branch at the end of the first if can be 67 /// revectored to the false side of the second if. 68 /// 69 class JumpThreading : public FunctionPass { 70 TargetData *TD; 71 LazyValueInfo *LVI; 72#ifdef NDEBUG 73 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 74#else 75 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 76#endif 77 public: 78 static char ID; // Pass identification 79 JumpThreading() : FunctionPass(&ID) {} 80 81 bool runOnFunction(Function &F); 82 83 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 84 if (EnableLVI) 85 AU.addRequired<LazyValueInfo>(); 86 } 87 88 void FindLoopHeaders(Function &F); 89 bool ProcessBlock(BasicBlock *BB); 90 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 91 BasicBlock *SuccBB); 92 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 93 const SmallVectorImpl<BasicBlock *> &PredBBs); 94 95 typedef SmallVectorImpl<std::pair<ConstantInt*, 96 BasicBlock*> > PredValueInfo; 97 98 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 99 PredValueInfo &Result); 100 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB); 101 102 103 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 104 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 105 106 bool ProcessBranchOnPHI(PHINode *PN); 107 bool ProcessBranchOnXOR(BinaryOperator *BO); 108 109 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 110 }; 111} 112 113char JumpThreading::ID = 0; 114INITIALIZE_PASS(JumpThreading, "jump-threading", 115 "Jump Threading", false, false); 116 117// Public interface to the Jump Threading pass 118FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 119 120/// runOnFunction - Top level algorithm. 121/// 122bool JumpThreading::runOnFunction(Function &F) { 123 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 124 TD = getAnalysisIfAvailable<TargetData>(); 125 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0; 126 127 FindLoopHeaders(F); 128 129 bool Changed, EverChanged = false; 130 do { 131 Changed = false; 132 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 133 BasicBlock *BB = I; 134 // Thread all of the branches we can over this block. 135 while (ProcessBlock(BB)) 136 Changed = true; 137 138 ++I; 139 140 // If the block is trivially dead, zap it. This eliminates the successor 141 // edges which simplifies the CFG. 142 if (pred_begin(BB) == pred_end(BB) && 143 BB != &BB->getParent()->getEntryBlock()) { 144 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 145 << "' with terminator: " << *BB->getTerminator() << '\n'); 146 LoopHeaders.erase(BB); 147 DeleteDeadBlock(BB); 148 Changed = true; 149 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 150 // Can't thread an unconditional jump, but if the block is "almost 151 // empty", we can replace uses of it with uses of the successor and make 152 // this dead. 153 if (BI->isUnconditional() && 154 BB != &BB->getParent()->getEntryBlock()) { 155 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 156 // Ignore dbg intrinsics. 157 while (isa<DbgInfoIntrinsic>(BBI)) 158 ++BBI; 159 // If the terminator is the only non-phi instruction, try to nuke it. 160 if (BBI->isTerminator()) { 161 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 162 // block, we have to make sure it isn't in the LoopHeaders set. We 163 // reinsert afterward if needed. 164 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 165 BasicBlock *Succ = BI->getSuccessor(0); 166 167 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 168 Changed = true; 169 // If we deleted BB and BB was the header of a loop, then the 170 // successor is now the header of the loop. 171 BB = Succ; 172 } 173 174 if (ErasedFromLoopHeaders) 175 LoopHeaders.insert(BB); 176 } 177 } 178 } 179 } 180 EverChanged |= Changed; 181 } while (Changed); 182 183 LoopHeaders.clear(); 184 return EverChanged; 185} 186 187/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 188/// thread across it. 189static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 190 /// Ignore PHI nodes, these will be flattened when duplication happens. 191 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 192 193 // FIXME: THREADING will delete values that are just used to compute the 194 // branch, so they shouldn't count against the duplication cost. 195 196 197 // Sum up the cost of each instruction until we get to the terminator. Don't 198 // include the terminator because the copy won't include it. 199 unsigned Size = 0; 200 for (; !isa<TerminatorInst>(I); ++I) { 201 // Debugger intrinsics don't incur code size. 202 if (isa<DbgInfoIntrinsic>(I)) continue; 203 204 // If this is a pointer->pointer bitcast, it is free. 205 if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) 206 continue; 207 208 // All other instructions count for at least one unit. 209 ++Size; 210 211 // Calls are more expensive. If they are non-intrinsic calls, we model them 212 // as having cost of 4. If they are a non-vector intrinsic, we model them 213 // as having cost of 2 total, and if they are a vector intrinsic, we model 214 // them as having cost 1. 215 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 216 if (!isa<IntrinsicInst>(CI)) 217 Size += 3; 218 else if (!CI->getType()->isVectorTy()) 219 Size += 1; 220 } 221 } 222 223 // Threading through a switch statement is particularly profitable. If this 224 // block ends in a switch, decrease its cost to make it more likely to happen. 225 if (isa<SwitchInst>(I)) 226 Size = Size > 6 ? Size-6 : 0; 227 228 return Size; 229} 230 231/// FindLoopHeaders - We do not want jump threading to turn proper loop 232/// structures into irreducible loops. Doing this breaks up the loop nesting 233/// hierarchy and pessimizes later transformations. To prevent this from 234/// happening, we first have to find the loop headers. Here we approximate this 235/// by finding targets of backedges in the CFG. 236/// 237/// Note that there definitely are cases when we want to allow threading of 238/// edges across a loop header. For example, threading a jump from outside the 239/// loop (the preheader) to an exit block of the loop is definitely profitable. 240/// It is also almost always profitable to thread backedges from within the loop 241/// to exit blocks, and is often profitable to thread backedges to other blocks 242/// within the loop (forming a nested loop). This simple analysis is not rich 243/// enough to track all of these properties and keep it up-to-date as the CFG 244/// mutates, so we don't allow any of these transformations. 245/// 246void JumpThreading::FindLoopHeaders(Function &F) { 247 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 248 FindFunctionBackedges(F, Edges); 249 250 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 251 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 252} 253 254/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 255/// if we can infer that the value is a known ConstantInt in any of our 256/// predecessors. If so, return the known list of value and pred BB in the 257/// result vector. If a value is known to be undef, it is returned as null. 258/// 259/// This returns true if there were any known values. 260/// 261bool JumpThreading:: 262ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 263 // If V is a constantint, then it is known in all predecessors. 264 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 265 ConstantInt *CI = dyn_cast<ConstantInt>(V); 266 267 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 268 Result.push_back(std::make_pair(CI, *PI)); 269 return true; 270 } 271 272 // If V is a non-instruction value, or an instruction in a different block, 273 // then it can't be derived from a PHI. 274 Instruction *I = dyn_cast<Instruction>(V); 275 if (I == 0 || I->getParent() != BB) { 276 277 // Okay, if this is a live-in value, see if it has a known value at the end 278 // of any of our predecessors. 279 // 280 // FIXME: This should be an edge property, not a block end property. 281 /// TODO: Per PR2563, we could infer value range information about a 282 /// predecessor based on its terminator. 283 // 284 if (LVI) { 285 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 286 // "I" is a non-local compare-with-a-constant instruction. This would be 287 // able to handle value inequalities better, for example if the compare is 288 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 289 // Perhaps getConstantOnEdge should be smart enough to do this? 290 291 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 292 BasicBlock *P = *PI; 293 // If the value is known by LazyValueInfo to be a constant in a 294 // predecessor, use that information to try to thread this block. 295 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB); 296 if (PredCst == 0 || 297 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst))) 298 continue; 299 300 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), P)); 301 } 302 303 return !Result.empty(); 304 } 305 306 return false; 307 } 308 309 /// If I is a PHI node, then we know the incoming values for any constants. 310 if (PHINode *PN = dyn_cast<PHINode>(I)) { 311 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 312 Value *InVal = PN->getIncomingValue(i); 313 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 314 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 315 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 316 } 317 } 318 return !Result.empty(); 319 } 320 321 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 322 323 // Handle some boolean conditions. 324 if (I->getType()->getPrimitiveSizeInBits() == 1) { 325 // X | true -> true 326 // X & false -> false 327 if (I->getOpcode() == Instruction::Or || 328 I->getOpcode() == Instruction::And) { 329 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 330 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 331 332 if (LHSVals.empty() && RHSVals.empty()) 333 return false; 334 335 ConstantInt *InterestingVal; 336 if (I->getOpcode() == Instruction::Or) 337 InterestingVal = ConstantInt::getTrue(I->getContext()); 338 else 339 InterestingVal = ConstantInt::getFalse(I->getContext()); 340 341 // Scan for the sentinel. If we find an undef, force it to the 342 // interesting value: x|undef -> true and x&undef -> false. 343 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 344 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) { 345 Result.push_back(LHSVals[i]); 346 Result.back().first = InterestingVal; 347 } 348 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 349 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) { 350 // If we already inferred a value for this block on the LHS, don't 351 // re-add it. 352 bool HasValue = false; 353 for (unsigned r = 0, e = Result.size(); r != e; ++r) 354 if (Result[r].second == RHSVals[i].second) { 355 HasValue = true; 356 break; 357 } 358 359 if (!HasValue) { 360 Result.push_back(RHSVals[i]); 361 Result.back().first = InterestingVal; 362 } 363 } 364 return !Result.empty(); 365 } 366 367 // Handle the NOT form of XOR. 368 if (I->getOpcode() == Instruction::Xor && 369 isa<ConstantInt>(I->getOperand(1)) && 370 cast<ConstantInt>(I->getOperand(1))->isOne()) { 371 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result); 372 if (Result.empty()) 373 return false; 374 375 // Invert the known values. 376 for (unsigned i = 0, e = Result.size(); i != e; ++i) 377 if (Result[i].first) 378 Result[i].first = 379 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first)); 380 return true; 381 } 382 } 383 384 // Handle compare with phi operand, where the PHI is defined in this block. 385 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 386 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 387 if (PN && PN->getParent() == BB) { 388 // We can do this simplification if any comparisons fold to true or false. 389 // See if any do. 390 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 391 BasicBlock *PredBB = PN->getIncomingBlock(i); 392 Value *LHS = PN->getIncomingValue(i); 393 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 394 395 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD); 396 if (Res == 0) { 397 if (!LVI || !isa<Constant>(RHS)) 398 continue; 399 400 LazyValueInfo::Tristate 401 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 402 cast<Constant>(RHS), PredBB, BB); 403 if (ResT == LazyValueInfo::Unknown) 404 continue; 405 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 406 } 407 408 if (isa<UndefValue>(Res)) 409 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 410 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 411 Result.push_back(std::make_pair(CI, PredBB)); 412 } 413 414 return !Result.empty(); 415 } 416 417 418 // If comparing a live-in value against a constant, see if we know the 419 // live-in value on any predecessors. 420 if (LVI && isa<Constant>(Cmp->getOperand(1)) && 421 Cmp->getType()->isIntegerTy() && // Not vector compare. 422 (!isa<Instruction>(Cmp->getOperand(0)) || 423 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB)) { 424 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 425 426 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 427 BasicBlock *P = *PI; 428 // If the value is known by LazyValueInfo to be a constant in a 429 // predecessor, use that information to try to thread this block. 430 LazyValueInfo::Tristate 431 Res = LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 432 RHSCst, P, BB); 433 if (Res == LazyValueInfo::Unknown) 434 continue; 435 436 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 437 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), P)); 438 } 439 440 return !Result.empty(); 441 } 442 } 443 return false; 444} 445 446 447 448/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 449/// in an undefined jump, decide which block is best to revector to. 450/// 451/// Since we can pick an arbitrary destination, we pick the successor with the 452/// fewest predecessors. This should reduce the in-degree of the others. 453/// 454static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 455 TerminatorInst *BBTerm = BB->getTerminator(); 456 unsigned MinSucc = 0; 457 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 458 // Compute the successor with the minimum number of predecessors. 459 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 460 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 461 TestBB = BBTerm->getSuccessor(i); 462 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 463 if (NumPreds < MinNumPreds) 464 MinSucc = i; 465 } 466 467 return MinSucc; 468} 469 470/// ProcessBlock - If there are any predecessors whose control can be threaded 471/// through to a successor, transform them now. 472bool JumpThreading::ProcessBlock(BasicBlock *BB) { 473 // If the block is trivially dead, just return and let the caller nuke it. 474 // This simplifies other transformations. 475 if (pred_begin(BB) == pred_end(BB) && 476 BB != &BB->getParent()->getEntryBlock()) 477 return false; 478 479 // If this block has a single predecessor, and if that pred has a single 480 // successor, merge the blocks. This encourages recursive jump threading 481 // because now the condition in this block can be threaded through 482 // predecessors of our predecessor block. 483 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 484 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 485 SinglePred != BB) { 486 // If SinglePred was a loop header, BB becomes one. 487 if (LoopHeaders.erase(SinglePred)) 488 LoopHeaders.insert(BB); 489 490 // Remember if SinglePred was the entry block of the function. If so, we 491 // will need to move BB back to the entry position. 492 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 493 MergeBasicBlockIntoOnlyPred(BB); 494 495 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 496 BB->moveBefore(&BB->getParent()->getEntryBlock()); 497 return true; 498 } 499 } 500 501 // Look to see if the terminator is a branch of switch, if not we can't thread 502 // it. 503 Value *Condition; 504 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 505 // Can't thread an unconditional jump. 506 if (BI->isUnconditional()) return false; 507 Condition = BI->getCondition(); 508 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 509 Condition = SI->getCondition(); 510 else 511 return false; // Must be an invoke. 512 513 // If the terminator of this block is branching on a constant, simplify the 514 // terminator to an unconditional branch. This can occur due to threading in 515 // other blocks. 516 if (isa<ConstantInt>(Condition)) { 517 DEBUG(dbgs() << " In block '" << BB->getName() 518 << "' folding terminator: " << *BB->getTerminator() << '\n'); 519 ++NumFolds; 520 ConstantFoldTerminator(BB); 521 return true; 522 } 523 524 // If the terminator is branching on an undef, we can pick any of the 525 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 526 if (isa<UndefValue>(Condition)) { 527 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 528 529 // Fold the branch/switch. 530 TerminatorInst *BBTerm = BB->getTerminator(); 531 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 532 if (i == BestSucc) continue; 533 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD); 534 } 535 536 DEBUG(dbgs() << " In block '" << BB->getName() 537 << "' folding undef terminator: " << *BBTerm << '\n'); 538 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 539 BBTerm->eraseFromParent(); 540 return true; 541 } 542 543 Instruction *CondInst = dyn_cast<Instruction>(Condition); 544 545 // If the condition is an instruction defined in another block, see if a 546 // predecessor has the same condition: 547 // br COND, BBX, BBY 548 // BBX: 549 // br COND, BBZ, BBW 550 if (!LVI && 551 !Condition->hasOneUse() && // Multiple uses. 552 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 553 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 554 if (isa<BranchInst>(BB->getTerminator())) { 555 for (; PI != E; ++PI) { 556 BasicBlock *P = *PI; 557 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator())) 558 if (PBI->isConditional() && PBI->getCondition() == Condition && 559 ProcessBranchOnDuplicateCond(P, BB)) 560 return true; 561 } 562 } else { 563 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 564 for (; PI != E; ++PI) { 565 BasicBlock *P = *PI; 566 if (SwitchInst *PSI = dyn_cast<SwitchInst>(P->getTerminator())) 567 if (PSI->getCondition() == Condition && 568 ProcessSwitchOnDuplicateCond(P, BB)) 569 return true; 570 } 571 } 572 } 573 574 // All the rest of our checks depend on the condition being an instruction. 575 if (CondInst == 0) { 576 // FIXME: Unify this with code below. 577 if (LVI && ProcessThreadableEdges(Condition, BB)) 578 return true; 579 return false; 580 } 581 582 583 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 584 if (!LVI && 585 (!isa<PHINode>(CondCmp->getOperand(0)) || 586 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) { 587 // If we have a comparison, loop over the predecessors to see if there is 588 // a condition with a lexically identical value. 589 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 590 for (; PI != E; ++PI) { 591 BasicBlock *P = *PI; 592 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator())) 593 if (PBI->isConditional() && P != BB) { 594 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 595 if (CI->getOperand(0) == CondCmp->getOperand(0) && 596 CI->getOperand(1) == CondCmp->getOperand(1) && 597 CI->getPredicate() == CondCmp->getPredicate()) { 598 // TODO: Could handle things like (x != 4) --> (x == 17) 599 if (ProcessBranchOnDuplicateCond(P, BB)) 600 return true; 601 } 602 } 603 } 604 } 605 } 606 } 607 608 // Check for some cases that are worth simplifying. Right now we want to look 609 // for loads that are used by a switch or by the condition for the branch. If 610 // we see one, check to see if it's partially redundant. If so, insert a PHI 611 // which can then be used to thread the values. 612 // 613 Value *SimplifyValue = CondInst; 614 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 615 if (isa<Constant>(CondCmp->getOperand(1))) 616 SimplifyValue = CondCmp->getOperand(0); 617 618 // TODO: There are other places where load PRE would be profitable, such as 619 // more complex comparisons. 620 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 621 if (SimplifyPartiallyRedundantLoad(LI)) 622 return true; 623 624 625 // Handle a variety of cases where we are branching on something derived from 626 // a PHI node in the current block. If we can prove that any predecessors 627 // compute a predictable value based on a PHI node, thread those predecessors. 628 // 629 if (ProcessThreadableEdges(CondInst, BB)) 630 return true; 631 632 // If this is an otherwise-unfoldable branch on a phi node in the current 633 // block, see if we can simplify. 634 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 635 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 636 return ProcessBranchOnPHI(PN); 637 638 639 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 640 if (CondInst->getOpcode() == Instruction::Xor && 641 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 642 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 643 644 645 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 646 // "(X == 4)", thread through this block. 647 648 return false; 649} 650 651/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 652/// block that jump on exactly the same condition. This means that we almost 653/// always know the direction of the edge in the DESTBB: 654/// PREDBB: 655/// br COND, DESTBB, BBY 656/// DESTBB: 657/// br COND, BBZ, BBW 658/// 659/// If DESTBB has multiple predecessors, we can't just constant fold the branch 660/// in DESTBB, we have to thread over it. 661bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 662 BasicBlock *BB) { 663 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 664 665 // If both successors of PredBB go to DESTBB, we don't know anything. We can 666 // fold the branch to an unconditional one, which allows other recursive 667 // simplifications. 668 bool BranchDir; 669 if (PredBI->getSuccessor(1) != BB) 670 BranchDir = true; 671 else if (PredBI->getSuccessor(0) != BB) 672 BranchDir = false; 673 else { 674 DEBUG(dbgs() << " In block '" << PredBB->getName() 675 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 676 ++NumFolds; 677 ConstantFoldTerminator(PredBB); 678 return true; 679 } 680 681 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 682 683 // If the dest block has one predecessor, just fix the branch condition to a 684 // constant and fold it. 685 if (BB->getSinglePredecessor()) { 686 DEBUG(dbgs() << " In block '" << BB->getName() 687 << "' folding condition to '" << BranchDir << "': " 688 << *BB->getTerminator() << '\n'); 689 ++NumFolds; 690 Value *OldCond = DestBI->getCondition(); 691 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 692 BranchDir)); 693 // Delete dead instructions before we fold the branch. Folding the branch 694 // can eliminate edges from the CFG which can end up deleting OldCond. 695 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 696 ConstantFoldTerminator(BB); 697 return true; 698 } 699 700 701 // Next, figure out which successor we are threading to. 702 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 703 704 SmallVector<BasicBlock*, 2> Preds; 705 Preds.push_back(PredBB); 706 707 // Ok, try to thread it! 708 return ThreadEdge(BB, Preds, SuccBB); 709} 710 711/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 712/// block that switch on exactly the same condition. This means that we almost 713/// always know the direction of the edge in the DESTBB: 714/// PREDBB: 715/// switch COND [... DESTBB, BBY ... ] 716/// DESTBB: 717/// switch COND [... BBZ, BBW ] 718/// 719/// Optimizing switches like this is very important, because simplifycfg builds 720/// switches out of repeated 'if' conditions. 721bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 722 BasicBlock *DestBB) { 723 // Can't thread edge to self. 724 if (PredBB == DestBB) 725 return false; 726 727 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 728 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 729 730 // There are a variety of optimizations that we can potentially do on these 731 // blocks: we order them from most to least preferable. 732 733 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 734 // directly to their destination. This does not introduce *any* code size 735 // growth. Skip debug info first. 736 BasicBlock::iterator BBI = DestBB->begin(); 737 while (isa<DbgInfoIntrinsic>(BBI)) 738 BBI++; 739 740 // FIXME: Thread if it just contains a PHI. 741 if (isa<SwitchInst>(BBI)) { 742 bool MadeChange = false; 743 // Ignore the default edge for now. 744 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 745 ConstantInt *DestVal = DestSI->getCaseValue(i); 746 BasicBlock *DestSucc = DestSI->getSuccessor(i); 747 748 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 749 // PredSI has an explicit case for it. If so, forward. If it is covered 750 // by the default case, we can't update PredSI. 751 unsigned PredCase = PredSI->findCaseValue(DestVal); 752 if (PredCase == 0) continue; 753 754 // If PredSI doesn't go to DestBB on this value, then it won't reach the 755 // case on this condition. 756 if (PredSI->getSuccessor(PredCase) != DestBB && 757 DestSI->getSuccessor(i) != DestBB) 758 continue; 759 760 // Do not forward this if it already goes to this destination, this would 761 // be an infinite loop. 762 if (PredSI->getSuccessor(PredCase) == DestSucc) 763 continue; 764 765 // Otherwise, we're safe to make the change. Make sure that the edge from 766 // DestSI to DestSucc is not critical and has no PHI nodes. 767 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 768 DEBUG(dbgs() << "THROUGH: " << *DestSI); 769 770 // If the destination has PHI nodes, just split the edge for updating 771 // simplicity. 772 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 773 SplitCriticalEdge(DestSI, i, this); 774 DestSucc = DestSI->getSuccessor(i); 775 } 776 FoldSingleEntryPHINodes(DestSucc); 777 PredSI->setSuccessor(PredCase, DestSucc); 778 MadeChange = true; 779 } 780 781 if (MadeChange) 782 return true; 783 } 784 785 return false; 786} 787 788 789/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 790/// load instruction, eliminate it by replacing it with a PHI node. This is an 791/// important optimization that encourages jump threading, and needs to be run 792/// interlaced with other jump threading tasks. 793bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 794 // Don't hack volatile loads. 795 if (LI->isVolatile()) return false; 796 797 // If the load is defined in a block with exactly one predecessor, it can't be 798 // partially redundant. 799 BasicBlock *LoadBB = LI->getParent(); 800 if (LoadBB->getSinglePredecessor()) 801 return false; 802 803 Value *LoadedPtr = LI->getOperand(0); 804 805 // If the loaded operand is defined in the LoadBB, it can't be available. 806 // TODO: Could do simple PHI translation, that would be fun :) 807 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 808 if (PtrOp->getParent() == LoadBB) 809 return false; 810 811 // Scan a few instructions up from the load, to see if it is obviously live at 812 // the entry to its block. 813 BasicBlock::iterator BBIt = LI; 814 815 if (Value *AvailableVal = 816 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 817 // If the value if the load is locally available within the block, just use 818 // it. This frequently occurs for reg2mem'd allocas. 819 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 820 821 // If the returned value is the load itself, replace with an undef. This can 822 // only happen in dead loops. 823 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 824 LI->replaceAllUsesWith(AvailableVal); 825 LI->eraseFromParent(); 826 return true; 827 } 828 829 // Otherwise, if we scanned the whole block and got to the top of the block, 830 // we know the block is locally transparent to the load. If not, something 831 // might clobber its value. 832 if (BBIt != LoadBB->begin()) 833 return false; 834 835 836 SmallPtrSet<BasicBlock*, 8> PredsScanned; 837 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 838 AvailablePredsTy AvailablePreds; 839 BasicBlock *OneUnavailablePred = 0; 840 841 // If we got here, the loaded value is transparent through to the start of the 842 // block. Check to see if it is available in any of the predecessor blocks. 843 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 844 PI != PE; ++PI) { 845 BasicBlock *PredBB = *PI; 846 847 // If we already scanned this predecessor, skip it. 848 if (!PredsScanned.insert(PredBB)) 849 continue; 850 851 // Scan the predecessor to see if the value is available in the pred. 852 BBIt = PredBB->end(); 853 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 854 if (!PredAvailable) { 855 OneUnavailablePred = PredBB; 856 continue; 857 } 858 859 // If so, this load is partially redundant. Remember this info so that we 860 // can create a PHI node. 861 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 862 } 863 864 // If the loaded value isn't available in any predecessor, it isn't partially 865 // redundant. 866 if (AvailablePreds.empty()) return false; 867 868 // Okay, the loaded value is available in at least one (and maybe all!) 869 // predecessors. If the value is unavailable in more than one unique 870 // predecessor, we want to insert a merge block for those common predecessors. 871 // This ensures that we only have to insert one reload, thus not increasing 872 // code size. 873 BasicBlock *UnavailablePred = 0; 874 875 // If there is exactly one predecessor where the value is unavailable, the 876 // already computed 'OneUnavailablePred' block is it. If it ends in an 877 // unconditional branch, we know that it isn't a critical edge. 878 if (PredsScanned.size() == AvailablePreds.size()+1 && 879 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 880 UnavailablePred = OneUnavailablePred; 881 } else if (PredsScanned.size() != AvailablePreds.size()) { 882 // Otherwise, we had multiple unavailable predecessors or we had a critical 883 // edge from the one. 884 SmallVector<BasicBlock*, 8> PredsToSplit; 885 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 886 887 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 888 AvailablePredSet.insert(AvailablePreds[i].first); 889 890 // Add all the unavailable predecessors to the PredsToSplit list. 891 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 892 PI != PE; ++PI) { 893 BasicBlock *P = *PI; 894 // If the predecessor is an indirect goto, we can't split the edge. 895 if (isa<IndirectBrInst>(P->getTerminator())) 896 return false; 897 898 if (!AvailablePredSet.count(P)) 899 PredsToSplit.push_back(P); 900 } 901 902 // Split them out to their own block. 903 UnavailablePred = 904 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 905 "thread-pre-split", this); 906 } 907 908 // If the value isn't available in all predecessors, then there will be 909 // exactly one where it isn't available. Insert a load on that edge and add 910 // it to the AvailablePreds list. 911 if (UnavailablePred) { 912 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 913 "Can't handle critical edge here!"); 914 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 915 LI->getAlignment(), 916 UnavailablePred->getTerminator()); 917 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 918 } 919 920 // Now we know that each predecessor of this block has a value in 921 // AvailablePreds, sort them for efficient access as we're walking the preds. 922 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 923 924 // Create a PHI node at the start of the block for the PRE'd load value. 925 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 926 PN->takeName(LI); 927 928 // Insert new entries into the PHI for each predecessor. A single block may 929 // have multiple entries here. 930 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 931 ++PI) { 932 BasicBlock *P = *PI; 933 AvailablePredsTy::iterator I = 934 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 935 std::make_pair(P, (Value*)0)); 936 937 assert(I != AvailablePreds.end() && I->first == P && 938 "Didn't find entry for predecessor!"); 939 940 PN->addIncoming(I->second, I->first); 941 } 942 943 //cerr << "PRE: " << *LI << *PN << "\n"; 944 945 LI->replaceAllUsesWith(PN); 946 LI->eraseFromParent(); 947 948 return true; 949} 950 951/// FindMostPopularDest - The specified list contains multiple possible 952/// threadable destinations. Pick the one that occurs the most frequently in 953/// the list. 954static BasicBlock * 955FindMostPopularDest(BasicBlock *BB, 956 const SmallVectorImpl<std::pair<BasicBlock*, 957 BasicBlock*> > &PredToDestList) { 958 assert(!PredToDestList.empty()); 959 960 // Determine popularity. If there are multiple possible destinations, we 961 // explicitly choose to ignore 'undef' destinations. We prefer to thread 962 // blocks with known and real destinations to threading undef. We'll handle 963 // them later if interesting. 964 DenseMap<BasicBlock*, unsigned> DestPopularity; 965 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 966 if (PredToDestList[i].second) 967 DestPopularity[PredToDestList[i].second]++; 968 969 // Find the most popular dest. 970 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 971 BasicBlock *MostPopularDest = DPI->first; 972 unsigned Popularity = DPI->second; 973 SmallVector<BasicBlock*, 4> SamePopularity; 974 975 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 976 // If the popularity of this entry isn't higher than the popularity we've 977 // seen so far, ignore it. 978 if (DPI->second < Popularity) 979 ; // ignore. 980 else if (DPI->second == Popularity) { 981 // If it is the same as what we've seen so far, keep track of it. 982 SamePopularity.push_back(DPI->first); 983 } else { 984 // If it is more popular, remember it. 985 SamePopularity.clear(); 986 MostPopularDest = DPI->first; 987 Popularity = DPI->second; 988 } 989 } 990 991 // Okay, now we know the most popular destination. If there is more than 992 // destination, we need to determine one. This is arbitrary, but we need 993 // to make a deterministic decision. Pick the first one that appears in the 994 // successor list. 995 if (!SamePopularity.empty()) { 996 SamePopularity.push_back(MostPopularDest); 997 TerminatorInst *TI = BB->getTerminator(); 998 for (unsigned i = 0; ; ++i) { 999 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 1000 1001 if (std::find(SamePopularity.begin(), SamePopularity.end(), 1002 TI->getSuccessor(i)) == SamePopularity.end()) 1003 continue; 1004 1005 MostPopularDest = TI->getSuccessor(i); 1006 break; 1007 } 1008 } 1009 1010 // Okay, we have finally picked the most popular destination. 1011 return MostPopularDest; 1012} 1013 1014bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) { 1015 // If threading this would thread across a loop header, don't even try to 1016 // thread the edge. 1017 if (LoopHeaders.count(BB)) 1018 return false; 1019 1020 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 1021 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues)) 1022 return false; 1023 assert(!PredValues.empty() && 1024 "ComputeValueKnownInPredecessors returned true with no values"); 1025 1026 DEBUG(dbgs() << "IN BB: " << *BB; 1027 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1028 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "; 1029 if (PredValues[i].first) 1030 dbgs() << *PredValues[i].first; 1031 else 1032 dbgs() << "UNDEF"; 1033 dbgs() << " for pred '" << PredValues[i].second->getName() 1034 << "'.\n"; 1035 }); 1036 1037 // Decide what we want to thread through. Convert our list of known values to 1038 // a list of known destinations for each pred. This also discards duplicate 1039 // predecessors and keeps track of the undefined inputs (which are represented 1040 // as a null dest in the PredToDestList). 1041 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1042 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1043 1044 BasicBlock *OnlyDest = 0; 1045 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1046 1047 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1048 BasicBlock *Pred = PredValues[i].second; 1049 if (!SeenPreds.insert(Pred)) 1050 continue; // Duplicate predecessor entry. 1051 1052 // If the predecessor ends with an indirect goto, we can't change its 1053 // destination. 1054 if (isa<IndirectBrInst>(Pred->getTerminator())) 1055 continue; 1056 1057 ConstantInt *Val = PredValues[i].first; 1058 1059 BasicBlock *DestBB; 1060 if (Val == 0) // Undef. 1061 DestBB = 0; 1062 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1063 DestBB = BI->getSuccessor(Val->isZero()); 1064 else { 1065 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 1066 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 1067 } 1068 1069 // If we have exactly one destination, remember it for efficiency below. 1070 if (i == 0) 1071 OnlyDest = DestBB; 1072 else if (OnlyDest != DestBB) 1073 OnlyDest = MultipleDestSentinel; 1074 1075 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1076 } 1077 1078 // If all edges were unthreadable, we fail. 1079 if (PredToDestList.empty()) 1080 return false; 1081 1082 // Determine which is the most common successor. If we have many inputs and 1083 // this block is a switch, we want to start by threading the batch that goes 1084 // to the most popular destination first. If we only know about one 1085 // threadable destination (the common case) we can avoid this. 1086 BasicBlock *MostPopularDest = OnlyDest; 1087 1088 if (MostPopularDest == MultipleDestSentinel) 1089 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1090 1091 // Now that we know what the most popular destination is, factor all 1092 // predecessors that will jump to it into a single predecessor. 1093 SmallVector<BasicBlock*, 16> PredsToFactor; 1094 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1095 if (PredToDestList[i].second == MostPopularDest) { 1096 BasicBlock *Pred = PredToDestList[i].first; 1097 1098 // This predecessor may be a switch or something else that has multiple 1099 // edges to the block. Factor each of these edges by listing them 1100 // according to # occurrences in PredsToFactor. 1101 TerminatorInst *PredTI = Pred->getTerminator(); 1102 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1103 if (PredTI->getSuccessor(i) == BB) 1104 PredsToFactor.push_back(Pred); 1105 } 1106 1107 // If the threadable edges are branching on an undefined value, we get to pick 1108 // the destination that these predecessors should get to. 1109 if (MostPopularDest == 0) 1110 MostPopularDest = BB->getTerminator()-> 1111 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1112 1113 // Ok, try to thread it! 1114 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1115} 1116 1117/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1118/// a PHI node in the current block. See if there are any simplifications we 1119/// can do based on inputs to the phi node. 1120/// 1121bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1122 BasicBlock *BB = PN->getParent(); 1123 1124 // TODO: We could make use of this to do it once for blocks with common PHI 1125 // values. 1126 SmallVector<BasicBlock*, 1> PredBBs; 1127 PredBBs.resize(1); 1128 1129 // If any of the predecessor blocks end in an unconditional branch, we can 1130 // *duplicate* the conditional branch into that block in order to further 1131 // encourage jump threading and to eliminate cases where we have branch on a 1132 // phi of an icmp (branch on icmp is much better). 1133 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1134 BasicBlock *PredBB = PN->getIncomingBlock(i); 1135 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1136 if (PredBr->isUnconditional()) { 1137 PredBBs[0] = PredBB; 1138 // Try to duplicate BB into PredBB. 1139 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1140 return true; 1141 } 1142 } 1143 1144 return false; 1145} 1146 1147/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1148/// a xor instruction in the current block. See if there are any 1149/// simplifications we can do based on inputs to the xor. 1150/// 1151bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1152 BasicBlock *BB = BO->getParent(); 1153 1154 // If either the LHS or RHS of the xor is a constant, don't do this 1155 // optimization. 1156 if (isa<ConstantInt>(BO->getOperand(0)) || 1157 isa<ConstantInt>(BO->getOperand(1))) 1158 return false; 1159 1160 // If the first instruction in BB isn't a phi, we won't be able to infer 1161 // anything special about any particular predecessor. 1162 if (!isa<PHINode>(BB->front())) 1163 return false; 1164 1165 // If we have a xor as the branch input to this block, and we know that the 1166 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1167 // the condition into the predecessor and fix that value to true, saving some 1168 // logical ops on that path and encouraging other paths to simplify. 1169 // 1170 // This copies something like this: 1171 // 1172 // BB: 1173 // %X = phi i1 [1], [%X'] 1174 // %Y = icmp eq i32 %A, %B 1175 // %Z = xor i1 %X, %Y 1176 // br i1 %Z, ... 1177 // 1178 // Into: 1179 // BB': 1180 // %Y = icmp ne i32 %A, %B 1181 // br i1 %Z, ... 1182 1183 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues; 1184 bool isLHS = true; 1185 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) { 1186 assert(XorOpValues.empty()); 1187 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues)) 1188 return false; 1189 isLHS = false; 1190 } 1191 1192 assert(!XorOpValues.empty() && 1193 "ComputeValueKnownInPredecessors returned true with no values"); 1194 1195 // Scan the information to see which is most popular: true or false. The 1196 // predecessors can be of the set true, false, or undef. 1197 unsigned NumTrue = 0, NumFalse = 0; 1198 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1199 if (!XorOpValues[i].first) continue; // Ignore undefs for the count. 1200 if (XorOpValues[i].first->isZero()) 1201 ++NumFalse; 1202 else 1203 ++NumTrue; 1204 } 1205 1206 // Determine which value to split on, true, false, or undef if neither. 1207 ConstantInt *SplitVal = 0; 1208 if (NumTrue > NumFalse) 1209 SplitVal = ConstantInt::getTrue(BB->getContext()); 1210 else if (NumTrue != 0 || NumFalse != 0) 1211 SplitVal = ConstantInt::getFalse(BB->getContext()); 1212 1213 // Collect all of the blocks that this can be folded into so that we can 1214 // factor this once and clone it once. 1215 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1216 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1217 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue; 1218 1219 BlocksToFoldInto.push_back(XorOpValues[i].second); 1220 } 1221 1222 // If we inferred a value for all of the predecessors, then duplication won't 1223 // help us. However, we can just replace the LHS or RHS with the constant. 1224 if (BlocksToFoldInto.size() == 1225 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1226 if (SplitVal == 0) { 1227 // If all preds provide undef, just nuke the xor, because it is undef too. 1228 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1229 BO->eraseFromParent(); 1230 } else if (SplitVal->isZero()) { 1231 // If all preds provide 0, replace the xor with the other input. 1232 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1233 BO->eraseFromParent(); 1234 } else { 1235 // If all preds provide 1, set the computed value to 1. 1236 BO->setOperand(!isLHS, SplitVal); 1237 } 1238 1239 return true; 1240 } 1241 1242 // Try to duplicate BB into PredBB. 1243 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1244} 1245 1246 1247/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1248/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1249/// NewPred using the entries from OldPred (suitably mapped). 1250static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1251 BasicBlock *OldPred, 1252 BasicBlock *NewPred, 1253 DenseMap<Instruction*, Value*> &ValueMap) { 1254 for (BasicBlock::iterator PNI = PHIBB->begin(); 1255 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1256 // Ok, we have a PHI node. Figure out what the incoming value was for the 1257 // DestBlock. 1258 Value *IV = PN->getIncomingValueForBlock(OldPred); 1259 1260 // Remap the value if necessary. 1261 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1262 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1263 if (I != ValueMap.end()) 1264 IV = I->second; 1265 } 1266 1267 PN->addIncoming(IV, NewPred); 1268 } 1269} 1270 1271/// ThreadEdge - We have decided that it is safe and profitable to factor the 1272/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1273/// across BB. Transform the IR to reflect this change. 1274bool JumpThreading::ThreadEdge(BasicBlock *BB, 1275 const SmallVectorImpl<BasicBlock*> &PredBBs, 1276 BasicBlock *SuccBB) { 1277 // If threading to the same block as we come from, we would infinite loop. 1278 if (SuccBB == BB) { 1279 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1280 << "' - would thread to self!\n"); 1281 return false; 1282 } 1283 1284 // If threading this would thread across a loop header, don't thread the edge. 1285 // See the comments above FindLoopHeaders for justifications and caveats. 1286 if (LoopHeaders.count(BB)) { 1287 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1288 << "' to dest BB '" << SuccBB->getName() 1289 << "' - it might create an irreducible loop!\n"); 1290 return false; 1291 } 1292 1293 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1294 if (JumpThreadCost > Threshold) { 1295 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1296 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1297 return false; 1298 } 1299 1300 // And finally, do it! Start by factoring the predecessors is needed. 1301 BasicBlock *PredBB; 1302 if (PredBBs.size() == 1) 1303 PredBB = PredBBs[0]; 1304 else { 1305 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1306 << " common predecessors.\n"); 1307 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1308 ".thr_comm", this); 1309 } 1310 1311 // And finally, do it! 1312 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1313 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1314 << ", across block:\n " 1315 << *BB << "\n"); 1316 1317 // We are going to have to map operands from the original BB block to the new 1318 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1319 // account for entry from PredBB. 1320 DenseMap<Instruction*, Value*> ValueMapping; 1321 1322 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1323 BB->getName()+".thread", 1324 BB->getParent(), BB); 1325 NewBB->moveAfter(PredBB); 1326 1327 BasicBlock::iterator BI = BB->begin(); 1328 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1329 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1330 1331 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1332 // mapping and using it to remap operands in the cloned instructions. 1333 for (; !isa<TerminatorInst>(BI); ++BI) { 1334 Instruction *New = BI->clone(); 1335 New->setName(BI->getName()); 1336 NewBB->getInstList().push_back(New); 1337 ValueMapping[BI] = New; 1338 1339 // Remap operands to patch up intra-block references. 1340 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1341 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1342 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1343 if (I != ValueMapping.end()) 1344 New->setOperand(i, I->second); 1345 } 1346 } 1347 1348 // We didn't copy the terminator from BB over to NewBB, because there is now 1349 // an unconditional jump to SuccBB. Insert the unconditional jump. 1350 BranchInst::Create(SuccBB, NewBB); 1351 1352 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1353 // PHI nodes for NewBB now. 1354 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1355 1356 // If there were values defined in BB that are used outside the block, then we 1357 // now have to update all uses of the value to use either the original value, 1358 // the cloned value, or some PHI derived value. This can require arbitrary 1359 // PHI insertion, of which we are prepared to do, clean these up now. 1360 SSAUpdater SSAUpdate; 1361 SmallVector<Use*, 16> UsesToRename; 1362 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1363 // Scan all uses of this instruction to see if it is used outside of its 1364 // block, and if so, record them in UsesToRename. 1365 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1366 ++UI) { 1367 Instruction *User = cast<Instruction>(*UI); 1368 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1369 if (UserPN->getIncomingBlock(UI) == BB) 1370 continue; 1371 } else if (User->getParent() == BB) 1372 continue; 1373 1374 UsesToRename.push_back(&UI.getUse()); 1375 } 1376 1377 // If there are no uses outside the block, we're done with this instruction. 1378 if (UsesToRename.empty()) 1379 continue; 1380 1381 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1382 1383 // We found a use of I outside of BB. Rename all uses of I that are outside 1384 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1385 // with the two values we know. 1386 SSAUpdate.Initialize(I); 1387 SSAUpdate.AddAvailableValue(BB, I); 1388 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1389 1390 while (!UsesToRename.empty()) 1391 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1392 DEBUG(dbgs() << "\n"); 1393 } 1394 1395 1396 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1397 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1398 // us to simplify any PHI nodes in BB. 1399 TerminatorInst *PredTerm = PredBB->getTerminator(); 1400 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1401 if (PredTerm->getSuccessor(i) == BB) { 1402 RemovePredecessorAndSimplify(BB, PredBB, TD); 1403 PredTerm->setSuccessor(i, NewBB); 1404 } 1405 1406 // At this point, the IR is fully up to date and consistent. Do a quick scan 1407 // over the new instructions and zap any that are constants or dead. This 1408 // frequently happens because of phi translation. 1409 SimplifyInstructionsInBlock(NewBB, TD); 1410 1411 // Threaded an edge! 1412 ++NumThreads; 1413 return true; 1414} 1415 1416/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1417/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1418/// If we can duplicate the contents of BB up into PredBB do so now, this 1419/// improves the odds that the branch will be on an analyzable instruction like 1420/// a compare. 1421bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1422 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1423 assert(!PredBBs.empty() && "Can't handle an empty set"); 1424 1425 // If BB is a loop header, then duplicating this block outside the loop would 1426 // cause us to transform this into an irreducible loop, don't do this. 1427 // See the comments above FindLoopHeaders for justifications and caveats. 1428 if (LoopHeaders.count(BB)) { 1429 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1430 << "' into predecessor block '" << PredBBs[0]->getName() 1431 << "' - it might create an irreducible loop!\n"); 1432 return false; 1433 } 1434 1435 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1436 if (DuplicationCost > Threshold) { 1437 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1438 << "' - Cost is too high: " << DuplicationCost << "\n"); 1439 return false; 1440 } 1441 1442 // And finally, do it! Start by factoring the predecessors is needed. 1443 BasicBlock *PredBB; 1444 if (PredBBs.size() == 1) 1445 PredBB = PredBBs[0]; 1446 else { 1447 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1448 << " common predecessors.\n"); 1449 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1450 ".thr_comm", this); 1451 } 1452 1453 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1454 // of PredBB. 1455 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1456 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1457 << DuplicationCost << " block is:" << *BB << "\n"); 1458 1459 // Unless PredBB ends with an unconditional branch, split the edge so that we 1460 // can just clone the bits from BB into the end of the new PredBB. 1461 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1462 1463 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { 1464 PredBB = SplitEdge(PredBB, BB, this); 1465 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1466 } 1467 1468 // We are going to have to map operands from the original BB block into the 1469 // PredBB block. Evaluate PHI nodes in BB. 1470 DenseMap<Instruction*, Value*> ValueMapping; 1471 1472 BasicBlock::iterator BI = BB->begin(); 1473 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1474 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1475 1476 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1477 // mapping and using it to remap operands in the cloned instructions. 1478 for (; BI != BB->end(); ++BI) { 1479 Instruction *New = BI->clone(); 1480 1481 // Remap operands to patch up intra-block references. 1482 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1483 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1484 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1485 if (I != ValueMapping.end()) 1486 New->setOperand(i, I->second); 1487 } 1488 1489 // If this instruction can be simplified after the operands are updated, 1490 // just use the simplified value instead. This frequently happens due to 1491 // phi translation. 1492 if (Value *IV = SimplifyInstruction(New, TD)) { 1493 delete New; 1494 ValueMapping[BI] = IV; 1495 } else { 1496 // Otherwise, insert the new instruction into the block. 1497 New->setName(BI->getName()); 1498 PredBB->getInstList().insert(OldPredBranch, New); 1499 ValueMapping[BI] = New; 1500 } 1501 } 1502 1503 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1504 // add entries to the PHI nodes for branch from PredBB now. 1505 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1506 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1507 ValueMapping); 1508 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1509 ValueMapping); 1510 1511 // If there were values defined in BB that are used outside the block, then we 1512 // now have to update all uses of the value to use either the original value, 1513 // the cloned value, or some PHI derived value. This can require arbitrary 1514 // PHI insertion, of which we are prepared to do, clean these up now. 1515 SSAUpdater SSAUpdate; 1516 SmallVector<Use*, 16> UsesToRename; 1517 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1518 // Scan all uses of this instruction to see if it is used outside of its 1519 // block, and if so, record them in UsesToRename. 1520 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1521 ++UI) { 1522 Instruction *User = cast<Instruction>(*UI); 1523 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1524 if (UserPN->getIncomingBlock(UI) == BB) 1525 continue; 1526 } else if (User->getParent() == BB) 1527 continue; 1528 1529 UsesToRename.push_back(&UI.getUse()); 1530 } 1531 1532 // If there are no uses outside the block, we're done with this instruction. 1533 if (UsesToRename.empty()) 1534 continue; 1535 1536 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1537 1538 // We found a use of I outside of BB. Rename all uses of I that are outside 1539 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1540 // with the two values we know. 1541 SSAUpdate.Initialize(I); 1542 SSAUpdate.AddAvailableValue(BB, I); 1543 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1544 1545 while (!UsesToRename.empty()) 1546 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1547 DEBUG(dbgs() << "\n"); 1548 } 1549 1550 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1551 // that we nuked. 1552 RemovePredecessorAndSimplify(BB, PredBB, TD); 1553 1554 // Remove the unconditional branch at the end of the PredBB block. 1555 OldPredBranch->eraseFromParent(); 1556 1557 ++NumDupes; 1558 return true; 1559} 1560 1561 1562