JumpThreading.cpp revision 1e452650c6044735b6aa922d41736bda5721adcc
1//===- JumpThreading.cpp - Thread control through conditional blocks ------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements the Jump Threading pass. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "jump-threading" 15#include "llvm/Transforms/Scalar.h" 16#include "llvm/IntrinsicInst.h" 17#include "llvm/LLVMContext.h" 18#include "llvm/Pass.h" 19#include "llvm/Analysis/InstructionSimplify.h" 20#include "llvm/Analysis/LazyValueInfo.h" 21#include "llvm/Transforms/Utils/BasicBlockUtils.h" 22#include "llvm/Transforms/Utils/Local.h" 23#include "llvm/Transforms/Utils/SSAUpdater.h" 24#include "llvm/Target/TargetData.h" 25#include "llvm/ADT/DenseMap.h" 26#include "llvm/ADT/Statistic.h" 27#include "llvm/ADT/STLExtras.h" 28#include "llvm/ADT/SmallPtrSet.h" 29#include "llvm/ADT/SmallSet.h" 30#include "llvm/Support/CommandLine.h" 31#include "llvm/Support/Debug.h" 32#include "llvm/Support/ValueHandle.h" 33#include "llvm/Support/raw_ostream.h" 34using namespace llvm; 35 36STATISTIC(NumThreads, "Number of jumps threaded"); 37STATISTIC(NumFolds, "Number of terminators folded"); 38STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 39 40static cl::opt<unsigned> 41Threshold("jump-threading-threshold", 42 cl::desc("Max block size to duplicate for jump threading"), 43 cl::init(6), cl::Hidden); 44 45// Turn on use of LazyValueInfo. 46static cl::opt<bool> 47EnableLVI("enable-jump-threading-lvi", cl::ReallyHidden); 48 49 50 51namespace { 52 /// This pass performs 'jump threading', which looks at blocks that have 53 /// multiple predecessors and multiple successors. If one or more of the 54 /// predecessors of the block can be proven to always jump to one of the 55 /// successors, we forward the edge from the predecessor to the successor by 56 /// duplicating the contents of this block. 57 /// 58 /// An example of when this can occur is code like this: 59 /// 60 /// if () { ... 61 /// X = 4; 62 /// } 63 /// if (X < 3) { 64 /// 65 /// In this case, the unconditional branch at the end of the first if can be 66 /// revectored to the false side of the second if. 67 /// 68 class JumpThreading : public FunctionPass { 69 TargetData *TD; 70 LazyValueInfo *LVI; 71#ifdef NDEBUG 72 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 73#else 74 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 75#endif 76 public: 77 static char ID; // Pass identification 78 JumpThreading() : FunctionPass(&ID) {} 79 80 bool runOnFunction(Function &F); 81 82 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 83 if (EnableLVI) 84 AU.addRequired<LazyValueInfo>(); 85 } 86 87 void FindLoopHeaders(Function &F); 88 bool ProcessBlock(BasicBlock *BB); 89 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 90 BasicBlock *SuccBB); 91 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 92 const SmallVectorImpl<BasicBlock *> &PredBBs); 93 94 typedef SmallVectorImpl<std::pair<ConstantInt*, 95 BasicBlock*> > PredValueInfo; 96 97 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 98 PredValueInfo &Result); 99 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB); 100 101 102 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 103 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 104 105 bool ProcessBranchOnPHI(PHINode *PN); 106 bool ProcessBranchOnXOR(BinaryOperator *BO); 107 108 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 109 }; 110} 111 112char JumpThreading::ID = 0; 113static RegisterPass<JumpThreading> 114X("jump-threading", "Jump Threading"); 115 116// Public interface to the Jump Threading pass 117FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 118 119/// runOnFunction - Top level algorithm. 120/// 121bool JumpThreading::runOnFunction(Function &F) { 122 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 123 TD = getAnalysisIfAvailable<TargetData>(); 124 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0; 125 126 FindLoopHeaders(F); 127 128 bool Changed, EverChanged = false; 129 do { 130 Changed = false; 131 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 132 BasicBlock *BB = I; 133 // Thread all of the branches we can over this block. 134 while (ProcessBlock(BB)) 135 Changed = true; 136 137 ++I; 138 139 // If the block is trivially dead, zap it. This eliminates the successor 140 // edges which simplifies the CFG. 141 if (pred_begin(BB) == pred_end(BB) && 142 BB != &BB->getParent()->getEntryBlock()) { 143 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 144 << "' with terminator: " << *BB->getTerminator() << '\n'); 145 LoopHeaders.erase(BB); 146 DeleteDeadBlock(BB); 147 Changed = true; 148 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 149 // Can't thread an unconditional jump, but if the block is "almost 150 // empty", we can replace uses of it with uses of the successor and make 151 // this dead. 152 if (BI->isUnconditional() && 153 BB != &BB->getParent()->getEntryBlock()) { 154 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 155 // Ignore dbg intrinsics. 156 while (isa<DbgInfoIntrinsic>(BBI)) 157 ++BBI; 158 // If the terminator is the only non-phi instruction, try to nuke it. 159 if (BBI->isTerminator()) { 160 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 161 // block, we have to make sure it isn't in the LoopHeaders set. We 162 // reinsert afterward if needed. 163 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 164 BasicBlock *Succ = BI->getSuccessor(0); 165 166 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 167 Changed = true; 168 // If we deleted BB and BB was the header of a loop, then the 169 // successor is now the header of the loop. 170 BB = Succ; 171 } 172 173 if (ErasedFromLoopHeaders) 174 LoopHeaders.insert(BB); 175 } 176 } 177 } 178 } 179 EverChanged |= Changed; 180 } while (Changed); 181 182 LoopHeaders.clear(); 183 return EverChanged; 184} 185 186/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 187/// thread across it. 188static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 189 /// Ignore PHI nodes, these will be flattened when duplication happens. 190 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 191 192 // FIXME: THREADING will delete values that are just used to compute the 193 // branch, so they shouldn't count against the duplication cost. 194 195 196 // Sum up the cost of each instruction until we get to the terminator. Don't 197 // include the terminator because the copy won't include it. 198 unsigned Size = 0; 199 for (; !isa<TerminatorInst>(I); ++I) { 200 // Debugger intrinsics don't incur code size. 201 if (isa<DbgInfoIntrinsic>(I)) continue; 202 203 // If this is a pointer->pointer bitcast, it is free. 204 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType())) 205 continue; 206 207 // All other instructions count for at least one unit. 208 ++Size; 209 210 // Calls are more expensive. If they are non-intrinsic calls, we model them 211 // as having cost of 4. If they are a non-vector intrinsic, we model them 212 // as having cost of 2 total, and if they are a vector intrinsic, we model 213 // them as having cost 1. 214 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 215 if (!isa<IntrinsicInst>(CI)) 216 Size += 3; 217 else if (!isa<VectorType>(CI->getType())) 218 Size += 1; 219 } 220 } 221 222 // Threading through a switch statement is particularly profitable. If this 223 // block ends in a switch, decrease its cost to make it more likely to happen. 224 if (isa<SwitchInst>(I)) 225 Size = Size > 6 ? Size-6 : 0; 226 227 return Size; 228} 229 230/// FindLoopHeaders - We do not want jump threading to turn proper loop 231/// structures into irreducible loops. Doing this breaks up the loop nesting 232/// hierarchy and pessimizes later transformations. To prevent this from 233/// happening, we first have to find the loop headers. Here we approximate this 234/// by finding targets of backedges in the CFG. 235/// 236/// Note that there definitely are cases when we want to allow threading of 237/// edges across a loop header. For example, threading a jump from outside the 238/// loop (the preheader) to an exit block of the loop is definitely profitable. 239/// It is also almost always profitable to thread backedges from within the loop 240/// to exit blocks, and is often profitable to thread backedges to other blocks 241/// within the loop (forming a nested loop). This simple analysis is not rich 242/// enough to track all of these properties and keep it up-to-date as the CFG 243/// mutates, so we don't allow any of these transformations. 244/// 245void JumpThreading::FindLoopHeaders(Function &F) { 246 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 247 FindFunctionBackedges(F, Edges); 248 249 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 250 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 251} 252 253/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 254/// if we can infer that the value is a known ConstantInt in any of our 255/// predecessors. If so, return the known list of value and pred BB in the 256/// result vector. If a value is known to be undef, it is returned as null. 257/// 258/// This returns true if there were any known values. 259/// 260bool JumpThreading:: 261ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 262 // If V is a constantint, then it is known in all predecessors. 263 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 264 ConstantInt *CI = dyn_cast<ConstantInt>(V); 265 266 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 267 Result.push_back(std::make_pair(CI, *PI)); 268 return true; 269 } 270 271 // If V is a non-instruction value, or an instruction in a different block, 272 // then it can't be derived from a PHI. 273 Instruction *I = dyn_cast<Instruction>(V); 274 if (I == 0 || I->getParent() != BB) { 275 276 // Okay, if this is a live-in value, see if it has a known value at the end 277 // of any of our predecessors. 278 // 279 // FIXME: This should be an edge property, not a block end property. 280 /// TODO: Per PR2563, we could infer value range information about a 281 /// predecessor based on its terminator. 282 // 283 if (LVI) { 284 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 285 // "I" is a non-local compare-with-a-constant instruction. This would be 286 // able to handle value inequalities better, for example if the compare is 287 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 288 // Perhaps getConstantOnEdge should be smart enough to do this? 289 290 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 291 // If the value is known by LazyValueInfo to be a constant in a 292 // predecessor, use that information to try to thread this block. 293 Constant *PredCst = LVI->getConstantOnEdge(V, *PI, BB); 294 if (PredCst == 0 || 295 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst))) 296 continue; 297 298 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), *PI)); 299 } 300 301 return !Result.empty(); 302 } 303 304 return false; 305 } 306 307 /// If I is a PHI node, then we know the incoming values for any constants. 308 if (PHINode *PN = dyn_cast<PHINode>(I)) { 309 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 310 Value *InVal = PN->getIncomingValue(i); 311 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 312 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 313 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 314 } 315 } 316 return !Result.empty(); 317 } 318 319 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 320 321 // Handle some boolean conditions. 322 if (I->getType()->getPrimitiveSizeInBits() == 1) { 323 // X | true -> true 324 // X & false -> false 325 if (I->getOpcode() == Instruction::Or || 326 I->getOpcode() == Instruction::And) { 327 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 328 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 329 330 if (LHSVals.empty() && RHSVals.empty()) 331 return false; 332 333 ConstantInt *InterestingVal; 334 if (I->getOpcode() == Instruction::Or) 335 InterestingVal = ConstantInt::getTrue(I->getContext()); 336 else 337 InterestingVal = ConstantInt::getFalse(I->getContext()); 338 339 // Scan for the sentinel. If we find an undef, force it to the 340 // interesting value: x|undef -> true and x&undef -> false. 341 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 342 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) { 343 Result.push_back(LHSVals[i]); 344 Result.back().first = InterestingVal; 345 } 346 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 347 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) { 348 Result.push_back(RHSVals[i]); 349 Result.back().first = InterestingVal; 350 } 351 return !Result.empty(); 352 } 353 354 // Handle the NOT form of XOR. 355 if (I->getOpcode() == Instruction::Xor && 356 isa<ConstantInt>(I->getOperand(1)) && 357 cast<ConstantInt>(I->getOperand(1))->isOne()) { 358 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result); 359 if (Result.empty()) 360 return false; 361 362 // Invert the known values. 363 for (unsigned i = 0, e = Result.size(); i != e; ++i) 364 if (Result[i].first) 365 Result[i].first = 366 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first)); 367 return true; 368 } 369 } 370 371 // Handle compare with phi operand, where the PHI is defined in this block. 372 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 373 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 374 if (PN && PN->getParent() == BB) { 375 // We can do this simplification if any comparisons fold to true or false. 376 // See if any do. 377 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 378 BasicBlock *PredBB = PN->getIncomingBlock(i); 379 Value *LHS = PN->getIncomingValue(i); 380 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 381 382 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD); 383 if (Res == 0) { 384 if (!LVI || !isa<Constant>(RHS)) 385 continue; 386 387 LazyValueInfo::Tristate 388 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 389 cast<Constant>(RHS), PredBB, BB); 390 if (ResT == LazyValueInfo::Unknown) 391 continue; 392 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 393 } 394 395 if (isa<UndefValue>(Res)) 396 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 397 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 398 Result.push_back(std::make_pair(CI, PredBB)); 399 } 400 401 return !Result.empty(); 402 } 403 404 405 // If comparing a live-in value against a constant, see if we know the 406 // live-in value on any predecessors. 407 if (LVI && isa<Constant>(Cmp->getOperand(1)) && 408 Cmp->getType()->isInteger() && // Not vector compare. 409 (!isa<Instruction>(Cmp->getOperand(0)) || 410 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB)) { 411 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 412 413 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 414 // If the value is known by LazyValueInfo to be a constant in a 415 // predecessor, use that information to try to thread this block. 416 LazyValueInfo::Tristate 417 Res = LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 418 RHSCst, *PI, BB); 419 if (Res == LazyValueInfo::Unknown) 420 continue; 421 422 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 423 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), *PI)); 424 } 425 426 return !Result.empty(); 427 } 428 } 429 return false; 430} 431 432 433 434/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 435/// in an undefined jump, decide which block is best to revector to. 436/// 437/// Since we can pick an arbitrary destination, we pick the successor with the 438/// fewest predecessors. This should reduce the in-degree of the others. 439/// 440static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 441 TerminatorInst *BBTerm = BB->getTerminator(); 442 unsigned MinSucc = 0; 443 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 444 // Compute the successor with the minimum number of predecessors. 445 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 446 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 447 TestBB = BBTerm->getSuccessor(i); 448 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 449 if (NumPreds < MinNumPreds) 450 MinSucc = i; 451 } 452 453 return MinSucc; 454} 455 456/// ProcessBlock - If there are any predecessors whose control can be threaded 457/// through to a successor, transform them now. 458bool JumpThreading::ProcessBlock(BasicBlock *BB) { 459 // If the block is trivially dead, just return and let the caller nuke it. 460 // This simplifies other transformations. 461 if (pred_begin(BB) == pred_end(BB) && 462 BB != &BB->getParent()->getEntryBlock()) 463 return false; 464 465 // If this block has a single predecessor, and if that pred has a single 466 // successor, merge the blocks. This encourages recursive jump threading 467 // because now the condition in this block can be threaded through 468 // predecessors of our predecessor block. 469 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 470 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 471 SinglePred != BB) { 472 // If SinglePred was a loop header, BB becomes one. 473 if (LoopHeaders.erase(SinglePred)) 474 LoopHeaders.insert(BB); 475 476 // Remember if SinglePred was the entry block of the function. If so, we 477 // will need to move BB back to the entry position. 478 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 479 MergeBasicBlockIntoOnlyPred(BB); 480 481 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 482 BB->moveBefore(&BB->getParent()->getEntryBlock()); 483 return true; 484 } 485 } 486 487 // Look to see if the terminator is a branch of switch, if not we can't thread 488 // it. 489 Value *Condition; 490 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 491 // Can't thread an unconditional jump. 492 if (BI->isUnconditional()) return false; 493 Condition = BI->getCondition(); 494 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 495 Condition = SI->getCondition(); 496 else 497 return false; // Must be an invoke. 498 499 // If the terminator of this block is branching on a constant, simplify the 500 // terminator to an unconditional branch. This can occur due to threading in 501 // other blocks. 502 if (isa<ConstantInt>(Condition)) { 503 DEBUG(dbgs() << " In block '" << BB->getName() 504 << "' folding terminator: " << *BB->getTerminator() << '\n'); 505 ++NumFolds; 506 ConstantFoldTerminator(BB); 507 return true; 508 } 509 510 // If the terminator is branching on an undef, we can pick any of the 511 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 512 if (isa<UndefValue>(Condition)) { 513 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 514 515 // Fold the branch/switch. 516 TerminatorInst *BBTerm = BB->getTerminator(); 517 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 518 if (i == BestSucc) continue; 519 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD); 520 } 521 522 DEBUG(dbgs() << " In block '" << BB->getName() 523 << "' folding undef terminator: " << *BBTerm << '\n'); 524 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 525 BBTerm->eraseFromParent(); 526 return true; 527 } 528 529 Instruction *CondInst = dyn_cast<Instruction>(Condition); 530 531 // If the condition is an instruction defined in another block, see if a 532 // predecessor has the same condition: 533 // br COND, BBX, BBY 534 // BBX: 535 // br COND, BBZ, BBW 536 if (!LVI && 537 !Condition->hasOneUse() && // Multiple uses. 538 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 539 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 540 if (isa<BranchInst>(BB->getTerminator())) { 541 for (; PI != E; ++PI) 542 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 543 if (PBI->isConditional() && PBI->getCondition() == Condition && 544 ProcessBranchOnDuplicateCond(*PI, BB)) 545 return true; 546 } else { 547 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 548 for (; PI != E; ++PI) 549 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator())) 550 if (PSI->getCondition() == Condition && 551 ProcessSwitchOnDuplicateCond(*PI, BB)) 552 return true; 553 } 554 } 555 556 // All the rest of our checks depend on the condition being an instruction. 557 if (CondInst == 0) { 558 // FIXME: Unify this with code below. 559 if (LVI && ProcessThreadableEdges(Condition, BB)) 560 return true; 561 return false; 562 } 563 564 565 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 566 if (!LVI && 567 (!isa<PHINode>(CondCmp->getOperand(0)) || 568 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) { 569 // If we have a comparison, loop over the predecessors to see if there is 570 // a condition with a lexically identical value. 571 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 572 for (; PI != E; ++PI) 573 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 574 if (PBI->isConditional() && *PI != BB) { 575 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 576 if (CI->getOperand(0) == CondCmp->getOperand(0) && 577 CI->getOperand(1) == CondCmp->getOperand(1) && 578 CI->getPredicate() == CondCmp->getPredicate()) { 579 // TODO: Could handle things like (x != 4) --> (x == 17) 580 if (ProcessBranchOnDuplicateCond(*PI, BB)) 581 return true; 582 } 583 } 584 } 585 } 586 } 587 588 // Check for some cases that are worth simplifying. Right now we want to look 589 // for loads that are used by a switch or by the condition for the branch. If 590 // we see one, check to see if it's partially redundant. If so, insert a PHI 591 // which can then be used to thread the values. 592 // 593 Value *SimplifyValue = CondInst; 594 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 595 if (isa<Constant>(CondCmp->getOperand(1))) 596 SimplifyValue = CondCmp->getOperand(0); 597 598 // TODO: There are other places where load PRE would be profitable, such as 599 // more complex comparisons. 600 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 601 if (SimplifyPartiallyRedundantLoad(LI)) 602 return true; 603 604 605 // Handle a variety of cases where we are branching on something derived from 606 // a PHI node in the current block. If we can prove that any predecessors 607 // compute a predictable value based on a PHI node, thread those predecessors. 608 // 609 if (ProcessThreadableEdges(CondInst, BB)) 610 return true; 611 612 // If this is an otherwise-unfoldable branch on a phi node in the current 613 // block, see if we can simplify. 614 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 615 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 616 return ProcessBranchOnPHI(PN); 617 618 619 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 620 if (CondInst->getOpcode() == Instruction::Xor && 621 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 622 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 623 624 625 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 626 // "(X == 4)", thread through this block. 627 628 return false; 629} 630 631/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 632/// block that jump on exactly the same condition. This means that we almost 633/// always know the direction of the edge in the DESTBB: 634/// PREDBB: 635/// br COND, DESTBB, BBY 636/// DESTBB: 637/// br COND, BBZ, BBW 638/// 639/// If DESTBB has multiple predecessors, we can't just constant fold the branch 640/// in DESTBB, we have to thread over it. 641bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 642 BasicBlock *BB) { 643 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 644 645 // If both successors of PredBB go to DESTBB, we don't know anything. We can 646 // fold the branch to an unconditional one, which allows other recursive 647 // simplifications. 648 bool BranchDir; 649 if (PredBI->getSuccessor(1) != BB) 650 BranchDir = true; 651 else if (PredBI->getSuccessor(0) != BB) 652 BranchDir = false; 653 else { 654 DEBUG(dbgs() << " In block '" << PredBB->getName() 655 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 656 ++NumFolds; 657 ConstantFoldTerminator(PredBB); 658 return true; 659 } 660 661 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 662 663 // If the dest block has one predecessor, just fix the branch condition to a 664 // constant and fold it. 665 if (BB->getSinglePredecessor()) { 666 DEBUG(dbgs() << " In block '" << BB->getName() 667 << "' folding condition to '" << BranchDir << "': " 668 << *BB->getTerminator() << '\n'); 669 ++NumFolds; 670 Value *OldCond = DestBI->getCondition(); 671 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 672 BranchDir)); 673 ConstantFoldTerminator(BB); 674 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 675 return true; 676 } 677 678 679 // Next, figure out which successor we are threading to. 680 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 681 682 SmallVector<BasicBlock*, 2> Preds; 683 Preds.push_back(PredBB); 684 685 // Ok, try to thread it! 686 return ThreadEdge(BB, Preds, SuccBB); 687} 688 689/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 690/// block that switch on exactly the same condition. This means that we almost 691/// always know the direction of the edge in the DESTBB: 692/// PREDBB: 693/// switch COND [... DESTBB, BBY ... ] 694/// DESTBB: 695/// switch COND [... BBZ, BBW ] 696/// 697/// Optimizing switches like this is very important, because simplifycfg builds 698/// switches out of repeated 'if' conditions. 699bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 700 BasicBlock *DestBB) { 701 // Can't thread edge to self. 702 if (PredBB == DestBB) 703 return false; 704 705 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 706 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 707 708 // There are a variety of optimizations that we can potentially do on these 709 // blocks: we order them from most to least preferable. 710 711 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 712 // directly to their destination. This does not introduce *any* code size 713 // growth. Skip debug info first. 714 BasicBlock::iterator BBI = DestBB->begin(); 715 while (isa<DbgInfoIntrinsic>(BBI)) 716 BBI++; 717 718 // FIXME: Thread if it just contains a PHI. 719 if (isa<SwitchInst>(BBI)) { 720 bool MadeChange = false; 721 // Ignore the default edge for now. 722 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 723 ConstantInt *DestVal = DestSI->getCaseValue(i); 724 BasicBlock *DestSucc = DestSI->getSuccessor(i); 725 726 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 727 // PredSI has an explicit case for it. If so, forward. If it is covered 728 // by the default case, we can't update PredSI. 729 unsigned PredCase = PredSI->findCaseValue(DestVal); 730 if (PredCase == 0) continue; 731 732 // If PredSI doesn't go to DestBB on this value, then it won't reach the 733 // case on this condition. 734 if (PredSI->getSuccessor(PredCase) != DestBB && 735 DestSI->getSuccessor(i) != DestBB) 736 continue; 737 738 // Do not forward this if it already goes to this destination, this would 739 // be an infinite loop. 740 if (PredSI->getSuccessor(PredCase) == DestSucc) 741 continue; 742 743 // Otherwise, we're safe to make the change. Make sure that the edge from 744 // DestSI to DestSucc is not critical and has no PHI nodes. 745 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 746 DEBUG(dbgs() << "THROUGH: " << *DestSI); 747 748 // If the destination has PHI nodes, just split the edge for updating 749 // simplicity. 750 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 751 SplitCriticalEdge(DestSI, i, this); 752 DestSucc = DestSI->getSuccessor(i); 753 } 754 FoldSingleEntryPHINodes(DestSucc); 755 PredSI->setSuccessor(PredCase, DestSucc); 756 MadeChange = true; 757 } 758 759 if (MadeChange) 760 return true; 761 } 762 763 return false; 764} 765 766 767/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 768/// load instruction, eliminate it by replacing it with a PHI node. This is an 769/// important optimization that encourages jump threading, and needs to be run 770/// interlaced with other jump threading tasks. 771bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 772 // Don't hack volatile loads. 773 if (LI->isVolatile()) return false; 774 775 // If the load is defined in a block with exactly one predecessor, it can't be 776 // partially redundant. 777 BasicBlock *LoadBB = LI->getParent(); 778 if (LoadBB->getSinglePredecessor()) 779 return false; 780 781 Value *LoadedPtr = LI->getOperand(0); 782 783 // If the loaded operand is defined in the LoadBB, it can't be available. 784 // TODO: Could do simple PHI translation, that would be fun :) 785 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 786 if (PtrOp->getParent() == LoadBB) 787 return false; 788 789 // Scan a few instructions up from the load, to see if it is obviously live at 790 // the entry to its block. 791 BasicBlock::iterator BBIt = LI; 792 793 if (Value *AvailableVal = 794 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 795 // If the value if the load is locally available within the block, just use 796 // it. This frequently occurs for reg2mem'd allocas. 797 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 798 799 // If the returned value is the load itself, replace with an undef. This can 800 // only happen in dead loops. 801 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 802 LI->replaceAllUsesWith(AvailableVal); 803 LI->eraseFromParent(); 804 return true; 805 } 806 807 // Otherwise, if we scanned the whole block and got to the top of the block, 808 // we know the block is locally transparent to the load. If not, something 809 // might clobber its value. 810 if (BBIt != LoadBB->begin()) 811 return false; 812 813 814 SmallPtrSet<BasicBlock*, 8> PredsScanned; 815 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 816 AvailablePredsTy AvailablePreds; 817 BasicBlock *OneUnavailablePred = 0; 818 819 // If we got here, the loaded value is transparent through to the start of the 820 // block. Check to see if it is available in any of the predecessor blocks. 821 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 822 PI != PE; ++PI) { 823 BasicBlock *PredBB = *PI; 824 825 // If we already scanned this predecessor, skip it. 826 if (!PredsScanned.insert(PredBB)) 827 continue; 828 829 // Scan the predecessor to see if the value is available in the pred. 830 BBIt = PredBB->end(); 831 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 832 if (!PredAvailable) { 833 OneUnavailablePred = PredBB; 834 continue; 835 } 836 837 // If so, this load is partially redundant. Remember this info so that we 838 // can create a PHI node. 839 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 840 } 841 842 // If the loaded value isn't available in any predecessor, it isn't partially 843 // redundant. 844 if (AvailablePreds.empty()) return false; 845 846 // Okay, the loaded value is available in at least one (and maybe all!) 847 // predecessors. If the value is unavailable in more than one unique 848 // predecessor, we want to insert a merge block for those common predecessors. 849 // This ensures that we only have to insert one reload, thus not increasing 850 // code size. 851 BasicBlock *UnavailablePred = 0; 852 853 // If there is exactly one predecessor where the value is unavailable, the 854 // already computed 'OneUnavailablePred' block is it. If it ends in an 855 // unconditional branch, we know that it isn't a critical edge. 856 if (PredsScanned.size() == AvailablePreds.size()+1 && 857 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 858 UnavailablePred = OneUnavailablePred; 859 } else if (PredsScanned.size() != AvailablePreds.size()) { 860 // Otherwise, we had multiple unavailable predecessors or we had a critical 861 // edge from the one. 862 SmallVector<BasicBlock*, 8> PredsToSplit; 863 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 864 865 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 866 AvailablePredSet.insert(AvailablePreds[i].first); 867 868 // Add all the unavailable predecessors to the PredsToSplit list. 869 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 870 PI != PE; ++PI) 871 if (!AvailablePredSet.count(*PI)) 872 PredsToSplit.push_back(*PI); 873 874 // Split them out to their own block. 875 UnavailablePred = 876 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 877 "thread-pre-split", this); 878 } 879 880 // If the value isn't available in all predecessors, then there will be 881 // exactly one where it isn't available. Insert a load on that edge and add 882 // it to the AvailablePreds list. 883 if (UnavailablePred) { 884 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 885 "Can't handle critical edge here!"); 886 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 887 LI->getAlignment(), 888 UnavailablePred->getTerminator()); 889 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 890 } 891 892 // Now we know that each predecessor of this block has a value in 893 // AvailablePreds, sort them for efficient access as we're walking the preds. 894 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 895 896 // Create a PHI node at the start of the block for the PRE'd load value. 897 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 898 PN->takeName(LI); 899 900 // Insert new entries into the PHI for each predecessor. A single block may 901 // have multiple entries here. 902 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 903 ++PI) { 904 AvailablePredsTy::iterator I = 905 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 906 std::make_pair(*PI, (Value*)0)); 907 908 assert(I != AvailablePreds.end() && I->first == *PI && 909 "Didn't find entry for predecessor!"); 910 911 PN->addIncoming(I->second, I->first); 912 } 913 914 //cerr << "PRE: " << *LI << *PN << "\n"; 915 916 LI->replaceAllUsesWith(PN); 917 LI->eraseFromParent(); 918 919 return true; 920} 921 922/// FindMostPopularDest - The specified list contains multiple possible 923/// threadable destinations. Pick the one that occurs the most frequently in 924/// the list. 925static BasicBlock * 926FindMostPopularDest(BasicBlock *BB, 927 const SmallVectorImpl<std::pair<BasicBlock*, 928 BasicBlock*> > &PredToDestList) { 929 assert(!PredToDestList.empty()); 930 931 // Determine popularity. If there are multiple possible destinations, we 932 // explicitly choose to ignore 'undef' destinations. We prefer to thread 933 // blocks with known and real destinations to threading undef. We'll handle 934 // them later if interesting. 935 DenseMap<BasicBlock*, unsigned> DestPopularity; 936 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 937 if (PredToDestList[i].second) 938 DestPopularity[PredToDestList[i].second]++; 939 940 // Find the most popular dest. 941 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 942 BasicBlock *MostPopularDest = DPI->first; 943 unsigned Popularity = DPI->second; 944 SmallVector<BasicBlock*, 4> SamePopularity; 945 946 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 947 // If the popularity of this entry isn't higher than the popularity we've 948 // seen so far, ignore it. 949 if (DPI->second < Popularity) 950 ; // ignore. 951 else if (DPI->second == Popularity) { 952 // If it is the same as what we've seen so far, keep track of it. 953 SamePopularity.push_back(DPI->first); 954 } else { 955 // If it is more popular, remember it. 956 SamePopularity.clear(); 957 MostPopularDest = DPI->first; 958 Popularity = DPI->second; 959 } 960 } 961 962 // Okay, now we know the most popular destination. If there is more than 963 // destination, we need to determine one. This is arbitrary, but we need 964 // to make a deterministic decision. Pick the first one that appears in the 965 // successor list. 966 if (!SamePopularity.empty()) { 967 SamePopularity.push_back(MostPopularDest); 968 TerminatorInst *TI = BB->getTerminator(); 969 for (unsigned i = 0; ; ++i) { 970 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 971 972 if (std::find(SamePopularity.begin(), SamePopularity.end(), 973 TI->getSuccessor(i)) == SamePopularity.end()) 974 continue; 975 976 MostPopularDest = TI->getSuccessor(i); 977 break; 978 } 979 } 980 981 // Okay, we have finally picked the most popular destination. 982 return MostPopularDest; 983} 984 985bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) { 986 // If threading this would thread across a loop header, don't even try to 987 // thread the edge. 988 if (LoopHeaders.count(BB)) 989 return false; 990 991 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 992 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues)) 993 return false; 994 assert(!PredValues.empty() && 995 "ComputeValueKnownInPredecessors returned true with no values"); 996 997 DEBUG(dbgs() << "IN BB: " << *BB; 998 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 999 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "; 1000 if (PredValues[i].first) 1001 dbgs() << *PredValues[i].first; 1002 else 1003 dbgs() << "UNDEF"; 1004 dbgs() << " for pred '" << PredValues[i].second->getName() 1005 << "'.\n"; 1006 }); 1007 1008 // Decide what we want to thread through. Convert our list of known values to 1009 // a list of known destinations for each pred. This also discards duplicate 1010 // predecessors and keeps track of the undefined inputs (which are represented 1011 // as a null dest in the PredToDestList). 1012 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1013 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1014 1015 BasicBlock *OnlyDest = 0; 1016 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1017 1018 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1019 BasicBlock *Pred = PredValues[i].second; 1020 if (!SeenPreds.insert(Pred)) 1021 continue; // Duplicate predecessor entry. 1022 1023 // If the predecessor ends with an indirect goto, we can't change its 1024 // destination. 1025 if (isa<IndirectBrInst>(Pred->getTerminator())) 1026 continue; 1027 1028 ConstantInt *Val = PredValues[i].first; 1029 1030 BasicBlock *DestBB; 1031 if (Val == 0) // Undef. 1032 DestBB = 0; 1033 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1034 DestBB = BI->getSuccessor(Val->isZero()); 1035 else { 1036 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 1037 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 1038 } 1039 1040 // If we have exactly one destination, remember it for efficiency below. 1041 if (i == 0) 1042 OnlyDest = DestBB; 1043 else if (OnlyDest != DestBB) 1044 OnlyDest = MultipleDestSentinel; 1045 1046 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1047 } 1048 1049 // If all edges were unthreadable, we fail. 1050 if (PredToDestList.empty()) 1051 return false; 1052 1053 // Determine which is the most common successor. If we have many inputs and 1054 // this block is a switch, we want to start by threading the batch that goes 1055 // to the most popular destination first. If we only know about one 1056 // threadable destination (the common case) we can avoid this. 1057 BasicBlock *MostPopularDest = OnlyDest; 1058 1059 if (MostPopularDest == MultipleDestSentinel) 1060 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1061 1062 // Now that we know what the most popular destination is, factor all 1063 // predecessors that will jump to it into a single predecessor. 1064 SmallVector<BasicBlock*, 16> PredsToFactor; 1065 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1066 if (PredToDestList[i].second == MostPopularDest) { 1067 BasicBlock *Pred = PredToDestList[i].first; 1068 1069 // This predecessor may be a switch or something else that has multiple 1070 // edges to the block. Factor each of these edges by listing them 1071 // according to # occurrences in PredsToFactor. 1072 TerminatorInst *PredTI = Pred->getTerminator(); 1073 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1074 if (PredTI->getSuccessor(i) == BB) 1075 PredsToFactor.push_back(Pred); 1076 } 1077 1078 // If the threadable edges are branching on an undefined value, we get to pick 1079 // the destination that these predecessors should get to. 1080 if (MostPopularDest == 0) 1081 MostPopularDest = BB->getTerminator()-> 1082 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1083 1084 // Ok, try to thread it! 1085 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1086} 1087 1088/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1089/// a PHI node in the current block. See if there are any simplifications we 1090/// can do based on inputs to the phi node. 1091/// 1092bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1093 BasicBlock *BB = PN->getParent(); 1094 1095 // TODO: We could make use of this to do it once for blocks with common PHI 1096 // values. 1097 SmallVector<BasicBlock*, 1> PredBBs; 1098 PredBBs.resize(1); 1099 1100 // If any of the predecessor blocks end in an unconditional branch, we can 1101 // *duplicate* the conditional branch into that block in order to further 1102 // encourage jump threading and to eliminate cases where we have branch on a 1103 // phi of an icmp (branch on icmp is much better). 1104 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1105 BasicBlock *PredBB = PN->getIncomingBlock(i); 1106 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1107 if (PredBr->isUnconditional()) { 1108 PredBBs[0] = PredBB; 1109 // Try to duplicate BB into PredBB. 1110 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1111 return true; 1112 } 1113 } 1114 1115 return false; 1116} 1117 1118/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1119/// a xor instruction in the current block. See if there are any 1120/// simplifications we can do based on inputs to the xor. 1121/// 1122bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1123 BasicBlock *BB = BO->getParent(); 1124 1125 // If either the LHS or RHS of the xor is a constant, don't do this 1126 // optimization. 1127 if (isa<ConstantInt>(BO->getOperand(0)) || 1128 isa<ConstantInt>(BO->getOperand(1))) 1129 return false; 1130 1131 // If the first instruction in BB isn't a phi, we won't be able to infer 1132 // anything special about any particular predecessor. 1133 if (!isa<PHINode>(BB->front())) 1134 return false; 1135 1136 // If we have a xor as the branch input to this block, and we know that the 1137 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1138 // the condition into the predecessor and fix that value to true, saving some 1139 // logical ops on that path and encouraging other paths to simplify. 1140 // 1141 // This copies something like this: 1142 // 1143 // BB: 1144 // %X = phi i1 [1], [%X'] 1145 // %Y = icmp eq i32 %A, %B 1146 // %Z = xor i1 %X, %Y 1147 // br i1 %Z, ... 1148 // 1149 // Into: 1150 // BB': 1151 // %Y = icmp ne i32 %A, %B 1152 // br i1 %Z, ... 1153 1154 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues; 1155 bool isLHS = true; 1156 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) { 1157 assert(XorOpValues.empty()); 1158 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues)) 1159 return false; 1160 isLHS = false; 1161 } 1162 1163 assert(!XorOpValues.empty() && 1164 "ComputeValueKnownInPredecessors returned true with no values"); 1165 1166 // Scan the information to see which is most popular: true or false. The 1167 // predecessors can be of the set true, false, or undef. 1168 unsigned NumTrue = 0, NumFalse = 0; 1169 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1170 if (!XorOpValues[i].first) continue; // Ignore undefs for the count. 1171 if (XorOpValues[i].first->isZero()) 1172 ++NumFalse; 1173 else 1174 ++NumTrue; 1175 } 1176 1177 // Determine which value to split on, true, false, or undef if neither. 1178 ConstantInt *SplitVal = 0; 1179 if (NumTrue > NumFalse) 1180 SplitVal = ConstantInt::getTrue(BB->getContext()); 1181 else if (NumTrue != 0 || NumFalse != 0) 1182 SplitVal = ConstantInt::getFalse(BB->getContext()); 1183 1184 // Collect all of the blocks that this can be folded into so that we can 1185 // factor this once and clone it once. 1186 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1187 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1188 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue; 1189 1190 BlocksToFoldInto.push_back(XorOpValues[i].second); 1191 } 1192 1193 // If we inferred a value for all of the predecessors, then duplication won't 1194 // help us. However, we can just replace the LHS or RHS with the constant. 1195 if (BlocksToFoldInto.size() == 1196 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1197 if (SplitVal == 0) { 1198 // If all preds provide undef, just nuke the xor, because it is undef too. 1199 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1200 BO->eraseFromParent(); 1201 } else if (SplitVal->isZero()) { 1202 // If all preds provide 0, replace the xor with the other input. 1203 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1204 BO->eraseFromParent(); 1205 } else { 1206 // If all preds provide 1, set the computed value to 1. 1207 BO->setOperand(!isLHS, SplitVal); 1208 } 1209 1210 return true; 1211 } 1212 1213 // Try to duplicate BB into PredBB. 1214 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1215} 1216 1217 1218/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1219/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1220/// NewPred using the entries from OldPred (suitably mapped). 1221static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1222 BasicBlock *OldPred, 1223 BasicBlock *NewPred, 1224 DenseMap<Instruction*, Value*> &ValueMap) { 1225 for (BasicBlock::iterator PNI = PHIBB->begin(); 1226 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1227 // Ok, we have a PHI node. Figure out what the incoming value was for the 1228 // DestBlock. 1229 Value *IV = PN->getIncomingValueForBlock(OldPred); 1230 1231 // Remap the value if necessary. 1232 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1233 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1234 if (I != ValueMap.end()) 1235 IV = I->second; 1236 } 1237 1238 PN->addIncoming(IV, NewPred); 1239 } 1240} 1241 1242/// ThreadEdge - We have decided that it is safe and profitable to factor the 1243/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1244/// across BB. Transform the IR to reflect this change. 1245bool JumpThreading::ThreadEdge(BasicBlock *BB, 1246 const SmallVectorImpl<BasicBlock*> &PredBBs, 1247 BasicBlock *SuccBB) { 1248 // If threading to the same block as we come from, we would infinite loop. 1249 if (SuccBB == BB) { 1250 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1251 << "' - would thread to self!\n"); 1252 return false; 1253 } 1254 1255 // If threading this would thread across a loop header, don't thread the edge. 1256 // See the comments above FindLoopHeaders for justifications and caveats. 1257 if (LoopHeaders.count(BB)) { 1258 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1259 << "' to dest BB '" << SuccBB->getName() 1260 << "' - it might create an irreducible loop!\n"); 1261 return false; 1262 } 1263 1264 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1265 if (JumpThreadCost > Threshold) { 1266 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1267 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1268 return false; 1269 } 1270 1271 // And finally, do it! Start by factoring the predecessors is needed. 1272 BasicBlock *PredBB; 1273 if (PredBBs.size() == 1) 1274 PredBB = PredBBs[0]; 1275 else { 1276 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1277 << " common predecessors.\n"); 1278 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1279 ".thr_comm", this); 1280 } 1281 1282 // And finally, do it! 1283 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1284 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1285 << ", across block:\n " 1286 << *BB << "\n"); 1287 1288 // We are going to have to map operands from the original BB block to the new 1289 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1290 // account for entry from PredBB. 1291 DenseMap<Instruction*, Value*> ValueMapping; 1292 1293 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1294 BB->getName()+".thread", 1295 BB->getParent(), BB); 1296 NewBB->moveAfter(PredBB); 1297 1298 BasicBlock::iterator BI = BB->begin(); 1299 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1300 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1301 1302 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1303 // mapping and using it to remap operands in the cloned instructions. 1304 for (; !isa<TerminatorInst>(BI); ++BI) { 1305 Instruction *New = BI->clone(); 1306 New->setName(BI->getName()); 1307 NewBB->getInstList().push_back(New); 1308 ValueMapping[BI] = New; 1309 1310 // Remap operands to patch up intra-block references. 1311 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1312 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1313 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1314 if (I != ValueMapping.end()) 1315 New->setOperand(i, I->second); 1316 } 1317 } 1318 1319 // We didn't copy the terminator from BB over to NewBB, because there is now 1320 // an unconditional jump to SuccBB. Insert the unconditional jump. 1321 BranchInst::Create(SuccBB, NewBB); 1322 1323 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1324 // PHI nodes for NewBB now. 1325 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1326 1327 // If there were values defined in BB that are used outside the block, then we 1328 // now have to update all uses of the value to use either the original value, 1329 // the cloned value, or some PHI derived value. This can require arbitrary 1330 // PHI insertion, of which we are prepared to do, clean these up now. 1331 SSAUpdater SSAUpdate; 1332 SmallVector<Use*, 16> UsesToRename; 1333 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1334 // Scan all uses of this instruction to see if it is used outside of its 1335 // block, and if so, record them in UsesToRename. 1336 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1337 ++UI) { 1338 Instruction *User = cast<Instruction>(*UI); 1339 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1340 if (UserPN->getIncomingBlock(UI) == BB) 1341 continue; 1342 } else if (User->getParent() == BB) 1343 continue; 1344 1345 UsesToRename.push_back(&UI.getUse()); 1346 } 1347 1348 // If there are no uses outside the block, we're done with this instruction. 1349 if (UsesToRename.empty()) 1350 continue; 1351 1352 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1353 1354 // We found a use of I outside of BB. Rename all uses of I that are outside 1355 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1356 // with the two values we know. 1357 SSAUpdate.Initialize(I); 1358 SSAUpdate.AddAvailableValue(BB, I); 1359 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1360 1361 while (!UsesToRename.empty()) 1362 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1363 DEBUG(dbgs() << "\n"); 1364 } 1365 1366 1367 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1368 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1369 // us to simplify any PHI nodes in BB. 1370 TerminatorInst *PredTerm = PredBB->getTerminator(); 1371 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1372 if (PredTerm->getSuccessor(i) == BB) { 1373 RemovePredecessorAndSimplify(BB, PredBB, TD); 1374 PredTerm->setSuccessor(i, NewBB); 1375 } 1376 1377 // At this point, the IR is fully up to date and consistent. Do a quick scan 1378 // over the new instructions and zap any that are constants or dead. This 1379 // frequently happens because of phi translation. 1380 SimplifyInstructionsInBlock(NewBB, TD); 1381 1382 // Threaded an edge! 1383 ++NumThreads; 1384 return true; 1385} 1386 1387/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1388/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1389/// If we can duplicate the contents of BB up into PredBB do so now, this 1390/// improves the odds that the branch will be on an analyzable instruction like 1391/// a compare. 1392bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1393 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1394 assert(!PredBBs.empty() && "Can't handle an empty set"); 1395 1396 // If BB is a loop header, then duplicating this block outside the loop would 1397 // cause us to transform this into an irreducible loop, don't do this. 1398 // See the comments above FindLoopHeaders for justifications and caveats. 1399 if (LoopHeaders.count(BB)) { 1400 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1401 << "' into predecessor block '" << PredBBs[0]->getName() 1402 << "' - it might create an irreducible loop!\n"); 1403 return false; 1404 } 1405 1406 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1407 if (DuplicationCost > Threshold) { 1408 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1409 << "' - Cost is too high: " << DuplicationCost << "\n"); 1410 return false; 1411 } 1412 1413 // And finally, do it! Start by factoring the predecessors is needed. 1414 BasicBlock *PredBB; 1415 if (PredBBs.size() == 1) 1416 PredBB = PredBBs[0]; 1417 else { 1418 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1419 << " common predecessors.\n"); 1420 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1421 ".thr_comm", this); 1422 } 1423 1424 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1425 // of PredBB. 1426 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1427 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1428 << DuplicationCost << " block is:" << *BB << "\n"); 1429 1430 // Unless PredBB ends with an unconditional branch, split the edge so that we 1431 // can just clone the bits from BB into the end of the new PredBB. 1432 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1433 1434 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { 1435 PredBB = SplitEdge(PredBB, BB, this); 1436 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1437 } 1438 1439 // We are going to have to map operands from the original BB block into the 1440 // PredBB block. Evaluate PHI nodes in BB. 1441 DenseMap<Instruction*, Value*> ValueMapping; 1442 1443 BasicBlock::iterator BI = BB->begin(); 1444 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1445 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1446 1447 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1448 // mapping and using it to remap operands in the cloned instructions. 1449 for (; BI != BB->end(); ++BI) { 1450 Instruction *New = BI->clone(); 1451 1452 // Remap operands to patch up intra-block references. 1453 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1454 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1455 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1456 if (I != ValueMapping.end()) 1457 New->setOperand(i, I->second); 1458 } 1459 1460 // If this instruction can be simplified after the operands are updated, 1461 // just use the simplified value instead. This frequently happens due to 1462 // phi translation. 1463 if (Value *IV = SimplifyInstruction(New, TD)) { 1464 delete New; 1465 ValueMapping[BI] = IV; 1466 } else { 1467 // Otherwise, insert the new instruction into the block. 1468 New->setName(BI->getName()); 1469 PredBB->getInstList().insert(OldPredBranch, New); 1470 ValueMapping[BI] = New; 1471 } 1472 } 1473 1474 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1475 // add entries to the PHI nodes for branch from PredBB now. 1476 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1477 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1478 ValueMapping); 1479 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1480 ValueMapping); 1481 1482 // If there were values defined in BB that are used outside the block, then we 1483 // now have to update all uses of the value to use either the original value, 1484 // the cloned value, or some PHI derived value. This can require arbitrary 1485 // PHI insertion, of which we are prepared to do, clean these up now. 1486 SSAUpdater SSAUpdate; 1487 SmallVector<Use*, 16> UsesToRename; 1488 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1489 // Scan all uses of this instruction to see if it is used outside of its 1490 // block, and if so, record them in UsesToRename. 1491 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1492 ++UI) { 1493 Instruction *User = cast<Instruction>(*UI); 1494 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1495 if (UserPN->getIncomingBlock(UI) == BB) 1496 continue; 1497 } else if (User->getParent() == BB) 1498 continue; 1499 1500 UsesToRename.push_back(&UI.getUse()); 1501 } 1502 1503 // If there are no uses outside the block, we're done with this instruction. 1504 if (UsesToRename.empty()) 1505 continue; 1506 1507 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1508 1509 // We found a use of I outside of BB. Rename all uses of I that are outside 1510 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1511 // with the two values we know. 1512 SSAUpdate.Initialize(I); 1513 SSAUpdate.AddAvailableValue(BB, I); 1514 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1515 1516 while (!UsesToRename.empty()) 1517 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1518 DEBUG(dbgs() << "\n"); 1519 } 1520 1521 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1522 // that we nuked. 1523 RemovePredecessorAndSimplify(BB, PredBB, TD); 1524 1525 // Remove the unconditional branch at the end of the PredBB block. 1526 OldPredBranch->eraseFromParent(); 1527 1528 ++NumDupes; 1529 return true; 1530} 1531 1532 1533