JumpThreading.cpp revision 77beb479f556093c5f1b4854fcbb095cda34f202
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 BasicBlock *PredBB); 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); 106 107 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 108 }; 109} 110 111char JumpThreading::ID = 0; 112static RegisterPass<JumpThreading> 113X("jump-threading", "Jump Threading"); 114 115// Public interface to the Jump Threading pass 116FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 117 118/// runOnFunction - Top level algorithm. 119/// 120bool JumpThreading::runOnFunction(Function &F) { 121 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 122 TD = getAnalysisIfAvailable<TargetData>(); 123 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0; 124 125 FindLoopHeaders(F); 126 127 bool Changed, EverChanged = false; 128 do { 129 Changed = false; 130 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 131 BasicBlock *BB = I; 132 // Thread all of the branches we can over this block. 133 while (ProcessBlock(BB)) 134 Changed = true; 135 136 ++I; 137 138 // If the block is trivially dead, zap it. This eliminates the successor 139 // edges which simplifies the CFG. 140 if (pred_begin(BB) == pred_end(BB) && 141 BB != &BB->getParent()->getEntryBlock()) { 142 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 143 << "' with terminator: " << *BB->getTerminator() << '\n'); 144 LoopHeaders.erase(BB); 145 DeleteDeadBlock(BB); 146 Changed = true; 147 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 148 // Can't thread an unconditional jump, but if the block is "almost 149 // empty", we can replace uses of it with uses of the successor and make 150 // this dead. 151 if (BI->isUnconditional() && 152 BB != &BB->getParent()->getEntryBlock()) { 153 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 154 // Ignore dbg intrinsics. 155 while (isa<DbgInfoIntrinsic>(BBI)) 156 ++BBI; 157 // If the terminator is the only non-phi instruction, try to nuke it. 158 if (BBI->isTerminator()) { 159 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 160 // block, we have to make sure it isn't in the LoopHeaders set. We 161 // reinsert afterward if needed. 162 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 163 BasicBlock *Succ = BI->getSuccessor(0); 164 165 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 166 Changed = true; 167 // If we deleted BB and BB was the header of a loop, then the 168 // successor is now the header of the loop. 169 BB = Succ; 170 } 171 172 if (ErasedFromLoopHeaders) 173 LoopHeaders.insert(BB); 174 } 175 } 176 } 177 } 178 EverChanged |= Changed; 179 } while (Changed); 180 181 LoopHeaders.clear(); 182 return EverChanged; 183} 184 185/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 186/// thread across it. 187static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 188 /// Ignore PHI nodes, these will be flattened when duplication happens. 189 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 190 191 // FIXME: THREADING will delete values that are just used to compute the 192 // branch, so they shouldn't count against the duplication cost. 193 194 195 // Sum up the cost of each instruction until we get to the terminator. Don't 196 // include the terminator because the copy won't include it. 197 unsigned Size = 0; 198 for (; !isa<TerminatorInst>(I); ++I) { 199 // Debugger intrinsics don't incur code size. 200 if (isa<DbgInfoIntrinsic>(I)) continue; 201 202 // If this is a pointer->pointer bitcast, it is free. 203 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType())) 204 continue; 205 206 // All other instructions count for at least one unit. 207 ++Size; 208 209 // Calls are more expensive. If they are non-intrinsic calls, we model them 210 // as having cost of 4. If they are a non-vector intrinsic, we model them 211 // as having cost of 2 total, and if they are a vector intrinsic, we model 212 // them as having cost 1. 213 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 214 if (!isa<IntrinsicInst>(CI)) 215 Size += 3; 216 else if (!isa<VectorType>(CI->getType())) 217 Size += 1; 218 } 219 } 220 221 // Threading through a switch statement is particularly profitable. If this 222 // block ends in a switch, decrease its cost to make it more likely to happen. 223 if (isa<SwitchInst>(I)) 224 Size = Size > 6 ? Size-6 : 0; 225 226 return Size; 227} 228 229/// FindLoopHeaders - We do not want jump threading to turn proper loop 230/// structures into irreducible loops. Doing this breaks up the loop nesting 231/// hierarchy and pessimizes later transformations. To prevent this from 232/// happening, we first have to find the loop headers. Here we approximate this 233/// by finding targets of backedges in the CFG. 234/// 235/// Note that there definitely are cases when we want to allow threading of 236/// edges across a loop header. For example, threading a jump from outside the 237/// loop (the preheader) to an exit block of the loop is definitely profitable. 238/// It is also almost always profitable to thread backedges from within the loop 239/// to exit blocks, and is often profitable to thread backedges to other blocks 240/// within the loop (forming a nested loop). This simple analysis is not rich 241/// enough to track all of these properties and keep it up-to-date as the CFG 242/// mutates, so we don't allow any of these transformations. 243/// 244void JumpThreading::FindLoopHeaders(Function &F) { 245 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 246 FindFunctionBackedges(F, Edges); 247 248 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 249 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 250} 251 252/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 253/// if we can infer that the value is a known ConstantInt in any of our 254/// predecessors. If so, return the known list of value and pred BB in the 255/// result vector. If a value is known to be undef, it is returned as null. 256/// 257/// This returns true if there were any known values. 258/// 259bool JumpThreading:: 260ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 261 // If V is a constantint, then it is known in all predecessors. 262 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 263 ConstantInt *CI = dyn_cast<ConstantInt>(V); 264 265 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 266 Result.push_back(std::make_pair(CI, *PI)); 267 return true; 268 } 269 270 // If V is a non-instruction value, or an instruction in a different block, 271 // then it can't be derived from a PHI. 272 Instruction *I = dyn_cast<Instruction>(V); 273 if (I == 0 || I->getParent() != BB) { 274 275 // Okay, if this is a live-in value, see if it has a known value at the end 276 // of any of our predecessors. 277 // 278 // FIXME: This should be an edge property, not a block end property. 279 /// TODO: Per PR2563, we could infer value range information about a 280 /// predecessor based on its terminator. 281 // 282 if (LVI) { 283 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 284 // "I" is a non-local compare-with-a-constant instruction. This would be 285 // able to handle value inequalities better, for example if the compare is 286 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 287 // Perhaps getConstantOnEdge should be smart enough to do this? 288 289 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 290 // If the value is known by LazyValueInfo to be a constant in a 291 // predecessor, use that information to try to thread this block. 292 Constant *PredCst = LVI->getConstantOnEdge(V, *PI, BB); 293 if (PredCst == 0 || 294 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst))) 295 continue; 296 297 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), *PI)); 298 } 299 300 return !Result.empty(); 301 } 302 303 return false; 304 } 305 306 /// If I is a PHI node, then we know the incoming values for any constants. 307 if (PHINode *PN = dyn_cast<PHINode>(I)) { 308 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 309 Value *InVal = PN->getIncomingValue(i); 310 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 311 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 312 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 313 } 314 } 315 return !Result.empty(); 316 } 317 318 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 319 320 // Handle some boolean conditions. 321 if (I->getType()->getPrimitiveSizeInBits() == 1) { 322 // X | true -> true 323 // X & false -> false 324 if (I->getOpcode() == Instruction::Or || 325 I->getOpcode() == Instruction::And) { 326 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 327 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 328 329 if (LHSVals.empty() && RHSVals.empty()) 330 return false; 331 332 ConstantInt *InterestingVal; 333 if (I->getOpcode() == Instruction::Or) 334 InterestingVal = ConstantInt::getTrue(I->getContext()); 335 else 336 InterestingVal = ConstantInt::getFalse(I->getContext()); 337 338 // Scan for the sentinel. 339 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 340 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) 341 Result.push_back(LHSVals[i]); 342 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 343 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) 344 Result.push_back(RHSVals[i]); 345 return !Result.empty(); 346 } 347 348 // Handle the NOT form of XOR. 349 if (I->getOpcode() == Instruction::Xor && 350 isa<ConstantInt>(I->getOperand(1)) && 351 cast<ConstantInt>(I->getOperand(1))->isOne()) { 352 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result); 353 if (Result.empty()) 354 return false; 355 356 // Invert the known values. 357 for (unsigned i = 0, e = Result.size(); i != e; ++i) 358 if (Result[i].first) 359 Result[i].first = 360 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first)); 361 return true; 362 } 363 } 364 365 // Handle compare with phi operand, where the PHI is defined in this block. 366 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 367 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 368 if (PN && PN->getParent() == BB) { 369 // We can do this simplification if any comparisons fold to true or false. 370 // See if any do. 371 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 372 BasicBlock *PredBB = PN->getIncomingBlock(i); 373 Value *LHS = PN->getIncomingValue(i); 374 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 375 376 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD); 377 if (Res == 0) { 378 if (!LVI || !isa<Constant>(RHS)) 379 continue; 380 381 LazyValueInfo::Tristate 382 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 383 cast<Constant>(RHS), PredBB, BB); 384 if (ResT == LazyValueInfo::Unknown) 385 continue; 386 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 387 } 388 389 if (isa<UndefValue>(Res)) 390 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 391 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 392 Result.push_back(std::make_pair(CI, PredBB)); 393 } 394 395 return !Result.empty(); 396 } 397 398 399 // If comparing a live-in value against a constant, see if we know the 400 // live-in value on any predecessors. 401 if (LVI && isa<Constant>(Cmp->getOperand(1)) && 402 Cmp->getType()->isInteger() && // Not vector compare. 403 (!isa<Instruction>(Cmp->getOperand(0)) || 404 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB)) { 405 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 406 407 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 408 // If the value is known by LazyValueInfo to be a constant in a 409 // predecessor, use that information to try to thread this block. 410 LazyValueInfo::Tristate 411 Res = LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 412 RHSCst, *PI, BB); 413 if (Res == LazyValueInfo::Unknown) 414 continue; 415 416 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 417 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), *PI)); 418 } 419 420 return !Result.empty(); 421 } 422 } 423 return false; 424} 425 426 427 428/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 429/// in an undefined jump, decide which block is best to revector to. 430/// 431/// Since we can pick an arbitrary destination, we pick the successor with the 432/// fewest predecessors. This should reduce the in-degree of the others. 433/// 434static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 435 TerminatorInst *BBTerm = BB->getTerminator(); 436 unsigned MinSucc = 0; 437 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 438 // Compute the successor with the minimum number of predecessors. 439 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 440 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 441 TestBB = BBTerm->getSuccessor(i); 442 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 443 if (NumPreds < MinNumPreds) 444 MinSucc = i; 445 } 446 447 return MinSucc; 448} 449 450/// ProcessBlock - If there are any predecessors whose control can be threaded 451/// through to a successor, transform them now. 452bool JumpThreading::ProcessBlock(BasicBlock *BB) { 453 // If this block has a single predecessor, and if that pred has a single 454 // successor, merge the blocks. This encourages recursive jump threading 455 // because now the condition in this block can be threaded through 456 // predecessors of our predecessor block. 457 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 458 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 459 SinglePred != BB) { 460 // If SinglePred was a loop header, BB becomes one. 461 if (LoopHeaders.erase(SinglePred)) 462 LoopHeaders.insert(BB); 463 464 // Remember if SinglePred was the entry block of the function. If so, we 465 // will need to move BB back to the entry position. 466 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 467 MergeBasicBlockIntoOnlyPred(BB); 468 469 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 470 BB->moveBefore(&BB->getParent()->getEntryBlock()); 471 return true; 472 } 473 } 474 475 // Look to see if the terminator is a branch of switch, if not we can't thread 476 // it. 477 Value *Condition; 478 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 479 // Can't thread an unconditional jump. 480 if (BI->isUnconditional()) return false; 481 Condition = BI->getCondition(); 482 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 483 Condition = SI->getCondition(); 484 else 485 return false; // Must be an invoke. 486 487 // If the terminator of this block is branching on a constant, simplify the 488 // terminator to an unconditional branch. This can occur due to threading in 489 // other blocks. 490 if (isa<ConstantInt>(Condition)) { 491 DEBUG(dbgs() << " In block '" << BB->getName() 492 << "' folding terminator: " << *BB->getTerminator() << '\n'); 493 ++NumFolds; 494 ConstantFoldTerminator(BB); 495 return true; 496 } 497 498 // If the terminator is branching on an undef, we can pick any of the 499 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 500 if (isa<UndefValue>(Condition)) { 501 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 502 503 // Fold the branch/switch. 504 TerminatorInst *BBTerm = BB->getTerminator(); 505 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 506 if (i == BestSucc) continue; 507 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD); 508 } 509 510 DEBUG(dbgs() << " In block '" << BB->getName() 511 << "' folding undef terminator: " << *BBTerm << '\n'); 512 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 513 BBTerm->eraseFromParent(); 514 return true; 515 } 516 517 Instruction *CondInst = dyn_cast<Instruction>(Condition); 518 519 // If the condition is an instruction defined in another block, see if a 520 // predecessor has the same condition: 521 // br COND, BBX, BBY 522 // BBX: 523 // br COND, BBZ, BBW 524 if (!LVI && 525 !Condition->hasOneUse() && // Multiple uses. 526 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 527 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 528 if (isa<BranchInst>(BB->getTerminator())) { 529 for (; PI != E; ++PI) 530 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 531 if (PBI->isConditional() && PBI->getCondition() == Condition && 532 ProcessBranchOnDuplicateCond(*PI, BB)) 533 return true; 534 } else { 535 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 536 for (; PI != E; ++PI) 537 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator())) 538 if (PSI->getCondition() == Condition && 539 ProcessSwitchOnDuplicateCond(*PI, BB)) 540 return true; 541 } 542 } 543 544 // All the rest of our checks depend on the condition being an instruction. 545 if (CondInst == 0) { 546 // FIXME: Unify this with code below. 547 if (LVI && ProcessThreadableEdges(Condition, BB)) 548 return true; 549 return false; 550 } 551 552 553 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 554 if (!LVI && 555 (!isa<PHINode>(CondCmp->getOperand(0)) || 556 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) { 557 // If we have a comparison, loop over the predecessors to see if there is 558 // a condition with a lexically identical value. 559 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 560 for (; PI != E; ++PI) 561 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 562 if (PBI->isConditional() && *PI != BB) { 563 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 564 if (CI->getOperand(0) == CondCmp->getOperand(0) && 565 CI->getOperand(1) == CondCmp->getOperand(1) && 566 CI->getPredicate() == CondCmp->getPredicate()) { 567 // TODO: Could handle things like (x != 4) --> (x == 17) 568 if (ProcessBranchOnDuplicateCond(*PI, BB)) 569 return true; 570 } 571 } 572 } 573 } 574 } 575 576 // Check for some cases that are worth simplifying. Right now we want to look 577 // for loads that are used by a switch or by the condition for the branch. If 578 // we see one, check to see if it's partially redundant. If so, insert a PHI 579 // which can then be used to thread the values. 580 // 581 Value *SimplifyValue = CondInst; 582 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 583 if (isa<Constant>(CondCmp->getOperand(1))) 584 SimplifyValue = CondCmp->getOperand(0); 585 586 // TODO: There are other places where load PRE would be profitable, such as 587 // more complex comparisons. 588 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 589 if (SimplifyPartiallyRedundantLoad(LI)) 590 return true; 591 592 593 // Handle a variety of cases where we are branching on something derived from 594 // a PHI node in the current block. If we can prove that any predecessors 595 // compute a predictable value based on a PHI node, thread those predecessors. 596 // 597 if (ProcessThreadableEdges(CondInst, BB)) 598 return true; 599 600 // If this is an otherwise-unfoldable branch on a phi node in the current 601 // block, see if we can simplify. 602 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 603 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 604 return ProcessBranchOnPHI(PN); 605 606 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 607 // "(X == 4)", thread through this block. 608 609 return false; 610} 611 612/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 613/// block that jump on exactly the same condition. This means that we almost 614/// always know the direction of the edge in the DESTBB: 615/// PREDBB: 616/// br COND, DESTBB, BBY 617/// DESTBB: 618/// br COND, BBZ, BBW 619/// 620/// If DESTBB has multiple predecessors, we can't just constant fold the branch 621/// in DESTBB, we have to thread over it. 622bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 623 BasicBlock *BB) { 624 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 625 626 // If both successors of PredBB go to DESTBB, we don't know anything. We can 627 // fold the branch to an unconditional one, which allows other recursive 628 // simplifications. 629 bool BranchDir; 630 if (PredBI->getSuccessor(1) != BB) 631 BranchDir = true; 632 else if (PredBI->getSuccessor(0) != BB) 633 BranchDir = false; 634 else { 635 DEBUG(dbgs() << " In block '" << PredBB->getName() 636 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 637 ++NumFolds; 638 ConstantFoldTerminator(PredBB); 639 return true; 640 } 641 642 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 643 644 // If the dest block has one predecessor, just fix the branch condition to a 645 // constant and fold it. 646 if (BB->getSinglePredecessor()) { 647 DEBUG(dbgs() << " In block '" << BB->getName() 648 << "' folding condition to '" << BranchDir << "': " 649 << *BB->getTerminator() << '\n'); 650 ++NumFolds; 651 Value *OldCond = DestBI->getCondition(); 652 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 653 BranchDir)); 654 ConstantFoldTerminator(BB); 655 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 656 return true; 657 } 658 659 660 // Next, figure out which successor we are threading to. 661 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 662 663 SmallVector<BasicBlock*, 2> Preds; 664 Preds.push_back(PredBB); 665 666 // Ok, try to thread it! 667 return ThreadEdge(BB, Preds, SuccBB); 668} 669 670/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 671/// block that switch on exactly the same condition. This means that we almost 672/// always know the direction of the edge in the DESTBB: 673/// PREDBB: 674/// switch COND [... DESTBB, BBY ... ] 675/// DESTBB: 676/// switch COND [... BBZ, BBW ] 677/// 678/// Optimizing switches like this is very important, because simplifycfg builds 679/// switches out of repeated 'if' conditions. 680bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 681 BasicBlock *DestBB) { 682 // Can't thread edge to self. 683 if (PredBB == DestBB) 684 return false; 685 686 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 687 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 688 689 // There are a variety of optimizations that we can potentially do on these 690 // blocks: we order them from most to least preferable. 691 692 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 693 // directly to their destination. This does not introduce *any* code size 694 // growth. Skip debug info first. 695 BasicBlock::iterator BBI = DestBB->begin(); 696 while (isa<DbgInfoIntrinsic>(BBI)) 697 BBI++; 698 699 // FIXME: Thread if it just contains a PHI. 700 if (isa<SwitchInst>(BBI)) { 701 bool MadeChange = false; 702 // Ignore the default edge for now. 703 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 704 ConstantInt *DestVal = DestSI->getCaseValue(i); 705 BasicBlock *DestSucc = DestSI->getSuccessor(i); 706 707 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 708 // PredSI has an explicit case for it. If so, forward. If it is covered 709 // by the default case, we can't update PredSI. 710 unsigned PredCase = PredSI->findCaseValue(DestVal); 711 if (PredCase == 0) continue; 712 713 // If PredSI doesn't go to DestBB on this value, then it won't reach the 714 // case on this condition. 715 if (PredSI->getSuccessor(PredCase) != DestBB && 716 DestSI->getSuccessor(i) != DestBB) 717 continue; 718 719 // Do not forward this if it already goes to this destination, this would 720 // be an infinite loop. 721 if (PredSI->getSuccessor(PredCase) == DestSucc) 722 continue; 723 724 // Otherwise, we're safe to make the change. Make sure that the edge from 725 // DestSI to DestSucc is not critical and has no PHI nodes. 726 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 727 DEBUG(dbgs() << "THROUGH: " << *DestSI); 728 729 // If the destination has PHI nodes, just split the edge for updating 730 // simplicity. 731 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 732 SplitCriticalEdge(DestSI, i, this); 733 DestSucc = DestSI->getSuccessor(i); 734 } 735 FoldSingleEntryPHINodes(DestSucc); 736 PredSI->setSuccessor(PredCase, DestSucc); 737 MadeChange = true; 738 } 739 740 if (MadeChange) 741 return true; 742 } 743 744 return false; 745} 746 747 748/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 749/// load instruction, eliminate it by replacing it with a PHI node. This is an 750/// important optimization that encourages jump threading, and needs to be run 751/// interlaced with other jump threading tasks. 752bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 753 // Don't hack volatile loads. 754 if (LI->isVolatile()) return false; 755 756 // If the load is defined in a block with exactly one predecessor, it can't be 757 // partially redundant. 758 BasicBlock *LoadBB = LI->getParent(); 759 if (LoadBB->getSinglePredecessor()) 760 return false; 761 762 Value *LoadedPtr = LI->getOperand(0); 763 764 // If the loaded operand is defined in the LoadBB, it can't be available. 765 // TODO: Could do simple PHI translation, that would be fun :) 766 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 767 if (PtrOp->getParent() == LoadBB) 768 return false; 769 770 // Scan a few instructions up from the load, to see if it is obviously live at 771 // the entry to its block. 772 BasicBlock::iterator BBIt = LI; 773 774 if (Value *AvailableVal = 775 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 776 // If the value if the load is locally available within the block, just use 777 // it. This frequently occurs for reg2mem'd allocas. 778 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 779 780 // If the returned value is the load itself, replace with an undef. This can 781 // only happen in dead loops. 782 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 783 LI->replaceAllUsesWith(AvailableVal); 784 LI->eraseFromParent(); 785 return true; 786 } 787 788 // Otherwise, if we scanned the whole block and got to the top of the block, 789 // we know the block is locally transparent to the load. If not, something 790 // might clobber its value. 791 if (BBIt != LoadBB->begin()) 792 return false; 793 794 795 SmallPtrSet<BasicBlock*, 8> PredsScanned; 796 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 797 AvailablePredsTy AvailablePreds; 798 BasicBlock *OneUnavailablePred = 0; 799 800 // If we got here, the loaded value is transparent through to the start of the 801 // block. Check to see if it is available in any of the predecessor blocks. 802 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 803 PI != PE; ++PI) { 804 BasicBlock *PredBB = *PI; 805 806 // If we already scanned this predecessor, skip it. 807 if (!PredsScanned.insert(PredBB)) 808 continue; 809 810 // Scan the predecessor to see if the value is available in the pred. 811 BBIt = PredBB->end(); 812 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 813 if (!PredAvailable) { 814 OneUnavailablePred = PredBB; 815 continue; 816 } 817 818 // If so, this load is partially redundant. Remember this info so that we 819 // can create a PHI node. 820 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 821 } 822 823 // If the loaded value isn't available in any predecessor, it isn't partially 824 // redundant. 825 if (AvailablePreds.empty()) return false; 826 827 // Okay, the loaded value is available in at least one (and maybe all!) 828 // predecessors. If the value is unavailable in more than one unique 829 // predecessor, we want to insert a merge block for those common predecessors. 830 // This ensures that we only have to insert one reload, thus not increasing 831 // code size. 832 BasicBlock *UnavailablePred = 0; 833 834 // If there is exactly one predecessor where the value is unavailable, the 835 // already computed 'OneUnavailablePred' block is it. If it ends in an 836 // unconditional branch, we know that it isn't a critical edge. 837 if (PredsScanned.size() == AvailablePreds.size()+1 && 838 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 839 UnavailablePred = OneUnavailablePred; 840 } else if (PredsScanned.size() != AvailablePreds.size()) { 841 // Otherwise, we had multiple unavailable predecessors or we had a critical 842 // edge from the one. 843 SmallVector<BasicBlock*, 8> PredsToSplit; 844 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 845 846 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 847 AvailablePredSet.insert(AvailablePreds[i].first); 848 849 // Add all the unavailable predecessors to the PredsToSplit list. 850 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 851 PI != PE; ++PI) 852 if (!AvailablePredSet.count(*PI)) 853 PredsToSplit.push_back(*PI); 854 855 // Split them out to their own block. 856 UnavailablePred = 857 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 858 "thread-pre-split", this); 859 } 860 861 // If the value isn't available in all predecessors, then there will be 862 // exactly one where it isn't available. Insert a load on that edge and add 863 // it to the AvailablePreds list. 864 if (UnavailablePred) { 865 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 866 "Can't handle critical edge here!"); 867 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 868 LI->getAlignment(), 869 UnavailablePred->getTerminator()); 870 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 871 } 872 873 // Now we know that each predecessor of this block has a value in 874 // AvailablePreds, sort them for efficient access as we're walking the preds. 875 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 876 877 // Create a PHI node at the start of the block for the PRE'd load value. 878 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 879 PN->takeName(LI); 880 881 // Insert new entries into the PHI for each predecessor. A single block may 882 // have multiple entries here. 883 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 884 ++PI) { 885 AvailablePredsTy::iterator I = 886 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 887 std::make_pair(*PI, (Value*)0)); 888 889 assert(I != AvailablePreds.end() && I->first == *PI && 890 "Didn't find entry for predecessor!"); 891 892 PN->addIncoming(I->second, I->first); 893 } 894 895 //cerr << "PRE: " << *LI << *PN << "\n"; 896 897 LI->replaceAllUsesWith(PN); 898 LI->eraseFromParent(); 899 900 return true; 901} 902 903/// FindMostPopularDest - The specified list contains multiple possible 904/// threadable destinations. Pick the one that occurs the most frequently in 905/// the list. 906static BasicBlock * 907FindMostPopularDest(BasicBlock *BB, 908 const SmallVectorImpl<std::pair<BasicBlock*, 909 BasicBlock*> > &PredToDestList) { 910 assert(!PredToDestList.empty()); 911 912 // Determine popularity. If there are multiple possible destinations, we 913 // explicitly choose to ignore 'undef' destinations. We prefer to thread 914 // blocks with known and real destinations to threading undef. We'll handle 915 // them later if interesting. 916 DenseMap<BasicBlock*, unsigned> DestPopularity; 917 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 918 if (PredToDestList[i].second) 919 DestPopularity[PredToDestList[i].second]++; 920 921 // Find the most popular dest. 922 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 923 BasicBlock *MostPopularDest = DPI->first; 924 unsigned Popularity = DPI->second; 925 SmallVector<BasicBlock*, 4> SamePopularity; 926 927 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 928 // If the popularity of this entry isn't higher than the popularity we've 929 // seen so far, ignore it. 930 if (DPI->second < Popularity) 931 ; // ignore. 932 else if (DPI->second == Popularity) { 933 // If it is the same as what we've seen so far, keep track of it. 934 SamePopularity.push_back(DPI->first); 935 } else { 936 // If it is more popular, remember it. 937 SamePopularity.clear(); 938 MostPopularDest = DPI->first; 939 Popularity = DPI->second; 940 } 941 } 942 943 // Okay, now we know the most popular destination. If there is more than 944 // destination, we need to determine one. This is arbitrary, but we need 945 // to make a deterministic decision. Pick the first one that appears in the 946 // successor list. 947 if (!SamePopularity.empty()) { 948 SamePopularity.push_back(MostPopularDest); 949 TerminatorInst *TI = BB->getTerminator(); 950 for (unsigned i = 0; ; ++i) { 951 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 952 953 if (std::find(SamePopularity.begin(), SamePopularity.end(), 954 TI->getSuccessor(i)) == SamePopularity.end()) 955 continue; 956 957 MostPopularDest = TI->getSuccessor(i); 958 break; 959 } 960 } 961 962 // Okay, we have finally picked the most popular destination. 963 return MostPopularDest; 964} 965 966bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) { 967 // If threading this would thread across a loop header, don't even try to 968 // thread the edge. 969 if (LoopHeaders.count(BB)) 970 return false; 971 972 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 973 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues)) 974 return false; 975 assert(!PredValues.empty() && 976 "ComputeValueKnownInPredecessors returned true with no values"); 977 978 DEBUG(dbgs() << "IN BB: " << *BB; 979 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 980 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "; 981 if (PredValues[i].first) 982 dbgs() << *PredValues[i].first; 983 else 984 dbgs() << "UNDEF"; 985 dbgs() << " for pred '" << PredValues[i].second->getName() 986 << "'.\n"; 987 }); 988 989 // Decide what we want to thread through. Convert our list of known values to 990 // a list of known destinations for each pred. This also discards duplicate 991 // predecessors and keeps track of the undefined inputs (which are represented 992 // as a null dest in the PredToDestList). 993 SmallPtrSet<BasicBlock*, 16> SeenPreds; 994 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 995 996 BasicBlock *OnlyDest = 0; 997 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 998 999 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1000 BasicBlock *Pred = PredValues[i].second; 1001 if (!SeenPreds.insert(Pred)) 1002 continue; // Duplicate predecessor entry. 1003 1004 // If the predecessor ends with an indirect goto, we can't change its 1005 // destination. 1006 if (isa<IndirectBrInst>(Pred->getTerminator())) 1007 continue; 1008 1009 ConstantInt *Val = PredValues[i].first; 1010 1011 BasicBlock *DestBB; 1012 if (Val == 0) // Undef. 1013 DestBB = 0; 1014 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1015 DestBB = BI->getSuccessor(Val->isZero()); 1016 else { 1017 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 1018 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 1019 } 1020 1021 // If we have exactly one destination, remember it for efficiency below. 1022 if (i == 0) 1023 OnlyDest = DestBB; 1024 else if (OnlyDest != DestBB) 1025 OnlyDest = MultipleDestSentinel; 1026 1027 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1028 } 1029 1030 // If all edges were unthreadable, we fail. 1031 if (PredToDestList.empty()) 1032 return false; 1033 1034 // Determine which is the most common successor. If we have many inputs and 1035 // this block is a switch, we want to start by threading the batch that goes 1036 // to the most popular destination first. If we only know about one 1037 // threadable destination (the common case) we can avoid this. 1038 BasicBlock *MostPopularDest = OnlyDest; 1039 1040 if (MostPopularDest == MultipleDestSentinel) 1041 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1042 1043 // Now that we know what the most popular destination is, factor all 1044 // predecessors that will jump to it into a single predecessor. 1045 SmallVector<BasicBlock*, 16> PredsToFactor; 1046 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1047 if (PredToDestList[i].second == MostPopularDest) { 1048 BasicBlock *Pred = PredToDestList[i].first; 1049 1050 // This predecessor may be a switch or something else that has multiple 1051 // edges to the block. Factor each of these edges by listing them 1052 // according to # occurrences in PredsToFactor. 1053 TerminatorInst *PredTI = Pred->getTerminator(); 1054 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1055 if (PredTI->getSuccessor(i) == BB) 1056 PredsToFactor.push_back(Pred); 1057 } 1058 1059 // If the threadable edges are branching on an undefined value, we get to pick 1060 // the destination that these predecessors should get to. 1061 if (MostPopularDest == 0) 1062 MostPopularDest = BB->getTerminator()-> 1063 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1064 1065 // Ok, try to thread it! 1066 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1067} 1068 1069/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1070/// a PHI node in the current block. See if there are any simplifications we 1071/// can do based on inputs to the phi node. 1072/// 1073bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1074 BasicBlock *BB = PN->getParent(); 1075 1076 // If any of the predecessor blocks end in an unconditional branch, we can 1077 // *duplicate* the conditional branch into that block in order to further 1078 // encourage jump threading and to eliminate cases where we have branch on a 1079 // phi of an icmp (branch on icmp is much better). 1080 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1081 BasicBlock *PredBB = PN->getIncomingBlock(i); 1082 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1083 if (PredBr->isUnconditional() && 1084 // Try to duplicate BB into PredBB. 1085 DuplicateCondBranchOnPHIIntoPred(BB, PredBB)) 1086 return true; 1087 } 1088 1089 return false; 1090} 1091 1092/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1093/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1094/// NewPred using the entries from OldPred (suitably mapped). 1095static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1096 BasicBlock *OldPred, 1097 BasicBlock *NewPred, 1098 DenseMap<Instruction*, Value*> &ValueMap) { 1099 for (BasicBlock::iterator PNI = PHIBB->begin(); 1100 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1101 // Ok, we have a PHI node. Figure out what the incoming value was for the 1102 // DestBlock. 1103 Value *IV = PN->getIncomingValueForBlock(OldPred); 1104 1105 // Remap the value if necessary. 1106 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1107 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1108 if (I != ValueMap.end()) 1109 IV = I->second; 1110 } 1111 1112 PN->addIncoming(IV, NewPred); 1113 } 1114} 1115 1116/// ThreadEdge - We have decided that it is safe and profitable to factor the 1117/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1118/// across BB. Transform the IR to reflect this change. 1119bool JumpThreading::ThreadEdge(BasicBlock *BB, 1120 const SmallVectorImpl<BasicBlock*> &PredBBs, 1121 BasicBlock *SuccBB) { 1122 // If threading to the same block as we come from, we would infinite loop. 1123 if (SuccBB == BB) { 1124 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1125 << "' - would thread to self!\n"); 1126 return false; 1127 } 1128 1129 // If threading this would thread across a loop header, don't thread the edge. 1130 // See the comments above FindLoopHeaders for justifications and caveats. 1131 if (LoopHeaders.count(BB)) { 1132 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1133 << "' to dest BB '" << SuccBB->getName() 1134 << "' - it might create an irreducible loop!\n"); 1135 return false; 1136 } 1137 1138 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1139 if (JumpThreadCost > Threshold) { 1140 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1141 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1142 return false; 1143 } 1144 1145 // And finally, do it! Start by factoring the predecessors is needed. 1146 BasicBlock *PredBB; 1147 if (PredBBs.size() == 1) 1148 PredBB = PredBBs[0]; 1149 else { 1150 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1151 << " common predecessors.\n"); 1152 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1153 ".thr_comm", this); 1154 } 1155 1156 // And finally, do it! 1157 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1158 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1159 << ", across block:\n " 1160 << *BB << "\n"); 1161 1162 // We are going to have to map operands from the original BB block to the new 1163 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1164 // account for entry from PredBB. 1165 DenseMap<Instruction*, Value*> ValueMapping; 1166 1167 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1168 BB->getName()+".thread", 1169 BB->getParent(), BB); 1170 NewBB->moveAfter(PredBB); 1171 1172 BasicBlock::iterator BI = BB->begin(); 1173 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1174 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1175 1176 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1177 // mapping and using it to remap operands in the cloned instructions. 1178 for (; !isa<TerminatorInst>(BI); ++BI) { 1179 Instruction *New = BI->clone(); 1180 New->setName(BI->getName()); 1181 NewBB->getInstList().push_back(New); 1182 ValueMapping[BI] = New; 1183 1184 // Remap operands to patch up intra-block references. 1185 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1186 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1187 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1188 if (I != ValueMapping.end()) 1189 New->setOperand(i, I->second); 1190 } 1191 } 1192 1193 // We didn't copy the terminator from BB over to NewBB, because there is now 1194 // an unconditional jump to SuccBB. Insert the unconditional jump. 1195 BranchInst::Create(SuccBB, NewBB); 1196 1197 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1198 // PHI nodes for NewBB now. 1199 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1200 1201 // If there were values defined in BB that are used outside the block, then we 1202 // now have to update all uses of the value to use either the original value, 1203 // the cloned value, or some PHI derived value. This can require arbitrary 1204 // PHI insertion, of which we are prepared to do, clean these up now. 1205 SSAUpdater SSAUpdate; 1206 SmallVector<Use*, 16> UsesToRename; 1207 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1208 // Scan all uses of this instruction to see if it is used outside of its 1209 // block, and if so, record them in UsesToRename. 1210 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1211 ++UI) { 1212 Instruction *User = cast<Instruction>(*UI); 1213 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1214 if (UserPN->getIncomingBlock(UI) == BB) 1215 continue; 1216 } else if (User->getParent() == BB) 1217 continue; 1218 1219 UsesToRename.push_back(&UI.getUse()); 1220 } 1221 1222 // If there are no uses outside the block, we're done with this instruction. 1223 if (UsesToRename.empty()) 1224 continue; 1225 1226 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1227 1228 // We found a use of I outside of BB. Rename all uses of I that are outside 1229 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1230 // with the two values we know. 1231 SSAUpdate.Initialize(I); 1232 SSAUpdate.AddAvailableValue(BB, I); 1233 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1234 1235 while (!UsesToRename.empty()) 1236 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1237 DEBUG(dbgs() << "\n"); 1238 } 1239 1240 1241 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1242 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1243 // us to simplify any PHI nodes in BB. 1244 TerminatorInst *PredTerm = PredBB->getTerminator(); 1245 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1246 if (PredTerm->getSuccessor(i) == BB) { 1247 RemovePredecessorAndSimplify(BB, PredBB, TD); 1248 PredTerm->setSuccessor(i, NewBB); 1249 } 1250 1251 // At this point, the IR is fully up to date and consistent. Do a quick scan 1252 // over the new instructions and zap any that are constants or dead. This 1253 // frequently happens because of phi translation. 1254 BI = NewBB->begin(); 1255 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) { 1256 Instruction *Inst = BI++; 1257 1258 if (Value *V = SimplifyInstruction(Inst, TD)) { 1259 WeakVH BIHandle(BI); 1260 ReplaceAndSimplifyAllUses(Inst, V, TD); 1261 if (BIHandle == 0) 1262 BI = NewBB->begin(); 1263 continue; 1264 } 1265 1266 RecursivelyDeleteTriviallyDeadInstructions(Inst); 1267 } 1268 1269 // Threaded an edge! 1270 ++NumThreads; 1271 return true; 1272} 1273 1274/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1275/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1276/// If we can duplicate the contents of BB up into PredBB do so now, this 1277/// improves the odds that the branch will be on an analyzable instruction like 1278/// a compare. 1279bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1280 BasicBlock *PredBB) { 1281 // If BB is a loop header, then duplicating this block outside the loop would 1282 // cause us to transform this into an irreducible loop, don't do this. 1283 // See the comments above FindLoopHeaders for justifications and caveats. 1284 if (LoopHeaders.count(BB)) { 1285 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1286 << "' into predecessor block '" << PredBB->getName() 1287 << "' - it might create an irreducible loop!\n"); 1288 return false; 1289 } 1290 1291 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1292 if (DuplicationCost > Threshold) { 1293 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1294 << "' - Cost is too high: " << DuplicationCost << "\n"); 1295 return false; 1296 } 1297 1298 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1299 // of PredBB. 1300 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1301 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1302 << DuplicationCost << " block is:" << *BB << "\n"); 1303 1304 // We are going to have to map operands from the original BB block into the 1305 // PredBB block. Evaluate PHI nodes in BB. 1306 DenseMap<Instruction*, Value*> ValueMapping; 1307 1308 BasicBlock::iterator BI = BB->begin(); 1309 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1310 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1311 1312 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1313 1314 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1315 // mapping and using it to remap operands in the cloned instructions. 1316 for (; BI != BB->end(); ++BI) { 1317 Instruction *New = BI->clone(); 1318 New->setName(BI->getName()); 1319 PredBB->getInstList().insert(OldPredBranch, New); 1320 ValueMapping[BI] = New; 1321 1322 // Remap operands to patch up intra-block references. 1323 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1324 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1325 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1326 if (I != ValueMapping.end()) 1327 New->setOperand(i, I->second); 1328 } 1329 } 1330 1331 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1332 // add entries to the PHI nodes for branch from PredBB now. 1333 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1334 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1335 ValueMapping); 1336 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1337 ValueMapping); 1338 1339 // If there were values defined in BB that are used outside the block, then we 1340 // now have to update all uses of the value to use either the original value, 1341 // the cloned value, or some PHI derived value. This can require arbitrary 1342 // PHI insertion, of which we are prepared to do, clean these up now. 1343 SSAUpdater SSAUpdate; 1344 SmallVector<Use*, 16> UsesToRename; 1345 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1346 // Scan all uses of this instruction to see if it is used outside of its 1347 // block, and if so, record them in UsesToRename. 1348 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1349 ++UI) { 1350 Instruction *User = cast<Instruction>(*UI); 1351 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1352 if (UserPN->getIncomingBlock(UI) == BB) 1353 continue; 1354 } else if (User->getParent() == BB) 1355 continue; 1356 1357 UsesToRename.push_back(&UI.getUse()); 1358 } 1359 1360 // If there are no uses outside the block, we're done with this instruction. 1361 if (UsesToRename.empty()) 1362 continue; 1363 1364 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1365 1366 // We found a use of I outside of BB. Rename all uses of I that are outside 1367 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1368 // with the two values we know. 1369 SSAUpdate.Initialize(I); 1370 SSAUpdate.AddAvailableValue(BB, I); 1371 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1372 1373 while (!UsesToRename.empty()) 1374 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1375 DEBUG(dbgs() << "\n"); 1376 } 1377 1378 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1379 // that we nuked. 1380 RemovePredecessorAndSimplify(BB, PredBB, TD); 1381 1382 // Remove the unconditional branch at the end of the PredBB block. 1383 OldPredBranch->eraseFromParent(); 1384 1385 ++NumDupes; 1386 return true; 1387} 1388 1389 1390